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.pcbi.1002733 | The Laminar Cortex Model: A New Continuum Cortex Model Incorporating Laminar Architecture | Local field potentials (LFPs) are widely used to study the function of local networks in the brain. They are also closely correlated with the blood-oxygen-level-dependent signal, the predominant contrast mechanism in functional magnetic resonance imaging. We developed a new laminar cortex model (LCM) to simulate the amplitude and frequency of LFPs. Our model combines the laminar architecture of the cerebral cortex and multiple continuum models to simulate the collective activity of cortical neurons. The five cortical layers (layer I, II/III, IV, V, and VI) are simulated as separate continuum models between which there are synaptic connections. The LCM was used to simulate the dynamics of the visual cortex under different conditions of visual stimulation. LFPs are reported for two kinds of visual stimulation: general visual stimulation and intermittent light stimulation. The power spectra of LFPs were calculated and compared with existing empirical data. The LCM was able to produce spontaneous LFPs exhibiting frequency-inverse (1/ƒ) power spectrum behaviour. Laminar profiles of current source density showed similarities to experimental data. General stimulation enhanced the oscillation of LFPs corresponding to gamma frequencies. During simulated intermittent light stimulation, the LCM captured the fundamental as well as high order harmonics as previously reported. The power spectrum expected with a reduction in layer IV neurons, often observed with focal cortical dysplasias associated with epilepsy was also simulated.
| Local field potentials (LFPs) are low-frequency fluctuations of the electric fields produced by the brain. They have been widely studied to understand brain function and activity. LFPs reflect the activity of neurons within a few square millimeters of the cerebral cortex, an area containing more than 10,000 neurons. To avoid the complexity of simulating such a large number of individual neurons, the continuum cortex model was devised to simulate the collective activity of groups of neurons generating cortical LFPs. However, the continuum cortex model assumes that the cortex is two-dimensional and does not take into account the laminar architecture of the cerebral cortex. We developed a three-dimensional laminar cortex model (LCM) by combining laminar architecture with the continuum cortex model. This expansion enables the LCM to simulate the detailed three-dimensional distribution of the LFP within the cortex. We used the LCM to simulate LFPs within the visual cortex under different conditions of visual stimulation. The LCM reproduced the key features of LFPs observed in electrophysiological experiments. We conclude that the LCM is a potentially useful tool to investigate the underlying mechanism of LFPs.
| Neuronal activity changes the distribution of electric potentials in the brain [1], [2]. Local field potentials (LFPs) are the low-frequency (<100 Hz) fluctuations in electric potentials in the extracellular space of the brain [2], [3]. They represent a weighted average of the potential changes produced by neuronal activity in a small volume around the measuring electrode [4], [5]. Concurrent electrophysiological and functional MRI experiments have also demonstrated that LFPs are correlated with signal change in functional magnetic resonance imaging, a method of detecting neuronal activity through changes in blood-oxygen-level-dependent signal [6]–[8]. Previous electrophysiological experiments investigating the neuronal processes underlying LFPs have measured the membrane potential of neurons and extracellular field potentials simultaneously [9], [10]. A major difficulty with this paradigm is that LFPs reflect the activity of more than 10,000 neurons [11] within 250 micrometres of the recording electrode [4], [5]. Simultaneous measurement of such a large number of neuron activities has not been achieved to date. Furthermore, multiple concurrent processes contribute to LFPs, including action potentials, synaptic transmission, glial activity, and even extracellular space diffusion [12] and are difficult to disambiguate.
Computer simulations have widely been adopted to predict changes in neuronal activity associated with corresponding LFPs. A previous study simulated the membrane potential changes of a large number of individual neurons as means of reconstructing the LFP [for example, see 13]. Simulating the dynamics of a large number of neurons faces the challenge of specifying the physiological parameters in large, inhomogeneous populations with diverse physiological properties [14]. An alternative approach to simulating individual neuronal activity has been to simulate the activity in an ensemble of neurons. An example of this is the continuum cortex model, developed by Wright et al [15]–[18], which has been used to simulate ensemble activity at different scales [17]. Existing continuum cortex models do not take into account the laminar architecture of the cerebral cortex. They are, therefore, limited in their ability to model the distribution of electric potential of the brain in three dimensions. Cortical neurons are organized in columns comprising as many as 20,000 neurons [19], [20]. Functionally, neurons in a column display similar responses to specific stimuli [21]. In this paper, we build on this notion to expand the continuum cortex model by incorporating the laminar connection architecture of the cortex and simulating the collective of neuronal ensembles within cortical columns. We have used the new laminar cortex model (LCM) to simulate LFPs within the visual cortex under different conditions of visual stimulation.
We give a brief overview of the continuum cortex model for completeness, but for specific details refer to [17]. The continuum cortex model simulates the collective electrophysiological activities of the cerebral cortex. A population approximation is used to overcome the difficulty of simulating a large number of individual neurons, and to capture the essential aspects of cortical dynamics [15], [16]. The continuum cortex model divides the simulated cortical area into a grid of elements, where is an integer. Each element consists of two populations of neurons: excitatory and inhibitory [17]. Each population is treated as a single entity capable of receiving spikes, changing membrane potential, and generating and propagating spikes [17].
The numbers of spikes propagating between neurons of two groups at any one time varies. In the continuum cortex model the effects of action potential shape and its temporal evolution are ignored. Instead, the average afferent spike rate () is used to measure interaction between the two groups of neurons. The spike rate is defined as the average number of spikes a neuron of one group receives from a neuron of the other group per unit time.
The continuum cortex model contains four main components: 1) spike generation, 2) spike propagation, 3) generation of the postsynaptic potential, and 4) membrane potential aggregation. The equations describing each component are provided in Text S1 and were developed either by using theoretical approaches or by experimentally fitting observed data using an appropriate function. The mean field approximation was employed during this procedure [17].
The LCM exploits the laminar architecture of the cortex. Five cortical layers (layer I to VI) are considered (cortical layers II and III are combined). Each layer is simulated using the continuum cortex model, and the layers are connected by laminar synaptic connections (see Figure 1). A synaptic connection map is created and used to control the connection between and within cortical layers (see Table S1 in Text S2). This connection map was based on empirical observations of the number of synapses formed between different types of neurons by Binzegger [22] (see Text S2). The connection map classifies the afferent synapses on each group of cortical neurons into three categories: 1) intracortical synapses, from within the visual cortex (), 2) cortico-cortical synapses, from other cortical areas (), and 3) thalamic synapses, projections from neurons in the lateral geniculate nucleus (LGN, ).
The LCM allows simulation of centimeter and column scale (micrometer) cortical regions [17]. Since the grid elements of the centimeter scale model correspond to the size of cortical columns, the connections between cortical laminae are assumed to be local. This means that elements in the same horizontal position of all cortical layers are connected vertically (see Figure 1B). In contrast, the column scale implementation is approximately the size of one cortical column. Therefore, connections between cortical layers are global, and the average spike rate of a cortical layer is the input to other cortical layers. The work here is focused on simulating LFPs produced in the visual cortex. Hence, results are limited to the application of the centimeter scale model.
We simulated the effect of visual stimulation on LFPs using the LCM. Different forms of visual stimulation were assumed to form different spike trains projecting from the LGN to deeper cortical layers of the visual cortex (Layer IV, V and VI, see Table S1 in Text S2). Three states of visual stimulation were examined in the model: 1) spontaneous activity without visual stimulation, 2) constant visual stimulation, and 3) intermittent light stimulation. As illustrated in Figure 2, these conditions correspond to afferent spike trains with the shape of small amplitude white noise, large amplitude white noise (the random number generator from [23] was adopted), and recurring Gaussian peaks, respectively.
Apart from the synapses projecting from neurons in LGN and the visual cortex, there are also a large number of synapses originating from other cortical areas (see Table S1 in Text S2). We assume that spikes from these synapses contribute to background noise, which was modeled as low-amplitude white noise.
The LCM has over 150 parameters, which fall into four categories relating to: 1) electrophysiological properties of neurons, 2) spike propagation, 3) synaptic transmission, and 4) connections between cortical laminae. Most of these parameters were estimated from experimental data, while others were left as free parameters. However, the cortex is complex, to the extent that our simplified parameters may not represent its physiology, morphology and architecture exactly. We found that a small deviation of the parameter values do not change the results reported here significantly. This is because a similar LFP outcome can be achieved by tuning free parameters.
Parameters relating to the electrophysiological properties of neurons are well established in the literature. We used the same values, derived from experimental data, as the continuum cortex model [17].
Spike propagation parameters and their values used are listed in Table 1. The propagation speed of spikes in the horizontal (lateral) direction () was set to 0.24 m/s, which is consistent with experimental measurements of the speed of spread of spikes in the cortex [24], [25]. Since collateral branches are usually smaller in diameter than the main axon, the speed of vertical (inter-laminar) propagation of spikes () was set to 1.2 times the speed of horizontal propagation. The spike propagation range parameters were set to the similar values as continuum cortex model [17].
There is a wide range of published values for synaptic transmission parameters [26], [27]. We chose the middle parameter value when a range was provided and the average when multiple values were reported. The excitatory and inhibitory synaptic gains and , were treated as free parameters. Their values were determined by fitting experimental data to the LFPs generated using the LCM.
The best set of parameter values was selected as those fulfilling the following criteria: 1) the LFP power spectrum fitted the function with [28]. 2) with simulated visual stimulation, there was an increase in gamma frequency in the power spectrum; 3) membrane potentials of neuron groups were less than 10 mV above their resting membrane potentials [29].
The simulation program was written using the ANSI C language and compiled with the Intel C compiler (http://software.intel.com/intel-compilers/). The program was compiled and executed on a Linux workstation (Dell® Precision T7500) with Ubuntu version10.10 (×86_64, http://www.ubuntu.com). OpenMP (http://www.openmp.org), a shared-memory parallel programming library, was used to parallelize the code to speed up program execution.
In this paper, the LCM was used to simulate a cortical area of size cm2. The domain was discretized to a grid. At the beginning of each execution of the program, the simulation time was initialized to zero, and every neuron state variable was set to its resting state value (see Text S2). The iteration time step was one millisecond. After initialization, the program executed without particular visual stimulation for 60 seconds at which time the system is assumed to have reached steady state. Constant visual stimulation or intermittent light stimulation was then applied for 20 seconds (time = 60–80 sec). LFPs were simulated for conditions of spontaneous activity and for each mode of visual stimulation.
In the simulation, the membrane potentials of all neuron groups in the middle element of a layer are recorded during the entire execution. Data of the last 1.024 second prior to visual stimulation and after stimulation were used for frequency spectrum analysis.
For comparisons with experimental data, the LFPs of the simulated cortical area are assumed to be the average of neuronal membrane potentials of the central elements of all layers, stated as:(1)where are the numbers of neurons in the central element of layer and are the potentials of the central elements of layer , which is the average of membrane potentials of neurons in the element, that is(2)where , are the numbers of excitatory and inhibitory neurons and and are the (average) membrane potentials of excitatory and inhibitory neuron populations respectively. The frequency spectrum of the LFPs was computed using the fast Fourier transform as implemented in MATLAB 2010a (http://www.mathworks.com). The LFP frequency power spectra were compared with experimentally measured data.
LFPs produced by LCM were also used to estimate current source density. The standard one-dimension current source density calculation method was used [30], [31](3)Here is electric conductance of the cortex, and was set to 0.3 S/m, is the potential at the point, and is the distance between two adjacent points. To reduce spatial noise, the three-point Hamming filter was applied [32], [33](4)
We examined the behaviour of the LCM using different parameter values. For each parameter combination, around 100 executions of the LCM were conducted, and the average LFP frequency spectrum was computed.
Figure 3 shows the power spectra of the LFPs obtained with different synaptic gains. The LCM was able to generate LFPs with different types and envelopes of oscillation, depending on the combination of excitatory and inhibitory synaptic gains used in the simulation. For example, when either excitatory or inhibitory synaptic gain was small, the frequency spectrum of background activity had an inverse-frequency shape. Stimulation resulted in an increase in gamma frequency. In contrast, when the excitatory and inhibitory synaptic gains were both large, particular frequency peaks dominated the LFP power spectra. Thus, variations of synaptic gains had a strong impact on LFP frequencies.
For large synaptic gains, the peaks in the power spectra did not change position with variation in synaptic gain. Dependence of peak position on other parameters was also examined by generating LFP power spectra with different parameter values. The time course of the postsynaptic potential (PSP) was found to be strongly correlated with the positions of the peaks. Peak frequency decreased with increasing PSP time course. (Four examples of LFP power spectra with different PSP time courses are shown in Figure S2). This suggests that the dominant oscillation frequency is controlled by the feedback between excitatory and inhibitory neurons.
The shape of the power spectrum of LFPs generated by the LCM is controlled by the balance between excitatory and inhibitory postsynaptic potentials (PSPs). These are influenced by many parameters simultaneously, including synaptic gains, spike propagation ranges and synapse numbers. Changes in PSPs caused by variation of one parameter could be compensated by other parameters. For example, increase of synaptic gains may not change PSP when the corresponding synapse number is decreased. Therefore, the LCM could produce similar LFPs using different combinations of parameter values.
Experimental models of neocortical epileptic foci suggest that reduced synaptic inhibition in layer IV plays an important role in epileptogenesis [34], [35]. Focal cortical dysplasias characterized by an absence or significant reduction in layer IV are also very frequently associated with epilepsy [36]. Figure 4 shows the LFP power spectrum shapes generated by the LCM when the numbers of synapses formed with presynaptic neurons in layer IV are decreased by 50%. Compared to Figure 3A, the power spectra show a small shift to small inhibitory gain. For example, for LFPs produced using excitatory and inhibitory synaptic gains of V/spike, the power spectrum changed from a frequency-inverse shape to one with spectral peaks as would be expected with seizures when presynaptic neurons of layer IV decrease by 50%. This suggests that, changes in neuron or synapse density may change the way LFPs oscillate dramatically. These alterations in dynamics may increase our understanding of how abnormalities in cortical architecture lead to seizures.
Figure 5 shows the time courses of membrane potentials in a single run of the LCM. We found that in every cortical layer, membrane potentials oscillated with amplitudes of 0.05–0.2 mV; the amplitudes are much larger in layers IV and VI (around 0.1 mV) than in other layers (around 0.05 mV). During stimulation, the membrane potentials and its oscillation amplitudes increased in all layers except layer I. The power spectra in all layers, as provided in Figure 5, all showed inverse-square decreasing frequency background activities, which is observed experimentally [37]. Stimulation also increased high-frequency membrane potential oscillation of all deep layers.
The laminar distribution of the LFP power spectrum amplitude was examined. Figure 5C shows the laminar distribution of the average of the LFP power distribution in the gamma frequency (30–100 Hz) and sub-gamma frequency (5–20 Hz) ranges for spontaneous activity and general stimulation. Higher frequency powers were observed in layers IV and VI. This is in agreement with experimentally measured laminar LFP amplitude profiles in the primary visual cortex [38]. Since layers IV and VI are the main layers of the visual cortex receiving and sending projections to the LGN, the observed variation in LFP power spectra amplitudes between layers most likely results from these projections. We simulated the propagation of one spike source in the cortex using LCM. In Figure 6 we provide the result when a spike source is placed in the four central elements of layer IV for 20 milliseconds after 60 seconds of spontaneous activity. Following spike onset, a strong potential is observed in the center of all cortical layers except layer I. The potential is decreased in elements surrounding the source, simulating surround inhibition. We display the temporal profiles of current source density along a transverse line through the central point in layer IV and for the central elements of each cortical area in Figure 7.
Many electrophysiological experiments have demonstrated that with intermittent light stimulation, neuronal activity in the visual cortex synchronizes with stimulus frequency [39]–[42]. Furthermore, EEG responses are enhanced at this frequency (fundamental harmonics), as well as at half the stimulus frequency (first sub-harmonic), and at multiples of the stimulus frequency (multiple harmonics). The responses to visual stimulation at specific frequencies, termed steady-state visual evoked potentials (SSVEPs), can be observed on both scalp EEG recordings [39] and invasive recordings of LFPs [40]. We used SSVEPs to examine the effect of cortical architecture on LFPs.
The LCM was used to simulate LFPs with 10 Hz intermittent light stimulation represented by a Gaussian distribution of spike rates for neurons projecting from the LGN to the visual cortex. The peak and standard deviation of the Gaussian shape was 30 spikes/second and 6.25 milliseconds, respectively (see Figure 2). Figure 8 shows the variation of LFPs with time and the associated power spectra. Simulations using the LCM reproduced the power spectra reported in experimental data [39]. The LFP power spectrum had peaks at frequencies that were multiples of the stimulus frequency (i.e. capturing multiple harmonics). Notably, the amplitude of fundamental harmonic (i.e. frequency peak at 10 Hz) was smaller in layer II/III than other layers. This is probably because there are fewer projections from LGN to layer II/III than other layers. Experimentally observed sub-harmonics were not obvious in simulations using the LCM [39].
This paper introduces the LCM and describes its use to simulate LFPs in the primary visual cortex. The LCM has the advantage that it incorporates the architecture of the visual cortex allowing the simulation of LFPs with high spatial and temporal resolution. We were able to simulate the membrane potential in each cortical layer, as well as its temporal variations. We used the LCM to investigate the relationship between visual stimulation and LFPs. We validated the model using two different experimental simulations: constant visual stimulation and intermittent light stimulation. Our results were comparable to relevant experimental measurements. We also simulated the effects of changes in neuronal density in layer IV, often observed in epileptic cortical dysplastic tissue. For certain parameter combinations the changes in the power spectra were those expected in seizures. CSD maps showed comparable features to experimental data and intralaminar CSD profiles following transient LGN input had the appearance of surround inhibition.
With constant visual stimulation, the LCM produced LFPs oscillating in two different ways determined by the combination of parameters used in the simulation. When the cortex was activated with low levels of background noise and stimulus input (small synaptic gains), the LFP oscillation was governed by the pool of excitatory neurons. Synaptic transmission acts as a filter due to the convolution in the membrane potential aggregation function of LCM (refer to Equation S1.8 in Text S1). Effectively, this dampens high frequency oscillations and results in an inverse-squared decreasing LFP spectrum. However, when the cortex is highly activated, inhibitory neurons play a more dominant role, resulting in oscillations in which initial activation of inhibitory neurons leads to suppression of the membrane potential of all neurons, including the inhibitory pool followed by a burst of activity cause by excitatory input.
The LFPs produced using low synaptic gains are comparable to experimentally observed LFPs in the normal brain, while the LFPs obtained with large synaptic gains are similar to those measured during seizures [37]. This suggests that changes in neuronal physiology can result in a change in the LFP power spectrum and may help to explain frequency changes in the EEG observed in certain neurological disorders. There are some differences between LFPs from the LCM and experimentally measured LFPs. The amplitude of low frequency (<10 Hz) LFPs produced by the model is lower than measured experimentally. A possible explanation is that the low frequency oscillation results from feedback loops between the visual cortex and other brain areas [43], which are not considered in the LCM. The gamma frequency (40–200 Hz) power of stimulated LFPs is also smaller than experimental measurements. We postulate that this is because extracellular potential changes caused by synaptic activities and spike conduction are not included in the calculation of LFPs. These are reported to have a greater influence on high frequency LFPs [44]–[46]. The LCM simplifies synaptic processes and spike propagation to a signal delivery level. It does not simulate the burst of synaptic transmission and spikes.
The CSDs calculated from LCM recreates several features from experimental observations [47]. Within layers, the CSD profile simulated surround inhibition [48]. Across cortical layers, the temporal profile of CSDs was similar to those observed by Schroeder et al. [47] with transition from sink to source following stimulation.
We used SSVEPs, to test the effects of incorporating cortical architecture on simulation output. In our intermittent light stimulation study, we used the LCM to reproduce the behaviour of SSVEPs. The fundamental and high order harmonics were apparent in the visual cortex. The first sub-harmonics, shown to be present empirically [39], may be brought about by feedback loops between the primary visual cortex and other visual cortical areas. These connections are not included in the LCM.
The LCM may be used to simulate abnormal responses to intermittent light stimulation such as the photoparoxysmal response observed in forms of genetic generalized epilepsy. This can be achieved by varying LCM parameters, and comparing the simulation output with measured EEG data. This has the potential to generate testable hypotheses relating to underlying neurophysiological mechanisms.
Although we showed that LCM is able to reproduce some of the results of electrophysiological experiments, it has some limitations. Firstly, only two populations of neurons (excitatory and inhibitory) are considered. The behaviour of excitatory neurons may not be best captured by a single category. For example, fast-spiking neurons generate spikes differently from other excitatory neurons [27]. In future work we will extend the LCM to include multiple categories of excitatory neurons. Secondly, simulation of neurotransmission in the LCM may be oversimplified. For example, in its current form it cannot simulate the effects of activating fast (AMPA) and slow (NMDA) excitatory glutamatergic receptors on LFPs. Thirdly, the physiological parameters used in our simulation were obtained from the results of experiments conducted in different species. In our simulations, LFPs were calculated as the aggregate membrane potential dynamics of populations of neurons, an approach commonly employed in simulation studies e.g. [49]. This approach may be inaccurate because it does not take into account the filtering properties of the neural membrane [44], [45]. Methods based, for example, upon summation of conductance of synapses to pyramidal neurons [2], [45], [50] are inapplicable to the LCM, which simulates the collective activity of neuron groups. A future hybrid model is required to link continuum cortical models and models based on simulating the properties of individual neurons.
|
10.1371/journal.pntd.0001903 | Projecting the Long-Term Impact of School- or Community-Based Mass-Treatment Interventions for Control of Schistosoma Infection | Schistosomiasis remains a significant health burden in many areas of the world. Morbidity control, focused on limiting infection intensity through periodic delivery of anti-schistosomal medicines, is the thrust of current World Health Organization guidelines (2006) for reduction of Schistosoma-related disease. A new appreciation of the lifetime impact of repeated Schistosoma infection has directed attention toward strategies for greater suppression of parasite infection per se, with the goal of transmission interruption. Variations in drug schedules involving increased population coverage and/or treatment frequency are now undergoing field trials. However, their relative effectiveness in long-term infection suppression is presently unknown.
Our study used available field data to calibrate advanced network models of village-level Schistosoma transmission to project outcomes of six different community- or school age-based programs, as compared to the impact of current 2006 W.H.O. recommended control strategies. We then scored the number of years each of 10 typical villages would remain below 10% infection prevalence (a practicable level associated with minimal prevalence of disease). All strategies that included four annual treatments effectively reduced community prevalence to less than 10%, while programs having yearly gaps (‘holidays’) failed to reach this objective in half of the communities. Effective post-program suppression of infection prevalence persisted in half of the 10 villages for 7–10 years, whereas in five high-risk villages, program effects on prevalence lasted zero to four years only.
At typical levels of treatment adherence (60 to 70%), current WHO recommendations will likely not achieve effective suppression of Schistosoma prevalence unless implemented for ≥6 years. Following more aggressive 4 year annual intervention, some communities may be able to continue without further intervention for 8–10 years, while in higher-risk communities, annual treatment may prove necessary until eco-social factors fostering transmission are removed. Effective ongoing surveillance and locally targeted annual intervention must then become their mainstays of control.
| Debate persists about how best to prevent disease caused by Schistosoma parasites. Current guidelines focus on suppressing morbidity by limiting average intensity of infection during childhood. However, this may not be sufficient to cure infection or prevent reinfection, leaving risk for sub-clinical morbidities such as growth stunting and anemia. More intensive programs involving broader coverage and/or more frequent treatments could potentially cure most infections and even prevent their return. Because such programs' effectiveness is not currently known, we used computer simulation (grounded by past treatment program results) to project short- and long-term impact in communities where Schistosoma are common. We estimated that 4 annual treatments (delivered community-wide or only to school-age children and high-risk adults) could effectively reduce local prevalence below 10%. Programs with gap years were less effective, particularly in high-risk communities. If a program was successful, infection stayed <10% for 7–10 years in low risk communities. However, rapid resurgence (within 1–5 years) was projected for high risk villages. We conclude that, given the networked transmission of Schistosoma, annual treatment programs of sufficient duration can have secondary benefits, i.e., long-term suppression in some areas. However, high risk areas will need continuing surveillance and frequent retreatment to truly minimize their risk for disease.
| Schistosomiasis is an environmentally transmitted parasitic disease that results in increased morbidity and mortality among millions of people living in tropical and subtropical regions [1]–[3]. Different control initiatives have had success in reducing either the prevalence or the mean intensity of Schistosoma infections in a number of individual countries [4]–[9] although, for other settings, the potential effectiveness of such strategies remains mostly unknown. With combined approaches of snail control, chemotherapy, health education, and hygienic improvement, China has made substantial progress against S. japonicum, which has been successfully eliminated in several provinces [10]–[13]. In Brazil, mass chemotherapy programs have achieved at least a 50% to 70% decrease in S. mansoni prevalence, with a marked reduction in infection-related morbidity and hospitalization in many states [14]. However, the effects of those initiatives have been uneven because of the regional- and country-specific differences in local transmission, in available resources, and in overall program commitment [12], [15].
In the meantime, schistosomiasis remains a major public health problem in sub-Saharan Africa [3], [8]. Possible measures for controlling African schistosomiasis include chemotherapy (primarily praziquantel therapy), snail control (by mollusciciding or habitat modification), provision of safe water alternatives, and/or education with behavior modification to limit exposure [4]. Among these, chemotherapy-based mass treatment remains the most widely used at present [8]. ‘Morbidity control’, focused on limiting infection intensity through periodic delivery of anti-schistosomal medicines, is the thrust of the World Health Organization (WHO) schistosomiasis control guidelines [16]. However, a new appreciation of the lifetime impact of repeated Schistosoma infections [17], [18] has directed attention toward strategies for greater suppression of parasite infection per se, with the ultimate goal of transmission interruption as a more effective elimination of disease risk. However, experience in schistosomiasis control programs suggests that current broad-based drug delivery campaigns may not consistently reduce local prevalence or prevent rapid reinfection [19], [20].
Randomized field trials are now underway to examine the relative impact of increased population coverage and/or treatment frequency in community-wide or school age-targeted treatment programs for Schistosoma control [21]. However, their effectiveness relative to existing WHO protocols is not yet known. In the present study, we used a computer simulation based on a recently developed, stratified worm burden (SWB) model [22] of multi-village Schistosoma transmission to project programmatic success for each of seven different mass treatment control strategies in a typical Schistosoma-endemic region of sub-Saharan Africa.
In Gurarie, et al. 2010 [22], we developed a new approach to predicting schistosomiasis transmission and community-level infectious burden using models based on a Stratified Worm Burden (SWB) formulation, combined with calibration using data from a completed multi-year pilot schistosomiasis control program. In this approach, a host population is subdivided into burden strata in terms of individual parasite load, with the ith stratum carrying (i−1) Δw to iΔw worms, where the size of the stratum, iΔw, is prescribed e.g., at increments of an additional 10 worm pairs per person. The transitions among strata depend on (i) force of infection (determined by the transmission environment and human risk factors); (ii) worm attrition (natural or drug-induced); and (iii) human demographics (birth, aging, mortality, growth sources). Details are presented in Supplement S1 and [22].
The strength of this SWB model analysis is that it more accurately reflects many of the underlying complexities of Schistosoma species transmission, while yielding a more usable projection of community infection prevalence (rather than mean worm burden) after a campaign. The SWB includes and tracks differences in infection risk between children and adults, the highly skewed distribution of infection intensity among infected individuals, and the networked meta-population aspects of transmission among neighboring villages sharing common water sources [23]. This is an advantage where decision-making policy is made on the basis of local population or school age prevalence of Schistosoma infection [16] After the model was calibrated with available field data from pilot control programs [24], [25], it was used to project the likely outcomes of different targeted- or population-wide mass treatment programs now being introduced into endemic areas in sub-Saharan Africa [21], [26]. The model can realistically account for inevitable program limitations in program resources, delivery efficiency, etc., whereby drug treatment cannot cover every infected human in the treatment area and the frequency of the treatment cannot be too high. Non-stationary and nonlinear models of this type are difficult or impossible to solve through classical equation analysis. However, their performance and outputs can be estimated for different chemotherapy scenarios through numerical analysis [22] in the programmed simulations presented here.
Taking burden strata as model variables instead of the population mean worm burden, SWB carries detailed information of the infection status of a host population and reflects its highly aggregated distribution (typical of Schistosoma and other helminth infections [27]), without imposing a priori assumptions about that distribution. Statistical moments (variance and/or aggregation) of such a distribution can include human population dynamics and, in its most simplified format, even reflect the predictions of the classical Mean Worm Burden (MWB) equations of MacDonald [28]. Following the SWB approach, we were able to develop calibration procedures [22] that could readily take advantage of available prevalence and intensity data from published treatment trials [24], [25]. The calibration methodology was, first, to estimate local force of infection in each village through fitting of SWB equilibrium equation parameters to prevalence data, then using reduced MWB and snail prevalence systems to estimate transmission parameters and the net effects of other unmeasured biotic and abiotic variables associated with transmission at local water sites [22]. These local factors include, in aggregate, water quality, rainfall, vegetation, seasonal water level persistence, human contact and sewage contamination. This model development allowed prediction of treatment outcomes across distributed village systems (Figure 1), including the significant effects of having multiple human age-strata sharing multiple water contact sites.
For the present study, our base-case study system consisted of 10 neighboring villages and 5 shared water (snail) sites, where the village level infection exposures are linked through the shared water sites (see Figure 1 for schematic representation of the transmission network). The calibration is taken from a well-studied Schistosoma haematobium-endemic region of the Msambweni District in coastal Kenya [24], [25], [29]–[31]. The 10 villages used to calibrate this study are located 50 km southwest of Mombasa, Kenya in an agricultural region that produces rice, coconuts, sugar cane, cassava, and maize. The area has a monsoon-type climate, with the period of January to April–May being hot and dry. The mean monthly temperature varies from 26.3°C to 26.6°C, with lows of 23.5°C (July) to 26.9°C (March–April) and highs from 26.7°C (August) to 36.3°C (February). Annual precipitation varies from 900 mm to 1,500 mm, with large yearly and monthly fluctuations. There are usually two rainy seasons, the long rains starting around March and continuing until July, and the short rains beginning in October and lasting through November. In 1984, the total population of this area was determined by household census to be 8,957. It was 16,790 by repeat census in 2002–3. Surveys of water use in the villages indicates that exposure to cercariae-infected waters comes through the practice of using dammed streams and seasonal ponds for washing and bathing, despite the availability of piped water (from kiosks) and borehole wells in most villages. No other schistosomiasis control programs were implemented in the study area during the course of these studies; routine chemotherapy dispensed by the local hospital was monitored, and there was no interval change in the level of such treatment.
As is typical for schistosomiasis [32], pre-treatment levels of Schistosoma infection school age prevalence varied significantly (23% to 74%) across the landscape, even within a span of 5 km [33]. Data available for model calibration included: (i) demographic data, i.e., human population numbers in different villages divided into child and adult age groups and, in addition, snail population densities at georeferenced exposure sites [34], [35]; (ii) infection, in terms of individual and mean egg counts for children in each village, and also the density of susceptible, infected, and shedding adult snails in all monitored water sites [34], [35], [36]; (iii) behavioral data on water contact exposures for each village population among their adjacent water contact/snail-infected sites [24], [31]. Full human and snail data were collected in 1983–1987 and again in 2000–2003 [24], [31], [33]–[35]. Human prevalence data for 2 villages were collected in 2006 and 2009 [37]. In all human surveys infection prevalence and mean egg-counts were based on standard 10 mL filtration of two urine samples [38] collected from the resident populations surveyed. The transmission coefficients from snail sites to villages and from villages to snail sites and the infection (egg excretion) rates for children and adults were estimated based on model fitting to these data. The calibrated parameters in [22] consisted of per capita transmission rates A (snail-to-human) and B (human-to-snail), partitioned among different human villages i = {1,2,…,10}, snail sites j = {1,…,4}, and demographic groups (“children”, adults”). All other inputs (demographic, environmental) are then entered into the model as additional factors. For instance, having human population (“children”+“adult”) Hi = Hic+Hia at village “i”, that carry MWB {wic, wia}, and snail number/infection prevalence {Nj, yj} at site “j”, we estimate the forces of infection between “i” and “j” asin terms of per-capita rates {Aijc/a}, {Bjic/a}. The overall force of human infection at each site λ = λc/a, along with “young/adult” demographics, will determine the SWB distribution for each group through equilibrium relations (see Fig. 2 of [22]).
Our basic assumption is that per-capita rates remain unchanged in time. So having calibrated model with the 1983–1987, and 2000–2006 data, we can apply it then for any other period, changing demographics and starting infection levels ad hoc for a given time period. Because local environment remains stable, where transmission decreases it is due to the reduced number of infected humans. For the current analysis, we have taken 2009 data for demographics and human infection [37] as the starting point (initial model inputs) and have numerically simulated the effect of different treatment strategies over long (multi-year) periods, taking into account projected population changes (growth parameters {Hic/a, Nj} - as prescribed functions of time), grounded on most recent human census and snail recovery trends.
The effect of drug treatment on the prevalence of Schistosoma infection was implemented in the model by instantaneously shifting humans at higher burden levels (stratum) to lower levels according to the established efficacy of praziquantel [25], [39]. As an example, within the model, a treatment session with an efficacy of 90% (i.e., killing 90% of worms) would bring a person in the 10th stratum, who is carrying between 90 and 100 worm pairs before therapy, down to the 0th strata (carrying 0 to 10 worms), as the remaining number of worm pairs is expected to be between 9 and 10 of worms after the treatment. Because egg output at this level of infection is, in practice, undetectable [40], all persons in the 0th stratum are considered ‘uninfected’, at least in terms of their contribution to transmission risk. In each year, community percent prevalence is estimated as [1−(fraction in the 0th stratum)]×100. This is a more realistic effect than modeled in the past, in which a treatment term has typically been added to the natural mortality of worms to indicate an extra loss. In fact, the drug-killing effect on worms occurs more quickly (in less than one month) than natural death (∼4–5 years [27]), and this more flexible scheme allows better prediction for the various options in timing and coverage among suggested control strategies [16], [21]. In our model's simulations we assumed at-random treatment coverage, in the sense that a random subset of each population was treated each time [41], [42]. In this sense, the ‘coverage’ of a control program means the fraction of all people treated in each treatment session, regardless of their past participation. Annual population growth for study villages was estimated to be 1.5%, based on the smoothed averages for household census surveys taken in 1984, 1987, 2000, 2003, and 2006.
Currently, the ‘standard-of-care’ for population-based drug treatment of Schistosoma haematobium and S. mansoni is based on the latest 2006 WHO guidelines [16]. Under these guidelines, populations are divided into 3 classes according to local prevalence of infection among children. These are: high-risk communities (>50% children infected), moderate-risk communities (10–49% infected), and low-risk communities (<10% prevalence), respectively (Table 1). For highly infected population, it is recommended that children between 5 and 15 year old, as well as adults, be treated annually; for moderately infected populations, children between 5 and 15 yr of age and high-risk adults are treated on an every 2 year schedule; for populations with low prevalence of infections, it is recommended that school children be treated twice, upon their entry into school and at school completion (Table 2).
Since 2010, new large-scale operational research trials have been initiated in seven schistosomiasis-endemic locations in sub-Saharan Africa, situated in Cote d'Ivoire, Niger (2 sites), Kenya (2 sites), Tanzania, and Mozambique [21]. These studies seek to identify more cost-effective approaches to regional and national schistosomiasis control, within the present-day framework of mass treatment campaigns for helminth control [26], [43]. The randomized SCORE trials involve implementation in 25 villages per study arm, and will compare the relative costs and effectiveness of different drug delivery strategies for population-based control of either S. haematobium or S. mansoni infection. Two different kinds of SCORE studies will compare relative effectiveness of program implementation in communities having either low-moderate or higher prevalence of infection, defined as 10–24% prevalence among school-age children (low) vs. ≥25% (high), respectively. For the selected SCORE villages, there was village-level randomized assignment to one of several different treatment coverage and delivery options: Some communities will receive community-wide treatment (CWT), which will include treatment for all consenting adults as well as school age children; other communities will receive the more standard school-age targeted treatments (SBT) along the lines of current W.H.O. recommendations [16]. With each type of delivery, some communities will receive yearly therapy, some will transition from CWT to SBT mid-way through the project, while others will transition to every other year treatments or to ‘drug holidays’. CWT will involve the most extensive treatment coverage, aiming to include as many people as possible, whereas SBT will target schools (with non-attenders of school age also encouraged to participate), while drug holiday means there will be no treatment for the year. For communities with high infection prevalence (>25%), there are 6 proposed strategies over the 4-year study period:
For populations with lower infection prevalence (10 to 24%), no CWT is included in the SCORE trials, and treatment will be either school–based every year [S-S-S-S], two years of SBT followed by two year holiday [S-S-H-H], or SBT every two years [S-H-S-H], with resulting community prevalence determined in the 5th year.
An important consideration in choosing to implement a more intensive drug treatment programs will be the incremental cost-effectiveness relative to standard-of-care [44]. For this paper's analysis, we focused on two major costs: the cost for community screening (in determining the prevalence of active Schistosoma infection in the population) and the cost of drug treatment (drug costs+delivery costs). In the SCORE program, screening and assignment is performed only at the beginning of the 4 year treatment regimen, so that screening costs are the same in each arm. However, for comparison among current WHO-recommended strategies, we have also modeled the possible impact of rescreening and reassigning treatment regimens on a yearly basis, and for the current WHO approaches (not actually researched by SCORE) we estimated the potential cost-savings and change in effectiveness using repeated rapid screening among school age children and reassignment of treatment strategies based on those annual results [45]–[47]. [Although it requires extra expense, screening before treatment has the potential to reduce the drug cost by adapting to a less expensive strategy when prevalence of infection becomes less intense.] We took the cost of initial community sensitization and screening to be approximately one U.S. dollar (USD) per person, based on in-country costs documented by the Partnership for Child Development and other national control programs [48], The cost of annual drug dose was estimated at 0.25 USD per child and 0.50 USD per adult. Rapid program screening is typically carried out by sampling 50–100 persons for each village [45], so 100 USD was the maximum estimated screening cost per each village, no matter how large the population. The calculated cost for drug depended on the number of people treated, so information on total population, coverage for each treatment strategy, and number of villages treated was tailored to the intended coverage of each specific strategy. Taking the end of 2009 as the model's time baseline, we then projected the following metrics of each drug treatment strategy: i) the number of years needed to bring down the every village prevalence to a low-risk level (<10%) and ii) the number of years a village was likely to remain at this safe level following the termination of treatment.
One of the limitations of treatment delivered in population-based deworming campaigns is the risk of reinfection, sometimes referred to as ‘reworming’ [49]. The complex nature of Schistosoma parasite transmission can leverage parasite persistence in both snail and human hosts, while incomplete adherence with program-delivered treatments allows for persistent egg contamination by untreated residents [50]. Our previous analysis of a multi-year school-based drug treatment program for S. haematobium control indicated that the median time to reinfection can vary significantly depending on village of residence [20], [24]. Over an 8 year observation period (1983–1991), median time to reinfection could vary significantly between 2 to 8 years in adjoining villages [20]. The transmission potential within each community appeared fairly resilient to the ‘perturbation’ caused by targeted drug administration in schools, where ‘time since treatment’ and the specific study year did not significantly alter annual reinfection hazard [20]. Further, in 2000, we observed that after an 8-year period during which control had lapsed, communities that had had the highest levels of infection in 1983 (before any therapy had been given) were the same ones that had the highest levels of infection after the 8-year pause (Figure 2) (Rank test rho = 0.927, p<0.01). While, in general, school-age prevalence in each community, when we revisited in 2000, was about 32 percentage points lower than pre-control values, three of ten communities had reverted to the WHO ‘high prevalence’ category (≥50%, see Table 1 for definitions), and the remaining seven remained in the ‘moderate prevalence’ group (10–49%), while none were in the desired low prevalence category (<10%) associated with lowest morbidity risk after 8 years without control (Figure 2).
Prior to predictive simulation of treatment program outcomes, we used the most recent 2009 infection prevalence data from the same region of Kenya to check the accuracy of the SWB model's predictions about infection prevalence following perturbation by community-based drug therapy. Treatment interventions were previously implemented in that region in 2000, and on a limited basis in 2003 and 2006 [33], [51]: in 2000, 79% of those infected in the 10 village area were treated; in 2003 and 2006, only the most heavily infected villages, villages 6 and 7, were treated, with a coverage of 41–53% of all those infected in the area. Our new data from follow-up 2009 surveys [37] showed that the six-year post-treatment prevalence infection among children was again high: 61.7% for village 6 and 62% for village 7. Figure 3 shows the comparison of observed 2009 prevalence (black dots) and model predicted 2009 prevalence point estimates (gray dots) among school age children for villages 6 and 7. In sensitivity analysis, allowing model input parameters to vary randomly across a range of ±20% of their base-case values, observed 2009 prevalence values were well enclosed by the inter-quartile range of model outputs, indicating that its predictions were not extremely sensitive to changes in base case input parameters in this setting, and that our model had good predictive accuracy for this setting.
We next used our calibrated SWB model to project the likely outcomes of implementing the current WHO treatment guidelines (Tables 1 and 2) for high-, moderate-, and low-risk communities in the networked 10-village Schistosoma-endemic area (Figure 1). Using inputs based on recent community prevalence, we projected the number of years that would be required to bring all communities to a safer, low level prevalence category (<10%), as a function of community uptake (treatment adherence) across the program (Figure 4). In doing so, we had each community continue on its original treatment strategy assignment until every community achieved the <10% prevalence level. In our simulations, coverage for high-risk adults among high-prevalence populations was assumed to be at least 60%. As our model was able to target different burden strata, we interpreted “high risk groups” as those adults in the 5 highest worm burden strata. Figure 4 shows our model's estimates of the number of years the WHO regimen would have to be implemented to bring all villages under control as a function of adherence among treated children. At high levels of uptake (80–90% adherence) control was achieved in 4 years, whereas at lower levels of adherence (60%) control took twice as long (8 years).
We next asked whether repeated annual community screening and reassignment of treatment strategies (again based on the WHO recommendations in Table 2) could offer a more cost-effective approach to infection control. Table 3 indicates the results of following a rescreen/reassignment strategy for up to 8 years in the same communities. Whereas continuation of the originally assigned treatment regimen was capable of lowering prevalence to <10% among children in each individual community, the strategy of reassigning treatment regimens based on yearly community screening was much less effective. As noted in Table 3, the rescreen/reassign approach was unable to effectively suppress infection prevalence below 15%, even at the highest level of treatment uptake (90% adherence), likely related to the early switch to less intensive coverage as prevalence dropped. If adherence were less good (i.e., at a level of 60%), continuation of initial treatment strategies was less costly and more effective than the rescreen/reassign strategy. Notably, at that level of coverage, the final school age prevalence with the rescreening and reassignment approach was projected to remain at 35% at the end of 8 years (i.e., well shy of the 10% goal). If, however, adherence were >70%, rescreen/reassign became a less costly option but with, again, conspicuously subpar effects on suppression of infection prevalence (Table 3).
In examining the success of WHO treatment regimens of differing frequencies and coverage, it was apparent that village level force of transmission (i.e., its reinfection potential) played a large role in whether the program achieved or failed to achieve the goal of <10% prevalence. Local transmission factors also influenced whether a successfully treated community would stay in the ‘safe zone’ of <10% prevalence for any significant period of time after the mass treatment campaign was ended. Figure 5 demonstrates the projected prevalence outcomes for treatment of two modeled villages, one of high initial prevalence and continuing high transmission potential, and one of moderate prevalence and lower risk of reinfection. Both communities respond well to the first two rounds of drug therapy, quickly dropping to <10% prevalence. However, village A relapses to >25% prevalence after a 3 year hiatus in mass treatment, and requires ‘consolidation’ treatments to regain the <10% status. If consolidation is suspended after 4 years of annual treatments, infection prevalence rebounds to the moderate level (>10%) in two years. If consolidation runs for 8 years, community prevalence approaches it lowest levels, yet prevalence is expect to rebound over 10% in 3–4 years. By contrast, in the moderate prevalence village B having lower transmission potential, once local prevalence is well suppressed (after the second round of mass CWT treatment) then prevalence is expected to remain in the ‘safe’ zone (<10%) for 4–5 years. In this case, suppression of infection is used to indicate a persistent reduction of prevalence below 10% (the WHO-recommended threshold for low-risk communities), but this does not imply an interruption of transmission.
Figure 6 indicates the relative likelihood that a program based on WHO-recommended treatment schedules (without reassignment) would lead to successful long-term suppression of infection prevalence after mass treatment is suspended. The duration of continued suppression varied substantially between the modeled villages, and was much shorter in the communities with the highest levels of starting prevalence and greater links to high risk water sites (6–10 years suppression in lower prevalence communities vs. zero up to 3–4 years suppression in the highest risk communities). Results also varied the considerably according to the local adherence to treatment (60–90%, Figure 6). Where adherence was quite high (90%) suppression of infection with program intervention was much greater, suppressing local prevalences down to <5% (Table 3). Following this more aggressive level of suppression, rebound of infection prevalence (to >10%) took substantially longer in all villages (Figure 6).
We next examined the potential of the six modified treatment regimens currently being researched by SCORE, in terms of their ability to achieve and maintain very low prevalence of Schistosoma infection. Final outcomes were quite different in the participating villages following 4 years of treatment in community-based or school-age targeted programs (including strategies having community = >school-age crossover in coverage (C-C-S-S), and those with gaps in treatments in different years (‘drug holidays’), i.e., C-C-H-H, S-S-H-H and S-H-S-H). Figure 7 shows the prevalence of each village after four years participation in each of the 6 candidate strategies. Strategies C-C-C-C, C-C-S-S, and S-S-S-S brought the prevalence of all villages well below the 10% prevalence objective after 4 years' treatment. By contrast, strategies C-C-H-H, S-S-H-H, and S-H-S-H achieved <10% in only half of the modeled villages, and these ‘successful’ villages were the villages with starting school-age prevalences of ≤36%. The projected prevalence outcomes at the end of C-C-C-C, C-C-S-S, or S-S-S-S treatment were similar, suggesting that addition of treatment for adult populations had relatively limited incremental benefits in terms of reducing local prevalence of infection. Strategies C-C-H-H, S-S-H-H, and S-H-S-H were not able to bring all villages to a safe level of infection, with post-program prevalences of 12–31% in 5 out of 10 targeted villages after the 4-year intervention period, so further treatment would be necessary in these villages (as in discussed earlier for Figure 5).
Figure 8 indicates the projected number of ‘safe years’ (local prevalence <10%) after the 4 yearly rounds of successful area-wide C-C-C-C, C-C-S-S, or S-S-S-S intervention at an adherence level of 70%. At this level of adherence, each of these three strategies resulted in more post-program low prevalence years than projected for the six-year WHO 2006 strategy (for the latter, when assigned according to initial high/moderate/low prevalence category and without rescreening/reassignment in subsequent years). This difference was noted in the six villages having the lowest pre-treatment prevalence. By contrast, among the highest prevalence (45–69% pre-treatment) villages, neither the C-C-C-C, C-C-S-S, S-S-S-S, nor WHO strategies yielded a long-term, post-treatment suppression of infection prevalence. Reinfection was rapid, and there was no obvious advantage among the four. In terms of relative costs, among the three most effective strategies for infection suppression (i.e., C-C-C-C, C-C-S-S, or S-S-S-S), the S-S-S-S strategy was calculated to cost the least, based on differences in overall numbers treated (S-S-S-S = $4185, compared to $8455 and $6273 for C-C-C-C and C-C-S-S, respectively).
Results of our model's simulations indicate that in many moderate-to high prevalence areas, broad-based drug treatment programs will need to be sustained for four or more years to effect optimal suppression of local prevalence to <10%. Within the sub-district scale of our simulation, there were important differences in how the individual (albeit networked) communities responded to the treatment strategies that were implemented, and how long the impact of prevalence reduction could be maintained after suspension of treatment intervention. Programs with gaps in annual delivery (‘drug holidays’) were effective only in the lower prevalence villages, and did not fare as well as annual treatment programs (C-C-C-C, C-C-S-S, or S-S-S-S) in terms of long-term post-treatment suppression of infection prevalence.
For neglected tropical diseases ‘deworming’ campaigns, there has been a recognized need for optimal community uptake, sustainability, and continued political will to achieve the objectives of the program [52]. The findings of our calibrated simulation support this thinking. In particular, our analysis suggests that schistosomiasis control programs should anticipate the need for multi-year and even multi-decade programs in some villages; initially, to achieve maximal suppression of local prevalence (particularly in high-risk, high initial prevalence areas), and later, to maintain very low prevalence and maximal reduction in Schistosoma-associated disease. Intensive intervention in the initial years may result in prolonged suppression of infection in communities having low risk for reinfection, during which treatment (but, importantly, not surveillance) could be suspended. Program resources could then be re-purposed, and turned toward identification and retreatment of the highest risk villages. For now, these high risk villages are readily identified by their rapid re-emergence of infection prevalence (i.e., in 1–4 years). These same communities could also become the focus of non-drug interventions aiming at reduction of local transmission potential—in effect, re-forming high-risk villages into villages that have much lower risk of transmission and reinfection. Such efforts could include habitat modification at transmission sites, provision of safe alternatives for water use and recreation, and behavior change interventions focused on limiting water contamination and high-risk exposures.
Based on our model simulation, we found that a policy of annual rescreening of villages (with possible treatment reassignment according to current WHO 2006 recommendations [16]) before each yearly treatment can appear less costly if treatment adherence is initially high (>70%), but the associated results in terms of infection suppression are much less good (Table 3). This phenomenon is apparently related to a premature transition to alternate year therapy in truly high risk communities as they lower their prevalence into the ‘moderate’ category after the first round of treatment. For such high-risk communities, repeated annual treatment (for at least 4 years) appears to be necessary to reach fully effective suppression of infection prevalence.
There are clearly limitations to our analysis. Although our projections are based on model simulations that use an advanced stratified model that appears well calibrated for a specific region of southeastern Kenya, its projections may not be fully generalizable to other endemic areas. However, we believe that the modeled mosaic of high- and moderate- prevalence villages within a sub-district area is typical of many territories found within Schistosoma-endemic areas, and that the results of our simulation will prove valid for many other locations where schistosomiasis control proves to be challenging. Wherever school age children have the most important role in disease transmission in a given area, it is likely that the predictions for other geographic locations would be qualitatively similar to ours. Among the modeling assumptions that might influence the accuracy of our predictions are the following: i) we assume that treatment uptake is uniform across the program area and consistent from year to year. In fact, regular refusal or inability to participate, or progressive increases in non-adherence after initial years of treatment, could allow a core of untreated individuals to perpetuate transmission in a given sub-location [50], [53], [54]; ii) it is possible that repeated annual treatments may provide periodic boost to anti-schistosomal immunity and relative resistance to reinfection [55]. This feature, which could gradually reduce individual risk of reinfection, was not included in our study; iii) likewise, pond snail abundance was assumed to be stable each year over the period of the simulations. In drier landscapes where Schistosoma transmission is a rare event that is associated only with episodic flooding events, program response to therapy in high prevalence villages could be expected to be dramatically better than in our simulation, unless or until a flooding/transmission event again occurs. Of note, we have not modeled any effects of adjuvant snail control or other interventions aimed at reducing environmental transmission (this will be the focus of a forthcoming study). The focus of our present analysis was only the outcomes of different possible drug-based treatment interventions.
Given these caveats and limitations, we draw the following conclusions based on our modeling analysis:
While these projected outcomes may not be fully realized in the seven ongoing SCORE operational trials, we feel that, for now, they offer useful, evidence-based, estimates of program outcomes where anti-schistosomal control programs are now being implemented. In terms of policy discussions and program design, the simulation results raise several new topics for consideration–There is clearly a need for ready identification of villages at high risk for reinfection. For now, annual rescreening of school age prevalence provides a basic marker of risk, but identification of other (proxy) features of high-risk villages could aid significantly in year-to-year planning for program deployment. In addition, beyond continuing surveillance, programs will need to decide how to manage their low-risk villages that no longer require therapy, and decide how best to bring them to the very desirable goal of complete transmission interruption.
|
10.1371/journal.ppat.1006427 | Wolbachia elevates host methyltransferase expression to block an RNA virus early during infection | Wolbachia pipientis is an intracellular endosymbiont known to confer host resistance against RNA viruses in insects. However, the causal mechanism underlying this antiviral defense remains poorly understood. To this end, we have established a robust arthropod model system to study the tripartite interaction involving Sindbis virus and Wolbachia strain wMel within its native host, Drosophila melanogaster. By leveraging the power of Drosophila genetics and a parallel, highly tractable D. melanogaster derived JW18 cell culture system, we determined that in addition to reducing infectious virus production, Wolbachia negatively influences Sindbis virus particle infectivity. This is further accompanied by reductions in viral transcript and protein levels. Interestingly, unchanged ratio of proteins to viral RNA copies suggest that Wolbachia likely does not influence the translational efficiency of viral transcripts. Additionally, expression analyses of candidate host genes revealed D. melanogaster methyltransferase gene Mt2 as an induced host factor in the presence of Wolbachia. Further characterization of viral resistance in Wolbachia–infected flies lacking functional Mt2 revealed partial recovery of virus titer relative to wild-type, accompanied by complete restoration of viral RNA and protein levels, suggesting that Mt2 acts at the stage of viral genome replication. Finally, knockdown of Mt2 in Wolbachia uninfected JW18 cells resulted in increased virus infectivity, thus demonstrating its previously unknown role as an antiviral factor against Sindbis virus. In conclusion, our findings provide evidence supporting the role of Wolbachia–modulated host factors towards RNA virus resistance in arthropods, alongside establishing Mt2’s novel antiviral function against Sindbis virus in D. melanogaster.
| Effective vector control is critically important to reduce the incidence of diseases caused by arthropod transmitted viruses. One proposed strategy involves the use of endosymbiotic bacteria Wolbachia pipientis as a novel biocontrol agent to prevent RNA virus transmission in mosquitoes. Previous work in the field suggests that the presence of this bacterium induces virus resistance within the host. However, the underlying mechanism of this antiviral phenotype is poorly understood, impeding its widespread use. Using the alphavirus, Sindbis as our model, we explored the tripartite interaction between the virus and the endosymbiont within its natural host, the fruit fly Drosophila melanogaster. In this study, we show that Wolbachia negatively influences multiple important aspects of the virus life cycle, extending our current understanding of the molecular nature of this interaction. We also provide evidence highlighting the role of a host gene, Mt2, in Wolbachia–mediated antiviral resistance, while uncovering its previously unknown role as an antiviral host factor against Sindbis virus.
| Heritable symbioses are pervasive in nature and exceedingly common in the insect world, where many endosymbiotic associations have been described [1]. Wolbachia pipientis is an alpha-proteobacterial, maternally transmitted endosymbiont that invades insect host populations by manipulating host reproduction, favoring infected females [2]. It is present in approximately 40% of insects, including several species of Drosophila and important disease vectors such as Aedes albopictus and Culex species of common house mosquitoes [3]. In D. melanogaster, Wolbachia strain wMel exhibits weak reproductive manipulation [4], but recent work has shown that this Wolbachia strain can also protect D. melanogaster from common viral pathogens such as Drosophila C virus (DCV), cricket paralysis virus (CrPV), and Flock House virus (FHV), as evidenced by increased survival and delay in virus accumulation [5, 6]. Not surprisingly, Wolbachia–mediated antiviral protection, so-called “pathogen blocking” is considered an exciting phenomenon that could be leveraged to reduce arthropod-transmitted diseases [7].
While A. albopictus infected with its native Wolbachia strain (wAlbB) has little to no effect on RNA virus replication (such as Dengue), transinfection of non-native Wolbachia strains like wMel into this mosquito species, as well as the naturally uninfected Aedes aegypti, have been shown to induce pronounced antiviral resistance [8, 9]. Moreover, field trials conducted as a part of the global Eliminate Dengue project have demonstrated that these transinfected A. aegypti mosquitos can invade native A. aegypti mosquito populations, persisting over the seasons [10, 11]. However, although Wolbachia’s pathogen-blocking ability is currently being implemented in vector control, we know little to nothing about the mechanism underlying Wolbachia-induced antiviral resistance. Mosquitos as a model suffer from a lack of widely accessible genetic tools, coupled with the added complexity of introducing Wolbachia transinfections. Conversely, the antiviral phenotype was originally characterized in Drosophila and the genetic tractability of this model organism makes it an ideal candidate platform for a mechanistic dissection of the tripartite interaction [5, 6]. To this end, we have leveraged the power of Drosophila genetics, combined with the molecular tractability of a parallel tissue culture system, to probe previously uncharacterized aspects of Wolbachia-mediated resistance against the prototype alphavirus, Sindbis virus (SINV). In this study, we show that presence of Wolbachia in D. melanogaster results in reduced SINV infectivity, viral RNA replication and protein synthesis. Furthermore, we have identified the DNA/RNA methyltransferase gene, Mt2 as a potential host factor responsible for Wolbachia-mediated antiviral resistance. Based on these results, we conclude that resistance towards SINV occurs at an early stage of virus replication that further affects subsequent stages of the virus life cycle and show evidence supporting the hypothesis that Wolbachia modulates the expression of a host methyltransferase gene (Mt2) to target virus RNA synthesis.
We investigated the effect of Wolbachia on SINV titer in wild-type flies infected or uninfected with Wolbachia strain wMel, which we refer to as Wolb+ and Wolb- respectively from this point onwards (Fig 1A). We did not observe death as a consequence of virus infection in Wolb+ or Wolb- flies. This was expected due to the non-pathogenic nature of SINV infection in Drosophila. Flies were collected 48 hours post-injection and virus titer was determined using end-point dilution assay on BHK cells. We found that infectious virus titer was significantly reduced in the Wolb+ individuals compared to their Wolb- counterparts.
We and others have previously shown that SINV infectivity, measured by the ratio of total virus particles produced to infectious units, is influenced by the host cell environment in which the virus is cultivated [12]. To determine whether the presence of Wolbachia inside the host cell changes SINV infectivity, we used a cell culture based system comprised of Drosophila melanogaster-derived JW18 cell line infected with Wolbachia strain wMel and an antibiotic-treated Wolbachia free control cell line, which we refer to as JW18-dox. Cells were infected with SINV at an MOI of 100 and samples of growth medium were taken at 48, 72 and 96 hours post infection (Fig 1B). We performed conventional end point dilution assay to quantify infectious particles released during infection and a previously described qRT-PCR based method to quantify the total copies of virus genome present in the growth medium over the course of the infection, which is equivalent to the total number of released virus particles [12].
We found that virus derived from Wolb+ cells had a significantly high particle: infectious unit ratio relative to virus derived from Wolb- cells, indicating that on a particle basis this virus was less infectious than that derived from the Wolb- cells (Fig 1B). We have previously shown that SINV infectivity changes over the course of infection in both mammalian and mosquito cells, and also varies in a host cell dependent manner, i.e. virus grown in one cell line may be more infectious on a per particle basis than virus grown in a different cell line [12]. In the current study we found that infectivity of particles produced deteriorated over time during infection in the Wolb+ cells.
These data show that not only does the presence of Wolbachia lead to a decrease in virus produced, but also a decrease the infectivity of the SINV particles produced during infection.
During SINV infection, the incoming 49S genomic RNA functions as a mRNA to encode the viral replication complex and later acts as a template for the synthesis of minus-strand RNA. The minus strand RNA itself then serves as a template for the synthesis of nascent full length 49S genomic RNA and the transcription of smaller 26S sub genomic RNA, which encodes the viral structural proteins. Synthesis of these different viral RNA species and their subsequent translation is required for the formation of virus particles. Given that we found that the presence of Wolbachia resulted in fewer infectious virus progeny, we hypothesized that viral RNA synthesis is inhibited when Wolbachia is present at the time of SINV infection.
To test this, wild-type Wolb+ and Wolb- flies were challenged with SINV. Infection was allowed to progress for 48 hours before the flies were collected and snap-frozen. Following tissue homogenization and RNA extraction, both relative, as well as absolute quantities of the viral RNAs were determined by qRT-PCR (Fig 2). We found that presence of Wolbachia is accompanied by 2.5-fold reductions in both viral genome (quantified using primers to nsP1 coding sequence) and subgenome (quantified using primers to E1 coding sequence) RNA levels, suggesting that viral RNA synthesis is inhibited in the presence of the endosymbiont (Fig 2A).
We subsequently repeated this experiment in our D. melanogaster derived cell culture system to further determine the absolute copies of plus strand and minus strand RNA produced during virus infection. JW18 and JW18-dox cells were infected with SINV at an MOI of 100. Infection was allowed to last for 96 hours before cells were collected for lysis and subsequent RNA extraction and absolute quantities of viral RNAs were determined using qRT-PCR (Fig 2B). In line with our previous observation in flies, we found an average 10-fold reduction in total copies of virus plus and minus strand RNA.
Given our observed reduction in virus titer and RNA levels, we expected a concomitant reduction in virus protein synthesis. During infection, SINV non-structural proteins are translated as a polyprotein from the first open reading frame present in the 49S genomic RNA. Sometime later during infection, the 26S subgenomic RNA is translated as a polyprotein that subsequently gives rise to the virus structural proteins [13]. Therefore, expression of a luciferase reporter present within any non-structural or structural protein can be used as a proxy for the net translational activity from the genomic and subgenomic RNAs respectively. To this end, we used a luciferase based viral translation assay to quantitatively determine translation of SINV non-structural and structural proteins. We used SIN- nsP3-nLuc virus, in which nanoluciferase (nLuc) has been translationally fused to the hypervariable domain of SINV nsP3 protein, and SIN-cap-nLuc virus, which has nLuc translationally fused to the C-terminus of SINV capsid protein. Wild-type Wolb+ and Wolb- flies were infected independently with each of the aforementioned viruses. Flies were collected 48hr post-infection and snap frozen and nLuc activity was measured post-homogenization (Fig 3A).
Our results indicated an average 10-fold reduction in nLuc activity in tissue derived from Wolb+ individuals challenged with SIN-nsP3-nLuc virus, suggesting inhibition of SINV non-structural protein synthesis (Fig 3A). In contrast, quantification of nLuc activity in tissues derived from individuals challenged with SIN-cap-nLuc virus revealed a greater (~50-fold) Wolbachia–mediated reduction in nLuc activity (Fig 3A). These data indicate that Wolbachia causes a reduction in the synthesis of SINV non-structural and structural proteins and that expression of viral proteins synthesized off the subgenome are affected to a greater extent. However, it was not clear whether this was a consequence of decreased translational efficiency or simply a consequence of reduced quantities viral RNA available for translation (Fig 2).
To determine the cause of reduced viral protein synthesis, we extracted total RNA from the tissue homogenates of individuals infected with SIN-nsP3-nLuc or SIN-cap-nLuc viruses and determined absolute quantities of viral genomic and subgenomic RNAs using qRT-PCR as before. Following the calculation of viral protein, expressed in terms of luciferase activity (RLU), to viral RNA (copies/ug), we failed to find any significant difference in the protein-to-RNA ratios between Wolbachia–infected and—uninfected individuals, suggesting that viral protein translation is reduced due to scarcity of viral transcripts and that the smaller number of viral transcripts produced in the presence of the bacterium can be translated with regular efficiency (Fig 3B). Taken together, our data strongly indicates that Wolbachia-mediated inhibition of SINV infection occurs at an early stage of viral RNA replication that subsequently results in reduced viral protein synthesis and virus titer, which is consistent with previous reports regarding SFV infection in JW18 cells [14].
Based on evidence from other systems, we investigated whether the antiviral resistance mediated by Wolbachia could be explained by modulation of host immune gene expression by the bacterium [15–19]. We tested this hypothesis by profiling the transcriptional activity of several candidate genes spanning different pathways that have been previously implicated as being a part of the host antiviral defense. The initial examination of the effect of Wolbachia on host gene expression was performed in the absence of a viral infection. We reasoned that it is important to consider the cellular environment that the virus is initially exposed to upon entry into the cell, not necessarily one that it encounters later during infection. In addition to canonical innate immune pathway components we examined the expression of Mt2, a gene encoding a nucleic acid methyl transferase previously shown to be required for DCV resistance in Drosophila [20]. Host gene expression was determined by qRT-PCR in Drosophila in the absence or presence of Wolbachia.
We observed a general increase in expression of genes associated with canonical innate immune pathways such as Imd and Toll in the presence of Wolbachia (S1 Fig). Interestingly, we also observed a significant increase in the expression of the host methyl-transferase gene Mt2, with an average of 7-8-fold increase in transcript levels in the presence of Wolbachia (S1 Fig, Fig 4A). While a number of previous studies have discounted the role of canonical immune pathways in Wolbachia-induced pathogen blocking, the role of Mt2 in this process is less clear. Interestingly, this elevated expression of Mt2 in Wolb+ flies decreased following SINV infection (Fig 4A).
To investigate the role of Mt2 in Wolbachia–mediated inhibition of SINV infection, we looked at the effect of Wolbachia on virus infection in a previously characterized, homozygous loss-of-function mutant of Mt2 (Mt2 -/-) [21]. SINV infection was established as before in Wolbachia infected and uninfected flies that were either wild-type or Mt2 -/- and the infection progressed for 48 hours before samples were collected, snap-frozen and virus titer was determined using end-point dilution assay on BHK cells (Fig 4B). SINV titer was, on average, 10-fold higher in Wolb+ Mt2 -/- mutants compared to Wolb+ wild-type flies. Previous studies have established a direct correlation between Wolbachia density and the degree of antiviral resistance [22–24]. In light of this we examined whether the loss of function in the Mt2 gene in the Mt2 -/- mutants is accompanied by a reduction in Wolbachia titer, which could potentially explain the observed loss in virus inhibition. Quantification of DNA collected from samples using qPCR showed no reduction in Wolbachia titer in the Mt2 -/- mutants compared to the wild-type, indicating that Wolbachia titer is not responsible for the observed reduction in virus inhibition (S2 Fig).
Similar results were obtained following shRNA-targeted knockdown of Mt2 expression in two genetically distinct Wolbachia–infected fly lines, with each of the two shRNAs targeting a different area of the Mt2 gene (Fig 5A and 5B, S3 Fig). Interestingly, the degree to which virus resistance was lost seemed to correlate roughly with the extent to which Mt2 expression was reduced (Fig 5A–5C, S3 Fig).
Virus titer from Wolbachia uninfected wild-type flies were still found to be higher compared to the Wolbachia infected Mt2 -/- mutants (Fig 4B) although the data failed to meet the threshold for statistical significance. Additionally, comparison of virus titer from Mt2-/- flies with and without Wolbachia indicated that pathogen blocking is mediated by factors in addition to Mt2. However, once again the difference observed failed to reach statistical significance.
Given the fact that our initial characterization of Wolbachia–mediated SINV resistance showed viral inhibition occurred at an early stage of RNA synthesis and consequently resulted in reduced viral protein levels, we next sought to determine the effect of Mt2 at these particular stages of virus replication. Quantification of SINV transcripts was performed using qRT-PCR as described above for absolute viral genome quantification (Fig 4C). Compared to Wolbachia–infected wild-type individuals, loss of Mt2 resulted in an average of 2-3-fold increase in viral transcript levels. In contrast, we did not see any significant difference in the transcript levels between Wolbachia infected Mt2 -/- mutants compared to uninfected wild-type and Mt2 -/- mutants, showing that SINV RNA synthesis can be almost fully restored following the loss of Mt2 function in the presence of Wolbachia.
While the data presented above indicated that inhibition at the stage of SINV RNA synthesis previously observed in the presence of Wolbachia is significantly reduced in Mt2 -/- mutant, it did not answer whether or not viral protein levels were restored in these mutants. We utilized our previously described SIN-cap-nLuc virus and luciferase based viral translation assay to quantify levels of SINV structural proteins in wild-type and Mt2 -/- mutants either in the presence or absence of Wolbachia. We found significantly higher expression of SINV structural proteins in the Mt2 -/- background (Wolb+ and Wolb-), relative to that observed in their wild-type counterparts (Fig 4D). This result implies that the antiviral effect of Mt2 targets a stage of viral RNA synthesis, with the loss of Mt2 restoring the RNA and consequent protein levels to wild-type Wolb- levels.
We next examined whether the antiviral effect of Mt2 against SINV could be independent of Wolbachia, as it has been reported before to be antiviral in the context of native DCV infection in D. melanogaster [20]. To test whether overexpression of the Mt2 gene by itself resulted in SINV resistance, we used Gal4-UAS expression system to overexpress Mt2 in Wolbachia uninfected flies. While we were only able to achieve modest levels of overexpression, following infection with SINV, Mt2 overexpressing individuals were found to accumulate virus at an average of 2-fold lower titer than their wild-type sibling controls (S4A Fig). Levels of viral RNA in Mt2 overexpressing flies were, on average, 2.5 fold lower compared to wild-type sibling controls (S4B Fig).
To further investigate the antiviral role of Mt2 against SINV, we asked whether Mt2 is responsible for our initial observation regarding the reduction of virus particle infectivity in the presence of Wolbachia. To this end, we utilized our Wolbachia uninfected JW18-dox cell culture model to knock down expression of Mt2 using targeted dsRNA against the methyltransferase gene. Cells were transfected with either Mt2-specific, or non-targeting dsRNA. SINV infection was established 48-hours post-transfection at an MOI of 100 and infection was allowed to last for 96 hours before cells and media were harvested separately.
Following RNA extraction from harvested cells, qRT-PCR based quantification of Mt2 gene expression revealed an average 50% knockdown relative to non-targeting controls (Fig 6A). Probing the effect of Mt2 knockdown on SINV RNA synthesis revealed viral RNA levels to be around 14-fold higher in cells transfected with Mt2 dsRNA (Fig 6B), further confirming the results obtained in our animal model (Fig 4C).
Effect of Mt2 knockdown on SINV particle infectivity in JW18-dox cells was performed as described previously by determining the ratio of total particles to the total number of infectious units. Virus titer was calculated using standard end-point dilution assay on BHK-21 cells. Relative fold change in virus titer from cells transfected with Mt2 dsRNA was found to be on average, 200-fold higher compared to cells treated with non-specific dsRNA (Fig 6C). Consequently, virus particle infectivity was found to increase for virus derived from cells in which Mt2 was knocked-down as evidenced by the 70 percent reduction in the particle-to-TCID50 ratio, indicating that Mt2 plays a role in regulating virus particle infectivity (Fig 6D). Taken together, these results show Mt2 to possess antiviral activity against SINV.
Arboviruses (arthropod-borne viruses) represent significant and imminent threats to public health across the globe due to the lack of commercially available vaccines or antiviral therapeutics. Several of the classic arboviral diseases are characterized by their long periods of absence, followed by sudden cases of re-emergence which may occur at unpredictable intervals, often decades apart [25, 26]. Therefore, until there are measures to inhibit viral replication and improve disease outcome, vector transmission control remains as the most effective strategy for preventing the spread of arboviral and other vector-borne pathogens. However, the increasing burden of diseases associated with dengue, chikungunya and Zika viruses reflect the ineffectiveness of traditional vector control strategies, thereby promoting the need towards developing unconventional, yet viable alternatives [27]. One promising approach involves harnessing the pathogen blocking properties exhibited by the maternally inherited endosymbiont Wolbachia pipientis.
At this time the causal mechanism behind Wolbachia induced RNA virus resistance in insects is poorly understood. Despite being observed in both native and transiently established Wolbachia infections [5, 6, 9–11, 14, 20, 22, 28–33], a considerable amount of phenotypic variability exists between different Wolbachia strains, hosts and viruses [22, 28–33]. In some cases, differences in titers of endogenous versus transinfected Wolbachia strains have been cited as a potential determining factor, while others have suggested a mechanistic divide between different host systems [8, 22]. However, it is important to note that despite these differences, antiviral resistance induced by Wolbachia is specific to RNA viruses, and this fact could help to narrow the underlying mechanism of virus inhibition. The data presented above demonstrate Wolbachia’s capacity to inhibit SINV replication in an arthropod host. We observed a significant decrease in the infectivity of SINV particles produced from Wolbachia infected cells. That is, while a significant number of virus particles were being produced in the presence of Wolbachia they were less infectious on a per particle basis than those produced in the absence of Wolbachia. This is important, as these results may help explain recent observations regarding the lack of detectable infectious virus particles present in salivary glands of wMel-infected A. aegypti mosquitoes harboring other RNA viruses [31–33]. Finally, we also showed that this block happens early during infection at the level of viral RNA replication and that differential expression of the host methyltransferase gene, Mt2, plays a significant role in blocking SINV infection.
The physiology of the host cell harboring a virus influences the titer of the resulting virus progeny. Several factors, such as the stage of infection, for example, influence virus titer, whereby titer improves temporally as virus accumulates with the progress of infection. On the other hand, virus titer is also dependent on the physiological attributes of the cell from which the virus originates [12]. Given that the presence of Wolbachia induces resistance against different RNA viruses, it would be logical to think that the intracellular endosymbiont alters cellular physiology, creating an antiviral state that consequently influences the titer of the progeny virus produced. If this is indeed the case, where the induced antiviral effect is cell autonomous, then virus resistance would unconditionally require the presence of Wolbachia inside the cell. Furthermore, varying densities of Wolbachia in different tissues would then affect the degree of antiviral effect from cell to cell [34]. Indeed, Wolbachia is known to influence host gene expression [35, 36] and in certain cases, altered host gene expression negatively influences the development of malarial parasites in Wolbachia–infected vector species such as A. aegypti and Anopheles gambiae [15, 16]. In further support of an altered cellular environment upon Wolbachia infection, the bacterium was found to inhibit translation of SFV in JW18 cells as early as 7hpi with an absence of active transcriptional change in Wolbachia genes in response to virus infection [14]. This led to the conclusion that Wolbachia’s antiviral mechanism is either fast acting or is already present upon virus infection [14]. These results, and varying degrees of evidence, support the idea that Wolbachia might prime the host immune system, including factors such as the induction of ROS and alterations of host miRNA profiles, to induce antiviral resistance [17, 37 – 39]. However, several studies have discounted the role of canonical Drosophila innate immune components like RNAi as well as pathways such as Toll and Imd [19, 39]. In contrast, little attention has been given towards investigating the role of non-canonical host genes that may be involved in resistance against viruses. We identified the DNA/RNA methyltransferase Mt2 as a host factor that restricted SINV infection in Drosophila melanogaster and established a link between its antiviral function and Wolbachia–mediated host resistance against SINV.
D. melanogaster Mt2 gene encodes for a 40 kDa Dnmt2 protein which belongs to the most widely conserved Dnmt protein family and is the denoted as the lone cytosine methyltransferase present in many arthropods, including flies and mosquitoes [40, 41]. Its function is unconventional, given that Dnmt2 has been shown to methylate both DNA and tRNAAsp at the cytosine-5 position (m5C). Dnmt2-directed cytosine methylation of genomic DNA in D. melanogaster is limited to specific loci on selected retrotransposons [42–44]. However, its robust RNA (tRNA) methyltransferase activity has been demonstrated both in vivo and in vitro, as evidenced by C38 methylation of tRNAAsp and additional tRNAs [21, 45]. This is significant because Dnmt2 is required for DCV resistance in Drosophila and RNA immunoprecipitation experiments demonstrated interaction of Dnmt2 with DCV RNA during infection [20]. Here, we extend the antiviral role of Dnmt2 in SINV inhibition, both in the presence and absence of Wolbachia. Our results show that Wolbachia infection leads to an increase in Mt2 expression prior to virus infection, thus the virus is being introduced to a system in which Mt2 expression is high. Interestingly SINV infection led to a reduction of Wolbachia-induced Mt2 expression. This is important to note as a prior study found no effect of Wolbachia on Mt2 expression, but in that study the analysis of expression levels was performed after virus infection [22]. Additionally, we show that loss of Mt2 results in a significant increase in SINV transcript levels as well as a concomitant increase in the synthesis of virus structural proteins and virus titer. We hypothesize that Mt2 may function to inhibit viral RNA synthesis, which is in line with its previously described RNA binding properties.
Unlike traditional DNA methyltransferases, whose subcellular localizations are predominantly in the nucleus, evidence has been provided that Dnmt2 is distributed in the nucleus and the cytoplasm where SINV replication takes place. Indeed, fractionation studies of protein extracts from Drosophila embryos and localization of fluorescently tagged Dnmt2 in DCV infected fly tissues have confirmed the presence of Dnmt2 in the cytoplasm [44, 20]. The fact that cytoplasmic relocalization of Dnmt2 in response to cellular stress has been observed in other systems is significant, since elevated oxidative stress in the presence of Wolbachia is correlated with antiviral resistance in both flies and mosquitoes [37, 38, 46, 47]. In contrast to the data presented here Zhang et al. found that mosquito (A. aegypti) Dnmt2 promoted dengue virus replication [38]. It is not clear how to reconcile these contrasting observations, however it is not surprising that viruses from different viral families are affected differently by a host factor that is elevated in response to Wolbachia. Although further research is required to determine the exact nature of interaction between SINV RNA and Dnmt2, previous work involving DCV revealed significantly higher virus induced mortality in flies carrying a catalytically inactive mutant of Dnmt2, which suggests that enzymatic activity of Dnmt2 is required for its function as an antiviral factor [20]. Future work will focus on determining the RNA methylation status of the SINV genome in the presence or absence of Wolbachia and the relative importance of Dnmt2’s RNA methyltransferase activity in SINV inhibition.
Although we discovered a link between Mt2 and Wolbachia’s pathogen blocking, the majority of the mechanistic details regarding this phenomenon remain unknown, and certainly other Wolbachia induced factors are likely to be involved. At this time, it is unknown whether Dnmt2 alters the methylation of SINV RNA. Given what is known about the functional implications of mRNA modifications in eukaryotic organisms, it is possible that the viral RNA acts as a target for cellular methyltransferases that subsequently influences the function of that RNA either positively negatively. Indeed, evidence of eukaryotic host methylation machinery acting as an either proviral or antiviral factor has been reported in the field [48–53]. Recent viral RNA methylation studies have found conserved, m6A modified regions spanning the RNA genomes of multiple members of Flaviviridae family, including Hepatitis C Virus (HCV), Zika Virus (ZIKV), Dengue Virus (DENV), Yellow Fever Virus (YFV) and West Nile Virus (WNV) [51, 52].
In the context of Wolbachia infection in insects, various degrees of evidence support the role of the endosymbiont in modulating the cytosine methylation profile of the host. In D. melanogaster, presence of native Wolbachia wMel is correlated with increased methylation across all cytosine residues in male testes [54]. In addition, transcriptome-wide methylation profiling of virulent wMelPop infected A. aegypti mosquitoes reveal mostly random, yet widespread changes in the host cytosine methylation profile compared to Wolbachia uninfected individuals [55]. It is important to note that similar to D. melanogaster, mosquitoes such as A. aegypti and A. gambiae possess a single Dnmt2 like cytosine methyltransferase [56, 57]. Therefore, given the selective nature of Wolbachia induced resistance towards viruses possessing RNA genomes, as well as growing evidence in the field that viral genome methylation regulates RNA virus infection, our data support the idea of Wolbachia induced modification of SINV RNA.
Whether or not Wolbachia-induced Mt2 expression changes the methylation state of the viral RNA it is apparent that it has a biological effect on virus replication. Our data indicate this effect is exerted at an early time in infection and results in reduced levels of RNA synthesis. Future work fill focus on understanding how Dmnt2 interaction with viral RNA influences RNA function. The observation that infectious unit production is more severely inhibited than particle production Wolbachia infected hosts indicates that the genomes in the particles are less capable of initiating infection. It is known that methylation of RNA can change the stability of that RNA [48] which, if true for SINV RNA, could result in decreased viral RNA synthesis. Alternatively, Wolbachia-mediated Dmnt2 activity may lead to downstream alterations in the use of the genome as a template for translation and minus-strand RNA synthesis.
Continued use of Wolbachia as an emerging vector control agent requires better understanding of the molecular mechanism underlying its antiviral resistance in the context of a tractable arthropod system. We provide evidence that the presence of Wolbachia reduces SINV particle infectivity and results in reduced virus titer. Furthermore, our data indicate a reduction in viral RNA synthesis, accompanied by decreased viral protein synthesis. Further characterization of early infection events is required to determine the exact nature of Wolbachia–mediated inhibition of SINV genome replication. Importantly, our data clearly demonstrate antiviral activity of the host RNA methyltransferase gene Mt2 in the presence and absence of Wolbachia. Given the data provided in this study, it is likely that the effect of Mt2 lies at the stage of viral RNA synthesis/stability. Future work will therefore focus on characterizing the nature of interaction between Mt2 and SINV RNA. The fact that both Wolbachia mediated antiviral resistance and the antiviral effect of Mt2 extends to members of different single stranded RNA virus families is significant, suggesting that existence of a common mechanism [5, 6, 9–11, 14, 20, 22].
Drosophila fly stock 6326 (of a W1118 genetic background), carrying Wolbachia was obtained from the Bloomington Stock Center (BDSC), and used as the infected wild-type background. A corresponding Wolbachia–uninfected line was created through tetracycline treatment (approx. 20 ug/mL fed in the fly media for 3 generations). Wolbachia infection status was subsequently confirmed through quantitative PCR using published primer sets (S1 Table)[58]. The flies were repopulated with a wild-type microbiota post tetracycline treatment through culture in bottles previously occupied by untreated male flies of the same background (stock 6326). UAS-Mt2 and Mt2 loss of function flies (provided by S. Bordenstein) [54] were used to examine the role of the Drosophila DNA methyltransferase gene on the pathogen blocking phenotype. The Mt2 loss of function mutation is in a W1118 background as described by LePage et al. [54]. Wolbachia-infected TRiP mutant stocks 38224 (y1 sc* v1; P {TRiP.HMS01667} attP40) and 42906 (y1 sc* v1; P {TRiP.HMS02599} attP40) were used for shRNA-targeted knock-down of Mt2 gene expression by driving Mt2 shRNA expression using previously described Act5C-Gal4 driver males (provided by Brian Calvi) y1 w*; P{Act5C-GAL4}25FO1/CyO, y+ (Fig 5) or y1 w*; P{w[Act5C-GAL4}17bFO1/TM6B, Tb1 (S3 Fig). For overexpression of Mt2, crosses were performed between UAS-Mt2 males and virgin Act5C-Gal4 driver females. Among the resulting progeny, straight-winged flies were considered to be overexpressing while siblings exhibiting Cyo phenotype were used as the wild-type control. All fly stocks were and maintained on standard cornmeal-agar medium supplemented with P/S at 25°C on a 24-hour light/dark cycle. In order to establish a systemic virus infection in vivo, flies were anesthetized with CO2 and injected in the thorax with 50nL of approximately 1010 PFU/mL of pelleted virus or control PBS using a glass capillary needle. Flies were collected two days post-infection, snap-frozen in liquid N2 and stored at -80°C for downstream processing.
In all stocks harboring Wolbachia infection, the Wolbachia strain was wMel2, confirmed by genotyping as shown in S5 Fig using primers described in Riegler et.al. 2005 [59] (S1 Table).
Wolbachia-infected Drosophila melanogaster JW18 cells and corresponding doxycycline-treated Wolbachia-uninfected JW18_dox cells (a generous gift from W. Sullivan) were maintained at 24°C in Shields and Sang media (Sigma), supplemented with 10% heat-inactivated fetal bovine serum (FBS) and 1% Antibiotic-Antimycotic (Gibco). BHK-21 cells (American Type Culture Collection) were grown at 37°C under 5% CO2 in MEM (CellGro) supplemented with 1% L-Gln, 1% Antibiotic-Antimycotic (Gibco), 1% non-essential amino acids and 10% heat inactivated Fetal Bovine Serum (FBS). JW18 cells were seeded in a 6-well plate at a density of approximately 2.26x106 cells/well 24h prior to infection. Infection was carried out using virus derived from BHK-21 cells, titered using a standard plaque assay. Virus was diluted in Shields and Sang media (Sigma) and the infection was established at an MOI = 100. For doxycycline cleared JW18-dox cells, around 50–55% infection was observed at 96 hpi. Mock infections were carried out by treating the cells equivalently, without the addition of virus.
Quantification of viral genome and subgenome translation was performed by introducing SINV luciferase reporter viruses nsP3-nLuc and cap-nLuc, respectively into 2-day old virgin female flies as described. The samples were homogenized in 1X Cell Culture Lysis Reagent (Promega) after they were collected 2-days post infection and clarified via centrifugation at 16,000 × g for 2 min. The samples were then mixed with luciferase reagent (Promega), and luminescence was recorded using a Synergy H1 microplate reader (BioTech instruments). In all cases, the luciferase readings were normalized to the levels of SINV genomic and sub-genomic RNA present in the assayed samples, determined by using qRT-PCR methods with ΔΔCT calculation of transcript levels.
Wolbachia density from fly homogenates and tissue culture cells were determined via qPCR on whole DNA using an Applied Biosystems StepOne Real-time PCR system (S1 Table) and SYBRGreen Chemistry (Applied Biosystems), previously described in [58]. Wolbachia density (and infection status) in JW18 cells was determined using DAPI staining, where Wolbachia were visualized in the form of cytoplasmic foci within the cells.
Knockdown of Mt2 expression was achieved in doxycycline treated JW18 cells using target dsRNA against the Mt2 gene. Mt2 dsRNA was synthesized from corresponding dsDNA generated using self-annealing primer sets. First, custom oligonucleotide was designed with a 5’ T7-Polymerase binding site (GAATTAATACGACTCACTATAG) followed by a 3’ target sequence specific to the Mt2 coding region. Similarly, another oligonucleotide was designed with a similar 5’ T7 polymerase binding site followed by a 3’ end complementary to the 3’ end of the previous oligo (S1 Table). Polymerase chain reaction was carried out using 100uM of primers and Q5 High-fidelity Master Mix [NEB] to produce dsDNA. dsRNA was synthesized via in-vitro transcription of this dsDNA, using T7-RNA Polymerase in the presence of the 5'cap analog 7'G5'ppp5'G [New England Biolabs], followed by transfection of 500 ng dsRNA into JW18-dox cells using Lipofectamine LTX [Thermo Fisher Scientific]. Maximum knockdown was achieved at 48 hours post transfection.
Whole flies were homogenized in TRIzol reagent, followed by RNA extraction. cDNA was synthesized using MMulV Reverse Transcriptase (New England Biolab) with random hexamer primers (Integrated DNA Technologies). Negative (no RT) controls were performed for each target. Quantitative RT-PCR analyses were performed using Brilliant III SYBR green QPCR master mix (Agilent) with gene-specific primers according to the manufacturer's protocol and with the Applied Bioscience StepOnePlus qRT-PCR machine (Life Technologies). The expression levels were normalized to the endogenous 18S rRNA expression using the delta-delta comparative threshold method (ΔΔCT). Fold changes were determined using the comparative threshold cycle (CT) method (S1 Table).
P-values were calculated as described in the individual Fig legends. The average fold change (FC) in each experiment was calculated using the variable bootstrapping method, measuring the fold change between each potential pair of flies to determine the variability of the mean [12]. 95% confidence intervals (CI) were calculated using one sample t-test of log2FC values to determine the significance of distribution of the mean relative to the null using IBM SPSS Statistics Software 24 [60].
|
10.1371/journal.ppat.1003130 | Memory T Cells in Latent Mycobacterium tuberculosis Infection Are Directed against Three Antigenic Islands and Largely Contained in a CXCR3+CCR6+ Th1 Subset | An understanding of the immunological footprint of Mycobacterium tuberculosis (MTB) CD4 T cell recognition is still incomplete. Here we report that human Th1 cells specific for MTB are largely contained in a CXCR3+CCR6+ memory subset and highly focused on three broadly immunodominant antigenic islands, all related to bacterial secretion systems. Our results refute the notion that secreted antigens act as a decoy, since both secreted proteins and proteins comprising the secretion system itself are targeted by a fully functional T cell response. In addition, several novel T cell antigens were identified which can be of potential diagnostic use, or as vaccine antigens. These results underline the power of a truly unbiased, genome-wide, analysis of CD4 MTB recognition based on the combined use of epitope predictions, high throughput ELISPOT, and T cell libraries using PBMCs from individuals latently infected with MTB.
| Mycobacterium tuberculosis is one of the most life-threatening pathogens of all time, having infected one-third of the present human population. There is an urgent need for both novel vaccines and diagnostic strategies. Here, we were able to identify the targets most dominantly recognized by latently infected individual that successfully contain infection. These targets are contained in three broadly genomic antigenic islands, all related to bacterial secretion systems and composed by several distinct ORFs. Thus, our results suggest that vaccination with one or few defined antigens will fail to replicate the response associated with natural immunity. Our analysis also pinpoints that the Th1 cells dominating the response are associated with novel and well-defined phenotypic markers, suggesting that the response is molded by unique MTB associated factors. This study demonstrates further that the approach combining peptide binding predictions with modern high throughput techniques is generally applicable to the study of immunity to other complex pathogens. Together, our data provide a new angle in the worldwide fight against M. tuberculosis and could be used for diagnostic or vaccine developments.
| Tuberculosis is one of the major causes of death from infectious disease. Current diagnostics do not distinguish active and latent infection, and the only available vaccine has limited efficacy. Hence, there is an urgent need for both novel vaccines and diagnostic strategies.
Human T cell responses to MTB involve CD4, CD8, CD1 and γ∂ T cells. Seminal studies showed that human memory T helper 1 (Th1) cells directed against the purified protein derivative (PPD) of MTB secreted IFN-γ [1]. IFN-γ has an essential role in the protective immunity to mycobacteria, as individuals with genetic defects in the IFN-γ receptor has an increased susceptibility to mycobacteria [2]. Th1 cells mainly express the chemokine receptors CCR5 and CXCR3 [3], while Th17 cells co-express CCR6 and CCR4 and Th22 cells co-express CCR6 and CCR10 [4], [5].
While several studies have reported the identification of MTB antigens, from abundant or easily purified proteins [6], [7], a truly genome-wide study to identify antigens is lacking. In only a few cases have techniques allowing ex vivo detection and/or characterization of MTB-specific T cells, prior to any in vitro expansion and manipulations, been utilized [8], [9], [10].
A key issue relating to MTB immunity is whether different classes of antigens elicit responses that have the same or diverse functional characteristics. MTB antigens described so far are predominantly secreted MTB proteins [11], Some of which are not essential for bacterial survival [12]. As a result, it was hypothesized that secreted proteins might act as decoy antigens, diverting the immune response from recognizing more relevant MTB proteins [13].
In this regard, two intriguing MTB protein categories are the PE/PPE proteins, and the Esx protein family, which have been shown to elicit B and T cell responses [14], [15]. The function(s) of PE/PPE proteins are not fully understood but data indicates that they influence antigen presentation and host cell apoptosis [15]. The PE/PPE genes encode almost 200 proteins (4% of the total open reading frames (ORFs)) [16], unique to Mycobacteria and most prevalent in pathogenic strains. While PE/PPE proteins are mainly located within the bacterial cell wall and cell surface, some are also secreted [17], [18].
PE/PPE genes are closely related to the Esx regions [19]. These regions encode Type VII secretion systems (T7SS), also known as Esx secretion systems. Five related, but functionally distinct and non cross-complementing T7SS (Esx 1–5), are present in MTB [20]. The best characterized is Esx-1, which encodes the Rv3874 (culture filtrate protein 10 kDa, CFP10) and Rv3875 (early secretory antigenic target 6 kDa, ESAT-6) antigens [20]. The genes encoding the Esx proteins, are arranged in tandem pairs (EsxA-W) at 11 genomic loci [21]. Esx secreted proteins have been detected from Esx-1, -3 and -5 indicating that these are functional secretion systems [22].
T cell epitopes have been described from all main MTB protein categories, indicating that protein function or cellular location per se does not determine which proteins can be recognized. Previous studies in complex pathogen systems demonstrated that immune responses are directed against a relatively large fraction of the genome [23], [24]. However, epitope reactivity is currently described only from about 4% of the approximately 4,000 ORFs of the MTB genome ([11] (IEDB, www.iedb.org)). Hence, we hypothesized that a genome-wide probe of the immunogenicity of MTB ORFs would reveal a large number of novel antigens. Defining the breadth of responses is key for the design of vaccination strategies that mirror natural immunity [25], evaluation of disease the performance of vaccine candidates and the development of diagnostics.
By combining HLA class II peptide binding predictions with modern high throughput techniques such as ex vivo ELISPOT analysis, HLA class II multimers, and the screening of T cell libraries [26], we were able for the first time to identify and characterize the genome-wide antigen response in latently infected individuals.
Protein sequences from five complete MTB genomes (CDC1551, F11, H37Ra, H37Rv and KZN 1435) and sixteen draft assemblies from the NCBI Protein database (Table S1) were aligned. The binding capacity of all possible 15-mer peptides (n = 1,568,148) was predicted for 22 HLA DR, DP and DQ class II alleles (Figure S1 and Table S2) most commonly expressed in the general population [27], to select peptides predicted to bind multiple HLA class II alleles (promiscuous epitopes). This approach identifies the most dominant and prevalent responses, corresponding to approximately 50% of the total overall response [27].
A total of 20,610 peptides (with a range of 2 to 10 per ORF, and an average of 5), including 1,660 variants not totally conserved amongst the genomes considered in the analysis, were synthesized and arranged into 1,036 peptide pools of 20 peptides (Figure S1). The ex vivo production of IFN-γ by PBMCs from 28 LTBI donors induced by each of the 1,036 pools was measured utilizing ELISPOT. Pools recognized by ≥10% of donors were deconvoluted, and 369 individual MTB epitopes were identified (Table S3). Individual donors recognized, on average, 24 epitopes, underlining the large breadth of response to MTB.
Epitope responses were ranked on the basis of magnitude to assess their relative dominance. The top 80 epitopes accounted for 75% of the total response and the top 175 epitopes accounted for 90% of the total response (Figure 1A). Only occasional weak responses were detected in 28 TB uninfected/non-Bacille Calmette-Guérin (BCG) vaccinated control donors, thus demonstrating that these responses were LTBI-specific (Figure 1A). The epitopes were mapped to individual MTB antigens using the H37Rv as a reference genome. A total of 82 antigens were recognized by more than 10% of LTBI donors (Figure 1B). These 82 antigens accounted for approximately 80% of the total response in LTBI donors (Figure 1C). Responses to the epitopes from the most frequently recognized antigens were further characterized utilizing PBMCs depleted of either CD4 or CD8 T cells. The majority (97%) of these epitopes were recognized exclusively by CD4 T cells (Table S3), as expected because of their identification on the basis of predicted HLA class II binding capacity.
Comparing these 82 most prevalently recognized antigens with antigens for which similar ex vivo epitope reactivity has been described (IEDB), we found that the majority (61/82 antigens, 74%) was novel. While a given antigen might not have been analyzed in sufficient detail to lead to the description of defined epitopes, it might nevertheless have been described as a target of T cell responses. Therefore we performed a literature search for each individual antigen to further categorized them as novel, or as targets of CD4 T cells, CD8 T cells or undefined T cell type. This revealed that 41% of the antigens we identified had not previously been described as T cell targets (Figure S2A and Table 1). The responses to novel antigens, in terms of both response frequency and magnitude, are comparable to those directed against previously known T cell targets (Table S4).
Further analysis of the IEDB data revealed a limited overlap, (18%; 28/158) between antigens identified in this study and antigens known as sources of HLA class I epitopes (Figure S2B). Finally, no significant correlation was found with the antigens recognized by serological responses from the MTB proteome [28] (Figure S2C).
Next, using the TubercuList database [16], we determined the protein category to which the identified antigens belong (Figure 2). As expected, the identified antigens were associated with almost every category, with the exception of regulatory proteins and proteins of unknown function. The significant overrepresentation of PE/PPE proteins was notable, as well as the underrepresentation of proteins in the conserved hypotheticals, cellular metabolism and respiration categories.
The localization of antigens recognized was next visualized by plotting the recognition data on a linear map of the MTB genome. Analysis of either percent of donors responding or percent of total response revealed striking clusters of reactivity within certain regions of the genome (Figure 3A). When the MTB genome was parsed into 5-gene windows, significant antigenic clusters (defined by minimum 4 proteins within the 5-gene window being recognized by 7.1% of LTBI donors) could be identified using binomial distribution probability and Bonferroni correction. Three significant antigenic islands (Figure 3B), encoding 0.55% of the total ORFs, accounted for 42% of the total response (Table 2). One of the islands (Island 3) contains the well-known Rv3875 and Rv3874 antigens, which is an Esx protein pair secreted via a T7SS. Strikingly, the other two islands also contain Esx protein pairs. Moreover, two of the antigenic islands are part of the known T7SS systems Esx-1 (Island 3) and Esx-3 (Island 1). It is noteworthy that the proteins recognized included not only the proteins believed to be secreted, but also the proteins forming the actual secretion apparatus (Island 1). Indeed, the antigens identified within these islands correspond to proteins from several different protein categories, mostly assigned to the cell wall and cellular processes and the PE/PPE category, which is not surprising since several of these proteins are part of the T7SS. Additionally, Rv3615c [29], which is functionally linked to Esx-1 [30], was also prevalently recognized. However, it stands as a single antigen and not as part of an antigenic island.
To dissect whether the main determinant of immunodominance was related to a given antigen being contained within an antigenic island or belonging to PE/PPE and Esx proteins families, we calculated the percentage of the total response for different groups of proteins as well as the percentage of the MTB genome associated with these protein groups (Table 2). To compare different protein groups we calculated the ratio between % of response and % genome, as a percent enrichment.
The PE/PPE proteins were responsible for 19% of the total response, and when divided into PE/PPE proteins within an island compared to non-island, the island PE/PPE were more predictive of immunogenicity than the non-island ones (Table 2). Also, in the case of Esx proteins and T7SS, proteins within the antigenic islands were more likely to be immunogenic than those outside the islands. Proteins not in the antigenic islands, and not belonging to PE/PPE and T7SS categories, were responsible for 14% of the total response (Table 2). Thus, these data show that the antigenic islands identified are highly predictive of immunogenicity, and that to be contained within the antigenic islands is the most reliable predictor of the immunodominance of PE/PPE and Esx proteins.
It has been proposed that some of the responses against secreted MTB proteins act as decoys [13], thereby supporting bacterial persistence. It has also been proposed that T cells differing in their degree of multifunctionality might differ in terms of protective potential, or have a role in pathology [31], [32], [33], [34]. Definition of dominant antigens allows testing the validity of these hypotheses. To address these issues we detailed responses against PE/PPE, Esx and other proteins expressed in the three major antigenic islands, or elsewhere, by a variety of approaches, including multiparameter intracellular cytokine staining (ICS) assays, tetramer staining and T cell libraries.
The frequency of IFN-γ, TNFα, and IL-2 expressing CD4 T cells elicited by proteins from the PE/PPE and cell wall and cell processes category, and from within an island versus non-island, induced similar cytokine expression patterns (Figure 4A and C; gating strategy in Figure S3). The vast majority of CD4+ T cells were IFN-γ+TNFα+IL-2+ or IFN-γ+TNFα+, followed by TNFα+ single producing CD4+ T cells. To a lesser extent, TNFα+IL-2+, single IFN-γ+, and single IL-2+ cells were also detected (Figure 4A and C). Triple cytokine producers were found in 27–40% of cytokine-expressing CD4+ T cells, 30–43% expressed any 2 cytokines, and 23–44% produced a single cytokine (Figure 4B and D). We did not observe any donor-, antigen- or epitope-specific pattern of cytokine production (Figure 4E).
CD4+ T cells were stained with selected HLA-epitope tetramer reagents and tetramer+ cells were enriched [8], [35]. Epitope-specific T cell responses were detected in 16 donors at frequencies 0.25 to 24.3% (median of 3.8, interquartile range 1.5–15.3) for seven different HLA/T cell epitope tetramer combinations (Figure 5A). Only a small number of tetramer-positive cells were detected with the epitope-specific tetramers in donors with a HLA mismatch (Figure 5A), which confirmed that tetramer specificity was derived from the epitope and HLA molecule combination. Epitope tetramer combinations were selected based on the number of donors responding, HLA restriction, and the availability of corresponding reagents for tetramer production. Memory subset phenotypes were determined using Abs to CD45RA and CCR7. Similar to the multifunctionality phenotype, we did not observe any differences in memory phenotype when comparing proteins from within an island vs. non-island (Figure 5B and C). Rv0129c/Rv1886/Rv3804, Rv3418c and Rv1195 epitope-specific tetramer+ T cells predominantly consisted of CD45RA−CCR7+ central memory T cells in all donors analyzed, followed by effector memory (CD45RA−CCR7−). Percentages ranged between 70.1 and 91.3% (median 85.0, interquartile range (77.7–86.8)) for central memory T cells and 8.6–26.8% (13.3 (10.2–19.0)) for effector memory T cells. Only a minor fraction appeared to be naïve (CCR7+CD45RA+) or effector T cells (CCR7−CD45RA+). For Rv0288/Rv3019c the percentages ranged between 49.5 to 84.5% (56.8 (52.0–74.7)) for central memory T cells, 9.8–37.1% (25.9 (13.3–33.8)) for naïve and 4.8–17.2% (10.0 (7.4–16.8)) for effector memory T cells. Again, a minor fraction of the tetramer+ cells appeared to be effector T cells (Figure 5B and C).
To measure frequency and distribution of MTB-specific T cells, we used the T cell library method [26]. The majority of epitope-specific tetramer+ cells were found to be CD45RA−. We therefore stained CD45RA−CD25− CD4 T cells from donors latently infected with TB (LTBI) with antibodies against chemokine receptors preferentially expressed on functionally distinct memory T cell subsets [36]. Five Th cell subsets were sorted: 1) CXCR3+CCR6−; 2) CXCR3+CCR6+, both enriched in Th1 cells; 3) CCR4+CCR6− (Th2); 4) CCR4+CCR6+ (Th17); and 5) CCR6+CCR10+ (Th22) [5]. MTB-specific T cells were almost exclusively found in the CXCR3+CCR6+ subset, while Flu-specific T cells were in the CXCR3+CCR6− and CXCR3+CCR6+ subsets, and Candida albicans-specific T cells were most prominent in the CCR4+CCR6+ subset, enriched in Th17 cells, but positive cultures were also detected in libraries from subsets enriched in Th1, Th2 and Th22 cells (Figure 6A and B). The narrow distribution of antigen-responding T cells in the CXCR3+CCR6+ subset was peculiar to MTB since Streptococcus pyogenes- or Staphylococcus aureus-specific T cells were found in both CXCR3+CCR6+ and CCR4+CCR6+ subsets (not shown).
Based on these results, we sorted three memory CD4 Th cell subsets (Figure 7A and B): 1) CCR6+CXCR3−, accounting for 24.1% (21.8–27.0) (median (interquartile range), n = 4) of the memory CD4+ T cell pool; 2) CCR6+CXCR3+ (32.0% (28.0–32.4)) and 3) CCR6− (37.0% (34.4–42.0)). For each donor a T cell library of 288 cultures was established. MTB-responding T cells were highly enriched in cultures derived from the CCR6+CXCR3+ T cell subset, and present at much lower frequency in the CCR6+CXCR3− and the CCR6− subsets (Figure 7C). This pattern of distribution was remarkable consistent: in all 4 donors analyzed more than 80% of the MTB-reactive memory CD4 T cell response resided in the CXCR3+CCR6+ subset (Figure 7D).
Next, we set up T cell libraries from 4 representative donors and the CXCR3+CCR6+ subset were directly stimulated, after expansion, with 59 representative peptide pools. The results of this analysis are shown in figure 8. Using this approach we were able to demonstrate that the results obtained with the MTB lysate also extended to responses specific for the various epitopes, and to confirm with a complementary approach the results of the ex vivo IFN-γ ELISPOT analysis utilizing the library of predicted HLA class II binding epitopes.
Individual MTB proteins have been studied to identify novel vaccine candidates, with several studies focused on culture filtrate proteins [6], [7]. Other studies utilized bioinformatic approaches to select a subset of genes as antigen candidates [37], [38]. However, the lack of a true genome-wide characterization has hindered a complete understanding of the mechanisms and specificity of the immune response to MTB.
This study provides the first in-depth truly genome-wide description of human T cell responses to MTB. We characterized and isolated T cells directly ex vivo, thus avoiding biases introduced as a result of in vitro restimulation and expansion of T cells before analysis. This approach should be generally applicable to the study of immunity to other complex pathogens. The HLA alleles were chosen to allow coverage of the most frequent DP, DQ and DR specificities in the general population [39]. However, we readily acknowledge that this selection has potential limitations and may bias the results toward the epitopes recognized by these alleles.
In terms of T cells recognizing MTB we found that the T cell response to MTB antigens in LTBI donors is strongly biased towards a subset of CXCR3+ Th1 cells that co-express CCR6 [4]. Interestingly, this narrow distribution was only seen for MTB and not other pathogens such as S. pyogenes and C. albicans within the same donor. The origin of CCR6+ Th1 cells and their differentiation requirements remains to be defined; they may represent a separate Th1 lineage, or they may differentiate from plastic CCR6− Th1 cells or CCR6+ Th17 cells [40]. Future studies will examine whether this highly focused response is key to MTB containment by examining patients who remain healthy vs. patients who progress to active disease.
Striking levels of heterogeneity of responses were detected. This expands previous observations using smaller subsets of antigens [6], [41], and a genome-wide screen of antibody responses [28]. The observed heterogeneity might reflect differences in MTB strains, bacillary load, and metabolic state, resulting in qualitative or quantitative differences in antigen expression [42], [43]. In any case, since natural immunity to MTB is multiepitopic and multiantigenic, and more than 80 antigens are necessary to capture 80% of the T cell response, vaccination strategies including one or a few antigens are unlikely to replicate natural immunity. Likewise, monitoring the immune response to one or a few antigens in the setting of clinical trials might yield a severely incomplete and biased picture of immune reactivity.
Several antigens from the DosR regulon, as well as resuscitation- and reactivation-associated antigens have been described as preferentially recognized by individuals with latent infection using long-term T cell cultures [44], [45]. We observed reactivity to two of these proteins, Rv2031c (2 donors) and Rv2627c (1 donor), and no significant association with proteins from the DosR regulon or latency-associated antigens, similar to previous observations [46].
Numerous tuberculosis vaccine candidates are currently in clinical trials, these candidates are based on 11 MTB antigens, 7 of which were prevalently recognized in this study; Rv3804c in MVA85A [47] and Aeras-402 [48], Rv1886c in Aeras-402, H1 [49] and HyVac4 [50], Rv0288 in Aeras-402 and HyVac4, Rv3875 in H1 and H56 [51], Rv1196 in Mtb72f/AS02A [52], Rv2608 and Rv3619 in ID93 [53]. Of the remaining 4 antigens Rv3620c in ID93 was also recognized whereas Rv2660c in H56, Rv0125 in Mtb72f/AS02A and Rv1813c in ID93 were not.
We identified three antigenic islands within the MTB genome map as main determinants of immunodominance. Remarkably, the majority of the novel antigens identified are associated (contained within or in close proximity to) these islands, which all contain Esx protein pairs and PE/PPE proteins, and are part of a putative secretory system. Our analysis demonstrated that these factors synergistically contribute to determining immunodominance and confirms the importance of PE/PPE and Esx proteins, but suggests that their immunodominance is perhaps mostly determined by their location within these antigenic islands.
Two main hypotheses can be put forth to explain the mechanism by which these features determine immunodominance. First, secreted proteins may act as decoys to divert the immune response from recognizing nonsecreted MTB proteins [13], thus favoring bacterial persistence. The second hypothesis envisions that antigenic islands are dominant because they are intrinsically immunogenic, and because they perform key biological functions necessary to maintain MTB persistence.
The decoy hypothesis has two predicated features; either secreted proteins result in diversion of the immune system from the bacteria itself, or the decoy effect is achieved by inducing an immune response to decoy antigens that are functionally distinct from non-decoy antigens. In the first case, we note that both secreted proteins and proteins involved in the secretion apparatus are equally recognized. Indeed, immune reactivity towards proteins involved in the secretion system apparatus has previously been described for T3SS and inflammasome activation by flagellin and the T3SS rod proteins [54], [55]. Furthermore, we were unable to detect a functionally distinct immune response in terms of multifunctionality, memory phenotype and T cell subsets, and independent of island vs. non-island localization and secretion status of the antigen recognized. Taken together, these observations argue against the decoy hypothesis. T cells that secrete multiple cytokines are a potential correlate of protection, but have also been implicated in pathology [31], [32], [33], [34]. Whatever their role might be, the majority of epitope-specific CD4+ T cell responses were multifunctional, with no differences between antigens from islands vs. non-islands, and between the PE/PPE vs. cell wall and cell processes categories. In terms of T cell phenotypes and T cell subsets a similar picture emerged, with epitope specific CD4+ T cells being predominantly CD45RA−CCR7+ central memory cells, in agreement with previous studies [26]. For some epitope specific CD4+ T cells a large fraction were CD45RA+CCR7+, a phenotype traditionally regarded as naïve. Such T cells have previously been reported [56], [57], and might reflect early differentiation into antigen-specific cells. Additional studies would be required to investigate this further.
The available data favors the second hypothesis, that the three antigenic islands are dominant because they perform key biological functions and are necessary to maintain MTB persistence. The most prevalently recognized island is Esx-3, which is controlled by the iron-dependent regulator IdeR and the zinc uptake regulator Zur [58], [59], suggesting its involvement in fundamental biological processes such as metal iron homeostasis. In addition, Esx-3 is essential for in vitro growth, and is conserved in a wide range of mycobacterial species [20], [60]. Furthermore, the Esx-3 system contributes to immune protection against MTB challenge in mice of the IKEPLUS strain [61] in a HLA class II dependent fashion.
Genes from island 2 are, like island 1 (Esx-3), regulated by Zur [58], providing a possible functional link between them. While island 2 is not part of an Esx secretion system per se, it is believed to originate from a duplication of the Esx-3 system [19]. Esx-1 and Esx-3 also appear to be linked, since Rv3873 interacts with Rv0288 [62].
Secretion systems similar to the T7SS associated with two of the three antigenic islands are also found in other bacteria, such as Listeria monocytogenes, S. aureus, and Bacillus anthrax [63]. Secretion of the substrates from T7SS are not dependent on interaction with host cells, unlike other bacterial secretion systems such as T3SS and T4SS, which are switched on upon host-cell contact. This suggests that T7SS, while essential for pathogenicity, may fulfill more general physiological roles than strictly host-cell oriented functions.
This study was completed in a non-TB-endemic population. Ongoing studies include a larger study population from different ethnicities and geographic locations, as well as patients with different disease states and BCG vaccination status. This will provide answers for different HLA phenotypes, as well as whether patients from an endemic area or with different disease states show a different recognition pattern.
In conclusion, this study describes the immunological footprint of MTB CD4 T cell recognition to an unprecedented level of detail. The high throughput cellular screens utilized here to analyze the human immune response to MTB provides information on the specificity, frequency and class of memory T cells, as well as on the individual variability in magnitude and quality of the response. As a result, 34 novel antigens and three broadly immunodominant antigenic islands were defined. The study of the class of proteins recognized, together with the phenotype of responding T cells, disproves the notion that responses against secreted antigens are a decoy utilized to favor bacterial persistence, and rather suggest that these proteins, together with those that are part of their general secretion apparatus, are targeted by fully functional T cell responses. More broadly, this study provides proof of principle of how such high throughput techniques can be applied to other complex pathogen systems. In terms of potential practical applications, the novel T cell antigens identified could be of potential use for diagnostic or vaccine purposes. Indeed, the heterogeneity of responses demonstrated herein suggests that a too narrow focus for vaccine evaluations will not replicate natural immunity. Finally, the antigens and epitopes identified can also be used as tools for identifying biomarkers to provide correlates of risk for, or protection against, tuberculosis disease.
Research conducted for this study was performed in accordance with approvals from the Institutional Review Board at the La Jolla Institute for Allergy and Immunology. All participants provided written informed consent prior to participation in the study.
Leukapheresis samples from 28 adults with LTBI and 28 control donors were obtained from the University of California, San Diego Antiviral Research Center clinic (age range 20–65 years). Subjects had a history of a positive tuberculin skin test (TST). LTBI was confirmed by a positive QuantiFERON-TB Gold In-Tube (Cellestis), as well as a physical exam and/or chest X-ray that was not consistent with active tuberculosis. None of the study subjects endorsed vaccination with BCG, or had laboratory evidence of HIV or Hepatitis B. The control donors had a negative TST, as well as a negative QuantiFERON-TB. Approval for all procedures was obtained from the Institutional Review Board (FWA#00000032) and informed consent was obtained from all donors.
Proteins from the 21 MTB genome projects available from the NCBI Protein database were downloaded into an in-house MySQL database. Of these, 5 were complete (CDC1551, F11, H37Ra, H37Rv, KZN 1435) and 16 were draft assemblies (Table S1). The protein sequences were parsed into all possible 15mer peptides (n = 1,568,148), for each of which binding to 22 different HLA DR, DP and DQ class II alleles most commonly expressed in the general population (Table S2) was predicted using the IEDB HLA class II ‘consensus’ prediction method [64]. The sequences of the H37Rv strain were used as a reference sequence. For each H37Rv protein, alignments were made of all orthologs identified in other genomes, as determined by a BLAST search. Because of the overall high sequence conservation among the proteins from all the 21 genomes, 1,220,829 (91.4%) of 15mers were completely conserved among all of the strains. For each protein, the best-predicted binders, as ranked by consensus percentile, were selected for synthesis. In order to ensure coverage of each of the proteins, the number of peptides selected per protein was no less than 2 and no more than 10, depending upon protein length (18,950 peptides). Any variants among the orthologs at the selected positions were also selected (1,660), for a total of 20,610 peptides.
Sets of 15-mer peptides synthesized by Mimotopes (Victoria, Australia) and/or A and A (San Diego) as crude material on a small (1 mg) scale were combined into pools of 20 peptides. Peptides utilized for tetramers were synthesized as purified material (>95% by reversed phase HPLC). The IEDB submission number for the peptides is 1000505.
PBMCs were obtained by density gradient centrifugation (Ficoll-Hypaque, Amersham Biosciences) from 100 ml of leukapheresis sample, according to manufacturer's instructions. Cell were suspended in fetal bovine serum (Gemini Bio-products) containing 10% dimethyl sulfoxide, and cryo-preserved in liquid nitrogen.
CD4 T cells were isolated from PBMCs by positive selection with microbeads (Miltenyi Biotec). Memory CD4+ T cell subsets were sorted with a FACSAria (BD Biosciences) to over 98% purity excluding CD45RA+, CD25+, CD8+, CD19+, and CD56+ cells. Antibodies used for positive selection were: anti-CCR6-PE or biotinylated (11A9; BD Biosciences) followed by streptavidin-allophycocyanin (APC) (Invitrogen) or streptavidin-APC-cyanine7 (APC-Cy7) (BD Biosciences); anti-CCR10-PE (314305, R&D Systems), anti-CCR4-PE-Cy7 (1G1; BD Pharmingen) and anti-CXCR3-APC (1C6; BD Pharmigen). Cells were cultured in RPMI 1640 medium supplemented with 2 mM glutamine, 1% (vol/vol) nonessential amino acids, 1% (vol/vol) sodium pyruvate, penicillin (50 U/ml), streptomycin (50 µg/ml) (all from Invitrogen) and 5% heat-inactivated human serum (Swiss Red Cross). T cells (1,000 cells/well) were stimulated polyclonally with 1 µg/ml PHA (Remel) in the presence of irradiated (45 Gy) allogeneic feeder cells (1.0×105 per well) and IL-2 (500 IU/ml) in a 96-well plate format and T cell lines were expanded as previously described [26]. Library screening was performed at day 14–21 by culturing extensively washed T cells (∼2.5×105/well) with autologous monocytes (2.5×104), either unpulsed or pulsed for 3 h with MTB whole cell lysate (5 µg/ml, BEI Resources) or control antigens. In some experiments, T cells were cultured with peptide pools (2 µg/ml). Proliferation was measured on day 2–3 after 16 h incubation with 1 µCi/ml [methyl-3H]-thymidine (Perkin Elmer). Precursor frequencies were calculated based on numbers of negative wells according to the Poisson distribution and expressed per million cells.
PBMCs incubated at a density of 2×105 cells/well were stimulated with peptide pools (5 µg/ml) or individual peptides (10 µg/ml), PHA (10 µg/ml) or medium containing 0.25% DMSO (corresponding to percent DMSO in the pools/peptides, as a control) in 96-well plates (Immobilon-P; Millipore) coated with 10 µg/ml anti-IFN-γ (AN18; Mabtech). Each peptide or pool was tested in triplicate. After 20 h incubation at 37°C, wells were washed with PBS/0.05% Tween 20 and incubated with 2 µg/ml biotinylated anti-IFN-γ (R4-6A2; Mabtech) for 2 h. The spots were developed using Vectastain ABC peroxidase (Vector Laboratories) and 3-amino-9-ethylcarbazole (Sigma-Aldrich) and counted by computer-assisted image analysis (KS-ELISPOT reader, Zeiss). Responses were considered positive if the net spot-forming cells (SFC) per 106 were ≥20, the stimulation index ≥2, and p<0.05 (Student's t-test, mean of triplicate values of the response against relevant pools or peptides vs. the DMSO control). For experiments utilizing depletion of CD4+ or CD8+ T cells, these cells were isolated by positive selection (Miltenyi Biotec) and effluent cells (depleted cells) were used for experiments.
The response frequency was calculated by dividing the no. of donors responding with the no. of donors tested. The magnitude of response (total SFC) was calculated by summation of SFC from responding donors.
PBMCs were cultured in the presence of 5 µg/ml MTB peptide and 4 µl/ml Golgiplug (BD Biosciences) in complete RPMI medium at 37°C in 5% CO2. Unstimulated PBMCs were used to assess nonspecific/background cytokine production. After 6 h, cells were harvested and stained for cell surface antigens CD4 (anti-CD4-PerCPCy5.5, OKT-4) and CD3 (anti-CD3-EFluor450, UCHT1). After washing, cells were fixed and permeabilized, using a Cytofix/Cytoperm kit (BD Biosciences) and then stained for IFN-γ (anti-IFN-γ-APC, 4S.B3), TNFα (anti-TNFα-FITC, MAb11) and IL-2 (anti-IL-2-PE, MQ1-17H12). All antibodies were from eBioscience. Samples were acquired on a BD LSR II flow cytometer. The frequency of CD4+ T cells responding to each MTB peptide was quantified by determining the total number of gated CD4+ and cytokine+ cells and background values subtracted (as determined from the medium alone control) using FlowJo software (Tree Star). A cut-off of 2 times the background was used. Combinations of cytokine producing cells were determined using Boolean gating in FlowJo software.
HLA class II tetramers conjugated using PE labeled streptavidin were provided by the Tetramer Core Laboratory at Benaroya Research Institute. CD4 T cells were purified using the Miltenyi T cell isolation kit II according to manufacturer's instructions. Purified cells (∼10×106) were incubated in 0.5 ml PBS containing 0.5% BSA and 2 mM EDTA pH 8.0 (MACS buffer) with a 1∶50 dilution of class II tetramer for 2 h at room temperature. Cells were then stained for cell surface antigens using anti-CD4-FITC (OKT-4), anti-CD3-Alexa Fluor 700 (OKT3), anti-CCR7-PerCPEFluor710 (3D12), anti-CD45RA-EFluor450 (HI100) (all from EBioscience) and Live/Dead Yellow (Life Technologies) to exclude dead cells. Tetramer-specific T cell populations were enriched by incubating cells with 50 µl of anti-PE microbeads (Miltenyi Biotech) for 20 min at 4°C. After washing, cells were resuspended in 5 ml MACS buffer and passed through a magnetized LS column (Miltenyi Biotec). The column was washed three times with 3 ml of MACS buffer, and after removal from the magnetic field, cells were collected with 5 ml of MACS buffer. Samples were acquired on an BD LSR II flow cytometer and analyzed using FlowJo software.
The identified epitopes were compared for sequence homology and the weakest epitopes sharing >90% homology were eliminated. The epitopes were mapped to the H37Rv genome allowing 1 substitution per peptide, to identify antigens. IEDB queries utilized criteria matching the experimental study (organism; MTB, host organism; human, latent disease, ex vivo, HLA class II). Epitopes were then mapped as above. To capture the most frequently recognized antigens the response frequency score (no. donors responded – Square root of no. donors responded/no. donors tested), was utilized [65].
|
10.1371/journal.pgen.1002638 | Hypoxia Disruption of Vertebrate CNS Pathfinding through EphrinB2 Is Rescued by Magnesium | The mechanisms of hypoxic injury to the developing human brain are poorly understood, despite being a major cause of chronic neurodevelopmental impairments. Recent work in the invertebrate Caenorhabditis elegans has shown that hypoxia causes discrete axon pathfinding errors in certain interneurons and motorneurons. However, it is unknown whether developmental hypoxia would have similar effects in a vertebrate nervous system. We have found that developmental hypoxic injury disrupts pathfinding of forebrain neurons in zebrafish (Danio rerio), leading to errors in which commissural axons fail to cross the midline. The pathfinding defects result from activation of the hypoxia-inducible transcription factor (hif1) pathway and are mimicked by chemical inducers of the hif1 pathway or by expression of constitutively active hif1α. Further, we found that blocking transcriptional activation by hif1α helped prevent the guidance defects. We identified ephrinB2a as a target of hif1 pathway activation, showed that knock-down of ephrinB2a rescued the guidance errors, and showed that the receptor ephA4a is expressed in a pattern complementary to the misrouting axons. By targeting a constitutively active form of ephrinB2a to specific neurons, we found that ephrinB2a mediates the pathfinding errors via a reverse-signaling mechanism. Finally, magnesium sulfate, used to improve neurodevelopmental outcomes in preterm births, protects against pathfinding errors by preventing upregulation of ephrinB2a. These results demonstrate that evolutionarily conserved genetic pathways regulate connectivity changes in the CNS in response to hypoxia, and they support a potential neuroprotective role for magnesium.
| How hypoxia damages the developing human brain is poorly understood, despite being a major cause of life-long neurologic and psychiatric problems. Premature infants are especially at risk for these problems, with increased rates of attention-deficit disorder, autism, cerebral palsy, epilepsy, psychiatric disorders, and cognitive impairment. It is unknown whether hypoxia can cause errors in the connections of neurons in the vertebrate nervous system. We used zebrafish, a vertebrate model animal, to answer this question. We found that hypoxic injury causes errors in how neurons connect. We went on to determine that a specific genetic pathway, the hif1 pathway, is activated by hypoxia and turns on downstream genes, which cause the connection problems. One of the genes activated by hif1, ephrinB2a, is responsible for many of the connection problems. Importantly, magnesium, used as a treatment for some preterm births, is able to help protect against the neuron connection errors. Our results show that hypoxia in vertebrates does cause errors in neuron connections and that magnesium can help prevent this.
| Hypoxic injury to the developing human brain is a major cause of both acute and chronic neurodevelopmental impairments. Premature infants, particularly those characterized by very-low birth weights (VLBW; less than 1,500 g) are the population at greatest risk for chronic hypoxic injury and for adverse neurocognitive outcomes [1]. Causes of hypoxia in these infants include chronic lung disease, pulmonary hypertension, congenital heart disease, and placental insufficiency. Up to 35% of VLBW infants experience neurodevelopmental impairments including attention-deficit disorder, autism, cerebral palsy, epilepsy, psychiatric disorders, and mental retardation/cognitive impairment [2]–[5]. While survival rates have improved dramatically for premature infants, neurodevelopmental outcomes have not [6], [7]; in fact, the total number of VLBW infants has increased over the past decade [8].
Strategies to protect against the effects of prematurity and chronic hypoxic injury to the central nervous system (CNS) have been limited since the pathophysiology that leads to the adverse neurodevelopmental outcomes is uncertain. Indirect measures in humans have shown altered connectivity in the brains of children born prematurely [9], [10], and the period between 20 weeks gestation and term birth (40 weeks gestation) is a period of extensive commissural and projection axon extension [11]. Most VLBW infants are born between 23 and 28 weeks gestation, and both pre-term and term infants experience episodic, often unrecognized, hypoxemia [12].
The mechanisms by which hypoxia disrupts connectivity in vertebrates are not known. Recent work in the invertebrate C. elegans has shown that hypoxia causes discrete axon pathfinding errors in certain interneurons and motorneurons by increased expression of the Eph receptor vab-1 [13]. However, it is unknown whether similar pathfinding errors, and similar genetic pathways and molecular mechanism, also occur in the more complex vertebrate CNS.
We hypothesized that hypoxia might specifically affect axon pathfinding and the development of CNS connectivity in vertebrates. To test this we developed a zebrafish (Danio rerio) model for investigating hypoxia at different stages of CNS development, and generated transgenic lines to test the genetic pathways involved. We analyzed both genetic and chemical modifiers of the molecular response of hypoxia on pathfinding, including cloning zebrafish hif1α and generating a constitutively active form of it. We found that hypoxia disrupts axon pathfinding in vertebrates through an evolutionarily conserved mechanism, by activation of the hif1α pathway and increased expression of ephrinB2. Further, we tested the role of magnesium sulfate, which is known to improve neurodevelopmental outcomes when given to mothers of infants at risk for premature delivery [14]. We found that magnesium sulfate helps reduce hypoxia-induced upregulation of ephrinB2, and decreases the frequency of pathfinding errors.
We developed a system in zebrafish to examine the effects of hypoxia on the development of CNS connectivity. We used a small, airtight plexiglass chamber in which zebrafish embryos were placed at different developmental stages. Hypoxia was induced by use of a digital controller that regulated nitrogen gas flow, and we pre-equilibrated solutions to either normoxia or hypoxia for at least 4 hours before use. Previous reports have shown that early zebrafish embryos tolerate anoxia, especially before 24 hours post-fertilization (hpf), but the anoxia can slow development [15], [16]. Because we wanted to model hypoxic insults to CNS development, we tested a wide range of hypoxic conditions at different times to determine the degree of hypoxia that embryos could tolerate with minimal mortality (Table 1). We assessed mortality at 72 hpf, at which point major stages of CNS development have occurred, including neurogenesis, cell-type specification, axon pathfinding, and synaptogenesis, following either 12 or 24 hour periods of hypoxia, using at least 30 embryos per condition. We found that zebrafish embryos can tolerate long periods of stringent hypoxia, especially at earlier developmental stages. Because we noted dysmorphic development in embryos exposed to 0.5% hypoxia (0.5% pO2) or lower, we used 1% hypoxia (1% pO2; normoxia is 21% pO2) for experiments.
To visualize effects of hypoxia on pathfinding, we evaluated several different transgenic lines as well as a pan-axonal antibody against acetylated tubulin. Using anti-acetylated tubulin, we found at low frequency subtle pathfinding errors. Further, in less than 5% of embryos we found severe axon pathfinding errors, most often of commissural axon tracts (Figure 1A, 1B). However, because of the low frequency of severe errors and difficulty visualizing and quantifying the subtle errors using anti-acetylated tubulin, we generated a transgenic line in which a small subset of axons expressed membrane-targeted GFP: Tg(foxP2-enhancerA.2:egfp-caax). This line precisely labeled a few distinct neuron types, with retinal, commissural, and longitudinal axons [17] (Figure 1C).
We examined axon tracts in Tg(foxP2-enhancerA.2:egfp-caax) embryos following 12-hour periods of hypoxia, and found a reproducible defect in formation of the tract of the commissure of the posterior tuberculum (TCPT) [18]. This defect was never seen in wild-type embryos (Figure 1D), and there was no overall difference in fluorescence intensities of normoxic compared to hypoxic transgenic embryos. Following hypoxia, the TCPT commissure (TCPTc) either did not form, or had fewer axons crossing.
We noted significant variability in the severity of the phenotype, with some embryos lacking the TCPTc, some with a normal appearance to the TCPTc, and some embryos with intermediate phenotypes. The intermediate phenotypes were embryos in which the number of TCPTc axons was reduced. The percentage of embryos with loss or near-complete absence of the TCPTc following 1% pO2 from 24–36 hpf was 60% (Figure 1E), with no significant effects of hypoxia when examined at other time-points. Because of the potential for subjectivity in scoring intermediate phenotypes of disrupted TCPTc, and in order to increase our ability to detect more subtle changes in the number of TCPTc axons crossing, we developed a quantitative measure of TCPTc axon crossing. We compared fluorescence intensity ratios of commissural to longitudinal axons of the TCPT axons, in normoxic versus hypoxic embryos (C/L ratio; Figure 1F; Methods). The average ratio for normoxic embryos was 0.639, whereas for hypoxic embryos (1% pO2 from 24–36 hpf) the average was statistically different at 0.225 (p<0.0001 for two-tailed t test) (Table 2). Thus, a decrease in the C/L ratio represents a decrease in the number of axons crossing in the TCPTc.
We examined the effects of hypoxia exposure at different developmental stages (Figure 1G), with at least 24 embryos for each period, and analysis at 72 hpf. We found that hypoxia during 24 36 hpf disrupted TCPTc formation, as shown by the statistically significant decrease in the C/L ratio. Increased duration of hypoxia up to 36 hours did not worsen the C/L ratio. Further, when we analyzed the TCPTc at 96 hpf, following hypoxia from 24–36 hpf, there was persistent failure of TCPTc crossing (Figure 1H, 1I). Our hypoxia conditions for 12 hours were therefore followed by either 36 or 60 hours of recovery in normoxia. This demonstrates that the pathfinding errors are not due to a simple maturational delay in axon extension, and that the TCPT axons do not then re-cross the midline.
To determine whether the observed pathfinding phenotype following hypoxia correlates with the timing of TCPTc formation, we examined normal development of TCPT axons. The first axons project by 24 hpf (Figure 2A), and by 36 hpf axons are crossing the midline and forming the TCPTc (Figure 2B). Therefore, our observation that maximal effects of hypoxia occur when embryos are exposed from 24 to 36 hpf is consistent with the timing of axon pathfinding.
We then examined the fate of the aberrant axons following hypoxia. Normally, the TCPT axons in Tg(foxP2-enhancerA.2:egfp-caax) embryos split (Figure 2C) into commissural and longitudinal portions, with the majority of axons crossing. In contrast, following hypoxia, most axons fail to cross the midline and turn to aberrantly follow a longitudinal pathway (Figure 2D). High-resolution pictures of the TCPTc following hypoxia (Figure 2E–2G′″) show that the TCPTc axons destined to cross the midline arise dorsally and remain tightly bundled as they cross the midline (Figure 2E–2E″). In contrast, following hypoxia, the TCPT axons split, with some crossing the TCPTc, whereas others turn caudally and join the longitudinal, more ventral tracts (Figure 2F–2G′″). Thus, the TCPTc hypoxia phenotype is characterized by axons specifically turning to follow an erroneous pathway.
Possible explanations for the observed disruption in axon pathfinding could be a general effect on CNS patterning; on neuron specification; or on apoptosis or cellular proliferation. To address these possibilities, we compared hypoxic embryos to normoxic embryos using a variety of cellular markers (Figure 3). We did not find any significant changes in a variety of markers, including dlx2 for CNS forebrain and diencephalon patterning, TH antibody staining for cell-type specification of tyrosine hydroxylase-expressing neurons, acridine orange for apoptosis, and anti-phosphohistone H3 antibody for proliferation. We quantified apoptosis at 72 hpf in a region of the telencephalon that includes TCPT neuron cell bodies, and found no significant difference between normoxic and hypoxic embryos (cell counts 6.5 in normoxia, 7.1 in hypoxia, standard deviation 4.6, n = 30 embryos, p = 0.6; Figure 3; Methods).
Cellular responses to hypoxia are coordinated by activation of the hif1 pathway [19]. hif1α is a basic helix-loop-helix transcription factor ubiquitously expressed, but which is normally hydroxylated and degraded under normoxic conditions. In hypoxia hif1α hydroxylation is inhibited, and hif1α is able to activate a downstream genetic pathway of target genes that modify an organism's response to hypoxia, for example, by increased angiogenesis, [20]. We wished to determine whether hif1 pathway activation was mediating the TCPTc pathfinding errors. First, we wanted to establish whether hif1 pathway activation was occurring from our hypoxia model. We decided to examine expression of igfbp-1, a known downstream transcriptional target of hif1 pathway activation from hypoxia in vertebrates, including zebrafish and humans [21]–[24]. Following hypoxia from 24–36 hpf, igfbp-1 expression was increased (Figure 4A, 4B). To demonstrate a role for hif1α, we used dimethyloxaloglycine (DMOG) to activate hif1α by inhibition of prolyl hydroxylase or factor inhibiting hypoxia-inducible factor [25], in the absence of hypoxia. Normoxic embryos were exposed to varying DMOG amounts from 24–36 hpf. Increasing amounts of DMOG led to increasing expression of igfbp-1 (Figure 4C); and increased pathfinding errors of the TCPTc (n>26 embryos for all conditions) (Figure 4D; Table 2). Further, an inhibitor of hif1α transcription CAY10585 [26] was able to reduce the C/L ratio in hypoxia (Figure 4K). These results suggest that the TCPT commissure errors due to hypoxia are caused by activation of the hif1 pathway.
To determine whether hif1α's role on the TCPT commissure is cell-autonomous, we decided to misexpress hif1α or hif1αmut, a constitutively active form of hif1α, using the UAS/GAL4 system [27]. We cloned the zebrafish hif1α cDNA, and then made hif1αmut by changing proline 621 to alanine in the conserved LXXLAP motif of HIF-1α [28]. We generated stable transgenic lines expressing hif1α or hif1αmut downstream of UAS, with a viral 2A peptide fused to GFP-caax or RFP-caax to generate a bicistronic message [29]. The use of the bicistronic message allowed us to monitor whether expression of hif1α or hif1αmut was occurring from the UAS, by the presence of fluorophore expression. We drove expression either in the TCPT neurons using Tg(foxP2-enhancerA.2:Gal4-VP16), or pan-neuronally using Tg(elavl3:Gal4-VP16) (Figure 4E–4K). We did not observe any TCPTc pathfinding errors when we expressed either hif1α or hif1αmut in the TCPT neurons (n>30 embryos for each genotype). In contrast, when we expressed hif1αmut but not hif1α using Tg(elavl3:Gal4), we observed TCPTc errors (Figure 4G, 4H). For both hif1α and hif1αmut we screened and isolated two different independent UAS transgenic alleles each, and examined fluorescence expression when crossed to a Gal4-driver line; all subsequent experiments were then based on use of a single allele with robust expression. Quantification of C/L ratios in Tg(elavl3:Gal4); Tg(foxP2-enhancerA.2:egfpcaax); Tg(UAS: hif1αmut-2A-TagRFP) embryos (n = 25 embryos) compared to Tg(elavl3:Gal4); Tg(foxP2-enhancerA.2:egfpcaax); Tg(UAS:hif1α-2A-TagRFP) embryos (n = 10) was statistically significant (Figure 4K). We confirmed that the hypoxia pathway is activated by expression of hif1αmut by showing that igfbp-1 is up-regulated (Figure 4I, 4J).
To try to further localize the site of action of hypoxia pathway activation, we expressed hif1αmut in neurons neighboring the TCPT axons. We crossed Tg(otpb.A:Gal4-VP16) [30] to Tg(foxP2-enhancerA.2:egfpcaax), and injected embryos with a plasmid carrying UAS:hif1αmut-2A-TagRFP. In these embryos hif1αmut is expressed neighboring the TCPT axons as they extend longitudinally, prior to decussating. We did not observe any defects in formation of the TCPTc (0%, n = 50 embryos, Figure S1). While we can not exclude the possibility that levels of hif1αmut were insufficient in these experiments to disrupt pathfinding, we think it is more likely that (if a suitable Gal4 line were available) expression targeted to the CNS midline and/or at the site of the decussation would be effective. This is based on the expression pattern of a potential receptor guiding TCPTc axons (below, FFure 6). Thus, the pathfinding errors in TCPT axons resulting from hypoxia can be mimicked by activation of the hif1 pathway pan-neuronally, but not in TCPT neurons alone.
Our finding that hif1 pathway activation has a non-cell-autonomous effect on TCPT axon pathfinding suggested that hypoxia affected pathfinding through effects on cell-cell signaling. The Eph-ephrin signaling system has conserved roles in axon pathfinding in both vertebrates and invertebrates and controls aspects of commissural axon pathfinding [31], [32]. Further, in C. elegans hypoxia upregulates the ephrin receptor vab-1, while knockdown of its ligand efn-2 prevents hypoxia pathfinding defects [13]. ephrinB2a is the zebrafish gene with greatest sequence conservation to efn-2 and further is expressed in telencephalic neurons in the zebrafish embryonic CNS [33].
TCPT neurons and axons express ephrinB2a during 24–36 hpf (Figure 5A), including in the TCPTc (Figure 5B). Hypoxia caused increased expression of ephrinB2a (Figure 5C–5E; Figure 6). Knock-down of ephrinB2a using a translation-blocking morpholino [33] led to loss of ephrinB2a expression (Figure 6A–6D), and protected against hypoxia pathfinding errors in the TCPT axons (Figure 5F, 5G). Further, injection of ephrinB2a morpholino into Tg(elavl3:Gal4); Tg(foxP2-enhancerA.2:egfpcaax); Tg(UAS:hif1αmut-2A-TagRFP) embryos led to rescue of TCPTc errors caused by misexpression of constitutively active hif1αmut (Figure 4K). These results suggest that ephrinB2a mediates some of the disruptive effects of hypoxia on pathfinding.
To determine how signaling by ephrinB2a is necessary for the pathfinding by TCPT axons, we expressed a form of ephrinB2a lacking the cytoplasmic signaling domain, UAS:ephrinB2a(Δcytoplasmic) [34] in TCPT neurons using Tg(foxP2-enhancerA.2:Gal4). Transient injections of UAS:ephrinB2a(Δcytoplasmic), but not full-length UAS:ephrinB2a, prevented hypoxia-induced pathfinding errors (Figure 5H; Table 2). There was no effect on the TCPTc under normoxic conditions by expressing either truncated or full-length ephrinB2a (Table 2). This demonstrates that ephrinB2a lacking its cytoplasmic domain is able to decrease guidance errors cell-autonomously, perhaps by acting as a dominant negative to interfere with signaling coming from its “ligand” receptor-tyrosine kinase EphA4a or EphB4. This is consistent with ephrinB2a using a “reverse” signaling mechanism for TCPTc pathfinding [35].
We examined the pathfinding of TCPTc axons relative to the expression of ephA4a (Figure 6E–F″), a known receptor for ephrinB2a, that can act during midline commissure formation. We found that the TCPTc axons travel along the edge of the midline expressing ephA4a during the period of initial axon extension at 24 hpf (Figure 6E–6E″). As the TCPTc forms, the axons cross the midline, again avoiding the ephA4a expression domain (Figure 6F–6F″). These results suggest that the ephrinB2a-expressing TCPTc axons are responding to a signal from the ephA4a-expressing midline cells. Hypoxic up-regulation of ephrinB2a may disrupt the normal balance of signaling between the ligand/receptor pair of ephrinB2a/ephA4a, and thereby disrupt TCPTc formation.
Few neuroprotective agents have been demonstrated to affect neurodevelopmental outcomes of premature infants. Administration of magnesium sulfate to mothers of infants at risk for premature delivery improves neurodevelopmental outcomes and reduces rates of cerebral palsy [14], although the mechanism is unknown. We sought to determine whether magnesium could reduce pathfinding errors in hypoxic conditions. Embryos tolerated magnesium sulfate concentrations ranging from 25 to 250 mM during 24 to 36 hpf with no apparent morphologic defects, although above 250 mM there was approximately 30% mortality. Increasing concentrations of magnesium led to rescue of hypoxia TCPTc pathfinding errors (Figure 7A). We determined that magnesium can reduce activation of the hif1 pathway, assayed by in situ for igfbp-1 (Figure 7B, 7C). However, magnesium was unable to rescue commissure errors in embryos expressing activated hif1αmut. Namely, in Tg(elavl3:Gal4); Tg(foxP2-enhancerA.2:egfpcaax); Tg(UAS:hif1αmut-2A-TagRFP) embryos, magnesium did not rescue pathfinding (Figure 4K). Magnesium did normalize Efn levels following hypoxia (Figure 6). Interestingly, magnesium was able to rescue pathfinding errors in the presence of DMOG (Figure 7D). These results show that magnesium does rescue the effects of hypoxia on pathfinding. Further, given the results with DMOG and hif1αmut, it suggests that magnesium's primary mode of action might be through the prolyl hydroxylase or factor inhibiting hypoxia-inducible factor (FIH) pathways.
The increasing numbers of preterm births and associated adverse neurodevelopmental outcomes has highlighted the limited understanding of basic mechanisms underlying this problem [6]–[8]. Indirect radiological evidence in infants born prematurely shows altered CNS connectivity [9], [10], and neurodevelopmental disorders such as autism are associated with changes in axonal and synaptic gene expression [36].
We investigated whether hypoxia, a known complication of prematurity, can disrupt axon pathfinding. We found that hypoxia disrupts connectivity in the vertebrate CNS, leading to precise and reproducible errors in axon pathfinding. The disruption of pathfinding we observed is not due to apoptosis or broad effects on CNS development and specification. We have found that the pathfinding errors are due to hypoxia activation of the hif1 pathway, and can be mimicked by chemical activators of hif1α, or by misexpression of a constitutively active form of hif1α. Further, the pathfinding errors are mediated by ephrinB2a reverse signaling, and can be rescued by knock-down of ephrinB2a. Finally, we have found that magnesium sulfate, which is used as a neuroprotective agent for impending preterm births, can reduce activation of the hif1 pathway and decrease pathfinding errors following hypoxia.
While hif1α is broadly upregulated following hypoxia, we determined that activation of the hif1 pathway disrupts pathfinding through a non-cell-autonomous effect. This may be due to hypoxia interfering with the balance of ligand-receptor signaling necessary for normal pathfinding. Thus, the same effect, disrupting axon guidance, can be achieved by at least one of two mechanisms: by increasing Eph receptor expression through hif1 pathway activation in neighboring cells; or by increasing EphrinB2a expression in TCPT neurons. Usually, EphrinB2a acts as a ligand for one of the receptor tyrosine kinases (RTK) of the EphA3, EphA4, or EphB4 families, which in turn sets off an intracellular signaling cascade in the RTK-expressing cell [37]. However, EphrinB2 is also able to “reverse” signal, usually by tyrosine phosphorylation of its own cytoplasmic domain [35].
We found an evolutionary conservation of the molecular mechanisms mediating hypoxia's effects on pathfinding. Pocock and Hobert [13] described pathfinding defects in the invertebrate C. elegans caused by hif1 pathway activation and increased expression of the Eph receptor VAB-1. Prior to our study there has only been indirect evidence showing disruption of axon pathfinding from hypoxia in vertebrates, and whether the same genetic pathways as in C. elegans would be involved was unknown. There is a marked increased in complexity of CNS structures, and increase in number of the members in gene families, from invertebrates to vertebrates. Thus, it is notable both that hypoxia can specifically disrupt vertebrate axon pathfinding, and that ephrinB2a, the closest zebrafish homolog to VAB-1, appears to be a primary mediator of hif1 pathway activation.
In C. elegans, hypoxia leads to increased expression of the Eph receptor VAB-1 and aberrant crossing of the midline by PVQ interneurons and HSN motorneurons, a phenotype which can be rescued by knock-down of the ligand efn-2 [13]. In contrast, we observed a failure of commissural axon crossing, even though we observed an up-regulation of the corresponding zebrafish ligand ephrinB2a. The reason for this difference between the two species, of increased crossing versus failure to cross the midline, is not clear, although there are several possible answers. First, it may reflect a bias in neurons assayed. Most of our analysis was done with an enhancer for axons that normally cross the midline. It is possible that with a different enhancer, for example of longitudinal axons, we might observe aberrant crossing. Second, some of the C. elegans neurons affected by hypoxia are not strictly ipsilateral in their normal projections. The PVQ-L crosses the “midline” to travel with the PVQ-R axon before switching back to the left side, while the HSN motorneurons project through the nerve ring to the opposite side of the nervous system [38]. Thus, the crossing of the midline may reflect a disorganization of pathfinding, rather than a loss of signals preventing midline crossing. Third, there are significant differences in midline guidance in the nervous systems of zebrafish and C. elegans. The C. elegans midline is defined by a guidepost function of midline motor neurons [39], whereas in zebrafish and other vertebrates glial bridges provide some of this role [40]. VAB-1 over-expression in C. elegans is able to produce axon guidance defects [41], so perhaps differences in the midline structures of the nervous systems results in different expression patterns and usage of the eph/ephrin signaling pairs.
We found that hypoxia caused increased expression of ephrinB2a by hypoxia, and subsequent failure of TCPTc axons to cross the midline. In mammals, ephB and ephA receptor family members, and ephrinB ligands, guide development of the corpus callosum, by a complex series of actions including effects on midline glial growth, on modulating axon responsiveness, and on reverse signaling in axons as they cross the midline [42], [43]. In zebrafish, the ligand ephrinB2a is expressed in telencephalic neurons and their axons as they extend towards the midline [33, this study], and we have found that the receptor ephA4a is expressed in the midline of the CNS during axon extension. These results support a model where signaling between ephrinB2-expressing axons and ephA4a-expressing midline cells helps guide the axons as they cross the midline. Alterations in levels of ligand or receptor, for example by hypoxia-induced up-regulation of ephrinB2, could thus impair normal TCPTc formation.
Comparing degree of hypoxia in different animal species is difficult, and extrapolating 1% hypoxia (1% pO2) in zebrafish to equivalent effects in premature infants is not straightforward. Since zebrafish can survive extended periods of hypoxia and even anoxia [this study]; [ 15,16], 1% hypoxia might be equivalent to relatively “mild” hypoxia in human infants. Premature infants at 36 weeks gestation have been found to spend more than 8 hours/day at less than 90% oxygen saturation [12], and infants with chronic lung disease or congenital heart disease have more significant decreases in oxygenation. Finally, using 1% hypoxia to examine the mechanisms underlying hypoxia's effect allowed us to examine the genetic pathways and more noticeable effects on CNS pathfinding. It seems likely that less stringent hypoxia would have similar, but less marked effects. Similarly, since typical human serum magnesium levels are ∼1 mM, the human physiological corollary to bathing zebrafish in 250 mM magnesium sulfate is unclear. However, the current magnesium dose given to expectant mothers is a single bolus dose of 4 g, followed by a continuous infusion of 2 g/hour until delivery [14]. Further, it is not known what fetal magnesium levels are following the dose, and whether there are differences in neurodevelopmental outcomes depending on overall dose or post-natal levels. While there are obvious limitations to extrapolating our findings to human infants, it is possible that increasing magnesium levels in human infants could have protective effects [44]. For these reasons we think that the mechanisms and genetic pathways activated by hypoxia in zebrafish may also be relevant in human development.
In addition to conservation of molecular mechanisms, is there conservation of effects of hypoxia on specific subsets of axon pathfinding in humans? Our work found that only a subset of axons was affected by hypoxia. For example, using our transgenic reporter line, we found that commissural axons extending from forebrain neurons were disrupted, whereas we did not observe problems in the longitudinal axons or in the optic chiasm. In humans, prematurity has been shown to affect connectivity of commissural structures, including the corpus callosum [10], as well as non-commissural axon tracts such as the internal capsule, superior fasciculus, uncinate fasciculus, and external capsule [10], [45]. These changes in connectivity are correlated with decreased overall intelligence quotient (IQ) score [46], with particular correlations noted between corpus callosum disruption and IQ as well as attention-deficit hyperactivity disorder [47], [48]. In addition, autism is associated with prematurity and disruptions of both intra-hemispheric and inter-hemispheric connectivity [49], [50]. A significant caveat to these studies, however, is the indirect nature of the measurements, relying upon fractional anisotropy based on water diffusion measured by magnetic resonance imaging, and that differences in cortical volume and/or synaptic changes could also impact the clinical findings.
This study raises several issues for future study. First, we found considerable variability in the degree of disruption of axon pathfinding from animal to animal. Following hypoxia, some animals had no TCPT commissural axons cross the midline, while in other animals the TCPTc was hardly or not affected. This suggests that other homeostatic mechanisms, as yet unknown, may help prevent or ameliorate the pathfinding errors. A second question concerns why certain axons (and neurons) are particularly susceptible to hypoxia. While we were found that the TCPTc was affected following hypoxia, in other axon groups, for example the optic tracts, we never observed pathfinding errors following hypoxia. Is this due to some intrinsic feature of different neuron types, or because certain molecules, such as ephrinB2a, are expressed only in certain neurons? Since we did observe errors in other axon tracts when we used pan-axonal immunohistochemistry, it is likely that a wider subset of neurons and axon tracts are affected. Third, does hypoxia affect the development of the other main determinant of connectivity, namely synapses? Microarray data in rodents has shown that hypoxia causes altered expression of multiple synaptic genes [51]. Additional studies will need to characterize effects of hypoxia on synapse development. Fourth, magnesium was protective against the hypoxic pathfinding disruption. Magnesium sulfate is currently used as a single-dose agent immediately prior to preterm births [14], but future work could examine whether elevated magnesium levels are associated with improved neurodevelopmental outcomes, as well as what the molecular target of magnesium is.
Our finding of disrupted connectivity in the brain following hypoxia and the involvement of a conserved genetic pathway suggests one mechanism that may contribute to the diverse neurodevelopmental impairments seen in premature infants. Premature infants have elevated rates of attention-deficit disorder, autism, cerebral palsy, epilepsy, psychiatric disorders, and cognitive impairment [2]–[5], and perhaps disruptions of connectivity might be responsible for some of these outcomes. Although cognitive impairment is the most common of these chronic neurodevelopmental problems [4], it is not known, for example, why certain infants develop autism as opposed to attention-deficit disorder, or what characteristics of the hypoxia lead to different clinical outcomes. The identification of the precise effects and molecular mediators of hypoxic injury in the developing vertebrate brain offers the possibility for improved understanding and novel therapeutic approaches.
All zebrafish experiments were performed with supervision and in strict accordance of guidelines from the University of Utah Institutional Animal Care and Use Committee (IACUC), regulated under federal law (the Animal Welfare Act and Public Health Services Regulation Act) by the U.S. Department of Agriculture (USDA) and the Office of Laboratory Animal Welfare at the NIH, and accredited by the Association for Assessment and Accreditation of Laboratory Care International (AAALAC).
Adult fish were bred according to standard methods. Embryos were raised at 28.5°C in E3 embryo medium and staged by time and morphology [52]. For in situ staining and immunohistochemistry, embryos were fixed in 4% paraformaldehyde (PFA) in PBS overnight (O/N) at 4°C, washed briefly in PBS with 0.1% Tween-20, dehydrated, and stored in 100% MeOH at −20°C until use.
Transgenic fish lines and alleles used in this paper were the following: Tg(foxP2-enhancerA.2:egfp-caax)zc69; Tg(foxP2-enhancerA.2:Gal4-VP16413–470)zc72; Tg(elavl3:Gal4-VP16413–470); Tg(UAS:hif1α -2A-egfpcaax)zc73; Tg(UAS:hif1α -2A-TagRFP)zc74; Tg(UAS: hif1αmut-2A-egfpcaax)zc75; Tg(UAS: hif1αmut-2A-TagRFP)zc76, where 2A is a viral hydrolase peptide sequence [29]; and Tg(otpb.A:Gal4-VP16413–470, myl7:EGFP)zc57 [30]. Injection of DNA constructs and generation of stable transgenic lines was performed essentially as described [17]. Lines are available upon request.
To induce hypoxia, embryonic zebrafish were placed in a sealed plexiglass chamber connected via a controller that monitored and adjusted nitrogen gas flow to a desired pO2 set point (Biospherix Ltd.). We observed that equilibration of oxygen partial pressures in water could take several hours measured with a dissolved oxygen water meter (Control Company). Therefore we pre-equilibrated all solutions to either normoxia or hypoxia for at least 4 hours before use, and transferred embryos into and out of pre-equilibrated solutions. At the desired time, embryos were placed into media that had been equilibrated to the hypoxic conditions. To terminate hypoxia, embryos were returned to media kept in normoxic conditions. Morphological staging was used to help determine age at fixation for analyses.
To assay whether a TCPT pathway had decreased commissural crossing, we measured the total fluorescence intensity (average intensity x area) of a rectangular area placed over the commissure or longitudinal tract. A confocal z-stack was taken of the region using identical confocal settings (20× objective, laser power 10%, gain 1.25%, offset 2%, PMT 400, speed 2.5 µs/pixel). We used ImageJ to calculate an average intensity projection of 10 slices (step size 2.8 µm), then measured total fluorescence intensity in a rectangle of set size (0.12×0.25; 10 µm×30 µm) over the midline of the commissural TCPT pathway or over the longitudinal axons prior to their decussation into the TCPT. A ratio of the commissural vs. longitudinal axon intensity was calculated (C/L ratio) (Figure 1F). Some experimental variation was noted, and so results were only directly compared for experiments performed on the same day. For determination of C/L ratios of hypoxia effects at different ages (Figure 1G), 24 or more embryos were imaged per age.
Immunohistochemistry was performed as previously described [17], [30]. Antibodies used were: mouse anti-acetylated tubulin 1∶250, rabbit polyclonal anti-tyrosine hydroxylase 1∶400 (Millipore), mouse monoclonal anti-GFP 1∶250 (Millipore), goat polyclonal anti-EphrinB2a 1∶20 (R&D Biosystems), rabbit anti-phosphohistone H3 1∶500 (pH 3 polyclonal, Upstate Biotechnology), Cy-3 anti-rabbit 1∶400, Alexa 488 donkey anti-mouse 1∶400 (Invitrogen), Alexa 555 rabbit anti-goat 1∶100 (Invitrogen).
Double immunohistochemistry for GFP and EfnB2a was performed by permeabilization using 10 µg/ml Proteinase K in PBST, re-fixation for 10′ with 4% PFA, further permeabilization with 0.1M citrate and 0.1% Triton in PBS, incubation with mouse anti-GFP and goat anti-ephrinB2a, followed by washing with PBST/1% DMSO and incubation with donkey anti-mouse Alexa 488 and rabbit anti-goat Alexa 555.
Whole-mount in situ labeling for igfbp-1 was performed using an anti-sense probe generated by NotI digestion and transcription using T3 polymerase from pCR4 Blunt-ZF IGFBP-1 [22], as previously described [53].
Expression clones were built using the Tol2 kit [54] and recombination reactions with Gateway (Invitrogen) plasmids. For clones lacking an expressed fluorescent marker, either a cmlc2:EGFP (official nomenclature myl7:EGFP) or cmlc2:TagRFP transgenesis marker was used in the final construct. The identity of constructs was confirmed by restriction enzyme digests and by sequencing on both strands (for coding sequences) or by partial end-sequencing (for enhancers). Specific plasmids used for cloning were p5E-foxP2-enhancerA.2; p5E-elavl3; p5E-10xUAS; pME-basEGFP-caax (middle entry clone with EGFP-caax preceded by minimal promoter); pME- basGal4-VP16413–470; pME- hif1α (no stop) or pME- hif1αmut (no stop); p3E-pA; p3E-2A-eGFPcaax-pA; p3E-2A-TagRFP-pA; into either pDestTol2pA2, or pDestTol2CG2/pDestTol2CR3 (pDestTol2pA3 with cmlc2:EGFP or cmlc2:TagRFP transgenesis marker, respectively) [17], [29], [30], [54].
We cloned and sequenced the zebrafish hif1α cDNA (hif1αβ). We generated a constitutively active form of hif1α, hif1αmut, by mutating proline 621 to alanine in the conserved LXXLAP motif on HIF-1 [28]. This mutation prevents the hydroxylation of HIF-1 and thereby its subsequent proteosomal degradation. Zebrafish hif1α was cloned based on the GenBank sequence AY326951, adding attB1F and attB2R sequences for cloning using the Gateway system (Invitrogen): forward primer: bp 230–250 (hif1α sequence italicized, attB1F regular type, start codon underlined) 5′-GGGGACAAGTTTGTACAAAAAAGCAGGCTACCCAGGAATGGATACTGGAG-3′; reverse primer bp 2587–2606 5′-GGGGACCACTTTGTACAAGAAAGCTGGGTGGAAGAGTGTCCGCAGTTGC-3′. PCR was performed using Tuebingen wild-type cDNA, and the resulting 2.4 kb fragment was recombined into pDONR221 and sequenced. To generate the hif1αmutconstruct, the proline at position 557 was changed to alanine with site-directed mutagenesis by overlap PCR (CCT to GCT at nucleotide 1677) and the final clone was confirmed by full-length sequencing. C-terminal fusion constructs of hif1α were made by overlap PCR changing the stop codon (TGA) to glycine (GGA).
Image acquisition and analysis were performed as described previously [17]. Immunostained embryos were transferred step-wise into 80% glycerol/20% PBST, mounted on a glass slide with a #0 coverslip placed over a well made using electrical tape, and imaged on a confocal microscope. Confocal stacks were projected in ImageJ, and images composed with Adobe Photoshop and Illustrator. For imaging fluorescent immunohistochemistry and bright-field in situ staining for Figure 6 E″ and F″, the “Exclusion” function on Adobe Photoshop was used to combine the layers, with the “Exposure” set at −4.0 for the bright-field image.
Determination of α-Efn staining intensity was calculated using identical confocal settings for imaging all of the embryos (laser power 10%, gain 1.25%, offset 2%, PMT 400, speed 2.5 µs/pixel). Intensities were calculated in a rectangle of set size (0.12×0.25; 10 µm×30 µm) over the midline of the telencephalon using maximum intensity projections of 5 z-slices (step size 2.8 µm) (Figure 6B).
One-cell stage embryos were injected with 1 nl of 3 ng/nl EfnB2a translation-blocking morpholino [33] as previously described [17].
Live control and hypoxic embryos (72 hpf) were collected into 1.5 mL tubes, stained for 30′ in 5 µg/mL acridine orange in E3 embryo medium, then washed for 30′ at RT in E3 on a nutator. Embryos were then anesthetized and mounted in low-melt agarose containing tricaine. Live imaging was performed using a confocal microscope (PMT 581V, scan size 640×480, speed 1.79 s/scan, UPLFL 20× objective, z-step size 2.8 µm, laser intensity 15.0%)
ImageJ was used to score 20 z-slices (step size 2.8 µm) counting dorsally form the equator of the lens. A square with an area of 100 µm2 was drawn with its top boundary immediately caudal to the olfactory pits, sides immediately medial to the eyes, and lower boundary at the center (rostral to caudal axis) of the lens (Figure 3E, E′). Apoptotic cells were counted that lay within the square including those that contacted the boundaries.
Embryos were manually dechorionated prior to 24 hpf. For Mg2+ treatment, E3/PTU-containing solution was mixed with the appropriate concentration of magnesium sulfate (25–250 mM); for dimethyloxalylglycine (DMOG) (Sigma) or CAY10585 (Cayman Chemical), the drug was diluted in 0.1% DMSO and added to E3/PTU to make final concentrations of 100–750 µM (DMOG) or 20 µM (CAY10585). Embryos were incubated during an exposure period from 24–36 hpf, and at 36 hpf embryos were returned to E3/PTU. Fixation with 4% PFA occurred at 72 hpf.
|
10.1371/journal.ppat.1004243 | Arabidopsis LIP5, a Positive Regulator of Multivesicular Body Biogenesis, Is a Critical Target of Pathogen-Responsive MAPK Cascade in Plant Basal Defense | Multivesicular bodies (MVBs) play essential roles in many cellular processes. The MVB pathway requires reversible membrane association of the endosomal sorting complexes required for transports (ESCRTs) for sustained protein trafficking. Membrane dissociation of ESCRTs is catalyzed by the AAA ATPase SKD1, which is stimulated by LYST-INTERACTING PROTEIN 5 (LIP5). We report here that LIP5 is a target of pathogen-responsive mitogen-activated protein kinases (MPKs) and plays a critical role in plant basal resistance. Arabidopsis LIP5 interacts with MPK6 and MPK3 and is phosphorylated in vitro by activated MPK3 and MPK6 and in vivo upon expression of MPK3/6-activating NtMEK2DD and pathogen infection. Disruption of LIP5 has little effects on flg22-, salicylic acid-induced defense responses but compromises basal resistance to Pseudomonas syringae. The critical role of LIP5 in plant basal resistance is dependent on its ability to interact with SKD1. Mutation of MPK phosphorylation sites in LIP5 does not affect interaction with SKD1 but reduces the stability and compromises the ability to complement the lip5 mutant phenotypes. Using the membrane-selective FM1–43 dye and transmission electron microscopy, we demonstrated that pathogen infection increases formation of both intracellular MVBs and exosome-like paramural vesicles situated between the plasma membrane and the cell wall in a largely LIP5-dependent manner. These results indicate that the MVB pathway is positively regulated by pathogen-responsive MPK3/6 through LIP5 phosphorylation and plays a critical role in plant immune system likely through relocalization of defense-related molecules.
| Pathogen- and stress-responsive mitogen-activated protein kinases 3 and 6 (MPK3/6) cascade plays an important role in plant basal resistance to microbial pathogens. Here we showed that Arabidopsis MPK3 and MPK6 interact with and phosphorylate the LIP5 positive regulator of biogenesis of multivesicular bodies (MVBs), which are unique organelles containing small vesicles in their lumen. Disruption of LIP5 causes increased susceptibility to the bacterial pathogen Pseudomonas syringae. Compromised disease resistance of the lip5 mutants is associated with competent flg22- and salicylic acid-induced defense responses but compromised accumulation of intracellular MVBs and exosome-like paramural vesicles, which have previously been shown to be involved in the relocalization of defense-related molecules. Phosphorylation by MPK3/6 increases LIP5 stability, which is necessary for pathogen-induced MVB trafficking and basal disease resistance. Based on these results we conclude that the MVB pathway is positively regulated by pathogen-responsive MPK3/6 through LIP5 phosphorylation and plays a critical role in plant immune system probably through involvement in the relocalization of defense-related molecules.
| Endosomes traffic molecules from the plasma membrane to intracellular compartments and transport molecules from the biosynthetic apparatus to the sites of action [1], [2]. Several different endosomes have been described based on biochemical composition, morphology, and function. Multivesicular bodies (MVBs) are late endosomes that contain intraluminal vesicles generated when the limiting membrane of the endosome invaginates and buds into its own lumen, thereby allowing cargo-containing intraluminal vesicles to be delivered into and degraded upon fusion with lysosomes or vacuoles [1], [2]. Those proteins retained in the limiting membrane of MVBs, on the other hand, can be delivered to the membrane of lysosomes or vacuoles, or sort back to the plasma membrane or other cellular compartments [1], [2]. Protein sorting into MVBs is highly regulated and is dependent on the action of three distinct protein complexes named ESCRT-I, II and III (Endosomal Sorting Complex Required for Transport) [3]. Ubiquitinated membrane proteins are first recognized by ubiquitin-binding proteins such as the TOM1 families of proteins, which also recruit ESCRT-I components from the cytoplasm. ESCRT-II and ESCRT-III complexes then transiently assembly on the endosomal membrane for cargo sorting, concentration and vesicle formation. For sustained protein trafficking through the MVB pathway, it is necessary that the ESCRT complexes are dissociated and disassembled from the membrane and recycled back into the cytoplasm. The Vps4p/SKD1 AAA ATPase together with its positive regulator Vta1/LIP5 catalyzes the process of ESCRT disassembly in an ATP-dependent reaction [4], [5], [6], [7], [8]. Studies in both yeast and mammalian cells indicate that both Vps4p/SKD1 and Vta1/LIP5 are critical players during MVB biogenesis [5], [9], [10], [11]. In Arabidopsis, disruption of the SKD1 gene is lethal and expression of an ATPase-deficient version of SKD1 causes alterations in the endosomal system and ultimately cell death [12]. Arabidopsis LIP5 interacts strongly with SKD1 and increases in vitro the ATPase activity of SKD1 by 4–5 fold [12]. However, disruption of LIP5 in Arabidopsis causes no phenotypic alterations under normal growth conditions, indicating that the basal levels of the SKD1 ATPase activity are sufficient for plant growth and development [12].
Plants respond to pathogens using two innate immune systems: PTI (pathogen-associated molecular pattern- or PAMP-triggered immunity) and ETI (effector-triggered immunity) [13]. PTI is activated by PAMPs such as bacterial flagellin through mitogen-activated protein kinase(MAPK)-dependent and MAPK-independent signaling pathways (Pitzschke et al., 2009). To suppress PTI, pathogens deliver effectors to plant cells, which may be recognized by plant resistance (R) proteins and activate ETI [13]. ETI is often manifested as hypersensitive responses (HR) associated with rapid programmed cell death [13]. Studies over the past decade have provided increasing evidence for association of vesicle trafficking with plant innate immune systems. In Arabidopsis, pattern-recognition receptor FLS2 confers immunity against bacterial infection through recognition of bacterial flagellin. Following flagellin binding, activated FLS2 undergoes endocytosis and accumulates in late endosomes/MVBs before degradation [14], [15]. Endocytosis of FLS2 functions as a molecular mechanism not only for the attenuation of FLS2 activation but probably also for signaling required for efficient PTI [15], [16], [17]. In addition, N-terminal motifs of a number of NB-LRR R proteins are associated with endomembrane and contribute to disease resistance. Potato R protein R3a relocates from the cytoplasm to late endosomes/MVBs when co-expressed with its cognate effector [18]. Inhibition of the relocalization of R3a to endosomes attenuates the R3a-mediated HR, indicating that relocalization to vesicle in the endocytic pathway is necessary for effector recognition and HR signaling by the R protein [18]. In the penetration resistance of cereal plants against powdery mildew fungal pathogens, which is conferred by local cell wall appositions (papillae), electron or confocal microscopy detected trafficking molecules through late endosomes/MVBs for delivering defense-related materials to papillae, thereby executing a timely and localized defense response to invading pathogens [19], [20], [21], [22], [23]. Similar relocalization of defense-related molecular such as the PENTRATION RESISTANCE 3 (PEN3) ATP binding cassette transporter for cell surface defense in response to conserved pathogen elicitors has also been observed in Arabidopsis [24]. In spite of the extensive microscopic data, genetic analysis of the role of MVBs in plant immune system is not straightforward because mutants for genes essential for MVB biogenesis are often lethal [12], [25]. Previously, the barley GTPase ARFA1b/1c has been localized to MVBs and shown to be important for callose-deposition and penetration resistance of barley [21]. However, the MVB localization of the ARF1 factor was later disputed and evidence was presented for localization of the GTPase to the Golgi and trans-Golgi network (TGN) [26]. Therefore, there is still no compelling genetic evidence for a critical role of the MVB pathway in plant immune system, much less its regulation during plant-pathogen interactions.
MAPK cascade is involved in transduction of pathogen signals to defense responses in plants [27]. In tobacco, Arabidopsis, rice and other plants, stress/pathogen-responsive MAPKs have been identified and extensively studied. In tobacco, WOUND-INDUCED PROTEIN KINASE (WIPK) and SALICYLIC ACID-INDUCED PROTEIN KINASES (SIPK) are activated in resistant tobacco by tobacco mosaic virus and are involved in pathogen-induced HR [28]. In Arabidopsis, functionally redundant MPK3 and MPK6 (orthologs to tobacco WIPK and SIPK) are also responsive to pathogens and pathogen elicitors and functional analyses using both loss- and gain-of-function approaches indicates their critical roles in plant immune responses including PTI, pathogen-induced phytoalexin biosynthesis and stomatal immune responses [27]. Pathogen-responsive MAPKs mediate activation of plant immune responses through phosphorylation of their downstream targets, thereby affecting their activity, stability and other molecular/biochemical properties. In Arabidopsis, MPK3 and MPK6 promote ethylene production through phosphorylation and stabilization of ACS2 and ACS6, two isoforms of the ethylene biosynthetic 1-aminocyclopropane-1-carboxylic acid synthase [29]. MPK3 and MPK6 also phosphorylate WRKY33, a transcription factor important for pathogen-induced expression of camalexin biosynthetic genes [30]. MPK3 and MPK6 also regulate Arabidopsis defense gene expression and disease resistance through phosphorylation of ethylene response factors [31], [32]. Other substrates of MPK3 and MPK6 have also been identified using a variety of approaches including proteomic and bioinformatics procedures but only a few of them have been functionally analyzed [33], [34], [35], [36]. However, our knowledge about the molecular mechanisms underlying the important biological functions of MPK3/MPK6 in plant immune responses is still limited.
In this study, we report identification of Arabidopsis LIP5, a positive regulator of SKD1 AAA ATPase of MVB biogenesis, as an interacting protein and a substrate of pathogen-responsive MPK6/MPK3. Functional analysis with lip5 T-DNA insertion mutants indicates that LIP5 plays a critical role in pathogen-induced MVB trafficking and in basal resistance to Pseudomonas syringae strains. The critical role of LIP5 in plant immune system is dependent on its ability to interact with SKD1. Further analysis reveal that LIP5 is expressed at low levels in healthy plants but its protein levels can be substantially elevated through phosphorylation by the pathogen-responsive MPK cascade. Mutation of MPK phosphorylation sites in LIP5 does not affect its interaction with SKD1 but reduces its stability and, as a result, compromises its ability to complement the basal resistance of the lip5 mutant plants. These results provide genetic evidence for a critical role of induced MVB biogenesis in plant basal resistance and establish an important mechanism for the regulation of vesicle trafficking during plant-pathogen interactions.
Pathogen-responsive MPK3 and MPK6 positively regulate pathogen-induced expression of WRKY33, which encodes a WRKY transcription factor important for plant resistance to necrotrophic pathogens and pathogen-induced phytoalexin biosynthesis [30]. MPK3 and MPK6 activate WRKY33 through phosphorylation and activated WRKY33 recognizes its own promoter and activates its own expression [30]. Without realizing the positive feedback mechanism of WRKY33 induction by MPK3/6, we were initially interested in identifying substrates of the pathogen-responsive MAPKs that may mediate pathogen induction of WRKY33. To identify possible substrates of the MAPKs, we cloned MPK3 and MPK6 full-length coding sequence (CDS) in frame into pBD-GAL4 plasmid and used them as baits for yeast two-hybrid screens. After screening 2×106 independent transformants of an Arabidopsis cDNA prey library, we identified positive clones by prototrophy for His and by LacZ reporter gene expression through assays of β-galactosidase activity. One of the positive clones identified with MPK6 as bait encodes LYST-INTERACTING PROTEIN 5 (LIP5, At4g26750). As the clone identified from the library screening contains only the 3′ terminal part of LIP5, we cloned its full-length coding sequence (CDS) into pAD-GAL4 to generate the pAD-LIP5 fusion construct and retested the interaction in yeast. Yeast cells co-transformed by pBD-MPK6 and pAD-LIP5 were able to grow on -His selective media (data not shown) and were positive for LacZ reporter gene expression based on the β-glucosidases activity (Figure 1A).
To determine whether MPK6 and LIP5 interact in vivo, we conducted BiFC (bimolecular fluorescence complementation) in transgenic Arabidopsis plants. We fused MPK6 and LIP5 to the N- and C-terminal yellow fluorescent protein (YFP) fragments to generate MPK6-N-YFP and LIP5-C-YFP fusion constructs, respectively. The fusion constructs under control of the CaMV 35S promoter were transformed into Arabidopsis plants and positive transformants were identified by RNA blotting and crossed to generate transgenic lines that co-expressed MPK6-N-YFP and LIP5-C-YFP. In these co-expressing transgenic lines, BiFC signals were detected in both the cytoplasm and the nucleus (Figure 1B). Control lines in which MPK6-N-YFP was co-expressed with unfused C-YFP or unfused N-YFP was co-expressed with LIP5-C-YFP did not show BiFC signals (Figure 1B).
No positive LIP5 clones were identified from our yeast two-hybrid screens with MPK3 as bait. However, in a published protein-protein-interaction map of Arabidopsis generated by testing all pairwise combinations of a collection of approximately 8,000 Arabidopsis open reading frames with an improved high-throughput binary interactome mapping pipeline based on the yeast two-hybrid system, one of the interacting partners of MPK3 is LIP5 [37]. Using both yeast two-hybrid assays and BiFC, however, we found that the interaction, if any, between MPK3 and LIP5, was much weaker than that between MPK6 and LIP5 (Figure 1B). The discrepancy could be caused by the transient nature of the interaction between LIP5 and MPK3 or the less accessible interaction domain of the MPK3 fusion proteins in our yeast two-hybrid assays and BiFC. As will be described later, both in vitro and in-gel assays showed that LIP5 could be phosphorylated not only by MPK6 but also by MPK3. Based on these results, we conclude that LIP5 is capable of interacting with MPK6 and, to a less extent, with MPK3.
Signaling through MPK3/6 in Arabidopsis and their orthologs in other plants plays critical roles in plant immune system [27]. As an interacting partner and potential substrate of the pathogen-responsive MAPKs, LIP5 may act downstream of the MAPKs in plant responses to pathogens. To determine the role of LIP5 directly, we characterized two T-DNA insertion mutants for LIP5. The lip5-1 null mutant (SAIL_854_F08), which contains a T-DNA insertion in the last exon of LIP5 (see Figure S1A), has been previously isolated and characterized with no apparent phenotype under normal growth conditions [12]. The lip5-2 mutant (GABI_351F05) contains a T-DNA insertion in the last intron (see Figure S1A) and also appears to be null based on qRT-PCR (see Figure S1B). Although there was no major phenotype in plant morphology throughout their entire life cycle, the growth of both lip5-1 and lip5-2 mutants under our normal growth conditions were slightly slower and their leaves were slightly paler green and flatter than those of wild-type plants (see Figure S1C). We also observed that the seed yields of both lip5-1 and lip5-2 mutants were about 70% of those of wild-type plants under normal growth conditions.
To test possible change in plant disease resistance, we first compared the lip5 mutants with Col-0 wild-type plants for response to the virulent P. syringae pv. tomato strain DC3000 (PstDC3000). As controls, we also included sid2-3 and npr1-3 mutants, which are deficient in SA biosynthesis and signaling, respectively [38], [39]. When inoculated with the virulent bacterial pathogen, lip5 mutants developed severe chlorosis based on both visual appearance (Figure 2A) and chlorophyll levels (see Figure S2A) as observed in the sid2 and npr1 mutants, while wild-type plants displayed only very mild disease symptoms at 3–4 days post inoculation (dpi). Furthermore, the levels of the growth of the virulent bacterial pathogen in the lip5, sid2 and npr1 mutants were about 10–20 times higher than those in the wild-type plants (Figure 2B). Thus the lip5 mutants were as susceptible to the virulent bacterial pathogen as sid2 and npr1 mutants.
We also compared wild type and lip5 mutant plants for responses to avirulent strains of the bacterial pathogen. We first tested the plants for HR development after inoculation with a high dose (OD600 = 0.1) of PstDC3000 carrying avrRpm1, avrB or avrRpt2. Visible tissue collapse and cell death were already developed in wild-type leaves inoculated with PstDC3000(avrRpm1), PstDC3000(avrB) or PstDC3000(avrRpt2) by 6 hours post inoculation (hpi) (see Figure S3A). The lip5-1 mutant leaves were normal in HR development after infiltration with PstDC3000(avrRpm1) or PstDC3000(avrB) (see Figure S3). However, no visible tissue collapse was developed in the lip5-1 mutant plants infiltrated with PstDC3000(avrRpt2) at 6 hpi as in wild-type leaves, although the mutants had well-developed HR at 24 hpi (see Figure S3A). Thus, HR development induced by one of the three tested avirulent strains of PstDC3000 was delayed, but not abolished in the lip5-1 mutant plants. Similar delays in HR development after infiltration with PstDC3000(avrRpt2) was also observed in mutants defective in SA biosynthesis or signaling [40] (see Figure S3A). We also inoculated the wild type and lip5-1 mutants with a low dose (OD600 = 0.0002) of the avirulent strains and analyzed the growth of the avirulent bacterial pathogens. As shown in Figure S3B, at 5 dpi, the lip5-1 mutant had ∼10 times higher levels of avirulent bacteria than wild-type plants. Similar increases in bacterial growth upon inoculation of a low dose of the avirulent PstDC3000 strains were also previously observed in Arabidopsis mutants defective in SA biosynthesis or signaling [40].
LIP5 has been shown to be a positive regulator of SKD1, a regulator of MVB biogenesis [12]. Both yeast two-hybrid and in vitro pull-down assays have shown that Arabidopsis LIP5 is a strong SKD1 interactor and stimulates the SKD1 ATPase activity by 4–5 times [12]. To determine whether the role of LIP5 in plant disease resistance is due to its action as a positive regulator of SKD1, we performed genetic complementation of lip5-1 mutant with LIP5 genes. Structural analysis of yeast LIP5 (Vta1) has reveals that the C-terminal domain (CTD) of LIP5 mediates LIP5 dimerization and both subunits are required for interaction with SKD1(VPS4) and for its function as a positive SKD1 regulator [41]. When two conserved CTD residues of yeast Vta1, Tyr-303 and Tyr-310, were mutated into Ala residues, the mutant proteins were deficient in interacting with VPS4 but were normal in maintaining their dimeric structure [41]. The Y303A and Y310A mutant Vta1 protein also failed to stimulate the ATPase activity of VPS4 [41]. The corresponding residues in Arabidopsis LIP5 for Tyr-303 and Tyr-310 of yeast Vta1 are Phe-388 and Phe-395, respectively (see Figure S4). We generated three mutant LIP5 (F388A, F395A and F388A/F395A) in which either or both of the Phe residues were mutated into Ala residues. Using yeast two-hybrid assays, we showed that while wild-type LIP5 is a strong interactor of SKD1, LIP5F388A interacted weakly with SKD1 based on quantitative assays of the LacZ reporter gene expression (see Figure S5A). By contrast, no interaction of LIP5F395A or LIP5F388A/F395A with SKD1 was detected in yeast cells using the LacZ reporter gene assays (see Figure S5A). As with the corresponding yeast Vtal mutant proteins, dimerization of the Arabidopsis LIP5 mutant proteins was normal (see Figure S5B).
To perform genetic complementation, we placed myc-tagged wild-type and mutant LIP5 genes into a plant transformation vector under control of the CaMV 35S promoter and transformed into the lip5-1 mutant plants. Transformant lines were identified and those expressing similar levels of the LIP5 transgenes based on western blotting were compared for responses to the virulent PstDC3000 strain. As expected, the lip5-1 mutant plants were hyper-susceptible to the virulent bacterial pathogen based on the disease symptom development (Figure 2C), chlorophyll contents (see Figure S2B) and bacterial growth (Figure 2D). Transformation of either wild-type LIP5 or LIP5F388A restored disease resistance of lip5-1 (Figures 2C and 2D). In contrast, transformation of mutant LIP5F395A or LIP5F388A/F395A failed to restore disease resistance of the lip5 mutant as indicated from both the severe disease symptoms (Figure 2C) and high bacterial growth (Figure 2D). LIP5F388A, a weak SKD1 interactor (see Figure S5A), also complemented the lip5 mutant for resistance to the bacterial pathogen (Figures 2C and 2D), most likely due to its overexpression driven by the strong CaMV 35S promoter in the transgenic plants. These results strongly suggest that interaction with SKD1 is necessary for the critical role of LIP5 in plant resistance to the bacterial pathogen.
Physical interaction of LIP5 with MPK6 and MPK3 makes LIP5 a possible substrate of the pathogen-responsive MAPKs. Survey of LIP5 protein sequence revealed six proline-directed serine or threonine residues that may act as potential MAPK phosphorylation sites (Ser-73, Thr-153, Ser-254, Ser-285, Ser-307 and Ser-323) (see Figure S4B). To examine phosphorylation, we generated a mutant LIP5 in which all six proline-directed Ser or Thr residues were mutated into Ala (LIP56A). Yeast two-hybrid assays indicated that mutations of the phosphorylation sites of LIP5 didn't alter its interaction with SKD1 or MPK6 (see Figure S6). We first generated His-tagged LIP5WT and LIP56A recombinant proteins for in vitro phosphorylation assays. As shown in Figure 3A, activated MPK3 and MPK6 both phosphorylated LIP5WT. Without co-incubation with the constitutively active MKK4DD/MKK5DD for activation, neither of the two MPKs was able to phosphorylate LIP5WT (Figure 3A). By contrast, LIP56A was not phosphorylated by activated MPK3 or MPK6 (Figure 3A).
We also performed in-gel kinase assay to determine phosphorylation of LIP5WT by the native MPKs. In the assay, we embedded recombinant LIP5WT in the SDS-PAGE gel instead of commonly used myelin basic protein. Phosphorylation of embedded LIP5WT was analyzed using total protein extracted from wild type, lip5, mpk3 and mpk6 mutant seedlings with or without prior treatment with flg22, which activates MPK3 and MPK6. As shown in Figure 3B, a low basal level of phosphorylation of LIP5WT was detected with protein extracts from untreated wild-type seedlings. Greatly Increased phosphorylation of LIP5 by both MPK3 and MPK6 was observed with protein extracts from flg22-treated plants in a concentration-dependent manner (Figure 3B). Similar patterns of phosphorylation of LIP5 were also observed using protein extracts from lip5 mutants, indicating that the mutants are normal in flg22-induced MPK3 and MPK6 activation. The loss of the kinase bands in the mpk3 and mpk6 mutants confirmed the phosphorylation of LIP5 by the respective MPKs in the wild type and lip5 mutants (Figure 3B). In addition to MPK3 and MPK6, we detected two kinases of approximately 65 and 85 kD that phosphorylates LIP5 at very low levels. Unlike MPK3 and MPK6, the activities of the two unknown kinases were unchanged in flg22-treated plants (Figure 3B).
To determine in vivo phosphorylation of LIP5 by MPK3 and MPK6, we subcloned myc-tagged LIP5WT and LIP56A into plant transformation vector under control of the CaMV 35S promoter and transformed into wild-type plants. These transgenic myc-LIP5WT and myc-LIP56A lines were then crossed with the dexamethasone (DEX)-inducible promoter-driven gain-of-function NtMEK2DD (GVG-NtMEK2DD), which, like Arabidopsis AtMKK4DD and AtMKK5DD, can activate MPK3 and MPK6 through phosphorylation [42], [43], [44]. Transgenic myc-LIP5/NtMEK2DD lines containing similar levels of LIP5 transgene transcripts were identified by RNA blotting using myc-tag DNA as probe (Figure 4). Total proteins were also isolated from these transgenic plants and analyzed by western blotting using anti-myc antibody after separation on a SDS-PAGE gel. As shown in Figure 4A, even with similar transcript levels, the protein levels of myc-LIP5WT were higher than those of myc-LIP56A even before DEX induction of NtMEK2DD. After DEX treatment, the protein levels of myc-LIP5WT were further increased while those of myc-LIP56A remained unchanged (Figure 4A). By 24 hours after DEX treatment, the protein level of myc-LIP5WT was at least 5–6 times higher than that of myc-LIP56A (Figure 4A). These results suggest that the protein stability of LIP5 is positively regulated by the basal activities of MAPKs and is further enhanced by increased MAPK activation by gain-of-function NtMEK2DD.
To provide direct evidence that LIP5 is phosphorylated in vivo, we performed the Phos-tag mobility shift assay, which is based on the reduced mobility of phospho-proteins due to their binding to the Phos-tag reagents in the SDS-PAGE gel matrix [45]. Detached leaves were treated with DEX and protein samples collected at different time points were separated on Phos-tag Acrylamide and myc-LIP5 proteins were detected by western blotting using anti-myc antibody. Without DEX treatment, we detected two major bands of LIP5WT and a minor band that was most retarded on the gel (Figure 4A). After DEX treatment, the intensities of the two retarded bands increased greatly, although the level of the least retarded band also increased significantly (Figure 4A). By contrast, we detected two bands for LIP56A and their intensities didn't change significantly after DEX treatment (Figure 4A). Based on these results, the least retarded band is mostly likely the unphosphorylated LIP5 while the two retarded bands are phosphorylated LIP5. Since even the LIP56A had a retarded band, there appears to be other sites in LIP5 besides the 6 proline-directed Ser and Thr that are subjected to phosphorylation by unknown kinases. Phosphorylation of proline-directed Ser/Thr residues by activated MAPKs caused further phosphorylation of LIP5WT, leading to further reduction in mobility on the Phos-tag gel (Figure 4A). Furthermore, increased phosphorylation of LIP5WT after DEX treatment was accompanied by a substantial increase in the protein level as indicated by the combined intensities of the three detected protein bands (Figure 4A). No such increase in protein levels was observed in DEX-treated plants expressing the mutant LIP56A protein (Figure 4A).
To determine whether LIP5 was also phosphorylated in vivo after pathogen infection, we inoculated with PstDC3000 transgenic myc-LIP5WT and myc-LIP56A lines expressing similar levels of the respective myc-LIP5 transgenes based on RNA blotting (Figure 4B). Total proteins were again isolated from these transgenic plants and analyzed by western blotting using anti-myc antibody after separation on a SDS-PAGE gel. As shown in Figure 4B, despite similar transcript levels, the protein level of myc-LIP5WT was substantially higher than that of myc-LIP56A even before pathogen infection. After PstDC3000 inoculation, the protein levels of myc-LIP5WT were further increased while those of myc-LIP56A were unchanged (Figure 4B). At 24 hpi, the protein level of myc-LIP5WT was at least 4–5 times higher than that of myc-LIP56A (Figure 4B). Furthermore, the Phos-tag mobility shift assay revealed increased phosphorylation of LIP5WT after PstDC3000, with the highest levels at 24 hpi (Figure 4B). No such increase in phosphorylation of LIP56A was observed after PstDC3000 (Figure 4B). These results indicate that the protein stability of LIP5 is positively regulated by its basal level of phosphorylation and can be further enhanced by increased phosphorylation after pathogen infection.
To confirm that retarded LIP5 protein bands on the Phos-tag gels were resulted from phosphorylation, we treated protein extracts from DEX-treated or pathogen-infected transgenic myc-LIP5WT plants with calf intestinal phosphatase (CIP) prior to electrophoresis. The treatment of CIP led to almost complete collapse of the retarded bands and the collapse could be blocked by inclusion of a protein phosphorylation inhibitor (see Figure S7). This result indicates that retarded migration of LIP5WT is due to phosphorylation.
To determine the importance of LIP5 phosphorylation by the pathogen-responsive MPKs, we compared the ability of myc-LIP5WT and myc-LIP56A in complementing the lip5 mutant phenotype. We initially used the 1.5 kb upstream promoter sequence of LIP5 for driving the myc-LIP5WT and myc-LIP56A transgene. Surprisingly, unlike the constitutive CaMV 35S promoter-driven myc-LIP5WT construct, the LIP5 promoter-driven myc-LIP5WT construct (PLIP5::myc-LIP5WT) failed to complement the lip5 mutant phenotype. RNA blotting using LIP5 as probe failed to detect the LIP5 transcripts and western blotting detected no accumulation of myc-LIP5WT in the lip5-1/PLIP5::myc-LIP5WT lines (data not shown). This result indicated that additional non-coding sequences other than the 1.5 kb upstream sequence are necessary for sufficient levels of LIP5 expression. For this reason, we again used the CaMV 35S-driven LIP5 (35S::LIP5WT and 35S::LIP56A) constructs and obtained more than 30 independent lines for each construct in the lip5-1 mutant background. Using RNA blotting with the myc tag DNA as probe, we identified three types of transgenic lines for each construct that contained high (H), medium (M) and low (L) levels of the myc-LIP5 transgene (Figure 5A). Protein immunoblotting using an anti-myc antibody revealed that there were similarly high levels of myc-LIP5 proteins in the transgenic lip5-1/35S::LIP5WT-H and lip5-1/35S::LIP56A-H lines, which contained relatively high levels of myc-LIP5 transcripts (Figure 5A). When these lip5 transgenic lines were inoculated with PstDC3000, the resistance was fully restored to the levels of wild-type plants based on both disease symptom development and bacterial growth (Figures 5B and 5C). On the other hand, when comparing the transgenic lines that contained medium levels of myc-LIP5 transcripts, we observed that the protein levels of myc-LIP5WT in the lip5-1/35S::LIP5WT-M lines were substantially higher than those of myc-LIP56A in the lip5-1/35S::LIP56A-M lines (Figure 5A). Inoculation with PstDC3000 further showed that while the disease resistance was fully restored in the lip5-1/35S::LIP5WT-M lines, it was only partially restored in the lip5-1/35S::LIP56A-M lines (Figure 5B and 5C). Finally, when comparing the transgenic lines that contained low levels of myc-LIP5 transcripts, we again observed that the protein levels of myc-LIP5WT in the lip5-1/35S::LIP5WT-L lines were substantially higher than those of LIP56A in the lip5-1/35S::LIP56A-L lines (Figure 5A). Furthermore, while the disease resistance was restored in the lip5-1/35S::LIP5WT-L lines, it was not restored at all in the lip5-1/35S::LIP56A-L lines (Figure 5B and 5C). Thus, when the LIP5 transgene was expressed at medium or low levels, its enhanced stability by phosphorylation became increasingly important for its full ability to complement the lip5 mutant phenotype. It should be noted that based on RNA blotting using LIP5 DNA fragment as probe, the transcript levels of the native LIP5 gene in pathogen-infected wild-type plants were even lower than those of myc-LIP5 in the lip5-1/35S::LIP5WT-L and lip5-1/35S::LIP56A-L lines (see Figure S8). Therefore, the results from the comparative analysis of the transgenic lip5-1/35S::LIP5WT-L and lip5-1/35S::LIP56A-L lines for the role of phosphorylation-regulated LIP5 stability is applicable to the native LIP5 protein in wild-type plants.
To study how lip5 mutants are compromised in disease resistance, we compared the mutants with wild type for PTI and SA-mediated defense, both of which are critical for resistance to the bacterial pathogen [46], [47]. To test whether lip5 mutants had normal onset of PTI, we employed flg22, a 22-amino-acid PTI peptide elicitor from bacterial flagellin. Application of flg22 induces transcriptional and translational reprogramming and cellular responses that prime the defense pathways in Arabidopsis [48]. Pre-application of flg22 can substantially decrease the growth of P. syringae in Arabidopsis leaves, and prolonged treatment of flg22 of Arabidopsis seedlings results in growth inhibition [49]. As shown in Figure 6A, prolonged incubation in flg22 led to inhibition of seedling growth of wild-type and lip5 mutants in a concentration-dependent manner. By contrast, no inhibition of seedling growth by flg22 was observed in the fls2-2 mutant for the Arabidopsis receptor required for perception of the bacterial flagellin elicitor [50] (Figures 6A). Likewise, pre-application of flg22 reduced the growth of the virulent bacterial pathogen by ∼15 fold in both the wild type and lip5 mutants but not in the fls2-2 mutant (Figure 6B). PTI is also associated with increased callose deposition [51]. To test whether lip5 mutant plants were normal in flg22-induced callose deposition, we treated wild-type, lip5-1 and fls2-2 mutant seedlings with 0, 0.2 and 1 µM flg22 and compared the numbers of callose deposition. As shown in Figure 6C, both wild-type and lip5-1 mutant seedlings showed similar concentration-dependent increases in callose deposition following flg22 treatment. By contrast, no flg22-induced increase in callose deposition was observed in the fls2-2 mutant (Figure 6C). Induction of a number of early PTI WRKY marker genes and PATHOGENESIS-RELATED PROTEIN (PR) late marker genes by flg22 was also largely normal in the lip5 mutants (see Figure S9). These results indicated that lip5 mutants were normal in PTI.
SA-mediated defense plays an important role in Arabidopsis resistance to P. syringae [52], [53], [54]. SA-mediated defense is associated with induction of PR genes, which are observed in both SA-treated and PstDC3000-infected plants [52]. The sid2 and npr1 mutants defective in SA biosynthesis and signaling, respectively, are compromised in pathogen-induced PR gene expression [38], [39]. To test if SA pathway is compromised in lip5 mutants, we analyzed induction of SA-regulated PR1 expression after infiltration with 10 mM MgCl2 (mock inoculation) or with virulent PstDC3000. No PR1 transcript was detected in either healthy wild-type or lip5-1 mutant plants (Figure 7A). In wild-type plants, little PR1 transcripts were detected after mock inoculation but were induced after inoculation with the bacterial pathogen (Figure 7A). In the lip5-1 mutant plants, intriguingly, PR1 transcripts were elevated substantially even in mock-inoculated leaves (Figure 7A). Elevated PR1 transcripts in mock-inoculated lip5-1 mutant was likely caused by infiltration-caused wounding as spraying or soaking of the leaves of the same mutant led to only slight induction of PR1 (see Figure S9). Furthermore, induction of PR1 was faster and to significantly higher levels in the lip5-1 mutant plants than in wild-type plants following PstDC3000 inoculation (Figure 7A). By contrast, pathogen-induced PR1 expression was greatly reduced in both sid2-3 and npr1-3 mutants (Figure 7A). To determine whether pathogen-induced PR1 proteins are normally secreted in the lip5 mutant plants, we generated a construct in which the tobacco acidic NtPR1 gene is under control of the Arabidopsis PR1 gene promoter. The construct was then transformed into both wild type and lip5-1 mutant. Pathogen-induced production of tobacco PR1 proteins and their accumulation in intercellular wash fluid were determined by western blotting using a monoclonal antibody (33G1) recognizing tobacco PR1 (Chen et al., 1993). As shown in Figure 7B, secretion of transgenic tobacco PR1 protein was normal in lip5-1 mutant, as shown in western blot of intercellular wash fluid after 1 mM SA treatment (Figure 7B). These results indicated that the lip5 mutant plants are competent in SA-mediated PR1 expression and PR1 protein secretion.
To further determine the responsiveness of the lip5 mutant plants to SA, we compared them with wild-type plants for SA-induced disease resistance. In wild-type plants, pretreatment of 1 mM SA reduced the growth of the virulent PstDC3000 strain by ∼15-fold (Figure 7C). Similarly reduced growth of the bacterial pathogen was observed in the lip5-1 mutant plants after SA treatment (Figure 7C). These results support that the lip5 mutants are competent in SA-mediated defense.
As a positive regulator of SKD1 in MVB biogenesis, the critical role of LIP5 in plant immune system is likely through regulation of cellular vesicle trafficking during plant-pathogen interactions. To test this, we compared wild-type and lip5 mutant plants for pathogen-induced endocytosis using the styryl dye FM1–43 as a fluorescent endocytosis marker. The membrane-selective FM1–43, which fluoresces significantly only in a lipid-rich membrane, is unable to cross membrane because of the amphiphilic nature but can enter the cells by endocytic vesicles invaginated from the plasma membrane [55], [56]. In uninfected wild-type plants, FM1–43 strongly labeled the plasma membrane as a result of the association of the dye with the lipid phase (Figure 8A). After PstDC3000 inoculation, the sharp fluorescence at the plasma membrane was reduced and became diffusive (Figure 8A). In addition, there was approximately 4-fold increase in the number of internalized fluorescent vesicles in pathogen-infected wild-type plants when compared to uninfected plants (Figure 8A). In the lip5-1 mutant plants, strong labeling of plasma membrane and low numbers of internalized fluorescent vesicles were also observed as in wild-type plants (Figure 8A). Unlike in wild-type plants, however, pathogen-induced reduction of fluorescence at the plasma membrane and increase in internalized fluorescent vesicles were almost completely abolished in the lip5-1 mutant plants (Figures 8A). These results indicate that pathogen infection stimulates endocytosis and vesicle trafficking in host cells in a LIP5-dependent manner.
To further analyze the role of LIP5 in pathogen-induced MVB formation, we generated transgenic wild-type and lip5-1 mutant plants expressing a GFP-fused MVB marker protein, the ARA6/RabF1 GTPase [12]. The transgenic plants were inoculated with the virulent PstDC3000 bacterial pathogens and observed for the ARA6-GFP signals at 0 and 24 hpi. At 0 hpi, we detected similar levels of punctate ARA6-GFP signals in the wild type and lip5 mutant plants (Figure 8). After PstDC3000 inoculation, the numbers of punctate ARA6-GFP signals were doubled in wild-type plants but were little changed in the lip5 mutant plants (Figure 8B). These results support that pathogen infection induces MVB biogenesis in plant cells in a LIP5-dependent manner. Interestingly, the numbers of ARA6-GFP-labeled vesicles were substantially higher than those of internalized FM1–43-stained vesicles at 0 hpi and, as a result of the higher basal levels, the induction of ARA6-GFP-labeled vesicles was smaller than that of internalized FM1–43-stained vesicles after PstDC3000 infection (Figure 8A & 8B). This discrepancy might result from the fact that FM1–43 stained mostly endocytic vesicles invaginated from the plasma membrane, which are apparently highly responsive to pathogen infection, while constitutively expressed ARA6-GFP would label not only MVBs derived from endocytic vesicles but also MVBs derived from constitutive vesicle trafficking pathways invaginated from other subcellular compartments.
To determine whether internalized FM1–43 fluorescent vesicles, which increased in wild-type plants after PstDC3000 infection, are related to MVBs, we analyzed co-localization of the vesicles with those labeled by ARA6-GFP. Transgenic plants expressing the ARA6-GFP MVB marker were inoculated with PstDC3000 and analyzed for increased vesicle trafficking using FM1–43. As shown in Figure 8C, internalized FM1–43 fluorescent vesicles were observed in pathogen-infected plants and approximately 40% of these FM1–43 punctate spots were also labeled with ARA6-GFP.
To analyze the role of LIP5 interaction with SKD1 and LIP5 phosphorylation in pathogen-induced vesicle trafficking, we examined internalized FM1–43 fluorescent vesicles in the lip5-1 mutant complemented with the LIP5F395A or LIP56A mutant gene. As shown in Figure 8D, will-type LIP5 at various levels could fully complement the lip5 mutant for pathogen-induced formation of FM1–43-stained vesicle structures. Likewise, mutant LIP56A at high levels (LIP5-6A-H) could fully restore the vesicle numbers in pathogen-inoculated lip5 mutant (Figure 8D). Mutant LIP56A at medium levels (LIP5-6A-M), on the other hand, could only partially restore the vesicle numbers in pathogen-inoculated lip5 mutant (Figure 8D). By contrast, mutant LIP5F395A even at high levels and mutant LIP56A at low levels (LIP5-6A-L) were ineffective in restoring the vesicle numbers in pathogen-inoculated lip5 mutant (Figure 8D). These results indicate that both the interaction with SKD1 and phosphorylation-dependent stability of LIP5 are critical for pathogen-induced vesicle trafficking. We also found that both the mks3 and mks6 mutants were normal in pathogen-induced formation of FM1–43-stained vesicles (Figure 8D). It is most likely that lack of compromised phenotypes in pathogen-induced vesicle trafficking in the mpk3 and mpk6 mutants was due to the functional redundancy of the two kinases as previously observed [30], [57], [58].
A number of reported studies have implicated MVBs in a cell wall-associated defense responses in barley leaves to the pathogenic powdery mildew fungus through fusion with the plasma membrane to release internal vesicles into the paramural space. To further analyze pathogen-induced, LIP5-regulated MVB trafficking, we employed transmission electron microscopy (TEM) for detection and ultrastructural characterization of MVBs or related vesicle bodies in PstDC3000-infected Arabidopsis plants. As shown in Figure 9, TEM revealed occurrence of spherical MVBs containing small vesicles in the lumen in both the cytosol and vacuole. The TEM study also revealed exosome-like paramural vesicles situated between the plasma membrane and the cell wall, similar to the paramural bodies (PMVs) observed in barley leaves attacked by the pathogenic powdery mildew fungus (Figure 9). To quantify the effects of PstDC3000 infection and disruption of LIP5 on the occurrence of the vesicular bodies, we estimated the average numbers of MVBs and PMBs per 10 sectioned cells in both wild plants and lip5-1 mutants at both 0 and 48 hpi. At 0 hpi, both wild-type and lip5 mutant plants had 1–2 MVBs or PMBs per 10-sectioned cells. At 48 hpi, the numbers of MVBs and PMBs per 10-sectioned cells in wild-type plants increased to 4 and 14, respectively (Figure 9). In contrast, the numbers of MVBs and PMBs per 10-sectioned cells in the lip5 mutant plants were only 1 and 3, respectively (Figure 9). Thus, PstDC3000 infection increased the number of MVBs and to a greater extent, PMVs and this pathogen-induced MVB/PMB formation was largely LIP5-dependent.
We monitored the expression level of LIP5 and its activating target SKD1 during PstDC3000 infection using RNA blotting. In wide-type plants, the levels of LIP5 transcript were slightly and transiently increased after mock inoculation (see Figure S10). In PstDC3000-inoculated plants, the increase in the levels of LIP5 transcript was even stronger, particularly at 2 and 3 dpi (see Figure S10). The levels of SKD1 transcripts was unchanged in mock-inoculated plants but also significantly elevated in PstDC3000-infected plants (see Figure S10). Thus, expression of both LIP5 and SKD1 was responsive to pathogen infection.
To examine the subcellular localization of LIP5, we generated a LIP5-GFP fusion gene and transformed into the lip5-1 mutant. Western blotting using an anti-GFP monoclonal antibody detected a protein band in the transgenic plants but not in untransformed plants with a molecular weight expected for that of LIP5-GFP (see Figure S11A), suggesting that the fusion protein was intact. Transformation of the LIP5-GFP gene completely restored the disease resistance of lip5-1 (see Figure S11B & S11C), indicating the fusion protein is functional. In uninfected Arabidopsis plants expressing LIP5-GFP, fluorescent signals were observed predominantly in the cytoplasm but also in the nuclei (Figure 10A). To confirm the nuclear localization of LIP5-GFP, we peeled off the leaf epidermal layer of the transgenic plants for DAPI staining and surprisingly found that wounding alters subcellular localization of LIP5 because many fluorescent punctate signals were observed after the leaf epidermis was peeled off (see Figure S12). There was also a 3–5-fold increase in the number of fluorescent punctate signals in the cells after infection with PstDC3000 (Figure 10A). Thus, both wounding and pathogen infection could alter subcellular localization of LIP5.
To examine the effect of the SKD5-LIP5 interaction on the subcellular localization of LIP5, we transiently co-expressed LIP5-GFP with SKD1 in Nicotiana benthamiana. Unlike in tobacco leaves expressing only LIP5-GFP, in which the fluorescent signals were largely cytosolic and diffusive as found in Arabidopsis, co-expression of LIP5-GFP with wild-type SKD1 resulted in a large number of punctate fluorescent structures (Figure 10B). No such punctate fluorescent structures were observed when LIP5F395A-GFP was co-expressed with SKD1 in tobacco leaves (Figure 10B). To further analyze this, we co-expressed LIP5-GFP with a mutated version of SKD1, SKD1E232Q, which is unable to hydrolyze ATP and when overexpressed in plant cells induces dominant-negative endosomal sorting defects including enlarged MVBs [12]. Indeed, co-expression of LIP5-GFP with SKD1E232Q led to a substantial number of enlarged, intensely fluorescent punctate structures, concomitant with reduced fluorescent intensity in the cytoplasm in infiltrated tobacco leaf cells (Figure 10B).
To determine whether the LIP5-GFP punctate signals generated after coexpression with SKD1 are MVBs, we analyzed their colocalization with ARA6 through coexpression of the LIP5-GFP, ARA6-mRFP and SKD1 in N. benthamiana. As observed with GFP-SKD1 from a previous study [12], the colocalization analysis between LIP5-GFP and ARA6-mRFP was difficult because of the strong fluorescent signals of LIP5-GFP. Careful examination revealed that about 50% of the LIP5-GFP punctate signals were labeled with ARA6-mRFP (Figure 11A). Interestingly, almost all of the yellow fluorescent punctate signals generated from dimerization of LIP5 through BiFC were labeled with ARA6-mRFP (Figure 11B). This difference in the extent of colocalization of LIP5 and ARA6 could be due to the differential association of the two proteins with MVBs at various stages of biogenesis. It might be possible, for example, that dimerized LIP5 proteins are preferentially associated with ARA6-positive MVBs whereas undimerized LIP5 proteins are associated with ARA6-negative vesicles.
The extensively characterized MPK3/MPK6 signaling cascade plays a crucial role in plant innate immunity [59]. A number of Arabidopsis proteins including WRKY33, ACS2, ACS6 and ERF6 have been identified as substrates of MPK3/6 and functionally analyzed for their roles in plant defense against necrotrophic pathogens [29], [30], [32]. Here, we report that LIP5, a positive regulator of the SKD1 AAA ATPase in MVB pathway, is another substrate of the pathogen-responsive MAPKs and plays an important role in plant basal resistance to the hemibiotrophic bacterial pathogen P. syringae. Through comprehensive genetic, molecular and biochemical analysis, we further demonstrated that the critical role of LIP5 in plant basal resistance is dependent on both its interaction with the SKD1 AAA ATPase in MVB biogenesis and its phosphorylation-enhanced stability. Our study provides genetic evidence for the critical role of MVB biogenesis in plant innate immunity and established an important mechanism for the regulation of vesicle trafficking during plant immune responses.
It has been reported that the previously isolated lip5-1 mutant showed no apparent phenotypic defects in normal growth conditions [12]. Although we observed no major phenotypes either with the lip5-1 mutant, there was a slight but detectable reduction in the size of plants (see Figure S1C). Similar reduction in plant growth was also observed with the lip5-2 mutant (see Figure S1C). The major phenotype of the lip5 mutants, however, was increased susceptibility to the bacterial pathogen P. syringae (Figure 2 & Figure S3B). Based on both bacterial growth and symptom development, the lip5 mutants were as susceptible to a virulent strain of the bacterial pathogen as the sid2 and npr1 mutants defective in SA biosynthesis and signaling, respectively (Figure 2). These results indicated that while largely dispensable for plant growth and development under normal growth conditions, LIP5 is important in plant immune system. Using a combination of FM1–43 staining, ARA6-GFP imaging and TEM, we further observed that the basal SKD1 activity without stimulating LIP5 is sufficient for the basal endocytic activities in plant cells, consistent with the absence of major growth phenotypes of the lip5 mutants. However, pathogen-induced endocytic activities were increased in wild-type plants but not in the lip5-1 mutant plants (Figure 8). This observation suggested that LIP5 is necessary for pathogen-induced vesicle trafficking.
A multitude of plant defense mechanisms including PTI and SA-mediated defense are important in Arabidopsis immune system against P. syringae [46], [47]. An increasing number of studies have implicated intracellular protein trafficking in these defense responses. In flg22-triggered immunity, activated pattern-recognition receptor FLS2 undergoes endocytosis upon flg22 binding not only for the attenuation of FLS2 activation but also for signaling required for efficient PTI [15], [16]. In both PTI and SA-mediated defense, Arabidopsis cells rely on various vesicle trafficking pathways for secretion of antimicrobial compounds (e.g. callose) or proteins (e.g. PR proteins) to combat the extracellular bacterial pathogen. A number of plant and pathogen components participate in or affect PTI or SA-mediated defense against P. syringae through directing or influencing vesicle trafficking. For example, the AvrPto effector protein from P. syringae increases bacterial growth in correlation with suppressed callose deposition and other cell wall-related defense [60]. Genes encoding components in the plant secretory pathway are up-regulated during SA-regulated systemic acquired resistance (SAR) and their mutations compromise SA-induced PR proteins and establishment of SAR [61]. The HopM1 effector protein interact with and cause degradation of the Arabidopsis endosomal protein HopM interactor 7 (AtMIN7), an ADP ribosylation factor quinine nucleotide exchange factor protein that modulates vesicle trafficking and polar deposition of callose in response to bacterial pathogens [62]. AtMIN7 has an important role in both PTI and SA-regulated immunity [63]. Strikingly, despite the high susceptibility to P. syringae, the lip5 mutant plants were responsive to flg22 and SA for induced disease resistance (Figures 6B and 7C). In addition, flg22-induced callose deposition and secretion of SA-induced PR1 were normal in the lip5 mutants (Figures 6C and 7B). The intriguing phenotypes of the lip5 mutants suggest that LIP5-regulated MVB pathway plays a distinct role in plant immune responses.
In the post-Golgi endosomal systems of plant cells, there are two types of well-described endosomal compartments, the TGN/early endosome (TGN/EE) and the MVB [2]. Unlike LIP5, which is associated with MVBs, other endosomal proteins including AtMIN7 [62], [63], KEEP ON GOING (KEG) [64], Exo70B2 and Exo70H1 exocyst subunits (Pecenkova et al., 2011) and syntaxin SYP132 (Kalde et al., 2007) with roles in PTI, ETI and SA-regulated immunity against P. syringae localize to the TGN/EE or plasma membrane. TGN/EE is a highly dynamic endosomal compartment that function as a major hub for both secretary and endocytic pathways [2]. In Arabidopsis, pathogen-induced callose is synthesized by PMR4, a plasma membrane integral enzyme that is dependent on the cellular secretory pathway for correct subcellular localization [65], [66]. Likewise, SA-regulated PR proteins enter the secretory pathway for their accumulation in the extracellular space [61], [67]. Mutations of genes encoding TGN/EE- or plasma membrane-localized proteins with critical roles in the secretory pathway would compromise production or secretion of defense-related compounds and proteins for establishment of cell-wall-based defense associated with PTI and SA-regulated SAR. On the other hand, as a regulator of the MVB biogenesis, a late endosomal compartment, LIP5 primarily affects the endocytic pathway. Therefore, critical components of flg22- and SA-induced defense mechanisms such as callose deposition and PR protein secretion may involve primarily secretory pathways that are not subject to regulation by LIP5. Second, LIP5 is a positive regulator, but not an essential component of MVB biogenesis. In the lip5 mutants, the basal levels of the SKD1 AAA ATP activity is sufficient for normal plant growth and development as indicated from its wild-type growth phenotypes when grown under normal growth conditions [12]. Even if some of the flg22- and SA-mediated defense mechanisms require vesicle trafficking through MVBs, basal SKD1 AAA ATPase activity may be sufficient for the required levels of MVB biogenesis for flg22- and SA-mediated trafficking of defense-related molecules.
Even though the lip5 mutants were responsive to flg22 and SA, they were hyper-susceptible to PstDC3000 when compared with wild-type plants with or without prior flg22 or SA treatment (Figures 2, [6] and [7]). Thus the role of LIP5-regulated MVB biogenesis in plant immunity is critical and cannot be fully compensated by flg22- or SA-induced defense mechanisms. MVBs contain endocytosed cargoes and deliver them either to vacuolar compartment for degradation or fuse with the plasma membrane to release the internal vesicles (exosomes) into the extracellular space. Indeed, the TEM study revealed that PstDC3000 infection resulted in increased numbers of MVBs and PMBs in wild-type plants and this pathogen-induced MVB/PMB formation was largely LIP5-dependent (Figure 9). A number of molecular processes could take place through activated MVB biogenesis and contribute to plant innate immunity. First, trafficking molecules through late endosomes/MVBs could play a direct role in plant immune systems by executing a timely, focal defense response to invading pathogens. In cereal plants attacked by powdery mildew fungal pathogens, penetration resistance is conferred by local cell wall appositions (papillae) deposited by the epidermal cells between the cell wall and plasma membrane [21]. Papillae contain callose and extracellular membrane materials and it has been suggested that these defense-related materials are loaded via MVBs into barely papillae based primarily on the accumulation of MVBs around papillae and presence of extracellular exosome-like vesicles beneath sites of fungal attack [19],[20],[22]. In spite of extensive microscopic results, there has been no direct genetic evidence for a critical role of MVBs in plant immune responses as mutations of genes essential for MVB biogenesis are often lethal [12], [25]. The demonstrated role of LIP5, an established regulator of MVB biogenesis, in plant basal resistance to P. syringae (Figure 2) and in pathogen-induced MVB/PMB formation (Figure 9) strongly suggests that reorganization of defense-related materials via MVBs may play a critical role in focally directing extracellular immune responses to invading pathogens including bacterial pathogens. Second, internalization of ligand-activated plasma membrane receptors by endocytosis and their localization to MVBs en route to vacuoles can lead to attenuation, stabilization or even stimulation of signaling [68]. Arabidopsis contain a large number of membrane receptors including more than 600 transmembrane receptor-like proteins (RLKs) with versatile N-terminal extracellular domains and C-terminal intracellular kinases [69]. These RLKs participate in regulatory signaling of plant growth, development and stress responses and may influence, directly or indirectly, plant immune responses in broad and diverse manners. Pathogen-induced endocytosis of membrane receptors may, therefore, constitute part of the global signaling reprogramming required for effective defense responses. Third, the composition of integral membrane proteins including nutrient transporters, ion channels and signaling receptors is critical for the growth, differentiation and survival of eukaryotic cells. As a result, the complex process of remodeling cell surface protein composition is necessary, particularly under stress conditions [70]. Removal of nutrient transporters by pathogen-induced endocytosis and MVB biogenesis, for example, may result in reduced export of necessary nutrients for the growth of extracellular pathogens. The structures and functions of integral membrane proteins can also be altered or damaged in pathogen-infected plants. Pathogen-induced endocytosis and MVB biogenesis can act as an important quality control mechanism that removes damaged and potentially toxic integral membrane proteins to promote health of plant host cells. The isolated lip5 mutants will be valuable for identifying and dissecting the mechanisms by which defective in pathogen-induced endocytosis and MVB biogenesis contribute to plant immune responses.
LIP5 interacted with MPK6 (Figure 1) and probably with MPK3 as well [37]. Furthermore, phosphorylation of LIP5 by both MPK3 and MPK6 was demonstrated based on three different types of assays. Recombinant LIP5WT but not LIP56A could be phosphorylated in vitro by activated recombinant MPK6 and MPK3 (Figure 3A). Recombinant LIP5 could also be phosphorylated in vitro by native MPK3 and MPK6 from flg22-treated plants (Figure 3B). In addition, myc-tagged LIP5WT but not LIP56A expressed in transgenic plants underwent increased phosphorylation in vivo upon pathogen infection or upon induced expression of gain-of-function NtMEK2DD (Figure 4), which phosphorylates and activates Arabidopsis MPK3 and MPK6 [71]. In these transgenic plants, importantly, LIP5WT accumulated to higher levels than LIP56A even in uninfected plants and the difference became more pronounced following activation of the MPKs through induced expression of gain-of-function NtMEK2DD or pathogen infection (Figure 4). The differential accumulation of myc-LIP5WT and myc-LIP56A was correlated with the differential phosphorylation of the expressed LIP5 proteins (Figure 4). These results indicated that phosphorylation of LIP5 by the pathogen-responsive MPKs increased its stability. Similarly, phosphorylation of Arabidopsis ACS2, ACS6 and ERF6 by MPK3/MPK6 increase the stability of the substrates (Li et al., 2012; Meng et al., 2013). Therefore, phosphorylation-dependent regulation of protein substrate stability appears to be a common mechanism through which MPK3/MPK6 regulates plant stress and defense responses.
LIP56A was as active as LIP5WT in interaction with both SKD1 and MPK6 (see Figure S6). Furthermore, the phosphorylation mutant LIP56A was fully capable of complementing the lip5-1 mutant for disease resistance when it was expressed at high levels in the transgenic plants (Figure 5). By contrast, mutant LIP5F395A incapable of interacting with SKD1 failed to complement the lip5-1 mutant plant even when it was expressed at high levels (Figure 2C & 2D). These results strongly suggested that phosphorylation of LIP5 by the pathogen-responsive MPKs affects primarily the stability but not the other critical properties of LIP5. When expressed at medium or low levels, LIP5WT but not LIP56A was fully able to complement the lip5-1 mutant for resistance to PstDC3000, in correlation with the differential accumulation of the LIP5WT and LIP56A proteins in the transgenic plants (Figure 5). Thus, the protein level of LIP56A in the transgenic lip5 mutant plants is the primary determinant for its ability to complement the lip5 mutant phenotypes.
Expression of LIP5 appeared to be relatively low based on RNA blot analysis of its transcripts in both healthy and pathogen-infected plants (see Figure S10). The expression of LIP5 also appeared to be under control of not only its promoter but also its other noncoding sequences because we failed to detect accumulation of either the LIP5 transgene transcripts or the gene product when a myc-tagged LIP5 transgene under control of its native promoter was transformed into the lip5 mutant plants. As a result, we used the constitutive CaMV 35S promoter in the LIP5 expression constructs and relied on the variation in transgene expression levels for testing the effects of protein phosphorylation on LIP5 protein accumulation and its ability to complement the lip5 mutant phenotypes (Figure 5A). To ensure that the results with the CaMV 35S promoter are physiologically relevant, we compared the levels of the LIP5WT and LIP56A transgene transcripts in the lip5 mutant background with those the native LIP5 gene in pathogen-infected wild-type plants. These experiments showed that expression level of the native LIP5 was even lower than those of the LIP5WT and LIP56A transgenes in the lip5-1/35S::LIP5WT-L and lip5-1/35S::LIP56A-L lines (see Figure S8). Accordingly, we expect that the stability and, consequently, the protein level of the native LIP5 will be even more strongly affected by protein phosphorylation than the myc-tagged LIP5 proteins in the lip5-1/35S::LIP5WT-L and lip5-1/35S::LIP56A-L lines. Therefore, the levels of LIP5 appear to be under stringent control at multiple levels. In healthy plants, the levels of LIP5 are maintained at low levels through low expression. Upon pathogen infection, the LIP5 levels are elevated by significantly elevated expression and, more importantly, increased stability through protein phosphorylation. Interestingly, constitutive expression of high levels of LIP5 or induced expression of the gain-of-function NtMEK2DD without pathogen infection did not lead to constitutive induction of MVB formation (data not shown), indicating that phosphorylation-dependent LIP5 stability is necessary but not sufficient for increased MVB formation.
In summary, we have demonstrated that Arabidopsis LIP5, a positive regulator of MVB biogenesis, is a critical target of pathogen-responsive MPK3/6 cascade in plant basal defense. LIP5 is expressed at very low levels in healthy plants but its protein levels can be substantially elevated through phosphorylation by pathogen-responsive MPKs to promote pathogen-regulated vesicle trafficking. Disruption of the LIP5 gene compromised pathogen-induced MVB and PMB formation and rendered plants highly susceptible to P. syringae. The critical role of LIP5 in plant basal resistance is dependent on both its ability to interact with SKD1 and its increased stability through protein phosphorylation. Despite their high susceptibility to P. syringae, the lip5 mutants were responsive to flg22 and SA and were normal in flg22- and SA-induced disease resistance, indicating that LIP5-regulated MVB pathway plays a critical and unique role in plant immune system.
The Arabidopsis mutants and wild-type plants used in the study are all in the Col-0 background. The lip5-1 and npr1-3 mutants have been previously described [12], [38]. Homozygous lip5-2 (GABI_315F05), sid2-3 (Salk_133146), fls2-2 (Salk_062054), mpk3 (Salk_151594) and mpk6 (Salk_062471) mutants were identified by PCR using primers flanking the T-DNA insertions listed in Table S1. Arabidopsis and N. benthemiana plants were grown in growth chambers at 22°C, 120 µE m−2 light on a photoperiod of 12-hour light and 12 h dark.
To identify MPK-Interacting proteins, we used the Gal4 based yeast-two-hybrid system as described by the manufacturer (Stratagene). Full-length MPK6 sequence was PCR-amplified using gene-specific primers (5′-atcgtcgacatggacggtggttcaggt-3′ and 5′-atcgtcgacctattgctgatattctggattgaaa-3′) and cloned into pBD-GAL4 vector to generate the bait vector. The Arabidopsis HybridZAP-2.1 two-hybrid cDNA library was prepared from Arabidopsis plants as previously described [72]. The bait plasmid and the cDNA library were used to transform yeast strain YRG-2. Yeast transformants were plated onto selection medium lacking Trp, Leu and His and confirmed by β-galactosidase activity assays using ONPG (o-nitrophenyl-β-D-galactopyranose) as substrate.
For assaying SKD1-LIP5 or LIP5-LIP5 interactions in yeast cells, full-length SKD1 and LIP5 coding sequences were PCR-amplified using gene-specific primers (SKD1: 5′-agcgaattcatgtacagcaatttcaaggaac-3′ and 5′-agcctcgagtcaaccttcttctccaaactcc-3′; LIP5: 5′-atgcgaattcatgtcgaacccaaacgaac-3′ and 5′-agctgtcgactcagtgaccggcaccggccga-3′) and cloned into pBD-GAL4 or pAD-GAL4 vectors as appropriate. Various combinations of bait and prey constructs were cotransformed into yeast cells and interactions were analyzed by assaying LacZ β-galactosidase activity.
The BiFC vectors pFGC-N-YFP and pFGC-C-YFP have been previously described [73]. The full-length LIP5, MPK3 and MPK6 sequences were PCR-amplified using gene-specific primers (LIP5: 5′-agctctagatctatgtcgaacccaaacga-3′ and 5′-atcggatccgtgaccggcaccggccga-3′; MPK3: 5′-atcggatccatgaacaccggcggtgg-3′ and 5′-atctctagaaccgtatgttggattgagtgct-3′; MPK6: 5′-atcggatccatggacggtggttcaggt-3′ and 5′-atctctagattgctgatattctggattgaaagca-3′) and cloned into pFGC-C-YFP or pFGC-N-YFP, as appropriate. The plasmids were introduced into Agrobacterium tumefaciens (strain GV3101) and transformed into Arabidopsis plants. Positive T1 transformants were identified by RNA blotting and crossed to generate transgenic plants containing a pair of BiFC constructs. Fluorescence and DAPI were visualized under a Zeiss LSM710 confocal microscope and images were superimposed with Zeiss LSM710 software.
Total RNA isolation and quantitative real-time PCR analysis of LIP5 transcripts in the lip5 mutants was performed using LIP5-specific primers (5′-aggctgctagattcgctgtg-3′ and 5′-ggccgatggatttgttagc-3′) and Arabidopsis ACTIN2 gene was used as internal control as previously described [74].
PstDC3000 inoculation was performed as previously described [40]. For tittering bacteria, inoculated leaves were homogenized in 1 ml 10 mM MgCl2 and diluted before plating on King's B Agar with appropriate antibiotics. Colony forming units were counted 2–3 days after bacteria growth at room temperature. Chlorophyll contents of inoculated leaves were determined as previously described [75]. Analysis of hypersensitive cell death after infection of avirulent strains of PstDC3000 including trypan blue staining was performed as previously described [40], [76].
The flg22 peptide was first dissolved in water to make a 10 mM stock solution. For testing flg22-induced growth inhibition, seeds of different genotypes were germinated and grown on ½ MS agar for 5 days before moving to ½ MS liquid medium containing flg22 at various concentrations in 12-well plates. Seedlings were further grown for 7 days before measuring seedling fresh weight. For determining flg22-induced disease resistance, 1 µM flg22 was pre-infiltrated into 3 leaves per plant one day before pathogen infection. For examining flg22-induced callose deposition, 10-days old Arabidopsis seedlings were transferred to ½ MS liquid medium containing 0, 0.2 or 1 µM flg22. Seedlings were stained by aniline blue 24 hours later as previously described [77]. Briefly, leaves were cleared with alcoholic lacto-phenol to remove chlorophyll, and washed first in 50% ethanol and then in water. Leaves were stained in a solution containing 0.01% aniline blue (Sigma-Aldrich) in 150 mM K2HPO4, pH 9.5 in dark for 30 min before being mounted in 50% glycerol. Pictures were taken with Nikon eclipse E800 epifluorescence microscopy and callose deposition was quantified from more than 10 125×100 µm microscopic fields per treatment per genotype. For analyzing flg22-induced defense genes, plants were treated with H2O2 or 100 nM flg22 and the samples collected after 30-minute treatment were used for analysis of WRKY gene expression and samples collected at 24-hour treatment were used for analysis of PR gene expression.
Mutant genes for SKD1E232Q, LIP5F388A, LIP5F395A, LIP5F388A/F395A and LIP56A were generated by QuickChange site-directed mutagenesis (Stratagene) or overlapping PCR using the primers listed in Table S2. The mutations were confirmed by DNA sequencing.
To generate myc-tagged wild-type and mutant LIP5 genes for overexpression in Arabidopsis, we first amplified these LIP5 coding sequences using a pair of LIP5-specific primers (5′-gcacatatgtcgaacccaaacgaacca-3′ and 5′-atcggatcctcagtgaccggcaccggccga-3′). The amplified LIP5 fragments were fused with a 4xmyc tag sequence in a modified pBlueScript vector [30]. The myc-tagged wild-type and mutant LIP5 genes were then subcloned into a modified pBI121 binary vector [30]. The resulting constructs were introduced into the Agrobacterium tumefaciens (strain GV3101) and transformed into Arabidopsis by floral dip method [78]. Transformants were identified by kanamycin resistance and expression of the transgenes was analyzed by RNA blotting or immunoblotting using an anti-myc monoclonal antibody (Sigma-Aldrich).
Total RNA isolation, separation and blot analysis using 32P-labeled DNA probes were performed as previously described [79].
Protein extraction, separation and blot analysis were performed as previously described [80]. The protein concentration was determined using the Bio-Rad protein assays kit using BSA as standard. Detection of myc-tagged proteins was performed using an anti-myc monoclonal antibody (Sigma-Aldrich).
A 1.5 kb promoter sequence of Arabidopsis PR1 gene was PCR-amplified using the following primers (5′-gttagcacaagcttgttttaacttataaaa-3′ and 5′-atcggatccttttctaagttgataatggttattgttgtg-3′) and cloned into a modified pCAMBIA1300p plant transformation vector. Tobacco acidic PR1a coding sequence was PCR amplified using gene-specific primers (5′-agcccatgggatttgttctcttttcaca-3′ and 5′-agctctagattagtatggactttcgcctct-3′) and placed behind the Arabidopsis PR1 promoter. The resulting construct, designated as PAtPR1::NtPR1 was transformed into Arabidopsis plants. Preparation of total proteins and IWF were performed as previously described (Wang et al., 2005). Immunoblot detection of NtPR1 was performed using the 33G1 monoclonal antibody against tobacco PR1 (Chen et al., 1993). Possible contamination of intracellular proteins in IWF was examined using an anti-catalase monoclonal antibody [81].
FM1–43 (Sigma-Aldrich) was dissolved in H2O as a 4 mM stock solution. Arabidopsis leaves were cut into smaller pieces and stained in 20 µM FM1–43 for 1 hour at room temperature, washed and examined under a Zeiss LSM710 confocal microscope. Internalized fluorescent puncta were identified as endocytic vesicles and counted from more than 20 60×60 µm microscopic fields per treatment per genotype. Full-length ARA6/RabF1 gene was PCR-amplified using the ARA6-specific primers (5′-agcgaattcatgggatgtgcttcttctct-3′; 5′-agcggatcctgtgacgaaggagcaggacgag-3;) and fused to the GFP gene in a binary plant transformation vector and transformed into Arabidopsis plants. The transgenic Arabidopsis plants were used for imaging under a confocal microscope following PstDC3000 infection and stained with FM1–43 for analysis of colocalization of FM1–43 and ARA6-GFP. Because of significant overlap between the FM1–43 and GFP fluorophores, colocalization of ARA6-GFP and FM1–43 staining was done by fluorescence unmixing. Briefly, FM1–43 stained ARA6-GFP transgenic plant leaves were excited at 488 nm. Emission was collected by the use of 15-channel fluorescence imaging, each channel encompassing 10 nm wide from 500 nm to 650 nm. Fluorescence unmixing of the image data was performed using Spectral Unmixing Tools with GFP-only spectrum and FM1–43-only spectrum (lipid) as references for the reliable separation of overlapping fluorescence signals and colocalization analysis.
The electron microscope was done at the Purdue Bindley Bioscience Center - Imaging Facility. Samples were fixed by the microwave under low vacuum. Briefly, leaf samples were cut into pieces 1–2 mm long and fixed with 2% paraformaldehyde (v/v) and 2.5% glutaraldehyde (v/v) in 0.1 M cacodylate buffer, pH 6.8, and rinsed with 0.1 M cacodylate buffer, pH 6.8. Tissues were further treated with 1% OsO4 (v/v) and 1.5% K3Fe(CN)6 (v/v) and washed. The samples were then dehydrated in a grade ethanol series and 100% propylene oxide at final change. Infiltration was done on the bench by 25% Spurr (overnight) and 50% (without accelerator), 75% Spurr with Accelerator (overnight) and 100% Spurr for 6 hours. Tissues were embedded in a flat mold with resin and polymerized at 60°C for 48 hours.
Samples were cut with a diamond knife using an ultramicrotome (LEICA EM UC6), sections were collected on copper grids coated with formvar and carbon, then poststained 5 min with 2% uranyl acetate in 70% methanol and Reynold's lead citrate for 3 min. Samples were imaged using a Philips CM-100 TEM (FEI Company, Hillsboro, OR) operated at 100 kv, spot 1, 200 µm condenser aperture and 70 µm objective aperture. Images were captured on an SIA L3-C digital camera. MVBs and PMBs were manually quantified from randomly selected cells in each genotype/treatment.
Full-length LIP5 gene was fused to the GFP gene behind the CaMV 35S promoter in a modified pCAMBIA1300 plant transformation vector and transformed into Arabidopsis plants. Standard confocal laser microscopy of stably transformed Arabidopsis leaves was performed for imaging of GFP.
For transient co-expression of LIP5-GFP with SKD1 and SKD1E232Q in N. benthamiana, the wild-type and mutant SKD1 genes was PCR-amplified using SKD1-specific primers (5′-agcctcgagatgtacagcaatttcaaggaac-3′ and 5′-agctctagatcaaccttcttctccaaactcc-3′) and cloned into pTA7002 under control of a dexamethasone (DEX)-inducible promoter. Agrobacterium cells containing LIP5-GFP and SKD1 constructs to be co-expressed were co-infiltrated into N. benthamiana leaves. Two days after infiltration, the leaves were infiltrated with 30 µM DEX to induce SKD1 expression and imaging of GFP was examined under a Nikon eclipse E800 epifluorescense microscope one day after DEX treatment.
For colocalization analysis of LIP5 with MVB marker ARA6, LIP5-GFP and SKD1 transgenes were coexpressed with the ARA6-mRFP marker gene [12] in N. benthamiana. Dimerized LIP5 fluorescent signals were generated from complementation of coexpressed LIP5-N-YFP and LIP5-C-YFP constructs in N. benthamiana. Imaging of coexpressed GFP, YFP and mRFP signals was performed with standard confocal laser microscopy.
Full-length LIP5WT and LIP56A coding sequences were PCR amplified and cloned into pET32a vector. Preparation of recombinant LIP5 and the in vitro phosphorylation assay were performed as previously described [30].
For preparation of native MPK3 and MPK6, total protein extracts were isolated from flg22-treated seedlings. In gel kinase assay was performed as previously described [30], using recombinant LIP5 proteins as substrates.
For examining phosphorylation of LIP5 by pathogen-responsive MPK3/6, which are activated by the gain-of-function NtMEK2DD, transgenic Arabidopsis harboring a 35S::myc-LIP5 construct was crossed with transgenic lines containing NtMEK2DD driven by a DEX-inducible promoter in pTA7200 [42], [82]. Progeny plants containing both constructs were examined for myc-LIP5 gene expression and assays of in vivo phosphorylation as previously described. Phos-tag Acrylamide (NARD Institute) was used for phospho-protein mobile shift assay to detect in vivo phosphorylation of LIP5 protein. Briefly, total proteins were separated in a 10% SDS-PAGE gel containing 100 µM Phos-tag and 200 µM MnCl2. myc-LIP5 protein shifts were detected by western blotting with anti-myc antibody.
For dephosphorylation assays, protein extracts were isolated from DEX-treated transgenic NtMEK2DD/myc-LIP5WT or pathogen-infected lip5-1/myc-LIP5WT plants and treated at 37°C for 45 minutes with CIP (0.4 U/µl) in the absence or presence of phosphatase inhibitors (10 mM NaF, 7 mM β-glycerophosphate and 5 mM Na-pyrophoshate). The protein extracts were subsequently separated on the regular SDS-PAGE and Phos-tag gels for immunoblot analysis using an anti-myc monoclonal antibody.
Arabidopsis Genome Initiative numbers for the genes discussed in this article are as follows: MPK3 (At3g45640), MPK6 (At2g43790), LIP5 (At4g26750), NPR1 (At1g64280), SID2 (At1g74710), FLS2 (At5g46330), PR1 (At2g14610) and SKD1 (At2g27600).
|
10.1371/journal.ppat.1002249 | Structural and Functional Analysis of Laninamivir and its Octanoate Prodrug Reveals Group Specific Mechanisms for Influenza NA Inhibition | The 2009 H1N1 influenza pandemic (pH1N1) led to record sales of neuraminidase (NA) inhibitors, which has contributed significantly to the recent increase in oseltamivir-resistant viruses. Therefore, development and careful evaluation of novel NA inhibitors is of great interest. Recently, a highly potent NA inhibitor, laninamivir, has been approved for use in Japan. Laninamivir is effective using a single inhaled dose via its octanoate prodrug (CS-8958) and has been demonstrated to be effective against oseltamivir-resistant NA in vitro. However, effectiveness of laninamivir octanoate prodrug against oseltamivir-resistant influenza infection in adults has not been demonstrated. NA is classified into 2 groups based upon phylogenetic analysis and it is becoming clear that each group has some distinct structural features. Recently, we found that pH1N1 N1 NA (p09N1) is an atypical group 1 NA with some group 2-like features in its active site (lack of a 150-cavity). Furthermore, it has been reported that certain oseltamivir-resistant substitutions in the NA active site are group 1 specific. In order to comprehensively evaluate the effectiveness of laninamivir, we utilized recombinant N5 (typical group 1), p09N1 (atypical group 1) and N2 from the 1957 pandemic H2N2 (p57N2) (typical group 2) to carry out in vitro inhibition assays. We found that laninamivir and its octanoate prodrug display group specific preferences to different influenza NAs and provide the structural basis of their specific action based upon their novel complex crystal structures. Our results indicate that laninamivir and zanamivir are more effective against group 1 NA with a 150-cavity than group 2 NA with no 150-cavity. Furthermore, we have found that the laninamivir octanoate prodrug has a unique binding mode in p09N1 that is different from that of group 2 p57N2, but with some similarities to NA-oseltamivir binding, which provides additional insight into group specific differences of oseltamivir binding and resistance.
| The influenza neuraminidase (NA) enzyme is the most successful drug target against the seasonal and pandemic flu. The 2009 H1N1 flu pandemic led to record sales of the NA inhibitors oseltamivir (Tamiflu) and zanamivir (Relenza). Recently, a new drug, laninamivir (Inavir), has been approved for use in Japan can also be administered effectively using a single dose via its octanoate prodrug (CS-8958), however its effectiveness against oseltamivir-resistant influenza infection has not been demonstrated in clinical studies. In this study we comprehensively evaluate the effectiveness of laninamivir and its prodrug using NA from different groups with different active site features. We expressed and purified a group 2 NA from the 1957 pandemic H2N2, an atypical group 1 NA from the 2009 H1N1 pandemic and a group 1 NA from avian H12N5. NA inhibition was assayed and NAs were further crystallized with each inhibitor to determine the structural basis of their action. We found that laninamivir inhibition is highly potent for each NA, however binding and inhibition of laninamivir and its prodrug showed group specific preferences. Our results provide the structural and functional basis of NA inhibition using classical and novel inhibitors, with NAs from multiple serotypes with different properties.
| The 2009 pandemic swine origin influenza A H1N1 virus (pH1N1) has reminded the world of the threat of pandemic influenza [1], [2], [3]. In 2009, the total sales of Tamiflu (oseltamivir phosphate) increased to over 3 billion US dollars (Annual General Meeting of Roche Holding Ltd, 2 March 2010). The total sales of Relenza (zanamivir) in 2009 were over 1 billion (GlaxoSmithKline Quarter 4 Report, 4 February 2010). Additionally, 5.65 million packs of Tamiflu were donated to the WHO in 2009 to replenish their stockpiles (Roche, Annual General Meeting of Roche Holding Ltd, 2 March 2010). Since the WHO has downgraded the threat of pH1N1 from the pandemic level in August 2010, there have still been ongoing reports of pH1N1 outbreaks in south-eastern states of the USA, India and New Zealand (US CDC). Furthermore, a new variant of pH1N1 has even been detected in Singapore, New Zealand and Australia [4]. Throughout the world, vaccinations have still been strongly advocated and stockpiles of oseltamivir and zanamivir are on reserve in case of another severe influenza outbreak in the near future. Both oseltamivir and zanamivir are excellent examples of modern structure-based drug-design and function as competitive inhibitors of the influenza neuraminidase (NA), and are by far the most commonly used influenza drugs [5], [6], [7], [8].
Influenza A virus contains two proteins on its surface in addition to the ion channel M2: hemagglutinin (HA) and NA [9]. Both M2 and NA are targets for clinically-available influenza drugs, however M2 drugs are rarely used anymore because M2 develops drug-resistant mutations very easily [10]. In the influenza virus infection life cycle, HA binds to terminally linked sialic acid receptors on the surface of host cells, allowing the virus to gain entry. In order for the influenza virus to efficiently break free from already infected cells and to continue replicating, sialic acid containing HA receptors must be destroyed. NA, which is a sialidase, catalyzes hydrolysis of terminally linked sialic acid and functions as the receptor-destroying element of influenza A and B viruses.
Influenza A NA has been grouped into 9 different serotypes, N1-N9, based upon antigenicity [11]. Additionally, influenza A NA is further classified into two groups: group 1 (N1, N4, N5 and N8) and group 2 (N2, N3, N6, N7 and N9), based upon primary sequence [12]. Group 1 NAs contain a 150-cavity (formed by amino acids 147–151 of the 150-loop) in their active site, whereas group 2 NAs lack this cavity [12]. Coordination of the 150-loop with the 430-loop appears to be critical for the formation of the 150-cavity [13], [14]. Soaking experiments of typical group 1 NAs with inhibitors often result in the closure of the 150-cavity and indicates some flexibility of the 150-loop [12], [15]. Molecular dynamics simulations also indicate some differences in the flexibility of the 150-loop between group 1 and group 2 NAs [14], [16]. Structural studies reveal that Asp151 and Arg152 of the 150-loop form key interactions with the 4-group and N-acetyl group of common NA ligands, respectively. These two residues move away from the substrate in the open conformation of the 150-loop and closer upon ligand binding [17]. Therefore the 150-loop plays an essential role in substrate and inhibitor binding [15]. Furthermore, the 150-cavity is currently being successfully explored as a target for novel NA inhibitors [12], [18], [19], [20].
The design of NA inhibitors is a classic example of structure-based drug discovery, pioneered by Mark von Itzstein and colleagues with the advent of the N2, N9 and influenza B NA structures [5], [6], [8], [21], [22], [23], [24]. Currently there are four NA-targeting inhibitors that have been approved for use: zanamivir, oseltamivir, peramivir and laninamivir (laninamivir has recently been approved in Japan). Additionally, there are many more NA inhibitors under clinical trials or under vigorous development due to the public threat of seasonal and pandemic flu and the rise of drug-resistant viruses [18], [19], [25], [26], [27], [28]. However, previous results have indicated that inhibitors which are highly similar to the natural NA ligand, sialic acid, are less susceptible to the problem of drug-resistance [29], [30], [31]. This suggests that drugs like zanamivir, that are similar to sialic acid and its transition state analogue 2-deoxy-2,3-dehydro-N-acetylneuraminic acid (Neu5Ac2en or DANA), have an advantage over oseltamivir, which is less similar (Figure 1). However, zanamivir must be administered twice daily over 5 consecutive days to attain its maximum effect. Therefore, the development of novel inhibitors that possess long term efficacy and that are also effective against oseltamivir-resistant influenza viruses is in great demand.
Laninamivir (R-125489) is a very promising, novel influenza NA inhibitor with high potency and the ability to efficiently inhibit common oseltamivir-resistant viruses, including those with the ubiquitous His274Tyr substitution [32], [33], [34]. Recently, laninamivir and its prodrug, laninamivir octanoate (CS-8958) have been approved for use in Japan as Inavir (Daiichi Sanko Press Release, 10 Sept. 2010). Clinical studies have confirmed that the prodrug, laninamivir octanoate, is effective in both children and adults, however laninamivir octanoate has not yet been demonstrated to be more effective than oseltamivir against oseltamivir-resistant His274Tyr H1N1 infection in adult patients [35], [36], [37]. Like zanamivir, the core structure of laninamivir is Neu5Ac2en, the NA transition state analogue (Figure 1). Both laninamivir and zanamivir contain a 4-guanidino group that is not present in Neu5Ac2en and laninamivir also contains an additional 7-methoxy group (Figure 1). Laninamivir octanoate is the octanoyl prodrug of laninamivir (Figure 1). In a similar manner that oseltamivir is processed to oseltamivir carboxylate in the liver, it has been demonstrated that laninamivir octanoate is processed to laninamivir in the lung [33]. The laninamivir 7-methoxy and its prodrug octanoyl ester increase the ability to be retained in the lungs and to function effectively in a single inhaled dose [32], [33], [34], [37], [38]. Moreover, the high similarity of laninamivir to the NA transition state analogue, Neu5Ac2en, allows for an effective response against oseltamivir-resistant NA [32], [33], [34], [35]. For these reasons, laninamivir and laninamivir octanoate offer advantages over both oseltamivir and zanamivir.
In order to comprehensively assess the effectiveness of the novel NA inhibitors, laninamivir and laninamivir octanoate, in comparison to oseltamivir and zanamivir, and to reveal the structural basis of their inhibition, we utilized: 1) pandemic A/RI/5+/1957 H2N2 N2 (p57N2) as a typical group 2 NA, 2) p09N1 as an atypical group 1 NA, and 3) avian H12N5 NA (N5) as a typical group 1 NA. Soluble, active p57N2, p09N1 and N5 were expressed in a baculovirus expression system and purified based upon previously reported methods [13], [39], [40]. NA inhibition assays were carried out and complex crystal structures were solved for laninamivir, laninamivir octanoate, zanamivir and oseltamivir in order to elucidate the structural basis of their inhibition. Our results indicate that laninamivir is potent against all 3 NAs with a similar binding mode to zanamivir. Laninamivir and zanamivir were more effective against group 1 N5, with a 150-cavity, than atypical group 1 p09N1 and group 2 p57N2, with no 150-cavity. This indicates that the ability of the bulky 4-guanidino group of zanamivir and laninamivir to become buried deep beneath the 150-loop is an important factor for their group-specific binding and inhibition. Furthermore, we confirm the binding of the prodrug, laninamivir octanoate, to p57N2, with a similar binding mode to laninamivir. Surprisingly, the p09N1-laninamivir octanoate complex shows a completely different binding mode: p09N1 adopts a Glu276-Arg224 salt bridge in its laninamivir octanoate complex, forming a hydrophobic pocket that is also necessary to accommodate oseltamivir. The observation of different Glu276 rotation in p09N1 and p57N2 offers insight into the group specific differences of oseltamivir binding and resistance.
In our previous studies, we have successfully obtained both soluble p09N1 and N5 using a baculovirus expression system originally developed by Xu et al. [13], [15], [39]. To determine the functional and structural basis of NA inhibition and binding by laninamivir and its prodrug, we first expressed and purified a new group 2 member from the 1957 pandemic H2N2 virus, p57N2, using similar methods. In this way, three major types of known NAs are covered in this comprehensive analysis: typical group 2 p57N2, atypical group 1 p09N1 and typical group 1 N5.
p57N2 was crystallized and its structure was solved by molecular replacement using A/TOKYO/3/1967 (H2N2) N2 (PDB code: 1IVG) as a search model [41]. As expected, the active site of p57N2 is highly similar to other group 2 NAs in that it has no 150-cavity (Figure 2 - upper left). Like the available group 2 A/TOKYO/3/1967 (H2N2) N2 and A/Memphis/31/98 (H3N2) N2 structures [42], [43], p57N2 also contains a 150-cavity deficient active site with a salt bridge between Asp147 and His150, confirming the presence of a stable, closed 150-loop (Figure 2 - upper left; Figure 3A).
Although the atypical group 1 p09N1 also has a 150-cavity deficient active site (Figure 2 - upper right), the 150-loop is quite different from that of p57N2. The p09N1 150-loop sequence (residues 147–150) is GTIKD, however p57N2 contains DTVHD with 3 polymorphic amino acids. The p09N1 therefore contains no Asp147-His150 salt bridge, but instead contains Ile149, which is commonly found in group 2 NAs, and Ile149 is able to rest closer to the hydrophobic Pro431 than Val149 is [13]. N5 on the other hand contains Val149 with no 147–150 salt bridge and displays a 150-cavity like all other structure-known NAs with Val149 and no 147–150 salt bridge (Figure 2 - lower left) [15]. Therefore, NAs with the three major styles of the 150-loop are covered in our comparative analysis.
All NA proteins produced in the baculovirus expression system displayed stable sialidase activity. IC50 values and 95% confidence intervals (CIs) are given in Table 1. Oseltamivir inhibited the activity of N5, p09N1 and p57N2 with IC50 values of 0.83 nM, 0.54 nM and 0.79 nM, respectively. Laninamivir inhibition was best for group 1 N5, followed by atypical group 1 p09N1 and worst for group 2 p57N2. Zanamivir was also more effective against N5 than p09N1 and p57N2, however the difference of zanamivir inhibition between p09N1 and p57N2 is not statistically significant (Table 1). Zanamivir inhibited N5, p09N1 and p57N2 with IC50 values of 0.59 nM, 1.11 nM and 1.36 nM, respectively. Laninamivir was in a similar range with zanamivir for N5, p09N1 and p57N2 with IC50 values of 0.90 nM, 1.83 nM and 3.12 nM, respectively. However inhibition of laninamivir was 1.53, 1.65 and 2.29 fold lower than zanamivir for N5, p09N1 and p57N2, respectively. Inhibition of N5, p09N1 and p57N2 by laninamivir octanoate was not as efficient, with IC50 values of 389 nM, 947 nM, and 129 nM, respectively. Hence inhibition of p57N2 by laninamivir octanoate was much better than for p09N1.
To determine the structural basis of the inhibition of laninamivir relative to zanamivir, we solved the very first complex structures of laninamivir with p57N2, p09N1 and N5 at resolutions of 1.8 Å, 1.8 Å and 1.6 Å, respectively; and the complex structures of zanamivir with p57N2, p09N1 and N5 at 1.9 Å, 1.9 Å and 1.6 Å, respectively [15]. Like zanamivir, the binding mode of laninamivir to all 3 NAs is highly similar to that of the NA transition state analogue, Neu5Ac2en. Some minor differences in the NA-inhibitor interactions between laninamivir and zanamivir are observed within each of the 3 NAs due to the additional hydrophobic 7-methoxy group of laninamivir; however all of the laninamivir complex structures highly resemble zanamivir binding (Table 2, Figure 3).
Due to the similar binding modes of zanamivir and laninamivr, we first carried out a detailed analysis of interactions with the 150-loop in each inhibitor complex. In all of the zanamivir and laninamivir structures, the 4-guanidino group is buried deep beneath the 150-loop where it forms many key hydrogen bonds with Glu119, the Trp178 peptide carbonyl, Glu227, and the Asp151 side chain and peptide carbonyl (Figure 3, Table 2). This 4-guanidino group is the most buried part of the inhibitor in the structure (Figure 3), which is emphasized by the absence of any water molecules beneath the 150-loop and surrounding the 4-guanidino group. Although the 4-guanidino plays an essential role for the high affinity of laninamivir and zanamivir to NA, accessibility of the 4-guanidino to its binding site deep below the 150-loop is a crucial factor for the laninamivir and zanamivir binding process. The typical group 1 N5 contains a 150-cavity in its uncomplexed structure and inhibition of N5 by laninamivir and zanamivir was better than inhibition of p09N1 and p57N2, which contain no 150-cavity in their uncomplexed structures (Figure 3, Table 1). Therefore, our data indicate that the group specific accessibility of the laninamivir and zanamivir 4-guanidino to the NA active site is a key factor in determining their effectiveness.
Slight differences in the interactions between the binding of laninamivir and zanamivir were observed due to the additional laninamivir 7-methoxy group (Table 2). Although this laninamivir 7-methoxy group is oriented away from its own ring oxygen and is pointed toward the hydrophobic Ile222 side chain, its distance is relative far at over 5 Å. Interactions between Arg371 and the inhibitor carboxylate were always highly consistent; however the carboxylate-Arg118 interactions are closer in zanamivir than laninamivir in every NA complex (Table 2). On the other hand, the carboxylate-Arg292 interactions are further in zanamivir than laninamivir in every NA complex (Table 2). Unlike p09N1 and p57N2, N5 contains Tyr347, which forms an additional hydrogen bond with the carboxylate of each inhibitor (Figure 3C).
Laninamivir octanoate complex structures with p09N1 and p57N2 (Figure 4) were solved at 1.6 Å and 2.2 Å, respectively, demonstrating that the laninamivir octanoate prodrug can also directly inhibit NA without further processing. In p57N2, laninamivir octanoate binds in a similar manner to laninamivir with an additional, novel hydrogen bond between the 9-ester carbonyl and Arg224 (Figure 4A). p09N1, on the other hand, has a totally different binding mode where the prodrug's ester is oriented toward Asn294 rather than Arg224 (Figure 4B). p09N1 Glu276 is also in a different orientation in the laninamivir octanoate complex structure than in the zanamivir or laninamivir complex structures and forms a salt bridge with Arg224 in the same manner as oseltamivir binding (Figure 4 and 5). The rotation of p09N1 Glu276 places it out of range for hydrogen bonding with the 8-OH and 9-ester-O of laninamivir octanoate. Instead, the p09N1-laninamivir octanoate 9-ester-O forms a unique hydrogen bond with Asn294 (Figure 4B). Additionally, 09N1 Ser247 forms another hydrogen bond with the laninamivir octanoate 9-ester-O at 3.4–3.5 Å. Still, the Glu276 rotation results in less hydrogen bonding in the p09N1-laninamivir octanoate complex compared to p57N2 (Figure 4C, Table 2).
In both structures, there is no observed electron density corresponding to the octanoyl carbon chain indicating that this part of the molecule is highly flexible and does not engage many stable hydrophobic interactions with p09N1 or p57N2. Still, electron density surrounding the entire ester can be observed in both complex structures. Furthermore, in p09N1, the position 7-methoxy of laninamivir octanoate is also oriented slightly away from its N-acetyl group relative to laninamivir and there is additional electron density pointing toward the ring, indicating lower stability of the p09N1-laninamivir octanoate complex (Figure 4). In all of our NA complex structures, bond distances in each molecule of the asymmetric units are very similar, however in the p09N1-laninamivir octanoate structure some greater differences are observed between molecule A and B in the asymmetric unit, which further reflects the lower stability of the prodrug's octanoyl ester in p09N1. In p09N1 molecule A, the distance between the 9-ester-O and Asn294 is 2.64 Å, however in molecule B the distance is much further at 3.93 Å (Table 2).
To our surprise, the binding mode of the p09N1-laninamivir octanoate complex structure is similar to all known NA-oseltamivir complex structures with respect to the Glu276-Arg224 interactions. Therefore, we also solved the p09N1 oseltamivir complex structure at a resolution of 1.7 Å. As observed in the other available oseltamivir-NA complex structures, in the p09N1-laninamivir octanoate complex structure, Glu276 indeed forms a salt bridge with Arg224, creating a hydrophobic pocket which accommodates the hydrophobic oseltamivir pentyl ether group (Figure 5) [7], [29], [44], [45]. This hydrophobic side chain of oseltamivir is favorably parallel to the Cβ and Cγ of Glu276 on one end and at the other end is pointed toward the hydrophobic Ile222 side chain, which contributes significantly to the high level of oseltamivir inhibition.
Recent studies have demonstrated that the novel influenza A virus NA inhibitors, laninamivir and laninamivir octanoate, are highly effective and have some advantages over zanamivir and oseltamivir [33], [34], [35], [37], [38]. In this study we verify that laninamivir, which highly resembles the NA transition state analogue, Neu5Ac2en, is indeed effective at inhibiting highly purified p57N2, p09N1 and N5, representing the three major types of all structure-known NAs with distinct 150-loop properties.
Zanamivir and laninamivir are clearly more similar to sialic acid and Neu5Ac2en, than oseltamivir, which renders zanamivir and laninamivir less susceptible to drug-resistance and effective against many oseltamivir-resistant viruses [29], [30], [31]. The high degree of similarity between the binding modes of zanamivir and laninamivir in all of the NA complex structures indicates that zanamivir and laninamivir should be effective against the same drug-resistant mutations. However, laninamivir contains an additional 7-methoxy group which is oriented toward Ile222. Although the distance between the laninamivir 7-methoxy and Ile222 is relative far (over 5 Å), laninamivir may be susceptible to Ile222Arg, a rare drug-resistant substitution [46]. Moreover, the additional 7-methoxy group of laninamivir may disrupt hydrogen bonding of the 7-O with water and likely contributes to a slightly lower inhibition of laninamivir compared to zanamivir that was observed for all 3 NAs here (Table 1).
Although zanamivir and laninamivir are highly similar to Neu5Ac2en, they both contain an artificial bulky 4-guanidino group. Upon binding, this 4-guanidino group becomes buried deep beneath Asp151 of the closed 150-loop and forms many hydrogen bonds which contribute to the high affinity of zanamivir and laninamivir to NA. However, the bulky 4-guanidino must be able to clear the 150-loop in order to bind NA and therefore a closed 150-loop may hinder the entry of zanamivir and laninamivir into the NA active site. In the open state of the 150-loop, when the 150-cavity is formed, Asp151 shifts over 1.5 Å (the Asp151 Cγ is shifted over 2 Å in N5) away from the ligand binding site, which should facilitate entry of inhibitors like zanamivir and laninamivir [12], [15], [17]. A similar model has recently been proposed by Wang et al., however this was based on a computer simulation using only the group 1 H5N1 NA structure [47].
The group specific 150-loop accessibility, based upon our structures of p57N2, p09N1, and N5, is consistent with the inhibition efficiency of laninamivir and zanamivir. Group 2 p57N2 contains an Asp147-His150 salt bridge, limiting the flexibility of its closed 150-loop and inhibition of p57N2 by laninamivir was the lowest (Figure 3A). p09N1 is an atypical group 1 with no 150-cavity, but no Asp147-His150 salt bridge, and inhibition of p09N1 by laninamivir was better than p57N2 (Figure 3B). The typical group 1 N5 contains a 150-cavity in its uncomplexed structure and inhibition of N5 by both laninamivir and zanamivir was the highest (Figure 3C). Therefore, we provide structural and functional evidence that the open 150-loop of a typical group 1 NA may facilitate the entry of the 4-guanidino group of zanamivir and laninamivir into the NA active site, relative to the closed 150-loop of group 2 NAs. The additional hydrogen bond between Tyr347 and the inhibitor carboxylate is also a key factor in explaining the higher N5 inhibition relative to p09N1 and p57N2. However, like the closed 150-loop, this residue also makes the active site cavity smaller and in this way may also limit access of inhibitors to the N5 active site. Furthermore, this residue is found only in group 1 NAs which contain an open 150-loop cavity [12]. Thus, Tyr347 may compensate for the open 150-loop in regards to substrate binding.
The complex structure of p57N2 with the laninamivir octanoate prodrug has a similar binding mode to laninamivir and zanamivir, however laninamivir octanoate in complex with p09N1 is completely different. This is the first time, as far as we know, that the same inhibitor has been observed to bind in two completely different conformations to influenza NAs. Additionally, p57N2 Arg224 forms a unique hydrogen bond with the laninamivir octanoate 9-ester carbonyl, and p09N1 Asn294 and Ser247 form unique hydrogen bonds with the laninamivir octanoate 9-ester-O. However, the novel conformation of the laninamivir octanoate-p09N1 complex disrupts any hydrogen bonding with Glu276. The overall lack of hydrogen bonds and instability in the p09N1-laninamivir octanoate structure relative to p57N2 provides the structural basis for higher laninamivir octanoate inhibition of p57N2 observed in our study and a previous report demonstrating better laninamivir octanoate inhibition of H2N2 and H3N2 viruses over H1N1 viruses [34]. The absence of any electron density surrounding the octanoyl carbon chain of laninamivir octanoate indicates that it is unable to take part in any favorable interactions with p57N2 and p09N1. The disordered octanoyl carbon chain likely destabilizes the interactions between the NA active site and the laninamivir octanoate 8-OH and 9-ester, which is indicated by the lower electron density surrounding the 9-ester. Therefore, the lower inhibition efficiency of laninamivir octanoate relative to laninamivir is not surprising.
Binding of oseltamivir to p09N1 was indeed highly similar to the binding observed in previous reports and is also similar to the binding mode of laninamivir octanoate to p09N1. Oseltamivir contains a 4-amino group, instead of the 4-guanidino group found in zanamivir and laninamivir, and is actually more similar to the natural ligand in this regard. Therefore, the orientation of the 150-loop during oseltamivir binding is not a major factor. Instead, the binding preference of oseltamivir for p09N1 over p57N2 and N5 may be instead explained by the ability of Glu276 to adopt the conformation that is critical to accommodate the osetalmivir pentyl ether side chain, which replaces the glycerol moiety of zanamivir, laninamivir and sialic acid. The observation that this Glu276 conformation occurs in the p09N1-laninamivir octanoate complex, but not the p57N2-laninamivir octanoate complex may indicate that this conformation is more stable in p09N1 after ligand binding which may explain why inhibition of p09N1 by oseltamivir was the best relative to N5 and p57N2.
In addition, this observation of different Glu276 dynamics in group 1 p09N1 compared to group 2 p57N2 offers some new insights into the group specificity of the oseltamivir-resistant His274Tyr substitution. The His274Tyr mutation is easily selected for N1 viruses, however cannot be selected for N2 virus types as N2 His274Tyr binding to oseltamivir is not impaired [48]. In group 2 NAs, Tyr274 is able to move away from Glu276 due to a small neighboring Thr252 residue, and oseltamivir can still bind for His274Tyr [30]. The native His274 is also further away from Glu276 in our p57N2-laninamivir octanoate structure and does not hydrogen bond with it. In group 1 NA, Tyr274 is not able to move away from Glu276 because of the bulky neighboring Tyr252 side chain, which prevents it from accommodating oseltamivir [30]. In a similar manner, the group 1 Tyr252 side chain promotes the native His274 to occupy a position where it can participate in a hydrogen bond network with Glu276 and Arg224 as observed in our 09N1-laninamivir octanoate structure (Figure 4C).
Recently, a clinical study has shown that laninamivir octanoate is not significantly better than oseltamivir against oseltamivir-resistant His274Tyr H1N1 infection in adult patients [36]. Since the laninamivir octanoate prodrug binds to p09N1 in a similar manner to oseltamivir, this may offer some explanation as to why laninamivir octanoate has a similar effect as oseltamivir against His274Tyr H1N1. However, this may indicate that the laninamivir octanoate is not processed, or processed slowly, to laninamivir in the adult patients from this study, since laninamivir has been clearly demonstrated to be effective against the oseltamivir-resistant His274Tyr influenza A viruses [32], [33], [34]. Further investigation into the efficacy of laninamivir octanoate in adults in clearly needed.
The results from this comprehensive analysis of group 2 p57N2, atypical group 1 p09N1 and typical group 1 N5 support the hypothesis that influenza NA inhibitors which more closely resemble the NA transition state analogue, Neu5Ac2en, are more likely to remain effective against NAs from both groups and with various drug-resistant amino acid substitutions. Most importantly, we provide mechanisms to explain the group 1 preference of laninamivir and zanamivir and the differential binding of the octanoate prodrug to group 1 p09N1 and group 2 p57N2 derived from pandemic influenza viruses.
Methylumbelliferyl-N-acetylneuraminic acid (MUNANA) was purchased from J&K Scientific Ltd. Sialic acid (Neu5Ac) was purchased from Sigma (Cat. No. 855650) and used without further purification. Laninamivir, laninamivir octanoate, zanamivir and oseltamivir were readily synthesized according to the relevant literatures [49], [50], [51], [52], [53]. All products were characterized by their NMR or MS spectra.
NA was prepared in a baculovirus expression system according to methods based on an original method reported by Xu et al. [39]. Both N5 and p09N1 were prepared as previously described in our laboratory [13], [15]. For p57N2, the cDNA encoding amino acid residues 83–469 were recombined into the baculovirus transfer vector pFastBac1 (Invitrogen), with a GP67 signal peptide, a 6X his-tag, a tetramerizing sequence and a thrombin cleavage site at the N-terminus. Recombinant baculovirus was prepared based on the manufacturer's protocol (Invitrogen). Sf9 suspension cultures were grown in Sf-900 II SFM serum-free media (GIBCO) at 28°C and 120 RPM and transfected with high-titer recombinant baculovirus. After growth of the transfected Sf9 suspension cultures for 3 days, centrifuged media were applied to a HisTrap FF 5 mL column (GE Health) which was washed with 20–50 mM imidazole, and then NA was eluted using 200–300 mM imidazole. After dialysis, thrombin digestion (Sigma, 3 U/mg NA; overnight at 4°C) and gel filtration chromatography using a Superdex-200 10/300 GL column (GE Healthcare), NA fractions were analyzed by SDS-PAGE. High-purity NA fractions were pooled and concentrated using a membrane concentrator with a molecular weight cutoff of 10 kD (Millipore). A buffer of 20 mM Tris-HCl, 50 mM NaCl, pH 8.0 was used for gel filtration and protein concentration.
A neuraminidase inhibition assay using MUNANA was performed as described by Potier et al. with modifications [54]. Briefly, 10 uL of purified, recombinant NA (10 nM) was mixed with 10 uL of inhibitor and incubated for 30 min at room temperature. NA and inhibitors were carefully diluted in fresh PBS buffer. At least 5 concentrations of each inhibitor at an appropriate range were used for each repeat. Following incubation, 30 uL of 166 uM MUNANA in 33 mM MES and 4 mM CaCl2 (pH 6.0) was added to the solution to start the reaction using a 12-tip pipette (Eppendorf). A positive and a negative control were included in each 12-well lane. After starting the reaction for each lane on the plate, the reaction mixture was immediately loaded on a SpectraMax M5 (Molecular Devices) where fluorescence was quantified over the course of 30 min at an excitation wavelength of 355 nm and an emission wavelength of 460 nm. Single time points were chosen where the positive control produced a fluorescence signal of approximately 1,000. All assays were done in triplicates and IC50 values for each inhibitor were calculated with sigmoidal fitting of the log[inhibitor] vs. inhibition percentage using GraphPad Prism.
NA crystals were grown using the hanging-drop vapor diffusion method. Initial screening was performed using a commercial kit (Hampton Research). Diffraction quality crystals of p57N2 were obtained by mixing 1 uL of the concentrated protein at 10 mg/mL in 20 mM Tris, pH 8.0, and 50 mM NaCl with 0.1M BIS-TRIS propane (pH 9.0), 10% v/v Jeffamine ED-2001 (pH 7.0). N5 crystals were obtained using 0.1 M HEPES (pH 7.5), 12% w/v polyethylene glycol 3,350 at 18°C [15]. Quality p09N1 crystals were obtained as described previously using 0.16 M calcium acetate hydrate, 0.08 M sodium cacodylate trihydrate, pH 6.5, 14.4% polyethylene glycol 8000, 20% glycerol at 18°C [13]. NA protein crystals were first incubated in mother liquor containing 20 mM of inhibitor, and then flash-cooled at 100 K. Diffraction data for the p57N2 and N5 were collected at KEK beamline Ne3A, while p09N1 data were collected at SSRF beamline BL17U.
Diffraction data were processed and scaled using HKL2000 [55]. Data collection and processing statistics are summarized in Table 3. The structure of p57N2 was solved by molecular replacement method using Phaser [56] from the CCP4 program suite [57] with the structure of A/TOKYO/3/1967 H2N2 N2 (PDB code: 1IVG) as the search model [41]. Initial restrained rigid-body refinement and manual model building were performed using REFMAC5 [58] and COOT [59], respectively. Further rounds of refinement were performed using the phenix.refine program implemented in the PHENIX package with coordinate refinement, isotropic ADP refinement and bulk solvent modeling [60]. The stereochemical quality of the final model was assessed with the program PROCHECK [61]. The final models have 84% of the residues in the most favored region of the Ramachandran plot [62] and no residue in disallowed regions. Structures of p09N1 and N5 were solved as described previously [13], [15].
All crystal structures have been deposited into the Protein Data Bank (PDB, www.pdb.org) with the following PDB codes: N5-laninamivir - 3TI8, p09N1-zanamivir - 3TI5, p09N1-laninamivir - 3TI3, p09N1-laninamivir octanoate - 3TI4, p09N1-oseltamivir - 3TI6, p57N2-zanamivir - 3TIC, p57N2-laninamivir - 3TIA, and p57N2-laninamivir octanoate - 3TIB.
|
10.1371/journal.pcbi.1004611 | Unbiased Rare Event Sampling in Spatial Stochastic Systems Biology Models Using a Weighted Ensemble of Trajectories | The long-term goal of connecting scales in biological simulation can be facilitated by scale-agnostic methods. We demonstrate that the weighted ensemble (WE) strategy, initially developed for molecular simulations, applies effectively to spatially resolved cell-scale simulations. The WE approach runs an ensemble of parallel trajectories with assigned weights and uses a statistical resampling strategy of replicating and pruning trajectories to focus computational effort on difficult-to-sample regions. The method can also generate unbiased estimates of non-equilibrium and equilibrium observables, sometimes with significantly less aggregate computing time than would be possible using standard parallelization. Here, we use WE to orchestrate particle-based kinetic Monte Carlo simulations, which include spatial geometry (e.g., of organelles, plasma membrane) and biochemical interactions among mobile molecular species. We study a series of models exhibiting spatial, temporal and biochemical complexity and show that although WE has important limitations, it can achieve performance significantly exceeding standard parallel simulation—by orders of magnitude for some observables.
| Stochastic simulations (simulations where randomness plays a role) of even simple biological systems are often so computationally intensive that it is impossible, in practice, to simulate them exhaustively and gather good statistics about the likelihood of different outcomes. The difficulty is compounded for the observation of rare events in these simulations; unfortunately, rare events, such as state transitions and barrier crossings, are often those of particular interest. Using the weighted ensemble (WE) method, we are able to enhance the characterization of rare events in cell biology simulations, but in such a way that the statistics for these events remain unbiased. The histogram of outcomes that WE produces has the same shape as a naive one, but the resolution of events in the tails of the histogram is greatly improved. This improved resolution in rare event statistics can be used to infer unbiased estimates of long timescale dynamics from short simulations, and we show that using a weighted ensemble can result in a reduction in total simulation time needed to sample certain events of interest in spatial, stochastic models of biological systems.
| Stochastic effects are of crucial importance in many biological processes, from protein dynamics [1], to gene expression [2], to phenotypic heterogeneity [3]. Unfortunately, due to the high computational cost of simulating complex stochastic biological systems, the effects of stochasticity on system response remain under-studied in realistic biological models.
From molecular to cellular scales, simulations of biological systems push the limits of our computational resources [4, 5]. Compromising between sampling power and model complexity will be a trade-off for the foreseeable future; for example, at atomistic resolution even the most powerful, specially designed supercomputers can simulate only modestly sized proteins at timescales that approach sufficiency for adequate sampling [6]. Similarly, models of cellular processes, though they omit entirely molecular-level details, are also constrained in complexity and realism by the need to perform adequate amounts of simulation in order to gather useful statistics [7]. Mixing scales in a simulation, though perhaps necessary for capturing the coupling across multi-scale networks, only makes this problem worse.
Enhanced sampling algorithms offer an attractive proposition: instead of compromising on model complexity in order to achieve well-sampled results, rather use simulation resources more effectively and extract more information given the same resources. Not surprisingly, there has been significant interest in sampling algorithms in the field of atomistic protein simulation, including umbrella and histogram sampling [8–10], path sampling methods, [11–17], and various flavors of replica exchange [18–21]. Arguably, such approaches have transformed the field of molecular simulation [6, 22, 23].
The essence of the present study is the extension of one successful enhanced sampling strategy for molecular simulation to spatially resolved cell-scale systems. Specifically, the weighted ensemble approach is a scale-agnostic method that is able to facilitate the enhanced sampling of a wide spectrum of stochastic simulations and non-Markovian processes [17], including Brownian dynamics [13], molecular dynamics [24], Monte-Carlo simulations of atomistic and coarse-grained protein dynamics [25, 26], chemical reaction networks [27], and as we demonstrate here, the spatially resolved stochastic reaction-diffusion processes used to simulate cellular processes. Weighted ensemble achieves its enhanced sampling by dividing up a model’s state-space into bins and maintaining an ensemble of trajectories with different weights that evenly sample these bins. This weighted ensemble is created by resampling the distribution of trajectories at fixed time intervals, spawning new simulations from trajectories that have wandered into unexplored regions and pruning them away if a region is overpopulated, in order to maintain even coverage of the space. This resampling process is exact, in the sense that it induces no bias in the estimates of equilibrium and non-equilibrium observables [17, 28]. Resampling at fixed time intervals lends the method some key benefits: it is trivially parallelizable, since trajectories run independently aside from interacting infrequently during resampling, and it is modular, needing no “under the hood” interaction with the underlying dynamics, rather requiring only intermittent reports of a progress coordinate.
Spatial heterogeneity can be crucial to accurately capturing the behavior of cell-scale biological systems, for instance in models of neuromuscular junction dynamics studied below [29]. Although simple models of biological signaling, where the molecules of interest are spatially homogeneous, or “well-mixed” are very common [30, 31], the assumption of spatial homogeneity may not always be justified; certain biological systems, while suitable for ignoring molecular structure, are not amenable to being modeled as spatially homogenous. Indeed, high resolution microscopy images of single cells show distinct patterns of localization for a wide variety of biomolecules [32–34], leading one to speculate if the well-mixed regime is the exception rather than the rule.
Here, we apply the weighted ensemble sampling procedure to decrease the cost of simulating spatial stochastic systems. After introducing our methodology, we present results for a toy diffusive binding system and two more complex systems: a cross-compartmental signal transduction model in a realistic cellular geometry and a model of an active zone in a frog neuromuscular junction. The flexibility and power of the WE method make it ideally suited for enhancing the sampling of these three diverse models.
We employ the weighted ensemble sampling algorithm to manage multiple instances of particle-based kinetic Monte Carlo simulations of a given spatially resolved model of cellular signaling. We make use of a variety of software packages in our work, all of which are freely available via MMBioS.org.
All simulations in this report employ spatially resolved particle-based kinetic Monte Carlo dynamics, implemented in the MCell software package. MCell (Monte Carlo Cell) is an open source program (MCell.org) that uses spatially realistic 3D cellular models and specialized Monte Carlo algorithms to simulate the movements and reactions of molecules within and between cells, or what is referred to as “cellular microphysiology” [38]. MCell has been used to study a wide range of neuroscience questions such as neurotransmitter diffusion in the brain [39], the structure and function of synapses in the central [40] and peripheral [29] nervous system, and the effect of drugs on nervous system function [41]. MCell has also been employed to investigate general cellular phenomena such as calcium signaling [42] and the role of diffusion in cellular transport [43].
MCell combines rigorously validated and highly optimized stochastic Monte Carlo algorithms, particle-based random walk diffusion of (point particle) molecules in space and on surfaces, and stochastic biochemical state transitions. MCell models can contain arbitrarily complex 3D mesh geometries representing the biological system under consideration. These geometries are typically derived from reconstructions of biological tissue (typically from electron microscopy data) [44], or created in silico based on average geometries [29], e.g. via CellBlender software (github.com/mcellteam/cellblender) [45]. MCell features a flexible model description language and has the ability to checkpoint simulation trajectories at arbitrary output intervals or times.
MCell is a kinetic Monte Carlo scheme, in the sense that the time evolution of the system is explicitly modeled. The Monte Carlo moves that the system makes are not arbitrary trial moves, but are rather chosen according to the reaction and diffusion rates of the molecules being simulated. A constant time-step is employed in these simulations, during which the likelihood of reaction and diffusion processes are computed and stochastically sampled; by using appropriate time-steps, the dynamics of the underlying processes are faithfully recapitulated (for further details, see [38, 46, 47]).
The construction of large, complex spatial models is facilitated by a combination of software that specializes in separate aspects of this task.
One of the limiting factors in performing spatially realistic cell simulations is the difficulty of obtaining cell geometries. This limitation can be addressed by learning generative models of cell organization directly from microscope images; these can be used to synthesize an unlimited number of realistic geometries. For instance, in the complex model in a realistic cellular geometry studied below, biochemical reaction networks, with corresponding compartments for organelles, are constructed using BioNetGen software [48, 49], combined with cell geometry models generated by CellOrganizer software [50–58] using CellBlender [45] to create the MCell spatial simulations [59]. More information about this process of generating cellular instances with realistic cellular and subcellular organizations/morphologies is given below. The WESTPA software in turn manages ensembles of the MCell simulations, for either weighted ensemble or brute-force sampling.
CellOrganizer (CellOrganizer.org) is an open source tool for learning conditional generative models of cellular organization from images [50–58]. From these models, new cellular geometries can be generated from different parts of the “shape space” of the system. Currently CellOrganizer supports models for cell shape, nuclear shape, vesicle frequency, location and size, and microtubule length, number and distribution. Important for this work is CellOrganizer’s ability to produce realistic geometric instances of cells and subcellular components for use in modeling using the experimental spatial extension of the Systems Biology Markup Language (SBML) [60].
Biochemical reaction networks in our model of signaling in a realistic cellular geometry are built with the BioNetGen software package (BioNetGen.org), which is a framework for specifying and simulating rule-based models of biochemical kinetics [48]. The rule-based approach allows combinatorially large chemical reaction networks to be compactly described using a small set of rules that define the underlying molecular interactions [49]. Indirect simulation of rule-based models requires automated generation of the reaction network implied by the rule set. The generated reaction network can then be simulated using a variety of approaches including ordinary differential equations and stochastic simulation. BioNetGen has previously been used to model a wide range of processes including signal transduction, metabolic pathways, and genetic regulatory networks [49]. BioNetGen enables the cellular topology to be defined via compartments [61], but it does not provide for the specification of more detailed geometric information about these compartments or molecule locations. An automated process converts these rules to an exhaustive network of chemical reactions representing the chemical kinetics of the system (see Fig 3).
The reaction network from BioNetGen is fed into CellOrganizer to obtain an appropriate cellular geometry, and the network and geometry are combined using the CellBlender package. In CellBlender, the reactions and geometry are merged, and exported to MCell. The system is then simulated as usual in MCell, either using weighted ensemble to manage the trajectories, or via brute-force.
We investigate three spatial models of cellular function: (1) a toy model of diffusive binding, (2) an idealized model of cellular signaling, and (3) a realistic model of a neuromuscular junction. All three particle-based kinetic Monte Carlo models are simulated in MCell (version 3.2.1), and are available in the supporting information.
A highly simplified model of diffusive binding was constructed as an initial test of the utility of weighted ensemble sampling in a spatial system. The model geometry is depicted in Fig 4.
In this toy model of diffusive binding, we define a cubical volume, of side length 2 microns, on the top of which 1000 ligands are initially bound to 1000 receptors at time t = 0. The volume also contains 1000 receptors at the bottom of the cube that are initially unbound. The ligands are then free to unbind (with a constant of 103/sec), diffuse around the volume (with a diffusion constant of 10−6cm2/sec), and re-bind to receptors at the top, or to receptors at the bottom (with a constant of 108/M/sec). We examine the probability density for the number of receptors at the bottom of the volume bound by ligands after simulating 10 milliseconds of dynamics.
There is significant interest in the variation of cellular morphology and its association with cell fate/function [59, 62–66], and here we employ a model that is a prototype for computationally investigating the effect of a specific geometry upon biological function. The system models protein production in response to an extracellular signal and highlights interesting aspects of signal transduction through different subcellular components, such as transport across membranes and feedback between molecules in different subcellular locations [59]. The model contains on the order of 105 reactive molecules, situated in a realistic cellular geometry. Because creating robust, high-quality complex models of cells is itself a challenging endeavor, we employ the model generation pipeline through BioNetGen and CellOrganizer described in the Methods section and Sullivan et al. [59].
We use the geometry shown in Fig 5, which is derived from three-dimensional images of HeLa cells using CellOrganizer. This geometry contains topologically distinct partitions: the extracellular region, the cytoplasm, the nucleus, and approximately 500 endosomes. The geometry also includes the membranes that partition these compartments, through which molecules must be transported when appropriate. Further details are included in the Supporting Information.
We use the reaction schema illustrated in Fig 6 to describe the reaction kinetics of the model. The BioNetGen rules for this model are included in the Supporting Information, and they produce a network of 354 chemical reactions between 78 species [61]. Briefly, the signaling network functions as follows. The system is initialized in a state of unbound receptors, and free extracellular ligands. The extracellular ligand binds to receptors on the cell membrane, facilitating receptor dimerization, which can be internalized to the endosomes. In the endosomes, receptor dimers can become phosphorylated and recruit a transcription factor, which upon phosphorylation can also dimerize and migrate to the nucleus. In the nucleus, the transcription factor initiates the transcription of mRNA1, which, when it migrates to the cytoplasm, produces protein P1. P1 can then migrate to the nucleus and act as a transcription factor for mRNA2, which, when it migrates to the cytoplasm, produces the final species in the cascade, protein P2. Although this reaction network is idealized, it embodies key aspects of the complexity expected in real signaling processes.
The third model we study represents a single active zone of a frog neuromuscular junction (NMJ). Synapses are of crucial physiological importance in neural function, yet their detailed molecular behavior, particularly the way in which calcium triggers synaptic vesicle fusion still lacks a complete, molecular level, characterization. This is mainly due to the lack of experimental approaches that can probe synapses at the required spatial and temporal resolution. Computational models can provide critical microscopic insight into how calcium binding triggers vesicle fusion and release [29].
The geometry of the frog NMJ active zone model is detailed in Fig 7 and has been described previously [29]. The active zone model consists of a double row of 26 synaptic vesicles and two rows of 26 voltage gated calcium channels (VGCCs) in the space between vesicles (see Fig 7). Thus each synaptic vesicle is associated with a single VGCC.
The system is initialized from a state of no free calcium in the active zone. During a simulation, VGCCs open stochastically, driven by a time-dependent action potential waveform [29]. Once open, VGCCs conduct calcium ions into the presynaptic space. Calcium ions can then freely diffuse and either bind to ∼106 static buffer molecules or one of eight calcium sensor proteins (synaptotagmin) on the synaptic vesicles. Since each synaptotagmin protein has five calcium binding sites, each synaptic vesicle contains a total of 40 calcium binding sites. A synaptotagmin protein is activated after binding at least two calcium ions, and vesicle fusion is triggered once three out of its eight synaptotagmin proteins have been activated. For each simulation we track the calcium binding events to synaptotagmin sites on synaptic vesicles and can thus determine the number of released vesicles and the time of release.
The NMJ model differs crucially from the two other systems studied here in that it possesses rate “constants” that vary in time. Specifically, the rates for the opening of and calcium conduction through VGCCs in the model are time dependent and are parameterized according to an experimentally measured action potential waveform. This time-dependent nature of vesicle release in synapses is critical for their physiological function [29]. Thus, the model, with its time-varying kinetics, cannot be treated using steady-state or equilibrium approaches and is only usefully simulated, even with a weighted ensemble, out of equilibrium and for a predetermined period of time.
We sampled the three spatially resolved cell-scale models of varying complexity using the weighted ensemble approach. The results from all three models demonstrate the ability of WE to sample rare events in models of varying spatial and biochemical complexity. The application of WE sampling to the NMJ model generated novel data about vesicle release in regimes of calcium concentration too difficult to sample well with conventional methods.
Our studies of rare event sampling in spatial stochastic systems start with the toy model shown in Fig 4 and described in detail in the Models section. Briefly, we simulate diffusing ligands unbinding from the top of a cubical volume and binding to the bottom for a short amount of time. In this time-span, it is rare for a large number of the ligands to bind at the bottom of the volume. Indeed, when we simulate the system 611 times via brute-force, we see that in most cases only about 10–20 receptors are bound at the bottom after 10 milliseconds. We simulated 611 brute-force trajectories in order to make a fair comparison of weighted ensemble sampling to a brute-force approach; the single weighted ensemble simulation we performed required computational resources equivalent to 610.7 brute-force simulations. Looking more closely at Fig 8 (see inset), we see that it will be impossible to adequately characterize events rarer than 1/611 via the brute-force ensemble of simulations, since the rarest event one can see with brute-force is equal to the inverse of the number of trajectories. On the other hand, the weighted ensemble approach is able to sample the distribution over many orders of magnitude of probability with an equal amount of computational effort as the brute-force ensemble.
Since a toy model even this simple is too complex to solve exactly, we compare the data from both the single weighted ensemble simulation and the equivalent brute-force simulations to a more authoritative estimate of the probability distribution obtained by exhaustive (weighted ensemble) simulation. To obtain this reference value, we performed 64 independent weighted ensemble simulations with the same parameters as the single “test” weighted ensemble run (blue circles, Fig 8), except that each of the 64 runs had 32 trajectory segments per bin, rather than 16 for the test run (i.e. approximately 128 times the sampling power of the single run). From the 64 independent runs (gray circles, Fig 8), we then computed the 95% confidence interval for the mean probability distribution using 10,000 bootstrap samples at each progress coordinate, from 0 to 70. Even though the exhaustive weighted ensemble runs and the single test run use different weighted ensemble parameters (i.e trajectories per bin), this difference does not substantially affect the sampling quality of the ensembles. Note that the 95% confidence interval for the mean of the true distribution is significantly tighter than the variance of the distribution of weighted ensemble samples of that distribution; the fact that the single run falls outside this interval is typical of the stochastic noise inherent in a single WE sample.
As explained in the Methods section, weighted ensemble is able to sample more of the complete distribution by efficiently spreading out the sampling power of the ensemble of trajectories, allowing the characterization of rare-events by sacrificing some accuracy in the regime where brute-force samples well (see Fig 2). Examining Fig 8, we see that the brute-force distribution is smoother at the peak of the distribution—indicating less uncertainty—but only marginally so; the weighted ensemble estimate of the peak of the distribution is also reasonably smooth. By sacrificing unneeded resolution at the peak, WE is able to instead spread that sampling power more evenly throughout the state-space of the model, using it to sample the full probability distribution more comprehensively.
The model of cellular signal transduction shown in Figs 5 and 6 contains ∼105, reactive molecules in a realistic geometry, and demonstrates the ability of the weighted ensemble sampling approach to scale to large, complex systems. We focus on characterizing the synthesis of protein P2. The production of P2, the last step in the cascade shown in Fig 6, is challenging to sample via brute-force. Nonetheless, it is a crucial quantity to calibrate if one is interested in the effects of spatial heterogeneity on the model, and we do so using weighted ensemble.
To begin our exploration of the signaling model, we initially examine the production of the protein P2 after 400 seconds of simulation (see Fig 9). The weighted ensemble data was produced by two independent runs, and the two resulting independent histograms are shown together. The independent runs allow us to roughly characterize the uncertainty in the estimated probability distribution by simply inspecting the vertical spread in the results.
Detailed exploration of the tail of a probability distribution, as shown in Fig 9, can be interesting in its own right, for instance to detect multimodality, or otherwise explore the state-space for rare but important events. We are also interested in using the high resolution characterization of the tail of the P2 distribution as leverage with which to facilitate estimation of the mean time to the production of five P2 molecules. The target of five P2 molecules was chosen to represent a modest but non-trivial level of P2 production.
To extract information about average P2 production time from short simulations, we work in a steady-state framework, as described in the Methods section. Using this methodology, we are able to infer the mean time to the creation of five P2 molecules, a relatively long timescale, from a weighted ensemble of short simulations. Shown in Fig 10 is probability flux arriving at the target state of five P2 molecules at each WE iteration, as well as a running average of those instantaneous measurements, made using the most recent half of the data up to that time. When the system reaches a steady-state, the inverse of the probability flux into the target state, shown for on the right vertical axis of Fig 10, is equal to the mean time to reach the target state. In Fig 10, we see that the estimated time to the production of five P2 molecules is on the order of 5,000 seconds. This estimate will converge, within stochastic noise, to the true MFPT of the system when the flow of probability induced by the recycling process has relaxed to a steady state; see the Methods section for details.
Although WE is extremely efficient at characterizing the P2 distribution (Fig 9), its performance for estimating the MFPT is not exceptional in this case. The two WE runs require 31,328 seconds (run 1) and 27,408 seconds (run 2) of aggregate simulation to reach the relatively steady estimation shown in Fig 10 at t = 400 seconds. By comparison, to obtain five to ten events for estimating the MFPT by brute-force sampling would require ∼25,000 to ∼50,000 seconds based on the estimated MFPT of ∼5,000 seconds. Note that such long runs would not be able to benefit from parallelization.
The efficiency of the steady-state approach to measuring the mean first passage time depends on the time to convergence, and the noise of the sampling, once converged. The noise of the sampling can be reduced by a more densely sampled weighted ensemble, but the time to convergence is more difficult to characterize. In the approach used here, the latter timescale depends on the waiting time to typical transition events (e.g. about 200 seconds in Fig 10), and the time it takes the system to relax to a steady-state. If these timescales (multiplied by the number of WE trajectories) are close to the timescale of the mean first passage time, then the estimate may not be particularly efficient. It will, however be less variable than a brute-force estimate of equivalent sampling power, and more convenient, in that it explores events of very different likelihood, and efficiently explores the state-space while estimating a key observable.
Finally, we apply weighted ensemble sampling to a model of the active zone of a frog neuromuscular junction. This system, shown in Fig 7, and described in detail in the Models section, simulates the dynamics of vesicle fusion in the presynaptic terminal. The MCell model used in this study is identical to the one described previously [29]. Briefly, calcium molecules are released into the active zone, and are free to diffuse and bind to the calcium binding sites on the synaptic vesicles in response to an action potential. When enough calcium binds to a vesicle in the proper arrangement, the vesicle is considered to have fused.
Calibrating and validating the response of the model against experimental data is of crucial importance, but at low calcium concentrations, it becomes highly inefficient to perform brute-force simulation to gather good statistics. In the neuromuscular junction system, the probability of vesicle fusion depends on the external calcium concentration and falls off sharply as the calcium concentration is decreased.
Fig 11 shows the distribution of times to first fusion in the model, or first passage time (FPT) distribution to fusion, when the external calcium concentration is 0.5 mM and 0.3 mM. At each concentration, we plot the averaged results of 100 weighted ensemble runs, each of which was performed as specified in the Models section, as well as the averaged results of brute-force simulations, which in total required the same computational effort to simulate as the 100 weighted ensemble simulations (7545 brute-force simulations for the 0.3 mM system, 3513 for the 0.5 mM system). The difference with which the two approaches—weighted ensemble and brute-force —are able to capture the shape of the distribution, and the uncertainty in the estimation of it, is striking. We are unaware of any definitive methods of estimating error when the sample yield is extremely low, and hence have omitted error bars when only one or two samples were obtained.
At low calcium concentrations, the overwhelming majority of simulations do not result in a vesicle release, which is why brute-force sampling is so ineffective. Notice that the total area (i.e. the total probability of vesicle fusion) in the histogram for the 0.3 mM condition is only on the order of 10−4. One would have to perform on the order of 100/10−4 = 106 simulations to start gathering meaningful statistics (100 samples) with which to compute the fusion time distribution. This amount of computing (∼20 years running in serial, if each simulation only takes a minute, or ∼20 weeks, running in parallel on a 48-core machine) is unfeasible to perform even once, let alone at all the different settings of model parameters of interest. Using weighted ensemble, however, it becomes practical to sample this model in the low-calcium regime, providing critical information for model validation and fitting purposes. The weighted ensemble sampling for the 0.3 mM condition shown in Fig 11 took time equivalent to 7545 brute-force simulations, and runs in matter of hours in parallel on 48 cores.
Fig 12 summarizes NMJ results at five different experimentally relevant calcium concentrations. The data are a striking recapitulation of an experimentally demonstrated power-law dependence of probability to fuse as a function of calcium ion concentration [67]. Validating the model in low calcium regimes has been intractable with traditional sampling approaches. Using weighted ensemble, we are able to sample the model at all concentrations of interest.
Spatial models of stochastic reaction-diffusion processes have found widespread use as tools in understanding the mechanics of biological processes on the cellular level and beyond [68–72]. Unfortunately, the effective sampling of large, realistic models, and the extraction of well-sampled values of experimentally relevant quantities are often beyond the realm of computational feasibility. We use a weighted ensemble approach to overcome this impediment and demonstrate speedups of orders of magnitude in sampling some observables in complex models of cellular behavior with spatial dependence. Weighted ensemble is an ideal approach to employ in addressing the issue of difficult to sample stochastic systems, and because of its efficiency and ease of use, we anticipate many further applications.
The multi-scale modeling problem posed by constructing accurate, physically realistic models of cellular level processes is considerable. We have demonstrated the utility of sampling spatially inhomogeneous stochastic simulations of cellular processes using a weighted ensemble (WE) approach. Although WE cannot estimate every quantity with high efficiency, estimates for some observables were obtained using orders of magnitude less overall computing than would have been required with conventional parallelization. We hope that these initial results will facilitate the study of more realistic and physically accurate spatial models of biological systems. As an ambitious example, integrating spatial models of stochastic processes with microscopy data of protein localization to predict phenotypic response to the perturbations of interactome networks is an attractive prospect for in silico drug development and personalized medicine. Currently, the bottlenecks in such a scheme are the lack of accurate models and the computational resources with which to simulate them. We hope that this work will contribute to the development of truly physiological computational models.
|
10.1371/journal.ppat.1000569 | Plasmodium falciparum Heterochromatin Protein 1 Marks Genomic Loci Linked to Phenotypic Variation of Exported Virulence Factors | Epigenetic processes are the main conductors of phenotypic variation in eukaryotes. The malaria parasite Plasmodium falciparum employs antigenic variation of the major surface antigen PfEMP1, encoded by 60 var genes, to evade acquired immune responses. Antigenic variation of PfEMP1 occurs through in situ switches in mono-allelic var gene transcription, which is PfSIR2-dependent and associated with the presence of repressive H3K9me3 marks at silenced loci. Here, we show that P. falciparum heterochromatin protein 1 (PfHP1) binds specifically to H3K9me3 but not to other repressive histone methyl marks. Based on nuclear fractionation and detailed immuno-localization assays, PfHP1 constitutes a major component of heterochromatin in perinuclear chromosome end clusters. High-resolution genome-wide chromatin immuno-precipitation demonstrates the striking association of PfHP1 with virulence gene arrays in subtelomeric and chromosome-internal islands and a high correlation with previously mapped H3K9me3 marks. These include not only var genes, but also the majority of P. falciparum lineage-specific gene families coding for exported proteins involved in host–parasite interactions. In addition, we identified a number of PfHP1-bound genes that were not enriched in H3K9me3, many of which code for proteins expressed during invasion or at different life cycle stages. Interestingly, PfHP1 is absent from centromeric regions, implying important differences in centromere biology between P. falciparum and its human host. Over-expression of PfHP1 results in an enhancement of variegated expression and highlights the presence of well-defined heterochromatic boundaries. In summary, we identify PfHP1 as a major effector of virulence gene silencing and phenotypic variation. Our results are instrumental for our understanding of this widely used survival strategy in unicellular pathogens.
| Plasmodium falciparum causes the most severe form of malaria in humans. The high virulence of this unicellular parasite is in part related to the selective expression of members of falciparum-specific gene families. These genes encode proteins that are exported into the cytoplasm and onto the surface of infected red blood cells. To avoid recognition by the host's immune system, P. falciparum employs sequential expression of antigenically different variants of these surface proteins. While the epigenetic mechanisms responsible for such clonal expression have been studied in some detail for the major virulence gene family var, the regulation and function of other exported protein families remain elusive. Here, we identify P. falciparum heterochromatin protein 1 as a major structural component of virulence gene islands throughout the parasite genome. This factor binds specifically to a reversible histone modification, which marks these virulence loci for transcriptional silencing. Our observations suggest a unifying epigenetic strategy in the regulation of host–parasite interactions and immune evasion in P. falciparum. Furthermore, these findings have important implications for the future study of hitherto uncharacterized exported proteins with roles in parasite virulence.
| Plasmodium falciparum causes the most severe form of malaria in humans with over one million deaths annually [1]. Severe and fatal outcomes of infections with this protozoan parasite result from a multitude of syndromes triggered by repeated rounds of asexual reproduction within erythrocytes. After invasion into red blood cells (RBCs), the parasite initiates a dramatic host cell remodeling process, culminating in the export of parasite virulence factors onto the surface of infected RBCs (iRBCs) [2]. The majority of these proteins is encoded by species-specific subtelomeric gene families, some of which underwent massive expansion during the evolution of the P. falciparum lineage [3]. One of the direct consequences of their concerted expression is the sequestration of iRBCs in the microvasculatory system, a process that is linked to severe complications including cerebral and placental malaria [4]–[6]. Sequestration occurs due to interactions of P. falciparum erythrocyte membrane protein 1 (PfEMP1) with various receptors on endothelial cells and uninfected erythrocytes [7]–[10]. 60 PfEMP1 variants are encoded by individual members of the var gene family [11]–[13]. Importantly, only one var gene is transcribed by a single parasite (mutual exclusion) [14]. Switches in var gene transcription occur in situ in absence of any apparent recombination events and result in antigenic variation of PfEMP1 [15]. This clonal phenotypic variation allows the parasite to evade variant-specific humoral immune responses and to sequester in various tissues [16]. Members of some other gene families (rif, stevor, Pfmc-2tm) are also expressed in a restrictive manner [17]–[19], however, their role in parasite biology and the underlying regulatory mechanisms remain unclear.
Multiple var genes are located in most subtelomeric regions directly downstream of telomere-associated repeat elements (TAREs) and in internal clusters on some chromosomes [20]. At either location var genes occur in close association with other variably expressed multi-gene families [13]. Several recent studies investigating the nature of the epigenetic mechanisms involved in the control of mono-allelic var gene transcription revealed an important role of the conserved promoter and intron sequences [21]. var promoters are silenced by default and, notably, activation of an episomal var promoters caused silencing of the entire repertoire of native var genes [22]–[24]. In addition to the 5′ upstream sequences the var gene intron is involved in silencing [25] and, although its exact role in this process remains controversial, this finding has been confirmed several times [26]–[28]. Fluorescent in situ hybridisation (FISH) experiments revealed that P. falciparum chromosome ends occur in perinuclear clusters [29],[30]. Consequently, subtelomeric var genes are inherently positioned at the nuclear periphery which is linked to enhanced transcriptional silencing in other eukaryotes [31]–[33]. Moreover, this spatial association was also demonstrated for chromosome-internal var genes [23],[34]. Transgenes inserted into subtelomeric repeat regions were transcriptionally silenced in a metastable fashion [35], similar to position-effect variegation in yeasts and higher eukaryotes [36]. The process of var gene activation thus appears to be linked to nuclear re-positioning of a silenced locus into a transcriptionally active zone lending support for the existence of a specialized perinuclear region dedicated to mutually exclusive var transcription [23],[35],[37]. The perinuclear location of var genes is clearly independent of their transcriptional state, however, conflicting results exist as to whether var gene activation occurs within or outside chromosome end clusters [23],[30],[34],[35],[37]. Together, this complex architectural setup provides a dynamic foundation for the heritable and variegated silencing of var genes and other variably expressed gene families by epigenetic control processes.
The current data is consistent with facultative heterochromatin-based silencing where dynamic alterations in local chromatin structure act as the major regulatory mechanism. Many of the reversible histone modifications characteristic for active or silenced chromatin in other eukaryotes have been described in P. falciparum [38]–[40]. Recent studies addressed their role in var gene regulation and uncovered the first indications for a distinct histone code linked to variegated var gene expression [41]. Active var genes are associated with histone 3 acetylation (H3K9ac) and methylation (H3K4me2, H3K4me3) and silenced var genes are enriched in H3K9me3 [39],[42]. The P. falciparum genome also encodes a core set of histone-modifying enzymes including histone deacetylases and methyltransferases (HMTs) [43],[44]. Interestingly, different subsets of var and rif genes show increased transcription in mutant parasite lines lacking either one of the two PfSIR2 paralogs, demonstrating a direct role for these histone deacetylases in virulence gene silencing [35],[45]. Recently, H3K9me3 has been mapped to the full set of var genes and to members of other subtelomeric gene families on a genome-wide scale [37],[46]. The presence of H3K9me3 at silenced var loci provides important clues about the possible control strategy underlying epigenetic var regulation. This particular histone modification is a conserved hallmark of heterochromatic silencing and serves as a docking site for heterochromatin protein 1 (HP1) to nucleate the propagation of heterochromatin along the chromosome fiber [47],[48]. Indeed, PfHP1, the P. falciparum ortholog of HP1, binds to H3K9me3 and was shown to be involved in variegated expression of a particular var gene [49]. Although these findings justify the assumption that PfHP1 may be associated with H3K9me3-enriched loci, the genome-wide localization of PfHP1 is unknown and cannot be inferred directly from these data; HP1 proteins interact with a multitude of other proteins [50], proteins other than HP1 bind to H3K9me3 [51], H3K9me3 is not necessarily sufficient to recruit HP1 [52],[53], and HP1 can be recruited to heterochromatin in an H3K9me3-independent manner [54],[55].
In this study, we provide a comprehensive analysis of PfHP1 on the level of protein function and genome-wide distribution. We show that PfHP1 binds specifically to H3K9me3 but not to other repressive histone methylation marks. Using nuclear fractionation and detailed immuno-localization experiments, we demonstrate that PfHP1 constitutes a major heterochromatin component with confined localization to perinuclear foci. High-resolution analysis by genome-wide chromatin immunoprecipitation (ChIP-on-chip) revealed a striking PfHP1 occupancy pattern restricted to 425 genes, most of which are members of P. falciparum-specific exported virulence families. The majority of these genes were also enriched in H3K9me3 underscoring the biological significance of this interaction in virulence gene expression. In addition, we detected 38 PfHP1-bound genes not enriched in H3K9me3. Many of these genes code for invasion proteins or proteins specifically expressed in different life-cycle stages suggesting a previously unrecognized role for PfHP1 in invasion pathway switching and life cycle progression. Furthermore, we show that PfHP1 is not associated with centromeric regions implying important differences in centromere biology compared to other eukaryotes. Consistent with a role of PfHP1 in virulence gene silencing we find that over-expression of PfHP1 leads to downregulation of 78 genes, the majority of which are located in heterochromatic domains. In summary, our results attribute an important role to PfHP1 in parasite biology and suggest a unifying PfHP1-dependent mechanism by which P. falciparum regulates the variegated expression of proteins involved in virulence and phenotypic variation.
HP1 belongs to a family of conserved chromatin proteins found from fission yeast to humans and is characterised by an N-terminal chromodomain, which binds H3K9me3, and a C-terminal chromoshadow domain implicated in homo- and heterodimerisation [56]. The P. falciparum genome encodes three putative chromodomain proteins amongst which PFL1005c was recently identified as the P. falciparum ortholog of HP1 (PfHP1) [49]. A multiple sequence alignment based on structural information on HP1 revealed a remarkable similarity between the chromodomains in PfHP1 and those in HP1 from various eukaryotes, including conservation of residues critical for interaction with H3K9me3 [57],[58] (Figure S1). Similarly, most of the conserved residues important for the C-terminal chromoshadow domain and homo- and hetero-dimerisation are also present [57], [59]–[61]. To confirm that PfHP1 indeed displays these structural and functional HP1-like properties we expressed PfHP1 in E. coli. Using pull down experiments, we showed that PfHP1-HIS binds specifically to H3K9me3 but not to unmethylated H3 (Figure S2). Furthermore, we experimentally verified homo-dimerisation of PfHP1 by mixed incubation of PfHP1-HIS with nuclear extracts prepared from 3D7/HP1-Ty parasites expressing a 2×Ty-tagged version of PfHP1 (Figure S2). These findings provide an independent confirmation of the results obtained by Perez-Toledo and co-workers [49] identifying PFL1005c as the P. falciparum ortholog of heterochromatin protein.
To further test the specificity of the PfHP1-H3K9me3 interaction, we investigated binding of PfHP1 to a set of alternative histone modifications. Methylation of two other lysine residues in H3 and H4, H3K27me3 and H4K20me3, respectively, are commonly associated with transcriptional repression [62],[63]. Furthermore, phosphorylation of serine 10 adjacent to K9me3 (H3K9me3S10p) was found to counteract the repressive effect of H3K9me3 by preventing HP1 binding [64],[65]. Using a peptide competition assay, we show that PfHP1 interacts specifically with H3K9me3, whereas H3K9me3S10p, H3K9ac and H4K20 peptides were unable to interact with PfHP1 (Figure 1A). H3K27me3 interfered weakly with PfHP1-binding to H3K9me3 which was also observed for mouse HP1β [58]. The H3K9me3S10p and H3K27me3 marks have not been detected in P. falciparum to date [40] and therefore the biological significance of these findings remains to be determined. However, it is tempting to speculate that P. falciparum may use phosphorylation of H3S10 to reverse H3K9me3-mediated repression.
Consistent with its localization to heterochromatic regions, PfHP1 was insoluble after serial extraction of isolated parasite nuclei with low and high salt buffers (Figure 1B). After complete digestion of native chromatin with micrococcal nuclease (MNAse) or DNAseI a substantial fraction of PfHP1 remained associated with the high salt-insoluble pellet. Interestingly, the association of PfHP1 with the insoluble nuclear fraction was not sensitive to treatment with RNAse suggesting that in contrast to other eukaryotes, binding of PfHP1 to chromatin does not require RNA components [66]–[68]. These results demonstrate a tight association of PfHP1 with highly compact heterochromatic structures and/or the nuclear matrix.
Since var genes are dynamically associated with chromosome end clusters, we expected PfHP1 to be located in such defined perinuclear regions. To test this hypothesis we used three independent transgenic parasite lines. 3D7/HP1-GFP expresses a PfHP1-GFP fusion protein from its endogenous promoter, and 3D7/HP1-HA and 3D7/HP1-Ty express epitope-tagged versions of PfHP1 from episomal plasmids. Live cell imaging revealed that PfHP1-GFP located to the nucleus in a punctate perinuclear pattern reminiscent of chromosome end clusters (Figure 2A). 3D reconstruction verified the position of PfHP1 foci to the nuclear periphery (Video S1). Indirect immunofluorescence assays (IFA) confirmed these results and accentuated the localization of PfHP1 to discrete and well-defined regions at the nuclear periphery (Figures 2B and C). On average we observed 3.6 PfHP1-HA signals per trophozoite stage nucleus (3.64 (mean)±1.34 (s.d.); 302 nuclei counted). A detailed IFA experiment confirmed this restricted localization pattern in parasites carrying a single or multiple nuclei throughout intra-erythrocytic development (Figure S3). Next, we used immunoelectron microscopy to characterise PfHP1 localization at the ultrastructural level in 3D7/HP1-GFP parasites. PfHP1 was dominantly located at the nuclear periphery within and adjacent to previously described electron-dense regions reflecting perinuclear heterochromatin (Figure 2D) [34]. These high-resolution immunolocalizations corroborated and extended the light microscopy results, and together these data support the presence of PfHP1 at heterochromatic regions in the nuclear periphery. The PfHP1 localization pattern is consistent with multiple punctate foci of HP1 observed in S. pombe, plant and mammalian nuclei [69]–[71]. Finally, we asked if this discrete perinuclear localization pattern corresponds to chromosome end clusters by testing for a direct association of PfHP1 with subtelomeric DNA using combined IFA/FISH. In greater than 80% of cases PfHP1-GFP signals occurred directly adjacent to, or overlapped with, the subtelomeric repeat probe rep20 suggesting that PfHP1 is a major component of telomeric clusters (Figure 3). Surprisingly, we obtained a higher number of signals per nucleus for rep20 (5.02 (mean)±1.66 (s.d.)) compared to PfHP1 (3.3±1.21). In summary, our microscopy-based data identify PfHP1 as a major component of chromosome end clusters suggesting a preferential association of PfHP1 with P. falciparum subtelomeric regions.
In light of the above results and the recently published genome-wide H3K9me3 patterns [37],[46] it was tempting to speculate that PfHP1 plays a dominant role in subtelomeric virulence gene silencing. To test this hypothesis and to identify additional potential PfHP1 target loci we performed genome-wide chromatin immunoprecipitation (ChIP-on-chip) using a high-density whole genome tiling array (NimbleGen Systems Inc.) [46]. We observed a striking association of PfHP1 with subtelomeric regions on all parasite chromosomes (Figure 4A, Tables S1 and S2). These domains covered TARE repeat blocks and extended inwards including all subtelomeric var genes. Additional internal PfHP1-bound islands were identified on chromosomes 4, 6, 7, 8 and 12, which in most cases defined chromosome-central var gene clusters. Hence, PfHP1 binds to the full complement of subtelomeric and chromosome-internal var genes. PfHP1 occupancy was not restricted to var loci but instead covered extended regions including all members of other gene families shown to be expressed in a clonally variant manner, including rif, stevor and Pfmc-2tm [17]–[19] (Figure 4B). This peculiar association is even more striking considering that PfHP1 was hardly detected at loci falling outside these chromosomal regions. In fact, the immediate boundaries between PfHP1-occupied and PfHP1-free regions delineate the sites of species-specific indels or synteny breakpoints between P. falciparum and other Plasmodium species [3],[72]. 95% of all 425 PfHP1-bound genes code for P. falciparum-specific proteins most of which are involved in host-parasite interactions (Figures 4C, S4 and Table S2). Almost all of these genes are members of subtelomeric gene families coding for proteins exported to the erythrocyte (var, rif, stevor, pfmc-2tm, surfin, pfacs, fikk kinases) or predicted to be exported (phista, phistb, phistc, dnajI, dnajIII, a/b hydrolases, hyp1 to hyp17) (Figure 4C and Table 1).
In addition, some other lineage-specific genes were also occupied by PfHP1: invasion-related genes Pfrh1 [73] and Pfrh3 (pseudogene) [74], liver stage antigen 1 (lsa1) [75], the gametocyte-specific gene Pf11-1 [76], non-syntenic tRNA and rRNA loci, and a number of genes coding for hypothetical proteins. Surprisingly few PfHP1-bound loci code for proteins with orthologs in other Plasmodium species. These include members of the rhoph1/clag family (clag2, clag3.1, clag3.2) involved in erythrocyte invasion [77]; genes implicated in the development of sexual stages: ccp1 [78],[79], pfs230 [80],[81], putative dynein heavy chains [82]; crmp1 and crmp4 expressed in sporozoites [83]; and PFL1085w, coding for an ApiAP2 transcription factor [84].
Compared to Salcedo-Amaya and colleagues [46], who used the same NimbleGen array for genome-wide H3K9me3 mapping, we found a high level of local as well as genome wide correlation between the presence of this repressive mark and PfHP1 (R2 = 0.72) (Figure 4A and B). Out of 425 PfHP1-bound genes, we found only 44 genes that were not classified as H3K9me3-enriched, and one H3K9me3-associated gene was devoid of PfHP1-binding. In contrast, 155 PfHP1-associated genes were not classified as H3K9me3-enriched in the study by Lopez-Rubio et al. (68 of which were not represented on their array) [37], while all but six genes enriched in H3K9me3 were also occupied by PfHP1. By lowering the enrichment threshold for the latter study, PfHP1-occupancy correlated well with both genome-wide H3K9me3 localization datasets showing that 387 out of 425 PfHP1-bound loci were consistently enriched in H3K9me3 (Table S2). Most of these genes are members of gene families coding for proteins exported to the erythrocyte and implicated in parasite virulence (Figure 4C). Surprisingly, about half of the genes bound by PfHP1 but devoid of the H3K9me3 mark are single copy genes and members of small gene families coding for invasion proteins or proteins expressed in different life-cycle stages (Figure 4C and Table S2). This raises the interesting possibility that PfHP1 may be recruited to these loci in an H3K9me3-independent manner. Alternatively, this discrepancy may be related to overexpression of PfHP1, or to the use of different parasite lines and/or ChIP protocols.
To validate the ChIP-on-chip results we investigated the association of PfHP1 with individual loci by ChIP-qPCR. We targeted ten and twelve randomly selected loci, which showed either a negative or a positive association with PfHP1 in the ChIP-on-chip experiment, respectively (Figure 5 and Figure S5). These results confirmed the ChIP-on-chip findings in all instances, showing that PfHP1 is associated with subtelomeric and internal virulence genes but not with genes that showed no association with PfHP1 in the ChIP-on-chip experiment (Figure 5A). No chromatin fragments from transgenic parasite lines were recovered with rabbit IgG control antibodies or anti-HA/anti-GFP antibodies used on 3D7 wild-type parasites (data not shown), demonstrating the specificity of these results. Importantly, and consistent with a role of PfHP1 in stably inherited heterochromatic silencing, PfHP1-occupancy was present at the same loci in next generation ring stage parasites (Figure 5B). As expected, PfHP1-positive genes were also enriched in H3K9me3, which confirms the genome-wide colocalization and underscores the in vivo relevance of the PfHP1/H3K9me3 interaction in virulence gene silencing and the mutually exclusive presence of H3K9ac and H3K9me3 (Figure 5C). In contrast, genes not bound by PfHP1 were generally enriched in H3K9ac although this association did not necessarily correlate with active transcription of these loci. This is not surprising in light of recent findings demonstrating that H3K9ac occupancy did not differ markedly between the coding regions of active and inactive genes in P. falciparum [46].
In summary, our results demonstrate an extraordinarily confined localization of PfHP1 throughout the genome, and an extensive colocalization with the repressive histone mark, H3K9me3. This implies an important role for PfHP1 in epigenetic regulation of exported virulence factors and indicates that variegated expression and phenotypic variation may represent a general, rather than exceptional, feature of most P. falciparum-specific or expanded gene families.
In other eukaryotes like S. pombe and D. melanogaster, HP1 is an important factor in centromere function and a major constituent of pericentromeric heterochromatin [48]. We observed no evident presence of PfHP1 in these domains, albeit the average level of PfHP1 ChIP-on-chip signal over genes directly adjacent to centromeres was somewhat higher as compared to the rest of the genome (Figure S6 and Table S1). To test if PfHP1 was indeed enriched at centromeres we performed ChIP-qPCR experiments targeting the centromeres on eight chromosomes and six genes directly up- or downstream of centromeres which displayed low-level PfHP1-binding in ChIP-on-chip. We were unable to detect binding of PfHP1 to these regions as none of the loci tested showed any sign of PfHP1 enrichment (Figure S6). These findings are in line with the observed absence of H3K9me3 marks at centromeric regions and suggest that P. falciparum centromere biology and chromosome segregation are independent of PfHP1.
HP1 has been implicated in gene silencing [85],[86] and hence we were interested in testing this proposed function of PfHP1 in P. falciparum. Although recent whole transcriptome analyses strongly suggest that most subtelomeric gene families are expressed in a restricted manner [87],[88], a formal demonstration of a direct link between PfHP1 and gene expression is lacking. We therefore focused on a possible genome-wide correlation by transcriptional profiling using RNA isolated at four consecutive timepoints across the intra-erythrocytic developmental cycle (IDC) from two biological replicates of the PfHP1-overexpressing line (Table S3). PfHP1 target genes displayed significantly lower absolute expression levels as compared to all other genes at all IDC stages (p<0.001, Wilcoxon ranksum test) (Figure S7). It is noteworthy that many PfHP1-negative genes are also weakly or not expressed during the IDC, stressing the notion that PfHP1 is not a general marker for inactive genes and that other processes such as gene-specific regulation participate in developmental and cell-cycle-dependent transcriptional control.
We were also interested in testing the effect of perturbations in PfHP1 expression on global gene transcription. Several attempts to generate a PfHP1-null mutant failed suggesting an essential role for this protein in parasite biology. We therefore investigated if over-expression of PfHP1 had any effect on gene transcription by comparing mRNA levels of all genes in the transfected lines to those in a control line. We detected 78 genes that were consistently down-regulated in two biological replicates and none that were upregulated (Figure 6). Of these, 50 are members of PfHP1-demarcated gene families, 28 of which showed a greater than three-fold enrichment for PfHP1 in the ChIP-on-chip experiment (p-value 7.76E-36), including nine of the ten pfmc-2tm family members. Importantly, this analysis identified additional PfHP1 target genes that were either not detected, or classified as below three-fold enriched, in the ChIP-on-chip experiment and showed no sign for H3K9me3 enrichment. These include all members of the hyp5 family and additional members of the crmp, ccp, eba and dynein heavy chain families. At this stage, however, it remains unknown if the down-regulation of these genes is due to a direct or indirect effect of PfHP1 over-expression. Hence, increased levels of PfHP1 enhanced silencing of variegated genes and had only minor effects on global gene transcription. These results are consistent with a dosage-dependent effect of PfHP1 and suggest further that unwanted heterochromatin spreading is efficiently prevented by defined boundary structures. Furthermore, our approach demonstrates that over-expression studies by transcriptional profiling may be employed to investigate the function of regulatory proteins in P. falciparum gene expression.
In this study we present a comprehensive analysis of P. falciparum HP1 and describe the first genome-wide binding profile of a non-histone chromatin component in this important pathogen. Our findings reveal important insights into the regulatory strategy employed to control the variegated expression of a large class of highly specialized virulence genes. This knowledge will be instrumental for future investigations to understand parasite virulence and survival.
We have shown that PfHP1 binds specifically to H3K9me3 and forms stable homodimers in vitro, which are both conserved features of HP1 in other eukaryotes. A recent study used similar approaches to demonstrate these biochemical features for PfHP1 [49]. In vivo, PfHP1 associates extensively with subtelomeric repeats and genes encoding virulence factors in both subtelomeric and chromosome-internal loci. The conserved organisation of subtelomeric regions into blocks of distinct subtelomeric repeat units followed by multiple members of various gene families is a hallmark feature of P. falciparum chromosome ends. Unknown protein(s) mediate physical linking of chromosome ends in the formation of telomeric clusters [30]. The findings presented here identified PfHP1 as a major constituent of these chromosome end clusters. Our nuclear fractionation results suggest that PfHP1 is a candidate protein responsible, at least in part, for the physical clustering of chromosome ends through interactions between the chromoshadow domain and other structural components.
Whether chromosome-internal heterochromatic domains are an integral part of chromosome end clusters, or rather represent physically distinct entities at the nuclear periphery remains a matter of debate. Our ChIP-on-chip results demonstrated that PfHP1 associates with all subtelomeric regions. Hence, if chromosome-internal PfHP1-enriched regions form entities distinct from telomeric clusters one would expect a higher number of perinuclear PfHP1 domains compared to the four to seven chromosome end clusters usually detected by FISH [29],[30], which we never observed. Furthermore, in our IFA/FISH experiments the average number of rep20 signals was higher than that of PfHP1. At this stage, we don't know if this observation is due to the actual absence of PfHP1 from some chromosome end clusters or to differential sensitivities of IFA- and FISH-based target detection. However, both results clearly argue against a location of central heterochromatic domains separate from telomeric clusters. On the other hand, our IFA/FISH results may also be consistent with the idea that the few perinuclear PfHP1 foci not co-localising with rep20 reflect chromosome-internal heterochromatic regions that are physically distinct from chromosome end clusters, as has been suggested by others [37]. We believe that both opposing hypotheses are related to technical limitations inherent to FISH and IFA/FISH experiments resulting in a failure to detect the full complement of DNA and protein targets simultaneously. To know which scenario reflects the in vivo situation, more refined approaches such as locus tagging, confocal microscopy and/or 3C and 4C chromosome conformation capture techniques need to be applied.
We have shown that the majority of heterochromatic protein-coding genes are located subtelomerically directly adjacent and in smooth transition to the non-coding TARE region. A number of additional PfHP1 islands are also found at chromosome-internal clusters. In total, PfHP1 binds to 425 genes reflecting 7.5% of the parasite's coding genome. Notably, all heterochromatic coding domains are contained within sharply defined boundaries, which in most cases reflect non-syntenic regions. In other words, nearly all of PfHP1-bound genes code for proteins that do not have orthologs in other organisms and are thus specific to the P. falciparum lineage. This set of PfHP1-bound genes compares well with the genome-wide pool of H3K9me3-enriched loci described recently [37],[46]. This strong correlation is highly relevant for our understanding of the in vivo role of the PfHP1/H3K9me3 interaction and underscores its significance in P. falciparum virulence gene silencing. Some genes associated with PfHP1 were enriched in H3K9me3 in the Salcedo-Amaya study [46] but not in the Lopez-Rubio study [37]. These include members of the surfin, fikk kinase and gbph families as well as members of uncharacterised gene families such as hypxx and exported co-chaperones [3] (see also Table 1). We attribute this discrepancy to an improved performance of native versus formaldehyde-crosslinked H3K9me3 ChIP rather than to the actual absence of this histone mark at these loci.
The vast majority of PfHP1-demarcated gene families code for proteins that are exported into the host erythrocyte to participate in the processes of host cell remodeling, immune evasion and cytoadherence [3],[89],[90]. A hallmark in the epigenetic regulation of var gene transcription is the strict mutually exclusive expression of a single family member. The demonstrated association of PfHP1/H3K9me3 with var genes significantly advances our knowledge of the mechanisms underlying mutually exclusive var gene expression and may serve as a model system to understand the regulation and biological role of other virulence gene families. Clonal variation was also experimentally demonstrated for expression of a subset of PfHP1-bound gene families including rif, stevor, pfmc-2tm and surfin [17]–[19],[91]. Furthermore, several transcriptional profiling studies indicate restricted transcription of additional exported gene families that are also associated with PfHP1/H3K9me3. In view of the well-described role of HP1 in regional gene silencing, this remarkable association hints at an overall strategy to control phenotypic variation of a large pool of protein families that evolved to facilitate survival in a hostile environment. The expansion of lineage-specific exported protein families is much more pronounced in P. falciparum compared to other Plasmodia [3]. This observation is most likely related to the trafficking of PfEMP1 and other proteins to the erythrocyte surface and probably associated with the high virulence of P. falciparum. It is therefore tempting to speculate that the continuous expansion of P. falciparum exported gene families from single ancestral gene types ultimately required the parallel evolution of an epigenetic system to ensure phenotypic variation and avoid premature exhaustion of the antigenic repertoire. It is noteworthy that all but one (PFI1780w, phistc) of the genes coding for the core complement of 36 exported proteins shared between different Plasmodium species [3] are not associated with PfHP1. This is indicative for conserved essential functions of these ancestral proteins in the trafficking of exported proteins that evolved before lineage-specific expansion of virulence gene families.
The multi-step process of merozoite invasion into erythrocytes is characterised by the sequential action of apically located proteins encoded by gene families such as eba, Pfrh and rhopH/clag [92]. Variations in the expression of these genes are linked to alternative invasion pathways involving different ligand-receptor interactions. For instance, in isogenic 3D7 lines using either a sialic-acid dependent or independent invasion pathway, clag2, clag3.1 and clag3.2 show clonal variation and, for the latter two genes, are transcribed in a mutually exclusive manner [93]. Similarly, members of the Pfrh and eba families were shown to be differentially expressed in different parasite strains [94]–[97]. The association of PfHP1 with invasion gene families provides an explanation for these observations and important information about the epigenetic mechanisms responsible for invasion pathway switching. Our results further indicate that some genes expressed at different life cycle stages are controlled by PfHP1, including genes important during gametocyte maturation, sporozoite targeting to the salivary glands, or intra-hepatic development. However, these genes reflect only a small fraction compared to the full complement of developmentally regulated genes in P. falciparum indicating that life cycle stage conversion is mostly regulated by other mechanisms.
Telomeric and pericentromeric heterochromatin represents a conserved structural feature important in genome stability, proper segregation of chromosomes and prevention of telomere fusions in S. pombe and higher eukaryotes [98]. At both locations, heterochromatin is strictly dependent on repetitive DNA and physical interactions with siRNA and the RNA-induced transcriptional silencing (RITS) complex [99]–[101]. Our and other recent findings suggest that pericentromeric heterochromatin does not exist in P. falciparum. First, while PfHP1 is clearly enriched in subtelomeric repeats, we were unable to detect PfHP1 at centromeres or in pericentromeric regions by ChIP-qPCR, and H3K9me3 is also absent [37],[46]. Second, we provide evidence that the binding of PfHP1 to chromatin is not mediated by an RNA component. Third, components of the RNAi machinery and the RITS complex are absent in the P. falciparum proteome [102]. Therefore, although transcription of non-coding RNAs from P. falciparum centromeric regions has been reported [103], their role in formation of pericentromeric heterochromatin in the absence of PfHP1, H3K9me3 and a discernable RITS complex seems unlikely. Together, these observations indicate that PfHP1 plays no role in centromere biology and chromosome segregation and suggests a striking difference between P. falciparum and other eukaryotes including the human host.
We have shown that the presence of PfHP1 is directly linked to low expression of target genes. Furthermore, most of the genes down-regulated upon PfHP1 over-expression are members of variegated gene families, an effect that was also reported in D. melanogaster [104],[105]. These findings highlight the role of PfHP1 in variegated virulence gene expression. The sudden drop in PfHP1 occupancy at the boundaries of all heterochromatic regions is striking and supports the notion of functional genome partitioning in P. falciparum to secure expression of essential genes outside of these domains. The cis-acting sequences involved in the formation of these boundaries are unknown, but fine-mapping of the regions identified in this study will help to identify such elements. In S. pombe, tRNA loci are implicated in the boundaries of pericentromeric heterochromatin [106]. Interestingly, the three tRNA loci where ChIP-on-chip data were available were all enriched in PfHP1, two of which map exactly to the borders of heterochromatic domains on chromosome 7. Similarly, two tRNA loci on chromsomes 4 and 13 were shown to be enriched in H3K9me3 [37]. We were unable to analyse a PfHP1 loss-of-function phenotype due to the refractoriness of PFL1005c to gene disruption. It is noteworthy that compared to S. pombe, which encodes two HP1 variants and is viable after deletion of Swi6/HP1 [56], P. falciparum encodes a single HP1 protein only, suggesting that PfHP1 is essential for parasite survival. It is conceivable that one deleterious effect of PfHP1 removal would be caused by flooding of the parasite with large numbers of exported proteins. However, we also predict essential roles for PfHP1 in the aforementioned aspects of genome organisation and telomere integrity.
To the best of our knowledge, the remarkable distribution of PfHP1 to a large complement of functionally clustered genes has not been described in any other organism. In D. melanogaster, HP1 and the cognate histone methyltransferase SU(VAR)3-9 are associated with genes specifically expressed during embryogenesis or male development [107]. Likewise, genome-wide Swi6/HP1-association mapping in S. pombe identified only a small number of mostly meiotically expressed target genes [106]. In conclusion, the important role of PfHP1 in controlling parasite virulence uncovers a novel aspect of HP1 function. We hypothesise that other Plasmodium species use a similar strategy, and it will be interesting to see if other apicomplexan parasites and pathogenic fungi also employ HP1 for variegated expression of contingency gene families. Furthermore, our results set the stage for the identification of additional heterochromatin components and regulatory factors involved in epigenetic control of P. falciparum virulence gene expression. Detailed knowledge of these processes will be important for our understanding of this widely used survival strategy of pathogens and may uncover novel ways to interfere with pathogenesis and disease.
P. falciparum 3D7 parasites were cultured, synchronised and transfected as described [23]. Transfection constructs were generated according to standard procedures (Protocol S1). GFP-tagged endogenous PfHP1 was obtained by single crossover integration. Parasites were cloned by limiting dilution.
Sequences were amplified from 3D7 gDNA and cloned into pET24a(+) (Novagen). PfHP1-HIS was expressed in E. coli Tuner (DE3) (Novagen) at 37°C in TB containing 1% glucose and induced at OD = 1.0 for 4 hrs with 1 mM IPTG. Soluble extracts were prepared by freeze/thaw lysis.
E. coli lysate containing PfHP1-6×HIS was incubated with biotinylated H3 peptides (Upstate) immobilized on streptavidin agarose beads (Pierce) at room temperature for 1 hr in 250 µl BB (20 mM Tris-HCl (pH 8), 150 mM NaCl, 1 mM EDTA, 0.1% Triton X-100). After six wash steps in BB (250 mM NaCl) bound proteins were eluted in Laemmli buffer and analysed by SDS-PAGE. For peptide competition assays, PfHP1-HIS bound to biotinylated H3K9me3 peptide immobilized on streptavidin beads was eluted using a 10-fold excess of non-biotinylated histone peptides H3K9me3, H3K9me3S10p, H3K9ac, H3K27me3, H4K20me3 (Diagenode, sp-056-050, sp-128-050, sp-004-050, sp-069-050, sp-057-050) in buffer BB (250 mM NaCl). After peptide elution, beads were treated with 1 M NaCl to elute remaining PfHP1-HIS and all supernatants were analysed by SDS-PAGE and Western. Homo-dimerisation of PfHP1 was tested by co-incubation of a 1 M KCl nuclear extract prepared from 3D7/HP1-Ty with E. coli lysate containing PfHP1-6×HIS or non-transformed E. coli lysate in presence of 1% sarkosyl. Proteins were diluted six times with DB (20 mM Tris-HCl (pH 8), 250 mM NaCl, 20 mM imidazol, 0.5% Tween20), and combined with 10 µl Ni-agarose. Beads were washed four times in DB and bound proteins were eluted as above.
Nuclear fractionation involved sequential extraction of proteins from isolated parasite nuclei with low salt, digestion with either DNAseI, MNAse, or RNAseA, followed by extraction in 1 M KCl and finally 2%SDS (for details see Protocol S2). The nuclear fractions were analysed by Western blot.
Primary antibody dilutions were: anti-HA 3F10 (Roche Diagnostics) 1∶1,000; anti-Ty BB2 (kind gift of K. Gull) 1∶5,000; anti-6×HIS (R&D Systems) 1∶5,000; anti-H3 (Abcam, ab1791) 1∶40,000; anti-H4 (Abcam, ab10158) 1∶10,000.
Methanol-fixed cells were analysed using rat anti-HA 3F10 (1∶100) or mouse anti-Ty (1∶1,000). Alexa-Fluor® 568-conjugated anti-rat IgG (Molecular Probes) 1∶500; FITC-conjugated anti-mouse IgG (Kirkegaard Perry Laboratories) 1∶300. Images were taken on a Leica DM 5000B microscope with a Leica DFC 300 FX camera and acquired via the Leica IM 1000 software. Nuclei of unfixed 3D7/HP1-GFP cells were stained with DAPI before mounting onto a glass slide. Images were taken on a Carl Zeiss Axioskop microscope with a PCO SensiCam camera. 3D nuclear reconstruction was achieved by taking sequential z-stack series using a Carl Zeiss Axiovert 200 M microscope with an AxioCam MRm camera. Images were deconvolved and 3D reconstruction was performed using Axiovision v4.2 software. Images were processed using Adobe Photoshop CS2.
Parasites were fixed with 4% formaldehyde and 0.0075% glutaraldehyde and permeabilized with 0.1% Triton-X. Cells were incubated with anti-GFP antibody (kind gift of E. Handman) (1∶200) followed by anti-rabbit IgG-fluorescein (Invitrogen) (1∶200). FISH was carried out using a rep20 probe as previously described [30],[35].
3D7/HP1-GFP were fixed in 1% glutaraldehyde for 1 h at 4°C, dehydrated, and embedded in LR Gold resin (Electron Microscopy Sciences, Fort Washington, PA). Ultrathin sections were cut using a Leica Ultracut R microtome, labeled with polyclonal rabbit anti-GFP (Sigma) and 10 nm colloidal gold goat-anti-rabbit IgG (SPI). Sections were poststained with uranyl acetate and lead citrate and observed using a Philips CM120 BioTwin Transmission Electron Microscope.
ChIPs were carried out using formaldehyde crosslinked chromatin, except for anti-H3K9me3, which was analyzed using native MNase-digested chromatin [46]. Cross-linked chromatin was prepared by adding 1% formaldehyde to synchronized parasite cultures (5×10E8 schizonts or 2×10E9 ring stages) and incubated for 10 min at 37°C. Crosslinking was terminated by addition of 0.125 M glycine final concentration. After saponin lysis, nuclei were separated using a 0.25 M sucrose buffer cushion and sheared by sonication in a Bioruptor UCD-200 (Diagenode) for 15 min at 30 sec intervals (size range 100–500 bp). 400–500 ng DNA-containing chromatin was incubated with 1 µg antibody (anti-HA 3F10 (Roche); anti-GFP (AbCam, ab290); H3K9ac (Diagenode); IgG rabbit polyclonal (UpState 12–370) (used as a negative control)) in presence of 10 µl A/G sepharose beads (Santa Cruz Biotechnology) overnight at 4°C. After extensive washes immunoprecipitated chromatin was eluted with 1% SDS and 0.1 M NaHCO3 and de-crosslinked at 65°C for 4 hrs. DNA was purified using PCR purification columns (Qiagen).
Native chromatin was prepared from freshly isolated nuclei by MNase digestion and subsequent extraction with salt-free buffers (10 mM Tris pH 7.4, 1 mM EDTA; 1 mM Tris pH 7.4, 0.2 mM EDTA). Chromatin was diluted in 2×ChIP incubation buffer (100 mM NaCl, 20 mM Tris pH 7.4, 6 mM EDTA, 1% Triton X-100, 0.1% SDS). 400 ng DNA-containing chromatin was incubated with 1 µg antibody (anti-H3K9me3 rabbit polyclonal #4861 [108] overnight at 4°C followed by the addition of 10 µl A/G beads and further incubation for 2 h. After washing with buffers containing 100, 150 and 250 mM NaCl, immuno-precipitated DNA was eluted and purified as described above (without de-crosslinking).
The efficiency of ChIP at specific genomic locations was tested by quantitative PCR (qPCR) (MyIQ sequence detector, BioRad) using primer sets described in Table S4. “Negative” and “positive” genes were randomly selected from the group of genes that showed ChIP/input ratios lower or greater than 1.6 (log 2) in the ChIP-on-chip experiment. The amount of target DNA recovered after immuno-precipitation was directly compared to a ten-fold dilution series of input DNA, and defined as percentage of input for each locus.
For genome-wide analysis 24 individual anti-HA (AbCam, ab9110) ChIP reactions were combined using formaldehyde-crosslinked material. Immunoprecipitated DNA was amplified by a modified T7 linear amplification method [46]. Briefly, DNA fragments were dephosphorylated using calf intestinal alkaline phosphatase (NEB) and subsequently G-tailed with terminal transferase (NEB). T7 promoter was incorporated using Klenow polymerase (NEB) and T7C9B primer. RNA was synthesized by T7 polymerase (Ambion T7 megascript kit) and subsequently reverse transcribed using N6 primers. Amplified dsDNA was labeled with Cy3- (ChIP) or Cy5- (input) coupled random heptamers and hybridized to a tiling array (based on the May 2005 NCBI sequence of the P. falciparum genome; 385,000 probes with a median spacing of 48 bp, Roche NimbleGen) [46]. Log2 ratios were computed for each sample pair and after Tukey bi-weight normalization visualized by SignalMap software (Roche NimbleGen). Probes were mapped to the latest genome assembly and visualized in the context of PlasmoDB v5.5 genome annotation.
Growth of two independant transfectants over-expressing PfHP1 (PfHP1-A and -B) and the mock transfectant 3D7/camHG (K. Witmer et al., unpublished), was tightly synchronised in parallel three times by sorbitol treatment to achieve a 10 hr growth window. Total RNA was isolated at four timepoints across the IDC at early ring stages (4–14 hours post-invasion (hpi)), late ring stages (14–24 hpi), trophozoites (24–34 hpi) and schizonts (32–42 hpi) by lysis of pelleted RBCs in TriReagent (Sigma). RNA samples were analyzed using a P. falciparum microarray as previously described [109]. RNA from each time point and parasite line was labeled with Cy5 and hybridized against a RNA pool assembled from equal amounts of total RNA collected from the 3D7 strain at every 8 hrs. Absolute transcript abundance was determined as a mean of the sums of median Cy5 signal intensity on each microarray gene spot in all four time points. The relative abundance of individual transcripts was analyzed by Significance Analysis for Microarrays (SAM) as implemented by the MEV version 4.3 [110]. SAM (delta = 0.75) revealed 217 and 599 genes downregulated in PfHP1-A and PfHP1-B, respectively. Only two genes were found up-regulated in PfHP1-B.
The raw ChIP-on-chip data discussed in this publication have been deposited in NCBI's Gene Expression Omnibus [111] and are accessible through GEO Series accession number GSE17029 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE17029). Data remapped to the latest genome annotation has been submitted to PlasmoDB (www.plasmodb.org).
The PlasmoDB accession numbers for genes and proteins discussed in this publication are: PfHP1 (PFL1005c); rh1 (PFD0110w); rh3 (PFL2520w); eba-165 (PFD1155w); lsa1 (PF10_0356); pf11-1 (PF10_0374); clag2 (PFB0935w); clag3.1 (PFC0120w); clag3.2 (PFC0110w); ccp1 (PF14_0723); pfs230 (PFB0405w); dynein heavy chains (PFI0260c, MAL7P1.162); crmp1 (PFI0550w); crmp4 (PF14_0722); ApiAP2 protein (PFL1085w)
|
10.1371/journal.pgen.1004064 | Natural Selection Reduced Diversity on Human Y Chromosomes | The human Y chromosome exhibits surprisingly low levels of genetic diversity. This could result from neutral processes if the effective population size of males is reduced relative to females due to a higher variance in the number of offspring from males than from females. Alternatively, selection acting on new mutations, and affecting linked neutral sites, could reduce variability on the Y chromosome. Here, using genome-wide analyses of X, Y, autosomal and mitochondrial DNA, in combination with extensive population genetic simulations, we show that low observed Y chromosome variability is not consistent with a purely neutral model. Instead, we show that models of purifying selection are consistent with observed Y diversity. Further, the number of sites estimated to be under purifying selection greatly exceeds the number of Y-linked coding sites, suggesting the importance of the highly repetitive ampliconic regions. While we show that purifying selection removing deleterious mutations can explain the low diversity on the Y chromosome, we cannot exclude the possibility that positive selection acting on beneficial mutations could have also reduced diversity in linked neutral regions, and may have contributed to lowering human Y chromosome diversity. Because the functional significance of the ampliconic regions is poorly understood, our findings should motivate future research in this area.
| The human Y chromosome is found only in males, and exhibits surprisingly low levels of genetic diversity. This low diversity could result from neutral processes, for example, if there are fewer males successfully mating (and thus fewer Y chromosomes being inherited) relative to the number of females who successfully mate. Alternatively, natural selection may act on mutations on the Y chromosome to reduce genetic diversity. Because there is no recombination across most of the Y chromosome all sites on the Y are effectively linked together. Thus, selection acting on any one site will affect all sites on the Y indirectly. Here, studying the X, Y, autosomal and mitochondrial DNA, in combination with population genetic simulations, we show that low observed Y chromosome variability is consistent with models of purifying selection removing deleterious mutations and linked variation, although positive selection may also be acting. We further infer that the number of sites affected by selection likely includes some proportion of the highly repetitive ampliconic regions on the Y. Because the functional significance of the ampliconic regions is poorly understood, our findings should motivate future research in this area.
| The Y chromosome has often been used as a marker for studying human demographic history [1], but one implicit assumption in these analyses is that the Y chromosome is not affected by natural selection at linked sites [2]. However, formal tests of models of selection have been lacking. In part, this has been due to a paucity of resequencing data for many male human genomes, where autosomal, X, Y and mtDNA for the same individuals could be compared. Such data eliminate one source of sampling variance that could influence comparisons between genomic regions, and also allow for chromosome-wide estimates of genetic diversity on the Y, which is often ignored in whole-genome analyses [3]–[5]. Under simple neutral models with constant and equal male and female population sizes, diversity is expected to be proportional to the relative number of each chromosome in the population: X diversity is expected to be three-quarters autosomal diversity (because there are three X chromosomes for every four autosomes) and both the Y and mtDNA diversity are expected to be one-quarter autosomal diversity [6].
The Y chromosome does not undergo homologous recombination, except in the small pseudoautosomal regions [7]. In general, diversity is reduced in genomic regions or genomes with little or no recombination [8]–[11]. Similarly, previous studies of small segments of the human Y chromosome have found low levels of genetic diversity, but multiple theories exist to explain this reduction [12]–[16].
Because the Y chromosome is found only in males, low diversity on the Y could result from neutral processes if, for example, the effective population size of males is reduced relative to that of females. One factor that can reduce the male population size is high variance in the number of offspring. Differences in the variance in reproductive success between the sexes, will cause differences in effective population sizes, even when the actual number of males and females is approximately the same [4], [13]. Based on comparing patterns of genetic variation on the X chromosome and the autosomes, several recent studies have found evidence of sex-biased demographic processes during human history [3]–[5], [17]–[20], often suggesting that the effective population size of females was higher than that of males throughout recent human history (Nf>Nm, if Nf represents the effective number of breeding females and Nm represents the effective number of breeding males).
Alternatively, purifying selection acting to remove new deleterious mutations on the Y chromosome, will affect diversity at linked neutral sites through a process called background selection. Background selection refers to the reduction in genetic diversity at sites that are themselves neutrally evolving, but are linked to other sites where deleterious mutations occur [21]–[24]. Background selection may be particularly potent on the Y chromosome, because there is no recombination on the Y chromosome. As such, deleterious mutations in one area of the chromosome could reduce levels of genetic diversity across the entire chromosome [12], [14]–[16]. However, the strength of selection is also important. Several weakly deleterious mutations may interact resulting in a Hill-Robertson interference [25], whereby interference among linked sites weakens their effects on linked neutral sites [26]. Similarly, positive selection, acting on beneficial mutations is expected to decrease diversity at linked neutral sites. Given the unique gene content and lack of recombination on the Y chromosome, it is likely to have experienced a complex evolutionary history.
Here, using genome-wide analyses of X, Y, autosomal and mitochondrial DNA, in combination with extensive population genetic simulations, we show that low observed Y chromosome variability is not consistent with a purely neutral model. Instead, we show that models of purifying selection and background selection affecting linked neutral sites are consistent with observed Y diversity. Further, the number of sites estimated to be directly under purifying selection greatly exceeds the number of Y-linked coding sites, suggesting the importance of the highly repetitive ampliconic regions [27]–[29]. Because the functional significance of the ampliconic regions is poorly understood, our findings should motivate future research in this area.
Analyzing complete genomic sequence data from 16 unrelated males (Table S1), we observe that normalized diversity on the human Y is extremely low compared to expectations from other genomic regions (Figure 1; Table 1). By analyzing resequencing data for the autosomes, X chromosome, Y chromosome, and mitochondria from the same individuals, we reduce sampling variance that might otherwise confound comparisons between regions of the genome. Here diversity is measured as the average pairwise differences per site, π, in the sample, and is normalized using divergence between humans and outgroup species (see Materials and Methods). The purpose of this normalization is to account for the possibility that different parts of the genome may have different mutation rates. The mutation rates could systematically differ across chromosomal types because the different chromosomes spend different amounts of time in the male and female germlines and the male germline has a higher mutation rate than the female germline [30]. Because the low diversity on the Y chromosome persists after this normalization, it cannot be explained by a correspondingly low mutation rate on the Y chromosome (Table S2; Figure S1). Further, the highly repetitive ampliconic regions of the Y were not assembled by Complete Genomics, and so are not analyzed here (Materials and Methods). Diversity on the Y chromosome is likely not being under-estimated due to the inability to call variants in haploid regions of the genome because diversity on the X measured in females, where the X is diploid, is nearly identical to diversity on the X measured in males, where the X is haploid (Figure S2). The pattern of reduced diversity on the Y chromosome is observed in both Africans and Europeans, suggesting that the effect is not population-specific, and holds regardless of whether the neutral sequence analyzed is near or far from genes (Table 1). Previous analyses of portions of the Y reported low Y diversity [12]–[16], but measuring divergence-normalized π per site at 0.0018 for Africans and 0.0024 for Europeans, we observe that chromosome-wide Y diversity is an order of magnitude lower than the equilibrium neutral expectation of one-quarter the autosomal level of diversity (Figure 1). Conversely, mitochondrial diversity is not reduced compared to expectations under neutrality (Figure 1). Additionally, our estimates of diversity on the X chromosome are consistent with previous estimates from Africans [5], [17] and Europeans [3], [5]. These trends held for all populations sampled in the public Complete Genomics data (Figure S3).
In contrast to diversity in other genomic regions, we observe that diversity is lower on the Y chromosome for the African populations in our sample than for the European populations in our sample (Table 1). Previous studies of Y chromosome diversity have also suggested that the difference in diversity on the Y is small between Africans and Europeans [31], [32], or that it may, as we observe, be higher in Europeans than some African populations [15], [33]. For example, haplotype diversity was found to be higher across Europeans than Africans (0.852 versus 0.841) [33]. Similarly, when the African populations are broken down into Sub-Saharan Africans versus North Africans (the Complete Genomics samples are Western/Northern Africans), European diversity falls in between these two, with European diversity on the Y chromosome actually higher than diversity in North Africans [33]. Other studies have observed slightly higher diversity in Africans than Europeans, but include a much more diverse group of Africans. For example, variation on the Y chromosome has been reported previously to be only slightly higher on the Y for African versus Non-African populations, even though the population of Africans is much more diverse (including Bakola from Cameroon, Dogon from Mali, Bantu from South Africa and Khoisan from Namibia and South Africa) [32] than the population we analyze. The uncorrected levels of diversity reported here for the Y chromosome (Table S2), differ from some previous studies [15], [31], [34], but are not directly comparable to these studies because: 1) they were based on genetic markers that were chosen specifically because they have high mutation rates [15], [31], [34]; and, 2) the populations are different than the ones available for this study [34]. The absolute number of SNPs identified here is not reduced relative to other sequencing platforms [35]. In fact, overall diversity is similarly observed to be low on the Y using this other technology, but a larger TMRCA is estimated [35], perhaps because the Y seems to harbor pockets of hidden diversity [36].
We next consider several possible models that could explain this unexpectedly low amount of diversity found on the Y chromosome relative to other genomic regions. Such models include differences in the variance in reproductive success between males and females, purifying selection on the Y chromosome, and positive selection on the Y chromosome.
In principle, a greater variance in male reproductive success than female reproductive success (Nf>Nm) could result in a lower than expected effective population size of the Y chromosome. In fact, previous studies have suggested that increased variance in offspring number has reduced the effective population size in human males versus females and might explain the reduced variability on the paternally inherited Y chromosome [4], [13]. To test the hypothesis that sex-biased demography explains the decreased Y chromosome diversity, we modeled increasingly skewed sex ratios using coalescent simulations, taking into account the complex demography of the populations analyzed here (Figure 1; Table S3; Methods). We use the case where Nm = Nf as the null model. As expected, decreases in the male effective population sizes (Nm/Nf<1) decrease expected Y diversity. However, we find that the reduction in the male effective population size required to explain the observed Y chromosome data, predicts levels of normalized autosomal, X and mtDNA diversity that are not consistent with the data in these markers (Figure 1; Table S3). This effect can also be illustrated by considering ratios of normalized diversity in each type of marker relative to autosomes. A skew in the sex ratio large enough to explain the observed reduction in Y/autosome diversity would also cause increases in X/autosome and mtDNA/autosome diversity that are incompatible with observations (Figure 1; Table S4). Thus, by analyzing all classes of genomic sequences, we are able to reject extreme sex-biased processes as the sole explanation for patterns of low observed Y variability.
Natural selection has also been suggested to play a large role in reducing diversity on the Y chromosome [12], [14]–[16], and works within the context of the demographic history of the populations. Purifying selection can reduce genetic variation at linked neutral sites via a process called background selection, which has received extensive theoretical treatment in the literature [21], [22], [26], [37]–[41]. Purifying selection has already been documented for the mtDNA [42]. Due to the lack of homologous recombination throughout most of the Y chromosome, background selection is expected to have a particularly strong effect, severely reducing diversity on the Y chromosome. Two factors determine the overall effect of background selection on reducing neutral diversity in non-recombining regions: 1) The strength of selection, and 2) the number of sites subject to selection. At approximately 60 million base pairs, there are orders of magnitude more sites that may be subject to selection on the human Y chromosome than on the mtDNA. Selection may actually be quite weak on individual mutations that occur on the Y chromosome, but in the absence of recombination, if many sites are possible targets of this weak selection, this can lead to a strong reduction in diversity among Y chromosomes.
Here, we performed forward simulations with purifying selection to assess whether background selection could reduce diversity at neutral sites on the Y chromosome to the levels observed in our data. We study purifying selection under different assumptions of the variance in male reproductive success. We chose to use forward simulations, rather than using standard analytical background selection models, which assume the effect of background selection is a simple reduction in effective population size, for several reasons. First, the standard formulas were derived for equilibrium demographic models, but human populations have a more complex demographic history with unknown effects on the process of background selection. Second, many mutations have been shown to be weakly deleterious and may persist in the population due to genetic drift [37], [38]. The standard theory does not allow for this. Finally, simulations studies suggest that the standard theory can over-predict the reduction in genetic diversity due to background selection if there are many weakly selected linked mutations [26]. The forward simulations that we conducted address all of these concerns.
We first evaluated whether purifying selection acting only on new nonsynonymous mutations in the coding regions of the Y chromosome could reduce levels of genetic diversity at linked neutral sites to the levels detected in our observed Y chromosome data. To do this, we performed forward simulations using realistic demographic models for the populations where only new nonsynonymous mutations were subjected to purifying selection (see Methods). We find that models of selection acting only on coding sites cannot sufficiently reduce expected diversity at linked neutral sites through background selection on the Y chromosome. Under the assumption of equal sex ratios, regardless of the mean selection coefficient used, all models result in levels of diversity at linked neutral sites that are significantly higher than the observed values for both Africans (P<0.001) and Europeans (P<0.025, Figure S4).
In principle, models with a larger female effective population size could explain the low diversity observed on the Y chromosome. However, we have demonstrated that such models cannot match the levels of genetic diversity observed on the X chromosome, mtDNA, and Y chromosome together. However, sex-biased demography along with purifying selection acting on new nonsynonymous mutations in the coding regions of the Y chromosome could reduce levels of diversity at linked neutral sites. To evaluate the joint effects of sex-biased demography and purifying selection, we used levels of putatively neutral diversity (i.e., diversity far from genes) on the X chromosome and the autosomes to estimate the degree of sex-biased demography for the populations in our study (Table 2). We find that Nm/Nf = 0.335 in the African population which is concordant with estimates from previous studies [4], [20], [35]. Under an assumption of an extremely reduced male effective population size, relative to females (Nm/Nf = 0.335) which matches patterns of diversity on the X chromosome, predicted diversity at linked neutral sites, from models including purifying selection only on nonsynonymous mutations, is still significantly higher than the observations in Africans (P<0.001, Figure S4). In Europeans, we estimate that that Nm/Nf = 1 (Table 2). These results hold for a wide range of the mean strength of selection (Methods; Figure S5).
Given its unique structure, it is possible that purifying selection acts on more than just the nonsynonymous sites on the Y chromosome. Specifically, in addition to the approximately 100,000 single copy coding sites (predicted from annotated coding genes [43]; Methods), the Y also contains 5.7 Mb of highly repetitive ampliconic regions, composed of long palindrome “arms”, each with nearly-identical sequences [27], [28]. Genes in these ampliconic regions are expressed exclusively in the testis [27], [28], and so may be under selection related to male fertility. Further, it has been hypothesized that, in the absence of homologous recombination with the X, intra-chromosome pairing and the resulting gene conversion between palindrome arms may reduce the mutational load on the Y, and so these palindromes themselves, as a means of allowing intra-chromosome recombination, may be subjects of selection [27]–[29].
Thus, we developed a novel approximate likelihood approach to estimate the number of sites affected by purifying selection (L) required to reduce diversity at linked neutral sites to the low values observed on the Y (Methods). Simulations show that our method can accurately estimate L (Methods; Table S5). Assuming an equal sex ratio, the maximum likelihood estimate of the number of sites subjected to purifying selection on the Y is as much as 30 fold higher than the number of coding sites, for both Africans and Europeans (Figure 2). Relaxing the assumption of an equal sex ratio to allow many fewer males relative to females (to the ratio of the number of males to the number of females that fit neutral diversity on the X and autosomes, Nm/Nf = 0.335 [4], [20]), and to an extreme bias in male reproductive success of Nm/Nf = 0.1, slightly decreases the estimates of the number of sites directly affected by purifying selection. However, the estimate from the African sample is still significantly greater than the number of coding sites. Our results strongly support the hypothesis that at least some of the ampliconic regions evolve under the direct effects of purifying selection, where new mutations in these regions are deleterious.
The above estimates assume that the selection coefficients of the deleterious mutations on the Y chromosome are the same as those estimated from nonsynonymous mutations on the autosomes, with appropriate re-scaling to account for the differences in Ne and ploidy on the autosomes and the Y chromosome (see Methods). However, it is possible that the strength of selection acting on noncoding mutations on the Y chromosome could be different than that acting on nonsynonymous mutations on the autosomes. It is unclear whether this difference in the strength of selection could bias our estimates of the number of sites directly under selection. To address this concern, we extended our approach to jointly estimate the number of sites directly affected by purifying selection (L) as well as the mean strength of selection (see Methods). Even when considering a range of different strengths of selection, we find that the estimates of the number of sites to be directly under the effect of purifying selection are largely insensitive to the mean strength of selection, and are still more than the number of X-degenerate coding sites (Figures S5 and S6). This suggests that content recruited to the Y chromosome after X–Y recombination was suppressed, including the high-copy-number ampliconic regions, as well as any transcription factor binding sites, may be subject to purifying selection that, due to the lack of homologous recombination, acts to reduce diversity on the human Y chromosome.
We found that a population expansion model matched the average observed levels of autosomal, X and mtDNA polymorphism in the African populations, and a bottleneck model matched the observed levels of polymorphism in the European population (Figure 1, Tables S4, S5 and S7). Several publications have documented various signatures of background selection throughout the genome [17], [44]–[47]. If background selection had reduced average levels of diversity across the genome (previous work suggests around a 6% reduction in diversity [24]), this would mean that the demographic parameters that fit the data were not truly reflective of population history, but instead reflected both population history and background selection. Thus, even if background selection is operating on the putatively neutral genomic regions we analyze here, the reduction in diversity on the Y chromosome is still too extreme to be consistent with that level of background selection. Rather, additional background selection, as we have modeled here, would be required.
Although models of purifying selection are consistent with the low observed diversity, it is also possible that positive natural selection may also be driving low diversity on the human Y via selective sweeps [48], [49], when neutral variation is removed due to the fixation of an advantageous mutation. Although it can be difficult to distinguish between genetic signals of background selection versus positive selection with few nucleotide polymorphisms, as is the case with the Y chromosome, we analyzed the data using two additional measures. First, we computed the folded site frequency spectrum for Y chromosome SNPs across all unrelated Y chromosomes in the Complete Genomics dataset (Figure S8). The abundance of low frequency SNPs is consistent with both positive selection and purifying selection (Figure S8), and the low overall number of SNPs makes further distinctions between the two models difficult. Second, we built a neighbor-joining tree for all unrelated Y haplotypes in the Complete Genomics dataset using phylip [50], then branch lengths were computed using a molecular clock in paml [51]. There is not an overarching star phylogeny, which would be indicative of a single selective sweep (Figure S9). While we cannot rule out such a scenario directly, we note that previous studies also found little or no evidence of selective sweeps [52] or gene-specific positive selection [53], [54] on the Y chromosome. However, one might conceive of a complex evolutionary history involving several instances of positive selection along different Y lineages that could result in the observed haplotype topology. Given recent findings of pockets of Y haplotype diversity, it is possible that recurrent positive selection may contribute to reduced Y diversity [36].
We observe that diversity across the entire human Y chromosome is extremely low. We find that neutral models with sex-biased demography may contribute to low Y diversity. However, models of extreme differences in reproductive success between males and females are insufficient as the sole explanation for patterns of genome-wide diversity. Alternatively, then, natural selection appears to be acting to reduce diversity on the Y. We show that models of purifying selection affecting Y chromosome diversity are consistent with low observed diversity, if purifying selection acts on more than the few coding regions left on the Y chromosome. Thus, our results suggest that selection may also act on the highly repetitive ampliconic regions, and support arguments for the functional importance of these regions [29]. Further strong purifying selection acting on the human Y is consistent with reports of the conservation of both the number and the type of functional coding genes on the Y chromosome in humans [12] and across primates [55], [56]. It is also possible that positive selection has been acting to reduce diversity on the Y chromosome, but this explanation would require multiple independent selective sweeps across populations.
Although positive selection is expected to confound evolutionary relationships, if purifying selection is the dominant force on the Y chromosome, the topology of the tree should remain intact, but the coalescent times are expected to be reduced. This means that the Y chromosome, keeping in mind that it is a single marker without recombination, may actually provide a more useful marker for inferring phylogeographic patterns than other markers. Indeed, recent resequencing efforts of the Y chromosome identified a single mutation that resolves a previously unresolved trifurcation of lineages, and reports monophyletic groupings of Y chromosomes from distinct populations [35]. While it a combination of factors influence genome-wide estimates of diversity, and variance in male reproductive success still affects patterns of autosomal, X, Y and mtDNA diversity, selection clearly affects levels of diversity on the Y, and so should be considered when drawing conclusions regarding demography and population history based on patterns of Y-linked markers.
We analyzed unrelated, high quality, publicly available whole genomes generated by Complete Genomics assembly software version 2.0.0 [57] (Table S1). Next generation sequence data often suffer from sequence errors, assembly errors and missing information, and non-reference alleles will be less likely to be mapped [58]. However, the Complete Genomics dataset overcomes many of these errors by using very high coverage (>30X [57]). Additionally, to be conservative, we only consider sites with data called in all individuals in each population. We removed putatively functional and difficult to assemble regions including: RefSeq known genes, CpG islands, simple repeats, repetitive elements (RepeatMasker), centromeres, and telomeres, downloaded from the UCSC Genome browser [43], and filtered using Galaxy [59]. We also excluded the hypervariable regions on the mtDNA [60], which might inflate estimates of mitochondrial diversity, and analyzed only the X-degenerate regions of the human Y [27], because diversity might be reduced in the pseudoautosomal or ampliconic regions. Divergence was computed from number of nucleotide differences per site between pairwise human and chimpanzee reference sequence alignments for autosomes, X, and mtDNA downloaded from the UCSC genome browser [43], and for the Y from ref [28]. The total number of SNPs called on the Y chromosome in the Complete Genomics dataset does not appear to be lower than other chromosome-wide assessments of Y variation. Of the SNPs across 16 individuals that overlap between the 1000 genomes (252 SNPs) and Complete genomics dataset (6236), there are only 12 sites called in the 1000 genomes dataset that are not called in the Complete Genomics dataset; all of these are singletons, and many have missing data across several individuals (Table S7). Further, the geographic distribution of Y chromosome sampled for the Complete Genomics dataset does not appear to be wider for the European versus the African populations [61]. The per generation per site mutation rates estimated from human-chimpanzee alignments, assuming a divergence time of 6 million years and 20 years per generation, are 2.11×10−08 for the autosomes, 1.65×10−08 for chromosome X, and 3.42×10−08 for chromosome Y. For mtDNA we use the mutation rate reported of 1.7×10−08 for the mtDNA [62]. The recombination rates used were 1 cM/Mb and (2/3)*(1 cM/Mb), for the autosomes and X, respectively. Diversity is measured using, π, the average number of nucleotide differences per site between all pair of sequences. For the inference of the number of sites under selection, we summarize the genetic variation data by S, the number of segregating sites, because the distribution of S, conditional on the underlying genealogy, is known (Poisson, see below). We do not directly analyze the ampliconic regions, as they were not assembled in the Complete Genomics data.
All estimates of diversity, and human-chimpanzee divergence used for normalization are reported in Table S2. Human-orangutan estimates of divergence could not be used because no whole Y chromosome sequence currently exists for orangutan. Although the Y chromosome sequence was recently published for the rhesus macaque, the sequence has diverged and degraded so much between human and macaque that very little of the noncoding regions are alignable [55], preventing us from reliably correcting for divergence across all chromosome types using human-macaque divergence.
Population genetics parameters used in coalescent [63] and forward simulations [64] for Europeans and Africans are similar to previously published estimates [65], [66]. We use a simple model of drift, which assumes purely random (Poisson) variation in offspring numbers for both males and females, and non-overlapping generations. For Africans, the neutral model is of an expansion from 10,000 to 20,000 individuals 4,000 generations ago. For Europeans the neutral model is of a bottleneck from 10,000 to 1,000 individuals 1,500 generations ago, followed by an expansion to 10,000 individuals 1,100 generations ago (Table S6).
Neutral expectations under equal and skewed sex ratios were modeled using coalescent simulations implemented in ms [63], assuming the population-specific demographic models described above, and allowing for recombination on the autosomes and X chromosome, but not the Y or mtDNA.
The effective population sizes for each chromosome type (Nauto, NchrX, NchrY, and NmtDNA), for given male and female effective population sizes (Nm and Nf) are (see e.g., ref [67]):For a fixed ratio and males to females (R = Nm/Nf), and fixed total effective population size (Nauto), we then calculate the male and female effective population sizes as:Using these equations we can use standard neutral coalescent simulations implemented in ms to simulate data for the four chromosome types, while varying R, but keeping Nauto constant. We keep Nauto constant to mimic the real data, as the demographic parameters were originally estimated from autosomal markers. Further details about the values used for simulations can be found in Table S8. Complete commands for ms simulations are given in Note S1.
We modeled purifying selection using forward simulations implemented in SFS_CODE [64]. The exact commands used in the SFS_CODE simulations are given in Note S1. Similar to the coalescent simulations, we modeled the African and European populations separately, using the population-specific demographic models described above, the Y chromosome per generation per base pair mutation rate, and sampling 8 chromosomes per simulation to match the sample size of our observed data. However, unlike ms, which scales parameters by the current population size and moves backward in time, SFS_CODE starts with the ancestral number of chromosomes and simulates a haploid population forward in time. Thus, when rescaling the effective population size from the autosomal estimates, for SFS_CODE we used the same diploid autosomal ancestral effective population size for both populations (N = 10,000). The Y chromosome effective size was then found using the same process described above for the neutral coalescent simulations.
To investigate purifying selection acting only on new nonsynonymous mutations, we simulated 60,041 nonsynonymous sites (90,062 coding sites are estimated from the union of all exons from X-degenerate, non-pseudoautosomal genes on the Y chromosome [43]) at which new mutations are expected to be subject to purifying selection. To assess the effect of background selection, each simulation also contained 500 kb of linked neutral sequence from which we calculated diversity.
The effect of background selection is a function of the distribution of selection coefficients for new, deleterious mutations, and can be modeled by varying the mutation rate, the number of sites affected by selection (L), and the selection coefficient acting on new mutations (s) [21]. When evaluating models with different strengths of purifying selection, we assumed that selection coefficients for the nonsynonymous sites were drawn from a gamma distribution. Previous studies found this distribution to fit the observed autosomal frequency spectrum well [37], [38], [68], [69], and there is little reason to believe that the shape of the gamma distribution varies across chromosomes. However, although the X- and Y-linked genes are often highly diverged in sequence and function, the remaining X-degenerate Y-linked genes are likely highly constrained in order to have survived on the Y [70]. Thus, it may not be precise to assume X-degenerate Y-linked genes evolve under similar selective constraints as autosomal genes. To address this, we investigate a wide range of scale parameters of the gamma distribution. For a fixed value of the shape parameter of the gamma distribution, the mean strength of selection can be changed by modifying the scale parameter of the gamma distribution. Thus, we fixed the shape parameter to 0.184 (as estimated by refs [37], [69]) and performed simulations using mean selection coefficients ranging from 0.0001 to 0.09 (Figure S4).
We ran 1,000 replicates for each set of selection parameters in each population. For each replicate we calculated π*, the simulated per site nucleotide diversity (average number of pairwise differences) normalized by the per site human-chimp divergence (0.02051; Table S1). The similarly calculated observed Y diversity is denoted πobs. For each set of parameter values we then calculated P1, the proportion of simulation replicates with π*>πobs was used to calculate a 2-sided P-value by P2 = 1−2×|P2−0.5|. Models with could not be rejected and were considered to fit the observed data.
To estimate the number of sites directly affected by purifying selection on the Y chromosome (defined as L) from looking at the levels of diversity at linked neutral sites, we developed a novel approximate likelihood approach [65], [71], [72] using the observed number of segregating sites, Sobs, in neutral regions, as a summary statistic. We then define the likelihood function for L in a neutral region as:where is the number of segregating sites in neutral regions of the observed data, is the sum of all the branch lengths of the genealogy in units of generations, and refers to all of the other fixed parameters in the model (e.g., the demographic history and distribution of selection coefficients). Under the infinite sites model, the conditional distribution of Sobs given T is Poisson (see e.g., [73]):where μ is the neutral mutation rate per generation over the entire region. This relationship holds even if the underlying genealogy has been affected by natural selection or other non-stationary demographic processes, as long as the individual mutations being analyzed are neutral. Then, the number of sites affected by purifying selection, L, enters the likelihood function by the effect that selection has on the genealogy. is the distribution (density) of the sum of the branch lengths over the entire genealogy under the particular model of demography and selection, with L sites directly affected by purifying selection. This distribution is difficult to calculate directly, and in general, the integral given above cannot be solved analytically. However, it could be approximated using simulation approaches that keep track of the genealogy as part of a forward simulation method [74]. If we could simulate from, then the distribution of could be approximated as the sum:However, even such an approach is cumbersome and slow because of the overhead involved in keeping track of a genealogy in simulations with multiple loci under selection. We instead employ an approximate approach using forward-simulations implemented in SFS_CODE [64]. For a simulation replicate producing variable sites, and with a simulated value of equal to T*,Therefore, a simulation consistent estimator of can be obtained from the number of segregating sites in a simulated sample. In other words, if we simulate enough sites in each replicate, the total tree length can be approximated using the number of segregating sites (Table S5; Figure S5). The aforementioned integral in the likelihood function can therefore be approximated stochastically by simulating data sets using SFS_CODE, with Si*, = 1, 2,…k, segregating sites, and each with a neutral mutation rate of μsim, and then evaluating,as an estimator of the likelihood function for L based on Sobs. The number of neutral base pairs on the Y chromosome with sufficient sequencing data was 7,758,906 and 7,974,045 bp for the African and European populations respectively. Assuming a neutral mutation rate of 3.42×10−08 per base pair per generation, μ = 0.265 for the African population and μ = 0.273 for the European population. However, forward simulations of >7 Mb of sequence are extremely time consuming. Thus, for computational efficiency, we simulated 500 kb of neutral sequence, giving μsim = 0.0171. We accounted for the fact that we simulated fewer neutral sites than in the actual data by including the ratio of the two per region mutation rates (μ/μsim), in our likelihood function represented above. We chose to simulate 500 kb of neutral sequence because a region of that size is small enough to be computationally efficient while still allowing an accurate approximation of T (Figure S5). Using this method we optimized the likelihood function over a grid of values for L ranging from below the number of coding sites, 50 kb, to more than the number of ampliconic regions, 6 Mb.
The population scaled selection coefficient (Ns) acting on a particular deleterious mutation was drawn from a gamma distribution, with the parameters estimated in Boyko et al. [37], including the same shape parameter (0.184) used above. However, because the Boyko et al. [37] model was developed for the autosomes, and assumes semi-dominant effects, we rescaled the mean strength of selection for a haploid model to represent Y evolution. The scale parameter of the Boyko et al. model (8200) was divided by the ratio of the number of chromosomes used in the original model (51272) to the number of Y chromosomes used in our simulations (5000), then multiplied by 2 because the original model described the fitness of a mutation in the heterozygous state, and all mutations on the Y chromosome will immediately be exposed to selection. Thus, our model used the resulting scale parameter (1600).
We also jointly estimated the number of sites directly under selection (L) and the mean strength of selection by looking at neutral diversity levels on the Y chromosome. We employed an approximate likelihood approach similar to that described above. However, here we investigated a two-dimensional grid of different values for L and a grid of different scale parameters for the gamma distribution of selective effects. Because we kept the shape parameter fixed at 0.184, changing the scale parameter changed the mean strength of selection. We found that our estimates of L were largely insensitive to the mean strength of selection. The profile likelihood curve shown in Figure S7 is remarkably similar to the likelihood curve shown in Figure S6, when the mean strength of selection was held constant.
Asymptotic approximate 95% confidence intervals included all points in the log-likelihood curve that fell within 1.92 log-likelihood units from the MLE (Note S1; Figure S6). Linear interpolation was used to find the appropriate cutoff in between grid points. SFS_CODE commands used for this section are given in Note S1.
We performed simulations to evaluate the performance of our approximate likelihood approach to estimate L by simulating 1,000 Y chromosome datasets using SFS_CODE under models of African and European demographic history. No recombination was allowed on the Y chromosome. Each simulation replicate, or simulated dataset, included 7.5 Mb of neutral sequence (equivalent to the size of our observed data) linked to 2 Mb of sites (i.e., L = 2 Mb) where new mutations were subjected to purifying selection (with selection coefficients drawn from the gamma distribution as discussed in Methods). For each simulated region, the approximate likelihood approach was used to estimate L based on the number of segregating sites within the neutral region. The distribution of selection coefficients used in the inference procedure was the same distribution used to simulate the data. The mean and median of the maximum likelihood estimates (MLEs) as well as the coverage properties of the asymptotic 95% confidence intervals (CIs) are shown in Table S5. The asymptotic 95% CIs contain the true value of L 96.6% of the time for the African simulations and 98.3% of time for the European simulations (rather than 95% of the time), suggesting that they are slightly conservative.
We repeated our analyses of whether purifying selection on coding sites can explain the low diversity on the Y chromosome and our estimation of the number of sites affected by purifying selection taking into account unequal male and female population sizes. We also evaluated whether the low diversity on the Y chromosome could be accounted for by purifying selection combined with unequal male and female population sizes. In particular, Hammer et al. [17] and Lohmueller et al. [20] estimate that there were roughly 2.63 females reproducing for each male that reproduces. In other words, Nm = 0.38Nf. Additionally, we performed our own estimate of Nm/Nf from the levels of diversity at putatively neutral sites (those >100 kb from genes) on the X chromosome and the autosomes and estimate Nm = 0.3352N (Table 2). We have shown (Figure 1) that demographic models with an autosomal ancestral effective population size of roughly 10,000 individuals fit the autosomal levels of diversity reasonably well (Table S3). We compute the effective population size of males under a skewed sex ratio by inputting the previously observed Nm/Nf ratio of 0.3352, and the autosomal size of 10,000 individual, in the equation [67]:We then repeated the forward simulations and analyses described above using this value for Nm.
|
10.1371/journal.ppat.1004203 | Toxoplasma gondii Profilin Promotes Recruitment of Ly6Chi CCR2+ Inflammatory Monocytes That Can Confer Resistance to Bacterial Infection | Ly6C+ inflammatory monocytes are essential to host defense against Toxoplasma gondii, Listeria monocytogenes and other infections. During T. gondii infection impaired inflammatory monocyte emigration results in severe inflammation and failure to control parasite replication. However, the T. gondii factors that elicit these monocytes are unknown. Early studies from the Remington laboratory showed that mice with a chronic T. gondii infection survive lethal co-infections with unrelated pathogens, including L. monocytogenes, but a mechanistic analysis was not performed. Here we report that this enhanced survival against L. monocytogenes is due to early reduction of bacterial burdens and elicitation of Ly6C+ inflammatory monocytes. We demonstrate that a single TLR11/TLR12 ligand profilin (TgPRF) was sufficient to reduce bacterial burdens similar to T. gondii chronic infection. Stimulation with TgPRF was also sufficient to enhance animal survival when administered either pre- or post-Listeria infection. The ability of TgPRF to reduce L. monocytogenes burdens was dependent on TLR11 and required IFN-γ but was not dependent on IL-12 signaling. TgPRF induced rapid production of MCP-1 and resulted in trafficking of Ly6Chi CCR2+ inflammatory monocytes and Ly6G+ neutrophils into the blood and spleen. Stimulation with TgPRF reduced L. monocytogenes burdens in mice depleted with the Ly6G specific MAb 1A8, but not in Ly6C/Ly6G specific RB6-8C5 depleted or CCR2−/− mice, indicating that only inflammatory monocytes are required for TgPRF-induced reduction in bacterial burdens. These results demonstrate that stimulation of TLR11 by TgPRF is a mechanism to promote the emigration of Ly6Chi CCR2+ monocytes, and that TgPRF recruited inflammatory monocytes can provide an immunological benefit against an unrelated pathogen.
| Toxoplasma gondii is an apicomplexan parasite that can infect all warm blooded animals, but rodent species are considered the primary reservoirs. Mice that are infected with T. gondii become more resistant to lethal infection with other pathogens. Ly6C+ inflammatory monocytes are innate immune cells that are critical for defense against T. gondii and other infections. Mice with defects in the ability to recruit inflammatory monocytes fail to control T. gondii replication and succumb to overwhelming inflammation. In this study we used a co-infection model to explain why T. gondii-infected mice are more resistant to the bacterium Listeria monocytogenes. We show that stimulation of the rodent specific Toll-like receptor TLR11 by the T. gondii ligand profilin can recruit inflammatory monocytes, and that these monocytes can protect the host against L. monocytogenes. These findings make profilin an important tool for the study of monocyte biology during T. gondii infection of rodents and are especially interesting given that TLR11 is nonfunctional in humans and other vertebrates.
| Toxoplasma gondii is an obligate intracellular Apicomplexan parasite that can infect nearly any nucleated cell of all warm blooded animals. Within warm blooded hosts, T. gondii replicates as a fast growing tachyzoite form, which disseminates throughout the body during acute infection. Over time and under immune pressure, the parasite differentiates into an encysted bradyzoite form within the central nervous system and muscle tissue, which establishes a life-long chronic infection. Approximately 30% of humans are infected with T. gondii but the infection may be asymptomatic in immunocompetent hosts.
T. gondii infection is characterized by a highly polarized Th1 type immune response associated with production of IL-12 by dendritic cells (DCs), neutrophils, and macrophages which drives T and NK cell production of IFN-γ, long regarded as the main mediator of acute and chronic defenses against the parasite [1], [2], [3]. One of the T. gondii proteins known to stimulate IL-12 production is T. gondii profilin (TgPRF), which is required for parasite actin remodeling during host cell invasion and egress, and is also a ligand for TLR11 and TLR12 [4], [5], [6], [7]. Another critical factor for innate defenses are a class of Gr-1+ Ly6C+ monocytes that produce nitric oxide (NO) and TNF-α, and are recruited in a CCR2 dependent manner in response to both oral and parenteral T. gondii infections [8], [9], [10], [11]. MCP-1−/− and CCR2−/− mice do not recruit Ly6C+ monocytes to the lamina propria in response to oral infection, leading to a higher influx of neutrophils and death from intestinal necrosis and inflammation [8], [9]. Similarly, MCP-1−/− and CCR2−/− mice fail to recruit inflammatory monocytes to the peritoneal cavity following i.p. inoculation leading to increased mortality and parasite burdens [10]. Thus, Ly6C+ monocytes are necessary for early control of T. gondii replication and to prevent immune pathology. However, the specific parasite factors that elicit Ly6C+ monocytes during T. gondii infection have not been identified. Ly6Chi monocytes are also recruited during infections with other protozoan and bacterial pathogens, including Listeria monocytogenes [12], [13], [14], [15], [16].
T. gondii sexual reproduction occurs exclusively in the intestines of the feline definitive hosts, making the rodents they prey on key intermediate hosts in the T. gondii lifecycle. T. gondii infection has been shown to alter rodent aversion to cat urine and fear avoidance behaviors in ways that increase the odds of predation and thus parasite reproductive success [17], [18]. Previous studies have also reported that mice infected with T. gondii are more resistant to secondary infections with unrelated pathogens, including L. monocytogenes, Salmonella typhimurium, mengo virus, Cryptococcus neoformans, Besnoita jejuni, Moloney leukemia virus and Schistosoma monsoni [19], [20], [21], [22], [23], [24], which may also serve to increase predation. We have recently shown that stimulation with soluble T. gondii antigens (STAg) reduced viral titers and conferred a survival advantage in mice infected with highly pathogenic H5N1 avian influenza virus [25], demonstrating that treatment with STAg can stimulate immunity against unrelated pathogens. In order to further investigate the mechanisms conferring this immunological benefit, we used a highly tractable L. monocytogenes infection model.
L. monocytogenes is a Gram positive facultative intracellular bacteria commonly associated with outbreaks of the foodborne illness listeriosis. In mice, intravenous inoculation with L. monocytogenes causes highly predictable infection, involving both innate and adaptive immune responses that ultimately clear the bacteria [26], [27]. Before the onset of adaptive immunity, bacteria replicate primarily in infectious foci within cells of the spleen and liver where innate immune responses are critical for controlling early bacterial growth to prevent dissemination and lethal systemic infection. Increased early bacterial burdens in the spleen and liver correlate with the severity and outcome of infection.
Ly6Chi CCR2+ inflammatory monocytes mediate critical innate control of early bacterial replication. During L. monocytogenes infection, Ly6Chi CCR2+ cells emigrate from the bone marrow in a CCR2-dependent manor, and traffic to sites of bacterial infection to differentiate into CD11C+ TNF-α and inducible nitric oxide synthase (iNOS) producing DCs (TipDCs) that enhance bacterial clearance [12], [15], [28]. Emigration of Ly6Chi CCR2+ cells from the bone marrow is directed by MCP-1 and MCP-3, which is mainly produced by non-hematopoietic cells during infection and can be produced by bone marrow mesenchymal stem cells (BMSCs) in response to circulating TLR ligands [12], [16], [29], [30]. Accordingly, CCR2−/− mice have reduced numbers of circulating Ly6Chi monocytes, reduced numbers of TipDCs in the spleen and liver, reduced TNF-α production and are more susceptible to L. monocytogenes infection [12], [15], [16], [28]. IFN-γ and TNF-α are essential to the innate response as mice lacking either cytokine rapidly succumb to L. monocytogenes infection [31], [32], [33].
In this study we show that chronic T. gondii infection or stimulation with STAg provides resistance against L. monocytogenes bacterial infection by reducing bacterial burdens in the major sites of bacterial replication, the spleen and liver. We also show that stimulation with the TgPRF is sufficient to induce this resistance independent of IL-12, T and NK1.1+ cells but cannot completely overcome the requirement for IFN-γ mediated defenses. Most importantly, we show that TgPRF induces production of MCP-1, which results in the trafficking of Ly6Chi CCR2+ inflammatory monocytes into the blood and spleen, and that CCR2-dependent recruitment of these cells is essential to the TgPRF-induced anti-bacterial response. These results demonstrate that stimulation of TLR11 by TgPRF is sufficient to promote recruitment of Ly6Chi CCR2+ inflammatory monocytes, and that these monocytes can provide and immunological benefit against other infections.
Previous research has shown that mice with a chronic T. gondii infection had greater survival or delayed time to death when challenged with a lethal L. monocytogenes infection [19]. Further experiments showed that this protective effect was not transferrable in the serum, and thus was likely a cell mediated response [34]. Although the specific bacterial burdens were not determined for the animals in these studies, early innate control of L. monocytogenes replication correlates well with severity of infection in mice: animals that maintain low bacterial numbers generally go on to clear the infection, whereas failure of innate immunity is associated with high numbers of bacteria, overwhelming sepsis and inevitable death. We hypothesized that the enhanced survival of T. gondii infected mice was due to innate control of L. monocytogenes replication.
To test this hypothesis, we infected naïve and T. gondii chronically infected mice with a lethal inoculum of L. monocytogenes, and then we determined the number of viable bacteria in the spleens and livers 72 hours later. T. gondii-infected mice had significant ∼3.6 log reductions in bacterial burdens in the spleen ∼4.5 log reductions in the liver compared to uninfected controls (Fig. 1A). In our experience, mice with bacterial burden less than 6 log10 CFU/g in the spleen and liver at 72 hours post infection typically remain asymptomatic and survive L. monocytogenes infection; whereas those with higher bacterial burdens usually succumb to infection. As the bacterial burdens in T. gondii infected mice were consistently less than 6 log10 CFU/g in both organs (Fig. 1A), these results suggest that survival of T. gondii infected mice reported previously [19] was due to early reductions in the numbers or replication of L. monocytogenes bacteria.
Our previous work with influenza virus [25] had shown that the protective effects of T. gondii infection could be replicated by treating mice with STAg, a non-infectious lysate of soluble antigens from sonicated T. gondii tachyzoites. STAg contains many T. gondii proteins, including profilin [5], and previous work has shown that STAg can stimulate immune responses similar to those induced by live parasites, including induction of IL-12, TNF-α, IFN-γ, IL-1β, IL-10 and MCP-1 in vivo or in vitro [35], [36], [37]. Consistent with these data, we observed increased levels of IL-12, TNF-α, IFN-γ and MCP-1 in the serum of STAg-stimulated mice within 24 hours (data not shown). We hypothesized that STAg treatment would reduce the bacterial burdens of L. monocytogenes infected mice as well as chronic T. gondii infection. Mice treated with 1 µl of STAg (approximately 1 µg total protein) 24 hours prior to infection with L. monocytogenes had ∼2.5 log reductions in bacterial burden in the spleens and ∼3.8 log reductions in the liver compared to PBS-treated controls (Fig. 1B). These effects were similar to the reduction in bacterial burdens we observed in T. gondii infected mice (Fig. 1A). STAg stimulated mice also experienced significantly less weight loss than PBS treated controls at 72 hours post infection (Fig. 1B). STAg stimulation was effective for reducing bacterial burdens and weight loss when given 2 or 6 hours post L. monocytogenes infection, although the reduction in bacterial burdens began to decline at 6 hours (data not shown).
To determine if the protective components in STAg were protein or other molecules such as RNA or DNA, we subjected STAg to proteinase K digestion. Proteinase K-digested STAg did not reduce bacterial burdens in L. monocytogenes infected mice (Fig. S1A), which suggested that the protective component(s) were protein. To identify the specific protein(s), we subjected STAg to ammonium sulfate (AS) precipitation and assayed the fractions for their ability to reduce the bacterial burdens. The AS precipitation fraction containing the proteins that remained soluble at AS concentrations >60% reduced the bacterial burdens similar to STAg (Fig. S1B). When we subjected these fractions to western blotting with antibodies against several T. gondii proteins, we saw TgPRF was present in the AS >60% fraction (Fig. S1C).
TgPRF is an actin-binding protein involved in parasite gliding motility, host cell invasion and egress, and is known for inducing IL-12 production through stimulation of TLR11 and TLR12 expressed on DCs and macrophages [4], [5], [6]. In order to determine if TgPRF was sufficient to confer protection against L. monocytogenes we stimulated mice with purified recombinant N-terminal his-tagged TgPRF (rPRF) (Fig. 2A). Mice stimulated with 100 ng rPRF 4 hours prior to L. monocytogenes infection had a significant ∼3.4 log reduction in bacterial burdens in the spleen and ∼4 log reduction in the liver compared to PBS-treated animals (Fig. 2A). rPRF-treated mice did not exhibit weight loss in contrast to PBS-treated controls which lost 17% of their starting weight by 72 hours post infection (Fig. 2A).
Because stimulation with rPRF was sufficient to reduce bacterial burdens similar to T. gondii infection (Fig. 1A), we expected rPRF to enhance survival of L. monocytogenes infected mice in our model (Fig. 2B). All (8/8) PBS-treated mice rapidly succumbed to L. monocytogenes infection within 7 days, with the majority of mice succumbing by day 5. In contrast, 100% (8/8 for each group) of mice stimulated with rPRF 4 hours prior to, or 4 hours after, L. monocytogenes infection survived for 30 days, at which point the experiment was terminated. These results demonstrate that rPRF-stimulation is sufficient to reduce bacterial burdens and confer a long-term survival advantage during L. monocytogenes infection.
Although TgPRF can be recognized by TLR11 and TLR12 [5], [6], [7], the ability of rPRF to reduce bacterial burdens was strictly dependent on TLR11. In multiple experiments, TLR11-deficient (TLR11−/−) mice treated with 40-fold more protein (4 µg rPRF) 4 hours prior to L. monocytogenes infection had no reduction in bacterial burden in either the spleen or liver compared to PBS-stimulated controls (Fig. 2C). rPRF-stimulated TLR11−/− mice also showed equivalent weight loss as the PBS controls (Fig. 2C). These results demonstrate that the effects of rPRF are dependent on recognition by TLR11 and that potential contaminants such as LPS do not contribute to the effect.
To determine whether TgPRF was the major T. gondii factor in STAg responsible for the resistance to L. monocytogenes, we stimulated TLR11−/− mice with STAg. Doses of STAg up to 200 µl did not result in significant reductions in bacterial burdens or reduced weight loss during L. monocytogenes infection (Fig. S2 and data not shown). We did observe modest but statistically significant reductions in bacterial burdens in the spleen (∼20-fold) and liver (∼80-fold) with 200 µl of STAg generated from twice the normal number of parasites, or 400 µg of protein (Fig. S2). These results were in contrast to WT mice, in which 1 µg of STAg reduced bacterial burdens by up 300-fold in the spleen and 6,000-fold in the liver (Fig. 1B). It is possible that the TLR11 independent effects of STAg could be due to parasite derived TLR ligands such as nucleic acids and GPI moieties, or other parasite derived proteins. However, it is unlikely that such large doses of STAg, equivalent to material from 1.6×108 lysed parasites, are relevant during natural infection and thus TgPRF is likely to be the main factor in STAg responsible for the resistance to L. monocytogenes.
STAg has been shown to induce cytokines and chemokines including IL-12, TNF-α, IFN-γ, IL-1β, IL-10 and MCP-1 [35], [36], [37]. TgPRF has been shown to induce IL-12 by classes of DCs and macrophages, IFN-α by CD11c+ spleenocytes, and to promote IFN-γ production by NK1.1+ cells [6]. To determine if TgPRF could induce production of other anti-listerial cytokines, we stimulated mice with rPRF then analyzed serum 2 or 24 hours later. rPRF stimulation induced significant production of IL-12 and MCP-1 at 2 and 24 hours, and IFN-γ and TNF-α by 24 hours (Fig. 3). These results show that TgPRF can stimulate the production of multiple cytokines and chemokines in addition to IL-12.
IL-12 mediates defenses against T. gondii by inducing IFN-γ production from NK and T cells, which in turn helps to activate macrophage effector functions, enhancing antigen presentation, and by promoting the differentiation of Th1 cells [2]. IL-12 plays a similar and critical role in L. monocytogenes infections [26], [27]. We hypothesized that the ability of rPRF to reduce the bacterial burdens would require IL-12 signaling. However, we determined that IL-12 signaling was not required for rPRF-induced resistance to L. monocytogenes infection using IL-12Rβ1 deficient (IL-12Rβ1−/−) mice (Fig. 4A). Compared to PBS-treated controls, rPRF-treated IL-12Rβ1−/− mice had significant ∼2.6 log and ∼2.8 log reductions in bacterial burdens in the spleen and livers, respectively. rPRF-treated IL-12Rβ1−/− mice exhibited only mild weight loss of 1.5%, in contrast to PBS-treated controls which lost 14% of their weight. IL-12Rβ1 is also a component of the IL-23 receptor, so these results indicate that both IL-12 and IL-23 signaling are not required for rPRF-induced resistance to L. monocytogenes infection.
IFN-γ is a critical mediator of innate defenses against both L. monocytogenes [31], [32] and T. gondii [1], [3]. Our previous work with STAg and influenza virus found that STAg-induced IFN-γ from NK cells was required to mediate protection against influenza virus [25]. To determine the role of IFN-γ mediated defenses in rPRF-induced protection against L. monocytogenes we treated IFN-γ deficient (IFN-γ−/−) mice with rPRF (Fig. 4B) then infected them with a low but lethal dose of L. monocytogenes, 200 CFU/animal, to account for the extreme susceptibility imposed by IFN-γ deficiency [31], [32]. rPRF-treated IFN-γ−/− mice had a slight but statistically significant 6-fold reduction in bacterial burden in the spleen and 10-fold reduction in the liver compared to PBS-treated controls. Although the bacterial burdens were still high, rPRF-treated IFN-γ−/− mice experienced less weight loss than PBS-treated controls. While all rPRF-treated IFN-γ−/− mice did succumb to L. monocytogenes infection within eight days, the delay was significant relative to PBS-stimulated animals (Fig. 4C). These results suggest that IFN-γ is at least partially required for rPRF-induced protection against L. monocytogenes and that rPRF stimulation cannot overcome the requirement for IFN-γ mediated defenses even at the low infectious doses used.
The major sources of IFN-γ are T cells and NK cells. NK1.1+ cells are the critical source of IFN-γ for early defense against T. gondii [38] and for STAg-induced protection against influenza virus [25]. Similarly, the majority of IFN-γ during early L. monocytogenes infection is produced by NK1.1+ cells [39], [40]. However, T cells can also produce IFN-γ in early responses to T. gondii [41] and L. monocytogenes infection [40], [42]. To determine if either NK1.1+ or T cells were required for rPRF-induced for protection against L. monocytogenes, we created mice deficient in both T and NK cells by depleting Rag1 deficient (Rag1−/−) mice with PK136 (anti-NK1.1) monoclonal antibody. In contrast to IFN-γ−/− mice, Rag1−/− NK1.1-depleted mice had no increase in susceptibility to L. monocytogenes infection and rPRF-stimulation was highly effective in Rag1−/− NK1.1-depleted mice infected with 6×104 CFU/animal, the same dose used for experiments with WT animals (Fig. 4D). rPRF-treated mice had a ∼3.7 log reduction in bacterial burden in the spleen and ∼1.8 log reduction in the liver, compared to PBS-treated controls. rPRF-treated Rag1−/− NK1.1-depleted mice also did not show weight loss (Fig. 4D). Similar results were observed with rPRF treatment in singly deficient Rag1−/− or wild-type NK1.1-depleted mice (data not shown). These data suggest that neither T nor NK cells are required for rPRF-induced reduction in bacterial burdens and survival. However, in the absence of T and NK cells, mice may develop compensatory defense mechanisms, so it is possible the factors required in these animals are different than in WT mice.
During L. monocytogenes infection, MCP-1 and MCP-3 signals promote emigration of TipDC precursors, Ly6Chi inflammatory monocytes, out of the bone marrow and into circulation in a CCR2-dependent manner [16]. Because serum levels of MCP-1 in rPRF-stimulated mice were significantly increased within 2 hours (Fig. 3) and because T. gondii infection is also known to elicit a population of Ly6C+ monocytes via CCR2 [8], [9], [10], we examined the ability of rPRF to promote emigration of Ly6Chi monocytes. Within four hours after rPRF stimulation, there was an ∼3 fold average increase in the frequency of CD11b+ Ly6Chi monocytes in both the blood and spleens of TgPRF stimulated animals (Fig. 5A and B).The Ly6Chi monocyte population expressed CCR2 (data not shown), consistent with an inflammatory monocyte and TipDC precursor populations described previously [9], [10], [11], [12], [13], [14], [15]. There was also an ∼2.7 fold average increase in the frequency of neutrophils (CD11b+ Ly6Cint Ly6G+) in the blood and a ∼2.5 fold average increase in the spleens of rPRF stimulated mice (Fig. 5A and B). To confirm that these results were specifically attributable to TgPRF, we measured monocyte and neutrophil recruitment in TLR11−/− mice. As expected, there was not an increase the percentage of Ly6Chi monocytes or neutrophils in TLR11−/− mice stimulated with 100 ng rPRF compared to PBS stimulated controls (Fig. S3), demonstrating the specificity of the TgPRF-TLR11 interaction in monocyte and neutrophil recruitment.
Ly6Chi CCR2+ monocyte emigration from the bone marrow into circulation is CCR2-dependent [9], [12]. To determine if Ly6Chi CCR2+ cells recruited in response to rPRF were essential for the reductions in bacterial burdens, we rPRF-stimulated CCR2 deficient (CCR2−/−) mice (Fig. 6A). rPRF-stimulated CCR2−/− mice did not have large reductions in bacterial burdens compared to PBS-treated controls, 2-fold in the spleen and 10-fold in the liver. Although the reductions were statistically significant, they are not likely biologically relevant given the overall high burdens. In addition, both groups experienced equal weight loss.
Although CCR2−/− mice have diminished levels of circulating Ly6Chi monocytes, they have increased numbers in the bone marrow at rest, and large numbers of activated TNF-α producing Ly6Chi monocytes accumulate in the bone marrow during infection [12]. Thus, the small reduction in bacterial burden we saw in rPRF-stimulated CCR2−/− mice could still be dependent on Ly6Chi CCR2+ monocytes, either by activation of a limited number of cells in circulation, or via soluble cytokines such as TNF-α produced by those cells restricted to the bone marrow. To deplete Ly6Chi CCR2+ monocytes, we treated mice with the anti-Gr-1 MAb RB6-8C5, which recognizes a common epitope shared by Ly6C and Ly6G [43]. Depletion with MAb RB6-8C5 reduced neutrophils in the spleens of rPRF stimulated mice by ∼95% and inflammatory monocytes by ∼85% (data not shown). We consistently observed that rPRF-stimulation did not offer any protection in RB6-8C5 depleted mice. There were no significant difference in bacterial burdens between rPRF- and PBS-stimulated mice in either the spleens or livers (Fig. 6B), and both groups experienced equal weight loss (Fig. 6B). Because TLR11 and TLR12 are expressed on macrophages and DCs [7], which may express Ly6C and thus would be depleted by RB6-8C5, we tested the ability of RB6-8C5 depleted mice to respond to profilin by measuring serum cytokine levels 2 hours post rPRF stimulation. rPRF-stimulated RB6-8C5 depleted mice produced significant amounts of IL-12 and MCP-1 (Fig. S4) at levels similar to WT mice at the same timepoint (Fig. 3). rPRF-stimulated RB6-8C5 depleted mice also produced significant levels of TNF-α (Fig. S4).This suggests that the cell population required for recognition of profilin and production of MCP-1 is not subject to depletion by RB6-8C5 MAb.
Because RB6-8C5 significantly depletes Ly6G+ neutrophils as well as Ly6Chi monocytes, we also depleted mice with the Ly6G specific MAb 1A8 [44] to establish the relative contribution of Ly6G+ cells. In contrast to CCR2−/− and RB6-8C5-depleted mice, 1A8-depleted rPRF-stimulated animals were consistently protected against L. monocytogenes infection (Fig. 6C). rPRF-stimulation reduced bacterial burdens in the spleens of 1A8-depleted mice by ∼3 logs and in the livers by ∼2.3 logs, although bacterial burdens in the livers of all 1A8 depleted mice were highly variable. This observation along with the fact that rPRF stimulated CCR2−/− mice had a 10-fold reduction in liver bacterial burdens may indicate that neutrophils play a minor role in defense in this organ. rPRF-stimulated 1A8-depleted mice also did not show weight loss in contrast to PBS-stimulated controls which lost significantly more weight (Fig. 6C).
Together, these results indicate that although rPRF stimulates a large influx of Ly6Cint LyG+ neutrophils into the blood and spleen, these cells are largely dispensable for rPRF induced protection and reduction of bacterial burden in the spleen and liver. While Ly6G+ neutrophils may have a small contribution in the liver following rPRF-treatment, CCR2-dependent recruitment of Ly6Chi CCR2+ inflammatory monocytes plays the central and essential role in rPRF-induced clearance of L. monocytogenes.
In this study we investigated how chronic infection with T. gondii protects the rodent host against unrelated pathogens [19], [20], [21], [22], [23], [24], [25]. Because rodents are the primary reservoir for T. gondii, elucidating the key ligand/receptor interactions is essential for understanding host defense. Our work identifies TgPRF as a T. gondii factor that recruits inflammatory monocytes and demonstrates that stimulation of TLR11 or TLR11/TLR12 heterodimers provides an immunological benefit to a T. gondii-infected host against another pathogen. Stimulation with TgPRF results in production of MCP-1 and recruitment of Ly6Chi CCR2+ inflammatory monocytes and Ly6G+ neutrophils into the blood and spleen, although only Ly6Chi CCR2+ inflammatory monocytes and CCR2-signaling are essential to reduce bacterial burdens. These data have significant implications for our understanding of the biology of T. gondii infection and the evolutionary maintenance of TLR11 in rodents.
Ly6Chi CCR2+ inflammatory monocytes were first identified in L. monocytogenes infection, where they differentiate into TipDCs at the sites of bacterial infection and are essential for early control of bacterial replication. Emigration of these cells out of the bone marrow is directed by the chemokines MCP-1 and MCP-3 and their receptor CCR2 [12], [16]. Accordingly, CCR2 mice have diminished numbers of TipDCs in the spleen and are highly susceptible to L. monocytogenes infection [15]. Ly6Chi monocytes have also been implicated in defense against many other pathogens, including T. gondii. In both oral and parenteral T. gondii inoculation, Gr-1+ Ly6C+ monocytes are recruited to sites of infection and are critical for acute survival [8], [9], [10], [11]. These cells have been shown to produce TNF-α and iNOS, and their emigration is dependent on MCP-1 and CCR2 consistent with inflammatory monocytes or TipDC precursor populations, although interestingly they do not appear to acquire CD11c [8], [9], [10], [11]. MCP-1−/− and CCR2−/− mice, which fail to recruit inflammatory monocytes, have enhanced mortality, greater parasite burdens, and die of pathological inflammation and intestinal necrosis [8], [9], [10], [11]. These studies show that Ly6C+ monocytes are essential for early control of T. gondii replication and to prevent immune pathology. However, the parasite factors that elicit Ly6C+ monocytes had not been identified. Here we identify TgPRF as a mechanism by which T. gondii can elicit emigration of a Ly6Chi CCR2+ inflammatory monocyte population and show that these cells are required for TgPRF to confer resistance to L. monocytogenes infection. In this study stimulation by TgPRF was associated with production of the CCR2 ligand MCP-1 but we did not examine production of other notable CCR2 ligands such as MCP-3. Presumably MCP-3 is also involved in CCR2 dependent inflammatory monocyte recruitment during T. gondii infection as mortality and defects in monocyte recruitment and are less severe in MCP-1−/− than CCR2−/− mice [10], although no specific studies have addressed the role of this chemokine. We also did not determine if TgPRF recruited monocytes acquire CD11c or differentiate into TipDCs during the context of L. monocytogenes infection.
Stimulation with TgPRF also results in trafficking of Ly6Cint Ly6G+ neutrophils into the blood and spleen. Early work suggested that neutrophils were the major cells responsible for controlling the early growth and dissemination of L. monocytogenes [45], [46]. These observations were based mainly on studies using an anti-granulocyte receptor-1 (Gr-1) MAb, which is now known to recognize both neutrophils (Ly6Cint Ly6G+) and non-neutrophil Ly6C+ cells, including subsets of monocytes, macrophages, DCs and lymphocytes [43]. Recent work has suggested that Ly6G+ neutrophils are largely dispensable for innate defenses [47] while others have shown that these cells contribute to significant anti-listerial defenses in the liver [48], [49]. Consistent with these findings, we observed that 1A8 depleted mice are slightly more susceptible to L. monocytogenes than WT mice (lethal dose 1×104 versus 6×104 CFU), although not as susceptible as CCR2−/− (8×103 CFU) or RB6-8C5 depleted animals (200 CFU). The fact that rPRF-stimulated 1A8 depleted mice are resistant to L. monocytogenes infection demonstrates that rPRF-recruited Ly6G+ neutrophils are dispensable for TgPRF-induced protection. Rather, Ly6Chi CCR2+ inflammatory monocytes and TipDCs play the predominant role in TgPRF-mediated defenses.
There are several mechanisms by which TgPRF recruited monocytes may contribute to early control of L. monocytogenes and that could also account for the requirement for IFN-γ. First, inflammatory monocytes may be directly bactericidal. Inflammatory monocytes recruited to the peritoneal cavity during T. gondii infection express iNOS and have enhanced parasite killing in vitro [11], so it is reasonable to infer that TgPRF recruited monocytes would display enhanced activity against L. monocytogenes as well. However, rPRF treatment effectively reduced bacterial burdens in L. monocytogenes infected iNOS deficient mice (data not shown) suggesting that NO production is unlikely to be a primary mechanism of killing. The impaired protection we observed in IFN-γ−/− mice could be due to generalized defects in antimicrobial effector mechanisms dependent on IFN-γ that stimulation with TgPRF cannot overcome or because the development of Ly6Chi inflammatory monocytes into TipDCs and inflammatory DCs during L. monocytogenes and T. gondii infections is largely dependent on NK1.1+ cell derived IFN-γ [50], [51].
Noncognate antigen driven proliferation and activation of memory T cells and innate NK cells could also mediate a degree of resistance dependent on IFN-γ and explain the IFN-γ dependence of TgPRF induced protection. Memory T cells can proliferate, produce IFN-γ and acquire effector cell functions during bacterial infection, which contributes to IFN-γ mediated defenses [52], [53], [54]. Activation is driven by IL-15 and IL-18 production by inflammatory monocytes and CD8α+ DCs, dependent on inflammasome activation, type I interferon and TLR priming [53], [54]. TgPRF could contribute to induction of noncognate memory T cell responses by increases in the number of inflammatory monocytes or serving as the TLR-based priming signal via stimulation of TLR11. Activation of transferred IFN-γ sufficient memory T cells mediated a ∼100-fold reduction in L. monocytogenes bacterial burden in IFN-γ-/- mice, but only modest 3-fold reduction in mice with intact IFN-γ responses [52], [53], [54]. Thus, it is unclear if the 2,500- to 30,000-fold reductions we describe in T. gondii infected or rPRF stimulated IFN-γ sufficient mice can be entirely attributed to cognate antigen independent induction of IFN-γ by memory T cells. Inflammatory monocytes can also induce IFN-γ production by NK cells [53]. Along these lines, TgPRF has been shown to stimulate IFN-γ production by NK1.1+ cells [6] and NK1.1+ derived IFN-γ is required for T. gondii induced protection against influenza [25]. In our model however, NK1.1+ cells do not appear to be essential for TgPRF mediated defenses against L. monocytogenes. The increased importance of NK cells in defense against influenza may be attributable to the comparatively increased role of NK cells in viral infections and killing of virus infected cells.
All of these mechanisms are unable to fully account for the fact that stimulation with rPRF was able to reduce L. monocytogenes bacterial burdens in Rag1−/− NK1.1 depleted mice. It is possible that in the absence of T and NK cells, alternative mechanisms leading to production of IFN-γ may be induced. Neutrophils could be an important source of IFN-γ independent of T and NK cells in our model. Recent evidence has clearly shown that IFN-γ producing neutrophils are present in the peritoneal cavity during T. gondii infection of WT and TLR11−/− mice and are a biologically relevant source of IFN-γ [55]. Neutrophil derived IFN-γ is produced independent of IL-12 [55], which is consistent with our results showing that neither T cells, NK1.1+ cells, nor IL-12 are required for TgPRF-induced resistance to L. monocytogenes. The fact that stimulation with TgPRF elicited a significant number of neutrophils suggests that IFN-γ producing neutrophils could provide a relevant source of non NK1.1+ derived IFN-γ in our model. Future studies will determine if TgPRF elicits these IFN-γ producing neutrophils during T. gondii infection of mice.
The identification of TgPRF as a T. gondii factor that elicits Ly6Chi inflammatory monocytes and neutrophils is especially important for our understanding of T. gondii infection given that humans presumably lack functional TLR11 and TLR12 receptors for TgPRF, yet inflammatory monocytes are critical for innate defenses against T. gondii. In mice, Ly6C+ and Gr-1+ cells are recruited to sites of T. gondii infection in a CCR2 dependent manner and produce TNF-α and iNOS [8], [9], [10], [11]. CCR2−/− and MCP1−/− mice fail to control parasite replication and are highly susceptible to both oral and parenteral T. gondii infection [8], [9], [10]. Lack of inflammatory monocytes is associated with severe inflammation at the sites of T. gondii infection, including increased numbers of neutrophils, intestinal necrosis and CNS pathology [8], [9], [10]. However, the beneficial versus detrimental role of TgPRF is unclear. Similar to mice lacking inflammatory monocytes, lack of TLR11 during systemic T. gondii infection is associated with inappropriate inflammation [56], which suggests a role for TgPRF recruited monocytes in the regulation of systemic immunopathological responses. In contrast, recognition of TgPRF is detrimental during oral T. gondii infection, likely because gut commensal bacteria stimulate anti-parasitic immune responses [57]. WT, but not TLR11−/−, mice develop acute ileitis and liver pathology suggesting that additional parasite or bacterial factors may be sufficient to direct recruitment of inflammatory monocytes in the absence of TLR11, but concurrent stimulation by gut microbes, TgPRF and other T. gondii molecules promotes overwhelming pathological inflammation. The detrimental effects of TgPRF recognition may also be due to TgPRF mediated recruitment of neutrophils, which lead to mucosal pathology [8], [9], [58] and contribute to parasite spread within the intestine [59]. Even so our work presented here shows that recognition of TgPRF and subsequent recruitment of inflammatory monocytes provides a host the benefit of innate defense against an unrelated pathogen. It is possible that carriage of T. gondii may have driven the maintenance of TLR11 specifically in rodent hosts due to this property, and that the interaction of TgPRF with TLR11 or TLR11/TLR12 heterodimers may be critical for this beneficial host-microbe interaction.
Other microbes are known to confer symbiotic-like protection against unrelated pathogens. Latent infection with the murine γ-herpesvirus MHV68 and the β-herpes virus MCMV conferred protection against the bacterial pathogens L. monocytogenes and Yersinia pestis [60]. Protection resulted in increased survival and correlated with 100-fold reductions in L. monocytogenes burdens in the spleen and liver, similar to the results we observed with chronic infection by T. gondii and stimulation with TgPRF. MHV68 infection also confers enhanced resistance to influenza A virus infection associated decreased viral titers, similar to previous results we reported for T. gondii infection [25], [61]. γHV68-induced protection against both L. monocytogenes and influenza was associated with elevated IFN-γ and increased numbers of activated macrophages with enhanced antibacterial activity [60], [61]. These results suggest that herpes virus and T. gondii exploit similar mechanisms to enhance antibacterial innate immunity.
Inflammatory monocytes and TipDCs play key roles in defense against several other pathogens. Ly6C+ monocytes are recruited in CCR2 dependent manner and help initiate protective T cell responses following infection with Mycobacterium tuberculosis, Leishmania major, and Cryptococcus neoformans [14]. The importance of Ly6C+ monocytes against C. neoformans infection may explain prior observations that chronic T. gondii infection confers a survival benefit during co-infection with this pathogen [21]. Ly6C+ monocytes have been shown to reduce Plasmodium chabaudi circulating parasitemia in a mouse model of malaria and to enhance clearance of West Nile Virus [14]. Inflammatory monocytes and TipDCs may play a more limited or even detrimental role in other infections. TNF-α and nitric oxide produced by TipDCs contribute to tissue injury and liver necrosis during infection with Trypanosoma brucei [14]. TipDCs are recruited to the bladder via CCR2 during uropathogenic E. coli infection but are dispensable for bacterial clearance [62]. Future studies will examine the role of TgPRF recruited inflammatory monocytes during T. gondii and other infections.
Animals were housed under conventional, specific-pathogen-free conditions and were treated in compliance with guidelines set by the Institutional Animal Care and Use Committee of the University of Wisconsin School of Medicine and Public Health (IACUC), according to IACUC approved protocol number M01545. This protocol adheres to the regulations and guidelines set by the National Research Council. The University of Wisconsin is accredited by the International Association for Assessment and Accreditation of Laboratory Animal Care.
Unless indicated otherwise, all mice used in this study were on a C57BL/6 background and used at 6–8 weeks of age. Wild-type (WT) mice were purchased from National Cancer Institute – Harlan, Frederick, MD. IL-12Rβ1−/− (002984, B6.129S1-Il2rb1tm1Jm/J), IFN-γ−/− (002287, B6.129S7-Ifngtm1Ts/J), Rag1−/− (002216, B6.129S7-Rag1tm1Mom/J), CCR2−/− (004999, B6.129S4-Ccr2tm1Ifc/J) mice were purchased from Jackson Laboratory (Bar Harbor, ME). TLR11−/− mice were a generous gift from Felix Yarovinsky [5] and were rederived at the University of Wisconsin. A/J mice (National Cancer Institute) were used for protein purification experiments because they more susceptible to L. monocytogenes infection than C57BL/6 mice [63], which allowed us to detect subtle changes in bacterial burdens in partially active fractions. All animals were housed and bred under specific pathogen free conditions at an AALAC accredited facility at the University of Wisconsin School of Medicine and Public Health. All experiments were conducted in accordance with an IACUC approved protocol.
L. monocytogenes strain EGD was a kind gift from C. Czuprynski. Mice were anesthetized with an isofluorane vaporizer connected to an IVIS 200 imaging system (Caliper Life Sciences, Hopkington, MA) then infected via retro-orbital i.v. injection with an appropriate number of bacteria to cause lethal infection as indicated. Animals were monitored daily for clinical signs of disease (ruffled fur, hunched posture, paralysis, etc.) and were euthanized if moribund. At 72 hours post infection, weight loss and bacterial burdens (CFU/g) in the spleen and liver were determined. WT mice were infected with approximately 6×104 CFU (∼6 LD50's), which consistently resulted in death or euthanasia of 100% of control animals. TLR11−/−, IL-12Rβ1−/−, IFN-γ−/−, Rag1−/− NK1.1-depleted, WT RB6-8C5 (Ly6C/Ly6G)-depleted, CCR2−/−, and WT 1A8 (Ly6G)-depleted mice were infected with approximately 4×104 CFU, 1×104 CFU, 200 CFU, 8×104 CFU, 200 CFU, 8×103 CFU and 8×103 CFU respectively. These doses were chosen because they resulted in bacterial burdens and weight loss similar to lethally infected WT mice.
10 week old WT mice were injected i.p. with 250 tachyzoites of the T. gondii strain PruΔ. In order to increase the number of animals that survived greater than 30 days into chronic infection, T. gondii-infected and control uninfected mice were all fed a diet containing sulfadiazine (1,365 ppm) and trimethoprim (275ppm) (TD.06596, Harlan Teklad, Madison, WI) from days 9 through 14 post T. gondii infection, then returned to a normal diet on day 15 through the duration of the experiment.
For experiments in WT mice, soluble T. gondii antigens (STAg) was made from sonicated tissue culture grown tachyzoites (4×108/ml) essentially as described previously [25] and typically had a protein concentration of ∼1 mg/ml. For experiments in TLR11−/− mice, double the amount of parasites (8×108/ml) were used.
Purified recombinant his-tagged T. gondii profilin (rPRF) was a kind gift from F. Yarovinsky [5]. rPRF preparations used in this study had endotoxin levels of 6×10−4 EU or 1×10−2 EU per 100 ng dose (estimated 0.06 and 1 pg endotoxin respectively) as measured by LAL assay (Pierce, Rockford, IL).
Blood was collected from mice via the lateral tail vein 2 or 24 hours post stimulation with rPRF as indicated. Serum was frozen in aliquots at −80°C and then analyzed using a Mouse Inflammation Cytometric Bead Array kit (BD Biosciences, San Jose, CA) according to the manufacturer's instructions.
Spleens were dissociated by mechanical disruption and digested with collagenase/dispase (20 ug/ml, Roche, Indianapolis, IN) and DNAse I (300 ug/ml, Roche) for 30 min at 37°C and passed through a 70 um cell strainer (BD Biosciences, San Jose, CA). Heparinized blood was collected via cardiac puncture and RBCs were removed by dextran sedimentation. Remaining RBCs were lysed with ammonium chloride. Cells were stained at 4°C in PBS with Live/Dead Violet Fixable Stain kit (Invitrogen, Carlsbad, CA), washed, then stained in PBS with 0.5 mM EDTA, 0.2% BSA, 0.09% azide, and 2% normal rat serum (Jackson ImmunoResearch, West Grove, PA). Anti-mouse CD45 (30-F11) APC, anti-mouse CD11b (M1/70) PerCpCy5.5 (eBioscience, San Diego, CA), anti-mouse Ly6C (AL-21) PE, anti-mouse CCR2 (475301) Fluorescein (R&D Systems, Minneapolis, MN), and anti-mouse Ly6G (1A8) PE-Cy7 antibodies were purchased from BD Biosciences except as indicated. Anti-Rat/Hamster CompBeads (BD Biosciences) were used to set compensation. Data were collected on an LSRII cytometer (BD Biosciences) and analyzed with FlowJo 7.6.1 (TreeStar, Ashland, OR).
All antibodies used for depletions were purchased from BioXCell (West Lebanon, NH). To deplete NK cells, mice were treated with the anti-NK1.1 MAb PK136 (250 µg/animal). To deplete both Ly6Chi inflammatory monocytes and Ly6Cint Ly6G+ neutrophils, mice were treated with anti-Gr-1 MAb RB6-8C5 (250 µg/animal) which recognizes a common epitope shared by Ly6C and Ly6G [43]. To deplete only neutrophils, mice were treated with anti-Ly6G MAb 1A8 (250 µg/animal) which has been shown to deplete neutrophils in the spleen, liver and blood [44], [48], [49]. All depletion treatments were administered in PBS via i.p. injection beginning 48 hours prior to stimulation with rPRF, then continued every 48 (1A8) or 96 (PK136 and RB6-8C5) hours after for the duration of the experiment.
Graphs and statistical analysis were made using Graph Pad Prism (San Diego, CA).Graphs represent means and error bars represent standard deviation except where noted otherwise. Bacterial burden data and cytokine were analyzed with the two-tailed student's t-test, and survival data were analyzed using the Log-rank (Mantel-Cox) method. p-values are represented by asterisks in figures as follows: *p<0.05, **p<0.01, ***p<0.001, and ****p<0.001. We consider all p-values <0.05 to be significant.
|
10.1371/journal.pntd.0003088 | MicroRNA-30e* Suppresses Dengue Virus Replication by Promoting NF-κB–Dependent IFN Production | MicroRNAs have been shown to contribute to a repertoire of host-pathogen interactions during viral infection. Our previous study demonstrated that microRNA-30e* (miR-30e*) directly targeted the IκBα 3′-UTR and disrupted the NF-κB/IκBα negative feedback loop, leading to hyperactivation of NF-κB. This current study investigated the possible role of miR-30e* in the regulation of innate immunity associated with dengue virus (DENV) infection. We found that DENV infection could induce miR-30e* expression in DENV-permissive cells, and such an overexpression of miR-30e* upregulated IFN-β and the downstream IFN-stimulated genes (ISGs) such as OAS1, MxA and IFITM1, and suppressed DENV replication. Furthermore, suppression of IκBα mediates the enhancing effect of miR-30e* on IFN-β-induced antiviral response. Collectively, our findings suggest a modulatory role of miR-30e* in DENV induced IFN-β signaling via the NF-κB-dependent pathway. Further investigation is needed to evaluate whether miR-30e* has an anti-DENV effect in vivo.
| Dengue is one of the most prevalent mosquito-borne viral diseases; though it is caused by the Dengue virus (DENV) in tropical/subtropical areas, it has shown tendency toward becoming a global public health concern, with estimated annual numbers of 50–100 million dengue infection cases and 500,000 people with severe disease who require hospitalization worldwide. Thus far no licensed vaccine or specific anti-DENV treatment for dengue is clinically available. Understanding the interaction of DENV with their human hosts is key to identifying potential therapeutic targets. In this work, we found that microRNA miR-30e* significantly suppressed DENV replication by promoting NF-κB-dependent IFN-β production. Our findings identifies miR-30e* as a possible restriction host factor for DENV infection, via positively modulating the antiviral innate immune response. Thus, this work broadens the understanding of the pivotal roles of miR-30e* in the interaction between DENV and the host.
| Dengue is an important mosquito-borne viral disease affecting humans, characterized by a spectrum of symptoms ranging from relatively mild dengue fever (DF) to more severe, and commonly lethal, dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS) [1], [2]. Dengue virus (DENV), the causative agent of dengue, is a positive-polarity, single-stranded RNA virus belonging to the Flaviviridae family. DENV is divided into 4 antigenically related but distinct serotypes, types 1–4 (DENV1–DENV4). An estimated more than 50 million people contract dengue virus annually, leading to approximately 500,000 hospitalizations and 25,000 deaths, particularly among children [3]. Despite an urgent need for effective counter-DENV strategies, thus far neither effective vaccine nor specific antiviral treatment exists for dengue.
The host innate immune system acts as the first line of defense against viruses, and establishment of viral infection requires the pathogen to antagonize such innate immunity [4]. Type I interferons (IFNs), which mainly include IFN-α and IFN-β, are vital components of the anti-viral innate immune system. Rapid synthesis and secretion of these cytokines is critical for host cells to establish an antiviral state. The initial induction of type I interferon is dependent on the recognition and activation of pathogens by pattern-recognition receptors, which further activates transcription factors, such as NF-κB. Under basal conditions, the NF-κB is retained in the cytoplasm by IκBα, which are subject to IκB kinase (IKK)-mediated phosphorylation under stimulation, resulting in degradation of IκBα and translocation of NF-κB into the nucleus [5], [6]. Activation of NF-κB in turn leads to the gene encoding IFN-β (Ifnb1) transcription and IFN-β production combined with IFN regulatory factor 3 (IRF3) [7], which ultimately rendering the cell to establishment of an antiviral state by increasing a subset of IFN-stimulated genes coding antiviral proteins or microRNAs [8]. NF-κB activation is positively regulated by various signaling molecules involved in the repression of its natural inhibitors, such as IκBα. Recent study revealed that zinc finger protein ZBTB20 promotes toll-like receptor-triggered innate immune responses by repressing IκBα gene transcription [9]. Our previous work also demonstrated that miR-30e* promoted nuclear localization and activation NF-κB, via directly interacted with IκBα 3′-UTR and suppresses IκBα expression [10].
microRNAs (miRNAs) are ∼22 nucleotide (nt) short non-coding RNAs (ncRNAs) that modulate gene expression at post-transcriptional level by targeting mRNAs for degradation or by inhibiting translation [11]. Increasing evidence indicates that miRNAs are not only involved in maintenance of normal cell functions, but also participate in host-virus interactions and play a pivotal role in the regulation of viral replication [12]. For example, miR-122, a liver-specific miRNA, facilitates the replication of the viral RNA by targeting the 5′ nontranslational region of hepatitis C virus (HCV) genomic RNA [13]. In addition, miR-323, miR-491, and miR-654 are reported to inhibit replication of the H1N1 influenza A virus through binding to the PB1 gene [14]. Previous studies also provide evidence that modulating cellular miRNAs may be one of the mechanisms that interferon system combat viral infection. Wang et.al demonstrated that cellular inducible miR-155 feedback positively regulates host antiviral innate immune response by promoting type I IFN signaling via targeting suppressor of cytokine signaling 1 (SOCS1) [15].
In the present study, we identified that cellular miR-30e* was up-regulated by DENV infection. Further investigation indicated that miR-30e* suppressed DENV replication by promoting IFN-β production. Additionally, we found that the antiviral effect of miR-30e* is mainly dependent on targeting IκBα in DENV-permissive cells. Therefore, our data suggest that miR-30e* might be an effective approach for improvements of nucleic acid inhibitors of DENV and implies a new therapeutic strategy for DENV infection in humans.
The human monocyte cell line U937 was cultured in RPMI-1640 medium (Invitrogen, Carlsbad, CA) supplemented with 10% fetal bovine serum (FBS) (GIBCO, Carlsbad, CA). The HeLa cell line was cultured at 37°C and 5% CO2 in Dulbecco's modified Eagle's medium (DMEM) (Invitrogen, Carlsbad, CA) supplemented with 10% FBS, 2 mM L-glutamine, 100 µg/ml streptomycin and 100 units/ml penicillin (Invitrogen, Carlsbad, CA). C6/36 Aedes albopictus cells (ATCC, CRL-1660) were maintained at 28°C and 5% CO2 in DMEM supplemented with 10% FBS. The Dengue 1 virus Hawaii strain, Dengue 2 virus New Guinea C strain and Dengue 3 virus H241 strain were kindly provided by the Guangzhou Center for Disease Control [16], [17] and propagated in the mosquito cell line C6/36. Viral stocks were stored at −80°C and titrated on C6/36 cells. For isolation of peripheral blood mononuclear cells (PBMC), whole blood was collected and subjected to Ficoll–Hypaque density gradient centrifugation according to the manufacturer's instruction (Lymphoprep kit, Nycomed, Oslo, Norway) to obtain purified PBMC [18], which were then resuspended and cultured in RPMI-1640 medium (Invitrogen, Carlsbad, CA) supplemented with 10% fetal bovine serum (FBS) (Hyclone, Logan, UT), 15 mM HEPES, 2 mM L-glutamine, 100 µg/ml streptomycin and 100 units/ml penicillin (Invitrogen, Carlsbad, CA).
Total RNA was extracted with Trizol reagent (Invitrogen, Carlsbad, CA) according to the manufacturer's instructions [19], [20]. For the first-strand cDNA synthesis, 500 ng of total RNA was reverse transcribed using random hexamer primer. qPCR reactions were carried out using Fast Start Universal SYBR Green Master Mix (Roche, Basel, Switzerland) and performed on Bio-Rad CFX96 real-time Detection System (Bio-Rad, Hercules, CA). All readings were normalized to the level of GAPDH mRNA. miRNA qRT-PCR was performed using the miRNA-specific TaqMan MicroRNA Assay kit (Applied Biosystems, Grand Island, NY) according to the manufacturer's instructions. miRNA expression was normalized to internal control U6 RNA. The primers sequences are shown in Supplemental Table S1.
Cells (1×104 cells/well) in growth medium were seeded in 96-well flat-bottom plates (in triplicates), and transfected with synthetic miR-30e* mimics or negative control (NC) mimics at a final concentration of 20 nM, or a synthetic specific miR-30e* inhibitor or inhibitor negative control (inhibitor NC) at a final concentration of 50 nM, for additional 48 h. Cell viability was measured by using the MTS (3-(4,5-dimethylthiazol-2-yl)-5-(3- carboxymethoxyphenyl)-2-(4- sulfophenyl)-2H-tetrazolium) assay to monitor cell viability, according to the manufacturer's recommendations. Briefly, 20 µl MTS solution (CellTiter 96Aqueous One Solution reagent, Promega, Madison, WI, USA) was added to each well and incubated for an additional 4 h at 37°C. The absorbance was measured at 490 nm using a microplate reader (Bio-Tek Synergy 2, Winooski, VT, USA).
The pGL3-IκBα-3′-UTR reporter plasmid was based on the pGL3 vector and described previously [10]. Cells were seeded in a 24-well plate 24 h prior to transfection, and 100 ng of pGL3-IκBα-3′-UTR reporter construct along with 10 ng of the control plasmid (pRL-TK Vector; Promega) and miRNA at indicated concentrations were cotransfected into the cells using the Lipofectamine 2000 reagent (Invitrogen, Carlsbad, CA). Twenty-four hours after transfection, the whole cell lysates were harvested and assayed with a Dual-Luciferase Reporter Assay System kit (Promega, San Luis Obispo, CA) to measure the luciferase activity according to the manufacturer's instruction. The IκBα-ORF was generated by subcloning PCR-amplified full-length human IκBα open reading frame (without 3′-UTR) into the pcDNA3.1 vector as previously described [10]. The miR-30e* mimics, negative control (NC) mimics, miR-30e* inhibitor and inhibitor negative control (Inhibitor NC) were purchased from RiBoBio (RiBoBio Inc., Guangzhou, China).
Western blot analysis was performed as described previously [10], [21], using the following primary antibodies: anti-DENV antibody D1-11 (anti-DENV2 E, monoclonal) (Santa Cruz Biotechnology, Santa Cruz, CA), anti-DENV prM antibody (GeneTex, Alton Pkwy Irvine, CA), anti-actin antibody (Sigma-Aldrich, St. Louis, MO) and anti-IκBα antibody (Cell Signaling Technology, Danvers, MA). Protein bands were revealed by horseradish peroxidase-conjugated antibody and enhanced chemiluminescence using a commercial kit (Thermo Fisher Scientific, Rockford, IL) by following the manufacturer's suggested protocols. Immunofluorescence staining was carried out using anti-DENV antibody D1-11 (anti-DENV2 E, monoclonal) (Santa Cruz Biotechnology, Santa Cruz, CA) and Rhodamine-conjugated secondary antibody (Jackson ImmunoResearch Laboratories Inc, West Grove, PA), and the images were captured using the AxioVision Rel.4.6 computerized image analysis system.
U937 and HeLa cells were seeded in a 6-well plate and transfected with the indicated miRNA (20 nM) for 24 h. The supernatants of treated cells were assayed for IFN-β protein release using the Human Interferon-β ELISA Kit (USCN Life Science, Wuhan, China) according to the manufacturer's instruction [22], [23]. Absorbance at 450 nm was read on microplate reader by using a Bio-Tek Synergy 2 microplate reader (Winooski, VT, USA).
Results are expressed as mean ± standard deviations (SD). Statistical analyses were performed on triplicate experiments using two-tailed Student's t test.
To investigate the role of miR30e* in DENV infection, HeLa cells were infected with DENV1, DENV2 and DENV3, respectively, and analyzed for miR-30e* expression by real-time RT-PCR. The results showed that DENV1 infection of HeLa at MOI of 1 for 6 h led to a transcriptional induction of miR-30e* (Figure 1). Similar results were obtained when cells were infected with DENV2 and DENV3 (Figure 1). Taken together, these results suggest that expression of miR-30e* could be induced by DENV infection.
We next examined whether miR-30e* has any effect on DENV replication. U937, HeLa or PBMC cells were transfected with synthetic miR-30e* mimics or negative control (NC) mimics. As shown in supplemental Figure S1A, our results revealed no or little inhibitory effects of miR-30e* or NC mimics on either tested cell lines or primary cells at dose of 20 nM. At 24 h after transfection, cells were challenged with DENV2 at MOI of 1, and cellular and supernatant viral RNA was collected quantified by real-time PCR. miR-30e* overexpression in both cell lines was verified by real-time RT-PCR (Figure 2A). Our results showed that miR-30e* caused a significant reduction of DENV2 RNA in U937, HeLa, and PBMC cells (Figure 2, B and C). Moreover, at the protein level, immunoblotting analysis showed that the expressions of DENV2 prM and envelop protein (E) were markedly suppressed by miR-30e* (Figure 2D), and staining experiments revealed results in accordance with those of the immunoblotting analysis (Figure 2E), suggesting a potent inhibitory effect of host miR-30e* on DENV2 RNA and protein synthesis.
To further investigate whether endogenous miR-30e* was involved in modulating virus replication, U937 and HeLa cells were transfected with a synthetic specific miR-30e* inhibitor or inhibitor negative control (Inhibitor NC). As shown in supplemental Figure S1B, our results also revealed no or little inhibitory effects of miR-30e* inhibitor or inhibitor NC on either U937 or HeLa cell line at dose of 50 nM. At 24 h after transfection, cells were challenged with DENV2 at MOI of 1, and the cellular viral RNA was quantified by real-time RT-PCR. miR-30e* repression in both cell lines was verified by real-time RT-PCR (Figure 3A). As shown in Figure 3B, miR-30e*-inhibited cells exhibited increased replication of DENV2. These results indicated that endogenous miR-30e* functions to suppress DENV2 propagation.
As type I IFN plays a pivotal role in the host antiviral innate immune response, we wondered whether elevated IFN production was responsible for the inhibition of virus replication in miR-30e*-overexpressing cells. Our results showed that miR-30e* significantly induced mRNA and protein expression of IFN-β in U937, HeLa and PBMC cells (Figure 4, A and B). Furthermore, IFN-inducible genes, including OAS1, MxA and IFITM1, were induced by miR-30e* transfection in both U937 and HeLa cells (Figure 4, C and D). These data suggested that miR-30e* suppression of DENV2 infection was closely associated with IFN-β production.
As NF-κB pathway plays an important role in regulating IFN-β production, and our previous studies showed that miR-30e* could activate NF-κB by directly targeting the IκBα 3′-UTR [10], to further investigate the underlying mechanisms responsible for the elevated IFN-β production induced by miR-30e*, we then investigated whether miR-30e* could promote IFN-β production by activating NF-κB. We used the luciferase reporter plasmid containing the 3′-UTR sequences of IκBα mRNA and determined whether miR-30e* could directly target the 3′-UTR sequences in U937 and HeLa cells. In consistence with our previous finding with glioma cells, the results showed that cotransfection of miR-30e* was able to inhibit the luciferase reporter activity (Figure 5A). Furthermore, the protein expression of endogenous IκBα was significantly repressed by miR-30e* in both U937 and HeLa cells (Figure 5B).
In order to understand the role of IκBα in miR-30e*-induced antiviral effect, we studied the effect of ectopically overexpression of IκBα ORF (without 3′-UTR) in miR-30e*-overexpressed cells, and found that it could significantly restore the cellular IκBα protein level (Figure 5C). Our data also indicated that concomitant overexpression of the IκBα ORF (without 3′-UTR) and miR-30e* in U937 and HeLa cells resulted in suppressed IFN-β production (Figure 5D) and robustly abrogated enhanced antiviral effect of miR-30e* (Figure 5E), suggesting that miR-30e* directly targets the 3′-UTR sequences of IκBα, thus enhancing IFN-β production and suppressing DENV replication.
Effective activation of antiviral innate immune responses is essential for the host antiviral defense, which is tightly regulated by a variety of molecular regulators, including miRNAs. Recent evidence indicates that some viruses encode miRNAs that dampen host antiviral immunity, and on the other hand, cellular miRNAs coded by the host can be antiviral via targeting host genes or viral coding sequences [24]. In this work, we report that in DENV-infected cells, inducible miR-30e* restores IFN-β production and inhibits DENV replication, presumably through targeting IκBα and subsequent activation of NF-κB. Our study identifies miR-30e* as a possible restriction host factor for DENV infection via positively modulating the antiviral innate immune response. Further in vivo studies will be required to determine the potential clinical significance of the proposed role of miR-30e* in modulating host cell response to DENV infection, although it remains highly challenging to establish an applicable animal model mimicking DENV pathogenesis, especially immunopathogenesis [25], [26].
Previous studies have indicated that miR-30e* might be a multifunctional microRNA. It is likely miR-30e* might be involved in maintenance of physiological conditions such as heart development [27] and adipogenesis [28], as well as in pathogenesis of diseases such as neural tube defects [29], cancer [10], [30]–[33] and trauma [34]). It was previously shown that the expression levels of miR-30 family were higher in PBMCs collected from patients with chronic hepatitis C compared with those from healthy individuals [35]. Additionally, Pedersen et.al reported that the miR-30 family could be induced by type I IFN in Huh7 cells and primary hepatocytes [36]. However, the functions of miR-30 and the underlying molecular mechanism of this process remain unclear. The key finding of our present study is the identification of inducible miR-30e* in the modulation of DENV multiplication by restoring the innate immune response via activating NF-κB signaling. Since it is not yet totally clear how miRNA is involved in the intrinsic immune system's neutralization of virus threat, this work provides an example of possible mechanisms via which a host cell miRNAs participate in modulating DENV-triggered innate immunity. As IFNs are main mediators of the host antiviral defense system, it would be interesting and important to illustrate the significance of miR-30e* for other viruses, which is under active investigation in the laboratory.
Host innate immunity is the first line of antiviral defense, functional to recognize viral components and produce type I IFN and other proinflammatory cytokines. Type I IFN is extensively employed in clinical therapy of viral infection. However, the efficacies of IFN therapy vary with different viruses, disease stages and the other host factors that influence host responses to IFN [37]. The long co-evolutionary history of viruses and their hosts leads to co-development of the antiviral capability of hosts and counter-antiviral strategies of viruses. Previous reports have demonstrated that DENV is usually a weak inducer of type I IFN responses [16]. It has been recognized that DENV-encoded nonstructural protein NS2B3 physically targets human mediator of IRF3 activation (MITA), and the interaction and cleavage of MITA could block IFN production and subverts the host innate immunity [38]. Our current results extend previous investigations into the modulation of the IFN system and specifically, the ability of a host cellular miRNA, namely, miR-30e*, to upregulate and restore type I IFN production. While our current study has shown that miR-30e* might be a physiologically relevant regulator of IFN function in response to DENV infection, further investigation is needed in the future to evaluate whether it can be of therapeutic significance in the context of in vivo DENV infection. And it requires to be clarified whether promotion of NF-κB dependent innate immunity against DENV infection could simultaneously cause immunopathologic events associated with cytokine overproduction. Furthermore, it is also of note that by showing that viral replication remained relatively unchanged after miRNA and siRNA production was globally abrogated through knocking down Dicer, Bogerd et al recently reported that many viruses are refractory to miRNA or small interfering (siRNA) modulation in the host cells [39], [40]. This study is important because it raises the question whether application of miRNA- or siRNA-based strategies could be therapeutically effective in suppressing viral infection in the clinic. On the other hand, however, as the above study was performed via knocking out Dicer and global endogenous miRNA/siRNA production, it remains to be clarified whether using a high dose, exogenous miRNA could be effective to suppress DENV infection in vivo.
Taken together, our study shows that miR-30e*, as a positive regulator, participates in antiviral innate immune responses once induced in cells upon DENV challenge. This work broadens the understanding of the roles of miR-30e* in the interaction between host and DENV, and further studies to decipher the biological basis for the antiviral activities of miR-30e* will be of theoretical as well as practical importance in developing useful antiviral strategies.
|
10.1371/journal.pmed.1002178 | Educational Outreach with an Integrated Clinical Tool for Nurse-Led Non-communicable Chronic Disease Management in Primary Care in South Africa: A Pragmatic Cluster Randomised Controlled Trial | In many low-income countries, care for patients with non-communicable diseases (NCDs) and mental health conditions is provided by nurses. The benefits of nurse substitution and supplementation in NCD care in high-income settings are well recognised, but evidence from low- and middle-income countries is limited. Primary Care 101 (PC101) is a programme designed to support and expand nurses’ role in NCD care, comprising educational outreach to nurses and a clinical management tool with enhanced prescribing provisions. We evaluated the effect of the programme on primary care nurses’ capacity to manage NCDs.
In a cluster randomised controlled trial design, 38 public sector primary care clinics in the Western Cape Province, South Africa, were randomised. Nurses in the intervention clinics were trained to use the PC101 management tool during educational outreach sessions delivered by health department trainers and were authorised to prescribe an expanded range of drugs for several NCDs. Control clinics continued use of the Practical Approach to Lung Health and HIV/AIDS in South Africa (PALSA PLUS) management tool and usual training. Patients attending these clinics with one or more of hypertension (3,227), diabetes (1,842), chronic respiratory disease (1,157) or who screened positive for depression (2,466), totalling 4,393 patients, were enrolled between 28 March 2011 and 10 November 2011. Primary outcomes were treatment intensification in the hypertension, diabetes, and chronic respiratory disease cohorts, defined as the proportion of patients in whom treatment was escalated during follow-up over 14 mo, and case detection in the depression cohort. Primary outcome data were analysed for 2,110 (97%) intervention and 2,170 (97%) control group patients. Treatment intensification rates in intervention clinics were not superior to those in the control clinics (hypertension: 44% in the intervention group versus 40% in the control group, risk ratio [RR] 1.08 [95% CI 0.94 to 1.24; p = 0.252]; diabetes: 57% versus 50%, RR 1.10 [0.97 to 1.24; p = 0.126]; chronic respiratory disease: 14% versus 12%, RR 1.08 [0.75 to 1.55; p = 0.674]), nor was case detection of depression (18% versus 24%, RR 0.76 [0.53 to 1.10; p = 0.142]). No adverse effects of the nurses’ expanded scope of practice were observed. Limitations of the study include dependence on self-reported diagnoses for inclusion in the patient cohorts, limited data on uptake of PC101 by users, reliance on process outcomes, and insufficient resources to measure important health outcomes, such as HbA1c, at follow-up.
Educational outreach to primary care nurses to train them in the use of a management tool involving an expanded role in managing NCDs was feasible and safe but was not associated with treatment intensification or improved case detection for index diseases. This notwithstanding, the intervention, with adjustments to improve its effectiveness, has been adopted for implementation in primary care clinics throughout South Africa.
The trial is registered with Current Controlled Trials (ISRCTN20283604)
| Non-communicable diseases (NCDs) are the leading cause of deaths worldwide, even in low- and middle-income countries (LMICs) that continue to battle to control communicable diseases like HIV and tuberculosis (TB).
Effective and affordable treatments prevent complications from NCDs like heart attacks and strokes, but access is limited by the variable availability and limited capacity of primary care health workers to detect and effectively manage these conditions. In many LMICs, non-physicians such as nurses provide primary care for NCDs.
Over the past 16 years, we have developed, evaluated, and refined integrated clinical management tools and training programmes that employ problem-based approaches to common symptoms like cough and priority health conditions including TB, HIV, asthma, and emphysema. We have shown them to be effective in improving the quality and outcomes of care for communicable diseases.
We have expanded this programme to include almost all NCDs and mental health. This study evaluated the impact, both benefits and harms, of introducing the expanded programme, called Primary Care 101 (PC101), in terms of the quality of primary care for four common chronic diseases: hypertension, diabetes, chronic respiratory disease, and depression.
We compared the care offered to patients with one of these four chronic diseases in 18 clinics in which primary care health workers were trained in the use of PC101 with that in 18 clinics where nurses continued to use the predecessor tool, which focused on communicable diseases.
The trial had a pragmatic design, meaning it was conducted under usual conditions of health system operational constraints. Clinics in urban and rural areas serving people living in socio-economically deprived areas of South Africa were selected.
We enrolled 4,393 patients with one or more of the NCDs of interest and followed them up for 14 mo after introduction of PC101 at the intervention clinics. The primary outcome of interest was intensification of treatment (or diagnosis, in the case of depression) for the four NCDs, analysed separately.
The results confirmed very high rates of multimorbidity (patients having more than one condition at a time), under-diagnosis, under-treatment, and poor disease control.
Introducing PC101 did not result in intensification of treatment for the four NCDs, but neither was there evidence of harm from the nurses’ expanded scope of practice.
The trial confirmed that multimorbidity and poor detection and control of NCDs and depression are common in this setting. Interventions are necessary to limit the impact of these conditions on people’s health and quality of life.
PC101 offered a practical and acceptable tool to help expand the scope of practice of non-physician clinicians to include NCD care, but we were not able to show improvements in care, as we have previously done for communicable diseases.
The study illustrates the limitations of trials designed to study the effects of complex system interventions in real life, where even small changes across many endpoints, as seen in our study, may be useful to decision-makers under pressure to respond constructively to the rise of multimorbidity and NCDs.
PC101 has been adopted for country-wide implementation in primary care clinics in South Africa.
| South Africa is facing a quadruple burden of disease: HIV and tuberculosis (TB); non-communicable diseases (NCDs), including mental health conditions; injury and violence; and maternal, neonatal, and childhood illnesses [1]. The past 15 years have seen concentrated efforts to strengthen the capacity of the public health system to treat HIV and TB. These investments seem at last to be paying off, with a rise in life expectancy, a decline in mortality [2], and fewer new HIV infections [3]. Yet the burden of NCDs and mental health remains unchecked; cardiovascular disease is now the second leading cause of death in South Africans after communicable diseases [4,5].
In South Africa, responsibility for the detection and treatment of NCDs lies at the primary care level, with nurses seeing nine out of ten patients, most of whom have more than one presenting condition [6]. However, the quality of NCD care is generally poor, characterised by under-diagnosis, under-treatment, and poor clinical control [1,7,8]. We have previously successfully piloted and trialled task-sharing interventions for communicable diseases, increasing the capacity of nurses to take on assessment and prescribing roles for HIV and TB previously restricted to doctors [9–15]. This programme has been scaled up throughout South Africa as part of the national government’s accelerated response to HIV and TB launched in 2010 [16]. A similar programme has been developed for use in other countries including Malawi, Botswana, Brazil, and Mexico [17]. We have since expanded this programme, now called Primary Care 101 (PC101), to include NCDs and mental health, hoping to leverage the health system reforms that accompanied the scale-up of antiretroviral therapy (ART) to improve the quality of primary care for other priority conditions.
These integrated programmes of care seek to overcome the limitations of vertical services that tend to neglect multimorbidity [18–23], and to expand the roles of nurses, increasing the number and distribution of health workers providing treatment for common NCD conditions.
While the benefits of nurse substitution and supplementation for a limited number of NCDs in high-income settings are well recognised [24], evidence from low- and middle-income countries (LMICs) is sparse and limited to a few pilot studies [25–28]. Fewer studies still have sought to improve care across several NCDs simultaneously. Meta-analyses of complex interventions in health systems confirm only small effect sizes (ranging from 0.4% to 6.3%) for carer behaviour (improved care), but given the size of the populations affected, these effect sizes are considered important, provided the interventions are introduced without harm. We report here the findings of the PC101 Trial, a pragmatic cluster randomised study evaluating the effectiveness of the PC101 intervention, which combines provision of an integrated management tool with educational outreach to nurses. The primary outcomes of interest were intensification of treatment for hypertension, diabetes, and chronic respiratory disease and case detection of depression in overlapping cohorts of patients with these conditions.
Ethical approval for the trial was obtained from the University of Cape Town Human Research Ethics Committee (reference number 119/2010) and the Western Cape Department of Health.
This was a pragmatic, parallel-group cluster randomised controlled trial performed in the Eden and Overberg districts of the Western Cape Province. Clusters were public sector primary healthcare clinics randomised within six sub-district strata. Outcome measures in each of four cohorts were assessed in individual patients. Patient cohorts overlapped; patients with more than one condition of interest were included in each applicable cohort, and cohorts were powered and analysed separately. This study design, with multiple cohorts, each with its own primary outcome evaluated simultaneously, aimed to reflect the realities in primary care clinics that nurses are required to diagnose and manage a wide range of conditions, that NCDs are associated with multimorbidity, and that a focus on one condition may compromise the management of others [29]. The Western Cape Department of Health provided consent for the inclusion and randomisation of clinics, before randomisation was performed. Patients provided written consent for data collection after randomisation of clinics and prior to data collection.
Fieldworkers recruited from local communities were trained to collect the trial data. They invited patients seated in the waiting rooms to be considered for the study and screened them using a structured questionnaire. Patients who met the eligibility criteria (Table 1) and provided informed consent were enrolled in the trial and completed the baseline questionnaire in Afrikaans, isiXhosa, or English, administered by the fieldworker using a handheld electronic device. Anthropometry (weight, height, waist circumference) and blood pressure were recorded [47]. Patients were asked to attend a follow-up interview 14 mo after their baseline interview. The lengthy period between interviews was intended to allow adequate opportunity for health workers to intervene in the care of trial patients, given that chronic disease patients are seldom reviewed at clinics more often than every 3–6 mo.
The questionnaire included questions on medical history, smoking status, mental health, health-related quality of life, and socio-economic status. The severity of respiratory symptoms among patients in the respiratory cohort was assessed using the symptom and activity domains of the St George’s Respiratory Questionnaire [48]. Patients who chose to complete the interview in isiXhosa were excluded from this section of the interview as there is no tested isiXhosa translation of this instrument. The presence of symptoms of depression was assessed with the CESD-10, administered to all patients enrolled in the study [32].
Depression treatment was defined as having received counselling, having been referred to psychiatric services, or being on an antidepressant at a therapeutic dose. Low-dose amitriptyline and imipramine are widely prescribed in South Africa for pain management or insomnia. We therefore defined antidepressant use at a therapeutic dose as prescription of amitriptyline or imipramine ≥50 mg daily and/or any other antidepressant. Counselling was defined as “talking with someone in a way that helps to find solutions to problems, or receive emotional support, and not just receiving advice on how to take medication.”
Fieldworkers extracted and photocopied patients’ prescription charts from their folders, clinic stores, and pharmacies for the year preceding the baseline interview. The medically qualified trial manager (N. F.) analysed all prescription charts and recorded prescriptions of chronic medication for each patient at the time of their interview. A data capturer entered the prescription data (medication, dose, and frequency) into a database, and the total daily dose for each medication was calculated. Prescription, interview, and laboratory data were imported and stored in a SQL server database, and a single longitudinal record constructed for every patient by the study database scientist (V. T.).
Reminder letters and cell phone text messages were sent to patients in the month preceding their scheduled follow-up interview. Patients who failed to attend this appointment were traced by phone or home visit. Patients received a gift voucher for a local grocery store with a value of ZAR100 (US$12.25) on completion of the follow-up interview, to compensate for travel costs and time. The follow-up questionnaire was similar to the baseline questionnaire, and fieldworkers repeated the anthropometry and blood pressure measurements. At follow-up, prescription data for the period since baseline were extracted, photocopied, analysed, and documented in the same way as at baseline.
Quality control measures included supervision of fieldworkers, electronic alert messages for fieldworkers if unusually high or low values were entered into the electronic questionnaire, monitoring of the data to identify unusual values or trends, and double entry of prescription data. At follow-up, prescription data were queried if they were missing, if the date of the prescription fell outside of a 1-mo window period based on the scheduled re-interview date, or if cohort-specific medications were excluded.
Blinding of the intervention was not possible at the clinic level due to the nature of the intervention.
The primary outcome for hypertension, diabetes, and chronic respiratory disease was treatment intensification, reflected by an increase in dose or number of medications or change in medication class. This outcome was chosen after considering research identifying clinician inertia as a key reason for failure to control these conditions [49,50]; treatment intensification is associated with improved control [51–53]; was likely appropriate for the study population, where under-treatment was highly prevalent [1,7,8]; fitted well with the focus of the intervention on the clinical practices of nurses and the expansion of their prescribing with training; and could be applied across three of the four chronic conditions of interest. Definitions of treatment intensification by cohort are summarised in Table 1. For the depression cohort, case detection was selected as the primary outcome because depression is recognised to be under-diagnosed and under-treated in primary care [54].
Secondary outcome measures were as follows: disaggregation of primary outcomes by type of medication; cardiovascular disease risk and risk factors such as blood pressure, body mass index (BMI), and smoking status; health-related quality of life measured using the EuroQol-5D [55] and the World Health Organization Disability Assessment Schedule 2.0 (WHODAS 2.0) [56]; mortality; and healthcare utilisation. These last four outcomes were designed to detect evidence of harm resulting from shifting clinical responsibility from doctors to nurses, an often overlooked consideration in evaluations of task-shifting [57].
The study was powered to detect clinically important differences in primary outcomes within each cohort, accounting for the cluster randomisation design. With 38 clinics available for randomisation, we calculated the number of patients needed per clinic for each cohort to detect differences in primary outcomes of between 10% and 15%, with 90% power, 5% significance, and intraclass correlations of outcome based on previous studies, and assuming 20% loss to follow-up (Table 1). Baseline rates of treatment intensification were not available in South Africa, and so we used rates from studies completed in high-income settings [50,58].
HbA1c was measured as part of the pre-planned blood sampling strategy in a subgroup of clinics because resource limitations meant that we could not measure it in all diabetic patients in all 38 clinics. We estimated that HbA1c tests were needed from 30 diabetic patients in 10 clinics in each group (i.e., 600 diabetic patients from 20 clinics in total) in order to a show a difference of 0.5% (HbA1c of 8.8% in the control group versus HbA1c of 8.3% in the intervention group, assuming a standard deviation of 3.4%).
We compared baseline clinic and patient characteristics between treatment groups. All clinics and patients were analysed in the treatment group to which they were randomly assigned. Primary and secondary outcomes were analysed at the patient level, separately within each cohort. No adjustment was made for the multiple disease-specific primary outcomes. The cluster randomisation design was accounted for using robust cluster variance-covariance estimates. Intervention effects were estimated using binomial regression models with treatment as the main effect, adjusted for stratification, and are reported with 95% confidence intervals. Secondary analyses were further adjusted for potentially confounding baseline characteristics such as treatment status and disease control at baseline, smoking status, age, sex, and co-morbidity with one of the study diseases.
We carried out pre-specified subgroup analyses of the primary outcomes stratified by baseline level of disease control using binomial regression models including baseline disease control as a covariate. Baseline disease control of hypertension was defined as blood pressure < 140/90 (or, in patients with diabetes or a history of cardiovascular disease, <130/80), and for diabetes, as HbA1c < 7%. For depression, since the outcome was detection, “control” was defined as any patient receiving treatment for depression as follows: being on antidepressant medication at therapeutic dosage or having received counselling in the past year or having been referred to psychiatric services in the last year. No definition of disease control was applied to patients with chronic respiratory disease. Heterogeneity of the intervention effect was assessed by looking at the interaction between treatment and baseline disease control. In addition, we pre-specified secondary analyses of the primary outcomes disaggregated by component. For the primary outcomes, missing data were considered not to have occurred.
We used linear regression to compare changes between baseline and follow-up in blood pressure, waist circumference, weight, BMI, HbA1c, and health status measures between the treatment groups, adjusted for stratification. Similarly, we used ordinal logistic regression to compare readiness to quit smoking, and Poisson regression to compare rates of healthcare utilisation between the treatment groups. Stata version 13.0 statistical software was used for all analyses.
Fig 1 shows the trial profile. All 38 randomised clinics completed the trial. In all, 4,904 patients were screened, of whom 4,393 patients met the eligibility criteria and were enrolled in the trial. Recruitment targets were exceeded for all cohorts except for diabetes, where recruitment fell short of targets. Enrolment of patients took place between 28 March 2011 and 10 November 2011 and was completed in intervention clinics before educational outreach sessions to nurses began. Follow-up data collection began on 21 May 2012 and ended on 13 December 2012.
In all, 1,927 patients in the intervention group were interviewed at follow-up (1,927/2,166; 89%), and 2,050 in the control group (2,050/2,227; 92%). Reasons for not being re-interviewed were similar between groups: death (63 in the intervention group versus 54 in the control group); relocation (42 in the intervention group versus 26 in the control group); too ill to be re-interviewed (two in the intervention group versus zero in the control group); and could not be traced (132 in the intervention group versus 97 in the control group). Prescription charts could be traced, and thus the primary outcome ascertained, for 206 patients who were not re-interviewed in the intervention group, and 151 in the control group, accounting for the very high rates of patients contributing data to the primary endpoint analysis (Fig 1).
Baseline patient characteristics are presented in Table 2 and detailed in a separate publication [47]. Baseline clinic characteristics are provided in Table A in S1 Appendix. Intervention and control clinics had similar numbers of nurses and doctors. Control clinics tended to be larger and, by chance, had more psychiatric services and on-site pharmacy facilities.
Baseline patient characteristics were generally well balanced between arms. Seventy-three percent of patients were women, and the median age was 52 y. There were high levels of unemployment and receipt of social welfare grants. Multimorbidity was common: 42% of patients had two conditions, and 26% more than two. The percentage of patients with a single condition of interest was as follows: hypertension, 20% (630 of 3,227); depression, 20% (489 of 2,466); diabetes, 8% (148 of 1,842); and chronic respiratory disease, 12% (135 of 1,157). A quarter of patients reported established cardiovascular disease. Eleven percent reported previous TB, and 2% reported being on ART. There were signs of under-treatment and under-diagnosis, with 18% of hypertensive patients reporting no or only one current antihypertensive medication, only 51% of diabetic patients receiving statins, only 50% of those with chronic respiratory disease or symptoms receiving any respiratory medication, and only 25% of those who screened positive for depression reporting some form of relevant treatment for the condition.
There was poor control of hypertension and diabetes despite treatment: blood pressure was ≥140/90 mm Hg in 59% of hypertensive patients, and HbA1c was ≥7% in 77% of those with diabetes in whom HbA1c was measured at baseline (704/1,842; 38%).
Treatment intensification in the hypertension and diabetes cohorts across both the intervention and control groups was common during the study period (Table 3), slightly favouring the intervention group (44% versus 40% for hypertension and 57% versus 50% for diabetes), although these differences were not significant when adjusted for stratification by sub-district and clustering. For hypertension, the risk ratio (RR) was 1.08 (95% CI 0.94 to 1.24; p = 0.252); for diabetes, the RR was 1.10 (95% CI 0.97 to 1.24; p = 0.126). Rates of treatment intensification in the chronic respiratory disease cohort were low (14% in the intervention group versus 12% in the control group) and not significantly different between groups (RR 1.08; 95% CI 0.75 to 1.55; p = 0.674). Fewer patients who screened positive for depression in the intervention group reported receiving treatment for depression at follow-up than their control group counterparts (18% versus 24%), but there was no difference between groups after adjustment for the trial’s design (RR 0.76; 95% CI 0.53 to 1.10; p = 0.142). Adjustment for baseline characteristics (Table 2) did not materially alter these results. The full regression models are presented in Table D in S1 Appendix.
Pre-specified subgroup analyses by baseline level of disease control (Table 4) showed that, in the diabetic cohort, the intervention was associated with treatment intensification only among patients with baseline HbA1c of 7%–10% (RR 1.30; 95% CI 1.16 to 1.47; p-value for interaction = 0.010). In the other cohorts, there were no significant differences in effectiveness between subgroups. However, treatment intensification tended to be more common, in both arms, in subgroups with poorer control at baseline. The non-significant difference in depression treatment, which favoured the control group, was mostly among those already receiving treatment for depression at baseline.
Disaggregated primary outcomes are presented in Table E in S1 Appendix. Notable findings include apparently significantly higher rates of aspirin initiation among patients with hypertension and diabetes attending intervention clinics, even though aspirin prescribing was restricted to doctors. Angiotensin-converting enzyme (ACE) inhibitor use was significantly higher among intervention group patients with known cardiovascular disease, as was sulphonylurea use among intervention group diabetic patients with BMI ≥ 30 kg/m2. In the depression cohort, the higher rate of depression treatment in the control arm was because more control group patients reported receiving counselling (15% in the intervention arm versus 22% in the control arm) and referral to psychiatric services (5% in the interventional arm versus 9% in the control arm). There was no significant difference between groups in the use of antidepressants, which was very low (<5%).
Table 5 reports differences in cardiovascular risk factors between baseline and follow-up. There were no differences between groups in terms of blood pressure, waist circumference, BMI, or HbA1c. Smoking quit rates were high overall, but similar between groups. However, readiness to quit smoking was significantly higher in the intervention group (odds ratio 1.73; 95% CI 1.17 to 2.57).
There were no differences between groups in health outcomes measured with the EuroQol-5D [55], CESD-10 [32], or World Health Organization Disability Assessment Schedule 2.0 [56] (Table 6). Mortality did not differ between groups (Table 6). Healthcare utilisation, as measured by clinic visits and hospital admissions during the 3 mo before the follow-up visit, was similar between groups, but there was a statistically non-significant higher number of hospital admissions in the intervention group (Table 7).
This paper reports our evaluation of the clinical effectiveness of a complex health systems intervention, based on task-shifting by adding nurse-led NCD and depression care to a proven effective, and scalable, integrated care model for nurse-led care of communicable diseases, in the context of limited availability of physicians to treat a high burden of multimorbid and poorly controlled NCDs in a middle-income country.
The primary analyses found no statistically significant effects of the intervention on the primary outcomes for any of the four disease cohorts. These cohorts were analysed separately, equivalent to four parallel trials; adjustment for having four primary outcomes instead of one would only have decreased statistical differences. Health status outcomes also did not differ between the intervention and control groups. But neither was there evidence of harm for any of these endpoints, or in terms of reduced well-being or excess hospitalisations or deaths. In addition, the intervention was not associated with higher healthcare utilisation at the primary care or hospital level. A pre-planned subgroup analysis by baseline level of diabetes control showed a benefit of the intervention in the subgroup of patients with moderately uncontrolled diabetes (HbA1c 7%–10% at baseline), but the two other pre-specified subgroup analyses (for hypertension and depression by baseline level of disease control) did not show a significant difference between groups.
While no primary outcomes showed a significant benefit of the intervention, the upper confidence limits included the possibility of meaningful clinical improvements, and the direction of results in three of the four primary endpoints in the study was consistent and positive. Also, the pre-specified secondary analysis of patients with diabetes and uncontrolled HbA1c measurements at baseline demonstrated a positive effect. After disaggregation of the disease groups, other significant findings were higher rates of aspirin initiation among patients with hypertension and diabetes, higher use of ACE inhibitors in patients with known cardiovascular disease, and more prescriptions of sulphonylureas in patients with diabetes and a high BMI (Table E in S1 Appendix).
The non-significant findings for the primary outcomes contrast with positive findings in our three previous pragmatic randomised controlled trials using a similar integrated management tool and the same training approach, focused on a narrower range of mainly communicable conditions [9–15,30,42,59]. These trials showed modest, but consistent, improvements across a range of process indicators and health and healthcare utilisation outcomes.
There are several potential reasons for the non-significant findings on the primary outcomes of our study. One is the level of uptake of PC101 into daily clinical practice. Owing to limited research funding, a complete and suitably detailed process evaluation of the uptake of PC101 into clinical practice was not possible. However, limited focus group discussions and observations in clinics by members of the research team confirmed heterogeneous uptake of PC101 within and between clinics, as might be expected in a pragmatic trial intervention. Overall low levels of uptake would seem unlikely, given the enthusiastic response and high uptake of the method by clinic staff reported in our previous implementation studies with the PALSA PLUS management tool [12,42]. Other factors should be considered, such as training. The addition of NCD care to the training programme may have proved a step too far—the content of the PC101 management tool was twice as substantial as that of the PALSA PLUS tool—and potentially overwhelming for nurses who were still learning to implement nurse-initiated and -managed antiretroviral treatment when the trial started. Furthermore, NCDs have long been managed by nurses in primary care clinics throughout South Africa, albeit with minimal training or intervention. As seen in the baseline characteristics, poor NCD care may have become entrenched, and markers of poor disease control routinely ignored [60]. The challenge of “undoing” these clinical habits and effecting a change in clinical behaviour is well described and may take repeated training sessions to achieve. Although training was provided throughout the trial, the comprehensive nature of PC101 made it difficult to cover the curriculum for NCDs sufficiently within the time frame of the study. Owing to limited research funding, but consistent with a pragmatic trial design, formal assessments of adequacy of training and uptake (use) of PC101 were not performed.
A further potential reason for the failure to show differences between groups was the effect of a co-intervention, the concurrent Chronic Disease Season campaign, instituted by the clinic managers in both control and intervention clinics. The impact of this unforeseen development is seen in the higher rates of treatment intensification for hypertension and diabetes (the focus of the campaign) than for chronic respiratory disease or depressive symptoms in both the intervention and control clinics. Whereas only 13% of patients with chronic respiratory disease and 3% of those with depression had medication intensified at follow-up, nearly half of those with hypertension and diabetes had intensified treatment (42% and 53%, respectively). These rates of intensification of antihypertensive and diabetic medications are similar to or slightly higher than those reported in high-income country settings [50,58].
Another consideration concerns methodology. We recruited all patients with the diseases of interest rather than only those requiring treatment intensification, and failed to assess adherence and exclude patients who did not adhere to previously prescribed medications and who might therefore have been less likely to have been prescribed additional treatment. However, the eligibility criteria were adopted on the assumption that decision-makers wanted evidence of effectiveness of the intervention across broad groups of patients, rather than for subgroups, and that, as lack of disease control was highly prevalent at baseline, the majority of patients would qualify for treatment intensification.
Other limitations of the study design include dependence on self-reported diagnoses for inclusion in the patient cohorts, reliance on process outcomes, and insufficient resources to measure important health outcomes, such as HbA1c, at follow-up. Also, the duration and timing of the follow-up data collection might not have been optimal for a study of chronic diseases, where follow-up visits being only every 3–6 mo limited opportunities for treatment intensification. This is illustrated by the low number of clinic visits during the follow-up period, a mean of around 2.5 per patient over a period of 14 mo (Table 7).
The main strength of the study was that it was a pragmatic trial, implemented under routine circumstances in a real-world setting with the intervention delivered by usual health department trainers, with minimal research-related distortions of care delivery. Observing this real-world implementation appears to have given relevant policy-makers sufficient confidence to make a decision on the suitability of the intervention for their health systems. Other strengths of the study include the cluster randomised design (appropriate to reduce the risk of contamination in an intervention directed at groups of nurses working in clinics), high follow-up rates for both patient interviews and prescription data, the inclusion of four different chronic diseases in a context characterised by high rates of multimorbidity, and identification and follow-up of patient participants by fieldworkers independent of clinical care.
So what are the implications of the trial for decision-makers in South Africa and other LMICs who are faced with overstretched health services and the need to address NCDs and mental health? In October 2013, even before the trial results were finalised, decision-makers were increasingly enthusiastic about the PC101 intervention, and both the Western Cape Department of Health and the National Department of Health in South Africa elected to commence implementation. Later dissemination of the trial findings on the effectiveness of this intervention to these local and national policy-makers did not change this decision. The decision, we were told, was much more influenced by demand from frontline clinicians and managers for what was perceived to be a highly feasible and acceptable approach to expanding skills for NCDs. Further factors that may have influenced decision-makers were the benefits of the new mode of clinician training reported in our prior studies [9,13,15], an independent report supporting the integrated Chronic Care Model as a feasible component of health system reform in South Africa [61], and the findings of a non-randomised evaluation of PC101 performed in 42 primary care clinics in three additional health districts [62]. The PC101 management tool is correctly seen as a means of overcoming the “silo” approach to individual disease management in which recommendations for different conditions may vary and even conflict and, more importantly, ensures that NCDs and mental health are not overlooked because of prioritisation of communicable diseases. For us, as researchers who look to rigorous research methods to guide health system development, this has been a powerful lesson in understanding that evidence of effectiveness is only one element under consideration by decision-makers [63]. Given clinicians’ strong attraction to the ease of integrating PC101 into clinic practice and the positive system effects of our intervention mentioned above, it might have been more useful to focus our primary analysis on lack of harm. For example, the study was not powered to test for differences in healthcare utilisation and reasons for referrals and hospitalisations. Thus, it is not possible to evaluate the significance of the small imbalance in numbers of hospital admissions between the intervention group and the control group, since an increase in hospitalisations reflecting more appropriate referrals from primary care may be interpreted as favourable rather than as a treatment failure. Specifically designed trials are required.
We now consider that it is our responsibility as health system researchers to invest in improving the effectiveness of this intervention. There are patterns in the data from the trial that provide reassurance that the intervention is not harmful and that, with further optimisation, might demonstrate improvements in effectiveness. Several adjustments have been made to the programme that is being scaled up with the aim of increasing its impact on skills, clinician confidence, and quality of care. The PC101 content has been broken down into four training modules (communicable diseases, NCDs, mental health, and women’s health) to allow staff to become familiar with one area at a time and embed changes into their clinical practice before moving to the next. We now also explicitly aim PC101 training at doctors, through dedicated workshops for professionals who would otherwise miss regular onsite training due to the sessional nature of their work. Implementation workshops, with an extra day aimed at meeting the needs of facility and middle managers, are included in the training of nurse trainers, and appointment of clinical governance teams within sub-districts allows local troubleshooting of barriers to implementation and inclusion of non-clinicians in the day-to-day running of the programme. A further cluster randomised trial in the North West province of South Africa (ClinicalTrials.gov NCT02407691) is currently evaluating the effect of the mental health module when combined with the provision of manualised depression counselling by lay health workers delivered to ART patients with co-morbid depression. A second study is evaluating this mental health module in patients with hypertension and co-morbid depression [64]. This expansion of human resources to include lay health workers is based on our experience from the PC101 trial that nurse training alone is insufficient to close the gap in depression care when there is limited access to treatment in the form of counselling services or antidepressant prescriptions (prescribing currently restricted to doctors).
Although it will not be possible to conduct another randomised controlled trial of the adapted PC101 implementation as it is scaled up, we plan to conduct such trials for future national and international adaptations of this programme [17]. Ease of implementability appears to be a major feature for policy-makers, and we will include proxies, such as acceptability to frontline clinicians, as outcome measures in future trials.
In conclusion, this pragmatic cluster randomised trial of the effects of an integrated management tool implemented using educational outreach to nurses showed no effect on treatment intensification in patients with NCDs or on case detection of depression. But neither was there evidence of harm. Despite this lack of positive clinical outcomes, decision-makers were disposed to view PC101 as a coherent, feasible, and acceptable extension of a programme of integrated care previously shown to be effective in the South African health system, and health authorities have committed to a national rollout of an improved version of the PC101 programme. The disjuncture between the clinical outcomes of our study and the policy choice exposes the different responsibilities of researchers and decision-makers in a health system. For us, as intervention developers, this focuses our attention on longer term improvements to strengthen components of the programme in order to achieve clinical impact on care for NCDs, while, as evaluators, we see the need for ongoing audit and further randomised pragmatic controlled trials to evaluate the effectiveness of these improvements.
Health systems research and development is an interactive and deliberative process. Perhaps the greatest contribution of this study lies in the relationships developed between our team and health system decision-makers, during a series of five large randomised evaluations of health systems interventions that responded to decision-maker-defined health systems needs over 16 years [17]. To this process we have each brought our different skills and perspectives, and together have developed, and are scaling up, an iteratively improved, evidence-informed approach to nurse-led primary care that strengthens human resources and health systems, and brings better care to South Africans, as well as models that can be applied in other low- and middle-income country settings.
|
10.1371/journal.pgen.1000733 | Mu Transposon Insertion Sites and Meiotic Recombination Events Co-Localize with Epigenetic Marks for Open Chromatin across the Maize Genome | The Mu transposon system of maize is highly active, with each of the ∼50–100 copies transposing on average once each generation. The approximately one dozen distinct Mu transposons contain highly similar ∼215 bp terminal inverted repeats (TIRs) and generate 9-bp target site duplications (TSDs) upon insertion. Using a novel genome walking strategy that uses these conserved TIRs as primer binding sites, Mu insertion sites were amplified from Mu stocks and sequenced via 454 technology. 94% of ∼965,000 reads carried Mu TIRs, demonstrating the specificity of this strategy. Among these TIRs, 21 novel Mu TIRs were discovered, revealing additional complexity of the Mu transposon system. The distribution of >40,000 non-redundant Mu insertion sites was strikingly non-uniform, such that rates increased in proportion to distance from the centromere. An identified putative Mu transposase binding consensus site does not explain this non-uniformity. An integrated genetic map containing more than 10,000 genetic markers was constructed and aligned to the sequence of the maize reference genome. Recombination rates (cM/Mb) are also strikingly non-uniform, with rates increasing in proportion to distance from the centromere. Mu insertion site frequencies are strongly correlated with recombination rates. Gene density does not fully explain the chromosomal distribution of Mu insertion and recombination sites, because pronounced preferences for the distal portion of chromosome are still observed even after accounting for gene density. The similarity of the distributions of Mu insertions and meiotic recombination sites suggests that common features, such as chromatin structure, are involved in site selection for both Mu insertion and meiotic recombination. The finding that Mu insertions and meiotic recombination sites both concentrate in genomic regions marked with epigenetic marks of open chromatin provides support for the hypothesis that open chromatin enhances rates of both Mu insertion and meiotic recombination.
| Genomic insertion sites of Mu transposons were amplified and sequenced via next generation technology, revealing more than 40,000 non-redundant Mu insertion sites that are non-uniformly distributed across the maize genome and within genes. Along chromosomes, frequencies of Mu transposon insertions are strongly correlated with recombination rates. Although both Mu and recombination occur preferentially in genes, gene density does not fully explain these patterns. Instead, the finding that Mu insertions and meiotic recombination sites both concentrate in genomic regions marked with epigenetic marks of open chromatin provides support for the hypothesis that open chromatin enhances rates of both Mu insertion and meiotic recombination.
| Gene knockouts are indispensable tools for genetic and functional genomics. The maize Mutator (Mu) transposon is the most active DNA transposon in plants [1]. In maize, a model species for which transformation can be achieved at only a low efficiency, Mu insertion mutagenesis has been an important tool for cloning genes due to its high copy numbers and high rate of germinal transposition [1],[2],[3]. In addition, because Mu elements do not exhibit a preference for transposition to nearby sites [4], as is the case for Ac/Ds transposons [5], they are ideally suited for genome-wide mutagenesis screens.
The Mutator transposon family is a two-component system. MuDR controls the transposition of itself and the other classes of the 12 nonautonomous Mu elements that have been reported so far [6]. All Mu elements share highly similar ∼215 bp terminal inverted repeats (TIRs) and upon insertion generate 9-bp target site duplications (TSDs) directly flanking Mu elements. Mu exhibits a preference for insertion in genes [7],[8],[9]. In addition, a few case studies reported a preference for insertion within 5′-UTRs or exons of genes [7],[8],[9],[10].
Although many investigations have been conducted on Mutator transposons, little is known about the genome-wide distribution of Mu insertions sites and the mechanisms by which these sites are selected. In this study, ∼965,000 Mu flanking sequences (MFSs) were obtained from 454 pyrosequencing libraries generated via Digestion-Ligation-Amplification [11], a novel approach for amplifying unknown sequences flanking known sequences. Analyses of these MFSs revealed 21 novel Mu TIR sequences and 324 genic Mu insertion hotspots that each contains ≥9 independent Mu insertions. Within genes, the Mu insertions exhibited a pronounced preference for 5′-ends with the strongest preference near transcription start sites. Additionally, regions close to the ends of chromosomes experience more Mu insertions than do peri-centromeric regions. This non-uniform pattern is similar to chromosomal distributions of recombination events and gene density. However, gene density does not fully explain the non-uniformity in genome distribution of Mu and recombination. Analyses using both cytosine methylation and histone modification data [12],[13] revealed a strong correlation between Mu insertion and cytosine methylation, H3K4me3 and H3K9ac modifications. Mu insertions and meiotic recombination sites both concentrate in genomic regions marked with epigenetic marks of open chromatin. We, therefore, hypothesize that open chromatin structure plays a key role in determining site selection of both Mu insertions and meiotic recombination events.
DLA is a PCR-based method to amplify unknown sequences flanking known sequences [11]. DLA was adapted to sequence Mu flanking sequences using 454 pyrosequencing, a strategy that is termed DLA-454 [11]. DLA is a novel adaptor-mediated PCR-based method that uses a single-stranded oligo as an adaptor and the conserved ∼215 bp TIRs of Mu transposons as primer binding sites to amplify MFSs. In DLA-454, 6-bp barcodes [14] are inserted between the 454 primer A and a Mu-specific primer, while an adaptor primer, Nsp-P, is appended to the 454 primer B. The resulting library is sequenced using 454 primer A. By doing so, sequencing reads should begin at the barcode, followed by the Mu-specific primer and a portion of the TIR (pTIR), and end with the MFS or in cases of short MFSs the Nsp-P primer. From two technically replicated 454 GS-FLX runs, ∼964,808 reads were obtained. 99% of these sequences can be unambiguously categorized using the barcodes because the first 6 bp of each read exactly matched one of the barcode sequences. A two-step trimming strategy (Methods) was applied to remove barcodes, Mu primer, amplified Mu TIR, 454 primer B and the Nsp-P adaptor primer to obtain MFSs. Based on the results of this two-step trimming process, almost all (99.7%) reads include the Mu-specific primer and over 94% carry amplified Mu TIR sequences, demonstrating that most reads are generated from sites that contain a Mu insertion. Those trimmed MFSs (638,492) that were associated with TIR sequences were aligned to the maize B73 reference genome (B73 RefGen_v1) provided by the Maize Genome Sequencing Project (MGSP) using BLASTN (Figure S1). 58% (370,632/638,492) of the trimmed MFSs satisfied our stringent alignment cut-offs (Methods). This rate of mapping is comparable to that obtained by aligning Mo17 reads (sequenced by Joint Genome Institute using 454 pyrosequencing) to the B73 RefGen_v1 using the same criteria (data not shown). Of the aligning MFSs, 98.6% (365,600/370,632) could be uniquely mapped to a single position on the B73 RefGen_v1 and the positions of the corresponding Mu insertions determined. SNP identified between the MFSs and the sequences of the B73 RefGen_v1 were used to distinguish independent Mu insertions in different plants at the same genomic positions.
About 70% (524,696/755,329) of the 454 reads that resulted from the first trimming contained 34 bp pTIR sequences that perfectly matched known pTIRs. pTIRs from all but one of the previously described Mu elements were detected. Assuming the frequency at which pTIRs were recovered is correlated with the frequency of the corresponding classes of Mu elements in the Mu stocks, we can conclude that Mu1 and MuDR have the highest copy numbers (Figure 1A). Only a few 454 reads contained pTIRs from Mu12 and none contained Mu10 pTIRs. The two TIRs (left and right) of most Mu elements are not perfectly conserved. This allowed us to determine that TIRs from both sides of six classes of Mu elements (Mu1, Mu3, Mu4, Mu7, Mu8 and MuDR) could be successfully amplified via DLA-454. MFS from only one side of four classes of Mu elements (Mu2, Mu5, Mu11 and Mu12) were recovered in the DLA-454 data set (Figure 1).
Approximately, 31% of the DLA-454 reads contain pTIRs that do not perfectly match any known pTIRs. These novel sequences could be the result of sequencing errors or be evidence for the presence of novel pTIRs. We stringently required 34-bp pTIRs to have a minimum edited distance (MED) of at least 2 relative to all known pTIRs before classifying them as potentially novel pTIRs (Methods). A total of 21 novel pTIRs each of which has at least 100 supporting reads were identified (Figure 1B, Table S1). Eight of the Mu elements associated with these novel pTIRs were PCR amplified using the TIR primer in combination with primers designed based on the MFSs associated with the novel pTIR. Seven of the PCR products were successfully sequenced using Sanger technology. All seven novel pTIRs contained the expected polymorphisms relative to known pTIRs, suggesting that this data set has defined 21 novel Mu TIRs. Among the 21 novel TIRs (nTIRs), 13 were associated with multiple independent MFSs (and one, nTIR14, was associated with over 100 independent MFSs), suggesting that they are or were mobile.
It has previously been established that Mu insertions exhibit a preference for typically low-copy genes as compared to non-genes [7],[8],[9]. Our first observation in support of this preference was that only ∼6% of all trimmed MFSs (600,139/638,492) contain repeat sequences as per Emrich et al., 2004 [15]. In addition, more than 98% of mappable MFSs (365,600/370,632) could be uniquely mapped to the genome even though up to 80% maize genome is repetitive [16],[17],[18]. To more directly test whether this preference of Mu elements to insert into genes holds true in our data set, we examined the numbers of Mu insertions in all of the 32,540 annotated genes in the MGSP's “filtered gene set” [18]. Even though the filtered gene set comprises only 7.5% of the genome, almost 75% of the mapped insertions are located within the 13,307 filtered genes. Similar results were obtained when these analyses were repeated with less stringently called gene sets. We therefore conclude that consistent with prior studies, Mu exhibits a strong preference for genic regions.
We then asked whether certain genes are “hotspots” for insertion. To do so, we used a simulation to determine that the probability of one or more genes acquiring nine or more insertions would be rare (p<0.05) if all genes were equally likely to acquire Mu insertions (see Methods). In the experimental data, 1% (324/32,477) of the filtered gene set had nine or more Mu insertions. Variation in gene length was not considered in this simulation because the correlation between Mu insertions and gene length is very low (r = 0.1). We used this set of genic “hotspots” to test the hypothesis that genes that experience high frequencies of Mu insertions are expressed at higher than average levels. Gene expression levels were estimated using mRNA-seq data from several tissues (Methods). Both hotspot genes (≥9 Mu insertions) and all genes that contained 1–8 Mu insertions have significantly higher levels of gene expression than genes without Mu insertions (Wilcoxon-test, all p-values<0.001, Table 1). Hotspot genes exhibit higher levels of gene expression than those genes with 1–8 Mu insertions (Wilcoxon-test, all p-values<0.001, Table 1). This relationship was observed consistently using data from each of three independent mRNA-seq experiments conducted using different tissues. Hence, we conclude that genes that experience elevated rates of Mu tend to be expressed at higher than average levels.
Previously, several studies identified a tendency for Mu insertions to target the 5′ ends of genes. For example, Hardeman and Chandler 1989 [19] reported a preference for the first two exons of bronze1 gene and Dietrich et al. 2002 [10] reported a pronounced preference for the 5′ UTR of the glossy8 gene. This pattern has been confirmed later in multiple genes [12],[20]. To explore the distribution of Mu insertions within genes in our data set we used a set of full-length cDNAs [21] to generate a set of genes (N = 15,050) whose complete structures could be defined (the “flcDNA gene set”; Methods). After aligning the MFSs to the flcDNA gene set, the average numbers of Mu insertions per Mb were computed for each genic region (e.g., 5′ and 3′ UTRs, exons and introns) across all genes in the flcDNA gene set. A Pearson's Chi-square test (see legend of Figure S2) supported the hypothesis that frequencies of Mu insertions vary significantly across genic regions (χ2 = 16,375, df = 7, p-value<2.2e-16). Overall, the 5′ most exons of genes had the highest frequencies of Mu insertions per Mb, particularly the 5′ UTRs and regions further upstream (Figure S2). In contrast, the 3′ portions of genes had relatively low frequencies of Mu insertions. Similar results were obtained using the MGSP's “filtered gene set”. Mu insertions occur in exons at much higher rates than in introns, which is consistent with a previous report using an engineered Mu transposon [7]. But not all exons have higher rates of insertion than introns. Indeed, the previously reported preference of Mu insertion of exons [7] can probably be explained simply by the preference of Mu to insert into 5′-most exons.
Further analyses were performed to understand the pattern of Mu insertion within genes without considering gene structure. Each gene, beginning at the transcription start site and ending at the polyA site, was split into 20 equally sized bins. The number of Mu insertions was counted in each of the 20 bins across all 15,050 genes. Figure 2A reveals a pronounced preference for insertion in the 5′-most bin and decreasing frequencies from 5′ to 3′. This pattern is observed even when using other numbers (viz., 10 and 50) of bins, indicating that Mu transposons exhibit a preference for insertion into the 5′ ends of genes.
To more specifically map the positions of preferred sites for Mu insertion, 400-bp sequences (200 bp each side) surrounding the transcription start sites (TSS) and translation start sites (ATG) were extracted from each of the 15,050 genes. The extracted 400-bp sequences were each divided into 20 bins and the numbers of Mu insertions in each bin counted and plotted, revealing that Mu exhibits a preference for regions 5′ of ATGs (Figure 2B) at or slightly 5′ of the TSS (Figure 2C).
Access to the B73 reference genome allowed us to examine the distributions of Mu insertions across chromosomes. Plotting numbers of Mu insertions per Mb reveals a non-uniform distribution on each of the chromosomes (Figure 3A). Chi-square tests provided strong evidence for non-uniformity on each chromosome (Pearson's test, all p-values <2.2×10−16). To better visualize these non-uniform patterns of Mu insertion, LOWESS curves were plotted (Figure 3B) [22]. For each chromosome, a pronounced “bowl-like” trend was observed in which frequencies of Mu insertions are higher at the ends of chromosomes than in peri-centromeric regions.
Because this trend is reminiscent of the distribution of meiotic recombination events in maize and other species [23],[24],[25], we were interested in examining the genome-wide distribution of meiotic recombination events per Mb and comparing these distributions to those of Mu insertion. We generated a combined genetic map containing 10,143 sequence-based genic markers (Methods). The sequences of 6,362 of these genetic markers could be uniquely aligned to the B73 RefGen_v1 and have consistent positions on both the genetic and physical maps (Methods; Figure S3). Using data from Figure S3, rates of recombination per Mb were calculated for each 1-Mb window and LOWESS curves of rates of recombination per Mb were plotted versus the physical coordinates of the B73 RefGen_v1 (Figure S4).
Each chromosome exhibits a “bowl-like” pattern of recombination per Mb similar to the frequencies of Mu insertions per Mb, which is consistent with previous cytogenetic observations [26]. The similarity between the distributions of Mu insertions and recombination events is also evident at greater granularity; viz., the numbers of Mu insertions and meiotic recombination sites in 1-Mb bins are well correlated genome-wide (r = 0.6).
Because both meiotic crossovers and Mu insertions exhibit preferences for genes, we wondered whether the bowl-like patterns simply reflected gene density. To test this hypothesis we used the MGSP's “filtered gene set” to plot the number of genes (and bp of genic sequence) per Mb across the ten chromosomes. Similar “bowl-like” patterns were observed for the distributions of annotated genes/Mb and the proportion of genic DNA/Mb (data not shown). Even so, after expressing numbers of Mu insertions and recombination rates on a per gene or per bp of genic sequence basis, the bowl-like patterns persist (Figure 4, Figure S5, and Figure S6), demonstrating that the number of Mu insertions per gene (or per bp of genic regions) is generally greater at the ends of chromosomes than near the centromeres. Therefore, gene density can not per se fully explain the “bowl-like” patterns of Mu insertions and recombination.
Weak consensus sequences for Mu insertion sites have been identified by several groups [7],[10],[27],[28]. To extend these results, we identified 2,217 non-redundant Mu insertion sites for which both the left and right MFSs were available and at which the expected 9 bp TSD could be detected. The mid-point of the TSD was set as position zero. The TSD is located at positions −4 to +4. GC content across the 2,217 sequences was calculated for each position (−12 through +12) separately. The null hypothesis that the GC content at each position does not differ from random can be rejected for positions −6 to −3 and +3 to +6 (p-values<0.01; Methods) (Figure 5A). The consensus sequence for position −6 to +6 is “SW::SWNNNNNWS::WS” (consistent with terminology of Dietrich et al. 2002 [10], the TSD is flanked by pairs of double colons), where S represents G or C, W represents A or T, and N can be any base. During Mu insertion the 9-bp TSD is thought to arise via the introduction and subsequent repair of staggered single-strand cuts before and after positions −4 and +4, respectively. These cuts are presumably generated by the Mu transposase. Position −6 to −3 and +3 to +6 have the pattern “SW::SW” and “WS::WS” (subsequently referred to as SWSW), respectively. We therefore hypothesize that SWSW is the preferred sequence of the Mu transposase binding/cutting site.
Although SWSW is the preferred pattern, it is not required for Mu insertion, because only 61 of the 2,127 Mu insertion sites have the exact “SWSW” pattern. The 2,127 Mu insertions are a subset of the whole set of Mu insertions (N = 42,948). The remaining Mu insertion data (N = 40,821) were used to cross-validate the consensus. In this independent data set, 13-bp windows surrounding Mu insertion sites with fewer mismatches relative to the consensus were over-represented (data not shown).
To rule out the possibility that an ascertainment bias explains these results, i.e., that “SWSW” is simply enriched in those genic regions that experience the highest frequencies of Mu insertions (e.g., promoters or 5′-UTRs), similar analyses were performed on regions surrounding certain genic landmarks: viz., transcription start sites (TSS), translation start sites (ATG), translation stops sites (STOP) and transcriptional end sites (END) from the flcDNA genes (Figure 5B, Methods). Surrounding each of these genic landmarks, the frequency of Mu insertions generally decreases as the number of mismatches increases. Interestingly, even after controlling for the number of mismatches in the 13-bp window, frequencies of Mu insertions are highest in the TSS, demonstrating that the TSS is enriched for Mu insertions for reasons other than having a high frequency of the putative transposase binding sites. Hence, the distribution of the consensus sequence is not sufficient to explain the non-uniform distribution of Mu insertion sites within genes.
It has been hypothesized that frequencies of Mu insertions are associated with chromatin structure [7],[29]. To test this hypothesis, frequencies of Mu insertions in single-copy regions of the entire genome associated with various types of histone modifications (Table 2, Methods) were compared. The average number of Mu insertions per Mb was significantly greater for regions that contained H3K4me3 modifications than for regions that contained no H3K4me3 modification (Wilcoxon rank-sum p-value <0.0001). The same held true for H3K9ac and H3K36me3 modifications. However, for H3K27me3 modification, the situation was reversed in that the presence of H3K27me3 modifications was associated with a statistically significant decrease (Wilcoxon rank-sum p-value <0.0001) in the average number of Mu insertions per Mb.
To check for possible interactions among histone modifications with respect to Mu insertions, a linear model with the number of Mu insertions per Mb as the response variable and the presence or absence of each histone modification and all possible interactions as explanatory variables was fit to the data. Each term in the model was significant at the 0.01 level except for two of the four three-way interactions and the four-way interaction among the four indicator variables corresponding to the four histone modifications. Thus, there is good evidence that the effects of the histone modifications on Mu insertion rates are not simply additive.
Table S2 shows that average number of Mu insertions per Mb for all 16 possible combinations of the four histone modifications. A second linear model was fit to the data used to generate Table S2. This model allowed each of the 16 possible histone modification patterns to have a different underlying Mu insertion rate. After testing for differences between each pair of patterns using the Tukey-Kramer method [30] for all pairwise comparisons, many significant differences across histone modification patterns were identified. In particular, regions with all histone modifications except H3K27me3 had a significantly higher average number of Mu per Mb than each of the other 15 patterns. Generally speaking, the H3K9ac or H3K4me3 modifications were most associated with elevated frequencies of Mu insertions among four examined histone modifications. H3K27me3 regions had relatively low frequency of Mu insertions even when other modifications were co-located. In contrast, H3K36me3 regions with H3K4me3 and/or H3K9ac co-located had much higher frequencies of Mu insertions than did H3K36me3 regions without H3K4me3 and/or H3K9ac co-located (Table S2). This illustrates one example of the type of interaction identified in our initial linear model analysis.
We also examined the relationship between DNA methylation and Mu insertion rates. This was done by identifying single-copy regions that either have evidence of containing DNA methylation (McrBC sensitive) or evidence of being hypomethylated (methylation filtration data). Regions with evidence of being methylated and being hypomethylated have frequencies of Mu insertions per Mb 5× lower or 5× higher than the WGS-GSS control, indicating that at least in single-copy regions methylation is strongly negatively correlated with frequencies of Mu insertions.
The stocks used in this study exhibited a high rate of Mu element transposition based on standard Robertson's seedling test assays [31]. Consequently, most Mu insertions investigated in this study were recently generated. Most of them are far less likely to have been subjected to selection than are less active elements. Hence, their distributions are expected to better reflect insertion site preferences than do the distributions of “ancient” transposons detected during genome sequencing projects.
Although there are some exceptions (viz., the long arms of Chr 2 and 5), overall the distributions of Mu insertions and meiotic recombination are similar, suggesting that common features may be involved in site selection for both kinds of events. Both types of events cluster in genes (this study and [32]), but we have demonstrated that gene density is not sufficient to explain the non-uniform distributions of Mu insertions and recombination events along chromosomes.
Both types of events share a preference for GC-rich regions. Meiotic recombination events cluster in GC-rich regions in humans and yeast [33],[34]. Similarly, our data are consistent with a prior report [7] that Mu insertion sites exhibit a bias toward GC-rich regions. The average GC content of the 100-bp intervals surrounding Mu insertion sites is 56% in our data set versus the average 47% GC in the filtered gene set. In addition, like the P elements of Drosophila [35],[36] to which they have been compared mechanistically [37], Mu transposons exhibit a strong preference for 5′-ends of genes, which are typically enriched for GC relative to 3′ ends [38].
What is not clear from these data is whether the high GC content of preferred Mu insertion sites is a cause or an effect of Mu insertion site preference. The finding that the preferred Mu insertion site consensus sequence does not exhibit a strong GC signal (50% GC) suggests that although Mu transposon exhibits a preference for regions that happen to have a high GC content, they do so for reasons other than the GC content of these regions. The overall frequency of Mu insertions in a given sequence is related to its similarity to the preferential consensus sequence (SW::SWNNNNNWS::WS). However, in different genic contexts (e.g., 5′-UTRs, 3′-UTRs) this preferential sequence has dramatically different impacts on Mu insertion frequency, strongly indicating that Mu insertion site selection is more dependent on genic context than on DNA sequences per se.
If DNA sequences are not a major factor in site selection for Mu insertions, what are the causal factor(s)? Frequencies of Mu insertions and meiotic recombination are both low in pericentromeric regions, which are rich in heterochromatin, suggesting an association with chromatin structure. Indeed, our genome-wide results demonstrate that various types of epigenetic modifications are differentially correlated with frequencies of Mu insertions. In addition, the 5′-ends of genes, which are preferred sites for both Mu insertion in maize and meiotic recombination in yeast and maize [25],[33],[34],[35],[36] have distinctive epigenetic modifications. We demonstrated that Mu insertions particularly favor regions surrounding the TSS, which also exhibit strong signal for both H3K4me3 and H3K9ac [13],[39] and exhibit low levels of cytosine methylation [13],[40],[41]. In addition, H3K4me3 modifications which cluster at DSB hotspots in yeast [42] were highly correlated with Mu insertions in this study. Both of these epigenetic modifications are associated with open chromatin structure [39],[40],[43],[44],[45],[46]. Hence, it is likely that chromatin structure plays a key, and perhaps even a causal role, in site selection for both Mu insertion and meiotic recombination. A number of other transposons (e.g., Ac/Ds, P elements, MITEs, and Tos17) exhibit preferences for low-copy genic regions [3],[29],[35],[47],[48],[49],[50],[51], which cluster in the euchromatin. It is therefore possible that these transposons may share a common mechanism of insertion site selection. To further test the hypothesis that chromatin structure is a common feature driving transposon insertion site preferences, the co-localization of new transposon insertion sites and/or transposase binding site with epigenetic marks could be assayed in mutants such as mop1 [52] that would be expected to alter the distribution of epigenetic marks.
The Mu activity of Mu stocks was determined using Robertson's standard seedling test assays [31] that detects the effects of Mu transposition via the appearance of new mutations. The progeny of crosses between Mu active lines and various inbreds and hybrids were self-pollinated to produce the Mu stocks used in this study. The resulting kernels were planted and leaves harvested for genomic DNA isolation. Mu flanking sequences were amplified from these genomic DNAs via DLA [11].
Genetic mapping was conducted using (N≤357) of the IBM (Intermated B73×Mo17) Recombinant Inbred Lines (RILs) as shown in Figure S7 and listed in Table S3. The IBM RILs are the standard mapping population for the maize genetics community.
DLA-454 was conducting following the protocol in [11]. As compared to Liu et al., additional barcode primers were used:
AaMu
5′ GCCTCCCTCGCGCCATCAGTCTGAGGCCTCYATTTCGTCGAATC
AbMu
5′ GCCTCCCTCGCGCCATCAGGTTAGCGCCTCYATTTCGTCGAATC
AcMu
5′ GCCTCCCTCGCGCCATCAGGGTACTGCCTCYATTTCGTCGAATC
AdMu
5′ GCCTCCCTCGCGCCATCAGCATGTGGCCTCYATTTCGTCGAATC
AhMu
5′ GCCTCCCTCGCGCCATCAGATTCTGGCCTCYATTTCGTCGAATC
The positions of histone modifications (H3K4me3, H3K9ac, H3K36me3 and H3K27me3) and cytosine methylation relative to BAC sequences from [13] were downloaded from http://www.ncbi.nlm.nih.gov/geo/ and mapped to the B73 RefGen_v1 (100% identity, 100% coverage). The 80–90% of sequences from each data set that uniquely mapped to the genome were used for further analysis. Whole genome shotgun genomic survey sequences (WGS-GSS, N = 17,232) and methylation filtration (MF) sequences (N = 349,950) [12] (http://magi.plantgenomics.iastate.edu/) [66] were mapped to the B73 RefGen_v1 (≥98% identity, ≥95% coverage). Only reads with a single best alignment (lowest e-value) were considered for the further analysis, resulting in the mapping of 81% (13,938/17,232) and 87% (304,490/349,950) of the WGS-GSS and MF reads, respectively.
Sequence read archive accession no: SRX007377, SRX007378.
|
10.1371/journal.pgen.1003626 | miR-133a Regulates Adipocyte Browning In Vivo | Prdm16 determines the bidirectional fate switch of skeletal muscle/brown adipose tissue (BAT) and regulates the thermogenic gene program of subcutaneous white adipose tissue (SAT) in mice. Here we show that miR-133a, a microRNA that is expressed in both BAT and SATs, directly targets the 3′ UTR of Prdm16. The expression of miR-133a dramatically decreases along the commitment and differentiation of brown preadipocytes, accompanied by the upregulation of Prdm16. Overexpression of miR-133a in BAT and SAT cells significantly inhibits, and conversely inhibition of miR-133a upregulates, Prdm16 and brown adipogenesis. More importantly, double knockout of miR-133a1 and miR-133a2 in mice leads to elevations of the brown and thermogenic gene programs in SAT. Even 75% deletion of miR-133a (a1−/−a2+/−) genes results in browning of SAT, manifested by the appearance of numerous multilocular UCP1-expressing adipocytes within SAT. Additionally, compared to wildtype mice, miR-133a1−/−a2+/− mice exhibit increased insulin sensitivity and glucose tolerance, and activate the thermogenic gene program more robustly upon cold exposure. These results together elucidate a crucial role of miR-133a in the regulation of adipocyte browning in vivo.
| Global obesity and associated health issues have raised the significance of adipocyte biology. Adipose tissues are classified as brown and white adipose. White adipose tissues store lipids, leading to overweight, obesity, insulin resistance and Type2 diabetes. In contrast, brown adipose tissues use lipid storage to generate heat, increase insulin sensitivity, and are negatively correlated with the incidence of Type2 diabetes. Recent studies indicate that white adipose is plastic and contains an intermediate type of adaptive adipocytes (so-called beige/brite adipocytes) that have the energy-dissipating properties of brown adipocytes. Prdm16 is a key molecule that determines the development of both brown and beige adipocytes. Thus, Prdm16 represents a novel molecular switch that expands brown/beige adipocytes and increases energy expenditures. However, how Prdm16 is regulated has been unclear. Here we report that the microRNA miR-133a specifically targets Prdm16 at the posttranscriptional level. Inhibition or knockout of miR-133a significantly increases Prdm16 expression and the thermogenic gene program in white adipose tissues, resulting in dramatically enhanced insulin sensitivity in animals. Our results suggest that miR-133a represents a potential drug target against obesity and Type2 diabetes.
| Adipose tissues are classified as brown (BAT) and white (WAT), and an intermediate category of “brite” or “beige” adipocytes exist within subcutaneous WAT [1], [2]. WAT are located in multiple subcutaneous or visceral locations of body, in the form of distinct fat depots, and contribute to overweight, obesity, insulin resistance and Type 2 diabetes. Brown adipocytes contain more mitochondria and express high levels of Ucp1, a mitochondria inner-membrane channel uncoupling ATP production with oxidative phosphorylation, therefore producing heat and dissipating chemical energy. Due to the global concern of obesity, browning of the white adipocytes, the induction of white adipocytes into beige adipocytes, has become a research focus. The signaling pathway that determines the developmental commitment and differentiation of brown and beige adipocytes is therefore crucial for understanding the process and importance of adipose browning.
Prdm16 is a critical regulator of brown adipocyte development and determines the thermogenic gene program in SAT. Downregulation of Prdm16 in brown adipocytes promotes their fate switch to myoblasts [3]. Conversely, ectopic overexpression of Prdm16 and its co-activator C/EBPβ in myoblasts or fibroblasts transdifferentiated them into brown adipocytes [3], [4]. Similarly, overexpression of Prdm16 in the stromal vascular fraction (SVF) cells of SAT led to the browning of white adipocytes [5]. Mechanistic studies have shown that Type 2 diabetic drug Rosiglitazone can stabilize Prdm16 protein, which activates PPARγ2 and initiates a brown adipocyte gene program that converts white adipocytes to beige adipocytes [6]. The signals that regulate Prdm16 transcription and post-transcriptional modification may offer new strategies for clinical applications and drug discoveries.
MicroRNAs are small non-coding RNAs that negatively regulate mRNA stability or protein translation through targeting the 3′untranslated regions (UTR) of mature mRNA. Previous studies have demonstrated that several myogenic microRNAs (i.e. miR-1, miR-206 and miR-133) are enriched in BAT in relative to WAT [7]. In addition, the cluster of miR-193b and miR-365, downstream signals of Prdm16, are required for brown adipocyte differentiation [8]. Moreover, miR-196a mediates the browning of white adipocytes through targeting Hoxc8, a repressor of brown adipogenic marker C/EBPβ [9]. These studies indicate that microRNAs play important roles in brown adipose development and the browning of white adipocytes.
In the present study, we examined the expression of over 30 microRNAs in the anterior subcutaneous WAT (asWAT) and inguinal WAT (ingWAT) that expressed relatively low and high levels of Prdm16, respectively. We identified several microRNAs whose expression is inversely correlated to Prdm16 expression. Based on this discovery, we conducted gain- and loss- of function studies to demonstrate that miR-133a regulates brown adipocyte biogenesis and browning of white adipocytes through the repression of Prdm16. Analysis of miR-133a knockout mice confirmed the in vivo function of this microRNA in regulating the adaptive plasticity of white adipocytes. We conclude that miR-133a plays a repressive role in adipocyte browning.
In the course of adipogenic marker screening among SAT depots, we found that Prdm16 is expressed at much higher levels in the ingWAT compared to the asWAT (Fig. 1A). Interestingly, we identified four miRNAs (miR-1, miR-206, miR-133a and miR-128) that are expressed at significantly lower levels in the ingWAT compared to the asWAT (Fig. 1B, Fig. S1). The strong inverse correlation between the expression of Prdm16 and miRNAs implies that Prdm16 may be regulated by these miRNAs. Within the 3′ UTR of Prdm16, there are putative target sites for miR-1, miR-206, miR-133a and miR-128 (Fig. 1C), raising the possibility that these miRNAs target Prdm16 mRNA. Using classical luciferase assay in HEK293 cells, we found that miR-133a indeed repressed the luciferase activity by 20% at 1 nM and over 50% at 10–100 nM (Fig. 1D). Mutation of the miR-133a target sequence in the 3′ UTR of Prdm16 totally abolished the repression of luciferase activity by miR-133a (data not shown). miR-128 also repressed the luciferase activity by ∼50% at 10–100 nM (Fig. 1E). Both miR-1 and miR-206 failed to repress the luciferase activity. These results suggest that miR-133a and miR-128 targets the 3′ UTR of Prdm16.
Prdm16 is a transcriptional regulator that controls brown adipocyte fate determination [3]. As miR-133a targets the 3′ UTR of Prdm16, we sought to examine if miR-133a is involved in Prdm16-mediated brown adipocyte commitment and differentiation. To separate committed preadipocytes from more primitive progenitors in the SVF of BAT, we used the aP2-Cre/mTmG mouse model, in which aP2 lineage cells show green fluorescence (mG+) and non-aP2 derived cells exhibit red fluorescence (mT+). Previous studies reported that aP2 expression marks adipocyte progenitors but not bipotential stem cells [10], [11], [12]. The SVF cells of BAT was sorted based on mT and mG fluorescence and subjected to gene expression analyses (Fig. 2A).
Compared to the non-committed (mT+) cells, committed (mG+) preadipocytes completely lost the expression of Pax7 and MyoD, two myogenic genes expressed by BAT, though the expression of dual BAT/muscle marker Myf5 was the same in both populations (Fig. 2B). These data indicate that committed preadipocytes concomitantly downregulate the expression of myogenic genes. The adipogenic commitment of mG+ cells is further supported by the elevated expression of the adipogenic markers, including Pparγ2, Prdm16, Ucp1 and Cidea, compared to the mT+ cells (Fig. 2C). Notably, qPCR results demonstrate that miR-133a is decreased by 80%, but miR-128 is not significantly altered, in the mG+ cells (Fig. 2D), suggesting that miR-133a is more likely to target Prdm16 in vivo. By contrast, the expression of miR-193b and miR-365, previously shown to be direct downstream targets of Prdm16 and required for BAT differentiation [8], were increased in mG+ compared to mT+ cells (Fig. 2D). The mG+ cells also expressed higher levels of miR-143, miR-145 and miR-455 (Fig. 2D), known as adipogenic miRNAs [7]. We further compared the relative expression of miRNAs and BAT related genes in adipose progenitor cells (APC, Fig. 2E) and mature adipocytes collected from the floating fractions of enzymatically digested BAT. Consistently, miR-133a but not miR-128 was significantly downregulated in the differentiated mature adipocytes compared to APC (Fig. 2F). By striking contrast, Prdm16 and other adipogenic markers including aP2, Pgc1α, Cidea and Ucp1 were all dramatically upregulated in mature adipocytes (Fig. 2G–K). Together, these data indicate that miR-133a downregulation along the commitment and differentiation of brown adipocytes might play a role in Prdm16 upregulation during brown adipogenesis.
To examine if miR-133a directly regulates Prdm16 and plays a role in BAT adipogenesis, we overexpressed miR-133a in cultured BAT APCs (Fig. 3A). Electroporation-mediated gene transfer resulted in 213-fold increase in the expression of miR-133a (Fig. 3B). As a consequence, Prdm16 mRNA was downregulated by 48% (Fig. 3C), and the BAT markers Ucp1 and Cidea were downregulated by ∼70% (Fig. 3D). Other adipogenic genes Pparγ2 and Pgc1α were moderately decreased, by 25%∼30% (Fig. 3D). Importantly, the effects of miR-133a overexpression were totally reversed by concomitant overexpression of Prdm16, and even led to ∼3 fold increase (overshoot) in adipogenic marker expression (Fig. 3D–E). This complete reversal and overshoot can be explained as overexpression of the miR-133a insensitive Prdm16 cDNA (lacking 3′ UTR) overrides the repression of miR-133a on endogenous Prdm16. Conversely, we used antisense oligonucleotide LNAs to specifically inhibit miR-133a in BAT APCs. Downregulation of miR-133a led to 40% upregulation of Pparγ2 and ∼3-fold increases of Prdm16, Ucp1, and Cidea (Fig. 3G). Our data suggest that BAT adipogenesis is inhibited by overexpression, and promoted by inhibition, of miR-133a.
Similarly, we overexpressed miR-133a in SAT preadipocytes (Fig. S2A). A 95-fold overexpression of miR-133a led to 38% downregulation of Prdm16 (Fig. S2B), accompanied by 50% downregulation of Pparγ2 and ∼70% downregulation of Ucp1, Cidea and Lhx8 (Fig. S2C). Our results together suggest that miR-133a represses BAT adipogenesis and WAT browning through targeting Prdm16.
miR-133a has two alleles, miR-133a1 and miR-133a2, which have identical sequences and are located in different chromosomes. To investigate the function of miR-133a in BAT and SAT, we examined miR-133a double knockout mice (dKO, miR-133a1−/−a2−/−), generated by intercrossing mice with the genotype of miR-133a1−/−a2+/−. Previous study indicates that compared to wildtype mice, knockout of either miR-133a1 or miR-133a2 led to a 40%–50% downregulation of miR-133a in skeletal and cardiac muscles [13], [14]. We examined the expression of various BAT and mitochondria markers in the dKO mice using miR-133a1−/−a2+/+ littermates as the control. As expected, miR-133a levels were reduced by 80%–98% in BAT, asWAT and ingWAT in the dKO compared to the controls (Fig. 4A, 4D, 4G).
Surprisingly, neither BAT markers (Prdm16, Ucp1 and Cidea) nor mitochondria and lipolysis markers (Cox8a, Hsl, Atgl and Cpt2) were significantly affected in BAT tissue of miR-133a dKO mice (Fig. 4B–C). In striking contrast, the dKO mice had ∼1.5–2 fold elevated expression of the brown adipose markers including Prdm16, Ucp1 and Cidea, and the mitochondria/lipolysis markers including Cox8b, Hsl, Atgl and Cpt2, both in asWAT (Fig. 4E and 4F) and ingWAT (Fig. 4H and 4I). As SAT is responsible for cold- and hormone-induced browning, these data suggest that miR-133a represses browning of white adipocytes in vivo under physiological conditions.
Due to high perinatal lethality (76%) and cardiac myopathy-related postnatal sudden death of the few surviving miR-133a dKO mice [13], [14], we used in the subsequent studies miR-133a1−/−a2+/− mice that had three out of the four miR-133a alleles knocked out but had normal cardiac and skeletal muscles. We used age, gender and genetic background matched WT mice (miR-133a1+/+a2+/+) as control. We reasoned that if we can detect phenotype in mice with 75% reduction of miR-133a, then there should be even more robust effects if miR-133a is completely knocked out.
To examine if the observed upregulation of brown adipose and mitochondrial specific genes in SAT of miR-133a mutants are associated with browning of white adipose, we conducted histological analysis. The ingWAT of miR-133a1−/−a2+/− mice appeared to be browner than that of wildtype mice (Fig. 5A). Western blots confirm that UCP1 protein is indeed upregulated in the ingWAT of miR-133a1−/−a2+/− mice compared to the wildtype mice (Fig. 5B). H&E staining reveals the appearance of numerous multilocular brown adipocyte-like cells in the ingWAT of miR-133a1−/−a2+/− mice, but not wildtype mice (Fig. 5C–D). Immunohistochemical staining with brown adipocyte specific UCP1 antibody indicates that these multilocular brown adipocytes are UCP1+ (brown signal), and the UCP1 immunoreactivity is much more abundant in the ingWAT of miR-133a1−/−a2+/− mice compared to the wildtype mice (Fig. 5E–F).
To directly test how miR-133a affects insulin sensitivity and glucose metabolism, we conducted glucose tolerance test (GTT) and insulin tolerance test (ITT). Strikingly, GTT indicates that the miR-133a mutants had ∼50% lower overnight fasting glucose levels than WT mice (Fig. 5G). The mutants also had much improved glucose tolerance at all the time points examined (Fig. 5G). Similar results were observed by ITT. Upon I.P. administration of insulin (0.75 U/Kg BW), blood sugar dropped much more rapidly and remained lower during 2 h examination period in the miR-133a1−/−a2+/− compared to the wildtype mice (Fig. 5H). These results indicate that reduced miR-133a level is associated with improved insulin sensitivity and glucose disposal in vivo.
Subcutaneous white adipose is capable of thermogenesis under cold exposure. The adaptive thermogenesis capacity is correlated to the level of Prdm16 expression. To investigate if inhibition of miR-133a promotes the adaptive thermogenesis of white adipose, we exposed miR-133a1−/−a2+/− and wildtype mice to cold environment. After 5 d exposure at 4°C, the level of miR-133a in miR-133a1−/−a2+/− ingWAT is about 2% of that in wildtype ingWAT (Fig. 6A), but the level of Ucp1 is about 130 times higher in miR-133a1−/−a2+/− ingWAT (Fig. 6B). The expression levels of Pparγ2, Prdm16, Pgc1a, and Cidea genes were 1.6-, 2.5-, 5-, and 10-fold higher in the ingWAT of miR-133a1−/−a2+/− mice compared to wildtype mice (Fig. 6C). Accordingly, genes related to mitochondrial function (Cox8b and Cpt2) and lipolysis (Hsl and Atgl) were also upregulated in the miR-133a1−/−a2+/− ingWAT (Fig. 6D). Consistent with the relative mRNA levels, UCP1 protein levels in asWAT and ingWAT of the miR-133a1−/−a2+/− mice are obviously higher than those of the wildtype mice at room temperature and after cold exposure (Fig. S3). Therefore, reduced level of miR-133a promotes the activity of cold-inducible thermogenesis gene program in vivo.
The adaptive thermogenesis of subcutaneous WAT has been shown to be mediated by a population of beige adipocytes [15]. We examined the ingWAT of miR-133a1−/−a2+/− and wildtype mice to address if reduction of miR-133a predisposes white preadipocytes to become beige cells that express unique beige markers and common BAT markers [15]. The expression level of miR-133a is reduced by 95% in the ingWAT of miR-133a1−/−a2+/− mice compared to the wildtype mice (Fig. 7A). Accordingly, the BAT-specific genes Prdm16, Cidea and Ucp1 were expressed at 2.5-, 5.5- and 8.4-fold in the ingWAT of miR-133a1−/−a2+/− mice compared to the wildtype mice (Fig. 7B). Importantly, beige adipocyte specific CD137 and Tmem26 genes were also expressed at higher levels (∼6 times) in the miR-133a1−/−a2+/− ingWAT compared to the wildtype (Fig. 7C). We further isolated and differentiated stromal vascular preadipocytes from subcutaneous WAT of miR-133a1−/−a2+/− and wildtype mice. Adipocytes differentiated from the miR-133a1−/−a2+/− preadipocytes expressed 23 times more Ucp1 than wildtype adipocytes (Fig. 7D). Accordingly, expression of other BAT-specific genes Prdm16, Pgc1α, Pparα, Pparγ2, and Cpt2 were also upregulated in the in vitro differentiated miR-133a1−/−a2+/− adipocytes (Fig. 7E). Together, the in vivo and in vitro gene expression analysis demonstrate that inhibition of miR-133a predispose white preadipocytes to become adaptive beige adipocytes upon differentiation.
We identified miR-133a as a regulator of Prdm16 in vivo. Based on the mutual exclusion model of miRNA-mRNA interactions, the cells that express high levels of miRNAs should have less expression of their targets, and vice versa [16]. However, we found both miR-133a and Prdm16 are expressed at very high levels in BAT compared to WAT. This paradox led us to hypothesize that within the BAT, there are different populations of cells that express high levels of miR-133a or Prdm16, respectively, with the notion that miR-133a is highly expressed in cells expressing low levels of Prdm16, and vice versa. In the course of brown adipocyte commitment and differentiation, aP2 expression marks more committed progenitors and preadipocytes, whereas aP2− cells contain more primitive adipocyte progenitors, mesenchymal stem cells and other cell types [12]. Compared to the more primitive cells (aP2−), aP2+ cells express increased Prdm16 and decreased miR-133a. Orchestrated with this notion is the observation that compared to APCs, differentiated brown adipocytes nearly lost the expression of miR-133a. These data imply that miR-133a-mediated Prdm16 repression occurs mainly in uncommitted stem cells to restrict their differentiation towards brown fat, and maintain their multipotency. The luciferase reporter assay, gain- and loss-of-function studies provide direct evidence that miR-133a target Prdm16. In consistency with our study, two recent studies demonstrated that miR-133 can target Prdm16 in both satellite cells and brown adipose cell lines [17], [18].
Interestingly, miR-133a dKO mouse has adipocyte browning in SAT but has no overt phenotype in BAT. Several possibilities might have led to this observation. First, miR-133b, another miR-133 family member, is also highly expressed in BAT (than in WAT) and maintains its expression in the miR-133a dKO BAT. Notably, miR-133b is dramatically downregulated in the asWAT and ingWAT of miR-133a dKO mice for unknown reasons. The loss of both miR-133a and miR-133b in SAT might have led to the upregulation of Prdm16 and activation of the BAT and thermogenic gene program. Second, the loss of miR-133a may be insufficient to further upregulate Prdm16, which is already highly expressed in the BAT. By contrast, Prdm16 is expressed at levels several fold lower in the SAT and the loss of miR-133a can therefore lead to an upregulation of Prdm16.
It has been reported that 76% of miR-133a1−/−a2−/− dKO mice die prior to P10 and the few surviving mice are subjected to sudden death due to cardiomyopathy [13]. Due to the extremely low survival rate of the miR-133a1−/−a2−/− dKO mice [13], [14], we used miR-133a1−/−a2+/− that had three out of the four miR-133a alleles knocked for most in vivo studies. We found that the ingWAT of miR-133a1−/−a2+/− mice contains numerous multilocular adipocytes that express UCP1. Our observation that beige adipocyte specific CD137 and Tmem26 genes are highly upregulated in the miR-133a1−/−a2+/− compared to the wildtype ingWAT suggests that the multilocular adipocytes in the miR-133a1−/−a2+/− WAT are probably the cold-inducible adaptive beige adipocytes. Consistent with this notion, we found that miR-133a1−/−a2+/− mice activate the thermogenic gene program much more robustly than the wildtype mice upon cold exposure. More importantly, ITT and GTT demonstrate that the miR-133a1−/−a2+/− mice exhibit increased sensitivity to insulin and glucose tolerance compared to WT controls. These results together provide strong in vivo evidence that miR-133a regulates the normal physiological function of adipose tissues.
Upon catecholamine hormone stimulations, WAT depots undergo lipid mobilization, a lipolysis process that hydrolyzes triglycerides of white adipocytes [19]. Free fatty acids released from white adipocyte lipolysis undergo beta-oxidation in the mitochondria of brown adipose tissue and provide energy. The Fatty acids also activate special Pparγ2 complex which directly activates Ucp1 expression and dissipates chemical energy [20]. Prdm16 is a Pparγ2 coactivator that drives the brown adipocyte gene program [3]. Here we showed that miR-133a inhibits white adipocyte browning, it would be interesting to study if miR-133a is involved in the repression of hormone stimulated adipocyte browning process.
Obesity has disrupted catecholamine signals, leading to excess fat accumulation and multiple metabolic diseases [19]. Prdm16 drives expression of the browning and thermogenic gene program [2], [5]. Overexpression of Prdm16 in adipose lineage resulted in large number of beige cell formation in SAT and more energy expenditure, which improved glucose metabolism and enhanced insulin sensitivity [2]. However, there is no report to show the anti-obese and anti-diabetic role of Prdm16, or adipocyte browning, in obese and diabetic background. It would be interesting to examine if overexpression of Prdm16 in the ob/ob mice or the db/db mice can ameliorate excessive fat accumulation and improve system insulin sensitivity. In this study we showed that miR-133a negatively regulates Prdm16 and miR-133a KO mice have dramatic phenotype including adipocyte browning, improved glucose metabolism and insulin sensitivity. Consistent with our observation, blockage of endogenous miR133 by antisense nucleotides in mice can greatly lower blood glucose levels [18]. It remains to be investigated if inhibition of miR-133a can increase system energy expenditure in the ob/ob and db/db background.
Interscapular brown adipose is detectable in the newborn humans to maintain body temperature but its mass gradually decreases in the postnatal life [21]. Recent studies have demonstrated that adult humans develop active brown adipocytes in response to cold exposure and the amount of BAT is inversely correlated with body weight [22], [23], [24]. Detailed molecular signature analysis suggested that the adult human brown adipocytes are more similar to murine beige cells, but not the classical interscapular BAT cells [15], [25]. Our study revealed that miR-133a represses white adipocyte browning and beige adipocyte formation in the mouse model. It remains to be investigated if miR-133a also plays a crucial role in the naturally occurred adipocyte browning in humans. As the increased beige adipocytes absorb more glucose and increase insulin sensitivity, it will be interesting to investigate if miR-133a could be a potent drug target for clinical purposes.
All procedures involving the use of animals were performed in accordance with the guidelines presented by Purdue University's Animal Care and Use Committee. mTmG and aP2-Cre mice were from Jackson Lab under stock# 007576 and 005069, respectively. miR-133a knockout mice were previously described [13]. For glucose tolerance test (GTT), 7-week-old mice were fasted overnight and injected with 1.5 mg/g glucose/body weight. For insulin tolerance test (ITT), 2–3 month old mice were fasted for 4 hours and injected with 0.75 U/Kg insulin/body weight. The blood glucose levels were monitored at 30 min intervals for 2 h with an ACCU-CHEK Active Blood Glucose System (Roche) using tail tip blood samples. For cold exposure, mice in their regular filter-top cages with double bedding and nesting materials were placed in ventilated plastic bins and housed in a 4°C room for 5 days.
Primary adipocyte cultures were performed as previously reported [26]. Interscapular BAT and various WAT depots were collected, minced and digested with isolation buffer for proper time at 37°C on a shaker. The isolation buffer contains 123 mM NaCl, 5 mM KCl, 1.3 mM CaCl2, 5 mM Glucose, 100 mM HEPES, 4% BSA, 1%P/S and 1.5 mg/ml Collagenase I. The digestion was stopped with DMEM containing 2%FBS and 1% HEPES, filtered through 100 µm filters, and cells were pelleted at 450× g for 5 min. The cells were cultured in growth medium containing DMEM, 20% FBS, 2% HEPES and 1% P/S at 37°C with 5% CO2 for 3 days, and then fresh media was changed every 2 days. Upon confluence, cells were exposed to induction medium for 4 days and then differentiation medium for several days until adipocytes mature. The induction medium contains DMEM, 10% FBS, 2.85 µM insulin, 0.3 µM dexamethasone (DEXA) and 0.63 mM 3-isobutyl-1-methylxanthine (IBMX) (Sigma), and the differentiation medium contains DMEM, 10% FBS, 200 nM insulin and 10 nM T3.
Plasmids carrying Renila luciferase gene linked to a fragment of Prdm16-3′UTR harboring miR-133a putative binding sites were cotransfected to HEK293 cells, along with control miRNA or miR-133a mimic (Invitrogen). The mutant 3′ UTR of Prdm16 was performed by mutagenesis of the miR-133a recognized sequences from GGACCAA into TTGGTCC. Samples were collected at 48 h post-transfection. Luciferase activity was measured with the use of the Dual Luciferase Assay System (Promega), and the relative luciferase activities were normalized to firefly luciferase. Plasmids carrying firefly luciferase gene linked to fragments of Prdm16-3′UTRs harboring the putative target sites of miR-206, miR-1, or miR-128 were co-transfected to HEK293 cells along with control miRNA, miR-206 mimic, miR-1 mimic, or miR-128 mimic (Invitrogen). The relative luciferase activities were normalized to Renila luciferase.
The stromal vascular fraction of adipose tissues was isolated as described above and cells were filtered through 30 µm sterile nylon mesh. The BAT SVF cells from aP2-mTmG are selected on the basis of fluorescence characteristics. mTmG adipocytes were used as a positive control for gating RFP+ cells. Cell debris and dead cells were removed by staining of dead cell dye. The adipose progenitor cells were sorted out from SVF cells of wildtype mice by antibodies against CD31-PE-Cy7, CD45-PE-Cy7, Ter119-PE-Cy7, CD34-FITC, and Sca1-Pacific blue (eBioscience). After sorting, cells were collected for RNA extraction right away or cultured in CO2 incubator at 37°C for differentiation or transfection experiments.
The transfection was performed by Neon electroporation system (Invitrogen). Final concentration of 500 nM miRNA mimics or miRNA LNAs were incubated with the indicated cells (50,000–100,000 cells) on ice for 5 min and the electroporation is performed under 1150 voltage, 20 msec intervals and 2 pulses. The cells were then seed on 12-well plates. After 12 hours the transfection complex was replaced with fresh adipogenic induction medium. After 4 days of induction, the medium was replaced with adipogenic differentiation medium and the cells were collected for RNA analysis after an additional 4 day differentiation.
The plasmids pMSCV-Prdm16 or the empty vector was transfected by lipofactamine 2000, along with the packing vector pEco to 10-cm Hek293 cells. Freshly isolated 48 h supernatants containing retrovirus particles were filtered and mixed with 4 ug/ml Polybrene. The mixtures (1 ml) were added to each well of BAT APCs, which have been recovered for 8–12 hours after the electroporation of miR-133a. Fresh culture medium was added 12 hours later and was replaced by BAT induction medium after additional 12 hours. Induction and differentiation were same as described above.
RNA was extracted and purified from mature adipocytes of adipose tissues or cell cultures with Trizol and contaminating DNA was removed with DNase I. Random hexamer primers were used to convert RNA into cDNA. For microRNA qPCR, multiple adenosine nucleotides were added to 3′ end of RNAs by E. coli DNA polymerase and cDNAs were synthesized with a specific RT primer [27]. QPCR was performed by using a light cycler 480 (Roche) machine for 40 cycles and the fold change for all the samples was calculated by 2−ΔΔct methods. 18s was used as housekeeping gene for mRNA expression analysis. 18s and U6 mRNA was used as housekeeping gene for microRNA expression analysis.
Serial sections of white fat were cut at 4 µm thick, de-paraffinized, and rehydrated through xylene, ethanol, and water by standard methods. For antigen retrieval, slides were submerged in 0.01 mol/L sodium citrate (pH 6.0) and heated to 96°C for 20 minutes in a laboratory microwave (PELCO). Immunohistochemistry was performed on a Dako Autostainer (Dako, Carpinteria, CA). Slides were incubated with 3% hydrogen peroxide and 2.5% normal horse serum (S-2012, Vector), followed by incubation with rabbit polyclonal anti-UCP-1 primary antibody (ab23841, Abcam) diluted 1∶200 in 2.5% normal horse serum (Vector, S-2012) for 60 minutes. Primary antibody binding was detected with an anti-rabbit horseradish peroxidase (HRP)–ImmPRESS Anti-Rabbit Ig (peroxidase) Polymer Detection Kit (MP-7401, Vector). Labeling was visualized with 3, 3′-diaminobenzidine (DAB) as the chromogen (SK-4105, Vector). Slides were counterstained with Harris hematoxylin (EK Industries, Joliet, IL) and whole slide digital images were collected at 20× magnification with an Aperio ScanScope slide scanner (Aperio, Vista, CA).
The data are presented with mean ± standard error of the mean (SEM). P-values were calculated using two-tailed student's t-test. The ones with P-value less than 0.05 were considered as statistic significant.
|
10.1371/journal.pgen.1004766 | Plasmid Flux in Escherichia coli ST131 Sublineages, Analyzed by Plasmid Constellation Network (PLACNET), a New Method for Plasmid Reconstruction from Whole Genome Sequences | Bacterial whole genome sequence (WGS) methods are rapidly overtaking classical sequence analysis. Many bacterial sequencing projects focus on mobilome changes, since macroevolutionary events, such as the acquisition or loss of mobile genetic elements, mainly plasmids, play essential roles in adaptive evolution. Existing WGS analysis protocols do not assort contigs between plasmids and the main chromosome, thus hampering full analysis of plasmid sequences. We developed a method (called plasmid constellation networks or PLACNET) that identifies, visualizes and analyzes plasmids in WGS projects by creating a network of contig interactions, thus allowing comprehensive plasmid analysis within WGS datasets. The workflow of the method is based on three types of data: assembly information (including scaffold links and coverage), comparison to reference sequences and plasmid-diagnostic sequence features. The resulting network is pruned by expert analysis, to eliminate confounding data, and implemented in a Cytoscape-based graphic representation. To demonstrate PLACNET sensitivity and efficacy, the plasmidome of the Escherichia coli lineage ST131 was analyzed. ST131 is a globally spread clonal group of extraintestinal pathogenic E. coli (ExPEC), comprising different sublineages with ability to acquire and spread antibiotic resistance and virulence genes via plasmids. Results show that plasmids flux in the evolution of this lineage, which is wide open for plasmid exchange. MOBF12/IncF plasmids were pervasive, adding just by themselves more than 350 protein families to the ST131 pangenome. Nearly 50% of the most frequent γ–proteobacterial plasmid groups were found to be present in our limited sample of ten analyzed ST131 genomes, which represent the main ST131 sublineages.
| Plasmids are difficult to analyze in WGS datasets, due to the fragmented nature of the obtained sequences. We developed a method, called PLACNET, which greatly facilitates this analysis. As an example, we analyzed the plasmidome of E. coli ST131, an ExPEC clonal group involved in human urinary tract infections and septicemia. Relevant variation within this clone (e.g., antibiotic resistance and virulence) is frequently caused by the acquisition and loss of plasmids and other mobile genetic elements. Nevertheless, our knowledge of the ST131 plasmidome is limited to a few antibiotic resistance plasmids and to identification of replicons from known plasmid groups. PLACNET analysis extends the number of sequenced plasmids in ST131, which can be used for comparative genomics, from 11 to 50. The ST131 plasmidome is seemingly huge, encompassing roughly 50% of the main plasmid groups of γ–proteobacteria. MOBF12/IncF plasmids are apparently the most active players in the dissemination of relevant genetic information.
| Clinical microbiology is being transformed by whole genome sequencing (WGS) [1]. A case in point is Escherichia coli: there were 1,618 E. coli projects submitted to NCBI compared to just 68 complete genomes by year 2013. Within the realms of clinical and environmental microbiology, plasmid analysis is increasingly used to track the dissemination of genes encoding virulence, resistance to antibiotics, heavy metals and biocides [2]–[4] and, to a lesser extent, to analyze differences in the adaptive evolution of certain clonal backgrounds [5], [6]. Hybridization with specific probes [7], amplification of plasmid replication initiator proteins (RIP) [8]–[10], and relaxases (REL) [11] allow preliminary identification of plasmid families. In addition, plasmid MLST (pMLST) is used for epidemiological surveillance, but is restricted to individual plasmids of a few plasmid families of Enterobacteriaceae (http://pubmlst.org/plasmid/). This precludes the detection of plasmid mutations or rearrangements, as well as the identification of conjugative plasmids not represented in the pMLST database and of most mobilizable plasmids [11]. Finished plasmid/genome sequencing provides accurate and non-biased information, but is still expensive and thus seldom used specifically for plasmid analysis. Draft WGS dramatically cut down cost and analysis time. Although it allowed rapid and cheap data acquisition, WGS datasets typically result in more than a hundred contigs for a given genome, due to the short read lengths generally obtained. Genome fragmentation makes it difficult to distinguish between physical units, that is, between chromosome and plasmid sequences, as well as between different plasmids that usually coexist in bacterial cells. Several strategies can be followed to analyze WGS genome sequences, the workflow described by [12] being a typical example. There are also applications to identify plasmids in WGS sequences, such as PlasmidFinder (http://cge.cbs.dtu.dk/services/PlasmidFinder/), which identifies plasmids according to PCR-based replicon typing (PBRT) [8]–[10] and the subtyping scheme included in the pMLST web page (http://pubmlst.org). PlasmidFinder is limited by its inability to reconstruct the sequences of entire plasmids, underscoring the urgent need for improvement over existing tools.
E. coli ST131 is a successful high-risk clonal complex of pandemic distribution, able to cause extraintestinal infections in humans [13]-[18]. The increasing recovery of ST131 isolates from hospitalized and non-hospitalized individuals and, more recently, from companion and foodborne animals [17], [19]–[25], sewage and main rivers of large European cities [26], [27] highlights the rapid spread and local adaptation to different habitats of this lineage. ST131 is characterized by high metabolic potential [28] and a variable number of virulence factors, including adhesins, siderophores, toxins, polysaccharide coats (capsules and lipopolysaccharides), protectins and invasins [19], [29], [30], mostly acquired by recombination and by the interplay of mobile genetic elements (MGEs) [16]. Such traits, which are common among different lineages of the E. coli B2 phylogroup [31], [32], enable strains to colonize mucosal surfaces, invade tissues, foil defence mechanisms and yield injurious inflammatory responses in the host. E. coli populations identified as ST131 by the widely used ‘Achtman scheme’ of multilocus sequence typing (MLST) [33] (http://mlst.warwick.ac.uk/mlst/dbs/Ecoli), split in diverse clusters or subclones on the basis of genomic profile, serotype, content of virulence factors, antibiotic susceptibility pattern and the presence of certain fimH alleles [21], [29], [34]–[36]. The most prevalent ST131 clonal sublineage (H30) is characterized by the presence of a fimH30 allele, serotype O25:H4 and a specifically conserved gyrA/parC allele combination that confers fluoroquinolone resistance (FQ-R). Most human infections caused by ST131 are due to isolates of the H30 sublineage [13], [16], [37]–[39], many of them carrying the blaCTX-M-15 gene which is responsible for resistance to third generation of cephalosporins. Some authors suggested differences between CTX-M-15 and non-CTX-M-15 producers, referred to as H30-R and H30-Rx sublineages, respectively [13], [35], [37], [38]. Currently, diverse O25b:H4 ST131 variants (e.g. fimH22, fimH30) or O16:H5 (e.g. fimH41) seem also to be widely spread [13]-[16], [40]. Full genome sequencing of several ST131 E. coli genomes, most of them H30-Rx variants [16], [41]–[44], revealed further differences among strains, mainly chromosomal SNPs, indels and plasmid variations [16], [43], [44]. Heterogeneity of MGEs has been reported in other relevant E. coli clones, mainly Shiga-toxin producing E. coli (STEC) as O157:H7, O104:H4 or O26:H11 [5], [6], [45]–[47], often associated with ecological diversification of E. coli populations that can influence host-pathogen interactions [48], [49]. Recently, International and European organisations including European Food Safety Agency, EFSA; European Centre for Disease Control, ECDC; Food Drug Administration, FDA; Centre for Diseases Control, CDC) and national food safety authorities underscored the need to identify clonal variants with enhanced transmissibility or pathogenicity as well as to infer the evolutionary history of pathogens of interest in Public Health (http://www.efsa.europa.eu/en/events/event/140616.htm). Because relevant adaptive traits are plasmid located, there is an urgent need to consider MGEs in population genetic studies.
In this work we describe PLACNET, a method to reconstruct plasmids from WGS datasets, and its application to the comprehensive analysis of bacterial plasmidomes. As a specific example, we describe the ST131 plasmidome and discuss its possible impact in the diversification of this clinically important lineage. PLACNET allows the identification of plasmids currently circulating among E. coli and other enterobacterial species that may be underestimated, thus providing a useful tool to approach comprehensive plasmid population genetic studies.
We analyzed ten E. coli genomes, classified as ST131 according to the Achtman scheme (http://mlst.warwick.ac.uk/mlst/dbs/Ecoli), which branch in three main clusters identified as ST43, ST9 and ST506 (Fig. 1) according to the cgMLST Pasteur Institute scheme (http://www.pasteur.fr/recherche/genopole/PF8/mlst/EColi.html). The use of these two schemes is widely accepted in epidemiology [50] and increasingly used for E. coli typing. The ST43 branch contains isolates of the H30 lineage, which split in three subclusters (four strains of virotype C, two of virotype A, one of virotype B). The ST9 branch corresponds to isolates of the H22/H324 sublineage (virotype D). The most distal branch to the main cluster is represented by the commensal strain SE15, a member of sublineage H41 identified as ST506 [16]. It does not contain any marker used for the virotype subtyping method described by Blanco et al (afa, sat, ibeA, iroN) [36], [51], [52]. Thus, the sample analyzed in this work includes representatives of all ST131 branches described to date [13], [16]. The core genome of the 10 strains encompasses 3.6 Mb (Fig. 1 inset). As can be seen, the phylogenetic tree of ST131 genomes can be rooted at the commensal strain SE15. It should be noted, however, that SE15 is not necessarily the ancestor of the pathogenic lineages, as inferred by recent evidence [16]. The divergence of SE15 from the other ST131 strains is of about 3,000 SNP/Mb, a measure of the depth of the ST131 phylogenetic branch (<0.3% divergence in the core genome). There are only 650 SNPs among the genomes of cluster C lineage (i.e., <200 SNP/Mb), indicating their close phylogenetic relationship. There are <300 SNPs within a given virotype. The average distance between clades A and B is of about 4,600 SNPs (i.e., 1,300 SNP/Mb).
The PLACNET protocol was used as explained in Materials and Methods. We proceeded with plasmid reconstruction, as exemplified in Fig. 2 for the reconstruction of the E61BA genome (ST9/H324/virotype D). When we applied the rules for reference homology, scaffold links and plasmid protein tagging, the E61BA network shown as “original network” was produced. Obviously, this network was not neat enough to allow plasmid reconstruction. Expert pruning of the network consisted on several steps. First, contigs smaller than 200 bp were eliminated. Second, hubs were identified (see arrows in the original network of Fig. 2), duplicated and assigned to separate disjoint connected components. Scaffold links and coverage information, as well as score values of conflict edges, were used to decide on valid component assignment. Inspection of the coding potential of hubs usually showed them to correspond to ISs, transposons or other known repeated elements (as shown in S9 and S10 Figs.). As a result, a pruned network was reconstructed as shown in Fig. 2. Differential coloring of disjoint connected components in the pruned network thus displayed the final network of plasmids (as contig constellations). In PLACNET Cytoscape representation, most plasmids can be identified by their RIP and/or REL proteins. Thus, the reconstructed E61BA genome contains seven plasmids: a 134 kb MOBF12/IncF plasmid (pE61BA-1), a 37.7 kb MOBP6/IncI2 plasmid (pE61BA-7), a 24.5 kb MOBC12 plasmid (pE61BA-2), a 18 kb MOBP11/IncP1 plasmid (pE61BA-4), two MOBP5/ColE1-like plasmids of 6.6 and 6.9 kb (pE61BA-5 and pE61BA-6, respectively) and one MOBQ12 5.0 kb plasmid (pE61BA-3). Only plasmid pE61BA-2 could be closed, the remaining contained at least two contigs. Thus, their reported sizes are minimum sizes, since they might include small repeated sequences that were taken out of the analysis during network pruning. Two contigs remained as “not assigned” to any physical unit in this particular genome because they did not show any reference or scaffold link that bind them to other contigs: a 2,953 bp contig (containing a putative DNA primase and a lytic transglycosylase) and a 1,301 bp contig (containing two conjugation-related genes: trbI and a partial traB gene).
The same procedure was applied to the three other strains sequenced for this work as well as to the four genomes obtained from public DBs as Illumina reads. The plasmid content of the four strains sequenced in this work was confirmed by the analysis of S1-digested genomic DNA profiles by PFGE. This analysis fully confirmed the presence of plasmids of similar size to those identified by PLACNET (S2 Text and S4 Table), In the case of strain E35BA, in which PLACNET identified two IncF plasmids that could not be separated (totaling 211 kb), S1-PFGE identified two plasmids of 140 kb and 75 kb. As a result of PLACNET analysis, we obtained the plasmid constellation networks shown in S1 to S8 Figs. A summary of the results, i.e., the reconstructed plasmids, is shown in Table 1, which includes also the plasmids of the ST131 reference strains JJ1886 and SE15. As can be seen, the number of plasmids in the ST131 genomes is variable, even from strains belonging to the same ST131 sublineage, ranging from just one plasmid in HVH177 (clade B/ST9/fimH22) or SE15 (clade A/ST506/fimH41) to seven plasmids in E61BA (clade B/ST9/fimH22), to give an average of 4 plasmids per genome. There is not a single plasmid group that appears specific of a particular sublineage. S1 Table contains the complete list of contigs assigned to each plasmid or chromosome.
There are two aspects of this work that will focus the discussion. On one side, the applicability, usefulness and limitations of PLACNET will be discussed. On the other, the plasmidome of E. coli ST131 genomes that were reconstructed by PLACNET will be analyzed as an example of the applicability of the method. Analysis of the individual reconstructed plasmids, meant for plasmid specialists, is expanded in S1 Text.
Most bacterial genomes contain more than one physical unit of DNA. Besides the main chromosome, some bacteria contain additional chromosomes and most contain plasmids. We propose that PLACNET can be used as a new method to analyze bacterial genomes. It allows the assignment of chromosomes and plasmids as separate physical units within a genome. Visual representation of the network in Cytoscape, in which plasmids appear as constellations in a starry sky, allows user-friendly apprehension of that genome constitution. We applied PLACNET in this work to analyze the plasmidome of E. coli ST131 genomes, but it has been shown to work also for a series of prototypic bacterial genera with different GC content and genome architecture, such as Salmonella, Klebsiella, Agrobacterium, Staphylococcus or Bacillus. As an example, the PLACNET representation of the genome of Staphylococcus aureus strain 118 (ST772) (GenBank acc number AJGE00000000) is shown in S23 Fig. PLACNET scope of application also includes multi-chromosome bacteria like Vibrio or Brucella, where it correctly predicts both chromosomes present in these species. One Vibrio cholerae Pacini 1854 genome (Bioproject ID: PRJEB2215) is shown in S24 Fig. as an example. Once contigs belonging to each plasmid are defined, classical plasmid analysis ensues, as explained in the Results section. Contigs selected as part of a single plasmid are taken together and its overall proteome used to build a clustering dendrogram with reference plasmids present in the network. The dendrogram tree gathers plasmids according to the number of homologous proteins they share, providing an indication on prototype plasmids closely related with the query plasmid. There are two issues in PLACNET analysis that require additional work and for which additional improvement can be expected:
HGT plays a critical role in shaping bacterial lineages, especially those of multi-environment opportunistic pathogens. Comprehensive characterization of plasmidomes has been impeded by methodological limitations, although they are essential for multilevel population genetics analysis, an approach necessary to explain selection and diversification of bacterial populations and to understand the reservoir dynamics of antibiotic resistance and virulence genes [69]. The application of PLACNET to ST131 genomes allowed the detection of emerging plasmid variants, important for the evolutionary history of this ExPEC lineage, which constitutes an outstanding example of a “high risk clonal complex”, a concept increasingly important in Public Health [69].
The evolutionary processes of main bacterial pathogens are often discussed in the context of lineage-associated acquisition of a specific virulence gene set. The present study demonstrates how E. coli ST131 strains, even when they are practically identical in their core genomes, contain a striking variety of different plasmids. Many of them remain unnoticed, since they are apparently cryptic. Prevalent plasmids, such as IncFs, undergo frequent recombination, continuously resulting in novel gene repertoires. Our results shed light on the role of plasmids in E. coli ST131 evolution. Horizontal transmission of plasmids that carry not only antibiotic resistance and virulence genes, but also other poorly analyzed functions (metabolic genes, colicins and as yet cryptic functions) is common in the ST131 plasmidome and results in frequent and rapid adaptive changes. Arrival to these conclusions has been made possible by the application of PLACNET, a plasmid reconstruction method for WGS datasets.
Comprehensive plasmidome analysis was carried out for 10 E. coli ST131 genomes, representing main ST131 sublineages described to date [13], [16]. They include strains coming from Spain (three fimH30, one fimH324), USA (three fimH30), Australia (one fimH30), Denmark (one fimH22) and Japan (one fimH41). The fimH30 strains from Spain were CTX-M-15 producers and belonged to the H30-Rx sublineage (additionally, one strain was also CTX-M-14), while those collected in the USA were KPC-2 producers. The four strains from Spain represent predominant ST131 variants on the basis of PFGE patterns and the presence/absence of four putative virulence markers (afaFM955459, encoding an Afa/Dr adhesion; sat, secreted autotransporter toxin; ibeA, invasion of brain endothelium; and iroN, salmochelin siderophore receptor) [36], [51], [52] and sequenced for this work.
The ST131 isolates studied represent epidemic variants exhibiting particular combination of putative virulence traits and were previously designed as distinct “virotypes” by capital letters A to D [36]. It should be noted that no correlation exists between these “virotype” designations and “ST131 strain designation” in other studies that also used capital letters to distinguish among ST131 clonal variants [16]. Other genome datasets were taken either from Bioproject NCBI database (https://www.ncbi.nlm.nih.gov/bioproject/), or from NCBI genomes database (E. coli JJ1886 [90] and E. coli SE15 [91]). In addition, fully sequenced plasmids pEK499, pEK204, pEK516, pJIE186-2 and pJIE143, previously found in other ST131 isolates [56], [92]–[94], were used for plasmid comparisons. Information about all genomes is detailed in Table 3.
Total DNA from E. coli ST131 strains FV9873, E35BA, E2022 and E61BA was extracted with QIAmp DNA Mini Kit (Qiagen). DNA concentration was measured with Nanodrop 2000 (Thermo Scientific) and Qubit 2.0 Fluorometer (Life Technologies). 1.0 µg DNA was sonicated (20 cycles of 30 s at 4°C, low intensity) with Bioruptor Next Generation (Diagenode). Sample quality was checked in a Bioanalyzer 2100 (Agilent Technologies). DNA samples were preconditioned for sequencing by using the TruSeq DNA Sample Preparation Kit (Illumina) and quantified with Step One Plus Real-Time PCR System (Applied Biosystems). Flow-cells were prepared with TruSeq PE Cluster Kit v5-CS-GA (Illumina). Sequencing was carried out using a standard 2×71 base protocol (300-400 bp insert size) in a Genome Analyzer IIx (Illumina, San Diego, CA) at the sequencing facility of the University of Cantabria. The main statistics of the eight sequence datasets analyzed are shown in Table 4.
The ST131 core genome was defined as the collection of genes present in the ten ST131 genomes analyzed, with more than 90% similarity and 90% coverage. CD-HIT-EST [95] was used to cluster genes. A homemade Perl script was created to parse the cluster and define the core genome set. All core genes were concatenated and aligned with progressive Mauve [96]. A tabular list of SNPs was extracted from the Mauve alignment by applying the SNP export tool of Mauve GUI. The tabular list of polymorphic sites was parsed by a homemade script. A given position was counted as an SNP if it varied between two given sequences. The number of SNPs was added for each pair of strains to give the final SNP count. Polymorphic sites with gaps were removed from the SNP count matrix. The Mauve alignment was curated by trimAl [97]. RAxML [98] was used to build the core genome phylogenetic tree, using 100 replicates for bootstrap determination.
PLACNET was developed to associate contigs with specific physical DNA units in WGS experiments. Networks are powerful models that allow visualization and analysis of sequence information. PLACNET networks are composed of two types of nodes (contigs and reference genomes) and two types of edges (similarity to reference sequences and scaffold links). Commonly, network layout algorithms simulate repulsion forces between nodes and attraction forces by the edges that link two nodes. Thus, node distribution in the network will depend on the intensity of forces that define the edges. In such network model, a plasmid will be represented by a connected component (a set of linked nodes) or, in other words, a constellation of contigs. Different physical units (plasmids and chromosomes) should be represented by disjoint connected components (separate constellations). The workflow (Fig. 6) involves the following steps:
|
10.1371/journal.pgen.1003973 | An Lmx1b-miR135a2 Regulatory Circuit Modulates Wnt1/Wnt Signaling and Determines the Size of the Midbrain Dopaminergic Progenitor Pool | MicroRNAs regulate gene expression in diverse physiological scenarios. Their role in the control of morphogen related signaling pathways has been less studied, particularly in the context of embryonic Central Nervous System (CNS) development. Here, we uncover a role for microRNAs in limiting the spatiotemporal range of morphogen expression and function. Wnt1 is a key morphogen in the embryonic midbrain, and directs proliferation, survival, patterning and neurogenesis. We reveal an autoregulatory negative feedback loop between the transcription factor Lmx1b and a newly characterized microRNA, miR135a2, which modulates the extent of Wnt1/Wnt signaling and the size of the dopamine progenitor domain. Conditional gain of function studies reveal that Lmx1b promotes Wnt1/Wnt signaling, and thereby increases midbrain size and dopamine progenitor allocation. Conditional removal of Lmx1b has the opposite effect, in that expansion of the dopamine progenitor domain is severely compromised. Next, we provide evidence that microRNAs are involved in restricting dopamine progenitor allocation. Conditional loss of Dicer1 in embryonic stem cells (ESCs) results in expanded Lmx1a/b+ progenitors. In contrast, forced elevation of miR135a2 during an early window in vivo phenocopies the Lmx1b conditional knockout. When En1::Cre, but not Shh::Cre or Nes::Cre, is used for recombination, the expansion of Lmx1a/b+ progenitors is selectively reduced. Bioinformatics and luciferase assay data suggests that miR135a2 targets Lmx1b and many genes in the Wnt signaling pathway, including Ccnd1, Gsk3b, and Tcf7l2. Consistent with this, we demonstrate that this mutant displays reductions in the size of the Lmx1b/Wnt1 domain and range of canonical Wnt signaling. We posit that microRNA modulation of the Lmx1b/Wnt axis in the early midbrain/isthmus could determine midbrain size and allocation of dopamine progenitors. Since canonical Wnt activity has recently been recognized as a key ingredient for programming ESCs towards a dopaminergic fate in vitro, these studies could impact the rational design of such protocols.
| To achieve exquisitely complex behavior, the mammalian CNS is comprised of numerous neuron types, each with different functions. These distinct neuron types are produced from neural progenitors during embryonic development. How the embryonic neural progenitors are programmed to produce distinct neuron types, in the correct position and number, is a central question in developmental neuroscience. We focused on studying the embryonic production of a key neuron type, the midbrain dopamine neuron (mDA), which is particularly vulnerable in Parkinson's disease (PD). Previous works from our lab and others have shown that Wnt signaling is critical for dopamine neuron production. Here we provide a mechanism for how Wnt signaling is initiated, and then downregulated. Key to initiating this process is a transcription factor, Lmx1b, whereas important to the downregulation process is a newly characterized microRNA, miR135a2. The quantitative balance of these factors determines how many dopamine neurons are produced during embryonic development. These studies will have direct implications for efficiently programming dopamine neurons from stem cells, a key goal of regenerative approaches for PD.
| MicroRNAs regulate gene expression in various aspects of central nervous system (CNS) and peripheral nervous system (PNS) development and function, including neurogenesis, glial differentiation, fate specification, synaptogenesis, spine formation and plasticity [1]–[7]. Less studied is their role in modulating the most critical developmental signaling molecules in the embryonic CNS – morphogens. Recent studies have suggested that morphogen function is not simply based on a concentration gradient, but rather an integral of concentration as well as the time of exposure [8]–[10]. Thus, mechanisms must exist to control the dose and time of morphogen expression and function. MicroRNAs have been shown to target key elements of morphogen pathways in the early embryo [11]. We considered it plausible that microRNAs may be involved in modulating morphogen function in the developing CNS.
Wnts are key morphogens in the developing and adult CNS that are involved in proliferation, survival, patterning, and neurogenesis [12]–[14]. Wnt1 is the prototypical canonical Wnt and its function has been documented particularly in the midbrain region [14]. Wnt1 is dynamically expressed in the midbrain, being expressed in a broad swath at 8.5 days post coitum (dpc) and ultimately restricting to the Roof Plate (RP), Isthmic Organizer (IsO) and Floor Plate (FP) regions. Loss of Wnt1 leads to a drastic decrease in midbrain size, as well as reduction and misspecification of midbrain dopamine neurons (mDAs), and this is exacerbated by loss of Wnt5a [15], [16]. Studies that have interrupted Wnt/beta-catenin signaling have revealed that this pathway is critical for specification and neurogenesis of mDAs [17]–[19]. Wnt signaling is required for the expression of key mDA determinants Lmx1a, Otx2, and Ngn2 and for the downregulation of Shh [17]. Counterintuitively, excessive Wnt signaling is also detrimental for mDA production [20], adding to the general notion that morphogen dosage must be tightly regulated [21].
In the ventral midbrain, the Foxa2+, Shh+ FP is roughly allocated into two main progenitor domains: a medial Lmx1a/b+/Msx+ progenitor domain that gives rise to many mDAs (this domain may be further subdivided [22], [23]), and a lateral Nkx6.1+/Sim1+ domain that gives rise to many Brn3a+ neurons, populating among others, the red nucleus [22], [24]–[28]. During early development, however, Nkx6.1 is expressed at the midline and Lmx1b is expressed more broadly than Lmx1a/Msx. Between 9.0–9.5 dpc, Lmx1a/Msx expression is initiated, and expands laterally [27]. At the midline, Lmx/Msx ultimately subsumes Nkx6.1, resulting in a medial Lmx/Msx and lateral Nkx6.1/Sim1 domain demarcated by a sharp boundary. Lineage-based progenitor labeling studies have suggested that the expansion of the Lmx1a domain, in part occurs by inductive mechanisms [22], [28]. This induction is likely, at least in part, mediated by Wnt1/beta-catenin signaling, which is important in the ventral midbrain [17], [29]–[31], and both necessary and sufficient for Lmx1a expression in the FP [17], [18], [32]. Indeed, this capacity to elicit drastic gene expression changes makes the ventral midbrain a particularly good model to interrogate Wnt1/Wnt signaling and its modulators.
In this study we have identified an autoregulatory loop involving Lmx1b and miR135a2 that is critical for determining mDA allocation. We show that Lmx1b promotes mDA progenitor fate, whereas miR135a2 delimits the mDA domain. Forced maintenance of Lmx1b results in expanded mDA progenitors whereas loss of Lmx1b results in diminished mDA progenitors. MicroRNA studies show the opposite effects. Conditional removal of Dicer1 from ESCs results in expanded mDA progenitors at the expense of Nkx6.1 progenitors. In contrast, increased miR135a2 levels, only during an early window, result in a reduction in the proportion of mDA progenitors. In addition to progenitor allocation defects, we observed changes in midbrain size in these mutants. Both progenitor allocation and midbrain size phenotypes may be caused, at least in part, by alterations in Wnt1/Wnt signaling. While Lmx1b promotes Wnt1/Wnt signaling, miR135a2 appears to negatively regulate Lmx1b/Wnt1/Wnt signaling in the context of the embryonic midbrain.
In order to identify microRNAs (miRs) in the Wnt1–rich mDA progenitor domain, ventral midline and dorsal lateral tissues were microdissected from 11.5 dpc mouse embryos and used to perform a microRNA array, followed by qRT-PCR validation with select individual TaqMan assays (Applied Biosystems). We reasoned that functionally relevant microRNAs would be 1) robustly expressed 2) differentially expressed and 3) bioinformatically predicted to target genes in the Lmx/Wnt axis. From five candidates that fit these criteria, we focused on mmu-miR-135a (miR135a), which was robustly expressed and showed a greater than 3-fold increase in the ventral midline compared to dorsal lateral tissue. The closely related mmu-miR-135b (miR135b) was also increased in the ventral midbrain, but this increase was not statistically significant (Figure 1A). Ventral midbrain enrichment of miR135a was further confirmed by Locked Nucleic Acid (LNA, Exiqon) in situ hybridization, which is designed to specifically detect mature microRNAs (Figure 1B). In addition, miR135a was predicted to target Lmx1b and several genes of the Wnt pathway through evolutionarily conserved binding sites in the 3′UTR [33](Table 1), thereby warranting further study.
To extend the expression data, we designed a reporter transgene (“sensor”) to verify the functional activity of miR135a in the ventral midbrain [34]. This transgene, comprised of eGFP with several sequences complementary to miR135a in the 3′UTR, is designed to broadly express eGFP. In regions of high miR135a activity, however, the eGFP levels should be suppressed. Since we did not design bulges in the microRNA binding sites, this transgene will not serve as a microRNA “sponge”, but only as a “sensor”. A control transgene, comprised of tdTomato with no complementary miR135a sites in the 3′UTR, was designed to broadly express tdTomato regardless of microRNA activity (Figure 1C). The transgenes were co-injected and transient transgenic embryos were harvested at 11.5 or 12.5 dpc. In ventral midline progenitors, eGFP was markedly reduced in the region predicted to have high miR135a activity, equivalent to that detected by the 135a LNA probe. In contrast, tdTomato showed little to no reduction at the ventral midline compared to neighboring regions (Figure 1D–G).
In mice, there are two miR135a family members, mmu-miR135a-1 (miR135a1) and mmu-miR135a-2 (miR135a2). Below we provide evidence suggesting that miR135a2, is expressed in the midbrain. Although miR135a2 was predicted to be intergenic on the miRBase Sequence Database and UCSC genome browser, a separate screen [35] coupled with further bioinformatic analysis revealed that miR135a2 was likely located between two exons of a previously uncharacterized gene. Based on its proximity to the 3′ end of nearby non-coding RNA, Rhabdomyosarcoma 2 associated transcript (Rmst)(Ensembl browser), which is known to be expressed in the midbrain [36], we hypothesized that miR135a2 was embedded in this gene. Thus, we performed RT-PCR on 11.5 dpc ventral midbrain RNA using a forward primer in Rmst and a reverse primer in the downstream flanking exon of miR135a2. This experiment yielded two predominant bands of approximately 700 bp and 900 bp (Figure S1A), indicating possible splice variants that will be further characterized in a future study. The most prominent fragment was sequenced and a BLAST search revealed that a) a variant of the Rmst transcript exists, which excludes exon 13, and has at least three additional exons and b) miR135a2 is located in the final detected intron of this transcript (Figure 2 – top panel). Moreover, in situ hybridizations with two separate probes (probe A and probe B) designed against this region showed similar expression in the midbrain and hindbrain FP, RP from the hindbrain to the telencephalon, and the IsO (Figure 2H; Figure S1B), equivalent to that detected by the 135a LNA probe (Figure 1B). Together, these results suggest that miR135a2 is coexpressed with Rmst in the midbrain. Although a separate internal promoter for the microRNA remains an alternative possibility, it is likely processed from an intron, a finding common for more than 50% of microRNAs [37].
The presence of miR135a2/Rmst in the FP, RP, and IsO resembles previously described expression patterns of two important genes, Lmx1b and Wnt1 [38], [39]. Thus, we compared the spatio-temporal relationship of these genes throughout early midbrain development. We observed similar and widespread expression with all three probes, as well as the miR135a LNA probe at 8.0 dpc, likely in the prospective midbrain, based on pan-midbrain reporter expression in Wnt1::Cre fate maps (Figure S1C). By 9.5 dpc all three genes were no longer detected throughout the midbrain, but rather were restricted to the FP, RP, and IsO (Figure 2A–L). The expression of these three genes correlates until ∼11.5 dpc, but at this age miR135a2/Rmst is also observed, at lower levels, in cells appearing to exit the ventricular zone throughout the midbrain (Figure S1B). This expression likely contributes to the low levels of miR135a observed in the dorsolateral samples in Figure 1A and is likely independent of Lmx1b. Between 12.5 dpc and 14.5 dpc, the microRNA transcriptional unit is maintained in dopamine progenitors, whereas Lmx1b and Wnt1 are severely downregulated, consistent with the phenomenon of “temporal exclusion” that has been described for many microRNAs and their cognate targets (Figure 3) [40]. On the other hand, Lmx1a, which is closely related to, and has partially overlapping function with Lmx1b, is not an in silico predicted target of miR135a2 and remains easily detectable in dopamine progenitors (Figure 3D–F).
These dynamic and tightly correlated expression patterns suggested miR135a2/Rmst as a potential component of the Lmx1b/Wnt1 regulatory network. At this point, we designed experiments to elucidate the details of this network by testing a) the hierarchical relationship between Lmx1b, miR135a2/Rmst, and Wnt1 b) the role of Lmx1b in midbrain development, and c) whether microRNAs, and specifically miR135a2 levels, were important for early midbrain development.
Given the overlapping expression patterns of Lmx1b, miR135a2/Rmst and Wnt1 in the FP, RP and IsO, we first tested the hierarchical relationship between these genes. To do this, we utilized a mouse strain designed to conditionally express Lmx1b coding sequence, but lacking greater than 95% of the 3′UTR, under control of robust CAG regulatory elements. Using this strain, we generated embryos in which Lmx1b was conditionally activated throughout the midbrain and rhombomere 1 from ∼8.0 dpc onward, with Engrailed 1 (En1) Cre recombination [41], [42]. Forced maintenance of Lmx1b in En1Cre/+; Rosa26Lmx1b/+ (En1::Cre;Lmx1bOE) led to ectopic mir135a2/Rmst and Wnt1 expression in the midbrain, although both of these mRNAs were expressed more robustly in dorsal compared to ventral regions (Figure 4A–H). To determine whether changes in Wnt1 expression correlated with changes in Wnt signaling, we employed an Axin2::d2eGFP reporter allele that provides a transcriptional readout of this pathway [43], [44]. In En1::Cre;Lmx1bOE,Axin2::d2eGFP embryos d2eGFP fluorescence is detectable throughout the midbrain, whereas in controls it is predominantly confined to the ventral midline and the region surrounding the roof plate (Figure 4I–J and data not shown).
Conversely, in embryos wherein Lmx1b was conditionally deleted (En1::Cre;Lmx1bcKO) [45], miR135a2/Rmst and Wnt1 were significantly reduced in mDA progenitors (Figure 4K–N), undetectable in the isthmus (Figure 4Q–T), and mildly reduced in the dorsal midbrain (Figure S2A–D). Axin2::d2eGFP transgene expression was drastically reduced in the ventral midbrain (Figure 4O–P) and barely detectable in the isthmus (Figure 4U–V), but appeared only slightly diminished in the dorsal midbrain (Figure S2E–F). These data suggest that Lmx1b is upstream of both genes and promotes Wnt signaling, but that in regions where Lmx1a is expressed, it can at least partially compensate for Lmx1b. This finding is consistent with in vitro studies wherein both Lmx1a and Lmx1b were shown to drive Wnt1 expression [46].
Finally, to determine whether the observed changes in miR135a2/Rmst corresponded to the mature microRNA, we used qRT-PCR to quantify mature miR135a levels. In En1::Cre,Lmx1bOE midbrain, mature microRNA was increased 2-fold, whereas in En1::Cre,Lmx1bcKO midbrain it was modestly reduced (Figure S2G). These results suggest that miR135a2 is coexpressed with the Rmst transcript and responsive to Lmx1b manipulations.
To determine the role of Lmx1b in midbrain development, we first investigated the functional consequences of maintaining Lmx1b in the early embryonic midbrain. Coronal sections through En1::Cre;Lmx1bOE embryos revealed an overall increase in third ventricle size and morphogenetic abnormalities (Figure S3A–B, G–H). In the ventral midbrain of 9.5–11.5 dpc En1::Cre;Lmx1bOE embryos we observed the dorsal-ventral (DV) extent of the FP marker Foxa2 to be significantly expanded, although depressed in level. The DV extent of the transcription factor Lmx1a was also significantly expanded throughout most of the Foxa2 progenitor domain. Lmx1a levels at the midline were consistently reduced, and in lateral regions were at even lower and graded levels (Figure 5A–F, Figure S3C–D, and data not shown). This result is consistent with ectopic Wnt signaling in the midbrain, in which Lmx1a is initially induced but ultimately attenuated [18]. Alternatively, progenitors might compensate for the overexpression of Lmx1b by downregulating the closely related Lmx1a.
To determine whether these domain changes were merely a consequence of overall increase in midbrain size, we measured the DV extent of the Lmx1a progenitor domain, the Foxa2 progenitor domain, and the length of the third ventricle (3V). By normalizing the DV extent of Lmx1a to the DV extent of Foxa2, and the DV extent of Foxa2 to the length of the third ventricle we determined that these domains are specifically expanded, rather than a consequence of the general increase in midbrain size (Figure 5A–I; Foxa2/3V shows a 20% increase, n = 3, control mean = 0.29±0.006, mutant mean = 0.35±0.02; p-value = .009). Moreover, the entirety of the Nkx6.1 progenitor domain is expanded, but intermingled with Lmx1a progenitors in these mutants (Figure 5J–K). As reported in previous studies [27], [47], [48] Lmx1a/b expansion led to repression of Nkx6.1, but likely because Lmx1a levels in lateral regions were not as robust as at the midline, this repression was only partial (i.e. several progenitors coexpressed Lmx1a and Nkx6.1). As a result, ectopic Lmx1a+ neurons, which appear to have increased Lmx1a levels after exiting the ventricular zone, were seen emanating from lateral aspects of the Foxa2 domain in addition to Nkx6.1+ neurons (Figure 5J–K). Further, we observed an increase in the DV extent of Shh, but a slight reduction in levels (Figure 5L–M). This result is consistent with an increase in Wnt signaling (see Figure 4) [17].
The early expansion of the Lmx1a/Foxa2 domain ultimately led to severe disruptions in ventral midbrain neuron types at later stages. Quantification of 13.5 dpc sections showed a 194% increase in Lmx1a+ cells and an 80% increase in TH+ mature mDAs. Brn3a+ red nucleus neurons were increased by 60%, likely because the overall Nkx6.1 progenitor domain was also expanded despite partial repression by Lmx1a. In contrast, a dramatic loss (90%) of Islet+ oculomotor neurons was observed (Figure 5N–U).
At 14.5 dpc, in addition to ectopic TH+ neurons in lateral regions, we observed a decrease in TH+ neurons at the midline, particularly in rostral sections (Figure S3I–R). Many cells at the midline appear to be stalled at the Nurr1+TH− state. This is likely because excess Lmx1b leads to increased Wnt1/Wnt signaling, too much of which is detrimental for normal dopamine neuron differentiation [20]. Alternatively, since a small Nurr1+/TH− population does exist in the wildtype postnatal midbrain, the increase in this population could indicate a change in fate. These data collectively demonstrate that failure to restrict Lmx1b during early embryogenesis drastically increases the third ventricle size and alters patterning of the midbrain. Further, excessive Lmx1b within the dopamine progenitor domain (see Figure S3E–F) is detrimental for normal dopamine differentiation, suggesting a need for careful modulation of its expression level.
Next, to determine whether Lmx1b was required for normal midbrain development, we examined En1::Cre;Lmx1bcKO embryos. Such embryos generated a complementary phenotype to the En1::Cre;Lmx1bOE in that the length of the third ventricle and midbrain size were reduced (Figure S4K–L). This reduction is at least in part due to apoptosis, as activated Caspase-3+ cells were increased, particularly in lateral midbrain regions (Figure S4A–B). Additionally, in En1::Cre;Lmx1bcKO mutants some Otx2+ cells were detected across the isthmic boundary in the hindbrain (Figure S4M–N). Fgf8 was drastically reduced in the isthmic region (Figure S4O–P).
Upon examination of the ventral midbrain of En1::Cre;Lmx1bcKO embryos, we observed that the DV extent of the Shh+ FP domain was reduced (Figure 6A–B). However, Shh appeared to be maintained rather than downregulated at the midline in mutant embryos, in accordance with the fact that Wnt1/Wnt signaling is reduced (see Figure 4N and P) [17]. Further, the DV extent of the Foxa2+ FP was reduced (Figure 6C–E, Figure S4C–D). After normalizing the DV extent of Foxa2 to the 3V, this reduction was found to be proportionate to the reduction in midbrain size (Foxa2/3V shows a 12% decrease, n = 3, control mean = 0.31±0.01, mutant mean = 0.28±0.01; p-value = .066). Within this domain, the Lmx1a+ dopamine progenitor domain was selectively reduced (Figure 6F–K, Figure S4E–H), similar to mice with defects in Wnt1/Wnt signaling [15], [17]. Likely, however, due to some residual Wnt1/Wnt signaling in En1::Cre;Lmx1bcKO, the morphological changes seen in the Wnt1 knockout and Shh::Cre;Ctnnb1cKO are not observed, and the distribution of remnant Lmx1a+ cells is different. Moreover, outside the main mDA progenitor domain, some Lmx1a+ cells were stranded within Nkx6.1+ territory, likely reflecting a failed attempt to establish a broader dopamine progenitor domain in this mutant. The Nkx6.1 territory was also reduced albeit not as drastically (22% reduction, n = 3, p-value = 0.06)(Figure 6L–M).
The early reduction of the Lmx1a/Foxa2 domain ultimately led to the diminution of many ventral midbrain neuron types at later stages. TH, a mature mDA marker, was drastically reduced, and quantification of 13.5 dpc sections revealed a 68% reduction of Lmx1a+ nascent mDAs. Brn3a+ red nucleus neurons showed a milder 39% reduction, likely reflective of a slight decrease in Nkx6.1+ progenitors. Only a few Islet+ oculomotor neurons were detected at 9.5 dpc, and these were virtually undetectable by 13.5 dpc (Figure 6N–T, Figure S4I–J). Altogether these data suggest that in the early embryo, Lmx1b is a key determinant of midbrain size, isthmic integrity, FP size, mDA progenitor domain size, and ventral neuron numbers.
Since the proper dosage of transcription factors in the FP is imperative for determining progenitor allocation between the Lmx and Nkx6.1 domains, we next tested whether microRNAs played a role in regulating this process. To do this, we used an ESC line that harbored a CAG::CreERT2 construct and Dicer1floxed/floxed alleles, such that Dicer1, the key microRNA processing enzyme, could be deleted upon 4-hydroxy tamoxifen (4OHT) administration [5]. We next developed an optimized protocol to derive dopamine progenitors from embryoid body aggregates (Figure 7A) [49]. Remarkably, in controls, we were able to achieve conditions that mimic the in vivo midbrain FP, wherein most progenitors were Foxa2+, and of these roughly equal numbers were Lmx1a/b+ and Nkx6.1+. We quantified the proportion of Foxa2+ cells that were Lmx1a+, Lmx1b+ and Nkx6.1+ in controls and mutants. In both controls and mutants, large numbers of Foxa2+ cells were observed. In controls, the proportion of cells that were Lmx1a+, Lmx1b+, or Nkx6.1+ was roughly equivalent. In 4OHT treated cultures, however, the proportion of Lmx1a/b+ cells was drastically increased, while the proportion of Nkx6.1+ cells was decreased (Figure 7B–C). These data suggest that microRNAs are involved in progenitor cell allocation between the Lmx1a/b+ and Nkx6.1+ domain in ESC derived cultures, and under these conditions, de-repression of target genes through the loss of microRNAs expands mDA progenitors.
We sought to determine if miR135a2 is sufficient to repress Lmx1b, and can thereby modulate midbrain development. To test this hypothesis, we first performed a luciferase assay in HEK293 cells, and found that miR135a2 was able to repress a construct harboring a fragment of the Lmx1b 3′UTR, but not constructs harboring mutations within the evolutionarily conserved miR135a2 binding site of the Lmx1b 3′UTR (Figure 8A–B and data not shown).
Next, we generated transgenic mice that conditionally express a mmu-miR-135a-2 precursor under control of CAG elements (CAG-loxP-STOPr-loxP-miR135a2-IRESeGFP; see Materials and Methods for description); in dissected 8.5 dpc midbrain, miR135a was detected in controls, and was approximately 3 fold increased in En1::Cre;135a2OE mutants (Figure S5D), although eGFP activity was not detectable (see Materials and Methods). In situ hybridizations showed a reduction in the domain size and levels of Lmx1b (Figure 8C–F), which was confirmed by qRT-PCR (Figure 8G) and indicated that miR135a2 is sufficient to repress Lmx1b in vivo. Otx2, which harbors a binding site for the closely related miR135b, is also decreased by both in situ hybridization and q-RT-PCR (Figure S6I–K); however, it remains to be determined whether Otx2 levels are decreased due to direct microRNA mediated repression, or in response to reduced Wnt1/beta-catenin signaling, or both (Figure S7).
We reasoned that overexpressing miR135a2 should, at least in part, phenocopy Lmx1b-deficient embryos. Indeed, in En1::Cre;135a2OE embryos obtained from three separate transgenic lines (Figure S5A–C, and data not shown), we observed an overall reduction in midbrain size (Figure S6A–B and J–K), albeit not as drastic as En1::Cre;Lmx1bcKOs. In 9.5 dpc En1::Cre;135a2OE embryos, we detected an increase in the apoptotic marker activated Caspase-3 (Figure S6A–B). Large numbers of apoptotic cells were observed in lateral regions of the midbrain, whereas few were detected in the ventral midbrain. In 10.5 dpc and 11.5 dpc embryos, apoptosis is markedly reduced relative to 9.5 dpc embryos (data not shown). Increased apoptosis, particularly at 9.5 dpc, could at least in part underlie the size reduction of the midbrain.
We next examined the progenitor domains in the ventral midbrain of En1::Cre;135a2OE. The DV extent of the Shh domain was decreased, and in many mutants analyzed (n = 6/10), Shh was not as robustly downregulated at the midline as in controls (Figure 8H–I). In one particularly severe mutant, Shh was maintained at the midline in a manner identical to En1::Cre;Lmx1bcKO embryos (Figure S6P–Q). The DV extent of the Foxa2 domain was also decreased (Figure 8J–L and S6C–D). After normalizing the DV extent of Foxa2 to the 3V, this reduction was found to be proportionate to the reduction in midbrain size (Foxa2/3V revealed an 8% decrease, n = 4, control mean = 0.28±0.02, mutant mean = 0.24±0.01; p-value = .28). Further analysis of the progenitor populations within the Foxa2 domain showed disproportionate changes. We observed a dramatic reduction in the dimensions of the Lmx1a/b+ domain, both in width (i.e. dorsal – ventral dimension) and thickness (i.e. ventricular – pial dimension). This reduction in the Lmx1a domain was not solely due to a reduction in FP size. When measured relative to Foxa2, the Lmx1a domain still appeared selectively reduced, as the width of the neighboring Nkx6.1+ domain was largely unaffected (Figure 8M–T and S6E–H).
Since Lmx1b is upstream of Wnt1/Wnt signaling, and Wnt1/Wnt signaling is important in mDA progenitor specification, we examined alterations in Wnt1/Wnt signaling in En1::Cre;135a2OE embryos. From 9.5–11.5 dpc, the levels of Wnt1 and the size of the Wnt1 expression domain, which is included within the Lmx1b domain, appeared to be reduced (Figure 9A–D). Although the reduction was not as drastic as in the En1::Cre;Lmx1bcKO, and is unlikely to singularly explain this phenotype, these results suggested that the amount of available Wnt1 ligand in the ventral midbrain was restricted.
We next examined ongoing Wnt signaling in the ventral midbrain using the Axin2::d2eGFP allele described above, which provides a transcriptional readout of robust canonical Wnt signaling [43], [44]. In control 10.5 and 11.5 dpc embryos harboring an Axin2::d2eGFP transgene, d2eGFP is observed in the ventral midbrain in two stripes adjacent to the midline. At 10.5 dpc, GFP labeling is predominantly observed outside of the Lmx1a domain, although some is detected within the domain. At 11.5 dpc, GFP labeling still exceeds the Lmx1a domain by several cell diameters along the DV axis. In comparison to controls, En1::Cre;135a2OE,Axin2::d2eGFP mutants showed diminished GFP labeling intensity within the ventral midbrain, which was nearly extinguished in particularly severe embryos (Figure 9E–H). Further, consistent with the reduction in Wnt1 ligand, the total range of GFP labeled cells in En1::Cre;135a2OE,Axin2::d2eGFP embryos was reduced along the DV axis (Figure 9I). Little to no reduction of GFP was observed in the dorsal midbrain (data not shown). Although not as severe as the En1::Cre;Lmx1bcKO, reduced canonical Wnt signaling, which is critical for early activation and maintenance of Lmx1a [17], [46], could in part, explain the diminished mDA progenitor domain. At 11.5 dpc, reduced Wnt signaling is likely to be important for aspects of neurogenesis rather than the establishment of the mDA progenitor domain.
As Lmx1b and Wnt1 are also present in the isthmus, we examined this region in En1::Cre;135a2OE embryos. Analysis of the midbrain/hindbrain junction revealed differences between the En1::Cre;Lmx1bcKO and En1::Cre;135a2OE mutants. While Lmx1b, Wnt1, and Fgf8 were modestly, but consistently, narrower in En1::Cre;135a2OE mutants, the isthmic boundary, as determined by Otx2, was unchanged (Figure S8A–H). In contrast, in the isthmus of En1::Cre;Lmx1bcKO embryos, both Wnt1 (Figure 4T) and Fgf8 (Figure S4P) were drastically reduced, and Otx2+ cells were found in the rostral hindbrain (Figure S4N).
We next examined later stage embryos to determine the consequences of early reduction of the mDA progenitor population. In 13.5 dpc En1::Cre;135a2OE embryos we observed a drastic decrease of TH+ mDAs. Quantification revealed a 66% reduction in Lmx1a+ nascent mDAs, as well as a 48% reduction in Islet+ oculomotor neurons. In contrast, Brn3a+ neuron numbers in the red nucleus [24], were not altered (Figure 10A–I), likely reflective of the largely unaffected Nkx6.1+ progenitor domain. Moreover, several Brn3a+ cells were observed at the midline in the dopaminergic field (Figure 10F–G and Figure S9A–D), which is consistent with a reduction of the Lmx1b+ mDA progenitor domain [50], [51]. Further analysis at 16.5 dpc confirmed a drastic reduction of TH+ dopamine neurons throughout the anterior-posterior axis (Figure 10J–O). At 19.5 dpc we observed that the remnant neurons express mature mDA markers including Nurr1 and DAT (Figure 10P–S); this finding is different from reported En1::Cre;Lmx1bcKO studies [52] likely because in En1::Cre;135a2OE mutants, Lmx1b is reduced but not abolished. Consistent with the reduction of dopamine neuron numbers, a concomitant decrease of mDA projections to the striatum was observed (Figure 10T–U).
Next, we reasoned that because deletion of Lmx1b with a slightly later and more specific Shh::Cre driver reveals no changes in the size of the mDA domain [51], if miR135a2 is indeed acting through the Lmx1b/Wnt axis, then using Shh::Cre to overexpress miR135a2 should have no effect on mDA domain size. To determine whether miR135a2, could elicit mDA progenitor domain changes when activated at later stages (i.e. after the early restriction to the RP, FP, and IsO) we used both Shh::Cre and Nestin (Nes)::Cre recombination. When miR135a2 overexpression was initiated at ∼8.5–9.5 dpc with Shh::Cre or ∼10.5 dpc with Nes::Cre, the DV extent of the Foxa2 and Lmx1a domains were unchanged (Figure 11).
To interpret these results, we carefully considered the spatiotemporal expression of the different Cre drivers. En1 expression is initiated at the 2 somite stage and encompasses the prospective midbrain and rhombomere 1 regions, inclusive of the IsO, by the 4–6 somite stage (8.0 dpc) [53], [54]. In contrast, Shh is initiated slightly later at the 8-somite stage (8.5 dpc) [55], specifically at the ventral midline. At this early stage, however, Shh expression is confined to a small group of medially located progenitors, which does not encompass the entire prospective mDA progenitor domain [22], [28]. Subsequently, between 8.75 and 9.5 dpc, Shh expression flares out laterally, and encompasses the entire mDA domain [22]. If one considers the recombination of the entire prospective mDA progenitor domain, there is a significant time difference between En1::Cre and Shh::Cre recombination. Nestin::Cre is active throughout this region, including the IsO, between ∼10.5–11.5 dpc [56]. Considering these recombination kinetics, our results indicate that increased miR135a2 levels are sufficient to restrict the mDA progenitor domain size only during an early critical window. Since Lmx1b is broadly expressed at the ventral midline during this time window [27], we posit that miR135a2 expression aids in the DV restriction of Lmx1b and the refinement of the mDA progenitor domain; within the mDA domain, the outer edges appear more sensitive to increased miR135a2 levels. An alternative interpretation of these results is that the spatial differences in recombination between En1::Cre and Shh::Cre, rather than solely the timing differences, lead to the differing effects on the mDA progenitor domain. Particularly, since En1::Cre recombines the entire midbrain-rhombomere 1 region, while Shh::Cre does not, it is possible in En1::Cre;135a2OE repression of the Lmx1b/Wnt1 axis results in a partial breakdown of the Wnt-Fgf loop [18], [57]. Thus, early defects in IsO activity cannot be ruled out, and could also contribute to the phenotype observed in En1::Cre;135a2OE.
By eight criteria: total midbrain size, apoptosis, Wnt1/Wnt signaling, DV extent of the FP (Foxa2/Shh domain size), mDA progenitor domain size, mDA numbers, oculomotor neuron numbers, and temporal sensitivity, the miR135a2OEs, at least partially resemble the Lmx1bcKOs (Figure S10). In several criteria tested, the En1::Cre;Lmx1bcKO is more affected than the En1::Cre;135a2OE, likely because Lmx1b levels are not completely abolished in the latter. Together with the in vitro luciferase assays, our data suggests that the En1::Cre;135a2OE phenotype is at least in part dependent on Lmx1b repression in the early embryo. Further, the En1::Cre;Lmx1bcKO and the En1::Cre;135a2OE phenotypes are similar because they may both, in part, be due to net deficits in Wnt signaling. In the En1::Cre;Lmx1bcKO, reductions in Wnt signaling are likely due to a massive deficit of Wnt1 ligand (Figure 4N). In the En1::Cre;135a2OE, reductions in Wnt signaling are likely due to modest reductions in several targets including multiple levels of the Wnt pathway (Figure 9, Figure S7, and Table 1), altogether resulting in a significant net Wnt signaling deficit.
In this study we have identified a novel transcription factor/microRNA negative feedback loop that critically impacts Wnt1/Wnt signaling and midbrain development. Feedback circuitry inherently possesses a powerful buffering capacity, such that fluctuations in gene expression can be stabilized and protein expression in dynamically changing environments can be fine-tuned [58]. In this case, widespread Lmx1b/Wnt1/Wnt signaling must be sharply restricted over a short period (between 8.0 and 9.5 dpc), as maintenance of Lmx1b/Wnt1/Wnt signaling leads to several unwanted consequences. Here, we provide evidence that Lmx1b is upstream of miR135a2/Rmst, and that miR135a2 represses Lmx1b/Wnt1/Wnt signaling (Figure 12A). Thus, in the early midbrain, these factors may determine mDA progenitor allocation and midbrain size.
Our data indicate that microRNAs play a critical role in determining the size of the mDA progenitor pool. We found that loss of the key microRNA processing enzyme, Dicer1, in embryoid body aggregates skews the proportion of Foxa2+ progenitors in favor of an Lmx1a/b+ mDA fate over an Nkx6.1+ fate. Conversely, we found that early miR135a2 overexpression in vivo led to a disproportionate reduction in the size of the Lmx1a/b+ mDA progenitor domain. We propose that miR135a2 might influence net Lmx stoichiometry in two ways: first, by directly repressing Lmx1b, and second, by lowering Wnt1/Wnt signaling and therefore Lmx1a levels. In the early embryo, these alterations most prominently affect the outer edges of the Lmx domain, which mainly express Lmx1b, until Lmx1a is induced in this region [27]. Additionally, it is possible that lowering Wnt signaling alters proliferation, and that can in part account for the reduced mDA progenitor domain size.
We postulate that because of the reduction in Lmx1b and Wnt1/Wnt signaling in En1::Cre;135a2OE mutants, the Lmx1a+ domain fails to expand, ultimately resulting in a region where both Lmx1a and Lmx1b are extinguished and instead occupied by Nkx6.1. Supporting the notion that the boundaries are most vulnerable to changes in Lmx levels, even complete removal of Lmx1b in the En1::Cre;Lmx1bcKO results in the selective failure to establish the outer edges of the mDA progenitor pool. Taken together, our data indicate that the precise balance of Lmx1b and miR135a2 at early time points in the embryonic midbrain determines the size of the dopamine progenitor domain (Figure 12B), and ultimately affects the organization and number of neurons found within different ventral midbrain populations.
Our data show that both increased or decreased Lmx1b expression in the early embryo results in reduced oculomotor neuron numbers. In the early midbrain, Lmx1b is transiently expressed in the oculomotor primordium [50]. Based on this, one possible result was that prolonged maintenance of Lmx1b would lead to increased oculomotor neurons in En1::Cre;Lmx1bOE mutants, but we found the opposite (Figure 5N–P). Thus, we conclude that Lmx1b downregulation is critical for production of normal oculomotor numbers. miR135a2 may play a role in the timely regulation of Lmx1b in this context.
Precise and dynamic expression of the Lmx1b-miR135a2 duo is also important for determining overall midbrain size. Failure to restrict Lmx1b leads to an enlargement of the third ventricle and morphogenetic abnormalities, while excess miR135a2-mediated repression or loss of Lmx1b, leads to a reduction in third ventricle and midbrain size. Since Lmx1b drives Wnt1 expression, the phenotypes in En1::Cre;Lmx1bOE, En1::Cre;Lmx1bcKO and En1::Cre;135a2OE mutants can at least in part be attributed to alterations in early Wnt signaling. In accordance with this notion, Wnt1/Wnt signaling has been shown to determine the overall size of the midbrain by influencing cell survival and proliferation, [38], [54], [59]–[62]. The potency of Wnt1/Wnt signaling in midbrain development suggests that a method to control its stoichiometry is critical; we propose that the dose and spatiotemporal expression of Wnt1/Wnt signaling is, at least in part, determined by Lmx1b and miR135a2.
miR135a2 is predicted to regulate a large set of target genes in addition to Lmx1b. Lmx1a, a related transcription factor, is not a predicted target by any algorithm we encountered. However, some of the genes in the predicted miR135a2 target set, including Wnt1, Wnt5a, several molecules in the canonical Wnt signaling pathway (Figure 9, Figure S7, and Table 1), and Msx1/2 (Anderegg, unpublished observations), may play a critical role in midbrain development and contribute to the observed phenotypes. Moreover, Otx2, a transcription factor also upstream and downstream of Wnt1/beta-catenin signaling [17], [29], [63], whose dosage determines the size of the mDA progenitor pool [63], is a predicted target of the closely related miR135b (Anderegg, unpublished observations). Since the seed is identical in these two microRNAs, it is possible that Otx2 dosage is also controlled, in part, by miR135a2, although our luciferase assays could not detect a significant interaction (see Figure S7B).
It also remains possible that miR135a2 acts on other signaling pathways important for midbrain development including TGF-beta/Bmp [64], [65], Shh [66], and Fgf [54], [67], either directly or indirectly through points of crosstalk. In fact, the mDA phenotype observed in En1::Cre;135a2OE mutants bears some resemblance to that of En1::Cre;Fgfr triple knockout mice [67]. However, parallel data obtained from miR135a2 overexpression in the forebrain resulting in phenotypes that overlap Wnt mutants, supports the notion that defects in Wnt signaling contribute to the observed midbrain phenotype (Caronia-Brown, unpublished observations). Additionally, in colorectal cancers, miR135a targets the Wnt pathway, although in that context, with net positive effect [68]. By potentially targeting multiple levels of the Wnt pathway (see Table 1), including upstream transcription factors (Lmx1b and Otx2), ligands, positive and negative modifiers, and downstream targets it is likely that miR135a2 confers precision to the stability of the overall early Wnt regulatory network responsible for midbrain size and dopamine progenitor specification. In this context, it is conceivable that the modest increase in miR135a2 levels in En1::Cre;135a2OE modulates multiple Wnt pathway targets resulting in a net negative Wnt signal; in contrast, in En1::Cre;Lmx1bcKO, a net negative Wnt signal is primarily obtained by severe downregulation of Wnt1.
Our model is based on several lines of evidence derived from the mutants presented here. Our data suggest that miR135a2 is sufficient to modulate canonical Wnt signaling and the mDA progenitor phenotype, but it does not address whether miR135a2 is necessary for regulation of these phenotypes. Future loss of function studies are important to address this question and are currently underway. Since single microRNA knockouts often have subtle or no phenotypes, a double knockout of miR135a2 and closely related miR135b will likely be required to unmask the functions of this microRNA family [69]. Given the importance of the Wnt pathway in disease and cancer, however, the gain of function experiments described here are an important advance, as they provide a novel modality for targeting this pathway.
The Lmx1b-miR135a2 pair appears to be an integral component in the molecular circuitry governing establishment of mDA progenitors. Here we have demonstrated that this pair determines early aspects of mDA progenitor domain allocation, likely via direct effects on the ventral midbrain but possibly through IsO activity as well. However, after the size and spatial boundaries of the dopamine progenitor domain are firmly established, it is likely that miR135a2 has a distinct, later role of tuning, and ultimately downregulating Lmx1b/Wnt1 expression within the mDA progenitor pool (see Figure 3). This later role could be to tune optimal levels of Wnt signaling in the mDA domain, as well as aspects of neurogenesis. Thus, understanding this circuit may shed light on dopamine neuron numbers, and may therefore be relevant for understanding dopamine related disorders and susceptibilities to these conditions.
Generating large numbers of bona fide mDAs is a key goal of stem cell based approaches towards Parkinson's disease treatment. Previously, this process was limited by the inability to produce authentic mDAs that survived grafting. Recently this problem was overcome, using Wnt1/Wnt agonists to program hESCs towards a bona fide mDA fate [70], [71]. In this context, understanding agents that modulate Wnt1/Wnt signaling is of critical importance. It is conceivable, that just as iPS cells can now be programmed using microRNAs [72], mDAs will be derived from stem cells with the aid of a rationally designed cocktail of microRNAs and anti-microRNAs.
All mouse work was done in accordance with Northwestern University's ACUC guidelines.
Tissue was microdissected from the ventral midline and dorsal lateral midbrain of unfixed, coronally sectioned 11.5 dpc Swiss Webster embryos. Total RNA, including small RNAs, was extracted using the mirVana kit (Ambion). The TaqMan Rodent MicroRNA A+B Cards Set v2.0 (4400239) was used to perform the array and validation of individual miRs was done with TaqMan PCR Assays for miR135a (ID 000460), miR135b (ID 002261), and normalized to snoRNA202 (ID 001232)(Applied Biosystems)(see below).
For miR135a analysis in various mutant lines, tissue was dissected from littermates and snap frozen on liquid nitrogen. The meso-diencephalic region was dissected from older embryos (9.75 dpc En1::Cre;Lmx1bOE, 9.75 dpc En1::Cre;Lmx1bcKO, and 11.5 dpc Nes::Cre;135a2OE), while the whole head was used for younger embryos (8.75 dpc En1::Cre;135a2OE). Total RNA, including small RNAs, was extracted using the mirVana kit (Ambion). The TaqMan PCR Assay for miR135a (ID 000460) was used and normalized to snoRNA202 (ID 001232)(Applied Biosystems).
For transcription factor analysis, ∼8.5 dpc En1::Cre;135a2OE littermates were used. A cut was made just caudal to the future rhombomere 1, and the tissue from the head was snap frozen on liquid nitrogen. Total RNA was extracted using the mirVana kit (Ambion) then the High Capacity cDNA Reverse Transcription Kit (Applied Biosystems) was used to generate cDNA. The TaqMan gene assay for Lmx1b (Mm00440209_m1) was used to analyze RNA from 8–14 somite embryos, and the TaqMan gene assay for Otx2 (Mm00446859_m1) was used to analyze RNA from 13–23 somite embryos. Gene expression was normalized to Gapdh (Mm99999915_g1)(Applied Biosystems).
For all qRT-PCR experiments, the amount of total RNA per reaction was the same for each sample. The delta Ct method was used to calculate the relative fold changes in gene expression. All statistical values were calculated using a two-sample equal variance, two tailed, Student's t-Test.
Ventral midbrain tissue was dissected from 11.5 dpc Swiss Webster embryos and flash frozen on liquid nitrogen. Total RNA, including small RNAs, was extracted using the mirVana kit (Ambion). A cDNA library was synthesized using SuperScript II Reverse Transcriptase (Invitrogen) and random hexamers. PCR was performed with the following primers:
exon 12 (F) GCTAGCTCGAGAGCCACGCTCTTTCCCAACAC
exon 15 (F) GCTAGCTCGAGATGAATGTTGTCGACACTGTCCCATC
exon 16 (R) TCAGGAAGCTTTCTTTCTCAAAGGTCCAGCTTAGATCC
For locked-nucleic acid (LNA) in situ, the GEISHA project microRNA detection protocol version 1.1 (http://geisha.arizona.edu) was used. For whole mount in situ, embryos were harvested and fixed in 4% DEPC PFA in PBS overnight, rinsed in PBS, put through a methanol series, and then stored in 100% methanol at −20°C until used. For section in situ, embryos were harvested and fixed in 4% DEPC PFA in PBS overnight and prepared for cryosectioning at 20 µm. In situ hybridization was performed with single-stranded digoxigenin labeled riboprobes directed against the following mRNAs: FGF8 (C. Tabin), Lmx1a (K. Millen), Lmx1b (C. Tabin), LNA probe 135a (Exiqon 39037-01), Otx2 (C. Tabin), miR135a2/Rmst probe A (Genbank AI853140; B. Harfe), miR135a2/Rmst probe B (Genbank BI734849; B. Harfe), Shh (A. McMahon), and Wnt1 (A. McMahon).
An optimized protocol for dopamine neuron differentiation from embryoid body aggregates was modified from motor neuron differentiation protocol [49] and as described below. An ESC line that harbored a CAG::CreERT2 construct and Dicer1floxed/floxed alleles [5] was transiently treated with the FGF inhibitors PD173074 (Sigma, 50 nM as final concentration) from days 2–4 of culture. Then, on day 4 of culture, Smoothened agonist (EMD 566660, 15 nM as final concentration) and mouse recombinant FGF8b protein (Cell Signaling, 100 ng/ML as final concentration) were added for 4 days. Concomitantly, either control solvent (Ethanol) or 4-hydroxy tamoxifen (4OHT) (Sigma, 250∼500 nM as final concentration) was added for 2 days. On day 9, embryoid bodies were fixed in 4% PFA, cryosectioned, and immunolabeled. Images were collected on a Zeiss LSM510 confocal microscope. MetaMorph Software (Molecular Devices) was used to quantify Lmx1a, Lmx1b, and Nkx6.1 levels in Foxa2+ cells of 15 embryoid bodies (5 each from 3 independent experiments, n = 3).
The stem-loop of mmu-miR-135a-2 (accession MI0000715) plus ∼100 bp 5′ and 3′ flanking sequence was PCR purified from genomic DNA and then cloned into the pCAG-RFP-int vector (Addgene plasmid 19822). A fragment of each target gene 3′UTR, which corresponds to the region predicted to contain a mmu-miR-135a-2 seed match, was PCR purified from genomic DNA and then cloned into the pmirGLO vector (Promega E1330). The predicted miR135a2 binding site in the Lmx1b 3′UTR was directly mutated by using the Phusion site-directed mutagenesis kit (Thermo Scientific) to create constructs with scrambled or deleted binding sites. The success of mutagenesis was confirmed by sequencing. The following primers, including restriction sites, were used for cloning:
135a2 (F) GCTAGGCGATCGCGTGTGCTTTGTGTCCCTTACATGTAGC
135a2 (R) TTAGGACGCGTGACACTCAAGGAACACCAAAGAGG
Lmx1b (F) GGATCTAGATCCATGCAGAGCTCCTACTTTG
Lmx1b (R) CAGTCTAGAAAGTGACTGTCCAAGAGCTCTGGGTC
Lmx1b-M1 (F): TCAGACTCTTCAGACCAATCAGCGGTGCCCTCCCCT
Lmx1b-M1 (R): GCGGGTGGTGGGCTGGGGG
Lmx1b-M2 (F): CGCTCAGACTCTTCAGACTAGGCTACGGTGCCCT
Lmx1b-M2 (R): GGTGGTGGGCTGGGGGGCC
Lmx1b-del (F): ACGGTGCCCTCCCCTCGGCCAGCC
Lmx1b-del (R): GTCTGAAGAGTCTGAGCGGGTGGTGGGCT
Ccnd1 (F): GCTAGGCTAGCGGTCTGCTTGACTTTCCCAACC
Ccnd1 (R): TCAGGCTCGAGTGACAGGACGATCGCCATCAG
Gsk3b (F): GCTAGGCTAGCAAGGATCATGGCAGGATCCCAG
Gsk3b (R): TCAGGCTCGAGTGTGGAGTGGGCAAAGGTGC
Otx2 (F): GCTAGGCTAGCAGGTTTTGTGAAGACCTGTAGAAGC
Otx2 (R): TCAGGCTCGAGTAGGACAATCAGTCGCACAATC
Tcf7l2 (F): GCTAGGCTAGCCTTGCTGTACCTGTATGTCCGTCC
Tcf7l2 (R): TCAGGCTCGAGCACAGGGCAGTTGACTAGGAGGT
Wnt1 (F): GCTAGCTCGAGTTCTGCACGAGTGTCTATGAGGTG
Wnt1 (R): TTAGGGTCGACAAGGGCGCCTATGAGAAGCTG
The plasmids were co-transfected into 24-well plates of HEK293 cells using Lipofectamine 2000 (Invitrogen #11668). After 24 hours, the cells were harvested for a luciferase assay (Promega E1910) and measured with the Clarity Luminescence Microplate Reader (Bio-Tek). Each Firefly reading was normalized to Renilla and replicated 8 times each in 2 separate experiments.
To generate the 135a2OE transgene, we isolated a 324 bp fragment of genomic DNA that contains the precursor sequence of miR135a2, using the 135a2 forward and reverse primers described above, and cloned it into different mammalian expression vectors. Ultimately, after testing four different vectors in vitro, CAG-loxP-STOPr-loxP-miR135a2-IRESeGFP was determined to have the least leakiness even though the STOP cassette was in the reverse orientation. The transgene was microinjected into B6SJL at Northwestern University's Transgenic and Targeted Mutagenesis Laboratory. Six founder lines were obtained, and three separate lines (En1::Cre;135a2OE #2, #3, and #5) were used for experiments. All three show a very similar phenotype, in terms of overall midbrain size as well as reduced mDA progenitors, thus ruling out the possibility of site-of-integration dependent phenotypes.
To assess the miR135a expression levels in different mutant scenarios we performed qRT-PCR (see above). When En1::Cre is used to recombine the transgene, we detect a ∼3 fold increase of miR135a in 8.75 dpc mutants compared to littermate controls. When Nes::Cre is used to recombine the transgene a few days later, we detect modest overexpression of miR135a in 11.5 dpc mutants (See Figure S5D), but the quantification in mDAs is complicated by high endogenous levels of miR135a in controls at this stage, particularly in cells exiting the ventricular zone throughout the midbrain (See Figure 4K and Figure S1B). Separate analysis indicates that endogenous miR135a levels increase ∼8 fold between 8.5 dpc control samples and 11.5 dpc control samples (data not shown). eGFP activity was undetectable from this transgene in all scenarios tested, as it is likely degraded during microRNA processing.
For the miR135a2 “sensor” experiment, we designed a transgene, CAG-eGFP-WPRE-5XmiR135a2rc, which contained five exact miR135a2 reverse complement repeats, each separated by a GGCCGGCC spacer. For a control, we designed a similar transgene, CAG-tdTomato-WPRE with no complementary miR135a2 sites in the 3′UTR. The transgenes were co-injected into BL6SJL embryos and transient transgenics were harvested at 11.5 dpc or 12.5 dpc.
For Lmx1b overexpression experiments, male mice harboring an allele in which Cre recombinase was knocked in to the En1 locus [41] were bred to mice harboring the Rosa26Lmx1b/+ allele [42]. For Lmx1b deletion experiments, mice harboring an allele of Lmx1b with loxP sites flanking the homeodomain in exons 4–6 were used [45]. En1::Cre;Lmx1bF/+ males were bred to Lmx1bF/F females to obtain mutant embryos. Both En1::Cre- and En1::Cre+;Lmx1bF/+ littermates were used as controls, as no phenotype was observed in heterozygotes. For miR135a2 overexpression experiments, we bred females harboring the 135a2OE allele (described above) with males harboring 1) the En1::Cre knock-in allele, 2) an allele in which a GFP-Cre fusion cassette was knocked in to the Shh locus [8] or 3) a Nes::Cre transgene, expressed under control of the rat Nestin promoter and enhancer [73]. To assess Wnt signaling, we used mice harboring an allele in which destabilized eGFP was placed under control of the Axin2 promoter, exon 1 and intron 1 (Axin2::d2eGFP) [43]. For all matings, the morning when a vaginal plug was detected was designated as 0.5 days post coitum (dpc). Mice were maintained and sacrificed according to the protocols approved by the Northwestern University Animal Care and Use Committee.
Embryos were harvested and fixed in 0.2%–4% PFA in PBS for various amounts of time depending upon embryonic ages, and sectioned at 20 µm. Tissue sections were post-fixed in 1%–4% PFA in PBS, rinsed in PBS, antigen retrieved depending on the antibody (Vector Labs H-3301), blocked in 5% donkey serum, 0.1% Triton X-100 in PBS, and incubated overnight at 4°C with primary antibodies diluted in blocking solution: rabbit Active Caspase-3 (Cell Technology; 1∶1000), rat DAT (Santa Cruz; 1∶50), goat Foxa2 (Santa Cruz; 1∶50), rabbit GFP (Molecular Probes; 1∶1500), mouse Islet-1 (Developmental Studies Hybridoma Bank; 1∶100), guinea pig Lmx1a (Y-C. Ma; 1∶20,000), rabbit Lmx1b (custom; 1∶5,000), mouse Nkx6.1 (Developmental Studies Hybridoma Bank; 1∶100), rabbit Nurr1 (Santa Cruz; 1∶500), rabbit Pitx3 (Zymed; 1∶500), and sheep TH (Pel Freeze; 1∶250). Sections were rinsed in PBS and incubated with appropriate Alexa 488, 555, 647 (Molecular Probes) or Cy3 and Cy5 (Jackson Immuno Research) secondary antibodies diluted 1∶250 in blocking solution, rinsed in PBS, covered with DAPI (1 mg/mL; Sigma) in PBS, rinsed in PBS, and coverslipped followed by epifluorescent (Leica) or confocal microscopy (Zeiss LSM 510 META laser scanning or Leica DM6000 CFS). Images were processed in Adobe Photoshop CS2.
To measure the intensity of eGFP and tdTomato fluorescence in 11.5 dpc transient transgenic embryos, images of unstained coronal sections were taken with a Leica DM6000 CFS microscope. After alignment in Adobe Photoshop, such that the sections were equally matched along the anterior-posterior axis, a representative rostral, midlevel and caudal section was selected from each embryo. Using the line scan measurement tool in the Leica Application Suite Advanced Fluorescence Software, two lines were drawn in the medial and lateral domain of each section and the mean fluorescence intensity was recorded. For measurements of the third ventricle and overall midbrain area, coronal sections from En1::Cre;Lmx1bOE, En1::Cre;Lmx1bcKO, or En1::Cre;135a2OE littermates were aligned in Adobe Photoshop, such that they were equally matched along the anterior-posterior axis. For four representative sections, the edges of the ventricular and pial surfaces were traced, and then both the perimeter and area were calculated using the “Analyze” function in NIH ImageJ. To calculate tissue area, the ventricular area was subtracted from the pial area. To determine FP and mDA progenitor domain size, coronal sections from En1::Cre;Lmx1bOE, En1::Cre;Lmx1bcKO, En1::Cre;135a2OE, Shh::Cre;135a2OE, or Nes::Cre;135a2OE littermates were immunostained for Foxa2/Lmx1a/Nkx6.1 and then aligned in Adobe Photoshop, such that they were equally matched along the anterior-posterior axis. For every third section, the distances along the ventricular surface from 1) the ventral midline to the dorsal edge of the Lmx1a domain, and 2) the ventral midline to the dorsal edge of the Foxa2 domain were calculated using the “segmented line measure” function in ImageJ. To account for changes in overall midbrain size, the dorsal-ventral extent of the Lmx1a domain was normalized to the extent of the Foxa2 domain by dividing the first measurement by the second. To quantify the amount of different ventral neuron types in En1::Cre;Lmx1bOE, En1::Cre;Lmx1bcKO, and En1::Cre;135a2OE littermates, coronal sections were immunostained with combinations of Brn3a/Islet/Lmx1a/TH and then aligned in Adobe Photoshop, such that they were equally matched along the anterior-posterior axis. One-half of a representative rostral, midlevel and caudal section were chosen for manual counting from each embryo, except where noted below. Islet was counted from every third section in En1::Cre;Lmx1bOE and En1::Cre;135a2OE, and Brn3a was counted from every sixth section in En1::Cre;Lmx1bOE. All statistical values were calculated using a two-sample equal variance, two tailed, Student's t-Test.
To determine miR135a targets, we used TargetScan, MicroCosm, miRanda, and miRDB algorithms. Wnt pathway genes were obtained from the following references [13], [74], [75].
|
10.1371/journal.pntd.0003928 | Development of Resistance to Pyrethroid in Culex pipiens pallens Population under Different Insecticide Selection Pressures | Current vector control programs are largely dependent on pyrethroids, which are the most commonly used and only insecticides recommended by the World Health Organization for insecticide-treated nets (ITNs). However, the rapid spread of pyrethroid resistance worldwide compromises the effectiveness of control programs and threatens public health. Since few new insecticide classes for vector control are anticipated, limiting the development of resistance is crucial for prolonging efficacy of pyrethroids. In this study, we exposed a field-collected population of Culex pipiens pallens to different insecticide selection intensities to dynamically monitor the development of resistance. Moreover, we detected kdr mutations and three detoxification enzyme activities in order to explore the evolutionary mechanism of pyrethroid resistance. Our results revealed that the level of pyrethroid resistance was proportional to the insecticide selection pressure. The kdr and metabolic resistance both contributed to pyrethroid resistance in the Cx. pipiens pallens populations, but they had different roles under different selection pressures. We have provided important evidence for better understanding of the development and mechanisms of pyrethroid resistance which may guide future insecticide use and vector management in order to avoid or delay resistance.
| Successful population control of mosquitoes is key to preventing transmission and epidemics of mosquito-borne diseases. This strategy relies heavily on insecticides, particularly pyrethroids. However, widespread pyrethroid resistance has hindered vector control implementation and sustainability. Generally, pyrethroid resistance in insects is mainly caused by target-site insensitivity and metabolic resistance. Although studies on the two resistance mechanisms have provided insights, how the target-site and metabolic resistance mechanisms jointly confer the resistance phenotype has remained unclear. Understanding the mechanism of resistance to insecticides is essential for vector control measures. Our study investigated the development of resistance to a pyrethroid in three mosquito lines by varying the intensity of selection between lines (intense selection, mild selection and no selection) and then tracking changes in kdr frequency and the activities of three families of metabolic detoxification enzymes over time. These findings may lead to the development of more appropriate use of insecticides and more accurate resistance monitoring systems in the field.
| Culex mosquitoes are important vectors responsible for transmission of lymphatic filariasis (LF) and several viral pathogens to millions of people worldwide, including St. Louis encephalitis, West Nile encephalitis, eastern equine encephalitis, Venezuelan equine encephalitis and Japanese encephalitis [1]. The World Health Organization estimated over 120 million cases of LF and about 40 million disfigured and incapacitated by the disease [2]. Globally, nearly 1.4 billion people in 73 countries worldwide are currently threatened by LF. Mosquito-borne diseases dramatically affect public health and pose a major burden in terms of economy and development worldwide. Mass drug administration in combination with alternative vector control methods have proven to be more effective and practical in avoiding the re-emergence and re-introduction of LF [3, 4]. Insect control is the primary intervention available for some of the most devastating mosquito-borne diseases, particularly those lacking vaccines such as malaria, dengue and LF [5]. Most vector control programs largely rely on the application of chemical insecticides by the use of outdoor spraying, insecticide-treated nets (ITNs) or indoor residual spraying (IRS) [6]. Because of the relatively low mammalian toxicity and rapid knockdown effect on insects, pyrethroids are the most commonly used insecticides and constitute the only recommended class of insecticides for ITNs. However, insecticide exposure is a potent selective force, presenting a risk of generating resistance that would threaten the efficacy of control programs. Hence, preventing or delaying the emergence and development of resistance to pyrethroids is very important for vector control efforts. Improving vector management involves a better understanding of resistance mechanisms.
Global surveys have indicated that resistance of mosquitoes to pyrethroids mainly occurs through increased detoxification, as well as target site insensitivity [7]. ‘Metabolic resistance’ usually results from enhanced detoxification enzyme activity in resistant organisms [8]. Detoxification enzymes typically linked to insecticide resistance mainly include three major gene families, cytochrome P450 monooxygenases (P450s or CYPs), carboxyl/choline esterases (CCEs) and glutathione-S-transferases (GSTs). Numerous studies have associated these detoxification enzymes with pyrethroid resistance in mosquitoes [9–11]. The primary target sites of pyrethroids, well known as knockdown resistance (kdr), encode voltage-gated sodium channels, and single or multiple substitutions in the sodium channel gene can reduce neuronal sensitivity to pyrethroids [12]. To date, More than 30 unique resistance-associated kdr mutations or combinations of mutations have been detected in several insect species [13]. Among them, the most common kdr mutations are the leucine to phenylalanine (Leu→Phe) substitution and the leucine to serine substitution (Leu→Ser) substitution at codon 1014 in the S6 hydrophobic segment of domain II in the sodium channel gene [14, 15]. These two common mutations have been shown to reduce the pyrethroid sensitivity of sodium channels in Xenopus oocytes, confirming their role in kdr [16, 17]. Many studies have revealed that all these mechanisms can occur simultaneously in resistant populations with cumulative phenotypic effects leading to resistance to a single or multiple insecticides [18, 19]. However, the relative contribution of the metabolic resistance and knockdown resistance in conferring the resistance phenotype has remained elusive.
High levels of resistance to pyrethroids in Culex mosquitoes has been widely reported [20, 21]. Here, we studied Cullex pipiens pallens because it is the most prevalent and important vector in China with high population density. We also chose deltamethrin, a representative pyrethroid insecticide, to explore the evolutionary mechanism of insecticide resistance, as well as the relative contributions of the target-site and metabolic resistance in the development of pyrethroid resistance. We detected kdr and activities of three types of detoxification enzymes (P450s, CCEs, and GSTs) in the same mosquito sample through large-scale population studies. Knowledge of the changing trends and patterns of insecticide resistance may impact future predictions and monitoring of pyrethroid resistance in mosquitoes and other arthropod pests and disease vectors.
A population of Cx. pipiens pallens was collected from natural habitats (Tangkou, Shandong Province of China) in September 2009. Mosquitoes were reared in standard insectary conditions (28°C, 14 h:10 h light/dark period, 75% relative humidity) with tap water (larvae) and net cages (adults). Adults were supplied with a 10% sucrose solution and blood fed on adult mice. Larvae were fed with fish food.
The population was selected with the pyrethroid insecticide deltamethrin (Jiangsu Yangnong Chemical Group Co., Jiangsu, China) in the laboratory. Selection was performed by exposing each generation of fourth-stage larvae for 24 h to a 50% lethal concentration (LC50) of deltamethrin. The LC50 was determined by a larval bioassay. Initially, the fourth-stage larvae were exposed to a wide range of test concentrations of deltamethrin. After determining the mortality of larvae in this wide range of concentrations, a narrower range (of 5 concentrations, yielding between 5% and 95% mortality in 24 h) was used to determine LC50 values. Five concentrations of deltamethrin and 3 replicates of 20 fourth-instar larvae per concentration were used. A control group was measured using 20 larvae in tap water without any insecticides. Numbers of dead and surviving larvae were recorded after 24 h. All surviving larvae were transferred to tap water, fed with larval food and allowed to emerge. Adults were fed to obtain eggs for the next generation.
After six generations, three strains under different insecticide selection pressures were established. The first strain, designated the ‘intense selection’ (IS) strain, was selected with the LC50 of deltamethrin for each generation causing 50–60% larval mortality. The second selected strain, termed the ‘mild selection’ (MS) strain, was exposed to the constant concentration of 0.05 ppm (LC50 for generation 6) in the subsequent deltamethrin selection. Deltamethrin exposure was withdrawn from the last strain, which was designated the ‘no-selection’ (NS) strain. All three strains were selected for 24 generations with three replicate groups, and each generation of every replicate group was initiated with more than 1000 adult females in order to limit bottleneck effects. The actual doses used for deltamethrin selection per generation were reported in the supporting information (detailed results are shown in S1 Table). A laboratory deltamethrin-susceptible strain of Cx. pipiens pallens (S-LAB) was used as a reference strain.
In every replicate group of each strain, more than 50 female adult mosquitoes, at post-emergence ages of 3–4 d, were collected for determinations of kdr alleles and metabolic enzyme activities. Generations 1, 6, 10, 14, 18, 22, 26 and 30 were analyzed. Two legs of each adult female were removed and preserved individually in 95% alcohol for subsequent DNA analysis. The remaining mosquito body was immediately used for metabolic enzyme activity determination. More than 3,000 female adult mosquitoes were used for analysis in this study.
Genomic DNA was extracted individually from two legs of each adult female by a fast tissue-to-PCR kit (Fermentas, K1091). The region containing kdr mutations within the para sodium channel gene was amplified by PCR with primers Cpp1 (5’-CCTGCCACGGTGGAACTTC-3’) and Cpp-2 (5’-GGACAAAAGCAAGGCTAAGAA-3’). The PCR primers were designed based on the cDNA sequence of Cx. quinquefasciatus para-sodium channel gene alpha subunit (Genbank accession number BN001092) [22]. PCR amplification was carried out in a volume of 20 μl, including 4 μl genomic DNA, 10 μl Tissue Green PCR Master Mix (Fermentas), 10 pmol primers Cpp1 and Cpp2 and 4 μl nuclease-free water. Amplification was performed with the following cycling conditions: initial denaturation at 95°C for 3 min, 35 cycles of 95°C for 30 s, 55°C for 30 s and elongation at 72°C for 40 s, followed by extension at 72°C for 3 min. PCR products were purified using the QIAquick PCR purification kit (Qiagen) and then sent for sequencing (BGI, Shanghai, China). In total, 2,427 female adult samples were successfully sequenced in the study. Genotype frequencies were calculated, and deviation from the Hardy–Weinberg equilibrium was analyzed by the web-based program GENEPOP [23].
Metabolic enzyme activity was measured in individual female mosquitoes by using the method described by Daibin et al. [24] with a slight modification. Every individual females was homogenized in a 1.5-ml tube with 300 μl of 0.25 M phosphate buffer (pH 7.2) and diluted by adding 300 μl of phosphate buffer. The tube was mixed and centrifuged, and the same supernatant was used to test the activity of GSTs, P450s and CCEs simultaneously. All assays were carried out in duplicate, and the protein content of the supernatant was measured by the Bradford method [25]. For the GSTs activity assay, a total of 90 μl of reduced glutathione solution (Sigma, G4251) and 90 μl of 1-chloro-2,4'-dinitrobenzene (cDNB) solution was added to 90 μl of mosquito supernatant. The absorbance was measured immediately using a microplate reader at 340 nm and then detected every 2 min for five times, using 0.25 M phosphate buffer as the negative control. For the P450s activity assay, a total of 10 μl of the 60mM 7-ethoxycoumarine (7-EC) solution was added to 100 μl of mosquito supernatant, and samples were incubated at 30°C for 5.5 h. The reaction was stopped by the addition of 150 μl of glycine buffer (pH 10.4, 1 mM), and sample absorbance was measured using a microplate reader at 450 nm with 0.25 M phosphate buffer as a negative control. The OD values were converted into concentrations by using standard regression based on a serial dilution of 7-hydroxycoumarin and its relevant OD values. The content of P450s was calculated for each mosquito. The OD values were converted into concentrations by using standard regression based on OD values of serial dilutions of 7-hydroxycoumarine. For the CCEs activity assay, the 0.1 mM β-nitrophenyl acetate (Sigma, N8130) solution was placed for 5 min in 30°C water bath first, and then a total of 220 μl of the β-nitrophenyl acetate solution was added to 50 μl of mosquito supernatant. The absorbance was measured using a microplate reader at 405 nm every min for five times with 0.25 M phosphate buffer as the negative control. All measurements were performed in duplicate. In total, we successfully detected 1,541 female adult samples for P450s, 1,368 samples for CCEs and 1,665 samples for GSTs.
The LC50 was calculated using Probit analysis [26] and Abbott’s correction for the mortality rate in the control group [27]. The resistance ratio (RR) is the ratio of the LC50 of the selected strain to the LC50 of the laboratory deltamethrin-susceptible strain S-LAB. Regression analysis (Curve estimation) was used to determine correlation coefficients between the deltamethrin susceptibility and generations under different selection. The DNA sequences of kdr were assembled and prealigned by BioEdit, aligned in ClustalW implemented in MEGA5 and the alignment was refined manually [28]. Then we used DnaSPv5 to estimate the number of haplotypes (h), the haplotype diversity (Hd) and nucleotide diversity (Pi). The kdr allele frequency in three selected strains was calculated, and statistically significant differences among strains were examined using the t-test. One-way analysis of variance (ANOVA) was used to examine whether metabolic enzyme activity varied in populations with different selection pressures. The generation was treated as a random factor. Metabolic enzyme activity data were square root transformed. The t-test was used for comparison of different generations when appropriate. Linear correlation analysis was used to study the correlation of the frequency of kdr and the metabolic enzyme activity with the degree of resistance.
Cx. quinquefasciatus para-sodium channel gene alpha subunit: BN001092.
Cx. pipiens pallens kdr haplotypes: GU198936- GU198938, GU339221.
A field population of Cx. pipiens pallens was collected, and three strains were established after being placed for 30 generations (~28 months) under different insecticide selection pressures. After selection for six generations, the RR increased from 1.69 at generation 1 (LC50 = 0.0206 ppm) to 4.11 at generation 6 (LC50 = 0.0501 ppm) (detailed results are shown in S2 Table). The resistance to insecticide was increased from generation to generation with exposure to insecticide, while it was reduced without exposure (Fig 1). Regression analysis showed that the level of resistance grew exponentially in IS and NS strain (Fig 2). The degree of insecticide resistance in the IS strain increased more rapidly than that in the MS strain. At generation 30, the RR of the IS strain had increased significantly to 79.61 (LC50 = 0.9713 ppm) (P < 0.05), and the RR of the MS strain was also increased significantly to 35.80 (LC50 = 0.4367 ppm) (P < 0.05). Notably, the level of insecticide resistance increased faster after generation 14 of the IS strain and generation 18 of the MS strain. By contrast, the level of insecticide resistance in the NS strain was reduced slowly without insecticide selection from generation 6, and significant differences were observed after generation 22. The RR of the NS strain was reduced significantly from 4.11 at generation 6 (LC50 = 0.0501 ppm) to 1.98 at generation 30 (LC50 = 0.0241 ppm) (P < 0.05), and the LC50 of generation 30 was still higher than that of the first generation (P < 0.05) (Fig 1).
A 480-bp fragment of the para-type sodium channel gene including codon 1014 was sequenced from 2,457 individual Cx. pipiens pallens of the three strains. The wild-type kdr codon sequence spanning position 1014 was TTA (L1014). The two most common types of kdr mutations detected were TTT (L1014F) and TCA (L1014S). A total of six genotypes were identified in all strains: L1014/L1014, L1014F/L1014F, L1014S/L1014S, L1014/L1014F, L1014/L1014S and L1014F/L1014S (Fig 3). The field population of Cx. pipiens pallens carried a high kdr mutation frequency (86.49%) of both the L1014F (72.64%) and L1014S (13.85%) mutations (Table 1). Under insecticide selection, the kdr mutation frequency increased significantly and reached 100.00% at generation 6. But the response of these two mutations to deltamethrin selection were different (Table 1 and Fig 4). Under insecticide selection, the frequency of L1014F was increased and faster in IS strain than MS strain. By contrast, the frequency of L1014S was decreased and slower in MS strain than IS strain. In the IS strain, the frequency of the L1014F mutation became fixed, and the strain became homozygous for kdr (genotype: L1014F/L1014F) at generation 14. In the MS strain, the rate of increase in the allele frequency for this mutation was slower than observed for the IS strain, and the MS strain became homozygous (genotype: L1014F/L1014F) at generation 26. Interestingly, the level of insecticide resistance was increased most significantly after the population became a homozygous population for the kdr resistance gene, suggesting the existence of other mechanism which could induce deltamethrin resistance. Without insecticide selection, the frequency of L1014S increased and L1014F decreased. In the NS strain, the frequency of L1014S increased significantly from 8.21% at generation 6 to 17.73% at generation 30, and the frequency of L1014F declined significantly but was still higher than that of the first generation (72.64%). These results suggest that the L1014F mutation is more closely associated with deltamethrin resistance in the Cx. pipiens pallens population than the L1014S mutation. The wild-type L1014 allele could not be detected from the sixth generation and became extinct with insecticide selection. Because our studied population carried a high kdr mutation frequency and was a closed population, the wild-type sequence was recovered infrequently. The three replicate groups had the same trends, and the detailed results are shown in S1 Fig.
The haplotype number (K) and haplotype diversity (H) are informative statistics for describing the distribution of haplotypes under an infinite-sites model [29]. During a selective sweep, the reduction in variation around a naturally selected locus will reduce the impact of reshuffling by recombination producing new haplotypes, and recombination is more likely to occur between two copies of the same haplotype [30]. In this study, the intron region downstream of L1014 in the kdr gene was cloned, sequenced then analyzed using DNAspv5 software. The results showed a significant reduction of K and H in the kdr gene under insecticide selection, and the nucleotide diversity (Pi) and haplotype diversity (Hd) decreased with the increase of resistance (Table 2). Pi and Hd were reduced to 0 at generation 26 in the IS strain and at generation 30 in the MS strain, but they were increased after withdrawal of deltamethrin.
The field population of Cx. pipiens pallens was selected with deltamethrin in the laboratory due to unknown use of insecticides at the sampling site. Metabolic enzyme activities were detected after deltamethrin selection for six generations. We analyzed in total 1545 samples for P450s, 1371 samples for CCEs and 1653 samples for GSTs (Table 3).
The results showed that the P450s activities of the population changed with different selection conditions (Table 3 and Fig 5). In the IS strain, significant among-population variation in different generations was found (one-way ANOVA, F6, 560 = 31.312, P < 0.0001), and the P450s activities increased significantly with the development of deltamethrin resistance. The P450s activities were increased by 2.21-fold at generation 30, and analysis by t-test showed significant variation after generation 14 when compared with activities in generation 6. The P450s activities increased more rapidly after generation 26. Likewise in the MS strain, significant among-population variation in different generations was observed (one-way ANOVA, F6, 561 = 8.841, P < 0.0001), and P450s activities also increased significantly with the development of deltamethrin resistance, although the increasing trend was slower than that in the IS strain. P450s activities increased by 1.61-fold at generation 30, and analysis by t-test showed significant variation in P450s activities after generation 18. Significant among-population variation was seen in different generations of the NS strain as well (one-way ANOVA, F6, 510 = 7.495, P < 0.0001). Without deltamethrin selection, P450s activities were reduced at each generation, decreasing by 0.81-fold at generation 30, and a significant reduction in those activities after generation 26 was found by t- test analysis.
Changes of CCEs activities were similar to those of P450s (Table 3 and Fig 5). In the IS strain, significant among-population variation in different generations was observed (one-way ANOVA, F6, 491 = 6.957, P < 0.0001), and the CCEs activities also increased significantly with the development of deltamethrin resistance. At generation 30, the activities increased by 1.55-fold, and activities of CCEs increased significantly as shown by t- test analysis after generation 18 when compared with those in generation 6. In the MS strain, no significant among-population variation in different generations was detected (one-way ANOVA, F6, 503 = 0.710, P = 0. 642 > 0.0001). The CCEs activities were not significantly greater at generation 30 (independent samples test, t = 1.178, P = 0.241 > 0.05). In the NS strain, significant among-population variation in different generations was found (one-way ANOVA, F6, 450 = 4.151, P < 0.0001). Without deltamethrin selection, CCEs activities were also reduced with each generation, decreasing by 0.59-fold at generation 30, and t- test analysis showed significant variations after generation 18.
Activities of P450s and CCEs both increased with the level of deltamethrin resistance, and they decreased after withdrawal of deltamethrin. However, changes of GSTs activities were not associated with deltamethrin resistance (Table 3 and Fig 5). The three replicate groups showed similar trends, and the detailed results are shown in S2 Fig. These results indicated that the enhanced activities of P450s and CCEs led to metabolic resistance in the population, and P450s may hold a prominent role in metabolic resistance.
Under insecticide selection, a very strong positive correlation was observed between metabolic enzyme activity and the degree of resistance (Table 4 and Fig 6). The activity of P450s and CCEs were significantly correlated. The correlations between metabolic enzyme activity (P450s and CCEs) and LC50 in the IS strain were high and slightly higher than correlations in MS strain. We also analyzed the relationship between the frequency of L1014F and the degree of resistance (Table 4). It was noted that the significant correlation between the frequency of L1014F and LC50 was only present in MS strain. There were no significant correlations among metabolic enzyme activity, the frequency of L1014F and the degree of resistance in the NS strain.
Resistance of mosquitoes to pyrethroids appears to rely mainly on target-site and metabolic resistance mechanisms. The two mechanisms can occur singly or simultaneously in resistant populations. A growing number of studies have found both metabolic- and kdr-based resistance mechanisms in most mosquito species [18, 24, 31]. Most researchers found that metabolic detoxification was the most important mechanism for the development of resistance in the mosquito population, whereas the target site played a less important role [32, 33]. Preliminary investigations of underlying resistance mechanisms of the pyrethroid resistance in field populations of Anopheles funestus in southern Africa indicated that a P450-based metabolic resistance was the main mechanism with no kdr mutation identified yet in this species [34]. Ochomo et al. found that phenotypic resistance to permethrin in An. gambiae s.s. was attributed to elevated expression of β-esterase and oxidase enzymes and the presence of kdr alleles at the voltage-gated sodium channel locus, but target-site mechanisms was detected in phenotypic resistance to deltamethrin solely [35]. It was noted that the different mechanisms occurred in the same resistance population. Although studies on the two resistance mechanisms have provided insights, how the target-site and metabolic resistance mechanisms differentially contribute to the resistance phenotype has remained unclear. To tackle these issues, our study detected kdr and three detoxification enzyme activities in the same mosquito sample and dynamically monitored the trends of resistance in populations with different insecticide selection pressures. We found that kdr and metabolic resistance both contributed to deltamethrin resistance in the Cx. pipiens pallens populations, but they had different roles under different selection pressures. The P450s activities increased significantly after generation 14 for IS strain and 18 for MS strain, and only the CCEs increased significantly after generation 18 in IS strain. In both the IS and MS groups, resistance to insecticide and frequencies of the kdr mutation L1014F increased before the detoxification enzyme activities were significantly increased. This phenomenon indicated that the target-site mechanisms was important under a relatively low insecticide selection pressure. After the population became homozygous for the L1014F mutation, the level of resistance grew along with the increase in detoxification enzyme activities. Linear correlation analysis showed that the significant correlation between the frequency of L1014F and LC50 was only present in MS strain, but the correlations between metabolic enzyme activity (P450s and CCEs) and LC50 was significantly high both in IS and MS strain, and slightly higher in IS strain. This finding seemed to indicate that the metabolic resistance increased with the increase of selection pressure, and played a main role in causing a high level of resistance under high insecticide selection pressure. Individual organisms with a low fitness cost will survive under selection. Numerous disruptive mutations can confer resistance (whether through suppression, down-regulation or gene-silencing), and a high dose of a selective agent may overcome fitness costs associated with disruption and thus favor a large pool of normally deleterious mutations [36, 37]. However, fitness costs also select against such alleles in the absence of toxicant selection. Metabolic resistance is mainly caused by the up-regulation of detoxification enzymes with high fitness costs. Some studies suggest that P450s overproduction decreases the fitness of individual organisms that carry them because the overproduced P450s can metabolize hormonal endogenous molecules. However, amino-acid substitutions may possibly involve fewer disturbances to the fitness of the individual [38]. Therefore, the low insecticide selection pressure preferred selection of individuals with kdr. When the selection pressure was increased via a high dose of insecticide, organisms with metabolic resistance were primarily selected.
Metabolic resistance to pyrethroid is known to be associated with three types of detoxification enzymes, P450s, CCEs and GSTs; however, the relative importance and pattern of these enzymes in the development of resistance are still disputed. Pyrethroid resistance is thought to be mediated essentially by P450s. Elevated levels of P450s activity are frequently observed in mosquito species as major malaria vectors in Africa [11, 39, 40]. In whole-genome microarray studies, members of the cytochrome P450 class of metabolic enzymes are frequently up-regulated in pyrethroid-resistant mosquitoes [41]. CCEs are also believed to act as a cause of metabolic resistance in some instances. Recently, the capacity of Aedes aegypti CCEs to metabolize pyrethroids leading to detoxification has been demonstrated in vitro [11]. GSTs are regularly found overexpressed in pyrethroid-resistant mosquitos. The potential role of GSTs in pyrethroid resistance is likely associated with protection against oxidative stress and sequestration of pyrethroids [42, 43]. Detoxification of pyrethroids by P450s either alone or in combination with CCEs and/or GSTs has been suggested to play a role in pyrethroid resistance [8, 44, 45]. In the same population, the metabolic resistance was also different. Ochomo et al. showed the association of elevated activities of β-esterases and P450s with permethrin resistance, but there was no elevated expression of detoxifying enzymes in phenotypic resistance to deltamethrin [35]. The different conclusions among diverse studies confirmed that the development of metabolic resistance is a rather complicated process, which may be affected by the species, strain (field or lab) or insecticides. In our study, the same species and strain were exposed to the identical insecticide, so we could detect the effects of selection pressure on metabolic resistance. Our study found that different selection pressures led to different levels of metabolic resistance with potentially different mechanisms. Metabolic resistance was mainly mediated by P450s under low insecticide selection pressure, but a high level of metabolic resistance was related to P450s and CCEs under high insecticide selection pressure. Results of the metabolic enzyme assay showed that the development of deltamethrin resistance was accompanied by the significant increase in activities of P450s and CCEs in the IS strain, and only P450s were raised in the MS strain. No correlations between GSTs and the level of deltamethrin resistance were found here. The resistance levels and P450s activity were similar at generation 30 for MS strain and generation 26 for IS strain, but the CCEs activity was significantly elevated in IS strain (Table 3 and S2 Table). CCEs activity appeared higher in IS strain, but CCEs activity made substantial jumps from generation 22 to generation 30 in the MS strain. These results suggested that insecticide selection had a finite amount of genetic variation in the base population that could be selected to generate resistant phenotypes. Under mild selection (in MS strain), the P450s activity responded first, then the CCEs activity was necessary and some of the genetic variants responsible for higher CCEs activity were selected from generations 22 to achieve high levels of resistance. Under intense selection pressure (in IS strain), perhaps P450s alleles alone could not segregate fast enough to produce the required resistance level, so CCEs alleles were selected concurrently. As we did not examine the members of the P450s and CCEs families involved in metabolic resistance, further investigations are needed to identify the associated enzymes and their mechanisms in the metabolic-based resistance.
When genomic regions are subjected to strong and recent selection pressure, the adjacent sequences extending outward from the site of selection with reduced diversity provide evidence of a selective sweep [41, 46]. Our study showed the diversity of the kdr gene was reduced with increased resistance under deltamethrin selection by cloning and sequencing the intron region downstream of L1014. These results confirmed that kdr led to the deltamethrin resistance in this population. In our study, a higher association of the L1014F mutation than the L1014S mutation was found with deltamethrin resistance in the Cx. pipiens pallens population. A similar conclusion was made in our previous research [22]. In the current study, we only detected the region around L1014 in the kdr gene, and whether other kdr mutation alleles were related to pyrethroid resistance in this Cx. pipiens pallens population was unclear.
Current chemical control programs are largely dependent on pyrethroids, and their efficacy is now threatened by the rise of resistance in target populations. Therefore, formulating a new insecticide use strategy to delay the development and/or spread of pyrethroid resistance is a priority. Based on results of our study, we propose some considerations for insecticide use. It is a must to avoid long-term exposure of mosquito populations to a constant low concentration of insecticide, because it can lead to significantly increased resistance in populations. In this study, when we selected the population (MS strain) with a constant concentration of deltamethrin (0,05ppm) for 24 generations, the level of resistance grew exponentially and the RR was increased more than eight times. This result suggested that the low concentration of insecticide in environment may affect the vector control, such as insecticides used in agriculture. So reducing pesticide residues and accelerating degradation in the environment may delay the development of resistance. Pesticide use should suit to local conditions, because the kdr and metabolic resistance had different roles under different selection pressures. Our study found that metabolic resistance mainly played a main role under high insecticide selection pressure. So the combination of insecticide and synergists of metabolic enzymes could delay the increase of insecticide in these areas with high resistance. For example the synergist PBO (piperonyl butoxide), a known inhibitor of P450s and esterase activity, has a significant role in vector control and is encouraged by WHO [47]. The NS strain in our study showed a reduction of RR by ~2-fold after 24 generations, but it was still higher than that at the first generation. This indicated that the high frequency of resistance alleles in a population might result in the slow recovery of sensitivity. The kdr alleles were fixed by generation 14 in the NS strain but the RR was still falling. Combined with the results of the steady decline in P450s and CCEs activities (Fig 5), the changes in RR might be caused by reductions in metabolic enzyme activities. However, the RR was still higher than the first generation, it might indicate that fitness costs of metabolic resistance were low. It needs further investigation. Whether in IS or MS, the level of resistance grew exponentially after long term exposure to pesticides. It is a must to rotate the insecticides promptly before the occurrence of high resistance. The insecticides adopted for rotations should have different modes of action to avoid cross-resistance. We demonstrated that the kdr gene and metabolic enzymes played different roles in the development of pyrethroid resistance. Hence, the target sites of insecticides and the resistance induced by metabolic enzymes should both be taken into account in rotations. In addition, as the pattern of decline in resistance can be slow, insecticide withdrawal must be maintained for an extended period of time. A study by Raghavendra et al. showed that persistence of resistance to DDT and malathion was respectively observed 30 yr and 9 yr after withdrawal from IRS, although reversal of deltamethrin resistance was observed relatively rapidly within 2–3 yr after its withdrawal from IRS [48]. Conclusions from our study may provide a reference for future management of insecticide resistance in mosquitoes and other arthropod pests and disease vectors.
|
10.1371/journal.pbio.1000055 | Regulation of Embryonic Cell Adhesion by the Prion Protein | Prion proteins (PrPs) are key players in fatal neurodegenerative disorders, yet their physiological functions remain unclear, as PrP knockout mice develop rather normally. We report a strong PrP loss-of-function phenotype in zebrafish embryos, characterized by the loss of embryonic cell adhesion and arrested gastrulation. Zebrafish and mouse PrP mRNAs can partially rescue this knockdown phenotype, indicating conserved PrP functions. Using zebrafish, mouse, and Drosophila cells, we show that PrP: (1) mediates Ca+2-independent homophilic cell adhesion and signaling; and (2) modulates Ca+2-dependent cell adhesion by regulating the delivery of E-cadherin to the plasma membrane. In vivo time-lapse analyses reveal that the arrested gastrulation in PrP knockdown embryos is due to deficient morphogenetic cell movements, which rely on E-cadherin–based adhesion. Cell-transplantation experiments indicate that the regulation of embryonic cell adhesion by PrP is cell-autonomous. Moreover, we find that the local accumulation of PrP at cell contact sites is concomitant with the activation of Src-related kinases, the recruitment of reggie/flotillin microdomains, and the reorganization of the actin cytoskeleton, consistent with a role of PrP in the modulation of cell adhesion via signaling. Altogether, our data uncover evolutionarily conserved roles of PrP in cell communication, which ultimately impinge on the stability of adherens cell junctions during embryonic development.
| Unlike conventional pathogens, prions are infectious particles devoid of nucleic acids and composed entirely of a misfolded host protein, PrP. It is widely assumed that the neurodegeneration observed in prion disorders may be related to an aberrant function of PrP in the misfolded state. However, the normal physiological function of PrP remains poorly understood, mainly owing to the absence of clear phenotypes in mice lacking PrP. Here, we show that when PrP is depleted in zebrafish embryos, dramatic phenotypes ensue, severely affecting the development of early and late (neural) structures. We examined the mechanisms responsible for some of these defects, and found that fish and mammalian PrPs play conserved roles in cell–cell communication, by directly mediating cell adhesion and by triggering cellular signals that further modulate the function of other adhesion molecules. In the early zebrafish embryo, these activities control not only tissue integrity and cell morphology, but also the complex cellular movements that give rise to germ layers. This study describes—to our knowledge—the first known in vivo function of PrP and its molecular cellular basis, which may provide helpful insights into the role of PrP in the adult brain and its proposed connections to prion-induced neurotoxicity.
| The prion protein (PrP) is a membrane-anchored glycoprotein, best known for its unique ability to undergo structural conversion from a normal “cellular” isoform (PrPC) into a pathogenic conformer known as “scrapie” (PrPSc) [1]. The accumulation of scrapie aggregates—prions—in the brain is a distinctive feature of transmissible spongiform encephalopathies, a group of lethal neurodegenerative diseases that include Kuru and Creutzfeldt-Jacob disease in humans, scrapie in sheep, and mad cow disease in cattle [2]. While much has been learned about the pathogenic properties of PrP, its normal physiological role remains elusive [3,4]. We previously identified PrP orthologs in fish and proposed that the conservation of their protein domain architecture reflects the maintenance of an ancient and important biological role of PrP across vertebrates [5].
Although PrP is widely expressed in mouse embryos [6], PrP knockout mice are surprisingly viable and show no major physical or behavioral abnormalities [7]. For the last 17 years, this lack of in vivo phenotypes has precluded PrP from genetic functional analysis, raising the intriguing question of whether its unknown physiological function is necessary or dispensable for the organism, and also whether prion neurotoxicity may be a consequence of PrP mis- or loss-of-function. So far, diverse roles have been proposed for PrPC, including signal transduction [8], cell adhesion and protection from apoptosis and oxidative stress [4], as well as neurogenesis [9,10], axonal growth [11], hematopoietic stem cell self-renewal [12], and lymphocyte activation [13,14]. However, these potential PrP functions do not seem to share a common molecular basis, and their in vivo relevance remains to be clarified.
Here, we show that early down-regulation of PrP impairs cell adhesion in the zebrafish embryo, disrupting morphogenetic cell movements and ultimately causing developmental arrest. Using aggregation assays, we established that PrP subserves complex roles in both Ca+2-independent and Ca+2-dependent cell adhesion. Our analyses of morphant embryonic cells revealed that PrP is required for the proper membrane localization of E-cadherin adhesion complexes. We also carried out experiments in Drosophila S2 cells to demonstrate that PrP itself induces homophilic cell adhesion, and that its accumulation at cell contacts leads to the recruitment of microdomain-associated proteins, eliciting signal transduction and rearrangement of the actin cytoskeleton. Finally, we found that the roles of PrP in cell adhesion and signaling are conserved across vertebrate classes, and that PrP interactions can take place even between mouse and fish orthologs. Our results contribute novel molecular and cellular aspects of PrP function in vitro and in vivo, which may be of relevance to understanding its long-sought physiological roles in the mammalian brain, as well as the potential link between PrP loss-of-function and prion-induced neurodegeneration.
Because zebrafish contain duplicated PrP genes [5], PrP-1 and -2, we assessed their degree of functional overlap by examining their patterns of embryonic expression. RNA in situ hybridization shows strong and ubiquitous distribution of maternal PrP-1 transcripts at early midblastula stages (2.5 h postfertilization [hpf], Figure 1A); this is followed by a sharp decrease after gastrulation to reach minimal levels in the forebrain and eyes at the pharyngula stage (30 hpf, Figure 1B). In contrast, PrP-2 transcripts remain undetectable from 2.5 hpf (Figure 1C) until somitogenesis but reach high levels by 30 hpf, especially in the brain (telencephalon and diencephalon), as well as in discrete neuronal populations of the central nervous system (CNS) (trigeminal ganglia and neuromeres) and peripheral nervous system (PNS) (lateral line ganglia and Rohon-Beard neurons) (Figure 1D, arrows). These complementary expression patterns, which were not evident in previous studies [15], indicate that fish PrP-1 and -2 carry out specialized roles during distinct developmental stages: PrP-1 transcripts are abundant at early stages characterized by active cell division and migration in the entire embryo, whereas PrP-2 is specifically up-regulated later in the developing nervous system, particularly in neurogenic placodes.
The developmental expression patterns of PrP-1 and -2 indicated that they perform early ubiquitous and late neural embryonic roles, respectively. To test this prediction, we blocked the translation of each gene by microinjecting antisense morpholinos into one- to four-cell stage embryos. Such knockdown of PrP-1 produced a strong early phenotype characterized by the failure to carry out gastrulation beyond 6 hpf (shield stage), resulting in a large proportion of arrested embryos at 9 hpf (∼95%, n = 200) (Figure 2B) compared to control embryos (5%, n = 200) (Figure 2A, Table S1). Western blot analysis of 6-hpf embryos confirmed that PrP-1 expression levels were effectively suppressed by morpholino microinjection (Figure 2G). On the other hand, PrP-2 knockdown embryos showed normal gastrulation (Figure 2C) and survived into early larval stages (≥7 d postfertilization [dpf]), but presented morphological defects in the head region, particularly malformed brains and eyes (∼65%, n = 200) (Figure 2K and 2M). These early and late phenotypes correlate with the developmental expression patterns of PrP-1 and -2, respectively. Thus, the gastrulation arrest in PrP-1 morphant embryos reveals that this protein is essential for epiboly (the spreading of the blastodisc from the animal to the vegetal pole), whereas the malformations observed in PrP-2 morphant embryos are consistent with a specific role of this protein in neural differentiation and brain morphogenesis.
Given their shared protein domain composition, PrP-1 and -2 are likely to have similar biological activities, despite their distinct amino acid sequences and developmental expression patterns. To examine the degree of functional relatedness between zebrafish (and mouse) PrPs, we tested their ability to rescue the PrP-1 loss-of-function phenotype. Embryos were coinjected with various combinations of PrP-1 morpholino and in vitro synthesized PrP mRNAs lacking the morpholino binding sites. As expected, the severity of the PrP-1 early phenotype could be rescued with PrP-1 mRNA (21% of arrested embryos, n = 200) (Figure 2D): coinjected embryos overcame the gastrulation arrest and survived into larval stages. Remarkably, partial rescue of the PrP-1 knockdown was also observed upon coinjection of PrP-2 (38% of arrested embryos, n = 200) (Figure 2E) and even mouse PrP mRNAs (47% of arrested embryos, n = 200) (Figure 2F). Differences in rescue efficiency between these mRNAs were also seen in the variation of the degree of epiboly attained at 9 hpf: while control embryos and PrP-1 morphants reached about 90% (Figure 2A, arrowheads) and 50% epiboly (Figure 2B, arrowheads), respectively, the zebrafish (Figure 2D and 2E, arrowheads) and mouse PrP (Figure 2F, arrowheads) rescues attained about 90% and 70% epiboly at this time, respectively (Table S1). Rescues using the corresponding EGFP-PrP fusion mRNAs (Figure 3A) produced similar results (Figure 2H and Table S2) and allowed us to visualize ubiquitous expression of the rescuing fusion proteins (Figure 2H and 2I). Furthermore, control mRNAs coding for only the PrP leader and GPI-anchor peptides (Figure 3A, controls) could not revert the PrP-1 phenotype (87.5% of arrested embryos, Figure 2I and Table S2), confirming that the rescue ability depends on the presence of the PrP cores (repetitive, hydrophobic, and globular domains). In contrast, the PrP-2 phenotype could not be rescued due to the technical limitation of having to inject the mRNAs at the one- to four-cell stages: rescuing mRNAs were inevitably expressed at blastula stages, causing early ectopic overexpression before the endogenous PrP-2 could actually be transcribed and therefore targeted by the morpholino. Interestingly, ectopic (ubiquitous) overexpression of zebrafish or mouse PrP mRNAs produced similar morphological phenotypes (Figure S1): asymmetric epiboly and severe defects in eye and brain morphology. Thus, although not identical, PrP down-regulation and overexpression phenotypes converge at the same developmental processes (gastrulation and neural development) where a basic cellular function shared by fish and mouse PrPs is required.
To gain preliminary insight into the cellular mechanisms responsible for the phenotypes observed, we analyzed the heterologous expression of zebrafish PrPs in mouse neuroblastoma 2a (N2a) cells, a neuronal cell line routinely used to study the functional and pathogenic properties of PrP. To overcome the lack of anti–zebrafish-PrP antibodies suited for immunofluorescence, we used EGFP-PrP fusion constructs (Figure 3A). In these experiments, expression of zebrafish and mouse EGFP-PrP constructs in N2a cells led to strong protein accumulation at cell contacts (Figure 3B, 3D, and 3F, arrowheads and fluorescence profiles). This phenomenon was not observed upon surface expression of control EGFP constructs (Figure 3C, 3E, and 3G), indicating that the PrP leader and GPI-anchor peptides are sufficient for protein targeting and attachment to the cell membrane, but that the accumulation at cell contacts is dependent on the presence of the PrP cores. Moreover, PrP accumulation was observed only when both cells forming the contact expressed the PrP construct (Figure S2A), suggesting that PrPs might engage in homophilic trans-interactions. Interestingly, while mouse PrP and zebrafish PrP-2 were observed along the entire cell membrane (Figure 3B and 3F), PrP-1 localized almost exclusively at cell contacts (Figure 3D), suggesting a contact-dependent regulation of PrP-1 membrane positioning. PrP accumulation at N2a cell contacts could also be observed by immunostaining endogenous PrP with a specific monoclonal antibody (Figure S2B), as well as by using DsRed-monomer constructs (Figure S2C–S2H), and in HeLa cells (data not shown). To examine zebrafish and mouse PrP expression in vivo, we microinjected the corresponding EGFP-PrP mRNAs into zebrafish embryos. The localization patterns observed at 6 hpf (Figure 3H–3K) were consistent with those seen in N2a cells, including the relatively homogeneous membrane distribution of PrP-2 and mouse PrP (Figure 3I and 3J), the local accumulation of PrP-1 in patches at cell contacts (Figure 3H), and even the loss of discrete accumulation by a control construct lacking the PrP-1 core (Figure 3K). Similar to their mammalian counterparts, attachment of zebrafish PrPs to N2a cell membranes via GPI-anchors was greatly reduced by PI-PLC treatment (Figure S2I and S2J); likewise, N-glycosylation could be demonstrated by PNGase F digestion (see Western blots in Figure S2K). We also generated mutant constructs for PrP-1 and PrP-2 in which the putative N-glycosylation residues were point mutated to glutamine. Western blot analysis of these constructs confirmed that, like mammalian PrPs, zebrafish PrP-1 and -2 are mono- and di-glycosylated at the expected residues (Figure 3L). These experiments strengthen the concept that the cellular function of PrP is conserved between fish and mammals.
The specific accumulation of PrP at cell contacts suggested that the zebrafish PrP phenotypes could be explained by defects in cell–cell communication. Given the relative simplicity and ease of manipulation of the early zebrafish embryo, we focused our analysis on the cellular and molecular characterization of the PrP-1 phenotype. Morphological examination of PrP-1 knockdown embryos at 6 hpf revealed that the developmental arrest was preceded by a marked decrease in tissue integrity and compactness; as they detached, deep cells in the morphant embryo lost their otherwise polygonal shape and became round (Figure 4A and 4B). The progressive loss of cell adhesion was clearly not a consequence of cell death, as in control embryos, death at this stage usually leads to generalized cell lysis within a few minutes. In contrast, round morphant cells survived at least until 12 hpf. Moreover, TUNEL (terminal deoxynucleotidyl transferase-mediated dUTP nick-end labeling) and DAPI stainings of 6-hpf PrP-1 morphant embryos showed no signs of apoptotic DNA fragmentation (Figures 4D–4F and S3A–S3D). Thus, the loss of embryonic cell adhesion is a specific effect of PrP-1 knockdown, which can be rescued by the local accumulation of exogenous PrPs at cell contacts sites (Figure 4C, arrowheads).
To quantitatively assess the cell adhesion defect in PrP-1 knockdown embryos, we prepared primary cultures of dissociated zebrafish blastomeres (single-cell suspensions of 6-hpf embryos) and tested their reaggregation potential in the presence of Ca+2. After 45 min in suspension, control cells formed cell aggregates with an average size of 4.5 ± 0.2 cells/aggregate (maximum size = 29 cells/aggregate), whereas PrP-1 morphant cells formed significantly smaller aggregates (p < 0.001) with an average size of 2.7 ± 0.1 cells/aggregate (maximum size = 9 cells/aggregate) (Figure 5A and 5B). Moreover, when dissociated control and morphant cells were cocultured, compact aggregates of control cells formed rapidly (within 5 min), from which loose morphant cells were often excluded (Figure 5C, arrowhead).
During gastrulation, cell adhesion is dynamically maintained by cadherin homophilic interactions [16]. This raised the possibility that the PrP-1 knockdown phenotype could be due—at least partly—to the misregulation of cadherin function, which is Ca+2 dependent [17]. To test this hypothesis, we performed aggregation assays with dissociated control and PrP-1 morphant cells in the presence and absence of Ca+2. When the assay was performed in the Ca+2-containing medium, PrP-1 morphant cells underwent a significant decrease in the number of large (>10 cells) and small (<10 cells) aggregates (100%, p = 0.0004; and 36% reduction, p = 0.003, respectively), compared to control cells (Figure 5D, + bars). The same relative effect was observed when the assay was performed in Ca+2-free medium, indicating that PrP-1 is required for the formation of Ca+2-independent cell clusters, mostly of small size (Figure 5D, − bars). On the other hand, aggregation of PrP-1 morphant cells in the presence or absence of Ca+2 showed no significant differences, implying that the formation of large clusters is mediated by cadherins and controlled by PrP-1. Interestingly, when the assay was performed with PrP-1 overexpressing cells in the presence of Ca+2, a dramatic increase in the number of large clusters was recorded (200% increase, p = 4 × 10−7, Figure 5D, + bars); in the absence of Ca+2, PrP-1 overexpressing cells virtually formed no large clusters (as expected), and the number of small clusters was larger than that of PrP-1 knockdown cells, and comparable to that of control cells (Figure 5D, − bars). These results are consistent with a complex role of PrP-1 in the maintenance of embryonic cell adhesion via cell-autonomous interactions at the plasma membrane. To test for cell autonomy, we transplanted small groups of deep cells from 4-hpf blastulae treated with PrP-1 and control morpholinos into 4-hpf control blastulae (Figure 5G). After 2 h, control cells established normal cell contacts within the control host embryo and acquired polygonal morphology (10 out of 12 experiments, Figure 5E). In contrast, PrP-1 morphant cells remained round and loose within the control host embryo (10 out of 10 experiments, Figure 5F). Thus, the adhesion defect in PrP-1 morphant cells is cell-autonomous and cannot be corrected by the cellular environment of the control host embryo. Unfortunately, transplantation of control cells into a morphant host was not informative because PrP-1 morphant embryos did not withstand manipulation.
Our aggregation assays indicated that PrP-1 can modulate Ca+2-dependent cell adhesion in the embryo. Therefore, we investigated whether PrP-1 knockdown would affect the expression and subcellular localization of E-cadherin. Since cadherin homophilic interactions are anchored to the actin cytoskeleton via catenins [18], we also analyzed PrP-1–mediated changes in the distribution of β-catenin and F-actin. Antibody and phalloidin stainings revealed the typical cell surface localization of these molecules in 6-hpf control embryos (Figure 6A–6C). In contrast, E-cadherin and β-catenin appeared largely intracellular in PrP-1 morphant cells, and the distribution of F-actin was disorganized (Figure 6D–6F). This apparent intracellular accumulation of E-cadherin could be due to increased E-cadherin endocytosis and/or degradation, or to deficient trafficking to the plasma membrane. To address these possibilities, we first carried out Western blot analysis on cell extracts from 6-hpf control and PrP-1 morphant embryos. Notably, PrP-1 morphant cells showed an almost complete reduction in the levels of the 120-kDa polypeptide reported to be the active membrane-bound form of E-cadherin [19], as well as a slight increase in the levels of the 140-kDa immature form of E-cadherin [19], thought to be abundantly stored in intracellular compartments [20] (Figure 6G). Previous studies have identified the recycling endosome, with its associated small GTPase Rab11, as an intermediate compartment that regulates post-Golgi trafficking and exocytosis of E-cadherin to the plasma membrane [21]. We reasoned that if the increased intracellular distribution of E-cadherin in PrP-1 morphant cells was due at least in part to deficient delivery of E-cadherin to the plasma membrane, then E-cadherin would be seen to accumulate in Rab11-positive intermediate compartments. Therefore, we analyzed the subcellular distribution of E-cadherin and Rab11 in 6-hpf embryos by double immunostainings, and quantitatively assessed changes in their degree of colocalization upon PrP-1 knockdown. We found that the marked intracellular distribution of E-cadherin in morphant embryos was accompanied by a significant increase (p < 0.001) in the number of E-cadherin/Rab11 double-positive vesicles (Figure 6H–6N). These experiments suggest that PrP-1 can modulate the function of E-cadherin by regulating its processing and/or transport from intracellular stores to the plasma membrane.
To verify that PrP-1 can specifically modulate the stability of adhesion complexes at discrete cell–cell contacts (as opposed to within a tissue), we also carried out immunostainings on dissociated blastomeres that had been allowed to reaggregate. Comparison of control and morphant cells showed that the local accumulation of E-cadherin and β-catenin at newly formed cell contacts requires PrP-1 (Figure 6, compare 6S and 6T with 6O and 6P). Moreover, we observed PrP-1–dependent accumulation of Fyn tyrosine kinase and tyrosine phosphorylated proteins at cell contacts (Figure 6, compare 6U and 6V with 6Q and 6R), suggesting that the regulation of E-cadherin localization by PrP-1 involves the activation of Src-related kinases and downstream targets. We asked next whether the regulatory role of PrP-1 over E-cadherin could be further confirmed by showing that the two molecules interact at the genetic level. To test for synergistic interactions, we microinjected embryos with low doses of PrP-1 morpholino, E-cadherin morpholino, or the two morpholinos together, and scored the number of embryos with arrested gastrulation at 6 hpf. Our results show a statistically significant increase (p < 0.005, n = 3, ∼400 embryos per experiment) in the percentage of arrested embryos for the PrP-1/E-cadherin double knockdown (88.75 ± 1%), compared to the PrP-1 (46.02 ± 0.8%) or to the E-cadherin (38.16 ± 0.46%) single knockdowns, or to the control morpholino (1.29 ± 0.23%). These data strongly suggest that PrP-1 and E-cadherin act synergistically to regulate embryonic cell adhesion.
Kane et al. [22] have demonstrated that E-cadherin mutants fail to carry out epiboly because radial intercalation—a morphogenetic cell movement crucial for the spreading of the blastoderm over the yolk cell—is impaired. In these mutants, cells from the interior layer of the epiblast move normally to the exterior layer but fail to integrate and become restricted, thereby blocking the expansion of the blastoderm. Because E-cadherin adhesion is affected in PrP-1 morphants, we hypothesized that their epibolic arrest could be explained by defective radial intercalation. To confirm this, we carried out time-lapse recordings of cell behavior in the epiblast. In control embryos (Figure 7A), intercalating cells (in blue) entered the exterior layer and flattened out within approximately 20 min, effectively increasing their area. During this process, cells from the exterior layer that were initially in contact with the intercalating cell (in green), or that established new contacts with it (in red), remained in stable contact. In PrP-1 morphant embryos (Figure 7B), intercalating cells entered the exterior layer and increased their area, but did not completely flatten out: after approximately 10 min, they eventually reduced their area and left the exterior layer (deintercalation). Moreover, not all exterior cells that were in contact with the intercalating cell remained in contact with it (see green cells 2 and 4, Figure 7), and other cells that became in contact with it (red cells) did not maintain stable cell contacts. Overall, deintercalation events occurred frequently in morphant embryos but were not observed in control embryos, suggesting that the developmental basis for the epibolic arrest of PrP-1 morphant embryos is the impairment of morphogenetic cell movements directly controlled by E-cadherin. In agreement with this, the corresponding time-lapse videos show that PrP-1 morphant cells have a reduced ability to maintain tissue cohesion and to migrate in a coordinated fashion (compare Videos S1 and S2). Interestingly, and despite these defects, morphant cells seemed quite active in generating processes. To further clarify this issue, we made time-lapse recordings of single dissociated blastomeres (8 hpf), which show that their intrinsic motility is not affected by PrP-1 knockdown (compare Videos S3 and S4). To assess whether PrP-1 knockdown might affect other cellular events relying on proper cell–cell communication, we also controlled for changes in mitosis rates upon PrP-1 knockdown. Interestingly, we found a partial reduction in the numbers of dividing blastomeres in morphant embryos (Figure S3E); however, neither cell size nor overall cell density appeared to be significantly compromised in these embryos.
It has previously been reported that mutation or down-regulation of E-cadherin specifically affects adhesion in the deep cell layer but not in the enveloping layer (EVL) of the zebrafish gastrula [23,24]. Similarly, the PrP-1 adhesion phenotype reported here appears restricted to deep cells, supporting the notion that PrP-1 modulates E-cadherin function, and that cell adhesion in the EVL may be controlled by additional mechanisms, such as the use of different types of cellular junctions. To further investigate this, we studied the distribution of markers for adherens and tight junctions in the EVL at 6 hpf. In PrP-1 morphants, EVL cells showed similar alterations in the distribution of E-cadherin, β-catenin, and F-actin as in deep cells (Figure 8A–8C and 8F–8H; for a Z-stack reconstruction of an embryo showing the size and morphology of both types of cells, see Video S5), indicating that PrP-1 regulates the stability of adherens junctions in both cell layers. In contrast, the membrane distribution of classical tight junction markers like occludin and ZO-1 remains unaffected in morphant EVL cells (Figure 8D, 8E, 8I, and 8J), whereas their presence in deep cells (control or morphant) could not be detected. Hence, the loss of PrP-1 and E-cadherin function may not significantly impair EVL cell adhesion due—at least in part—to its different cell junction composition.
The formation of PrP-1-dependent small cell clusters in the absence of Ca+2 (Figure 5D) indicates that PrP might have its own adhesive properties. To clarify this, we employed nonadhesive Drosophila S2 cells, an experimental paradigm classically used to demonstrate the adhesive properties of membrane proteins [25]. Notably, S2 cells transfected with mouse and zebrafish EGFP- and DsRed-monomer-PrP constructs acquired the ability to aggregate and accumulate PrP at cell contacts (Figures 9A–9C and S4; Video S6). In contrast, control EGFP constructs did not accumulate even at fortuitous cell contacts (Figure 9F). Since S2 cells lack endogenous PrP [26] and do not express adhesion molecules, these experiments show that PrP expression leads to cell aggregation. Similarly, aggregation of cells transfected with frog and chicken PrP constructs indicates that the PrP-mediated adhesion is conserved across vertebrate classes (Figure 9D and 9E, arrowheads). To corroborate that PrP accumulates between distinct cells (as opposed to between dividing daughter cells), we reproduced these results using mixed S2 cell populations separately transfected with EGFP- and DsRed-monomer-PrP constructs (Figure 9G, arrowheads). In these experiments, untransfected cells did not form aggregates and remained excluded from the mouse PrP-transfected aggregates (data not shown), suggesting that cell aggregation is due to homophilic affinity between PrPs on apposing cell membranes. Interestingly, cross-interactions involving zebrafish and mouse PrPs also triggered aggregation of S2 cells (47 ± 1% aggregated cells compared to 83 ± 2% between mouse PrPs and 94 ± 2% between zebrafish PrP-2s, Figure 9H, arrowheads), revealing that, unlike the species restrictions that limit prion propagation [27], functional interactions can take place even between distantly related PrPs.
Our present work in zebrafish embryos revealed that PrP can regulate the function of E-cadherin; similarly, other studies have reported functional interactions between PrP and NCAM [10]. To exclude the possibility that PrP might cause S2 cell aggregation by indirectly activating these endogenous adhesion molecules, we immunostained S2 cells for DE-cadherin, DN-cadherin, and fasciclin II (Drosophila NCAM), and confirmed that these molecules are not re-expressed upon transfection with PrP constructs. To further test whether PrP itself can act as an adhesion molecule, we asked whether an anti-PrP antibody could interfere with the aggregation process. Indeed, clustering of cells transfected with mouse EGFP-PrP was drastically reduced by incubation with M-20, an anti-mouse PrP polyclonal antibody (Figure 9I, fluorescence images, arrowheads). The reduction in the percentage of aggregated cells was concentration-dependent, going from 83 ± 6% (control without antibody) to 48 ± 6% and 18 ± 5% in the presence of 2 and 4 μg/ml M-20, respectively. In contrast, incubation with 4 μg/ml of control antibody did not significantly affect cell clustering (82 ± 7%) (Figure 9I, histogram). Moreover, PrP-mediated aggregation of S2 cells was not affected by the presence of 0.5 mM EGTA (data not shown). Thus, we conclude that PrP can mediate homophilic cell adhesion in a Ca+2-independent manner.
Using T cells, we previously have shown that PrP can elicit signal transduction and reorganization of the actin cytoskeleton via Src-related tyrosine kinases, and that these events take place at specialized membrane microdomains defined by the presence of reggie/flotillin scaffolding proteins [14]. The present study shows that similar cellular signals between zebrafish blastomeres are inhibited upon PrP-1 knockdown (Figure 6T and 6U). Therefore, we investigated whether such signaling events are also induced in S2 cells upon the formation of Ca+2-independent, PrP-mediated contacts. In fact, activated Src-kinase (Figure 9J, arrowheads) and phosphotyrosine staining (Figure S5A–S5C), as well as coclustering of reggie/flotillin membrane microdomains (Figure 9L, arrowheads, and Figure S6A–S6E), and accumulation of F-actin (Figure 9N) could be seen to colocalize with PrP accumulation at cell contacts. Expression of control EGFP constructs did not induce such effects (Figure 9K, 9M, and 9O), strongly suggesting that the signaling observed is concomitant with PrP-mediated cell adhesion.
The present study shows that the loss of PrP function in a vertebrate can produce clear phenotypes amenable to cellular and molecular characterization. Our experiments reveal important roles of PrP during zebrafish development, PrP-1 regulating embryonic cell adhesion during gastrulation, and PrP-2 affecting later stages of neural development. Interestingly, the developmental expression and cellular localization patterns of PrP-2 suggest that it might play a role closer to that of mammalian PrP in the nervous system. In fact, mouse PrP is expressed throughout the developing nervous system in a pattern analogous to that of zebrafish PrP-2 [6]. Nevertheless, the differences in expression patterns and knockdown phenotypes between PrP-1 and −2 are most likely of transcriptional regulatory nature, because both proteins share important biological properties with mouse PrP, such as membrane anchorage, posttranslational processing, subcellular localization, and the ability to revert the PrP-1 knockdown phenotype. The strength of the PrP knockdown phenotypes in zebrafish is in sharp contrast with the lack of significant defects in PrP knockout mice. This striking difference may be due to the activation of gene compensatory mechanisms in the embryonic stem (ES) cells selected to derive the mouse knockouts [2]. For instance, a potential role of PrP in supporting axonal growth can be compensated by up-regulation of integrins in PrP knockout mice [11]. Hence, clear PrP phenotypes in mammals might become visible only upon replacement of the PrP gene with a dysfunctional (i.e., truncated) copy, such as in the “Shmerling phenotype” [28]. Examination of the immediate changes in gene expression upon PrP knockout in mammalian ES cells and embryos might help clarify this issue.
While PrP may establish interactions in cis or trans with adhesion molecules like NCAM [10], the specific accumulation of PrPs at the contacts between transfected N2a cells suggests that PrPs can also establish trans-interactions on apposing plasma membranes, as it has been hypothesized for brain endothelial cells [29]. In fact, our aggregation assays with zebrafish blastomeres and Drosophila S2 cells show that PrP itself can mediate Ca+2-independent homophilic cell adhesion, and that this adhesive property is conserved across vertebrate classes. Moreover, our results also demonstrate that PrP interactions play a regulatory role in vivo, by eliciting the signal transduction events necessary to modulate Ca+2-dependent cell adhesion in the zebrafish gastrula. In particular, we show that PrP-1–mediated signaling influences proper processing and/or trafficking of E-cadherin from storage vesicles to adherens junctions at the plasma membrane. This upstream regulatory role of PrP-1 over of E-cadherin is underscored by the remarkable similarities between the developmental roles of these two molecules. For instance, zebrafish E-cadherin is maternally expressed at adherens junctions and required to regulate embryonic cell adhesion of deep but not EVL cells; mutation or knockdown of E-cadherin induce epibolic arrest and disaggregation of blastodermal cells [20,22,24]. In this study, we further strengthened the functional connection between PrP-1 and E-cadherin, by showing that they interact genetically to control distinct cell morphogenetic movements required for zebrafish gastrulation.
Our experiments suggest that the regulation of E-cadherin by PrP-1 is likely to occur indirectly via signal transduction, and not through direct physical interaction. Accordingly, PrP colocalizes, but does not physically interact, with E-cadherin in cell junctions of human enterocytes [30]. Likewise, the different embryonic localization patterns of PrP-1 (in discrete membrane patches) and E-cadherin (along entire cell–cell contacts) argue against a necessary physical interaction between the two. Moreover, discrete localization in patches at cell contacts, and regulation of E-cadherin function via signaling have also been described for the wnt11 signaling pathway [31,32]. Interestingly, we find that PrP-1 knockdown also results in increased cytoplasmic accumulation of β-catenin. Since the loss of β-catenin signaling is known to increase neuronal apoptosis in Alzheimer's disease patients [33], it would be interesting to study whether β-catenin or wnt signaling are affected along with PrP function during prion-induced neurodegeneration.
Our previous work on T cells revealed that PrP-mediated signaling can trigger activation of Src-related kinases (such as Fyn and Lck) and elevation of intracellular Ca+2 levels, along with reorganization of the actin cytoskeleton [14]. Here, we show that similar signaling events are induced upon PrP accumulation at cell contacts. Interestingly, Src-related kinases are known to regulate cell adhesion via direct phosphorylation of p120 and β-catenins [34,35]. Moreover, two Src-related kinases, Fyn and Yes, are required for Ca+2 signaling and for regulation of the actin cytoskeleton during zebrafish epiboly [36,37]. Thus, PrP signaling may modulate embryonic cell adhesion and actin cytoskeleton dynamics through the activation of Src-related kinases and associated targets.
Analysis of the PrP-2 phenotype is beyond the focus of this study. However, ongoing studies in our lab (E. Málaga-Trillo and L. Luncz, unpublished data) indicate that PrP-2 plays a role in the proliferation and differentiation of developing neurons in vivo, similar to what has been shown for mouse embryonic cells [9]; the precise signaling pathways involved in these processes remain to be clarified.
Altogether, our experiments reveal evolutionarily conserved roles of PrP in the maintenance of Ca+2-dependent and Ca+2-independent embryonic cell adhesion. On one hand, we showed that PrP can directly mediate homophilic cell adhesion and signaling via Src-related kinases. On the other hand, we uncovered a functional link between the activity of PrP at cell contacts and the regulation of cadherin-mediated cell adhesion. Furthermore, we have demonstrated that these roles of PrP are required in vivo to regulate morphogenetic movements that drive early zebrafish development. The implications of these findings for mammalian prion biology await further elucidation; however, they open new avenues for the study of PrP function and prion-induced neurodegeneration across vertebrate models.
The leader, core, and GPI-anchor signal fragments (144, 1,677, and 120 bp for PrP-1; 219, 1,485, and 138 bp for PrP-2; 105, 660, and 87 bp for mouse PrP; 126, 447, and 75 bp for Xenopus PrP; and 129, 588, and 84 bp for chick PrP, respectively) containing the necessary restriction sites were generated by PCR and cloned into pCRII-TOPO as separate fragments, before subcloning into the end vectors. The corresponding EGFP-PrP constructs were engineered through conventional cloning procedures, by inserting leader cDNAs into the NheI/AgeI; and core and/or GPI-anchor cDNAs into the BglII/EcoRI sites of pEGFP-C1 (BD Biosciences Clontech). Zebrafish PrP N-glycosylation mutants were engineered by introducing point mutations (asparagine to glutamine) in residues 509 and/or 514 of PrP-1, and 438 and/or 443 of PrP-2. DsRed-monomer-PrP constructs were generated by replacing the EGFP ORF with the DsRed-monomer sequence (BD Biosciences Clontech). For Drosophila S2 cell transfection experiments, the EGFP- and DsRed-PrP constructs were subcloned into the XbaI/ApaI sites of the pAc5.1/V5-HisA vector (Invitrogen, provided by V. Katanaev). For morpholino rescue experiments, PrP ORF cDNAS were subcloned into the EcoRI site of pCS2+ [38] (provided by Z. Varga) and transcribed in vitro using the mMESSAGE mMACHINE SP6 kit (Ambion). For colocalization studies in S2 cells, rat reggie-1 and −2, and zebrafish reggie-2a were cloned into the EcoRI/BamHI sites of the pDsRed-Monomer-N1 (BD Biosciences Clontech), and further subcloned into the EcoRI/NotI sites of the pAc5.1/V5-HisA vector.
Zebrafish developmental stages are indicated after Kimmel [39] and in hours postfertilization (hpf). Whole-mount in situ hybridization was performed as described in http://zfin.org/zf_info/zfbook/chapt9/9.8.html. The globular domain regions of zebrafish PrP-1 and PrP-2 were cloned in pCRII-TOPO (Invitrogen) and used as templates for the synthesis of RNA in situ hybridization probes with the DIG RNA Labeling Kit (Boehringer). Transcription patterns were visualized on an Axioplan 2 compound microscope (Carl Zeiss) using Nomarski optics, photographed with a Zeiss Color Axiocam, and further processed with Adobe Photoshop 8.0.
The following morpholinos were purchased from Gene Tools and designed to target two independent sequences at the 5′ UTRs of each zebrafish PrP: MO-PrP1–1 (5′-TGA GCA GAG AGT GCT GCG GGA GAG A-3′), MO-PrP1–2 (5′-CGC TTC TTC AAC CTT TTT ATG GAC C-3′), MO-PrP2–1 (5′-CCA AGG GAC AAC AAT CGC CCA AGA G-3′), MO-PrP2–2 (5′-AGG ACT CGC TTA AAA CAG CCC GAA G-3′), Control (5′-CCT CTT ACC TCA GTT ACA ATT TAT A-3′), and MO-Cdh1, targeting the first 25 coding base pairs of zebrafish E-cadherin [20]. All microinjections were performed at early cleavage stages (one- to four-cell stage) using a manual micromanipulator (Narishige) coupled to a Transjector 5246 (Eppendorf) under a Stemi 2000 stereomicroscope (Zeiss). After running specificity and dose-dependency controls, morpholinos were injected at a concentration of 0.8 ng/nl in 1× Danieau buffer (58 mM NaCl, 0.7 mM KCl, 0.4 mM MgSO4, 0.6 mM Ca(NO3)2, 5.0 mM HEPES [pH 7.6]) and 0.125% Phenol Red (Sigma); both PrP-1 morpholinos produced the same phenotype (Table S1). For double-knockdown experiments, low doses (0.4 ng/nl) of PrP-1 morpholino, E-cad morpholino, or both morpholinos were microinjected; the numbers of embryos with arrested gastrulation were given as the percentage of total embryos treated, and statistically analyzed with an unpaired t-test (two-tailed distribution; average ± standard error of the mean [SEM]; n = 3). For rescue experiments, morpholinos at 1.6 ng/nl in 1× Danieau buffer were coinjected with capped mRNAs at 80 pg/nl at a 1:1 ratio in 0.05 M KCl and 0.125% Phenol Red; for overexpression experiments mRNAs were microinjected at or 40 pg/nl. At least 300 embryos were microinjected per experiment (5-nl injection volume) and kept in E3 medium (5 mM NaCl, 0.17 mM KCl, 0.33 mM CaCl2, 0.33 mM MgSO4) at 28 °C; quantitation of phenotypes was carried out for 200 embryos per experiment. Phenotypes were photographed with Zeiss color and black & white Axiocam cameras on an Axioplan 2 microscope using Nomarski optics. Images were further processed with Adobe Photoshop 8.0. Apoptotic cells in fixed embryos were stained with the TUNEL method using the In situ cell death detection kit, AP (Roche), images were acquired on a LUMAR.V12 (whole mounts) or Axioplan 2 (flat mounts) microscopes (Zeiss). Embryos treated with 10% ethanol for 5 min were used as positive apoptotic controls. In addition, embryos were stained with DAPI (100 ng/ml) at room temperature (RT) for 30 min, and examined for their nuclear morphology and the presence of apoptotic bodies.
To analyze gastrulation cell movements, control and morphant embryos at 75% epiboly (8 hpf) were mounted and recorded essentially as previously described [22,40], using an Axioplan 2 microscope under Nomarski optics. For analysis of single-cell behavior, isolated blastomeres were obtained as described below, mounted live in Ringer's solution and similarly recorded. Subsequently, the recordings were analyzed and converted to movies using Axiovision 4.6. To illustrate differences in radial intercalation, selected images from the time-lapse sequence were imported into Adobe Photoshop and pseudocolored to facilitate visualization.
N2a cells were maintained in 10% FCS MEM (Invitrogen), supplemented with l-glutamine, pyruvate, and penicillin-streptomycin at 37 °C and 5% CO2. Cells were grown on poly-lysine–coated coverslips for 24 h prior to transient transfection using Lipofectamine 2000 (Invitrogen). S2 cells were maintained in 10% FCS Schneider's Medium (AMIMED), supplemented with l-glutamine and penicillin-streptomycin at 24 °C. Cells were grown for 24 h prior to transient transfection using Effectene (QIAGEN). Analyses were performed 20 h (N2a) and 24 h (S2) after transfection.
To assay for functional GPI-anchoring, transiently transfected N2a cells were treated with PI-PLC (Roche) as previously described [41], and visualized by fluorescence microscopy. To determine glycosylation states, N2a cells were grown in six-well plates, transiently transfected, lysed, and incubated with N-Glycosidase F (Roche) as reported before [42]; samples were then analyzed by western blot using a monoclonal anti-GFP antibody (Roche). Additionally, N2a cells were transiently transfected with the zebrafish PrP N-glycosylation mutants, lysed, and analyzed by western blot using the anti-GFP antibody.
S2 cells were transfected with mouse EGFP-PrP, DsRed-monomer-PrP, or zebrafish EGFP-PrP-2, and after 24 h, incubated in 0.05% trypsin in PBS for 5 min at RT. After washing, cells were resuspended in 10% FCS Schneider's Medium alone or supplemented with M-20 polyclonal anti-mouse PrP antibody (Santa Cruz Biotechnology), control antibody (mouse IgG; BD Biosciences), or EGTA at the concentrations indicated in the main text. After 2 h, cells were mounted for quantification. Three low-magnification fields of equal cell density were randomly taken from each experiment, and the cell clusters were scored (groups of three or more fluorescent cells). Cell contacts were quantified and given as the percentage of total transfected cells (average ± SEM; n = 3, ∼200 transfected cells per experiment; one-way ANOVA test).
Control zebrafish embryos, as well as embryos injected with lissamine-tagged PrP-1 morpholino or PrP-1 mRNA, were staged and collected in groups of approximately 50 individuals, dechorionated with pronase (2 mg/ml; Sigma) and mechanically dissociated to a single-cell suspension by pipetting for 5 min in Ringer's solution (116 mM NaCl, 2.9 mM KCl, and 5 mM HEPES [pH 7.2]) supplemented with 5 mM EDTA and 0.5 mM EGTA. The dissociated cells were collected by centrifugation, washed twice, and used for western blot analysis, or resuspended in Ringer's solution with or without 1.8 mM CaCl2 to test for Ca+2 dependence. Control, PrP-1 morphant cells, a 1:1 mixture of both, or PrP-1 overexpressing cells were then transferred to microfuge tubes, allowed to aggregate for various periods of time up to 45 min at 28 °C, and mounted for visualization and quantitative evaluation. The number of single cells and cells in aggregates were pooled and given as the percentage of total cells: approximately 200 cells were counted per experiment, eight independent experiments were considered (n = 8, average ± SEM), and statistically analyzed with a one-way ANOVA test. Drosophila S2 and zebrafish embryonic cells were imaged using Plan-NEOFLUAR 20× or 40× objectives and an AxioCam HRm on an Axioplan 2 microscope. Images were further processed with Corel PHOTO-PAINT 11.
Twenty to 30 cells from donor embryos labeled with lissamine-tagged morpholinos were transplanted into unlabeled host embryos essentially as described before [43] using the Transferman NK 2 and CellTram Vario micromanipulators (Eppendorf) on an Axiovert 200 microscope (Zeiss), and monitored on a LUMAR.V12 stereomicroscope before being fixed and mounted for observation. An unrelated morpholino that binds to the 5′ leader sequence of the pCS2+ vector was used as a specificity control.
N2a cells were grown on polylysine-coated coverslips and fixed for 15 min in 4% paraformaldehyde (PFA) 24 hours after transfection. S2 and zebrafish blastomeres cells were immobilized on Alcian blue–coated coverslips, fixed in 4% PFA for 15 min and mounted, or permeabilized with 0.1% Triton X-100 in PBS, probed for 1 h at RT with primary antibody or stained with 1:1,000 Alexa-488 or −568 Phalloidin (Molecular Probes), followed by incubation in 1:1,000 diluted Cy-3 or Alexa-488 conjugated goat anti-rabbit or donkey anti-mouse secondary antibodies (Jackson ImmunoResearch), also for 1 h at RT. The following primary antibodies were used: polyclonal anti-phospho-Src (Tyr416; Cell Signaling Technology) diluted 1:1,000, polyclonal anti-phosphotyrosine antibody (PY350; Santa Cruz Biotechnology) diluted 1:500, monoclonal anti-phosphotyrosine (P-Tyr-100; Cell Signaling Technology) diluted 1:500, prion monoclonal antibody (6H4; Prionics) diluted 1:1,000, and polyclonal anti-Fyn (FYN3; Santa Cruz Biotechnology) diluted 1:500. For analysis of PrP-1 and E-cadherin expression levels, ten control and morpholino embryos were dechorionated, deyolked, lysed, and analyzed by western blot using a purified mouse monoclonal anti–E-cadherin antibody (610182; BD Biosciences) diluted 1:2,000; a purified rabbit polyclonal anti–PrP-1 serum (generated in our lab) diluted 1:4,000; and a goat polyclonal IgG against γ-tubulin (C-20: sc-7396; Santa Cruz Biotechnology) diluted 1:200 as loading control. The anti–PrP-1 serum was not suited for immunofluorescence.
Zebrafish embryos were staged and fixed in 4% PFA overnight at 4 °C. After three washes in PBS-T (0.1% Triton in PBS) and a 1-h incubation in PBS-DT (1% DMSO in PBS-T), they were blocked for 4 h at RT in PBS-DT containing 10% goat serum, incubated with primary antibody (or stained with 1:100 Alexa-488 Phalloidin; Molecular Probes), washed three times in PBS-T, incubated with secondary antibody, and washed three more times in PBS-T. All washes were performed for 5 min at RT; antibody incubations were carried out overnight at 4 °C. The following primary antibodies were used: zebrafish cdh1 rabbit antiserum [20], purified mouse anti–E-cadherin (BD Biosciences), rabbit polyclonal anti–β-catenin (C2206; Sigma), rabbit polyclonal anti-phosphohistone H3[pSer10] (HO412; Sigma), and rabbit polyclonal Rab11 antibody (ab3612; Biozol) at 1:1,000 dilutions; Alexa-488 conjugated goat anti-rabbit secondary antibody at 1:1,000 (Jackson Immunoresearch). Quantification of E-cadherin/Rab11 colocalization was carried out on double-immunostained embryos. For each type of cell (EVL or deep cell [DC]), approximately 25 cells per embryo and approximately five control or morphant embryos were analyzed per experiment. The number of colocalizations per cell were pooled, and the results were statistically analyzed with an unpaired t-test (n = 3, two-tailed distribution). Quantification of immunostained mitotic cells/embryo was carried out on flat mounts of 15 control and 15 morphant embryos, and statistically analyzed with an unpaired t-test (two-tailed distribution; average ± SEM; n = 15). Visualization was carried out on Axioplan 2 and confocal LSM 510 laser-scanning microscopes (Zeiss). Images and fluorescence profiles were obtained with LSM 510 software (Zeiss) and further processed using Corel PHOTO-PAINT 11 and Adobe Photoshop 8.0.
|
10.1371/journal.pntd.0002746 | Development of a Novel, Single-Cycle Replicable Rift Valley Fever Vaccine | Rift Valley fever virus (RVFV) (genus Phlebovirus, family Bunyaviridae) is an arbovirus that causes severe disease in humans and livestock in sub-Saharan African countries. Although the MP-12 strain of RVFV is a live attenuated vaccine candidate, neuroinvasiveness and neurovirulence of MP-12 in mice may be a concern when vaccinating certain individuals, especially those that are immunocompromised. We have developed a novel, single-cycle replicable MP-12 (scMP-12), which carries an L RNA, M RNA mutant encoding a mutant envelope protein lacking an endoplasmic reticulum retrieval signal and defective for membrane fusion function, and S RNA encoding N protein and green fluorescent protein. The scMP-12 underwent efficient amplification, then formed plaques and retained the introduced mutation after serial passages in a cell line stably expressing viral envelope proteins. However, inoculation of the scMP-12 into naïve cells resulted in a single round of viral replication, and production of low levels of noninfectious virus-like particles. Intracranial inoculation of scMP-12 into suckling mice did not cause clinical signs or death, a finding which demonstrated that the scMP-12 lacked neurovirulence. Mice immunized with a single dose of scMP-12 produced neutralizing antibodies, whose titers were higher than in mice immunized with replicon particles carrying L RNA and S RNA encoding N protein and green fluorescent protein. Moreover, 90% of the scMP-12-immunized mice were protected from wild-type RVFV challenge by efficiently suppressing viremia and replication of the challenge virus in the liver and the spleen. These data demonstrated that scMP-12 is a safe and immunogenic RVFV vaccine candidate.
| Rift Valley fever virus (RVFV) is a mosquito-borne zoonotic pathogen, which causes febrile illness, encephalitis and fatal hemorrhagic fever in humans and severe hepatic disease with high mortality and spontaneous abortion rates in ruminants. RVFV is endemic to the African continent. Because many different mosquito species support RVFV replication, the virus has the potential to spread to other areas of the world, such as North and South America, Asia, and Europe and could cause serious public health problems and economic losses. Consistent with this concern, RVFV has caused epidemic disease in the Arabian Peninsula. Currently, there is no approved vaccine suitable for mass vaccination programs of humans. Although the MP-12 strain of RVFV is a live attenuated vaccine candidate, its neuroinvasiveness and neurovirulence in mice are areas of concern, especially when considering the immunization of children and the immunocompromised. In this study, we present a novel MP-12-based, single-cycle replicable RVF vaccine candidate. This vaccine candidate was not neurovirulent in mice and was effective in protecting immunized mice from wild-type RVFV challenge, indicating its potential to be developed as a safe vaccine for use in both humans and livestock.
| Rift Valley fever virus (RVFV), a member of the genus Phlebovirus within the family Bunyaviridae, carries a tripartite, single-stranded and negative–sense RNA genome [1]–[3]. The L RNA encodes the L protein, a viral RNA-dependent RNA polymerase; the M RNA encodes four proteins, including two accessory proteins, the NSm and 78-kDa proteins, and two major viral envelope proteins, Gn and Gc (Gn/Gc); and the S RNA uses an ambisense strategy to express the N protein and an accessory protein, NSs. In infected cells viral RNA synthesis occurs in the cytoplasm, while viral assembly and budding take place at the Golgi apparatus, where Gn/Gc accumulates.
The virus is transmitted by mosquitoes and is maintained in nature, in sub-Saharan Africa, at least in part, by transovarial transmission. RVFV is able to infect various species of mosquitoes [4] and has the potential to spread to other areas of the world. Indeed, RVFV has already spread outside of the African continent to the Arabian Peninsula. The intentional spread of RVFV is also a serious national biosecurity concern. Human infection usually results in febrile illness, but may also cause viral hemorrhagic syndrome, encephalitis, and ocular disease [5]–[7]. RVFV also infects domestic ruminants and causes high mortality and spontaneous abortion rates with severe hepatic disease [8]. Introduction of RVFV to other areas of the world, including North and South America, Asia, and Europe, could cause serious public health problems and economic losses.
RVFV spread can be prevented by the effective vaccination of animals and humans [1]. RVFV is considered to be serologically monotypic [9]–[11], and humoral immunity, particularly neutralizing antibodies that recognize Gn/Gc, is important for protection [12]–[20]. Although a good human RVFV vaccine is urgently needed, there is no approved vaccine that can be adapted to massive vaccination programs. The MP-12 strain of RVFV [21], which was developed by the serial passage of wild-type (wt) RVFV strain ZH548 in the presence of the mutagen 5-fluorouracil, is markedly attenuated and yet retains its immunogenicity [22]–[28]; hence, MP-12 is a promising live vaccine candidate for both human and veterinary use. However, intraperitoneal (i.p.) inoculation of young mice with MP-12 can result in efficient virus replication in the central nervous system (CNS) (J. Morrill et al, unpublished data). Furthermore, i.p. inoculation of SCID mice with MP-12 results in the development of neurological signs and death of all mice [29]. These data suggest that MP-12 can invade the CNS and undergo efficient replication in immunocompromised animals, and may potentially do so in immunocompromised humans as well. However, neurovirulence tests in rhesus macaques show MP-12 to be less neuroinvasive and neurovirulent than acceptable lots of yellow fever or measles vaccine (28). Even so, neuroinvasiveness and neurovirulence is of concern when considering RVFV immunization of the general public, given the diversity of ages, health statuses and genetic backgrounds. Thus, it is important to develop highly immunogenic RVFV vaccines with reduced or no neurovirulence.
To develop a safe and immunogenic RVF vaccine, we have generated a novel, single-cycle replicable MP-12 (scMP-12), which does not cause systemic infection in immunized hosts, while resulting in expression of all viral structural proteins and production of noninfectious, virus-like particles (VLPs) in naïve cells infected with scMP-12. The scMP-12 did not show any sign of neurovirulence after intracranial inoculation into suckling mice, demonstrating its safety. scMP-12-immunized mice elicited neutralizing antibodies and were efficiently protected from wt RVFV challenge by inhibiting wt RVFV replication in various organs and viremia. Our data suggest that scMP-12 has excellent potential to be developed as a safe RVF vaccine.
All mouse studies were performed in facilities accredited by the Association for Assessment and Accreditation of Laboratory Animal Care in accordance with the Animal Welfare Act, NIH guidelines and U.S. federal law. The animal protocol was approved by the UTMB Institutional Animal Care and Use Committee. The wt RVFV ZH501 strain was used in an enhanced ABSL-3 laboratory within the Galveston National Laboratory at UTMB in accordance with NIH guidelines and U.S. federal law.
Vero E6 cells and BSR-T7/5 cells [30], the latter of which stably express T7 RNA polymerase, were maintained as described previously [31], [32]. BHK-21 cells were maintained in minimal essential medium (MEM) α medium (Gibco) supplemented with 5% fetal bovine serum (FBS). The MP-12 strain of RVFV was generated by reverse genetics [31].
A standard PCR-based method, in which pProT7-M encoding antiviral-sense M RNA [31] served as a template, was used to generate pProT7-M-Gn/GcΔ5, which expresses M-Gn/GcΔ5 RNA carrying a deletion between nucleotide positions 3597 and 3611 in the M segment. A Quickchange II site-directed mutagenesis kit (Agilent Technologies) was used to obtain pProT7-M-Gn/GcΔ5-derived mutants, each of which carried an amino acid substitution(s) within a putative fusion peptide. Plasmid pCAGGS-bla-G was constructed by inserting the Not I-EcoR V fragment of pCX4-bsr [33], which contains the encephalomyocarditis virus internal ribosomal entry site and blasticidin-resistant gene, into the Not I and Stu I sites of pCAGGS-G, which carries the entire open reading frame (ORF) of MP-12 M RNA encoding 78-KDa, NSm, Gn and Gc proteins. The sequences of all of the constructs were confirmed not to contain unwanted mutations. MP-12, scMP-12, and MP-12-based, 2-segmented virus replicon particles (VRP) were generated by using a reverse genetics system [31]. Briefly, BSR-T7/5 cells were co-transfected with plasmids encoding the L, N, and Gn/Gc proteins, and anti-viral sense L, M, and S RNAs for MP-12 recovery. scMP-12 recovery was performed by using a similar method with the following modifications: a plasmid expressing the S RNA carrying an N gene and green fluorescent protein (GFP) (S-GFP RNA) was used in place of that expressing the S RNA; a plasmid encoding M-Gn/GcΔ5 RNA with two amino-acid substitutions, F826N and N827A, was used in place of that expressing the M RNA; and a plasmid expressing the MP-12 Gn/Gc optimized for bovine codon usage was used in place of that expressing the MP-12 Gn/Gc to prevent or minimize homologous RNA recombination events between expressed mRNA encoding Gn/Gc and the replicating M RNA mutant. For VRP recovery, a plasmid expressing S-GFP RNA was used in place of the plasmid encoding the S RNA and the plasmid encoding the M RNA was eliminated. Culture fluid was collected at 5, 10 and 10 days post transfection for MP-12, scMP-12, and VRP, respectively.
Vero E6 cells were transfected with pCAGGS-bla-G, and incubated in the presence of 20 µg/ml of blasticidin from 1 day post-transfection. After obtaining blasticidin-resistant cell clones by limiting dilution, each cell clone was tested for Gn protein expression by indirect immunofluorescence with an anti-Gn monoclonal antibody (R1-4D4) [34], and a cell clone expressing highest levels of Gn was selected and designated as Vero-G cells.
A standard plaque assay was used to determine the infectivity of MP-12 [31]. For determining the infectivity of scMP-12 and VRP, Vero-G cells in 6-well plates were inoculated with 400 µl of serially diluted samples and incubated for 1 h at 37°C. After removal of the inocula, cells were incubated with MEM containing 0.6% Tragacanth gum (MP Biomedicals), 5% FBS, and 5% tryptose phosphate broth at 37°C. After 3 days incubation, cells were washed with phosphate-buffered saline (PBS) and fixed with PBS containing 4% paraformaldehyde for 20 min at room temperature. After removing paraformaldehyde and overlays, the cells were permeabilized with 0.1% Triton-X100 and incubated with anti-N rabbit polyclonal antibody, which was generated by injecting a purified, bacterially-expressed fusion protein consisting of glutathione-S-transferase and full-length MP-12 N protein into rabbits, followed by incubation with horseradish peroxidase-conjugated, anti-rabbit IgG antibody. The plaques were visualized with Nova RED peroxidase substrate (Vector Laboratories, Burlingame, CA). This modified plaque assay was also used for observing plaque morphologies of MP-12 in Vero-G cells.
The cell fusion assay was performed as previously described [35], [36] with some modifications. Briefly, BSR-T7/5 cells were co-transfected with plasmids encoding the Venus, N, and L proteins, and M-Gn/GcΔ5 RNA or M-Gn/GcΔ5 RNA with single amino acid substitutions, and incubated at 37°C for 24 h. To initiate cell fusion, the cells were washed with Mg2+- and Ca2+-containing acidic PBS (pH adjusted to 5.2 with citric acid) and treated with the acidic PBS for 5 min., and then incubated in complete medium at 37°C for 60 min. GFP signals in the cells were observed under a fluorescence microscope (Zeiss).
BSR-T7/5 cells were co-transfected with plasmids encoding the N and L proteins, and M-Gn/GcΔ5 RNA or its mutant. Twenty-four hours after transfection, cells were fixed with 4% paraformaldehyde and permeabilized with 0.1% Triton-X100, or not permeabilized. Cells were incubated with the primary monoclonal antibody that recognizes Gn (R1-4D4) or Gc (R1-5G2) [37] for 1 h at room temperature and with the Alexa-594-conjugated secondary antibody for 1 h at room temperature, and observed under a fluorescence microscope.
Cells were harvested by using a cell scraper and washed with PBS. After incubation of the harvested cells on ice for 20 min in cell lysis buffer (20 mM Tris-HCl, 150 mM NaCl, 1% Triton X-100), the cell lysate was centrifuged at 2,000 rpm for 3 min by using a microcentrifuge. The resultant supernatant was mixed with the same amount of 2× sample buffer and boiled for 5 min. Equal amounts of samples were subjected to SDS-polyacrylamide gel electrophoresis. Proteins were electroblotted onto polyvinylidene difluoride membranes (Millipore). After blocking the membrane with 1% bovine serum albumin for 1 h, the membranes were incubated with the primary antibody for 1 h at room temperature. After incubation with the secondary antibody for 1 h at room temperature, the blots were developed by using an ECL kit (GE Healthcare). Anti-MP-12 mouse polyclonal antibody [31] was used to detect the virus-specific proteins.
Total RNAs were extracted by using TRIzol reagent (Invitrogen) and subjected to Northern blot analysis as described previously [38]. Viral-sense-specific, digoxigenin-labeled RNA probes [31] and a digoxigenin system (Roche) were used for the detection of viral RNAs. The L RNA probe hybridizes with viral-sense L RNA at nucleotide positions 19–756, the M RNA probe at nucleotide positions 1297–2102, and the S RNA probe at nucleotide positions 39–776 from the 3′ ends of the viral-sense RNA segments. The probe that hybridizes with anti-viral sense S RNA binds at nucleotide positions 39–776 from the 5′ end of the anti-viral-sense S RNA segment.
Culture medium harvested from plasmid-transfected cells or scMP-12-infected cells was clarified by centrifugation at 3,000 rpm for 15 min by using a tabletop centrifuge. The clarified supernatant was layered on top of a step sucrose gradient consisting of 20, 30, 50, and 60% sucrose (wt/vol) and centrifuged for 3 h at 26,000 rpm at 4°C using a Beckman SW28 rotor [32]. The particles at the interface of 30 and 50% sucrose were collected, diluted and subjected to a second sucrose gradient centrifugation consisting of 20, 30, 50, and 60% sucrose for 18 h at 4°C. The particles at the interface of 30 and 50% sucrose were collected and pelleted down through a 20% sucrose cushion at 38,000 rpm for 2 h at 4°C using a Beckman SW41 rotor.
scMP-12 was serially passaged 10 times in Vero-G cells under the following three conditions for each passage: inoculation without sample dilution and harvest at 4 days p.i., inoculation after 10 times sample dilution and harvest at 5–6 days p.i., and inoculation after 100 times sample dilution and harvest at 7 days p.i. We visually inspected for an increase in the number of GFP-positive cells every day. Each of the culture fluids collected was also inoculated into Vero E6 cells, and the GFP signal was examined daily up to 5 days p.i.
Two-day-old CD1 mice were intracranially inoculated with 104 PFU of MP-12, scMP-12, or the same volume of Hank's balanced salt solution (HBSS). We monitored the mice for survival for 21 days.
CD1 mice (5-week-old females) were intramuscularly immunized with 104 PFU of MP-12, 105 PFU of MP-12, 105 PFU of scMP-12, or 105 PFU of VRP. Thirty-six days later, blood was collected from the retro-orbital venous plexus of the mice. Forty days after vaccination, the immunized mice were challenged subcutaneously with 103 PFU of the virulent RVFV strain ZH501, which was equivalent to approximately 1,000 times the 50% minimal lethal dose (LD50). The animals were observed for survival and clinical signs of disease for 21 days post-challenge. To determine the effect of immunization on virus replication, sera and specimens of liver, spleen and brain were harvested from randomly selected animals at 3, 6, 9 and 11 days post-challenge. Sera and 10% tissue homogenates were tested for virus presence and titer in Vero E6 cells, as previously described (39).
Serum neutralizing antibody titers were determined by using an 80% plaque-reduction neutralization test (PRNT80), as previously described [39].
We designed the scMP-12 system as shown in Fig. 1A. scMP-12 carries a membrane-fusion defective mutant of Gn/Gc and is rescued by using a modified MP-12 reverse genetics system [31], in which BSR-T7/5 cells stably expressing T7 polymerase [30] are co-transfected with three RNA-expression plasmids expressing the L RNA, a mutant M RNA encoding a membrane-fusion defective mutant of Gn/Gc, and a S-GFP RNA encoding the N and GFP proteins, as well as three protein expression plasmids encoding the L, N, and Gn/Gc proteins. The scMP-12 that is produced is infectious due to the presence of Gn/Gc and undergoes amplification in Vero-G cells stably expressing Gn/Gc. Inoculation of the amplified scMP-12 into naïve cells results in viral RNA synthesis, expression of viral proteins, including L, N and the fusion-defective Gn/Gc, and production of noninfectious VLPs containing the fusion-defective Gn/Gc. In immunized hosts, scMP-12 undergoes single cycle replication in infected cells, resulting in the intracellular accumulation of all of the viral structural proteins and the production of noninfectious VLPs; scMP-12 particles in the inoculum, viral proteins accumulated in scMP-12-infected cells and released noninfectious VLPs all serve as immunogens to elicit immune responses to RVFV proteins. Due to its characteristic single-cycle replication, it is highly unlikely that the scMP-12 can cause systemic infection or invade the CNS of immunized animals or humans.
To isolate fusion-defective Gn/Gc mutants suitable for scMP-12, we first developed a cell-to-cell membrane fusion assay. Phlebovirus glycoprotein-induced, virus-cell membrane fusion requires a low pH (∼pH 5.4) environment [35]. Exposure of cells expressing RVFV Gn/Gc to low pH conditions does not induce cell-to-cell fusion due to the absence of Gn/Gc at the plasma membrane; RVFV Gn/Gc accumulates at the Golgi apparatus and the endoplasmic reticulum (ER) in infected cells and in expressed cells. Phleboviruses have an ER retrieval signal of ∼5 amino acids in the cytoplasmic tail of Gc [40], and removal of this signal in the Gc of Uukuniemi virus (a Phlebovirus) results in an accumulation of expressed Gn/Gc at the Golgi apparatus and plasma membrane [40]. Likewise, mutant MP-12 Gn/Gc lacking the terminal C-terminal 5-amino-acid residues of the Gc (Gn/GcΔ5) primarily accumulated at the Golgi apparatus when expressed, and some mutant glycoprotein was translocated to the plasma membrane (Fig. 2A). Exposure of the cells expressing Gn/GcΔ5, but not those expressing wt Gn/Gc, to low pH conditions induced cell-to-cell membrane fusion; fusion was not observed at neutral pH conditions for cells expressing Gn/GcΔ5 (Fig. 2B). These data suggest that Gn/GcΔ5 protein that localized to the plasma membrane was fusion-competent only under low pH conditions.
We sought to generate fusion-defective Gn/Gc mutants by altering amino acids in the putative fusion peptide, which was previously predicted by computational studies and structural analysis [41], [42]. Alignment of the predicted fusion peptide sequences of several Phleboviruses revealed the presence of a highly conserved cysteine residue at position 825 (C825), which is involved in a disulfide bond in the Gc [41], and a phenylalanine residue at position 826 (F826) (Fig. 2C). Because hydrophobic residues are important for the insertion of fusion peptides into the cell plasma membrane [43], we tested the fusion competence of a series of Gn/GcΔ5-derived mutants, in which the F826 was changed to a hydrophilic residue, or its surrounding hydrophobic residues and hydrophilic residues were changed to hydrophilic residues or hydrophobic residues, respectively (Fig. 2D). While the V828N and P830N mutants retained fusion activity, the other mutants lost such activity (Fig. 2D). Anti-Gn monoclonal antibody recognized all of the Gn/Gc mutants, while the anti-Gc monoclonal antibody R1-5G2 failed to detect the W821N and C825A mutants, implying an alteration of the Gc conformation occurred from these mutations. From the C823A, F826N and N827A mutants, all of which lost fusion activity and were detected by R1-5G2, we selected F826N and N827A mutants for subsequent studies.
Because development of scMP-12 is aimed at improving RVF vaccine safety, it is important to prevent the generation of infectious viruses in scMP-12-immunized hosts, as well as during scMP-12 preparation in cell culture. Hence, we tested several M RNA mutants, each encoding Gn/GcΔ5, with different combinations of fusion peptide mutations and chose an M RNA mutant encoding Gn/GcΔ5 with the F826N and N827A mutations (scMP-12 M RNA)(Fig. 1B) for scMP-12 preparation primarily due to its excellent genetic stability. BSR-T7/5 cells were co-transfected with three protein expression plasmids expressing the L, N, and Gn/Gc proteins, and three RNA expression plasmids encoding the L, scMP-12 M, and S-GFP RNAs. The GFP signal generated in scMP-12-infected cells facilitated the monitoring of scMP-12 replication. We also generated a VRP, an MP-12-based virus replicon particle (VRP) carrying only the L and S-GFP RNAs. Because other groups have reported the generation of a VRP (also called RVFV replicon particles) carrying the L and S-GFP RNAs derived from wt RVFV [44], [45], we refer to the wt virus-based VRP as VRPwt to distinguish between it and the MP-12-based VRP used in this study. MP-12 was rescued as previously described [31], and used as a positive control. Culture fluids from MP-12 samples were collected at 5 days post-transfection, while those from the scMP-12 and VRP samples were collected at 10 days post-transfection; these samples were defined as P0 samples.
To amplify and titrate the scMP-12 samples, we generated Vero-G cells stably expressing MP-12 Gn/Gc, and found the expression levels of Gn/Gc in Vero-G cells to be roughly one-fourth of the levels for MP-12-infected Vero cells at 12 h post-inoculation (p.i.) (Fig. 3A). Like MP-12-infected Vero E6 cells, Gn and Gc signals primarily accumulated in perinuclear regions of Vero-G cells (Fig. 3B). We independently inoculated the P0 samples of scMP-12 and VRP into Vero-G cells and obtained passage 1 (P1) samples after 10 days p.i. These P1 samples were predominantly used for subsequent studies. MP-12, scMP-12 and VRP formed large, medium and small plaques, respectively, in Vero-G cells, in which plaques were visualized by anti-N protein antibodies (Fig. 4A). Inoculation of MP-12, scMP-12 or VRP into Vero-G cells at a multiplicity of infection (MOI) of 0.05 showed efficient MP-12 replication with maximum titers ∼108 PFU/ml at 3 days p.i. (Fig. 4B). scMP-12 replicated to ∼106 PFU/ml at 2–3 days post-infection, whereas the titers of the VRP were roughly 5–10 times lower than those of scMP-12 (Fig. 4B). As expected, we observed efficient accumulation of the three viral RNA segments in Vero-G cells infected with MP-12 or scMP-12, and L and S-GFP RNAs in VRP-infected Vero-G cells (Fig. 4C). We purified the particles produced from Vero-G cells infected with scMP-12, MP-12 or VRP by sucrose gradient centrifugation. Western blot analysis of purified particles using anti-MP-12 antibody showed the presence of Gn/Gc and N proteins in all samples (Fig. 4D). The origin is unknown for two bands found, one that migrated more slowly and the other faster than the Gn/Gc of the MP-12 sample in the gel. Northern blot analysis of viral RNAs extracted from the purified particles showed packaging of three viral RNAs in MP-12 and scMP-12 samples and that of the L and S-GFP RNAs in the VRP sample (Fig. 4D); the abundance of each of the viral RNAs was roughly proportional to the titers of MP-12, scMP-12 and VRP at day 3 (Fig. 4B). These data show that the scMP-12 underwent efficient replication and amplification in Vero-G cells.
To examine scMP-12 replication in naïve cells, we inoculated MP-12, scMP-12 or VRP into naïve BHK cells and examined the accumulation of viral proteins and RNAs (Fig. 5A). Efficient accumulation of the Gn/Gc and N proteins occurred in MP-12-infected cells. Accumulation of the N and Gn/Gc proteins also occurred in scMP-12-infected cells, with lower levels of Gn/Gc accumulation as compared to MP-12-infected cells. VRP-inoculated cells showed an accumulation of the N protein, but not the Gn/Gc protein. Northern blot analysis showed that the three viral RNAs replicated in MP-12-infected cells and in scMP-12-infected cells, and L and S-GFP RNAs replicated in VRP-infected cells. An RNA probe that specifically binds to anti-viral-sense S RNA clearly demonstrated N mRNA synthesis in these RNA samples (Fig. 5A, right panels). Thus, the scMP-12 underwent efficient viral RNA synthesis and viral protein accumulation in infected naïve cells.
We next purified the particles released from scMP-12-infected BHK cells by sucrose gradient centrifugation and detected viral proteins in the purified particles (Fig. 5B). The purified particles produced from MP-12-infected cells and VRP-infected cells served as a positive control and a negative control, respectively. Western blot analysis showed the production of MP-12 particles in the positive control by demonstrating the N and Gn/Gc proteins. No Gn/Gc signal was detected in the VRP sample, whereas the scMP-12 sample showed a low level of Gn/Gc signal. Both scMP-12 and VRP samples showed low levels of the N protein signal. Because synthesis of the Gn/Gc proteins did not occur in VRP-infected cells, it is highly unlikely that the N protein in the VRP sample represents released VRP. Continuous sucrose gradient centrifugation of culture fluid of MP-12-infected cells showed sedimentation of N protein with the purified virions as well as to lower sucrose density fractions [46], suggesting the release of N protein which is not associated with virus particles from infected cells. Furthermore, release of N protein not associated with viral envelope proteins was reported in studies of RVFV VRP [45] and Crimean-Congo hemorrhagic fever virus [47]. Hence, the N protein signal in the VRP sample most probably represents the N protein that was not associated with virus particles. Likewise, most of the N signal in the scMP-12 sample was probably derived from the non-VLP-associated N protein. Nonetheless, the Gn/Gc signal in the scMP-12 sample suggests the occurrence of low levels of VLP production from scMP-12-infected naïve cells. Inoculation of supernatant from MP-12-infected BHK cells, but not from scMP-12-infected BHK cells or VRP-infected BHK cells, into fresh BHK cells resulted in viral RNA synthesis (Fig. 5C), demonstrating that the VLP produced from scMP-12-infected naïve cells was not infectious.
To evaluate the genetic stability of scMP-12, we performed 10 serial passages of scMP-12 in Vero-G cells under three different conditions, as described in Materials and Methods, and tested the generation of infectious viruses that undergo multiple cycles of replication in naïve cells. Multiple cycles of the scMP-12 amplification in Vero-G cells resulted in an increase in the numbers of GFP-positive cells during incubation in each passage, whereas an increase in the numbers of GFP-positive cells did not occur after inoculation of any of the passage samples in Vero cells, suggesting the absence of infectious viruses in all of the passaged samples. Also plaque assays using Vero cells did not show the presence of infectious viruses in any of the samples. Sequence analysis of the PCR products of scMP-12 M RNA showed that scMP-12 retained the introduced mutations after 10 passages under the three different conditions. These results demonstrate that scMP-12 stably retained the introduced mutations.
We tested the neurovirulence of scMP-12 by intracranially inoculating 1.0×104 PFU of scMP-12 into 2-day-old CD1 mice and monitoring for survival and clinical signs for 21 days p.i. As controls, HBSS and the same titer of MP-12 were inoculated. All MP-12 infected mice died by 3 days p.i., whereas all mice inoculated with scMP-12 or HBSS survived and did not show any clinical signs of disease (Fig. 6), demonstrating the absence of detectable levels of neurovirulence in scMP-12.
We intramuscularly inoculated 5-week-old female CD1 mice once with 105 PFU of scMP-12 and determined the PRNT80 titers at 36 days p.i. As controls, mice were inoculated with 105 PFU of VRP, 105 PFU of MP-12, 104 PFU of MP-12, or HBSS (Fig. 7). HBSS-inoculated mice had no detectable neutralizing antibody titers, while mice inoculated with 105 PFU of MP-12 and 104 PFU of MP-12 had a mean PRNT80 titer of 1∶1,477 and 1∶310, respectively. The mean PRNT80 titers of the mice immunized with 105 PFU of scMP-12 and 105 PFU of VRP were 1∶238 and 1∶38, respectively; the difference in the PRNT80 titers was statistically significant. Thus, the mice immunized with 105 PFU of scMP-12 elicited neutralizing antibody titers that were statistically higher than those immunized with 105 PFU of VRP and were comparable to those immunized with 104 PFU of MP-12.
We next tested whether scMP-12 immunization protects mice from wild-type RVFV challenge. Five-week-old female CD1 mice were intramuscularly inoculated once with 105 PFU of scMP-12, 105 PFU of VRP, 105 PFU of MP-12, 104 PFU of MP-12, or HBSS. At 40 days post-immunization, the mice were challenged subcutaneously with 1.0×103 PFU of the ZH501 strain of RVFV and their survival was monitored for 21 days p.i. (Fig. 8). All HBSS-inoculated mice died by 10 days p.i., whereas all mice immunized with 105 PFU of MP-12 survived. Most of the mice immunized with 104 PFU of MP-12 or 105 PFU of scMP-12 survived, yet 1 of the 19 MP-12-immunized mice died at day 11, and 3 of the 29 scMP-12-immunized mice died, one at day 10 and two at day 19, respectively. In contrast, 45% of the VRP-immunized mice died by day 12 p.i., demonstrating that scMP-12 immunization protected most of the mice from wt RVFV challenge, and scMP-12-induced protection was better than the VRP-induced protection.
To study the extent to which scMP-12-induced immune responses suppressed wt virus replication upon challenge, HBSS-inoculated mice and mice immunized once with 105 PFU scMP-12, 105 PFU VRP, 105 PFU MP-12, or 104 PFU MP-12 were challenged with the ZH501 strain of RVFV, as described above, and the virus titers in serum, liver, spleen and brain were determined at days 3, 6, 9 and 11 post-challenge (Fig. 9). At day 3 p.i., 4 out 5 HBSS-inoculated mice had >105 PFU/ml of viremia, and one and three mice showed virus replication in the liver and the spleen, respectively. Efficient virus replication in the brain also occurred in HBSS-inoculated mice from days 5 to 9 p.i. In contrast, mice immunized with VRP or MP-12 showed neither viremia nor virus replication in the liver, spleen or brain. scMP-12 immunization also prevented viremia and virus replication in the liver and spleen, while two mice, one having no detectable PRNT80 titer and the other having a PRNT80 titer of 1∶20, showed virus replication in the brain at day 9 p.i.
By using a membrane fusion assay and newly established Vero-G cells, we generated scMP-12 and tested its potential as a safe and immunogenic RVFV vaccine. scMP-12 amplified efficiently in Vero-G cells and stably retained the introduced mutations in ten serial passages in this cell line under three different experimental conditions. In infected naïve cells, scMP-12 underwent efficient viral RNA synthesis and accumulated viral proteins, including Gn/Gc, and produced low levels of non-infectious VLPs. The scMP-12 did not show any sign of neurovirulence after intracranial inoculation into 2-day-old mice, demonstrating excellent safety. scMP-12 immunization in mice induced neutralizing antibodies, whose titers were higher than those in VRP-immunized mice, and protected most of them from wt RVFV challenge by suppressing viremia and wt RVFV replication in the liver and the spleen. Taken together, we consider that scMP-12 has an excellent potential to be developed as a novel safe RVFV vaccine.
We examined effects of mutations within the putative fusion peptide for membrane fusion (Fig. 2). The crystal structure of RVFV Gc suggested that V828 and hydrophobic residues W821 and F826 within the putative fusion peptide serve as a membrane anchor during the pre-fusion step [41]. By substituting the hydrophobic residues for hydrophilic residues in the putative fusion peptide, we experimentally demonstrated that an F826N mutation, but not V828N mutation, abolished membrane fusion. Anti-Gc monoclonal antibody did not recognize Gc carrying W821N, and possibly a van der Waals interaction between W821 and F826 was disrupted by this mutation, leading to Gc structural alteration [41]. C825 is highly conserved among Phleboviruses (Fig. 2C), and the C825A mutant was defective for the fusion function. Because C825 is involved in a disulfide bond in Gc [41], and the anti-Gc monoclonal antibody did not recognize the C825A mutant, a lack of fusion function in this mutant was probably due to the structural alteration of Gc. Because other Phleboviruses also encode the ER retrieval signal in the Gc cytoplasmic tail, development of similar membrane fusion assays for other Phleboviruses would be possible. Recently, others also reported the utility of the RVFV membrane fusion assay that uses a plasmid transfection method [48]. Experiments using such fusion assays, which employ a conventional plasmid transfection, will be valuable for further understanding of the membrane fusion mechanism in Phleboviruses, and identification and evaluation of antivirals that suppress viral membrane fusion activity [48].
The data that RVFV spread can be prevented by effective vaccination of animals and humans [1] and that neutralizing antibodies, the majority of which recognize Gn/Gc protein, play a critical role in protection [12]–[20] led to development of several different types of RVFV vaccine candidates that primarily aim to elicit high titers of neutralizing antibodies. Formalin-inactivated RVFV vaccine requires several immunizations to induce and maintain protective immunity [49], [50]. In contrast, several attenuated RVFV mutants, including MP-12, MP-12-derived mutants carrying a modified cellular gene in place of the NSs gene [51], and a wt RVFV-derived avirulent mutant lacking NSs and NSm genes, both of which are viral virulence factors [52]–[55], demonstrated excellent protective immunogenicity against wt RVFV after a single immunization of animals [56]. Examples of other vaccine candidates are VLPs [57]–[59], recombinant vaccinia viruses encoding Gn and Gc proteins [29], alphavirus encoding the Gn protein [60], alphavirus replicon encoding the Gn protein [61], and a soluble ectodomain of the Gn protein [62]. Most of these vaccine candidates have used multiple dose immunization protocols to confer complete protection to immunized rodents against wt RVFV challenge. Single immunization of mice with scMP-12 (Fig. 8), VLP expressing low levels of viral N protein in infected cells [59] or VRPwt expressing both L and N proteins in inoculated cells [44], [63] showed good protection of the immunized mice from lethal wt RVFV challenge. This finding may imply that the expression of the N protein and probably also the L protein in immunized animals facilitated development of strong protective immune responses. In addition, viral-replicating, single-stranded RNA and the incoming RNA virus nucleocapsids activate the innate immune system through interaction with the host pattern recognition receptor, e.g. RIG-I [64]–[69], and potentiates the adaptive immune responses [70]. Moreover, viral RNAs in virus particles have an adjuvant effect for augmenting host-adaptive immune responses through a Toll-like receptor 7 signaling pathway in dendritic cells [71], [72]. Therefore, it is likely that incoming nucleocapsids of scMP-12, intracellular viral RNAs accumulated in scMP-12-infected naïve cells, and viral RNAs in the released VLPs all contributed to enhancement of the host immune response, making scMP-12 highly immunogenic. Importantly, scMP-12 was more immunogenic than VRP (Fig. 7), and the scMP-12-immunized mice were protected from wt RVFV challenge more efficiently than the VRP-immunized mice (Fig. 8); hence, the expression of Gn/Gc in cells supporting scMP-12 replication and viral RNA containing VLPs produced by cells in which scMP-12 replicated augmented the protective immune response.
A lack of neurovirulence and the characteristic single-cycle replication property of scMP-12 demonstrate that scMP-12 is superior to MP-12 in safety, as MP-12 killed all of the 2-day-old mice following intracranial inoculation, whereas scMP-12 was less immunogenic than MP-12; neutralizing antibody titers in mice immunized with 105 PFU of scMP-12 were comparable to those immunized with 104 PFU of MP-12 and lower than those immunized with 105 PFU of MP-12. Improvement of scMP-12 immunogenicity may be possible by generating a scMP-12 variant that produces a high abundance of VLPs following scMP-12 replication. Because substitution of several histidines in RVFV Gc with alanine inhibits membrane fusion activity but does not interfere with virion assembly [73], the efficient production of noninfectious VLPs may occur in cells supporting replication of scMP-12 variants carrying some of these mutations.
The finding of efficient scMP-12 amplification in Vero-G cells suggests that a scMP-12-based vaccine stock can be prepared in Vero-G cells or their equivalent without plasmid transfection, thereby allowing the production costs of the scMP-12-based vaccine to be comparable to those for MP-12. scMP-12 and VRP produced plaques in Vero-G cells, showing the utility of Vero-G cells for easy titration and characterization of RVFV mutants lacking functional Gn/Gc proteins. We noted that MP-12 replicated roughly 10 times better in Vero-G cells than in Vero E6 cells (Fig. 4B and [31]), which led us to suggest that higher levels of intracellular Gn/Gc accumulation augments MP-12 production. Likewise, an increase in the abundance of intracellular Gn/Gc in scMP-12-replicating cells may also enhance scMP-12 titers. Hence, the development of another Vero cell clone, in which expression levels of Gn/Gc are comparable to those in MP-12-infected Vero cells, would contribute to mass immunization programs using an scMP-12-based vaccine. The absence of infectious virus after 10 serial passages of scMP-12 in Vero-G cells under three different conditions demonstrated that homologous RNA recombination that can eliminate the mutations in scMP-12 M RNA did not occur between replicating scMP-12 M RNA and expressed mRNA encoding Gn/Gc in Vero-G cells, further indicating the utility and safety of Vero-G cells for preparation of the scMP-12-based vaccine. Lastly, we found that scMP-12 replicated ∼10 times better than did the VRP in Vero-G cells (Fig. 4). These data were consistent with the notion that M RNA serves important roles in viral RNA co-packaging [32].
Expression of GFP from the S-GFP RNA of scMP-12 facilitated easy monitoring of scMP-12 replication and generation of infectious viruses in scMP-12 preparations. VRPwt also used S-GFP-type RNA for easy monitoring of VRPwt replication [44], [45]. However, vaccines encoding a foreign reporter gene, such as GFP, may not be appropriate for human use. Therefore, before we can develop a scMP-12-based human vaccine, it is necessary to test the replication competence, safety, and immunogenicity of scMP-12-based vaccine candidates lacking the NSs gene or of those carrying RVFV Clone 13-type S RNA lacking ∼70% of the NSs gene [74].
Our study was primarily aimed at the development of a safe and immunogenic human RVF vaccine, yet scMP-12 may be further developed as a veterinary vaccine. Others have reported that MP-12 is teratogenic in some cases [75]. Considering that scMP-12 only undergoes a single cycle of replication, it is unlikely cause disease in immunized animals. Vaccines that are compatible with a differentiation of infected and vaccinated animals (DIVA) are suitable for use as animal vaccines. Examples of replication-competent RVF DIVA vaccine candidates are RVFV Clone 13 lacking ∼70% of the NSs gene [74], MP-12 lacking NSm, which elicited high titers of neutralizing antibodies in sheep and calves [76], and wt RVFV-derived mutant virus lacking NSm and NSs, which induced protective immunity in immunized sheep [56]. The data that scMP-12, which lacks an NSs gene, protected immunized mice from wt RVFV challenge (Fig. 8) and that VRPwt, which also lacks an NSs gene, can induce protective immunity in sheep [63] indicate a potential for a scMP-12-based DIVA vaccine to reduce the incidence of RVF among humans and animals and to control this important pathogen [8].
|
10.1371/journal.pcbi.1003284 | A Brain-Machine Interface for Control of Medically-Induced Coma | Medically-induced coma is a drug-induced state of profound brain inactivation and unconsciousness used to treat refractory intracranial hypertension and to manage treatment-resistant epilepsy. The state of coma is achieved by continually monitoring the patient's brain activity with an electroencephalogram (EEG) and manually titrating the anesthetic infusion rate to maintain a specified level of burst suppression, an EEG marker of profound brain inactivation in which bursts of electrical activity alternate with periods of quiescence or suppression. The medical coma is often required for several days. A more rational approach would be to implement a brain-machine interface (BMI) that monitors the EEG and adjusts the anesthetic infusion rate in real time to maintain the specified target level of burst suppression. We used a stochastic control framework to develop a BMI to control medically-induced coma in a rodent model. The BMI controlled an EEG-guided closed-loop infusion of the anesthetic propofol to maintain precisely specified dynamic target levels of burst suppression. We used as the control signal the burst suppression probability (BSP), the brain's instantaneous probability of being in the suppressed state. We characterized the EEG response to propofol using a two-dimensional linear compartment model and estimated the model parameters specific to each animal prior to initiating control. We derived a recursive Bayesian binary filter algorithm to compute the BSP from the EEG and controllers using a linear-quadratic-regulator and a model-predictive control strategy. Both controllers used the estimated BSP as feedback. The BMI accurately controlled burst suppression in individual rodents across dynamic target trajectories, and enabled prompt transitions between target levels while avoiding both undershoot and overshoot. The median performance error for the BMI was 3.6%, the median bias was -1.4% and the overall posterior probability of reliable control was 1 (95% Bayesian credibility interval of [0.87, 1.0]). A BMI can maintain reliable and accurate real-time control of medically-induced coma in a rodent model suggesting this strategy could be applied in patient care.
| Brain-machine interfaces (BMI) for closed-loop control of anesthesia have the potential to enable fully automated and precise control of brain states in patients requiring anesthesia care. Medically-induced coma is one such drug-induced state in which the brain is profoundly inactivated and unconscious and the electroencephalogram (EEG) pattern consists of bursts of electrical activity alternating with periods of suppression, termed burst suppression. Medical coma is induced to treat refractory intracranial hypertension and uncontrollable seizures. The state of coma is often required for days, making accurate manual control infeasible. We develop a BMI that can automatically and precisely control the level of burst suppression in real time in individual rodents. The BMI consists of novel estimation and control algorithms that take as input the EEG activity, estimate the burst suppression level based on this activity, and use this estimate as feedback to control the drug infusion rate in real time. The BMI maintains precise control and promptly changes the level of burst suppression while avoiding overshoot or undershoot. Our work demonstrates the feasibility of automatic reliable and accurate control of medical coma that can provide considerable therapeutic benefits.
| Medically-induced coma (also referred to as medical coma) is a drug-induced state of profound brain inactivation and unconsciousness used to treat refractory intracranial hypertension and status epilepticus, i.e., epilepsy that is refractory to standard medical therapies [1]–[3]. Following a traumatic brain injury, an anesthetic drug such as a barbiturate or propofol, is administered continuously to provide brain protection by decreasing the cerebral metabolism and blood flow, and thereby, intracranial hypertension [2]. In the treatment of status epilepticus the anesthetic is administered to directly inhibit activity in the seizure foci [3]. For treating both refractory intracranial hypertension and status epilepticus, the state of medical coma is achieved by continually monitoring the patient's brain activity with the electroencephalogram (EEG) and titrating the anesthetic drug infusion rate to maintain a specified level of burst suppression. Burst suppression is an EEG pattern characterized by intervals of electrical bursts that alternate with isoelectric or quiescent intervals termed suppressions [4], [5] and is an EEG marker of profound brain inactivation. In most cases, once burst suppression is achieved, it can be controlled by decreasing or increasing the infusion rate of the anesthetic to decrease or increase the suppression level.
No guidelines have been set to define what level of burst suppression should be achieved to maintain a medical coma [3]. A common practice is for the intensive care unit team to agree upon a target level of burst suppression, monitor continually the EEG and adjust manually the infusion rate of the anesthetic to maintain the target level. In most cases, the medical coma is required for at least 24 hours and frequently longer. It is not realistic to expect intensive care unit staff to maintain reliable and accurate control of a patient's brain state for such a long period by manually changing the infusion rate of the anesthetic in response to changes in the EEG observed in the bedside monitor. A more rational approach would be to define numerically a target level of burst suppression and implement a computer controlled system or a brain-machine interface (BMI) that monitors the actual level of burst suppression based on the brain's EEG activity and adjusts the rate of the anesthetic infusion pump as needed in real time to maintain the target level.
When used to control the delivery of anesthetic drugs, BMIs are often termed closed loop anesthetic delivery (CLAD) systems. During the last 60 years considerable work has been done on the development of CLAD systems for maintenance of general anesthesia and sedation (see Discussion). Interest in CLAD systems has grown driven by attempts to design more efficient, cost-effective ways to administer anesthesia care. To date, no CLAD system has been developed to manage medical coma. Systems to control burst suppression have only been studied in rodent models. Vijn and Sneyd implemented a CLAD system in a rodent model to establish a paradigm for testing new anesthetics [6]. Cotten and colleagues used the Vijn and Sneyd paradigm to study new etomidate-derived anesthetics in a rodent model [7]. Both studies reported average control results rather than results for individual animals and controlled constant target levels of burst suppression rather than time-varying target levels. Here we present a BMI using a stochastic control framework for control of time-varying burst suppression target trajectories in individual rodents. Our study uses a rodent model to establish the feasibility of automatic control of burst suppression as a way to eventually achieve real-time control of medical coma for therapeutic purposes in humans. We show that for individual rodents the BMI enables accurate maintenance of multiple desired target levels within the same experimental session, enables prompt transitions between target levels without overshoot or undershoot, and allows specific constraints to be formally imposed over the infusion rates or the vital states (see Discussion).
The presented BMI applies an EEG-guided, closed-loop infusion of propofol to control the level of burst suppression in medically-induced coma in a rodent model using a stochastic control framework. In this framework, we use the concept of the burst suppression probability (BSP) to define the brain's instantaneous probability of being in the suppressed state and quantify the burst suppression level. We use a two-dimensional linear compartment model to characterize the effect of propofol on the EEG. For each animal, we estimate the parameters of the compartment model by nonlinear least-squares in an experiment prior to initiating control. The BMI consists of two main components: an estimator and a controller. We derive a two-dimensional state-space algorithm to estimate the BSP in real time from the EEG thresholded and segmented into a binary time-series. Taking the BSP estimate as the control signal, we derive controllers using both a linear-quadratic-regulator (LQR) and a model predictive control strategy. We first verify the performance of the developed stochastic control framework in a simulation study based on the model parameters estimated from the actual experimental data. We then illustrate the application of our BMI system by demonstrating its ability to maintain precise control of time-varying target levels of burst suppression and to promptly transition between changing target levels without overshoot or undershoot in individual rodents.
Animal studies were approved by the Subcommittee on Research Animal Care, Massachusetts General Hospital, Boston, Massachusetts, which serves as our Institutional Animal Care and Use Committee. Animals were kept on a standard day-night cycle (lights on at 7:00 AM, and off at 7:00 PM), and all experiments were performed during the day.
Surface EEG recordings were collected using extradural electroencephalogram electrodes that were surgically implanted at the following 4 stereotactic coordinates relative to lambda: A (Anterior) 0 mm L (Lateral) 0 mm, A6L3, A6L-3, and A10L2 [6], [15], [16]. During implantation, general anesthesia was induced with isoflurane. At the above four stereotactic coordinates, four holes were made using a microdrill (Patterson Dental Supply Inc., Wilmington, MA). An electrode with mounting screw and socket (Plastics One, Roanoke, VA) was screwed into each of these four holes. The sockets were in turn inserted in a pedestal. Dental acrylic cement was used to permanently fix the screws, sockets and pedestal. Recording began after at least 7 days of recovery following implantation.
During the experiment, the potential difference between electrodes A0L0 and A6L3 was recorded and the signal was referenced to A10L2 and recorded using a QP511 Quad AC Amplifier System (Grass Instruments, West Warwick, RI) and a USB-6009 14-bit data acquisition board (National Instruments, Austin, TX). The binary signal was acquired at a sampling rate of 500 Hz and fed into our BMI. Our algorithm was implemented in a simulink-matlab framework on a HP Probook 5430 s laptop. This setup controlled a Physio 22 syringe pump (Harvard Apparatus, Holliston, MA) to deliver the propofol infusion rate. A 24 gauge intravenous catheter was placed in a lateral tail vein during brief general anesthesia with isoflurane (2% to 3%) in oxygen, and then the animal was allowed to fully recover from the isoflurane general anesthetic in room air before the start of the experiment. The temperature of the animal was monitored and maintained in the normothermic range for the duration of the experiment.
For all experiments, the magnitude of the raw EEG signal was low-pass filtered below 5 Hz and then thresholded to convert it into a binary signal. At the start of an experiment, the threshold level was empirically chosen based on visual inspection of the BSP and the corresponding binary data and based on the values of the filtered EEG over the bursts and suppressions. Figure 1b shows the burst-suppression raw EEG, filtered EEG and threshold, and the resulting binary signal. The segmentation algorithm was run in real time. Several preliminary boluses of propofol were administered to each rat and the obtained BSP traces were used for system identification in each animal (see System Identification section below). The experiment was then conducted by giving the rat a manual propofol bolus to induce a burst suppression state, and the real-time BMI control experiment started once the BSP dropped to a level of 0.1–0.3. In the real-time BMI experiments, the goal was to acquire, maintain, and transition between three target BSP levels (low, medium, high). The order of the target levels was randomized. Each real-time BMI control experiment was conducted for an average of 62 min. Three rats were available for the experiments, weighing 366, 391, and 422 gr respectively. Each rat was used for two real-time experiments, resulting in six real-time experiments.
Our system identification procedure is conducted prior to real-time BMI control for each animal in a preliminary experiment and consists of two steps. First, a BSP signal is estimated from the binary thresholded EEG trace using a special case of our recursive Bayesian estimator in which we take the state to be the scalar variable . Hence the corresponding state model in the estimator imposes a smoothness constraint on using a first-order linear Gaussian process [17]. Specifically, we use a special case of (10) as . Second, the corresponding BSP estimate is fitted using a non-linear least-squares procedure to minimize the sum-squared-error between the model predicted BSP and the estimated BSP. The system parameters are thus the solution to(29)where is the model predicted BSP given the values for the system parameters (see Results and Figure 2).
To characterize the performance of the BMI at steady state, we compute the error between the target BSP at each time, , and the controlled BSP, , as(30)
We use the error to calculate multiple standard metrics [18] of performance. These metrics are the median absolute deviation (MAD)(31)the median prediction error (MDPE)(32)and the median absolute performance error (MDAPE)(33)The MDPE is a measure of bias at steady state and the MDAPE is a measure of normalized error. We compute these metrics for low, medium, and high target BSP levels and across all levels for each experiment. The median is computed across data points at steady state. Finally we compute the median of all these measures across all experiments to quantify the overall performance of the BMI.
To characterize the performance of the BMI in transitioning between target BSP levels, we calculate the rise time for an upward transition and the fall time for a downward transition. These are defined as the time it takes, once the target is changed, for the BSP to reach within 0.05 BSP units of the new target BSP. We then find the rate of BSP change defined as(34)and calculate the median of this rate across all upward transitions and also across all downward transitions.
In addition to calculating the steady-state error metrics above for the low, medium, and high levels in each experiment, across all levels for each experiment, and across all experiments (Table 1), we also performed specific statistical assessments based on the error distribution at each level to examine the reliability of the BMI overall [19]. In particular, we considered the BMI to be reliable at each level if its absolute error measure, , was lower than a specified value with high probability. Experimentally we found that the absolute error at any time step in our BMI system was almost always below . Therefore we considered the BMI to be reliable at a given level if its absolute error at that level was less than 0.15 with probability ≥0.95 and highly reliable if the absolute error at each level was less than 0.10 with probability ≥0.95. This is equivalent to the 95th percentile of the absolute error distribution at a given level being less than 0.15 and 0.10, respectively. Hence we can compute from the absolute error distribution at each level the 95th percentile and consequently identify the BMI performance at each level as reliable or not.
After evaluating the reliability of the BMI at each level separately, we use a Bayesian analysis to identify the reliability of the BMI across all levels. To do so, we combine the results of the reliability assessments across all levels to compute an overall assessment of reliability for the experiment. In our experiments, we tested the BMI over 20 levels with the time duration at each level between target transitions being 18.6 minutes on average. For the purpose of steady-state error calculation, we remove 5 minutes of data after an upward transition and 7 minutes of data after a downward transition to ensure that the system is at steady-state and to ensure approximate independence between levels. The independence assumption between levels is justified because if we assume even a high first-order serial correlation of 0.98 between adjacent data points separated by one second and we allow between 5 to 7 minutes for the transition between levels before making the steady-state error calculations, then the maximum correlation between the closest two points in immediately adjacent levels is between , where min data point per minute and minutes data points per minute. Because these maximum correlations are effectively 0, assuming independence between levels is highly reasonable (we acknowledge that lack of correlation is not equivalent to independence). Hence the data between levels within animals are approximately independent so that the 20 levels serve as the unit of analysis in the overall assessments of reliability.
Denoting the probability that the BMI system is reliable at any level by , the total number of reliably controlled levels, , is binomially distributed with success probability out of independent levels. The number of successes is in turn equal to the number of levels for which the BMI is identified as reliable as described above. Given the binomial likelihood and taking the prior distribution for to be the uniform distribution on the interval (0, 1), it follows that the posterior distribution for is the beta distribution with parameters and [15], [20]. We thus estimate as the mode of this beta distribution and consider the BMI system reliable overall if the leftmost point of the 95% credibility interval for is greater than 0.
To test our closed-loop BMI system for control of medical coma, we perform both simulation-based verification as well as real-time in vivo experiments in rats. In both cases, we implement the recursive Bayesian estimator combined with both the bounded LQR controller as well as the model predictive controller. Using both validation methods, we show that the closed-loop BMI system can accurately control time-varying target levels of burst suppression in real time.
For each experiment, we first performed the system identification step for each animal using the scalar filtering and the nonlinear least-squares model fitting (see Materials and Methods). Figure 2 shows two sample BSP traces in response to boluses of propofol administered in preliminary experiments prior to BMI control, and the fitted system response of the second-order system in (2). The estimated parameters for Figures 2a and 2b are and , respectively. Once the system model was fitted, the real-time BMI control experiments were conducted. We use the fitted system model in Figure 2b for our simulation-based verification below.
We first perform a set of simulations to verify the performance of the closed-loop BMI system. In our simulations, we assume that the anesthesia drug delivery period is a total of 45 minutes and that the goal is to keep the BSP at three desired target levels, 0.4, 0.7, 0.9, each for 15 minutes. We simulate all 6 possible order permutations of these levels. To run the simulations, we use the estimated system model in Figure 2b. Note that all the fitted system models in our experiments were controllable.
To specify the cost function (see (27) and Supporting Text (S.18)), we take . We choose this since the main goal is to have the effect-site concentration close to the target value and since the effect-site concentration is the observable through the EEG. The choice of in turn depends on how fast we desire the controller response to be. Smaller values of result in faster controller response since the cost on the amount of drug infusion is reduced. Here we pick for our desired response.
We take the discretization step to be sec. This means that the closed-loop system updates its estimate of the BSP and its drug infusion rate every second. To simulate a trial of the closed-loop controlled system response, at each time we use and to find using (2) with initial condition , . To get the binary output of the thresholded EEG within this time step, we generate a realization of the binomial distribution in (7) with mean and using a sampling rate of 10 Hz (i.e., taking ). Given this binomial realization, we use the recursive Bayesian estimator to estimate the concentration state , and then use this estimate as feedback in the controller to decide on the infusion rate .
We impose the constraints on the control (i.e., drug infusion rate) by first finding the unconstrained control solution from (25) and then using the closest value to it in the constrained feasible set . For example, to impose positivity and for negative control solutions we use zero instead. We can similarly do this for constraints on the maximum drug infusion rate.
Figure 3 shows sample closed-loop controlled BSP traces for each of the 6 possible permutations of the desired target trajectories. Here the only imposed constraint is positivity of the drug infusion rates. In each subfigure, the top panel shows the BSP traces and the bottom panel shows the drug infusion rate. The stochastic control framework can achieve successful control of burst-suppression. The framework is particularly successful in changing the BSP level without overshoot or undershoot.
We also tested the model predictive controller with various time horizons, . In the model predictive controller, we impose the constraints on the control inputs (i.e., drug infusion rates) explicitly in the formulation and thus find the constrained (approximately) optimal solution. Since our goal is to compare the bounded LQR and MPC control strategies in this set of simulations, we assume that both controllers know the BSP perfectly at each time (i.e., we use the true as feedback in the controller). We compare the MPC drug infusion rate with the bounded LQR infusion rate in Figure 4, where the constraint is positivity on the drug infusion rate. As we increase the optimization horizon, the two solutions converge. This shows that, in this problem, solving the unconstrained LQR and then bounding it is approximately optimal. The controlled BSP in Figure 3 is noisier than in Figure 4 because in the former the BSP is estimated from a stochastic binary time-series emulating the segmented EEG (Figure 1b) and in the latter BSP is assumed to be perfectly known to the controllers. We can also show that, similarly, when an upper-bound on the drug infusion rate is desired, the two solutions again converge (Figure 5). It is important to note, however, that in our problem no constraints are placed on the state. Our recursive Bayesian estimator combined with the implemented real-time MPC can extend our framework to solving more complex problems with constraints also on the state variables, such as blood pressure (see Discussion).
Even though simulation-based validations are helpful in assessing the behavior of the BMI, the true test of the BMI is in in vivo experiments as we present below.
We implemented our BMI in experiments with rodents and tested it for controlling the level of burst suppression in real time. The BMI used the recursive Bayesian estimator combined with either the bounded LQR controller or the MPC. The BMI in both cases could successfully and accurately control the BSP level in rodents in real time.
The control sessions lasted an average of 62 minutes and consisted of at least 3 target BSP levels, thus requiring at least 3 transitions. Figure 6 shows the BSP and the drug infusion rate in 6 closed-loop BMI sessions that were run in real time in rodents (see also Supporting Figure S1 that shows the evolution of in these experiments). Figure 6a–e were run with the bounded LQR controller and Figure 6f was run with the MPC. All experiments except for the one in Figure 6e consisted of 3 target levels, identified as low, medium, and high levels for the purpose of metric calculation. The experiment in Figure 6e consisted of 5 target BSP levels and hence we identify the lowest two levels as the low level and the highest two levels as the high level to calculate the metrics.
As is evident in Figure 6, the BMI could successfully and promptly transition between levels and accurately maintain the BSP at a desired target level. At steady state, the BMI-controlled BSP closely followed the target BSP level. The real-time variations in the drug infusion rate at higher levels of BSP, e.g., at 0.9, were larger than at the lower levels since larger amounts of propofol are needed to keep the EEG in suppression 90% of the time while allowing for the bursts 10% of the time (this can also be seen from (1) by observing that is monotonically increasing with ). The MDAPE (measure of normalized error) across all experiments for the low, medium, and high target BSP levels was only 7.32%, 3.02%, and 2.78%, respectively. When considering all levels, the MDAPE was only 3.61% (Table 1). Moreover, the deviation between the target BSP level and the BMI-controlled BSP, measured through MAD, was 0.03 BSP units or less for any level. Across all levels the MAD was only 0.02 BSP units (Table 1), a negligible error in practice. Finally, the MDPE was small across all levels. Together, these results demonstrate that the BMI achieved precise control of multiple target burst suppression levels at steady state within the same experimental session.
We also performed a Bayesian analysis to assess overall reliability of the BMI based on the steady state error distributions at each of the 20 levels used in the experiments (Materials and Methods). The data at different levels within animal are approximately independent so that the 20 levels serve as the unit of analysis in the overall assessment of reliability. The 95th percentile of the absolute error distribution at each of the 20 levels was less than 0.15 giving a mode of the posterior density for (probability that the BMI is reliable at any level) of and a 95% credibility (Bayesian confidence) interval for of (0.87 to 1.00) (Figure 7). The lower bound of the 95% credibility interval of 0.87 is well above 0, the point of no control. These findings establish that the system is reliable. In addition, for 17 out of 20 of the levels the 95th percentile of the absolute error distribution was less than 0.1, giving a mode of the posterior density for of and a 95% credibility interval for of (0.67 to 1.00) (Figure 7). This finding suggests that furthermore the BMI system meets our definition of being highly reliable overall. We therefore conclude that the BMI system is highly reliable for real-time control of medical coma using burst suppression across dynamic targets.
In addition to accurate and reliable control at steady state, the BMI was especially successful in promptly transitioning between target BSP levels. The BMI could increase the level of BSP rapidly, while avoiding overshoot. To increase the BSP, the BMI immediately increased the drug infusion rate once the target was increased, and then gradually reduced the infusion rate until the BSP approached the new target level. The rate at which the BMI increased the BSP level was 0.32 BSP units per minute. The median rise time in our experiments was under a minute (49 seconds).
The BMI was also able to decrease the BSP level without undershoot. To decrease the BSP, the BMI first stopped the drug infusion and then gradually restarted it once the BSP approached the lower target BSP level. The rate at which the BMI could decrease the BSP level was 0.11 BSP units per minute. In decreasing the level of BSP, the time response of the BMI is mainly governed by the clearance rate in the pharmacokinetic model of the rat. Hence although the controller stopped the drug infusion immediately once the target was dropped, it took a few minutes for the BSP to go down to the desired target level. The median fall time in our experiments was 4.45 minutes.
These results thus demonstrate the feasibility of automatic reliable and accurate control of medically-induced coma using a BMI.
To study the feasibility of automating control of medically-induced coma, we developed a BMI to control burst suppression in a rodent model. Our BMI system reliably and accurately controlled burst suppression in individual rodents across dynamic target trajectories. The BMI promptly changed the BSP in response to a change in target level without overshoot or undershoot and accurately maintained a desired target BSP level with a median performance error of 3.6% and a percent bias of -1.4%.
Our work contributes to the extensive BMI research in anesthesiology aimed at controlling brain states under general anesthesia. This field began in the 1950s [21]–[23] and developed further in the 1980s [24]. BMI systems for control of sedation are now commercially available [25] and have been recently approved for use in the United States. The most commonly used control signal is the Bispectral Index (BIS) [26]–[40]. Other control signals include an auditory evoked potential index [41], the spectrogram median frequency [24], [42], [43], a wavelet-based index [44] and an entropy measure [45]. Both standard and non-standard control paradigms [24], [27], [35], [41], [45] have been used in these systems with the principal objective being control of unconsciousness [24], [26]–[32], . A recent report controlled both antinociception and unconsciousness [45]. Although several criteria have been established for successful control, a criterion used in BIS studies has been maintaining BIS not at a specific value but in the broad range between 40 to 60 [26]–[39]. Vijn and Sneyd [6] and Cotten et al. [7] controlled constant target levels of burst suppression in rodent models and reported average control results over rodents. Schwilden demonstrated control of median frequency in individual human subjects [24]. None of these studies considered control of dynamic time-varying trajectories.
We developed a BMI for real-time control of burst suppression across time-varying target levels in individual rodents using a stochastic control framework. Our stochastic control framework consists of a two-dimensional state estimator and an optimal feedback controller. In our formulation, we assumed a stochastic form of the log transformed version of our system to incorporate both the two-dimensional system model and noise in our estimates and to ensure non-negative concentration estimates (Eqs. (2) and (6)). This model-based two-dimensional state estimator is one major reason that the current BMI largely avoided overshoot and undershoot. By incorporating the two-dimensional stochastic dynamic model and computing both and at each update (Eqs. (14)–(17)), the estimator predicted the effect of the real-time drug infusion rate on the BSP. In upward transitions, this avoided underestimating the BSP in response to drug infusion that would result in overestimating the required amount of drug and hence in an overshoot. This similarly prevented undershoot in downward transitions. Our framework is thus analogous to maintaining control in a navigation system by estimating both position and velocity.
In addition to the two-dimensional estimation algorithm, the BMI consists of LQR and MPC controllers. Controllers using MPC and LQR strategies have been used successfully in many applications. We recently demonstrated the success of a LQR paradigm to control a motor neuroprosthetic device using point process observations of spiking activity and a linear Gaussian kinematic state model [46]–[50]. MPC has been widely used in process control and chemical industries [51]–[53] and has been previously applied to closed-loop administration of analgesics [54], for sedation control using the BIS as the control signal [55] and in a simulation study for control of BIS during surgery [40]. The LQR and MPC controllers are both formulated in an optimal feedback control framework [8]. They specify the control objective as a cost function to be minimized by selecting the optimal infusion rates. We can therefore adjust the behavior of these controllers, for example the speed of transitions, by adjusting the penalty on various terms in the cost function. While our LQR implementation imposes constraints on the drug infusion rates by bounding the control solution, our MPC implementation allows us to impose explicitly any required constraints on both the states and the drug infusion rates, such as non-negative or bounded infusion rates, by solving a constrained optimization problem at each time step in real time. For example, if the BMI system always needed to keep the drug infusion rate below a specified maximum level, the MPC controller could impose this explicitly in the solution. Since the only constraints in our problem were on the control variable (i.e., the infusion rate), the LQR and MPC strategies performed similarly (Figure 5). However, the recursive Bayesian estimator combined with the real-time MPC strategy can be used to solve problems that require constraints on the state variables as well. This situation could arise in problems requiring joint control of multiple state variables, such as controlling simultaneously the anesthetic level and other physiological variables such as blood pressure and heart rate.
Other approaches can also be used for anesthesia control. We recently reported successful control of burst suppression using a proportional-integral (PI) controller in simulated rodent [11], simulated human [11], and actual rodent experiments [19]. The experimental studies differed from the ones presented here in that the transitions between target levels were carried out in 5 to 10 minutes ramps. Also, the BSP estimation algorithm in those studies was one rather than two dimensional. Given the stochastic two-dimensional dynamic model and the EEG signal, here we used a stochastic control paradigm consisting of a two-dimensional estimator and an optimal feedback controller in place of the one dimensional estimator and the deterministic PI controller. The model-based two-dimensional state estimator in our framework is one major reason that the current BMI can both make prompt and reliable transitions between levels (median rise time of 49 seconds and fall time of 4.45 minutes) and avoids BSP overshoots and undershoots in any transitions. The stochastic control framework also offers tremendous flexibility. In particular, the MPC allows us to extend the BMI to control with constraints on the control variable and a vital variable such as blood pressure. The stochastic optimal formulation also provides a framework for adjusting the behavior of the BMI by simply modifying the cost function. Finally, while the mathematical derivation for the stochastic framework may be more complex, the final infusion rate solution is straight forward. The LQR solution is given simply by a linear function of the state estimate at each time. The MPC controller optimization problem is convex and can be solved using existing convex optimization software. Indeed we ran the MPC in real-time in our rodent experiments on a standard laptop (Figure 6f).
We chose levels of burst suppression as a control target because it is a physiologically defined brain state [4], [5] with a well-defined EEG signature that can be readily characterized in real time for the purpose of control. The linear two-dimensional state model (Eq. (2)) is the simplest pharmacokinetics representation for relating the concentrations of anesthetic in the blood and in the brain to BSP (Eq. (1)) computed from burst suppression in the EEG. This simplified two-compartment model was sufficient in our experiments to achieve reliable and accurate control of burst suppression.
Our Bayesian state estimator (Eqs. (14)–(17)) computes the central compartment and the effect-site concentrations in real time from the EEG converted into binary observations. Here, both the prior state model (Eq. (6)) and the binomial observation model (Eq. (7)) are non-linear functions of the state. We thus use two approximations at each time step to derive the estimator recursions, a linear approximation to the prior model at that step and a Gaussian approximation to the posterior model. Gaussian Laplace-type approximations have been successfully used in many applications for example in our previous work estimating states with linear prior models from point process observations of neural spiking activity [12], [13], [47]–[50], [56]–[59]. Our system identification procedure used the one-dimensional version of the binary filter, coupled with a non-linear least squares procedure to estimate model parameters (Eq. (2)) for each animal and thereby, implement individually tailored control strategies. Future work can extend this system identification procedure to an efficient expectation-maximization (EM) algorithm by replacing the one-dimensional binary filter algorithm with the current Bayesian state estimator [13], [60], or can design an adaptive estimator that not only computes the BSP but also updates the system parameters during the several hours of real-time control.
We demonstrated in a rodent model that the BMI achieved reliable and accurate control of burst suppression. It would also be valuable as a next step to test this BMI in a rodent model of refractory seizures or intractable intracranial hypertension prior to testing it in humans.
A BMI system to automatically control medically-induced coma could provide considerable cost-saving and therapeutic benefits. Although the state of medical coma is often required for several days, it is achieved by manually adjusting the anesthetic infusion rate to maintain a specified level of burst suppression assessed by continual visual inspection of the EEG. Automated control would allow much more efficient use of intensive care unit personnel as a single nurse per shift would not have to be solely dedicated to the task of manually managing the drug infusion of a single patient for several days. Hence even assuming the same patient outcomes between automated and manual control, there could be important savings in intensive care unit resources under the automated control regimen.
In addition to the inefficient use of the intensive care unit staff, manual manipulation of the infusion rate does not approximate the infusion rate changes of an automatic controller (Figure 6). Similarly, visual inspection of the EEG does not provide an accurate estimate of the state of burst suppression. The current work establishes the feasibility of implementing automated, accurate and reliable control of medical coma in a rodent model suggesting that a BMI could be developed to study whether such accurate control improves patient outcomes. For example, reliable and accurate control of medical coma could offer the possibility of ensuring adequate brain protection for intracranial hypertension and adequate therapy for status epilepticus while using the least amount of anesthetic and minimizing overshoots when transitioning to a desired level of burst suppression. Reliable and accurate control would also make it easier to induce periodic arousals to conduct neurological assessments and prevent anesthetic overdose syndrome [61]. To establish these potential therapeutic benefits of reliable and accurate control of medical coma, outcome studies in rodent models of intracranial pressure and status epilepticus will be required before proceeding to human investigations.
We have also shown that other states of general anesthesia have well defined EEG signatures [62], [63]. Therefore, the ability of our BMI to track accurately changing target levels of burst suppression further suggests that it could be adapted to control states of general anesthesia and sedation for patients requiring surgical or non-surgical procedures. Our stochastic estimation paradigm and model predictive controller could also be used to control jointly the state of general anesthesia and physiological variables such as blood pressure. These investigations will be the topics of future reports.
|
10.1371/journal.ppat.1006217 | Multi-layered control of Galectin-8 mediated autophagy during adenovirus cell entry through a conserved PPxY motif in the viral capsid | Cells employ active measures to restrict infection by pathogens, even prior to responses from the innate and humoral immune defenses. In this context selective autophagy is activated upon pathogen induced membrane rupture to sequester and deliver membrane fragments and their pathogen contents for lysosomal degradation. Adenoviruses, which breach the endosome upon entry, escape this fate by penetrating into the cytosol prior to autophagosome sequestration of the ruptured endosome. We show that virus induced membrane damage is recognized through Galectin-8 and sequesters the autophagy receptors NDP52 and p62. We further show that a conserved PPxY motif in the viral membrane lytic protein VI is critical for efficient viral evasion of autophagic sequestration after endosomal lysis. Comparing the wildtype with a PPxY-mutant virus we show that depletion of Galectin-8 or suppression of autophagy in ATG5-/- MEFs rescues infectivity of the PPxY-mutant virus while depletion of the autophagy receptors NDP52, p62 has only minor effects. Furthermore we show that wildtype viruses exploit the autophagic machinery for efficient nuclear genome delivery and control autophagosome formation via the cellular ubiquitin ligase Nedd4.2 resulting in reduced antigenic presentation. Our data thus demonstrate that a short PPxY-peptide motif in the adenoviral capsid permits multi-layered viral control of autophagic processes during entry.
| Cells have developed surveillance systems to detect invading pathogens, e.g. when they damage the membrane upon cell entry. Cells respond to membrane damage by activating selective autophagy to target pathogens for lysosomal degradation and pathogen removal. In this manuscript we show that endosome penetrating adenoviruses also activate an autophagic response upon entry and that this involves Galectin-8 mediated membrane damage recognition similar to pathogenic bacteria. In contrast adenoviruses avoid degradation by escaping into the cytosol. We show that they possess a small virion encoded PPxY-peptide motif, which they use to divert the autophagic response of the cell. This motif sequesters the cellular ubiquitin ligase Nedd4.2 resulting in limited autophagosome formation, which prevents viral degradation and antigenic presentation and ensures efficient endosomal escape and nuclear transport.
| Intracellular pathogens, such as viruses, penetrate the limiting membrane of the cell to access cellular functions for propagation support. In response, cells try to detect and eliminate entering viruses through multiple pre-existing defense mechanisms referred to as restriction factors or intrinsic immunity [1]. To establish productive infections, viruses thus have to divert, limit or control cellular intrinsic immunity. Adenoviruses (AdV) are amongst the most efficient viruses to enter cells and rapidly establish lytic infections after nuclear genome delivery. AdV are non-enveloped, double stranded DNA viruses, which enter cells by receptor-mediated endocytosis [2,3]. Uptake invokes structural changes in the capsid [4], which releases the membrane lytic internal capsid protein VI (PVI) to breach the endosomal membrane [5,6]. Following membrane rupture, AdVs escape to the cytosol and use microtubule dependent transport accessing the nucleus via the microtubule organizing center (MTOC) [7]. At the nucleus the capsid binds and disassembles at the nuclear pore complex followed by genome import into the nucleus [8].
Membrane penetration is an essential step in the infection process. This was shown in work using the temperature sensitive AdV mutant ts1 (TS1), which has a point mutation (P137L) in the viral protease gene preventing newly assembled virions from undergoing maturation cleavage at the non-permissive temperature. TS1 particles are hyper stable and enter cells via endocytosis but fail to release PVI. Subsequent absence of membrane penetration results in ts1 particles being sorted into lysosomes for degradation [9,10]. A key role in AdV cell entry is played by a highly conserved PPxY peptide motif (where x can be any amino acid) in PVI, which is exposed upon PVI release [11]. PPxY motifs bind to WW domains commonly found e.g. in the Nedd4 family of HECT-domain E3 ubiquitin ligases. Using recombinant proteins it was recently shown that PVI binds directly via the PPxY motif to the ubiquitin ligase Nedd4.2 [12]. Mutating the motif to PGAA impairs Nedd4.2 binding and abolishes PVI ubiquitylation. More significantly, although mutation of the PPxY motif does not decrease membrane rupture, for unknown reasons PVI-mutated viruses (M1) have a strong nuclear transport defect and fail to localize at the MTOC, resulting in an up to twenty-fold defect in specific infectivity compared to the wild type (WT) [11].
Membrane damage, caused by viruses or other pathogens, is perceived as a danger signal by the cell and recognized through the galectin system. Galectins are beta-galactoside-binding proteins that contain carbohydrate recognition domains, which sense and mark membrane damage through the abnormal cytosolic exposure of intra-lumenal glycans [13]. Several galectins including Gal3, Gal9 and Gal8 mark bacteria induced membrane damage although only Gal8 was shown to restrict bacterial replication [14,15]. Galectin positive membrane fragments can be subject to degradation via autophagy [15,16]. This is achieved e.g. through Gal8 mediated recruitment of the adapter protein NDP52 towards bacteria containing vacuoles. NDP52 in turn establishes the link to the autophagic machinery for cargo degradation exemplifying how the initial sensing via galectins leads to an active antimicrobial response [15,17].
Autophagy, or selective autophagy when the cargo is sequestered via receptors, is an evolutionary-conserved cytosolic, lysosome-dependent, degradation process [18]. It serves as a cellular survival pathway to maintain homeostasis e.g. by providing the cell with nutrients upon starvation. Autophagy is initiated through formation of an isolated membrane structure, the phagophore, via the concerted action of several factors belonging to the AuTophaGy-related genes (ATG) [19,20]. After initiation the cytoplasmic protein LC3 (Microtubule-associated protein 1A/1B-light chain 3) becomes lipidated and incorporated into the elongating membrane, which forms a double-membrane vesicular structure, the autophagosome, and closes around the cargo destined for degradation. In contrast autophagosome formation and maturation into autolysosomes is still not fully understood but may involve LC3 mediated transport from the site of origin towards the perinuclear region by using a retrograde, microtubule dependent and dynein-dynactin motor complex mediated transport. This active transport facilitates efficient fusion of closed autophagosomes with perinuclear lysosomes [21–23].
Selective autophagy is part of the intrinsic immune response to attenuate intracellular pathogen infections via lysosomal degradation [24]. Next to galectins, selective autophagy involves additional marker molecules such as ubiquitin and further adapter molecules [25]. We recently demonstrated that Gal3 also marks membrane damage caused by AdV and observed that AdV traffics in Gal3 positive membranes prior to endosomal escape [5,26] although a role in selective autophagy was not addressed. Still, several studies have reported that autophagy can promote or restrict viral infections including AdV [27–29]. Autophagy also indirectly restricts viral infection by increasing viral peptide presentation on MHC-II molecules to mount an adaptive immune response as recently shown for AdV [28] and other viruses [30,31]. Accordingly several viruses have developed efficient strategies to temper with the autophagic machinery [32].
Within this study we use AdV to show that selective autophagy mediated by Gal8 also targets endosomolytic viruses. Moreover we show that AdV escape from and limit the autophagic response through a conserved PPxY peptide motif encoded in the membrane lytic capsid protein VI. We show that this is achieved by restricting the formation of autolysosomes via a mechanism involving the cellular ubiquitin ligase Nedd4.2 resulting in reduced antigenic presentation. Furthermore we provide evidence that AdV hijacks the autophagic response to achieve efficient endosomal escape and accelerated nuclear transport and genome delivery showing a multilayered viral control of the autophagic machinery during virus entry.
We showed previously that E1/E3 deleted GFP expressing adenoviral vectors with the PPxY-motif (WT) in capsid protein VI mutated to PGAA (M1) have reduced transduction efficiency and altered nuclear transport [11]. Both vectors were used throughout this study and are referred to as “WT” and “M1” respectively. To verify the phenotype we infected cells with WT and M1 and quantified the transduction levels at different multiplicities of infection (MOI). The results confirmed the M1 phenotype in a variety of cell types suggesting a virion-associated mechanism (Figs 1A, 1B and S1). We then asked if the defect was linked to the role of PVI in membrane penetration. We infected cells for different times with WT and M1 and used Gal3 as a marker for AdV induced membrane damage. The results showed that both viruses efficiently induce Gal3 punctae (Fig 1C). We quantified both the number of Gal3 punctae per cell as a measure of membrane rupture and the colocalization of virus particles with Gal3 as a measure for endosomal escape. We observed no apparent difference in the kinetics of membrane lysis between WT and M1 (Fig 1D) and only minor differences in the absolute number of cell-associated virus over time (Fig 1E). In contrast, the M1 colocalized with Gal3 punctae to a much larger extend than WT (Fig 1F), suggesting a post-lysis endosomal escape defect. Previous in vivo imaging of AdV endosomal escape suggested a propelled and microtubule dependent escape mechanism [5,26] and both dynein and microtubule transport have been implicated in AdV entry [33,34]. To confirm that AdV utilize dynein to escape from ruptured endosomes we used Ciliobrevin D (CilioD), a reversible chemical inhibitor of the dynein AAA+ ATPase motor to inhibit dynein activity during viral transduction (Fig 2). At non-toxic CilioD concentrations we observed decreased GFP transgene expression compared to vehicle control treated cells (Fig 2A). Time course infection experiments with fluorescently labeled WT revealed no change in the initial number of Gal3 punctae upon CilioD treatment. However, Gal3 punctae increased over time (Fig 2B) and the WT remained associated with Gal3 (Fig 2C) in drug-treated cells suggesting involvement of dynein in virus escape and membrane damage removal but not in the initial lysis. Sequestration of the WT after CilioD treatment also prevented efficient virus translocation to the perinuclear area (Fig 2D) consistent with previous observations made for the M1 virus [11]. Taken together, these data show that the PVI PPxY-motif and dynein are critical for efficient escape from ruptured endosomes and subsequent transport to the nucleus.
We next asked if AdV, similar to invasive bacteria, would activate selective autophagy upon membrane rupture. To address this question we used three different AdV’s all based on genotype HAdV-C5, which share receptor-mediated endocytosis but differ upon endosomal sorting. Next to WT and M1 we infected cells with an E1/E3 deleted GFP expressing ts1 adenoviral vector grown at the non-permissive temperature [10]. This hyperstable ts1 mutant virus (“TS1” in this study) does not release PVI and does not lyse the endosome upon uptake. We then performed an infection time course (shown in Fig 3A) and quantified appearance of Gal8 punctae to mark membrane damage. We used Gal8 because of its functional role in selective autophagy [15]. Gal8 punctae formation occurred in WT and M1 but not TS1 infected cells (S2 Fig), was often linked to particles also positive for PVI (Fig 3B) and punctae formation and particle association was rapid (Fig 3C) and followed similar kinetics as Gal3 punctae formation. To show that AdV induced membrane damage activates autophagy we next stained for LC3 and quantified LC3 punctae reminiscent of autophagosome formation [35]. We observed LC3 punctae formation for WT and M1 but not for TS1 supporting that membrane lysis was required to initiate autophagy (Fig 3C). We confirmed the results in a different cell type (S2B Fig) and by showing the membrane damage-dependent accumulation of phosphatidyl-ethanolamin (PE) conjugated LC3-II by western blot in WT and M1 but not in TS1 infected cells (Fig 3E). LC3 punctae formation occurred rapidly within 15–30 min after virus uptake (Fig 3F) resembling PVI release kinetics [10]. We then asked if AdV induced autophagy was selective and mediated by adapter proteins. We repeated the time course experiment and analyzed recruitment of autophagy adapter p62, NDP52 and optineurin. Accumulation of p62 and NDP52 punctae was dependent on virus induced membrane damage (S2C and S2D Fig). Punctae appearance and association with WT and M1 particles followed similar kinetics (Fig 4A for NDP52 and 4C for p62) as previously observed for galectins and LC3. In contrast optineurin, was not detected in association with viral particles under our conditions. Co-staining for cellular marker molecules and viruses showed large overlap in colocalization suggesting functional links. E.g. several viruses positive for NDP52 also stained positive for Galectins (Fig 4B) and p62 positive viruses stained also positive for ubiquitin (Fig 4D) reminiscent of their recruitment towards invasive bacteria.
In summary our analysis suggested that PVI released from entering viral particles caused membrane damage to which the cell responds with selective autophagy. Interestingly we observed higher levels and more prolonged accumulation of LC3 punctae for the escape defective mutant M1 vs. WT (Fig 4F). We hypothesized that M1 virions maybe degraded by autophagy, which could explain their strongly reduced infectivity.
To determine if M1 viruses are degraded via autophagy we first visualized AdV association with LC3 in living cells. We infected cells stably expressing LC3-GFP [36] with Alexa594 coupled WT and M1 virions and imaged the cells using spinning disk confocal microscopy. We observed that within ~15–30 min several cell associated viruses turned LC3 positive, mostly occurring at the cellular periphery (Fig 3A, S1 Movie). At >1hpi most WT viruses showed no more association with LC3 and accumulated in the vicinity of the nucleus (Fig 3B and 3C, top). In contrast, M1 viruses lacked nuclear accumulation and were often found inside large LC3 positive structures reminiscent of autophagosomes (Fig 3B and 3C, bottom). This difference in dynamic LC3 association between WT and M1 viruses was confirmed using quantitative time resolved fluorescence analysis (Fig 3D). Autophagosomes are characterized by forming a double membrane around their substrate. To confirm that M1 particles associate with autophagosomes we next performed transmission electron microscopy (TEM). Cells were infected with fluorescently labeled WT or M1 virus, fixed at 30 min post infection and processed as detailed in the material and methods section. WT viruses were observed in the cytoplasm (Fig 5Ea and 5Eb) or at the nuclear pore complex (NPC, Fig 5Ec and 5Ed). In contrast M1 viruses could be observed inside vesicular structures (Fig 5Ee and 5Eg). At higher magnification we could observe that M1 particles seemed to be entrapped in partially ruptured vesicles resembling endosomes (Fig 5Ef and 5Eh). These virus containing vesicles were partially (f) or fully (h) engulfed by a double membrane vesicle resembling autophagosomes. Interestingly we observe a cleft between the endosome and the inner autophagosomal membrane filled with electron dense material, which would be the putative location for the linking autophagy receptors (marked with asterisk in Fig 5Ef and 5Eh). Our analysis thus strongly supports that M1 particles still inside the ruptured vesicle are sequestered by autophagy whereas WT viruses escape from the endosomal compartment prior to autophagic sequestration.
To investigate a potential autophagic degradation for the M1 virus we treated cells with 3-methyladenin (3-MA), which blocks autophagosome formation via the inhibition of class III PI3K [35]. Pretreated and vehicle control cells were transduced with WT or M1 viral vector expressing GFP and relative transduction levels were determined using fluorescence activated cell sorting (FACS). 3-MA treatment did not affect WT virus infectivity but significantly increased M1 virus infectivity restoring WT infectivity to nearly 75% (Fig 6A). In addition we found that the M1 virus located per se more than the WT virus with Pi3P positive compartments from which autophagosome formation originates (S3A Fig) [37] and that upon M1 infection Beclin-1, a subunit of the class III PI3K kinase complex, was enriched on cellular membranes (S3B and S3C Fig) suggesting differences in class III PI3K kinase complex activity between both viruses. We next inhibited the autophagic flux by blocking lysosome acidification with chloroquine (CQ, Fig 6B) [35]. CQ treatment increased M1 virus infectivity without affecting WT infectivity although the effects were more moderate (Fig 6B). Next we depleted cells of the LC3 conjugation factor ATG5 using SH-RNAs [35]. Again we observed a specific but moderate increase in M1 infectivity (Fig 6C). Because of the reduced impact in restoring M1 infectivity and because our depletion efforts yielded only partial removal of ATG5 we next infected ATG5 -/- MEFs (Mouse Embryonic Fibroblasts) with WT and M1. As shown in Fig 6D absence of ATG5 in MEFs fully restored the M1 infectivity although the M1 defect in the murine model was less pronounced than in human cells. Still this analysis strongly supported that autophagy was responsible for the M1 defect.
Together our results showed that M1 virus infectivity could be fully or partially restored by blocking autophagy at different steps, showing that the M1 virus but not the WT virus becomes an autophagic substrate. To exclude that the increase in M1 infectivity was not due to pleiotropic effects of autophagy inhibition we depleted cells of galectins or adapter molecules that link the virus induced membrane damage to the autophagic machinery (S4A Fig). M1 infectivity increased only after depleting Gal8 (Fig 7A) but not when depleting either Gal3 (Fig 7B) or Gal9 (Fig 7C). The rescue of M1 infectivity was strongly correlated to the depletion level of Gal8 (Fig 7A) and we were able to observe almost complete restoration of WT infectivity in our best depletion experiments. In contrast M1 infectivity was neither influenced by Gal3 depletion levels nor by Gal9 depletion levels. We next determined WT vs. M1 virus association with LC3 and lysosomes in control- and Gal8-depleted cells over time. Gal8 knockdown had no effect on LC3 colocalization with the WT, but colocalization of the M1 with LC3 decreased to WT levels (Fig 7D, top panel). Additionally, M1 colocalization with the lysosome marker Lamp1 significantly decreased in Gal8 knockdown cells strongly suggesting that Gal8 depletion prevented efficient targeting of the M1 virus for autophagic degradation (Fig 7D, bottom panel). Taken together our results identify Gal8 as an essential restriction factor for the M1 virus and show that Gal8 recruitment towards AdV-ruptured membranes is necessary to target virus-containing membranes for autophagosomal degradation.
We next depleted cells from the autophagy receptors p62, NDP52 and optineurin. We observed no rescue of M1 infectivity when depleting optineurin in agreement with the absence of optineurin recruitment to virus particles. M1 infectivity was only slightly but specifically increased upon depletion of the Gal8 specific autophagy adapter NDP52 (Fig 7E). In contrast, depletion of the autophagy adapter p62 slightly increased infectivity for both viruses implying a somewhat different role in AdV entry (Fig 7E). Thus we next asked if co-depletion of autophagy receptors would have a more pronounced effect and would overcome the M1 restriction. Efficient double depletion of NDP52 and p62 gave similar results as depletion of p62 alone excluding an additive effect (Fig 7F). Taken together this implied that none of the autophagy adapters tested plays a major role in M1 restriction, suggesting that Gal8 mediated restriction of bacteria and adenoviruses work at least in part via different pathways.
Autophagy inhibition had little impact on WT infectivity suggesting that AdV might actively target autophagic processes. To address this question we infected cells with WT and M1 and performed a time course western blot analysis to determine the levels of LC3-II. LC3-I conversion initiated within 15–30 min pi for both viruses and continued at least up to 2 hpi for the M1 virus (Fig 8A). In contrast LC3 conjugation induced by the WT virus was comparable in the beginning but declined at ~1hpi and returned to basal levels at 2hpi (Fig 8A). The same experiment performed following CQ treatment, which blocks lysosome acidification and stops the recycling of LC3-II showed LC3-II levels accumulating for the M1 virus but not for the WT virus suggesting that the autophagic flux in WT infected cells was impaired (S5A Fig). We next determined the ratio of autophagosomes to autolysosomes in WT and M1 infected cells at 1hpi. Our analysis showed that at 1hpi ~ 60% of LC3 positive structures in M1 infected cells were also positive for the lysosome marker Lamp2 indicating conversion into autolysosomes. In contrast in WT infected cells only ~40% were Lamp2 positive (Fig 8B and 8C) indicating a defect in autolysosome formation. To confirm that the WT impairs autolysosome formation we transduced cells with LC3 fused to GFP and RFP (LC3-GFP-RFP). Due to the pH-sensitivity of GFP, neutral autophagosomal membranes appear yellow while acidic autophagolysosomes appear red [35]. Quantification in WT and M1 infected cells at 1hpi confirmed a reduced level of autolysosome for the WT (S5B and S5C Fig). To understand how the WT prevents autolysosome formation we first ask if the virus controls the onset of autophagy. We induced non-selective autophagy through overnight starvation followed by infection with WT or M1 and quantified number and maturation state of LC3-positive structures at 1hpi (Fig 8D and 8E). Both viruses induced comparable levels of autophagy marked by an increase in LC3 positive punctae in starved cells (Fig 8D) while the ratio of lamp2 positive vs. lamp2 negative LC3-positive structures was reduced in WT infected cells vs. M1 (Fig 8E) showing that the WT controls autolysosome formation without impairing the onset of autophagy.
Autophagy promotes antigen presentation through fusion with MHC-II containing compartments including antigens incorporated into the AdV capsid [28,31]. Thus we asked if limiting autolysosome formation restricts AdV antigenic presentation. For this we infected mice with equal particle numbers of WT, M1 or PBS control. Ten days later, the mice were sacrificed and purified splenocytes were stimulated with either AdV capsids or GFP purified from E.coli. T-cell activation was measured by IFNγ ELISPOT (Fig 8F). We saw more IFNγ spot forming units (SFU) in splenocytes stimulated with AdV capsids from the M1 infected mice vs. WT infected mice, suggesting an increased display of AdV capsid antigens during M1 infection. An inverse effect was observed when the splenocytes were stimulated with purified GFP protein and compared to WT infected mice. Because the GFP is only expressed after adenovirus transduction of cells, this observation suggests that there is a defect in M1 infectivity in vivo. Additionally, we generated a human CD4+ T cell clone that responds to a conserved hexon epitope. We incubated these with syngenic APCs transduced with equivalent viral particles of either WT or M1, and measured T cell activation by IFNγ ELISA (Fig 8G). Again, we saw increased IFNγ levels from the T cells exposed to M1 vs. WT transduced APCs suggesting increased antigen presentation on MHC-II. No significant difference in MHC-II or CD86 expression was observed in human monocyte derived DCs between WT and M1 treated groups, suggesting that differences in the expression of these molecules between treatments did not account for differences in T-cell activation (S5D Fig). Taken together we show that AdV use the PVI PPxY motif to limit capsid epitope presentation on MHC-II presumably by restricting autolysosome formation.
While several Nedd4 ligases bind the PVI PPxY motif, we previously showed that only depletion of Nedd4.2 caused a transport defect and reduced infectivity [11]. To test if Nedd4.2 was also involved in AdV control of autophagy we infected Nedd4.2 depleted cells (S6A Fig) with WT virus and analyzed LC3-I conversion into LC3-II by western blot (Fig 9A). We observed increased and prolonged levels of LC3-II in infected cells following Nedd4.2 depletion, however, basal levels of LC3-II were also elevated. In individual cells (including non-infected cells) LC3 positive structures appeared more abundant, were larger in size and showed differences in subcellular distribution compared to control cells (Fig 9B and 9D). We next analyzed the percentage of autolysosomes by costaining LC3 with Lamp2 in control vs. Nedd4.2 depleted cells after AdV infection. In control depleted cells the proportion of autolysosomes in WT infected cells was reduced compared to M1 infected or non-infected control cells similar to our previous observation. In contrast upon depletion of Nedd4.2 all cells showed similar reduced levels of autolysosomes including M1 infected and non-infected control cells, without any further reduction in WT infected cells (Fig 9C). These results identified an important physiological role for Nedd4.2 in autolysosome formation and suggested that the WT virus interferes with Nedd4.2 to limit autolysosome formation. We next addressed the physiological role for Nedd4.2 in autophagy regulation in more detail. We analyzed autophagosome formation in Nedd4.2 depleted cells using starvation induced autophagy without viral infection. Under starvation conditions, depletion of Nedd4.2 prevented fusion of LC3 and Lamp2 positive structures (S6B and S6D Fig). The basal level of autophagosomes was higher in Nedd4.2 depleted cells compared to control cells. Starvation increased the number of autophagosomes in both, depleted and control cells (S6C Fig) showing that Nedd4.2 does not prevent autophagy induction. In contrast under both basal and starvation induced autophagy conditions the percentage of autolysosomes was lower in Nedd4.2 depleted cells than in control cells showing that Nedd4.2 is involved in autophagosome maturation into autolysosomes independent of virus infections (S6D Fig). We next investigated the change in subcellular distribution of autophagosomes upon infection. In WT infected cells at 1hpi LC3 positive structures clearly accumulated near the MTOC (Fig 9D) a distribution that is typical for AdV and other incoming viruses [7,38]. Interestingly, upon infection of Nedd4.2 depleted cells LC3 positive structures lost the MTOC accumulation and displayed a more perinuclear distribution (Fig 9D). We confirmed this observation by quantifying the proximity of LC3 positive structures to the MTOC (marked by pericentrin stain) in control and Nedd4.2 depleted cells at 1 hpi (Fig 9E). Because we previously observed a similar MTOC accumulation defect in Nedd4.2 depleted cells with WT viruses [11], we next asked if autophagy contributes to nuclear transport of AdV.
To address a possible role of autophagy in AdV nuclear transport we first infected cells with WT and M1 and analyzed the subcellular distribution of LC3 positive structures at 1hpi. This experiment should distinguish if MTOC accumulation of LC3 positive structures required recruitment of Nedd4.2 towards WT viruses or, in the case of the M1 virus, if availability of Nedd4.2 without capsid recruitment via protein VI was sufficient for MTOC targeting of LC3 (Fig 10A). WT infected cells showed increased MTOC accumulation of LC3 positive structures compared to M1 infected cells, which we confirmed by MTOC proximity quantification (Fig 10B). Because in vivo imaging of fluorescent viruses in LC3-GFP expressing cells often displayed virus mobility in association with LC3 (S7 Fig and S2 Movie) we asked if the autophagic machinery plays a more direct role in WT MTOC targeting. We repeated the above assay in ATG5 depleted cells and quantified the virus distribution around the MTOC. Most cells showed reduced MTOC accumulation for WT viruses in ATG5 depleted cells (Fig 10C), which we confirmed by MTOC proximity quantification. In contrast the more random distribution of M1 was not affected by ATG5 depletion (Fig 10D). Because the WT virus distribution in ATG5 depleted cells was very similar to the distribution of the M1 virus we asked if the endosomal escape was still functional. For this we compared PVI release between WT and M1 in ATG5 depleted cells vs. control cells as a measure of endosomal escape (Fig 10E). The analysis showed that ATG5 depletion did not impact on the initial release of PVI from either WT or M1 but delayed the separation of PVI from the WT (Fig 10F) showing that efficient endosomal escape requires ATG5.
Nuclear genome delivery can be considered as endpoint of AdV entry, which can be quantified by accumulation of genome associated protein VII dots in the nucleus [39] (Fig 10G). We quantified nuclear import of adenoviral genomes over time in ATG5 depleted and control cells for both viruses. Our analysis showed an initial delay for WT genomes in ATG5 depleted cells at 1 hpi that was fully compensated at 2 hpi confirming that fast nuclear genome transport relies on ATG5 while compensatory transport mechanisms exist (Fig 10G, left panel). It also confirmed that the M1 mutant virus was subject to autophagic degradation because ATG5 depletion increased nuclear genomes compared to control depleted cells suggesting that more virion particles survive the entry process and are able to deliver their genome. Nevertheless, nuclear genome arrival in ATG5 depleted cells showed the same delay as WT genomes (Fig 10G, right panel).
Taken together our study demonstrates how upon entry adenoviruses use a capsid encoded PPxY-peptide motif to exploit and control the autophagic machinery at multiple levels to secure efficient genome delivery.
Many invasive pathogens challenge the cellular membrane integrity to access the cytosol. Cells are able to sense pathogen invoked membrane damage via out-of-place detection of intra-lumenal glycans exposed to the cytosol. Thurston and co-workers showed that several galectins as well as poly-ubiquitin serve as danger sensor for ruptured membranes but only Gal8 via its specific adapter NDP52 restricts bacterial proliferation targeting the damaged vesicles for autophagic degradation [15]. Here we show that cells respond via a similar principle to incoming AdV demonstrating that non-enveloped endosomolytic viruses also activate selective autophagy through membrane damage. Under our assay conditions we show that release of the membrane lytic PVI from internalized viruses induced the transient association with galectins (3,8 and 9), ubiquitin, adapter molecules (p62 and NDP52) and LC3 within minutes of virus uptake. Using the control TS1 mutant virus, which lacks PVI release, we were able to pinpoint the membrane damage as the initiating event that causes this response. While WT viruses rapidly escaped from the ruptured endosome, and accumulate at the MTOC, the mutant M1 virus lacking a conserved PPxY motif in PVI remained associated with the ruptured (Gal3, Gal8 and PVI positive) vesicle and was subject to autophagic degradation. EM analysis revealed that during this process autophagosomal membranes engulf M1 viruses still associated with ruptured endosomes. Accordingly degradation through autophagy was able to explain the M1 infectivity defect and autophagy inhibition restored M1 infectivity. Thus we identified the capsid encoded PPxY motif as molecular determinant that allows the WT virus to subvert the cellular antiviral response providing the first example of a virus encoded motif to combat cellular antimicrobial autophagy. M1 virus removal occurred through a pathway involving Gal8 detection of virus induced membrane damage corroborating that autophagy recruitment through Gal8 is part of a conserved cellular pathway for the detection and removal of membrane damage evoked by microbes as previously suggested [15]. Gal3, which was also recruited to AdV ruptured membranes was not restricting the M1 mutant virus despite its recently described link to autophagy [40]. Depletion of the Gal8 binding autophagy receptor NDP52 as well as p62, which are both involved in restricting invasive bacteria was much less effective than autophagy inhibition or Gal8 depletion in protecting M1 from degradation. This observation suggests the existence of additional and/or alternative pathways to link AdV induced membrane damage via Gal8 towards autophagy clearance. These pathways may include the use of alternative autophagy receptors such as TAXBP1 recently shown to be involved in Salmonella clearance although no Gal8 binding was reported [41].
Our work with the M1 virus underpins the power of selective autophagy as antimicrobial mechanisms and shows how AdV have evolved capsid encoded evasive mechanism. This may not be true for all endosomolytic viruses. A recent report showed that picornavirus induced membrane damage also activates Gal8 mediated autophagy. In this case however the virus uses a cellular factor, the lipid-modifying enzyme PLA2G16, to counteract selective autophagy [42]. Remarkably, escaping endosomes would suffice for AdV to evade autophagic degradation. Nevertheless we observe that the WT virus uses the PPxY motif in PVI also to prevent efficient formation of autolysosomes. This effect is not linked to the onset of autophagy because LC3 punctae are induced and LC3-II conversion takes place upon WT infection. However, unlike with the M1 virus, we see much less large autophagosomal structures in WT infected cells and LC3-II levels rapidly return to basal levels. One possible explanation is that the WT virus interferes with the elongation process of the autophagosomal membrane to prevent autophagosome formation. This would make sense because only fully formed autophagosmes can fuse with lysosomes [43]. Depletion of Nedd4.2 removes the ability of the WT to interfere with autolysosome formation showing that the recruitment of Nedd4.2 through the PPxY motif in capsid protein PVI is central to AdV autophagy evasion. Murine Nedd4.2 was recently identified as factor that promotes autophagy using knockdown approaches [44]. In our hands Nedd4.2 depletion increases basal autophagy levels and also inhibits autolysosome formation independently of infection suggesting a key role for Nedd4.2 in physiological autophagy regulation. This observation is supported by recent work showing that Nedd4.2 is involved in controlling ULK1 levels under stress conditions to limit autophagy [45]. Diverting Nedd4.2 from its physiological role could therefore be a major function of the PVI PPxY motif. Another Nedd4 ligase, Nedd4.1, was recently shown to also regulate autophagy [46]. In addition, Nedd4.1 (but not Nedd4.2) was shown to control cellular Beclin-1 levels, a subunit of the class III phosphatidylinositol 3-kinase complex, which is crucial for the phagophore formation and membrane elongation [47,48]. We observe a stronger association of the M1 virus with PI3P positive membranes coinciding with enrichment of Beclin-1 on cellular membranes, suggesting that the WT selectively alters the function of the class III PI3K complex during entry although we were unable to functionally link this observation to Nedd4.2. Still, this seems to be a common viral strategy as several enveloped viruses target Beclin-1 and the class III PI3K complex to interfere with autophagic processes [49] including autophagosome maturation [50].
One attractive reason for viruses to control autolysosome formation is to limit antigenic presentation via the MHC-II pathway, which can be fed through autophagy [51]. E.g. Herpes simplex virus (HSV-1) employs the viral γ34.5 protein to bind and inactivate Beclin-1 to limit antigenic presentation [52]. AdV down regulates the antigen presenting machinery (MHC-I) During replication via regulatory E3 proteins [53,54]. Currently there is not a known mechanism underlying control of MHC-II presentation once transcription has begun although it might be an important obstacle. Our observations would suggest that using a capsid-encoded determinant, the PPxY motif, allows viral control of MHC-II antigen presentation upon entry preceding viral gene expression, which gives the virus an advantage over its host. In turn it identifies M1 capsids as potential tools for enhanced MHC-II exposure, which maybe of relevance for vaccination or other therapeutic approaches.
Another important observation in our study is the role of autophagy in virus transport towards the nucleus. We show that MTOC targeting for the WT requires dynein motors, Nedd4.2, the PVI PPxY motif and the LC3 conjugation system (ATG5). Removal of either component prevents MTOC accumulation, including a delay in genome delivery presumably because the WT becomes entrapped at the site of membrane rupture similar to the M1. WT and M1 virus do not differ in the kinetics with which they release PVI or acquire membrane damage sensors (Gal3, Gal8) nor does the turnover of membrane damage appear accelerated in either case. An attractive explanation for our observations would be that the LC3 conjugation system aids in motor recruitment towards the entrapped WT virus permitting efficient endosomal escape while Nedd4.2 recruitment via the PPxY motif in PVI interferes with the autophagosome maturation process initiated by the detection of damaged endosomes. A direct role for LC3 in microtubule transport during autophagosomal maturation and trafficking was recently established [21–23]. Thus PVI exposure could provide a higher affinity-binding site for Nedd4.2, which might redirect LC3 acquisition by molecular motor or motor scaffolding proteins for autophagic transport and maturation in favor of viral escape. Escape of the virus would not end the autophagosome maturation defect because (some) PVI remains within the ruptured membrane although the exact fate of the ruptured membrane including PVI after viral escape remains unclear at this stage.
Our preliminary assessment using in vivo imaging analysis suggest that WT virus may move from the membrane penetration site to the MTOC in association with LC3. Furthermore the depletion of LC3 conjugation impairs the MTOC accumulation of viral particles. It is thus possible that AdV acquire LC3 for its own transport to the nucleus. While recently autophagosomes where suggested to support AdV endosomal escape by fusing with endosomes [27] future work has to clarify if LC3 plays a more active role during endosomal escape e.g. as part of the motor complex, which extracts the virus from the endosome and/or as part of the complex that drives cytosolic transport of escaped viruses. Such a tentative model as detailed in Fig 11 would accommodate our observation that the WT on one hand limits autolysosome formation upon membrane damage while at the same time depletion of ATG5 prevents efficient endosomal escape and nuclear transport suggesting a reprogramming of autophagy to benefit the virus. It would also be in agreement with the observation that subsequent transport and genome delivery is delayed (but not abolished) in ATG5 depleted cells without affecting the overall accumulated infectivity measured at 24hpi. Delayed nuclear transport could also (in part) explain our previous observation how protein VI can affect transcription initiation of incoming viral genomes [55].
In summary our work has highlighted how a minimum of genetic conservation through preserving a short capsid encoded peptide motif allows incoming viruses (and presumably other pathogens) to subvert host cell defense mechanisms to profit at multiple layers. In this specific case the accumulated benefits for AdV include limited antigenic presentation paired with accelerated genome delivery showing that autophagy may act as a pro-viral mechanism upon entry of non-enveloped viruses. This is also the first direct demonstration how a viral pathogen overcomes the cellular response to membrane breach. Given the high prevalence of capsid encoded PPxY motifs in other viral systems (including HSV-1) [56–58] and a common association of viruses with the MTOC in entry it may be worthwhile to investigate if diverting Nedd4.2 during entry is part of a broader viral strategy for immune evasion and transport, which would make for an excellent drug target.
All autophagy related assays were done in accordance to the “guidelines for the use and interpretation of assays for monitoring autophagy” (3rd Edition) [34]
All cells were grown under standard conditions (detailed in SI). ATG5 -/- and control MEFs were a kind gift from R. Duran, Institute Bergonié, Bordeaux, France. Amplification of human recombinant Ad5-VI-WT, Ad5-VI-M1 viruses and their E1-deleted GFP-transgene expressing vector counterparts (including Ad5-ts1-GFP) was done in HEK293 (Human Embryonic Kidney 293 cells, ATCC CRL-1573, kindly provided by G Nemerow, Scrips research institute, La Jolla, USA) cells and purified using double CsCl2-banding [10,11]. Virus particle to cell ratios were calculated based on the estimated copy numbers of viral genomes. Copy numbers were calculated according to the OD260 method (1 OD260 = 1.16×1012 particles/ml) [59]. Viruses were labeled by using the Alexa Fluor Microscale labeling kit (LifeTechnologies) as detailed in [26].
Plasmid transfections was done in 12-well dishes using 1×105 U2OS (Human bone Osteosarcoma Epithelial cells, ATCC HTB-96, kindly provided by M. Piechaczyk, IGMM, Montpellier, France) cells using Lipofectamine 2000 and OptiMEM (LifeTechnologies) according to the manufacturer’s instructions. RNAi mediated depletions were performed in 12-well dishes using 5×104 U2OS cells. Following optimization cells were transfected after 24 and 48 h with 20 to 50 pmol of each siRNA constructs using Lipofectamine RNAimax and OptiMEM (LifeTechnologies). Depletion using shRNA-encoding lentiviruses was done by overnight incubation of 1×105 U2OS cells (in 12-well dishes) with lentiviral particles at MOI of 5. The following day lentiviruses were removed and replaced by fresh complete DMEM and transduced cells were selected 24h later by adding puromycin (2μg/ml, Invitrogen). Sequences are listed in the SI.
Cell transduction with adenoviral vector was performed using 1×105 U2OS cells seeded into 24-well plates. Cells were pre-treated with 100 μM CilioD for 30min or with 10mM of 3’MA or 50μM of Chloroquine for 3h or DMSO and DMF as vehicle control. Cells were transduced with 50 physical particles per cell of either GFP expressing WT or M1 in the presence of drugs followed by medium replacement with drug-free medium after 3h. Cells were analyzed 24h later by flow cytometry for GFP expression. Acquisitions were done on a FACSCantoII cytometer (BD Biosciences) and the data were processed and analyzed by the FACSDIVA software (BD Biosciences). U2OS depleted cells (siGAL3/8/9, siCTRL, shATG5/p62/NDP52 and shCTRL) were transduced following the same procedure.
For western blot analysis cell lysates were separated on size-resolution adapted SDS-PAGE and transferred to nitrocellulose membranes (cut off of 0.2μm). Extraction of membrane proteins was done using the Mem-PER™ plus membrane protein extraction kit (ThermoFisher). Membranes were blocked in TBS containing 10% of dry-milk and 0.01% of Tween 20 (Sigma) during one hour at room temperature, followed by over-night incubation at 4°C with primary antibodies and with HRP-conjugated secondary antibodies against rabbit, goat or mouse (Sigma) at a dilution of 1∶10 000 for 1 one hour at room temperature. Specific signals were revealed using the enhanced chemiluminescence detection system (Super signal West femto, Thermoscientific) and signals were acquired using an ImageQuant LAS 4010 system (GE Healthcare life Sciences). Antibodies are listed in the SI.
For immunofluorescence analysis 1×105 U2OS were grown on coverslips in a 12-well dish. Coverslips were incubated with 150μl of complete DMEM containing viruses (250 physical particles per cells) during 30 min at 37°C. Then viruses were removed and replaced by complete DMEM. Coverslips were fixed with 4% PFA at each time point (or in Methanol during 20 minutes at -20°C, for LC3 staining). Cells were blocked/permeabilized with IF-buffer (10% FCS in PBS and 0.5% Saponin). Primary antibody and secondary antibodies where applied to the coverslip in IF-buffer for 1 h at 37°C. Cells were mounted in DAKO mounting media containing DAPI and analyzed by confocal microscopy. Antibodies are listed in the SI.
Confocal images were taken on a Leica SP5 confocal microscope equipped with Leica software and analyzed as detailed in SI. For quantitative image analysis we analyzed for each condition/time point at least 10 cells with >50 virus particles each (n>500). Live cell imaging was performed as described previously [39]. Briefly, cells were seeded in ibidi μ-slide VI0.4 (Ibidi), and images were acquired using a Leica spinning-disk microscopy system equipped with an environmental chamber heating the whole optical system to 37°C. Frames were taken every second (x100 objective) for each color channel and recorded using MetaMorph software and assembled in ImageJ.
U2OS cells were grown to 70% confluency in 35mm glass bottom Grid-500 μ-Dish (Ibidi, Cat.No. 81168) and infected with Alexa594 fluorescently labeled WT or M1 viruses (250 physical particles per cell) at 37°C for 30 min. At 30 min post infection cells were fixed with 2.5% glutaraldehyde (GA) in PBS overnight at 4°C. Subsequently, the cells were washed with PBS, postfixed for 30 minutes with 1% OsO4 in PBS, washed with ddH2O, and stained with 1% uranyl acetate in water. The samples were gradually dehydrated with ethanol and embedded in Epon resin (Carl Roth, Germany) for sectioning. Ultrathin 50 nm sections were prepared using Ultracut Microtome (Leica Microsystems, Germany). The sections were poststained with 2% uranyl acetate. Electron micrographs were obtained with a 2K wide angle CCD camera (Veleta, Olympus Soft Imaging Solutions GmbH, Münster, Germany) attached to a FEI Tecnai G 20 Twin transmission electron microscope (FEI, Eindhoven, The Netherlands) at 80kv.
0.5-10x106 monocyte-derived dendritic cells were derived from healthy anonymous donors and transduced with virus or media alone for 2 hours. Cells were incubated with CD4+ T-cell clones that had been cultured for 4 weeks to recognize a conserved AdV hexon epitope at a 3:1 ratio overnight. Supernatants were collected and ELISA was performed according to manufacturer’s instructions (eBioscience). For ELISpot assays, B6 mice (The Jackson Laboratory, Barr Harbor, ME) were infected with 1010 viral particles, or PBS control, intramuscularly. Mice were sacrificed 10 days after infection, and whole spenocytes were collected. Splenocytes were plated into ELIispot plates (Millipore) and stimulated with Ad5-luc capsids or GFP purified from E. coli for 48 hours. ELIspot was performed per manufacturer’s instructions (Millipore) and spots were reported as the number of spot forming colonies (SFC)/10^5 PBMCs.
Human monocyte-derived dendritic cells were cultured from peripheral blood mononuclear cells (PBMCs) obtained from healthy donors. Briefly, PBMCs were isolated from fresh whole blood by centrifugation on Histopaque cell separation medium. Monocytes were selected by adherence to culture dishes for 1 h, washed, and cultured in Iscove's modified Eagle's medium plus 10% FBS plus 25 ng/ml recombinant human GM-CSF (PeproTech) for 7 days to obtain differentiated dendritic cells. Human CD4+ T-cells recognizing a conserved epitope in adenovirus hexon were cloned as described previously [60]. Briefly, PBMC were suspended in 200 μl R10 media and incubated with 100 μM peptide for 1 h. Cells were diluted to 2 ml with R10 and incubated in 24-well plate at 37 C in a 5% CO2 atmosphere. After 7 days, human recombinant IL-2 (Becton Dickinson, Bedford, MA, USA) was added to a final concentration of 20 U/ml and then every 3–4 days. T-cell clones (1x10(6) per well) were restimulated with peptide-loaded, irradiated autologous PBMC (3000 rad) or (x10(6) per well) every 7–10 days. CD4+ T-cells were further isolated by negative selection magnetic cell sorting using a commercially available kit (Miltenyi Biotec, cat# 130-096-533).
All experiments using animals and human cells were conducted in accordance with the guidelines and under approval of the IACUC of Loyola University Chicago Stritch School of Medicine (#2011020) and the Loyola University Chicago Stritch School of Medicine Institutional Review Board (IRB # is LU204021) in accordance with guidelines set forth by the USDA and PHS Policy on Humane Care and Use of Laboratory Animals under the guidance of the Office of Laboratory Animal Welfare (OLAW). Loyola University Chicago, Health Sciences Division has an Animal Assurance on file with the Public Health Service (#A3117-01 approved through 02/28/2018), is a fully AAALAC International accredited institution (#000180, certification dated 11/19/2013), and is a USDA registered/licensed institution (#33-R-0024 through 08/24/2017). Loyola University Chicago, Health Sciences Division’s Institutional Animal Care and Use Committee (IACUC) is responsible for reviewing all protocols involving living vertebrate animals ensuring compliance with federal regulations, inspecting animal facilities and laboratories and overseeing training and educational programs.
Data are presented as mean, error bars as standard deviation (STD) or standard error (SE) as indicated in the Fig. legend. Statistical analysis if not indicated otherwise was done using unpaired students t-test (NS: no significant, *:P<0.05; **:P<0.01; ***:P<0.001; ***: P<0.0001).
|
10.1371/journal.pntd.0004432 | Rabies in the Baltic States: Decoding a Process of Control and Elimination | Rabies is a fatal zoonosis that still causes nearly 70, 000 human deaths every year. In Europe, the oral rabies vaccination (ORV) of red foxes (Vulpes vulpes) was developed in the late 1970s and has demonstrated its effectiveness in the eradication of the disease in Western and some Central European countries. Following the accession of the three Baltic countries—Estonia, Latvia and Lithuania—to the European Union in 2004, subsequent financial support has allowed the implementation of regular ORV campaigns since 2005–2006. This paper reviews ten years of surveillance efforts and ORV campaigns in these countries resulting in the near eradication of the disease. The various factors that may have influenced the results of vaccination monitoring were assessed using generalized linear models (GLMs) on bait uptake and on herd immunity. As shown in previous studies, juveniles had lower bait uptake level than adults. For the first time, raccoon dogs (Nyctereutes procyonoides) were shown to have significantly lower bait uptake proportion compared with red foxes. This result suggests potentially altered ORV effectiveness in this invasive species compared to the red foxes. An extensive phylogenetic analysis demonstrated that the North-East European (NEE) rabies phylogroup is endemic in all three Baltic countries. Although successive oral vaccination campaigns have substantially reduced the number of detected rabies cases, sporadic detection of the C lineage (European part of Russian phylogroup) underlines the risk of reintroduction via westward spread from bordering countries. Vaccine induced cases were also reported for the first time in non-target species (Martes martes and Meles meles).
| This paper reviews ten years of rabies epidemiology in the three Baltic countries. Both surveillance efforts and oral rabies vaccination campaigns have resulted in the near eradication of the disease. Multivariate analysis assessed with generalized linear models (GLM) suggested lower oral vaccination effectiveness in raccoon dogs compared with red foxes, highlighting the importance of adapting oral vaccination strategy to each vector of the disease. Although eradication of the disease is almost achieved, the detection of some cases belonging with the Russian rabies lineage emphasizes a risk of rabies reintroduction in the Baltic States due to westward spread from bordering countries. This study show also the first vaccine-induced cases detected in non-target species (Martes martes and Meles meles).
| Rabies disease is a fatal mammalian encephalomyelitis caused by the rabies virus of the genus Lyssavirus (family Rhabdoviridae) [1]. The virus is distributed worldwide, with the exception of the Antarctic, Australia and several islands and although all species of mammals are susceptible to this virus, it infects principally carnivores and bats [2]. In Europe, the genus lyssavirus evolves through five virus species (four of them circulate in bats only): the classic rabies virus (RABV) affecting non-flying terrestrial mammals only, the european bat lyssaviruses type 1 and type 2 (EBLV-1 and EBLV-2) and the more recently detected Bokelo bat lyssavirus (BBLV) and Lleida bat lyssavirus not yet taxonomically assessed [3].
RABV has spread in Europe since antiquity as a dog and wolf-mediated disease [4]. In the 1940s, likely due to spillover from domestic animals, a new epizootic maintained by a single species, the red fox, emerged in Eastern Europe with an assumed ground zero in Kaliningrad [5].The front moved from Poland to Germany spreading through Europe with a speed of approximately 30–60 km per year, reaching France in 1968 and Italy in 1980 [6]. Large rivers, lakes and high mountain chains acted as obstacles to the spread; bridges facilitated the crossing of rivers. Intensive fox destruction campaigns alone cannot stop the spread of the virus [7], prompting oral rabies vaccination (ORV) programs that rapidly proved to be the only efficient technique for controlling the disease. The first ORV field trial was conducted in 1978 in Switzerland [8] and was gradually extended to surrounding countries, such as Belgium, France and Germany. In the 1980s, fox rabies control in European Union became a public health issue. Since 1989, the European Commission has provided funding to Member States for national eradication programmes, thereby improving surveillance and encouraging regular implementation of oral vaccination campaigns on large scales in coordination with neighbouring countries. This strategy leaded to the successful elimination of rabies in most Western and Central European countries [9,10]. In Europe, approximately half of the historical rabies endemic countries are now free of rabies (Austria, Belgium, Czech Republic, Finland, France, Germany, Italy, Luxembourg, Switzerland and the Netherlands). In the Baltics, the three countries were recently officially recognized free of rabies according to OIE (World Organisation for Animal Health) criteria [11–13]. In the last three years, some sporadic cases have been reported in some countries (Bulgaria, Hungary, Slovakia and Slovenia) and the disease is still endemic in several Eastern European countries (eastern Poland, Romania, Ukraine, Belarus and Russia, source: http://www.who-rabies-bulletin.org/Queries/Surveillance.aspx).
In the Baltic States, represented by Lithuania, Estonia and Latvia, sylvatic rabies emerged in the 1950s-1960s [14]. Since this time, a surveillance of the disease were progressively implemented and positive cases have been observed mainly in red foxes and raccoon dogs [15–17]. Although the red fox is known to be highly susceptible to RABV and is the main reservoir and vector of rabies throughout Europe, the Baltic countries has the particularity to host a second vector and reservoir, i.e. the raccoon dog [14]. Raccoon dog is one of the most successful alien carnivores in Europe. Native to East of Asia, this species was introduced in the eastern part of Russia via fur industry during the first half of the 20th century and has spread throughout Europe, becoming common in the Baltics and some other northeastern European countries. After it was first observed in the 1950’s in the Baltics, ten years were required to colonize the entire countries [18]. Foxes and raccoon dogs are both opportunistic omnivores, often share the same habitats and overlap their home ranges increasing the probability of contacts between the two species. Moreover, their combined densities could allow rabies epizootics to persist in a certain area [19]. The existence of this important rabies transmitter in this area challenged health authorities and questioned on its potential impact on the success of conventional ORV method used to control rabies in Western Europe.
ORV programs were experimented differently according to the Baltic State. While no ORV was implemented in Estonia until 2005, in Latvia, ORV was firstly initiated in 1991 using chicken head vaccine baits. ORV using manufactured baits started in 1998 and has been performed twice a year since 1999, but regular purchase of the necessary amount of vaccine baits for annual nationwide vaccination was not possible because of financial reasons. The vaccination area was enlarged every year to cover the whole territory by 2001–2003 and vaccines were distributed manually [20]. In Lithuania, ORV was tested for the first time in 1983 with fish or meat baits containing a vaccine made of a derived ERA (Evelyn Rokitnicki Abelseth) laboratory fixed virus strain produced in Russia. In 1993 ORV was occasionally assessed on three districts [21]. Between 1995 and 2000, following the Lithuanian National Rabies prevention programme, ORV was performed generally manually and a large range of vaccines was used (Street Alabama Duffering (SAD) Bern, SAD P5/88 (Rabifox), (Street Alabama Gif (SAG) 1) over variable geographic areas.
Following the accession of three Baltic countries to the European Union in 2004, subsequent financial support allowed the implementation of regular oral vaccination campaigns in the three countries since 2006 and ORV are still ongoing. This paper reviews ten years of surveillance efforts and oral vaccination campaigns conducted in the frame of European Commission programmes. Through the epidemiological analysis of rabies surveillance in these countries and an in-deep analysis of the ORV monitoring results, this paper emphasizes determinants of success and draws lessons for the future. These findings could provide valuable insights into the strategy required for rabies elimination and may help guide future implementation of oral vaccination programmes.
Covering approximately 175,000 km2, the Baltic States lie in the northeastern part of Europe and comprise the countries of Estonia (45,227 km2), Latvia (64,589 km2) and Lithuania (65,303 km2) (Fig 1). The Baltic States are bounded on the west and north by the Baltic Sea, which gives the region its name, on the east by Russia (511 km of common border), on the southeast by Belarus (818 km), and on the southwest by Poland (104 km) and an exclave of Russia named Kaliningrad (255 km). The topography of this area is relatively flat (culminating points in the three countries are around 300 m), characterized by numerous lakes and ponds, especially in the north, and hills in Lithuania. The most commonly encountered landscape is the temperate forest covering between 35 and 50% of the territories.
All suspect non-flying mammals exhibiting clinical signs suggestive of rabies or showing abnormal behaviour, animals found dead in the field including road kills and those to which humans have been exposed (bites, scratches, licking of wounds or contamination of mucous membranes with saliva) are defined as indicator animals and are submitted for diagnosis [19]. The sampling scheme focusing on these animals, covering the whole country territory, is herein defined by expert committees of the WHO (World Health Organization) and EFSA (European Food Safety Authority) as the surveillance system [2,19]. All collected samples were shipped and analyzed in the respective National Reference Laboratories of each Baltic country (Estonian Veterinary and Food Laboratory for Estonia; Institute of Food Safety, Animal Health and Environment "BIOR" for Latvia and National Food and Veterinary Risk Assessment Institute of Lithuania for Lithuania). Brain tissues were analyzed for viral antigens using the Fluorescent Anti-body Test (FAT) which is the gold standard technique for rabies diagnosis [22,23]. For all three countries, FAT-negative results of animals involved in human exposure and FAT-inconclusive results were confirmed using the rabies tissue culture infection test (RTCIT) [24], Reverse Transcription Polymerase Chain Reaction (RT-PCR) [25] or Real-Time Polymerase Chain Reaction (RT-qPCR) [26,27].
The first wildlife ORV campaign in Estonia was organized in autumn 2005 and covered 57% of Estonian lands in the northern part of the country as part of a PHARE Twining Light Project (Fig 1) [16]. Vaccination programmes covering the entire territory (excluding urban areas, roads, water bodies and wet fields) representing approximately 43,000 km2 were carried out from 2006 to 2010. Bait distribution was performed twice a year, in spring (May, early June) and in autumn (September, October) as recommended by WHO and EFSA [2,19]. Baits were distributed at a rate of 20 baits per km2 using small fixed-wing aircraft flying at an altitude of 100–150 m, speeds of 150–200 km/h and in parallel flight lines (global positioning system (GPS) routes followed by the plane) distanced of 600 m [16,28]. Since spring 2011, ORV campaigns have been conducted only in a buffer zone of 9,325 km2 adjacent to neighbouring infected countries (Russia and Latvia) to ensure a sufficient level of immunity among raccoon dog and red fox populations. No automatic dropping device was used in the airplanes and no additional manual distribution was carried out in the field. A single vaccine bait type was selected through a tendering process, the modified live attenuated SAG2 vaccine (RABIGEN, Virbac Laboratories, France) (Fig 2).
In Latvia, following a PHARE Twinning Light project, ORV campaigns were carried out in 2005 for the first time via aircraft in half of the country (the size of vaccination area was 28,000 km2 and it was delimited by natural barriers) twice a year with 23 baits per km2. Starting from 2006, two vaccination campaigns were implemented in the entire territory (64,589 km2) (except in 2008 and autumn 2011 when ORV campaigns were incomplete and in spring 2014 where no ORV was carried out). Since 2006, between 21 and 31 baits per km2 have been distributed using flight line distances of 1000 m until 2008, 1000 m and 500 m alternately between 2008 and 2011, and 500 m since 2011. An automatic dropping device has been used since 2007 to distribute the baits. The type of vaccine purchased varied according to the procurement procedure. In general, two vaccines were used within the period 2005–2011—SAD B19 vaccine (FUCHSORAL, IDT Biologica GmbH, Germany) and SAD Bern (LYSVULPEN, Bioveta, Czech Republic). Since 2012, only the Lysvulpen vaccine has been in use (Fig 2).
In Lithuania, ORV programmes have been implemented since 2006 using aircrafts over the whole country (65,000 km2) except lakes, urban areas and the Ignalina nuclear power-station. The no-fly area surrounding the Ignalina power plant was covered by manual distribution of baits. Like in other countries, the vaccination strategy has been implemented biannually (one vaccination in spring between March and May and one vaccination in autumn between October and December). Parallel flight lines generally separated by 1000 m (since 2011, 500 m in areas on the Belarus border) at an altitude of 150–200 m and speed 150–200 km/h were used to distribute 20 baits per km2 [29] (Fig 2). Since 2006, only the Lysvuplen vaccines have been distributed except in 2011 and 2012 when Fuchsoral vaccines were used.
In addition to the sampling scheme designed for rabies surveillance, a second sampling plan defined as monitoring of ORV was set up in vaccinated areas to evaluate the efficacy of ORV campaigns in terms of bait consumption (bait-uptake) and herd immunity [2,19,30]. This sampling focused on the collection of animals (red foxes and raccoon dogs) targeted by oral vaccines. These animals sampled by hunter associations are therefore considered as not suspected for rabies. Herd immunity level was assessed by enzyme-linked immunosorbent assays (ELISAs) [31]. Two commercial anti-rabies ELISA kit were used within the study: the BioPro ELISA (BioPro, Czech Republic) and the Platelia Rabies II kit (Bio-Rad, France). Their technologies differ by their coating aspect. The BioPro ELISA is a blocking ELISA using the crude glycoprotein to coat the plates and a positivity threshold (expressed as a percentage of blocking) of 40% [32,33]. The Platelia Rabies II kit is an indirect test using a purified rabies glycoprotein for the coating [34]. Serum titers were expressed as Equivalent Units per milliliter (EU/mL) with a cut-off of positivity fixed at 0.5 EU/mL in Estonia and Lithuania and 0.125 EU/mL in Latvia. The BioPro Rabies ELISA Ab kit was used in Latvia only.
Bait uptake was investigated by collecting red fox and raccoon dog jaws and by analysing the tetracycline (TTC) specific fluorescence in thin tooth sections under ultraviolet light [35,36]. Indeed, after its inclusion in the coating of the bait and its consumption by the targeted animal, the tetracycline molecule, used as a bait uptake marker, is incorporated into bones and teeth. This interaction creates a line that can be detected using epi-fluorescence microscopy. Each animal sampled for monitoring were analyzed for both serological analysis and tetracycline detection when possible (depending of the organs let intact by the shot of the hunter).
Studied animals from surveillance and monitoring scheme were originated from the field, died of natural causes and during the hunting/vaccination program developed and launched by the ministry of each country. These sampling processes were realised in compliance with the legislation of each country and under the recommendations of international institution (WHO [2] and EFSA [19]). In Europe, such process does not require any specific ethical approval as animals are received only dead in laboratories. Hunting plans are organised in the frame of control programmes of the disease and organised by Member States.
A panel of 165 field rabies viruses was collected in Baltic countries between 2004 and 2013 for this study. The isolates investigated from domestic and wild animals were extracted from brains of animals samples in Estonia (n = 43), Latvia (n = 42) and Lithuania (n = 80). The samples were isolated from 12 different wild and domestic animal species: Nyctereutes procyonoides (65), Vulpes vulpes (44), Canis familiaris (14), Bos taurus (11), Felis catus (ten), Procyon lotor (four), Meles meles (three), Equus caballus (two), Martes martes (two), Ovis aries (one), Canis (one), Lynx lynx (one) and six non-determined species (S1 Table). The samples were initially tested using the FAT prior to genetic characterization [23].
For all the samples, forward and reverse sequences were assembled and edited using the ContigExpress program of Vector NTI software, version 11.5.3. (Invitrogen, France). Alignments were edited using Genedoc software, version 2.7.000 [41]. The same software was used to translate the gene sequence. Percentage identities and similarity scores were determined using the BIOEDIT program version 7.2.5. [42]. After the alignment of sequenced amplified PCR products, 106 identical sequences (56 from Lithuania, 23 from Latvia and 27 from Estonia) showing 100% nucleotide identity for the N gene (460 nt) were removed from the phylogeny. Fifty-nine partial N gene sequences (24 from Lithuania, 19 from Latvia and 16 from Estonia) were available for subsequent analysis. The dataset contained 93 sequences (361 nucleotides, positions 109 to 470 compared with the challenge virus standard (CVS)-11 strain GenBank no. GQ918139). Fifty nine representative Baltic samples, eight isolates from neighbouring countries (six from Poland and two from Russia), two from Ukraine, seven from Europe, two fixed strains (D42112 and HQ829841), three representatives of rabies vaccine strains (EF206708, EF206709 and EF206719), six referenced Artic and Artic-like isolates and six reference strains used as outgroup were included in the dataset (S1 Table).
A total of 24,919 animals were diagnosed for rabies from 2005 to 2014 in the Baltics. Around 70 to 80% of all detected positive cases were found in red foxes and raccoon dogs (For Estonia, 35% foxes and 48% racoon dogs; for Latvia, 40% foxes and 30% raccoon dogs; for Lithuania 31% foxes and 40% raccoon dogs). In the three countries, the maximum number of detected rabies cases was observed during the 2005–2006 period (Fig 2). The highest number of detected cases was recorded at the same semester of the implantation of the first ORV in Estonia and Latvia, while one semester after the first ORV in Lithuania. The ORV induced indisputably a decrease of the number of positives cases in the three countries (excepted in Lithuania between the second semester of 2006 and the first semester of 2007). Regarding the maximum number of cases observed in each country, 90% reduction of detected cases was reached after two ORV campaigns in Estonia, eight in Latvia and four in Lithuania. The last rabies case (field strain) was notified in 2011 in Estonia, in 2012 in Latvia and in 2013 in Lithuania. When starting ORV, surveillance effort (number of indicator animals sampled per 100 km2 in the whole country) ranged from 1.7 to 1.3 in Estonia (2005 and 2006), 1.7 to 1.6 in Latvia (2005 and 2006) and 5.8 in Lithuania (but caution must be taken in the interpretation of this number because animals collected for monitoring were also included for this latter country). Since no rabies cases were detected, the pressure of surveillance appeared also comparable between the three countries ranging from 0.42–0.30 in Estonia (2012–2014), 0.42–0.30 in Latvia (2013–2014) and 0.48 in Lithuania (2014). These data thus support the comparability of the number of positive cases in the different countries in recent years.
The two phylogenetic analyses of the partial N gene sequences performed using either PhyML or Mr Bayes produced trees with identical topology. The phylogenetic analysis showed that 163 of the 165 studied Baltic sequences belonged to the lineage formed by the classical rabies virus within the cosmopolitan lineage, with a bootstrap value of 86% (Fig 8). No Arctic or Arctic-like variants were identified in the panel of viruses studied from the Baltic States.
The majority of the Baltic rabies isolates grouped with the North-East European lineage (NEE), forming one strongly supported group (bootstrap value of 83). The NEE group consisted of 52 samples from the Baltic States and 21 published viral sequences (S1 Table) [38,49–51]. Both wild and domestic species fell in the NEE group.
The NEE Group showed less than 1% nucleotide divergence and 3% amino acid divergence among all Baltic isolates. Nucleotide sequence analysis showed 100% of nucleotide identity between a red fox sample (no. 11584) isolated in 2006 in Estonia and three samples from Lithuania (a red fox isolated in 2007, a raccoon dog and a cattle both isolated in 2009). The same perfect identity was obtained between the Estonian isolate no. 11584 and two samples from Latvia (a raccoon dog and a dog both isolated in 2008).
Five sequences from the Baltic States clustered with C group [48] formed with two published sequences, one from Russia and one from Ukraine (bootstrap support of 96%). Four species were included in this group: red fox (n = 2), raccoon dog (n = 1), cattle (n = 1) and dog (n = 1). Within C group, sequences showed more than 99.9% of nucleotide similarity. 100% nucleotide identity was shown between a red fox sample (MT3-TA11-00267) isolated in Estonia in 2011 and two samples from Lithuania; a dog (no. 864) in 2012 and a raccoon dog (no. 4740) isolated in 2010.
The PhyML tree also showed that a badger (Meles meles) (no. DR 784) and a marten (Martes martes) (no. 24771) isolated in Latvia in 2013 and in Lithuania in 2008 respectively, belonged to the group formed by the rabies virus vaccines (bootstrap of 100) (Fig 8). The vaccine-induced case isolated in Lithuania was found in the Alytus district in the south of the country, an area vaccinated since 2006 with Lysvulpen baits, whereas the vaccine-induced case isolated in Latvia was found in the Aloja district, in the north of the country, an area also vaccinated with Lysvulpen baits since 2011. Nucleotide analysis of the partial N gene sequenced of the two isolates showed 100% of nucleotide identity with the two referenced SAD-derived laboratory vaccine virus strains (EF206719 and EF206709) and there was 99.4% nucleotide similarity with the SAD Bern vaccine strain (EF206708). The case found in Latvia was located in an area vaccinated with the Lysvulpen baits 25 km away from Estonia where the Rabigen baits were used. As the partial N gene sequence analysis did not discriminate among the two SAD-derived laboratory vaccine virus strains, the amplification of partial G gene sequence on the DR784 isolate was undertaken to identify the vaccine strain. The comparison between DR784 isolate and three vaccine strains, SAD B19 (EF206709), SAG2 (EF206719) and SAD Bern (EF206708), showed 100% nucleotide identity with SAD B19 and 99.8% of similarity with the SAG2 and SAD Bern vaccines. The isolate DR784 was characterized by the presence of arginine in codon 333 (G gene). The sequence was clearly distinct from SAG2 (EF206719), characterized by two mutations in codon 333 yielding glutamine (Gln) at this position instead of arginine (Arg).
Surveillance data indicated a drastic reduction (90%) in the number of detected cases following 1 to 4 years of ORV. These results corroborate those from other European countries where 90% reduction of rabies detected cases were observed within 10 years, and in many cases less than 5 years following first ORV [52–54]. Depending on the country, the time to complete elimination (i.e. remaining 10%) is more or less longer to achieve. While eradication requires an additional 10 or more campaigns until no more cases are detected in Freuling at al. [10] we found that 2 to 8 campaigns were necessary. Variation in the reduction of the number of cases detected after each ORV depends on multiple factors such as the geographical location of the infected country, the initial epidemiological situation, the tools and strategy used in the control programmes and indubitably the implementation of an appropriate surveillance scheme.
As soon as ORV was implemented on the whole territories, the proportion of positive cases started to decrease in the three countries. As suggested previously in Brochier et al. [55], and Aubert [56] for fox rabies and Townsend et al. [57] for dog rabies, an inadequate vaccination area can compromise success and considerably extend the time to elimination. For Lithuania, the animals collected in vaccinated areas for the monitoring of ORV were also diagnosed for rabies. Because this sampling focuses on the animal population targeted by oral vaccines and not suspected for rabies (in contrary to classic rabies surveillance plans where only suspect animals are collected), Lithuanian surveillance data probably overestimates the number of negative samples compare to other countries. The comparison of the percentage of positive cases between countries became consequently hazardous. For this reason, combining data issued from surveillance sampling and monitoring sampling should, insofar as possible, be avoided [2,22]. Appropriate surveillance schemes focus on indicator animals collected at anytime, anywhere throughout the country and no sample size can be defined for proving the absence or the presence of rabies in wildlife regardless of the reservoir species. In contrast, the monitoring schemes are based on sampling foxes and raccoon dogs shot by hunters in vaccinated areas after ORV campaigns [30].
The oral vaccines used at the present time in Europe for raccoon dogs were developed to control rabies specifically in foxes. An experimental study evaluated the safety and efficacy of SAG 2 baits on caged raccoon dogs [58]. Either direct instillation or bait ingestion using a virus dose containing at least 10 times the field vaccine dose proved vaccine safety during the 60 days of observation of animals. More than 6 months after oral vaccination with the field dose, all animals were challenged with a street rabies virus. All vaccinated animals developed high rabies neutralizing antibody titers and survived a virulent challenge, demonstrating the effectiveness of the vaccine bait according to the European Pharmacopeia monograph. These results suggest that SAG2 vaccine baits are suitable for this species. Another study conducted on the SADP5/88 vaccine (derived from SAD Bern and no longer in use) in which two different doses of the vaccine were administrated showed satisfactory protection of challenged animals [59]. Paradoxically, to our knowledge, there are very few experimental studies using vaccines used in Baltic countries on raccoon dogs to assess their efficacy and safety prior to their release in the field.
For the first time, bait uptake results suggest a significant difference in the frequency of uptake of red foxes and raccoon dogs, with a lower proportion of tetracycline-positive raccoon dogs compared with red foxes. This result can be attributed to the difference in behavior of the two species and particularly to the hibernation of raccoon dogs in the Baltics during the cold season (November–March) [60], which may influence the epidemiology of the disease and access to vaccines distributed during this period. The impact of hibernation was suggested in a model of rabies transmission in both raccoon dogs and red foxes [61]. As suggested by our results, strategies to control rabies in countries where this species is an important transmitter should better focus on the raccoon dog biology. As example, ORV could also target raccoon dogs after they emerge from hibernation.
All countries implemented ORV according to the EU recommendations[19]. Bait uptake levels in Baltic countries rapidly reached 80% of the target population. Our study showed a constantly increasing bait uptake throughout the study period, suggesting cumulative exposure to distributed baits [19]. Data analysis in Estonia and Lithuania confirmed previous studies, showing a significantly lower bait uptake in juveniles than in adults [28,62,63]. As a matter of fact, difficulties in reaching juveniles during ORV campaigns were observed. This was observed especially in early spring [19] because cubs are in dens or cannot be vaccinated because too young to eat the bait. Latvia has used two types of vaccines, Lysvulpen and Fuchsoral vaccines. Analysis of factors that potentially affect bait uptake showed a significant influence of the type of bait used, with higher bait uptake when the vaccine Lysvuplen was used. The type of bait influence was independent from the year as further analysis, omitting the first years of vaccination with Fuchsoral baits, still considered the bait type as a significant variable explaining the TTC variations. Given that, according to the manufacturer’s specifications, both vaccines contain 150 mg of tetracycline in the bait matrix, the reason for this difference is unknown. More investigations on bait matrix composition and palatability are needed.
Neutralizing antibodies are the most reliable parameter for assessing the efficacy of vaccination because they are closely correlated with protection against rabies infection [64]. The assessment of the rabies antibody level is theoretically the best means for evaluating vaccination coverage because individual variation in immune reactions is taken into account. ELISAs allow large-scale screening because they are rapid, easy to perform, do not require live rabies virus or cell culture, and can be performed in any laboratory. These tests have been demonstrated as particularly suitable for assessing the effectiveness of oral vaccination in field samples [31,65,66]. The evolution of herd immunity level did not show any specific pattern, showing an unsteady evolution in all three countries. The surprising absence of any immunological trend may reflect the lack of sensitivity or reliability of some commercial ELISA kits, as has already been demonstrated recently [34,67,68]. Although the overall average bait uptake in this study was 80%, seroconversion level was approximately 50%. The same large discrepancies observed between uptake and seroconverion were attributed the lack of sensitivity of a commercial kit on field samples compared to blood samples taken from experimentally infected foxes and raccoon dogs, probably due to the reduced quality of the sera (haemolysis, bacterial contamination due to field condition) [28,67,69].
Latvia was the only country that used two different kinds of ELISA kits (Bio-Rad and BioPro) to evaluate vaccination coverage in red foxes and raccoon dogs. Further analysis demonstrated that significantly different levels of seroconversion were found for the two different kits. BioPro ELISA results showed lower seroconversion level than those of the Bio-Rad ELISA kit. These discrepancies are inconsistent with previous studies in which the seroconverion were found lower using Bio-Rad compared to BioPro kits due to the lower sensitivity of the first test [33,70]. Our results may be explained by the fact that a different cut-off value from the 0.5 conventional one’s was used for the Bio-Rad kit (0.125 instead of the 0.5 used in Wasniewski et al.). These results must be also nuanced by the fact that Latvia encountered specific events in the same period when using the Bio-Pro test in 2012, a year during which an epidemic sarcoptic mange occurred. Immunological reactivity due to sarcoptic mange could potentially have interfered with the rabies vaccination, leading to a lower response and a decrease in the level of rabies antibodies.
A sharp decrease in the number of marked animals was observed in Estonia during the last four campaigns as soon as the ORV area was reduced to a buffer zone of 9 325 km2 (20 km along the Southern border and 30–50 km in eastern part of the country). This drop could be explained, inter alia, by an “edge effect” due to the small size of the vaccinated areas. The areas being small, the perimeter-to-surface ratio is higher and the probability of sampling an unvaccinated animal in bordering areas is higher than for a large ORV areas. Moreover, the proportion of raccoon dogs in the monitoring sample has increased every year, ultimately constituting more than ¾ of all animals tested. This example highlights the importance of considering the structure of the monitoring sample in the determinism of the overall and final bait uptake level. Thus, comparison of monitoring data between countries and their interpretation should be assessed by taking into account the species (raccoon dogs vs red foxes) and the age class (juvenile vs adult) of the sampled population.
The molecular epidemiology of RABV in the Baltic countries showed the presence of three types of RABV variants in the Baltic States, the North-East European group (NEE) (158/165 isolates), the Russian group (C) (5/165 isolates) and two vaccine-induced rabies cases. These results confirm that the terrain for rabies hosts infected with Baltic variants is broad [71], ranging from Eastern to the Central Europe. More precisely, the NEE group has been reported in Eastern part of Russia and from Finland to Romania [49] including the Baltic States [28,72,73], Slovakia, Poland and Ukraine [38,48,74], while C group has been reported from the European part of Russia [48] to different parts of Ukraine [74]. Although the C group is the most widely reported RABV variant in Russia [75] including regions of western Siberia, Kazakhstan and Tuva, four other variants have been previously described in Russia [48].
This study is the first to report the presence of C variant in North-East Europe with three cases in Lithuania reported between 2009 and 2013, one case in Estonia in 2011 and one case in Latvia in 2012. The occurrence of the C variant in Baltic States could be the result of a westward spread of rabies-infected hosts from Russia or from Belarus to the Baltic States. Animal-to-animal transmission of rabies virus or human-mediated transports of latently infected animals could explain the movement of rabies infected hosts across borders. There are numerous studies illustrating rabies virus transmission by human-mediated animal movements [76], wildlife-mediated movement of rabies [50] or movements of infected animals a cross frozen seas [77].
In Russia, six wild canid species (red fox, raccoon dog, artic fox, steppe fox, jackal and wolf) are vector of the disease. In Eastern Europe and in north-eastern Europe, most wildlife cases are reported in red foxes and raccoon dogs. The NEE variant is particularly associated with raccoon dogs in north-western Russia and north-eastern Europe, while C group were previously associated with the red fox and the steppe fox in Russia [48]. In this study, no phylogenetic distinction was reported between the red fox and raccoon dog isolates, whatever the variant analysed (C and NEE groups) and whatever the phylogenetic method used. Perfect identity observed between one isolate (red fox) in Estonia in 2006 and five strains (two raccoon dogs, one fox, one cattle and a dog) isolated in Latvia and in Lithuania between 2007 and 2009 suggests that the variant circulating in fox and raccoon dog populations have the same origin. Dogs may have served as an early reservoir for interspecies rabies virus transmission generating viral lineages that then spread to other species [78].
Due to the risk of residual pathogenicity of oral rabies vaccines related to the viral strain’s attenuation level, all rabies virus samples isolated in areas where attenuated rabies virus vaccines are used should be typed in order to distinguish between vaccine and field virus strains [2,19,22]. For the first time, we demonstrated that two field Baltic isolates (a marten from Lithuania in 2008 and a badger from Latvia in 2013), clustered with the group forming the rabies vaccines, SAG2, SAD B19 and SAD Bern. Clearly, the two vaccine-induced rabies cases were closely related to SAD B19 strains, although both cases were found in an area vaccinated with SAD Bern Lysvulpen baits. Previous study results indicated that the SAD Bern Lysvulpen vaccine shows higher similarity to the strains belonging to the SAD B19 vaccine [79]. Such findings led to a change in the viral strain description for the national marketing authorization dossier of this vaccine, http://www.uskvbl.cz/en/authorisation-a-approval/marketing-authorisation-of-vmps/list-of-vmps/authorised-by-national-and-mrdc-procedures.
This is also the first reporting of a vaccine-associated virus detected in badgers and in martens. To date, few vaccine-induced rabies cases have been documented in target species. Muller et al. [80] reported six vaccine-induced rabies cases in foxes caused by SAD B19 and SADP5/88 in vaccinated areas in Germany and Austria, respectively. In Slovenia, a young fox was also shown closely related to SAD B19 in 2012 [81]. The most likely explanation for these vaccine associated cases isolated in non target species is either residual pathogenicity of the virus vaccine despite vaccine attenuation or reversion to virulence. RNA viruses are known to have high mutation rates due to the lack of proofreading by RNA polymerases and could have occasionally reversed to more virulent viruses. The second hypotheses would be a transmission from a red fox or raccoon dog initially infected by a vaccine strain. Potential transmission of vaccine strain has indeed been recently questioned when finding vaccine strain in salivary gland of a naturally infected fox [81].
|
10.1371/journal.pntd.0002791 | The Challenge of Producing Skin Test Antigens with Minimal Resources Suitable for Human Application against a Neglected Tropical Disease; Leprosy | True incidence of leprosy and its impact on transmission will not be understood until a tool is available to measure pre-symptomatic infection. Diagnosis of leprosy disease is currently based on clinical symptoms, which on average take 3–10 years to manifest. The fact that incidence, as defined by new case detection, equates with prevalence, i.e., registered cases, suggests that the cycle of transmission has not been fully intercepted by implementation of multiple drug therapy. This is supported by a high incidence of childhood leprosy. Epidemiological screening for pre-symptomatic leprosy in large endemic populations is required to facilitate targeted chemoprophylactic interventions. Such a test must be sensitive, specific, simple to administer, cost-effective, and easy to interpret. The intradermal skin test method that measures cell-mediated immunity was explored as the best option. Prior knowledge on skin testing of healthy subjects and leprosy patients with whole or partially fractionated Mycobacterium leprae bacilli, such as Lepromin or the Rees' or Convit' antigens, has established an acceptable safety and potency profile of these antigens. These data, along with immunoreactivity data, laid the foundation for two new leprosy skin test antigens, MLSA-LAM (M. leprae soluble antigen devoid of mycobacterial lipoglycans, primarily lipoarabinomannan) and MLCwA (M. leprae cell wall antigens). In the absence of commercial interest, the challenge was to develop these antigens under current good manufacturing practices in an acceptable local pilot facility and submit an Investigational New Drug to the Food and Drug Administration to allow a first-in-human phase I clinical trial.
| Despite reaching the global elimination target for leprosy, the need for a diagnostic tool to detect pre-symptomatic disease remains. Transmission has not been completely intercepted despite over 30 years of extensive curative treatment. With limited resources, two new leprosy skin test antigens, MLSA-LAM and MLCwA, suitable for human application were developed and manufactured in a local pilot plant. Requirements for manufacturing and clinical testing were met and an Investigational New Drug was established with the Food and Drug Administration to test both antigens in a phase I clinical trial for safety in a non-endemic region for leprosy and a phase II clinical trial for safety and efficacy in an endemic region for leprosy.
| Detection of pre-symptomatic leprosy continues to be identified by the World Health Organization (WHO) as a priority [1]. With the introduction of multiple drug therapy (MDT) by the WHO in 1982 to subvert extensive resistance of Mycobacterium leprae resulting from years of dapsone monotherapy, the prevalence of leprosy began a dramatic decline [2]. Over the past 30 years, prevalence has dropped by about 98% from an estimated historical high of 11.5 million cases in 1983 [3] to the current figure of 192,246 registered cases [1]. Contrary to this remarkable achievement, leprosy incidence as defined by new case detection (NCD) remained relatively constant or increased slightly from 1985 at 555,188 new cases in the top 33 endemic countries [4] to 571,792 in 1990 [5] and 620,672 in 2002 [6]. A significant decline of 51.4% in new cases was then observed between 2002 and 2005 to 299,036, followed by another decline of 23.6% to the current figure of 228,474 [1]. A total decline of 58.8% in detection of new cases from 1985 to 2010 has been observed. Although many investigators have questioned the value of these numbers based on confounding operational factors [7], one fact remains; NCD has generally exceeded prevalence. Of added concern is the increased NCD of childhood leprosy signifying active and recent transmission of disease [8], [9]. These findings provide evidence that transmission of M. leprae from infected to susceptible individuals remains a problem.
Little is known of the extent of pre-symptomatic [10] leprosy in the endemic regions of the world, or reservoirs of infection, or bacterial or immunological basis of the distinctive pathogenesis of leprosy, notably nerve damage [11], [12]; however, we do know that early detection and treatment does reduce transmission [13] and disease sequelae [14]–[16]. Clinical leprosy is a bacteriological and pathological polar disease ranging from the tuberculoid (TT) to the lepromatous (LL) forms, with intermediate stages [17], [18]. This spectrum of disease is determined by the immunological status of the host [19]. The lepromatous forms are marked by high antibody titers, but T cell hyporesponsiveness (anergy) [20]; whereas the tuberculoid form show little evidence of M. leprae specific antibodies, but a vigorous Th1 response. Likewise, some household contacts (HC) of leprosy patients also demonstrate a specific T-cell response. These changes in the immune response along the continuum of disease suggest that a cell mediated immunity (CMI) test may be adequate to detect early leprosy infection.
Our approach toward development of an early diagnostic tool for leprosy has been focused on the delayed hypersensitivity (DTH) immune response, because it is considered to be sensitive, simple, cost-effective, and inexpensive when applied as Tuberculin Purified Protein Derivative (PPD), skin test for tuberculosis [21]. A DTH type IV reaction is initiated when antigen is injected into subcutaneous tissue and processed by antigen presenting cells. A Th1 effector cell recognizes the antigen and releases cytokines IL-2, IFN-γ, and TNF, which act on vascular endothelium causing erythema and recruitment of T-cells, phagocytes, fluid, and protein. This cascade of events causes a measurable induration response within 48–72 hours in humans. A lack of DTH response to recall antigen is evidence of anergy [22].
Early leprosy skin test studies with whole bacilli preparations, such as Lepromin-H (Mitsuda) [23] and Lepromin A [24], [25], had proven utility in classification of disease with the 21 day Mitsuda granulomatous reaction. However, the Lepromin antigen tends to prime the immune response and moreover is not specific for leprosy. Lepromin Dharmendra (Dharmendra) [26], Convit's Soluble Protein Antigen (SPA) or Leprosin, and Rees's M. leprae soluble antigen (MLSA) [27] evoked a 48 hour DTH reaction (the Fernandez reaction) [28]. TT leprosy patients had a characteristic DTH response to SPA and MLSA; whereas LL leprosy patients were anergic to these antigens, but not to other mycobacterial antigens such as PPD [29]. The DTH responses of borderline patients typically fell within the spectrum of their disease classification [30], [31]. Promising features of the MLSA and SPA included: neither had sensitizing potential [32]; both were potent immunologically; and, both were found to be safe in human vaccine trials in Venezuela, Malawi, and India [33], [34]. Shortcomings included inconsistent readings due to soft rather than hard DTH reaction in some individuals; variations in potency between batches due to quality control issues; and, lack of adequate sensitivity and specificity.
Two refined leprosy skin test antigens were identified [35]. The first antigen was a modified Rees antigen: MLSA-LAM (MLSA devoid of lipoglycans, primarily the immunosuppressive and cross-reactive lipoarabinomannan (LAM), and also lipomannan (LM), and phosphatidylinositol mannoside (PIM) and other lipids [36]–[38]). The second antigen was MLCwA (M. leprae cell wall antigen), consisting of the powerful immunogens of the cell wall devoid of lipoglycans [39],[40].
Active ingredients of these two intradermal skin test antigens are proteins of M. leprae. MLSA-LAM contains soluble protein antigens; over 100 individual proteins were initially recognized on two-dimensional gels, and about 30 of these had been sequenced and the immunological responses studied in part [41], [42]. Foremost among these antigens are the 70 kDa (DnaK), 65 kDa (GroEL), 45 kDa, 38 kDa, 35 kDa major membrane protein-I (MMP-I), 23 kDa superoxide dismutase (SOD), 18 kDa small heat shock protein (SmHSP), 18 kDa bacterioferritin (Bfr), 10 kDa (GroES), and the ribosomal proteins S7/S12 [43]–[48]. More recently, the full spectrum of proteins in soluble and insoluble subcellular fractions of M. leprae have been demonstrated and many more identified through the modern-day “proteomics” approach [49]–[51]. MLCwA contains many of the same proteins as MLSA-LAM, particularly the 70 kDa and 65 kDa and degradation products of these, the export/secretory proteins (notably the 30/31 kDa, multigene antigen 85 complex), and also some larger, uncharacterized proteins [50]. Details of the full spectrum of MLCwA constituent proteins have since been published [51].
Both leprosy antigens were chosen as skin test candidates based on adequate yield and biological justification with a robust DTH response in M. leprae sensitized compared to M. tuberculosis infected guinea pigs [52] and strong induction of lymphocyte proliferation and secretion of IFN-γ from TT leprosy patient immune cells [53], [54]. These early studies led to the development and manufacturing of these antigens [35]. The neglected tropical disease of leprosy is a disease of the poor, living in marginalized countries [55]; hence, commercial interest in the development of new products was lacking. Despite limited experience and resources, product development was implemented in this academic setting [56], The researchers overcame the challenges of developing and manufacturing skin test antigens suitable for human application [57].
Protocols for animal use were reviewed and approved by the Animal Care and Use Committee (ACUC) at Florida Institute of Technology (FIT) and Colorado State University (CSU). The CSU approved ACUC protocol number was 02-167A-02. The FIT Armadillo Facility was in-compliance with United States Department of Agriculture-American Public Health Association (USDA-APHA), United States Public Health Service-Office for Protection from Research Risks (USPHS-OPRR), and International ACUC (IACUC) standards. The CSU Laboratory Animal Facility followed IACUC regulations and guidelines. The Phase I trial (registration number: NCT01920750) and phase II trial (registration number: NCT00128193) were registered with ClinicalTrials.gov. The phase I trial was not registered prior to implementation, because the trial was completed (February, 1999), before ClinicalTrials.gov registry was made available to the public (February, 2000). Retrospective registration of the phase I trial was requested for publication. The clinical Phase I Protocol, Protocol S1, and Phase II Protocol, Protocol S2, are attached as Supportive Information; although details of the clinical study will follow in subsequent articles.
M. leprae cannot grow axenically, but can be propagated in the nine-banded armadillo, Dasypus novemcinctus [58]. At the Florida Institute of Technology (FIT), Melbourne, Florida, Eleanor. E. Storrs and subsequently Arvind Dhople et al. under National Institutes of Health (NIH), National Institutes of Allergy and Infectious Diseases (NIAID) with regard to support and authorization, captured armadillos from state or nationally managed land areas in Central Florida for propagation of M. leprae. Animals were treated for parasites, quarantined, and tested prior to release by: 1) acid fast staining of ear snips, nasal swabs, and blood for evidence of acid-fast bacilli; 2) culturing of blood samples for sterility in Trypticase Soy Broth and thioglycollate broth; 3) hematology; 4) serodiagnosis for IgM antibodies to phenolic glycolipid-I; and, 5) Lepromin test to determine susceptibility to M. leprae [58], [59].
The source of M. leprae was a untreated lepromatous leprosy individual from Guyana with large numbers of highly bacilliferous subcutaneous nodules and lepromas. Genetic evidence has since indicated that M. leprae isolates are antigenically homogeneous [60], [61]. Infected armadillos were sacrificed and the livers and spleens were homogenized and fractionated to separate M. leprae bacilli to serve as the Master Seed Stock in 2 ml volumes (3×108 bacilli/ml) frozen at −70°C. Subsequently infected armadillos with disseminated leprosy were sacrificed and the tissues (liver and spleen), aseptically removed. The infected armadillo tissues were shipped to the Pilot Plant Skin Test Antigen Facility at CSU.
A total of 242 g of M. leprae infected tissue (spleen, 19 g; liver, 223 g) from three infected armadillos [animal nos. A563 (19 g spleen, one preparation), A572 (109 g liver, divided into three preparations), and A581 (114 g liver, divided into three preparations)] were fractionated using a modified 3/77 Draper protocol [62] (Figure 1), except for omission of the step involving protease digestion with chymotrypsin and trypsin and alterations in buffer composition. Protease digestion of homogenate was removed since no difference was seen between treated and untreated tissue preparations in terms of purity, protein content, and immunological potency of the recovered M. leprae.
Tissue sections ranging from 19 g to 36.5 g were homogenized with 10 mM disodium ethylenediaminetetraacetic acid (EDTA, Sigma, St. Louis, Mo.), pH 8.0 at 3 ml/g of tissue. Homogenates were tested for sterility on brain heart infusion agar, blood agar, and Lowenstein-Jensen agar (BD, Franklin Lakes, NJ). Tissue fragments were pelleted and washed twice with 10 mM EDTA by centrifugation (Sorvall RC5, Thermo Fisher Scientific, Inc., Rockford, IL) at 15,000× g for 10 min at 4°C in 50 ml Teflon Oakridge tubes, followed by extraction with 0.1 M sodium hydroxide (Mallinckrodt Baker Inc., Phillipsburg, NJ) in 10 mM EDTA while stirring at room temperature for 2 h to remove pigment and to separate M. leprae from tissue. The suspension was pelleted and washed twice with 0.1 mM sodium phosphate/0.1% Tween 80 (Mallinckrodt/Fisher) followed by digestion with 20 mg collagenase (Sigma, St. Louis, Mo.) and 0.23 mM calcium chloride (Sigma) in 200 ml of the sodium phosphate Tween 80 buffer while stirring overnight at 37°C. The digest was again pelleted and washed prior to two-phase extraction with 6% polyethylene glycol 6,000 and 8% Dextran T-500 (Sigma) in 0.1 M sodium phosphate/150 mM sodium chloride at 10 ml/g of tissue in a separatory funnel. The upper phase containing bacteria was removed and an equal volume of 0.2% Tween 80 added prior to centrifugation at 27,000× g for 30 min at 4°C. Purified M. leprae was washed twice at 15,000× g with buffered water and the concentration of bacilli estimated with a 1∶100 and 1∶200 dilution by optical density at A540 using an empirically determined conversion factor of 0.362 based on dry weight, i.e., A540 of 1.0 = 0.362 mg M. leprae/ml multiplied by the dilution factor. Samples of the bacilli were tested for sterility by culturing on brain heart infusion agar, blood agar, and Lowenstein-Jensen agar. Purity was subjectively determined by acid fast staining using methlyene blue as a counterstain for residual tissue, with acceptance criteria of ≥90% [63], [64].
M. leprae (128.43 mg) from seven such preparations were pooled and washed twice with 25 ml phosphate buffered saline (PBS) by centrifugation at 27,000× g for 15 min at 4°C (Figure 2). Bacteria were suspended in 5 ml PBS and disrupted by sonication on cold packs with an ultrasonic processor (Sanyo Soniprep 150, MSE Ltd., Lower Sydenham, London) at 1.5 MHz, 50% duty, and 1 second pulse intervals over six 5 min cycles with 5 min cooling between each cycle. Pre and post-sonicated bacteria were stained using the TB Acid Fast Stain Kit (Thermo Fisher Scientific Inc.) for counting to verify greater than 80% breakage.
Disrupted bacteria were centrifuged at 27,000× g for 30 min. Supernatant consisting of cytosol and membrane was transferred to a fresh tube and centrifugation repeated. The pellet consisting of M. leprae cell wall was washed three times with 10 ml PBS. The cytosol/membrane containing supernatant was transferred to an Ultra Clear 5 ml (13×51 mm) tube and ultracentrifuged (Optima TLX 120, Beckman Coulter Inc., Brea, CA) at 100,000× g for 2 h at 4°C to pellet the membrane. To remove lipoglycans [65] cold 20% condensed Triton X-114 (Baxter, Deerfield, IL) was added to the supernatant (cytosol) to a final concentration of 4%, followed by rocking at 4°C overnight. The tube was placed in a beaker of water at 37°C for 10 min to condense the Triton X-114 followed by centrifugation for 15 min at 3,900× g at 22°C to separate detergent and aqueous layers. The top layer was transferred onto tandem 1 ml Extracti-gel D (Fisher) columns to remove residual detergent. Extraction and removal of residual detergent was then repeated.
Cell wall pellet was resuspended with 2 ml of 2% sodium dodecyl sulfate (SDS, Fisher)/PBS and stirred while heating at 56°C for 1 h followed by centrifugation for 15 min at 27,000× g at 22°C to remove the SDS solubilized M. leprae cell wall antigens; the residual M. leprae cell walls consisting of the mycolylarabinogalactan-peptidoglycan complex has been the subject of much research [66], [67]. The supernatant was transferred to a fresh tube and the extraction was repeated. The MLCwA preparation was passed over two 1 ml Extracti-gel D columns to remove residual SDS and finally subjected to two rounds of TX-114 extraction followed by removal of residual detergent as described above.
The protein concentration of each of the antigen preparations was assessed by the Bicinchoninic Acid assay (Fisher). Antigens were diluted with PBS containing 0.0005% Tween 80 to a final dosage of 10.0, 5.0, 2.5, 1.0, and 0.1 µg protein per 0.1 ml followed by 0.2 µm filtration to remove residual particulates. A total of 1 ml of each of the antigen doses was aliquoted into prewashed and sterilized 2 ml borosilicate vials with 13 mm silicon rubber stoppers and aluminum caps (Wheaton, Millville, NJ). Vials were labeled in accordance with Food and Drug Administration (FDA) labeling requirements, including the statement, “Caution: New Drug-Limited by Federal Law to Investigational Use” [68], autoclaved for 20 min at 121°C; cooled at room temperature, and placed at −70°C for storage as MLSA-LAM and MLCwA batch no. 23 and lot no. 051297. Vials used in the phase I clinical trial remained at CSU, while those used in the phase II clinical trial were sent to Fisher Bioservices Repository (Rockville, MD) for relabeling with randomly assigned codes and shipment to the phase II clinical site (Figure 3).
Product development began in 1992 with the immediate challenges of acquiring adequate expertise and funding, generally offered by an industrial partner. While maintaining a focus on the need for an early diagnostic test for leprosy, primary resources including regulatory, technical, and financial support were identified through government, professional, and industry contacts. Establishing a product development plan was also difficult, since the Product Development Roadmap [83] or FDA Translational Critical Path [84] had not yet been published, and experience with the complicated process was mostly found within pharmaceutical and biotechnology companies. To overcome these hurdles, regulatory and technical assistance were provided by NIH, NIAID, Division of Microbiology and Infectious Diseases (DMID) Regulatory Affairs Specialists, FDA representatives, professional organizations including the Parenteral Drug Association (PDA) and International Society for Pharmaceutical Engineering (ISPE), and quality system consultants. Finally, the mind-set in the research environment required a change from innovation to standardization to develop these two new antigens.
Options for manufacturing the two new leprosy skin test antigens under cGMP, suitable for human application, were limited. Costs for using a contract manufacturing organization (CMO) were prohibitive; it was difficult to find any with an open schedule, and few had biosafety level 2 (BSL-2)/cGMP clean rooms required for safe manufacturing of these antigens. In addition, service providers acknowledged that they were fearful of working with M. leprae. Consequently, a retired BSL-3 research laboratory was converted to a cGMP Pilot Facility (Figure 4) at CSU for the sole purpose of manufacturing these leprosy skin test antigens. To this end, the manufacturing and testing process for MLSA-LAM and MLCwA was developed to meet 21 CFR parts 210, 211 for current Good Manufacturing Practices (cGMP) [85], [86]. Details of Pilot Plant Facility renovation are available from the authors.
The Pilot Facility consisted of a suite of five rooms, 1) Gowning and Material Transfer Room, 2) Manufacturing Suite A, 3) Manufacturing Suite B, 4) Quarantine/Released Goods Room, and 5) Quality Control Laboratory. Both the manufacturing and quality control rooms were under positive pressure cascading from the innermost room to the entry foyer. Air was supplied by a dedicated heating ventilation air conditioning system with single pass air flow monitored with gauges in the entry room and an anemometer prior to entry of the manufacturing suite. High efficiency particulate air filters were positioned on both the supply and exhaust air streams to purify air entering and exiting the clean rooms. The manufacturing rooms were classified [87] as international standard organization (ISO) 7 clean rooms. The innermost manufacturing room was used for downstream processing (antigen purification, formulation, and vialing), while the outermost manufacturing room was used for upstream processing (tissue fractionation and bacteria sonication). The gowning and material handling room was classified as an ISO8 clean room for personnel aseptic Tyvek gowning, wipe down and transfer of materials and equipment into the manufacturing area, and entering and exiting of personnel. The innermost quality control room, an ISO8 clean room was used for testing raw materials, intermediate product, and final product, while the quarantine/released goods room was a clean, non-classified clean room used for quarantine and release of raw materials.
Commissioning of the cGMP Pilot Plant for manufacturing skin test antigens was performed. Rooms were decontaminated with para-formaldehyde. The Pilot Plant was cleaned and the environment was monitored on three consecutive days and three consecutive weeks following directive documents to assess the cleanliness of the facility. Monitoring viable airborne organisms was performed with the Rotary Centrifugal Air Sampler (Biotest Diagnostics, Brooklyn Park, MN) and settling plates, both using Trypticase Soy Agar strips/plates. Monitoring viable surface organisms was performed with Rodac plates containing Trypticase Soy Agar and neutralizer for cleaning agents. Isolates were identified to the genius and species level using API Test Kits (Biomerieux, Etolile, France; distributed by VWR). Total particle counts in each clean room were measured using a Particle Counter (Metone Instruments, Grants Pass, Oregon). Acceptance criteria were met with each test enabling release of the Pilot Plant for cGMP manufacturing.
A quality system [88] was created for processing and testing leprosy skin test antigens in the renovated pilot plant [89]. The documentation system addressed: facility and equipment, materials, production, product labeling, laboratory control, and quality [90]. Two batch records were written, one for fractionation of tissues and the other for bacteria. A total of 255 supporting standard operating procedures (SOPs) were written to cover the quality system and manufacture of antigens. Facility and equipment SOPs were written for operation, maintenance, and calibration of dedicated equipment. SOPs for directing and tracking the chain of custody for raw materials transferred through purchasing, receiving, quarantine, release, and storage were created. Process directives supporting environmental monitoring, gowning, transferring material, manufacturing, in-process testing, and release testing were written into SOP format with data forms to collect relevant information. Explicit details for product labeling were captured in the batch record. All levels of training, including equipment use, biosafety, good laboratory practice (GLP), cGMP, and good clinical practice (GCP) were directed through SOPs. Logs were created to track part numbers, documents, raw materials, sample submission, equipment and room usage. Documents were subjected to the mandated review and approval process prior to implementation [91].
The manufacture of antigens was a two step process beginning with receipt, tracking, and release of raw materials. The primary raw material was spleen and liver tissues laden with M. leprae propagated in armadillos at FIT. Upon aseptic harvest, tissues were tested for the presence of contaminating bacteria using microbiological medium and then sent to the Pilot Plant, where they were frozen at −70°C in a qualified freezer until use. Manufacturing reagents were United States Pharmacopeia grade or equivalent, if available; otherwise, the highest purity was specified. Each reagent was released for use based on a certificate of analysis provided by the vendor, per an approved in-house specification sheet. Materials were tracked using a receiving code and part number system.
Tissue fractionation under the respective batch record was performed to release and purify M. leprae from armadillo tissue. A total of seven tissue runs were performed to accumulate 100–150 mg bacteria. Tissue weights ranged from 19–36.5 g for manageability and to maximize yields. A total of 128.4 mg of M. leprae was purified from 242 g tissue, resulting in a yield of 0.05% (Table 1). Sterility testing was performed on each bacterial lot, and material was stored at −70°C until use. Bacterial fractionation under the respective batch record was performed using the pooled intermediate product. Totals of 4.6 mg of MLSA-LAM and 5.0 mg of MLCwA were obtained, representing a yield of 3.57% and 3.88% from intact bacteria, respectively.
Assays to assess MLSA-LAM and MLCwA critical quality attributes of identity, purity, sterility, potency, and safety were performed [92]. Ten vials of each antigen dose (2.5, 1.0, and 0.1 µg/0.1 ml) planned for clinical studies were tested on all assays with two exceptions. Identity testing by gel electrophoresis and immunoblotting was performed on samples taken prior to autoclaving, which degrades proteins resulting in smearing of bands on gels and immunoblots. A representative silver stained gel of both antigen preparations is shown in Figure 5. Immunoblotting results showed that neither antigen preparation had detectable armadillo tissue or LAM present, both contained MMP-I, and only MLSA-LAM contained GroES and SOD, while only MLCwA contained GroEL proteins. Purity testing for adventitious agents was performed on tissue homogenates and concentrated final product (10.0 µg and 5.0 µg/0.1 ml); both were free of detectable human viral pathogens. The presence of collagenase, Triton X-114, and SDS were tested and found to be less than the lower limit of detection. Extracti gel D ligand was not tested, because if released, it would be removed by filtration prior to vialing. Calcium chloride, polyethylene glycol, and Dextran T-500 were not tested, because following multiple washes, the calculated residual concentration in the purified bacteria suspension had decreased by 46-fold and was found to be harmless as demonstrated in animal safety studies.
Antigen preparations were found to be sterile under aerobic and anaerobic conditions and potent when assessed for a DTH response in guinea pigs sensitized with M. leprae or infected with M. tuberculosis. Stability, although not a critical quality attribute was assessed during product development using a research batch and prior to and during clinical testing, resulting in 4 years of satisfactory results. All test results were used to complete the regulatory package. The Lot Release Summary and stability results for both MLSA-LAM and MLCwA can be found in Table 2.
In 1994, a draft Investigational New Drug (IND) [93] was formulated and specific questions related to IND enabling studies, manufacturing, and phase I clinical trial design was sent to our NIH, NIAID, DMID program officer at the time (the late Dr. Darryl Gwinn) and Regulatory Affairs Specialist (Ms. Carol Manning) for submission to the FDA Center for Biologics and Evaluation Research (CBER) for preliminary review and comment.
A FDA Response Letter with a comprehensive list of queries was received. The first topic of focus was the armadillo infected tissue and included questions on the following subject matters: 1) the origin, isolation, and characterization of the M. leprae strain; 2) creation, storage, maintenance, and viability testing of the master seed stock; 3) armadillo quarantine, test for human pathogens, and general health status; 4) potential human infectivity of indigenous armadillo microorganisms; 5) armadillo inoculation procedures and biosafety procedures for staff; and 6) test for viral adventitious agents. The second topic of concern centered on the manufacturing and characterization process, including questions on: 1) procedural flow charts; 2) potential or known human toxicities and quantitative tests for reagents used in the manufacturing process; 3) qualitative compositional analyses for each skin test antigen; 4) presence of cross-reactive antigens; 5) level of host contamination, endotoxin, and sterility; 6) in-vitro and in-vivo potency assays conforming to intended clinical use in humans; 7) stability testing prior to clinical studies; and 8) preclinical testing of clinical lots for safety, activity, and skin test conversion in a dose ranging study. Further questions were raised regarding the clinical phase I study design: 1) clinical study details; 2) potential impact of anergy regarding leprosy and HIV patients; 3) consent form and Institutional Review Board for each study site; 4) Case Report Forms for data collection; 5) references supporting related antigens and clinical studies; and 6) distinguishing subjects that are infected or harboring live bacilli from those who are infected and cured. A reply to the FDA Response Letter was satisfactory and a Pre-IND Meeting followed to review details of the manufacturing and testing process.
Skin test antigens were manufactured in May, 1997. The IND chemistry, manufacturing, and control (CMC) section was then completed and our DMID Study Sponsor submitted the IND Application to CBER for review. In September, 1998, FDA allowed the clinical investigation of two new drugs, MLSA-LAM and MLCwA, to proceed each at 3 doses (2.5, 1.0, and 0.1 µg) initially in a phase I clinical trial with ten healthy subject residing in a non-endemic region for leprosy, and subsequently in a phase II clinical trial with healthy subjects, leprosy patients, leprosy patient contacts, and tuberculosis patients residing in an endemic region for leprosy.
A tool for the detection of pre-symptomatic leprosy is an urgent need [94], [95]. How to address the treatment of individuals with evidence of specific leprosy exposure is a matter of debate [10]; chemoprophylaxis is proving highly efficacious in the short term, as applied to household contacts [96]. Individuals positively identified as pre-symptomatic could be a tool in the identification and further management of the disease, particularly reduction of incidence, i.e. NCD. Serological and gene approaches had not proven satisfactory for the purpose of diagnosing leprosy [56]; although these and other test methods are continually being refined and evaluated, in particular: details of new M. leprae antibodies [97], [98], new approach in the application of M. leprae specific DNA polymerase chain reaction [99]–[101], and cell-mediated immune response assays primarily based on IFN-γ release [102], [103]. While tests for PGL-I IgM antibodies have found favor for certain applications, most are not suitable for epidemiological application [104]. However, the two new leprosy antigens described here, MLSA-LAM and MLCwA, showed promise in guinea pig DTH studies and IFN-γ release assays [53], [54], [105]. Skin testing is the only means for mass epidemiological screening. Antigens for this purpose were targeted for product development as new leprosy skin test antigens.
Notwithstanding significant challenges, the development and manufacturing of these two leprosy skin test antigens suitable for human application was successfully accomplished. Securing adequate funding, identifying a large team of experts, and establishing a product development plan were key achievements that benefited the entire development phase. Changing the focus and practices of the research staff from basic to applied research enabled production of the skin test antigens. The magnitude of effort necessary in meeting regulatory requirements, in particular, substantial documentation, compounded by limited staff, funding, and experience was demanding A special attribute of this undertaking was the oversight of NIH, NIAID, DMID sponsor who provided financial, technical, and regulatory assistance, and served as a conduit to the FDA for cGMP and IND related questions.
The positive impact of developing and manufacturing these two new leprosy skin test antigens in an academic setting was realized only after successful implementation. The effort produced knowledge, skill, and understanding of the product translational process at the academic institutional level. Students were a valuable asset and in return, gained a unique learning opportunity. Looking forward, this work provides a product development template for products of neglected tropical diseases. Academic institutions cannot carry the heavy load of full product development alone, but this prototype presents alternative opportunities to move viable product ideas from the bench to the clinic.
The outcome was two new leprosy skin test antigens, suitable for human application, produced in a setting inexperienced in the manufacture of products for human use. This was necessitated by our focus on one of the major neglected tropical diseases of our time, and one of little commercial value. A consequence of this effort was the establishment of a contract manufacturing organization, Biopharmaceutical Manufacturing in an Academic Research Center (BioMARC), at Colorado State University for developing and manufacturing biological products to test in early clinical studies.
|
10.1371/journal.ppat.0030069 | Structure of GrlR and the Implication of Its EDED Motif in Mediating the Regulation of Type III Secretion System in EHEC | Enterohemorrhagic Escherichia coli (EHEC) is a common cause of severe hemorrhagic colitis. EHEC's virulence is dependent upon a type III secretion system (TTSS) encoded by 41 genes. These genes are organized in several operons clustered in the locus of enterocyte effacement. Most of the locus of enterocyte effacement genes, including grlA and grlR, are positively regulated by Ler, and Ler expression is positively and negatively modulated by GrlA and GrlR, respectively. However, the molecular basis for the GrlA and GrlR activity is still elusive. We have determined the crystal structure of GrlR at 1.9 Å resolution. It consists of a typical β-barrel fold with eight β-strands containing an internal hydrophobic cavity and a plug-like loop on one side of the barrel. Strong hydrophobic interactions between the two β-barrels maintain the dimeric architecture of GrlR. Furthermore, a unique surface-exposed EDED (Glu-Asp-Glu-Asp) motif is identified to be critical for GrlA–GrlR interaction and for the repressive activity of GrlR. This study contributes a novel molecular insight into the mechanism of GrlR function.
| Attaching and effacing pathogens are a group of enteric pathogens that includes the closely related enterohemorrhagic Escherichia coli (EHEC) and enteropathogenic E. coli (EPEC). EPEC causes severe diarrhea in young children in developing countries, while EHEC is a causative agent of hemorrhagic colitis. A major infection mechanism employed by EHEC and EPEC is the type III secretion system (TTSS). TTSS is a syringe-like apparatus composed of approximately 20 proteins that serve to transfer virulence proteins from the bacteria directly into the host cytoplasm. The genes encoding for the TTSS components and related proteins are organized in several operons that are clustered in the locus of enterocyte effacement (LEE). GrlR and GrlA are LEE-encoded, newly identified, regulators that are common to all the attaching and effacing pathogens. This article reports the crystal structure of GrlR and explains how it can bind with GrlA to influence the activity of TTSS. Further, we have identified an EDED motif of GrlR crucial for the recognition of GrlA and activity. This study will help to understand the virulence determinants of E. coli, which is important for controlling the diseases caused by these organisms.
| The enterohemorrhagic Escherichia coli (EHEC) and enteropathogenic E. coli (EPEC) are closely related human enteric pathogens [1]. EPEC causes severe diarrhea in young children in developing countries, while EHEC is a causative agent of hemorrhagic colitis, which is more common in the industrialized world [2]. EPEC, EHEC, and the mouse pathogen Citrobacter rodentium (CR) belong to a group of pathogenic bacteria that are defined by their ability to form “attaching and effacing” (AE) histopathology on intestinal epithelia. This histopathology is characterized by localized destruction of apical microvilli, followed by intimate adhesion of bacteria to the cell plasma membrane [3]. A major virulence mechanism underlying AE-causing bacteria is the type III secretion system (TTSS), which is employed by the bacteria as a molecular syringe to inject (translocate) effectors into the host cell. These effector proteins subvert normal host cell functions to benefit the bacteria [4–6]. TTSS components and related proteins are encoded by 41 genes organized in five major operons, LEE1 through LEE5, and several additional transcriptional units, all clustered in the locus of enterocyte effacement (LEE) [7].
Under repressive conditions, the entire LEE is silenced by the histone-like nucleoid structuring protein (H-NS). Activation of most LEE promoters is dependent on Ler, an H-NS paralogue encoded by LEE1, which functions as anti–H-NS to alleviate the H-NS–mediated silencing of most of the LEE promoters [8–11]. Therefore, controlling the activity of the LEE1 promoter (PLEE1) is critical for initiating a cascade that mediates the expression of all of the LEE genes. GrlA and GrlR are two LEE-encoded regulators that are required to optimize PLEE1 activity [12]. These two proteins from EPEC and EHEC, respectively, exhibit about 98% identities. GrlA acts as a positive regulator for PLEE1; moreover, GrlA and Ler form a positive transcriptional regulatory loop acting synergistically to strongly activate ler expression [12]. It is suggested that in order to prevent the detrimental accumulation of Ler, the Ler–GrlA feedback loop is negatively modulated by two checkpoints: (1) When Ler reaches the threshold concentration, it represses ler transcription [13]. (2) GrlR, a negative regulator of ler expression [14,15], might act as anti-GrlA to establish an additional checkpoint that down-regulates the feedback loop, setting it back to the steady-state level. In agreement with this hypothesis, GrlR interacts with itself and also with GrlA to form a macromolecular assembly in the cytoplasm of AE pathogens [16]. It has been proposed that GrlR conveys a negative regulation through its interaction with GrlA and that this hetero-complex is functionally relevant [12,16]. The literature search and sequence analysis indicated a presence of a helix-turn-helix, a DNA recognition motif at the N-terminus of GrlA [14], and the C-terminal region may interact with GrlR.
Here, we report the crystal structure of GrlR from EHEC refined up to 1.9 Å resolution as well as structure-based functional studies on GrlR. GrlR has a typical β-barrel fold consisting of an internal hydrophobic cavity with a plug-like loop on one side of the barrel. Structure-based site-directed mutagenesis targeting the surface residues of GrlR showed that these residues are crucial for the ability of GrlR to bind GrlA and to carry out its regulatory function. In vitro and in vivo experiments further confirmed the vital role of these residues for the regulatory function of GrlR. Our finding represents a novel regulatory mechanism in the TTSS of pathogenic bacteria.
The structure of recombinant GrlR from EHEC was solved by the multi-wavelength anomalous dispersion method from synchrotron data. The model was refined to a final R-factor of 0.215 (Rfree = 0.269) at 1.9 Å resolution (Figure 1A) with good stereochemical parameters (Table 1). The GrlR model consists of residues from Met1 to Val111, with seven additional residues at the N-terminus (Gly-6, Leu-5, Val-4, Pro-3, Arg-2, Gly-1, and Ser0) resulting from the linker sequence of the (His)6 affinity tag. The C-terminal residues from Asn112 to Lys121 had no interpretable electron density and were not modeled. There are two molecules in the asymmetric unit (Figure 1B) and they are related by a 2-fold noncrystallographic symmetry approximately parallel to the a-axis. Interestingly, these two molecules are packed in a perpendicular fashion to each other, resulting in a maximum interaction (Figure 1B).
The GrlR molecule mainly consists of a single domain from residues Asp5 to Ile107 that forms a β-barrel. Residues Tyr59 to Asp70 form an extended loop that plugs one side of the cylindrical β-barrel structure. The β-barrel consists of eight anti-parallel β-strands running from one side of the molecule to the other. On one side of the β-barrel there are four long loops, including a plug-like structure, whereas on the other side, four tight β-turns are connecting the adjacent β-strands. Both ends of the β-barrel were closed off by the N-terminus (Met1 to Lys4) and plug-like loop residues Tyr59 to Asp70. The tip of the ten residue–long plug-like loop, which is highly hydrophobic, may close or open the cavity primarily by hydrophobic interactions. The β-barrel cavity is highly hydrophobic in nature with side chains from seven Tyr, six Ile, seven Leu, four Val, and two Phe residues lining the inner cavity surface (Figure 2A). The approximate dimensions of the β-barrel are 35.2 Å in height and 18.5 Å in diameter.
GrlR shares 93% to 100% identity between different AE-causing E. coli strains and EHEC (strain EDL933) and over 85% sequence identity (CLUSTAL W [17]) with that of CR. There is no significant overall sequence identity with any other protein in the National Center for Biotechnology Information (NCBI) database. A search for topologically similar proteins within the Protein Data Bank (PDB) database was performed with the program DALI [18]. The highest structural homology is observed between GrlR and the electron transport domain of quinohemoprotein amine dehydrogenase (PDB code 1jju; with 18% sequence identity; z-score 9.2 and 2.5 Å RMSD [root mean square deviation] for 90 Cα atoms). This is followed by a lipid-binding TTSS secretin pilotin protein, MxiM (PDB code 1y9t; with 16% sequence identity; z-score 4.90 and 2.7 Å RMSD for 77 Cα atoms), which has a cracked barrel structure [19].
During structure refinement, we noticed a small molecule composed of 12 atoms in the hydrophobic cavity of GrlR, which, subsequently, was identified to be the fragment of Triton-X100. It is worth mentioning here that the bacterial lysis buffer used for GrlR purification contained 1% (v/v) of Triton-X100. The detergent may have bound tightly to the hydrophobic cavity of GrlR during this stage, and had co-crystallized with GrlR. The bound detergent is situated at the center of the cavity and is parallel to its axis (Figure 1A); the interaction of the detergent with hydrophobic residues of the cavity may play a crucial role in increasing the solubility of GrlR (Figure 2A). The ligand molecule is well defined in the electron density map. Figure 2B shows the simulated annealing Fo-Fc omit map. The superimposed GrlR on lipid-bound MxiM [19] indicated that the probability of having a lipid molecule in the hydrophobic pore of GrlR is not ruled out. Based on the structural homology, the bound ligand, and the hydrophobic nature of the cavity of GrlR, we suggest that the cavity may recognize a specific small hydrophobic ligand and interact with side chains of the cavity residues. Exact roles of this cavity and the plug-like loop for the function of GrlR are not yet established.
GrlR was found to exist as a homodimer in solution, with an apparent molecular weight of 29 kDa, as determined by gel filtration and dynamic light scattering. The analytical ultra centrifugation experiments also revealed the dimeric nature of GrlR. These results were consistent with a dimeric arrangement observed in the crystal structure, with the dimer having approximate dimensions of 46.5 × 32.6 × 35.2 Å. The strong hydrophobic cluster at the dimeric interface is maintained by the side chains from residues Ile7, Ile23, Ile25, Val39, and Ile107 from both monomers of the dimer. In addition, six hydrogen bonding contacts (<3.2 Å) mainly from Gln41, Glu108, His55, Pro109, and Val 111 of both monomers, as well as numerous hydrophobic interactions, are maintaining the dimer architecture. Figure 3 shows the electrostatic surface representation of the dimeric GrlR (GRASP [20]). The observation of a dimeric GrlR for the wild-type as well as for mutants suggests a functionally important role for dimerization.
GrlR was shown to associate with itself and with GrlA to mediate the regulatory network [12,17]. GrlA is homologous to CaiF, a known DNA binding protein [21], and its sequence analysis identified a helix-turn-helix DNA recognition motif at the N-terminus [12]. The C-terminal region of GrlA is rich in basic residues (nine arginines and seven lysines), suggesting that it may have a role in the interaction with the acidic GrlR. A loop region in the crystal structure of GrlR, residues Glu46 to Asp49 (46EDED49), is highly exposed on the surface and is less well defined in electron density. It is worth noting here that GrlA and GrlR have extremely opposite charges: the calculated isoelectric points (pI) of GrlR and GrlA are 4.83 and 9.71, respectively. Taking together all of these facts, we now hypothesize that the negatively charged cluster (the EDED [Glu-Asp-Glu-Asp] motif) is involved in the GrlR–GrlA interaction and thus may play an important role for repressing the ability of EHEC to perform TTSS-mediated protein secretion.
To test the above hypothesis, EDED residues were mutated and interactions of different GrlR mutants with glutathione-S-transferase (GST)–GrlA were tested by pull-down assays followed by SDS-PAGE analyses. We found that GrlA was not stable on its own, but the GrlA fused with GST was sufficiently stable for pull-down experiments with both wild-type and mutants of GrlR to verify the binding between these two proteins. We have also confirmed that wild-type GrlR binds to GST-GrlA (Figure 4A), whereas no binding has been observed with GST alone. All of the four single mutants (E46A, D47A, E48A, and D49A) and the double mutant (E46A-D47A) of GrlR showed similar binding to GrlA in pull-down assays. However, the triple mutant (E46A-D47A-E48A) showed a significant drop in binding to GrlA. Whereas the EDED tetra mutant, with all of the four residues mutated to alanine, did not bind to GrlA, no protein band corresponding to GrlR was detected in the pull-down assay (Figure 4A). Bands corresponding to GST-fused GrlA and GrlR were analyzed with peptide mass finger printing and their identities were confirmed (unpublished data). To verify the integrity of secondary structures in the mutants, circular dichroism spectra were measured for wild-type GrlR as well as for all mutants. In all cases, the circular dichroism spectra showed the existence of similar β-sheet secondary structures, which is consistent with the crystal structure. These results indicated that the surface-exposed EDED motif is a key structural feature for the binding of GrlR to GrlA.
In order to elucidate the role of the wild-type and mutants of GrlR in the repression of TTSS, a protein secretion assay was carried out in EHEC. Bacteria containing different plasmids were grown in Dulbecco's Modified Eagle's Medium (DMEM), and total secreted extracellular proteins were recovered from the medium and compared by SDS-PAGE. We used EspB, which is a major secreted protein of EHEC, as a representative marker and compared the secretion of EspB in wild-type and mutant GrlR (Figure 5, upper panel). The wild-type GrlR repressed the secretion of EspB. Single (E46A, D47A, E48A, D49A) and double mutants (E46A-D47A) showed similar reduced secretion of this protein, whereas the triple mutant (E46A-D47A-E48A) significantly reduced GrlR repression and the secretion of EspB was increased. However, the tetra mutant (E46A-D47A-E48A-D49A) totally abolished the repression effect of GrlR and the secretion was restored to a level similar to that of the control. We have also verified the expression level of all of the GrlR mutants (Figure 5, lower panel). These results are consistent with our previous pull-down assays (Figure 4A), which showed that GrlR and GrlA binding was affected in the triple and tetra mutants. Our experiment demonstrated the importance of the EDED motif for carrying out the regulatory function of TTSS in EHEC.
A global regulator of TTSS, Ler, which is encoded by the first gene of LEE1, positively regulates several secreted proteins, including EspB [9,22]. It has been shown that GrlR overexpression suppresses production of Ler [12,14]. From our experiments, we have shown that overexpression of wild-type GrlR, but not mutated GrlR, affects the secretion of EspB.
The role of the EDED motif in repressing the activity of the PLEE1 was analyzed by a transcription kinetics assay. Since the signal generated by a single copy chromosomal gfp gene under the PLEE1 is too low to detect in EHEC due to intrinsic low expression levels of the LEE1 in EHEC, we performed our studies in the closely related EPEC, where the intrinsic level of LEE1 expression is higher. To this end, we have constructed an EPEC strain (GY2455) expressing GFP+ (green fluorescent protein) from the native LEE1 promoter (PLEE1). The fluorescence (amount of GFP) as well as OD600 (amount of bacteria) were determined, in real time, during growth. The presence of plasmid encoding GrlR did not affect gfp expression unless IPTG was added. Under our experimental conditions, GrlR expression resulted in attenuation of PLEE1 activity, but not complete repression. Similar results were seen with the plasmids encoding GrlR, which mutated at different residues of the EDED motif. Upon replacing the entire EDED motif with AAAA, GrlR was no longer capable of PLEE1 repression (Figure S2). These results support the hypothesis that the EDED motif is crucial for repression of PLEE1 by GrlR. Overall, results of this gfp assay, with the exception of small variations, are comparable to EHEC experiments. These observed variations may be due to the difference in strains as well as the experimental conditions.
The regulatory network that controls the expression of the virulence genes of AE pathogens is complex. Much of this complexity is merged at controlling the activity of the PLEE1 and ler expression. We, as well as other groups, demonstrated that Ler is a master regulator, turning on and off a large number of virulence genes, including espB [9,22,23]. The punctually temporal regulation of Ler and maintaining its accurate levels of activity are essential for the successful colonization of the host [12,16]. The GrlR−GrlA complex plays a key role in controlling Ler expression [12,14]. Iyoda et al. [24] reported recently that the GrlR–GrlA complex also controls the expression of FlhDC, the flagella master regulator. Thus, the GrlR–GrlA complex plays an important role in controlling the expression of two key master regulators, Ler and FlhDC. Structural study of GrlR identified the surface-exposed EDED motif and the importance of these residues was further investigated. Our in vivo and in vitro functional studies of wild-type and mutant GrlR showed that the EDED motif is crucial for the recognition of GrlA by GrlR and for the GrlR regulatory activity. To our knowledge, this is the first report of a key role of the EDED motif in bacterial regulation.
Based on the properties of these two proteins, the location of the EDED motif in the dimeric GrlR and the dimeric nature of most of the helix-turn-helix–containing DNA-binding proteins, we propose that the GrlA may also exist as a homodimer, and that dimers of GrlR and GrlA in combination are involved in the regulatory mechanism.
Our study provides a novel structural basis for an understanding of the regulatory mechanism of the GrlA–GrlR complex and thus provides new insight into the complex regulatory network that governs the virulence of AE pathogens.
The bacterial stains and the plasmids used in this study are listed in Table S1. The intact grlR gene was PCR amplified from EHEC EDL933 chromosomal DNA and cloned into a derivative of pET vector (pETM32) (Novagen, http://www.emdbiosciences.com/html/NVG/home.html) or pSA10 vector. The respective targeted residues were substituted with alanine. Plasmid pGEX-grlA was constructed by amplifying the grlA DNA fragments from EHEC EDL933 chromosomal DNA and cloning into pGEX-4T1 (Amersham Biosciences, http://www.gelifesciences.com). Construction of pGY2Ler was as follows. The bla gene in the suicide plasmid pGP704 [25] was replaced by tetAR, and the gfp+ gene [26] was cloned into the XbaI and SmaI sites generating pGY2. A fragment containing PLEE1-ler (the regulatory region of LEE1 and the first gene in LEE1-ler) was amplified by PCR, digested with BamHI and XbaI, and cloned into pGY2 digested with BglII and XbaI, generating pGY2Ler, in which gfp+ is transcriptionally fused to ler. pGY2Ler was introduced into E. coli SM10 λpir, which was further introduced into EPEC by mating. A trans-conjugant KanS TetR StrepR colony containing an integration of pGY2Ler into the EPEC chromosome was selected to form transcriptional fusion of the LEE1 regulatory region with ler and the gfp+ gene (Figure S1) and was termed strain GY2455.
Plasmid DNA was transformed into E. coli BL21 and the cells were grown in defined M9 medium [27] supplemented with 25 mg/l L-SeMet at 37 °C to 0.6 AU at OD600. One liter of culture was induced with 100 μM IPTG and continued to grow at 20 °C overnight. Cells were then harvested by centrifugation and resuspended in 40 ml of lysis buffer (50 mM Tris-HCL [pH 7.5], 0.4 M NaCl, 1 mM EDTA, 1% (w/v) Triton X-100, 5% (w/v) glycerol, 10 mM ßME, and one tablet of Complete protease inhibitors [Roche Diagnostics, http://www.roche.com]). The protein was purified in three steps using DEAE-Sepharose (Amersham Pharmacia, http://www.gelifesciences.com), NI-NTA (Qiagen, http://www.qiagen.com), and gel filtration (Superdex 75, Amersham Biosciences) columns, respectively. The His fusion tag was not cleaved. Drops containing 1 μl of protein solution (4 mg/ml) and 1 μl of reservoir solution were equilibrated by hanging drop vapor diffusion at 21 °C. The best crystals were grown from 25% ethylene glycol, 4% tert-butanol, and 4% trifluoroethanol (Hampton screens followed by additive screening), with the protein in 20 mM Tris-HCL (pH 7.5), 200 mM NaCl, and 5% (w/v) glycerol. Crystals measuring ~0.2 mm in length grown over the course of 3 d belonged to space group P212121 with a = 43.83 Å, b = 66.09 Å, c = 83.58 Å and contained two molecules in the asymmetric unit. The Matthews coefficient is 2.2 Å3/Da [28], giving a solvent content of 45%. The X-ray data collection and refinement statistics are given in Table 1.
Crystals were cryoprotected in the reservoir solution supplemented with 40% ethylene glycol and flash cooled at 100 K. The structure was determined using crystals of SeMet-labeled protein by multi-wavelength anomalous dispersion method [29]. X-ray diffraction data were collected at beamline X12C, Brookhaven National Laboratory (Upton, New York, United States), using a Quantum-4 CCD detector (ADSC, http://www.adsc-xray.com). Two data sets were collected at wavelengths corresponding to the peak and inflection point. All of the data sets were processed with HKL2000 [30]. All of the eight Se sites of an asymmetric unit were located by using the program BnP [31]. The phases were further improved by density modification using RESOLVE [32], which gave a final overall figure of merit of 0.70. Over 50% of the backbone atoms of the model were built by the RESOLVE iteration method. The remaining residues of the molecules were added after several cycles of manual model building using O [33] and followed by refinement using CNS [34]. Finally, 342 well-defined water molecules were added, and refinement was continued until the R-value converged to 0.215 (Rfree = 0.269) for reflections I>σ(I) to 1.9 Å resolution. The model had good stereochemistry, with all residues falling within the allowed regions of the Ramachandran plot (Table 1) analyzed by PROCHECK [35].
The plasmid pGEX-grlA was transformed into E. coli strain BL21 DE3 and overexpressed under IPTG induction. GrlA protein with GST tag was immobilized on GST sepharose beads (Amersham Biosciences) in Lysis buffer (20 mM TRIS [pH 7.5], 200 mM NaCl, 5% glycerol, 10 mM ßME ) and washed with wash buffer at various salt concentrations (200 mM, 500 mM, and 1 M NaCl) to remove non-specific bound protein from the beads. The beads with immobilized GrlA protein were checked for purity and quantified using SDS-PAGE and were subsequently used for performing the pull-down assay studies. Expression and purification of his-tagged GrlR fusion proteins were performed as described previously. Equal amounts of GrlR wild-type and mutant proteins were added to the GST sepharose beads with bound GrlA and incubated at 4 °C for 15 min (Figure 4B). The expression level of GrlR tetra mutant was low compared to that of the wild-type and other mutants; however, an approximately equal amount of tetra mutant was used in all of the experiments. The beads were washed twice with wash buffer and resolved in 12% SDS-PAGE along with the controls.
The oligomeric state of GrlR was investigated by monitoring its sedimentation properties in sedimentation velocity experiments using a Beckman Coulter Optima XL-A (http://www.beckmancoulter.com) equipped with absorbance optics. Sedimentation coefficients and molecular masses were determined by fitting using both the C(s) method [36] and Enhanced van Holde–Weischet Analysis [37] as implemented in UltraScan 7.3 [38,39].
Molecular weight determination was carried out with the aid of a Voyager STR MALDI-TOF mass spectrometer (Applied Biosystems, http://www.appliedbiosystems.com). For MS/MS analysis, sample digestion, desalting, and concentration steps were carried out by using the Montage In-Gel digestion Kits (Millipore, http://www.millipore.com). Protein spots were analyzed using an Applied Biosystems 4700 Proteomics Analyzer MALDI-TOF/TOF (Applied Biosystems). Data processing and interpretation was carried out using the GPS Explorer software (Applied Biosystems) and database searching was performed using the MASCOT program (Matrix Science, http://www.matrixscience.com). The National Center for Biotechnology Information database (http://www.ncbi.nlm.nih.gov) was used for the combined MS and MS/MS search.
Far UV spectra (260–190 nm) of GrlR wild-type and mutants were measured using Jasco J810 spectropolarimeter (Jasco, http://www.jascoinc.com) in phosphate buffer (pH 7.5) at room temperature using 0.1 cm path length, stoppered cuvettes. A total of three scans were recorded and averaged for each spectrum, and the baseline was subtracted.
To prepare the secreted protein of EHEC, overnight cultures of EHEC strains in LB were diluted at 1:50 into DMEM supplemented with 100 mM ampicillin and 0.1 mM IPTG, and incubated for 9 h at 37 °C in a 5% (v/v) CO2 atmosphere. Bacterial cells were removed from the culture by centrifugation (5,500g, 10 min, 4 °C) and the supernatants were collected and passed through a 0.22-μm filter and precipitated by 10% TCA as described previously [22]. The extracellular proteins were resolved in 12% SDS-PAGE and visualized by staining with commassie blue. Western blot analysis was carried out as described previously [40], and EHEC cells harboring various pSAgrlR plasmids were harvested and resolved by SDS-PAGE. Proteins were transferred to PVDF membrane and detected by anti-6His (Qiagen) antibody.
Plasmids (pSA10) expressing GrlR or various GrlR mutants from the Ptac promoter were introduced into EPEC GY2455. The generated strains were grown overnight under conditions that repress the activity of the PLEE1 (30 °C in LB) [26]. For activation of the PLEE1, the cultures were washed and 50 times diluted with Casamino-DMEM [26] supplemented or not supplemented with 0.25 mM IPTG. Immediately upon dilution, cultures, in 96-well plates, were placed in a microplate reader (SPECTRAFluor Plus; TECAN, http://www.tecan.com), pre-set at 37 °C and grown within the plate reader. The fluorescence intensity (filter set at 485-nm excitation wavelength and 535-nm emission wavelength) and optical densities at 600 nm (OD600) were automatically read during growth every 5 min, and data were collected by Magellan software (TECAN).
Coordinates of GrlR have been deposited in the Protein Data Bank (http://www.pdb.org [41]) under accession code 2OVS. |
10.1371/journal.pgen.1004501 | Determinative Developmental Cell Lineages Are Robust to Cell Deaths | All forms of life are confronted with environmental and genetic perturbations, making phenotypic robustness an important characteristic of life. Although development has long been viewed as a key component of phenotypic robustness, the underlying mechanism is unclear. Here we report that the determinative developmental cell lineages of two protostomes and one deuterostome are structured such that the resulting cellular compositions of the organisms are only modestly affected by cell deaths. Several features of the cell lineages, including their shallowness, topology, early ontogenic appearances of rare cells, and non-clonality of most cell types, underlie the robustness. Simple simulations of cell lineage evolution demonstrate the possibility that the observed robustness arose as an adaptation in the face of random cell deaths in development. These results reveal general organizing principles of determinative developmental cell lineages and a conceptually new mechanism of phenotypic robustness, both of which have important implications for development and evolution.
| It is widely believed that development plays an important role in the phenotypic robustness of organisms to environmental and genetic perturbations. But, the developmental process and cell fate are largely predetermined and fixed in some species, including for example mollusks, annelids, tunicates, and nematodes. How these organisms deal with perturbations that cause cell deaths in ontogenesis has been a long-standing puzzle. We propose and demonstrate that the developmental cell lineages of these species are structured such that the resulting cellular compositions of the organisms are only moderately affected by cell deaths. A series of highly nonrandom features of the cell lineages underlie their developmental robustness and these features likely originated as adaptations in the face of various disturbances during development. Our findings reveal important organizing principles of determinative developmental cell lineages and a conceptually new mechanism of phenotypic robustness, which have broad implications for development and evolution.
| Phenotypic robustness, often referred to as canalization, is the phenomenon that a phenotypic trait is invariant in the face of environmental or genetic perturbations [1]–[7]. Phenotypic robustness allows the maintenance of high fitness even under suboptimal conditions, which are not uncommon in nature [1]–[7]. Phenotypic robustness may also facilitate adaptation under certain conditions [8]. The genetic basis of phenotypic robustness has been of long-standing interest, and several underlying mechanisms have been elucidated [2], [3], [5]. For instance, capacitors such as molecular chaperones can buffer the disturbances from stressful environments and deleterious mutations; phenotypic variance is exposed upon the removal of capacitors [9], [10]. Functional redundancy in genetic systems is another cause of robustness because it renders the phenotype of an organism relatively invariant to the loss of a genetic component. Such redundancies are known to exist at both the individual gene level (e.g., between duplicate genes) [11] and the systems level (e.g., between alternative metabolic pathways) [5], [12]. Several evolutionary mechanisms explain the origin and maintenance of such functional redundancies and the resulting robustness [12]–[15]. Other proposed mechanisms of robustness include expression regulation via transcriptional regulatory networks [16], posttranscriptional regulation by microRNA [17], [18], and certain feedback/feed-forward circuits in signaling among cells [19].
It has long been recognized that ontogenesis, or the development of an organism from a fertilized egg to an adult, is a key component of phenotypic robustness [1]. But the mechanism underlying the ontogenic robustness is not well understood. Regulative development, where rescuing processes may be triggered in response to cell deaths caused by environmental or genetic perturbations, could ensure ontogenic robustness. However, regulative development usually accompanies massive cell rearrangements and migration before or during cell fate specification [20], which is not a desirable feature in species or tissues that have short developmental time, let alone the complex genetic or cell-cell communication network required for the regulation. In fact, no embryo displays only regulative development [20]. Even in largely regulative embryos, one finds determinative (also known as mosaic) development [20], where the developmental process and cell fate are fixed. In invertebrate embryos, especially those of mollusks [21], annelids [22], tunicates [23], and nematodes [24], [25], determinative development is extensively observed [20]. How do these species deal with environmental or genetic perturbations in ontogenesis? To answer this question, we investigate the ontogenic robustness of three invertebrates dominated by determinative development, using developmental cell lineages that describe the exact genealogical relations of all cells of an individual embryo or adult. We show that the determinative development of these invertebrates is highly robust to two types of cell death, which approximate the effects of random environmental disturbances (or somatic mutations) and genetic disturbances (i.e., germline mutations), respectively. We identify multiple extremely nonrandom features of the cell lineages that explain the ontogenic robustness, and show by evolutionary simulation that this characteristic can arise as an adaptation to certain disturbances in ontogenesis.
A typical developmental cell lineage takes the form of a binary tree composed of nodes and branches (Fig. 1A). The nodes represent cells, whereas the branches show descendant relationships among cells. There are two categories of nodes: terminal and internal. Terminal nodes represent cells at the final stage of the developmental process represented by the cell lineage, which may or may not be the final stage of development. Terminal cells are the ultimate product of the ontogenesis represented by the cell lineage. By contrast, internal nodes represent direct or indirect progenitors of the terminal cells, which are produced through differentiation and proliferation of the internal nodes. Here we consider only those cell lineages that start from the zygote.
In organisms such as the nematode worm Caenorhabditis elegans, cell fate determination is generally autonomous such that dead or degenerated cells are rarely replaced or compensated by other cells [25]. In other words, the cell lineage and the identity of each cell in the lineage are essentially invariable among individuals. We first classify the terminal cells in a cell lineage into different functional types [25], [26]; cells of different types perform distinct physiological functions whereas cells of the same type perform similar functions. From a cellular perspective, the ultimate consequences of both environmental and genetic disturbances to ontogenesis may be largely represented by the loss of certain terminal cells, although other consequences also exist (see Discussion). Because of the distinct functions of different terminal cell types, it is reasonable to assume, to a first approximation, that the probability of an organism to survive and reproduce is determined by the product of the weighted fraction of live terminal cells of each cell type. That is,(1)where T is the total number of terminal cell types except apoptotic cells, Ni is the number of terminal cells of type i in the absence of any disturbance, ni (ni≤Ni) is the actual number of live terminal cells of type i, and ai≥1 is an exponent reflecting the relative importance of cell type i to organismal growth and reproduction. For simplicity, we describe our results using ai = 1 for all i. Using other ai values yields similar results (Figs. S1A–F). Here f can be viewed as a measure of developmental robustness to cell deaths. While f is likely a component of Darwinian fitness, it is not equivalent to fitness.
Impacts of environmental and genetic perturbations on cell lineages manifest as the loss of internal and terminal cells. When an internal cell dies, all of its direct and indirect descendant cells are regarded as lost. We consider two types of perturbation. The first type is referred to as necrosis [27] or simply random cell death. This type of lineal perturbation mimics environmental disturbances or somatic mutations that lead to accidental deaths of individual cells. Note that our use of necrosis is different from that in some literature where it also includes cell death caused by germline mutations [27]. The second type of perturbation is referred to as division program failure (see Fig. 1B and below for the definition of cell division programs), which mimics germline mutations that cause the deaths of all cells that use a particular genetic program for cell division. We consider all internal cells that use the failed program to be arrested, resulting in the loss of all direct and indirect descendants of these internal cells. The two types of perturbation only approximate environmental disturbances (and somatic mutations) and genetic disturbances (i.e., germline mutations), respectively, because of several kinds of exceptions (see Discussion). Let fn and fp be the f value in the presence of necrosis and program failure, respectively. To estimate fn, we first calculate f when one non-root cell and all of its descendants are removed (Fig. 1C). We repeat this process for every non-root cell in the lineage and calculate the mean resulting f, which is the expected lineage robustness to the death of a randomly chosen non-root cell. A real cell lineage is said to be robust to necrosis if its fn is significantly higher than that expected from a randomized cell lineage that produces the same terminal cells.
To estimate fp, we follow a previous definition of cell division programs [26]. Every cell division in the lineage produces two daughter cells from a parental cell. The program used in the division is defined entirely by the types of the daughter cells. Here, a daughter cell may be terminal or internal, meaning that its type may be a terminal cell type or a division program (see node colors in Figs. 1A, B). Thus, two internal cells that give rise to the same types of daughter cells use the same program. For example, node I5 and I8 in Fig. 1A both use the program P3 (Fig. 1B). This definition is supported by the observation that the transcriptome of a cell is largely determined by the cell fate rather than the lineal history [28]. We traverse the entire cell lineage to define all division programs. If one program fails, all internal cells that use the program and all of their descendant cells are lost (Fig. 1D). We thus estimate fp by calculating the expected f using Eq. (1), with a specific per-program rate of failure being 1 over the number of internal cells (see Materials and Methods). A cell lineage is said to be robust to program failure if its fp is significantly greater than that expected from a randomized cell lineage that produces the same terminal cells.
We used determinative developmental cell lineages starting from the zygote and up to a >100-cell stage. To our knowledge, such lineages have been completely described in only three animal species: Caenorhabditis elegans, Pellioditis marina, and Halocynthia roretzi. Previously reported developmental cell lineages of other species are incomplete in internal or terminal cells on the lineage tree, their mother/daughter relationship, and/or functional categorization of terminal cells, and thus cannot be analyzed here (Table S2). Among the three species to be analyzed here, the nematode C. elegans is the first animal with its developmental cell lineage mapped at the single cell resolution [25]. Here we use the C. elegans cell lineage producing 671 terminal cells during the hermaphrodite embryogenesis, and categorize the terminal cells by standard anatomical descriptions [26], [29]. P. marina is another nematode, but lives in the sea. The cell lineage of P. marina, followed up to muscle contraction, has 638 terminal cells [24], [26]. H. roretzi, commonly known as the sea squirt, is an ascidian. We use the cell lineage of H. roretzi up to the 110-cell stage in this study [26], [30]. Hence, our analysis includes representatives from both Protostomia (C. elegans and P. marina) and Deuterostomia (H. roretzi), the two subgroups of triploblastic animals.
We estimated fn and fp of each of the three cell lineages by considering necrosis (Figs. 1E, G, I) and program failure (Figs. 1F, H, J), respectively. To compare with each real lineage, we generated 10,000 random cell lineages under the assumption that, at any developmental stage, all terminal cells have the same probability of cell division. Computationally, each of these random lineages was created by randomly coalescing the terminal cells of the real lineage exactly two cells at a time (Fig. S2). We found that fn and fp are significantly greater in each of the three real lineages than in their corresponding random lineages (P<0.0001; Figs. 1E–J), demonstrating that the three animal cell lineages are robust to both necrosis and program failure. The reduction in f (from 1) caused by necrosis is on average 19.6, 20.4, and 12.1% smaller in the three real lineages, compared with their respective random lineages. The corresponding numbers are 22.8, 22.2, and 27.8% for program failure. If the terminal cells of the same functional type are not equally important [31], one could divide a cell type into subtypes when estimating fn and fp, which would also result in more cell division programs. To evaluate the impact of considering subtypes of terminal cells and division programs on our results, we divided C. elegans neurons into five subtypes [26] and altered the definition of division programs correspondingly. The results, however, are qualitatively unaltered (Figs. S1G, H). We also examined an expanded hermaphroditic C. elegans cell lineage that includes post-embryonic cells, totaling 937 terminal cells, and the results are similar (Figs. S1I, J). One assumption made in our analysis is that all cells (or programs) have the same probability of necrosis (or failure). To investigate the impact of this assumption, we allowed different cells (or programs) to have different rates of necrosis (or failure) that follow an exponential distribution with the mean of the distribution identical to the constant rate used above. The obtained results are, however, similar (Fig. S1K, L), suggesting that our results are not sensitive to the assumption of equal necrosis (or program failure) rates. Because the overall results from the three species are similar, from now on, we present our findings from C. elegans in the main figures and those from the other two species in the accompanying supplementary figures.
To understand the underlying mechanisms of the observed lineage robustness to necrosis and program failure, we examined various characteristics of both the real and random cell lineages. Let us define the depth of a cell in a cell lineage by the number of cell divisions required to generate the cell from the root, which is the zygote in the lineages analyzed here. To a terminal cell, the smaller its depth, the lower the probability of its loss, because each cell division carries risks of necrosis and program failure. Hence, reducing the depths of terminal cells should improve f.
Let us define the maximum depth of a cell lineage by the largest depth among all terminal cells in the lineage. Indeed, the maximum depth is significantly smaller in the three real lineages, compared with their corresponding random lineages (Fig. 2A; Figs. S3A, B). The theoretical minimum of the maximum depth of a cell lineage with a total of L terminal cells is [log2L], where the square bracket represents the minimal integer that is no smaller than the number inside. The theoretical minimum of the maximum depth is 10, 10, and 7 for the three real lineages, respectively. The observed maximum depth is 12, 11, and 7, respectively, indicating that the real maximum depth is either identical to or close to the theoretical minimum. By calculating fn and fp of random lineages with different maximum depths, we found that, on average, fn and fp both increase as the maximum depth decreases (Figs. 2B, C; Figs. S3C–F), suggesting that reducing the maximum depth of a cell lineage tends to increase its robustness to necrosis and program failure. Intriguingly, the real lineages have significantly greater fn and fp than the random lineages with the same maximum depths (Figs. 2B, C; Figs. S3C–F), revealing the presence of additional factors that contribute to the high fn and fp of the real lineages.
We found the mean depth of all terminal cells in each of the three real lineages to be significantly smaller than that of their corresponding random lineages even when the maximum depths of these random lineages are fixed at the observed values (Fig. 2D; Figs. S3G, H). For each real lineage, we then generated 10,000 random lineages that have the maximum depth identical to the observed value and the mean depth similar to the observed value (See Materials and Methods). The mean depth is found to impact fn and fp negatively even when the maximum depth is fixed (Figs. 2E, F; Figs. S3I–L), confirming that a smaller-than-expected mean depth given the maximum depth is another contributor to the high fn and fp of the real lineages. Nonetheless, the real lineages still have greater fn and fp than the random lineages of the same maximum and mean depths (Figs. 2E, F; Figs. S3I–L), suggesting that additional factors contribute to the high fn and fp of the real lineages.
When the depths of all terminal cells are fixed, the only thing in a cell lineage that can vary is the lineal topology. Here, the term “topology” is equivalent to that in phylogenetics, including both the lineage tree structure when the terminal cells are unlabeled and the arrangement of the labeled terminal cells given the tree structure. To test the impact of lineal topology on robustness, for each real lineage, we generated 10,000 random lineages with varying topologies but the same depth distribution as in the real lineage for all terminal cells. Again, fn and fp are significantly greater in the real lineages than in their respective randomized lineages (Figs. 3A, B; Figs. S4A–D), demonstrating the contribution of lineal topology to the high robustness of real lineages.
Furthermore, fn and fp are significantly greater in each real lineage than in its corresponding random lineages that share the same tree structure, which were generated by relabelling the terminal cells in the real lineage (Figs. 3C, D; Figs. S4E–H). Thus, the organization of terminal cells contributes to the robustness of the real lineages.
What features of the terminal cell organization underlie the high robustness of the real lineages? As mentioned, terminal cells have been classified into functional types. We define the size of a cell type by the number of terminal cells belonging to the type. Cells of large cell types are referred to as common cells, whereas those of small cell types are referred to as rare cells. According to Eq. (1), loss of a rare cell has a larger adverse impact on f than the loss of a common cell. Because a low-depth terminal cell is less likely than a high-depth terminal cell to be lost, one strategy to improve f, given the lineage tree structure, is to arrange the terminal cells in such a way that the rare cells have relatively low depths and common cells have relatively high depths. Indeed, in each of the three real lineages, a positive correlation exists between the depth of a terminal cell and the size of its cell type (see binned results in Fig. 4A; Figs. S5A, B). This correlation (ρrare-early) is significantly stronger than the chance expectation, which is calculated using 10,000 lineages constructed by randomly relabelling the terminal cells of each real lineage (Fig. 4B; Figs. S5C, D). A comparison among the random lineages shows that fn and, to a much lesser degree, fp increase with ρrare-early (Figs. 4C, D; Figs. S5E–H), confirming the prediction that early appearances of rare cells in a lineage render the lineage more robust.
The above analysis depends critically on the classification of terminal cells. For instance, if the late-appearing neuron cells are divided into many subtypes, the rare-early correlation would be weakened. It is thus important to classify terminal cells objectively. To this end, we analyzed the recently published single-cell expression levels of 93 genes in 363 cells of the L1 stage larvae of C. elegans [28]. Three of the 363 cells are not terminal cells in the lineage considered here and are thus removed. We then classified the remaining 360 terminal cells into eight types (Fig. S5I) based on transcriptome similarities among cells (see Materials and Methods), because the terminal cells of the C. elegans lineage analyzed here were previously classified into eight functional types after the removal of apoptotic cells [26], [29]. Although the new classification differs substantially from the previous classification (Fig. S5I), the rare-early correlation in the C. elegans lineage remains highly significant under the new classification (P<0.0001, compared with 10,000 random lineages with the same lineage tree structure; Fig. S5J), and this result is insensitive to the number of cell types classified (from 4 to 40) (Fig. S5K). Because only 360 of the 671 terminal cells in the C. elegans lineage analyzed here have the single-cell gene expression data, the new cell type classification is incomplete and hence cannot be used to calculate fn and fp (see Table S1). But, because cell type reclassification does not alter the shallowness and tree structure of the lineage and because the rare-early correlation clearly remains unchanged, the reclassification should not qualitatively affect fn and fp.
Besides its contribution to robustness, the rare-early correlation has other potential implications. For instance, the risk of mutation is minimized for cells that appear early. In the case of the C. elegans cell lineage analyzed here, the early appearance of the two germ cells may have reduced the germline mutation rate per nematode generation. Note, however, that the rare-early correlation is unlikely to have been caused by natural selection for a low germline mutation rate, because the correlation remains strong (ρ = 0.515; P<10−38) even when the germ cells are not considered. A potential alternative explanation of the rare-early correlation is that the rare cells have physiological roles to support the developing embryo and hence need to be produced earlier. However, this hypothesis does not appear to be empirically supported. For example, the two germ cells that appear very early in the cell lineage have no physiological role in supporting the developing embryo.
Intuitively, one may think that cell types are clonal, meaning that all terminal cells of a cell type form a monophyletic or paraphyletic group in a cell lineage tree [32], [33]. This intuition, however, is incorrect. Studies in multiple animal species have shown that cell types are typically nonclonal or polyphyletic [34]–[41], meaning that the terminal cells of the same cell type are derived from multiple sublineages (e.g., see Fig. 1A). If we compare two cell lineages with the same tree structure, an internal cell death will kill the same number of terminal cells in the two lineages, but the dead terminal cells are more likely to be of the same type in the clonal lineage than in the non-clonal lineage. Because the loss of multiple terminal cells of the same type tends to result in a lower f than the loss of the same number of terminal cells distributed among several types (see Materials and Methods), we predict that clonality reduces lineage robustness while non-clonality improves lineage robustness.
To test the above hypothesis, let us measure the clonality of a cell lineage by(2)where T is the total number of terminal cell types, M represents all pairs of terminal cells, where a subscription of i limits the cell pairs within cell type i, and dj is the lineal distance between cell pair j, which is the number of edges on the shortest path connecting the cell pair (e.g., the lineal distance between the two T3 cells in Fig. 1A is 4). For a given lineage tree structure, in the equation is fixed, whereas decreases with clonality. Thus, the stronger the clonality, the larger the C.
To examine the role of clonality on robustness, we controlled all known lineage features that contribute to robustness. Specifically, from a real cell lineage, we rearranged the terminal cells without altering their respective depths and generated a series of random lineages with different levels of clonality (see Materials and Methods). We found that (i) C is lower in the three real lineages (triangles in Fig. 5A; Figs. S6A, B) than in the corresponding random lineages where all cell types are as clonal as possible given the constraints of the lineage tree structure and cell depths (purple dots; P<0.12, 0.02, and 0.02, respectively) and (ii) a decrease in C leads to an increase in lineage robustness to both necrosis (Fig. 5A; Figs. S6A, B) and program failure (Fig. 5B; Figs. S6C, D). These observations support our hypothesis that the non-clonality of the real lineages improves their robustness. Notably, the non-clonality, in combination with the previously identified lineage features, seems sufficient to explain the observed robustness of the real lineages (Fig. 5A, B).
A potential alternative explanation of the non-clonality of the real lineages is that it is dictated by some spatial requirements for terminal cells. Because not all cells of the same type are physically proximate (e.g., epithelial cells on the two hands of a person) and because cells may possess limited abilities to migrate, be costly to migrate, and/or have reduced migration under natural selection for rapid development [24], [42], some cell types are necessarily non-clonal. To examine whether some spatial requirements have dictated the reduction of the clonality of the real lineages and as a result created the observed robustness as a byproduct, we obtained the three-dimensional spatial coordinates of terminal cells in C. elegans [43]. We then calculated Spearman's rank correlation (ρp-l) between the physical distance and lineal distance for all pairs of terminal cells that belong to the same type (see Materials and Methods). The C. elegans lineage indeed has a much greater ρp-l (triangle in Fig. 5C) when compared with the above generated random lineages of similar C (dots in Fig. 5C). Nevertheless, the C. elegans ρp-l is apparently not maximized, because we were able to acquire an even higher ρp-l by rearranging the terminal cells of the C. elegans lineage within their respective depths (square in Fig. 5C). More importantly, a comparison among the above generated random lineages shows that ρp-l tends to increase with C, suggesting that lowering C does not help raise ρp-l. In other words, the observed low clonality cannot be explained by the spatial requirements.
Notwithstanding, the C value of each real lineage remains significantly greater than that when all of its terminal cells are completely randomly situated within their respective depths (dark blue dots in Fig. 5A; Figs. S6A–B; P<0.02 in all three species). Two factors may have constrained the further reduction of C and the further rise of robustness in the real lineages. The first potential constraint arises from a demand for spatial proximity of certain terminal cells of the same type. Because cell migration is limited [24], [42] and because some cell types exert their functions through physical connections such as in the neural system and muscles, there is a requirement for a large number of terminal cells of the same type to be produced next to one another, which hinders a further reduction of C. To test this hypothesis, we focused on pairs of terminal cells that share their immediate progenitor as well as their cell type. By analyzing the spatial coordinates of terminal cells, we found that such “twin” terminal cells are significantly closer physically to each other than each is to other cells of the same type (Z = −15.32, P<10−52; see Materials and Methods), suggesting that the existence of such “twin” terminal cells is in part a result of the spatial requirements. There are also significantly (P<10−4) more twin terminal cells in C. elegans (155) than the random expectation (59.4), which is estimated by randomly rearranging terminal cells within their respective depths. To examine if this spatial requirement constrains the reduction of clonality and the rise of robustness, we retained the relationship of twins while randomly rearranging the terminal cells within their respective depths. In support of our hypothesis, a greater C and smaller fn and fp are observed after this type of rearrangement (blue dots in Figs. 5D, E and Figs. S6E–H), compared with the rearrangement without the twin constraint (red dots).
The second potential constraint is lineage complexity [26], which is the number of division programs used in the lineage, as shown in Fig. 1B. Azevedo and colleagues suggested that lineage complexity has been selectively minimized in evolution [26]. Because lineage complexity negatively correlates with C (Fig. 5F; Figs. S6I, J), selection for lower complexity is expected to increase clonality and hence decrease lineage robustness.
Taken together, our analyses revealed lower-than-expected clonalities in the real cell lineages, which have likely resulted from the potential selection for lineage robustness. Why the clonalities are not further reduced may be explained by the spatial constraints of certain cells and the possible selection for lineage simplicity.
We have demonstrated the robustness of the three animal cell lineages to necrosis and program failure and have identified a number of lineage features or mechanisms that underlie the observed robustness. But how did such lineage robustness originate? It is not obvious that natural selection for high robustness during the evolutionary expansion of cell lineages will result in high robustness, because today's lineage is greatly restricted by its ancestral forms. For instance, 577 terminal cells have the same fate in the lineages of the two nematodes C. elegans (671 terminal cells) and P. marina (638 terminal cells) [26]. Below we investigate by computer simulation if the observed robustness of the three cell lineages is achievable simply by adaptation to necrosis or program failure during the evolutionary expansion of cell lineages.
Our simulation, named “macroevolution”, mimics the expansion of a cell lineage in macroevolution by stepwise additions of new terminal cells via divisions of existing terminal cells. That is, upon a division, one daughter cell inherits the identity of its parental cell while the other evolves into a new cell type (see Materials and Methods). Our simulated lineages are the same as the real lineages in the number and identities of terminal cells, but differ in lineal topology and terminal cell depths. Different intensities of selection for high fn or high fp are applied during the course of the macroevolution. The actual evolution of developmental cell lineages does not necessarily proceed as our macroevolution simulation, because cell fate may change in evolution [24]. In other words, the actual evolution of developmental cell lineages may be more flexible and our overly-constrained macroevolution likely reveals the lower limit of robustness achievable under natural selection for robustness.
We found that selection for high fn can result in lineages that have similar levels of fn as observed in the real lineages (when the selection intensity is 0.5), while removing the selection results in lineages of much lower fn (Fig. 6A; Figs. S7A, K). Compared with the lineages generated under no selection, those generated by selecting for high fn exhibit the features known to improve fn, including the lineage shallowness (although slightly less extreme than in the real lineages) and rare-early correlation (Figs. 6B–D; Figs. S7B–D, L–N). Clonality C is not compared because it is not comparable among lineages with different tree structures. Intriguingly, selection for high fn also increases fp (Fig. 6E; Figs. S7E, O). On the contrary, selection for high fp fails to recapitulate the observed level of fp or fn (Figs. 6F, J; Figs. S7F, J), except in the case of the H. roretzi lineage (Figs. S7P, T), which may be too small to be informative. While selection for high fp does result in shifts of the robustness-enhancing lineage features in the predicted directions, these shifts tend not to reach the levels observed in the real lineages (Figs. 6G–I; Figs. S7G–I, Q–S; see Discussion). These findings reveal the possibility that the higher-than-expected fn and fp in the real lineages are both due to natural selection for high fn. In other words, the observed high fp in the real lineages is probably a byproduct of the selection for high fn. Consistent with this conclusion is the finding that fn and fp are highly positively correlated among various randomized lineages generated from the real lineages (Figs. S7U–W). While our simulation does not prove that the higher-than-expected fn and fp in the three animal cell lineages are caused by selection for fn, it does provide strong evidence for the viability of the hypothesis.
When rearranging terminal cells of a real lineage within their respective depths, we showed that increasing lineage clonality reduces lineage robustness (Figs. 5A, B) and complexity (Fig. 5E), implying a positive correlation between robustness and complexity. A more extensive analysis, however, revealed that the correlation between robustness and complexity may be positive or negative, depending on the groups of random lineages compared (Figs. 7A, B). For example, comparing lineages generated by shuffling terminal cells within their respective depths (red dots in Figs. 7A, B and Figs. S8A–D) and those generated by completely shuffling all terminal cells (green dots), we found that the rare-early correlation increases robustness but decreases complexity. By contrast, when we compare the lineages generated by within-depth shuffling with (blue dots in Figs. 7A, B and Figs. S8A–D) or without (red dots) retaining twin terminal cells, we found a positive correlation between robustness and complexity, suggesting that the spatial constraint of terminal cells decreases complexity at the cost of robustness.
Because lineage robustness and complexity are sometimes negatively correlated, selection against complexity (or for simplicity) [26] may result in high robustness, and vice versa. We thus used our macroevolution simulation with relatively strong selection (intensity = 0.05) to investigate if selection for robustness to necrosis or simplicity alone can account for both the robustness and simplicity of the real lineages. To investigate the interplay between robustness and simplicity, we used different fitness functions in the macroevolution simulation with different weights for robustness (R) and simplicity (S) (dash-lined box in Figs. 7C, D; Figs. S8E–H, where the relative weights for R and S are reflected by the power of R; see Materials and Methods for details). Four observations were made (Figs. 7C, D; Figs. S8E–H). First, selection for simplicity alone enhances simplicity but not robustness (fn or fp), suggesting that robustness cannot be a byproduct of selection for simplicity (blue dots in the figures). Second, selection for robustness to necrosis alone enhances robustness (fn and fp) as well as simplicity (red dots). Nevertheless, the resulting lineages are still more complex than the real lineages, suggesting that the simplicity of the real lineages may be partially but not entirely due to selection for robustness to necrosis. Third, simultaneous selections for both robustness and simplicity can generate lineages that are similar to the real lineages in fn, fp, and simplicity (green, purple, and orange dots), supporting the actions of both selective forces in the evolution of the real lineages. Fourth, under the parameters used, selection for robustness alone generates lineages that exceed the real lineages in fn and fp, whereas this disparity disappears under simultaneous selections for robustness and simplicity, supporting the notion that a requirement for simplicity prevents a further increase in robustness.
Phenotypic robustness is an important characteristic of life, but its underlying mechanisms and evolutionary origins are not well understood. In this article, we demonstrated that determinative developmental cell lineages are robust to necrosis and cell division program failure, which approximately represent environmental (or somatic) and genetic (i.e., germline) disturbances, respectively. Although some germline mutations with incomplete penetrance may act like environmental perturbations and some environmental perturbations impact specific programs and hence behave like germline mutations, our analyses included both types of disturbances. We further showed by computer simulation that such robustness can arise from natural selection for robustness against necrosis during the evolutionary expansion of cell lineages. Our findings thus reveal a new mechanism of phenotypic robustness as well as its potential evolutionary origin. Different from almost all previously studied mechanisms of phenotypic robustness, which act at the subcellular or cellular levels, cell lineage robustness manifests at the supracellular level and hence is a unique feature of multicellular organisms. Our study also provides a novel explanation of the contribution of ontogenesis to phenotypic robustness.
Necrosis caused by environmental stresses or somatic mutations are unavoidable [27]. Similarly, cell division program failure is expected to occur occasionally due to germline mutations [27]. Although mechanisms buffering stresses and mutations may exist such that the rate of necrosis or division program failure is lowered [2]–[5], [44], cell lineage robustness is likely an important mechanism buffering the adverse effect of necrosis and program failure upon their occurrence. This mechanism is especially important for species whose development is predominantly determinative because the cell fate is largely fixed in these organisms. Cell lineage robustness, along with regulative development (see below), complements subcellular and cellular mechanisms of robustness to form a multi-layer defense system against environmental and genetic perturbations that are common in nature.
A highly related subject is the robustness of sublineages, which are lineages starting from non-zygote cells. Although our definition of robustness is directly applicable to sublineages, the expectation for sublineage robustness may vary, depending on the specific types of sublineages. For instance, regulative development occurs to a small number of cells during the late development of C. elegans, where cell losses may sometimes be compensated by additional divisions of neighboring cells [25]. The sublineages with the ability of such compensatory growth, conferred by regulative development, may not have the typical properties of robust lineages such as low cell depths. Compensatory growth may be further generalized to include sublineages that are populated by stem cells, commonly seen in arthropods and vertebrates [45]. Other than delaying the development, such compensatory growth can apparently solve the problems caused by necrosis [46]. The existence of compensatory growth demonstrates canalization mechanisms other than the cell lineage robustness revealed here, further supporting the importance of having a robustness developmental process. With a modification of our model, stem cells may be included such that the robustness of developmental cell lineages can be evaluated for organisms with prevalent compensatory growth.
Our findings of cell lineage robustness and its mechanisms offer important biological insights. For example, somatic-mutation-based analysis in mice revealed that cells of the same type from the same organ such as the kidney have several different lineal histories [38], [39]. We showed that this phenomenon of non-clonality is beneficial to lineage robustness, at the cost of lineage simplicity. Thus, the benefit of having increased robustness may exceed the cost of having extra division programs for these cell types. In general, features contributing to lineage robustness (e.g., rare cells are produced relatively early in ontogeny) may be found to be detrimental to other traits (e.g., common cells would have relatively high probabilities of death). Knowing the underlying tradeoffs helps understand the origins of such detrimental features, which may lead to potential solutions.
One intriguing finding from our macroevolution simulation is that both fn and fp can be raised by natural selection for high fn but not as effectively by selection for high fp. This disparity is not because the parameters we used render the cell death rate higher in the presence of necrosis than in the presence of program failure. In fact, the opposite is true (e.g., see Fig. 1E–J). The disparity may be related to the larger variation in fp than fn among individuals whose expected lineages under no cell death are identical. Even when the rates of necrosis and program failure are fixed, individuals with the same expected lineages can still have different f values because necrosis and program failure are stochastic. Based on the necrosis and program failure rates used here, we estimated in C. elegans that the standard deviation of fn among individuals is 0.1095, while that of fp is 0.1745. It is reasonable to assume that a larger variation in f translates into a larger variation in fitness. Because the larger the variation in fitness among (isogenic) individuals, the lower the efficacy of natural selection [47], selection for fp is less effective than selection for fn in raising lineage robustness. Nevertheless, whether the standard deviations of fn and fp are directly comparable is unclear and other explanations may exist.
Our finding that cell lineage robustness can result from natural selection for robustness, coupled with previous findings on the possibility of selection for genetic and environmental robustness [2], [3], [7], [48], suggests the likelihood that the observed cell lineage robustness against necrosis and program failure is a direct result of natural selection for robustness. Nevertheless, we cannot exclude the possibility that the observed robustness results partially or entirely as a byproduct of other selections or generative biases [49]. For example, we showed that a low maximum depth improves the robustness of a cell lineage and that the maximum depths of the three animal cell lineages are identical or close to their theoretical minimums. Although selection for robustness against necrosis may explain this phenomenon (Fig. 6B), selection for short developmental time is another possible explanation [50]. It was previously thought that the robustness-enhancing feature of non-clonality of the cell lineages of C. elegans and P. marina results from selection for rapid development that avoids the time-consuming cell migration [24]. But the recently determined nearly complete cell lineages of two other nematodes that develop slowly (Halicephalobus gingivalis and Rhabditophanes sp.) exhibit similar non-clonality [51], [52], suggesting that selection for rapid development is not the cause of the observed non-clonality in nematodes. Furthermore, selection for fast development cannot explain other robustness-enhancing features of cell lineages (e.g., early appearances of rare cell types). Given the presence of multiple robustness-enhancing features in animal developmental cell lineages, natural selection for robustness is a plausible and the most parsimonious explanation of the origin of lineage robustness. The generative bias hypothesis would assert that the observed high robustness of developmental cell lineages is caused by mutational biases. But it is difficult to imagine that such biases would generate multiple nonrandom features that all happen to increase lineage robustness.
Our estimation of cell lineage robustness is based on several assumptions. First, we used a necrosis rate of one cell per lineage and a program failure rate of 1/Ninternal per program, where Ninternal is the total number of internal cells in the lineage. These parameters are probably lower than the actual rates, rendering our estimates of cell lineage robustness conservative. Our results should be verified in the future with the real necrosis and program failure rates when they become available. It is also possible that the rate of necrosis and that of program failure vary among cells and programs [53]. Such information, when it becomes available, can be incorporated into our model (Fig. S1K, L) to achieve a more accurate estimate of cell lineage robustness.
Second, the robustness function, as described by Eq. (1) under ai = 1, may not be accurate, because of (i) variable importance of cells of different types, (ii) imprecise cell type classification, (iii) interactions among different cell types, and/or (iv) potential compensation of cell death by other mechanisms. Yet, our conclusion appears robust to variable impacts of different cell types, because the use of different ai values yielded similar results (Fig. S1A–F). Our conclusion is also highly robust to cell type classification, because transcriptome-based and function-based classifications yielded similar results that are invariant to the number of cell types classified (Figs. S5I–K). Further division of a cell type into subtypes did not alter our results (Fig. S1G, H). Our results also appear to be robust to the variation in the necrosis rate among cells or the failure rate among programs (Figs. S1K–L). While necrotic cell death is rarely compensated by regeneration of the corresponding live cells in C. elegans [25], this may not be the case in some complex organisms especially during late development [46]. Similarly, it is possible that only some but not all descendants of an affected internal cell get lost. By contrast, when cell induction or replacement occurs in cell fate determination, a cell death may lead to the loss of terminal cells that are non-descendants of the dead cell. Furthermore, a perturbation may induce the production of extra terminal cells (e.g., by blocking apoptosis [54]). These variations, as well as potential interactions among different cell types, have not been considered in our analyses, but can be studied in the framework developed here when detailed information about these processes becomes available.
Third, it is possible that the potential type of an internal cell (i.e., its division program) is limited genetically, but such a constraint is not explicitly modeled in our randomization of cell lineages. With a better understanding of this constraint, we can refine our randomization in the future. Nonetheless, it is worth mentioning that the expression profiles of 93 genes examined in C. elegans terminal cells were found to be largely determine by their cell types [28], suggesting that the division programs of internal cells are not so limited, because otherwise the transcriptomes of the terminal cells should be dictated by their lineal histories.
By progressively constraining various features of a cell lineage, we identified several contributors to lineage robustness, including terminal cell depths, lineal topology, early appearances of rare cells, and non-clonality of cell types. Although the impacts of any two of these characteristics to lineage robustness are not completely overlapping, it is important to note that they are not completely non-overlapping either. As such, it is difficult to assess the relative contributions of these characteristics to lineage robustness. Although all of these characteristics exist in the three animal cell lineages examined, variations are expected when additional species are examined. For instance, in most organisms, primordial germ cells are set aside early in development [55]. In mice, however, these cells appear at a much later stage [40], [56], which likely reduces the rare-early correlation and the cell lineage robustness. But, mouse blastomeres up to the eight cell stage are equipotent [57], which reduces clonality and increases lineage robustness.
Using the macroevolution simulation, we showed that, in the evolutionary expansion of cell lineages, adaptation to random necrosis can result in highly robust cell lineages. Our simulation is a coarse-grained approximation rather than a precise description of the evolution of developmental cell lineages. However, because our simulation explicitly models historical contingency, the constraint imposed by ancestral lineages on future lineages, our simulation is more realistic than that by swapping sublineages in a real or random lineage [26]. Our model can be further improved by including the genetic networks that underlie cell fate determination [49] when such information becomes available. It can also be improved by allowing lineages to expand through divisions of internal cells rather than only terminal cells.
Decoding the developmental cell lineages of the human and other organisms is a grand scientific challenge (http://www.lineage-flagship.eu/). With the rapid advancement of genomics [58], especially single-cell genome sequencing [59]–[61], it will not be long before one can use somatic mutations accumulated during ontogenesis to reconstruct the cell lineages of complex organisms such as mammals [62], [63]. Our computational analysis of the three determinative cell lineages provides experimentally testable [39] hypotheses on the organizing principles of developmental cell lineages and opens the door toward characterizing systemic properties of complex cell lineages, an area that promises to be of both theoretical and applied values in understanding evolution, development, and carcinogenesis.
To reliably evaluate lineage robustness as defined in Eq. (1) of the main text, the cell lineage data must satisfy the following four criteria (Table S1). First, the lineage has the form of a binary tree. Second, the lineage starts from the zygote and contains all cells up to a developmental stage with at least 100 cells. Third, all the terminal cells at this stage must be included in the lineage data. Fourth, the terminal cells should be functionally categorized because the impact of a cell death depends on the cell type. There are only three developmental cell lineages that meet all these criteria (C. elegans, P. marina, and H. roretzi) and they were retrieved from an earlier publication [26]. Several other well-known cell lineages do not satisfy one or more of the four requirements and thus cannot be used here (Table S2). In the C. elegans cell lineage, 671 terminal cells were categorized by standard anatomical descriptions [25] as: 39 blast, 113 death, 93 epithelial (arcade, hypodermis, pharyngeal structural, rectum, valves), 2 germ, 13 gland (coelomocytes, excretory system, and pharyngeal glands), 20 intestinal, 123 muscle (including the head mesodermal cell), 46 neural structural cells, and 222 neurons. To consider the potentially different importance of cells of the same type, we subdivided neurons into two excretory canal neurons, 95 interneurons, 45 motor neurons, 26 polymodal neurons, and 54 sensory neurons [26]. We also validated our primary result using an expanded post-embryonic hermaphroditic C. elegans cell lineage with a total of 937 terminal cells, including 5 blast, 131 death, 262 epithelial (arcade, hypodermis, pharyngeal structural, rectum, valves), 2 germ, 13 gland (coelomocytes, excretory system, and pharyngeal glands), 20 intestinal, 153 muscle (including the head mesodermal cell), 46 neural structural, and 305 neuronal cells [29], [64]. For P. marina, the cell lineage up to muscle contraction, containing 638 terminal cells, were classified as: 81 body muscle, 67 death, 2 germ, 131 hypodermis, 20 intestine, 195 nervous system, 112 pharynx, and 30 unknown fate [24]. For H. roretzi, the cell lineage up to the 110-cell stage was used. The terminal cells were classified according to the fates of their descendants [30] as: 12 endoderm, 50 epidermis, 6 mesenchyme (and trunk lateral cells), 10 muscle (and trunk ventral cells), 16 nervous system (brain, nerve cord, palps, primordial pharynx, and sensory pigment cells), 10 notochord, and 6 undifferentiated.
If the program failure rate is p, the number of programs that fail in a lineage is a random variable b following the binomial distribution B(Nprogram, p), where Nprogram is the total number of unique programs in the lineage. After randomly picking b failed programs, we calculated f using Eq. (1). The above step was repeated 10Nall times to calculate the expected f. Here Nall is the total number of cells in the lineage. We used p = 1/Ninternal, where Ninternal is the number of internal cells in the lineage. The above assumed relationship between p and Ninternal ensures that the expected number of internal cells whose division programs fail is the same between a real lineage and all of its randomized lineages, which is required for a fair comparison of their fp values. For two reasons, the stochasticity involved in the estimation of fp is unavoidable. First, despite the constancy in the expected number of cells with failed programs, the expected number of failed programs varies between a real lineage and its random lineages. Second, because b could be large, it is computationally impossible to explore all possibilities in the event of multiple program failures.
We generated random lineages under eight different constraints. (i) We randomly coalesced the terminal cells of a real lineage (Figs. 1E–J) using the following procedure. Suppose there are m terminal cells. We randomly pick two of them (regardless of their cell types) and coalesce them, meaning that they become sisters and share the immediate progenitor cell. There are now m-1 cells left (m-2 terminal cells and 1 progenitor cell). We then randomly pick two cells from these m-1 cells and repeat the coalescence process until there is only one cell left. This process generates a random cell lineage of the m terminal cells (Fig. S2). (ii) We constrained the random coalescent process such that the maximum depth is fixed at a predetermined value (Figs. 2B–D). (iii) We constrained the random coalescent process such that the maximum depth is fixed at a predetermined value and the mean depth is close to a predetermined value (i.e., cell depths in a random lineage is a bootstrap sample of the real depths) (Figs. 2E, F). (iv) We generated random lineages by constraining the distribution of the depths of all terminal cells as in a real lineage but allowing variation of the lineage tree structure and depths of individual cells. In procedures (iii) and (iv), when the set of cell depths is given, each terminal cell is randomly assigned with one of the depths. We then randomly paired-up the xy cells at the maximum depth y as sister cells, creating xy/2 internal cells at depth y-1. It is repeated at depth y-1 for the (xy/2+xy-1) cells, and then recursively at depth y-2, y-3, …, and 1 (Figs. 3A, B). (v) We generated random lineages that have the same topology as a real lineage and then randomly shuffled all the terminal cells (Figs. 3C, D; Figs. 4B–D; Figs. 7A, B). (vi) We shuffled all the terminal cells in a real lineage within their respective depths (Figs. 5D, E; Figs. 7A, B). (vii) In addition to the constraint in (vi), we further maintained the twin terminal cells as twins in shuffling (Figs. 5D, E; Figs. 7A, B). (viii) We first defined a random order of cell types. Within each depth, g percent of terminal cells in every type are picked and sorted by the predefined type order, while the remaining unsorted cells are randomly inserted into the sorted list of cells. The cells are then assigned in the order of appearance in the list to the terminal nodes at that given depth from the left to the right of the lineage. The procedure is repeated for every depth (with the same cell type order) to create a lineage whose clonality increases with g (Figs. 5A–C, F). For all except the second and eighth constraints, we generated 10,000 random lineages. For the second constraint, we set the maximum depth as small as Dmin and as large as Dreal+2(Dreal−Dmin+1), where Dmin is the theoretical lower limit of the lineage's maximum depth and Dreal is the observed maximum depth of the real lineage. We then generated 5,000 random lineages for each possible maximum depth between these two extremes. For the eighth constraint, we used g at every 5th percentile, and created 50 random lineages for each value of g. The source codes for generating the random lineages can be downloaded from http://code.google.com/p/eadlin/downloads/list.
Gene expression profiles of 93 genes in 363 cells at the C. elegans L1 stage [28] were retrieved. Three of the 363 cells are not terminal cells in the lineage considered here and are thus removed. We then used the hclust function in the R package to hierarchically cluster the 360 cells based on the pair-wise Euclidean distances in the expression levels of the 93 genes. The tree is cut at an appropriate height to acquire a designated number of groups of cells (e.g. cutting at the root will result in two groups); these groups are regarded as transcriptome-based cell types.
The three-dimensional spatial coordinates of 334 terminal cells in the C. elegans lineage were retrieved from a recent paper [43]. The physical distance between two cells is the Euclidian distance between the centers of their nuclei [43]. To search for a cell lineage with ρp-l greater than that (0.2533) observed in C. elegans, we randomly generated 100 lineages that differ from the C. elegans lineage by only one swap between two terminal cells of the same depth and type. We chose the lineage with the highest ρp-l among the 100 random lineages, and repeated this process 100 times to obtain a lineage with ρp-l = 0.4085 (the square in Fig. 5C).
Using terminal cells with three-dimensional coordinates, we calculated the mean physical distance between a pair of twin cells in C. elegans. We similarly calculated the mean physical distance between a pair of randomly picked terminal cells of the same type for the same number of pairs as twins. We repeated this calculation 100 times to estimate the mean and standard deviation. These values allowed the calculation of a Z-score for the observed value from the twins.
The macroevolution simulation is designed to mimic the evolutionary expansion of a cell lineage under the constraint of its ancestral forms. Basically, the evolution is modeled by repeated additions of terminal cells, and each individual addition is called a round of bifurcation. To ensure that the macroevolution generates a cell lineage that is comparable to a given real lineage, we first completely shuffled the terminal cells of the real lineage to obtain a randomized terminal cell sequence. Starting from the first cell in the sequence as the founder cell, a lineage of m terminal cells is evolved by m-1 rounds of bifurcation. In each round, one random terminal cell from the evolved lineage is chosen and divided into two daughter cells. After the division, one of the daughter cells inherits the cell type of its parental cell, whereas the other is assigned the type of the next cell in the predetermined terminal cell sequence. The parental cell then becomes an internal cell with a division program generating its original cell type and the new cell type. At each step of lineage expansion, 100 random bifurcations are examined. Among them, a random lineage is chosen from the top k robust lineages as the starter of the next round of expansion, where k is adjusted between 1 and 100 to represent different selection intensities. The smaller the k value, the stronger the selection. In Fig. 6 and Fig. S7, the presented selection intensity equals k/100; in Fig. 7 and Fig. S8, k equals 5. Regardless of the cell lineage size, we used an expected necrosis rate of 1 necrosis per cell lineage or an expected program failure rate of 1/Ninternal failure per program. For the macroevolution involving selection for simplicity S ( = 1/complexity), we calculated lineage complexity [26] at every round of bifurcation and combined it with the robustness (R) to define the fitness of a lineage. For instance, R5S (Fig. 7B) means that fitness = R5S. Here R equals fn defined in Eq. (1). Under each parameter set, we repeated the macroevolution 100 times to access variations. For the developmental cell lineage of H. roretzi, to retain its fully symmetric feature during macroevolution, bifurcations were carried out in one half of the lineage, but the robustness was calculated after mirroring the half lineage. The source code of the macroevolution simulation can be downloaded from http://code.google.com/p/eadlin/downloads/list.
Based on Eq. (1), it is clear that when different cell types have different numbers of (terminal) cells, a cell death that happens to a common cell type would have a smaller effect on f than a cell death that happens to a rare cell type. Now let us consider the scenario of T terminal cell types, each with exactly N terminal cells. Let h terminal cells to die, where 1<h<N. If all dead cells are of the same type, we have . If we arbitrarily assign h1 (0<h1<h) cell death events to another cell type, we have . It can be shown that . Assigning h1 cell deaths to a second type and h2 cell deaths to a third type (0<h2<h-h1) would result in . It can also be shown that . The same is true when the cell deaths are distributed among more cell types. Thus, the loss of multiple terminal cells of the same type tends to result in a lower f than the loss of the same number of terminal cells distributed among several types.
|
10.1371/journal.pgen.1007954 | The complex genetic architecture of shoot growth natural variation in Arabidopsis thaliana | One of the main outcomes of quantitative genetics approaches to natural variation is to reveal the genetic architecture underlying the phenotypic space. Complex genetic architectures are described as including numerous loci (or alleles) with small-effect and/or low-frequency in the populations, interactions with the genetic background, environment or age. Linkage or association mapping strategies will be more or less sensitive to this complexity, so that we still have an unclear picture of its extent. By combining high-throughput phenotyping under two environmental conditions with classical QTL mapping approaches in multiple Arabidopsis thaliana segregating populations as well as advanced near isogenic lines construction and survey, we have attempted to improve our understanding of quantitative phenotypic variation. Integrative traits such as those related to vegetative growth used in this work (highlighting either cumulative growth, growth rate or morphology) all showed complex and dynamic genetic architecture with respect to the segregating population and condition. The more resolutive our mapping approach, the more complexity we uncover, with several instances of QTLs visible in near isogenic lines but not detected with the initial QTL mapping, indicating that our phenotyping accuracy was less limiting than the mapping resolution with respect to the underlying genetic architecture. In an ultimate approach to resolve this complexity, we intensified our phenotyping effort to target specifically a 3Mb-region known to segregate for a major quantitative trait gene, using a series of selected lines recombined every 100kb. We discovered that at least 3 other independent QTLs had remained hidden in this region, some with trait- or condition-specific effects, or opposite allelic effects. If we were to extrapolate the figures obtained on this specific region in this particular cross to the genome- and species-scale, we would predict hundreds of causative loci of detectable phenotypic effect controlling these growth-related phenotypes.
| The question of the complexity of the genetic variants underlying diversity in plant size and shape is central in evolutionary biology to better understand the impacts of selection and adaptation. In this work, we have combined the high resolution of a robotized platform designed to grow Arabidopsis plants under strictly-controlled conditions and the power of quantitative genetics approaches to map the individual genetic components (the 'QTLs') controlling diverse phenotypes, and hence reveal the so-called 'genetic architecture' of these traits. We show that the more we increase our resolution to map QTLs, the more complex of a genetic architecture we reveal. For instance, by focusing all of our mapping power on a small region representing 2.5% of the genome in an unprecedented phenotyping effort, we reveal that several independent QTLs had remained hidden in this region beyond a major-effect QTL that is always clearly visible. If this region is representative of the genome, this means that our current understanding misses potentially hundreds of variants finely controlling traits of evolutionary or agronomical interest.
| Fine-tuning plant growth throughout development and in response to environmental limitations is a decisive process to optimize fitness and population survival in the wild. As a sessile organism, plants have to cope with environmental fluctuations and evolved a wide range of responses. This is well illustrated by their great phenotypic plasticity and their ability to colonize very diverse habitats, through intraspecific genetic diversity as revealed in most pathways [1]. Aerial and below-ground growth represent a balance between resource investment in the structures and resource acquisition (respectively photosynthesis and water / nutrient uptake). Thus, growth is a highly complex trait controlled by many genes with constitutive or more specific roles depending on developmental stage, tissue, timing, environment [2–7]. In this context, plant growth can be considered as a model complex trait to increase our knowledge in the genetics of evolution, as well as to improve plant performance.
Forward mutant analysis plays a central role in plant biology to blindly identify gene functions associated with a phenotype [8], but sometimes remains limiting to reveal genes with modest phenotypic effect, or when addressing genes from redundant families. With regard to growth and stress tolerance, these limitations are likely to be relevant given the multigenic nature of growth phenotypes, the low mean effect at each locus and/or epistatic interaction they involve [9, 10]. Thus, the use of naturally-occurring variation through quantitative genetic approaches designed to map quantitative trait loci (QTLs) is interesting notably to complement the search for alleles selected during evolution which may not be brought out with classical loss-of-function approaches. Linkage mapping and genome-wide association lead to the identification of large amount of alleles involved in intraspecific phenotype variation from different plant species [1, 11].
With the drop of sequencing and genotyping costs, phenotyping clearly is the limiting factor for quantitative genetics approaches [12]. However, the complexity of the genetic architecture of a given trait, which depends on the contribution and the number of loci controlling a trait and their interactions with the genetic background and the environment, has direct consequences on how much phenotyping remains limiting. Highly heritable traits with a limited number of contributing loci (in a given segregating material, or at the species scale) are more likely to be well understood than more complex traits. For instance, a large part of the phenotypic variation for flowering time in Arabidopsis thaliana maps to a limited number of loci [13–16], including FRIGIDA and FLC genes [17–19] and thus has a relatively simple genetic architecture, although many more loci make smaller contributions -at least in some environments- and allelic heterogeneity also interferes [15, 16, 20, 21]. By contrast, traits like fitness or growth can be expected to have a more complex basis as they integrate many upstream traits, and consequently many genes, each prone to residual variation and heterogeneity. Smaller contributions from individual loci means that, although one can still estimate total heritabilities, the accuracy and throughput of phenotyping will be limiting to confirm individual QTLs' contributions. Heritabilities for flowering time-related traits will often be above 80%, while biomass accumulation or fitness' heritabilities are essentially found in the range 20–60% [22–27].
Another factor that will influence the genetic complexity of a trait is its response to the environment through phenotypic plasticity [28]. Part of the environmental fluctuations may be controlled in an experimental design, while another part may contribute to the residuals. Whether the sensitivity of a pathway or trait to the environment depends on the number and architecture of the contributing loci remains an open question, however the relationships between higher plasticity and lower heritability are described [26]. Water availability is an environmental factor that varies through space and time and shows great heterogeneity which certainly constrains plant growth and shapes plant distribution in nature and in agricultural systems. Prevalence of drought episode is expected to increase with global climate change making the understanding of plant response to drought one of the major challenge of the next decades [29, 30]; this includes deciphering the genetic basis for variation in mechanisms such as drought escape, avoidance and tolerance [31]. Hence, this environmental parameter is definitely a good candidate to understand the genetic architecture of GxE. However, drought is both difficult to control and hard to predict, because of interactions with almost all other factors in the environment (temperature, air flow, light) and interplay with other constraints (especially nutrient-related or osmotic). The development of robotic phenotyping tools throughout the community makes it now feasible to acquire traits on hundreds or thousands of plants in precisely controlled and reproducible conditions [32–35], pushing a bit further one of the main limitation for a better decomposition of the genetic architecture of these complex traits.
Still, regarding plant growth variation in nature, mainly genomic regions with relatively large effect were identified in Arabidopsis and were often related to development, immunity or major hormones (for instance [36–43]). A limited number of non-theoretical studies seem to confirm that many genes with smaller effect–potentially involving epistasis and linked loci–would be responsible for part of the phenotypic variation of such complex traits (for instance [44–46]).
Here, we undertook a precise analysis of plant growth genetic architecture under both optimal watering condition and mild drought stress (as revealed under artificial conditions of reduced and constant soil water content), using a classical linkage mapping approach on 4 biparental segregating populations [47]. As a dynamic trait, we chose to follow growth during the vegetative phase using a high-throughput phenotyping robot (the Phenoscope [34]; https://phenoscope.versailles.inra.fr/) to map major- to small- effect QTLs as well as their interaction with drought stress. Zooming in on the loci, we use near-isogenic lines to validate these QTLs and reveal in more detail the genetics behind a single QTL peak. We then focus more precisely on a region where a major Quantitative Trait Gene (QTG) is segregating (= CRY2, a known polymorphic actor with major pleiotropic phenotypic consequences), and show that other loci with additive or opposite effect are also present in its vicinity, illustrating the complexity of growth genetic architecture.
The four RIL sets used in this study (BurxCol, CvixCol, BlaxCol, YoxCol) were conducted under well-watered (WW) and moderate water deficit (WD) conditions on our high-throughput phenotyping platform. Ensuring that growth occurs in a highly controlled and homogeneous environment, the Phenoscope records a number of image-based quantitative traits describing shoot development (Fig 1). Taking daily pictures gave access to cumulative (Projected Rosette Area, PRA) and dynamics (Relative Expansion Rate, RER) growth parameters for individual plants (Fig 1C & 1D) as well as other descriptive or derived traits (rosette morphology and RGB colour components) [48]. A principal component analysis (PCA) was performed using all picture-based phenotypes at the final day of the experiment, 29 Days After Sowing (DAS; hence 'PRA29' etc) and relative expansion rate calculated between 16 and 29 DAS (RER16-29; Fig 2). The first axis explained a major part of the total variance, essentially through final rosette size (PRA29) and expansion rate (RER16-29). However, PRA29 and RER16-29 variables were not perfectly correlated, with genotypes exhibiting moderate PRA29 despite high RER16-29. The red (Red29) and green (Green29) components colour phenotypes mainly contributed to the second axis and were positively correlated, and both were negatively correlated with rosette compactness at 29DAS (Compactness29) which was the main trait contributing to the third axis. Individual projection showed that the first axis strongly structured the individuals according to the watering treatment (WW versus WD) while the axes 2 and 3 represented cross (RIL set) effects, differentiating CvixCol and BurxCol (axis 2) and BlaxCol (axis 3) from the other RIL sets. PRA29, RER16-29 and Compactness29 are complementary growth phenotypes that were investigated further in this study to quantify different aspects of shoot development variation: final projected rosette area is a cumulative proxy for biomass and photosynthetically-active surface, rosette compactness is an informative parameter describing the rosette morphology, and relative expansion rate highlights the dynamics of growth.
Phenotypic distribution among the RILs compared to their parents (S1 Fig) revealed extensive transgressive segregation for most of the traits and the crosses studied. As expected, mild drought stress (WD) condition impacted the distribution of the RILs for PRA29 and RER16-29 with generally reduced values for both traits. Interestingly, compactness distribution was much more robust to stress, which indicated that overall the morphology of the rosette is less affected by mild drought. In order to estimate the part of phenotypic variation that is explained by genetic factors for each traits, heritabilities were calculated (S1 Table) and were essentially below 0.2 for RER (except in BurxCol where they were higher), essentially around 0.5 for cumulative PRA and generally above 0.5 for Compactness (this trait is certainly less sensitive to shifts in developmental stage that could be induced, for instance, by small changes in germination time). Overall, heritabilities were also lower in stress conditions than in control, as if the stress was inducing noisier phenotypic variations.
According to ANOVA analyses (S2 Table), all traits and RIL sets showed significant variation according to both the watering condition (WD versus WW) and the genotype (within each RIL set), except for Compactness29 response to stress in BlaxCol. There were also weaker (compared to the genotype and watering condition effects) but still significant biological replicates’ effects (i.e. independent Phenoscope experiments), but less genotype x experiment interactions (with the exception of PRA and RER traits in BlaxCol for instance). PRA is more prone to genotype x experiment interactions than other traits, especially in CvixCol and YoxCol. Genotype x condition interactions are often milder than genotype or condition effects, and overall compactness–or YoxCol–show much less genotype x condition interactions than other traits/sets.
The phenotypic values of each RIL were then corrected for inter-experiment differences (indicated by the significant biological replicates’ effect).
Our experimental design allowed the identification of many QTLs for all combinations of traits, conditions and RIL sets, and also for the GxE interaction term using genotype x condition effects from the ANOVA model for each trait (Fig 3 and S3 Table). Globally 112 QTLs were identified all along the genome when conditions are studied independently (62 under WW + 50 under WD) likely corresponding to at least 18 independent loci (genetically distant-enough from each other to be most likely considered as independent), yielding a median of 4 QTLs per modality of cross x trait x condition (ranging from 1 to 8 QTLs). QTL hotspots across RIL sets and traits were identified for instance at the beginning of chromosome 1, at the bottom of chromosome 2 and 5. These hotspots include very highly significant QTLs with LOD scores above 10, and up to 32. Chromosome 3 appeared to show less significant QTLs in all crosses, especially for PRA29 and RER16-29. Individual QTL contributions to phenotypic variance (R2) ranged from 1 up to 30%, and showed a L-shaped distribution of effect (S2 Fig). Using empirical significance boundaries according to the observed distribution of QTLs effects, ~10% of the QTLs could be considered as showing major effects and significance (R2>10% and/or LOD>15); ~25% of the QTLs could be considered as showing intermediate effects and significance (5%<R2<10% and/or 7<LOD<15); the remaining 2/3rd of the QTLs could be considered as showing minor effects and significance R2<5% and/or LOD<7). Many more potential QTLs not listed here were only suggestive with LOD score peaks just below our permutation-based threshold (<2.4 LOD).
Most of the QTL profiles are stable across the 2 watering conditions, especially for the major-effect loci. However, QTLs specific for one condition were detected, e.g for RER16-29 under WD in YoxCol on chromosome 4, and at the top of chromosome 1 under WW. We also mapped QTLs for the interaction term with the drought treatment, yielding 19 QTLs (Fig 3 and S3 Table). These essentially emphasize large effect QTLs showing a modulation of their effect in response to stress (especially for PRA, see chromosome 1 and 5), with no new loci revealed. For RER16-29 in YoxCol, this would confirm the interaction of the above-mentioned locus on chromosome 4 (although its exact position is questionable), but not for the one at the top of chromosome 1. There may be some power issues when comparing across conditions due to lower heritabilities of traits under WD.
Although derived from PRA29, Compactness29 showed an independent genetic basis, as exemplified by the major peaks on the bottom of chromosome 4 in BurxCol or 5 in BlaxCol. Although contributing to RER16-29, PRA29 does not always share the same contributing loci, for instance on the top of chromosome 1 in CvixCol and BurxCol. Other more complex cross x trait patterns are apparent, like at the bottom of chromosome 2 where a major QTL for Compactness29 in three of the four RIL sets seems to colocalize with a significant PRA29 and RER16-29 QTL (in the same direction), but only in one cross. It may be that these Compactness29 QTLs are actually independent in each RIL set.
A two-dimensions search for epistatic interactions was performed across all traits, conditions and RIL sets (S3 Fig). Overall, the BurxCol and CvixCol RIL sets showed more significant epistasis compared to the two other sets. Interestingly, pairwise interactions controlling growth phenotypes are overall quite different depending on RIL set and growth phenotypes. Shared epistasis effects are potentially detected in both watering conditions: they appear as symmetrical across the diagonal on S3 Fig. One of the most significant epistatic interaction was observed between 2 loci on the top and bottom of chromosome 4 in the BurxCol cross (S3 Fig; interaction component highly significant for RER16-29 in WW: 5.1 LOD). Positions and directions of effect match perfectly with the previously published SG3 x SG3i interaction known to segregate in this cross [49]. Based on its significance, another relevant interaction (3.34 LOD in WW and 3.12 LOD in WD) was observed for PRA29 in the BlaxCol cross between the bottom of chromosomes 4 and 5 (Fig 4). The effect of the bottom of chromosome 5 QTL on PRA29 is observed only when RILs carry Bla alleles at the bottom of chromosome 4. As a consequence of this epistasis, these QTLs appear barely significant in single QTL scans (Fig 3).
We also performed dynamic QTL detection on daily-recorded traits (PRA and Compactness) to reveal the evolution of QTL effect throughout the experiment on the Phenoscope: these interactive QTL profiles can be accessed at http://www.inra.fr/vast/PhenoDynamic.htm
Most of the PRA QTLs observed after 29 days of growth correspond to locus that become gradually significant across the experiment and are essentially not time-specific. These are most likely contrasted allelic effect on growth that cumulate their effect over time. There are only a few exceptions for PRA, like the bottom of chromosome 5 locus segregating in BlaxCol which remains significant only until 13DAS and thereafter is canceled out. For Compactness, the picture is rather different with numerous examples of QTLs that are essentially significant around specific time-points, even sometimes in successive waves of significance (BlaxCol, bottom of chromosome 5: QTL peaking at Days 11, 17 and 27 –providing that this is a unique locus).
In order to take advantage of the different crosses to Col-0 in our experimental setup, QTL mapping was performed on the whole dataset using MCQTL (Multi-Cross QTL) tool to compare allelic effects in a multicross design and potentially reveal shared QTLs. The combined QTL maps obtained (Fig 5 for PRA29, S4 Fig for RER16-29, S5 Fig for Compactness29) highlight 10 independent loci, including at least 4 regions with contrasting allelic effect on PRA29: for instance, the middle of chromosome 1 region shows contrasted phenotypic consequences in different crosses, particularly when comparing Bur and Yo alleles (with respect to Col). At this scale, it can be difficult to distinguish between different alleles at the same QTL and different QTLs. Conversely, combining the information of multiple crosses sometimes allows to predict narrower QTL intervals than with the initial QTL mapping, enabling the detection of distinct linked loci, for instance for PRA on the bottom half of chromosome 2; a location where the dynamic QTL analysis for PRA in CvixCol was already showing signs of 2 different segregating loci with slightly distinct dynamics over time. Another striking example is for Compactness on chromosome 5, with neighbouring QTLs predicted to show opposite allelic effects (e.g. BlaxCol).
To confirm and investigate further the complexity of the genetic architecture of these traits, we used near-isogenic lines to mendelize QTLs and assess in more details the role of smaller chromosomal regions. Using 81 independent Heterogeneous Inbred Families (HIFs) scattered across the genome or chosen to decompose candidate regions in specific crosses [50], we tested a total of 79 QTL effects from the 24 modalities of RIL set x trait x condition (Fig 6 for PRA29; S6 Fig for RER16-29 and Compactness29). Globally, 60% of the HIFs with a segregating region covering a candidate region previously identified showed significant effect with consistent direction, thus validating the QTL. Larger effect QTL were more often validated in HIF, with ~75% of the tested major or intermediate effect QTLs confirmed, compared to ~50% of the minor effect QTL. Specifically, for PRA29 trait in the two conditions studied (Fig 6), 28 QTLs were assessed with at least one HIF, among which 17 (60%) were confirmed. Some (minor-effect) QTL x condition interactions detected in the RIL set were also significantly confirmed using HIF, such as the top of chromosome 5 locus in BurxCol which has significance (PRA29) only under WW.
We are also interested in positive HIF results that do not match with the results of the initial QTL mapping. This was particularly possible in regions with high HIF coverage and shows that several LOD score peaks were explained by more than one underlying QTL, as also suggested by the MCQTL analysis. A good example lies in the CvixCol cross, where the bottom half of chromosome 4 seemed to control PRA29 as a single locus in the initial QTL analysis (Fig 3) and is now subdivided in at least 2 independent loci with opposite allelic effects after the HIF analysis (Fig 6). The 2 adjacent HIFs do not only show opposite allelic effect, but also different allelic effect amplitude on PRA29 (-1.4cm2 versus +0.9cm2), potentially explaining why only a Col-negative allelic effect was detected in the initial QTL mapping in this region, since the effect in this direction is stronger. Another example segregates in YoxCol at the top of chromosome 3 where 2 QTLs with opposite-effect on Compactness29 under WW conditions seem to localize closely according to the HIFs (S6 Fig) but wasn't detected at all in the QTL mapping at 29DAS (Fig 3) and only remained significant at intermediate stages around 16DAS according to the dynamic analysis. Here, the 2 adjacent HIFs show very similar allelic effect amplitude on Compactness29 (+0.0301 versus -0.0295), likely explaining why no QTL was detected in this region for this trait. These could be examples of a lack of power of the RIL design to detect complex patterns due to the confusing effect of linked loci; alternatively, the confusing effect of epistasis could also interfere when comparing QTL mapping results from RILs and HIFs, because of the specific genetic background of each HIF.
Finally, we can occasionally exploit the localization of the segregating region(s) of the tested HIFs to narrow down the candidate QTL region, like for a PRA29 QTL in CvixCol at the bottom of chromosome 2 which is narrowed down to the extremity of the chromosome and shown to be distinct (= confirmed in non-overlapping HIFs) from another nearby QTL segregating in the same cross (with allelic effect in the same direction), as predicted by the multicross analysis above. Here again the dynamic analysis also helped to distinguish these loci based on their effect through time. Still, the 'precision' remains approximately at the Mb level at this stage.
To tackle further the question of the complexity of the genetic architecture at a higher resolution than with simple HIFs, we decided to systematically dissect a region of 3Mb at the very beginning of chromosome 1 in CvixCol: in this region all previous approaches have predicted a single QTL with intermediate to major effect significance on PRA29 (Fig 3), which was confirmed in an HIF (Fig 6). This HIF was used to zoom in on the region across 30 bins (stairs) of ~100kb, defined by successive recombination breakpoints (S4 Table) and phenotypically evaluated individually in the same conditions as above (an approach coined 'microStairs'). We selected this interval to test if a region harbouring a large-effect growth QTL may typically also include independent loci, maybe of smaller effect or hidden by the main QTL. A pairwise comparison of their growth phenotypes allows to test the impact of Cvi versus Col alleles over relatively short physical intervals of expected average size 100kb, with a maximum ~200kb, depending on recombination breakpoints location and marker interval. We either compared only the 2 successive recombinants ('stair by stair'), or we took advantage of the support from all pairwise comparisons ('staircase') to increase our power to detect QTL (Fig 7 for PRA29 and S7 Fig for RER16-29 and Compactness29).
The strongest phenotypic effect for PRA29 in WW and WD (but also–to a lesser extent–for RER16-29 in WD) was fine-mapped down to the stair between recombinants 111 and 112, covering potentially a physical interval (bin) extending to the maximum between 1.070 and 1.291 Mb (an interval including 54 genes). In the middle of this region lies the obvious candidate gene CRYPTOCHROME 2 (AT1G04400; CRY2, a blue-light photoreceptor; S5 Table) that is known to harbour a functional variant in Cvi (a single amino-acid change) and impact plant photomorphogenic development especially in short days [51, 52]. This is a good test case of our approach, as it is very likely that CRY2 is primarily responsible for the growth difference overall observed in this HIF.
Still, on either side of this locus, a few other QTLs also affect growth, underlying a much more complex genetic architecture than expected after the initial HIF results: in both watering conditions, a milder PRA29 QTL, with opposite allelic effect than CRY2, was detected in the stair defined by recombinants 115 and 116 (1.495 to 1.703 Mb positions). Another PRA29 QTL was predicted in the bin 125/126 (2.509 to 2.710 Mb). For Compactness29 specifically under WD treatment, CRY2 did not seem to be causal and the causative locus for the observed HIF phenotype would most likely be in the bin 118/119 (1.801 to 2.027 Mb); this QTL was not clear from the initial QTL analysis for this trait on Day 29 (S6 Fig), but appeared significant earlier in the experiment (cf dynamic analyses for CvixCol around 14DAS and later). There are probably even more loci, especially at the very beginning of the region for PRA29 (stairs between recombinant 101 and 104 at least), but we reach the limits of our experimental design and phenotyping precision to be able to conclude accurately, with either too complex genetic architecture in this region or not enough recombinant lines to robustly support each intervals' effect.
We then looked for high impact polymorphisms (premature stop codon, frameshift or non-synonymous mutations) likely affecting gene function between Col-0 and Cvi-0 within the most promising bins to identify candidate genes. Because of the numerous non-synonymous changes between these accessions (S5 Table), we decided to arbitrarily filter the genes with the criteria of at least 3 non-synonymous mutations to increase our chance to detect genes with altered function or degenerated sequences after loss-of-function mutations. Here, we discuss some of those polymorphic genes where previous publications have indicated a putative function or effect that could relate to our phenotypes.
Within the bin 115/116 lie at least 2 interesting candidates:
URIDINE DIPHOSPHATE GLYCOSYLTRANSFERASE 74E2 (AT1G05680; UGT74E2) is an auxin glycosyltransferase whose overexpression was shown to modify plant morphology and the size of the petioles, to delay flowering, and to increase drought tolerance [53]. The gene is also known for ample natural variation in expression, including potentially cis-acting variants [54]: http://www.bioinformatics.nl/AraQTL/multiplot/?query=AT1G05680.
GLUTAMATE RECEPTOR 3.4 (AT1G05200; GLR3.4) is a calcium-dependent abiotic stimuli-responsive gene [55] expressed throughout the plant and impacting at least lateral root initiation [56]. It harbours several non-synonymous variants in Cvi-0 (compared to Col-0), but also higher expression in Col-0 as shown by several local-eQTLs in different crosses (http://www.bioinformatics.nl/AraQTL/multiplot/?query=AT1G05200) confirmed to be cis-regulated through Allele-Specific Expression assay [57] (reported in S6 Table).
Within the bin 125/126 lie at least 2 interesting candidates:
AT1G08130 encodes DNA LIGASE 1 (LIG1), a ligase involved in DNA repair, which mutation causes severe growth defects [58]. Its expression is also known to be controlled by a local-eQTL (most likely cis-acting) in LerxCvi:
http://www.bioinformatics.nl/AraQTL/multiplot/?query=AT1G08130
AT1G08410 is DROUGHT INHIBITED GROWTH OF LATERAL ROOTS 6 (DIG6), encoding a large 60S subunit nuclear export GTPase 1 that impacts several developmental processes regulated by auxin, including growth [59].
Further work is required to prove any link between the observed phenotypic variation and these candidate genes.
Owing to its fine regulation throughout development and in interaction with the environment, plant growth represents a highly complex trait potentially controlled by numerous factors and interactions. Little is known on the actual genetic architecture of plant growth natural variation, with essentially a few genes of major effect being identified until now and only a few exceptions of more complex genetics revealed [44]. Association genetics hold great promise to dissect the underlying molecular bases of complex traits [12], however one can wonder if the genetic architecture of highly integrative traits like growth or fitness is amenable to genome-wide association studies (GWAS) at the species scale: GWAS especially lacks power to decompose traits controlled by many loci of small effects when the underlying alleles have low frequency in the mapping population. For instance when exploiting worldwide collections of accessions, very little (if any) significant associations for growth parameters were found [60], even when using very anti-conservative thresholds [2, 61] or using morphological parameters with higher heritabilities [62]. Even with more targeted growth traits like root cell length, Meijon et al. did not detect any significant signal above the threshold, although the first peak just below the threshold identified a causal gene [63]. Overall, it is argued that linkage and GWA studies are complementary in the loci that they are able to reveal, depending on the genetic architecture of the trait in the population considered (e.g. [21, 64]). Here, by studying four different crosses to the reference Col-0, we find many cross-specific loci, especially of mild effects, several of which might correspond to low frequency-alleles that would not likely be pictured in GWAS (not enough power due to low frequency x effect size).
Whether linkage or association mapping, these approaches are both similarly phenotyping-intensive and prone to interaction with uncontrolled environmental parameters (increasingly so with the scale of the experiments). Using our high-throughput phenotyping robots to grow individual plants under tightly controlled conditions, we intended to dissect the genetic basis of plant growth under optimal and limiting watering conditions, to a level of accuracy rarely reached so far. We focused on vegetative growth from days 8 to 29 (after sowing) and selected three non-fully-correlated variables allowing to characterize plant growth dynamically: rosette area (PRA), relative growth rate (RER) and compactness. PRA is a typical cumulative trait: what is observed on day n is not independent of what has occurred from day 1 to n-1. Compactness, which basically represents a measure of PRA normalized by rosette width, is rather independent of cumulative phenomenons and hence shows much more age-specific QTLs; this is particularly striking when comparing the dynamic analyses for these traits. Finally, RER integrates growth rates over a specific period of time, and is independent of plant size (= relative). Because estimated during the exponential phase of growth, RER is very much similar to what is obtained by fitting exponential models to the PRA data and exploiting model parameters [2].
Hence, growth-related phenotypes, depending on how they are exploited, will present different genetic architectures throughout age, with individual loci making different contributions to cumulative or age-specific traits. It has previously been shown that heritabilities for growth-related traits change over time [6, 65]; one possible explanation is that QTLs are more or less likely to act at specific time points. For instance, studying growth dynamics in maize [66, 67] and root tip growth in Arabidopsis [68] allowed to identify marker-trait associations that would not be detectable by considering the cumulative trait only at a single (final) time point. At the other end of time-resolution for growth, going into much more details of the dynamics (several images per day) may result in noisy raw data requiring further treatment before exploiting, for example due to projected growth estimates interfering with circadian leaf movements [65].
Our work has been performed under two environmental conditions, a control condition and one that moderately limits growth due to water (but not nutrient) availability [34], i.e. a mild drought treatment. The QTL profiles obtained at the genomic scale are very robust to mild drought with most of the large effect QTL showing no clear signs of interaction with water availability in our conditions. Some of them still change their level of significance with conditions, but it is difficult to know if this is a real interaction with drought, a change in trait variance under stress, or a change in the rest of the genetic architecture of the trait (which will impact the significance of individual QTL). Condition-specific QTLs detected here always are of small effect, which also raises the question of the power of these comparisons across different conditions/experiments due to false negative in mapping QTL. Still, this result (a relatively smaller part of the phenotypic variation is plastic rather than constitutive) is similar to what was found previously for instance in linkage [69] and association mapping [61], showing an overlap in the network of genes that regulate plant size under control and mild drought conditions [60]. Drought stress might have pleiotropic effects on different tightly interrelated phenotypic traits and impose strong constraints on them, reducing trait variability [31]. Of course, this may highly depend on the type of stress (intensity, stage of application, stability …) that is applied. Here, we chose a mild stress intensity, to remain physiologically relevant and avoid the squeezing of trait variation concomitant to strong stress levels. Also, our robot is compensating (twice a day) the individual plant size effect on transpiration that may otherwise artificially increase stress intensity according to intrinsic plant size difference.
One advantage of using a star-like cross design, increasingly used in nested association mapping (NAM) populations, is to combine the power of individual crosses and take advantage of the comparison of multiple alleles, with respect to the reference allele in order to identify allelic variants more efficiently [70–73]. These could correspond to variants originating from Col-0 or to shared allelic variants among the other parents (same direction of the allelic effect in all crosses), or to allelic series (other parents could have divergent direction/intensity of effect with respect to the reference parent). This multi-population study has confirmed the effect of several loci across traits and environments, with particular power for compactness, and in some instances already allows to predict that independent loci actually underlie major peaks. However, this approach is still limited by the mapping resolution which makes it difficult to distinguish shared variants from linked loci.
Considering the precision of our phenotyping (which has an impact on the part of phenotypic variation that is amenable to genetic dissection), the need for higher-resolution approaches to better describe the trait's genetic architecture is obvious here. If the architecture of variation in our crosses is more complex than just a limited number of loci independently segregating for intermediate or major effects, then the density of recombination observed in a simple RIL set will not allow to decipher the full architecture [74, 75]. We first went deeper in resolution by phenotyping numerous pairs of near-isogenic lines (actually, HIFs) that each interrogate a specific portion of a chromosome (2-3Mb on average) in an otherwise fixed genetic background. This nicely confirms a majority of loci but also shows some effects that were unexpected after the initial QTL mapping results, already indicating more complex genetic architecture than anticipated; this includes single peaks splitting up in independent loci or complex patterns of linkage versus pleiotropy (when comparing different traits) and linkage versus GxE (when comparing traits in different conditions). It seems that QTL colocalization among crosses (linkage versus shared variants) is also often questioned, although this requires to be able to compare multiple HIFs showing positive and/or negative results, which can be difficult for several reasons. Indeed, many factors can explain that a QTL is not validated in an HIF: the QTL could be under epistatic interaction with another locus (thus, a specific HIF may not represent the adequate genetic background), the QTL could be mislocalised by QTL mapping (thus, out of the HIF segregating region) or the HIF harbours a more complex genotype than expected at the segregating region, such as a double recombined region (thus, the HIF doesn't actually allow to test the whole region). This makes the comparison between HIFs difficult, even in the same cross, and negative HIF results should particularly be interpreted carefully. In this context, the rate of QTL validation obtained here is rather satisfying. The difficulties to identify genes responsible for complex phenotypes also depends on their involvement in epistatic interactions. HIFs are particularly sensitive to epistasis (compared to traditional NILs) as they each harbour a different genetic background, which means that they allow epistasis to be interrogated providing that one can test enough independent HIFs, otherwise they have to be compared with care.
Epistatic interactions are detected in all crosses / traits (although not always with very high significance) even when the trait heritability or variation is not so high in a cross, illustrating another factor of the complex genetic architecture. Some interactions seem to be condition-specific, but power issues are likely to be limiting in understanding these patterns of GxGxE. Furthermore, HIF have an expected candidate segregating region of several Mb usually, so our observations are likely just a glimpse at the real complexity of growth as it is known that linkage and epistasis is also active at a very local scale [44].
Still, the sensitivity of our approach here is validated by the detection of QTL colocalizing with several already-known QTG expected to segregate in our crosses, like CRY2 as discussed above [52, 76], MPK12 which would explain nicely the bottom of chromosome 2 QTL in CvixCol [77, 78], or SG3 detected here through its epistasis with SG3i in BurxCol [49, 79].
To avoid genome-wide epistasis and better reveal local-scale architecture, we have investigated in further details a single HIF background for a specific 3Mb region containing a known QTG of large and pleiotropic effect (CRY2) in an original approach. Our analysis reveals that there are at least 3 other QTGs in this interval controlling one of the traits in at least one condition. For PRA29, the picture seems to be even more complex with traces of at least one more locus; here, it seems that phenotyping accuracy becomes again limiting after all. Obviously these loci with opposite allelic effects, different patterns of pleiotropy and interaction with the environment, and just a few hundreds of kb from each other, remain cryptic in simple QTL mapping. This major result of local-scale independent complex genetic architecture for different traits and conditions should lead us to a lot of caution when interpreting colocalizing QTLs from different traits / conditions / age, as these may very well be independent loci rather than a single pleiotropic locus, as shown here for PRA and Compactness. If we were to extrapolate the figures obtained on this specific region in this particular cross to the genome- and species-scale, we would expect hundreds of causative loci of detectable phenotypic effect controlling these growth-related phenotypes. One way to approach these individual loci would be to decompose their independent signature based on different dynamics or underlying traits (transcriptomics, metabolomics …) in a 'systems genetics' strategy [54, 60, 80].
Complex genetic architecture as revealed in this study has consequences on quantitative genetics experimental design and interpretation, arguing in favor of linkage mapping or GWAS depending on the balance between genetic complexity due to linked loci (where association is expected to behave better than linkage mapping) and genetic complexity due to small effect/rare alleles (where association will behave poorly). Other intermediate experimental designs like multiparental populations or nested-association mapping should bring more power [12, 81]. Resolution is improved by pushing recombination densities to its limits and it was shown to help resolve more complex genetic architecture in yeast [82]. In plants, using 'hyper-recombinant' mutations to generate new segregating populations could also be a strategy in the future [83].
The 4 RIL sets used for this work were generated at the Versailles Arabidopsis Stock Center, France (http://publiclines.versailles.inra.fr/) and were either described previously [47] or on the Publiclines website where all relevant information (description and associated data) is gathered. Versailles stock center ID are indicated as 'xxxAV' and 'xxRV' for Accessions and RILs respectively). They are derived from crosses between the following pairs of accessions, chosen to maximize genetic and phenotypic diversity [84]:
RIL set 'BlaxCol' (ID = 2RV): Bla-1 (76AV) x Col-0 (186AV) / 259 RILs
RIL set 'CvixCol' (ID = 8RV): Cvi-0 (166AV) x Col-0 / 358 RILs
RIL set 'BurxCol' (ID = 20RV): Bur-0 (172AV) x Col-0 / 283 RILs
RIL set 'YoxCol' (ID = 23RV): Yo-0 (250AV) x Col-0 / 358 RILs
HIFs have been selected in each RIL set to cover several regions of the genome, some of which are expected to segregate for QTLs while others were chosen at random locations. As described previously [85] they are derived from the progeny of one RIL which is heterozygous only at the locus of interest. Hence, one HIF family is composed of 2 or 3 lines fixed for one parental allele at the segregating region and 2 or 3 lines fixed for the alternate parental allele at the segregating region, in an otherwise identical genetic background. Each HIF (ID = 'xxHVyyy') is named after the RIL ID ('xxRVyyy') from which it has been generated ('xx' is the ID of the RIL set, 'yyy' is the ID of the RIL): for example, the family 2HV142 is fixed from the RIL 2RV142. The tentative QTL validation is based on the phenotypic comparison of these fixed lines within the HIF family. We have used 81 HIF families. The complete dataset gathering genotypic information on the RILs used to generate HIFs has been submitted to INRA institutional data repository (https://data.inra.fr) [50], with the genotypic conventions and ID from the Versailles Arabidopsis Stock Center (http://publiclines.versailles.inra.fr/).
The progeny of CvixCol RIL 8RV294 (also used to generate HIF family 8HV294), segregating for the first 3Mb of chromosome 1, has been screened to detect recombined individuals. 4000 individuals were genotyped with markers at the edge of the heterozygous region and then 29 evenly distributed recombinants were selected using markers spaced every ~100kb. These recombinants were genotyped and fixed for the remaining segregating region in such a way that they each differ genotypically from the next recombinant by a ~100kb bin on average (S4 Table). Similarly, as for the HIF, the descendance of 3 similarly-fixed lines are saved and phenotyped to account for possible maternal environmental effect. This is similar to the approach taken by Koumproglou et al. [86], but at a much finer scale, hence the name 'microStairs'.
Phenotyping was performed on the Phenoscope robots as previously described [34] (https://phenoscope.versailles.inra.fr/). Every RIL set and their respective parental accessions have been phenotyped in 2 independent Phenoscope experiments (= biological replicates), except for CvixCol (3 biological replicates), with 1 individual (plant) per RIL per condition. In short, the peatmoss plugs' soil water content (SWC) is gradually adjusted for each plant individually as a fraction of the initially-saturated plug weight. We worked at 2 watering conditions: 60% SWC for non-limiting conditions (called 'WW' for well-watered) and 30% SWC for mildly growth-limiting watering conditions (called 'WD' for water deficit). The growth room is set at a 8 hours short-days photoperiod (230 μmol m-2 sec-1) with days at 21°C/65%RH and nights at 17°C/65%RH. A picture of each individual plant is taken every day at the same day-time and a semi-automatic segmentation process (with some manual corrections when required) is performed to extract leaf pixels. From this we exploit different traits: Projected Rosette Area (PRA), circle radius, convex hull area, average Red, Green and Blue components (leaf pixels, RGB colour scale), and derived phenotypic traits are calculated, such as the compactness (ratio PRA / convex hull area) and the Relative Expansion Rate (RER) over specific time windows (Fig 1), as previously described [34]. The complete raw phenotypic dataset has been submitted to INRA institutional data repository (https://data.inra.fr) [48]. Principal component analysis (PCA) was performed using the ade4 R package based on phenotypic data from all the RIL sets in WW and WD conditions.
Heritabilities were calculated based on the mean squares (MS) of the following ANOVA model, used as estimators of genetic and residual variances:
Yij∼μ+αi+εij
where Yij is the phenotypic value, μ is the mean, αi is the Genotype factor and εij is the Residuals. The genetic variance VarG was estimated by (MSα - MSε)/n (where n is the number of replicates). The residual variance VarR was estimated by MSε. h2 = VarG/(VarG+VarR).
Then, for QTL detection and further analyses, the phenotypic values were corrected for experiment effects. Corrected phenotypic values were calculated using the intercept (μ), the condition (α), genotype (γ), and genotype*condition (σ) effects of the following linear model:
Yijkl∼μ+αi+βj+γk+δij+λjk+σik+εijkl
where Yijkl: phenotype; μ: mean; αi: effect of the condition; βj: effect of the experiment; γk: effect of the genotype; δij: effect of the interaction condition*experiment; λjk: effect of the interaction experiment*genotype; σik: effect of the interaction condition*genotype; εijkl: residuals. The interaction term (independently used for QTL mapping) refers to the condition*genotype interaction component.
Note that one of the replicates of the CvixCol phenotypic data was already analyzed for QTL mapping -with a different statistical model- in Tisné et al. [34].
Similarly, the set of near isogenic lines (microStairs) were phenotyped and analyzed from 3 full biological replicates in independent Phenoscope experiments.
QTL detections were performed using Multiple QTL Mapping algorithm (MQM) implemented in the R/qtl package [87, 88] using a backward selection of cofactors. At first, genotype missing data were augmented, then one marker every three markers were selected and used as cofactors. Important markers were selected through backward elimination. Finally, a QTL was moved along the genome using these pre-selected markers as cofactors, except for the markers in the 25.0 cM window around the region of interest. QTL were identified based on the most informative model through maximum likelihood. According to permutation results (computed with the mqmpermutation() function in R/qtl), a general LOD threshold of 2.4 was chosen for all QTL maps to ensure a FDR below 0.05 and remain conservative. Interactive QTL maps for time-course series were generated using the R/qtl charts package [89]. All QTL positions were projected on the consensus genetic map of the 4 crosses built with R/qtl (this is the map shown along the x axis on Figs 3, 5 and 6). A joint genotype dataset was constructed with ‘A’ alleles coding Col alleles (the common parent), ‘B’ alleles for non-Col alleles, and monomorphic markers in a cross coded as missing. The linkage groups were considered known from the individual maps and the physical position of markers, and a first marker order was calculated using orderMarkers() function. In case of conflicting marker order between individual, physical and consensus maps, the function switch.order() was used to retain the most probable order (i.e with the lowest number of recombination). The consensus genetic map markers' identities and positions are available together with the HIF dataset in Dataverse [50].
Epistatic interactions were identified using the scantwo() function of the R/qtl package. LOD scores were calculated for additive, interaction and full models for all pairwise combination of markers, except for adjacent markers and a general conservative LOD threshold (3 LOD) was determined from permutations. Effect plots for the pairs of markers were drawn using the R/qtl package.
QTL mapping in the multi-cross design was performed with the MCQTL package [90]. The model was described as additive (no dominance effect) and connected (Col-0 centered design) and the following 3 steps process was applied. Step 1: thresholds were calculated by trait on the whole genome using 1000 resampling replicates (PRA29_WW = 3.73; PRA29_WD = 3.84; RER16-29_WW = 4.22; RER16-29_WD = 3.95; Compactness29_WW = 3.64 Compactness29_WD = 3.30). Step 2: QTL detection was performed using iQTLm method with a general and conservative threshold of 4 LOD. To perform this detection, cofactors were automatically chosen by backward selection with a threshold of 2.8 LOD among a skeleton with a minimal inter distance of 10cM. Search for QTL was not allowed within +/-10cM window surrounding the QTL to avoid linked genetic regions. Step 3: model estimations were performed for each trait and condition using the shared QTL positions identified at step 2.
The phenotypes of the recombined HIFs lines were modeled using the following linear equation:
Yij∼μ+αi+βj+εij
where Yij is the value of the phenotype; μ is the mean of the phenotype; αi is the effect of the stair (bin) i; βj is the effect of the line j (maternal replicate within each stair) and εij is the residuals. An anova was performed with this model and the p-value of the stair effects were adjusted by a Benjamini-Hochberg correction.
Polymorphic candidate genes (Cvi versus Col) were listed for each PRA29 'microStairs' significant interval according to variants listed on the 1001Genomes website (http://1001genomes.org/), through the Polymorph1001 tool. Differentially cis-regulated variants were extracted from Cvi/Col Allele-Specific Expression (ASE) data [57] and from CvixCol local-eQTLs data [91] across the whole region.
Two datasets have been submitted to INRA Dataverse repository (https://data.inra.fr).
This dataset gathers the main raw phenotypic data obtained and exploited in Marchadier, Hanemian, Tisné et al. (2018). It contains data from 4 RIL sets across 9 Phenoscope experiments.
For each Phenoscope experiment, Recombinant Inbred Line (RIL) and Condition ('WW' = Well Watered / 'WD' = Water Deficit), the data set indicates the phenotypic value for 6 traits at 21 successive time points. 'Trait.XX' = Trait at XX days after sowing, with 'XX' = 09 to 29 and 'Trait' = PRA (Projected Rosette Area; in cm2), GreenMean / RedMean / BlueMean (rosette pixels' colour components; arbitrary unit), ConvexHullArea (area of the convex hull encompassing the rosette; in cm2) and CircleRadius (radius of the smallest circle encompassing the rosette; in cm). RIL set IDs and RIL IDs are according to Publiclines http://publiclines.versailles.inra.fr/rils/index
Each row represents a single HIF and the genotype of the F7 RIL it originates from is indicated along the chromosomes with RIL ID, markers and genotypic conventions from Publiclines http://publiclines.versailles.inra.fr/rils/index (i.e. 'A' = Col allele; 'B' = alternate parental allele; 'C' = heterozygous). The region highlighted in yellow is the segregating region that is tested in the HIF family through several fixed lines for each parental allele. For each of the 3 growth traits (Compactness29 = rosette compactness 29 days after sowing; PRA29 = Projected Rosette Area 29 days after sowing; RER16-29 = Relative rosette Expansion Rate between days 16 and 29 after sowing) in 2 conditions ('WW' = Well Watered / 'WD' = Water Deficit), whenever significant, the p-value of the comparison between allelic lines ('Pval') and the direction of the allelic effect ('sign' calculated as [Col-Xxx] where Xxx is the alternate parental allele) are indicated in the last columns of the table.
A specific webpage is associated with this work to display interactive graphes for dynamic QTL analyses at http://www.inra.fr/vast/PhenoDynamic.htm
|
10.1371/journal.pgen.1003303 | JNK-Interacting Protein 3 Mediates the Retrograde Transport of Activated c-Jun N-Terminal Kinase and Lysosomes | Retrograde axonal transport requires an intricate interaction between the dynein motor and its cargo. What mediates this interaction is largely unknown. Using forward genetics and a novel in vivo imaging approach, we identified JNK-interacting protein 3 (Jip3) as a direct mediator of dynein-based retrograde transport of activated (phosphorylated) c-Jun N-terminal Kinase (JNK) and lysosomes. Zebrafish jip3 mutants (jip3nl7) displayed large axon terminal swellings that contained high levels of activated JNK and lysosomes, but not other retrograde cargos such as late endosomes and autophagosomes. Using in vivo analysis of axonal transport, we demonstrated that the terminal accumulations of activated JNK and lysosomes were due to a decreased frequency of retrograde movement of these cargos in jip3nl7, whereas anterograde transport was largely unaffected. Through rescue experiments with Jip3 engineered to lack the JNK binding domain and exogenous expression of constitutively active JNK, we further showed that loss of Jip3–JNK interaction underlies deficits in pJNK retrograde transport, which subsequently caused axon terminal swellings but not lysosome accumulation. Lysosome accumulation, rather, resulted from loss of lysosome association with dynein light intermediate chain (dynein accessory protein) in jip3nl7, as demonstrated by our co-transport analyses. Thus, our results demonstrate that Jip3 is necessary for the retrograde transport of two distinct cargos, active JNK and lysosomes. Furthermore, our data provide strong evidence that Jip3 in fact serves as an adapter protein linking these cargos to dynein.
| To form and maintain connections, neurons require the active transport of proteins and organelles between the neuronal cell body and axon terminals. Inhibition of this “axonal” transport has been linked to neurodegenerative diseases. Despite the importance of this process, to date there was no vertebrate model system where axonal transport could be studied in an intact animal. Our study introduces zebrafish as such a model and demonstrates its power for the analysis of axonal transport. We used this system to 1) initiate a genetic screen to find novel mediators of axonal transport; 2) develop in vivo imaging strategies to visualize axonal transport in real time in the intact animal; and 3) discover, using these methods, that JNK interacting protein 3 (Jip3) is required for the transport of two cargos, a kinase and lysosomes, from axon terminals to the cell body (retrograde transport). In the absence of Jip3, these cargos accumulate and axon terminals become dysmorphic, though the retrograde transport of other cargos is normal. Interestingly, abnormal localization of these cargos has been linked to axonal disease states, but our work is the first to identify a specific adapter protein necessary for their transport from axon terminals.
| Active transport of proteins and organelles between the neuronal cell body and axon terminals is necessary for the formation and maintenance of functional neural circuits. Anterograde (to axon terminals) and retrograde (to the cell body) transport rely on motor proteins of the Kinesin and Dynein families respectively. These motors use the energy of ATP hydrolysis to walk along microtubule tracks, carrying cargo to its proper destination. Though 15 kinesin families exist in mammals [1], only 1 retrograde microtubule based motor protein, cytoplasmic dynein, is responsible for the majority of retrograde cargo transport in axons [2]–[4], leading to intriguing questions about the nature of dynein-cargo interaction specificity which have been largely unexplored [5].
The core cytoplasmic dynein motor is composed of an array of proteins that includes two motor domain-containing heavy chains, two intermediate chains, two light intermediate chains, and four light chains which bind the intermediate chains [6]. Though recombinant dynein heavy chain can function in microtubule sliding assays in vitro [7], dynein complex interacting proteins have been shown to be essential for the initiation of retrograde cargo movement in vivo. Dynactin, a large dynein-interacting protein complex, and Lis1 have been separately shown to be co-factors that are necessary for the initiation of retrograde transport [8]–[10]. Loss of either of these factors leads to decreased retrograde transport frequency of some cargo and can lead to the accumulation of dynein components as well as cargo in axon terminals [9]. Retrograde cargo is thought to either bind directly to the core dynein complex proteins or, alternatively, to additional adapter proteins. It is tempting to speculate that the use of distinct adapter proteins may confer specificity to motor-cargo interactions in the dynein motor system. Despite their importance for the understanding of dynein-based cargo transport, the identity of specific dynein cargo adapters is dramatically lacking [5].
We used the advantages of the zebrafish system, including its amenity to forward genetics and live imaging, to identify Jip3 (JNK-interacting protein 3) as a cargo-specific adapter for dynein-based axonal transport. Through a forward genetic screen, we isolated a mutant strain (jip3nl7) that exhibited swellings in axon terminals of long sensory axons, a potential sign of interrupted retrograde transport. jip3nl7 carried a mutation in Jip3, a scaffold protein shown previously to serve as an adapter and facilitator of synaptic cargo anterograde transport through its interaction with Kinesin-1 [11]–[13]. In addition to anterograde transport machinery, Jip3 interacts with components of the dynein motor complex and c-Jun N-terminal Kinase (JNK). Indeed, Jip3 was first identified as a scaffold protein that links JNK to its upstream activating kinases, facilitating JNK activation [14]. Interestingly, Cavalli and colleagues demonstrated that Jip3 and activated JNK (pJNK) colocalized with p150glued (dynactin complex protein) distal to sciatic nerve injury. Based on this data, they postulated that Jip3-JNK-dynein interaction may be important during retrograde damage signaling [15]. Furthermore, in this and other studies, Jip3 has been shown to biochemically interact with components of the retrograde motor complex, specifically p150glued [15] and dynein light intermediate chain (DLIC; [13]). Thus, an intriguing possibility is that Jip3 could serve as an adapter for dynein-mediated retrograde transport of JNK and other cargo; however, neither this hypothesis nor the possibility that Jip3 is required for retrograde transport of any cargo, has been directly addressed to date.
Our work reveals discrete and direct roles for Jip3 in the retrograde transport of two cargos, pJNK and lysosomes. Using an in vivo imaging technique we developed for use in the zebrafish, we found specific retrograde transport defects in jip3nl7: frequencies of lysosome and pJNK retrograde transport were decreased causing accumulation of both cargos in axon terminals. Further analyses showed that direct Jip3-JNK interaction was necessary for retrograde clearance of pJNK from axon terminals and provided evidence that increased levels of pJNK were directly responsible for axon terminal swellings. Surprisingly, JNK activity and Jip3-JNK interaction had no impact on lysosome localization. Rather, co-transport analysis of lysosomes with both Jip3 and DLIC provided strong evidence that DLIC-lysosome interaction during retrograde transport relies on Jip3. Thus, based on our data we posit that Jip3 serves as an adapter protein for the retrograde transport of two distinct cargos, pJNK and lysosomes, and that failed retrograde clearance of pJNK contributes to the dysmorphic axon terminals in jip3nl7 mutants.
jip3nl7 was isolated in a forward genetics screen for which we utilized the TgBAC(neurod:EGFP)nl1 transgenic zebrafish (hereafter referred to as neurod:EGFP; [16]). This transgenic strain expresses an EGFP reporter in the central and peripheral nervous systems, including the posterior lateral line (pLL) ganglion and the long sensory axons emanating from it (Figure 1A, 1B; for screen details consult the Materials and Methods). We focused our screen on the long sensory axons of the pLL because of their planar character and superficial localization. These axons originate from the pLL ganglion, located just posterior to the ear, and extend along the trunk, branching to innervate mechanosensory hair cells that reside within surface sensory organs called neuromasts (NMs; axon terminals innervating NM3 and terminal NMs are shown in Figure 1B′ and 1B″ respectively). Initial pLL nerve extension and NM formation is complete by 2 dpf (days post-fertilization), and by 5 dpf a functional neural circuit has developed between NM hair cells and afferent pLL axons [17]. The recessive jip3nl7 mutant (Figure 1C) was isolated because it displayed truncation of pLL axons (incomplete penetrance; Figure 1C″) and swollen axon terminals innervating all trunk NMs (penetrance 100%; NM3 in Figure 1C′ and data not shown). To determine if long central nervous system axons were also affected by loss of Jip3, we analyzed axons of the reticulospinal tract as well as the efferent axons that project from the CNS to innervate the pLL NMs by crossing the jip3nl7 mutation into the TgBAC(phox2b:EGFP)w37 transgenic line [18]. Similar to pLL afferents, both reticulospinal tract and pLL efferent axons were truncated in jip3nl7 mutants (Figure 1D, 1E). jip3nl7 mutants were homozygous viable and the pLL axonal phenotype did not have a maternal component, as progeny derived from homozygous crosses displayed identical phenotypes to that of progeny derived from heterozygous crosses (data not shown).
We used a positional cloning approach to isolate the genomic locus containing the jip3nl7 gene mutation. Zebrafish Jip3, which mapped to this locus, is similar to its mammalian orthologs and contains two coiled coil domains, one leucine zipper deemed integral for Kinesin Light Chain (KLC) and dynactin binding [19], [20], and a JNK binding domain (Figure 1F). Sequencing of jip3 from jip3nl7 mutants revealed a mutation at nucleotide 552 which created a premature stop codon, truncating the Jip3 protein at amino acid 184 (Figure 1F). In situ hybridization analysis showed that, similar to mouse [21], jip3 was expressed in the central and peripheral nervous systems of the zebrafish embryo (Figure 1G). jip3 expression was lost in jip3nl7, perhaps due to nonsense-mediated mRNA decay (Figure 1H). Consequently, jip3nl7 is likely a Jip3 null. Initial investigations revealed the pLL nerve phenotypes were not due to impaired pLL patterning, neuronal cell death, abnormal glial support/myelination, or gross cytoskeletal abnormalities (Figure S1). As Jip3 has been shown to interact with members of the anterograde and retrograde motor complexes [11]–[13], [22], [23] and interruptions in transport have been associated with axon swellings like those observed in jip3nl7 [24], [25], we next focused our investigations on the potential function of Jip3 in axonal transport.
To study the function of Jip3 in axonal transport, we developed methods to visualize microtubule-based axonal transport in the pLL system in vivo, in intact zebrafish embryos and larvae (Figure 1I). Zebrafish are ideal for such a preparation as they are transparent through early embryonic and larval development, facilitating in vivo live imaging, and transient transgenesis can be used reliably to express tagged cargos of interest mosaically. Using these advantages, we developed a protocol that requires no surgical or invasive techniques to visualize protein or organelle transport in the long and planar axons of the pLL. To image axonal transport in zebrafish pLL axons, zygotes are injected with DNA encoding a cargo of interest tagged with a fluorescent reporter. Expression of these constructs is controlled by a neuron-specific 5 kilobase portion of the neurod promoter (5kbneurod; [26]). This results in mosaic expression of the desired cargo in the pLL ganglion, which, in ideal preparations, labels 1 to 2 neurons. Neurons expressing cargo are then monitored for full axon extension, innervation of NMs, and the absence of cargo accumulation in neuronal cell bodies and axons to assess optimal concentrations of DNA for injection. Using this approach, cargo transport can be visualized in individual pLL axons during axon extension (1–2 dpf), post-extension (after 2 dpf), and after functional synaptic connections are established (5 dpf).
We first utilized this technique to observe the localization and transport of a Jip3-mCherry fusion in pLL neurons and their axons. During axon extension (30 hpf), Jip3-mCherry localized to the neuronal cell body and axon growth cones (Figure 1J, 1K), similar to Jip3 localization in cultured neurons [27]. We then visualized Jip3 transport at 2 dpf, just after pLL nerve extension completes, and analyzed transport parameters using kymograph analysis (Figure 1L, 1M and Video S1). Jip3-containing cargo traveled at average velocities of 1.60 µm/sec in the anterograde direction and 1.35 µm/sec when moving in the retrograde direction (N = 7 larvae); these parameters are consistent with fast anterograde and retrograde transport [1].
Next, we assayed the localization and transport of ssNPY-mCherry [28], a marker of Golgi-derived vesicles, to determine if loss of Jip3 affects the axonal transport of this generalized cargo. At 5 dpf, we observed large accumulations of mCherry positive puncta in axon terminals of jip3nl7 mutants but not in wildtype siblings (Figure S2A, S2B; for this and other experiments, mutants were identified using the genotyping protocol described in the Materials and Methods, except where otherwise indicated). In vivo imaging and kymograph analysis demonstrated bidirectional movement of mCherry-positive puncta in wildtype and jip3nl7 mutants (Figure S2C–S2F; Videos S2 and S3) with decreased frequency of anterograde and retrograde transport of this cargo in jip3nl7 at 2 dpf with a tendency toward a decrease at 5 dpf (Figure S2G). Neither distance nor velocity of cargo movement were altered (Figure S2H, S2I), potentially implicating Jip3 in cargo-motor attachment, rather than modulation of motor activity.
Next, we set out to determine the identity of the mCherry labeled retrograde cargo(s) by looking for accumulation of commonly transported retrograde cargos in jip3nl7 axon terminals using immunofluorescence [29], [30]. Neither late endosomes (Rab7-positive) nor autophagosomes (LC3-positive) accumulated in jip3nl7 axon terminals (Figure S3A–S3D). Consistent with a previous study on Jip3's role in anterograde transport of TrkB [13], TrkB levels were decreased in jip3nl7 axon terminals, as assayed by TrkB antibody labeling (Figure S3E, S3F). In contrast, the axon terminal swellings in jip3nl7 were rich in lysosomes that were visualized using two separate markers, Lamp1 (detected by immunofluorescence; Figure 2A, 2B) and Lysotracker red (vital dye; Figure S3G, S3H).
We then asked whether abnormalities in lysosomal transport caused lysosome accumulations in axon terminals by employing our in vivo imaging approach, using a Lamp1-mTangerine fusion [31] to mark lysosomes in pLL axons (Figure 2C–2F; see Videos S4 and S5). The ability of a Lamp1-EGFP fusion construct to label lysosomes was confirmed by double labeling with the vital dye Lysotracker red (Figure 2G). Similar to our immunolabeling results, Lamp1-mTangerine accumulated in the axon terminals of jip3nl7 mutants but not wildtype controls (Figure 2E, 2F). Live imaging analysis demonstrated that, though Lamp1-mTangerine transport parameters were not altered at 2 dpf, the number of lysosomes moving in the retrograde direction was significantly decreased at 3 dpf in jip3nl7 axons (Figure 2H–2J; WT = 15.08±2.71 vs. jip3nl7 = 5.14±2.71 particles/100 µm*min, p≤0.002; Wilcoxon rank-sum). A similarly reduced frequency of lysosome retrograde transport was also observed at 5 dpf, while distance and velocity of movement were largely unaffected at all stages (Figure 2K, 2L). These data show that retrograde lysosome transport relies on Jip3.
Jip3 has been shown to interact with components of the Kinesin-1 motor to regulate anterograde transport [11]–[13], but a role for Jip3 in retrograde transport has not been described previously. Therefore, we next sought to address how Jip3 functioned to regulate retrograde axonal transport. Jip3 was originally identified as a JNK-interacting protein and has been shown to facilitate JNK activation in vitro [14]. Thus, we would predict that loss of Jip3 would lead to decreased JNK activation. As JNK activity can impact numerous intracellular processes that could potentially affect axonal transport machinery [32], [33], we assayed levels and localization of active JNK (pJNK) using pan-pJNK immunolabeling. Surprisingly, instead of a decrease, we found elevated levels of pJNK in the mutant axon terminals innervating all NMs from 2 dpf onward (Figure 3A–3I and data not shown; see Materials and Methods for an explanation of fluorescent intensity measurement). In contrast, total JNK (tJNK) levels in jip3nl7 were comparable to controls (Figure 3J and Figure S4A–S4D). Western blot analysis of whole embryo extracts revealed no increase in overall tJNK or pJNK levels in jip3nl7 (Figure S4E, S4F), pointing to a change in localization of pJNK rather than overall JNK expression or activity.
Given the ability of Jip3 to bind components of the retrograde motor and pJNK [14], [15], we reasoned that Jip3 might directly mediate pJNK retrograde transport/clearance from axon terminals by attaching this active kinase to the dynein motor complex. To determine if Jip3 has a specific role in pJNK transport, we used two complimentary approaches. First, we developed an axon injury model for use in the zebrafish pLL nerve to indirectly assay pJNK transport, similar to a protocol previously used in mouse sciatic nerve (Figure 4A; see Materials and Methods for procedure details; [15]). Following injury, cargos that are transported in the anterograde direction will accumulate proximal to the injury site, whereas retrograde cargos will accumulate distal to the injury site. Severing the pLL nerve between NM2 and NM3 at 5 dpf resulted in accumulation of pJNK in the pLL nerve proximal and distal to the site of injury in wildtype larvae by 3 hours post-injury. In contrast, pJNK failed to accumulate distal to the site of injury in jip3nl7 mutants (Figure 4B–4E, 4J), indicating failed retrograde pJNK transport in mutant axons. Total JNK levels were not significantly different proximal or distal to injury site in jip3nl7 mutants (Figure 4F–4I, 4K), though there was a strong trend towards decreased levels of the tJNK anterograde pool (proximal to the injury site) in jip3nl7 mutants. This data supports the hypothesis that loss of Jip3 inhibits pJNK retrograde transport, which would lead to accumulations of this kinase in axon terminals.
Next, we asked whether dynein motor components were normally transported to axon terminals in jip3nl7 mutants, as the perturbation of this transport could indirectly affect retrograde cargo movement. Using immunolabeling for two components of the dynein complex (Dynein heavy chain and p150glued), we demonstrated proper localization of these core dynein motor proteins to jip3nl7 mutants, confirming that the retrograde motor can reach axon terminals in jip3nl7 mutants (Figure S5A–S5G). From this data, we can also infer that even in the absence of Jip3, the initiation of dynactin-mediated, dynein movement was intact since these retrograde motor components did not accumulate in axon terminals [9], [10].
Finally, we used our in vivo live imaging to concretely determine if retrograde JNK transport was impaired in jip3nl7 mutant pLL axons using transient expression of JNK3 tagged with mEos. We chose to use JNK3 for our in vivo analysis because Jip3 has been shown to bind most strongly to the JNK3 homolog [14], and jnk3 is strongly expressed in the zebrafish nervous system (Figure S6A, S6B). Phospho-JNK immunolabeling of embryos expressing JNK3-mEos driven by the 5kbneurod promoter in pLL axons demonstrated that a large portion of JNK3-mEos positive vesicles carried the active form of this kinase (Figure 5A). Live imaging experiments revealed JNK3-mEos positive puncta traveled bidirectionally in wildtype and jip3nl7 mutants at 2 dpf (Figure 5B, 5C; Videos S6 and S7). Using kymograph analysis (Figure 5D, 5E), we found a decrease in the number of JNK3-mEos positive puncta moving in the retrograde direction at 2 dpf in jip3nl7 mutants (Figure 5F; wildtype:2.99±0.48 vs. jip3nl7:1.15±0.58 particles/100 µm*min, p≤0.05; Wilcoxon rank-sum) while retrograde movement distance and velocity were largely unchanged (Figure 5G, 5H). Taken together with the results from our injury model, these data confirmed that the frequency of retrograde pJNK transport was hindered in jip3nl7 mutants.
Based on our data and previous work showing that Jip3 can bind components of the dynein motor complex [15], we hypothesized that direct Jip3-JNK interaction was necessary for the retrograde transport of pJNK. To address this, we first asked whether Jip3 and JNK3 were transported together in pLL axons using a dual cargo transport assay. We co-injected Jip3-mCherry and JNK3-mEos plasmids and identified embryos in which both constructs were expressed in the same pLL neuron. Notably, co-injection of these and other cargos used for dual transport analysis (see below) resulted in almost 100% co-expression. Sequential imaging of Jip3 and JNK3 positive vesicles at 2 dpf revealed a high degree of co-transport, primarily in the retrograde direction (Video S8). While only 16% of vesicles in the anterograde pool were positive for both Jip3 and JNK3, 87% of vesicles in the retrograde pool carried both proteins (N = 5 embryos). This data supported a role for Jip3 in the retrograde transport of activated JNK. Importantly, since mEos is a green to red photoconvertable molecule, we used extreme caution during these dual imaging experiments to prevent accidental photoconversion and noted no green to red shift in the vesicles imaged during these sessions (data not shown).
Next, we addressed whether the direct interaction between Jip3 and JNK was necessary for retrograde pJNK transport by asking whether the pJNK accumulation in jip3nl7 could be rescued with a Jip3 variant that lacked the JNK binding domain (Jip3ΔJNK: amino acids 202–214; [32]). DNA constructs were injected into zygotes to mosaically express Jip3-mCherry or Jip3ΔJNK-mCherry in individual pLL ganglion neurons. At 4 dpf, axon terminals expressing the respective fusions were imaged live and scored for axon morphology before larvae were individually immunolabeled for pJNK and the same axon terminals were re-imaged. As each NM is innervated by 2 axons and this innervation is segregated in space [34], we could use the non-expressing half of the NM to identify which larvae were jip3nl7 mutants as well as utilize it as a normalizing factor for the quantification of pJNK immunofluorescence. Though full-length Jip3 rescued axon terminal swellings and the accumulation of pJNK, Jip3ΔJNK was unable to rescue either phenotype (Figure 6A–6E). Importantly, expression of Jip3ΔJNK by mRNA injection rescued axon length, providing evidence that deletion of this region did not result in protein instability or failed processing, and pointing to a JNK-independent mechanism for Jip3's role in axon outgrowth (Figure S7). In summary, these data show that direct interaction between Jip3 and JNK is necessary for pJNK retrograde transport and also revealed a correlation between the accumulation of pJNK due to loss of Jip3-JNK interaction and the generation of axon terminal swellings.
To determine if high levels of pJNK in axon terminals were sufficient to cause axon terminal swellings, we conditionally and mosaically expressed a constitutively active form of JNK3 (caJNK3; [35], [36]) fused to EGFP under the control of a heat shock promoter in pLL neurons of wildtype larvae. Fifteen hours after activation at 4 dpf, we identified larvae that were expressing this construct in pLL axon terminals. Subsequently, these larvae were individually immunolabeled using anti-pJNK and anti-GFP antibodies to determine if caJNK3 could alter axonal morphology and additionally determine if axonal swellings correlated with elevated pJNK levels. Using this assay, we found that increased pJNK levels by expression of caJNK3 correlated with the presence of axon terminal swellings (Figure 6F). Interestingly, expression of caJNK3 did not always elevate pJNK levels (8 out of 17 larvae) and axon terminals were not swollen in these instances (data not shown). To test if axon terminal swellings were a result of JNK activity, we mutated the site phosphorylated by the upstream activating MAPKK to render caJNK3 inactive (caJNK3-IA; [37]). To assay the efficacy of the caJNK3 and caJNK3-IA constructs, we expressed both individually using RNA-mediated whole embryo expression and assayed phospho-cJun levels, a direct downstream JNK target, by Western blot analysis. As predicted, caJNK3 elevated levels of p-cJun (Figure 6H) while caJNK3-IA did not (Figure 6I). Induction of caJNK3-IA using a protocol identical to that used of caJNK3 did not cause axonal swellings in any of the 16 larvae we imaged (Figure 6G), confirming that JNK activity was indeed required for the generation of axon terminal swellings. These experiments demonstrated that high JNK activity is sufficient to induce axonal swellings and provided strong evidence that the axon terminal swellings in jip3nl7 mutants are due to increased pJNK levels at axon terminals.
Our data demonstrated that lysosomes accumulate in jip3nl7 mutant axon terminals (see Figure 2) and elevated pJNK levels cause axon terminal swellings (see Figure 6). Next, we asked whether elevated pJNK could cause lysosomal accumulation. To test this, we used the approach described above to conditionally expressed caJNK3 at 4 dpf in wildtype larvae. Larvae expressing caJNK3 in pLL neurons were immunolabeled with an anti-Lamp1 antibody and axon terminals were imaged. This analysis demonstrated that elevation of pJNK levels did not increase Lamp1 levels above controls (Figure 7A, 7B). Importantly, lysosome number and dynamics appeared normal in the presence of activated JNK, as Lysotracker red vital dye labeling was similar between caJNK3 expressing axons and non-expressing neighboring axons (Figure 7C, 7D).
Based on genetic work in Drosophila, JNK has been postulated to act as a “switch”, controlling anterograde vs. retrograde motor activity for cargo transport [38]. Thus, we asked whether Jip3-JNK interaction could be a potential regulator of directional lysosome transport. First, we used sequential imaging to determine if JNK3 and lysosomes were co-transported by co-expressing JNK3-mEos and Lamp1-mTangerine in pLL axons and imaging their transport at 2 dpf (N = 6; Video S9). This analysis demonstrated that only ∼19% of Lamp1-positive vesicles moving in the anterograde or retrograde direction were co-labeled with JNK3-mEos. Interestingly, 72% of JNK3 positive retrograde vesicles label with Lamp1-mTangerine, suggesting that, though lysosomes do not rely on JNK3 for their movement, JNK3 was transported with lysosomes towards the cell body.
Finally, we tested whether Jip3-JNK interaction had any function in lysosome transport, which, if disrupted, could lead to lysosome accumulation in axon terminals in the absence of Jip3. To address this, we assayed whether lysosome accumulation in jip3nl7 mutants could be rescued by expressing Jip3ΔJNK (and Jip3 as a control) by RNA injection. For this assay, RNA was co-injected with the Lamp1-mTangerine DNA construct to visualize lysosomes in individual axons (see Figure 2E, 2F). Rescue score was determined as the average of the scores recorded by 2 blind, independent raters and was based on the ratio of punctate lysosomes (similar to wildtype in Figure 2E) vs. aggregates (as in mutants in Figure 2F). This analysis determined that Jip3ΔJNK was as effective as full-length Jip3 at suppressing lysosome accumulation in jip3nl7 mutants (Figure 7E). We did not, however, observe complete rescue, potentially due to RNA degradation by 3 dpf. To complement this analysis, we implemented a DNA-based expression strategy that would allow expression of the rescue constructs at later stages. We expressed Jip3-mCherry and Jip3ΔJNK-mCherry in pLL axons using the 5kbneurod promoter and assayed larvae for lysosome accumulation using Lamp1 immunolabeling at 4 dpf. Larvae were imaged live at 4 dpf to identify the axon terminals expressing these constructs and to identify mutant and wildtype siblings based on axonal phenotype of mCherry negative axons. Subsequently, larvae were individually immunolabeled for pJNK and Lamp1 and the same axon terminals were reimaged. Consistent with our previous results (see Figure 6), Jip3ΔJNK failed to rescue axon terminal swellings or pJNK accumulation in jip3nl7 mutants but was capable of suppressing the elevation of Lamp1 levels similar to full-length Jip3 (Figure 7F–7I and data not shown; N = 5 out of 8 jip3nl7 mutants injected with Jip3ΔJNK showed full rescue). Together, these data argue that Jip3-JNK interaction is not necessary for retrograde lysosome transport and supports a JNK-independent role for Jip3 in lysosome clearance from axon terminals.
In cultured cells, DLIC, a dynein accessory protein, functions in dynein-dependent lysosome transport [30]. As Jip3 has been shown to interact with DLIC [22], we hypothesized that Jip3 might serve as an adapter for lysosome-DLIC attachment during retrograde lysosome transport in axons. To ascertain whether Jip3 co-localized with moving lysosomes and could function in such a direct role, we performed sequential imaging of axons expressing both Jip3-mCherry and Lamp1-EGFP cargos at 2 and 3 dpf. Co-transport analysis revealed that Jip3 is present on lysosomes moving in the retrograde direction at both time-points (Figure 8A–8E; Video S10). Interestingly, the percentage of lysosomes that were transported in the retrograde direction labeled with Jip3 was higher at 3 dpf than at 2 dpf (2 dpf: 15%±3.8%, N = 5 vs. 3 dpf: 37%±4.2%, N = 4). This may indicate a differential reliance on Jip3 for the transport of this organelle beyond 2 dpf, leading to the decrease in lysosome retrograde transport frequency only after 2 dpf in jip3nl7 (see Figure 2).
Finally, we co-expressed DLIC tagged with mTangerine (mTangerine-DLIC) and Lamp1-EGFP to characterize DLIC localization and co-transport with lysosomes and determine if this association is lost in jip3nl7 mutants. At 3 dpf, mTangerine-DLIC localized to discrete puncta along the axon and in axon terminals in wildtype larvae (Figure 8F). In contrast, in jip3nl7 mutants, DLIC accumulated in axon terminals, similar to lysosomes and pJNK (Figure 8G). Co-transport analysis of mTangerine-DLIC and Lamp1-EGFP cargos revealed a decrease in the ratio of DLIC-positive lysosomes moving in the retrograde direction in jip3nl7 mutants (Figure 8H–8M; Video S11). This observation points to a failure of lysosome-dynein interaction during transport with loss of Jip3. Interestingly, there was a slight decrease in DLIC-Lamp1 vesicle co-transport in the anterograde direction as well in jip3nl7 mutants suggesting that this complex may move bidirectionally. In summary, our data supports a model where the independent interaction of Jip3 with pJNK and lysosomes is required for the attachment of these cargoes to the dynein motor for clearance from axon terminals (Figure 9).
Our results revealed a novel role for Jip3 in retrograde axonal transport. We provided evidence that loss of Jip3 led to a decreased frequency of retrograde transport of an active kinase (pJNK) and lysosomes but not other components of the endosomal or autophagocytic system. We demonstrated that direct interaction of Jip3 and JNK was necessary to prevent pJNK accumulation and the axon terminal swellings characteristic of the jip3nl7 mutant but had no effect on lysosome accumulation. Additionally, exogenous expression of activated JNK phenocopied the jip3nl7 mutant axon terminal swellings but did not cause lysosome accumulation, providing evidence that high levels of active JNK cause this phenotype in a lysosome-independent manner. Finally, our co-transport analysis suggested that Jip3 directly facilitated lysosome interaction with the dynein motor through binding to the accessory protein DLIC. Given the decrease in frequency of cargo movement, the normal distribution of dynein components in jip3nl7 mutant axon terminals, and the high rate of Jip3-lysosome and Jip3-JNK3 co-transport, we posit that Jip3 likely serves as an adapter protein that mediates attachment of these cargos to the dynein motor (Figure 9).
Jip3 has been implicated in anterograde axonal transport in several studies through its interaction with both Kinesin light chain and Kinesin heavy chain components of the Kinesin-1 motor [12], [13], [23]. We became interested specifically in Jip3's function in retrograde transport as jip3nl7 demonstrated the unusual quality of extreme swellings in axon terminals, the end of the line for anterograde transport. A function for Jip3 in retrograde transport has indeed been posited by Cavalli et al. as they demonstrated that Jip3 co-localized with pJNK distal to nerve ligation and co-purified from similar membrane fractions as dynein components [15]; however, our study is the first to provide conclusive evidence that Jip3 is required for retrograde transport of pJNK, as pJNK accumulates in axon terminals in jip3nl7 mutants, Jip3 and JNK3 are co-transported, and direct Jip3-JNK interaction is functionally required for pJNK retrograde transport. Thus, our work identifies pJNK as a Jip3-dependent retrograde cargo. In addition, through the implementation of our in vivo imaging approach, we found that the frequency of retrograde JNK3 transport was decreased with loss of Jip3, but the processivity of the motor (reflected by run length) and velocity of movement were unchanged. This data, in combination with previous biochemical studies of Jip3-JNK and Jip3-dynein interaction [15], provide strong evidence that Jip3 functions as an adapter for pJNK, linking it to the dynein complex for transport, while not affecting motor movement itself.
Using a combination of immunolabeling and in vivo imaging techniques, we further show that Jip3 is necessary for retrograde transport of lysosomes through interaction with the dynein accessory protein DLIC. DLIC has been shown to be an important mediator of dynein-based lysosome movement in culture systems [30] and was shown to biochemically interact with Jip3 in another system [22]. Thus, Jip3 could provide a link between lysosomes and dynein through its interaction with DLIC. In support of this, Jip3 is co-transported with lysosomes, the retrograde transport velocities for Jip3 alone were highly similar to those observed for lysosomes, and DLIC-lysosome co-transport was significantly decreased in jip3nl7 mutants. Together, these data provides strong evidence that Jip3 serves as an important adapter protein for lysosome-DLIC interaction and subsequent retrograde lysosome transport. Notably, Jip3 was implicated in the anterograde transport of DLIC to axon terminals in C. elegans [22]. However, instead of a decrease, we observed increased levels of DLIC in jip3nl7 axon terminals, arguing that this Jip3 function may not be conserved in vertebrates or is compensated for by another member of the Jip family [39].
Elevated levels of activated JNK, lysosome accumulation and axonal dysmorphology have been co-associated with neurodegenerative disorders [40]. Interestingly, though our studies indicated that Jip3-JNK interaction was not required for lysosome retrograde transport, JNK3 was frequently present on lysosomes moving in the retrograde direction, suggesting that Jip3 could serve to attach both cargos to the dynein motor simultaneously. Furthermore, our results point to a lysosome-independent etiology of axon terminal swellings in jip3nl7 mutants. Evidence to support a lysosome-independent mechanism includes: 1) the ability to induce axonal swellings without lysosome accumulation by exogenous expression of constitutively active JNK; 2) the absence of axon morphological changes following expression of an inactivated form of the constitutively active JNK; and 3) rescue of lysosome accumulation, but not pJNK levels or axonal swellings, in jip3nl7 mutant axon terminals by Jip3ΔJNK expression. Thus, our work provides evidence that axonal swellings can occur downstream of this active kinase without causing concomitant accumulation of organelles in the autolysosomal pathway. The exact etiology of axonal swellings in jip3nl7 mutants due to elevated levels of activated JNK remains to be determined.
Importantly, jip3nl7 mutants did not exhibit a global disruption of retrograde axonal transport, which would indirectly lead to cargo accumulations. Evidence supporting the specificity of transport disruptions includes: 1) absence of the accumulation of other cargo (late endosomes, autophagosomes, and signaling endosomes) in jip3nl7 axon terminals; and 2) normal localization of dynein heavy chain and p150glued in jip3nl7 axon terminals, indicating that dynactin-based initiation of dynein transport is not hindered [9], [10]. Thus, our data supports a direct role for Jip3 as an adapter for the transport of two specific retrograde cargos, pJNK and lysosomes.
In summary, our data demonstrate novel and separate roles for Jip3 in the retrograde axonal transport of activated JNK and lysosomes. It is tempting to speculate that Jip3-dependent retrograde clearance of activated JNK may be a novel and crucial strategy for the removal of this active kinase from axon terminals, bypassing traditional phosphatase pathways. Furthermore, we show that enhanced JNK activity can indeed cause axon terminal swellings, similar to those observed in the jip3nl7 mutant, in the absence of lysosome accumulation. Thus, we have shown that there can be an independent etiology for these tightly coupled events observed in disease models. The similarities between the axonal swellings, high levels of pJNK, and accumulation of lysosomes in jip3nl7 and neurodegenerative diseases such as Alzheimer's Disease points to an intricate relationship between these phenotypes during pathogenesis. Our studies begin to unravel how Jip3-dependent regulation of retrograde axonal transport may underlie or modulate such disease states.
Adult *AB and WIK zebrafish and *AB/WIK hybrids were maintained at 28.5°C and staged as described [41]. Embryos were derived from natural matings or in vitro fertilization, raised in embryo media, and developmentally staged using previously established methods [42]. Strains utilized included TgBAC(neurod:EGFP)nl1 [16], TgBAC(phox2b:EGFP)w37 [18], TgBAC(neurog:DsRed)nl6, TgBAC(foxd3:EGFP)nl5 transgenics and mitfaw2 [43], and mapk8ip3nl7 (jip3nl7) mutants.
We used Escherichia coli-based homologous recombination to modify a neurog1- and foxd3-containing bacterial artificial chromosome (BAC) clones [44]. The neurog1 BAC clone zK171N3 contains 63.8 kb of upstream and 106.1 kb of downstream sequence of neurog1, while the foxd3 BAC clone zC137J12 contains 66.2 kb of upstream and 122.1 kb of downstream sequence of foxd3 (http://www.sanger.ac.uk/Projects/D_rerio/mapping.shtml). After recombination, the modified BAC clones contained DSRedExpress-1 and EGFP positioned at the endogenous start site of neurog1 or foxd3, respectively. The accuracy of recombination was evaluated by PCR, sequencing, and analysis of transient expression. To obtain germline transgenics, we microinjected 20–80 pg of BAC DNA (linearized with Srf I for neurog1 BAC and supercoiled for foxd3 BAC) into zebrafish zygotes, raised injected fish to adulthood, and screened their progeny for reporter gene expression. The germline transmission rate was 2.3% for neurog1 BAC and 1.4% for the foxd3 BAC. The TgBAC(neurog1:DSRed)nl6 and TgBAC(foxd3:EGFP)nl5 strains have been outcrossed for multiple generations and transfmitted the transgenes in a Mendelian manner.
The jip3nl7 mutant was identified in a standard three-generation N-ethyl-N-nitrosourea (ENU) mutagenesis screen [45], [46]. For this screen, TgBAC(neurod:EGFP)nl1-positive larvae were screened at 4 dpf for axon truncation and the presence of axonal swellings under epifluorescence. For genetic mapping, heterozygous carriers of jip3nl7 on a polymorphic *AB/WIK background were incrossed to produce homozygous, heterozygous and wildtype progeny. Initial chromosome assignment was done by bulk segregate analysis of DNA pools from 20 wildtype and 20 mutant individuals using microsatellite markers (http://zfin.org/ZFIN). Flanking regions were identified using individual wildtype and mutant larva and markers z15457, z21697, and a designed marker, CA50 (forward: 5′-TTACACACTTTCAGCCTGTC, reverse: 5′-CCTTTATGCCACGGTCACA).
Genomic DNA was isolated from larvae by incubating it overnight at 55°C in PCR Extraction Buffer (10 mM Tris pH = 8, 2 mM EDTA, 0.2% Triton X-100, 200 µg/ml Proteinase K). Total RNA was isolated from larvae using Trizol according to the manufacture's protocol (Invitrogen) and cDNA was generated using Superscript II reverse transcriptase and oligo dT primers (Invitrogen). The full mapk8ip3 (jip3) cDNA was amplified using following primers (forward: 5′-CGTTAAACGAGCTTCGGACA, reverse: 5′-GCGTGTCACTTTGAGTTTGG) based on the predicted sequence and subsequently entered into GenBank (KC170712). Full-length jnk3 was amplified using primers (forward: 5′-ATGAACAGACGTTTCTTATATAACTGC, reverse: 5′-CACGGCTGCACCTGCGCTG) designed against the annotated sequence (NM 001037701). Full-length dynein light intermediate chain was amplified using primers (forward: 5′-TGTCACTCAAGCCTGCGAAG, reverse: 5′-GGATTTGTCGTTTTCAGCAG) designed against the annotated sequence (NM 001017669). To genotype jip3nl7 carriers, a 385 bp region around the mutation was amplified from genomic DNA by PCR using annealing T = 55°C and the following primers (forward: 5′-TTTGTCTGTTGAAATTGCT, reverse: 5′-ACGGTCCATACCCATGATT). PCR products were then digested with SpeI, as the single nucleotide change generates this restriction site in the jip3nl7 allele, producing two bands, 243 and 142 bp.
RNA in situ hybridization was performed as described [47]. Digoxygenin-labeled antisense RNA probes were generated for jip3 and jnk3 using the full-length cDNA cloned. Whole mount immunohistochemistry was performed following established protocols [48]. The following antibodies were used: anti-GFP (1∶1000; Invitrogen #A11122), anti-pJNK (1∶100; Cell Signaling #9251S), anti-tJNK (1∶100; Cell Signaling #9252), anti-p150glued (1∶100; Signal Transduction Labs #610473), anti-dynein heavy chain (1∶100; gift of R. Vallee; [49]), anti-Rab7 (1∶100; Sigma #R8779), anti-Lamp1 (1∶100; Developmental Studies Hybridoma Bank), anti-LC3 (1∶100; Novus #NB100-2331), anti-TrkB (1∶100; Santa Cruz Biotechnology #sc-12) and Alexa-488/568/647 (1∶750; Invitrogen). Antibodies not used previously in zebrafish were validated by Western blot analysis (see below: Figure S3I–S3L; Rab7–24 kD, LC3–14.5 kD, TrkB-69 kD and 18 kD, Lamp1–27 kD). For TUNEL labeling, embryos were processed as previously described [50] with minor modifications according to the manufacturer's instructions (In situ cell death kit, Roche). For Lysotracker red vital dye staining, 4–5 dpf larvae were incubated in Lysotracker red (1∶10,000; Invitrogen) for 15 minutes in embryo media, washed briefly, embedded in 1.2% low-melt agarose, and imaged. All fluorescently labeled embryos were imaged using a FV1000 laser scanning confocal system (Olympus). Brightfield or Nomarski microscopy images were collected using a Zeiss Imager Z1 system. Images were processed using ImageJ software [51]. Brightness and contrast were adjusted in Adobe Photoshop and figures were compiled in Adobe Illustrator.
For western blot analysis, protein was isolated from wildtype and jip3nl7 3 dpf larvae by homogenizing individuals in extraction buffer (150 mM NaCl, 50 mM Tris pH = 7.4, 5 mM EDTA, 0.05% NP40, 25 mM NaF, 10 mM Na3VO4, 1 mM DTT, 10 µL/mL protease inhibitor) at a ratio of 4 µL buffer per embryo. The equivalent of 4 embryos was run in each lane on a 12% SDS-PAGE gel and transferred onto a PVDF membrane (Millipore). Primary antibodies were applied overnight: anti-pJNK (1∶1000; Cell Signaling #9251S), anti-tJNK (1∶1000; Cell Signaling #9252), anti-p150glued (1∶1000; Signal Transduction Labs #610473), anti-dynein heavy chain (1∶1000; gift of R. Vallee; [49]), anti-Rab7 (1∶2000; Sigma #R8779), anti-Lamp1 (1∶4000; Developmental Studies Hybridoma Bank), anti-LC3 (1∶500; Novus #NB100-2331), anti-TrkB (1∶100; Santa Cruz Biotechnology #sc-12), and anti-p-cJun (1∶1000; Cell Signaling 9164S). After washing, an anti-rabbit*HPR, anti-mouse*HRP, or anti-rat*HRP secondary (1∶10,000; Jackson Immuno) was applied for 90 minutes. Protein was visualized using SuperSignal West Pico Chemiluminescent Substrate according to the manufacture's specification (Thermo Scientific). If necessary, the blot was then stripped with 25 mM glycine (pH = 2.5) and re-probed with rabbit anti-α-actin (1∶10,000; Sigma).
To generate constitutively active JNK3 that could be activated in a temporally specific manner, we fused MKK7 to JNK3 and placed this fusion behind a heat-shock inducible promoter (hsp70:mkk7-JNK3-egfp, referred to as caJNK3 in the text). To generate an inactive form of the same construct (referred to as caJNK3-IA in the text), two amino acids were mutated (T221A and Y223F) to render JNK3 not able to be phosphorylated, which is required for its activity [37]. For induction of transcription of both constructs, 4 dpf larvae injected with 10 pg of the caJNK3 or caJNK3-IA constructs were heat-shocked at 38°C for 1 hour. Larvae were then transferred to 28.5°C prior to analysis.
Zygotes were injected with plasmid DNA encoding fluorescently tagged cargos of interest with expression driven by the 5kbneurod promoter [26]. At 30 hpf, 2 dpf, or 5 dpf, embryos or larvae were sorted under epifluorescence to identify individuals with tagged cargo expression in a few cells of the pLL ganglion. For imaging, embryos were mounted in 1.2% low melting point agarose on a glass coverslip, submerged in embryo media containing 0.02% tricaine and imaged using a 60X/NA = 1.2 water objective on an upright Fluoview1000 confocal microscope (Olympus). For each embryo, a region of interest (30–200 µm) was selected in the pLL nerve in which a long stretch of axon was observable in a single plane. Scans were taken at the fastest possible speed (3–5 frames per second) for 600 to 2500 frames. Embryos were subsequently released from agarose and processed for genotyping. For co-transport, embryos expressing both constructs in a single cell were selected and imaged as described above using sequential imaging of the 488 and 568 nm excitation channels. 600 frames were collected at 2–3 frames per second.
Transport parameters were analyzed using kymograph analysis in the MetaMorph software package (Molecular Devices, Inc.). Kymographs were generated from each imaging session and used to determine distance moved in individual bouts of movement (uninterrupted straight lines) and velocity of movement (slope of uninterrupted straight lines). Typically, 10–50 traces were analyzed in each kymograph and these were averaged within individual embryos for statistical analysis. The number of particles moving in each direction was estimated based on traces on the kymographs and then normalized to length of axonal segment and total imaging time.
Five day old zebrafish larva (neurod:EGFP carriers) were anesthetized in 0.02% tricaine (MS-222; Sigma) and embedded in 3% methylcellulose on a slide. Pulled thick-walled glass capillaries were used to sever the nerve between NMs 2 and 3. Slides were immersed in Ringer's solution (116 mM NaCl, 2.9 mM KCl, 1.8 mM CaCl2, 5 mM HEPES pH = 7.2, 1% Pen/Strep) and incubated at 28.5°C for 3 hours. Larva were then collected and immunolabeled for pJNK or tJNK and EGFP. Details of image and statistical analyses are described below.
For analysis of pJNK and tJNK intensity in axon terminals and after nerve injury, individuals were immunolabeled as described above. For consistency of labeling, larvae that were directly compared were processed in the same batch. Confocal Z-stacks (0.5 µm between planes) were taken of the area of interest using a 40X/NA = 1.3 oil objective with identical settings. Images were analyzed using ImageJ [51]. For fluorescent intensity measurements of pJNK or tJNK in wildtype and mutant axon terminals, summed projections of the regions of interest were generated only through regions that contained the neurod:EGFP signal and converted to 8 bit in ImageJ. In the pLL nerve injury analysis, a 30 µm, neurod:EGFP-positive region encompassing the proximal or distal edge of the severed axon was selected and summed projections through only this segment were compiled for analysis. By restricting our analysis to the neurod:EGFP axons we eliminated a majority of the fluorescent signal from the surrounding tissue. Prior to statistical comparison, the mean background fluorescent intensity, measured in a region adjacent to the NM axon terminal or injury site, was subtracted from the values generated.
For analysis of pJNK levels in the DNA rescue experiment, axon terminals expressing Jip3-mCherry or Jip3ΔJNK-mCherry (typically innervating half a NM) and control terminals not expressing these constructs (the alternate half of the NM) were outlined in similar summed confocal projections and the mean fluorescent intensity was measured. The ratio of pJNK fluorescence in the axons expressing the rescue construct to those not expressing the rescue construct were compared for statistical analysis.
Statistical analysis was performed using the JMP software package. Data suitable for parametric analysis were analyzed using ANOVA, with Tukey-Kramer highly significant difference post-hoc contrasts for more than two variables. Data not suitable for parametric analysis were analyzed using Wilcoxon rank-sum analysis. Ordinal data was analyzed using Chi Square test. In all cases, data from individual embryos was averaged prior to analysis making each N equivalent to an embryo.
All animal work was approved by and conducted according to guidelines of the Oregon Health & Science University IACUC.
|
10.1371/journal.pcbi.1005174 | Temperature-Dependent Model of Multi-step Transcription Initiation in Escherichia coli Based on Live Single-Cell Measurements | Transcription kinetics is limited by its initiation steps, which differ between promoters and with intra- and extracellular conditions. Regulation of these steps allows tuning both the rate and stochasticity of RNA production. We used time-lapse, single-RNA microscopy measurements in live Escherichia coli to study how the rate-limiting steps in initiation of the Plac/ara-1 promoter change with temperature and induction scheme. For this, we compared detailed stochastic models fit to the empirical data in maximum likelihood sense using statistical methods. Using this analysis, we found that temperature affects the rate limiting steps unequally, as nonlinear changes in the closed complex formation suffice to explain the differences in transcription dynamics between conditions. Meanwhile, a similar analysis of the PtetA promoter revealed that it has a different rate limiting step configuration, with temperature regulating different steps. Finally, we used the derived models to explore a possible cause for why the identified steps are preferred as the main cause for behavior modifications with temperature: we find that transcription dynamics is either insensitive or responds reciprocally to changes in the other steps. Our results suggests that different promoters employ different rate limiting step patterns that control not only their rate and variability, but also their sensitivity to environmental changes.
| Temperature affects the behavior of cells, such as their growth rate. However, it is not well understood how these changes result from the changes at the single molecule level. We observed the production of individual RNA molecules in live cells under a wide range of temperatures. This allowed us to determine not only how fast they are produced, but also how much variability there is in this process. Next, we fit a stochastic model to the data to identify which rate-limiting steps during RNA production are responsible for the observed differences between conditions. We found that genes differ in how their RNA production is limited by different steps and in how these are affected by the temperature, which explains why different genes respond differently to temperature fluctuations.
| Temperature is known to affect gene expression patterns in cells. This has profound effects, as changes in transcription and translation dynamics propagate to the behavior of genetic networks, which manifests in their sensitivity to temperature changes [1–3].
The expression patterns are not solely characterized by the rates at which the genes are expressed, but also by the associated stochasticity. The latter affects the phenotypic variability of populations of genetically identical cells [4–6] and the temporal variations in the behavior of the individual cells [7]. In unicellular organisms, such variations, even at the level of single molecules [8], can determine a cell fate.
In bacteria, much of the stochasticity in gene expression stems from transcription [9]. Live-cell measurements report that different promoters and intra- and extracellular conditions result in wide differences in transcription dynamics, both in rate and stochasticity [10]. Sub-Poissonian [11], Poissonian [12], and super-Poissonian [10, 13] dynamics (featuring less, equal to, and more variability than a corresponding Poisson process, respectively) have been reported, each resulting from a different combination of mechanistic properties that shape RNA production dynamics.
The way the effects of temperature changes on transcription kinetics propagate to the cellular behavior is still poorly understood. Rates of biochemical reactions are known to be affected by temperature changes, as dictated by the laws of physics, such as the Arrhenius law; however, biochemical laws have been found to not suffice to explain changes in more complex biological processes such as bacterial growth [14].
For example, it is expected that the number of RNA polymerases, the rate at which they work, and RNA lifetimes are affected by temperature. Also, at suboptimal temperatures, Escherichia coli shifts to specific expression patterns by changing in transcription factor numbers [15], by regulating of the relative σ-factor concentrations [16], and by affecting DNA conformation: negative supercoiling increases at low temperatures, and relaxes at high ones [17]. Despite these findings, quantitative information of the changes and on their contribution to the changes in transcription dynamics is still lacking.
Some progress toward these goals has been made through in vitro measurements of the closed and open complex formation dynamics [18–20]. Another study reports that RNA polymerase, rRNA, and tRNA concentrations, and the fraction of stable RNA remained constant, while the elongation rate, ppGpp concentration, and cell growth rate increase to up to 40°C, while after 40°C the changes become complex and, e.g. the growth rate decreases [21]. So far, these studies only identified changes in mean expression rates. Further information on other dynamical properties, such as stochasticity, is required. In addition, it is unclear to what extent these measurements reflect what occurs in live cells.
Here, we study the transcription dynamics of the Plac/ara-1 promoter in live E. coli at the single RNA level, under various inducer concentrations, for a wide range of temperatures. Using statistical models, we identify the most likely changes with temperature in the rate limiting steps of transcription initiation. This expands on our previous work [22–24] in that the analysis at different temperatures allows identifying more complex underlying details of the kinetics of transcription. Also, we test if similar changes are observed in the PtetA promoter. Finally, we use the inferred models to study the possible causes of the underlying changes.
We use the following model of transcription initiation [25], which combines the active-inactive promoter model [26] with a sequential model of transcription initiation [27, 28]:
Poff⇌koffkonPon→k˜1R︸pre-commit stepI1→k2I2→k3Pon+E︸post-commit step (1)
where Poff and Pon represent the promoter in an inactive (i.e. repressed) and active (i.e. free from repressors or bound by an activator) states, respectively, R represents RNA polymerases, and Ik are intermediate complexes of transcription initiation. Finally, the product E represents the elongation complex. As the RNA polymerase numbers are not observed, we let k 1 ≐ k ˜ 1 R to represent the effective forward rate for an active gene.
In the model, the promoter switches between being active and inactive (on/off) for transcription, i.e. whether an RNA polymerase can unobstructedly reach the start site and initiate transcription [9], depending on the binding and unbinding of regulatory molecules [9, 29]. In some cases, these molecules can also affect subsequent steps [28, 30].
Note that once the promoter is in a state that allows the RNA polymerase to bind and form the closed complex, not necessarily will this result in the production of an RNA, as the closed complex can reverse to the previous state [20, 31]. As such, the system is not yet fully committed to transcription. This “full commitment” only occurs once reaching the next state (I1). At this stage, it becomes highly unlikely that any reversion occurs (e.g. in the λ PR promoter, I2 is always unstable compared to the complete open complex state, even at 0°C [31]).
Once this commitment occurs, it is followed by a sequence of steps responsible for the formation of a stable open complex, which includes the isomerization steps [20, 28, 31, 32]. The last step in this model represents the complex escaping the start site, clearing the promoter region. Note that the reversibility of the closed complex formation [20, 24] does not reduce the applicability of the model [24, 25]. Also, regardless of this reversibility, the model can still be separated into an R-dependent (pre-commit) and R-independent (post-commit) parts.
The ability of regulatory molecules to create an on/off promoter dynamics allows bursty RNA production when fast production events are separated by long, random off periods (due to, e.g. slow repressor unbinding) [7]. If this process dominates RNA production, the transcription intervals are highly noisy (coefficient of variation cv ≥ 1). Meanwhile, if the subsequent sequential process dominates transcription, the intervals between RNA production events are more regular, resulting in less noisy transcription (cv ≤ 1).
Which and how many steps most contribute to the observed transcription dynamics appears to depend on the promoter and intra- and extracellular conditions [10, 11, 13, 22, 23]. In what concerns modeling this process, for example, if the promoter’s visit to the off-state or a sequential step are fast, these can be eliminated from the model, as they do not contribute significantly to the transcription dynamics. The steps with most influence on the transcription rate are called rate limiting [32].
Here we equate the transcription intervals with those of transcription initiation, which implies that transcription elongation is neglected. This is justified by the fact that, on average, elongation is not expected to affect the transcription intervals, as each transcript is expected to be delayed by a similar time. As a result, elongation primarily adds solely extra variance in the inter-transcription intervals. However, given the timescales of elongation and transcription initiation, this additional variance can be ignored: e.g. chain elongation at 50 nt/s [33] for a 5000 nt gene is expected to add a noise term with a standard deviation of 2 s [25]. Even considering elongational pauses under GTP-starved conditions [34], the elongation noise term is therefore likely negligible at the resolution of our measurements (cf. sampling interval of 60 s). Further, we note that no differences have been found between genes with elongation regions of different length [7, 35].
As each of the steps are complex processes rather than elementary chemical reactions, it is not yet well understood how the temperature affects of each of the steps in Eq (1) in live cells. For this reason, we let each of the model parameters change as a function of the temperature according to a polynomial function, rather than according to some biochemical model. As we have relatively few samples on the temperature axis, a quadratic equation was deemed sufficient to capture either temperature-independent or a linear and nonlinear temperature-dependent relationship:
k x ( T ) - 1 = a k x - 1 , 2 T 2 + a k x - 1 , 1 T + a k x - 1 , 0 (2)
where a k x - 1 , j are the order-j coefficients of the polynomial of the temperature-dependence of the parameter kx, and T is the temperature. While this model is not expected to provide particular insight of the mechanisms of temperature-dependence (i.e. a k x - 1 , j do not encode a particular physical meaning), it allows us to identify how each parameter likely responds to temperature changes, be it independent or dependent, linear or nonlinear, or monotonic or bitonic.
The parameters of the model can be estimated from the transcription intervals in a maximum likelihood sense [25]. Confidence in the model parameters can be estimated using the delta method [36]. Further, to determine if each of the steps in the above model play a significant role in the overall dynamics, we use the Bayesian information criterion (BIC) [37] as the model selection criterion. A difference in BICs (ΔBIC) of 0 to 2, 2 to 6, or ≥ 6 indicates weak, positive, or strong evidence, respectively, against the model with the greater BIC [37]. For the purposes of BIC, a censored sample is assumed to be worth 0.5 exact samples (the exact value varies depending on how badly the sample is censored). To avoid drawing false conclusions due to this approximation, we also compute a lower bound for ΔBICs when determining if the best models fits significantly better than the alternatives (see S1 Appendix).
Transcription intervals in individual cells were measured in live E. coli using the MS2-GFP RNA-tagging system [7]. The cells feature a multi-copy reporter gene expressing MS2-GFP and a single-copy target gene, controlled by the promoter of interest, containing 96 MS2-GFP binding sites. Shortly after a target RNA is produced, the binding sites are quickly occupied by the abundant GFPs, allowing the fully tagged RNAs to be visualized using fluorescence microscopy [38]. Once formed, the RNA-96-MS2-GFP complex remains fluorescent for much longer than the cell lifetime [38]. The constructs used here were engineered previously [7, 11]. An example microscopy image is shown in Fig 1.
The activity of the target genes was also analyzed by quantitative PCR (qPCR) as a function of the media composition. To quantify intracellular RNA polymerase concentrations in each condition, we measured the amount of RpoC subunits by western blot (WB), as these are the limiting factor in the assembly of the RNA polymerase holoenzyme [39, 40]. Measuring gene activity as a function of the RNA polymerase abundance allows extrapolating the relative duration of the pre- and post-commit steps [24, 32].
More details of the constructs, measurement procedures, and data acquisition methods is given in S1 Appendix.
We studied how the distribution of durations between consecutive RNA productions in multiple, individual cells (afterwards denoted by transcription intervals) of the Plac/ara-1 promoter differ with temperature. This study was conducted for each possible induction scheme of this promoter, so as to assess if the temperature-dependence of the initiation kinetics is inducer scheme-dependent. Those induction schemes are: a) 1 mM of IPTG and 0.1% of arabinose (denoted by Full), b) 1 mM of IPTG only (IPTG), and c) 0.1% of arabinose only (Ara). In each condition, we recorded multiple time series of 120 minutes in length with a 1 minute sampling interval for temperatures between 24 and 41°C (see S1 Appendix for details).
We first quantified how the mean and the standard deviation (sd) of the transcription intervals change with temperature. As the cell division time varies significantly between conditions, a procedure accounting for the uneven truncation of the intervals must be used. For this, the right-censored (unobserved) transcription intervals at the end of each time series are communicated to the estimator as in [25]. First, we used gamma distribution as the model, as it can have any mean and sd independent of each other (a model must be assumed in order to estimate moments from the censored data). The results are tabulated in S1 Table. The distributions are shown in Fig 2. These data suggest that, the transcription interval duration changes by about a 2× factor along the range of temperatures tested, for each induction scheme. In addition, these changes appear to be nonlinear, and even non-monotonic in some cases. Meanwhile, the standard deviation of the transcription intervals tends to follow the mean, resulting in a coefficient of variation (cv) of slightly above unity with a slight increase as temperature increases. The median-to-mean ratio is approximately constant with respect to temperature at about 0.61, which indicates that the temperature affects the whole distribution and not only the long intervals.
Next, we fit the data with the model of Eq (1) independently in each induction scheme and temperature, and tested if the on/off switching and which forward steps are most responsible for the observed transcription dynamics using the Bayesian information criterion (BIC) (see Materials and Methods). The results are summarized in S2 Table. In all cases, we found evidence only for a single sequential rate limiting step. While there is weak evidence that this is true in all cases (ΔBIC lower bound (ΔBIC LB; see S1 Appendix) against any multi-step model is no less than 1.15), strong (statistically significant) evidence exists that this holds at least in certain temperature ranges (for all induction schemes, there exists a temperature with ΔBIC LB no less than 5.16). In addition, we also found evidence for the presence of on/off switching in each induction scheme, as there is a condition where the lower bound for ΔBIC is no less than 9.20, providing strong evidence that the on/off switching plays a role, in at least certain temperature ranges, regardless of the induction scheme.
As the order of the steps with rates k1, k2, k3 cannot be identified from an individual measurement by this methodology (see e.g. [25]), at this stage we cannot resolve if the rate limiting step occurs prior or after commitment to transcription. To achieve this, we combine the data from different temperatures, as models with equivalent changes in their transcription intervals can be rejected on the basis that they would require too intricate changes in their parameters between conditions (e.g. multiple parameters co-varying in a complex manner). Using this procedure, the order of the last two steps remains unknown, but the changes in each can be identified.
To combine the data, we fit the data jointly in each of the three induction schemes, such that the parameters k on - 1, k off - 1, k 1 - 1, k 2 - 1, and k 3 - 1 are either a) constant, b) vary linearly, or c) vary quadratically as a function of temperature (the quadratic curve representing a nonlinear relationship). Meanwhile, no model is imposed on the changes as a function of induction scheme, as these are likely nonlinear. Again, model selection is used to determine which parameters change significantly with temperature.
We found the preferred model to be the one where kon, koff, k2, and k3 are constant and k 1 - 1 is a non-linear (Full and IPTG) or linear (Ara) as a function of the temperature. The ΔBICs between the second best and the best fitting models were about 1.32, 4.92, and 1.57 (ΔBIC LBs 1.31, 4.91, and 1.30) for Ara, IPTG, and Full, respectively. Meanwhile, we found similar results in models where the step with rate k3 is removed, with ΔBICs of 1.32, 4.39, and 1.51 for Ara, IPTG, and Full, respectively, indicating that using the higher-order model does affect the identification of the temperature-dependence of the parameters.
In the best fitting models, the relationship between k 1 - 1 and temperature differs under different induction schemes. However, we tested if a similar change in this parameter could explain the changes in all cases, by deriving a combined model where this parameter follows a single function, up to a scale, in the three induction schemes. We found this new model to have smaller BIC with a ΔBIC of 10.6 when compared to the best unconstrained model, (ΔBIC LB of 8.80), which provides strong evidence that the temperature affects the promoter by regulating k 1 - 1 in a similar, nonlinear fashion, regardless of the induction scheme.
Finally, we tested if the changes in k 1 - 1 could be explained solely by the changes in RNA polymerase numbers (cf. R in Eq (1)). For this, we quantified in an independent measurement the relative abundance of the RpoC subunits using western blot analysis for 24, 37, and 41°C. We found that the numbers were 0.415 and 0.562 at 24 and 41°C, respectively, relative to that of at 37°C. By plugging these numbers in our model, we found strong evidence that the parameter k1 does not follow the changes in RNA polymerase numbers (ΔBIC of 27.9, ΔBIC LB of 27.0), i.e. k ˜ 1 is not constant with temperature. Moreover, we found that if the above RNA polymerase numbers were the sole change caused by the temperature, the changes in the transcription intervals ought to be larger than what was observed in vivo. This suggests that there are other temperature-dependent changes in k ˜ 1, which attenuates the effects of RNA polymerases number changes.
The means and sds of the transcription intervals resulting from the best fitting model (quadratic k 1 - 1 with a similar pattern in each induction scheme) are shown in Fig 3. In this model, the steps prior to transcription commitment vary between temperature conditions from 3060 to 5000 s, from 2310 to 3770 s, and from 1580 to 2590 s, while the duration of the process after transcription commitment is a constant equal to 399, 7.79, and 1.50 s for Ara, IPTG, and Full induction schemes, respectively. Interestingly, in each case, the temperature-independent post-commit step tends to be no more than 12% of the total duration of the mean interval between consecutive RNA production. Consequently, as the temperature modulates the longer lasting step in transcription, it allows for large changes in the transcription interval with temperature.
To assess whether the data fit the model as expected, we also performed Monte Carlo simulations to determine if the empirical data fits the model significantly worse than that generated from the appropriate model. For this, we generated a transcription interval from the best fitting model to correspond each sample extracted from the measurements, which was subsequently censored using the same procedure that constrains the empirical data (i.e. observation time is limited by the remaining cell lifetime, and observations occur at intervals of 60 s). This simulation was repeated for 1000 rounds, from which likelihoods were calculated. The fraction of rounds where the simulated data fit worse than the empirical data (cf. a p-value) were 0.230, 0.161, and 0.040 for Ara, IPTG, and Full, respectively, and 0.095 for all cases combined. As such, we found no evidence to reject our model (typically done e.g. when this probability is less than 0.01). The mean and sd for a single simulation is shown in S1 Fig (cf. the equivalent figure for the empirical data in Fig 3).
We compared our results on the relative durations of the pre- and post-commit steps with an independent method, namely, by extrapolating the corresponding values from a Lineweaver-Burk plot [24, 32]. For this, we used qPCR to measure the target gene activity and western blot to estimate the RNA polymerase concentrations in different media conditions (see S1 Appendix for details). We found the post-commit step to take up only about 22.5% (sd 14.6%) of the transcription duration for Full induction at 37°C. This confirms the previous result, as the values reported above, derived from the microscopy measurements, are well within the 95% confidence interval (CI; [−6.08,51.1]%) of the qPCR/WB measurements.
The parameter values for the best fitting models are listed in S3 Table and some derived features of the kinetic of the on/off switching in S4 Table. These parameter values suggest that the promoter remains most of the time unavailable for RNA polymerases to bind (small duty cycle). After this, it produces a small (≪ 1) burst, whose size is temperature-dependent. The mean interval between bursts appears to be unaffected by the removal of arabinose (cf. IPTG versus Full; as kon is nearly constant and much smaller than koff), while removing IPTG has more complex effects. These results are consistent with observed changes in the median-to-mean ratio.
Two studies, one using fluorescent in situ hybridization (FISH) [41] and the other using a yellow fluorescent protein fusion library for E. coli [10] to quantify the cell-to-cell diversity in RNA numbers from several promoters, have reported Fano factors ranging from 1 to 3. Also, in the work by Golding and colleagues [7], it was estimated that the Fano factor in transcript production could be as large as 4.1. These results were taken as indicative of bursty RNA production. However, recent works have shown that much of this cell-to-cell diversity in RNA numbers can be explained by other factors [42], such as RNA degradation [43] and stochastic partitioning of RNA molecules in cell division [44]. As we base our study on the transcription intervals between consecutive RNA production events in individual cells, rather than the variability in RNA numbers, our analysis is less susceptible to such errors. In addition, only under specific conditions does a large Fano factor correspond to large burst sizes [45]. As such, we do not expect our estimate of the average burst size to be in contradiction with these studies.
Even though the Lac-Ara-1 Promoter initiation kinetics responded similarly to temperature changes under all induction schemes, this response might not generalize to different promoters with different rate-limiting step configuration. To test this, we studied the effects of temperature changes on the transcription dynamics of the PtetA promoter.
For this, using raw data from [11], we made use of the methodology in [25] in order to extract the single-cell transcription intervals of the PtetA promoter at 24 and 37°C with 15 ng of aTc (Full induction). In each case, the time series were of about 60 minutes in length with 1 minute sampling. S5 Table tabulates the mean, sd, and cv of the transcription intervals estimated from the data. The mean transcription interval changed by about a 2.5× factor.
Unlike the Plac/ara-1 promoter, the PtetA promoter appears to have sub-Poissonian dynamics, suggesting that its RNA production kinetics is mostly dictated by a sequence of multiple rate limiting steps, rather than the activation of the promoter. To confirm this, we fit the data with the model of Eq (1) and tested for on/off switching and number of rate limiting steps. The results, shown in S6 Table, provide strong evidence for two or three rate limiting steps (ΔBIC LBs against any single step model were no less than 19.0). Meanwhile, no evidence of significant on/off switching is present (ΔBICs against best on/off model are 1.14 (LB 0.387) for 24 and 5.74 (LB 5.34) for 37°C).
Finally, we fit the two data points jointly with a 3-step model, where one of the three parameters changes as a function of the temperature (more complex changes cannot be identified with only two temperature conditions). The results suggest that the two common rate limiting steps are about 121 s each in duration, while the changing step is 1910 and 543 s for 24 and 37°C, respectively. As there is no on/off switching, we cannot identify which steps are in the pre-commit and post-commit stages. It is however possible to determine that the relative duration of the post-commit step must be either 11.3% or 94.4% for 24°C and 30.9% or 84.6% for 37°C.
To determine whether the post-commit step is the longer or shorter one, we extrapolated the relative durations of the post- and pre-commit steps from a Lineweaver-Burk plot [32] based on qPCR and WB measurements. At 37°C, we found the duration of the post-commit step to be about 77.1% (sd 6.64%) of the transcription duration with a 95% CI of [64.1,90.1]%, suggesting that, in the PtetA promoter, the long, temperature-sensitive step occurs after transcription commitment.
Having measured the dynamics of transcription initiation of Plac/ara-1 and Ptet as a function of temperature and induction scheme, and assessed which parameter values are most responsible for the changes in the observed dynamics, we next make use of the model extracted from the data to investigate the ability of each parameter of the model in changing the dynamics of RNA production and the overall the range of behaviors possible with this model.
Finally, we explored, in terms of the model, why the analysis identified some parameters as the most likely candidates for being responsible for the observed changes in RNA production dynamics with temperature and induction scheme, while other parameters were identified as not fit to explain these changes. In other words, we investigate the limitations of each parameter in changing the dynamics of RNA production. In this regard, note that the fitting procedure does not equate the transcription interval mean, standard deviation, nor the coefficient of variation of the model and data in particular, but maximizes the likelihood that the selected model generates the data.
S4 Fig shows the mean, sd, and cv of the transcription intervals for the Plac/ara-1 as a function of temperature for the Full induction case, while Fig 5 shows the same variables in the relative parameter space about the parameters of the best-fitting model. The additional curves in Fig 5 detail the behavior of the mean, sd, and cv if other parameters were changing instead. For the Plac/ara-1 promoter, the mean, sd, and cv would not change significantly with changes in k2 and k3. This holds as k2 and k3 are relatively small (not rate limiting) and relatively large changes in them barely affect the transcription intervals. Meanwhile, while varying kon could result in the observed changes in the mean and sd, it could not explain the observed changes in the cv of the data. This is due to the fact that the model features a large k on - 1, implying that the noise can only be effectively tuned by varying the burst size k 1 - 1 / k off - 1. Consequently, the changes observed in the data must be explained by variations in either k off - 1 or k 1 - 1. We verified that these conclusions hold for the other model candidates as well (varying k 1 - 1, k off - 1, k 2 - 1, and k 3 - 1), as they could shift the model into different regions of the parameter space. In addition, we found the curves for IPTG to be similar, while for Ara, varying k2 can modulate mean and sd, but cannot explain the changes in cv.
Meanwhile, the corresponding curves displaying the changes in mean, sd, and cv of the transcription intervals for the PtetA promoter, are shown in S5 and S6 Figs. Here, both the mean and sd do not change significantly with changes in k1 and k3, the smaller two of the rate limiting steps. Again, this is due to the fact that these steps are relatively smaller. Interestingly, k1 and k3 could be used to manipulate cv appropriately, which is in contrast with the behavior observed in the Plac/ara-1 promoter, and enabled by the more even distribution of the transcription interval durations to the pre- and post-commit stages.
Another reason why changing the identified parameters could be preferred over changing the other feasible parameters is that the two have opposite effects as a function of increasing temperature. Such results are found both in Plac/ara-1, where koff and k1 have opposite effects on both mean and sd of the transcription intervals, and in PtetA, where k1 and k2 (or k3) have opposite effects on the cv. Alternatively, there could be some physical constraints e.g. on the parameters ranges which are not considered in our models.
We quantified how temperature affects the dynamics (rate, stochasticity, underlying steps, etc.) of the transcription initiation kinetics of the Plac/ara-1 promoter in live individual E. coli cells. This was performed in three differing induction schemes of IPTG and arabinose, for a wide range of temperatures above and below the optimal growth temperature.
For this, we used statistical methods to compare detailed stochastic models fit to the empirical data in maximum likelihood sense. The selected models inform on the most likely way the changes in RNA production kinetics with temperature and induction scheme emerge from the changes in the rate limiting steps underlying transcription initiation. To overcome the limitations of the presently available methods of observing transcription dynamics in live cells, we performed a differential analysis of several measurements under a diverse set of conditions.
We found that not all steps in transcription initiation are affected equally by temperature changes: varying only some of them suffices to explain the changes found in the measured transcription intervals, regardless of the induction scheme. Specifically, nonlinear changes in the closed complex formation alone suffice to explain the observed changes in transcription dynamics of the Plac/ara-1 promoter. By correlating these changes with variations in RNA polymerase numbers, we found that these can be only partly responsible for the observed changes in transcription dynamics, which indicates that temperature affects the interaction between the transcription start site and an RNA polymerase.
Next, we used similar methods to analyze PtetA promoter under two different temperatures. We found that this promoter rate-limiting events occur at different stages of transcription initiation, resulting in a different, less noisy transcription kinetics shape. In agreement, we found that its response to temperature is not explained by modulating the closed complex formation step as for the Plac/ara-1 promoter, but instead arises from changes in the open complex formation and/or promoter escape dynamics. Overall, this suggests that the patterns of rate-limiting step kinetics of E. coli’s promoters not only cause the genes to differ in RNA production rate and noise, but also in how responsive is their RNA production dynamics to temperature changes.
Finally, we used stochastic models to explore the possibilities of tuning transcription dynamics by varying each of the rate limiting steps. An advantage of modulating the identified steps over the other rate limiting steps was identified: it allows more flexibility in tuning both the mean and variance of the transcription intervals. In addition, in the region of the parameter space suggested by our data, the transcription dynamics is most sensitive to these particular changes, while other means of tuning the dynamics result in opposite changes in the response. This suggest that varying these particular steps might offer the promoters greater adaptability to temperature changes than if the transcription dynamics were tuned by other means.
Our study exemplifies how differential analysis of transcription intervals with statistical methods can inform on the underlying steps of transcription initiation which, at the moment, cannot be directly measured in live cells. We expect these techniques to be applicable, with small modifications only, to study similar processes such as translation or the behavior of genetic circuits.
|
10.1371/journal.pgen.1003979 | Genome-Wide Screen Reveals Replication Pathway for Quasi-Palindrome Fragility Dependent on Homologous Recombination | Inverted repeats capable of forming hairpin and cruciform structures present a threat to chromosomal integrity. They induce double strand breaks, which lead to gross chromosomal rearrangements, the hallmarks of cancers and hereditary diseases. Secondary structure formation at this motif has been proposed to be the driving force for the instability, albeit the mechanisms leading to the fragility are not well-understood. We carried out a genome-wide screen to uncover the genetic players that govern fragility of homologous and homeologous Alu quasi-palindromes in the yeast Saccharomyces cerevisiae. We found that depletion or lack of components of the DNA replication machinery, proteins involved in Fe-S cluster biogenesis, the replication-pausing checkpoint pathway, the telomere maintenance complex or the Sgs1-Top3-Rmi1 dissolvasome augment fragility at Alu-IRs. Rad51, a component of the homologous recombination pathway, was found to be required for replication arrest and breakage at the repeats specifically in replication-deficient strains. These data demonstrate that Rad51 is required for the formation of breakage-prone secondary structures in situations when replication is compromised while another mechanism operates in DSB formation in replication-proficient strains.
| Inverted repeats are found in many eukaryotic genomes including humans. They have a potential to cause chromosomal breakage and rearrangements that contribute to genome polymorphism and the development of diseases. Instability of inverted repeats is accounted for by their propensity to adopt DNA secondary structures that is negatively affected by the distance between the repeats and level of sequence divergence. However, the genetic factors that promote the abnormal structure formation or affect the ability of the repeats to break are largely unknown. Here, using a genome-wide screen we identified 38 mutants that destabilize imperfect human inverted Alu repeats and predispose them to breakage. The proteins that are required to maintain repeat stability belong to the core of the DNA replication machinery and to the accessory proteins that help replication fork to move through the difficult templates. Remarkably, when replication machinery is compromised, the proteins involved in homologous recombination promote the formation of secondary structures and replication block thereby triggering breakage at the inverted repeats. These results reveal a powerful pathway for the destabilization of chromosomes containing inverted repeats that requires the activity of homologous recombination.
| Long palindromic sequences (inverted repeats ∼100 bp or more each without a spacer or with a short spacer) present a threat to both prokaryotic and eukaryotic genome stability. In E. coli, long palindromes placed on plasmids are frequently excised and cause cell inviability when introduced to chromosome [1]. In yeast, they have been shown to drastically induce ectopic and allelic recombination and a variety of gross chromosomal rearrangements (GCRs) including deletions, translocations and gene amplification [2]–[9]. Long inverted repeats were demonstrated to undergo frequent deletions and induce gene conversion and intra-chromosomal recombination in mice [10]–[12]. Palindromic sequences have been found in the vicinity of chromosomal breakpoints of translocations in humans and are implicated in the pathogenesis of diseases. For example, palindromic AT-rich repeats (PATRRs) have been shown to induce both non-recurrent and recurrent translocations; the latter could result into Emanuel syndrome [13]–[18]. Palindrome-mediated large deletions and interchromosomal insertions are causative factors of several types of εγδβ thalassemia [19] and X-linked congenital hypertrichosis syndrome, respectively [20]. Also, palindromes are abundant in cancer cells and are associated with DNA amplification in colon and breast cancer, medulloblastoma and lymphoma [21]–[28].
Palindromic sequences can form hairpin and cruciform structures due to their intrinsic symmetry [1]. Formation of these aberrant structures has been considered to be responsible for the genetic instability associated with this sequence motif. Hairpins occurring on the lagging strand can interfere with DNA replication and be attacked by structure-specific nucleases leading to DSBs. In E. coli, hairpins formed during DNA replication at long palindromic repeats are cleaved by the SbcDC nuclease [29]–[33]. Similarly, in S. pombe, the nuclease activity of the Mre11/Rad50/Nbs1complex (Mre11/Rad50 is the homolog of SbcDC) was implicated in the generation of breaks at palindromes [8], [34]. However, Casper et al. (2009) showed that in S. cerevisiae, the Mre11 complex is not involved in breakage at a large inverted repeat consisting of two Ty1 elements with a ∼280 bp spacer in strains where DNA polymerase α was down-regulated [35]. We previously demonstrated that in S. cerevisiae, the Mre11/Rad50/Xrs2 complex does not initiate DSBs at closely spaced Alu inverted repeats (Alu-IRs) but is required along with Sae2 for processing breaks that have hairpin termini [5]. This disparity in the Mre11 complex's effect on DSB generation at palindromic sequences might be attributed to the difference in the formation of stable hairpins during replication and the inability of this complex to cleave hairpins with large loops. This conjecture, however, remains to be experimentally proven. These observations also point out the existence of an Mre11-independent pathway in generating DSBs at palindromic sequences. We proposed that in yeast, Alu-IR-mediated hairpin-capped breaks can result from the resolution of cruciform structures in which a putative nuclease cleaves symmetrically at the base of the two hairpins [5]. Cruciform resolution on plasmid in yeast was shown to be dependent on the structure-specific endonuclease Mus81/Mms4 [36], although chromosomal fragility at inverted repeats was not influenced by this complex [5]. Cruciform formation and resolution were also proposed to be the triggering events for translocations at PATRRs in human sperm cells [37]–[39]. Recently, in a plasmid transfection assay, the GEN1 nuclease was implicated in cruciform resolution in HEK293 cells, and the resultant hairpin-capped breaks were further processed by Artemis for DSB repair [40]. Whether this mechanism operates in PATRR-mediated chromosomal translocations remains to be established.
Although the formation of hairpin and cruciform structures is deemed as the key initiation event for fragility at inverted repeats, the pathways that predispose eukaryotic cells to or provide protection against chromosomal breaks are still not well defined. Previously, deficiencies in Pol1, Pol3 and Rad27 proteins responsible for synthesis of the lagging strand during DNA replication were found to augment instability at inverted repeats [3], [7], [9]. However, it is unknown if fragility is exclusively confined to deficiencies in lagging strand synthesis. In addition, it is important to identify mechanisms that facilitate or prevent instability of imperfect IRs that contain a spacer (quasi-palindrome) and are not fully homologous to each other, since these repeats prevail over perfect palindromes in the human genome [9], [41].
In this study, we carried out an unbiased genome-wide screen aimed at identifying the genetic factors controlling fragility of homologous and divergent Alu-quasi-palindromes in yeast. Using 12 bp-spaced Alu-IRs with either 100% or 94% homology between the two repeats, we analyzed the effects of deletions of around 4800 non-essential genes and downregulation of 800 essential genes on quasi-palindrome-mediated GCRs. In addition to defects in lagging strand synthesis, we found that deficiencies in proteins involved in replication initiation and leading strand synthesis, replication pausing checkpoint pathway, the Sgs1-Top3-Rmi1 dissolvasome, proteins involved in Fe-S cluster biogenesis or telomere maintenance augment breakage and GCRs induced by Alu-IRs. Replication block and fragility at inverted repeats in replication-deficient strains were abrogated upon deletion of RAD51, indicating an unexpected role for homologous recombination in the formation of cruciform structure at palindromic repeats when replication is compromised.
We systematically analyzed the effect of more than 6000 mutations on Alu-IR-mediated fragility using a genome-wide screen in the yeast S. cerevisiae (Figure 1 and Figure S1). The screen's scheme is based on the approach developed in Tong et al., 2001 [42] with modifications. In the query strains, a quasi-palindrome consisting of two 320 bp Alu elements in inverted orientation with a 12 bp spacer was placed telomere-distal to the counterselectable marker CAN1 on the left arm of chromosome V. The two Alu elements were either 100% or 94% homologous (100% Alu-IRs or 94% Alu-IRs). Breakage at the Alu-IRs and loss of the 40 kb telomere-proximal fragment results in canavanine-resistant colonies. The tester strains included a complete set of 4786 deletion mutations for non-essential genes (YKO strains) and two sets of 842 essential genes whose expression is either regulated by the doxycycline-repressible promoter (yTHC strains) or decreased due to mRNA perturbation (DAmP strains). An hphMX cassette was positioned telomere-proximal to the Alu-IRs, providing a marker for selecting the presence of the repeats during the screening and the testers were marked by a kanMX cassette. The schematics for combining the left arm of chromosome V containing the fragile motifs and the mutations have been previously applied to study instability of the trinucleotide GAA/TTC repeats and are described in detail in Zhang et al., 2012 [43]. Briefly, the query strains were crossed with each tester strain to get diploids, which then underwent sporulation. Haploids containing both the Alu-IRs and the mutation of interest were replica plated to canavanine-containing medium. Mutants with augmented repeat-induced GCRs exhibited increased number of canavanine-resistant papillae compared to the wild-type strains. Since the rate of canavanine-resistant colonies occurring due to GCR in the wild-type strain carrying 100% Alu-IRs is 10-fold higher (5×10−5) than in the strongest mutator Δmsh2 (6×10−6), the screen specifically identified hyper-fragility mutants.
We verified the effect of the identified mutants by recreating the hyper-fragile alleles in strains with the ADE2 gene inserted between CAN1 and Alu-IRs that allows differentiation of GCRs from mutations based on the color of canavanine-resistant clones [6] (Figure 1). To create the mutant alleles, the kanMX cassette was used to knockout non-essential genes and a tetO7 repressible promoter was used to replace the natural promoters of essential genes and regulate their expression [44]. The essential genes under the control of tetO7 promoter will be referred to as TET-ORFs in the following text.
38 mutants that exhibit a hyper-fragility phenotype in strains containing either 100% or 94% homologous Alu-IRs were identified from the screen (Table S1). 17 mutants belonged to the YKO collection, 17 mutants were uncovered from the yTHC collection and 4 mutants were identified from the DAmP collection. The mutants could be grouped into six classes of genes coding for the dissolvasome and proteins involved in replication, Fe-S cluster biogenesis, checkpoint response, telomere maintenance and DSB repair.
Previously, it has been shown that downregulation of or mutation in the DNA polymerases α and δ causes increased instability of inverted repeats [3], [5], [7]. Consistently, we found that TET-POL1 and TET-POL3 strains destabilize both 100% and 94% Alu-IRs and exhibit 11- to 20-fold higher fragility than the wild-type strains. This screen also revealed that downregulation or deletion of other key components of the DNA replication pathway, namely, the origin recognition complex ORC, the DNA helicase Mcm2-7, the DNA primase complex, the leading strand synthesis polymerase ε, the single-strand binding protein RPA, the polymerase sliding clamp PCNA, the clamp loader RFCs or the endonucleases Dna2 and Rad27 participating in Okazaki fragment maturation, induce fragility at Alu-IRs. Deficiencies in these proteins caused a 3- to 15-fold and a 3- to 34-fold increase in GCR rates for 100% Alu-IRs and 94% Alu-IRs, respectively. We also observed a 5- to 9-fold elevation of GCRs in strains carrying the defective replication checkpoint surveillance complex, Mrc1-Tof1-Csm3. This result prompted us to test if Mec1, which is recruited to stalled replication forks and phosphorylates Mrc1 in response to DNA replication stress [45], [46], senses inverted repeat-mediated replication impediment. Since Δmec1 is lethal, we assessed the effect of Δmec1 in Δsml1 background. We found that Δmec1Δsml1 but not Δsml1 led to a 5-fold increase in GCRs. These data demonstrate that intact replication machinery and replication checkpoint are required to prevent palindrome instability. Moreover, secondary structure formation and breakage are not only restricted to defects in lagging strand synthesis since fragility is also increased in strains where Polε and Mcm2-7 complex were downregulated.
Besides the replication checkpoint surveillance mutants, the screen also revealed that GCRs mildly increase (2- to 4-fold) in Δrad17, Δmec3, Δddc1 and Δrad24 mutants deficient in DNA damage checkpoint signaling [47]. As discussed below, this effect could be explained by the improved recovery of the broken chromosome when checkpoint activation is impaired.
The third group of mutants that amplify Alu-IRs fragility included members of the cytosolic iron-sulfur protein assembly targeting complex. TET-YHR122W led to a 3- and 8-fold increase in GCRs in 100% and 94% Alu-IRs, respectively. Yhr122w was shown to physically interact with Cia1 and Mms19 in the biogenesis of Fe-S clusters in various DNA repair and replication proteins [48], [49]. We found that disruption of MMS19 led to an 18- and 14-fold increase in GCRs in strains containing 100% and 94% Alu-IRs, respectively. This is also consistent with our previous finding that Δmms19 causes an increase in Alu-IR-induced homologous recombination [9].
The screen revealed that deletion of SGS1, the RecQ helicase homolog implicated in the dissolution of branched DNA structures and unwinding of CTG/CAG hairpins [50], [51], caused a 10- and 7-fold elevation in GCRs in 100% and 94% repeats-containing strains. Sgs1 interacts with Rmi1 and Top3 to form the dissolvasome complex [52]. Consistently, we found that deletion of RMI1 and of YLR235C that partially overlaps with TOP3 also led to hyper-fragility (Table 1 and Table S1). Our data suggest potential roles of Sgs1-Rmi1-Top3 in influencing palindrome stability through unwinding the hairpin or cruciform structures formed by the repeats.
The fifth group of hyper-fragile mutants consisted of TET-TEN1, TET-STN1 and TET-CDC13. The Ten1-Stn1-Cdc13 complex is involved in telomere maintenance and protection [53]. Downregulation of Ten1 resulted in a 3-fold elevation of fragility (Table 1). The TET-CDC13 strain demonstrated a similar increase in the level of arm loss. Notably, the closest telomere is about 40 kb away from the location of the inverted repeats. In another study, we found that downregulation of Ten1-Stn1-Cdc13 also predisposes the triplex-forming GAA/TTC repeats to breakage and expansions [43]. Taken together, these data suggest among other possibilities that this complex plays a role in helping replication machinery to move through difficult regions.
Previously, we demonstrated that the Mre11-Rad50-Xrs2 complex and the Sae2 protein are required to open hairpins to initiate DSB repair at inverted repeats [5]. We also showed that in Δmre11 mutants, GCR rates increased likely due to the inability of mutants to hold DSB ends together and open the hairpin termini, which therefore increase the probability of formation of dicentric chromosomes [6]. Predictably, the screen identified Δmre11 and Δrad50 as hyper-fragile mutants with a 10- and ∼44-fold increase in GCRs induced by homologous and homeologous Alu-IRs, correspondingly. This group therefore encompasses mutants that do not impact secondary structure formation and breakage, but rather increase probability of arm loss and recovery of the broken chromosome.
In the wild-type strain, DSBs induced by Alu-IRs have covalently-closed hairpin termini [5]. To determine if the nature of breaks in the identified hyper-GCR mutants was similar to the wild-type strain, we characterized DSB intermediates in a subset of mutants. In addition, estimation of the level of breaks provides a way to distinguish between mutants that facilitate formation or enhance stability of the secondary structures and mutants that increase the loss of the acentric DSB fragment (e.g. mrx mutants) or improve the recovery of the broken chromosome.
We compared the levels of chromosomal breaks in the wild-type strain containing 100% Alu-IRs with a subset of mutants from each group described in the previous section (Figure 2 and Figure S2). DSB detection was carried out in Δsae2 strains to prevent the opening of the hairpins and the resection of the broken fragments [5]. The lethality of Δsgs1Δsae2 can be rescued by the deletion of HDF1 [54]. Therefore, the effect of Δsgs1 on DSB formation was assessed in the Δsgs1Δsae2Δhdf1 triple mutant. DSBs were analyzed with a telomere-distal probe upon AflII digestion or a telomere-proximal probe using BglII digestion of chromosomal DNA embedded in agarose plugs. Upon AflII or BglII digestion, DSBs occurring inside the repeats were expected to be 1.3 kb or 3.3 kb, respectively. We also anticipated the appearance of inverted dimers that are double the size of the DSB intermediates (2.6 kb or 6.6 kb, correspondingly). These molecules resulting from replication of hairpin-capped breaks were previously detected in the wild-type strains [5].
No DSBs were observed in the presence of Sae2 in both wild-type and mutant strains, likely due to hairpin opening and robust resection of the breaks. However, DSBs were readily detected in Δsae2 background. In TET-POL3, TET-POL2, Δcsm3, Δsgs1Δhdf1 (Figure 2), Δmms19, TET-TEN1 (Figure S2) and Δsml1Δmec1 (Figure S4) mutants, there was a 2- to 15- fold increase in breaks in comparison with wild-type strains when the telomere-proximal or the telomere-distal fragments were probed, indicating that these mutations increase fragility at Alu-IRs by either facilitating secondary structure formation or stabilizing the structures. It is important to note that no increase in DSBs were detected in the Δsae2Δhdf1 and Δsae2Δsml1 mutants (Figure S3 and Figure S4) indicating that the increase in fragility is due to deficiencies in Sgs1 and Mec1, accordingly. In Δrad17, the amount of breaks was comparable to the wild-type strain, suggesting that DNA damage checkpoint-deficient mutants provide conditions for better recovery of the broken chromosomes, rather than affecting the formation and/or stability of the secondary structures. It is important to note that besides DSBs we could also detect dimers and no other intermediates were observed. The dependence of DSB detection on Δsae2 and the existence of dimers suggest that breaks in hyper-fragile mutants might contain hairpin termini similar to those in wild-type strains.
To test the premise of hairpin-capped breaks in the mutants experimentally, the DSB fragments in TET-POL3Δsae2 were analyzed via neutral/alkaline two-dimensional (2D) gel electrophoresis (Figure 3). We found that the 1.3 kb telomere-distal DSB fragment migrated as a 2.6 kb single-stranded DNA (ssDNA) fragment in the alkaline gel. Similarly, the 3.3 kb telomere-proximal DSB fragment migrated as a 6.6 kb ssDNA fragment under denaturing conditions. No additional bands (e.g. those corresponding to nicked hairpins) were seen, indicating that Alu-IRs generate covalently-closed hairpin-capped breaks in both TET-POL3 and wild-type strains.
The symmetry of the breaks and the presence of covalently-closed hairpins at the DSB termini suggest that the final steps in breakage in mutants and wild-type are the same and include cruciform formation and resolution.
The screen revealed that mutants deficient in the DNA replication pathway comprise the major group that augments fragility at Alu-IRs. Analysis of DSB intermediates indicated that cruciform resolution is the likely scenario for fragility in these mutants (Figure 2, 3). Generation of ssDNA due to replication defects in the leading or lagging strands might provide optimal conditions for the formation of hairpins, not cruciforms. We hypothesized that a deficiency in the DNA replication can lead to formation of the cruciform structure through template switching when the fork stalls at a hairpin. In another screen for factors that channel replication stress into fragility, we identified Rad51, a key protein in homologous recombination. In the Δrad51 background, the GCR rates of both TET-POL3 and TET-POL2 mutants decreased almost to the wild-type level (Table 2). Consistent with the reduction in GCRs, the amount of DSBs and inverted dimers in TET-POL3Δsae2 and TET-POL2Δsae2 significantly decreased upon deletion of RAD51. Notably, lack of Rad51 does not affect GCR rates or DSB formation in the wild-type strains (Table 2 and Figure 4). Consistently, deletion of RAD54, the auxiliary protein for strand invasion during recombination, in wild-type and TET-POL3 mutant had a similar effect on fragility (Table S2 and Figure S5) indicating that the involvement of homologous recombination in the induction of fragility is specific to conditions when replication is compromised.
To gain better insight into the mechanism underlying Alu-IR-induced fragility, we monitored replication progression through 100% homologous repeats in the wild-type strain and the replication-deficient mutant TET-POL3 using 2D gel electrophoresis and Southern hybridization. While replication progression was not hampered at the quasi-palindrome in the wild-type strain, the TET-POL3 mutant demonstrated a robust fork arrest at the repeats. The fact that the replication block in TET-POL3 is completely removed upon deletion of RAD51 (Figure 5) argues for Rad51-mediated template switching as the signal for replication pausing. These data, along with the observation that Δrad51 suppresses DSB formation in replication deficient strains, support the scenario where an attempt to bypass hairpin structures during compromised replication via Rad51-dependent template switching promotes the formation of cruciform structures behind the replication fork. These structures are further attacked by nucleases, resulting in DSBs (Figure 6). Although DSB formation in other hyper-fragile mutants in Δrad51 background was not analyzed, the fact that the GCR levels in these strains decreased as compared to their RAD51 counterparts strongly suggests that the same mechanism of break formation operates in these mutants (Table S2).
Overall, these data reveal an important role of homologous recombination in promoting DSB formation at inverted repeats, specifically in replication-deficient mutants.
Palindromic sequences are strong inducers of DSBs and rearrangements in both prokaryotes and eukaryotes. The two distinct events that trigger fragility are considered to be the formation of either hairpin or cruciform structures at the repeats. In this study we found that when replication is compromised, replication delay imposed by inverted repeats is channeled into cruciform resolution via the action of homologous recombination pathway. These data led us to propose that the transition from hairpin to cruciform formation through Rad51-mediated template switching is the mechanism for fragility operating in cells under replication stress.
Inverted repeat-induced GCRs can be augmented in mutants that either influence secondary structure metabolism or alter repair of the broken chromosome. Previous studies from our lab have demonstrated that Alu-IRs-induced DSBs have hairpin termini that are opened by the Mre11 complex and Sae2 to initiate resection [5]. Unprocessed hairpin-capped molecules lead to the formation of acentric and dicentric inverted chromosomes. Detailed analysis of GCR events showed that dicentric chromosomes are stabilized as a result of breakage in anaphase, followed by resection and repair preferentially via break-induced replication with non-homologous chromosomes. It is important to note that DSB resection that precedes the healing of the broken chromosome activates checkpoint signaling and is manifested as cells arrested in G2/M [6]. Previously, we found that GCR rates are elevated in mrx mutants. This increase is not due to frequent DSBs at Alu-IRs, but rather a result of more efficient formation of dicentric chromosomes and loss of the broken acentric fragments. Consistently, Δmre11, Δrad50, and Δxrs2 were identified in this genome-wide screen as hyper-fragile mutants (Table 1 and Table S1). Another group of mutants that do not increase breakage but amplify GCR rates are those defective in DNA damage checkpoint signaling (Δrad17, Δmec3, Δddc1, Δrad24) (Table 1, Table S1 and Figure S2). It is conceivable that in the absence of checkpoint activation after dicentric breakage, the rate of resection is decreased [55] and the broken chromosomes are replicated and segregated together to the daughter cells for several generations [56], [57], which improves their chances for repair.
The mutants identified in the screen that increase DSB formation and GCRs at Alu-IRs are deficient in DNA replication, replication-pausing checkpoint surveillance, Fe-S cluster biogenesis, telomere maintenance and protection, or the function of the Sgs1-Rmi1-Top3 dissolvasome. As discussed below, the impact of deficiencies in these different processes on fragility can be explained by an increase in the probability of formation or stability of secondary structures during replication.
The screen revealed that depletion of the major components of the replication fork responsible for synthesis of both leading and lagging strands increases Alu-IR-induced fragility. Our results are consistent with previous findings that mutations in the DNA polymerases α and δ promote excision of IRs and IRs-induced recombination and rearrangements [2], [3], [35], [38], [58]. It is possible that deficiencies in the synthesis of either the leading or lagging strand can lead to the generation of extensive single-stranded regions, thereby creating ideal conditions for the formation of hairpin structure, the initial event in Alu-IRs fragility (Figure 6). In replication-proficient mutants mismatches strongly suppress the fragility potential of inverted repeats which should be expected if cruciform extrusion is the initial step in breakage. However, in replication-deficient strains where transient hairpin structure probably precedes cruciform formation, mismatches in the inverted repeats are expected to have a lower impact on the formation of hairpin structure due to the presence of single stranded regions. This might explain the higher relative increase in fragility at imperfect repeats in comparison with repeats without heterology, for example in TET-RFA2, TET-POL2 and TET-PRI1 mutants (Table 1). Interestingly, downregulation of the helicase Mcm2-7 and ORC also led to hyper-fragility at the repeats. Although the MCM helicase is a part of the replication machinery, it travels ahead of the fork, therefore generation of ssDNA due to depletion of this helicase is unlikely. The effect of deficiencies in MCMs and ORC on fragility might be the consequence of the inability of the closest origin (ARS507) to fire since amounts of both protein complexes are important for regulating the timing of origin activation [59]. Replication forks traveling longer distances from the remote origins might be less processive and more prone to collapse upon encountering replication barriers. Downregulation of MCMs and ORC also increases instability at another fragile motif in yeast, the triplex-forming GAA/TTC repeats [43], indicating that this phenomenon might be universal in situations when the replication fork passes through difficult regions. Consistent with this assertion, in human cell lines that have different replication landscapes, fragility at FRA3B and FRA16D sites depends on the distance the replication fork travels [60]. Alternatively, increased fragility in mutants for MCMs and ORC might be due to the assembly of a hampered replisome that lacks components required for leading or lagging strand synthesis.
Deletion of MMS19 and downregulation of YHR122W, genes encoding proteins involved in Fe-S cluster biogenesis [48], [49], were also found to induce hyper-fragility at Alu-IRs. Recently, it has been shown that Mms19 and Yhr122W along with Cia1, are required for the transfer of Fe-S clusters to various proteins including polymerase δ DNA primase and Dna2 [48], [49], deficiencies in which were identified to augment fragility in the screen. The presence of the Fe-S clusters in the polymerases α and ε [61] and the fact that these proteins interact with Mms19 [48] also makes them likely substrates for the CIA targeting complex. The effect of mutation in this pathway on Alu-IRs-mediated fragility can therefore be attributed to the impaired maturation of the DNA replication machinery.
The deficiencies described above are expected to create optimal conditions for the formation of a hairpin that impedes replication progression. The hairpin might be formed at lower frequencies in replication-proficient cells as well. In both replication-proficient and -deficient strains, the secondary structure or the arrested fork might trigger the activation of checkpoint response required to recruit proteins to remove the hairpin and promote replication restart (Figure 6). The fact that deficiency of Mec1 and the Mrc1-Tof1-Csm3 complex leads to hyper-fragility implicates these proteins as possible sensors of secondary-structure-imposed replication arrest. However, the Mrc1-Tof1-Csm3 complex is also required to coordinate the Mcm2-7 helicase and DNA polymerase activities [53], [62]–[64], therefore, we cannot completely rule out that deficiencies in this complex affect the integrity of the replisome as well.
It seemed reasonable to suggest the existence of helicases recruited to remove hairpins at the arrested fork. Indeed, the screen identified the Sgs1-Top3-Rmi1dissolvasome. Although Δsgs1 does not affect the stability of short CAG/CTG repeats (less than 25 repeats), it increases the contraction and fragility rate of long CAG/CTG repeats (70 repeats), indicating that longer hairpins might be better substrates for Sgs1 activity [50], [65]. In addition, the Sgs1-Top3-Rmi1 complex is involved in the dissolution of double Holliday junctions [66]. Hence, it is probable that this complex also irons out long hairpins formed by Alu-IRs during replication.
An interesting group of mutants that destabilize Alu-IRs include TET-TEN1, TET-CDC13, and TET-STN1. The Cdc13-Stn1-Ten1 (CST) complex is involved in protection of chromosome ends, telomerase recruitment and telomere replication. Hyper-fragility at inverted repeats due to deficiencies in this complex can be explained by the sequestration of the Tof1-Mrc1-Csm3 complex from the replisome to the single-stranded regions at uncapped telomeres [67], [68]. Alternatively, this complex which is structurally similar to RPA [69] may facilitate replication progression through the hairpin. Dewar and Lydall, (2012) proposed that in mammalian cells the CST complex which is distributed throughout the genome [70], aside from its role in telomere metabolism, facilitates replication through difficult regions. Taking into account that downregulation of the CST complex also increases GAA/TTC-mediated fragility and expansions [43] and the physical interaction of this complex with Polα [71], [72], it is reasonable to suggest that the role of CST in DNA replication might be evolutionarily conserved.
In wild-type strains carrying inverted repeats, the deduced mechanism of breakage is cruciform-resolution by a putative nuclease that cuts symmetrically at the base of the two hairpins. This generates two hairpin-capped molecules that are present in equimolar ratios [5]. Since replication is a polar process, in replication-deficient strains, a nuclease attack on the accumulated hairpins or stalled replication fork would be expected to produce DSB intermediates different from those induced in the wild-type strains. Anticipated intermediates would include nicked hairpins, branched structures, or asymmetrical hairpin-capped breaks. Somewhat unexpectedly, we found that in the TET-POL3 strain in the absence of Sae2, the DSB intermediates were structurally identical to the replication-proficient strains: only covalently-closed hairpin-capped breaks and inverted dimers resulting from replication of the DSBs were detected. Accumulation of hairpin-capped intermediates on both sides of the break indicates that cruciform-resolution is the predominant pathway for fragility under replication stress. Since deletion of SAE2 leads to stabilization of hairpin-capped breaks in all mutants analyzed, we propose that this mechanism operates not only in the TET-POL3 strain, but also in other hyper-GCR mutants identified in the screen.
Based on our finding that deletion of RAD51 or RAD54 strongly decreases GCRs and breaks in replication-deficient strains (Table 2, Figure 4, Table S2 and Figure S5), we proposed that cruciform formation and resolution can result from the action of the homologous recombination machinery on intermediates present at the stalled replication fork. Consistent with this conjecture, replication arrest observed in TET-POL3 was also dependent on Rad51. We cannot completely rule out the possibility that homologous recombination proteins facilitate the hairpin formation. However, taking into account that Rad51 forms nucleoprotein filaments that are essential for the invasion step of homologous recombination and that Rad54 promotes strand exchange [73], we favor the explanation that Rad51 along with other components of the homologous recombination machinery promotes template switching when the replication fork encounters the hairpin structure. Synthesis of the hairpin-forming sequence on the unperturbed strand and reannealing of this newly synthesized DNA might allow formation of a cruciform structure which is resolved by a putative nuclease to give rise to hairpin-capped DSBs (Figure 6). In this case, the replication stalling observed in TET-POL3 would reflect the accumulation of arrested forks in response to template switching rather than inhibition of DNA synthesis by the hairpin structure. Rad51 was found to be present at unperturbed and stalled replication forks [74]–[77], and the involvement of recombination proteins in the fork restart and bypass of DNA lesions via template switching has been demonstrated in several studies [78]–[84]. Here, we show that the attempt of homologous recombination proteins to bypass the secondary-structure barrier may be detrimental and culminate in breaks and GCRs. We also observed that in the TET-POL2 mutant there was a stronger Rad51-dependent increase in dimer formation than in the accumulation of DSBs indicating that at least in the situation where the leading strand synthesis is compromised template switch might culminate in the inversion of the replication fork. This pathway previously described in S. pombe [80], [81], that does not generate DSBs might operate in parallel with the cruciform resolution pathway.
It is important to note that the Rad51 effect is specific in situations where replication is compromised. In replication-proficient strains, breaks and GCRs are not affected by Rad51 status, indicating that another mechanism for cruciform-formation exists. It is possible that in wild-type strains a homologous recombination-independent template switching mechanism leading to fragility operates, or that the cruciform formation is unrelated to replication. The latter hypothesis is supported by our recent finding that hairpin-capped breaks in the wild-type strain preferentially occur in G2 phase of the cell cycle (Sheng et al., in preparation).
Based on this study, we propose that in the human population, the carriers of hypomorphic alleles for the BLM-hTOPOIIIα-hRMI1-hRMI2 dissolvasome and proteins involved in DNA replication, replication-pausing checkpoint surveillance, Fe-S cluster biogenesis, telomere maintenance and protection might be susceptible to inverted repeat-induced breaks and carcinogenic GCRs. Importantly, the status of these proteins determines the stability of imperfect repeats with a spacer and divergent arms that are present in the human genome [41], [85]. At the same time, it is likely that homologous recombination can trigger chromosomal breakage at secondary structure-forming fragile sites and AT-rich palindromic sequences under conditions of replication stress. This detrimental role of homologous recombination in promoting chromosomal instability might contribute towards the development of diseases associated with fragile motifs. Homologous recombination-mediated chromosomal breakage and rearrangements might operate at secondary structure-forming fragile sites and AT-rich palindromic sequences under replication stress. This detrimental role of homologous recombination in promoting genome instability might contribute towards the development of diseases.
yTHC, DAmP and YKO collections were purchased from Open Biosystems. All other strains in this study are derivatives of BY4742 (Open Biosystems). The genotype of the query strains for the screen is: MATα, Δura3, Δleu2, Δhis3, Δlys2, rpl28-Q38K, Δ mfa1::MFA1pr-HIS3, V34205::lys2::Alu-IRs, V29617::hphMX. The 100% or 94% homologous inverted Alus were inserted into the LYS2 gene via the pop-in and pop-out method as previously described [9]. The detailed construction of the query strain can be found in Zhang et al., 2012 [43].
The effect of mutant alleles identified from the screen was verified in derivatives of YKL36 that carries the GCR assay and has the following genotype: MATa, Δbar1, Δtrp1, Δhis3, Δura3, Δleu2, Δade2, Δlys2, V34205::ADE2, lys2::Alu-IRs. To create the mutant strains, in the case of non-essential genes, the target gene was disrupted by the kanMX4 cassette [86]; in the case of essential genes, the repressible tetO7 promoter construct was PCR-amplified [44] from pCM225 (Euroscarf) and was used to replace the natural promoter of the gene to create the TET-alleles (Table S3).
In strains used for DSB analysis, SAE2 was disrupted by TRP1. For construction of the Δsgs1Δhdf1Δsae2 triple mutant, SGS1 was disrupted by the kanMX4 cassette, and HDF1 was knocked out by the hphMX cassette [87] (Table S3). To study the effect of RAD51 on Alu-IRs-mediated fragility, RAD51 was replaced by a hisG-URA3-hisG cassette [88].
The screen was carried out as described in Zhang et al., 2012 [43].
Yeast cells were grown on YPD plates for 3 days. For each strain, a minimum of 14 independent colonies were taken to perform fluctuation test to estimate GCR rates. Appropriate dilutions of cells were plated on YPD and canavanine-containing plates to determine the GCR frequency. The GCR rates were calculated using the formula μ = f/ln(Nμ) as described in Drake, 1991 [89]. 95% confidence intervals were calculated as described in Dixon, 1969 [90]. The canavanine-containing plates used for tests were made from arginine-drop out medium with low amount of adenine (5 mg/L) and supplemented with L-canavanine (60 mg/L).
Yeast cells from overnight cultures were embedded into 0.8% low-melting agarose plugs at a concentration of 24×108 cells/ml. The plugs were treated with 1.5 mg/ml lyticase for 3 hr, followed by overnight 1 mg/ml proteinase K treatment. For restriction digestion of the DNA, the plugs were washed twice with 1 X TE buffer (10 mM Tris-Cl [pH 8.0], 0.1 mM EDTA) for 30 min, treated with 1 mM PMSF for 1 hr, washed with distilled water for 1 hr and equilibrated with restriction buffer for 20 min. Each plug (∼40 µl) was digested with 50 units of AflII or BglII for 16 hr. Digested plugs were loaded in a 1% (AflII digestion) or 0.7% (BglII digestion) agarose gel, respectively, and run in 1 X TBE for 18 hr. The gels were treated with 0.25 N HCl for 20 min, alkaline buffer (1.5 M NaCl, 0.5 M NaOH) for 30 min and neutralization buffer (1.5 M NaCl, 1 M Tris [pH 7.5]) for 30 min. The gels were then transferred in 10 X SSC to charged nylon membrane for 2 hr through a Posiblotter (Stratagene). Southern hybridization was carried out using P32-labeled LYS2-specific probes at 67°C overnight. DNA membranes were washed twice for 15 min each in buffer containing 0.1% SDS and 1% SSC and the signals were detected by the typhoon phosphoimager (GE Healthcare Life Sciences). The hybridization signals were quantified using ImageJ software (NIH).
Yeast plugs were prepared and digested as described above. Neutral/neutral and neutral/alkaline gel analysis was performed as previously described with small modifications [5], [91]. In the first dimensional gel electrophoresis, the plugs were loaded in a 1% (AflII digestion) or 0.7% agarose (BglII digestion) gel, respectively, and run for 18 hr in 1 X TBE. The gel slices containing the bands of interest were then cut out for the second dimensional gel electrophoresis. For neutral/neutral gel analysis, the gel slices were loaded in 1% (AflII digestion) or 0.7% (BglII digestion) agarose gel made in 1 X TBE, run in 1 X TBE for 18 hr at 1.7 V/cm and then processed for Southern hybridization. For neutral/alkaline gel, the gel slices were treated with 10 mM EDTA for 30 min, 5 mM EDTA for 30 min and embedded in agarose gel made in buffer containing 50 mM NaCl, 1 mM EDTA. Next, the gels were soaked in 5 X alkaline buffer for 30 min, 1 X alkaline buffer (50 mM NaOH, 1 mM EDTA) for 30 min and cooled down in 1 X alkaline buffer at 4°C for 15 min. The gels were then run in 1 X alkaline buffer at 0.7 V/cm for 40 hr at 4°C and processed for Southern hybridization.
2D gel analysis was carried out as previously described in Brewer and Fangman, 1987 [92]. Overnight yeast cultures were synchronized in G1 with alpha factor (50 µg/107 cells) at OD600 = 0.8. 2 µg/ml doxycycline was added to the cultures to downregulate Polδ in the case of TET-POL3 and TET-POL3Δrad51 strains. Cells were then released into fresh YPD. 50 min after release, wild-type, TET-POL3, TET-POL3Δrad51 strains were harvested and their genomic DNA samples were prepared as described in Friedman and Brewer, 1995 [93]. For the first dimensional gel electrophoresis, AflII digested DNA samples were loaded in a 0.4% agarose gel and run in 1 X TBE at 1.7 V/cm for 22 hr. For the second dimensional gel electrophoresis, gel slices containing bands of interest were cut out and loaded into a 1.2% agarose gel supplemented with 0.3 mg/ml ethidium bromide. The gels were run in 1 X TBE at 6 V/cm for 11 hr. Gels were then processed for Southern hybridization. Images were quantified using ImageQuant TL software (GE Healthcare Life Sciences).
|
10.1371/journal.pntd.0004716 | Cathepsin B-Deficient Mice Resolve Leishmania major Inflammation Faster in a T Cell-Dependent Manner | A critical role for intracellular TLR9 has been described in recognition and host resistance to Leishmania parasites. As TLR9 requires endolysosomal proteolytic cleavage to achieve signaling functionality, we investigated the contribution of different proteases like asparagine endopeptidase (AEP) or cysteine protease cathepsins B (CatB), L (CatL) and S (CatS) to host resistance during Leishmania major (L. major) infection in C57BL/6 (WT) mice and whether they would impact on TLR9 signaling. Unlike TLR9-/-, which are more susceptible to infection, AEP-/-, CatL-/- and CatS-/- mice are as resistant to L. major infection as WT mice, suggesting that these proteases are not individually involved in TLR9 processing. Interestingly, we observed that CatB-/- mice resolve L. major lesions significantly faster than WT mice, however we did not find evidence for an involvement of CatB on either TLR9-dependent or independent cytokine responses of dendritic cells and macrophages or in the innate immune response to L. major infection. We also found no difference in antigen presenting capacity. We observed a more precocious development of T helper 1 responses accompanied by a faster decline of inflammation, resulting in resolution of footpad inflammation, reduced IFNγ levels and decreased parasite burden. Adoptive transfer experiments into alymphoid RAG2-/-γc-/- mice allowed us to identify CD3+ T cells as responsible for the immune advantage of CatB-/- mice towards L. major. In vitro data confirmed the T cell intrinsic differences between CatB-/- mice and WT. Our study brings forth a yet unappreciated role for CatB in regulating T cell responses during L. major infection.
| Cutaneous forms of leishmaniasis are characterized by lesions that progress over months or years and that often leave permanent scars. Toll like receptors play an important role in the recognition and initiation of immune responses, and the intracellular TLR9, a sensor of pathogen double-stranded DNA, plays a crucial role in host resistance to Leishmania parasites. To achieve functionality, proteolytic enzymes, like cathepsins B, L, or S or asparagine endopeptidase, must cleave TLR9. Using mice deficient for different cathepsins, we demonstrate that these cathepsins do not seem to be individually involved in TLR9 processing. Interestingly we observed that Cathepsin B-deficient mice were more resistant to Leishmania major infection, meaning they resolve lesions and reduce parasite burdens faster than wild-type C57BL/6 mice. We found that this resistance is based on adaptive rather than innate immunity, with a central role of Cathepsin B-deficient T cells that contribute to faster controls of leishmania probably by higher IFNγ production. Cathepsin B inhibitors were already shown to have favorable effect in leishmaniasis, but the mechanisms behind these effects remain unclear. Our study highlights a new role for cathepsin B at the T cell level and provides new clues to how targeting this molecule is beneficial for treating leishmania infections.
| A protective immune response against intracellular protozoan parasites of the genus Leishmania is characterized by the development of IFNγ-producing T cells. This supports macrophages in the induction of anti-leishmanial effector functions, such as production of nitric oxide [1,2]. IL-12, a cytokine produced largely by antigen-presenting cells (APCs), such as dendritic cells (DCs), contributes to immunity against Leishmania major (L. major) by both polarizing and supporting T helper (Th) 1 responses [3]. The capacity of DCs to produce IL-12 is directly conditioned by the recognition of pathogen associated molecular patterns (PAMPs). This is achieved through a variety of receptors, of which Toll-like receptors (TLRs) are by far the best characterized [4,5]. A large body of knowledge has been accumulated on the recognition of Leishmania by different TLRs [6,7]. We, and others, have previously described a critical role for intracellular TLR9, a sensor of pathogen double-stranded DNA, in recognition and host resistance to Leishmania parasites [8–12]. TLR9 requires a proteolytic cleavage step inside the endolysosome to achieve signaling functionality. TLR9 maturation was proposed to be a multistep process requiring, among other molecules, the contribution of asparagine endopeptidase (AEP) and other cysteine proteases such as cathepsins B (CatB), L (CatL) or S (CatS) [13–16]. Although analysis of TLR9 processing and signaling supported a role for both cathepsins and AEP in macrophages and DCs, there is no consensus on their contribution to TLR9 maturation and its consequences on innate immunity. In Leishmania infection, despite the known importance of DCs in polarizing Th responses and the role of cysteine proteases in modulating DC functions, the role of these proteins remains poorly understood. The importance of the Th1/Th2 balance for protective immunity in leishmaniasis is clearly illustrated by the susceptibility of the prototypical Th2 BALB/c mouse strain as opposed to the resistance of Th1-prone C57BL/6 or DBA/2 mice [1].
Cathepsins have been the subject of a few studies during Leishmania infection and use of specific inhibitors has been shown to skew Th responses [17–19]. Inhibiting CatB was suggested to favor the development of protective Th1 responses in BALB/c but not in DBA/2, whereas inhibition of CatL exacerbated the disease in both BALB/c and DBA/2 mice [18,19]. Still, further research is needed to elucidate the mechanisms behind these effects. The role of AEP and CatS however, has not been investigated in L. major infection. We thus set out to investigate how AEP, CatB, CatL and CatS affect the immune response during L. major infection with the aim of assessing if TLR9-dependent responses are affected by these proteases. Comparing mice deficient in AEP, CatB, CatL and CatS, we observed that only CatB-deficient mice were more resistant to infection, meaning they resolve lesions and reduce parasite burdens faster than C57BL/6 (WT) mice. We went on to test the involvement of CatB in TLR9-dependent DC responses and found no impact of CatB. We performed a thorough analysis of the local immune responses to L. major and found no major differences in the innate immune response to infection between WT and CatB-/- mice. We however found a clear difference in the Th1 response between the two mouse strains, observing a somewhat faster apparition and a consistently speedier decline of IFNγ levels in CatB-/- mice. Having observed no differences in innate responses or antigen presenting capacity of DCs we investigated whether CatB might affect lymphocyte populations, as previous studies have suggested [20–23]. Using a series of adoptive transfer and in vitro experiments we found that CD3+ T are intrinsically different in CatB mice and can confer an immune advantage to recipient alymphoid mice upon transfer, as compared to their WT counterparts. To our knowledge this is the first study to identify a T cell-intrinsic role for CatB in L. major infection and thus setting the stage for a new direction of research into the role of cysteine proteases in Leishmania protozoan infections.
Animals were housed in the Institut Pasteur animal facilities accredited by the French Ministry of Agriculture to perform experiments on mice in compliance of the French and European regulations on care and protection of the Laboratory Animals (EC Directive 86/609, French Law 2001–486 issued on June 6, 2001). The CETEA (Comité d'Ethique pour l'Expérimentation Animale—Ethics Committee for Animal Experimentation) “Paris Centre et Sud" reviewed and approved the animal care and use protocol under the approval number 2012–0059.
Six to 8 week-old female C57BL/6 mice were purchased from Charles River Laboratories. TLR9-/- mice, backcrossed to the C57BL/6 background for at least 10 generations were provided by S. Akira (Osaka University, Osaka, Japan). Mice deficient in asparagine endopeptidase, cathepsin B, cathepsin L, and cathepsin S, as previously described [24–27], and backcrossed to the C57BL/6 background for at least 10 generations were supplied by Dr. B. Manoury (Hospital Necker, Paris). These mice were bred in our facilities and housed under specific pathogen-free conditions. Genotypes of deficient mice were controlled by PCR on their genomic DNA. Rag2-/-γc-/- mice were kindly provided by Dr. A. Galgano (IP Paris). Promastigotes of L. major (LV39) were propagated in vitro in M199 medium supplemented by 10% of foetal calf serum (FCS) (Biowest, France). Killed or live L. major promastigotes were in stationary phase; the killed L. major promastigotes were heat killed at 56°C for 30 min.
For infection, mice were inoculated subcutaneously (s.c.) into the footpads (FPs) with 3x106 stationary phase L. major promastigotes. The size of the resulting lesions was assessed weekly by measuring the footpad size with a metric caliper (Kroeplin, Germany) and comparing it with the footpad thickness before infection. Each week after infection, blood was collected by retro-orbital bleeding for seric determinations, the draining lymph nodes and the footpads were harvested for cellular analysis and tissue parasite burden, determined by limiting dilution analysis as previously described [28].
Bone marrow derived dendritic cells (BMDCs) and bone marrow derived macrophages (BMDMs) were differentiated from bone marrow cells, obtained from the mice detailed above, as previously described [29]. In brief, cells were grown in complete RPMI 1640 with GlutaMAX (Gibco) with 10% FCS (Biowest), 1mM sodium pyruvate, HEPES 10mM, MEM non-essential amino acids, 2Mercaptoethanol 0,05mM and gentamicin 0,02mg/ml (all from Gibco). BMDCS and BMDMs were differentiated with GM-CSF from J558L cell line supernatants [30] or M-CSF from L929 fibroblast cell line supernatants, respectively. At day 8 and 7, respectively, BMDCs were 80% CD11c+CD11b+ and BMDMs were 95% CD11c-CD11b+. Total cells were harvested and cultured in 6-well plates (3x106/well) and stimulated for 6 h with killed or live L. major promastigotes at a parasite/cell ratio of 5:1, or 250 ng/ml of CpG 1826 or 100ng/ml of conventional LPS (E. coli, O111:B4, Sigma). Cells were harvested at 6h after stimulation for RNA extraction. At 24h, cell-free supernatants were harvested for cytokine determination by specific ELISA.
Cells were isolated from footpads, lymph nodes and spleens of naive or L. major-infected mice at various times after infection. The footpad tissue was cut into small pieces that were gently separated with a pestle and digested with liberase 1,5mg/ml in PBS at 37° for 1 h. The reaction was stopped with FCS 10% in RPMI. Then in a second step cells were separated from the remaining tissue by a digestion with collagenase type 4 (0.5 mg/ml) and DNase type I (40 ug/ml) (Boehringer-Mannheim) for 30 min stopped with PBS with 10% FCS. Cells were then washed, filtered through a 70μm cell strainer and used for FACS analysis. Single cell suspensions from draining popliteal lymph nodes or spleens were prepared by crushing organs through a 70μm cell strainer before staining for FACS analysis.
CD3 or CD4 T cells were isolated from the spleens of C57BL/6 and CatB-/- mice and washed in PBS. CD3 or CD4 isolation was achieved by specific negative enrichment kit (eBioscience and Miltenyi Biotec, respectively) according to manufacturer’s protocol. Purity of isolated cells was verified by flow cytometry and was always around 95%. To study the proliferation of CD4+ T cells, 2x106/ml purified cells were labelled with carboxyfluorescein succinimidyl ester (CFSE) at the concentration of 5μM/ml in PBS without any trace of EDTA or protein at 37°C during 10 minutes. The reaction was stopped with culture medium with 10% FCS and the labeling was controlled by FACS. The cells were resuspended in complete RPMI medium at 2x105/ml, anti-CD28 was added at (2,5 μg/ml) and the cells were plated on anti-CD3 or isotype-coated wells (5 μg/ml) with and incubated at 37°in CO2 5% incubator for 72h. CFSE dilutions were analysed by FACS. Naïve CD3+ purified cells from WT and CatB-/- were incubated at the concentration 2x105/ml at 37°C in complete medium with phorbol 12-myristate 13-acetate (PMA) 200 ng/ml and 500 ng/ml of ionomycine. Supernatants were harvested at 24, 48, and 72h and cytokine production was assessed by specific ELISA.
IL-1β, IL-6, TNF, IL-10, CXCL1, CCL2 and CCL5 were quantified in cell culture or footpad supernatants using ELISA kits (BD Biosciences). Weight-matched footpad tissues were dilacerated and incubated in 400 μl of PBS with protease inhibitor, 4 hours at room temperature to allow cytokines diffusion before harvesting the supernatants.
All ELISA procedures were performed according to the manufacturer’s protocol.
RNA from cells or from lymph-nodes and footpads tissue was extracted using respectively a microRNeasy extraction kit (Qiagen) or TRIzol Reagent (Thermo Fisher Scientific), according to the manufacturer’s protocols. RNA (2 μg) was reverse transcribed using (200 U) Moloney murine leukemia virus reverse transcriptase (SuperScript II, Invitrogen). Subsequent real time PCR was performed on Step One Plus (Applied Biosystems) using Taq polymerase (Taq-Man Universal PCR master mix, Applied biosystems). All PCR data values were normalized to the expression of the hypoxanthine phosphoribosyltransferase (HPRT) gene.
Single cell suspensions from footpads, draining lymph nodes or spleens were surface stained for phenotypic analysis. Cells were incubated for 30 minutes with appropriate amounts of antibodies in the presence of Fc receptor-blocking agent (Fc block from BD Bioscience), after which cells were washed and stained with a viability dye (eBioscience). Antibodies used were directed against mouse CD45 (GL2), CD11c (HL3), CD40 (3/23), CD3 (145-2CII), CD4 (RM4-5), CD8 (53–6.7), CD25 (PC65), CD19 (ID3), CD11b (M1/70), Ly6G (1A8-Ly6G), Ly6C (HK1.4) and NK1.1 (PK136) from BD Biosciences and eBioscience. For intracellular staining, surface-labeled cells were fixed and permeabilized with the Inside Stain Kit from Miltenyi Biotec for IFNγ labelling with an anti-IFNγ (XMG1.2) or fixed and permeabilized with the transcription factor permeabilization kit from eBioscience and stained for foxp3 (MF23 from BD Biosciences). All controls were stained with the respective isotypes. Flow cytometric data were acquired on a MACSQuant device (Miltenyi Biotec) and analysed using FlowJo software (TreeStar).
Soluble leishmanial antigens (SLA) were prepared using 109/ml stationary phase promastigotes of L. major as previously described [19]. The aliquots were stored at -800 and thawed only one time. SLA (2ug/ml in PBS) were covalently coated on to 96 well plates (Nunc Maxisorp Roskilde, Denmark) for 2 hours at room temperature and overnight at 4°C. After blocking with 10% FCS in PBS, the wells were incubated with serum samples serially diluted in the blocking buffer for 2 hours at room temperature. After washing the plates were incubated for 90 min at room temperature with horseradish peroxidase (HRP) -Goat anti-mouse IgG (H+L) (Invitrogen) diluted at 1/5000 or -Goat anti-Mouse IgG1 diluted at 1/4000 or -Goat Anti-Mouse IgG2a diluted at 1/1000 (both from Human ads-HRP Southern Biotech). After washing with Tween 0,1%, BSA 1%, bound antibodies were revealed with TMB (TMB Substrate Reagent Set from BD) and the colorimetric reaction was read at 450nm.
Spleen cells were isolated from C57BL/6 and CatB-/- mice and washed in PBS. Cell preparation were either kept on ice or processed for CD3 cell isolation before adoptive transfer. CD3 isolation was achieved by negative enrichment kit (eBioscience) according to manufacturer’s protocol. Purity of isolated cells was verified by flow cytometry and was always >94%. Splenocytes or purified CD3 cells were resuspended at a concentration of 106 cells/100 μl PBS and adoptively transferred by retro-orbital intravenous (i.v.) injection of 100 μl into recipient Rag2-/-γc-/- mice. On day 7 after adoptive transfer, mice were inoculated s.c. into the footpads with 6x106 stationary phase L. major promastigotes and lesions size was monitored as described above. The in vivo proliferation of T cells was done in Rag2-/-γc-/- mice, 21 days after CD3 transfer as described above but without any infection. On day 21, each mouse was injected in intrapertoneal (i.p.) with 1 mg of 5-Bromo-2′-deoxyuridine (BrdU), 12h later a second injection of 1 mg BrdU was performed, and at 16h a third injection [31]. One hour after the last injection, spleens were harvested and BrdU labelling was performed according to manufacturer’s instructions using the BrdU flow kit from BD Pharmingen.
Cells from draining lymph nodes of C57BL/6 or CatB-/- mice were harvested 21 or 28 days after L. major inoculation. Lymph node single cells suspensions or purified CD4+ T cells, at a purity of around 95%, were obtained by negative selection with a CD4+ T cell isolation kit (Miltenyi Biotec) using manufacturer’s instructions, then used as stimulus target cells. BMDCs from C57BL/6 or CatB-/- mice, 2,5x105 were cultured in 24 well plates at 37°C and 5% CO2 using complete RPMI 1640 and stimulated with 1,25x106 living or killed L. major promastigotes for 6h. After three washes in RPMI to remove free parasites, lymph node cells or purified CD4+ T cells were co-cultured with BMDCs at a ratio of 5:1 for 24h. For intracellular labelling of IFNγ, brefeldin A (5 μg/ml) was added for the last 4h of culture. Supernatants were harvested and frozen at -20°C for ELISA assays. Cells were used for intracellular detection of IFNγ by FACS, as described above. Controls included lymphocytes cultured alone, naïve splenic T cells, and co-cultures stimulated with PMA-ionomycine at 200 ng/ml and 500ng /ml, respectively.
Statistical significance was tested using Prism 5.0 Software (GraphPad). student t-test and 2-way ANOVA comparisons with post-hoc Bonferroni tests were used as statistical tests. Error bars in all figures represent SEM, with the midlines representing the mean value in scatter plots.
To investigate the involvement of AEP and cysteine proteases such as CatB, L and S on the development of cutaneous lesions following infection with L. major, we compared the footpad thickness between WT, AEP-/-, CatB-/-, CatL-/-, CatS-/- and TLR9-/- mice over a time-course of 49 days. Unlike TLR9-/- mice, which develop significantly larger lesions and higher parasite burdens, as we previously reported [8], AEP-/-, and CatL-/-mice had similar lesion sizes and parasite burdens as compared to WT mice (Fig 1A–1C and S1 Fig). Noteworthy, CatS-/- mice were slightly more susceptible to infection than WT mice (Fig 1D). These results suggest that AEP, CatL and CatS are not individually implicated in the maturation of TLR9 and do not play a major role in L. major infection. CatB-/- mice, however showed a better control of infection with L. major, with footpad lesions and parasite burden in the draining lymph nodes regressing significantly faster from day 21 post-infection (p.i.) as compared to WT mice (Fig 1E and 1F). Considering the different course of infection dynamics between CatB-/-, WT and TLR9-/-, we questioned the role of this protease in the immune response to L.major infection and if it could affect TLR9 activation. Previously, we showed the importance of TLR9 for the activation of DCs by L.major DNA [8]. We used L. major, CpG and LPS to test the capacity of the BMDCs from WT and CatB-/- mice to be activated by TLR9-dependent or independent stimuli. Stimulation of BMDCs from both strains of mice with CpG and LPS induced similar levels of TNF and IL-6 (Fig 2A). To assess whether cathepsin B expression is involved in the pivotal role of DCs and macrophages during Leishmania infection, we also stimulated these cells with heat-killed or live L. major promastigotes and found there was no significant difference in cytokine responses to L. major between cells from CatB-/- and WT (Fig 2B). The same results for TNF were obtained with BMDMs from WT or CatB-/- mice (S2A Fig). Additionally, we assessed the expression of IL-12p40, IL-12p35, IL-6, IFNβ, IL-10 and TNF transcripts in BMDCs and BMDMs from WT and CatB-/- mice and found no significant differences in response patterns to either L. major, CpG or LPS (Fig 2C and 2D, S2B and S2C Fig). These results suggest that the differences in control of cutaneous L.major infection observed between WT and CatB-/- mice most likely do not involve DCs or macrophages and exclude a role for CatB in TLR9-dependent responses in these cells.
In order to investigate the nature of the immune response during the course of L. major infection in the two strains of mice we performed a serial analysis of multiple parameters. We first analyzed the levels of key chemokines like CXCL1 (KC), that promotes neutrophil recruitment, as well as CCL2 (MCP1) and CCL5 (RANTES), both involved in recruitment and activation of monocytes and lymphocytes, in the footpads (FPs) of infected mice. As shown in S3 Fig, ELISA analysis confirmed that the production of chemokines was higher in WT mice at days 21 and 28 p.i.. These data point to a loosening of chemokine gradients in FPs of infected CatB-/- as compared to WT mice, which correlates with faster resolution. We also studied innate immune cell recruitment to the infected FPs by FACS (gating strategy depicted in Fig 3A). We found similar proportions of CD45+ cells in FPs of WT and CatB-/- mice. Among CD45+ cells we observed similar proportions of CD11b+CD11c-Ly6Chigh monocytes and higher frequencies of CD11b+CD11c-Ly6Cint macrophages in CatB-/- mice, compared to WT mice, but only at late time points (Fig 3B–3D). In contrast, we found differences at the CD11b+CD11c-Ly6Ghigh neutrophil level, with significantly higher percentages of cells persisting after day 21 in WT mice (Fig 3E). We did not find differences between the percentages of macrophages, monocytes and neutrophils in the draining popliteal lymph nodes (dLNs) of WT or CatB-/- mice (S4A–S4C Fig). Another innate cell population we analyzed in the dLNs were natural killer (NK) cells but again, we did not find significant differences between WT and CatB-/- mice (S4D Fig). We also investigated DC content and found no significant time-stable differences in percentages of either CD11c+CD11b+ or CD11c+CD11b- DCs in FPs or dLNs of WT and CatB-/- mice (S4F–S4H Fig). Moving further, in order to decipher the local inflammatory profile governing the anti-leishmania response we determined the levels of TNF and IL-1β in the FPs. We found significantly lower levels of TNF and IL-1β in FPs of CatB-/- mice at day 21 and 28 or day 28 and 35, respectively (Fig 4A and 4B). We also studied the level of expression of TNF and IL-6 mRNA in FPs and dLNs of infected mice. While we found no differences in TNF and IL-6 mRNA expression in the FPs during the whole observation period (Fig 4C and 4D), we did find significantly higher transcript levels for both cytokines in dLNs of CatB-/- mice at day 7 p.i. as compared to WT (Fig 4E and 4F). Starting from day 21, inflammation significantly decreases in comparison to WT mice to promote parasite clearance and decrease of lesion size.
Because proliferation of lymphocytes in dLNs remains an absolute requirement for adaptive immunity in cutaneous leishmaniasis, we studied the dynamics of T and B lymphocytes in WT and CatB-/- mice during the course of infection (gating strategy in Fig 5A). Cell numbers in dLNs, the majority of which are lymphocytic, increase dramatically during the course of infection. Especially in WT mice, dLN cell numbers were 50 to 100 times higher compared to uninfected mice. These numbers peaked at day 28 in WT mice and slowly decreased thereafter (Fig 5B). In contrast, cellularity of dLNs in CatB-/- mice peaked at day 21 and decreased at a faster rate compared to WT mice. Phenotypically, we observed a constantly lower percentage of CD19+ B cells in the dLNs of CatB-/- mice reaching significance from day 21 (Fig 5C). While CD3+ T cell percentages were not different between the two strains of mice, CatB-/- mice had higher CD4+ T cell percentages compared to WT (Fig 5D–5F). Among CD4+ T cells, we observed significant lower foxp3+CD25+ regulatory T cells (Tregs) in CatB-/- from day 21 onward (Fig 5G). To complete our investigation of adaptive responses during L. major infection, we explored the dynamics of T cell related cytokines in dLNs and FPs of infected mice. We compared the expression profile of Th1 (IFNγ) and Th2-associated (IL-4 and IL-10) mRNAs in dLNs of WT and CatB-/- mice during infection since we found no detectable transcripts in the FPs. While we observe increased IFNγ expression at early moments in CatB-/- mice, this trend is quickly lost past day 21 p.i. reaching significantly lower levels compared to WT from day 28 p.i. onward (Fig 6A). Interestingly, this tendency of CatB-/- to express higher levels of IFNγ like at day 14 p.i. was corroborated by a significantly lower expression of IL-4 transcripts in these mice at day 7 p.i. (Fig 6B). IL-4 expression further decreased and showed no significant differences during the rest of the observation period. We also observed a lower IL-10 transcript expression in CatB-/- mice around day 21–28 p.i. (Fig 6C) Since the IFNγ-inducible Nitric Oxide Synthase (iNOS) axis is well recognized as an effector mechanism in leishmaniasis we also monitored the levels of iNOS transcripts in dLNs and FPs of infected mice. CatB-/- mice had significantly lower levels of iNOS in the dLNs compared to WT from day 21 to 28 p.i. and a similar tendency further on (Fig 6D). Interestingly, iNOS transcripts in the FPs were highly upregulated throughout the infection (S5A Fig), with differences between mouse strains showing a similar trend as like in dLNs but without reaching significance.
Additionally, we performed a serial analysis of soluble Leishmania antigen (SLA)-specific immunoglobulins and observed an increased level of total IgGs, as well as increased levels of IgG1 and IgG2a in sera of CatB-/- mice as compared to WT, with differences reaching statistical significance at days 28 and 35 (Fig 6E and 6F). The IgG2a/IgG1 ratio however showed no significant differences between WT and CatB-/- mice throughout the observation period (S5B Fig). Altogether these data point to a faster and more efficient induction of a protective Th1 response in CatB-/- as compared to WT accompanied by a faster transit to a resolution phase which correlates with a lower presence of Tregs and IL-10.
While previous studies hypothesized that inhibition of CatB during L. major infection would lead to changes in Th cell polarization due to difference in antigen presentation [18,19], we sought to investigate whether L.major activated WT and CatB-/- DCs had the same ability to stimulate antigen-specific T cells from L. major infected mice. To this aim, we took dLNs from day 21 and 28 infected mice and we co-cultured these cells with L. major antigen-loaded BMDCs to determine IFNγ production by specific ELISA and intracellular staining. As shown in Fig 7A–7D, we observed similar IFNγ production regardless of the origin of the BMDCs used for stimulation. To restrict antigen presentation to class II MHC, we performed similar experiments using purified CD4+ T cells from dLNs at days 21 and 28 p.i. and again observed no BMDC-dependent differences between antigen-presentation capacity of WT or CatB-/- mice (Fig 7E–7H). We observed no detectable IFNγ secretion under control conditions with un-stimulated BMDCs or naïve CD4+ T cells. These data indicate that DCs from WT and CatB-/- mice are equally capable of presenting L. major antigens in a MHC II context, suggesting that antigen presentation does not likely contribute to the differences seen in infection clearance between the two strains.
Data presented so far do not support a role for APCs in the observed differences between CatB-/- and WT mice during L. major infection but rather one for lymphocytes, given the more efficient induction of a protective Th1 response and rapid decline of inflammatory cells in CatB-/- dLNs. To address this issue specifically we reconstituted RAG2-/-γc-/- mice with splenocytes (SplC) from either WT or CatB-/- mice (RAG2-/-γc-/- + WT or RAG2-/-γc-/- + CatB-/- SplC, respectively) and measured FP thickness after infection with L. major. RAG2-/-γc-/- + CatB-/- SplC showed less FP swelling compared with mice reconstituted with WT cells at days 21–35 p.i. (Fig 8A). These data argued for predominant role of lymphocytes in the immune advantage of CatB-/- mice during L. major infection, but could not pinpoint a specific cell population in cause. We thus went further and repeated the reconstitution experiments with purified CD3+ T cells (RAG2-/-γc-/- + WT or RAG2-/-γc-/- + CatB-/- CD3s, respectively). Similar to the splenocyte transfer experiments, we found that RAG2-/-γc-/- + CatB-/- CD3s had milder FP swelling compared to RAG2-/-γc-/- + WT CD3s arguing for a T cell-related difference (Fig 8B). Since RAG2-/-γc-/-mice do not have dLNs and leishmania parasites disseminate systemically in these mice [32], we assessed cellularity and parasite burdens in the spleens. At day 42 p.i. the number of viable parasites in spleens from RAG2-/-γc-/- + CatB-/- CD3s was 1–2 orders of magnitude lower than in RAG2-/-γc-/- + WT CD3s, whereas in the FP the differences were more than 3–4 orders of magnitude less in the same direction reaching statistical significance (Fig 8C). The cellular composition of the spleens in reconstituted RAG2-/-γc-/- mice reflected the origin of the transferred cells, but also mirrored the situation found in WT or CatB-/- mice post infection, with more CD4 and less CD25+CD4+ cells (Fig 8D). Interestingly, the spleens of naïve mice showed that CD8 cells are underrepresented in CatB-/- mice (S6 Fig), and we observed a significantly lower percentage of CD8 T cells in RAG2-/-γc-/- + CatB-/- CD3s.
Additionally in order to identify the T cell-intrinsic differences that could account for the observed phenotypes, we assessed cytokine-production and proliferation capacity of isolated T cells in vitro. Following stimulation with PMA, CatB-/- T cells produced significantly more IFNγ than WT T cells but significantly less IL-2 (Fig 9A and 9B). To assess proliferative capacity, we stimulated CFSE-labeled purified T cells with anti-CD3 for 72 hours and analyzed CFSE dilutions by FACS. As depicted in Fig 9C, CatB-/- CD4 cells have a lower proliferative capacity at 72h as compared to WT T cells. Moreover, we assessed the in vivo proliferation capacity purified T cells by measuring BrdU incorporation at 21 days after transfer into RAG2-/-γc-/- mice. We found significantly higher incorporation of BrdU in WT CD4 cells as compared to CatB-/-, but no differences at the CD8 level (Fig 9D).
In summary, our data sustain a role for cathepsin B at the T cell level, allowing a higher proliferation rate but a lower IFNγ secretion capacity. In the case of L. major infection, this might make WT mice develop a later but more sustained Th1 type response, maintaining inflammatory lesions for a longer time.
We previously showed that TLR9 plays a critical role in resistance to L. major that is dependent on a parasite DNA–TLR9 –DC–Th1 axis [8,29]. The work presented here aimed at investigating the functional consequences that different proteases, reported to intervene in the functional maturation of TLR9, might have on the evolution of cutaneous L. major infection. Using mice genetically deficient for AEP, CatB, CatL and CatS we found that only CatB-/- were significantly different from WT mice. These mice showed a faster dynamic and increased efficiency of cutaneous lesion resolution and clearance of L. major parasites. Previous studies on the role of cathepsins during L. major infection revealed different effects of cathepsin inhibition. Targeting cathepsin B by specific inhibitors was reported to be protective in BALB/c mice but with no effect on DBA/2 mice [18]. In agreement with these studies we report a protective phenotype in CatB-/- mice. CatL inhibitors on the other hand, were reported to exacerbate disease in susceptible BALB/c and resistant DBA/2 mice [17,33]. We find no effect of CatL, although factors like specificity of the inhibitors as opposed to genetic deletion might be responsible for the different results. Our results complete and expand on existing literature by the use of AEP and CatS knockout mice, where we find no major difference in disease progression and parasite clearance. To our knowledge this is the first report on the role of these proteases during L. major infection. The slightly higher susceptibility, which we observed in the CatS-/- mice, might be linked to an incomplete maturation of TLR9 or a breakdown in cleavage of the MHCII invariant chain, as previously reported for this cathepsin [14,34]. Since none of the AEP-/-, CatL-/-, CatS-/- or CatB-/- mice reproduced the susceptibility of TLR9-/- to L. major infection in terms of lesion size and parasite clearance we concluded that these proteases are not individually involved in TLR9 functional maturation. However, we can also not exclude a compensatory role for any given cathepsin in the absence of the other.
Looking for a role of cathepsin B in the response to TLR9 stimulation, we assessed APC cytokine responses of WT and CatB-/- mice to CpG and L. major. We found no influence of CatB on responsiveness of BMDCs or BMDMs to TLR9-dependent or independent stimulation in terms of cytokine transcription or production. In opposition to our finding, a recent study reported that CatB deficiency affects both antigen presentation and IL-12 production from APCs upon stimulation with L. major [19]. While we did not assess protein levels of IL-12, we did not observe significant differences between WT and CatB-/- DCs in terms of the transcriptionally regulated expression of IL-12p35 [35] to live or killed L. major or other stimulants like LPS and CpG. However minor variations in experimental conditions like composition of culture media, type of serum and growth factors used might also be responsible of the observed differences. We also find similarities with the aforementioned work, like no difference in IL-6 production in response to L. major. Differential harvest times might also explain why we see no differences in TNF response to LPS, as reported by Ha et al [36], however as the authors themselves mention, TNF can probably be secreted through a cathepsin B independent pathway. Interestingly, membrane-associated TNF was shown to be sufficient to protect mice during leishmania infection [37]. But a subsequent study showed that this was also dependent on the L. major strain used [38]. Taking this into account, the differences we find in relation to earlier work on the influence of CatB on the APC response to L. major [19], might also be due to the different strains used (LV39 vs MHOM/IL/81/FE/BNI). All in all, our results suggest that in our model, dendritic cell or macrophages are not likely to be the source of the differences observed in the control of L. major.
Cysteine protease cathepsin L and B function as antigen processing proteases and modulate the processing pattern of L. major antigen (SLA). It was shown that the digestion of SLA by a mitochondria /lysosome fraction was slightly influenced by the presence of an inhibitor for catepsine B or L but overall, the degradation profile of the antigen remains similar [17,18,33,39]. Indeed, degradation was very partial and the size of the fragments still very far from those of the peptides which are presented. To address this question, we assessed the antigen presenting capacity of BMDCs from WT and CatB-/- mice. We found no difference between the antigen presenting capacity of cells from both mice strains; BMDCs being as efficient at triggering IFNγ secretion from in vivo primed T cells. As originally described, we found that cathepsin B is dispensable for MHC class II presentation [25]. Once again we saw no indication of a role of APCs (DCs) in the difference in anti-leishmanial T cell responses, although we cannot exclude a difference in antigen presentation by other cells like B cells [40].
The resistant C57BL/6 genetic background of WT and CatB-/- mice to L. major infection allowed us to compare immune parameters over a long course of infection in a high-dose and relevant model of self-resolving leishmaniasis. We defined the soluble mediators and cellular components of the inflammatory response. Overall, WT mice responded to L. major with a delayed, enhanced and more sustained inflammation as compared to CatB-/-. This inflammatory pattern was characterized by higher local chemokine levels accompanied by a significantly increased infiltration with neutrophils in both FPs and dLNs that persists beyond day 21 of infection in WT mice. High neutrophil numbers at the site of inoculation reflect the inflammatory state but have also been shown to correlate with immune escape in leishmaniasis [41], neutrophils being considered as a Trojan horse used by the parasite to evade the immune system [42]. The higher levels of IL-6 and TNF we find in dLNs of CatB-/- at day 7 p.i. are interesting, yet difficult to correlate with a response that might favour Th1 priming. Although the role of IL-6 during L. major infection has been previously investigated, it was shown to have no major effect or at least not on early responses [43–45] TNF however, was shown to have a clear protective role in L. major infection, with higher TNF levels in C57BL/6 mice compared to BALB/c [37,46]. We did not observe impairment in IL-1β secretion in CatB-/- mice, these results being consistent with a cathepsin B-independent activation of IL-1β, especially since multiple cathepsins have been implicated in NLRP3-mediated IL-1β activation [47]. Our results thus point to a minor influence of CatB deficiency on the innate response during L. major infection.
A delicate balance between Th1 and Th2 is supported as the turning point for anti-leishmanial immunity and this conclusion has benefited greatly by studies in prototypical Th1 and Th2 mice, C57BL/6 and BALB/c, respectively [48,49]. An early Th2 has been reported to develop in C57BL/6 mice, which is afterwards shut down and dominated by a protective Th1 [50]. Accompanying the favorable phenotype of CatB-/- mice, we observed that an adaptive Th1 response started at earlier time-points and was resolved faster as compared to WT mice. Previous reports on use of cathepsin B inhibitors have shown that in vivo treatment of infected BALB/c mice can suppress the development of a deleterious Th2 [18]. Consistent with these early studies, we observed a tendency of CatB-/- mice to express higher levels of IFNγ and lower levels of IL-4 in the early phases of disease. This trend was lost after day 21 when CatB-/- Th1 markers fall below those of WT mice. We were not able to detect transcript for these cytokines in the FPs, probably because of the low numbers of T cell present at this site as compared to the dLNs. We did however find very high iNOS transcript upregulation in the FPs. While WT mice maintain higher levels of Th1 cytokines and increased cellularity in the dLNs, they also maintain a relatively high size of FP lesions and parasite burdens. These observations have already been described and attributed to anti-inflammatory mechanisms like Treg expansion and IL-10 secretion which allow persistence of the parasite [51]. We observed a difference in the representation of T cell populations in CatB-/- mice, characterized by a higher percentage of CD4, among which a lower percentage of cells were CD25+foxp3+ Tregs. This difference was present in naïve mice and maintained during the course of infection indicating that the equilibrium between effector and regulatory T cells as described by Belkaid et al. [52] is higher in C57BL/6 mice than in CatB-/-, which might explain the faster development and regression of the inflammatory response in these mice. To further evaluate the immune status of infected mice, we examined the proportion of L. major-specific IgG subclasses. Except for a higher abundance of antibodies in CatB-/- mice, we found no indication of a preferential skewing of the IgG2a/IgG1 ratio. Since we did not observe higher, but in fact lower percentages of B cells in CatB-/- mice during later stages of the disease, we believe that a qualitative difference in Th-responses at early moments might induce the differences we observe in IgG levels.
Cathepsin B is one of the most abundant cysteine proteases not only present in lysosomes of APCs but also in B and T lymphocytes [21,22,53–55], and has been described to be implicated in different inflammatory disease [56,57]. Lysosomal proteases were believed to be mainly involved in different types of cell death. Aside from the role of cathepsin B in promoting cell death, a different line of research suggested that surface cathepsine B protects cytotoxic lymphocytes and NK cells from self-destruction after degranulation [22]. These results were obtained with highly specific cathepsin B inhibitors, but were not confirmed by Baran et al [23], using CatB-/- mice. In addition it has been shown that in vitro supra optimal activation induces apoptosis of T cells due to release of catalytically active CatB and CatL in the cytosol [53].
Considering the possible effects of cathepsin B on lymphocytes, we turned to a system in which we could prove the involvement of lymphocytic cells in the CatB phenotype. We initially reconstituted RAG2-/-γc-/- mice with splenocytes from either WT or CatB-/- and observed FP thickness after infection with L. major. RAG2-/-γc-/- mice reconstituted with CatB-/- splenocytes developed significantly smaller lesions throughout the course of infection compared with mice reconstituted with WT cells. While this result argues for a mostly lymphocytic origin of the observed differences, we could not exclude the influence of small myeloid populations in the spleen. Further experiments led to similar observations with the use of isolated CD3+ T cells. These results argue that, at least in our RAG2-/-γc-/- reconstitution system, B cells and NK cells along with other non-T cell populations in the spleen, do not contribute much to the development of local inflammation during L. major infection. We found significantly lower parasite burdens in the FP of RAG2-/-γc-/- mice reconstituted with CatB-/- T cells as compared to WT T cells. As previously described [32], RAG2-/-γc-/- mice fail to contain systemic dissemination of leishmanial parasites due to lack of lymph nodes, which might also account for the lack of resolution we observed in these mice. Differences in cellular phenotype similar to those found between dLNs of WT and CatB-/- mice were also found in the spleens of reconstituted mice, as a higher percentage of CD4 + cells with significantly fewer CD4+CD25+ cells among them. In RAG2-/-γc-/- mice reconstituted with CatB-/- T cells we observed a lower percentage of CD8 + as compared to WT in spleen, which was in accordance to the original phenotype of the transferred cells. The lower level of CD8 could explain the better resolution of the inflammatory response as previously suggested by Belkaid et al, which showed that CD8 T cells contribute to inflammation in a RAG2-/- reconstitution model [58].
These points led us to study the cytokine production capacity and the ability to proliferate of T cells. We found that CatB-/- T cells have an intrinsic capacity to secrete more IFNγ but are less able to produce IL-2 and proliferate in vitro. Altogether, the data underline that intrinsic differences in the T cell compartment exist between WT and CatB-/- mice which evoke a precocious Th1 response, probably as a result of a higher IFNγ production capacity, and faster resolution of inflammation which could be attributed to the lower proliferative capacity of CatB-/- CD3 T cells. We cannot at this moment say if the differences in T cell reactivity imply a cell-intrinsic role for CatB or are the result of development and selection of T cells in a CatB-/- -environment. Given the ubiquitous expression of CatB and the many functions attributed to this protease, identifying a precise mechanism of action might prove very difficult. In addition to the differences in cytokine secretion capacity and proliferation we report for T cells, cathepsin B might affect other processes that may influence L. major disease progression like the reported role in cleavage of chemokines or IL-1 processing [47,59].
Unanswered questions remain on the source of these functional differences within the T cell population. The imbalance in CD8, CD4 and Treg might be a first reason to consider, followed by the individual competence of each of these populations. The increased inflammatory capacity correlates with a possible defect in Treg suppressive capacity.
Cathepsin B has already been proposed as a therapeutic target for inflammatory diseases, and for leishmaniasis, and our work brings forth new evidence on the effects that cathepsin B inhibition might have at the T cell level, and sets the stage for future investigation.
|
10.1371/journal.ppat.1000755 | Neutrophil-Derived CCL3 Is Essential for the Rapid Recruitment of Dendritic Cells to the Site of Leishmania major Inoculation in Resistant Mice | Neutrophils are rapidly and massively recruited to sites of microbial infection, where they can influence the recruitment of dendritic cells. Here, we have analyzed the role of neutrophil released chemokines in the early recruitment of dendritic cells (DCs) in an experimental model of Leishmania major infection. We show in vitro, as well as during infection, that the parasite induced the expression of CCL3 selectively in neutrophils from L. major resistant mice. Neutrophil-secreted CCL3 was critical in chemotaxis of immature DCs, an effect lost upon CCL3 neutralisation. Depletion of neutrophils prior to infection, as well as pharmacological or genetic inhibition of CCL3, resulted in a significant decrease in DC recruitment at the site of parasite inoculation. Decreased DC recruitment in CCL3−/− mice was corrected by the transfer of wild type neutrophils at the time of infection. The early release of CCL3 by neutrophils was further shown to have a transient impact on the development of a protective immune response. Altogether, we identified a novel role for neutrophil-secreted CCL3 in the first wave of DC recruitment to the site of infection with L. major, suggesting that the selective release of neutrophil-secreted chemokines may regulate the development of immune response to pathogens.
| When infectious agents enter our body, neutrophils are the first cells recruited to the scene. In addition to their capacity to kill microbes, neutrophils secrete molecules that attract other cells also involved in immune defense, such as dendritic cells (DCs). Here, we investigate the secretion of DC-attracting chemokines by neutrophils after inoculation of mice with Leishmania major, a protozoan parasite that can cause long-lasting, skin ulcers in man. Following parasite inoculation, most inbred strains of mice (e.g.C57BL/6) develop self-resolving lesions, while in a few strains (e.g. BALB/c) lesions fail to heal. We report here that in “healer” C57BL/6 mice, higher numbers of DCs were attracted at the site of infection than in “non-healer” BALB/c mice. DC recruitment correlated with secretion by neutrophils of the chemokine CCL3, as indeed a markedly decreased DC recruitment was observed in C57BL/6 mice depleted of neutrophils or deprived of the capacity to produce CCL3. DC recruitment was restored upon transfer of normal neutrophils to CCL3 deficient mice. Our results reveal a crucial role for neutrophil-secreted CCL3 in early recruitment of DCs in L. major-infected “healer” mice, and suggest that the type of chemokine secreted by neutrophils will have consequences in the launching of pathogen-specific immune response.
| Neutrophils rapidly accumulate at the site of microbial infection and recent evidence show that they play a major role in immunity to several pathogens. Neutrophils, through the early release of cytokines and chemokines, create a microenvironment critical for the shaping of the development of an antigen-specific immune response (reviewed in [1],[2]).
Analyzing the early mechanisms controlling dendritic cell migration to the skin will contribute to the understanding of the development of immunity against infections. To this end, the experimental murine model of infection with the protozoan parasite Leishmania major (L. major) was used. After parasite inoculation, most mouse strains, including C57BL/6 mice, are resistant to infection and develop a protective CD4+ Th1 immune response, while a few strains such as BALB/c mice are susceptible to infection and develop a CD4+ Th2 type of immune response (reviewed in [3]). Following infection with L. major, neutrophils are massively and equally recruited to the site of parasite inoculation in mice from both strains of mice [4],[5], and recently, early recruitment of neutrophils and their essential role in the development of L. major protective immune response was confirmed using mice infected in the ear through the bite of female sandflies [6]. Depletion of neutrophils prior to Leishmania inoculation was shown to modify the development of the CD4+ T helper immune response [5],[7],[8], however, the exact mechanism(s) involved in this early process remain(s) to be determined.
Once exposed to L. major promastigotes, neutrophils from mice resistant or susceptible to infection were reported to develop distinct phenotypes including differential expression of Toll-like receptors and cytokine secretion [9]. Neutrophils could therefore create a microenvironment in the skin and influence that of skin draining lymph node, determining the development of the antigen-specific immune response. Dendritic cells, the most efficient antigen presenting cells, will be critically involved in this process. Indeed, following infection with L. major, dendritic cells have been reported to be crucial in the resistance to infection [10],[11], reviewed in [12]. Thus, considering the massive presence of neutrophils recruited to the site of parasite inoculation within the first day of infection, we hypothesized that crosstalk between neutrophils and dendritic cells at the site of infection might shape the development of the L. major specific immune response. DCs present in the L. major inoculated skin are trafficking from the epidermis/dermis, or recruited from the blood or/and from the bone marrow. They may include Langerhans cells [13], dermal DCs [14], as well as the rapidly differentiating monocyte-derived DCs [15],[16].
In the present study, we have analyzed the role of neutrophil-derived chemokines in the recruitment/trafficking of dendritic cells in the skin, during the first days of infection with Leishmania major. Although several chemokines have been reported to attract immature DC, the particular role of neutrophil-secreted chemokines in this early process has been scantily investigated. We analyzed the secretion of CCL3, CCL4, CCL5 and CCL20, as these chemokines have been reported to be both secreted by neutrophils, and to recruit immature DCs [17],[18],[19],[20].
Our results indicate first, that neutrophils from C57BL/6 L. major resistant mice secrete significantly more CCL3 than BALB/c susceptible mice in response to L. major in vitro, and second, that CCL3 is the key chemokine involved in chemoattraction of immature DCs. Infected C57BL/6 mice displayed high levels of CCL3 one day post L. major inoculation in the ear dermis, and markedly more Langherans cells, dermal DCs and monocyte-derived DCs were recruited to the site of parasite inoculation than in infected BALB/c mice, an effect mediated by neutrophil-derived CCL3. The early neutralization of CCL3 or its absence in CCL3−/− mice resulted in a delay in development of IFNγ secreting-Th1 cells, correlating with transient higher parasite load and tissue damage, a phenotype more sustained and statistically significant in CCL3−/− mice. This identifies the CCL3 secreted by neutrophils during the first days of infection as a critical chemokine involved in the recruitment/trafficking of dendritic cells, which influences the subsequent development of the immune response.
Neutrophils have been reported to secrete chemokines in response to microbial stimuli [20]. First we determined whether L. major induced the transcription and secretion of DC-attracting chemokines in inflammatory neutrophils. L. major-recruited inflammatory neutrophils were purified by MACS and incubated with L. major and/or with IFNγ, an activator of neutrophils. Sixteen hours later, cells were collected for mRNA analysis, or twenty-four hours later, cell-free supernatant was analyzed for chemokines reported to attract immature DCs. Incubation of C57BL/6 neutrophils with L. major increased both CCL3 transcript levels and protein secretion, while only a very mild effect was seen in response to IFNγ alone (Figure 1A,B). In contrast, incubation of C57BL/6 neutrophils with L. major alone did not induce a significant increase of either CCL4 mRNA or protein, but incubation of neutrophils with IFNγ induced an increase in its mRNA and release (Figure 1A,B). IFNγ and to a lower extent L. major, independently induced CCL5 mRNA, with IFNγ alone inducing the highest release of CCL5. Interestingly, the presence of L. major significantly impaired the IFNγ-induced release of both CCL4 and CCL5 (Figure 1B).
To determine if L. major would induce similar transcription and secretion of DC-attracting chemokines from neutrophils in the L. major-susceptible BALB/c mice, L. major recruited inflammatory BALB/c neutrophils were treated and analyzed as described above. L. major did not induce significant level of chemokine transcription nor elevated release from BALB/c neutrophils (Figure 1A,B). Neither C57BL/6 nor BALB/c neutrophils secreted or transcribed CCL20 in response to L. major, IFNγ or both (data not shown).
The selective secretion of CCL3 by L. major peritoneally-induced inflammatory C57BL/6 but not BALB/c neutrophils (Figure 1) was also measured in inflammatory dermal neutrophils recruited in the ear dermis 24 hours after L. major inoculation (Figure S1).
Altogether, these results show that, once exposed to L. major promastigotes, C57BL/6 neutrophils secrete CCL3 a chemokine known to attract immature DCs, and that the production of chemokines by BALB/c neutrophils is significantly lower in response to the parasite.
We next tested the potential effect of neutrophil supernatant on chemoattraction of immature DCs using Transwell cell migration assays. Bone marrow-derived partially immature C57BL/6 and BALB/c DCs (MHCIIlow, CD40−, B7.1−, B7.2low) were deposited on the filter of a Transwell migration assay plate, the lower compartment containing the supernatants recovered from C57BL/6 or BALB/c neutrophils exposed or not to L. major. While robust chemo-attractive activity for BM-iDCs was detected in the supernatants of C57BL/6 neutrophils exposed to L. major, no similar strong chemo-attractive activity was detectable in the supernatants of BALB/c neutrophils exposed to L. major (Figure 1C).
To assess if this chemo-attractive activity was due to CCL3 two approaches were selected. First, CCL3 was depleted from supernatant of C57BL/6 mouse neutrophils exposed to L. major, and the CCL3-depleted supernatant monitored for its iDCs chemo-attractive activity, using the Transwell migration assay. Remarkably, a significant decrease of the BM-iDC migration towards the latter supernatant was measured (Figure 1D). Second, supernatants from CCL3−/− C57BL/6 mouse neutrophils exposed to L. major were similarly tested and shown to display reduced chemo-attractive activity for +/+ BM-iDCs (data not shown).
In contrast, CCL3 depletion of BALB/c supernatants showed only a mild and not statistically significant effect on DC chemoattraction, in line with the low chemokine secretion and chemoattraction of L. major-stimulated BALB/c neutrophil supernatants (Figure 1D).
These results demonstrate that the CCL3 present in the supernatant of C57BL/6 neutrophils exposed to L. major promastigotes is the main chemokine attracting BM-iDCs in vitro.
A significant difference in chemokine secretion was measured in our ex vivo/in vitro analysis between L. major-stimulated C57BL/6 and BALB/c neutrophils, with consequence on the recruitment of iDCs. This prompted us to investigate whether similar differences were observed in vivo. To evaluate how neutrophil and chemokine release influence dendritic cell recruitment in vivo, the model of ear skin explants was used. L. major was delivered intradermally in the ear and at different time points post inoculation, ears were recovered, and further processed as ear explants. This ex vivo approach allows the evaluation of DC recruitment in the skin dermis [21],[22].
C57BL/6 and BALB/c mice were infected i.d. in the ear with L. major, and the number of cells that migrated out of infected ear skin explants was analyzed by FACS and quantified during the first 48 hours following parasite inoculation.
A large number of neutrophils emigrated from the ear explant within hours of parasite inoculation and their number started to decrease 48 hours later (Figure 2A). As previously reported for L. major infected footpads [4],[5], at this early stage post parasite inoculation the neutrophil number did not differ between C57BL/6 and BALB/c mice. We then estimated the number of Langerhans cells (LC) and dermal DC (dDC), defined by FACS analysis by their high surface expression of CD11c, DEC205, and MHC Class II (Figure 2B). Already six hours after infection, the number of LC and dDC migrating out of the ear dermis was significantly higher in C57BL/6 than in BALB/c mice, with the highest difference occurring twenty-four hours post infection (Figure 2A). As monocyte-derived DC (MoDC) play an important role in L. major infection [23], this cell population, characterized as being Ly6G−, Ly6C+, CD11b+ and CD11cdim (Figure 2B), was also analyzed by FACS. Twenty-four hours post L. major inoculation, a significantly higher number of MoDCs emigrated from C57BL/6 ear explants as compared to BALB/c ear explants (Figure 2A). Most of the DCs emigrating from the ears resulted from the L. major presence, while only a small percentage of DCs emigrated from the ears was due to needle-dependent injury, as illustrated by the low DC recruitment measured when medium alone was injected (Figure S2).
These results reveal a significant difference in DC migration during the first day post infection with L. major between C57BL/6 and BALB/c mice.
In order to further investigate whether the difference in the number of DCs recruited during the first day post L. major inoculation in C57BL/6 and BALB/c mice correlated with their distinct neutrophil functional phenotype, mice of both strains were given i.p. a single injection of the neutrophil-depleting mAb NIMP-R14, and six hours later, L. major promastigotes were i.d. delivered in one of their ears. Twenty-four hours post L. major inoculation, the DC subsets sedimenting out of the ear explants were analyzed by FACS, quantified and compared to the ones sedimenting from ear explants prepared from mice that were given a control mAb. Depletion of neutrophil significantly decreased the number of LC and dDC and abolished the recruitment of MoDC in the ear skin dermis (Figure 3A,B). These results demonstrate an essential role for neutrophils in the early recruitment of DC following inoculation of L. major.
Since neutrophils appeared to be essential for the recruitment of dendritic cells in the first days following infection with L. major and as we have shown in vitro that the CCL3 secreted by neutrophils was critical for the recruitment of DCs, we next sought to document whether this chemokine could play an essential role in vivo.
To this end, the level of chemokine mRNA was measured in L. major infected ears during the first 48 hours post infection. In C57BL/6 infected mice, CCL3 mRNA was strongly induced within 24 hours of L. major inoculation (Figure 4A), while significantly less CCL3 mRNA was induced in L. major infected BALB/c mice. L. major induced only a small increase in CCL4 and CCL5 mRNA at the site of infection (Figure 4A), while infection did not induce CCL20 mRNA (data not shown).
To investigate if neutrophils were responsible for CCL3 transcription at the site of infection in C57BL/6 mice, neutrophils were depleted with an injection of the NIMP-R14 mAb 6 hours prior to infection, and chemokine transcript abundance was measured in infected ears 24 hours post parasite inoculation. In L. major infected mice injected with a control mAb, CCL3 mRNA levels were increased significantly 24 hours after L. major inoculation while in neutrophil-depleted mice, CCL3 mRNA levels were much lower (Figure 4B). Among the three other chemokine transcripts - CCL4, CCL5, and CCL20 – monitored, only the CCL4 followed a transcript profile similar to the CCL3 one though with lower amplitude (Figure 4A). These data demonstrate that twenty-four hours post parasite inoculation in the ear, neutrophils contribute to most of the CCL3 and part of the CCL4 present during the first day of infection.
In order to directly establish that CCL3 acts as a key chemokine accounting for the recruitment of DCs in the L. major-loaded dermis, CCL3 was depleted in mice by treatment with Evasin-1 [24], a highly selective neutralizing chemokine binding protein with high affinity for CCL3 and to weaker affinity for CCL4 [25]. Mice were given Evasin-1 two hours prior to i.d. L. major inoculation in the ear, and DC mobilization in the infected ear explants was quantified as described above. Injection of Evasin-1 had no significant effect on the number of neutrophils that migrated out of ear skin dermis 24 hours after infection (Figure 5A). This allowed the investigation of the role of neutrophils on DC recruitment at the site of infection, under conditions when CCL3 secretion was neutralized. While injection of Evasin-1 into C57BL/6 mice resulted in significant decrease of emigration of both LC/dDC and MoDC from the ear explants, injection of Evasin-1 into BALB/c mice did not result in any similar phenotypic changes (Figure 5A).
To confirm the role of CCL3 in DC recruitment in vivo, C57BL/6 mice or mice genetically deficient in CCL3 (CCL3−/−) on the C57BL/6 genetic background, were infected with L. major i.d., and emigration from the ear explants assessed as described above. As measured in mice treated with Evasin-1, neutrophil recruitment was not significantly decreased one day post infection. However, recruitment of both LC/dDC, and MoDC was significantly reduced in CCL3−/− mice (Figure 5B), confirming that CCL3 is a major chemokine involved in LC/dDC migratory properties as well as MoDC recruitment. These results strongly suggest that the CCL3 secreted by neutrophils contributes for a major part to early DC trafficking and/or recruitment in the dermis of L. major inoculated mice.
So far, our data suggest that neutrophils and CCL3 contribute significantly to the early DC trafficking/recruitment to the site of L. major inoculation in C57BL/6 mice. In order to monitor whether neutrophils were indeed the major source of CCL3 that mediates DC recruitment, we transferred either C57BL/6 wild type (WT), or CCL3−/− neutrophils into CCL3−/− mice that were given i.d. L. major, and twenty-four hours later, leucocytes emigrating from the ear explants were analyzed by FACS. The number of neutrophils did not differ significantly after injection of WT or CCL3−/− neutrophils into C57BL/6 and CCL3−/− mice (Figure 6A). However, injection of WT neutrophils into ears of CCL3−/− mice strongly increased the number of LC/dDC and MoDC migrating out of the the ear explants, to levels that were comparable to those measured in WT mice injected with WT or CCL3−/− neutrophils (Figure 6B,C right panels). Even though injection of CCL3−/− neutrophils into CCL3−/− mice also increased the number of LC/dDC and MoDCs, the levels attained were significantly lower than those obtained when CCL3−/− mice were injected with WT neutrophils. These results demonstrate that the CCL3 secreted by neutrophils plays an important role in the early trafficking/recruitment of DCs cells to the site of infection during the first days of L. major infection.
To investigate the role of CCL3 in the onset of the immune response in C57BL/6 mice, CCL3 was inhibited during the first five days post L. major inoculation by daily administration of Evasin-1. Fifteen and forty days post infection, the development of T helper immune response was assessed. L. major-specific IFNγ and IL-4 cytokine production was measured in draining lymph node CD4+ T cells, and immunoglobulin isotype switching was measured in the serum. Fifteen days post infection the transient depletion of CCL3 resulted in almost total abolition of IFNγ secretion, but no significant difference in IL-4 levels (Figure 7A). The low IFNγ levels correlated with a significant decrease of L. major-specific IgG2a and a milder but not statistically significant increase in IgG1 serum levels in Evasin-1 treated mice (Figure 7C). Low IFNγ secretion was also measured fifteen days post L. major inoculation in CCL3−/− mice (Figure 7A and 7C), in line with results from a previously published study [26], but IgG2a levels did not differ significantly from C57BL/6 controls, while IgG1 levels were slightly increased (Figure 7C). Of note, very low levels (<100pg/ml) of IL-13 and IL-17 were measured in supernatant of L. major-restimulated CD4+ T cells, with no difference between the groups (data not shown).
Six weeks post L. major inoculation, high levels of IFNγ with correspondingly high levels of L. major-specific IgG2a, and low levels of IL-4 and IgG1 characteristic of a Th1 immune response were measured, with no statistically significant difference being noted between the groups (Figure 7B, 7D). Mice treated with Evasin-1 developed slightly larger lesion than control mice but smaller lesions than CCL3−/− mice, with no statistically significant difference (Figure 7E). In contrast, development of lesion was markedly increased in CCL3−/− mice compared to C57BL/6 controls, and CCL3−/− mice harboured higher parasite load within their lesions, with a significant difference compared to Evasin-1 treated or control mice (Figure 7G).
In this study, we focused on the role of neutrophils in the recruitment of DCs to the site of L. major inoculation, one main question addressed being whether neutrophils could transiently and locally contribute to the Langherans cell/dermal dendritic cell trafficking as well as to the recruitment of monocytes, the latter being capable to be programmed to either macrophages or monocyte-derived DCs. Our results demonstrate a previously unappreciated role of primary neutrophil extravasation, and of the CCL3 released by neutrophils in the rapid recruitment and trafficking of LC/dDC as well as of monocyte-derived DC to the site of L. major inoculation. We report here for the first time that the selective secretion of CCL3 by neutrophils is critical in vivo for the recruitment of DCs to the site of Leishmania inoculation, as revealed by markedly reduced recruitment of DCs in mice with pharmacological neutralization or absence of CCL3, and in neutrophil-depleted mice. This decrease was restored in CCL3−/− mice by the co-injection of WT C57BL/6 neutrophils together with the parasite.
CCL3 has been reported to attract neutrophils, however, our data clearly show that this chemokine does not contribute significantly to the first wave of neutrophil migration following L. major inoculation, as the number of neutrophils recruited to the site of parasite delivery one day post infection was not significantly affected by either the absence or the neutralization of CCL3.
Previous studies performed in vitro reported the transcription or/and secretion of DC attracting chemokines by neutrophils. Human neutrophils were shown to secrete molecules involved in attraction of immature DCs such as defensins [27], and CCL3 following LPS stimulation in vitro [28]. Murine neutrophils were also reported to transcribe CCL3, CCL4, CCL5 and CCL20 mRNA in response to Toxoplasma gondii exposure in vitro, with the highest induction of CCL5 mRNA [29]. In the present study, L. major induced the highest transcription and secretion of CCL3, low level of CCL5, and no transcription nor secretion of CCL4 and CCL20 in vitro. Thus, distinct pathogens can elicit different chemokine transcription patterns in neutrophils, and as reported here, the same pathogen can induce distinct chemokine release depending on the genetic background of the host. In this line, induction of TLRs in response to L. major was previously reported to differ in neutrophils from C57BL/6 and BALB/c mice, respectively resistant or susceptible to L. major [9].
The coordinate expression of chemokines and their receptors has been shown to be important in protective immunity to infection with L. major. While most studies have focused on chemokine expression in draining lymph nodes, only a few have investigated chemokine expression at the site of parasite inoculation: CCL2 and CCL3 mRNA expression were reported to be elevated already one day post infection in L. major infected footpads of C57BL/6 mice [30]. CCR2, the receptor for CCL2 and other monocyte chemoattractant proteins, was thought to be required for the generation of a protective immune response against L. major, but the differential outcome observed in CCL2−/− and CCR2−/− mice following L. major infection suggests that ligands other than CCL2 were involved in this protection [26],[31]. We also observe that in the presence of L. major, low level of CCL5 is secreted by neutrophils, but we show that it is not the major neutrophil-secreted chemokine involved in the recruitment of DCs. Indeed, during L. major infection, CCL5 expression was reported to increase selectively in C57BL/6 compared to BALB/c mice, but mainly in the late phase of infection [32]. We also report here a small transient increase in CCL4 and CCL5 mRNA one day post L. major inoculation at the site of infection in C57BL/6 mice, but CCL3 clearly showed a much higher level of transcriptional induction in these mice. Altogether, these and our studies reveal the importance of the tight control and timing of chemokine secretion during the first days post L. major inoculation.
How can the different levels of CCL3 released by C57BL/6 and BALB/c neutrophils exposed to L. major, affect the subsequent development of the immune response? Our data demonstrate that the early secretion of CCL3 has an impact on the development of the adaptive immune response, first through the recruitment of DC, and second, possibly through their activation. Whether and how neutrophils impact on the subsequent DC functions that account for T lymphocyte signaling is currently under investigation. Neutrophils have indeed been reported to deliver maturation and activation signals to DCs in BCG and Toxoplasma gondii infection [29],[33], and neutrophils were shown to associate with immature DC through interactions between DC-SIGN on immature DC and specific glycans on neutrophils [34]. In addition, neutrophil-derived ectosomes have been reported to interfere with the maturation of MoDCs [35]. Whether these DC subsets reach the skin-draining lymph node remains to be established, but it is plausible that they do, providing signals to T lymphocytes. In this line, following L. major inoculation, the early absence of CCL3 had consequences on the Th1 immune response normally developing fifteen days post parasite inoculation in the draining lymph nodes of C57BL/6 mice, preventing most of the IFNγ secretion by draining lymph node CD4+ cells. Further investigations will allow deciphering the mechanism involved in the early though transient inhibition of Th1 response following early neutralization of CCL3, and in the stepwise neutrophil-dependent processes that allow DCs to traffic from the dermis to the skin- draining lymph node at the very early stage post L. major promastigote delivery.
BALB/c and C57BL/6 neutrophils exposed to L. major have been reported to differ in the induction of cytokine secretion, with C57BL/6 neutrophils secreting IL-12, and BALB/c neutrophils secreting IL-12 p40 homodimers, blocking IL-12 signalling [9]. We report here the selective secretion of CCL3, a chemokine reported to induce IL-12 secretion in macrophages, by C57BL/6 but not BALB/c neutrophils. These differences in cytokine and chemokine secretion in L. major-exposed C57BL/6 and BALB/c neutrophils explain, at least in part, their distinct contribution to the subsequent development of T helper cells in the draining lymph node.
Several DC subsets have been reported to be involved in the development of the L. major protective immune response [10],[11] (reviewed in [12]), and infection-induced inflammatory reactions include a sharp increase in DCs at the site of parasite inoculation. L. major has been shown to be phagocytosed by dermal DCs [36], and Leishmania antigens have been reported to be transported by dermal DCs rather than by LCs [37]. Recently, de novo differentiation of monocytes into DCs, and the crucial importance of these migratory dermal monocyte-derived DCs in controlling the development of a protective CD4+ Th1 type of immune response has been demonstrated in L. major infection [23], even if a role for resident lymph node DCs is not excluded [38]. Thus the role of neutrophil-derived chemokines, together with the important contribution of CCL3 in the early recruitment of MoDC at the site of infection reported in the present study, emphasize the importance of neutrophils in recruiting the cells contributing to the priming of CD4+ Th1 cells that are essential in efficient protection against L. major infection.
Natural transmission of L. major is occurring during the bite of an infected sandfly. When mouse ears are exposed to Leishmania-hosting sand flies, neutrophils are rapidly recruited to the site of parasite inoculation, an early phenotypic trait also observed after intradermal needle inoculation of L. major [6]. It will be important to compare the two experimental systems and to explore whether factors derived from the sandfly may contribute to the early and transient wave of leucocytes as well as to their short term functions.
In conclusion, neutrophil and neutrophil-produced CCL3 appear crucial in the early recruitment of dendritic cells in the dermis, that will further direct the development of an adaptive immune response to L. major. Therefore, strategies interfering with these factors could represent a novel way to shape immune responses to pathogens.
Female BALB/c and C57BL/6 mice were purchased from Harlan Olac Ltd. (Bicester, UK). CCL3−/− mice were purchased from Jackson laboratory (Bar Harbor,USA). All mice were bred in the pathogen-free facility at the BIL Epalinges Center and used at 6 week of age. L. major (LV 39 MRHO/Sv/59/P strain) were maintained and grown as previously described [2]. All animal experimental protocols were approved by the veterinary office regulations of the State of Vaud, Switzerland, authorization 1266.3 to FTC, and experiments were performed adhering to protocols created by this office.
Mouse inflammatory neutrophils collected by peritoneal lavage 4 hours post infection i.p. of 5.107 stationary phase L. major, were isolated and purified using MACS-positive selection as previously described [9]. Purity of neutrophils was >98% as assessed by FACS and Diff-Quick (Dade Behring) staining of cytospins. Each experiment was validated using FACS sorted neutrophils positively gated through 1A8 labeling and negatively gated with a cocktail of mAbs (against CD3, CD49b, B220, F4/80, and CD11c, see below).
Neutrophils were cultured in RPMI-1640 media supplemented with 10% FCS and antibiotics (2.5×106 cells/ml) in the presence or absence of L. major metacyclic promastigotes (at a 5∶1 parasite∶cell ratio). The protease inhibitor aprotinin (0.4 µg/ml, Sigma Chemical Co., St. Louis, MO, USA) was added to the culture to facilitate cytokine detection. Chemokine concentration in the culture supernatant was quantified by ELISA using kits from R&D systems.
Inflammatory neutrophils cultured in vitro under different conditions were harvested, mRNA extracted, cDNA synthesized and quantitative RT-PCR performed as previously described with a LightCycler system (Roche) [9]. Each cytokine transcript was normalized to the value of the hypoxanthine phosphoribosyltransferase endogenous control, represented as arbitrary values. The primers for the real time PCR were the following: CCL3: F 5′ CCA AGT CTT CTC AGC GCC AT 3′, R 5′ TCC GGC TGT AGG AGA AGC AG 3′, CCL4: F 5′ TCT TGC TCG TGG CTG CCT 3′, R 5′ GGG AGG GTC AGA GCC CA 3′, CCL5 [39], CCL20 F5′ CTT GCT TTG GCA TGG GTA CT 3′, R 5′ GTC TGT ATG TAC GAG AGG CA 3′.
CCL3 depletion in neutrophil supernatants. Neutrophil supernatant were placed on a plate coated with antibody from the CCL3 ELISA kit (R&D systems). Depletion in supernatant was checked by ELISA. For in vivo depletion, mice were treated with Evasin-1, a CCL3 blocking protein, engineered by Merck-Serono [24]. 10 µg of Evasin-1 were injected intraperitoneally 2h before injection of the parasite. As controls, mice were injected with a similar regimen of PBS.
Bone-marrow cells were cultured in RPMI-1640 media supplemented with 10% FCS, antibiotics and 30% GM-CSF for 6 days, as previously described [40]. At day 6, the maturation of DCs was checked by FACS, measuring the levels of MHC ClassII, CD40, B7-1 and B7-2 surface molecules.
Supernatant from neutrophils cultured under different conditions were placed in the lower compartment of a transwell plate (96 well plate, 3.2mm diameter 5um pore size, ChemoTx System, NeuroProbe, UK). 105 DCs were put on top of the filter. After 2h of incubation at 37°C, the number of DCs that migrate towards the supernatant were counted on Neubauer chambers using trypan blue.
Mice were given i.p. -6h before L. major inoculation 250 µg of the NIMP-R14 mAb, a rat IgG2b mAb that selectively binds to mouse neutrophils [41]. This treatment was previously reported to deplete selectively neutrophils for three days [8]. As controls, mice were given i.p. the RR3-16 mAb against the Vα3.2 chain of the T-cell receptor (RR3-16, gift of R. MacDonald, Ludwig Institute for Cancer Research, Epalinges, Switzerland).
Mice were given intradermally into ear PBS or 106 stationary phase L. major promastigotes. Six, 24, 48h post L. major inoculation mice were sacrificed, the ventral and dorsal sheets of the ear were separated with forceps, the two leaflets being transferred dermal side down in a plate containing RPMI-1640 media supplemented with 10% FCS and antibiotics at 37°C. The leukocyte populations emigrating spontaneously over 14 hours from the ear explants were then counted and stained for a FACS analysis [22]. In selected experiments, dorsal and ventral sheets of the ears were separated and the dermal side was digested with 0.1 mg/ml of Liberase TL (Roche) for 2h at 37°C. Ears from 7 mice were pooled, cut into pieces and filtered through a 40 µm filter, washed and processed for FACS sorting. Sorted neutrophils were further incubated in RPMI medium for 24 hours. Chemokine presence in neutrophil-free supernatant was then analyzed by ELISA.
Leukocytes emigrating from the ear explants were processed for cell surface staining. The mAb 24G2 was used to block FcRs. For analysis of cell populations several mAbs were used: PE conjugated anti-LY6G (clone 1A8), anti-CD3 (clone G4.18), anti-CD49b (clone DX5), anti-MHCII (clone M5/114.15.2), Cyc conjugated-strepatvidin, PE conjugated, APC and Cyc conjugated anti-CD11c (clone N418), FITC conjugated anti-LY6-C (clone AL-21), Cyc conjugated anti-CD11b (clone M1/70), all mAbs from e-bioscience, SanDiego, CA, US; anti-DEC205 (clone NLDC-145, AbB Serotec,UK, Ldt). Biotinilated anti-F4/80 (clone C1∶A3-1,CEDRALANE, Canada). Cells were analyzed with a FACScan (3 colors) or FACSCalibur (4 colors) (BD Biosciences, Mountain View, CA, USA) and analyzed with the program FlowJo (Tree Star. Inc., Ashland, OR, USA).
Four hours post i.p. inoculation of L. major peritoneal lavage neutrophils were purified by MACS. 106 C57BL/6 or CCL3−/− neutrophils were co-injected in the ear dermis with 106 stationary phase L. major promastigotes. Twenty-four hours later, mice were sacrified, ears were processed as above and emigrating cell populations analyzed by FACS.
3×106 L. major were inoculated in the footpad of either CCL3−/− mice or +/+ mice in which CCL3 was blocked during the first five days of infection by daily injection of Evasin-1. Mice were sacrificed 15 and 40 days post- L. major inoculation in the footpad. Draining lymph node CD4+ T cells were isolated by MACS (Miltenyi Biotec), and cultured in the presence of irradiated C57BL/6 splenocytes ± UV-irradiated L. major promastigotes. Cytokine levels were measured by ELISA. Sera obtained at different time points post L. major inoculation were tested for L. major-binding IgG1 and IgG2a, and parasite burden was determined by limiting dilution assay as previously described [42].
Data were analysed using the Student's t-test for unpaired data.
|
10.1371/journal.pgen.1005206 | Accounting for Experimental Noise Reveals That mRNA Levels, Amplified by Post-Transcriptional Processes, Largely Determine Steady-State Protein Levels in Yeast | Cells respond to their environment by modulating protein levels through mRNA transcription and post-transcriptional control. Modest observed correlations between global steady-state mRNA and protein measurements have been interpreted as evidence that mRNA levels determine roughly 40% of the variation in protein levels, indicating dominant post-transcriptional effects. However, the techniques underlying these conclusions, such as correlation and regression, yield biased results when data are noisy, missing systematically, and collinear---properties of mRNA and protein measurements---which motivated us to revisit this subject. Noise-robust analyses of 24 studies of budding yeast reveal that mRNA levels explain more than 85% of the variation in steady-state protein levels. Protein levels are not proportional to mRNA levels, but rise much more rapidly. Regulation of translation suffices to explain this nonlinear effect, revealing post-transcriptional amplification of, rather than competition with, transcriptional signals. These results substantially revise widely credited models of protein-level regulation, and introduce multiple noise-aware approaches essential for proper analysis of many biological phenomena.
| Cells respond to their environment by making proteins using transcription and translation of mRNA. Modest observed correlations between global steady-state mRNA and protein measurements have been interpreted as evidence that mRNA levels determine roughly 40% of the variation in protein levels, indicating dominant post-transcriptional effects. However, the techniques underlying these conclusions, such as correlation and regression, yield biased results when data are noisy and contain missing values. Here we show that when methods that account for noise are used to analyze much of the same data, mRNA levels explain more than 85% of the variation in steady-state protein levels. Protein levels are not proportional to mRNA levels as commonly assumed, but rise much more rapidly. Regulation of translation achieves amplification of, rather than competition with, transcriptional signals. Our results suggest that for this set of conditions, mRNA sets protein-level regulation, and introduce multiple noise-aware approaches essential for proper analysis of many biological phenomena.
| Cellular protein levels reflect the balance of mRNA levels, protein production by translation initiation and completion, and protein removal by degradation, secretion, and dilution due to growth [1–3](Fig 1A). A standard quantitative model for protein-level regulation [4, 5] is
∂ P i ∂ t = τ i M i - δ i P i (1)
where Pi is the cellular protein level (molecules per cell) of gene i, Mi is the mRNA level, and τi and δi are the mRNA translation and net protein removal rates, respectively. According to this model, at steady-state, protein levels will be proportional to mRNA levels with proportionality constants of τi/δi:
P i = τ i δ i M i (2)
such that if rates of translation and removal did not vary by gene, and in the absence of experimental noise or other variation, steady-state mRNA and protein levels would correlate perfectly [1]. Consequently, the mRNA–protein correlation observed in global measurements of mRNA and protein levels has been intensely studied, and deviations from perfect correlation used to quantify the contribution of post-transcriptional processes to cellular protein levels [1, 3, 6–9].
The consensus across these studies holds that, in a wide array of organisms, transcriptional regulation explains 30–50% of the variation in steady-state protein levels, leaving half or more to be explained by post-transcriptional regulatory processes [3, 7, 9–16]. Higher correlations are observed, generally for subsets of less than half the genome [1, 9, 17]. Low observed mRNA–protein correlations have motivated the search for alternate forms of regulation capable of accounting for the majority of protein-level variability [3, 9, 13]. In one proposal, mRNA levels serve mainly as an on-off switch for protein expression, imposing coarse control over protein levels which is then tuned by post-transcriptional mechanisms [9]. Recent studies have indeed uncovered wide between-gene variation in post-transcriptional features such as inferred translation rates [18] and protein degradation rates [3].
However, as frequently noted [1, 7, 9, 10, 19–21], noise in measurements can cause many of the observations attributed to post-transcriptional regulation. Here, noise encompasses variability due to cell-to-cell variation, growth conditions, sample preparation and other effects due to experimental design [22], and measurement biases and error [10, 21, 23]. Uncorrelated noise between mRNA and protein measurements will reduce the observed mRNA–protein correlation relative to the true value [24], while inflating the variation in measurements of translational efficiency and other post-transcriptional processes.
Most studies, particularly of protein levels, cover only a subset of known genes, due to factors such as signal-to-noise limitations, method biases, and continual revision of the coding-sequence annotations used to design and analyze assays. Limited and variable transcriptome and proteome coverage complicate analyses further, making it difficult to compare studies and to synthesize a holistic view of regulatory contributions. Missing data tends to reduce the precision of estimates, if data are missing at random (MAR). However, most quantification methods are biased toward detection of more abundant mRNAs and proteins [9]. Data which are not missing at random (NMAR) in this way have reduced variance or restricted range. Range restriction, in turn, tends to systematically attenuate (reduce in absolute magnitude toward zero) the observed correlations and regression coefficients relative to complete data [25, 26]. That is, biased detection produces biased estimates of the mRNA–protein correlation, leading to incorrect conclusions about regulatory contributions [27].
In many comparisons of the roles of transcriptional and post-transcriptional regulation, protein levels are correlated with or regressed on various predictors (mRNA level and half-life, codon usage, amino-acid usage, etc.) to determine relative contributions to protein-level variation [1, 3, 4, 14, 18, 21]. If mRNA levels are found to explain a certain percentage, say X, in protein levels, then the other predictors are asserted to explain no more than 100−X percent of the variance [3, 9, 21, 28, 29]. A basic assumption of such analyses is that transcriptional and post-transcriptional regulation vary independently between genes. Several of the same studies report that high-expression genes show signs of more efficient translation [3, 4, 18] (reviewed in [1]), raising concerns about the validity of this assumption.
A related assumption of these analyses, one encoded in the standard functional model above, is that mRNA and protein levels are proportionally or linearly related [1, 5]; the slope of this line is the mean number of proteins per mRNA. More often, the data are plotted on a log-log scale, where linearity appears as a slope of 1. Consistent with this, ordinary least-squares linear regression shows that the slope is quite close to 1 for E. coli (0.96) and budding yeast (1.08) [17], and estimates of proteins per mRNA have been reported roughly constant across mRNA expression levels in a prominent study [30].
However, like correlations, slopes estimated by standard linear regression are biased downward by noise in mRNA level measurements, an effect called regression dilution bias[31] which affects any regression where the independent variable is measured with error. A frequently encountered case is that, given two measurements X and Y, the slope from regressing Y on X is not the inverse of the slope of regressing X on Y[32–34]; this is regression dilution bias at work. Consequently, linear regression cannot be used to estimate the functional relationship between mRNA and protein levels, raising the question of what the true functional relationship is. Use of nonparametric methods avoids assumptions of linearity [1], at the cost of destroying genuine information about the dynamic range of gene expression and its determinants.
Analytical solutions to many of these problems exist—notably, Spearman introduced a correction for noise-induced attenuation of correlation estimates more than a century ago [24]—yet have largely failed to find their way into the hands of groups carrying out gene-regulation experiments and analyses (with a few exceptions [15]). Some problems remain almost entirely unaddressed, such as providing accurate estimates of the functional relationship between variables measured many times with correlated noise yielding variably and systematically missing values.
Here, we develop and integrate approaches to address all of these challenges, with the aim of providing more comprehensive and rigorous estimates of the relationship between mRNA and protein levels than have previously been possible. To do so, we take advantage of the rapid, continual progress made in global measurement of mRNA and protein levels by multiple methods [6, 17, 30, 35–41]. All of these methods were first employed at the genome scale in studies profiling gene expression during log-phase growth of budding yeast in rich medium, a de facto standard. These studies often compare results against previous studies, evaluating agreement, precision, coverage, and dynamic range while pointing out relative advantages of each approach (e.g.[17, 18, 30, 37, 38, 40]).
Our efforts to synthesize these data into a coherent whole are grounded in the stance that all these works constitute measurements of the same underlying quantities—average mRNA and protein levels in a large cell population prepared under narrowly defined conditions—whether or not such measurements were the study goal. Systematic differences between approaches due to experimental choices will introduce variation which may not be distinguishable from simple inaccuracy in measurement. We treat this variation as experimental noise without prejudice. Distinctions between biological variability, measurement error, method bias, and other sources of noise are of course important, particularly in deciding how to control or manage noise. These distinctions may also depend on one’s perspective. For example, unintentional differences in growth conditions may lead two groups following the same protocol to make measurements on samples which inevitably are, in truth, biologically different, such that error-free measurement would reveal differences in mRNA and protein levels. In one sense, these differences reflect biological variability; in an equally valid sense, they represent experimental noise. Similarly, intentional protocol differences that are not meant to alter measurement accuracy (such as use of new methods intended to make measurements more precise), yet carry known and unknown biases, may also introduce noise. Here, we take an empirical approach to noise which does not involve divining intent. Versions of this approach are taken, often implicitly, by the many previous analyses that integrate experiments from multiple groups [8, 9, 17, 18, 30].
Our results reveal that, once noise is accounted for, mRNA and protein levels correlate much more strongly under these experimental conditions than previously appreciated, with a correlation coefficient of r = 0.93. We find that protein levels are not proportional to mRNA levels, but instead are more steeply related, an effect we show is consistent with measurements of translational activity. Transcriptional and post-transcriptional regulation act in a concerted, non-independent manner to set protein levels, inconsistent with common attempts to divvy up and assign protein-level variance to each mechanism. As a byproduct, we generate what by several measures is the most complete and accurate quantitative transcriptome and proteome available, in average molecules per haploid cell, for this widely studied organism under these well-studied conditions. Finally, we highlight and introduce methods for analyzing correlations and functional relationships between measured data which may be used broadly.
We collected 38 measurements of mRNA levels and 20 measurements of protein levels from 13 and 11 separate studies respectively, each of haploid S. cerevisiae growing exponentially in shaken liquid rich medium with 2% glucose between 22°C and 30°C (Table 1). As described in the Introduction, we assume, for modeling purposes, that each replicate in each experiment constitutes a measurement of the true per-gene mean mRNA and protein levels under these narrowly defined conditions. These data cover varying amounts of the genome and display a wide range of correlations between studies (Fig 1B, Pearson correlations on log-transformed values with zeros and missing values omitted).
Although correlations of replicates within studies are quite high [9], with median r = 0.97 for mRNA and 0.93 for protein levels, between-study correlations are far more modest, r = 0.62 for mRNA measurements and 0.57 for protein measurements (Fig 1C). That is, data from a typical mRNA study explains 39% of the variance in another study (r2 = 0.39) and a typical protein study’s results explain only 32% in another study’s variance, consistent with previous studies reporting wide variation between studies [16]. Strong outliers indicate high reproducibility for a two pairs of studies (Fig 1C), but each such outlier is a correlation between separate studies done by the same research group, suggesting the presence of additional variability sources between groups. Coverage of the 5,887 verified protein-coding genes in yeast [42] also varies widely across pairs of studies (Fig 1C).
Coupled with high within-study reproducibility, the low between-study reproducibility indicates the presence of large systematic errors between studies. In a single study [38], mRNA levels in a commercially prepared sample were measured using two methods, a commercial microarray and single-molecule RNA sequencing. These measurements correlate with r = 0.86 (73% of the variance explained in one measurement by the other), quite similar to the r = 0.84 correlation of the single-molecule measurement with an independent RNA-Seq dataset on RNA from a different study [43]. These data hint, coupled with similar observations in other biological systems [44], that high within-study reproducibility is likely to reflect reproducible biases associated with use of a single measurement technique in addition to reproducible features of the biological sample.
Correlations are modest even between studies using similar methods (e.g., r = 0.81 between two RNA-Seq datasets using Illumina instruments [18, 43]). Comparing mRNA studies performed using similar or different methods on a shared set of 4,595 genes revealed a consistent bias toward higher median correlations between studies using similar methods, but these differences were not statistically distinguishable (Fig 1D, no t-test P < 0.05 for differences in correlation when comparing studies employing shared methods versus independent methods after false discovery rate correction).
Between-study correlations quantify the studies’ mean ratio of true variance to total variance, termed the reliability [15, 45, 46] (see Methods). In turn, setting aside sampling error, the maximum observable correlation between any two datasets is equal to the geometric mean of their reliabilities. Because virtually all reported global mRNA–protein correlations involve mRNA and protein levels measured in separate studies, between-study reliabilities are the relevant quantity. The modest reliability values—setting aside those of the same group reporting two studies, which we exclude from this analysis—sharply limit the maximum observable mRNA–protein correlations. This limit has startling consequences: if steady-state mRNA and protein levels actually correlated perfectly (true r = 1.0), then given the median observed between-study correlations in Fig 1C, we would expect to observe mRNA–protein correlations of only r = 0 . 57 × 0 . 62 = 0 . 60.
The data reveal a wide range of modest mRNA–protein correlations with a median of r = 0.54 (Fig 2A) quantified either by the Pearson correlation between log-transformed measurements or the nonparametric Spearman rank correlation (S1 Fig; both measures produce similar results and we employ the former throughout). The largest pair of datasets covers 4,367 genes and shows an mRNA–protein correlation of r = 0.618 (r2 = 0.38, 38% of protein-level variance explained by mRNA levels), close to consensus values [9]. The largest dataset containing replicated measurements of mRNA and protein in at least two studies yields similar correlation values; notably, averaging paired measurements together and correlating the averages increases the apparent correlation (Fig 2B).
This averaging effect has a simple explanation: if experimental noise drives down the mRNA–protein correlation, and noise is to some extent random between studies, then averaging together measurements from different studies will increase the correlation as random noise dilutes out and signal titrates in. However, exploiting averaging comes with hidden dangers when using these data. Averaging requires multiple measurements. Few protein datasets cover even half the genome, and incomplete data tend to be biased toward abundant proteins, as revealed by examining levels in a large dataset when restricted to proteins detected in smaller datasets (Fig 2C); it is plausible that higher-expression proteins correlate more strongly with mRNA levels. We therefore checked for an averaging effect using a subset of the data with a minimum level of reproducibility, at least eight mRNA and eight protein measurements, which includes 549 genes. This high-coverage gene subset does encode more highly abundant proteins relative to the rest of the genome as assessed by western blotting (Fig 2D). As a benefit, however, changes in correlation due to averaging within this subset do not merely reflect underlying systematic changes in the expression levels of the analyzed genes. In this subset, the observed mRNA– protein correlation rises markedly as more measurements are averaged together (Fig 2E), more than doubling in the apparent protein-level variance explained by mRNA level (from 33% to 72%) simply by averaging together more measurements of the same genes. These data strongly indicate that experimental noise substantially reduces the apparent correlation between mRNA and protein levels.
The foregoing analyses involve estimates uncorrected for noise, which as described in the Introduction do not properly estimate the true correlation between the variables being measured. We will first incorporate noise-aware estimates of the true correlation, and then address the more challenging problem of accounting for missing data to arrive at a true genome-scale estimate of the mRNA–protein correlation.
Reduction of correlations by noise can be corrected using information from repeated measurements, assuming the noise is uncorrelated across measurements [24, 46]. Quantitative corrections for correlation attenuation were first introduced more than a century ago by Spearman [24], are widely used in the social sciences [46–48], and have found recent applications in biology [15, 45, 49–52]. Given two measurements each of variables X and Y, each with uncorrelated errors, the true correlation can be estimated using only correlations between the four measurements X1, X2, Y1, Y2 (see Methods):
r ^ X Y true = r X 1 Y 1 r X 2 Y 2 r X 1 Y 2 r X 2 Y 1 4 r X 1 X 2 r Y 1 Y 2 (3)
The correction reflects a simple intuition: the denominator quantifies the reliabilities of the measurements, which determine the maximum observable correlation, and the numerator quantifies the observed correlation using a geometric mean of four estimates and is divided by this maximum value to yield an estimate for the true value. The estimate is not itself a correlation coefficient, and may take values outside (−1,1) due to sampling error [46]. Also note that there is no P-value associated with this estimate; statistical testing for significant association using uncorrected correlation measures remains valid.
To demonstrate and test Spearman’s correction, we applied it to simulated data generated to mimic key features of mRNA and protein data, but with a known underlying correlation and known measurement reliability. We generated data for 5,000 simulated genes with a range of correlations and fixed reliability; a fixed correlation and a range of reliabilities; and a fixed correlation and reliability with a range of data missing at random, or non-randomly, with a detection bias against low-expression genes. We then measured the observed correlation, uncorrected for noise, and used Spearman’s correction to estimate the true correlation. At each set of parameters, we generated 50 transcriptome/proteome pairs to assess reproducibility.
As shown in Fig 3A–3C, noise reduces correlations in a non-negligible way. Given an actual correlation of 0.9, and a reliability of 0.7, higher than the mean values for real data (cf. Fig 1C), the observed correlation has a mean of 0.631±0.009 (standard deviation), whereas Spearman’s correction yields a median value of 0.901±0.007, closely matching the true value. Spearman’s correction performs well over a wide range of reliabilities (Fig 3B) and when data are missing at random (Fig 3C), cases where observed correlations provide a wide range of estimates that are all systematically incorrect. Smaller datasets lead to increased variability of the Spearman estimate due to sampling error (Fig 3C). When faced with data biased toward detection of high-abundance mRNAs and proteins, Spearman’s correction systematically underestimates the true correlation (Fig 3D), as expected due to restriction of range effects.
Using Spearman’s correction on real data, we estimated mRNA–protein correlations for pairs of mRNA- and protein-level studies, obtaining a median corrected correlation of 0.92. Variability due to sampling error was large for small datasets as expected (cf. Fig 3C and 3D), and decreased as dataset size increased, with estimates stabilizing for large datasets (> 3000 genes) at a mean of r = 0.88±0.02 (Fig 2A). This value is echoed by consideration of the largest dataset with two mRNA [38, 43] and two protein [30, 40] measurements each (Fig 2B). For these data, the four observed mRNA–protein correlations are r = 0.60, 0.63, 0.62 and 0.64, and the correlation between mRNA and protein measurements are rmRNA = 0.86 and rprotein = 0.57 respectively, yielding the corrected estimate r ^ true = 0 . 60 × 0 . 63 × 0 . 62 × 0 . 64 4 0 . 85 × 0 . 57 = 0 . 89.
As demonstrated, Spearman’s correction, while useful, does not address biases due to data that are systematically missing. Spearman’s correction also assumes uncorrelated errors, and thus has no mechanism for handling correlated errors arising due to, for example, protocol similarities within a study or use of similar measurement techniques between studies. Actual datasets show evidence for all of these effects (Fig 1).
Extending estimates to the full genome, accounting for structured noise and non-randomly missing data, requires a more sophisticated approach. Even seemingly simple approaches to reduce noise, such as averaging measurements normalized to the same scale, are unworkable as strategies for estimating genome-scale mRNA–protein relationships: only 16 proteins are detected by all 11 protein quantification studies, and these proteins are all highly abundant. Throwing out smaller datasets discards potentially valuable measurements, and it is unclear when to stop, since all datasets are incomplete to some degree.
To address these challenges, we adapted structural equation modeling to admit nonrandomly missing data (see Methods). We introduce a structured covariance model (SCM), adapted with important modifications from recent work [27], that explicitly accounts for structured noise arising from replicates and use of shared measurement techniques, explicitly estimates noise at multiple levels and the nonlinear scaling factors linking underlying variables, and allows inferences of latent covariance relationships with imputation of missing data (Fig 4).
The SCM accurately estimates true correlations in simulated data when substantial data are missing nonrandomly, a case on which Spearman’s correction produces severely biased estimates (Fig 3D). Fitting the SCM to real data yields estimates of whole-genome steady-state mRNA–protein correlation of r = 0.926±0.004 across all 5,854 genes for which an mRNA has been detected in at least one of the 38 mRNA quantitation experiments (Fig 2A). That is, mRNA levels explain 86% of variation in protein levels at the whole-genome scale. We emphasize that the SCM does not involve any attempt to maximize the mRNA–protein correlation or any assumptions about the strength of the correlation.
To examine the influence of low-coverage datasets on the correlation estimate, we re-fit the SCM on data restricted to studies with no more than 60% or 80% missing values (cf. Table 1), resulting in essentially unchanged correlation estimates of r = 0.919 and r = 0.933, respectively. Including these smaller datasets does not alter these estimates significantly.
The SCM integrates all data to produces mean and variability estimates of mRNA and protein levels, yielding a dataset in which mRNA levels have been quantified for 5,854 genes and protein levels have been quantified for 4,990 genes in at least one study.
To evaluate the accuracy of these estimates, we linearly scaled them to molecules per haploid cell using high-quality published values for mRNA per cell and protein per cell. Estimates of the number of mRNA molecules per cell range from 15,000 to 60,000 molecules per cell [36, 53]. A more recent study argued that the earlier, lower estimate resulted from misestimation of mRNA mass per cell and average mRNA length, with 36,000 molecules per cell as a revised estimate also supported by independent measurements [54]. The higher estimate resulted from rescaling the lower estimate to match expression of five genes measured by single-molecule fluorescence in situ hybridization (FISH) [53]. We adopted the 36,169 mRNA molecules per cell estimate [54]. Scaled to 4μg of protein in 1.5×106 cells (2.7pg protein per yeast cell in cells roughly 30 μm3 in size) [55], SCM protein levels sum to just over 35 million protein molecules per haploid cell, similar to the 50 million molecules per cell estimated previously [20] within the variation in total protein extraction from haploid yeast cells (cf. [56], which estimates 4.95pg per cell).
Scaled SCM per-gene means provide the best point-estimates of molecules per cell (Fig 5A), although the correlation between estimates of means is necessarily higher than the estimated true correlation, since each estimate contains error. For a more representative global view of mRNA and protein levels, we draw a sample from the SCM estimates according to each gene’s mean and variance in levels (Fig 5B). Correlations between sampled mRNA and sampled protein levels (r = 0.923) are consistent with the inferred underlying correlation.
We then compared scaled SCM estimates to small-scale gold-standard, independent measurements of absolute mRNA and protein levels not used in our analysis. (No genome-scale gold-standard measurements of mRNA or protein levels exist for yeast or any other organism.) SCM estimates of absolute mRNA levels matched FISH measurements well [53] (average difference of 1.2-fold between estimated and measured levels [Fig 5B], with one outlier estimate overshooting the FISH value by 1.7-fold). Notably, these results demonstrate that the FISH estimates are compatible with roughly 36,000 mRNA molecules per cell during exponential growth as reported [54], and do not require the almost two-fold higher number of cellular mRNAs extrapolated in the FISH study.
Absolute protein levels for a set of 21 proteins differing up to 25,000-fold in cellular abundance have been measured using single-reaction monitoring (SRM) spiked with stable-isotope standards [57]. SCM estimates correlate better with these absolute levels (r = 0.94 between log-transformed values) than does any individual dataset. This includes the only study, using western blotting [30], which reports levels for all 21 proteins (r = 0.90) (Fig 5C, average difference of 1.4-fold between SCM estimates and SRM measurements, compared to 1.8-fold using western blotting). Relative protein levels estimated by integrating multiple datasets using an alternative approach in which noise is not modeled [16] correlate with absolute levels less well (r = 0.88) than do the SCM estimates. The structured covariance modeling approach thus estimates steady-state cellular mRNA and protein levels with an unmatched combination of completeness and accuracy.
To evaluate imputation of missing data, we focused on the 864 genes with a detected mRNA but no protein detected in any of the 11 studies. Some of these genes encode well-studied proteins such as the proteasomal regulator Rpn4p and the cyclin Cln3p, indicating clear false negatives. For a systematic evaluation, we turned to ribosome profiling studies [18], which quantify ribosome-protected mRNA fragments normalized for gene length (ribosome density), providing an estimate of the mRNAs being actively translated in vivo. At least one of five studies under compatible experimental conditions detects ribosomes in the coding sequence of 637 of these 864 genes, suggesting active translation. Normalized ribosome density for this restricted set of genes correlates with the imputed protein levels (Fig 5E, r = 0.55), despite the attenuating effect of range restriction. Because the missing protein data correspond to genes at the detection limit of these ribosome-profiling studies, we predict that many of the remaining genes will be found to produce proteins at low levels during exponential growth. The SCM estimates serve as predictions for the levels of these as-yet undetected proteins.
Our results indicate that the true correlation between steady-state mRNA and protein levels in exponentially growing budding yeast is far higher than previously recognized, explaining the vast majority of variation in protein levels on a log scale. In many previous analyses, this would be equivalent to demonstrating a minor role for other forms of regulation: if the variation in protein levels were a pie, and mRNA levels took a slice, other forms of variation would get only the leftovers. As we will show, such competition is largely illusory.
Positive evidence exists for strong post-transcriptional contributions to protein levels. The dynamic range of protein abundance is wider than mRNA abundance, which must reflect dynamic-range amplification by post- transcriptional regulation [9]. Indeed, wide per-gene variation exists in measurements of translational efficiency [18, 58, 59].
The report that translational activity, estimated by ribosome profiling, explained more than twice the protein-level variation than did measured mRNA levels [18] prompted us to more closely examine these results. We reproduced these comparisons, and found that subsequent ribosome-profiling studies [58–61] confirmed the strong predictive power of ribosome density for the protein levels originally employed, which came from a single study [40] (Fig 6A). We wondered whether these findings might reflect experimental noise that differed between the mRNA and ribosome-footprint measurements in the original study. Correlations using SCM-estimated protein levels are substantially higher for both SCM-estimated mRNA levels and ribosome density measured in all studies, consistent with reduction of noise in the SCM estimates (Fig 6A). SCM-estimated mRNA levels predict protein levels better (r = 0.926) than any of the individual ribosome profiling studies (Fig 6A). This likely reflects remaining noise and systematic bias in the profiling studies, since using Spearman’s correction to estimate the true correlation between ribosome density and protein level yields correlations of r = 0.88 against SCM-estimated protein levels and r = 0.91 using the largest two largest protein-level datasets, measured by mass-spectrometry and western blotting. These results suggest that, contrary to previous reports, measures of translation and mRNA level have essentially equivalent and quite strong predictive power for protein levels.
However, major contributions to protein levels from mechanisms other than mRNA level become obvious upon inspection of the data. The dynamic range of protein expression (from fewer than 50 to more than 1,000,000 molecules per cell [30, 57]) is wider than that of mRNA levels (e.g. from 0.1 to 89 molecules per cell in a landmark early study [36]). In the SCM estimates, the full range of mRNA expression is roughly 10,000-fold (0.02 to 253 molecules per cell on average), whereas the range of protein expression is more than 1,000,000-fold (an average of 0.4 molecules to 1.3 million molecules per cell). Since both mRNA and protein are roughly lognormally distributed, the ratio of log-transformed ranges, 1.6, yields a rough measure of relative variation. (This relative variation is unchanged when attention is restricted to the central 95% of the mRNA and protein distributions to mute outlier effects.) Individual mRNA and protein datasets vary but confirm similar differences in dynamic range (S3 Fig). We address more representative estimates of relative dynamic range below. As previously noted [9], this dynamic-range amplification must involve post-transcriptional variation.
The standard use of a logarithmic scale raises some questions about the interpretation of dynamic range. What does a ten-fold difference mean, if it is between 0.01 to 0.1 molecules per cell rather than between 1 and 10 molecules per cell? Are fractional numbers meaningful? We proceed as though they are. Fractional molecules per cell in a population average may indicate mRNAs or proteins present in only a fraction of cells in the population, which can arise in many ways, from conditional expression (e.g. during a segment of the cell cycle) to incomplete repression (leakiness). Here, estimates of levels reflect the measurements but confer no particular interpretation. We note that no obvious break or cutoff exists in the data or the SCM estimates to suggest a gene-expression threshold below which the biology changes qualitatively.
A consequence of two facts—the higher dynamic range of protein levels than of mRNA levels, and the strong log-log linear correlation between the two—is that steady-state protein levels cannot be (even noisily) proportional to steady-state mRNA levels at the genome scale. In the standard model (P i = τ i δ i M i with protein P and mRNA M for gene i, cf. Eq 2), steady-state protein levels will be roughly proportional to steady-state mRNA levels on a log-log scale assuming translation rates and degradation rates are uncorrelated with mRNA levels. This is most easily seen considering the case of constant translation and degradation rates (τi = τ and δi = δ, respectively) across all genes, such that P i = τ δ M i 1 where we have made explicit the exponent of 1. In this case, ln P i = 1 × ln M i + ln ( τ δ ). Deviations from proportionality can be captured by deviations from a log-log slope of 1.
As described in the Introduction, several studies have estimated slopes very near 1, but have not accounted for error-induced systematic underestimation of slopes due to regression-dilution bias [31]. We therefore used a noise-tolerant regression technique closely related to principal component analysis known as ranged major-axis (RMA) regression [33], which yielded a range of slopes systematically higher than the ordinary least-squares regression slopes (Fig 6B and 6C) and have a median of 1.54. Unlike OLS, RMA regression permits error in both variables and is symmetric, such that regression of Y on X produces the inverse slope to that obtained by regression of X on Y. Other techniques with the same symmetry property but different technical assumptions each yield slopes substantially larger than 1 and larger than OLS estimates (S2 Fig). The estimated slopes for individual pairs of datasets span a wide range, even using RMA and limiting attention to large datasets (Fig 6B), suggesting the existence of systematic biases, toward increased and decreased variance, separating these studies. The presence of such biases in protein-quantitation studies, though not their precise source, has been previously described [62].
The SCM approach, which accounts for both noise and missing data, yields an estimated slope of 1.69, compatible with the range of estimates from noise-aware methods on individual pairs of datasets (Fig 6B and 6C) and also similar to the expectation (1.6) derived from examination of the relative dynamic ranges above. Steady-state protein levels therefore reflect a dramatic multiplication of the transcriptional signal: rather than competing with transcriptional regulation as often assumed, post-transcriptional regulation cooperates.
If translational activity drives much of this cooperative amplification, higher-expressed mRNAs must tend to be more highly translated. Such an effect was noted in passing in the earliest ribosome-profiling study [18]. Several additional such studies satisfying our experimental criteria have been performed since [58–61], which allows us to more thoroughly quantify the relationship between levels of translation and expression. The coverage of these datasets is excellent, so we focus on the 4,435 genes for which all five studies report ribosome density measurements.
Using these data, we found a markedly supralinear relationship between relative translational activity (estimated by median ribosome density), and SCM-estimated mRNA levels (Fig 6D) with a log-log slope of 1.68. As this result implies, translational efficiency (TE) (median ribosome density divided by median normalized mRNA levels within these same studies [18, 58]) increases with SCM-estimated mRNA level (Fig 6E, Spearman rank correlation r = 0.65), with some evidence for a ceiling or saturation effect at high expression levels. These results provide strong evidence that highly expressed genes generate highly translated mRNAs.
RMA regression of ribosome density against SCM mRNA levels yielded a slope of 1.70, compared to a slope of 1.72 of SCM protein levels against mRNA levels (Fig 6F), suggesting that increases in translational activity accompanying elevated mRNA expression are sufficient to generate the broader dynamic range of protein levels relative to mRNA levels.
A subtle possibility is that the SCM estimates have a compressed dynamic range relative to true values, which would inflate both the slope of the translational-activity–mRNA relationship and the correlation between TE and mRNA levels. To address this possibility, we exploited the fact that three of the accompanying mRNA-level measurements in the ribosome-profiling studies [58, 60, 61] were not used in our SCM estimates and therefore constitute an independent, modern, replicated mRNA dataset. The median of these recent measurements correlate well with our SCM estimates (r = 0.90, Pearson correlation on log-transformed values) and the SCM and recent measurements have statistically indistinguishable distributions (S4 Fig). High- and low-expression genes deviate slightly consistent with experimental error in RNA-seq at the low end and compression of the SCM estimates at the very high end. We regressed ribosome densities and protein levels against these recent mRNA levels. Slopes were lower but substantially above 1.0 (1.46±0.02 and 1.49±0.03 [95% confidence intervals] for translational activity and SCM protein levels versus recent mRNA measurements, respectively, Fig 6F). Importantly, calculation of the slope of translational activity versus recent mRNA level does not involve our SCM measurements at all, and thus provides independent evidence that translational activity levels have a wider dynamic range than mRNA levels.
To provide an overall view of relative dynamic ranges, we plotted the distribution of estimated numbers of steady-state mRNAs and proteins per gene. We used ribosome density measurements to estimate the number of ribosomes engaged in translating each mRNA species in a typical haploid cell (Fig 6G), assuming 200,000 ribosomes per cell [63] of which 85% are engaged in active translation (see Methods). Both SCM and recent mRNA levels show a similar dynamic range for most mRNA species, and a narrower distribution than ribosome or protein levels.
In summary, measured variation in translational activity correlates strongly with mRNA level and is sufficient to quantitatively account for the strong nonlinear relationship between mRNA levels and protein levels.
The analysis above illustrates a fundamental asymmetry: although absence of post-transcriptional regulatory processes would produce a perfect mRNA–protein correlation [1], a perfect mRNA–protein correlation need not indicate a negligible post-transcriptional contribution to relative protein levels.
Contrary to the conclusions of many analyses, it is possible for mRNA levels and (for example) translation rates to each explain more than 50% of protein-level variation. Both processes could each contribute 100% of protein-level variation. All that is required is that their contributions not be independent.
To see this, consider the following toy model for regulation of protein levels which does not involve assuming that translation rates are independent of mRNA levels:
∂ P i ∂ t = τ i M i − δ i P i standard model , cf . Eq . 1
with
δi=δτi=α(Miϵi)γln ϵi~N(0,σ)constant degradation ratetranslation rate rises nonlinearly with mRNA levelnoisy evolved correlation between mRNA levels and translation rates
Despite appearances, the functional relationship between translation rates and mRNA levels does not imply or depend on mechanistic properties of transcription and translation. All variance in this model (as in all analyses in the present work) derives from differences between genes, so the functional relationship merely describes an empirical correlation. As described in more depth in the Discussion, such a correlation can arise if genes have evolved differential translational efficiencies tuned to multiply transcriptional signals.
In this toy model, with εi = 1 (or more generally σ = 0), translation rates and mRNA levels reinforce each other perfectly albeit nonlinearly. Under these conditions, steady-state mRNA levels explain 100% of the steady-state protein-level variation on a log scale. Translational regulation also explains 100% of the protein-level variation.
P i = α M i γ δ M i = α δ M i 1 + γ steady-state protein levels ln P i = ln α δ + ( 1 + γ ) ln M i log protein levels are linearly related to log mRNA levels = - ln (δα1/γ) + ( 1 + 1 γ ) ln τ i log protein levels are linearly related to log translation rates
Adding variation to translation rates (σ > 0) and fixing other parameters allows close reproduction of the SCM estimates on several dimensions (Fig 7A and 7B; source code including parameters presented in Methods). Both datasets have similar mRNA–protein correlations (r = 0.926 for experimental data, r = 0.922 for toy model), similar log-log slopes (1.69 for both), and similar dynamic ranges for mRNA and protein levels.
The critical difference between this model and the standard model for protein-level variation, Eq 1, is the evolved strong positive correlation between mRNA levels and translational efficiency. This, too, is evident in experimental data when calculating translational efficiency (Fig 7C, RMA slope = 0.71, Spearman r = 0.62). The correlation is mirrored by the toy model, where translation rate per mRNA and mRNA level can be directly compared (Fig 7D; RMA slope = 0.74, Spearman r = 0.74). The experimental data are substantially missing at the low end, which will tend to attenuate the correlations.
The toy model does not capture the apparent saturation of translational efficiency at high mRNA levels (Fig 7C). Assuming this effect is real, other mechanisms, such as decreased rates of protein turnover, must be added to the toy model to even better reflect the data, which we leave for future detailed modeling.
Our results demonstrate that the frequently reported result that steady-state mRNA levels explain less than half (30–50%) of the variation in protein levels constitutes a significant underestimate. In exponentially growing budding yeast, the best-studied system and source of many of these claims, we find that the true value at the whole-genome scale, taking into account the reductions in correlation due to experimental noise and missing data, is closer to 85%.
Many thoughtful studies have tackled this problem before, arriving at results that match ours on certain dimensions, but via quite different approaches. Previous work has employed versions of Spearman’s correction [15], contended with differences in dynamic range by adopting nonparametric approaches [1, 17], and integrated multiple datasets [8, 11, 16, 17]. All of these works have reached conclusions which differ from the portrait assembled here.
Our analysis transcends these studies on several fronts. The present study incorporates more measurements than any previous work. We distinguish between correlations between measurements and estimates of underlying correlations accounting for between-study reliability, a critical difference that has largely eluded previous work. The structured covariance model natively handles nonrandomly missing data to provide more complete and accurate molecules-per-cell estimates than previous studies. Most importantly, we have not relied on the common but mistaken assumption that different modes of regulation act independently.
A consistent approach in the literature has been to pit transcriptional and post-transcriptional variation against each other, both analytically and rhetorically (e.g., “transcriptional regulation is only half the story” [28]). As we have shown, the data do not fit this competitive paradigm, and even invalidate some of its analytical assumptions, such as independence and non-collinearity. The competitive versus cooperative aspects of post-transcriptional regulation come to the fore when considering the dynamic ranges of gene expression. A wider range of protein than mRNA levels is well-established in a range of organisms [3, 15, 64], and our results further cement this observation. However, dynamic-range variation could be achieved in different ways, captured by two extremes. At one extreme, post-transcriptional regulatory variation is uncorrelated with transcriptional regulation, reducing the contribution of mRNA levels to protein levels. At the other extreme, post-transcriptional variation correlates strongly with transcriptional regulation, multiplying the transcriptional signal with little interference. In both cases, post-transcriptional regulation amplifies the dynamic range of gene expression, but only in the latter case does it also faithfully amplify the mRNA signal itself. Our data clearly and convergently indicate that the biology, at least for this organism under these conditions, lies toward the latter, cooperative extreme.
Coordinated transcriptional and translational signal amplification may explain a range of other observations, particularly regarding proteins-per-mRNA (PPM) ratios, which are frequently used to isolate signs of post-transcriptional regulation. Because post-transcriptional amplification correlates strongly with mRNA levels, PPM will remain correlated with mRNA, and as a consequence, any sequence features correlated with mRNA will tend to correlate with PPM as well. As an example, amino-acid composition correlates with PPM in yeast [17], with valine/alanine/glycine frequencies higher in high-PPM sequences and leucine/asparagine/serine frequencies lower in high-PPM sequences. These are precisely the same amino acids previously shown to vary most strongly in frequency, in the same directions, with increasing mRNA abundance [65]. Similarly, many other correlates of PPM are also correlates of mRNA levels (codon bias, tRNA adaptation), including mRNA level itself [1, 11]. For features such as codon bias, which arises in response to selection for translational efficiency [66], association with increased PPM might seem an obvious causal link, but because codon bias strongly associates with mRNA level, the null expectation is that it will correlate with PPM even if codon bias had no effect on translational activity at all. Analyses of the determinants of protein levels must contend with the collinearity and non-independence of contributing processes.
The strong correlation between steady-state mRNA and protein levels may seem to validate the use of mRNA levels as relatively faithful proxies of protein levels. We urge caution, as a tempting conclusion—that mRNA changes serve as faithful proxies for protein changes—does not follow. Attempts to infer the correlation between mRNA and protein changes from steady-state mRNA–protein correlations confuse two distinct and complex phenomena. The genome-scale relationship between mRNA levels and protein levels is an evolved property of the organism, reflecting tuning by natural selection of each gene’s transcriptional and post-transcriptional controls, rather than a mechanistic input-output relationship between mRNA and protein mediated by the translational apparatus. Two genes with steady-state mRNA levels differing by 10-fold may have 500-fold differences in protein levels due to evolved differences in their post-transcriptional regulation. These evolved steady-state differences do not predict how the protein levels for these genes will change if both mRNAs are induced 10-fold, because evolution does not occur on this timescale; the changes in protein levels are instead dictated by the cellular mechanisms of translation.
An important intermediate case between the evolutionary and mechanistic cases is variation in mRNA and protein levels in individuals across a genetically diverse population. The potential for correlations between mRNA and protein relies upon substantial true variance in mRNA levels. In population-variation studies, one expects relatively few variants and resulting variation far lower than the orders of magnitude considered here. Correspondingly, in such studies mRNA-change–protein-change correlations may be low even given a strong underlying link between mRNA and protein levels.
If the nonlinear multiplication of mRNA levels into protein levels is an evolved property, what mechanism(s) has evolution exploited? The present work supports a particular class: the increased density of ribosomes on high-expression mRNAs, with variation sufficient to account for the nonlinearity, suggests increased rates of translation initiation as the major contributor. Correspondingly, recent work has shown that in yeast and a wide range of other organisms, the stability of mRNA structures in the 5’ region weakens as expression level increases, favoring more efficient translation initiation [67], and wide variation in heterologous protein levels can be achieved by varying mRNA stability near the initiation site [68, 69].
Several limitations still attend our approach. By assuming single multiplicative errors per experiment, we ignore variation in per-gene error which may be systematically different between low- and high-expression genes and/or systematically affect particular measurement techniques [62]. For example, limitations in the dynamic range of a measurement technique will tend to compress the resulting measurements, causing such systematic errors. Our model does not contend with distortions possibly imposed by alterations to 3’ regulatory signals (e.g. tagging with affinity epitopes [30] or fluorescent proteins [39] to enable protein detection), or with variability in quantification due to propensities of particular mRNAs to be more efficiently sequenced or for their protein products to be unusually amenable to mass-spectrometric detection. The lack of any gold-standard genome-scale measurements hinders detection of such biases. Our results underscore the urgent need for such standard measurements of absolute mRNA and protein levels to enable identification and correction of systematic errors in established and emerging gene-expression measurement techniques. More sophisticated models for experimental error at many levels, which would be informed by but need not wait for such gold-standard measurements, also promise to provide higher-fidelity biological estimates from existing data.
We infer a higher mRNA–protein correlation (r = 0.93) here than when using an earlier, related model [27] (r = 0.82), a difference we attribute to two factors. First, the present analysis stratifies by measurement technology using all data, whereas the previous estimate did not, although in that study, stratifying by technology on a reduced dataset yielded r = 0.86 [27]. Here, using all data and treating technology-related experimental noise separately from other sources of noise, we are able to average out more systematic technology biases, likely producing superior estimates of the associated measurement variability and reducing noise-induced attenuation of the mRNA–protein correlation. Second, in the present analysis, population-averaged protein levels and mRNA levels are constrained to each have a single underlying variance, whereas in the earlier study each experimental replicate had a separate variance. Inference of artificial experiment-specific variances spread variability across experiments (overfitting), where in the present analysis, we adopt the more biologically plausible stance that the true underlying mRNA and protein population-average distributions each have a characteristic variance which is measured by each experimental replicate. The present model, deprived of extra parameters, infers higher correlations.
Our study considers a single well-studied growth condition for a single well- studied organism, raising questions about how to generalize this work. The principles of accounting for noise, but not precise results, can and should be extrapolated to regulatory contributions in other settings and other organisms. An influential study on mouse fibroblasts measured mRNA and protein levels and degradation rates for thousands of genes [3], concluding that mRNA levels explained 41% of the variation in protein levels, with most variation instead explained by translational regulation. Our results indicate many ways in which the results of this study could be profitably revisited. Indeed, a recent follow-up study concluded that, once effects of error and missing data were accounted for, mRNA levels explain 75% or more of the protein-level variation in these data [21].
The protein regulatory environment of rapidly dividing cells differs from that of many other cellular states. The faster cells divide, the more rapidly protein molecules partition into daughter cells, adding an approximately constant amount to all protein removal rates and consequently reducing between-gene variation in these rates. This will tend to increase the dependence of protein levels on mRNA levels, and decrease the dependence on degradation rates, during proliferation.
In addition to cellular state, regulatory contributions depend on timescale. Post-transcriptional processes must dominate protein-level changes within seconds to a few minutes of a stimulus or signal; transcriptional responses, particularly in eukaryotes, where transcription and translation are uncoupled, are all but powerless at this timescale. As such, the notion of general determinants of protein levels without regard to timescale has questionable utility. A final theme emerging from our study is that careful empirical studies, coupled with noise-aware analyses, are needed to determine regulatory contributions for any cellular condition of interest at any timescale.
Let us assume we wish to measure latent variables ϕ and ψ but, due to noise, actually observe variables X = ϕ+εX and Y = ψ+εY where the random noise variables εX and εY are uncorrelated and mean zero. The reliability
α X = Var ( Φ ) Var ( X ) = Var ( Φ ) Var ( Φ ) + Var ( ϵ X ) (4)
quantifies the ratio of signal variance to total (signal plus noise) variance in X. Given two random variables X1 and X2 representing replicate measurements of ϕ, the latent (true) variance is Cov(X1,X2) = Cov(ϕ+εX1,ϕ+εX2) = Cov(ϕ,ϕ) = Var(ϕ), where the error terms vanish because they are uncorrelated by assumption. Thus, the Pearson correlation between replicates is
ρ X 1 , X 2 = Cov ( X 1 , X 2 ) Var ( X 1 ) Var ( X 2 ) = Cov ( Φ , Φ ) Var ( X 1 ) Var ( X 2 ) = Var ( Φ ) Var ( X 1 ) Var ( Φ ) Var ( X 2 ) = α X 1 α X 2 , (5)
which is the geometric mean of the reliabilities of the two measurements.
We wish to infer the Pearson correlation coefficient between latent variables ρ ϕ , ψ = Cov ( ϕ , ψ ) Var ( ϕ ) Var ( ψ ) but, due to noise, we observe random variables
ρ X , Y = Cov ( X , Y ) Var ( X ) Var ( Y ) = Cov ( Φ , ψ ) ( Var ( Φ ) + Var ( ϵ X ) ) ( Var ( ψ ) + Var ( ϵ Y ) ) ≤ ρ Φ ψ . (6)
with equality only when Var(εX) = Var(εY) = 0 (i.e. there is no noise).
Uncorrelated noise has no average effect on the numerator because errors cancel (see above), but the error terms in the denominator do not cancel. This effect additively inflates the variances in the denominator, biasing the observed correlations downward relative to the truth. Given the reliabilities αX and αY, Spearman’s correction is given by
ρ Φ ψ = ρ X Y α X α Y (7)
To estimate ρϕψ, we need estimates of ρXY, αX and αY. A natural estimator replaces these population quantities with the sample correlation coefficients, rxy, α ^ x and α ^ y with
α ^ x = r x 1 , x 2 α ^ y = r y 1 , y 2
where x1, x2 are realizations of X and y1,y2 are realizations of Y. These replicates are used to estimate reliabilities.
The true correlation, ρϕ,ψ, can then be estimated using only correlations between measurements:
r ^ Φ ψ = r x 1 y 1 r x 2 y 2 r x 1 x 2 r y 1 y 2 = r x 1 y 1 r x 2 y 2 α ^ x α ^ y
We extend this estimate to
r ^ Φ ψ = r x 1 y 1 r x 2 y 2 r x 1 y 2 r x 2 y 1 4 α ^ x α ^ y
which has the further desirable properties of exploiting all pairwise correlations and being independent of the choice of indices.
Taking this approach to its logical conclusion, given a set of N measurements of ϕ and M measurements of ψ, we propose the estimator
r ^ Φ ψ = ( ∏ i , j N , M r x i , y j ) 1 N × M ( ∏ i < i ′ N r x i , x i ′ ) 1 N ( N - 1 ) ( ∏ j < j ′ M r y j , y j ′ ) 1 M ( M - 1 ) ,
where the numerator is the geometric means of all pairwise correlations, and the demoninator is the square root of the product of the geometric means of the pairwise reliability estimates (correlations between measurements) for each variable.
We gathered 38 measurements from 13 studies measuring mRNA expression, and 20 measurements from 11 studies measuring protein concentrations, yielding a total of 58 high-throughput measurements of mRNA and protein levels from a maximum of 5,854 genes in budding yeast. The measurements were taken using different technologies including custom and commercial microarrays, competitive PCR, high-throughput RNA sequencing, flow cytometry, western blotting, scintillation counting of 35S-labeled protein on 2D gels, and liquid chromatography coupled to tandem mass spectrometry (LC-MS/MS) using a range of labeling and quantification techniques. All yeast cultures were haploid S.cerevisiae growing in shaken liquid rich medium with glucose between 22°C and 30°C and sampled during the exponential growth phase. Details of the datasets are summarized in Table 1.
For analytical purposes, we treat data from one study [38] which performed two independent measurements using different methods as two studies (RNA-Seq and microarray), one per method. This study’s RNA-Seq employed a single-molecule sequencing method, smsDGE; we treat this as an RNA-Seq dataset.
We downloaded ribosome-profiling data from the primary sources for five studies [18, 58–61]. Within-study replicates were averaged; for one study of translational inhibitors [59], the no-inhibitor and 1×-inhibitor replicates were averaged. Summary ribosome density and mRNA levels for these datasets were computed by log-transforming all instances of each type of measurement, subtracting the grand median value, and exponenentiating the per-gene median of the resulting values. To ensure the measurements were independent of SCM estimates, we excluded the mRNA levels from Ingolia and colleagues from the mRNA estimates, leaving three studies (one ribosome-profiling study did not report mRNA levels [59]). To preserve the measured dynamic range in the data, no scaling of variance was performed. Translational efficiency was computed as the median normalized ribosome density (five studies) divided by the median normalized mRNA level (three studies), ensuring these results are independent of the SCM estimates.
To estimate the number of ribosomes translating each mRNA species, we multiplied median ribosome densities (which are proportional to ribosomes per nucleotide) by gene length, then normalized the resulting distribution to sum to 200,000 ribosomes per haploid cell [63]. We further assumed that approximately 85% of ribosomes are engaged in active translation during rapid growth [70].
Raw data (with missing values), data normalized and imputed using the SCM, and merged molecules-per-cell estimates are archived in Dryad (http://datadryad.org) with DOI doi:10.5061/dryad.d644f.
All analyses were carried out using R [71] using custom scripts which may be downloaded from GitHub (http://github.com/dad/mrna-prot). Regression analyses using major-axis (MA), scaled major-axis (SMA), and ranged major-axis (RMA) regression were performed using the package lmodel2. RMA was performed using interval ranges.
The model has two components: an observation model p(Ii,j∣Xi,j), which provides the probability of observing a value for mRNA/protein i in replicate j, given the underlying mRNA/protein level, and a hierarchical model p(Xi,j∣…) for the underlying mRNA/protein levels themselves. The full model is specified as
Xi,j=Li,l[ j ]Gl[ j ]+Ti,t[ j ]+Ei,k[ j ]+Ri,j+νj
(8)
L i ∼ 𝓝 2 ( 0 , Ψ ) (9)
T i , t ∼ 𝓝 N T ( 0 , τ t ) (10)
E i , k ∼ 𝓝 ( 0 , ξ k ) (11)
R i , j ∼ 𝓝 ( 0 , θ j ) (12)
p ( I i , j = 0 | X i , j = x ) = 1 1 + exp ( - η k [ j ] 0 - η k [ j ] 1 X i , j ) . (13)
Random variables Li,l correspond to the true denoised protein (l = 1) and mRNA (l = 2) levels, for mRNAs and proteins i = 1, …, N, and Li = [Li,1,Li,2]′. The random variables Ti,t and Ei,k capture common technological variation and batch effects, respectively, t = 1, …, Nt, k = 1, …, NE. Ri,j are experimental noise for replicate j = 1, …, NR.
Both technology effects and batch effects between experiments are assumed to be independent, Cov(Ti1,t1,Ti2,t2) = 0 if t1 ≠ t2, and Cov(Ei1,k1,Ei2,k2) = 0 if k1 ≠ k2. Measurement noise is independent between replicates, Cov(Ri1,j1,Ri2,j2) = 0 if j1 ≠ j2.
The parameters νj corresponds to the normalizing constants of the mRNAs/proteins within a replicate (on the log-scale, normalizing constants become offsets). The coefficient Gl represents the log-variance of the denoised true mRNA or protein. The ratio A = G a b u n d G a b u n d represents the amount of post-transcriptional amplification of mRNA to protein. At steady state we expect
P i ≈ M i A
for protein Pi and mRNA Mi.
This model falls into the class of models that were extensively studied in an earlier work [27]. The results are largely insensitive to deviations from parametric modeling assumptions and to several details of prior specifications.
Below is R code to reproduce the toy model in Fig 7.
# Random number seed
set.seed(115)
# Number of genes
n <- 5854
# Exponent of empirical (evolved) relationship between steady-state
# mRNA levels and translation rates
gamma <- 0.56
# Scaling factor, 1/time
alpha <- 0.1
# Degradation rate, 1/time
delta <- 0.001
# Standard deviation of mean-zero variation added to log mRNA levels to yield
# unscaled log translation rates
te.variation <- 1.1
# Steady-state mRNA levels in molecules/cell (log-normal)
# Mean and variance are equal to those of the SCM mean estimates
log.m <- rnorm(n, mean = 1.09, sd = 1.25)
m <- exp(log.m)
# Translation rate -- add log-normal variation to, and scale, mRNA levels
tau <- alpha*exp(log.m + rnorm(n,mean = 0,sd = te.variation))^gamma
# Steady-state protein levels in molecules/cell (log-normal)
prot.variation <- 0.55
p <- (tau/delta)*exp(log.m + rnorm(n, mean = 0, sd = prot.variation))
# Plot protein vs. mRNA
plot(m, p, log = ‘xy’, pch = 16, las = 1,
xlab = ‘mRNA level (mol./cell)’, ylab = ‘Protein level (mol./cell)’)
# Plot translation rate vs. mRNA
plot(m, tau, log = ‘xy’, pch = 16, las = 1,
xlab = ‘mRNA level (mol./cell)’,
ylab = ‘Translation rate per mRNA (proteins/sec)’)
|
10.1371/journal.pntd.0003584 | Association of Symptoms and Severity of Rift Valley Fever with Genetic Polymorphisms in Human Innate Immune Pathways | Multiple recent outbreaks of Rift Valley Fever (RVF) in Africa, Madagascar, and the Arabian Peninsula have resulted in significant morbidity, mortality, and financial loss due to related livestock epizootics. Presentation of human RVF varies from mild febrile illness to meningoencephalitis, hemorrhagic diathesis, and/or ophthalmitis with residual retinal scarring, but the determinants for severe disease are not understood. The aim of the present study was to identify human genes associated with RVF clinical disease in a high-risk population in Northeastern Province, Kenya.
We conducted a cross-sectional survey among residents (N = 1,080; 1–85 yrs) in 6 villages in the Sangailu Division of Ijara District. Participants completed questionnaires on past symptoms and exposures, physical exam, vision testing, and blood collection. Single nucleotide polymorphism (SNP) genotyping was performed on a subset of individuals who reported past clinical symptoms consistent with RVF and unrelated subjects. Four symptom clusters were defined: meningoencephalitis, hemorrhagic fever, eye disease, and RVF-not otherwise specified. SNPs in 46 viral sensing and response genes were investigated. Association was analyzed between SNP genotype, serology and RVF symptom clusters. The meningoencephalitis symptom phenotype cluster among seropositive patients was associated with polymorphisms in DDX58/RIG-I and TLR8. Having three or more RVF-related symptoms was significantly associated with polymorphisms in TICAM1/TRIF, MAVS, IFNAR1 and DDX58/RIG-I. SNPs significantly associated with eye disease included three different polymorphisms TLR8 and hemorrhagic fever symptoms associated with TLR3, TLR7, TLR8 and MyD88.
Of the 46 SNPs tested, TLR3, TLR7, TLR8, MyD88, TRIF, MAVS, and RIG-I were repeatedly associated with severe symptomatology, suggesting that these genes may have a robust association with RVFV-associated clinical outcomes. Studies of these and related genetic polymorphisms are warranted to advance understanding of RVF pathogenesis.
| The underlying risk factors that lead to severe human Rift Valley Fever disease are unknown, but are likely multi-factorial. Host factors, such as innate immune genetic makeup, are likely important determinants of disease phenotype. This study investigated the association of 46 single nucleotide polymorphisms (SNPs) in genes encoding innate immune receptors, signaling pathways or mediators with RVF disease phenotype. Of the 46 SNPs tested, TLR3, TLR7, TLR8, MyD88, TRIF, MAVS, and RIG-I were repeatedly associated with severe RVF symptomatology, suggesting that these genes may have a robust association with RVFV-associated clinical outcomes.
| Rift Valley fever virus (RVFV) is a negative strand RNA virus of the family Bunyaviridae. Episodic epidemics of Rift Valley Fever (RVF) present a significant natural threat to human health in many countries of Africa and the Middle East, causing retinitis, encephalitis and hemorrhagic fever [1,2]. Epizootics of RVFV also seriously affect livestock, including sheep, cattle, goats, buffalo, and camels, creating serious economic disruption and risk of famine [3]. Two of the largest RVF outbreaks have occurred in Kenya over the last decade, the first in 1997–98 [4], and another more recently in 2006–2007 [5]. Both epidemic human disease, (including hemorrhagic fever), and enzootic livestock disease, (including excess mortality and miscarriage), are most prevalent in semi-arid areas that experienced prolonged excess rainfall during El Nino-Southern Oscillation (ENSO) weather anomalies [6]. Given the recent US experience with West Nile Virus, we could expect that, after either accidental or intentional introduction, RVFV will have the potential to become a widespread multi-state or multinational problem in North America.
Our ongoing field studies aim to better define the epidemiology of RVF viral transmission at the local community level [7–11]. However, little is known about the pathogenesis of the variable disease progression observed between different RVFV-infected human subjects. In communities where 20–30% of persons are exposed, only about 1% of infections progress to severe liver dysfunction and hemorrhagic disease, and late onset encephalitis is rare, although 10–30% develop some form of anterior or retinal eye disease [12,13]. Currently there is no specific treatment for RVF.
It remains unclear why the majority of infected humans recover from RVFV infection after only a brief febrile illness. Evidence from experimental animal models suggests that early activation of innate immunity provides the greatest protection against lethal RVFV infection [14,15], and that differences in interferon-mediated response pathways [16–18] could be responsible for resistance to lethal infection [18–20]. Later-onset, adaptive immunity (with the production of neutralizing antibody responses) likely also plays a role in modulation of RVFV-infection associated disease. However, given the rapid time course of lethal disease progression, with most major symptoms developing within the first week of illness [21], a study of the variability in innate immune responses appears to be the logical first step to elucidate inter-subject variation in disease progression.
Several classes of innate receptors are important in host anti-viral defense, including membrane bound Toll-like Receptors (TLRs) [22], cytoplasmic DExD/H box RNA helicases such as retinoic acid-inducible gene-I (RIG-I) and melanoma differentiation-associated gene 5 (MDA5) [23–26] and NOD-like receptors (NLRs) including inflammasomes [27–32]. TLRs are innate receptors that recognize specific structures expressed by microorganisms. Surface TLR2 and −4 recognize viral proteins including hemagglutinin of measles virus [33] and components of HSV [34] and CMV [35] and RSV [36]. Endosomal TLRs act as receptors for nucleic acids, including TLR3 (dsRNA [37]); TLR7 and TLR8, (single-stranded RNA [38,39]); and TLR9, (unmethylated CpG DNA motifs [40]). There is strong evidence in support of a role for endosomal TLRs in the detection of viruses including Influenza virus and Vesicular Stomatitis virus (VSV) (TLR7) and HSV (TLR9 in certain cell types) [38,41–43]. Signaling by all TLRs originates from a conserved intracellular domain (Toll–IL-1–resistance; TIR), which mediates recruitment of members of a family of adaptor molecules. Recruitment of the common adaptor, myeloid differentiation factor 88 (MyD88) [44,45] leads to the interaction and activation of the IRAK family members [46] and the subsequent activation of TRAF6 [47] resulting in NF-κB activation. Activation of the Interferon Regulatory Factors (IRFs), important mediators of IFN gene transcription also occurs downstream of the TLRs. Thus, these pathways could potentially play an important role in modulating the severity of RVFV-associated disease in humans. A role for TLR3 was demonstrated in a murine model of another phlebovirus, Punta Toro virus [48]. Other reports have shown a protective role for poly I-C (ligand for TLR3 and MDA5) when used as a pretreatment prior to RVFV infection with virulent ZH501 strain [14,49]. In our previous study, TLRs did not play a predominant role in IFN production [50]; however, their role in human innate responses and RVF disease pathogenesis remains unclear.
Viral nucleic acid recognition can also occur via the RNA helicases RIG-I and MDA5 [23,24]. Both proteins are expressed in the cytoplasm and contain caspase recruitment domains (CARDs) as well as a C-terminal region harboring ATP-dependent RNA helicase activity [24]. RIG-I and MDA5 activate downstream signaling via the adaptor MAVS (mitochondrial anti-viral signaling [51], also called IPS1 [52,53], CARDif [54] or VISA [55]), which relays signals to downstream kinases to trigger IFN gene transcription. RIG-I is required for triggering anti-viral responses against several classes of RNA viruses including (Flaviviridae, Paramyxoviridae, Orthomyxoviridae and Rhabdoviridae) [52], whereas MDA5 is required for the response against picornaviruses like encephalomyocarditis virus (EMCV) [56,57]. RIG-I senses viral and synthetic RNA containing 5’-triphosphate caps whereas MDA5 detects synthetic poly(I-C) in vivo, although the nature of the viral ligand for MDA5 remains unclear [53,56,58]. Our previous work emphasized the importance of the RNA helicase adaptor MAVS for RVFV induced type I IFN production in primary immune cells as well as for protection against mortality and morbidity during mucosal challenge in mice [50]. We showed that type I IFN responses were mediated through RIG-I in mice and in vitro human cell systems, although MDA5 also played a role at the earliest time points of viral entry.
In the present study, we hypothesized that distinct RVFV-associated clinical syndromes are related to differences in early innate host responses to viral infection, and that variation in these host responses may be associated to differences in the makeup of innate immune response pathways. In this first genetic epidemiologic study of human RVF, we sought to examine the association between genes in innate immunity pathways and clinical phenotypes linked to acute RVFV infection. This manuscript describes the genetic associations we discovered among variants within host immunologic pathways that may influence susceptibility to RVFV-associated disease.
The study protocol was approved by the University Hospitals Case Medical Center Institutional Review Board (IRB), Cleveland, Ohio (No. 11-09-01) and the Ethical Review Committee of the Kenya Medical Research Institute, Nairobi, Kenya (Non-SSC Protocol No. 195). Before participation, written informed consent was obtained from adult study subjects, and parents provided written informed consent for participating children; children over 7 years of age also provided individual assent.
Study participants were recruited from six villages located in the Sangailu area of Ijara District, located in Northeastern Province, Kenya (Fig. 1). All local residents were eligible for participation, with the exception of those living in the area for less than 2 years, and children < 1 year of age, who were excluded. After an initial demographic census was performed, consented subjects were enrolled, surveyed via structured interview for potential RVFV exposure history and past symptoms suggestive of RVF, and examined by a nursing officer with particular attention to current visual acuity and eye disease, as previously described [10,59]. Whole blood was collected by phlebotomy (∼ 5 mL venous blood samples from persons > 5 years of age and 1 mL from children under the age of 5). Individual sera and associated blood clots were separated and stored frozen at −80°C.
The study sample was representative of the local mix of 99% Somali ethnicity, and < 1% Bantu, Indian, or other Asian. Out of 1134 household residents identified, 1128 completed survey questionnaires (parents served as proxies for young children), and 1110 completed a basic physical examination and vision check. A total of 1114 provided blood for anti-RVFV antibody screening, and 1082 provided full data from survey, exam and serology testing. This paper's genetic analysis focuses on 363 individuals from an RVF-endemic study area, Ijara Constituency of Garissa County. The study group included 117 individuals who were antibody-seropositive for RVFV and had DNA available for analysis, and 246 unrelated local control subjects from nearby households or villages. The study subjects (seropositive and seronegative) were sampled from 251 households in Sangailu Location (centered around coordinates 1 deg. 19 min S, 40 deg. 44 min E). In this area, residents per household ranged from 1 to 9 (median = 3), and for our study 1 to 4 persons were tested per household (average = 1.5 per household).
To test for evidence of past RVFV infection, serum specimens were screened for the presence of anti-RVF IgG via ELISA [4,7]. Briefly, high protein binding plates (Corning) were coated with RVFV variant rMP-12 viral antigens prepared in Vero cell lysates, and blocked in 5% non-fat dairy milk. Serum samples were diluted 1:100 in PBS/5% milk solution and allowed to absorb for 1 h at 37°C. After washing, a HRP-conjugated secondary anti-human IgG antibody was applied, again in the milk solution at a 1:2000 dilution. Plates were incubated at 37°C for 1 h then developed using a TMB substrate (Thermo) and absorbance was read at 405nm. Each sample was run in duplicate, and OD values were normalized to background values from wells coated with uninfected Vero cell lysate and averaged. Samples were considered positive with OD values greater than the mean + 2 SD for pooled control sera obtained from unexposed North Americans.
Genomic DNA was isolated from frozen blood clots using a 96 well DNeasy Blood and Tissue kit (Qiagen) with some slight modifications. Approximately 500 μL of thawed blood clot material was homogenized in ALT lysis buffer using a mixer mill (Retsch). Proteinase K (Qiagen) was added to each tube and incubated at 56°C for 60 min with occasional agitation followed by pulse centrifuged at low speed to pellet debris. Supernatant was removed and DNA was eluted using spin columns according to the manufacturer’s recommendations. Total DNA was quantified using the Quant-iT PicoGreen dsDNA Assay Kit following the manufacturer’s recommendations (Life Technologies). The results were verified on random samples by spectrophotometry (NanoDrop 1000, Thermo Scientific). Low yielding samples were concentrated by re-precipitating the DNA in ethanol in the presence of 20mg/ml glycogen (Thermo Scientific) and re-suspended in TE buffer (Qiagen).
This study focused on genes encoding molecules likely to be involved in early innate immune responses to RVFV including: IL6, IL6R, TLR3, TLR7, TLR8, TRIF (TICAM1), MyD88, RIG-I (DDX58), LGP-2 (DHX58), MAVS, IFNAR1, IFNB1, Mda5 (IFIH1), CCR5, DC-SIGN (CD209), and CFH. Single nucleotide polymorphisms (SNPs) within these genes were chosen if they were in the promoter region, the coding region of the gene, or in the untranslated regions (3’UTR or 5’UTR) of the gene, and having at least a 10% minor allele frequency in the Maasai (MKK) or Luhya (LWK) Kenyan HapMap populations [60–62]. The exception to this was CD209, where an intergenic SNP was chosen, as no putatively functional SNPs met the allele frequency criteria in HapMap. A total of 48 SNPs were genotyped using the Illumina VeraCode platform. Two SNPs failed quality control because of poor intensity; the remaining SNPs were all in Hardy-Weinberg equilibrium. Allelic frequencies are included in S1 Table.
RVFV-related disease phenotypes were defined on the basis of subjects’ self-reported symptoms on the study intake questionnaires. The time period of interest for occurrence was any time during or after the 2006–2007 RVF epizootic in Northeastern Province. The nineteen questions included in the symptom review covered known RVF complications including isolated eye symptoms (eye pain, scleral injection, poor vision or blurry vision), central nervous system symptoms (photophobia, meningismus, vertigo/dizziness, reduced consciousness, confusion, coma, or seizures), symptoms of a hemorrhagic diathesis (bleeding gums, non-traumatic bruising, hematemesis, hematochezia), or non-focal systemic symptoms (fever, malaise, back ache, nausea, anorexia).
Presence vs. absence of the clinical phenotype clusters plus positive RVFV serology were the phenotypes of interest. SNP genotypes were analyzed both according to a dominant model with respect to the minor allele, and according to an additive model with increasing counts of the minor allele (e.g., AA = 0, AB = 1, BB = 2); these analyses were conducted using PLINK [64] (http://pngu.mgh.harvard.edu/∼purcell/plink/). If there were fewer than 10 rare homozygotes plus heterozygotes, the Fisher’s exact test was used to define related p values for significance testing. If there were fewer than 10 homozygotes for the rare allele in either the affected or unaffected group for a given trait, the results from the additive model were not considered. In addition, the categorical cluster score trait was analyzed using Kendall’s tau to model increasing severity of the phenotype in association with increasing counts of the minor allele, using dominant and additive models as before; the analyses were conducted using SPSS version 20.
Haplotype association analyses were conducted using the proxy association method in PLINK. The most likely haplotypes were evaluated using the EM algorithm. Two statistical tests are reported: an omnibus test, which compares the distribution of most likely haplotypes in cases versus controls, and haplotype-specific tests, comparing the presence versus absence of that specific haplotype in cases versus controls. Finally, gene-level and pathway-level analyses were conducted using PLINK.
Because this was an exploratory pilot study, results significant at α = 0.10 are presented. Of note, a large number of associations (46 SNPs) were tested in our exploratory analysis, and none of the results were significant after Bonferroni correction for multiple test comparisons.
The analysis included 363 individuals, of whom 219 (67.1%) were female (Table 1). A total of 117 (32.2%) were found to be seropositive for RVFV by ELISA.
Subject- reported RVFV-associated symptoms were clustered to system-related and severity-related groupings to facilitate association analysis with genetic polymorphisms (Table 1). Many individuals (72%) reported at least one non-specific RVF-associated symptom, such as past fever (66%) or malaise (51%). Meningoencephalitis symptoms (photophobia, meningismus, vertigo/dizziness, reduced consciousness, confusion, coma, or seizures) were common and at least one symptom was reported by 44.6% of individuals. Individuals reporting at least 3 symptoms of meningoencephalitis, a more stringent classification, were much less common (8.3%). Any hemorrhagic (HE) fever-associated symptom was reported by 9.4%; three or more HE symptoms were reported by only 4 subjects (1.1%). By contrast, eye disease symptoms were commonly reported (38.6% for any single symptom; 29.5% for 2 or more symptoms).
Our computer-assisted severity group clustering, based on overall number and type of symptoms per individual, classified 28/177 (24%) seropositive subjects as having been more mildly symptomatic, 44/117 (38%) as moderately symptomatic, and 45 (38%) as more severely symptomatic (Table 1).
To analyze associations of clinical disease with individual SNPs, a single major locus model was first applied, and significance considered at α = 0.10. Using a dominant model, a number of SNPs showed association with clinical phenotypes with p < 0.10 (Table 2). We looked at several polymorphisms in complement factor H (CFH), a gene previously found to be associated with eye disease in published genome-wide association studies [65,66] as well as host susceptibility to meningococcal disease [67]. A single polymorphism in CFH (rs1061147) was associated with the presence of any eye symptom (p = 0.059). However, two other SNPs in this gene (rs1065489; rs3753396) were not significantly associated with individual symptoms or with clusters of clinical symptoms (S3 Table).
A polymorphism in the gene of the pro-inflammatory cytokine interleukin-6 (IL-6) (rs2069849) was associated with presence of 3 or more non-specific symptoms (p = 0.025, Table 2). Two SNPs in the 3’ UTR region of the IL-6 receptor were associated with meningoencephalitis or hemorrhagic symptoms with a p < 0.10 (rs4072391; rs7514452).
Also detailed in Table 2, several single polymorphisms in the RNA helicase pathway showed associations with clinical symptoms. A polymorphism in DDX58 (RIG-I) (rs2274863) was associated with the subject report of 3 or more ME symptoms (ME3, p = 0.026), and with the past experience of any ME symptom (p = 0.03). A SNP in the 3’ UTR of the common adaptor MAVS (rs3746660) was significantly associated with the experience of any eye symptom (p = 0.041), any ME symptom (p = 0.059) and also with a history of having had two or more eye symptoms (eye2, p = 0.0598) and two different SNPs were associated with positive serology (rs7262903; rs7269320).
In the TLR pathway, TLR7 SNP rs864058 was associated with positive RVFV serology (p = 0.032) as well as a history of any hemorrhagic symptom (HE_any, p = 0.02803). TLR8 SNPs rs3747414 and rs5744088 were associated with having three or more meningoencephalitic symptoms (ME3) and positive serology, respectively. The SNP rs6853, in the adaptor molecule MyD88 which mediates signaling by both TLR7 and TLR8, showed association with the presence of at least one HE symptom (p = 0.01776). The adaptor TRIF (TICAM1; rs229151) was associated with ME3 (Fisher’s exact p = 0.002), as well as eye2, any eye, ME_any, and the presence of non-specific symptoms.
An additive analysis was next performed to examine the impact of multiple copies of the polymorphisms of interest. Because of the rarity of some of these phenotypes and SNP minor alleles, the dominant model (as shown in Table 2) was more significant, with a few SNPs showing robust associations in the additive model (Table 3). CCR5, RIG-I, LGP2 and IFNAR1 all had SNPs associated with clinical symptom traits at the p < 0.1 level (Table 3).
Next, we examined the association between the SNPs we selected for study and disease severity as determined by cluster analysis, again considering significance at α = 0.10 level. As shown in Table 4, rank correlation of SNP genotype with severity cluster scores revealed significant associations with exon SNPs in the LGP2 helicase (p = 0.08), the MAVS adaptor molecule (p = 0.047), and the interferon-receptor IFNAR1 (p = 0.013).
To better understand the effect of polymorphisms in the overall genes of interest (versus specific SNP associations), we conducted gene- and pathway-level tests of association. In this analysis, all of the single SNP associations within a gene or pathway were included and significance considered at the α = 0.10 level. As shown in Table 5, TLR7 variation was associated with the presence of at least one HE symptom (p = 0.062) and TLR3 gene was associated with presence of 2 or more HE symptoms (p = 0.022). In the pathway analysis, the TLR3-TRIF pathway was associated with HE2 (p = 0.035) and ME3 (p = 0.0176), the TLR7-MyD88 pathway and TLR8-MyD88 pathways were both associated with HE_any (p = 0.07; p = 0.039) and the combined TLR7-TLR8-MyD88 pathway was associated with presence of at least one HE symptom (p = 0.0458). The IL6-IL6R pathway was associated with presence of 3 or more non-specific symptoms (p = 0.065).
Finally, we conducted a haplotype analysis based on the most likely haplotype phases using the EM algorithm as implemented by the proxy association method in PLINK, considering significance at α = 0.10. First, we found that haplotypes in TLR3 had a different distribution in individuals with and without HE2 (Table 6); the overall distribution was significantly different (p = 0.0102). The TG haplotype was associated with a 5-fold increased risk of HE2 (p = 0.00775), and the AG haplotype had a significant protective effect (p = 0.0103). These results confirm the gene-level association between HE2 and TLR3 in Table 5. Second, we found haplotypes in MAVS were distributed differently in individuals with and without Any3 (omnibus p = 0.0514), with the GCT haplotype resulting in a 1.88 increased risk of Any3 (p = 0.00295). Though single SNP analyses of MAVS revealed associations with other clinical phenotypes, there were no associations observed with Any3, suggesting that an untyped polymorphism on the GCT haplotype may be associated with Any3.
In this study we examined polymorphisms in human genes of the innate immune system using diverse approaches and demonstrated association with a variety of clinical phenotypes in an ethnically Somali population of long-term residents in a RVFV endemic area. Our analysis included documentation of symptom recall using a structured interview administered by trained study personnel and we acknowledge that there may be inaccuracies with self-reported symptoms including memory lapses, selective recall of more severe symptomatology and other potential biases which are difficult to control in a retrospective study of this nature. Additional epidemiological factors associated with seropositivity and with severity of disease in this population are described elsewhere [11]. We analyzed a total of 46 SNPs in 16 genes (CFH, IL6, IL6R, IFIH1, DDX58, DHX58, MAVS, CCR5, TLR3, TLR7, TLR8, MYD88, TICAM1, IFNB, IFNAR1 CD209), and have identified innate immunity pathways that may play an important role in the pathogenesis of clinical RVF associated symptoms. These genes included those for the RNA helicases RIG-I (DDX58), LGP2 (DHX58) and their common adaptor MAVS (also called IPS-1) as well as endosomal Toll-like receptors TLR3, TLR7 and TLR8 and their signaling adaptors MyD88 and TRIF.
A strong association was observed in our analysis of the inflammatory cytokine IL-6 and its receptor, IL6R. We found association of the IL6 SNP and 3 of the 4 IL6R SNPs in our single gene analysis (Table 2). In an analysis of all SNPs in the IL6 and IL6R pathway, a significant association was found with non—specific symptoms including fever, anorexia, and backache. Therefore, although our data does not show a strong association between IL-6 and severe RVF symptoms, there is likely a role for IL-6 response, along with those for other inflammatory cytokines, in the pathogenesis of severe RVF. We have previously shown that IL-6 is one of several inflammatory responses to RVFV infection in a murine model of mucosal RVFV infection [50]. It is possible that robust IL-6 responses may lead to a cytokine “storm” via IL-6 receptor signaling, resulting in more severe clinical pathology such as hepatic inflammation, encephalitis, and risk for death.
One gene that was of interest, based on previously published GWAS studies was serum complement factor H. Several studies have shown an association of CFH mutations with age related macular degeneration [68–70] as well as with other eye diseases including uveitis [71]. Retinitis is a serious long-term complication of human RVF and we have previously observed prevalence as high as 21% in our study population [10,59]; therefore, we hypothesized that CFH may contribute to the pathogenesis of retinal disease. Surprisingly, we did not see a strong association of any of the 6 SNPs in the CFH gene with RVF specific eye disease symptoms, although one SNP (rs1061147) showed weak association with a cluster of general eye symptoms (Table 2).
As other viruses, including a related member of the Phlebovirus group, the ssRNA virus Punta Toro virus (PTV), have been associated with TLR activation [48,72], we decided to look for associations of clinical RVF symptoms with common polymorphisms in TLRs and signaling adaptor molecule genes. Although we did not find an association with individual SNPs in TLR3, we did find a TLR3 gene-level as well as haplotype association with having had two or more symptoms of hemorrhagic fever. Also, there was a pathway association between TLR3-TRIF SNPs and multiple symptoms of hemorrhagic fever, as well as 3 or more symptoms of meningoencephalitis, which suggests that this innate pathway is important in the pathogenesis of RVFV-associated severe disease. Although these association results were not all significant at α = 0.05, the consistency of association across phenotypes suggests the associations are robust and indicate a role in RVFV pathology. This was not surprising as there has been a clear association between TLR3 mediated innate responses and poor outcomes in a murine model of PTV [48]; however, in a murine model of mucosal RVFV infection we did not see an impact of the TLR3/TRIF pathway on severe disease or type I IFN responses [50]. A recent paper found Toll-7 dependence of RVFV induced autophagy in Drosophila and MyD88 dependence in a human osteosarcoma cell line [73], although the role of these pathways in human primary immune cell autophagy or immune responses remains unclear. In previous studies we also did not see TLR7 or TLR8 dependence for IFN and other cytokine responses to RVFV, yet in this genetic analysis we do see associations between polymorphisms in human TLR7 and TLR8 with clinical symptoms at the level of individual SNPs, as well as single gene and pathway analysis [50]. We conclude that the impact of the endosomal TLRs, including TLR3, TLR7 and TLR8 in the innate responses to RVFV and the pathogenesis of severe RVF is unclear. There may be differences between the utilization of endosomal TLRs for viral sensing at the cellular level and the impact of these important innate receptors at the whole organism level. Also, it is increasingly being recognized that there are important differences between mouse and human innate receptor activity in health and disease which may be contributing to the differences that we observe between laboratory studies and analysis of human field collected samples.
We have previously shown the importance of the RNA helicases RIG-I and MDA5 as well as the common signaling adaptor MAVS (also known as IPS-1) in RVFV induced IFN responses [50]. In a single gene analysis, we observed a trend towards a significant impact of one individual RIG-I (DDX58) SNP located in the exon of the gene. In the additive model, significant correlations were found between the helicase family member LGP2 (DHX58) and serology and between a 3’UTR SNP in RIG-I and any symptom. Interestingly, in a symptom cluster analysis both LGP2 (DHX58) and MAVS showed association and MAVS also showed association using a haplotype an analysis. These findings point out the challenge of single allelic association testing; whereas using more complex haplotype and cluster analysis demonstrated association of this well established viral innate sensing pathway with clinical phenotypes in RVF.
The type I interferon receptor, IFNAR, is formed by class II helical cytokine receptor family members IFNAR1 and IFNAR2 [74–76]. Although the role of type I IFN in host defense to multiple viruses is well established, and the role of IFNAR in modulating susceptibility and severity of disease has been established in multiple models of viral infection, including RVFV [77], no previous human studies have shown a correlation of genetic polymorphisms in the IFNAR genes with disease phenotype in RVF. Other groups have shown association with polymorphisms in IFNAR1 and Hepatitis B and C infection and disease [78–80]. Our current studies found significant associations between two polymorphisms (rs2257167, rs17875834) and disease phenotypes using an additive model and phenotype cluster / severity analysis.
Our findings point to important innate immune pathways in the pathogenesis of RVF associated symptoms. Polymorphisms in TLR3, TLR7, TLR8, MyD88, TRIF, MAVS, and RIG-I were repeatedly associated with severe symptomatology, suggesting that these genes may have a robust association with RVFV-associated clinical outcomes. Future studies to further explore the importance of these pathways in RVFV associated disease in different populations as well as correlation with in vivo and in vitro models of RVF are warranted.
|
10.1371/journal.pcbi.1002458 | What Can Causal Networks Tell Us about Metabolic Pathways? | Graphical models describe the linear correlation structure of data and have been used to establish causal relationships among phenotypes in genetic mapping populations. Data are typically collected at a single point in time. Biological processes on the other hand are often non-linear and display time varying dynamics. The extent to which graphical models can recapitulate the architecture of an underlying biological processes is not well understood. We consider metabolic networks with known stoichiometry to address the fundamental question: “What can causal networks tell us about metabolic pathways?”. Using data from an Arabidopsis BaySha population and simulated data from dynamic models of pathway motifs, we assess our ability to reconstruct metabolic pathways using graphical models. Our results highlight the necessity of non-genetic residual biological variation for reliable inference. Recovery of the ordering within a pathway is possible, but should not be expected. Causal inference is sensitive to subtle patterns in the correlation structure that may be driven by a variety of factors, which may not emphasize the substrate-product relationship. We illustrate the effects of metabolic pathway architecture, epistasis and stochastic variation on correlation structure and graphical model-derived networks. We conclude that graphical models should be interpreted cautiously, especially if the implied causal relationships are to be used in the design of intervention strategies.
| High-throughput profiling data are pervasive in modern genetic studies. The large-scale nature of the data can make interpretation challenging. Methods that estimate networks or graphs have become popular tools for proposing causal relationships among traits. However, it is not obvious that these methods are able to capture causal biological mechanisms. Here we address the power and limitations of causal inference methods in biological systems. We examine metabolic data from simulation and from a well-characterized metabolic pathway in plants. We show that variation has to propagate through the pathway for reliable network inference. While it is possible for causal inference methods to recover the ordering of the biological pathway, it should not be expected. Causal relationships create subtle patterns in correlation, which may be dominated by other biological factors that do not reflect the ordering of the underlying pathway. Our results shape expectations about these methods and explain some of the successes and failures of causal graphical models for network inference.
| Understanding the nature of cause and effect is fundamental to all fields of scientific investigation, but the concept of causality can present special difficulties in biology [1]. Experiments that utilize controlled interventions represent the most widely used approach to establishing causality. However, in his seminal work on experimental design, RA Fisher proposed that causation can be inferred from multi-factorial experiments performed with randomization [2]. An extension of this principle provides the foundation for computational approaches to network reconstruction in experimental genetic crosses, such as the recombinant inbred strain panel used in this study. Natural allelic variation is randomized during meiosis to generate a multi-factorial perturbation affecting multiple phenotypic outcomes. This meiotic randomization allows for the inference of quantitative trait loci (QTL) that are causal to phenotype [3].
Recent advances in high-throughput phenotyping technologies have made large-scale measurements of molecular traits possible. Expression QTL (eQTL), metabolic QTL (mQTL) and protein QTL (pQTL) can be used to link thousands of molecular phenotypes to genetic loci, as well as to clinical phenotypes [4]. A typical xQTL study will involve cross sectional sampling of a genetically variable population at a single time point. It is not immediately obvious that such data could provide insight into causal biological mechanisms, which derive from non-linear dynamic processes of gene expression and metabolism. However, a rich body of literature supports the idea that correlation structure in static data can provide insights into causal relationships among the measured variables [5], [6].
The interpretation of a directed edge between nodes and in a graphical model is that intervention on will alter , but intervention on will not alter . In a metabolic reaction, intervention on the substrate concentration will alter the product concentration. Reaction stoichiometry is often well understood [7]. Substrate molecules are converted by known enzymes into products, which in turn act as substrates for subsequent reactions. Reactions are organized into pathways which may converge, branch or intersect to form elaborate networks. More complex pathways involving feedback through allosteric interactions between enzymes and metabolites may also be present. It is not clear to what extent graphical models inferred from mQTL data capture these types of interactions.
Several algorithms have been proposed for the inference of causal relationships among phenotypes using genetic data [8]–[14]. These methods employ linear statistical models to infer the relationships between QTL and phenotypes, as well as relationships among phenotypes [15]. Causal edge detection is sensitive to subtle correlation patterns in the data. Inferences have been shown to be subject to a large proportion of false positive edges and can be skewed by environmental and experimental design factors that are not accounted for in the model [16], [17]. Agreement between the graphical model and the true underlying biology is a central goal of systems biology. The topology of networks inferred from xQTL data is often interpreted as a reflection of the underlying biological process - which may be metabolic or regulatory in nature, nonlinear, and involve the dynamic interaction of molecules within cells and tissues. However, the extent to which graphical models derived from static data capture these processes is not well understood, which makes the interpretation of edges challenging.
Deterministic models of cellular metabolism can be defined by ordinary differential equations (ODEs) derived from simple laws of mass-balance [18]–[21]. The reaction rates are modeled as non-linear processes, e.g. Michaelis-Menten kinetics and Hill functions, which depend on kinetic rate parameters [22]. Models of this type are powerful because of their ability to make in silico predictions of the response of a system to perturbations. We present a simulation study in which we generate synthetic mQTL data from dynamical models of pathway motifs with two sources of perturbation. We vary the rate parameters in a manner that mimics a genetic cross and we drive the simulations models with an input function that includes stochastic noise.
Glucosinolates are secondary metabolites that influence the interaction of plant and pest and have a wide range of important functions in human health [23]–[25]. The economic importance of glucosinolates has led to significant progress in understanding the biochemical pathways and genetics [26], [27]. Glucosinolate biosynthesis occurs in three well understood stages in which amino acids undergo (Figure 1): (1) chain-elongation, (2) formation of glucone moeity, and (3) side-chain modification. In this work, we examine mQTL data from a class of aliphatic glucosinolates in a highly replicated Arabidopsis BaySha recombinant inbred population [28]. The metabolites under investigation participate in side-chain reactions. Genetic analysis reveals shared QTL and wide-spread epistasis in the pathway [29].
In order to address these questions, we have inferred causal networks from mQTL data using simulated metabolic models of common pathway motifs and real data from a well characterized metabolic network. We demonstrate that correlation structure can be shaped by a variety of factors, including, genetic variation, pathway architecture, position in the pathway and feedback. Our results highlight the necessity of biological variation outside of the variation contributed by genetic factors for reliable network inference. Substrate-product relationships are not always reflected in the correlation structure of the system and recovery of the biochemical ordering of species should not be expected. Substrate inhibition, which is pervasive in metabolic pathways, can diminish or mask these relationships and lead to missing edges in network inference. An accurate genetic model is also critical to the inference process, especially when epistasis is involved. Our findings should temper expectations and provide new insights into the interpretation of causal genotype-phenotype networks.
Pathway motifs were constructed using ODEs (Figure 2). Flux rates, , were described with Michaelis-Menton kinetics. Simulations were performed under genetic perturbations, , with stochastic input, (Figure S1). The aliphatic glucosinolate biosynthetic pathway from an Arabidopsis BaySha population was also investigated (Figure 1). For each pathway, we carried out a three-step analysis: (1) QTL mapping for the metabolites in the pathway to identify the relevant genetic factors. (2) Metabolite correlations were calculated with and without conditioning on genetic factors. Correlation after conditioning represents the association between metabolites that is driven by sources outside of the genetic factors, e.g., propogation of random input fluctuations through the pathway. Correlation that disappears after conditioning implies an independent relationship between metabolites, e.g., and . We interpret the presence of correlation after conditioning as being indicative of either causal or reactive relationships, e.g., or . (3) We generated multiple causal networks from their posterior distribution, using a MCMC algorithm previously described [14] and summarized results across the ten top scoring networks.
In order to infer a causal relationship between a substrate and its product , non-genetic variation in substrate concentration has to propagate to the product. This is a necessary, but not sufficient condition for causal inference. To see this, suppose that one metabolite is causal to another, and that variation includes a genetic driver, . The linear equations for the causal graphical model can be written as: or equivalently:
Suppose there is no propagation of the non-genetic variation, , then:and the traits are conditionally independent given genotype, . It is clear from the equations that, is the term that carries the residual correlation between and . Therefore, variation in metabolites beyond that induced by genotype must be propagated through the biological pathway to create the correlation structure necessary for causal inference.
Consider the BaySha data example: , where denotes the QTL on Chrs 4, 5 and their interaction. There is a strong correlation between the residuals and () (Figure 7A), which is driven by the propagation of the non-genetic variation, . To see this dependency, we imputed data with no propagation of variation: and are approximately independent with negligible correlation (). A causal edge between and would not be detected with network inference (Figure 7B).
Graphical models provide a framework for estimating causal relationships between genotypes and phenotypes. Models of this type can be used to perform in silico experiments that predict responses to genetic and environmental perturbations. Ideally, these models should inform us about of the response to targeted interventions, such as a drug that alters the properties of a metabolic enzyme. There are numerous reasons for caution in such inferences. The inference models are linear, but the true relationships among relevant variables is likely to be driven by a non-linear dynamical process. It is not clear that these relationships should be captured by linear correlation. Correct interpretation is important, particularly if the graphical models are used to guide intervention strategies.
Several algorithms have been proposed for building graphical models in the context of genetic crosses [8]–[14]. These methods all derive models from the correlation and partial correlation structure in the data. We found that the available model building methods produced highly concordant results for models of the size and architectures considered here. Therefore we chose one specific MCMC algorithm to investigate the relationship between an inferred graphical model and the biochemical pathway that gave rise to the data. An advantage of the MCMC algorithm is the ability to sample multiple networks from a posterior distribution. This avoids reliance on a single network, which is problematic when two or more distinct networks can explain the data equally well. Sampling also provides a measure of uncertainty in the inferred network topology. Summarizing an ensemble of networks is challenging. We chose a consensus representation consisting of edges that occur most frequently in the sampled networks. If there is not enough information in the data to reliably establish the existence of an edge, this is reflected in low edge weights of the consensus network. Also, if we observe an edge that is present in most of the sampled networks but with opposing directions in different networks, we can conclude that the edge is present but there is insufficient data to resolve it direction (e.g., Figure 6C).
We analyzed metabolite data and from real and simulated pathways with known network stoichiometry. The Michaelis-Menton kinetics used in our simulated metabolic reactions are special cases of Hill functions and represent a rough approximation to actual enzyme reactions. Similar models have been used to describe gene regulatory networks and other biological phenomena, e.g. [19], [20], [30]. Constraint based modeling provide an alternative approach to delineate metabolic networks from steady-state data [31]. In the steady-state, the system of ODEs reduces to a linear system, but nonlinear relationships may arise between fluxes and pathways [32]. Investigation of the properties of constraint based and other non-correlation based methods for inference in dynamical systems remains an area of active research [33]–[36].
Correlation in metabolite data can be driven by a variety of factors that do not directly relate to the network stoichiometry. In order to capture the biochemical ordering of the pathway, noise has to propagate through the biochemical network. Many biological pathways are buffered by feedback or other stabilizing features that reduce noise propagation and mask the correlations that would imply causal connections. Failure to consistently observe substrate-product correlation may explain some of the differences observed between the plant data and simulations for matching pathway architectures. Our objective is not to confirm that our simulations accurately reflect the plant data or to make generalizations about certain pathway architectures. Rather, we seek to leverage real data from a well-studied biological system and simulated data from pathway motifs to explore a variety of architectures and conditions. A shortcoming of in silico models is their inability to fully capture the richly interconnected nature of biological systems. We considered simple motifs in isolation and modeled them with Michaelis-Menton kinetics. Correlation structure depends on the network architecture, the size and nature of the genetic perturbation, stochastic fluctuation, and enzyme kinetics. The advantage of this simulation is that no biological variation arises from factors outside of what is modeled. Whereas, metabolic systems in vivo contain mechanisms that make them robust, e.g., buffering, cycling and feedback, but may be impossible to pin-point with real data.
In the plant data, many of the substrate-product relationships remain intact after conditioning on QTL (Figure 5). This suggests that a real metabolic pathway may give rise to meaningful biological correlations that reflect the topology of the pathway despite the non-linear nature of the underlying processes. This is promising from the point of view of network reconstruction, but is not without limitation. The architecture of the homo-methionine side-chain was only partially captured, with an additional edge between Allyl and OHP3 that reflects the shunting of flux through the lower branch of the pathway (Figure 6A). The biochemical ordering of the dihomo-methionine side-chain was captured exactly (Figure 6B). We are only to able to detect an undirected connection between MT8 and MSO8 in the hexahomo-methionine side-chain (Figure 6C). Lack of a private QTL or a gradient in the effect size gives rise to likelihood equivalent models from which the direction of causality could not be distinguished. A similar situation was observed when a global model was estimated from the entire panel of metabolites and QTL (Figure 6D). The shared nature of the QTL hindered network reconstruction of the entire pathway. Most of the side-chain members were linked, but the direction of causality was not consistent with the pathway or with the networks constructed for each of the side-chains independently. Allyl and But-3-enyl are unlinked in the metabolic pathway, but are both products in reactions facilitated by AOP2. The causal link between them is likely driven by this co-regulation.
Conditioning on QTL genotypes strengthens the correlation among metabolites in most of the simulated pathway motifs (Figure 3). An exception occurs in the branching pathway with substrate inhibition which shows an almost complete loss of correlation between the branchpoint and upper branch metabolites and after conditioning (Figure 3F). In the linear pathway, when reaction rates are not operating at saturation and there are no branches to redirect the flux, any variation in the flux must propagate through each of the metabolites [37]. This results in a uniform correlation structure among the metabolites, which in turn yields weak causal linkages and order ambiguity among metabolite nodes in the graphical model. However, graphical models strongly and consistently associate metabolites to the QTL node controlling their downstream flux in linear pathways (Figure 3A, Text S1). The branching pathway is a linear pathway with a sink that represents demand on a metabolite from another reaction or pathway (Figure 2D). The stoichiometry of the branching pathway was captured exactly with the graphical model (Figure 3D). This suggests that the diversion of flux through side reactions is helpful in defining pathway order. For merging pathways, the correlation structure is dependent on the nature of the reaction at the merge point. When two pathways merge through a bi-substrate reaction (Figure 2B) there is strong association between the substrates that combine, but these are only weakly coupled to the downstream component of the pathway. On the other hand, when two pathways merge through independent reactions, the upstream metabolites and are only weakly correlated with each other, but the there is strong uniform correlation across the two linear components of the pathway (Figure 3C). Ordering metabolites in the independent merging pathway suffers from the same weaknesses as in the linear pathway. These results emphasize the influence of network stoichiometry on the correlation structure of the pathway.
Biosynthetic pathways, which often branch to produce two or more end products, are especially prone to inhibition [38]. We examined biosynthetic pathways that were inhibited in two ways: (1) loss of function in one pathway branch and (2) substrate inhibition. In the plant data, loss of function in AOP2 gave rise to an epistatic interaction between loci on Chr 4 and Chr 5 [28], [29]. Ignoring epistatic interactions and model fitting with only main-effect terms led to dense graphs that were difficult to interpret (data not shown). Substrate inhibition is estimated to occur in approximately 20% of enzymes [39]. This process can be viewed as a regulatory mechanism in which accumulation of a substrate represses the reaction velocity. In our simulation, the accumulation of metabolite inhibits the flux through a branched pathway (Figure 2E). The inhibition is reflected in the correlation structure, is negatively correlated with the other metabolites (Figure 3E). QTL disappears, suggesting that substrate inhibition can dominate the effects of genetic perturbations (Figure 3D–E). The correlation structure of this pathway was most sensitive to conditioning on QTL. When substrate inhibition is present, a loss of correlation and genetic control can occur, which makes two connected pathways look independent. These results highlight the importance of an accurate genetic model for network inference, especially in the presence of inhibition and epistasis.
Estimation of kinetic parameters in dynamic models requires time course data, which is often sparse, and the computations involved can be challenging [40]. The choice of experimental perturbations and design have been shown to have major influence on parameter estimation, and subsequently the accuracy of the computational model [41]. Complex models of biological systems exhibit parameter sensitivities that span several orders of magnitude [42]. Concentration profiles and model outputs are sensitive to small changes in kinetic rate parameters [43], [44]. The impact of parameter values on concentrations carries over into the correlation structure, and consequently, the downstream network inference. In our simulations, the perturbation is analogous to genetically determined non-competitive inhibition, where is genetically perturbed to be either high or low, thereby changing the flux capacity [45]. This strategy ensures that there is a significant difference between genotype groups and enables us to identify QTL. Random stochastic fluctuations were used as input and propagated through the pathway. Stochastic inputs allow us to examine the out of equilibrium dynamics of the system. The fluctuations themselves represent some of the randomness the pathway encounters from being part of a cellular system that is continuously changing [46], [47]. The models represent continuous excitation of the cell with the assumption that the intra-cellular dynamics can be faithfully modeled with ODEs. Examining system behavior over a spectrum of parameter values and stochastic inputs would offer additional insight into the sensitivity of the correlation structure.
Using both real data and simulated data, we tested the ability of graphical models to capture causal relationships between variables from from a variety of metabolic pathway topologies and conditions. We found that the use of linear statistical models to approximate relationships in dynamic non-linear systems from static data has some merit, but the results should be interpreted carefully. It is not realistic to expect to fully recover ordered pathway relationships with causal inference methods. Our results emphasize the necessity of biological variation beyond the genetic factors in the model for reliable network inference. We demonstrated that residual correlation induced between substrate and product in a metabolic reaction can be dominated by variety of factors, including, flux shunting, co-regulation, position in the pathway, genetic factors and inhibition. We found that inhibition can lead to missing edges in graphical models, washing out the genetic signal and making connected pathways look independent. An accurate genetic model is important, especially when epistasis is present. Taken together, these results temper our expectations and explain some of the success and failures of causal graphical models for genotype-phenotype inference.
Metabolic QTL data from a population of 403 Arabidopsis BaySha recombinant inbred lines (RIL) were examined in this study [28]. The data include measurements of aliphatic metabolites and genotypes from markers across the genome. A substantial number of samples have metabolite levels that are below the level of detection (Table S1). We applied a transformation to the scale . QTL mapping was performed for each metabolite with R/qtl [48]. Genome scans for single-QTL and two-QTL models were performed with Haley-Knot regression. The logarithm of odds (LOD) threshold for significance (P) was calculated from permutations [49].
Pathway motifs were used to define systems of ODEs that depend on flux rates, , modeled with Michaelis-Menten kinetics (Figure 2) [22]. If a substrate produces , then the rate of reaction is described by:where is the maximum rate of velocity and is substrate concentration at which half of is attained. When two substrates and combine to produce , , we write:When the accumulation of a metabolite feeds-back to inhibit a flux:where is an affinity constant. This flux form represents substrate inhibition which occurs at high substrate concentrations. As , the reaction flux is uninhibited and approaches standard Michalis-Menton form.
The dynamics of a substrate is described with the mass balance equations:where and denote the production and utilization of respectively, the stoichiometric coefficients are given as and and is the genotype. Genetic perturbations are made through the coefficients as either high or low, depending on the genotype or . For simplicity, we assume that each participates in a single reaction and that they are unlinked. In our simulations, we set , and . We also modeled a loss of function mutation by setting for certain genotypes (Figure S3) [50].
There are genotype combinations for each pathway of reactions. Each combination can be viewed as a sample from a randomized genetic population. For every unique genotype combination, we use an input flux that is perturbed by a random process, , modeled as a Brownian path over the interval [51]. The system is propagated, . The perturbations, , and the concentration levels at the end of the simulation are collected as data for correlation analysis and graphical model fitting. The output of each simulation can be viewed a sample in mQTL data. A schematic depicting the entire simulation process is shown in Figure S1.
The Pearson correlation is calculated for the variables in each pathway architecture. Residuals are estimated after each metabolite is conditioned on the QTL in the model. The residuals are used to calculate the conditional correlation of the metabolites given the genetic factors in the model. Directed graphical models are estimated using Bayesian Networks with a MCMC algorithm [14]. In pathways with epistasis, we include single degree of freedom variables that represent a composite genotype as variables for inference [52]. The sparsity parameter was set in the range . Each chain was run from two starting points, convergence was verified using correlation of edge weights (posterior probabilities) and the acceptance rate of each chain was in the range of 23%–45%. The results are based on the marginal summary over the ten graphs with the highest posterior probability. Alternative representations over the top and graphs and the four most probable graphs for each pathway are presented in Text S1.
|
10.1371/journal.pntd.0006164 | Changing demographics of visceral leishmaniasis in northeast Brazil: Lessons for the future | Visceral leishmaniasis (VL) caused by Leishmania infantum became a disease of urban areas in Brazil in the last 30 years and there has been an increase in asymptomatic L. infantum infection with these areas.
A retrospective study of human VL was performed in the state of Rio Grande do Norte, Brazil, for the period of 1990–2014. The data were divided into five-time periods. For all VL cases, data on sex, age, nutritional status and childhood vaccination were collected. Geographic information system tools and statistical models were used to analyze the dispersion of human VL. The mean annual incidence of VL was 4.6 cases/100,000 inhabitants, with total 3,252 cases reported. The lethality rate was 6.4%. Over time the annual incidence of VL decreased in the 0–4 years (p<0.0001) and 5–9 (p <0.0001) age groups, but increased in ages 20–39 (p<0.001) and >40 years (p<0.0001). VL occurred more often in males (β2 = 2.5; p<0.0001). The decreased incidence of VL in children was associated with improved nutritional status and childhood immunizations including measles, poliomyelitis, BCG, and hepatitis B. Human VL correlated temporally and geographically with canine L. infantum infection (p = 0.002, R2 = 0.438), with rainfall and with Lutzomyia longipalpis density (r = 0.762). Overall, the incidence of VL decreased, while VL-AIDS increased, especially between 2010–2014. VL was more frequently found in areas that lacked urban infrastructure, detected by lack of garbage collection and sewers, whereas HIV infection was associated with higher levels of schooling and evidence of higher socioeconomic status.
The demographics of VL in northeastern Brazil have changed. Disease incidence has decreased in children and increased in adults. They were associated with improvements in nutrition, socioeconomic status and immunization rates. Concurrent VL-AIDS poses a serious challenge for the future.
| We studied factors associated with the changing demographics of visceral leishmaniasis (VL) in Northeast Brazil, including environmental and socioeconomic determinants of disease, during the period 1990 to 2014. The incidence of VL was higher in urban areas, and regions with higher levels of canine L. infantum infection. Human males were more commonly affected than females. The incidence of VL in children under age 10 decreased during the period of study, simultaneous with increased incidence in adults. Reduction in the VL incidence among children was associated with improved socioeconomic status, administration of childhood vaccines and better nutritional status. Geographic areas with higher rainfall had higher densities of Lu. longipalpis, the primary vector of L. infantum in Brazil. VL was more frequently found in areas with indicators of poverty including sparse garbage collection and lack of urban infrastructure. During the past 25 years, HIV/AIDS has spread to areas where VL is endemic and has contributed to an increased incidence of VL-AIDS co-infection in adults.
| Visceral leishmaniasis (VL) is a life-threatening disease caused by L. infantum [1], which was first recognized in Brazil in 1932 [2–4]. It is likely the parasite initially arrived in northeastern Brazil with people and/or dogs previously infected with L. infantum in southern Europe or North Africa [5;6]. Lu. longipalpis is a competent vector for L. infantum, and it is found in most countries in Latin America. Dogs are considered the major reservoirs of L. infantum. At first, cases of VL in Brazil occurred sporadically in semi-arid, rural areas of the northeast region. Most cases occurred in children under 10 years of age [7;8]. However, in the late 1980’s and early 1990’s, urban outbreaks occurred in large cities in the northeast and other regions of Brazil [9–12]. Massive migration of the population to urban regions, adaptation of Lu. longipalpis to peridomestic environments, and transport of L. infantum infected dogs to urban areas occurred during this period. The temporal occurrence of human VL has assumed a variable pattern that correlates with environmental forces including el Niño/la Niña [13;14], which influence rainfall and humidity and thus the density of sand flies.
Several studies in endemic areas of Brazil have shown that most people infected with L. infantum remain asymptomatic, if they are not immunosuppressed [15–19]. It is not clear whether people with VL or with asymptomatic L. infantum infection serve as reservoirs and contribute to the long-term maintenance of this pathogen in endemic areas. In the past, the ratio of symptomatic VL to asymptomatic L. infantum infection among children was approximately one in every six, and for adults the ratio was 1 in 18. Risk factors for developing symptomatic VL in children included malnutrition [20–22], neoplastic disorders [19;23] and viral co-infection [24]. Brazil has had freely available vaccination for all age groups since early 1980 and this has had a coverage of over 90% for routine child immunizations. This improved vaccination rate has led to decreased and/or elimination of some of the common and potentially fatal diseases of children. For instance, it is known that measles infection predisposes to opportunistic infection for years [25;26] and measles has been eliminated from Brazil for almost a decade, although there was an outbreak in 2013, which was contained [27;28].
HIV/AIDS is known to increase the risk of developing symptomatic, rather than asymptomatic VL [29–31]. HIV co-infection with VL was first noted to contribute to the increased incidence of VL in adults in Europe in the mid-1980’s [32]. VL/AIDS was recognized in Brazil in early 90’s [32–34]. Thus, a new epidemiological pattern of VL is emerging because of spread of HIV to all regions of the country. Subjects with VL-AIDS have an increased risk of VL relapse and death. The goal of the current study was to identify demographic, spatial and socioeconomic factors associated with VL in northeastern Brazil between 1990 and 2014, using VL cases reported in the state of Rio Grande do Norte. We assessed the geographic distribution of VL and its association with the spatial distribution of HIV/AIDS, and with socioeconomic factors that could influence the outcome of L. infantum infection. A better understanding of the epidemiological dynamics of L. infantum and HIV co-infections, and their determinants is essential to guide new health policies.
The study was conducted in the state of Rio Grande do Norte in northeastern Brazil. The state has an area of 52,811,126 square kilometers, with a population of 3,408,510, 77.8% of whom now live in urban areas. Most of the state has a semi-arid climate, with rainfall less than 800 mm per year and an average temperature of 27°C. A more humid climate is found along the east coast of the state, which borders the Atlantic Ocean, where the rainfall indices are greater than 1,400 mm per year. The state is grouped in 19 micro-regions (MR), each with distinct climate, topography, hydrography, population density and economy. These micro regions served as units for the current analysis.
A retrospective study of VL cases diagnosed in the state of Rio Grande do Norte was performed, and correlated with demographic and epidemiologic factors in the corresponding regions of the state. The study was divided into five time periods: (1) 1990–1994, (2) 1995–1999, (3) 2000–2004, (4) 2005–2009 and (5) 2010–2014. The association between spatial patterns of VL and HIV-AIDS was assessed using quantitative data available from 1991, 2000 and the 2010 censuses. The incidence of VL, AIDS and VL-AIDS co-infection per 100,000 inhabitants was extracted from the state records listed below. Additional variables that were collected included: (1) dates of new cases of VL, AIDS and VL-/AIDS, (2) sex, (3) age, (4) nutritional status of children under 5 years, (5) vaccination rates for measles, poliomyelitis, BCG and hepatitis B in children under 5 years. The spatial and temporal correlation between human VL and canine infection was also assessed. Environmental or socioeconomic variables considered were: (1) the association of annual rainfall and the density of Lu. longipalpis, (2) socioeconomic data from the censuses included literacy rate, education, income, city water supply, waste disposal, septic tank and presence of sewage.
Data on human VL were obtained from the Notifiable Diseases Information System (SINAN). This federal government system catalogs the reported cases and coordinates investigations of diseases for which reporting is mandated by the Brazilian government, as defined by specific legislation. The data are captured in the Health Post Centers and/or hospitals and are sent to state Secretary of Health, whose office uploads the information in SINAN. The list of notifiable disease is updated as new outbreaks occur. For example, the recent epidemic of Zika virus infections led this virus to be added to the list of mandatory reportable diseases. Data on Lu. longipalpis density and rates of infection in domestic dogs were obtained from the Surveillance and Leishmaniasis Control Program, Secretary of Health. Data on HIV or AIDS were obtained by crossing information from SINAN with the Brazilian Mortality Information System database and release of medications for HIV. Nutritional data on children younger than 5 years were obtained from the Brazilian Minister of Health. Those data were available for all 19 micro-regions of the state of Rio Grande do Norte.
Data about immunization coverage were obtained from the Brazilian National Program of Immunization Information System. The vaccination coverage percentage was calculated per estimated population at the age group targeted to be vaccinated and the doses of vaccines administered in each municipality, and grouped into the 19 micro regions for this analysis.
Data variables gathered from censuses included education level, income, local health facility, piped water supply, garbage collection, street cleaning, sewer system, septic tank, urbanization and population density. Those data were collected from the Instituto Brasileiro de Geografia e Estatística (IBGE) website (http://www.ibge.gov.br/home/). Annual rainfall data in the municipalities were obtained from the State of Rio Grande do Norte Agricultural Research Company, EMPARN, (http://www.emparn.rn.gov.br/).
The effect of sex and micro-region (MR) on the temporal variation in the incidence VL from 1990 to 2014 were evaluated by a general linear model with categorical explanatory variables [35] according to the following formula: Yti = β0 + β1t + β2I(Sex) + θiMR(i|19) + error, (model 1) where the dependent variable was the incidence of VL (Yti) per 100,000 inhabitants, in the year t and in the micro-region i. The independent variables were the time (t) in years, the categorical variables were sex (1 if male) and micro-region (1–19) MR (i |19). The β1 coefficient measured the average annual increase in the incidence of VL, whereas β2 measured the differential incidence of VL between male and female. Micro-region 19 was considered the reference, since it was the site of the first VL outbreak in the state of Rio Grande do Norte. The micro-region parameter was considered when evaluating the existence of spatial aggregation. The temporal incidence of VL considering the patient's age was analyzed by linear regression using the following statistical model: Zt = β0 + β1t + error (model 2). The Zt is the rate of cases per 100,000 inhabitants in year t, and β1 is the slope of the adjusted line that defined the secular trend of the incidence of VL for the subsequent year. A rising trend in cases/unit of time unit was observed when β1>0, whereas a downward trend in case rates/time was observed when β1<0. The case rate over time was stationary if β1 = 0. The model was adjusted in each of the following age groups: 0–4 years, 5–9 years, 10–19 years; 20–39 years and ≥40 years. The incidence/time unit was calculated within each age group.
The impact of routine vaccination of the population, including vaccines for measles, poliomyelitis, BCG and hepatitis B, on the incidence of VL was analyzed in children under 5 years of age between 2000 and 2014. We evaluated these data for each micro-region. Vaccination rate data were superimposed on micro-region VL incidence data to determine whether vaccination coverage correlated with VL, particularly in children under age 5. We used an adjusted linear regression model as defined by the formula: Yti = β0 + β1t + β2Xti + θiMR(i|19) + error, (model 3). Similar to model 1, the independent variable Xti was the vaccination coverage in micro-region i during year t. Because they were strongly correlated, an analysis of principal components was made and the vaccination coverage was represented by the score of the first component.
The nutritional status was considered by the proportion pi of children whose weight for age was categorized as very low(p1), low(p2), appropriate(p3) and high(p4). From those, a Nutritional Status Index—NSI was built considering a weighted proportion with zero sum weights w = (-1, -1/3, 1/3, 1) defined by NSI=−1p1−(13)p2+(13)p3+1p4, ranging between -1 and 1. A value of NSI close to -1 point is a very low status while a value close to 1 point is a very high status. The influence of nutrition on VL development was assessed by a model similar to model 3, where Xti is NSIti. (Model 4).
The relationship between the incidence of human VL with canine VL in micro-regions/time was performed by adjusting a simple linear regression model defined by the formula: Yt = β0 + β1Xt + error, (model 5). The dependent variable, Yt was set to log (rate human VL+ 0.5) in the year t in which there was canine examination, and the independent variable was Xt the corresponding level of canine infection (LCI). The level of canine infection was defined as a weighted proportion of infected dogs, using weights ranging from zero to 100 depending on the total number of dogs examined versus the number that had Leishmania infection, model 6, as follows:
LCI=log[(100infectedexamined)(100examinedmax(examined))]
The spatial dependence and the association between response variables was performed by modeling the distribution of human VL case events with the spatial distribution of canine VL and social predictive factors, by adjusting mixed autoregressive spatial linear models (Spatial Lag Model Autoregressive-SAR) that captured the self-spatial correlation through a single ρ parameter (rho) added to the regression model; this was chosen to model factors in the same test: i.e., temporal variation in VL, effect of sex, and geographic micro-region. The equation was expressed as: y = ρXy + XBeta + error, (Model 7), where y is incidence of VL per 100,000 inhabitants at micro-region level. ρ (rho) measures the spatial dependence the VL incidence, W is the weight matrix modeling the spatial structure, X is the matrix of predictor variables, β is the regression coefficient vector which evaluated the association between Y and X, and error represents the residuals. The log transformation was applied to normalize the response distribution. The same analytical approach was applied to the spatial distribution of HIV/AIDS. The predictor variable data (X) was collected in the censuses 2000 and 2010. All statistical models tested herein are shown in S1 Supporting Information.
We used Excel 2013 in the construction of the database, Statistica StatSoft version 7.0, in the estimation of linear models (Models 1 to 7) and Quantum GI version 2.12.e-Lyon (http://www.gnu.org.licenses) in the construction of maps and R System version 3.2.2 (https://www.r-project.org) for the mixed autoregressive linear models.
The source of base layers used to build the figures was found at http://censo2010.ibge.gov.br/resultados and https://mapas.ibge.gov.br/bases-e-referenciais/bases…/. The softwares used to build the maps were QGis version 2.12.e-Lyon (https://www.gnu.org/licenses/) and R System version 3.2.2 (https://www.r-project.org)
This study was reviewed and approved by the Universidade Federal do Rio Grande do Norte Ethical review board CAAE12584513.1.0000.5537. Data were anonymous records and exempted from signed consent.
A total of 3,252 cases of VL were reported in the state of Rio Grande do Norte, northeast Brazil, between the years of 1990 and 2014. The mean annual incidence was 4.6 VL cases/100,000 inhabitants. Fig 1A shows the incidence of VL, AIDS and VL/AIDS by 5-year period. The overall VL lethality rate was 6.4%, with a total of 210 deaths (210/3,252). However, the highest lethality rate was 8.2% in the period 1990–1994 and was associated with the highest incidence of VL (Fig 1B).
There were 5,777 cases of AIDS with an average incidence of 8.1 cases/100,000 inhabitants (Fig 1A) between 1990 and 2014. During the study period, the average incidence of concurrent VL/AIDS was 0.16 per 100,000 inhabitants. However, on the 5th period (2010–2014) it reached 0.46/100,000 (Fig 1A).
VL cases were predominantly found between 1990–1994 in the eastern coastal region of Rio Grande do Norte, but the disease subsequently spread to the Northeastern Coast (θ16 = 5.983; p<0.0001) and to other areas (θ17 = 6.256; p<0.0001, model 1), Table 1 and Fig 2. Although there has been an increase in areas reporting VL, there was a mean decrement of 0.135 VL cases/per year (β1 = -0.135, p<0.0001) (Table 1).
Over time there was a decreasing trend in the incidence of VL in both sexes. However, the male incidence was uniformly higher by approximately 2.5 per 100,000 (β2 = 2.498; p<0.0001), (Table 1; Fig 3A). There were two major peaks of VL, the first in 1991–1992 and the second in the 1999–2000 (Fig 3A).
The temporal incidence of VL decreased significantly 0–4 and 5–9 among the age groups (β1 = -0.0117, p<0.0001 and β1 = -0.0042, p<0.0001, respectively) between 1990 and 2014, with significant increase in the 20–39 and >40 age groups (β1 = 0.0071, p<0.0001 and β1 = 0.0105, p<0.0001, respectively), Table 2. At the same time, the VL incidence was stationary in the 10–19 age group (β1 = -0.0016, p = 0.1320), (Table 2, Fig 3B). The mean age of VL increased linearly during the period of study, (age = (-1392.657) + 0.704 (year), p<0.001), with an annual increase of 0.704 years (8.4 months). The mean age of VL prior to 2000 was 12.9 ± 0.98 (SD) years, whereas from 2000 to 2014, the mean age was 21.7 ± 3.74 (SD) years (p<0.005), Fig 3B.
The adjustment of model 3 for spatial dispersion of VL showed that, by correcting for the effect of trends and differences between micro-regions, there was a strong negative association between the incidence of VL and the score of the vaccine coverage (β2 = -4.805; p = 0.0003). The score in year t and the micro-region i representing the vaccination coverage, at this time and place was calculated by Xti = 0.0989BCGti + 0.3993POLIOti + 0.3181MEASLESti + 0.3915HEPATITISti, obtained from the first principal component. This means that an increase in one unit in the vaccination coverage score was associated with a reduction of 4.8 in the incidence rate of VL in children younger than five years. There is a strong association between the two variables estimated by a third-degree polynomial relationship (Fig 4A).
The association between the Nutritional Status Index-NSI and the incidence rate of VL in children less than 5 years was evaluated by adjusting for other parameters, models 3 and 4. There was a trend toward an inverse correlation between better nutritional status and VL in children, but this did not reach statistical significance, (β2 = -141.76; p = 0.1136). In contrast, there was a positive linear correlation between the incidence of human VL and canine VL (Human VL = 0.4038 + 0.2889LCI; r = 0.456, p = 0.0758), evaluated by the level of L. infantum canine infection (LCI) per models 5 and 6, as described in the methods session (Fig 4B). An increase of one LCI unit was associated with an increase in the incidence rate of human VL.
There was a positive correlation between the density of sand flies and the rainfall index (r = 0.762, p<0.0001), according to model 5. In addition, an increase in annual rainfall index correlated with increased annual incidence of VL (r = 0.616, p = 0.005). The fluctuation in rainfall index explained 38% of the variation in the incidence of VL, with a100 mm increment in annual rainfall associated with an increase of 0.6 in VL/100,000 inhabitants (Fig 4C).
Among the socioeconomic variables that correlated with the incidence of VL was the percentage of households with garbage collection (β = -0.1684; p = 0.0116 in the 2000 census and β = -0.0341, p = 0.0153 in the 2010 census) and the percentage of households connected in the general water supply network (β = -0.3514; p = 0.100 in the 2010). In contrast, the incidence of AIDS correlated positively with garbage collection (β = 0.0358, p = 0.0005 in 2000 census and β = 0.0270, p = 0.0130 in the 2010 census), with higher sanitation level (β = 0.0407, p = 0.0007, in the 2000 census and β = 0.0418, p = 0.041in the 2010 census), with literacy rate (β = 0.0648, p = 0.0056 in 2010 census), with access to city water (β = 0.0431, p = <0.0001 in the 2000 census), adjusted model 7.
Transmission of VL occurs in settings where the infected sand fly vector lives in proximity to a mammalian reservoir and susceptible humans [36], or through other means of transmission such as blood transfusion [37]. Despite efforts of health officials to interrupt the routes of transmission of L. infantum, VL continues to be a major health problem in Brazil after Malaria [38]. The demographics have changed substantially since VL was first reported in the 1930’s [2]. During early years, the disease occurred predominantly in rural areas of the Northeast region, with most cases of VL occurring in children younger than 10 years [7]. Mass migration of the population to urban areas beginning in the 1980s was accompanied by a change in the pattern of transmission to peri-urban regions of large cities in the Northeast and the southeast regions of the country [39].
The state of Rio Grande do Norte in northeastern Brazil, provides an example of the changing epidemiology of VL. There was a significant increase in the age at disease diagnosis, with an increase in adult VL. The disease decreased in children under age 10 years and increased in adults, mainly from period 3 of this study. The average age at diagnosis of VL in Rio Grande do Norte rose from 12.9 years prior to 2000 to 21.7 years in 2014 (Fig 3B). An increase in the average age of VL has also been observed in other Brazilian states [39;40]. Human VL has been associated with poverty and malnutrition in children [20–22;41]. We hypothesized that multiple socioeconomic factors might contribute to the significant reduction in childhood VL in the less than 10 age groups. Since 1999, social programs to decrease poverty and economic measures to control inflation have been successfully implemented in Brazil, with a coincident improvement in many measures of health [42]. Interventions have included supplementation of micronutrients, including iron and vitamin A in pregnant women and children aged 6–18 months, as well as fortification of wheat and corn flours with iron and vitamins [43]. There is an increase in vaccine coverage, with more uniform administration of vaccines as polio, measles, BCG and others. Those measures have been associated with increased average birth weight [22;44], decreased childhood diarrheal diseases [45;46]. The improved health could lead to healthier gut brush border, better absorption of nutrients and protection against opportunistic pathogens such as Leishmania. As an example, studies have shown that the vaccine-preventable disease measles can induce immunosuppression for years [26;47].
In previous studies in the state of Rio Grande do Norte, we found that children and adults were infected at comparable rates with L. infantum, as detected by positive anti-leishmanial serology and/or positive skin test response to Leishmania antigens [15]. Since a majority of L. infantum infections are asymptomatic, it is likely that the above-mentioned health interventions have resulted in enhanced development of protective Type 1 immune responses to Leishmania spp. and other pathogens, with a decreased likelihood that L. infantum infection will progress to VL in young children. In addition, improved socioeconomic status, improved living conditions, and expansion of urban regions may be responsible for decreased sand fly density and transmission in some areas.
In the current study males accounted for most VL cases (67%). Greater susceptibility of males to VL has also observed in other human studies [8,40] and in experimental models of Leishmania infection, hamster and murine models of VL [48]. Higher levels of testosterone have been associated with increased risk of VL caused by L. donovani in India and Sudan [49;50] possibly mediated by increased IL-10 production and down regulation of Th1 responses.
Domestic dogs are thought to be the primary reservoir of L. infantum in Brazil [51]. A correlation between human and canine VL was observed in the current study, and in reports from other areas in Brazil and Northern Africa [52–54]. Further studies are needed to better define the roles of dogs and asymptomatically infected humans as reservoirs for L. infantum in the epidemiology of VL in Brazil.
Higher rainfall indices correlated temporally and geographically with a higher incidence of human VL, especially in areas close to the Atlantic Ocean. An association between VL with increased rainfall has been reported in other regions of Brazil [55;56]. However, some cases of VL occurred in areas with lower humidity, higher temperatures and lower rainfall indices. It is likely that variations in the microenvironment provided niches in which sand flies were nonetheless able to thrive in proximity to humans and a dog reservoir.
HIV/AIDS occurred predominantly in males in the initial stage of the pandemic, although recently more women have become infected with HIV in Brazil [57]. The highest incidence of HIV/AIDS occurs in urban regions, coinciding with regions that have a higher incidence of VL. HIV/AIDS has been expanding throughout Brazil since 1990, and has now spread to all areas of Rio Grande do Norte. Coinfection with HIV and Leishmania spp. has contributed to the increased incidence of VL in adults in southern Europe in Spain, France, Italy and Portugal [31;58;59]. Consistently, in this report we document the presence of VL/AIDS in Rio Grande do Norte since 1990, but there was a considerable increase in coinfection in the third period (2000–2004), presumably because HIV infections spread to areas that were endemic for L. infantum infection. A large number of individuals are asymptomatically infected with L. infantum in the state of Rio Grande do Norte [15;17;60], and people with asymptomatic L. infantum and HIV seem to be at greater risk of developing VL and of death [61]. Therefore, it is imperative that strategies and guidelines be developed to prevent the development of VL during HIV infection. In summary, the demographics of VL in northeastern Brazil have changed substantially over the past 25 years. The incidence has decreased in children in association with improved nutrition, socioeconomic status, childhood immunizations, and overall health. In contrast, the incidence of VL in adults has increased. The latter could be explained in part by failure to develop immunity to the parasite as a child, and the geographic coincidence of HIV infection and VL. The emergence of concurrent VL-AIDS poses a serious health challenge for the future.
|
10.1371/journal.pgen.1000294 | Effects of cis and trans Genetic Ancestry on Gene Expression in African Americans | Variation in gene expression is a fundamental aspect of human phenotypic variation. Several recent studies have analyzed gene expression levels in populations of different continental ancestry and reported population differences at a large number of genes. However, these differences could largely be due to non-genetic (e.g., environmental) effects. Here, we analyze gene expression levels in African American cell lines, which differ from previously analyzed cell lines in that individuals from this population inherit variable proportions of two continental ancestries. We first relate gene expression levels in individual African Americans to their genome-wide proportion of European ancestry. The results provide strong evidence of a genetic contribution to expression differences between European and African populations, validating previous findings. Second, we infer local ancestry (0, 1, or 2 European chromosomes) at each location in the genome and investigate the effects of ancestry proximal to the expressed gene (cis) versus ancestry elsewhere in the genome (trans). Both effects are highly significant, and we estimate that 12±3% of all heritable variation in human gene expression is due to cis variants.
| Variation in gene expression is a fundamental aspect of human phenotypic variation, and understanding how this variation is apportioned among human populations is an important aim. Previous studies have compared gene expression levels between distinct populations, but it is unclear whether the differences that were observed have a genetic or nongenetic basis. Admixed populations, such as African Americans, offer a solution to this problem because individuals vary in their proportion of European ancestry while the analysis of a single population minimizes nongenetic factors. Here, we show that differences in gene expression among African Americans of different ancestry proportions validate gene expression differences between European and African populations. Furthermore, by drawing a distinction between an African American individual's ancestry at the location of a gene whose expression is being analyzed (cis) versus at distal locations (trans), we can use ancestry effects to quantify the relative contributions of cis and trans regulation to human gene expression. We estimate that 12±3% of all heritable variation in human gene expression is due to cis variants.
| Admixed populations are uniquely useful for analyzing the genetic contribution to phenotypic differences among humans. Phenotypic differences that are observed among human populations may have systematic non-genetic causes, such as differences in environment [1],[2]. However, in an admixed population such as African Americans, such differences are minimized and the only systematic differences among individuals are in the proportion of European ancestry, which can be accurately inferred using genetic data. Several recent epidemiological studies in African Americans have taken advantage of this, showing that many phenotypic traits vary with the proportion of European ancestry [3]–[5]. Here, we apply this idea to analyze population differences in gene expression.
Gene expression is a fundamental determinant of cellular phenotypes, and understanding how gene expression variation is apportioned among human populations is an important aspect of biomedical research, as has been true for apportionment of human genetic variation at the DNA level [6]. Recently, four studies analyzed lymphoblastoid cell lines from HapMap samples and reported that a large number of expressed genes exhibit significant differences in gene expression among continental populations [7]–[10]. However, results of these studies may be affected by non-genetic factors such as differences in environment, differences in preparation of cell lines, or batch effects [2],[9],[11],[12]. In particular, a recent review article has suggested that much of the expression variation across populations is caused by environmental factors [13]. On the other hand, analyses of expression differences that are correlated to ancestry within an admixed population are robust to all of these concerns.
In this study, we analyzed lymphoblastoid cell lines from 89 African-American samples and investigated the relationship between expression levels of ∼4,200 genes and the proportion of European ancestry. We compared the results with those predicted from the differences in expression levels between 60 European samples (CEU from the International HapMap Project) and 60 African samples (YRI from HapMap) [6]. We confirmed the existence of heritable gene expression differences between CEU and YRI by showing a highly significant correspondence between observed CEU vs. YRI differences (i.e. differences between sample means) and the expression differences predicted by ancestry differences among African Americans. Notably, the correspondence holds regardless of whether differences between CEU and YRI are large or small. This suggests that the effects of heritable population differences on variation in gene expression are widespread across genes, mirroring population differences at the DNA level [6].
Heritable variation in gene expression may be due to cis or trans variants. Previous studies in humans have been successful in mapping both cis and trans effects, but the results they provide are far from complete, due to limited sample sizes [14], [15], [9], [16]–[20]. In particular, the relative number of cis vs. trans associations that were reported varies widely across these studies, perhaps due to differences in power or choices of significance thresholds [13]. Thus, the overall extent of cis vs. trans regulatory variation in human gene expression has not yet been established. Here, by measuring how gene expression levels across all genes vary with local ancestry (0, 1 or 2 European chromosomes) either proximal to the expressed gene (cis) or elsewhere in the genome (trans), we estimate that 12±3% of heritable variation in human gene expression is due to cis variants.
100 African-American (AA) samples from the Coriell HD100AA panel were genotyped on the Affymetrix SNP 6.0 GeneChip. Genotyping was conducted at the Coriell Genotyping and Microarray Center, and the genotype data was obtained from the NIGMS Human Genetic Cell Repository at Coriell (see Web Resources). In addition, genotype data from 60 European (CEU), 60 African (YRI), 45 Chinese (CHB) and 44 Japanese (JPT) samples was obtained from Phase 2 HapMap [6] (see Web Resources). We restricted all analyses to 595,964 autosomal markers with <5% missing data in AA samples and <5% missing data in Phase 2 HapMap samples, with A/T and C/G markers excluded so as to preclude any ambiguity in strand complementarity. Our analyses were not sensitive to the number of markers used. Two AA samples which we identified as cryptically related to other AA samples were excluded from the set of samples used for principal components analysis.
Local ancestry (0, 1 or 2 European chromosomes) at each location in the genome was estimated for each AA sample using the HAPMIX program, a haplotype-based approach that has been shown to attain an r2 of 0.98 between inferred local ancestry and true local ancestry in simulated African-American data sets (A.L.P., N.P., D.R. & S.M., unpublished data; see Web Resources, specifically http://www.stats.ox.ac.uk/̃myers/software.html). The HAPMIX program inputs AA genotype data and phased CEU and YRI data from Phase II HapMap [6], and outputs the estimated probability of 0, 1 or 2 European chromosomes at each location in the genome. The weighted sum of these probabilities (multiplied by 0.00, 0.50 or 1.00, respectively) forms an estimate of local % European ancestry. Genome-wide ancestry was computed as the average of estimated local ancestry throughout the genome.
Lymphoblastoid cell lines for 60 HapMap CEU, 60 HapMap YRI and the Coriell HD100AA samples were obtained from Coriell Cell Repositories (see Web Resources). Gene expression was assayed using the Affymetrix Genome Focus Array, as described previously [7]. We restricted our analysis to the 4,197 genes on the array that are expressed in lymphoblastoid cell lines [7]. The gene expression data is publicly available (GEO accession number GSE10824) (see Web Resources). For HD100AA samples, we excluded two cryptically related samples (see above), four samples identified as genetic outliers (see Results), and five samples for which gene expression measurements were not obtained, so that 89 AA samples were included in gene expression analyses.
For each gene g, we normalized gene expression measurements for CEU and YRI to have mean 0 and variance 1 across 120 CEU+YRI samples, and normalized gene expression measurements for AA by applying the same normalization for consistency. We implicitly assume an additive genetic model in which gene expression has genetic and non-genetic components, with part of the genetic component predicted by ancestry. Let egs denote normalized gene expression of gene g for sample (i.e. individual) s. Let θs denote the genome-wide European ancestry proportion of sample s, so that θs has value 1 for CEU samples and 0 for YRI samples as above, and fractional values for AA samples. We consider a model in which egs = agθs+νgs for CEU and YRI samples and egs = cagθs+νgs for AA samples, where c is a global parameter and νgs represents the residual contribution to gene expression that is not predicted by ancestry. Thus, the parameter c represents a validation coefficient measuring the aggregate extent to which the observed gene expression differences ag between CEU and YRI (differences between sample means) are heritable.
We implemented two different approaches for fitting the parameters c and ag of this model: (1) Starting with the initial guess c = 1, we alternated computing maximum likelihood estimates for ag (for all g) conditional on c, and computing a maximum likelihood estimate for c conditional on ag (for all g), and iterated to convergence. In each case, the maximum likelihood estimates were obtained via linear regression (with a separate linear regression for each g when estimating ag, and a single linear regression when estimating c). (2) For each g, we estimated values ãg,CEU+YRI by regressing egs against θs using CEU and YRI data only, and ãg,AA by regressing egs against θs using AA data only. We then regressed ãg,AA against ãg,CEU+YRI to obtain an estimate of c. In this computation, we scaled our estimates of ãg,CEU+YRI using the sampling error correction ξ (described below in Computation of QST) to remove the effect of sampling error on the denominator Σg(ãg,CEU+YRI)2 of our estimate of c. (On the other hand, we note that sampling noise in the AA data does not bias our computation of c, whose expected value does not change when noise is added to ãg,AA). We observed that approaches (1) and (2) produced identical estimates of c, indicating that both approaches are effective in finding the best fit to the model. We followed approach (2) to plot ãg,AA vs. ãg,CEU+YRI and to compute estimates of c specific to different values of |ãg,CEU+YRI|.
We repeated the above computation using genotype data instead of gene expression data. We restricted the analysis to markers in which the average of CEU and YRI frequencies was between 0.05 and 0.95. Although AA genotypes at each marker were used twice in this computation—both for estimating genome-wide ancestry using all markers and for measuring the effect of genome-wide ancestry on genotype at a specific marker—we note that with hundreds of thousands of markers, our estimate of genome-wide ancestry is negligibly impacted by data from a specific marker.
We investigated the effects of cis ancestry and trans ancestry on gene expression in AA. Roughly, we define cis ancestry as the local ancestry at the gene whose expression is being analyzed, and trans ancestry as the average ancestry at non-cis regions. We extended our above model by letting egs = ccisagγgs+ctransagθs+νgs for AA samples, where γgs denotes the estimated local ancestry of sample s at the SNP closest to the center of gene g (cis locus; average of transcription start and transcription end positions). We note that although trans ancestry is theoretically defined as the average ancestry at non-cis regions, this quantity is in practice virtually identical to θs because cis regions (regardless of the precise definition of cis) form an extremely small proportion of the genome. Because chromosomal segments of ancestry in AA typically span >10 Mb [21], it is nearly always the case that a gene lies completely within a single ancestry block, so that our analysis is not sensitive to the choice of genomic location used to define cis ancestry γgs. The probabilistic estimates of local ancestry produced by HAPMIX are extremely accurate (see above), so that γgs is typically close to 0.00, 0.50 or 1.00 (corresponding to 0, 1 or 2 copies of European ancestry). To avoid complications in local ancestry analyses on the X chromosome, we restricted this analysis to 4,015 autosomal genes. (Analyses involving global ancestry were not affected by inclusion or exclusion of genes on the X chromosome.) We estimated the global parameters ccis and ctrans as above, accounting for the correlation between genome-wide and local ancestry by using residual values of γgs (adjusted for θs) to compute ãcis,g,AA (and conversely for ãtrans,g,AA).
Let F denote the proportion of total variance in gene expression that is attributable to population differences. For quantitative traits with an additive genetic basis, the quantity that is analogous to single-locus estimates of FST is not F, but rather QST = F/(2−F) (reviewed in [22]). This is a consequence of the contributions of genetic variation on two distinct haploid chromosomes, magnifying the effect of population differences under an additive genetic model. We computed both F and QST. For each gene g, we normalized gene expression measurements for CEU and YRI to have mean 0 and variance 1 across 120 CEU+YRI samples. We defined the ancestry θs of sample s to be 1 if s is a CEU sample, and 0 if s is a YRI sample. As above, we modeled normalized expression of gene g for sample s as egs = agθs+νgs. Equivalently, under this definition, ag is equal to the difference in normalized gene expression between CEU and YRI samples. We defined F to be the quantity such that the true value of ag has mean 0 and variance 2F across genes [23]. For a specific gene, agθs has variance 0.25ag2 and νgs has variance 1–0.25ag2 across CEU+YRI samples (these variances have expected value 0.5F and 1–0.5F, respectively). Due to sampling error, the observed difference ãg in normalized gene expression between CEU and YRI samples (i.e. the coefficient obtained from a regression of egs on θs) has variance 2F+(1–0.5F)/30, where 1/30 is the sum of reciprocals of CEU and YRI sample sizes. We thus estimated mean F as (Varg(ãg) – 1/30)/(2 – 0.5/30). The ratio between mean F and Varg(ãg)/2 represents a sampling error correction that we call ξ. We estimated median F as the median value of ãg2/2 times ξ. The value of ξ was 0.93, indicating that the sampling error correction had only a minor effect on these computations. To account for differences between CEU and YRI due to non-genetic factors, we adjusted F by multiplying it by c. (We note that the scaled population differences cag have variance that is c2 times the variance of ag, but explain only the proportion c of the true component of variance that is attributable to ancestry.) We then computed QST = F/(2−F). We calculated the standard error of our estimate of F via jackknife, repeating the computation of F 120 times with one of the 120 CEU+YRI samples excluded in each computation, and estimating the standard error as the standard deviation of the 120 estimates times the square root of 120.
We analyzed Affymetrix 6.0 genotype data from the African-American panel of 100 samples from Coriell Cell Repositories, together with HapMap samples (see Materials and Methods). We first ran principal components analysis, using the EIGENSOFT software [24]. The top two principal components are displayed in Figure 1, in which most AA samples roughly lie on a straight line running from CEU to YRI (we excluded three genetic outliers with partial East Asian ancestry and one genetic outlier whose ancestry is very close to CEU from subsequent analyses). This suggests that the ancestry of the AA samples might be reasonably approximated as a mixture of varying amounts of CEU and YRI ancestry, as reported previously [21]. However, given the wide range of genetic diversity across Europe and particularly across Africa [23], we sought to test this hypothesis further. We removed related samples, genetic outliers, and samples without valid gene expression measurements to obtain a reduced set of 89 AA samples for subsequent analysis (see Materials and Methods). We computed FST values between the set of 89 AA samples and possible linear combinations αCEU+(1−α)YRI, adjusting for sample size. The lowest value of FST = 0.0009 was obtained at α = 0.21. Thus, the 89 AA samples are extremely well-modeled as a mix of CEU and YRI, with average ancestry proportions of 21% CEU and 79% YRI. Though this justifies our modeling approach using CEU and YRI, we caution against drawing historical inferences from this finding: because FST scales with the square of admixture proportion, it is possible that African Americans inherit a small percentage of their ancestry from a more diverse set of populations.
We estimated the genome-wide proportion of European ancestry for each the 89 AA samples (see Materials and Methods). Genome-wide ancestry proportions varied from 1% to 62% with a mean±SD of 21±14%; this ancestry distribution is similar to that in other AA data sets [21],[25]. Genome-wide ancestry estimates were strongly correlated (r2>0.99) with coordinates along the top principal component (eigenvector with largest eigenvalue) (Figure 1).
We measured gene expression in lymphoblastoid cell lines from 60 CEU and 60 YRI samples from HapMap and 89 AA samples from Coriell, using the Affymetrix Genome Focus Array (see Materials and Methods). Our basic approach was to validate observed differences between CEU and YRI (differences between sample means) by analyzing the correlation between the genome-wide proportion of European ancestry estimated from SNP genotyping and the gene expression levels we measured in the AA cell lines. A caveat is that the proportion of European ancestry in African Americans might in principle be correlated to environmental variables. However, such correlations would not affect our approach unless they specifically tracked environmental differences between CEU and YRI. An additional caveat is that the Coriell panel of AA samples is known to be sampled from several (unknown) cities in the United States; AA samples from different U.S. cities might differ systematically in both the average proportion of European ancestry [21],[26] and in the preparation of cell lines. However, ancestry differences among AA populations in different U.S. cities are usually relatively small (standard deviation of 1% in Table 2 of [21]; standard deviation of 6% in Figure 2 of [26]), and in any case would not affect our approach unless differences in cell line preparation specifically tracked differences between CEU and YRI.
Using the ancestry estimates and expression data at 4,197 genes for CEU, YRI and AA samples, we fit a model in which the effect of ancestry on gene expression at gene g is equal to ag per unit of European ancestry for CEU and YRI samples (so that ag is equal to the difference in mean expression level between CEU and YRI, which have ancestry 1 and 0 respectively), and equal to cag per unit of European ancestry for AA samples, where c is constant across genes (see Materials and Methods). Thus, the global parameter c measures the extent to which observed gene expression differences between CEU and YRI are validated in AA, and therefore heritable. If systematic differences observed between CEU and YRI were entirely due to genetic factors, we would expect to see the same ancestry effects in AA samples, so that c = 1. On the other hand, under the hypothesis that observed differences between CEU and YRI are entirely due to non-genetic factors, we would expect c = 0. We note that our procedure for estimating c accounts for both experimental noise and sampling noise in the measurement of gene expression levels. Thus, assuming analogous normalizations for CEU, YRI and AA samples, our estimate of c is not dependent on the accuracy of our measurements; it is also independent of sampling effects.
Fitting the above model, we obtained c = 0.43, the slope of the regression line in Figure 2. With 4,197 genes analyzed, this estimate of c is different from zero with overwhelming statistical significance (P-value<10−25; 95% confidence interval [0.38,0.47]). Thus, gene expression differences among AA samples of varying ancestry strongly confirm that heritable differences contribute to observed gene expression differences between CEU and YRI. Performing the analogous computation with genotype data, we obtained c = 0.96, confirming that c is close to 1 for genetic effects (see Figure 3) and that modeling AA as a mix of CEU and YRI is appropriate for our analyses. The deviation between c = 0.96 and the expected value of 1 is discussed in Text S1.
We investigated whether the correspondence between observed CEU vs. YRI gene expression differences and expression differences due to ancestry among AA is concentrated in genes with large differences between CEU and YRI. If only a fraction of genes were truly differentiated, as suggested by previous studies, then genes with large observed CEU vs. YRI differences would be more likely to be truly differentiated and would show stronger validation in AA. For example, when we simulated a mixture model in which c = 0.43 for the set of all genes but only 50% of genes are truly differentiated between CEU and YRI, we obtained a larger value of c = 0.53 for genes in the top 10% of observed CEU vs. YRI differences (see Text S1). However, Figure 2 shows no evidence of nonlinear effects. Indeed, we recomputed c using only genes in the top 10% of the magnitude of observed CEU vs. YRI differences, and obtained c = 0.44, which is similar to the value of 0.43 using all genes. These results suggest that population differences in gene expression are not restricted to a fraction of genes but in fact are widespread across genes, mirroring population differences at the DNA level [6].
We considered whether the alternative approach of analyzing the AA data independently, without regard to differences between CEU and YRI, would be informative about differences in gene expression due to ancestry. We determined that the AA data analyzed separately contains too much sampling noise for that approach to be useful here (see Text S1). A related observation is that efforts to estimate the proportion of genes with population differences in gene expression, for example using the previously described [27] lower bound statistic 1–π0, may produce substantial underestimates in the case of data sets affected by sampling noise (see Text S1).
The effect of ancestry on gene expression in African Americans may be due either to variation in regulatory variants proximal to the gene (cis) or to variants elsewhere in the genome (trans). We inferred the local ancestry of each AA sample at each location in the genome (see Materials and Methods). A description of how local ancestry varies across the genome (either across or within samples) is provided in Text S1. We quantified the extent to which the validation of CEU-YRI expression differences in AA was attributable to cis or trans effects in AA by computing validation coefficients ccis and ctrans (see Materials and Methods). We obtained ccis = 0.05 and ctrans = 0.38. As expected, the sum ccis+ctrans is very close to the validation coefficient c that was obtained using genome-wide ancestry only (see Text S1). Both ccis (P-value = 6×10−6; 95% confidence interval [0.03,0.07]) and ctrans (P-value<10−25; 95% confidence interval [0.33,0.43]) were significantly different from zero. Thus, only a small fraction of the effect of ancestry on gene expression is due to ancestry at the cis locus. On the other hand, performing the analogous computation with genotype data, we obtained ccis = 0.99 and ctrans = –0.03, indicating as expected that the effect of ancestry on genotype is entirely due to ancestry at the cis locus, and confirming the high accuracy of our estimates of local ancestry.
We estimate the proportion πcis of heritable gene expression variation between Europeans and Africans that is due to cis variants as ccis/(ccis+ctrans) = 12%, with a standard error of 3%. An important question is whether our estimate of πcis can be extended to all heritable variation in human gene expression. If the relative magnitude of cis vs. trans effects were different for all variation as compared to population variation—equivalently, if the relative magnitude of population variation relative to all variation were different for cis vs. trans effects—then the answer to this question would be no. To evaluate whether this is the case, we computed FST(CEU,YRI) for ∼3,000 unique cis eQTL SNPs and ∼700 unique trans eQTL SNPs identified in a recent study of gene expression in human liver [20]. We obtained FST values of 0.158 for cis eQTLs and 0.154 for trans eQTLs, which were not significantly different from 0.159 for all HapMap SNPs (P-values = 0.79 and 0.51 respectively), based on standard errors computed using the EIGENSOFT software [6],[24]. Although this analysis involved eQTLs for liver tissue rather than lymphoblastoid cell lines, a reasonable assumption is that the same result holds for other tissue types. Thus, population variation does not appear to differ for cis vs. trans effects, implying that our estimate of πcis = 12±3% applies to all heritable variation in human gene expression.
We estimated both the proportion of gene expression variation attributable to population differences, which we call F, and the quantity QST = F/(2−F) which is analogous to FST for genetic (allele-frequency) data (see Materials and Methods). We obtained a mean F = 0.20 and median F = 0.12, similar to the median F = 0.15 from a previous analysis of CEU and YRI gene expression [8]. A jackknife calculation indicated that the standard error in our estimate of mean F was 0.02, corresponding to a 95% confidence interval of [0.15,0.25]. In our initial calculation of F, we ignored the possibility of non-genetic contributions to population differences. However, the fact that c is smaller than 1 implies that not all of the observed CEU vs. YRI differences are reflected in differences due to ancestry among AA. Some of these differences must reflect non-genetic factors. We therefore adjusted our estimates of F by multiplying them by c = 0.43 (see Materials and Methods). After this adjustment, we obtained a mean F = 0.09 and median F = 0.05. These estimates of F are substantially lower than those reported previously [8]. Our mean F corresponds to a QST value of 0.05, which is lower than the FST of 0.16 that is observed in genetic data [6]. The lower value of QST as compared to genetic data is unsurprising since QST represents a proportion of total gene expression variation, which is expected to include both genetic and non-genetic components. We also note that if measurement variation is substantial, then the use of technical replicates to correct for the effects of measurement variation would lead to a higher value of QST.
We have shown how phenotypic variation in an admixed population can be coupled with variation in ancestry to shed light on differences between ancestral populations; our approach makes no assumptions about the population histories underlying the differences between the ancestral populations. We have applied this approach to gene expression in African Americans and shown that observed population differences (differences in sample means) between CEU and YRI in gene expression correspond, with overwhelming statistical significance, to differences among African Americans of varying ancestry, implying a substantial heritable component to the population differences. In reaching this conclusion via analysis of an admixed population, we eliminate confounding with non-genetic contributions to observed differences between the ancestral populations, which could result from differences in environment, differences in preparation of cell lines, or batch effects. The value of 0.43 for the “validation coefficient” c implies that both genetic and non-genetic effects contribute to observed population differences between CEU and YRI.
Interestingly, the validation coefficient c did not vary appreciably as a function of the magnitude of observed gene expression differences between CEU and YRI. This suggests that the effects of ancestry on gene expression are widespread across genes, as opposed to affecting only a fraction of genes. Although there exist genes for which the observed effect of ancestry on expression levels is close to zero (Figure 2), this does not rule out small ancestry effects at these genes, as similar results are observed in genetic data (Figure 3) in which it is commonly believed that ancestry affects 100% of common SNP frequencies. Indeed, if ancestry affects genotype and genotype affects gene expression (as indicated by previous studies reporting a substantial heritable component to gene expression [16],[17]), then the presence of ancestry differences at almost all expressed genes seems a not unreasonable hypothesis, and one with which our results are entirely consistent. However, just as with DNA variation, it is clear that population differences in gene expression represent only a small fraction of the overall variance, most of which is due to variation within populations.
In addition to validating the aggregate effects of ancestry on human gene expression, we were able to partition heritable variation into cis and trans effects, which would not be possible in a simple comparison of continental populations. Our admixture approach was fruitful despite the small magnitude of differences between human subpopulations. Our distinction between cis and trans effects is somewhat imprecise, due to the extended length (>10 Mb) of segments of continental ancestry in African Americans, but this has little effect on our conclusions, since a 10 Mb region represents a proportion of the genome that is much smaller than the 12% proportion of heritable variation in gene expression that we attribute to variation at the cis locus. Comparing our results to results obtained in other species, we note that two recent studies of gene expression in Drosophila also reported that cis effects explain a small fraction of heritable variation [28],[29], although previous Drosophila studies had suggested a larger role for cis effects [30],[31]. Our results have broad ramifications for future efforts to map the genetic regulation of gene expression. However, conclusions drawn from gene expression measured in lymphoblastoid cell lines do not necessarily extend to other tissue types, motivating further investigation. Going forward, admixed populations will continue to be useful for understanding and mapping gene expression and other phenotypes.
|
10.1371/journal.pgen.1007346 | Synergistic co-regulation and competition by a SOX9-GLI-FOXA phasic transcriptional network coordinate chondrocyte differentiation transitions | The growth plate mediates bone growth where SOX9 and GLI factors control chondrocyte proliferation, differentiation and entry into hypertrophy. FOXA factors regulate hypertrophic chondrocyte maturation. How these factors integrate into a Gene Regulatory Network (GRN) controlling these differentiation transitions is incompletely understood. We adopted a genome-wide whole tissue approach to establish a Growth Plate Differential Gene Expression Library (GP-DGEL) for fractionated proliferating, pre-hypertrophic, early and late hypertrophic chondrocytes, as an overarching resource for discovery of pathways and disease candidates. De novo motif discovery revealed the enrichment of SOX9 and GLI binding sites in the genes preferentially expressed in proliferating and prehypertrophic chondrocytes, suggesting the potential cooperation between SOX9 and GLI proteins. We integrated the analyses of the transcriptome, SOX9, GLI1 and GLI3 ChIP-seq datasets, with functional validation by transactivation assays and mouse mutants. We identified new SOX9 targets and showed SOX9-GLI directly and cooperatively regulate many genes such as Trps1, Sox9, Sox5, Sox6, Col2a1, Ptch1, Gli1 and Gli2. Further, FOXA2 competes with SOX9 for the transactivation of target genes. The data support a model of SOX9-GLI-FOXA phasic GRN in chondrocyte development. Together, SOX9-GLI auto-regulate and cooperate to activate and repress genes in proliferating chondrocytes. Upon hypertrophy, FOXA competes with SOX9, and control toward terminal differentiation passes to FOXA, RUNX, AP1 and MEF2 factors.
| In the development of the mammalian growth plate, while several transcription factors are individually well known for their key roles in regulating phases of chondrocyte differentiation, there is little information on how they interact and cooperate with each other. We took an unbiased genome wide approach to identify the transcription factors and signaling pathways that play dominant roles in the chondrocyte differentiation cascade. We developed a searchable library of differentially expressed genes, GP-DGEL, which has fine spatial resolution and global transcriptomic coverage for discovery of processes, pathways and disease candidates. Our work identifies a novel regulatory mechanism that integrates the action of three transcription factors, SOX9, GLI and FOXA. SOX9-GLI auto-regulate and cooperate to activate and repress genes in proliferating chondrocytes. Upon entry into prehypertrophy, FOXA competes with SOX9, and control of hypertrophy passes to FOXA, RUNX, AP1 and MEF2 factors.
| In the formation and longitudinal growth of endochondral bones, committed mesenchymal cells condense and differentiate into chondrocytes to form a growth plate, within which chondrocytes undergo coordinated and sequential differentiation phases of proliferation, cell cycle exit and hypertrophy, resulting in longitudinal bone growth[1, 2]. Endochondral bone formation requires tightly controlled proportions of the different chondrocyte populations, recognized by their distinct morphology, characteristic gene expression patterns, and organization into different zones. Firstly, round chondrocytes become proliferative, flattening to form columns. As proliferating chondrocytes (PCs) mature, they exit the cell cycle and enter a prehypertrophic phase. This phase is an important transition, which produces signals for maintaining proliferation on the one hand and on the other, to regulate the progression from proliferation to cell cycle exit, entry into a prehypertrophic state, followed by the final stages of differentiation in which the cells enlarge to form hypertrophic chondrocytes (HCs) and then become osteoblasts [1–4]. Disruption of the progression from one differentiation state to the next and the relative proportions results in skeletal defects such as chondrodysplasia [5, 6].
The sophisticated program of chondrocyte differentiation requires the activation or repression of many genes, which is strictly mediated by transcription factors (TFs), including the SOX (SOX5, SOX6 and SOX9)[7–13], GLI (GLI1, GLI2 and GLI3) [14–17], RUNX (RUNX2 and RUNX3)[18, 19], MEF2C[20], AP1[21] and FOXA[22] family members[23].
SOX9 is the master regulator of chondrocyte differentiation. Chondrocytes cannot form in Sox9 null mutants and heterozygous mutations in SOX9 severely disrupt skeletal development, causing campomelic dysplasia [7, 8, 10, 11, 13]. Shortly after mesenchymal condensation, SOX9 cooperates with SOX5 and SOX6 to activate the expression of cartilage matrix genes, e.g., Col2a1 and Aggrecan, positively regulating chondrocyte proliferation, while inhibiting both the progression of these cells to hypertrophy and the osteogenic program [11, 24–26].
The Hedgehog signaling pathway regulates chondrocyte proliferation and hypertrophy through a complex negative feedback loop with the PTHrP signaling pathway [27]. IHH secreted from prehypertrophic chondrocytes (PHCs) activates HH signaling and GLI transcription factors (GLI1/2/3) in proliferating chondrocytes (PCs) via its receptor PTCH1, and stimulates the expression of Pthrp. GLI1 functions as an activator which is highly expressed in PCs and perichondrium adjacent to the prehypertrophic and hypertrophic zones. GLI2 has both activator and repressor forms and GLI3 acts as a repressor. GLI2 is expressed in most chondrocytes, but at a lower level in HCs [28]. Binding of PTHrP to its receptor PPR results in activation of PKA and phosphorylation of SOX9 that enhances its transcriptional activity [29], indicating crosstalk between SOX9 and IHH signaling in regulation of chondrocyte proliferation and differentiation. RUNX and FOXA transcription factors are critical regulators of hypertrophic chondrocyte maturation. Mutations in RUNX2 cause the skeletal disorder, Cleidocranial dysplasia [30]. In mice, inactivation of Runx2 or FoxA2/FoxA3 causes severe defects in chondrocyte hypertrophy and bone formation [18, 22].
Although these TFs have been studied individually for their importance in chondrocyte differentiation, understanding of how they interact and integrate into a gene regulatory network (GRN) that acts genome wide, is still only emerging and largely limited to addressing control of individual gene expression (reviewed in [23]). For example, in vitro transactivation assays in cultured chondrocytes highlight a potential in vivo cooperation between GLI1/2 and RUNX2/SMADs in activating Col10a1 via interaction with its promoter [31]. The Notch signaling pathway transcriptional co-activator, Mastermind-like 1 (MAML1), was reported to enhance the transcriptional activity of RUNX2[32]. Recently genome-wide analyses of SOX9 binding peaks in chondrocytes [33] assisted the discovery that SOX9 and AP1 factors (Jun) co-activate Col10a1 in prehypertrophic chondrocytes to promote hypertrophy [21]. The SOX proteins are characterized by their dependence on partner factors in controlling cell differentiation [34]. Cooperation of the SOX trio proteins (SOX9, SOX5 and SOX6) controls sequential differentiation of chondrocytes [9, 11, 12, 35–37]. Cooperative interaction between SOX9 and other factors such as AP-1, NFAT, FOXA, RUNX and HOX has also been implicated because their binding motifs are enriched in the SOX9 peaks [21, 33]. It has been reported that GLI1 can regulate Sox9 via a far upstream enhancer [38] and SOX9 can regulate its own expression via another far upstream enhancer [39]. We previously implicated cooperation between SOX9 and GLI factors in repressing Col10a1 expression in proliferating chondrocytes [40]. Despite the wealth of information about the individual roles of SOX9 and GLI in regulating chondrocyte differentiation, it is not fully understood about how these TFs together mediate the transition from PCs and PHCs where SOX9 and GLI factors dominate, to HCs controlled by FOXA, RUNX and other factors [40]. Also little is known whether these two factors cooperate to activate chondrocyte genes and if they do, the breadth of potential genes that are cooperatively regulated by these factors.
The need to understand the regulatory mechanisms driving the phases of chondrocyte differentiation in the growth plate has prompted investigations to establish global transcriptomic analyses for gene signatures for the different populations. Prior transcriptomic studies on mouse chondrocyte populations had been narrowly focused on chondrogenic cell lines [41], early stage limb mesenchyme and E13.5 chondrocytes before hypertrophy [42], manually dissected tibia segments [43], or postnatal proliferating and hypertrophic chondrocytes without transition zones [44, 45]. A recent study on the transcriptomes of 217 single cells from the growth plate aimed to reconstruct the spatial and temporal pattern gene expression of individual chondrocytes [46]. However only a limited number of genes were mapped in that dataset and many genes important for chondrocyte differentiation were not detected (e.g. Ctnnb1, Gli2, Wnt5a, Wnt5b, etc.). The spatial information on gene expression is also lacking since the cells were not fractionated according to zones.
The limited information on the integration of the GRN that controls the important transitions from one differentiation phase to the next within the growth plate motivated us to develop a comprehensive atlas of gene expression for finely fractionated chondrocyte subpopulations in growth plate. We aimed to use this resource for the analyses and discovery of the complex molecular signatures, differential gene expression patterns, biological processes and pathways operating during the phases of chondrocyte differentiation, especially in the transition into prehypertrophy. We created a searchable library, GP-DGEL (URL: http://www.sbms.hku.hk/kclab/gp.html), that provides sequential and dynamic gene expression information encompassing growth plate chondrocytes at different stages from proliferative to prehypertrophy, early and late hypertrophy. By integrative analysis of the transcriptome and chondrocyte ChIP-seq datasets coupled with functional tests, we find evidence for a dominant role for SOX9-GLI cooperation in proliferating chondrocytes and identify new SOX9 GLI targets. Importantly we find evidence for a model of phase transition of the gene regulatory program from SOX9-GLI cooperation to SOX9-FOXA competition directing chondrocyte differentiation. In this SOX9-GLI-FOXA centric model GRN, SOX9-GLI preferentially activates sets of genes in proliferating chondrocytes. In maturing prehypertrophic chondrocytes the SOX9-GLI nexus begins to fade, so that in full hypertrophy, control is relayed to an alternative set of transcription factors including FOXA2, RUNX2, AP1 and MEF2C.
We generated GP-DGEL, by fractionating the mouse proximal tibial growth plate at postnatal day 10 (P10) into four zones representing chondrocyte sub-populations: PCs in the proliferating zone (PZ), PHCs in the pre-hypertrophic zone (PHZ), early HCs in the upper hypertrophic zone (UHZ) and late differentiated HCs in the lower hypertrophic zone (LHZ) (Fig 1A). Navigating the location and identity of these zones was guided by the morphologies and RT-PCR analyses for the expression of zone-characteristic markers (Col2a1, Col10a1, Ppr, Ihh and Mmp13; S1A–S1C Fig). Gene expression profiling data from the 4 fractions of biological triplicates were generated for further analysis (S1D Fig).
We defined the set of genes expressed at each stage: a total number of 4799, 4811 and 4879 genes were expressed in PZ, PHZ and HZ (average of the gene expression in UHZ and LHZ) respectively. We categorized all the expressed genes that were commonly or uniquely expressed in PZ, PHZ and HZ (see Methods and S1E Fig). Here, 4792 out of 4886 genes (98%) are commonly expressed showing the “On” state in all three zones. The remainder 94 genes are expressed in only one or two zones (S1 Table). Only one gene, Fzd9, was expressed only in PCs. No genes were specifically expressed only in PHCs, as might be expected for cells in transition from proliferation to hypertrophy, but 73 genes were HC-specific.
GP-DGEL allows the detection of variation in gene expression across different zones. Amongst all the expressed genes, 1891 genes (~37%) were differentially expressed with the Coefficients of Standard Deviation (CSD) over the 4 zones greater than 0.15 (S2A Table). The differentially expressed transcription factors included Sox9, Sox5, Gli1, Gli2, Runx3 and Mef2c in PZ and PHZ, which is consistent with their known roles in regulating endochondral ossification, affirming the reliability of the dataset. The remaining genes were constantly expressed across the zones, including Hif1a, which plays an important role in bone development but is regulated through post-translational modification [47–49].
Using k-means clustering analyses to categorize the patterns of the 1891 differentially expressed genes (DEGs), we identified four major clusters: Cluster I genes exhibit decreasing expression from PZ to HZ (654 genes); Clusters II to IV genes are typically most highly expressed in PHZ (299 genes), UHZ (31 genes) and LHZ (907 genes), respectively (Fig 1B; S2A Table). This categorization formed the basis for genome-wide discovery and identification of biological processes, pathways and GRNs that underlie these transition patterns. We tested GP-DGEL for its capacity as a resource for the discovery and functional analyses of signaling pathways, biological processes and transcriptional regulators of chondrocyte differentiation as follows.
To identify the enriched biological processes and signaling pathways for each cluster, we performed Gene Ontology enrichment analysis. Genes associated with biology process of “skeletal system development” were enriched in PZ, PHZ and HZ, establishing the authenticity of our data (Fig 1C; S3 Table). Genes associated with the processes of “regulation of Smoothened signaling pathway”, “transcription”, and “cell proliferation” were significantly enriched in the PZ (cluster I) and supported a significant role for IHH signaling (mediated by Smoothened and Gli) in proliferating chondrocytes. Genes associated with the processes of “sterol metabolic process”, “cell motility, “actin cytoskeleton regulation” and “cell growth” were most common in the PHZ (cluster II). This agrees with the dramatic changes in chondrocyte size and morphology observed during hypertrophy. Sterol (cholesterol) biosynthesis is required for the processing and maturation of hedgehog ligands and Hedgehog signaling [50]. In explant organ cultures, cholesterol was found to stimulate chondrocyte hypertrophy and bone growth through regulating the expression of Rora. Inhibition of cholesterol biosynthesis attenuates chondrocyte enlargement [51] and results in growth retardation with decreased chondrocyte proliferation and Ihh expression [52].
The UHZ (cluster III) was enriched for genes associated with the process of “extracellular matrix (ECM) organization”, which is consistent with the transition from synthesis of an ECM rich in collagen II to one where collagen X is the major component [53]. The LHZ (cluster IV) was enriched for “hydrogen transport”, “vascular development”, “cell redox homeostasis”, “phosphate metabolic process”, “ossification” and “regulation of apoptosis”, consistent with late-stage differentiation, cartilage calcification, degradation, vascular invasion, bone formation and chondrocyte to osteoblast trans-differentiation that occur at the chondro-osseous junction[2, 3] where cell cycle re-entry in the process has been implied[4]. The enrichment for “regulation of apoptosis” is intriguing: of the 37 genes highlighted, 16 genes were classed as contributing to "negative regulation of apoptosis" (p-value = 3.5e-18) and 9 genes as contributing to "positive regulation of apoptosis" (p-value = 1.9e-13) such as Cdkn1a (p21) a cell cycle regulator and its interacting pro-apoptotic factor Trp53inp1 which may imply complex control that balances apoptosis and survival and control of cell cycle re-entry in the transition from hypertrophic chondrocytes to osteoblasts. Overall the corroboration of genes associated with processes that occur in the relevant zones attests to the quality of the library.
Many signaling pathways are known to play key roles in coordinating chondrocyte proliferation and differentiation but their relative importance in each phase and sub-population is unclear. A gradient of BMP pathway gene expression has been reported for the rat postnatal growth plate [45] but how this compares with other pathways is not known. To gain global mechanistic insight into the relative scope of signaling action in each zone, we computed the enriched GO terms for the genes involved in canonical signaling pathways: WNT, BMP/TGFβ, FGF, Notch, IGF, Hippo and Hedgehog (S4 Table). We found the expression of the components in WNT, BMP/TGFβ, Hippo, FGF and Notch pathways was not significantly enriched in particular regions over the 4 different zones (Fisher’s exact test p-value > 0.05). In contrast, genes of Hedgehog and IGF pathways were preferentially expressed in PZ and PHZ, suggesting distinctive roles of these signaling pathways in regulating cell cycle progression and the initiation of chondrocyte hypertrophy.
We tested the capacity of GP-DGEL to identify potential associations of the differentially expressed genes with those implicated in mouse skeletal phenotypes and human skeletal diseases in MGI and OMIM databases. A subset of 396 genes, accounting for 20% of the whole list, was associated with abnormal skeletal phenotypes in mouse (S2B Table). 93 genes were associated with human skeletal disorders (S2C Table). To infer the functional significance of the phase-specifically expressed TFs on skeletal development, we ranked the TFs according to the CSD values over 4 zones (S5 Table). Of 76 phasic-specific TFs, 46 (including Sox9 and Trps1) were associated with human skeletal disorders and/or mouse skeletal defects. GP-DGEL can therefore be used to identify new candidate genes of skeletal disorders. An example worthy of further investigation is Srebf1, not known to cause skeletal disorders but is implicated as a regulator of cholesterol metabolism and apoptosis (OMIM 184756) [54].
The large sets of genes sharing phasic-specific expression patterns imply the presence of a coordinated transcriptional program at each phase. For an unbiased identification of phase specific transcriptional regulators, we performed de novo motif enrichment analysis in the promoter regions for the genes in each cluster using the computer program Discriminating Motif Enumerator (DME), MotifClass and MatCompare [55–58]. The most enriched TF binding motifs include SOX9 and GLI (GLI1, GLI2 and GLI3) in the PZ and PHZ; SOX9/FOXA and KLF4 in the UHZ; and MEF2C and FOXA motifs in the LHZ (Fig 2A). The core consensus binding motifs for FOXA factors (FOXA1, FOXA2 and FOXA3) and SOX9 comprise highly similar AT-rich sequences (ACAAA-like for FOXA; ATTGT-like for SOX), raising the possibility that these factors compete for binding in regulating gene expression [22].
The over-representation of a GLI motif in the PZ and PHZ genes agrees with the known action of Hedgehog signaling in PCs. Gli1, itself a target of Hedgehog signaling, is most highly expressed in the PZ, while the cytoplasmic GLI3 repressor may transform into an activator in the presence of IHH [59]. KLF4 motif enrichment in the UHZ cluster could imply a role in promoting hypertrophy which would be consistent with its capacity to reprogramme dermal fibroblasts in concert with SOX9 and cMYC [60]. In the LHZ cluster, we detected enrichment for the binding motif for MEF2C, a vital regulator of chondrocyte hypertrophy that is required for the proper expression of Col10a1, Runx2 and Vegf [20]. Beyond the chondrocyte fate, KLF4 and MEF2C may prime the lineage progression of hypertrophic chondrocytes to osteoblasts [3, 61, 62].
SOX9 functions as a dimer in chondrocyte differentiation [63]. To predict the degree to which SOX9 dimer/monomer binding motifs were utilized in chondrocyte gene regulation, we screened for evolutionarily conserved SOX9 binding sites located within 10kb from the transcriptional starting site (TSS). Using the MEME program [64] for long consensus motif analysis, we identified SOX9 dimer motifs with varied length of spacer sequences (Fig 2B). The most enriched SOX9 dimer motifs were identified for PZ and PHZ genes (p-value<1.0e-5). In proliferating chondrocytes, SOX9 dimer motifs were associated with the genes which it activates (e.g. Col2a1) or represses (e.g. Col10a1) [24, 40, 65]. The length of the spacer sequences in the dimer motifs ranges from 4 to 13-bp in the PZ and 4 to 16-bp in the PHZ genes (S6 Table), raising the question whether the variation in the linking sequences could confer different specificity of co-binding of partner factors with SOX9 dimers [66, 67]. For LHZ where SOX9 protein level dropped to undetectable level, no significant SOX9 dimer binding motifs were identified.
The enrichment of SOX9 binding motifs in the DEGs (Fig 2A and 2B) is consistent with the vital roles of SOX9 in regulating phasic gene expression and helps identification of target genes regulated by SOX9 at each stage of chondrocyte differentiation. We searched the evolutionarily conserved noncoding DNA elements across 30 vertebrates in gene promoter, intergenic, intronic and 3’- UTR regions for putative SOX9 monomer and dimer binding sites (S7A–S7D Table). To identify functional binding sites, we integrated the bioinformatics predictions with the SOX9 ChIP-seq dataset from mouse neonatal rib chondrocytes [33]. Overall, 503 genes out of 654 in the PZ cluster, 250 out of 299 in the PHZ cluster, 24 out of 31 in the UHZ cluster and 664 out of 907 in the LHZ cluster were found to harbor at least one SOX9 binding region (SBR) (S8A1–S8B2 Table), consistent with the major role of SOX9 in regulating chondrocyte differentiation. SOX9 can act as both an activator and a repressor in PCs [40]. Therefore the genes identified by this analysis could be either activated or repressed by SOX9.
We selected those genes with predicted SOX9 binding sites located within 250-bp from the SOX9 ChIP peaks as potential SOX9 targets. Multiple copies of monomer and dimer sites near the SOX9 binding peaks were identified for known SOX9 targets Sox9, Sox5, Sox6, Col2a1, Acan and Col10a1 (Fig 2C; S7E and S7F Table). In the Col10a1 locus, we found a SOX9 binding peak 4.4 kb upstream of the TSS (Fig 3H), where an element has been shown to mediate repression by SOX9 in non-hypertrophic chondrocytes [22, 40]. These data affirm the validity of our approach.
We identified several potential SOX9 targets (Zbtb20, Wwp2, Foxp2, Ppa1, Slc8a3, Bnip3 and Wnk4). By in situ hybridization or antibody staining on proximal tibia growth plate, we confirmed the expression patterns of these targets (Fig 3A–3I) as corresponding with the regions identified in GP-DGEL. These potential SOX9 targets may function in different steps of chondrocyte differentiation and endochondral bone formation. For instance, Wwp2 (Fig 3B) has been identified as a direct SOX9 target during palatogenesis [68]. Interestingly many genes encoding major components of the IHH signaling pathway were identified as potential SOX9 targets in our study, including Ihh, Ptch1, Gli1, Gli2 and Gli3 (Fig 2C), which is consistent with the enrichment for GLI binding motifs for the PZ and PHZ clusters (Fig 2A).
To validate SOX9 binding under the SOX9 peaks, we performed in vivo ChIP-qPCR assays on three candidates: Cyr61[69, 70], Trps1 [59, 71] and Ptch1 (S9 Table), with an Aggrecan enhancer situated 10kb up-stream of TSS as a positive control [12]. The ChIP-qPCR results showed that SOX9 binds to the promoters of Cyr61 and Ptch1, and the intron 1 of Trps1 (S2A–S2D Fig). Within the regions covering the validated binding sites, we detected SOX9 binding peaks (Fig 4A–4C), indicating that these genes are direct SOX9 targets.
To test whether the expression of these candidates is associated with SOX9 activity, we compared their expression levels in wild type (Sox9+/+) and heterozygous null (Sox9+/-) mutant littermates in embryonic day 13.5 (E13.5) limbs, when the limb abnormality is minimal [72]. Expression of known SOX9 targets (Sox9, Sox5, Sox6 and Col2a1) was down regulated in Sox9+/- mutants compared with wild type littermates (Fig 5A), consistent with the dosage requirement for SOX9. Expression of Cyr61, Trps1, Ptch1, Gli1 and Gli2 was reduced by approximately 50% in Sox9+/- mutants (Fig 5A), and Gli3 expression has been reported to be reduced in Sox9+/- mutants [35], indicating that SOX9 positively regulates these genes. Ihh is expressed in PHCs [14], and its expression is not significantly changed in Sox9+/- mice, consistent with the previous finding that the expression of Ihh is not affected by Sox9 heterozygous mutation [72].
Since Ptch1 is down-regulated in Sox9+/-mutants, we tested whether Ptch1 is transcriptionally regulated by SOX9. Using the chondrogenic cell line ATDC5, we found SOX9 transactivated a luciferase reporter driven by regulatory sequences in the Ptch1 promoter region containing a SOX9 binding peak in a dosage-dependent manner in (Fig 5C), indicating that SOX9 directly regulates the expression of Ptch1.
The enrichment for SOX9 and GLI motifs in the PZ and PHZ clusters (Fig 2A), the co-expression of Sox9 and Gli1 in the PCs (S3A and S3B Fig), the known GLI activation of Ptch1 [73] and the cooperative repression of Col10a1 by SOX9-GLI3R [40] raise the possibility of a substantial role for cooperation of SOX9 with GLI in activating gene expression. To investigate whether SOX9 and its targets are co-regulated by GLI factors, we screened the phasic DEGs for putative GLI binding sites (S7E Table). Abundant GLI consensus motifs were found near SOX9 peaks in Sox9 itself and SOX9 target, in particular Sox9, Sox5, Sox6 and Col2a1 (Fig 2C). Since it has been reported that cells derived from a common progenitor lineage share similar genome-wide epigenetics and TF binding profiles[74], we integrated the bioinformatics predictions with the SOX9 ChIP-seq from newborn rib chondrocytes (S8A1 and S8A2 Table), GLI1 (S8C1 and S8C2 Table) and GLI3 (S8E1 and S8E2 Table) ChIP-chip datasets from E11.5 developing limbs[73] to check whether in principle, these TFs could bind to the putative common target genes.
Binding regions for SOX9 (SBR), GLI1 and GLI3 (GBR) were found in 1426, 699 and 1421 phasic DEGs respectively (SOX9, S8B1 and S8B2 Table; GLI1, S8D1 and S8D2 Table; and GLI3, S8F1 and S8F2 Table). Among these, 721 of 1426 SOX9-targeted DEGs (51%) harbored at least one SOX9/GLI linked binding region (SGBR) with an inter-peak distance shorter than 250-bp (S10A–S10D Table). The genes that were most enriched for SGBRs include the known SOX9 targets Sox9, Sox5, Sox6 and Col2a1. Interestingly substantial over-representation of putative SOX9/GLI common targets was found for the PZ, PHZ and UHZ clusters compared to genes in the LHZ cluster (p-value<0.01), consistent with the expression pattern of SOX9 protein which spans the PZ, PHZ and persists into the UHZ [26]. The correlation of SGBRs with phasic gene expression decreased as the inter-peak distance increased (S10A1, S10D2 and S10D2 Table). Correlation was also found between phasic-specific genes and SGBRs that were located in the intergenic regions (p-value = 0.0032, S10A3 Table), suggesting SOX9-GLI may also mediate long-range regulation.
These bioinformatics predictions suggest that SOX9 and GLI factors cooperate to regulate common targets in PCs and PHCs. To test these predictions we examined the SOX9 ChIP-seq data and found that the SOX9-bound regions in Trps1 and Ptch1 loci were co-localized with the GLI1 binding peaks (Fig 4B and 4C). We tested the ability of SOX9 and GLI singly and in combination to transactivate expression of the Ptch1-luciferase reporter vector. Both SOX9 and GLI1 could drive the expression of the Ptch1-luciferase reporter (Fig 5C and S3C–S3F Fig). We also detected SOX9 binding peaks in the Gli1 and Gli2 loci (Fig 4D and 4E). In Sox9+/- mutants, the expression of Gli1 and Gli2 was down-regulated (Fig 5A), indicating that these genes may be regulated by SOX9.
Firstly we tested for the cooperative control of GLI on SOX9 targets. We found GLI1 peaks in SOX9 target genes, including Sox9, Col2a1, Sox5 and Sox6, which are close to SOX9 binding regions (Fig 4F–4I). To test whether the expression of SOX9 targets is affected by removal of the GLI activator, we compared their expression levels in Gli2+/+ and Gli2-/- littermates. In Gli2-null mutants, with the exception of Cyr61, the SOX9 targets Sox9, Col2a1, Sox5, Sox6 and Gli1, Trps1 and Ptch1 were markedly downregulated (Fig 5B), consistent with cooperative regulation by SOX9 and GLI.
We next tested the cooperative activity of SOX9 and GLI in regulating Sox9, Col2a1, Ptch1, Gli1 and Gli2 by transactivation assays using luciferase reporters driven by genomic fragments containing at least one SGBR. GLI1 and GLI2 transactivated the Ptch1, Gli1 and Gli2 reporters. This transactivation activity was significantly enhanced by SOX9 (Fig 5D–5F). SOX9-dependent transcriptional activation of Sox9 and Col2a1 reporters was enhanced by GLI1 and GLI2 (Fig 5G and 5H), confirming the addictive action of SOX9 and GLI activators in the regulation of common target genes.
While SOX9 and GLI factors play key roles in the GRN of PCs and PHCs, they are not expressed in hypertrophic chondrocytes in the LHZ (Fig 6A, S3A and S3B Fig). We therefore sought to gain insight into the GRN that mediates the transition from proliferation to prehypertrophy and hypertrophy. It is notable that Foxa2, a critical regulator of hypertrophy [22], is mainly expressed in PHCs and HCs (Fig 6B). Interestingly, SOX9 and FOXA2 are co-expressed in PHCs and early HCs (Fig 6C). We analyzed the published FOXA2 ChIP-seq dataset [75] which has been used in other studies to identify multiple target genes in the notochord [76], and found multiple FOXA2 binding peaks in the Sox9 promoter and distal gene regulatory elements (S8G Table). As Sox9 expression diminishes in the UHZ whilst Foxa2 is robustly expressed, it is possible that FOXA2 represses Sox9 expression.
As FOXA2 and SOX are co-localized in PHCs and early HCs and bind closely related AT-rich DNA elements [22], FOXA2 and SOX9 could compete for binding sites and alter the dynamics of the regulatory phase. To study the relationship between these two factors, we expressed SOX9 and FOXA2 in ATDC5 cells and examined their impact on the transactivation of two established SOX9 targets, Col2a1 and Col10a1, using promoter/enhancer-driven luciferase reporter expression. As expected, SOX9 transactivated the expression of a Col2a1-luciferase reporter (Fig 7A). However, this transactivation was progressively weaker with increasing amounts of FOXA2. Notably, FOXA2 alone did not transactivate the Col2a1 reporter. The reporter was progressively activated with increasing amounts of SOX9 (Fig 7B).
SOX9 represses the expression of Col10a1 by direct binding to the conserved regulatory region located between −4.3 and −3.6 kb of the mouse Col10a1 gene [40]. We tested the ability of FOXA2 to activate the expression of luciferase driven by the Col10a1 promoter (-832bp to +68bp) with the enhancer region (-4433bp to -3780bp). FOXA2 alone could transactivate the Col10a1-luciferase reporter and this activation was gradually dampened with increasing amounts of co-transfected SOX9 (Fig 7C). SOX9 alone did not transactivate the Col10a1 reporter and the repression was released with increasing amounts of FOXA2 (Fig 7D).
To test further the transcriptional competition between SOX9 and FOXA, we selected known regulatory regions from the Col2a1 and Col10a1 loci [24, 40, 77] containing SOX9/FOXA binding motifs and carried out electrophoretic mobility shift assay (EMSA) with homogenously purified FOXA and SOX9 protein constructs (Fig 7E, Gel I-IV). We found that both FOXA and SOX9 effectively associated with the Col2a1 and Col10a1 sequences with FOXA migrating as a monomer (Fig 7E, Gel I-IV, lane 7) whereas SOX9 migrated as a monomer or dimer under equilibrium conditions (Fig 7E, Gel I-IV, lane 8). The SOX9 monomer fraction predominates at the tested concentration suggesting that the homodimer cooperativity is profoundly weaker than for canonical SOXE DNA elements in the reverse-forward (ACAATGN3-5CATTGT) configuration [66]. On the Col2a1 element, SOX9/FOXA heterodimer fractions appeared under conditions when the FOXA monomer is also formed suggesting that FOXA is able to interact with DNA bound by SOX9 monomers (Fig 7E, Gel I, lane 2 and 3). At high FOXA concentration when DNA probes become limiting, the SOX9 monomer disappear and the FOXA monomer and SOX9/FOXA heterodimer become dominant (Fig 7E, Gel I, lane 1). The dimeric SOX9/DNA complex persisted even at very high FOXA concentrations suggesting a highly stable association of dimeric SOX9. In the inverse experiment when the FOXA concentration is fixed and SOX9 is increased, the SOX9/FOXA heterodimer is formed equally well and at high Sox9 concentration the Sox9 homodimer is formed at the expense of the FOXA monomer (Fig 7E, Gel II), suggesting that SOX9 and FOXA can from heterodimers on Col2a1 DNA in an un-cooperative fashion but SOX9 homodimers and FOXA monomers are incompatible and compete. On the Col10a1 element, SOX9 also forms a homodimer with similar efficiency as on Col2a1 DNA whilst FOXA binds monomerically. However, a SOX9/FOXA heterodimer is barely visible on this element (Fig 7E, Gel III and IV, lane 1–3). Interestingly, the presence of FOXA counteracts the formation of monomerically bound SOX9/DNA complexes but favors the formation of dimeric SOX9/DNA complexes (Fig 7E, Gel III and IV, compare lanes 1–3 with lanes 8). This indicates that dimeric SOX9 more effectively resists competition by FOXA than monomeric SOX9. Together, these findings demonstrate that FOXA and SOX9 possess the capacity to associate with highly similar DNA sequences and indicate that competition between SOX9 and FOXA is a plausible mechanism for the transcriptional switches during chondrogenesis.
In our study we have aimed to provide insights into the gene expression dynamics and gene regulatory network that guide chondrocytes through their phases of differentiation in the growth plate. Although the cells in each region were not pure populations, especially in the LHZ which is adjacent to the primary ossification center with vascular invasion, the expression profiles of many chondrogenic markers (Sox9, Sox5, Sox6, Wwp2, Col2a1, Col9a1, Acan, Comp, Ihh, Ptch1, Gli1, Gli2, Ppr, Fgfr3, Igf1, Bmp6, Wnt5b, Dkk1, Cdkn1c, Mef2c, Bmp2, Col10a1, Mmp9, Mmp13, et al) did show high consistency with the published data. Cognizant of the potential limitations we have validated the expression of the novel genes (Zbtb20, Foxp2, Slc8a3, Ppa1 and Bnip3). Therefore analysis of these microarray data still provides vast transcriptomic information on chondrocyte differentiation. Towards that end, we developed a library of differentially expressed genes, GP-DGEL that has fine spatial resolution and global transcriptomic coverage, allowing systematic analyses of the genes that regulate transition between these phases. GP-DGEL is a valuable resource to complement efforts to identify causative mutations in skeletal dysplasia and predict the underlying GRN. This is illustrated by our correlative analyses of the 1891 DEGs with the MGI and OMIM databases, which identified genes associated with mouse and human skeletal disorders and additional candidates (S2 Table). GP-DGEL has also enabled the identification of new gene signatures. Many of the DEGs remain poorly studied in chondrocytes. Integration of the dataset with global ChIP-seq data allows the identification of target genes for TFs, singly and in combination, thereby revealing cooperative activities. Using this approach we identified new targets for SOX9 and evidence for SOX9-GLI cooperation.
We validated several of the predicted SOX9 targets (Cyr61, Trps1, Ptch1, Gli1 and Gli2) by functional assays. The downregulation of Cyr61, Trps1 and Gli2 in Sox9+/- chondrocytes in a recent report [35] is in agreement with our data. The presence of SOX9 peaks associated with these genes in SOX9 ChIP-seq data from rat chondrosarcoma cells is also consistent with direct regulation [37]. We also confirmed the expression patterns of other potential SOX9 targets (Zbtb20, Wwp2, Foxp2, Ppa1, Bnip3, Slc8a3 and Wnk4) that were identified based on the presence of associated SOX9 binding peaks (Fig 3). These genes are candidates for functional studies. An example is Wnk4, which is expressed in late PHCs and early HCs (Fig 3G). WNK4 is the major regulator of the Na-Cl co-transporter in the kidney, a regulator of adipogenesis and energy metabolism and a causal gene for pseudohypoaldosteronism type II [78–80], but has no known role in chondrocyte hypertrophy.
A major outcome of the integrated approach is the identification of genes that are co-regulated by both SOX9 and GLI factors. Zbtb20, highly expressed in PHCs and downregulated in the UHZ, is a potential SOX9-GLI target since SOX9 and GLI binding peaks were identified in the locus (Fig 3A). Ablation of Zbtb20 in chondrocytes results in an expanded HZ, and delayed vascularization [81], consistent with a role downstream of SOX9 in regulation of the transition from prehypertrophy to hypertrophy [26]. This function may additionally require co-regulation of SOX9-GLI. Other predicted targets that can be followed up in functional analyses include Fgfr3, Igf1r, Bmp6, Wnt5a and Ror2, which are the direct targets of GLI1 and/or GLI3, among which Fgfr3, Igf1r and Bmp6 are also targeted by SOX9 (S8 Table), indicating SOX9-GLI interacts with FGF, IGF, BMP and WNT signaling for regulating chondrocyte proliferation and differentiation. Although in the luciferase assays, SOX9 show higher transactivation potential on Ptch1 reporter comparing to GLI1 and GLI2 (Fig 5D), their transcriptional activities are comparable on other Hedgehog targets including Gli1 and Gli2 (Fig 5E and 5F). Since the regions used in these experiments only comprise a short representative fragment containing SOX9 and GLI binding regions, the level of transactivation achieved should not be taken as an absolute quantification of the degree of activation of the gene in vivo. Therefore we cannot conclude that SOX9 is a more potent factor for Ptch1 then GLI1 and GLI2. But we can confirm that Ptch1 is a common target of SOX9 and GLI.
Our analyses of the transcriptome and functional assays have implicated SOX9, GLI and FOXA as key regulators mediating differentiation transitions in the growth plate. Building on this finding, we went further to infer the wider interaction network mediated by these factors. By integrating data in GP-DGEL with SOX9, GLI1 and GLI3 ChIP-seq datasets, we found evidence for a regulatory network centered on SOX9-GLI-FOXA (Fig 8A). The GRN presents the progressive changes in expression of TFs as chondrocyte transition from proliferating to late differentiated states. Since SOX9 and GLI are highly expressed and their DNA binding regions are most enriched in the cluster of PZ genes (Fig 2A), we placed SOX and GLI families in the center of the network (Fig 8A). The common targets of SOX9 and GLI harboring at least one SGBR were placed in the inner circle of the network, while the genes that were targeted singly by SOX9 or GLI were arranged in the outer rim. Multiple binding regions for SOX9 and GLI factors were identified in the promoters and the intergenic regions of Gli1 and Gli2 and also the SOX trio of Sox9, Sox5 and Sox6 (S8 Table). The model therefore predicts the existence of regulatory feedback loops in which Gli1 and Gli2 are the targets of SOX9 and vice versa.
We next focused on the key transition from PHZ to UHZ and LHZ. The expression of SOX9 single target genes without GLI binding peaks in this outer rim, including transcription factors Mef2c, Rora, and Tcf3, showed little change in the transition from PHZ to UHZ. Exceptions included Irx3, Irx5, Cebpb, Ets-1, Atf6, and Egr1, which were not expressed in the PZ and PHZ but progressively increased in the UHZ and LHZ as SOX9 expression was down-regulated in the LHZ. These genes may be negatively regulated by SOX9 since SOX9 is both an activator and a repressor. SOX9 represses Col10a1 expression in PCs[40] and osteoblastic gene expression in HCs[26]. The coincident increasing gradient of expression of Foxa2 in the PHZ and UHZ, and the shared binding consensus of SOX9 and FOXA2, are consistent with the role of FOXA2 in activating the hypertrophic program[22]. Whether the same motif mediates transactivation and repressor activities is unclear. The SOX9 motif was also identified in the LHZ where SOX9 protein is absent, suggesting this motif may mediate the activity of the FOXA factors (FOXA2 and FOXA3), which are expressed in HCs and may compete with SOX9 for the SOX/FOXA motif[22].
SOX9 is known to auto-regulate itself through several long-range enhancers [82] and GLI factors can activate Sox9 through direct binding to a far upstream enhancer[38]. From ChIP-chip data, GLI1 input was found in the Sox9 promoter region (Fig 4F), suggesting that proximal chromatin interaction by GLI activator may also contribute to Sox9 expression. Our data indicate that in the context of growth plate cartilage, IHH signaling targets Sox9 as well as its transcriptional targets through GLI factors, and vice versa, to promote stage-specific chondrocyte differentiation, consistent with previous studies which found that the expression of Sox9 was induced by HH signaling articular cartilage chondrocytes[83] and retinal explants[84].
The cooperative action of SOX9 and GLI factors may reflect a wider application of cooperation between SOX and GLI factors in other systems. It is interesting that in neural tube patterning, cell fate determination requires both SOX2 and GLI1 inputs [85]. In pancreatic ductal adenocarcinoma, both SOX9 and GLI1 are important to maintain the malignant phenotype of cancer stem cells. Suppression of either SOX9 or GLI1 by siRNA reduces the expression of Sox9, Gli1 and Gli2 [86]. In primary chondrocytes, SOX9 up-regulates the expression of PthrP through direct binding to its promoter region via collaboration with GLI2 [87]. Even when Hedgehog signaling was blocked by cyclopamine, overexpression of Sox9 could still increase the expression of Ptch1, which is consistent with our finding that SOX9 positively regulates the transcription of Ptch1 without affecting Ihh expression. This report also demonstrated that SOX9 can directly bind to GLI2 in vitro[87], thus their direct interaction for cooperation in vivo warrants further investigation.
We propose a SOX9-GLI-centric model in which SOX9 and IHH signaling initiate control of chondrocyte differentiation phases, especially in the PZ and PHZ (Fig 8B). Upon transition to PHC, down-regulation of GLI repressor (due to Hedgehog signaling) allows higher levels of activator SOX9 together with RUNX2, and increased expression of MEF2C and FOXA2/3 which promote hypertrophic differentiation exemplified by Col10a1 expression[26]. We demonstrated that FOXA proteins compete with SOX9 for the binding to regulatory elements derived from Col2a1 and Col10a1. These data imply FOXA2 competition with SOX9 is important for the switch to the down-regulation of Col2a1 and activation of Col10a1 in HCs. We propose that in the presence of high SOX9 level, particularly in its highly stable homodimeric form, FOXA is precluded from accessing this regulatory element (Fig 7E). However, when SOX9 level declines and FOXA is elevated, FOXA2 competes off SOX9, accesses this binding site and activates Col10a1 expression. The data are consistent with a model where FOXA2 and FOXA3 de-repress Col10a1 by binding the regulatory regions bound by SOX9[22]. FOXA has been proposed to act as a pioneer factor to displace the linker histone and keep the enhancer accessible for specific TFs to activate gene expression in liver cells[88]. FOXA factors might initiate the hypertrophic program by out competing with repressive SOX factors to keep the chromatin accessible for other synergistic factors including RUNX2, MEF2C and SMAD1/4[22]. We surmise that SOX9 expression levels need to be lowered for this competition to take place as SOX9 forms a stable dimer at high concentrations that cannot be easily displaced. RUNX2 interaction with SOX9 depletes the effective level of SOX9 [89]. Furthermore, ZBTB20 and TRPS1, which are induced by SOX9, may repress SOX9 in the HZ in a negative-feedback loop [81, 90]. In view of the reported SOX9/AP1 cooperation in transactivating Col10a1 expression, it is interesting to note that AP1 binding sites were also present close to the SOX9/FOXA binding motifs in the Col10a1 enhancer[21]. This raises questions about whether FOXA2 could also cooperate with AP1 factors in promoting hypertrophy and if FOXA factors can compete to modulate the cooperative action of SOX9-AP1.
We have illustrated the powerful utility of integrating GP-DGEL with other databases as a discovery strategy to determine which biological processes and pathways, transcriptional regulators and their potential cooperating partners are active, in the growth plate. Our phasic GRN featuring SOX9, GLI and FOXA represents an initial template for the construction of a more complete model of chondrocyte differentiation that should incorporate a more complete set of TFs. In particular, it would be important to understand how SOX9-FOXA competition integrates with the SOX9-AP1 cooperation in promoting hypertrophy[21]. The protein interactome, non-coding RNAs, epigenetic status and chromatin/super-enhancer organization should also be taken into account in the future. It would be too great a task for a single study to investigate and functionally validate all the different target genes and processes identified. Therefore the gained information is shared as a public resource to facilitate and inspire additional discovery (S4 Fig). Mining and integration of the information in GP-DGEL with other emerging genome-wide data on the binding profiles of other transcription factors will be essential to extend our understanding of the complex and dynamic GRN mediating the transition steps in chondrocyte differentiation.
Transverse sections of the proximal tibia of 10-day-old female F1 (offspring of CBA and C57BL/6) mouse were obtained from chondrocyte sub-populations by cryosectioning. 5-micron sections were prepared and pooled into fractions consisting of 10 sections per fraction to ensure separation of each cell type in the growth plate. Samples were dissolved in Trizol reagent (Invitrogen) for RNA extraction. To guide the sub-division of chondrocyte populations into zones, every 10th section was examined histologically and 10% of RNA was isolated from selected sections at regular intervals for RT-PCR analyses of known growth plate markers (S1C Fig), to guide the sub-division of chondrocyte populations into zones. Sox9 heterozygous null (Sox9+/-) mutants were generated by crossing Sox9-flox mice (gift of Andreas Schdel)[10] with Protamine-Cre transgenic mice (gift of Yelena Marchuk) [91]. Gli2 null (Gli2-/-) mice were a gift from C.C Hui[92].
Total RNA was isolated according to the instructions for RNA isolation (Invitrogen). Prior to microarray analysis, 50ng total RNA was used to generate cDNA from each fraction by reverse transcription using SuperScript II reverse transcriptase and random hexamers. Semi-quantitative PCR analysis was performed to detect the expression of chondrogenic markers to identify the subpopulations of chondrocytes. Quantitative PCR was performed using Syber-Green master mixture to compare the expression levels of SOX9 target genes in the chondrocytes of wild type and Sox9+/- mice.
RNA quality and integrity were analyzed using RNA 6000 Nano Kit (Agilent). The pooled RNA was amplified for one cycle using MessageAmpTM II-Biotin Enhanced Kit, then labeled and hybridized to Mouse Genome 430 2.0 GeneChip containing 45101 probe sets (Affymetrix) in the Centre for Genomic Sciences (the University of Hong Kong). All primary microarray data are deposited in the GEO website (GSE99306). Gene expression data for each zone in triplicate were normalized by using RMA algorithm in BioConductor software package [93]. The k-Means Clustering algorithm [94, 95] and Eisen software tools [96] were used to identify the distinct expression patterns of genes with Coefficient of Standard Deviation (C.S.D.) > 0.15 across 4 zones. For each gene, the C.S.D. value was calculated with the formula:
C.S.D.=S/X¯, where S is the standard deviation and X¯ is the mean expression level of the samples over the 4 zones. The Gene Ontology analysis was performed for each cluster of genes by using the Gene Ontology database [97] and the David Web Tools [98]. To identify differentiation phase-specific genes and differential patterns of expression across different zones, we defined a set of “On” and “Off” genes in the dataset. Sox9 mRNA is expressed in the PZ and PHZ and is down-regulated in UHZ and LHZ. In contrast, Col10a1 is exclusively expressed in PHCs and HCs during hypertrophic differentiation in the PHZ, UHZ and LHZ. Therefore, we used the expression level of Sox9 in HZ (356, the average from UHZ and LHZ) and that of Col10a1 in PZ (511) to set the threshold of “On-Off” states for each zone (S2A Table).
The DME analysis were performed by using the Software package CREAD [99] with input from the promoter sequences extracted from 1k upstream and 200 bp downstream of TSS of the genes in each cluster. The background sequence file was generated by using the computer program ‘fasta-shuffle-notryptic.pl’ in the Bioinformatics CPAN Perl module of InSilicoSpectro-Databanks version 0.0.43. The Matcompare program in the CREAD package was used to compute the similarity between the identified DME motif and those in the TRANSFAC, JASPAR and ENCODE databases [100–103]. The Position Frequency Matrices and the consensus DNA binding sequences of the transcription factors were compiled from TRANSFAC database and the literature. Foreground (FG) represents the number of occurrences of the identified DNA motifs in the set of promoter DNA sequences. Background (BG) represents the number of occurrences of the motifs in the randomly generated DNA sequences. The ratio of FG/BG indicates the fold enrichment of the identified motifs in the foreground over the background set of sequences [104].
Position Weight Matrix identified in DME promoter analyses and the functional SOX9_COL2C1, COL2C2 [24], COL2C3 [65] binding consensus (S7 Table) were utilized as the seed motif for screening of SOX9 monomer binding sites in the genomic region within 10kb from TSS of the zone DEGs. The DNA sequences of 25-bp flanking the identified SOX9 monomer site on both sides were retrieved from Reference Genome (mm9) after removing DNA repeats. The MEME program was run with the command:
$meme monomer_site_flanking_sequence.fq -dna -mod anr -nmotifs 2 -w 30 -oc meme_out_30bp -pal
The parameter on motif length was set to range from 10 to 30-bp with the palindrome search mode activated.
The genomic sequences of evolutionarily conserved non-coding elements in the promoter and intergenic regions of each gene were retrieved from the Mouse Reference Genome Sequence of NCBI build 37, mm9. The conservation scores of DNA sequences for 30-vertebrates and the genomic coordinates of the non-coding elements were obtained from UCSC Genome Browser Database [105]. The algorithm [106] was implemented to match the Position Weight Matrix of the transcription factors with the genomic DNA sequences for screening of their binding elements. A match between the TF and the target sequence was accepted if the sequence similarity score was over 85% and the UCSC phastCons DNA conservation score was over 300. For prediction of SOX9 COL2C1, COL2C2 and COL2C3 binding elements, we searched for the exact matches of the binding motifs in the sequence of evolutionarily conserved non-coding DNA elements. The Position Weight Matrices used for identification of SOX9 dimer and GLI binding elements were constructed from the SOX9 binding HMG core sequence and the GLI binding consensus in TRANSFAC database respectively.
Detailed methods of rib chondrocyte isolation and SOX9 ChIP-seq experiment were described as reported [33]. The GLI1 promoter and GLI3 Genome-Wide ChIP-chip datasets were downloaded from GEO database (GSE11062 and GSE11063)[73] and converted from mm8 to mm9 assembly by using the UCSC Toolkit [107]. The SOX9 and GLI binding regions were identified by applying the procedure for local maxima finding [108] with 25- and 50-bp genomic windows respectively.
Fisher’s exact test on two-tailed P value was performed for a 2x2 contingency table with GraphPad Software (GraphPad Software, Inc.), where group 1 represents the PZ, PHZ and HZ DEGs in the clustering heatmap, and group 2 represents the DEGs containing SGBR. The Odds Ratio number was calculated with the formula,
OddsRatio=PZ,PHZSGBR/HZSGBRPZ,PHZtotal/HZtotal
where PZ,PHZSGBR is the number of SOX9/GLI common target in the PZ, PHZ gene sets, and PZ,PHZtotal is the total number of the PZ, PHZ genes in the heatmap. Fisher’s Exact Test was performed with: (i) varied inter-peak distances between SOX9 and GLI binding regions; (ii) varied genomic distance between SGBRs and the target gene TSS as in previous studies[109, 110]. The intergenic region was defined by the two nearest genes located at the 5’-end and 3’-end of the gene in query. The gene annotation information was downloaded from UCSC Genome Database.
In situ hybridization was performed as previously described[111]. Hind limbs dissected from 10-day-old F1 littermates were fixed in 4% paraformaldehyde overnight at 4°C and decalcified in 0.5M EDTA for 24h before embedding in paraffin. [35S]UTP labeled probes for the selected genes were hybridized on the paraffin sections of the hind limbs.
The paraffin sections were dewaxed and rehydrated. For cryosection, tissues were fixed in 4% PFA overnight before immersed in 30% sucrose. Sections were blocked with blocking buffer (5% BSA or goat serum, 0.5% Tween20) for 1 hour at room temperature. The primary antibodies of rabbit anti-Foxp2 (1:400; Abcam), rabbit anti-SOX9 (1:500, Millipore), guinea pig anti-SOX9 (1:2000, gift from V. Lee, STEMCELL Technologies) and rabbit anti-FOXA2 (1:500, Millipore) were diluted in blocking buffer and applied on the sections at 4°C overnight. The signal was visualized by using 1:500 goat-anti-rabbit or donkey-anti-guinea pig antibodies and mounting with Vectashield® mounting medium containing DAPI.
EMSAs were performed as described [66]. DNA probes were prepared with cy5-label at the 5’ end of the forward strand and reverse strand unlabeled. Equimolar amounts of complementary strands were annealed at 95°C for 5 min and subsequent cooling to 4°C at 1°C /min. Reaction mixtures (60nM probes and varying concentrations of proteins) were incubated at 4°C in the dark for 4h and electrophoresed at 200V for ~40min at 4°C in the dark. The gels were imaged with a Typhoon FLA-7000 PhosphorImager (FUJIFILM).
ATDC5 cells were grown in DMEM/F12 supplemented with 5% fetal bovine serum, human transferrin (10μg/ml) and sodium selenite (5ng/ml), and seeded in 12-well plates. With ~70%–80% confluency on the following day, the cells were transiently transfected with pGL3-basic luciferase reporters containing different regulatory elements using Lipofectamine 2000 (Invitrogen). Luciferase activity was measured using the Dual luciferase reporter assay kit (Promega) according to the manufacturer's instructions. Luciferase expression is given as a fold-change relative to the activity of Renilla luciferase.
The work with the use of mice and their care was approved in accordance with our institutional guidelines (Committee for the Use of Live Animals in Research, the University of Hong Kong).
|
10.1371/journal.pcbi.0030084 | Enhancer Responses to Similarly Distributed Antagonistic Gradients in Development | Formation of spatial gene expression patterns in development depends on transcriptional responses mediated by gene control regions, enhancers. Here, we explore possible responses of enhancers to overlapping gradients of antagonistic transcriptional regulators in the Drosophila embryo. Using quantitative models based on enhancer structure, we demonstrate how a pair of antagonistic transcription factor gradients with similar or even identical spatial distributions can lead to the formation of distinct gene expression domains along the embryo axes. The described mechanisms are sufficient to explain the formation of the anterior and the posterior knirps expression, the posterior hunchback expression domain, and the lateral stripes of rhomboid expression and of other ventral neurogenic ectodermal genes. The considered principles of interaction between antagonistic gradients at the enhancer level can also be applied to diverse developmental processes, such as domain specification in imaginal discs, or even eyespot pattern formation in the butterfly wing.
| The early development of the fruit fly embryo depends on an intricate but well-studied gene regulatory network. In fly eggs, maternally deposited gene products—morphogenes—form spatial concentration gradients. The graded distribution of the maternal morphogenes initiates a cascade of gene interactions leading to embryo development. Gradients of activators and repressors regulating common target genes may produce different outcomes depending on molecular mechanisms, mediating their function. Here, we describe quantitative mathematical models for the interplay between gradients of positive and negative transcriptional regulators—proteins, activating or repressing their target genes through binding the gene's regulatory DNA sequences. We predict possible spatial outcomes of the transcriptional antagonistic interactions in fly development and consider examples where the predicted cases may take place.
| With the availability of complete genome sequences and quantitative gene expression data, it becomes possible to explore the relationships between sequence features of regulatory DNAs and the transcriptional responses of their associated genes [1–7]. Developmental genes regulated by multiple enhancer regions and their spatio–temporal dynamics of expression are of particular interest [8–11]. The enhancers of developmental genes, such as gap and pair-rule genes, interpret maternally deposited information and participate in the formation of progressively more complex expression patterns, thus increasing the overall spatial complexity of the embryo. In part, the information required to generate these downstream patterns (e.g., gap and pair-rule) is present in the enhancer sequences.
Much attention has been paid to the investigation of transcription factor binding motifs and motif combinations, and to interpreting their role in the formation of spatial gene expression patterns. [5,12,13]. However, some early enhancers of Drosophila contain virtually identical sets of binding motifs, yet they produce distinct expression patterns [6,14]. It has been argued extensively that binding site quality (affinity) and site arrangement within enhancers (grammar) contributes to the levels and precision of enhancer responses [6,15–21]. In fact, some experimental studies of differentially arranged binding sites confirm the dependence of enhancer response on distances between binding sites and on binding site orientation [6,16,22–24], and some structural enhancer features such as motif spacing preferences and characteristic binding site linkages. “Composite elements” and other syntactical features were identified in many model organisms using computational analyses of binding site distributions throughout entire genomes [5,25,26]. Recent studies involving in vivo selection of optimal binding-site combinations in yeast also revealed a number of preferred motif combinations and structural features [27]. Nevertheless, some phylogenetic studies indicate significant flexibility in the regulatory code [28–31].
The analysis of unrelated, structurally divergent, but functionally similar enhancers aids in defining the balance between the stringency of the functional cis-regulatory “code” and its flexibility as demonstrated by changes in primary enhancer sequence over the course of evolution. [6,18,32]. Requirements for multiple cofactors that influence transcription via protein–protein interaction complicate computational predictions and studies of enhancers. While known binding motifs are easy to find, most protein–protein interactions leave no clear footprints in the DNA sequence of enhancers—some developmental coregulators such as CtBP (C-terminal binding protein) and Groucho influence the transcriptional response through interactions with sequence-specific transcription factors (e.g., [33]). Finally, regulatory signals from enhancers must be transmitted to the basal transcriptional machinery; this involves enhancer–promoter communication of some sort, as well as the recruitment of mediator complexes [2,21,34–36]. Both aspects further complicate the in silico prediction and analysis of enhancer activity.
Until recently, most models explaining enhancer responses in development were largely qualitative [37,38]. Davidson's group [2,39] and Hwa's group [21] undertook quantitative modeling of enhancer–promoter interactions and investigated the responses of architecturally complex regulatory units. The elaborate nature of developmental enhancers in Drosophila was described in quantitative models introduced by Reinitz's group [1,7]. Here, we summarize some basic structural considerations and investigate mechanisms of enhancer regulation to demonstrate how such features may affect the transcriptional responses. Our quantitative analyses involve models based on the fractional occupancy of transcription factor binding sites present within enhancers [2,21,40,41]. On the one hand, the described models are similar to those developed by Hwa's group [21] as they consider structural enhancer details. On the other hand, the models include biological assumptions for developmental enhancers (i.e., quenching), similar to those introduced by Reinitz's group [7]. Technically, our models use a homotypic array (a unit containing a number of identical sites) of binding sites as an elementary unit for modeling.
Based on quantitative analysis of transcriptional responses, we analyze some models for developmental pattern formation. In particular, we explore the outcome of the interplay between two antagonistic transcription factors, an activator and a repressor. We demonstrate that a pair of antagonistic gradients with similar or even identical spatial distributions is sufficient to initiate stripes of expression of a downstream gene. Given that the antagonistic gradients may be deposited by the same localized or terminal signal (e.g., in the fly embryo) [42] or by a focal signal (e.g., in the case of a butterfly eyespot) [43], the models explain how initiation from a single point in space can lead to efficient gains in spatial complexity.
The transcriptional state of enhancers of developmental genes is among major factors in developmental pattern formation [6–8,10]. If a transcription factor is present in a concentration gradient, the probability of that factor occupying a binding site in a target enhancer at a given position along the gradient depends on the factor's concentration at that position (coordinate). This logic suggests that in the case of activator and repressor gradients, calculating the probability of activator, but not repressor, binding (i.e., the successful transcriptional state resulting in transcription) may serve well to model the spatial expression patterns of the early developmental genes.
Let us consider an elementary enhancer, which contains two binding sites: one for an activator and one for a repressor. Let us assume that binding of the activator A in the absence of the repressor R brings the elementary two-site regulatory unit i (the enhancer; see Figure 1A) into a successful transcriptional state.
The equilibrium probability of the successful state pi depends on the binding probabilities of A (pA) and R (pR), which depend on the concentrations of the regulators ([A] and [R]) and on the binding constants (KA and KR) of the binding sites for the corresponding transcription factors (see Equations S1–S5 in Protocol S1):
Extending this formula to multiple different activators or repressors may be easily obtained with the same logic (see Equation S6 in Protocol S1). Bintu and coworkers recently introduced a number of similar models, describing DNA–protein and protein–protein interactions on proximal promoters [21], where the authors used an “effective dissociation constant,” which is the inverse of the binding constant (K) used in this study.
Developmental enhancers usually contain homotypic or heterotypic binding site arrays for multiple activators and repressors [44]. The probability of achieving a successful transcriptional state for the binding site array (enhancer) i, containing M identical, noninteracting activator sites and N identical, noninteracting repressor sites, is equal to (see Equation S7 in Protocol S1):
Here, Ψ is the sum of the statistical weights of molecular microstates for a homotypic site array and the denominator ΨAMΨRN is the sum of the statistical weights for all microstates of the system (i.e., the partition function; see Protocol S1, “Binding site arrays”).
In such site arrays, bound transcription factors may cooperate or compete for binding. Let us consider a cooperative array as an element of enhancer architecture (Figure 1B). Assuming presence of lateral diffusion [41,45], equal binding affinities for all sites in the array and expressing cooperativity C as the ratio between the second and the first binding constants, one can approximate the sum of statistical weights Ψ of all possible molecular microstates for a cooperative array as follows (see Equations S8 and S9 in Protocol S1):
Binding sites for an activator and a repressor may overlap, and the corresponding proteins compete for binding. Well-known examples in Drosophila development include Bicoid and Krüppel [46], Caudal and Hunchback [44], and Twist and Snail [6]. The classic example outside Drosophila is the competition between CI and Cro in the phage lambda switch [47]. The sum of microstates for a competitive site array, containing M overlapping A/R binding sites (Figure 1C; also see Figure S1 and Equations S8–S12 in Protocol S1), can be approximated by:
In addition to competitive interactions, this model also includes homotypic cooperative interactions between the regulators (see Equations S10–S12 in Protocol S1).
Structural elements within an enhancer (single sites or entire site arrays) may be distributed over extended genomic regions (thousands of bases, e.g., the Drosophila sna enhancer) [48,49]. In these cases, the distant regulatory elements within the enhancer may represent relatively independent units—modules [15,26] (see Figure 1D). Each independent module may include a single binding site or a binding site array. Redundancy of the enhancer elements (binding sites and modules) is a well-known biological phenomenon [44]. If the modules within an enhancer are independent from one another, bringing any one module into a successful transcriptional state may be sufficient for bringing the entire enhancer into a successful state, even if another module(s) is repressed.
Given the probabilities pi of successful states of all i independent modules or enhancers (Equations 1–4), the probability PEnc of the multimodule enhancer being in a successful state is equal to:
This is the reverse probability of the enhancer being in an inactive state, which is the product of the probabilities of each independent module being in an inactive state (1 − pi); Reinitz's group [1,7] implemented similar expressions for the quenching mechanism. While distinct modules may provide simultaneous responses to different inputs, multiple equivalent modules may allow for the boosting of an enhancer's overall response to a single input [50] (see Figure S1E and S1F).
In practice, however, the modules may not be completely independent from each other. Short-range repression and other factors (discussed below) may be involved in distance-dependent module responses [22–24,48]. Let us consider an enhancer containing two modules, a and b. Module a contains an activator site and a repressor site; module b contains an activator site only (see Figure 1E). Potentially successful enhancer states include all combinations in which at least one activator molecule is bound. However, the mixed state KaA[A]KaR[R] is always inactive as the repressor, and the activator sites in the module a are “close”. If module b is not “too far” from module a, short-range repression from a may reach the activator site in b. We can account for this possibility (and for its extent) by introducing a multiplier δ, depending on distance between the modules a and b (see also Equations S14–S16 in Protocol S1):
In this formula, Ψab is the sum of weights for all microstates, and Ψaboff is the sum of weights for the microstates that are always inactive (see Protocol S1, Equation S14). If modules a and b are “far,” δ = 1; if they are “close,” δ = 0. If the distance between a and b is somewhere in between, so that a repressor bound in a partially affects the activator bound in b, we could introduce a distance function δ = f(x) (0 ≤ δ ≤ 1), where δ depends on the distance x between a and b (and perhaps other variables, such as the repressor type). However, all we currently know about the distance function is that short-range repression is effective at distances less than 150–200 bases, and long-range repression may spread through entire gene loci (i.e., 10–15 kb [23,24,48]). Without exact knowledge about the distance function, the module concept (Equation 5) allows modeling of distance-dependent responses, but only in a binary close/far (yes/no) fashion.
Most of the enhancer response models (Equations 1–6) consider inputs from two antagonistic gradients, but enhancers may be under the control of a larger number of regulators (see Figure 1D). However, gradients of some of these regulators may either have similar spatial distributions (e.g., Dorsal and Twist) [51], or non-overlapping spatial expression domains (e.g., Krüppel and Giant) [37]. Therefore, in many cases the combination of all inputs may be parsed down to one or more pairs of antagonistic interactions. Based on the described quantitative models approximating enhancer responses (see above), we analyzed possible spatial solutions produced by gradients of two antagonistic regulators.
The examples in Figure 2A–2C demonstrate that the spectrum of possible enhancer responses is quite rich. One surprising result of these simulations is that even identically distributed antagonistic gradients can yield distinct spatial expression patterns such as stripes (Figure 2B). We identified conditions for the “stripe” solutions using differential analysis of the site occupancy function shown in Equation 1. For example, if both regulators are distributed as identical gradients and if their concentrations and binding constants are equal (KA = KR; [A] = [R]), then it is sufficient to identify conditions for the maximum of a site-occupancy function y(x) depending on the spatial coordinate x:
In this variant of Equation 1, k is the product of absolute concentration of the regulators [Abs] and the binding constant KA (k = KA[Abs]). The function f(x) is the distribution of the relative concentration (0 ≤ f(x) ≤ 1) of the transcription factors along the spatial coordinate x (i.e., the embryo axis). The function's maximum y′(x) = 0; x > 0 is f (x) = 1/k. In the Gaussian, logistic, and exponential decay forms of the function f(x) (see details in [52]), the maximum 1/k exists only if KA[Abs] > 1 (i.e., if binding constants and/or the absolute concentrations are high) (see also Figure S2). In the simple case (Equation 7), the absolute value of the fractional occupancy at the maximum is not very high (0.25); adding more sites or modules (see Figure S1) allows for the function's values to approach 1 (see Figure 2B).
However, if the antagonistic gradients are not identical (e.g., if the activator gradient is “wider” than the repressor gradient), the solutions for the stripe expression are more robust (Figure 2A). Shifting the peak of the activator gradient relative to the repressor gradient produces even more robust stripe patterns, as in the case of classical qualitative models [37], where a repressor “splits” or “carves out” the expression of a target gene (Figure 2C).
The formation of distinct gene expression domains (e.g., stripes) in response to similarly or even identically distributed gradients is of interest because this mechanism can lead to the very efficient gain of spatial complexity in just a single step: based on primary sequence, enhancers of target genes can translate two similarly distributed gradients into distinct gene expression domains or stripes. Such similarly distributed antagonistic gradients may come about by induction due to a single maternal gradient or due to a terminal (focal) signal emanating from a discrete point or embryo pole. The general pattern formation mechanism in the case described can be represented as follows: (1) maternal/terminal signal initiates two antagonistic gradients; and (2) interactions between the two gradients produce multiple stripe patterns. In an extreme case (e.g., Figure 2B), the described “antagonistic” mechanism could use only a single gradient/polar signal to produce multiple stripes of target gene expression.
The interaction between two antagonistic gradients is an example of a feed-forward loop. Due to a cascade organization of the developmental transcriptional networks, feed-forward loops are among the most common network elements (network motifs); a detailed analysis of the feed-forward networks and potential solutions can be found in a recent work by Ishihara et al [53].
To explore the interplay of antagonistic gradients in detail, we considered particular examples, such as the regulation of rhomboid (rho) by gradients of Twist and Snail and the regulation of knirps by the maternal gradients of Hunchback and Bicoid [54].
The enhancer associated with rho directs localized expression in ventral regions of the neurogenic ectoderm (vNEs) [51]. The rho vNE enhancer, as well as enhancers of other vNE genes such as ventral nervous system defective (vnd), is activated by the combination of Dorsal and Twist, but is repressed by Snail in the ventral mesoderm [13,51]. Both Twist and Snail are targets of the nuclear Dorsal gradient, which is established by the graded activation of the Toll receptor in response to maternal determinants [55]. The Twist and Snail expression patterns occupy presumptive mesodermal domains in the embryo, yielding slightly distinct protein distributions. Our recent quantitative analysis indicates that the boundaries of rho and vnd expression are defined largely by the interplay of the two antagonistic Twist and Snail gradients (see Figure 2D and 2F) [6], and the expression patterns of rho and vnd resemble the predicted solutions shown in Figure 2A. The patterning mechanism in this case can be represented as follows: (1) a terminal signal (Toll/Dorsal gradient) initiates two similar antagonistic gradients, Twi and Sna; and (2) Twi and Sna gradients produce multiple (distinct) stripe patterns (rho, other vNE genes).
Another example of the interplay between an activator and a repressor gradient is the early expression of the gap gene knirps in response to maternal gradients of Bicoid and Hunchback. Bicoid and Hunchback are deposited maternally and have similar, but distinct distributions—high in the anterior and low in more posterior regions of the embryo (see Figure 2E). The graded drop-off of the knirps repressor Hunchback at 50%–60% egg length is steeper than that of the knirps activator Bicoid. This is similar to the theoretical case shown in Figure 2C, where a narrow repressor “splits” a wider activator expression domain, thus producing two peaks of expression of the downstream gene. Known enhancer elements of knirps drive kni expression in the anterior and the posterior embryo domains and contain binding sites for Bicoid, Hunchback, Caudal, Tailless, and Giant [44,56–58]. However, tailless, caudal, and giant are downstream of Bicoid; it is likely that these and some other genes participate in the later maintenance of kni expression. It has been extensively argued that gap genes (and hunchback) stabilize their patterns along the anterior–posterior axis by mechanism of mutual repression [49]. At later stages (after cycle 14), the inputs from Bicoid and Hunchback into knirps regulation may stabilize fluctuations in knirps expression and fluctuations in the entire gap gene network due to mutual repression. Dynamic models from Reinitz's group based on slightly different logistic response functions support the sufficiency of Bicoid and Hunchback in the establishment of the early knirps expression [59].
To explore the role of Bicoid and Hunchback interplay in the early expression of knirps, quantitative expression data for Bicoid, Hunchback, and Knirps were downloaded from the FlyEx database [60], and models simulating the knirps enhancer response were generated based on Equations 1–4. One model assumed that Bicoid and Hunchback bind independently from each other; another model assumed that there is an interference (possibly competition) between the Bicoid and the Hunchback sites (Equation 7: competitive binding). Fitting the available quantitative data with the models (see parameter values in Table 1) shows that both models are sufficient to explain the posterior expression of knirps. However, the competitive model (Figure 2G) also predicts the anterior expression of knirps. This result was especially striking, as the anterior knirps expression data were not included in some of the fitting tests. Bicoid and Hunchback motifs are quite different, so it is unlikely that this is a case of direct competition for overlapping binding sites. Other mechanisms may account for the negative interaction between the two regulators; for instance, binding of Bicoid may prevent Hunchback dimerization [61] and/or efficient binding.
Shifting the knirps expression data by more than 5% along the anterior–posterior axis (see Materials and Methods) results in reduction of the data-to-model fit quality for the posterior kni expression domain (see Table 1 for exact parameter values). The robustness of knirps regulation was emphasized earlier [59,62], and the present analysis using site occupancy confirms that the interplay of the two antagonistic gradients, Bicoid and Hunchback, is sufficient to explain the initial formation of both the anterior and the posterior strips of knirps expression.
To test the models describing gene response to antagonistic gradients, we introduced mutations in the rho enhancer and compared the expression patterns produced by the reporter gene in vivo with the simulated expression patterns simulated in silico (Equations 1–6). Specifically, the models for rho and vnd expression predicted the following [6]: (1) The position of the dorsal expression border of rho is highly sensitive to Twist and/or its cooperativity with Dorsal. Reducing Dl–Twi cooperativity or Twist–Twist cooperativity shifts the dorsal border ventrally. (2) The number of independent elements (groups of closely spaced Dorsal-Twist-Snail sites, or “DTS” elements) contributes to the expression pattern of rho and vnd according to Equation 5 (boost): a higher number of DTS elements in vnd is responsible for the shift of the ventral vnd expression border relatively to rho [6].
These two specific predictions, based on the model analysis and simulations, were tested by modifying the structure of the minimal rho enhancer. First, the distance between the Dorsal and the Twist sites in the DTS element was increased (see Figure 3). The increased distance between the two sites reduced the cooperative potential between the Dorsal and Twist sites. Indeed, the observed effect in vivo is consistent with the effect of the same mutation simulated in silico, causing a ventral shift of the dorsal border of the reporter gene expression (compare Figure 3E with 3A). An additional mutation eliminating the weaker Twist site from the DTS element affects Twist–Twist cooperativity in the enhancer and shifts the dorsal rho–lacZ expression border. In fact, the combined effect produced by these two mutations in vivo (Figure 3G; compare with 3C) and the deletion of the weak Twist site alone (Figure 3F; compare with 3B) demonstrate shifts of the dorsal expression border of the rho-lacZ transgene in concordance with the models. Last, a second DTS module was introduced into the rho enhancer in the context of the previous two mutations. The predicted in silico effect is a “boost” in expression, resulting in the shift of both ventral and dorsal expression borders. Again, the predicted changes in the expression pattern were observed in vivo—not only were the positions of the ventral and the dorsal border shifted (Figure 3H; compare with 3D), but the overall level of expression of this transgenic construct appears higher (unpublished data).
The described in vivo tests of the in silico predictions using site-directed mutagenesis of the rho enhancer have demonstrated that though the quantitative models based on fractional site occupancy are approximations, they can produce reasonable predictions for the response of complex regulatory units (such as fly enhancers) to gradients of transcriptional regulators.
Using transcriptional response models and quantitative expression data, we demonstrated how two similar terminal gradients can determine stripes of expression of downstream genes. Related examples are quite frequent in development. For instance, the posterior stripe of hunchback is the result of activation by Tailless and repression by Huckebein [63,64]. As in the case with Twist and Snail, the posterior gradient of Tailless is slightly broader than the gradient of Huckebein. Therefore, the mechanisms of posterior hunchback expression may be similar to the mechanisms shown in Figure 2A, 2B, 2D, and 2F. However, while the examples above involve direct transcriptional regulation in the embryonic syncytial blastoderm, extracellular morphogen gradients may produce similar outcomes if the cellular response is transcriptional in nature.
Formation of eyespot patterns in butterfly wings is an elegant example of axial (here focal) patterning in a cellular environment (see Figure 4A). The interplay between Notch and Distalless specifies the position of focal spots and intervein midline patterns in the butterfly wing [65]. Subsequent Hedgehog signaling from the focal spots is believed to induce the formation of concentric rings of gene expression and the pigmentation of the eyespots in the adult butterfly wing [66]. Known targets of the Hedgehog gradient are the butterfly homologs of engrailed and spalt [67]. Initially, both genes are expressed around the focal spot, but at later stages an external ring of engrailed expression appears around the spalt expression pattern (see Figure 4B and 4C). In the case of engrailed pattern formation, a simplified mechanism [67] may include elements of the following feed-forward network: (1) focal signal (focal spot/Hedgehog signaling) initiates two antagonistic gradients, the activator Engrailed and the repressor Spalt; and (2) subsequent interactions between Engrailed and Spalt produce multiple ring patterns. An extension of the model in Equation 1, (k is the rate of synthesis and c is the rate of decay; d[R]/dt = 0) reproduces the dynamic changes in the engrailed pattern (Figure 4A, 4D–4E):
Examples of axial or focal patterning using a single source of signaling or a combination of similar antagonistic gradients are common. The interplay between maternal hunchback and maternal nanos during development of the short germ-band insect Schistocerca is an example of axial patterning similar to the interplay between Bicoid and Hunchback [68]. Specification of segments during insect limb development is comparable to the mechanisms of Twist/Snail interplay and the butterfly eyespot formation [69]. Nature uses many combinations of signals and gradients in pattern formation, but the most effective mechanism/combination may be one that allows maximal informational gain in a minimal number of steps. From this perspective, the interplay between similar or identical gradients is of significant interest.
Quantitative distribution data for Dorsal, Twist, and Snail were published previously [6]. Quantitative expression data for mRNA levels of mutated rho enhancers were generated by in situ hybridization (the data are available at the DVEx database: http://www.dvex.org). Multiplex in situ hybridization probes were used for colocalization studies, including co-stainings for the endogenous mRNAs and lacZ reporter gene expression as described previously, and confocal microscopy and image acquisition were performed as described [6]. In short, signal intensity profiles of sum projections along the dorso–ventral axis of mid-nuclear cleavage cycle of 14 embryos were acquired using the ImageJ analysis tool (National Institutes of Health, http://rsb.info.nih.gov/ij). Background signals were approximated by parabolic functions and subtracted according to existing methods [70]. Online programs for the automated background subtraction and data alignment are available from the University of California Berkeley Web resource (http://webfiles.berkeley.edu/∼dap5). After background subtraction, the data were resampled and aligned according to the position of Snail gradient and the distribution of endogenous rho message. Expression datasets for anterior–posterior genes were downloaded from the FlyEx database (with options: integrated, without background) [60]. In all cases, signal amplitude was normalized to the 0–1 range, and the data was resampled to 1,000 datapoints along the coordinate of the corresponding axis. In all models, we used the relative concentration multiplied by a maximal absolute concentration. This absolute concentration is an independent unknown parameter (range, 10−8–10−9 M) equal for all reaction components.
The minimal rho enhancer [6] was mutated via site-directed mutagenesis in pGem T-Easy (Promega; http://www.promega.com) using the following primers: Dl-Twi distance, RZ65mut: 5′-GTTGAGCACATGTTTACCCCGATTGGGGAAATTCCCGG-3′; deletion of Twist site, RZ66mut: 5′-GGCACTCGCATAGATTGAGCACATG-3′; creation of a second DTS, RZ67mut: 5′-GCAACTTGCGGAAGGGAAATCCCGCTGCAACAAAAAG-3′; and RZ68mut: 5′-CACACATCGCGACACATGTGGCGCAACTTGC-3′.
Mutated enhancers were cloned into the insulated P-element injection vector E2G as described previously [13]: constructs were introduced into the D. melanogaster germline by microinjection as described previously [71]. Between three and six independent transgenic lines were obtained and tested for each construct; results were consistent across lines.
To fit our models with actual quantitative data, we maximized the agreement r (Pearson association coefficient) between the model output predictions and the observed (measured) expression patterns:
The best set of parameters X* from the parameter space I is defined by the binding constants, cooperativity values, and the number of binding sites. We used a standard hill-climbing algorithm (full neighborhood search) for the main parameter space (e.g., [72]). For each identified maximum, we measured the value of the site occupancy function and discarded maxima that produce site saturation values below selected thresholds, as well as such that are located beyond selected realistic parameter ranges for binding constants and cooperativity values. All maxima producing the highest data-to-model agreement were found multiple times, suggesting that exhaustive mapping of the parameter space was achieved. Fitting “shifted data” (wrong data) for Knirps was performed by exploring exactly the same parameter space and exactly the same number of seed points for each shift value. Quantitative gene expression data for dorso–ventral genes are available at http://www.dvex.org; the analysis tool “E-response,” fitting utilities, and online data-treatment programs are available at the University of California Berkeley Web resource http://webfiles.berkeley.edu/∼dap5.
|
10.1371/journal.ppat.1003811 | The Malarial Serine Protease SUB1 Plays an Essential Role in Parasite Liver Stage Development | Transmission of the malaria parasite to its vertebrate host involves an obligatory exoerythrocytic stage in which extensive asexual replication of the parasite takes place in infected hepatocytes. The resulting liver schizont undergoes segmentation to produce thousands of daughter merozoites. These are released to initiate the blood stage life cycle, which causes all the pathology associated with the disease. Whilst elements of liver stage merozoite biology are similar to those in the much better-studied blood stage merozoites, little is known of the molecular players involved in liver stage merozoite production. To facilitate the study of liver stage biology we developed a strategy for the rapid production of complex conditional alleles by recombinase mediated engineering in Escherichia coli, which we used in combination with existing Plasmodium berghei deleter lines expressing Flp recombinase to study subtilisin-like protease 1 (SUB1), a conserved Plasmodium serine protease previously implicated in blood stage merozoite maturation and egress. We demonstrate that SUB1 is not required for the early stages of intrahepatic growth, but is essential for complete development of the liver stage schizont and for production of hepatic merozoites. Our results indicate that inhibitors of SUB1 could be used in prophylactic approaches to control or block the clinically silent pre-erythrocytic stage of the malaria parasite life cycle.
| Malaria is caused by a single-celled parasite and is transmitted by the bite of an infected mosquito. The inoculated sporozoite forms of the parasite invade liver cells where they replicate, eventually releasing thousands of merozoites into the bloodstream to initiate the blood stage parasite life cycle which causes clinical malaria. The liver stage of the parasite life cycle is asymptomatic, so it is widely considered a potential target for prophylactic vaccine- or drug-based approaches designed to prevent infection. In this study, we use a robust, highly efficient gene engineering approach called recombineering, combined with a conditional gene deletion strategy to examine the function in liver stages of a parasite protease called SUB1, previously implicated in release of blood stage parasites. We show that SUB1 is expressed in the liver stage schizont and that the protease is essential for production of liver stage merozoites. Our results enhance our understanding of malarial liver stage biology, provide new tools for studying essential gene function in malaria, and suggest that inhibitors of SUB1 could be used as prophylactic drugs to prevent clinical malaria.
| Transmission of the malaria parasite to a vertebrate host is initiated by the bite of an infected Anopheline mosquito. The inoculated sporozoites migrate from the site of inoculation, enter the circulation, and are arrested in liver sinusoids where they traverse the vascular endothelium and invade hepatocytes, coming to rest within an intracellular membrane-bound parasitophorous vacuole (PV) [1], [2]. After an initial period of non-replicative development, which lasts around 24 h in the rodent malaria species Plasmodium berghei, the intracellular parasite - now known as an exoerythrocytic form (EEF) - initiates an asexual replicative program. This begins with several rounds of nuclear division to form a multinucleated syncytium or schizont, concomitant with a large increase in the size of the PV to accommodate the growing parasite. Approximately 55 h following hepatocyte invasion (in hepatoma cells) the single plasma membrane of the schizont begins to invaginate around groups of parasite nuclei to form the so-called cytomere stage [3], [4]. Subsequent further invagination of the parasite plasma membrane produces clearly defined individual merozoites tightly packed within the PV. Shortly thereafter, the PV membrane (PVM) disintegrates, releasing the merozoites to move freely within the host cell cytoplasm. PVM rupture triggers an unusual form of cell death in the host cell, involving DNA condensation, disintegration of host cell mitochondria and loss of plasma membrane proteins, but lacking certain other classical features of apoptosis such as caspase activation and loss of host plasma membrane phospholipid asymmetry [4], [5], [6]. In vitro, infected hepatoma cells such as HepG2 cells round up at this point and detach from their monolayers to float freely in the cultures [4], [5]. Just prior to detachment, merozoite-filled vesicles called merosomes, each surrounded by membrane of host cell origin, are extruded from the host cells. In vivo, these enter the lumen of the liver sinusoids from where they are carried to the pulmonary microvasculature to rupture, allowing egress of their merozoite cargo [4], [5], [6]. The merozoites invade erythrocytes to initiate the asexual blood stage cycle. The entire liver stage has a duration of between 2 and 15 days [7], [8], depending on the Plasmodium species, and culminates in the production and release of thousands of hepatic merozoites from each infected hepatocyte. Whilst not itself associated with any pathology, the liver stage and other pre-erythrocytic stages are a prerequisite to the asexual blood-stage cycle in a natural malarial infection, and so are potential targets for prophylactic immune-based or chemotherapeutic interventions designed to prevent disease.
Compared to Plasmodium asexual blood stages, liver stage malaria parasites are relatively difficult to access [7], [8] and so, despite these elegant and detailed morphological descriptions of the hepatic malaria life cycle, little is known of the signals and molecular players involved in liver stage merozoite development, PVM rupture, merosome formation and merozoite egress. The limited available data suggest that in many respects liver stage merozoites are probably very similar in makeup to their well-studied blood stage counterparts [9]. Elements of merozoite morphogenesis and egress are therefore likely shared between the liver and blood stages. As an example of this, treatment of mature hepatic or erythrocytic schizonts with the cysteine protease inhibitor E64 prevents PVM rupture [5], [10], [11], implicating a common role for cysteine protease(s) in merozoite release. The effects of E64 may result from inhibition of host cell calpain-1 activity, which has been implicated in egress [12], as well as of host cell cysteine proteases implicated in the parasite-induced cell death [13]. Alternatively or in addition, the target(s) of E64 may include members of the parasite serine repeat antigen (SERA) family, which are expressed in mature stages of blood schizonts [14], [15], [16], [17]. SERA proteins may play a role in egress [18], [19], and some of them have E64-sensitive cysteine protease activity [15]. In blood stages some or most SERA proteins are substrates of a conserved Plasmodium subtilisin-like serine protease called SUB1 that is discharged from specialised secretory organelles called exonemes into the PV lumen minutes before egress [20], [21], [22], [23]. SUB1 cleaves the SERA proteins to release their central papain-like domain [15], [20]. SUB1-mediated cleavage of P. berghei SERA3 (PbSERA3) has been shown to activate its protease activity [15], suggesting that one important role of SUB1 may be to initiate a protease cascade that leads to egress. Discharge of SUB1 into the PV also allows it to modify several other important merozoite proteins, including the major glycolipid-anchored merozoite surface protein MSP1 [24], [25], which is thought to act as an erythrocyte binding ligand [26], [27], [28]. SUB1 therefore likely plays a central role in both development and egress of blood stage schizonts. Intriguingly, whereas both MSP1 and members of the SERA protein family are expressed in liver stage schizonts [9], [11], it is not known whether SUB1 is expressed in liver stages, or whether it has a similarly important role in maturation and release of liver stage merozoites.
The study of genes in liver stages that are essential during the asexual erythrocytic phase of the life cycle requires an inducible or stage specific system for gene disruption. The site specific recombinase Flp is currently the only validated system [29] to knock out or knock down essential genes in liver stages. A panel of highly efficient deleter lines expressing a thermosensitive variant of the recombinase, FlpL, under different sporozoite specific promoters is now available and these have been used to characterise essential malarial gene functions in liver stages, including that of MSP1 [30] and the parasite cyclic GMP (cGMP)-dependent protein kinase, PKG [31]. To control a gene through Flp or FlpL it is necessary to introduce two 34 bp flippase recognition target (FRT) sites into the genome such that they flank a crucial part of the target gene, which becomes excised when Flp is expressed, thereby inactivating the gene of interest.
Placing FRT sites in a genetic modification vector remains a major challenge. To achieve a complete gene knock out upon activation of Flp, it would be desirable to flank the entire target gene with FRT sites. This requires large allelic exchange vectors with at least one very long homology arm comprising the target gene plus an additional 1 kb or more of upstream homologous sequence to achieve genomic integration of the FRT site most distant to the selection cassette. Constructing such large vectors in E. coli can be difficult, or even impossible, due to the high (>77%) AT content and repetitive nature of genomic DNA (gDNA) of most Plasmodium species, which causes instability in circular high-copy plasmids. Recently a P. berghei gDNA library with high sequence integrity, relatively large inserts (averaging ∼9.0 kb) and covering now >85% of genes was generated in a linear, low copy plasmid in E. coli [32]. Here we present molecular tools and protocols that exploit the efficiency, speed and robustness of recombinase mediated engineering in E. coli to convert a gDNA library clone with a 9.6 kb genomic insert containing the P. berghei sub1 (pbsub1) gene into a complex allelic exchange vector for the Flp mediated deletion of the entire pbsub1 gene. We generate a conditional knock out parasite in which we examine the expression and function of SUB1 in liver stages of P. berghei. We show that SUB1 is expressed in subcellular organelles of the liver stage schizont that resemble exonemes, and that expression of the protease is indispensable for completion of the liver stage of the parasite life cycle. Unexpectedly, parasites lacking SUB1 exhibit a defect in development prior to egress, indicating a hitherto unappreciated role for SUB1 in intracellular parasite growth.
A previous transcriptomic and proteomic analysis of the rodent malaria species P. yoelii indicated the presence of P. yoelii sub1 mRNA in liver stages but detected no SUB1 protein by mass spectrometry [33]. However, mass spectrometric analysis of blood-stage schizonts of the human malaria pathogen P. falciparum has detected only between 1 and 8 peptides [34], [35], [36], suggesting that – as with many enzymes - SUB1 is likely a poorly abundant constituent of the total proteome. To address the question of whether SUB1 is expressed in P. berghei liver stages, we produced a rabbit polyclonal antibody specific for the catalytic domain of P. berghei SUB1 (PbSUB1). Examination of P. berghei blood stage schizont extracts by Western blot using the antibody produced signals likely corresponding to the full-length and processed (mature) forms of PbSUB1 (Supplemental Figure S1A in Text S1), by analogy with the maturation profile previously observed with recombinant forms of SUB1 from P. berghei and three other Plasmodium species [37], [38], [39]. The anti-PbSUB1 antibodies were then used to examine P. berghei liver stage EEFs by immunofluorescence assay (IFA). Hepatoma cells infected with sporozoites of a drug selectable marker-free transgenic P. berghei clone that constitutively expresses GFP [40] were fixed 64 h post infection and probed with the antibodies. As shown in Figure S1B in Text S1, a punctate signal was obtained that is highly reminiscent of the exoneme-specific pattern previously observed in P. falciparum blood stage schizonts probed with polyclonal or monoclonal antibodies (mAb) against P. falciparum SUB1 [20], [22]. This suggested that PbSUB1 may be localised in similar subcellular organelles in liver schizonts.
To further test this interpretation, we generated a P. berghei line expressing epitope-tagged PbSUB1 (called PbSUB1-HA) using homologous recombination to modify the endogenous pbsub1 gene (PBANKA_110710) in the GFP-expressing parasite background (Figure S2 in Text S1). Western blot analysis of blood stage schizont extracts from PbSUB1-HA parasites using an anti-HA mAb detected a strong double band migrating slightly more slowly than that detected by the anti-PbSUB1 rabbit antibodies (Figure S3A in Text S1), consistent with the expected small increase in mass of the epitope-tagged PbSUB1 as a result of its fusion to the HA epitope tag. In contrast, Western blot analysis of PbSUB1-HA salivary gland sporozoite extracts with the anti-HA antibody, or with the anti-PbSUB1 antisera, detected no specific signal (not shown). IFA analysis of mature blood stage (Figure S3B in Text S1) or mature liver stage (Figure 1 and Figure S4 in Text S1) PbSUB1-HA schizonts with the anti-HA mAb again produced a clear punctate signal. The foci were associated with but distinct from individual merozoite nuclei, and again similar to the exoneme-specific signal previously observed in P. falciparum blood stage schizonts. No PbSUB1-HA IFA signal was detected in salivary gland sporozoites or early liver stage schizonts (Figure S3C in Text S1, Figure 1 top row and Figure S4 in Text S1) or at the cytomere stage when the parasite plasma membrane is just beginning to invaginate to surround groups of nuclei (Figure S5 in Text S1). Collectively, these results convincingly demonstrate that PbSUB1 is not expressed in sporozoites or early EEFs, but is expressed in mature liver stage schizonts, where it likely accumulates in subcellular organelles similar to the exonemes previously described in blood stage schizonts.
In our previous work on P. falciparum SUB1 [20] we were unable to obtain viable parasites in which the pfsub1 gene was disrupted, suggesting an essential role in blood stages. Attempts to disrupt the pbsub1 gene in P. berghei blood stages similarly failed (S. Yeoh, R. Tewari, O. Billker and M. Blackman, unpublished), suggesting that PbSUB1 too is indispensable in the erythrocytic parasite life cycle. To study the role of the pbsub1 gene in liver stages we therefore decided to exploit a recently-described conditional deletion approach [30], [41] in which stage-specific expression of the Saccharomyces cerevisiae site-specific recombinase Flp (or its thermosensitive variant FlpL) is employed to disrupt a target gene in the late mosquito stages of the parasite life cycle. Working from a P. berghei genomic DNA library clone from the PlasmoGEM resource (http://plasmogem.sanger.ac.uk/) containing 9.6 kb of the pbsub1 locus and neighbouring genes, we constructed allelic exchange vectors designed to flank the entire pbsub1 coding sequence with FRT sites, while at the same time inserting a C-terminal HA epitope tag into pbsub1 (Figure 2A). The 5′ FRT site was introduced into one of two alternative positions in the large upstream intergenic region together with a constitutive promoter sequence of the P. berghei hsp70 gene. A promoterless GFP coding sequence was positioned immediately downstream of the second FRT site (Figure 2A and Figure S6 in Text S1). Precise placement of the FRT sites and other exogenous sequence both upstream and downstream of pbsub1 was achieved in 4 steps by recombinase mediated genetic engineering in E. coli, using both the improved Red/ET recombinase system of lambda phage [42] and transient expression of Flp as described in Supplemental Methods and Figure S6 in Text S1. The constructs were designed such that, following integration by ends-out homologous recombination into the parasite genome, correct recombinase-mediated excision of the sequence lying between the FRT sites (which included the epitope-tagged pbsub1 gene) would reposition the GFP reporter adjacent to the hsp70 promoter, driving constitutive GFP expression only in those parasites in which deletion of the pbsub1 gene had occurred (Figure 2A and Figure S6 in Text S1). To prevent the FRT site from interfering with initiation of translation [30], the start codon for GFP expression was placed just upstream of the 5′ FRT site, such that after excision the FRT sequence would be translated into a 12 residue N-terminal extension of GFP. The final constructs, called pJazz-FRTed-pbsub1 and pJazz-FRTed-pbsub1short (which differed only in the placement of the 5′ FRT site at either ∼2.3 kb or ∼1.8 kb respectively upstream of the start ATG of the pbsub1 gene) contained ∼7 kb and 700 bp regions of homology respectively at their 5′ and 3′ ends for homologous integration into the P. berghei genome.
The constructs were transfected into the P. berghei UIS4/FlpL deleter clone [30] which expresses FlpL under control of the sporozoite stage-specific uis4 promoter. Transfected parasites were expanded under pyrimethamine treatment and cloned by limiting dilution to obtain 2 independent parasite clones (named condSUB1 clone A and B) transfected with the pJazz-FRTed-pbsub1 construct, and a single parasite clone transfected with the pJazz-FRTed-pbsub1short construct (called condSUB1short). The expected homologous integration event in each of the parasite clones was confirmed by diagnostic PCR, Southern blot and pulse-field gel analysis (Figure S6 in Text S1).
The condSUB1short and condSUB1 parasite clones did not express GFP, as expected, and exhibited no growth phenotype in blood stages (data not shown), indicating that the modifications to the pbsub1 locus resulting from integration of the targeting constructs did not affect parasite viability. To initially assess stage-specific deletion of the modified pbsub1 gene in the condSUB1 clones, Anopheles stephensi mosquitoes fed on mice infected with the condSUB1 clone A parasites were subjected to a temperature shift to 25°C at 18 days following transmission in order to enhance activity of FlpL. At day 26 post-transmission dissected midguts, salivary glands and salivary gland sporozoites were examined microscopically. GFP expression was observed in 65±19% of the oocysts as well as in the majority of sporozoites recovered from the condSUB1 clone A-infected insects, consistent with the expected FlpL-mediated excision event (Figure 2B). There was some variation between individual experiments (n = 4), but visual microscopic examination of the isolated day 26 condSUB1 sporozoites showed that the proportion of GFP-positive sporozoites was usually ∼90% (though this varied somewhat in subsequent experiments; see below), similar to previous findings of others using the Flp/FRT system in P. berghei [30], [41], [43]. Analysis by genotyping PCR of sporozoites collected at day 18 and day 26 following transmission (Figure 2C) confirmed efficient, time-dependent excision of the flirted pbsub1 gene, with undetectable levels of the non-excised pbsub1 locus in the day 26 sporozoites. To examine the behaviour of the pbsub1-deficient parasites throughout their lifecycle, condSUB1-infected mosquitoes were allowed to feed on naïve mice at 26 days following transmission (‘bite-back’ infection), by which point most of the sporozoites observed in the insect salivary glands were expressing GFP. The bite-back mice were monitored for the appearance of blood-stage parasites, which were then recovered and analysed by genotyping PCR. As shown in Figure 2C, despite the predominance of the excision event in the previous mosquito stages, only the non-excised pbsub1 locus was detectable in the erythrocytic parasites that appeared in the bite-back mice (observations from n = 7 independent experiments, each using 3–5 mice). These parasites did not express GFP (not shown). Identical results were obtained in an independent experiment with the condSUB1 clone B as well as the condSUB1short parasites (Figure S7 in Text S1). These results strongly suggested that those sporozoites in which the pbsub1 gene had been deleted were incapable of successfully establishing a blood stage infection.
To gain more insight into the nature of the defect resulting from pbsub1 deletion, we next examined whether the inability of PbSUB1-deficient sporozoites to establish a blood stage infection was a result of compromised hepatocyte invasion, although we considered this unlikely given our previous evidence that PbSUB1 is not expressed in sporozoites. To investigate this, we incubated hepatoma cells in vitro with 26 day condSUB1 clone A or clone B sporozoites and assessed their capacity to invade the cells. For this we used a modified differential staining assay [44] that distinguishes extracellular (i.e. residual cell surface-bound) sporozoites from intracellular parasites, combined with automated microscopy and high content image analysis software. Extracellular sporozoites were detected with an antibody against the circumsporozoite protein (CSP) on the parasite surface, whilst an antibody against GFP was used to detect all excised condSUB1 sporozoites. As a control for these assays we used sporozoites of the P. berghei UIS4/FlpL-F clone [41], which constitutively expresses GFP under the same hsp70 promoter as used in the condSUB1 and condSUB1short clones. These sporozoites were also obtained from insects that had been placed at 25°C at 18 days following transmission. As shown in Figure 3A, no significant differences were observed in the proportions of GFP-positive infected host cells detectable 2 h following addition of control or condSUB1 sporozoites, indicating equivalent invasive capacity. This observation was confirmed under in vivo conditions by using qRT-PCR to quantify parasite liver loads following infection of mice with condSUB1 or control sporozoites. As shown in Figure 3B, there was no significant difference in parasite liver loads measured 40 h after intravenous inoculation of 20,000 condSUB1 or UIS4/FlpL-F sporozoites. These results showed that condSUB1 sporozoites are fully competent to initiate and establish a liver infection.
Having determined that pbsub1-null sporozoites display no invasion phenotype in vitro or in vivo and are able to efficiently initiate intrahepatic growth, we next addressed whether the transmission defect observed in the excised condSUB1 parasites was due to a defect in subsequent liver-stage replication. Development of hepatic EEFs comprises a well-described set of morphological transitions, in which an early schizont gradually increases in size and passes through cytomere and merozoite formation stages before rupture of the PVM to allow release of the mature merozoites into the host cell cytosol. To initially assess expression of epitope-tagged PbSUB1 in the condSUB1 parasites and to attempt to confirm loss of PbSUB1 expression upon excision, we used IFA to compare the GFP-positive (excised) and GFP-negative (non-excised) EEFs obtained following infection of hepatoma cell cultures with condSUB1 clone A or B sporozoites, using antibodies against GFP and the HA epitope tag fused to PbSUB1. Whilst mature GFP-negative (non-excised) condSUB1 liver stage schizonts displayed the expected punctate IFA signal, as observed previously with the PbSUB1-HA clone, we were able to detect only immature forms of GFP-positive (excised) condSUB1 EEFs (Figure 4), with no mature forms visible. This unexpected result was explained by subsequent findings described below.
To investigate the intrahepatic development of PbSUB1-deficient parasites, the size of GFP-positive condSUB1 EEFs in infected hepatoma cell cultures was analysed at 28 h and 48 h post infection, comparing them with non-excised condSUB1 and control UIS4/FlpL-F EEFs using automated microscopy and image analysis software. Parasite identification in this case was achieved using antibodies to the P. berghei HSP70 heat-shock protein, whilst GFP expression was used to discriminate excised PbSUB1-deficient (GFP-positive) condSUB1 parasites from the minority of non-excised condSUB1 parasites. No significant differences in EEF dimensions were seen at the 28 h time point (data not shown). However, as shown in Figure 5, a small but significant reduction in the mean surface area (28±3%) of the PbSUB1-deficient parasites was observed at 48 h post infection compared to both controls, indicating a subtle defect in schizont development.
For a more detailed analysis of this phenotype, we investigated the appearance throughout EEF maturation of three parasite marker proteins with distinct subcellular locations. Starting from 24 h post invasion, infected hepatoma cells were examined by IFA using antibodies against the PVM protein EXP1, the soluble PV protein PbSERA3 (a late liver stage marker expressed from cytomere stage onwards that is eventually released into the host cell cytosol; [11]), and the plasma membrane protein MSP1 (present from cytomere stage onwards and involved in the formation of hepatic merozoites [30]). At time points up to and including cytomere stage, expression and localisation of EXP1 and PbSERA3 was normal in the PbSUB1-deficient parasites (Figure S8 in Text S1). However, at very late time points (from around 64 h onwards), whilst normal rupture of the PVM and associated release of PbSERA3 into the host cell cytoplasm was evident in the majority of the control infected cells, it was only very rarely detected in the PbSUB1-deficient parasites (Figure 6). Equally strikingly, whereas control parasites displayed as expected a clear MSP1 signal from cytomere stage onwards, which subsequently translocated to surround individual merozoites, the majority of the PbSUB1-deficient parasites completely lacked a detectable MSP1 signal (Figure 6 lower panels and Figure 7) and showed no signs of correct merozoite formation. Multiple nuclei were observed in the early PbSUB1-deficient schizonts (Figure 7), indicating normal nuclear replication, but in later stages many of the schizont nuclei appeared condensed and abnormal.
To produce a quantitative description of these observations, we used confocal microscopy in two independent experiments to categorise a total of 40–60 individual parasitised hepatoma cells at each of several time points following infection with either the excised condSUB1 parasites or the control UIS4/FlpL-F parasites (Figure 8). This analysis confirmed no significant difference in expression and localisation of the 3 marker proteins prior to 52 h post infection. Subsequent to this, however, the PbSUB1-deficient parasites began to display clear differences in the levels or expression pattern of the marker proteins, culminating in a nearly complete absence of formation of daughter merozoites.
Merozoite egress from infected hepatocytes is via extrusion of merosomes, vesicles filled with mature invasive merozoites. To address the consequences of PbSUB1 depletion specifically on merosome formation, hepatoma cell monolayers infected in vitro with condSUB1 or control UIS4/FlpL-F sporozoites were cultured to allow complete parasite development and then cell supernatants were repeatedly harvested between 62–70 h post infection. At these time points, the supernatants normally contain non-adherent rounded-up infected cells as well as merosomes (Figure S9 in Text S1). The detached cells were collected, fixed in suspension, gently centrifuged and processed for IFA, then counted. As shown in Figure S10 in Text S1, a clear deficiency in formation of detached cells and merosomes was observed in the case of the GFP-positive (excised) condSUB1 parasites. In contrast, the fraction of condSUB1 parasites in which pbsub1 excision had not occurred (and which therefore did not express GFP) produced ∼10-fold more merosomes despite representing only ∼20% of the input condSUB1 parasite population, providing an internal control demonstrating that pbsub1 deletion rather than the presence of the modified pbsub1 locus was responsible for the defect in merosome formation. To exclude the possibility that the excised condSUB1 parasites might simply exhibit a delay in merosome formation, cultures were further monitored for up to 85 h post infection; however, even following such prolonged culture, no GFP-positive merosomes were observed in the condSUB1-infected hepatoma cell supernatants (data not shown). Injection of the condSUB1 merosome preparations into naïve mice resulted in a blood-stage infection which contained only non-excised parasites (data not shown), mirroring the bite-back data described in Figure 2 and Figure S7 in Text S1. These results unambiguously confirmed that PbSUB1 is required for completion of liver stage development and formation of infectious liver stage merozoites.
Studying essential blood stage genes in Plasmodium liver stages requires regulatable genetic approaches. A tetracycline repressible transactivator has recently been shown to allow dynamic gene regulation in P. berghei blood stages in vivo and in liver stages in vitro [45], but it remains to be tested whether it can be used successfully on essential liver stage genes. While the potential to use the same transgenic parasite for studying essential genes at different life cycle stages would clearly be an attraction of a tetracycline regulatable system, the sequence-specific Flp recombinase system applied here already provides a powerful tool to study essential genes specifically in the sporozoite and liver stage. However, precise placement of FRT sites has remained a major challenge. In the mouse, target sites for recombinases can often be designed to flank a critical exon excision of which results in effective gene disruption (e.g. [46]). In malaria parasites, in contrast, where genes are relatively short and only around half have introns, flanking the entire protein coding sequence of a gene with FRT sites would be a good default strategy. However, the high AT content and repetitive nature of P. berghei genomic DNA has until now made it impractical to construct such genetic modification vectors for larger genes, since long genomic fragments are unstable in conventional circular high-copy plasmids in E. coli. Matters are further complicated by the observation that FRT sites positioned within 100 bp upstream of a start codon can interfere with gene expression in P. berghei [30]. As a consequence, recent studies have resorted to excising only the 3′ regulatory sequence of a gene [30], [41], [43], resulting in gene silencing due to destabilisation of the mRNA, rather than achieving a complete gene knock out. Whilst this approach can work well, it is not uniformly successful due to the potential for cryptic polyadenylation sites downstream of the modified gene that act to stabilise the mRNA [47], [48]. Furthermore, it is a drawback of this strategy that it sacrifices the key advantage of recombinases in providing the very tight regulation that stems from complete removal of a target gene.
We here have demonstrated that a large genomic insert from a P. berghei genomic DNA library in a low copy linear pJAZZ vector can be manipulated successfully within E. coli using sequence-specific and lambda Red/ET recombinases. This approach, which is well established for generating conditional knock out alleles for the mouse (e.g. [46]), allows FRT sites to be placed almost at will and without the need to manipulate Plasmodium DNA by conventional restriction/ligation cloning. We also generated some of the PCR templates and a Gateway donor cassette that can form part of a more generic strategy for turning large gDNA inserts into complex conditional knock out alleles for P. berghei genes. In the current study, to ensure positioning of the upstream FRT site well away from potentially important flanking promoter elements we chose a position ∼2.3 kb or ∼1.8 kb away from the pbsub1 ATG start codon, distances substantially larger than most Plasmodium promoter sequences mapped to date. We also chose to integrate a strong promoter 5′ to the FRT site in the upstream intergenic region of the target gene. This successfully allowed us to use expression of GFP to monitor excision of the pbsub1 gene in individual liver stage parasites. As a result, in our microscopic analyses of the pbsub1 knock out phenotype we were able throughout to distinguish excised parasites clearly from the minority of parasites in which excision had failed. While this strategy proved useful in the current study, where a large upstream intergenic region was available, for other genes integrating a long promoter sequence may interfere in unpredictable ways with expression either of the target gene itself, or of a neighbouring gene. In such cases it will be possible to omit the hsp70 promoter from the PCR amplicon in Step 1, leaving behind only the 34 bp FRT site after Step 2. Importantly, with the tools described here, constructs similar to that described can be produced and quality-control evaluated within as little as 16 working days. The robustness and speed of recombinase mediated engineering means it can be carried out in continuous liquid culture on 96 well plates. Insertion of an upstream FRT site in Steps 1 and 2 in such an optimised protocol could probably be completed within a week. Steps 3 and 4 can be adapted to use our existing pipeline for Red/ET mediated engineering on 96-well plates [32]. Such an optimised recombinase based strategy would greatly ease the generation of complicated conditional alleles and the study of essential genes in liver stages.
Excision of the flirted pbsub1 gene, as indicated by expression of the GFP reporter in our transgenic parasites, occurred efficiently at the oocyst stage despite expression of FlpL under control of the uis4 (maximally upregulated in infectious sporozoites) promoter. We were not overly surprised by this observation, since although this promoter is maximally active in salivary gland sporozoites [49] we are not aware of any previous evidence that it is entirely ‘off’ in oocysts. It is important to note that GFP expression in an oocyst does not imply that all the resident sporozoites have undergone excision, only that some have done so. On the other hand it is also conceivable that - given the mode of parasite replication in oocysts, in which sporozoites bud from a syncytial sporoblast containing multiple nuclei [50] - GFP produced by excised parasites could be incorporated into the cytosol of non-excised sporozoites developing within the same oocyst. This could lead to a degree of over-estimation of the excision rate in our study, which may explain the apparent small mismatch between the proportions of GFP-positive (excised) input sporozoites and GFP-negative (non-excised) EEFs observed in some of the in vitro hepatocyte infection experiments (e.g. Figure 4). Importantly, the stage specificity of the uis4 promoter in this system is supported by the fact that the transgenic condSUB1 and condSUB1short parasites showed no growth defect in the asexual blood stages, and no GFP expression was observed in those stages (although this does not rule out low level ‘leaky’ FlpL expression since excision of pbsub1 in asexual blood stages would likely produce non-viable parasites).
Our finding that hepatocyte invasion, and the early stages of liver stage growth were all normal in the PbSUB1-deficient condSUB1 parasites confirms that PbSUB1 does not play an important role in these phases of the parasite life cycle; indeed, an effect on hepatocyte invasion was not expected anyway since PbSUB1 expression was undetectable in salivary gland sporozoites, either using our polyclonal anti-PbSUB1 antibodies or by epitope-tagging. In contrast, investigation of more mature liver stages of the PbSUB1-deficient parasites revealed a clear defect in merozoite formation. Unexpectedly, this was associated with an apparent absence of expression of the glycolipid-anchored major plasma membrane protein MSP1. MSP1 is an established SUB1 substrate, but our previous studies in blood stages [20], [22] have demonstrated that SUB1-mediated proteolysis of MSP1 requires discharge of SUB1 from exonemes, consistent with the topology of MSP1 on the merozoite surface where it is effectively a component of the PV lumen. We therefore had no a priori reason to predict that an absence of SUB1 should affect MSP1 expression or trafficking. Similarly unexpected was the observation that the MSP1 expression defect in PbSUB1-deficient parasites, as well as a small but significant effect on the size of the intracellular EEFs, was evident well before the point at which expression of PbSUB1 was detectable by IFA. We interpret this result as indicating that PbSUB1 plays an important role in EEF development before significant accumulation of the protease in exonemes. It is conceivable that an absence of SUB1 impacts on EEF exoneme biogenesis, or even that SUB1 plays a general processing role in parasite protein trafficking, analogous to that of certain members of the subtilisin-like prohormone convertase family [51]. Further work will be required to explore this possibility. Whether these additional roles for SUB1 also operate in blood stages will require the application of a conditional expression strategy suitable for use in Plasmodium blood stages, such as the recently-described DiCre system [48], and this work is underway.
MSP1 has previously been shown to be important for merozoite development [30]. It is perhaps unsurprising then, that a block in merosome formation was evident in the PbSUB1-deficient parasites; this is likely a direct result of the defect in merozoite formation. It was also accompanied by a block in PVM rupture. In blood stages, members of the SERA family have been implicated in egress. Many or most blood stage SERA family members are substrates of SUB1 [20], [25], [38], and moreover at least one member of the SERA protein family in P. berghei, PbSERA3, undergoes proteolytic processing in liver stages [11]. It is therefore possible that the observed defects in PVM rupture and egress are due to an absence of PbSUB1-mediated processing of PbSERA3 and/or other SERA proteins.
Of the two other liver stage parasite proteins that have previously been implicated in egress using gene disruption approaches - liver-specific protein 1 (LISP1; [52]) and the parasite cyclic GMP-dependent protein kinase (PKG; [31]) – the latter has recently been shown to play a key role in regulating discharge of SUB1 into the PV in blood stage schizonts [22]. Given our definitive evidence here for expression of SUB1 in liver stages and its essential role, it is tempting to speculate that the egress defect observed in the PKG knockout reported by Falae et al. [31] is at least in part due to a resulting block in SUB1 discharge. Future work will focus on the fate of SUB1 in PKG knockout EEFs.
In conclusion, we have combined cutting-edge molecular tools with a conditional gene deletion strategy to obtain the first complete conditional deletion of a liver stage Plasmodium gene. The molecular strategies described here will render stage-specific regulation of gene expression in Plasmodium more accessible to the research community. Despite the absence of associated pathology, the liver stage of the malaria parasite life cycle acts as an important amplification step in the infection pathway and so has long been considered an attractive target for vaccine or drug-mediated prophylactic approaches to disease control. Our results show that inhibitors of SUB1 could be used in prophylactic approaches to control or block the pre-erythrocytic stage of the malaria parasite life cycle.
Research using animals was approved by the Ethical Review Committee of the Wellcome Trust Sanger Institute and was conducted in accordance with the UK Animals (Scientific Procedures) Act 1986, under licence number PPL 80/2158 issued by the UK Home Office.
Tuck-Ordinary outbred mice and C57BL/6N inbred mice were used as 6–8 week old females. All animal procedures were carried out in accordance with a valid UK Home Office project licence. Anopheles stephensi strain SD500 mosquitoes were allowed to feed on mice 3–4 days after injection of infected blood, then maintained on fructose at 20°C. Sporozoite numbers were determined from day 21 by homogenising dissected salivary glands and counting released sporozoites using a haemocytometer. For conditional gene disruption experiments, mosquitoes were placed at 25°C from day 17–18 post infection and sporozoite isolation was performed from day 26 post infection. Parasite transfection experiments used the GFP-expressing P. berghei ANKA clone 507m6cl1 [40], [53] (kindly provided by Chris Janse and Shahid Khan, University of Leiden) for epitope-tagging of pbsub1, and the P. berghei NK65 UIS4-FlpL deleter clone [30] (a kind gift of Robert Menard, Institut Pasteur, Paris) for generation of the condSUB1 clones. The P. berghei NK65 UIS4-FlpL-F clone [41], which constitutively expresses GFP under the control of the P. berghei hsp70 promoter, was used as a control for experiments with the condSUB1 clones; prior to isolation of control UIS4-FlpL-F sporozoites for hepatocyte infection experiments, mosquitoes infected with this line were also subjected to a temperature shift as described above for the condSUB1 parasites. Transfection of targeting constructs followed standard methodology [40]. Transgenic parasites were selected using pyrimethamine and cloned by limiting dilution.
DNA encoding the predicted catalytic domain (residues Ser196-Asn599) of PbSUB1 (PBANKA_110710) was amplified by PCR using primers F_CatDPbSUB1synth_BamHI and R_CatDPbSUB1synth_XhoI (Table S1) and cloned into the bacterial expression vector pGEX-His (a kind gift of Dominique Soldati-Favre, University of Geneva, Switzerland) for expression as a recombinant hexahistidine-tagged protein in E. coli BL21-Gold DE3 cells (Stratagene). Recombinant product was purified by nickel chelate chromatography and used to immunise a rabbit. Before use, sera were adsorbed against E. coli acetone powder (40 mg/ml serum) to deplete antibodies against bacterial proteins. Other primary antibodies used were: the anti-HA.11 mouse mAb 16B12 (Covance); the anti-HA rat mAb 3F10 (Roche); the P. berghei MSP1-specific mAb 25.1 [54] (a gift from Tony Holder, NIMR, London, UK); a polyclonal antiserum against the PVM protein EXP1 [55], [56] (a kind gift from Volker Heussler, University of Bern, Switzerland); a polyclonal anti-HSP70 mouse antibody (a gift from Kai Matuschewski, Max Planck Institute, Berlin, Germany); a polyclonal antibody specific for PbSERA3, raised against a recombinant fragment of PbSERA3 called PbS3C1 [15]; and chicken and rabbit polyclonal anti-GFP sera (Abcam). Western blot analysis of P. berghei schizont SDS extracts and IFA analysis of paraformaldehyde-fixed, permeabilized blood stage parasites were performed as described previously [15], [57]. For IFA of parasitised hepatoma cells, infected cells were fixed for 15 min in 3% paraformaldehyde in phosphate-buffered saline (PBS), then quenched in 50 mM ammonium chloride in PBS. Samples were permeabilized with 0.1% (v/v) Triton X-100, washed, then incubated for 1 h in blocking solution (2% (w/v) BSA in PBS). Samples were probed with relevant primary antibodies diluted in blocking solution for 1 h, then with appropriate secondary antibodies before counterstaining with DAPI and observation using a laser scanning confocal microscope (LSM510, Zeiss). Primary antibodies were used for IFA at dilutions ranging from 1∶400–1∶1000. Secondary antibodies (usually used at a 1∶1000 dilution) were Alexa Fluor 488, 555 or 633-conjugated antibodies against mouse, rabbit or chicken IgG (Invitrogen).
A 1,248 bp sequence corresponding to the 3′ coding sequence of the pbsub1 gene was amplified from P. berghei gDNA using primers F_KI_XhoI and R_KI_ApaI (Table S1) and cloned into the transfection vector pSD278-HA (a kind gift of Dominique Soldati-Favre, University of Geneva, Switzerland), to produce an in-frame fusion to a single HA epitope tag, followed by the pbdhfr-ts 3′ UTR and a hDHFR drug selection cassette. Before transfection, the plasmid was linearised at a unique Hind III restriction site that lies 682 bp upstream of the stop codon of the pbsub1 gene.
To construct a conditional deletion vector for pbsub1 we first identified a clone, PbG01-2474a09, from an arrayed and end-sequenced library of P. berghei ANKA cl15cy1 genomic DNA in E. coli [32] that carried pbsub1 (PBANKA_110710) and flanking genes on a 9.6 kb insert in a low copy linear plasmid [58]. This clone served as starting point for engineering an allelic exchange vector using a combination of site specific and lambda phage recombinases in four steps (Figure S11 in Text S1), which used protocols essentially as described [59]. Each intermediate product was fully sequenced before moving on to the next stage. For Step 1 (Figure S11 in Text S1) we used lambda Red/ET mediated recombination with zeocin selection to insert a PCR amplicon ∼2.3 kb or ∼1.8 kb upstream of the pbsub1 start codon that comprised an hsp70 promoter followed by a zeo-pheS cassette for positive and negative selection in E. coli. The zeo-pheS cassette was flanked by two FRT sites in the same orientation and its insertion point was still 800 bp away from the start codon of the upstream gene. Primers used were F_Step1_rec and R_Step1_rec or F_Step1short_rec and R_Step1short_rec (Table S1) and plasmid pColE1 5′hsp70-ATG-FRT-zeo-pheS-FRT served as a template (see Supplemental Methods). Multiple stop codons were present downstream of the inserted sequence. In Step 2, the FRTed zeo-pheS cassette was excised by inducing Flp-e recombinase expression in E. coli under negative selection against pheS, leaving behind only the hsp70 promoter and one FRT site in the upstream intergenic region. In Step 3, Red/ET mediated recombination was used under zeocin selection to insert immediately upstream of the pbsub1 stop codon a DNA fragment that introduced a single HA epitope tag followed by a generic 3′ UTR from the pbdhfr-ts gene as well as a zeo-pheS cassette flanked by attR recognition sites for Gateway clonase. The DNA fragment used in this step was release by a Hind III digest from plasmid pColE1 sub1-HA-attR1-zeo-pheS-attR2-3′sub1 (see Supplemental Methods). Finally, the product of Step 3 was subjected to an in vitro Gateway reaction under negative selection against pheS, which replaced the bacterial selection cassette with an FRT site immediately followed by a GFP coding sequence, an hsp70 terminator sequence and an expression cassette for hdhfr-yfcu for selection in P. berghei. In the final construct, FRT sites were positioned such that excision of pbsub1 would bring an ATG start codon 5′ of the upstream FRT site in frame with the gfp coding sequence, allowing expression of the fluorescent marker protein from the hsp70 promoter. The construct was tested by Flp-e activation in E. coli TSA cells, followed by PCR genotyping of the excised vector. See Supplemental Methods for more details on vector generation.
Genomic DNA was extracted from blood stage parasites, infected mosquito midguts or infected salivary glands. For Southern blot analysis, the DNA was digested with suitable restriction enzymes and separated by gel electrophoresis. Transfer of the DNA to a nitrocellulose membrane and hybridisation with gene-specific probes was performed according to standard procedures. Probes were labelled with α-[32P] adenosine triphosphate (Amersham Biosciences) by random priming using a Prime-It Random Prime kit (Stratagene).
To detect integration of the transfected pPbSUB1-HA construct, primers Fprom_PbSUB1, F2_PbSUB1and R_HA were used. The size of the expected PCR products - which could only be produced if integration occurred as predicted - were 1,942 bp and 1,288 bp respectively. Two other control PCRs were generated using primers F1_PbSUB1, Fprom_PbSUB1, R1_3′utr. These were expected to produce 1,937 bp and 2,036 bp products only from the unmodified wild type pbsub1 locus. For Southern blot analysis, parasite genomic DNA was digested with Pml I and Nhe I. An 841 bp probe annealing to the 5′ flanking sequence of pbsub1 was generated with primers F_prom_probe and R_prom_probe. This allowed detection of a 4.6 kb band for the wild type pbsub1 locus and a 10.6 kb band for the expected integration event.
To detect integration of the pJazz-FRTed-pbsub1 constructs, primer F_selection (forward) was used with reverse primers R_ext1, R_ext2, or R_ext3 to produce PCR products of 1,471 bp, 2,034 bp and 2,617 bp respectively. As control PCRs for detection of the unmodified pbsub1 locus primers F2_PbSUB1 or F3_PbSUB1 were used together with R1_3′utr. These were predicted to produce products of 1,382 bp and 598 bp only from the unmodified pbsub1 locus. For Southern blot analysis, parasite genomic DNA was digested using StuI and NciI, and a 1,246 bp probe annealing to the pbsub1 gene was generated with primers F_KI_XhoI and F_KI_ApaI. This allowed detection of a 3 kb band for the unmodified pbsub1 locus and an 8.4 kb band for P. berghei condSUB1 and condSUB1short parasites.
Intact P. berghei chromosomes were separated by pulsed-field gel electrophoresis on 0.8% agarose gels as described previously [32]. Gels were blotted and hybridised using the standard Southern blot protocol. For detection of integration in the correct chromosome a 452 bp probe hybridising to the pbdhfr-ts 3′ UTR was generated with primers F_3′utr_pbdhfr_probe and R_3′utr_pbdhfr_probe (Table S1).
Hepa1-6 mouse or HepG2 human hepatoma cells were cultured in Dulbecco's Modified Eagle Medium high glucose (DMEM; Gibco/Invitrogen) plus 10% (v/v) fetal calf serum (Invitrogen). For sporozoite infection rate assays and determination of parasite size at 28 and 48 h post infection, cells were seeded in 96-well culture plates (10,000–15,000 cells per well). The next day, 20,000 or 50,000 control or condSUB1 sporozoites were added to each well in complete medium containing 1% penicillin/streptomycin. Following incubation for 2–48 h, cells were washed four times with PBS then fixed and processed for IFA. For differential extracellular/intracellular staining (in/out assay), staining was performed as described previously [44] using a Cy3 conjugated mouse anti-CS antibody, a rabbit Alexa Fluor 488 anti-GFP antibody (Invitrogen) and Hoechst 33342 (Molecular Probes/Invitrogen).
For the in/out assay and the determination of parasite area at 28 h and 48 h post invasion, images of fixed, stained wells were automatically acquired with a Cellomics ArrayScan VTI HCS Reader (20× magnification; 100 or 155 fields per well of a 96-well plate) and analysed using the Colocalization BioApplication software (ThermoFisher). For the in/out assay, object recognition was based on the GFP staining and average intensity of the CSP-specific signal was used to distinguish extracellular (high) and intracellular (low) sporozoites. The number of cell nuclei was determined based on the Hoechst signal and used to calculate invasion rates. For the parasite size assay, number of cell nuclei and number and area of parasites were determined. Object recognition was based on the HSP70-specific staining, applying a minimum object size of 15 µm2 at 28 h post invasion and 50 µm2 at 48 h post invasion. The average intensity of the GFP staining was used to discriminate excised and non-excised condSUB1 parasites. The cell nuclei were counted with help of the Hoechst signal and numbers used to calculate infection rates. For each comparison, Student's t-test was performed and for the invasion rates the threshold of the p value was set to 0.05, whereas for parasite sizes a Bonferroni corrected threshold of the p value was chosen at 0.0167 (*) or 0.0033 (**).
To assess release of merosomes and detached cells in hepatoma cultures, 40,000–80,000 cells were seeded into 24-well plates one day before sporozoites were added at a multiplicity of infection of 1. Merosomes were collected from the culture supernatants between 65 and 69 h post infection, fixed in suspension, pelleted, resuspended in a small volume of medium, fixed onto poly-L-lysine coated microscope slides (Menzel-Gläser), and stained with anti-GFP, anti-MSP1, anti-EXP1 antibodies and DAPI before analysis by fluorescence microscopy.
C57BL/6N mice were injected intravenously with 20,000 P. berghei sporozoites and livers were collected at 40 h post infection. Livers were mechanically homogenized on ice with a Tissue Tearor (IKA Ultra Turrax T-10) in 4 ml denaturing solution (4 M guanidine thiocyanate, 25 mM sodium citrate pH 7, 0.5% N-Laurosyl-sarcosine, 0.1% β-mercaptoethanol) total RNA extracted using an RNeasy Mini Kit (Qiagen). Samples were treated with Turbo DNAse (Ambion) according to the manufacturer's instructions. One microgram of total RNA was reverse-transcribed using a Transcriptor First Strand cDNA Synthesis kit (Roche). Parasite 18S ribosomal RNA and mouse hypoxanthine guanine phosphoribosyltransferase (HPRT) cDNAs obtained from the reaction were quantified by real-time quantitative fluorogenic PCR using previously described primers respectively F_Pb18S and R_Pb18S for P. berghei 18S ribosomal RNA, and F_HPRT and R_HPRT for the Mus musculus housekeeping gene HPRT gene. To quantify gene expression, Power SYBR Green PCR Master Mix (Applied Biosystems) was used according to the manufacturer's instructions. The reaction was performed in an ABI Prism 7000 sequence Detection System (Applied Biosystems) with 2 µl of cDNA in a total volume of 25 µl and the following reaction conditions: 1 cycle of 2 min at 50°C, 1 cycle of 10 min at 95°C, 50 cycles of 15 sec at 95°C and 1 min at 60°C. Each sample was assayed in triplicate. Relative amounts of RNA were calculated using the ABI Prism 7000 SDS 1.2.3 Software, and normalised against expression levels of the mouse HPRT mRNA.
|
10.1371/journal.pcbi.1000696 | Parameter Estimation and Model Selection in Computational Biology | A central challenge in computational modeling of biological systems is the determination of the model parameters. Typically, only a fraction of the parameters (such as kinetic rate constants) are experimentally measured, while the rest are often fitted. The fitting process is usually based on experimental time course measurements of observables, which are used to assign parameter values that minimize some measure of the error between these measurements and the corresponding model prediction. The measurements, which can come from immunoblotting assays, fluorescent markers, etc., tend to be very noisy and taken at a limited number of time points. In this work we present a new approach to the problem of parameter selection of biological models. We show how one can use a dynamic recursive estimator, known as extended Kalman filter, to arrive at estimates of the model parameters. The proposed method follows. First, we use a variation of the Kalman filter that is particularly well suited to biological applications to obtain a first guess for the unknown parameters. Secondly, we employ an a posteriori identifiability test to check the reliability of the estimates. Finally, we solve an optimization problem to refine the first guess in case it should not be accurate enough. The final estimates are guaranteed to be statistically consistent with the measurements. Furthermore, we show how the same tools can be used to discriminate among alternate models of the same biological process. We demonstrate these ideas by applying our methods to two examples, namely a model of the heat shock response in E. coli, and a model of a synthetic gene regulation system. The methods presented are quite general and may be applied to a wide class of biological systems where noisy measurements are used for parameter estimation or model selection.
| Parameter estimation is a key issue in systems biology, as it represents the crucial step to obtaining predictions from computational models of biological systems. This issue is usually addressed by “fitting” the model simulations to the observed experimental data. Such approach does not take the measurement noise into full consideration. We introduce a new method built on the combination of Kalman filtering, statistical tests, and optimization techniques. The filter is well-known in control and estimation theory and has found application in a wide range of fields, such as inertial guidance systems, weather forecasting, and economics. We show how the statistics of the measurement noise can be optimally exploited and directly incorporated into the design of the estimation algorithm in order to achieve more accurate results, and to validate/invalidate the computed estimates. We also show that a significant advantage of our estimator is that it offers a powerful tool for model selection, allowing rejection or acceptance of competing models based on the available noisy measurements. These results are of immediate practical application in computational biology, and while we demonstrate their use for two specific examples, they can in fact be used to study a wide class of biological systems.
| Many biological processes are modeled using ordinary differential equations (ODEs) that describe the evolution over time of certain quantities of interest. At the molecular level, the variables considered in the models often represent concentrations (or number of molecules) of chemical species, such as proteins and mRNA. Once the pathway structure is known, the corresponding equations are relatively easy to write down using widely accepted kinetic laws, such as the law of mass action or the Michaelis-Menten law.
In general the equations will depend on several parameters. Some of them, such as reaction rates, and production and decay coefficients have a physical meaning. Others might come from approximations or reductions that are justified by the structure of the system and, therefore, they might have no direct biological or biochemical interpretation. In both cases, most of the parameters are unknown. While sometimes it is feasible to measure them experimentally (especially those in the first class), in many cases this is very hard, expensive, time consuming, or even impossible. However, it is usually possible to measure some of the other variables involved in the models (such as abundance of chemical species) using PCR, immunoblotting assays, fluorescent markers, and the like.
For these reasons, the problem of parameter estimation, that is the indirect determination of the unknown parameters from measurements of other quantities, is a key issue in computational and systems biology. The knowledge of the parameter values is crucial whenever one wants to obtain quantitative, or even qualitative information from the models [1],[2].
In the last fifteen years a lot of attention has been given to this problem in the systems biology community. Much research has been conducted on the applications to computational biology models of several optimization techniques, such as linear and nonlinear least-squares fitting [3], simulated annealing [4], genetic algorithms [5], and evolutionary computation [6],[7]. The latter is suggested as the method of choice for large parameter estimation problems [7]. Starting with a suitable initial guess, optimization methods search more or less exhaustively the parameter space in the attempt to minimize a certain cost function. This is usually defined as the error in some sense between the output of the model and the data that comes from the experiments. The result is the set of parameters that produce the best fit between simulations and experimental data. One of the main problems associated with optimization methods is that they tend to be computationally expensive and may not perform well if the noise in the measurements is significant.
Considerable interested has also been raised by Bayesian methods [8], which can extract information from noisy or uncertain data. This includes both measurement noise and intrinsic noise, which is well known to play an important role in chemical kinetics when species are present in low copy numbers [9]. The main advantage of these methods is their ability to infer the whole probability distributions of the parameters, rather than just a point estimate. Also, they can handle estimation of stochastic systems with no substantial modification to the algorithms [10]. The main obstacle to their application is computational, since analytical approaches are not feasible for non-trivial problems and numerical solutions are also challenging due to the need to solve high-dimensional integration problems. Nonetheless, the most recent advancements in Bayesian computation, such as Markov chain Monte Carlo techniques [11], ensemble methods [12],[13], and sequential Monte Carlo methods that don't require likelihoods [10],[14] have been successfully applied to biological systems, usually in the case of lower-dimensional problems and/or availability of a relatively high number of data samples. Maximum-likelihood estimation [15],[16] has also been extensively applied.
More recently, parameter estimation for computational biology models has been tackled in the framework of control theory by using state observers. These algorithms were originally developed for the problem of state estimation, in which one seeks to estimate the time evolution of the unobserved components of the state of a dynamical system. The controls literature on this subject is vast, but in the context of biological or biochemical systems the classically used approaches include Luenberger-like [17], Kalman filter based, [18]–[20], and high-gain observers [21]. Other methods have been developed by exploiting the special structure of specific problems [22]. State observers can be employed for parameter estimation using the technique of state extension, in which parameters are transformed into states by suitably expanding the system under study [22]–[24]. In this context extended Kalman filtering [25],[26] and unscented Kalman filtering [27] methods have been applied as well.
When the number of unknown parameters is very large, it is often impossible to find a unique solution to this problem. In this case, one finds several sets of parameters, or ranges of values, that are all equally likely to give a good fit. This situation is usually referred to as the model being non identifiable, and it is the one that's most commonly encountered in practice. Furthermore, it is known that a large class of systems biology models display sensitivities to the parameter values that are roughly evenly distributed over many orders of magnitude. Such “sloppiness” has been suggested as a factor that makes parameter estimation difficult [28]. These and similar results indicate that the search for the exact individual values of the parameters is a hopeless task in most cases [6]. However, it is also known that even if the estimation process is not able to tightly constrain any of the parameter values, the models can still be able to yield significant quantitative predictions [12].
The purpose of the present contribution is to extend the results on parameter estimation by Kalman filtering by introducing a procedure that can be applied to large parameter spaces, can handle sparse and noisy data, and provides an evaluation of the statistical significance of the computed estimates. To achieve this goal, we introduce a constrained hybrid extended Kalman filtering algorithm, together with a measure of accuracy of the estimation process based on a variance test. Furthermore, we show how these techniques together can be also used to address the problem of model selection, in which one has to pick the most plausible model for a given process among a list of candidates. A distinctive feature of this approach is the ability to use information about the statistics of the measurement noise in order to ensure that the estimated parameters are statistically consistent with the available experimental data.
The rest of this paper is organized as follows. In the Methods Section we introduce all the theory associated with our procedure, namely the constrained hybrid extended Kalman filter, the accuracy measure and its use in estimation refinement, and the application to the model selection problem. In the Results Section we demonstrate the procedure on two examples drawn from molecular biology. Finally, in the Discussion Section we summarize the new procedure, we give some additional remarks, and we point out how these findings will be of immediate interest to researchers in computational biology, who use experimental data to construct dynamical models of biological phenomena.
Throughout this paper, we will assume that the process of interest can be modeled by a system of ordinary differential equations of the form:(1)The state vector usually contains concentrations of certain chemical species of interest, such as mRNA or proteins. The input signal represents some kind of external forcing of the process, such as temperature changes, the addition or removal of certain chemicals or drugs, and so forth. The output signal represents the quantity or quantities we can measure experimentally. These are related to the state through the function , which we call the output function. The output function is to be determined from the design of the biological experiments that are used to get the measurements for parameter estimation. As an example, when measuring protein concentrations, in some biological experiments it is harder and/or more expensive and/or more time consuming to distinguish among different post-translational modifications of the same protein. This situation corresponds in our setting to choosing equal to the sum of two or more state variables, representing the total amount of protein.
The vector contains the unknown parameters that we seek to estimate. Note that, since the parameters are constants, it is always possible to consider them as additional state variables with a rate of change equal to zero. In this way, we treat them as constant functions of time as opposed to constant numbers. This technique is usually referred to as state extension. Our system (1) then becomes:(2)
Using state extension, the problem of parameter estimation is converted into a problem of state estimation, that is determining the state of a system from measurements of the output. More precisely, we are trying to determine the initial conditions that when used to initialize the system (2) generate the observed output . In the case of the parameters, since , it is obvious that for all .
Solving this problem requires answering the following two questions.
The first question is usually referred to as the problem of identifiability. In control theory, much work has been done in studying this property in terms of another one called observability [23],[24]. Roughly speaking, a system is observable if every set of initial conditions produces an output that is different from the one generated by every other set. Identifiability can also be studied a posteriori [6], by testing the reliability of the estimates after they have been computed. We will make use of this second approach.
To answer the second question, we need to show how to design an algorithm (or device) that can estimate and from measurements of , which, in general, will not be perfect, but noisy and sparse. Such algorithms, called state observers, can be formulated in a plethora of different ways, each of which is better suited for different applications. The observer we are going to use is based on extended Kalman filtering, and is described in detail in the next Section.
Extended Kalman filtering is considered to be the de-facto standard of nonlinear state estimation [29]. It found several applications in many different fields, such as positioning systems, robot navigation and economics. The Kalman filter is a set of equations that provides an efficient computational technique to estimating the state of a process, in a way that minimizes the covariance of the estimation error. The filter is very powerful in several aspects: it supports estimations of past, present, and even future states, and it can do so even when the precise nature of the modeled system is unknown. Unlike most of the classical parameter estimation methods, the Kalman filter is a recursive estimator. At each time step the filter refines the previous estimate by incorporating in it new information from the model and from the output.
The Kalman filter works in two steps: first it estimates the process state and covariance at some time using information from the model only (prediction); then it employs a feedback from the noisy measurements to improve the first estimates (correction). As such, the equations for the Kalman filter fall into two groups: time update equations for the prediction step and measurement update equations for the correction step. The time update equations are responsible for propagating forward (in time) the current state and error covariance estimates to obtain the a priori estimates for the next time step. The measurement update equations are responsible for the feedback, i.e. for incorporating a new measurement into the a priori estimate to obtain an improved a posteriori estimate. After each time and measurement update pair, the process is repeated with the previous a posteriori estimates used to predict the new a priori estimates.
In order to set these ideas in a more rigorous mathematical framework, consider the following system:(3)As we note from the structure of system (3), we are assuming that we have a continuous-time process which we want to estimate using discrete-time measurements of the output. This is the most common case when dealing with deterministic models of biological systems. These are usually of the form (1), therefore continuous-time. However, the measurements for estimation tend to be available only at discrete time instants. We will denote these instants , with being the corresponding values of the measurements. The output of the filter will then be the a posteriori estimates of the state corresponding to instants , which we will denote . We remark that after applying state extension as described in the previous Section, the unknown parameters are now part of the state of the system, therefore their estimates at time are components of . We also note that the output function in (3) is allowed to be different at different time step: this is very important e.g. when incorporating data from different measurements, because it allows the algorithm to use measurements of different species at different times.
The variable , usually called the process noise, represents the amount of confidence we have in our model. The process noise is assumed to be a Gaussian random variable with zero mean and covariance , where is a positive definite matrix. The noise that affects the different components of the state is assumed to be uncorrelated, so that is diagonal. Larger entries in correspond to lower confidence in the accuracy of the model. The variable is referred to as the measurement noise, and expresses the reliability of the measurements. The measurement noise is also assumed to be Gaussian with zero mean, and its covariance matrix will be denoted by . Again, is assumed to be a positive definite, diagonal matrix, since the noise that affects different measurements is assumed to be uncorrelated. Note that while is usually chosen by the user in order to tell the filter how much the model should be trusted, is fixed by the quality of the measurements. In other words, the statistics of the measurements noise are assumed to be known. This fact will be particularly important for the a posteriori reliability test described in the next Section.
The variation of the Kalman filter we present here is the one that is best suited for a system of the form (3), and it is usually referred to as the hybrid extended Kalman filter (HEKF). The word extended refers to the fact that it can deal with nonlinear systems, while hybrid indicates that it uses continuous-time process model and discrete-time measurements. We next describe the time update equations and measurement update equations of the HEKF.
First of all, we need some initial conditions to start the filter from. Ideally, we would like the initial conditions to be (the initial conditions of the process) but this is clearly not possible. Since we do not have any measurements available to estimate , it makes sense to take our initial estimate of equal to the expected value of the initial state . Therefore, we write:(4)It follows that the initial condition for the error covariance can be set as:(5)
We can now apply the time update equations to obtain the current a priori estimates. The current a priori state estimate, which we denote , is formed by integrating the continuous-time process in the time interval , using the previous a posteriori estimate as initial condition. The current a priori error covariance estimate, denoted , is formed by integrating a differential Lyapunov equation using the previous a posteriori error covariance as initial condition [29].(6a)(6b)
The matrix is the Jacobian of evaluated at the previous a posteriori state estimate. The structure of equations (6) shows a very important feature of the HEKF algorithm, i.e. its ability to deal with non-uniformly sampled data. As we will see in the examples in the Results Section, this is useful because it allows one to capture all the information about the evolution of a process using a minimum number of data points.
The measurement update equations are used to form the a posteriori estimates by incorporating information from the output of the system into the a priori estimates. The correction is based on the difference between the actual measurement and the predicted measurement, that is what the measurement would be if the real value of the state were exactly equal to its a priori estimate. Such difference is weighed by a gain, which takes into account the fact that the measurements are not perfect. The gain at time is given by:(7)where the matrix is the Jacobian of evaluated at the previous a posteriori state estimate. Given that, the current a posteriori state and error covariance estimates, denoted and respectively, are formed using the following equations:(8a)(8b)
We refer to [29] for a rigorous derivation of the equations we presented so far.
We remark that the algorithm we just introduced, as well as the ones employed in other works [25]–[27], provides unconstrained estimates. In some cases it is necessary to take into account equality or inequality constraints that prevent from assuming certain values. This can be important for the following reasons.
To cope with these issues, we apply the constrained estimation technique developed in [30],[31]. This is derived using the fact that the estimate is the value that maximizes the conditional probability of given the measurements up to time . Furthermore, and are jointly Gaussian, which means that is conditionally Gaussian given . Finally, if , and are jointly Gaussian, then is the conditional mean of given the measurements . These three properties, which are derived in [32], imply that the conditional probability of given can be written as:Now, suppose we have a set of linear constraints of the form , where is a constant matrix of suitable dimensions. If does not satisfy the constraints, we need to replace it with a constrained estimate . This can be obtained by maximizing subject to the constraints, or equivalently, by maximizing its natural logarithm. Therefore, the problem we need to solve can be cast as:(9)
Since is a covariance matrix, and it is therefore strictly positive definite, this is a strictly convex quadratic programming problem that can be easily solved using standard algorithms, such as reflective Newton methods [33] and active set methods [34].
While for linear models the Kalman filter has nice convergence properties, in the case of the extended Kalman filter for nonlinear systems no such properties have been proven yet. As it is well-known in the literature [29], sometimes the filter may diverge, or may give biased estimates. While the first situation is easily detected in any implementation, the second one is dangerous, because the algorithm appears to run normally but produces severely wrong results. It is therefore extremely important to have a test that allows us to assess the reliability of the estimates.
The test we present here is based on a simple estimation of the variance of a random variable. Consider again a continuous-time process which is measured at discrete time instants. Assuming we are able to measure different quantities, we can rewrite our model expanding the components of the output:(11)As in the previous Section, we assume that is a Gaussian random variable with zero mean and diagonal covariance matrix . This means that is a matrix, whose diagonal entries are the variances of each component of . What (11) says is that each output is a sampled version of the corresponding function of the state, with an additive noise superimposed to it.
Now, suppose that by running the HEKF we find an estimate of . Let be the solution of (11) corresponding to . If we accept as a good approximation of the real solution , then we can write estimates of each component of the noise as:(12)This equation, for , gives samples of Gaussian random variables with zero mean. The main idea behind the test is that if is close to , and consequently is close to , then the variance of will be close to the variance of .
Let be the variance of . We can use the samples (12) to build a point estimate of in the following way:(13)The random variable has a probability density function equal to the distribution with degrees of freedom [35].
Using this fact, we can form interval estimates of corresponding to different confidence coefficients . The confidence coefficient is a probability, so it takes values between 0 and 1. Common values for include 0.9, 0.95 and 0.997. Denote by the -th percentile of the distribution with degrees of freedom. Then, is in the interval(14)with a probability of .
It is then clear that if the real variance of does not lie in the interval indicated by (14), it is extremely unlikely that the measurements were generated by the set of parameters , given the fact that the noise has a variance of . Therefore, we can reject the estimate as wrong with a confidence of .
We remark that this test can be also used independently of the HEKF to validate/invalidate the estimates computed by any other parameter estimation method.
Although the HEKF can be applied to fairly large extended systems, when the parameter space is very large (and the extended system is therefore not observable) a single run of the filter will generally yield estimates that do not satisfy the identifiability test described in the previous Section. Also, the estimates will be characterized by large uncertainties, as one can see by inspecting the entries of the matrices. In this situation, the solution to the parameter estimation problem is not unique, therefore there will be infinite sets of parameters that are all equally likely to be correct. The best that one can do in this case is to find one or more values of such that the corresponding solutions are consistent with the experimental observations in the sense of the test.
In order to do that, we can make use of the probabilistic information we have about the measurement noise . In particular, we know that is a Gaussian random variable with zero mean and covariance . As we saw in the previous Section, given a certain estimated parameter set , we can construct samples of an estimate of through (12). It makes sense, then, to ask for which values of the mean and variance of will be close to zero and respectively. In other words, one can minimize the expected value and the difference between and by solving the following problem:(15)
The weights and can be chosen by the user to attribute different relative importances to the mean matching and to the variance matching parts of the cost function. The most appropriate choice can be different for different problems. Note the scaling that is introduced in the function, which ensures that all the measurements are equally weighted in the minimization process, regardless of their size. This problem will not have any special properties in general, so it can be solved with any general purpose minimization algorithm. The Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm [36] has proven to be a good practical choice.
We argue that this moment matching optimization is a better alternative than directly fitting the data points, as it guarantees that the result of the optimization process will be a statistically valid parameter set in the sense of the test (see the examples in the Results section).
To summarize, the proposed algorithm is a three-stage process. In the first stage, we run the constrained HEKF algorithm on the model to get a first estimate of the unknown parameters. In the second stage we study a posteriori the identifiabilty problem, by running the test. If the test is passed, the HEKF was able to recover the unique solution to the problem and the first estimate can be considered valid. If not, most likely no unique solution exists, and the first estimate needs to be refined by running the third stage, i.e. the moment matching optimization. The whole procedure is visualized in the flowchart of Figure 1.
One of the most interesting features of approaching the parameter estimation problem using state extension is that it allows for a simultaneous estimation of both the state and the parameters of the process under investigation. Therefore, the Kalman filter, together with the variance test we described, can also be used to address the problem of model selection.
Frequently, the structure of biochemical pathways is not completely known. One has an idea of the genes and proteins that play a role in a certain process, but the exact interconnections among such components are not fully elucidated. It may not be clear, for example, whether a certain gene is regulated using a positive feedback loop or a negative feedback loop, or if a certain reaction takes place with or without intermediate steps. In these scenarios, it is possible to write down different models corresponding to the different hypotheses and then use the Kalman filter to assess which one is the most likely to have generated the measurements that are observed in the experiments.
In order to simplify the presentation, suppose we have two different models of the form (3) for the same process. We can write them as:(16)The two models differ in everything except the measured data points and the statistics of the noise that is superimposed to them.
Running the HEKF for these models will give estimates of their states, which we will denote and . In analogy to what we did for the test, we can plug the estimates into and respectively. This will give two different estimates of the measurement noise :(17a)(17b)We can now form point estimates and interval estimates of the variance of each component of and using (13) and (14) respectively. Again, the main idea behind this test is that the estimated variances that are closer to the real variances of the measurement noise must come from the model that is more likely to have generated the measurements observed in the experiments. Moreover, if the real variances of do not lie in the interval estimates computed for a certain model, we can reject that model as wrong with a probability of , where is the confidence coefficient that was used for the test.
Note that the two estimates of the measurement noise (17) can also be formed by using the model solution. However, using the Kalman filter estimates of the states allows the procedure to be carried out even if the initial conditions are unknown.
The repressilator is a synthetic gene regulatory network, whose model is frequently used as an example for numerical algorithms [10],[27]. It consists of three genes connected in a feedback loop, where each gene transcribes the repressor protein for the next gene in the loop. The original model of Elowitz and Leibler [39] consists of six equations with four parameters, where all the three genes have the same production and degradation rates, and are affected in the same way by the corresponding repressor. Likewise, the three proteins have the same production and degradation rates.
In this example we consider a more general version of the repressilator, where each component is allowed to have different parameters. The model equations are as follows(20)for , with the convention that . The interactions of each gene/protein pair are characterized by 6 rates, therefore the total number of parameters to be estimated is 18.
We are assuming that we are able to measure the mRNA concentrations (), but not the protein concentrations (). We collect 30 equally spaced data points for each mRNA species. The noise in the measurements is assumed to have a power (i.e. variance) of of the mean of the signal. The parameters and the initial conditions to generate the simulated data are chosen so that the system displays a limit cycle behavior.
As in the large parameter space case for the heat shock model, a single run of the HEKF produces estimates that do not satisfy the identifiability test. Therefore, we apply the estimate refinement technique by minimizing (15) with . The results are presented in Figure 7. For the sake of brevity, we only show the and measurements. The measurement is presented in the supporting Figure S1.
We also compared the results of our method with a nonlinear Levenberg-Marquardt least-squares fitting and a genetic algorithm fitting directly on the data points. The results are summarized in Tables 5, 6 and 7. Only our method was capable of estimating a parameter set that was consistent with the simulated data in the sense of the test.
We have presented a novel approach for parameter estimation and model selection in computational biology. We have used this approach as a basis for a new algorithm for estimating parameters in models of biological systems from noisy and sparse experimental measurements. The approach is based on the combination of an extended Kalman filter algorithm, a statistical accuracy test, and a moment matching procedure. Furthermore, we have showed how the same tools can be used to discriminate among different candidate models of the same biological process. We have demonstrated the application of these ideas through two examples, a reduced order model of the heat shock response in E. coli and a generalized model of the repressilator (an additional example is available in the supporting file Text S1).
Parameter estimation using state observers in general, and the Kalman filter in particular, confers the significant advantage of fully exploiting the prior knowledge on the process that is encoded into the model. Observers are designed using the system's equations themselves, thus taking into account the system's dynamics. The Kalman filter has nice properties that are guaranteed to hold when the underlying dynamical system is linear and the noise statistics are Gaussian. In this case, the Kalman filter is the optimal state estimator, meaning that it produces the estimates with the smallest standard deviation of the estimation error. Even if the noise is not Gaussian, the Kalman filter is the optimal linear estimator. When the filter is extended for use with nonlinear dynamical systems through the time-varying linearization (10), such properties only hold in an approximate way, and one loses many of the theoretical guarantees that apply when the model is linear. However, in practice the extended Kalman filter has proven to be a successful choice in a wide range of applications, becoming the de-facto standard in nonlinear state estimation [29].
The Kalman filter approach to parameter estimation displays some features that make it particularly well suited to biological applications. For example, the hybrid Extended Kalman Filter (HEKF) is capable of estimating the parameters of continuous-time models with discrete-time measurements. This is important because most deterministic models of biological systems are continuous-time. However, most experimental techniques produce discrete-time data, often with large and non-uniform sampling intervals. The presented algorithm accommodates such situations without introducing any additional error due to a discretization of the system equations.
In spite of the above advantages, several challenges arise when using the Kalman filter for parameter estimation in a general nonlinear model. First, in the nonlinear setting, the Kalman Filter is not in general the optimal estimator. Moreover, if the initial estimates are too far off the filter may diverge, or converge to an estimate whose mean is different from the true mean. Additional factors can also be a source of error. State observers, as the name implies, were originally developed to estimate the state of a system – not its parameters. The state extension that becomes necessary to include the parameters into the estimation variables can introduce non-uniqueness of the solution (loss of observability), which can be problematic for the algorithms [29]. Furthermore, the covariance propagation equation in (6) is subject to numerical ill-conditioning, which can make the estimated error covariance matrices unreliable. These are some of the key reasons why the extended Kalman filter can produce unreliable estimates, and consequently, why a refined algorithm is needed for parameter estimation.
To alleviate some of the shortcomings of the HEKF in parameter estimation, we have proposed to augment the HEKF with an a posteriori statistical test and a subsequent optimization stage, both of which explicitly incorporate the information about measurement noise statistics into the estimation process. The test serves as a tool for the statistical reliability assessment of computed estimates, which validates the consistency of these estimates with respect to noise statistics. It also inspires a new technique for the discrimination between different candidate models for the same process. When the test shows that filter parameter estimates are inconsistent with the noise model, which can happen for any of the reasons mentioned in the previous paragraph, an estimate refinement step can become necessary. This takes the form of an optimization stage that begins where the HEKF left off. This proceeds until an estimate that satisfies the test is reached.
If the test for parameter estimates is the sole measure for accepting or rejecting a parameter estimate, then why not use it solely for parameter estimation by optimizing that measure directly, bypassing the Kalman Filter altogether? In the small parameter space case, numerical evidence suggests that if a unique solution to the parameter estimation problem exists, the HEKF is able to infer it with great speed and accuracy. This was seen in both the heat shock model and in the gene expression model, available in the supporting file Text S1. If the number of parameters is large and a good initial guess is not available, the HEKF is still able to run and provide a suitable initial guess for the subsequent refinement step, which can be expected to significantly reduce the running time of the moment matching optimization. Furthermore, the HEKF provides a computationally cheap algorithm, which scales much better than e.g. Bayesian methods and the particle filter. For these reasons, we believe that the HEKF represents a good choice as a first stage followed by moment matching optimization.
Coupling the Kalman filter with the statistical moment matching minimization presents a new way of thinking about optimization in parameter estimation. Classically, optimization is performed by trying to fit the model solution with the experimental data. While this is successful in some cases, it gives no guarantee that the parameters will produce a solution that is statistically consistent with the data. In the repressilator example, for instance, the classical least squares fitting produces for the state a variance that is too small compared to the one that was used to generated the simulated measurements (Table 6). In this situation, one runs into the issue of overfitting, in which the fitted model seems to replicate very well the behavior suggested by the data but fails to be robust to perturbations, so whenever it is used for further investigation, its behavior exhibits large inaccuracies. In contrast, the approach proposed here aims to match the mean and the variance of the measurement noise instead of the data points themselves, and is therefore able to “look beyond the noise” to recover the model parameters.
|
10.1371/journal.pgen.1000375 | The Individual Blood Cell Telomere Attrition Rate Is Telomere Length Dependent | Age-associated telomere shortening is a well documented feature of peripheral blood cells in human population studies, but it is not known to what extent these data can be transferred to the individual level. Telomere length (TL) in two blood samples taken at ∼10 years interval from 959 individuals was investigated using real-time PCR. TL was also measured in 13 families from a multigenerational cohort. As expected, we found an age-related decline in TL over time (r = –0.164, P<0.001, n = 959). However, approximately one-third of the individuals exhibited a stable or increased TL over a decade. The individual telomere attrition rate was inversely correlated with initial TL at a highly significant level (r = –0.752, P<0.001), indicating that the attrition rate was most pronounced in individuals with long telomeres at baseline. In accordance, the age-associated telomere attrition rate was more prominent in families with members displaying longer telomeres at a young age (r = –0.691, P<0.001). Abnormal blood TL has been reported at diagnosis of various malignancies, but in the present study there was no association between individual telomere attrition rate or prediagnostic TL and later tumor development. The collected data strongly suggest a TL maintenance mechanism acting in vivo, providing protection of short telomeres as previously demonstrated in vitro. Our findings might challenge the hypothesis that individual TL can predict possible life span or later tumor development.
| An age-dependent telomere shortening has been frequently observed in cross-sectional studies on human blood cells. Telomerase is an enzyme capable of lengthening telomeres, and it is activated in most tumor cells in order for them to become immortalized. This is one of the first longitudinal studies on telomere length, investigated in human blood samples taken at two occasions with approximately 10 years between them. An interesting finding was that the individual telomere length in the first blood sample was highly correlated with telomere attrition rate. Thus, individuals displaying the longest telomeres at baseline showed the most rapid telomere shortening over time and vice versa. This was also observed at the family level when exploring a multigenerational cohort. These results are in concordance with the fact that telomerase seems to preferentially act on the shortest telomeres in cultivated cells and provide fundamental knowledge for general telomere and cell biology. Because one part of the cohort developed tumors after the second blood draw, we had the opportunity to examine whether telomere attrition rate differed in tumor patients compared with controls, but no such indication was observed. However, for prostate cancer, short telomere length ≥9 years before diagnosis seemed to predict death.
| Telomeres are protective end structures of the chromosomes. Telomere length is dictated partly by hereditary [1]–[6] and partly by environmental [7],[8] and epigenetic factors [9]. The hereditary impact on TL has been estimated to range between 36–84% [3]–[6]. An equally strong telomere length inheritance was reported for monozygotic (MZ) as for dizygotic (DZ) twin pairs, indicating that the correlation in TL was mainly due to shared environmental factors [10]. In contrast, relatively minor environmental effects on TL during life were suggested in MZ twins where identical homologue telomeres differed less in TL compared to the two alleles within one individual [1],[2]. Regarding the influence of life style and environment on telomere maintenance, the published data are conflicting and no consensus has been reached concerning the impact of e.g. smoking, blood pressure or serum lipids on TL (literature overview in [11]).
Patients with smoking associated malignancies, such as human bladder, head and neck, lung, and renal cell cancers, have been shown to display shorter blood TL at diagnosis compared to controls [12],[13]. Short blood TL has therefore been suggested as a predisposition factor for these cancer types. For breast cancer, no difference in blood TL between patients and controls was found in one study [14], whereas we recently reported longer telomeres in peripheral blood cells of breast cancer patients and, furthermore, that long blood TL indicated a poor survival [15]. Numerous studies have shown an inverse correlation between blood cell TL and age [16]–[18]. Hence, it might be assumed that this characteristic is also true at the individual level. However, data are essentially lacking on individual telomere attrition rates and its relation to the occurrence of malignancy. Martin-Ruiz et al. did not find an association between telomere length at baseline and malignancy related mortality in a longitudinal study on individuals >85 years old [19].
In the present study, we have investigated individual blood cell telomere shortening in a large cohort of voluntarily donated samples. Our novel results show that the attrition rate was strongly correlated to telomere length at baseline, but unrelated to later tumor development.
In the study cohort of 959 individuals, investigated at two occasions with 9–11 year intervals, an overall TL shortening occurred with age as expected (r = −0.164, P<0.001) and women displayed longer telomeres than men (P = 0.052, after age-adjustments). However, about one third (34%) of all individuals demonstrated a stable TL or even elongated their telomeres over approximately a decade. There were very little differences between cases and controls (31.8% and 34.9%, respectively). When the individual telomere attrition per year was plotted against the relative TL (RTL) value of sample 1, a very strong and inverse correlation was found (r = −0.752, P<0.001; n = 959) (Figure 1A). This finding was also observed when analyzing tumor cases (r = −0.788, P<0.001; n = 314) and controls (r = −0.730, P<0.001; n = 645) separately (Figure 1B and 1C) To make sure that the strong correlation was not based on the very highest or lowest RTL values, the analysis was also made on individuals with RTL values <1 and >0.3. This subcohort showed the same statistical outcome (r = −0.623, p<0.001). Moreover, there was no gender difference (men: r = −0.781, p<0.001, women: r = −0.737, p<0.001). Hence, the attrition rate was most pronounced in individuals displaying the longest telomeres at baseline.
In a separate cohort of multigenerational families, we selected 13 families encompassing ≥10 members over at least three generations, and plotted RTL against age. A parallellity test revealed statistically significant differences in slope and intercept values between the 13 families (P<0.001), i.e. the families differed with regard to telomere shortening over time (Figure 2A). When the slope value from each family was plotted as a function of the corresponding intercept, it was found that RTL at young age (intercept estimated for the age of 14) was highly correlated with the telomere attrition rate (slope) (r = −0.691, P = 0.009) (Figure 2B).
RTL in sample 1 or 2 did not differ between cases and controls after age-adjustment (Figure 3), indicating that blood TL was not a potential biomarker for tumor development in prediagnostic samples. The telomere attrition rate was similar for cases and controls (P = 0.446 after age and sex adjustments) (Figure 4). As expected, prostate and breast cancer were the most common tumor types (the distribution of different types is given in Table 1), and due to the relatively short follow up time after diagnosis (0–113, mean 34 months) events of deaths were few. When all tumor cases were analyzed as one group, shorter than median RTL (for all tumor cases) in sample 1 (≥9 years before diagnosis) indicated a poor prognosis (not shown in figure). This is illustrated in the largest tumor group, prostate cancer (n = 81), where all deaths were found in the short RTL group in sample 1 (P = 0.004; cut off = median RTL value for prostate cancer cases) (Figure 5). When the same analysis was performed for RTL in sample 2 (collected 0–11 years before diagnosis) no significant prognostic difference was found between prostate cancer patients with long versus short telomeres (P = 0.174). Moreover, no association between telomere attrition rate and prognosis was found neither in the entire tumor group nor in the prostate cancer group (P = 0.266 and P = 0.889, respectively).
Age-associated telomere shortening is a generally accepted finding based on large cross-sectional studies. It has been assumed that this characteristic telomere attrition is true at the individual level as well. It has also been speculated that the telomere attrition rate in blood cells of tumor patients is higher compared to controls. The present longitudinal study cohort demonstrated the expected decline in telomere length by time, but we also observed large individual differences. Actually, in about one third of all individuals an elongation of the telomeres occurred over a decade. Most interestingly, individuals displaying the longest telomeres at the first blood draw demonstrated the most pronounced telomere shortening over time, and vice versa. The average coefficient of variation for the method was ∼6% and some of the variation in TL might be related to the technique itself. However, to make sure that the strong correlation was not based on the very highest or lowest RTL values, the correlation analysis was also made on individuals with RTL values <1 and >0.3. The result of the restricted correlation analysis was very similar to the result of the larger analysis, showing a strong correlation between telomere length at baseline and attrition rate.
A large variation in telomere attrition at the individual level has been observed in previous longitudinal studies on telomere length [19]–[21]. In a very recent study by Aviv et al. [21], TL was measured in leukocytes collected on two occasions from 450 whites and 185 African Americans, participating in the Bogalusa Heart Study. The median time period between the first and second blood sampling was shorter compared to our study (∼6 years vs. ∼10 years), and the participants were fewer and younger (age range: 20.0–40.0 years at baseline). Nevertheless, they found that the age-dependent TL attrition rate was proportional to TL at baseline, which is in accordance to our present study. The majority of participants in their study displayed TL shortening (85.9% of African Americans and 88.0% of whites), whereas the rest displayed a stable or increased TL. Similar to our observation, they also found that the rate of TL shortening varied considerably among individuals.
One explanation to the variations in attrition rate could be differences in epigenetic regulation, with secondary effects on telomere maintenance. Another reason might be that telomerase act preferentially on short telomeres, which has been shown in mice models and cell culture systems [22]–[26]. This would be in line with our observation of a very strong inverse correlation between individual TL at baseline and telomere attrition over time. Interestingly, we found a similar result in our separate family cohort. Hence, comparable data were obtained when analyzing telomere attrition rate at both the individual and at the family level. In the present study, individuals with the shortest TL actually elongated their telomeres over a decade, indicating that the TL maintenance machinery is focused on protecting the shortest telomeres. Nevertheless, other factors are likely to influence the TL attrition rate as well. In our study, the correlation value between blood RTL at baseline (sample 1) and attrition rate was r = −0.752 when analyzing the entire cohort. The corresponding r-squared value is hence 0.566, or ∼57%. This means that the telomere length at time point 1 could explain 57% of the variation in attrition rate. Thus, 43% of the variation might well be explained by other factors, such as life style, oxidative stress, inflammation etc. In the study by Aviv et al. [21], oxidative stress was proposed as a potential candidate for causing proportional telomere shortening. We agree that oxidative stress is likely to be important for telomere attrition, but the theory does not explain why a subset of the cohort demonstrated TL elongation. We suggest that a cellular TL regulating mechanism, rather than environmental/life style factors, is the major factor determining the rate of telomere attrition over time.
The working hypothesis that blood cell TL can indicate a later development of a malignant tumor was not supported in the present study. This hypothesis emanates from data showing altered TL in cases with a variety of malignancies. In urinary bladder, head and neck, lung, and renal cell cancers, shortened blood telomeres have been described at diagnosis, whereas data on breast cancer indicate unchanged or longer telomeres compared to controls [12]–[15]. Since no difference in TL existed between cases and controls, neither ≥9 (sample 1) nor 0–11 (sample 2) years before the appearance of a malignancy, we conclude that blood TL is not a prediagnostic biomarker for malignancy per se. However, our cases suffered from a variety of tumors and we cannot exclude that blood TL might be a biomarker for specific tumor types. A support for this is a recent study indicating that short blood telomeres were associated with a decreased risk for melanoma but also an increased risk for basal cell carcinoma, whereas there was no trend for squamous cell carcinoma [27].
In the largest tumor group in our material, prostate cancer, the blood TL ≥9 years before diagnosis seemed to indicate a poor prognosis. All prostate cancer cases with long blood telomeres (>median) were alive five years after diagnosis compared to <60% in the group with short telomeres. However, in the sample taken 0–11 years before prostate cancer diagnosis, TL did not give prognostic information. Thus, since the same prostate cancer patients were studied at different time points ahead of the cancer diagnosis, a shift had occurred indicating differences in the rate of telomere loss. This shift is interesting, since for breast cancer, blood TL at diagnosis seems to be a strong prognostic biomarker, with longer telomeres associated with a worse prognosis [15]. We have obtained similar data for renal cell carcinoma [Svenson et al., unpublished data] indicating that this feature is not unique for breast cancer. Unfortunately, in the present material the other cancer groups were too small to allow this type of statistical calculations. The biological background for these findings is unclear but it is tempting to speculate that factors responding to the presence of a tumor also have an impact on telomere maintenance, especially in immune reactive cells and their precursors. Due to few events in the cancer group a more definitive analysis of prediagnostic TL in relation to prognosis must await longer follow up times.
In conclusion, and similar to what have been observed in cultured cells in vitro, human blood cells in vivo seem to have a telomere maintenance system that gives priority to short telomeres. Based on human cross-sectional studies of age-associated telomere attrition, it has been speculated that TL at a certain age can predict a theoretical future life span. Our findings indicate that TL regulation through life might be more complex than previously known, complicating such life span predictions. We suggest that, at least in blood cells, the main TL regulator is a general mechanism that senses the telomere length similar to the counting mechanism demonstrated in cells from different species [28]–[31]. However, it might well be possible to avoid excessive telomere loss by living a healthy life as recently indicated [32]. Our data has important implications for our understanding of human telomere biology and for future analyses of telomere maintenance mechanisms in vivo.
The North Sweden Health and Disease Study (NSHDS) include the Västerbotten Intervention Project (VIP), launched in 1985 in the County of Västerbotten, the cardiovascular research program Monitoring of Trends and Determinants in Cardiovascular Diseases (MONICA) and the local Mammography Screening Project (MSP) [33]. At present the population based NSHDS collection contains samples from around 85000 individuals, nearly all Caucasians. Blood was drawn with anticoagulants, separated into plasma, erythrocyte and buffy coat fractions and stored at −80°C in small aliquots. In the NSHDS collection we identified >7000 individuals who had donated blood samples at a ∼10 year interval (9–11 years) and of these 343 persons had obtained a cancer diagnosis after the second blood sample (time from sample 2 to diagnosis: 0–11 years, mean 2.7) (Figure 6). From the same cohort, 686 age and sex matched controls were also selected. The age span was 30–61 years for sample 1 and 40–70 years for sample 2. Cancer cases were identified through record linkages with the regional Cancer Register. Due to insufficient amounts of buffy coat cells for DNA extraction or unsuccessful RT-PCR, 314 cases and 645 controls (totally 1918 samples) were included in the statistical analyses (cases: 176 men and 138 women; controls: 361 men and 284 women).
To permit analysis of a possible family linked pattern regarding TL attrition, a multifamily cohort was also utilized, initially aimed at studying genetic and environmental factors influencing heredity of personality traits, upbringing, general health and longevity (a study designed and conducted in the late 90's by the author RA). In total, whole blood was available from 962 individuals in 68 families (445 men and 517 women) with an age span of 0–102 years. Thirteen of these families could be selected for the purpose of this study (se statistics below).
The study was approved by the Umeå University Ethical Committee.
DNA was extracted from buffy coats and whole blood using conventional methods. Relative telomere length was measured using quantitative real-time PCR as described previously [34],[35]. In short, telomeres and a single copy gene (β2-globin) were amplified in all samples including an internal reference control cell line (CCRF-CEM) to which all samples were compared. The ΔΔCt method was used for calculation of RTL values and a standard curve was created in each PCR run to monitor the PCR efficiency. The mean inter-assay coefficient of variation for this method ranges between 4–8% in our laboratory.
Normality was shown regarding RTL distributions. Pearson partial correlation was performed to calculate age-adjusted correlations between continuous variables. ANCOVA was used for age and/or sex adjusted comparisons between groups. Cumulative survival for cancer patients with long vs. short telomeres was investigated using Kaplan-Meier with the log-rank test. Survival was defined as the number of months between diagnosis date to death or to last follow-up (Feb 2008).
To investigate whether the rate of telomere loss with age was linked to TL at a young age, 13 separate families in the multifamily cohort were studied. In each family, samples from 10 or more (maximum 28) related individuals, i.e. no in-laws, were available in at least three generations. The age of the individuals in the youngest generation varied between 14 and 32 years and in the oldest generation between 70 and 101 years. The number of men and women was similar within all families except for one which contained more women. The RTL values were plotted against age and linear regression was used to generate intercept (“starting RTL”) and slope (telomere loss) values for each family. The calculated intercepts corresponded to the estimated RTL value at the age of 14. The slope was then plotted as a function of the intercept and the correlation was examined using Pearson's Correlation Coefficient. MLwiN [36], a software for multilevel analysis, was used to test for parallellity between the 13 regression lines. All other statistics were analyzed in SPSS 15.0. A P-value ≤0.05 was considered to be significant.
|
10.1371/journal.pcbi.1003561 | The Spatial Resolution of Epidemic Peaks | The emergence of novel respiratory pathogens can challenge the capacity of key health care resources, such as intensive care units, that are constrained to serve only specific geographical populations. An ability to predict the magnitude and timing of peak incidence at the scale of a single large population would help to accurately assess the value of interventions designed to reduce that peak. However, current disease-dynamic theory does not provide a clear understanding of the relationship between: epidemic trajectories at the scale of interest (e.g. city); population mobility; and higher resolution spatial effects (e.g. transmission within small neighbourhoods). Here, we used a spatially-explicit stochastic meta-population model of arbitrary spatial resolution to determine the effect of resolution on model-derived epidemic trajectories. We simulated an influenza-like pathogen spreading across theoretical and actual population densities and varied our assumptions about mobility using Latin-Hypercube sampling. Even though, by design, cumulative attack rates were the same for all resolutions and mobilities, peak incidences were different. Clear thresholds existed for all tested populations, such that models with resolutions lower than the threshold substantially overestimated population-wide peak incidence. The effect of resolution was most important in populations which were of lower density and lower mobility. With the expectation of accurate spatial incidence datasets in the near future, our objective was to provide a framework for how to use these data correctly in a spatial meta-population model. Our results suggest that there is a fundamental spatial resolution for any pathogen-population pair. If underlying interactions between pathogens and spatially heterogeneous populations are represented at this resolution or higher, accurate predictions of peak incidence for city-scale epidemics are feasible.
| Fundamental spatial processes such as individuals' interactions and movement are not sufficiently well understood and yet they define the transmission of infectious diseases through populations. Spatial models of epidemics represent the region of interest (such as a city or country) as a collection of spatial units. To anticipate the magnitude and timing of peak incidence and to predict demand on health care resources in the region a clear understanding is needed of the relationship between the resolution of the representation (number and size of the pixels), the population interactions and the epidemic trajectories. We used a spatially explicit meta-population model of disease transmission to demonstrate that thresholds existed such that models with too low a resolution overestimated peak incidence, implying that ill-defined models may result in incorrect predictions. However, the results suggest that if population interactions are represented in sufficient detail, accurate estimates of peak demands on key health care resources are feasible.
| Novel respiratory pathogens continue to pose substantial public health challenges, not least because of the risk that large epidemics may overwhelm key health care resources such as vaccination stockpiles and intensive care facilities. Recent epidemics of concern include: SARS [1], influenza [2]–[4], H7N9 [5], [6] and MERS [7], [8]. During an epidemic it is important to accurately predict the impact of the epidemic over different spatial scales, where scale refers to the size of the region being monitored; such as a hospital, city, country or globally. Intervention policies should be defined relative to this spatial scale, for example taking account of how long it will take to vaccinate a whole city or to distribute a treatment country-wide. Those making decisions about intervention strategies need a clear understanding of the underlying epidemic process, so as to anticipate the magnitude and timing of peak incidence at their scale of interest and to effectively control the epidemic.
Spatially explicit transmission models are used frequently to increase understanding of the spread of epidemics caused by pathogens which transmit between individuals close in space. For example: influenza [9]–[11], measles [12]–[14], and smallpox [15], [16] have all been represented by spatially explicit epidemic models. All of these examples can be thought of as metapopulation models in which the population of interest is represented as a collection of sub-populations located in space, for example households [17]–[19], airports (GLEaM [20]) or districts/states [21]. The advantages of these models are that they can capture complicated mobility and mixing patterns and heterogeneous population density, without the complexity of an individual-based model. Also, model output can be easily reported for specific populations, such as counties or cities.
It is known that heterogeneity both in population density and typical mixing behaviour heavily influence disease spread. Both of these are defined according to the resolution of the population representation, where resolution defines the number and size of the pixels making up the “image” of the population within the model. A pixel is the smallest single component of an image. A high resolution representation will divide the region into many small pixels; a lower resolution uses fewer, larger pixels (Fig. S1). The resolution chosen is usually decided by the data available: for example, population and travel data may be defined at the ward or county level only. The level of mixing between individuals in distinct pixels is defined by mobility models, these are often fitted to travel data from censuses. Also, sometimes, resolution is limited by computational capacity.
The concepts in our paper require precise definitions of the terms: scale, resolution and pixel. The literature using these three words is somewhat ambiguous with the terms resolution and scale sometimes used interchangeably. Therefore, for clarity, we have included explicit definitions at first use of the words (above) and in Table 1.
We implemented a generic metapopulation model with arbitrary spatial resolution (see Methods) varying from approximately (30″ by 30″, the smallest unit representation) upwards (Fig. S1). We generated a theoretical population density in a region with total population just over 4 million and of size approximately (49×49 pixels). The region had three ‘urban’ areas where population density was generated using a 2-dimensional bivariate Gaussian and a ‘rural’ area, generated from a uniform distribution, Fig. 1G. We used this formulation to simulate the spread of a pathogen representative of influenza, with an SIR-like natural history, assuming that the generation time was 2.6 days and the basic reproductive number was 1.8. The epidemic was seeded with 10 individuals in a central region (Fig. 1G), simulations were repeated 25 times at each resolution.
The within-pixel contact rate was fixed for all pixels. Mobility between pixels was represented by a kernel with an offset power function. A kernel defines the relative probability of travelling between two pixels. The offset power function is an adaptation of the gravity model. The gravity model states that an individual's probability of mixing in a pixel different to their home pixel is inversely proportional to the distance apart of the pixels, to some power. The offset power function adds in an offset distance parameter, which means that pixels closer together than this distance mix fully. See Methods for a full definition of the kernel and the resulting mobility model.
Initially we considered three different kernels: we used an offset of 2 km and three different powers giving low, medium and high contact between pixels (the power, , was −6, −4 and −2 respectively), Fig. S2. The highest mobility kernel is in line with kernels fitted to commuter data in the UK and US [9]. However, our review of data on travel patterns found that only 15% of an average individual's journeys are commuting, making up just 19% of the total distance an average individual travels each year [22] (Table S1). Commuting data also excludes key at-risk groups – the under 17 s and over 70 s – who have lower mobility travel patterns compared to the 18–69 population [22]. Therefore, we explored more restrictive kernels than those estimated using commuting data to reflect shorter distances travelled and lower frequency travel in the most at-risk populations and the regular non-commuting travel of the wider population.
We confirmed that the overall cumulative attack rate (CAR) for our model was independent of the mobility kernel and the model resolution (Fig. 1A–F). This was by design: the model was constructed such that with the assumption of mass action mobility (the rate of contact between two groups is proportional to the size of each of the groups) the epidemic was identical at every resolution. This means that the next generation matrix at any resolution and for any mobility has the same spectral radius: was the same at all resolutions and contact levels and the local and global s were the same. A full proof that was constant with respect to resolution is given in the Text S1 in Supporting Information S1, and is similar to that in Ref [23]. Because was constant, if the mobility was such that there was contact between every pair of pixels, the final epidemic size was the same across all resolutions and in every pixel. If mobility was restrictive enough that some pixels were never infected the final CAR reflected this restriction. The full proof that attack rates were constant with respect to resolution is in the Text S2 in Supporting Information S1, and is similar to those in Refs [24], [25].
For the theoretical population density, the existence of a fundamental spatial resolution was apparent: at resolutions lower than this threshold, system-wide peak incidence was substantially over-estimated, obtaining a high peak incidence and fast spread similar to that obtained in a fully mixed model (the lowest resolution). However, at the fundamental resolution and above, consistent estimates of the peak attack rate were obtained (Fig. 1H). This was increasingly evident as mobility became more and more restricted: for the most localised mobility assumptions (low power), peak incidence in the fully mixed case was nearly double that at the highest resolution. At high resolutions, multiple small pixels containing low numbers of individuals and with a high heterogeneity in population size slowed the epidemic spread; resulting in a long epidemic duration and a low peak incidence compared to low resolution model scenarios.
Increased mobility reduced the effect of resolution on the epidemic trajectory. At medium mobility, peak incidence increased with decreasing resolution but there was no distinct threshold. At the highest mobility, peak incidence was unaffected by resolution: the high level of contact between pixels facilitated the quick spread of the epidemic, indicated by a short epidemic duration and a high peak incidence at every resolution (Fig. 1A–C).
Resolution and mobility remained important when the model was constructed with real population densities. We repeated the analysis (using the same three kernels) for four regions selected from LandScan data [26]: Guangzhou, Rio de Janeiro, Delhi and New York (Fig. 2A). The smallest LandScan unit is approximately 1 km2 (30″ by 30″) in size. The effect of resolution was most evident when mobility was more restricted, as with the theoretical population. In Guangzhou, Rio and New York, changing the spatial resolution had a significant effect on the peak incidence when mobility was at a low to medium level, though the effect was less clear in Delhi (Fig. 2B–E). The Delhi region had the largest total population size and the highest mean population density of all regions we considered (Table S2 and Fig. S3). Therefore, even at low mobility the numbers mixing will be relatively high, meaning that the disease spread will not be as restricted as it would be in a less densely populated region.
We used Latin Hypercube Sampling (LHS) [27] to determine whether the patterns we saw with the illustrative mobility kernels could be generalised within a wider parameter space of mobility functions. We varied the kernel parameters: the power, , between −6 and −2 (as discussed earlier, this selection gave a wide range of mobility levels) and the saturation distance between 1 and 10 km, choosing from a log scale (so smaller distances are more likely). We tested 50 parameter sets chosen using the LHS technique [27], with 10 separate realisations of each set for each region (variation in results from the stochastic model was low - see confidence intervals for 25 repeats in Fig. 2 for example). Kernels for the 50 sets are plotted in Fig. S4. The LHS results confirmed that the effect of resolution is most important in populations which are less mobile, Fig. 3. As mobility decreased (a combination of the offset and the power in the kernel) the difference in peak incidence between the lowest and the highest resolution increased. This was particularly true in Guangzhou, Rio and New York, but in Delhi the effect was reduced (due to Delhi having a very large population in comparison to the other regions).
Recently it has been suggested that the movement of individuals depends not only on the source and destination cities, but also on the population density of the surrounding area [28]. This model is called the radiation model and has been proposed as a distinct alternative to the gravity model. However, we calculated the actual number of individuals moving between pixels (the flux) and found the radiation model flux to be very close to the offset gravity model of medium mobility, particularly at the highest resolution, Fig. 4. Indeed the radiation model is always bounded by the three gravity models we use and our LHS models explore a large space around these. More generally, gravity-like models have been implemented with a number of different normalisation assumptions, some of which produce population flux patterns very similar to the radiation model [15], [17].
When managing epidemics it is desirable to know the size and duration of the epidemic and the magnitude and timing of the peak incidence over the spatial scale of interest [4], [29], [30]. This scale of interest may be a city, a region or a whole country. Resources such as treatment, vaccinations and diagnostic tests will take time to be deployed over this scale and it can take time to develop and generate enough of these resources for the whole affected population [31], [32]. Accurate predictions about the magnitude and timing of peak incidence would greatly enhance the ability of public health officials to effectively limit the impact of epidemics.
We have shown how the representation of population interactions can impact model estimates of key epidemic outcomes. We examined the effect of the resolution of the population density on the model predictions of epidemic spread over the scale of interest. We refer to resolution as defining the number and size of the individual pixels dividing the region; higher resolution representations use a higher number of smaller pixels. Our results imply that for plausible population densities and mobility patterns, fundamental resolutions exist for specific pathogens such that the detail of the population and their interactions must be represented faithfully if accurate epidemic trajectories are to be estimated.
The impact of model resolution was clear in models of less mobile populations: our results indicate that at lower mobility, low resolution representations overestimated the peak incidence, obtaining a high peak incidence and fast spread similar to that obtained in a fully mixed model. However, sufficiently high resolution representations gave lower and later peak incidences because of the delaying effect of multiple small pixels. Indeed at low mobility, clear thresholds existed for the resolution of the theoretical population density, such that models with resolutions below the threshold over-estimated the system-wide peak incidence. Similar thresholds existed for real population densities: Guangzhou, Rio, New York and Delhi. Increased mobility reduced the effect of resolution on the epidemic.
The kernels which were most affected by resolution were those which gave a lower mobility than that identified by commuting data (Table S1). Generally children are considered to cause the majority of transmission of pathogens like flu and measles, because their level of age group assortative mixing is very high [33], [34]. Children also travel less far than working adults [22]. Together, these imply that a kernel for children is likely to be more restrictive than those defined by commuting data alone. Therefore, our results indicate that the correct specification of population interactions and sufficient spatial resolution is particularly relevant for epidemics such as measles and flu - those in which children play a large role.
Although we have considered age effects implicitly by including lower mobility levels than are reported for commuting data (Table S1), the explicit representation of age within a similar modelling framework may lead to additional insight. For example, transmission dynamics at different scales may be driven by different age groups: the behaviour of more mobile adults may be disproportionately important in the seeding of nearby pixels. However, the slower than expected within-country spatial spread during 2009 [35] suggests that for pandemic influenza, population sub-groups with reduced mobility likely do define the fundamental resolution.
We have chosen to represent the real biological process by a high resolution metapopulation model. Although we have not been able to push the model to resolutions higher than 1 km by 1 km, we suggest it is reasonable to assume, for the mobility kernels considered here, that the thresholds observed for peak incidence would not change substantially were we to approach the resolution of an individual-based model.
The model used here was intended specifically to test only the changing resolution of the disease transmission process. By design we did not want to assume that transmissibility was intrinsically higher or lower in different parts of the population. In future work, we hope to calibrate this model structure using actual disease incidence data and (after a minor modification to the definition of the force of infection) test for the possibility that population density affects transmissibility.
Although it is somewhat reassuring that estimates of peak incidence are biased upwards if resolution is too low, the epidemic duration is underestimated. In order to avoid the effects of incorrect model specification, where possible, spatial resolution should be treated in a similar manner to temporal resolution in fixed-time-step models: neither the doubling nor halving of spatial resolution should have a substantive effect on key model outputs.
We defined a spatially explicit meta-population model as follows (similar to Ref. [21]). A given region of known population density was represented as pixels, such that each pixel (index ) is the same spatial size but the number of individuals in the pixel () varied according to location. Mixing between and within each pixel was determined by a mobility model, represented by a matrix such that an entry was equal to the probability that for an individual from pixel , given that the individual made a contact, this contact was with an individual from pixel (mobility was defined using a kernel, discussed later).
The rate at which susceptible individuals in pixel became infected depended on (1) their risk of infection from those in pixel , (2) the risk of infection from infected individuals in pixel who travelled to , (3) the risk of infection that susceptible individuals from encountered when they travelled to . Therefore, the force of infection or the average rate that susceptible individuals in pixel became infected per time-step was:(1)where was the total number of pixels and for any pixel , was the number of individuals, was the number of infected individuals and infectious contacts were made with other individuals present in the pixel with rate . Note that is the same across all pixels; in future it may be of interest to vary the transmissibility across pixels (so moves into the sum in Eqn. (1) as ).
The system of difference equations for a pixel in the stochastic SIR model was (with the condition that all classes hold a whole number of individuals):(2)where was the number of individuals in pixel , state (S, I or R) that experienced the event – infection or recovery – in time-step . We ignored death in this model as we considered fairly short timescales and a non-fatal strain of influenza.
Each time-step , the number of individuals experiencing each event () that occurred in pixel and state , with a population was determined in the following way:
We used a mobility model to determine the relative frequency of potentially infectious contact. This was represented as a matrix with entries , defined as the probability that for an individual from pixel , given that the individual made a contact, this contact was with an individual from pixel , so:(3)where was the total population in pixel , was the interaction kernel defining the effect of the distance between pixels and on the contact between them. The kernel defines the relative probability of travelling between two pixels and not the absolute flux, similar to [15], [17]. The factor normalised and ensured that the rows sum to 1. The matrix was used in the calculation of force of infection, Eqn. (1).
We used a variation of the offset power function for the kernel (similar to [9], [15]):(4)where was the distance below which the kernel function saturated, we used . The power determined the mixing between pixels, this was varied to give a range of mobilities but was always less than 0.
The next generation matrix, , for the model with pixels and force of infection (Eqn. (1)), can be defined (similar to [21]):(5)where was the time spent infected (which depended on recovery rate such that , same for all pixels), was the number of individuals in pixel , infectious contacts were made with other individuals present in the pixel with rate and was the mobility matrix defined earlier. Then was equal to the spectral radius of this matrix [36], [37]. For this model, , i.e. was independent of resolution and mobility; see Text S1 in Supporting Information S1 for full derivation.
|
10.1371/journal.pgen.1007430 | Drosophila species learn dialects through communal living | Many species are able to share information about their environment by communicating through auditory, visual, and olfactory cues. In Drosophila melanogaster, exposure to parasitoid wasps leads to a decline in egg laying, and exposed females communicate this threat to naïve flies, which also depress egg laying. We find that species across the genus Drosophila respond to wasps by egg laying reduction, activate cleaved caspase in oocytes, and communicate the presence of wasps to naïve individuals. Communication within a species and between closely related species is efficient, while more distantly related species exhibit partial communication. Remarkably, partial communication between some species is enhanced after a cohabitation period that requires exchange of visual and olfactory signals. This interspecies “dialect learning” requires neuronal cAMP signaling in the mushroom body, suggesting neuronal plasticity facilitates dialect learning and memory. These observations establish Drosophila as genetic models for interspecies social communication and evolution of dialects.
| In this study, we find that many different Drosophila species never having been exposed to parasitoid wasps can trigger caspase activation in the ovary and depress egg-laying when placed next to flies that had visual experience with wasps. Interestingly, when teacher flies of one species are placed with a student of a different species, communication exists, to varying degrees, which seems dependent on evolutionary relatedness. Cohabitation of two species that can partially communicate can learn each other’s “dialect”, yielding effective interspecies communication. There are various inputs involved in dialect learning, including the presence of visual and olfactory cues and memory functions, including genes implicated in social learning defects in murine models, such as PTEN. The neuroplasticity of adult Drosophila allows for learning of dialects, but the specific dialect learned is dependent on social interactions exclusive to a communal environmental context, which provides both visual and olfactory inputs. We find flies can communicate with one another about an anticipated danger, which is suggestive of a fly “language.” The presence of a neurologically plastic system, allowing for social learning, can subsequently lead to a dramatic physiological response, requiring active learning and memory formation through integration of multiple inputs.
| The ability to interpret environmental information is a phenomenon found throughout all life forms. From bacteria to plants and to mammals, communication occurs within as well as between species. In some cases, information that is being shared can be highly specific, such as in the case of honeybees communicating instructions on where to find nectar[1–3]. In other cases, opportunistic bystanders can also benefit from general information. For example, predator alarm calls generated as a warning are observed, where multiple species participate in repeating the alarm throughout the community[4–8]. In all cases, the information that is shared can be dependent on local environmental cues and experiences and the manner in which information is communicated is strongly influenced by past experiences of each individual. For example, birds, which live in geographically distinct populations, manifest unique song variants or regional dialects that can last for decades, but these animals are nevertheless still able to communicate with others of their species[9–11]. Because dialects are learned and therefore influenced[12] by specific local environmental differences, it suggests that both social and non-social experiences can have dramatic effects on cognitive development[13].
It is proposed that a myriad of environmental cues, both social and non-social, are critical to animal development in determining the ability to convey and receive specific types of information. However, there are many outstanding questions as a result of this proposition: What cues are important? When are these cues important? How can environmental cues interact with genetically determined developmental programs? Although social communication is most extensively documented in more derived species such as mammals and birds, insects can also display a broad range of behavioral tasks. Bees are known to be able to learn from non-natural sources in order to obtain a reward through social learning. Such information can be passed on to naïve, student bees through the use of visual cues[14,15]. Insect social learning extends to the genetic model system of Drosophila, where student, observer flies learn from a trained, teacher-fly, using visual cues. This has been shown in communication involving food sources and predator threats[16,17].
Chemical cues can serve as intra- and interspecies signals, such as fox and guinea pig urine affecting not only conspecific behavior, but also the behavior of other animals[18–20]. Sound can also be used, such as in bats and bottlenose dolphins, which are able to distinguish members of the community through the use of echolocation pitch recognition[21,22]. Plants have a vast arsenal of responses to pathogens[23], including communicating a threat to neighboring plants through the use of volatile organic compounds[24]. Plant interspecies[25–31] and intraspecies[32–34] communication occurs both in laboratory settings and in the wild[30,35].
Drosophila melanogaster and other Drosophila species have provided insights into mechanisms of learning, memory, and complex behaviors[36,37]. However, these behaviors and phenotypes have been studied almost exclusively in domesticated D. melanogaster lab monocultures, while D. melanogaster wild populations are surrounded by a broad range of predators, microbes, and other Drosophilids, highlighting a communal component of the organism’s life cycle[38]. This raises the possibility of behavioral phenomenon that have yet to be discovered and analyzed in domesticated lab monocultures[39–41]. Given the vast range of environmental inputs on a wild Drosophilid, a fly must be able to discern important information from extraneous inputs, while interacting with conspecifics and a variety of other species [42–46].
Although modes of intra- and interspecies communication are likely to be genetically limited, there is also value in learning to interpret signals from variable, local environments that may provide immediate survival benefits. How do genetically constrained neurological features and variable environmental factors interact to produce context-dependent, meaningful information? Under which environmental factors would information sharing between different species occur and be beneficial? In this study, we sought to begin to address these questions in the Drosophila model system by using a pan-Drosophila predator known to elicit social communication [17,47]. D. melanogaster presented with parasitoid wasps have multiple behavioral responses, including a reduction in oviposition (egg laying) through an increase in ovarian apoptosis [17,48–51]. After removal of the wasp, a wasp-exposed “teacher” fly can instruct a naïve “student” fly about the presence of the wasp threat through the exclusive use of visual cues, such that students now reduce their own oviposition by triggering ovarian apoptosis. Using this fly-fly social communication paradigm we asked (1) whether social communication is conserved among other Drosophila species, (2) if Drosophilids engage in interspecies communication, and (3) what environmental and genetic factors are required for interspecies communication.
We utilized the fly duplex, an apparatus with two transparent acrylic compartments to test whether different species respond to seeing predators (acute response) and if exposed “teacher” female flies can communicate this threat to naïve unexposed “student” female flies[17]. The duplex allows flies to see other flies or wasps in the adjacent compartment, without direct contact, making all communication only visual (Fig 1A). Ten female and two male flies are placed into one duplex compartment, with an adjacent compartment containing twenty female wasps. Following a 24-hour exposure, wasps are removed and acute response is measured by counting the number of eggs laid in the first 24-hour period in a blinded manner. Flies are shifted to a new duplex, with ten female and two male naïve student flies in the adjacent compartment (Fig 1A, see Methods). Following a second 24-hour period, all flies are removed and the response of both teacher and student is measured by counting the number of eggs laid in a blinded manner. The 24-48-hour period measures memory of teachers having seen the wasps and students having learned from the teachers. Using wild-type D. melanogaster, we find both an acute response and a memory response to the wasp in teacher flies and a learned response in naïve student flies (Fig 1B, S1A Fig, S1 File for all raw egg counts and p-values in this study) [17,50,51].
We then asked whether the acute, memory, and student social learning behaviors are conserved in other Drosophila species, with varying relatedness to D. melanogaster ranging from sister species, such as D. simulans, to very distantly related species, such as D. virilis. For each species, we tested a sister species as an additional way to validate our observations. Across a broad span of the genus Drosophila, we find the conservation of both the acute and memory responses in teacher flies in addition to the ability of teachers to communicate to conspecific student flies. (Fig 1C, S1B–S1H Fig). Some of these species have been previously shown to depress oviposition during wasp exposure [51]. Our experimental design allows for only visual cues to be detected from the wasps and from teachers to student flies. Thus, in all species tested, visual cues are sufficient for flies to detect wasps and for naïve flies to learn from wasp-exposed teacher flies. Conservation of these behaviors is especially impressive as the species tested are separated by millions of years of evolution, yet the basic behaviors observed in D. melanogaster are maintained. Moreover, this conservation further underscores the importance this innate behavior must have since even laboratory cultures that have not experienced wasp for many generations nevertheless exhibit a robust response. In particular, the conservation of the fly-fly communication behavior speaks to a presence of a conserved form of fly signaling and signal interpretation, which we suggest might be thought of as a “fly language” in this paradigm.
Oviposition reduction is modulated in part by the effector caspase Dcp-1[17]. In D. melanogaster, we observe overlapping staining of activated Dcp-1 with a punctate pattern of DNA staining with 4’, 6-diamidino-2-phenylindole (DAPI), indicative of oocyte specific apoptotic activity (Fig 1D–1K, S2 Fig). We performed immunofluorescence with antibodies specific to activated Dcp-1 across a broad range of Drosophila species, revealing cleaved caspase following wasp exposure in all 15 Drosophila species tested (S3–S16 Figs). We observed an increase in positive cleaved caspase oocytes following wasp exposure (S17 Fig), along with a decrease in total number of egg chambers (S18 Fig), suggestive of ovarian apoptosis and elimination of oocytes[17]. Phylogenetic trees shown are adapted from previous work [52].
Following the observation that an acute response to wasps and intraspecies communication is conserved across the genus, we asked whether the wasp threat could be communicated between two different species. We utilized 15 Drosophila species that respond to wasps to answer this question (S3–S18 Figs). The species were selected to span the phylogeny with different degrees of relatedness to D. melanogaster [52]. We find that D. melanogaster are able to communicate the threat to and receive communications from closely related species, such as D. simulans and D. yakuba, with oviposition of students paired with wasp-exposed teachers being ~10–30% compared to unexposed (Fig 2A and 2B, S19A–S19F Fig). Interestingly, species more distantly related to D. melanogaster, such as D. ananassae and its sister species, elicit a partial communication phenotype, with oviposition depression of students paired with wasp-exposed teachers being ~50–65% of unexposed flies (Fig 2C–2F, S19G–S19J Fig). A second strain isolate of D. ananassae also show partial communication with D. melanogaster (Fig 2C–2F, S19G–S19J Fig). Species more distantly related to D. melanogaster, such as D. willistoni, D. equinoxialis, and D. virilis, cannot communicate with D. melanogaster (Fig 2G–2L, S19K–S19P Fig). We statistically characterized these category assignments based on the criteria of mean value and statistical significance compared to unexposed in order to define efficient, partially, and lack of communication (S2 File, Methods). Collectively, the data suggest that evolutionary distance contributes to the efficiency of interspecies communication. D. ananassae show varying communication phenotypes with other Drosophila species, though the pattern of communication is different. For example, D. ananassae exhibit partial communication with D. simulans (S20A and S20B Fig), strong communication with its sister D. kikkawai (S20C and S20D Fig), and partial communication with D. equinoxialis and D. willistoni (S20E–S20H Fig). D. ananassae, in addition to D. melanogaster, are unable to communicate with the distantly related D. mojavensis and D. virilis (Fig 2I and 2J, S20I–S20L Fig). Species such as D. virilis, which were unable to communicate with D. melanogaster and D. ananassae, can communicate with other species, such as its sister species D. mojavensis (Fig 2K and 2L). Thus, although all species tested are capable of intraspecies communication, there is a fundamental, species-specific difference in communication mode or “fly language” when communicating wasp predator threat.
We wished to assay whether the observed communication behavior is hardwired in the fly brain, or if it had a level of plasticity as a result of socialization. Known learning and memory mutants have shown defects in socialization assays [53,54]. Additionally, the cuticular hydrocarbon composition on flies changes as a function of social, but not sexual, experience [55], though sexual experience is also affected by isolation [56]. Thus, we sought to assay whether socialization, which has been shown to have behavioral affects in other assays [57] including egg laying behavior [44,58], has an effect on intraspecies communication.
To ask this question, we collected L1 larvae and isolated each larva in a Falcon round-bottom polypropylene tube containing 1 mL standard Drosophila media. Larvae were allowed to pupate and eclose in isolation. Each tube was kept separate such that no visual information could be transferred between individuals in tubes. Following eclosion, 1 female aged 3–5 days old was used as the student, paired with one socialized female teacher (Fig 3A). This 1:1 ratio was first tested with D. melanogaster where both teachers and students were previously socialized, observing typical strong communication (Fig 3B). Interestingly, the flies raised in isolation presented with a partial communication phenotype, similar to when normally socialized D. melanogaster and D. ananassae are paired (Fig 3C). Larval isolation has been previously shown to have effects on cooperative larval behavior, and thus, we cannot rule out the possibility that isolation of larvae translates to behavioral defects in adult flies [59]. However, given the observation of this partial communication phenotype, we suggest that while the ability to communicate is hardwired in the fly brain, there exists a degree of plasticity that is dependent on previous socialization.
Given that our isolation experiments demonstrate a level of plasticity dependent on socialization, we wished to explore the possibility that partial communication between species might be alleviated as a result of socialization between two different species. Since closely related species can communicate the threat of a wasp, we postulated that environmental factors contributing to interspecies communication for distantly related species may be partially dependent on socialization. To test this idea, D. melanogaster were cohabitated with species capable of only partial communication (e.g. D. ananassae) (Fig 2C and 2D). Cohabitation lasted for one week in a single container, allowing for frequent and multiple channels of sensory interactions. Following a weeklong cohabitation period, the two species were separated and used as students paired with teachers of the other species (Fig 4A). In all experiments teachers had existed only as a monoculture, kept separate from other species monocultures, while all flies experiencing an interspecies cohabitation period were subsequently used only as students.
If cohabitation of different species results in exchange of information that later facilitates communication between these two species, then we predict that species capable of only partial communication may then be capable of full or enhanced communication. We find that cohabitation can greatly enhance communication between some species, suggesting that some form of training occurs during this period. After cohabitation, D. ananassae students learn very efficiently from D. melanogaster teachers, demonstrating that cohabitation of two species yields an expanded communication repertoire (Fig 4, S21 Fig). This observation indicates that poorly communicating species are not limited by structural barriers such as wing shape or olfactory capacity. Instead this suggests that, similar to local dialects in bird songs, Drosophila species-specific cues can be learned simply by repeated exposure to the “dialect”, and provides further evidence for the role of socialization in Drosophila communication (Fig 3). Thus, there exists a variation of signal among populations of different species of Drosophila, even though there exists a conserved fly “language” to communicate the threat of a wasp. Thus, we suggest this to be analogous to “dialects” as it reflects natural variations between a common mode of communication, which can be alleviated through socialization between species. Hereafter we refer to this cohabitation as a “dialect learning” period.
We observed dialect learning in two different D. ananassae strain isolates, and two additional sister species (Fig 4B–4E, S21A–S21G Fig), indicating that dialect learning is likely to be a wide-spread phenomenon in Drosophila. Interestingly, some distantly related species that were unable to communicate with D. melanogaster (i.e. D. willistoni, D. equinoxialis) acquired the ability to partially communicate following a cohabitation-training period (Fig 4F–4I, S21H and S21I Fig). This was not the case for very distantly related species (i.e. D. virilis, D. mojavensis), which showed no ability to communicate with D. melanogaster even after a week-long cohabitation (Fig 4J and 4K, S21J–S21M Fig). We also tested a transgenic D. melanogaster, to see if it was capable of teaching and dialect learning, and find such flies can teach their dialect to and learn the dialect from D. ananassae (S21N and S21O Fig).
Additionally, we tested whether D. ananassae communication could benefit from cohabitation-training with species other than D. melanogaster. We find efficient communication between D. simulans (S22A and S22B Fig), D. equinoxialis (S22C and S22D Fig), and D. mojavensis (S22E and S22F Fig) with D. ananassae following a cohabitation-training period. In contrast to the D. melanogaster results, we find communication with more distantly related species is altered after dialect training. With D. virilis and D. mojavensis, in the untrained states, we observe no ability to communicate (S20I–S20L Fig), but find a partial communication phenotype following cohabitation (S22G–S22J Fig). D. virilis and D. mojavensis, although capable of interspecies communication and dialect learning, cannot learn the D. melanogaster dialect, but can learn D. ananassae dialect. These results suggest that some interspecies communication barriers do exist while others can be overcome by a period of dialect training during cohabitation.
Given our observation that two species can learn dialects following a cohabitation-training period, we wondered whether having more species present during the dialect training period influences dialect learning. In nature, flies encounter many different species of Drosophila, and given this knowledge, we hypothesized that neuronal plasticity exists in the fly brain to allow flies to learn multiple dialects from a given training period that includes multiple species as inputs. To probe this question, D. melanogaster were cohabitated with species capable of only partial communication or no communication in the untrained state, but show efficient and partial communication after dialect training (i.e. D. ananassae and D. willistoni, respectively). These three species were cohabitated for one week in a single container. We then used the trained D. melanogaster as students with untrained D. ananassae and D. willistoni teachers (Fig 5A). We find that trained D. melanogaster are able to efficiently communicate with D. ananassae and partially communicate with D. willistoni (Fig 5B and 5C). These results mirror assays where these species were individually trained (Fig 4B, 4C, 4F and 4G), suggesting that flies can simultaneously make use of multiple inputs from multiple species and be able to learn and remember each unique dialect they encounter. Additionally, we tested D. ananassae and D. willistoni as students that were cohabitated with D. melanogaster. We find that D. ananassae can communicate efficiently with D. melanogaster and D. willistoni (Fig 5D and 5E), and that D. willistoni can partially communicate with D. melanogaster and effectively communicate with D. ananassae (Fig 5F and 5G). These data also mirror individual training (Fig 4B, 4C, 4F and 4G, S22E and S22F Fig). Collectively, these data demonstrate that a fly can have vast communication repertoires consisting of multiple dialects that it acquires.
Given the result above with multiple species being able to learn multiple dialects, we wondered the level of specificity and the level of generalization of dialect learning as a means to provide insight into the identity of the “signal” being transferred between species. To test this, we performed cohabitation of D. melanogaster and D. kikkawai, a sister species to D. ananassae. We then assayed the communication ability of D. melanogaster with either D. ananassae or D. willistoni (S23A Fig). We find that D. kikkawai trained D. melanogaster are able to effectively communicate with D. ananassae, suggesting that there is a generalizability to dialect learning (S23B Fig). We tested the ability of these flies to communicate with D. willistoni, as D. ananassae has an ability to communicate with D. willistoni in the naïve state, while D. melanogaster does not, allowing further analysis into the generalizability of the signal. We find that D. kikkawai trained D. melanogaster are unable to communicate with D. willistoni, suggesting that while dialect learning is generalizable in some instances, it also has a layer of specificity (S23C Fig).
In order to better understand dialect learning, we tested the roles of sensory cues and genetic factors during the dialect learning period. We measured dialect learning by quantifying improvement in interspecies partial communication between D. melanogaster and D. ananassae that normally exhibit efficient communication only after cohabitation. Given that in D. melanogaster, and in other species tested, we found visual cues to be sufficient for the teacher-student dynamic (Fig 1) [17], we asked if visual cues are sufficient and/or necessary for dialect learning. We approached this question by performing the dialect training in the fly duplex, such that the two species could only see each other (Fig 6A), or by performing the training in the dark, so that the two species could physically interact, but lacked visual cues (S24A Fig). We find that visual cues alone are not sufficient (Fig 6B and 6C), but are necessary (S24B and S24C Fig) for dialect learning. The observation that visual cues are necessary but not sufficient makes the dialect learning process different from the teacher-student dynamic that requires only visual cues[17]. Furthermore, we wondered if seeing another species altered the behavior of a fly to facilitate dialect learning. This hypothesis addresses the possibility that flies are passively acquiring information through eavesdropping and that the communication ability gained could be unidirectional. Blind D. melanogaster ninaB mutants do not function as students. Surprisingly, D. ananassae cohabitated with blind D. melanogaster do not learn the D. melanogaster dialect (S24D and S24E Fig). This result is striking because it suggests that there is an active learning component and a bidirectional exchange of information between fly species and not simply eavesdropping or mimicry. We also performed cohabitation training under two different, monochromatic light sources, and this resulted in only a partial communication between D. melanogaster and D. ananassae, (Fig 6D and 6E, S24F and S24G Fig). To exclude the possibility of a dimmer light source inhibiting dialect training under monochromatic settings, we repeated cohabitation-dialect-training in a full spectrum, lower light intensity setting, and found both species were able to learn the dialect (S24H and S24I Fig). Thus, full spectrum light is essential in dialect learning. Importantly, the observation that blind D. melanogaster do not allow wild-type D. ananassae to dialect learn suggests that visual inputs are critical to altering behavioral/chemical outputs required to facilitate dialect learning. This also suggests that during the dialect learning period, transfer of information may occur bidirectionally, if the visual input that is required is indeed provided by the other species.
Wing movement was shown to be required for teacher flies to instruct students in the teacher-student dynamic[17], raising the possibility that wing movement was also important for dialect learning. Therefore, we tested flies mutant in the erect wing gene (ewg), which impairs wing movement while maintaining morphologically normal wings. The allele tested has wild-type EWG protein expression in the nervous system, thus is only deficient in its non-neuronal functions, such as flight muscles [60]. We find that D. ananassae cannot dialect learn from ewgNS4 flies (Fig 6F), although ewgNS4 mutants have no dialect learning impairment (Fig 6G). This suggests that dialect learning by D. ananassae requires D. melanogaster to have mobile wings.
To test if olfactory cues play a role in dialect learning, we utilized D. melanogaster mutants defective in chemosensory signaling. The majority of olfactory receptors require a co-receptor for wild-type function, including Orco (Or83b) for odorant receptors [61] and Ir8a or Ir25a for ionotropic receptors [62]. Ir8a olfactory sensory neurons (OSNs) primarily detect acids and Ir25a OSNs detect amines, allowing us to probe specificity of detection. We find that D. ananassae are able to learn dialect from Orco1, Ir8a1, Ir25a2, single and Ir8a1;Ir25a2;Orco1 triple mutants and RNAi expressing D. melanogaster targeting each of these gene products (Fig 6H and 6J, S25A–S25L Fig). By contrast only Ir25a2 mutant and RNAi knockdown D. melanogaster were able to learn the D. ananassae dialect (Fig 6I and 6K, S25A–S25L Fig). These data demonstrate that Orco- and Ir8a-mediated olfactory inputs are required for dialect learning. This further suggests that multiple olfactory cues play important roles in the dialect learning period. We also find that D. melanogaster males and females are both required for dialect training D. ananassae (Fig 6L–6M, S25M and S25N Fig) and that the length of the training period is also critical, as 24 hours is insufficient a period for dialect learning (S25O and S25P Fig). Thus, although the exact olfactory molecule(s) critical during a dialect learning period are yet to be identified, we speculate that dialect learning is a complex process requiring visual, olfactory and sex specific cues.
To examine the possibility that dialect training involves active learning mediated by neurons of the mushroom body, we utilized the GAL4 Gene-Switch system to transiently express a transgene specifically in the mushroom body (MB). Using the GAL4 Gene-Switch ligand system, RU486 [63] activates the GAL4 transcription factor, while administration of the vehicle (methanol) does not [63]. RU486 was administered during the cohabitation period (or methanol for control), but not when flies were used as students, post-dialect training (Fig 7A). Feeding of RU486 to the MB switch driver line does not impair dialect learning (S26A Fig). We expressed the Tetanus toxin light chain (UAS-TeTx) specifically in the MB of D. melanogaster (to inhibit synaptic transmission during dialect training). We find that D. ananassae are able to learn the dialect of these MB inhibited flies (Fig 7B). However, D. melanogaster in which MB synaptic transmission is inhibited during the training period are unable to learn the D. ananassae dialect (Fig 7C). Control methanol-only conditions (i.e. no RU486 ligand) with flies of identical genotypes do not show this defect (S26B Fig). These data collectively indicate that visual and olfactory cues are required and possibly relayed to the MB, either directly or indirectly through a currently unknown circuit, to facilitate dialect learning. By contrast MB function does not appear to be important for D. melanogaster behavior(s) that enable D. ananassae to learn a dialect (S6B Fig). Consistent with this idea, although Orb2ΔQ mutants cannot function as students (Fig 7E) [17], D. ananassae nevertheless learns the D. melanogaster dialect from Orb2ΔQ mutants (Fig 7D).
Because MB function is necessary for dialect learning during dialect training, we tested the long-term memory proteins Orb2, FMR1, and phosphatase and tensin homolog (PTEN) [64,65] that are known to be required in the MB for memory formation. PTEN has been implicated in murine social learning models, though it has not been tested in a social learning assay in Drosophila [66]. We used the MB Gene-Switch to knockdown expression only during the cohabitation period, after which expression was allowed to resume. D. ananassae learn the dialect of each of these three knockdown lines, again suggesting that MB mediated processes in D. melanogaster are not necessary for D. ananassae dialect training (S26C–S26G Fig). However, under these conditions we find that functional Orb2 and PTEN are required for dialect learning in D. melanogaster, but FMR1 is dispensable (Fig 7F–7H). Orb2 and FMR1 were previously shown to be important in the teacher-student transmission of a wasp threat, and knockdown of either gene completely ablated students learning from teacher flies [17]. In this case, partial communication between D. ananassae teachers and D. melanogaster students can occur because Orb2 and PTEN expression is restored after the dialect training period, thus functioning as wild-type D. melanogaster. D. melanogaster flies having undergone knockdown of Orb2 and PTEN only during dialect training are able to function as students to wild-type D. melanogaster after the cohabitation period is completed, suggesting the partial communication phenotype observed with D. ananassae teachers is a result of gene knockdown during cohabitation and not a by-product of irreversible cellular damage or death caused by the RNAi treatments (S26H–S26L Fig). Collectively, these data show critical gene products are required to function in the MB for dialect learning during the training period. Importantly, although visual inputs are necessary MB function and active learning are not necessary in D. melanogaster in order to in turn provide cues enabling dialect learning by a wild-type D. ananassae student.
In this study, we present an evolutionarily conserved response to predatory wasps across the genus Drosophila, manifesting as egg laying depression coincident with an activated effector caspase, Dcp-1. These endoparasitoid wasps are ubiquitous keystone species in many ecosystems around the world, which prey on Drosophila larvae, with infections rates as high as 90% in natural populations [67–69]. We have shown that flies communicate a wasp threat through visual cues. We used a known generalist wasp species, Leptopilina heterotoma [70,71], suggesting that the communication observed may constitute a form of “social protection” against a pan-threat. Given the geographical distribution of this generalist wasp, species tested in this study have a high likelihood of wasp encounter [72–75]. The effects of other larval and pupal generalists, in addition to specialist wasps, are currently unknown, but may provide a fruitful avenue of study. The high infection rate and prevalence of parasitoids in nature suggest to us that other wasp strains and species may also induce intra- and interspecies communication.
Interspecies communication occurs to varying degrees, likely dependent on evolutionary relatedness. Closely related species, such as D. melanogaster and D. simulans, D. ananassae and D. kikkawai, and D. mojavensis and D. virilis, communicate as effectively as conspecifics. Species more distantly related to D. melanogaster exhibit only partial communication or lack the ability to confer predator information with D. melanogaster. Because of this natural variation in ability to communicate we suggest a useful analogy to language “dialects” that may hinder efficient communication between two dissimilar dialects of a common language. When two species are only able to partially communicate, they can learn each other’s dialect after a period of cohabitation, yielding interspecies communication enhanced to levels normally observed among conspecifics. Such signals benefiting two individuals has been modeled to be honest, and evolutionarily stable [76]. Although dialect learning facilitates interspecies communication across broad evolutionary distances, the ability to learn a specific dialect is dependent on relatedness of the two species (Fig 8A). This observation of the role of phylogenetic distance influencing dialect learning is true in cases both utilizing D. melanogaster and D. ananassae in combination with other species tested (Fig 8A, S27 Fig). The observation that different strains of the same species exhibit this partial communication that can then be enhanced by cohabitation, suggests that both social communication and dialect learning are innate behaviors conserved among all Drosophilids tested here (Fig 8A, S27 Fig). Multiple strains of D. melanogaster reared in the laboratory for many decades exhibit this behavior, supporting the idea that this is an innate behavior. However, flies reared in isolation from the larval stage result in compromised communication ability, suggesting that while the ability to communicate is hardwired, or innate, there is a socialization dependent input the facilitates efficient communication, even between conspecifics. Thus, adult Drosophila neuronal plasticity allows for learning of both the communication between conspecifics and of dialects, but the specific dialect learned is dependent on social interactions specific to a communal environmental context that provides both visual and olfactory inputs. This same plasticity allows for the learning of multiple dialects in a given environment. It is remarkable that communal rearing of two species can enhance communication about a predator that is yet to be experienced by either species. Furthermore, dialect learning does not trigger Dcp-1 activation and oviposition depression, suggesting that social communication about predator presence is different from social interactions that enable dialect learning that later enhances predator presence communication.
Understanding memory formation, storage and retrieval requires knowledge of the underlying neuronal circuits. In Drosophila, the mushroom body (MB) is the major site of learning and memory and we find that the MB is necessary for dialect learning [77,78]. We hypothesize that, given the large number of inputs required in dialect learning (olfactory, ionotropic, and visual cues), which are then relayed to the MB, the “dialect” may be implemented in the MB via several neuronal classes that are activated and deactivated [77]. We suspect that there are increases in MB output neurons (MBONs) that reinforce the memory following a sufficient amount of time of stimulation (i.e. greater than 24 hours in our assay). At the same time, we suspect there may be a decrease in inhibitory MBONs that may be responsible for ignoring other species. This increase/decrease would promote interactions and learning between species. Following this MBONs changing in synaptic strength, we suspect that dopaminergic neurons (DAN(s)) reinforce these signals in the appropriate MB lobes, similar to olfactory memories in other assays. We propose this given the need for olfactory reinforcement during dialect training, in addition to other necessary cues, which emulate the known MB circuitry [77].
We propose dialect learning to be a novel behavior requiring visual and olfactory inputs, perhaps integrated in and relayed through the MB, resulting in the ability to more efficiently receive information about a common predator. Without dialect learning, this information would otherwise be lost in translation or muddled, resulting in an inefficient behavioral response with significant survival disadvantages. Inhibiting synaptic transmission and knockdown of key learning and memory genes in the MB demonstrates that these inputs must be processed and consolidated in the MB, although input neuronal signaling is initiated from the visual and olfactory systems (Fig 8B). Given the need for multiple sensory inputs, dialect learning is fundamentally different from the previously described teacher-student paradigm, where visual cues are necessary and sufficient for information exchange[17]. Additionally, we suggest that this study also points to previously unappreciated functions of the Drosophila MB in integrating information from multiple olfactory and visual inputs [77]. Such cognitive plasticity that allows for dialect learning from many different species hints that adult behaviors could only emerge in a manner that is dependent on previous social experiences where relevant ecological pressures are ever present and multiple species co-exist in nature. Thus, there is a real benefit to cognitive plasticity, where sharing of information directly, or by coincident bystanders, could result in behavioral immunity to pan-specific threats.
The specific information shared by different species during dialect learning is not known. This study, however, provides important clues as the complex suite of sensory systems and cues that may be required for efficient dialect learning. We have presented an example of how interspecies social communication and dialect learning in Drosophila can lead to changes in germline physiology and reproductive behavior. What other ethological behaviors are modulated by MB functions and social interactions typically not revealed in laboratory monocultures? We suggest that the Drosophila MB may integrate a myriad of social and environmental cues in order to produce ethologically relevant behavior that is responsive and useful to local environmental conditions.
The D. melanogaster strains Canton-S (CS), Oregon-R (OR), white1118(w1118), and transgenic flies carrying Histone H2AvD-RFP (His-RFP) were used as wild-type strains. Experiments were primarily performed using CS as wild type flies except where otherwise indicated. Orco1(Or83b1), UAS-TeTx, UAS-Orb2RNAi, UAS-FMR1RNAi, UAS-FMR1RNAi, UAS-PTENRNAi, UAS-Ir8aRNAi, UAS-Ir25aRNAi, ninaBP315 were acquired from the Bloomington Drosophila Stock Center (stock numbers 23129, 28838, 27050, 27484, 34944, 25841, 25813, 43985, and 24776 respectively). Drosophila species were acquired from the Drosophila Species Stock Center (DSSC) at the University of California, San Diego. Flies and their respective stock numbers are listed: D. simulans (14021–0251.196), D. mauritiana (14021–0241.01), D. sechellia (14021–0248.25), D. yakuba (14021–0261.01), D. tsacasi (14028–0701.00), D. kikkawai (14028–0561.00), D. ananassae (14024–0371.13 and 14024–0371.11), D. pseudoobscura (14011–0121.00), D. neocordata (14041–0831.00), D. equinoxialis (14030–0741.00), D. willistoni (14030–0811.00), D. immigrans (15111–1731.08), D. mojavensis (15081–1352.22), and D. virilis (15010–1051.87). All experiments with D. ananassae used strain number 14024–0371.13 unless otherwise noted (S1 Table). All stocks were kept separate to prevent visual transfer of information that could confound experiments.
The ewgNS4 mutant line was kindly provided by Yashi Ahmed (Geisel School of Medicine at Dartmouth). The mushroom body Gene-Switch line was kindly provided by Greg Roman (Baylor College of Medicine). Ir8a1, Ir25a2, Ir8a>GAL4, Ir25a>GAL4 and Ir8a1;Ir25a2;Orco1 lines were kindly provided by Greg S. B. Suh (Skirball Institute at NYU). Flies aged 3–6 days post-eclosion on fresh Drosophila media were used in all experiments. Flies were maintained at room temperature with approximately 30% humidity. All species and strains used were maintained in fly bottles (Genesse catalog number 32–130) containing 50 mL of standard Drosophila media. Bottles were supplemented with 3 Kimwipes rolled together and placed into the center of the food. Drosophila media was also scored to promote oviposition. Fly species stocks were kept separate to account for visual cues that could be conferred if the stocks were kept side-by-side.
The Figitid larval endoparasitoid Leptopilina heterotoma (strain Lh14) was used in all experiments. L. heterotoma strain Lh14 originated from a single female collected in Winters, California in 2002. In order to propagate wasp stocks, we used adult D. virilis in batches of 40 females and 15 males per each vial (Genesse catalog number 32–116). Adult flies were allowed to lay eggs in standard Drosophila vials containing 5 mL standard Drosophila media supplemented with live yeast (approximately 25 granules) for 4–6 days before being replaced by adult wasps, using 15 female and 6 male wasps, for infections. These wasps deposit eggs in developing fly larvae, and we gave them access specifically to the L2 stage of D. virilis larvae. Wasp containing vials were supplemented with approximately 500 μL of a 50% honey/water solution applied to the inside of the cotton vial plugs. Organic honey was used as a supplement. Wasps aged 3–7 days post eclosion were used for all infections and experiments. Wasps were never reused for experiments. If wasps were used for an experiment, they were subsequently disposed of and not used to propagate the stock.
Briefly, fly duplexes were constructed (Desco, Norfolk, MA) by using three standard 25mm x 75mm pieces of acrylic that were adhered between two 75mm x 50mm x 3mm pieces of acrylic. Clear acrylic sealant was used to glue these pieces together, making two compartments separated by one 3mm thick acrylic piece. Following sealant curing, each duplex was soaked in water and Sparkleen detergent (Fisherbrand catalog number 04-320-4) overnight, then soaked in distilled water overnight and finally air-dried. This same cleaning protocol is used following usage of a duplex. The interior dimensions of each of the two units measured approximately 23.5mm (wide) x 25mm (deep) x 75mm (tall).
For experiments using Fly Duplexes (teacher-student interaction), bead boxes (6 slot jewelers bead storage box watch part organizer sold by FindingKing) were used to accommodate 12 replicates of each treatment group. Each compartment measures 32 x 114 mm with the tray in total measuring 21 x 12 x 3.5 mm. Each compartment holds 2 duplexes, and the tray in total holds 12 duplexes. Each bead box was soaked in water and Sparkleen detergent (Fisherbrand catalog number 04-320-4) overnight, then soaked in distilled water overnight and finally air-dried every time before and after use. Empty duplexes were placed into the bead box compartments. 50 mL standard Drosophila media in a standard Drosophila bottle (Genesse catalog number 32–130) was microwaved for 39 seconds. This heated media was allowed to cool for 2 minutes on ice before being dispensed. Each duplex unit was then filled with 5 mL of the media and further allowed to cool until solidification. The open end of the Fly Duplex was plugged with a cotton plug (Genesse catalog number 51-102B) to prevent insect escape. 10 female flies and 2 male flies were placed into one chamber of the Fly Duplex in the control, while 20 female Lh14 wasps were placed next to the flies in the experimental setting for 24 hours. After the 24-hour exposure, flies and wasps were removed by anesthetizing flies and wasps in the Fly Duplexes. Control flies underwent the same anesthetization. Wasps were removed and replaced with 10 female and two male “student” flies. All flies were placed into new clean duplexes for the second 24-hour period, containing 5 mL Drosophila media in a new bead box. For fly duplexes containing a subset of species, specifically D. mojavensis, D. immigrans, and D. virilis, 10 yeast granules were added to the standard Drosophila media after solidification of the food. This activated yeast was added to promote oviposition. Flies showed minimal oviposition in food lacking yeast. We speculate this was observed due to the fly food being optimized for D. melanogaster, which could be creating sensitized species to wasp presence. Plugs used to keep insects in the duplex were replaced every 24 hours to prevent odorant deposition on plugs that could influence behavior. The oviposition bead box from each treatment was replaced 24 hours after the start of the experiment, and the second bead box was removed 48 hours after the start of the experiment. Fly egg counts from each bead box were made at the 0–24 and 24-48-hour time points. 12 biological replicates were performed except where otherwise indicated.
All experimental treatments were run at 25°C with a 12:12 light:dark cycle at light intensity 167, using twelve replicates at 40% humidity unless otherwise noted. Light intensity was measured using a Sekonic L-308DC light meter. The light meter measures incident light and was set at shutter speed 120, sensitivity at iso8000, with a 1/10 step measurement value (f-stop). Fly duplexes and bead boxes soaked with distilled water mixed with Sparkleen after every use for and subsequently rinsed with distilled water and air-dried in the manner described above. To avoid bias, all egg plates were coded and scoring was blind as the individual counting eggs was not aware of treatments or genotypes/species.
Species were cohabitated in standard Drosophila bottles (Genesee catalog number 32–130) containing 50 mL standard Drosophila media. Three Kimwipes were rolled together and placed into the center of the food. Batches of 3 bottles were made per treatment. Two species were incubated in each bottle with 100 female and 20 males of each species per bottle. Every two days, flies were placed into new bottles prepared in the identical manner. Flies were cohabitation for approximately 168 hours (7 days), unless otherwise noted. Following cohabitation, flies were anesthetized and the two species were separated. The flies were then used as students to wasp or mock exposure teachers of the opposite species. For example, we cohabitated D. melanogaster and D. ananassae for one week. Following the weeklong cohabitation, we separated the dialect trained flies. Trained D. melanogaster were placed in duplexes next to D. ananassae either mock or wasp exposed. Trained D. ananassae were placed in duplexes next to D. melanogaster either mock treated or wasp exposed.
For experiments utilizing more than two species for dialect learning, species were cohabitated in standard Drosophila bottles (Genesee catalog number 32–130) containing 50 mL standard Drosophila media. Three Kimwipes were rolled together and placed into the center of the food. Batches of 3 bottles were made per treatment. The three species were incubated in each bottle with 100 female and 20 males of each species per bottle. Every two days, flies were placed into new bottles prepared in the identical manner. The three-fly species were cohabitation for approximately 168 hours (7 days), unless otherwise noted. Following cohabitation, flies were anesthetized and one of the three species was tested by pairing them with teachers of the other two species. For example, we cohabitated D. melanogaster, D. ananassae, and D. willistoni for one week. Following the weeklong cohabitation, we separated the dialect trained flies. Trained D. melanogaster were placed in duplexes next to either D. ananassae or D. willistoni, mock or wasp exposed.
For cohabitation experiments where two species were allowed visual only cues, the Fly Duplex was utilized. The two species were co-incubated side-by-side with 100 female and 20 males of each species per unit using the two chambers of the fly duplex such that the flies could only see each other. The fly duplex was placed into bead boxes, with each unit of the duplex containing 5 mL of standard Drosophila media. Every two days, flies were placed into new fly duplexes with fresh 5 mL standard Drosophila media. Following the weeklong co-incubation, flies were anesthetized and the two species were separated. The flies were then used as students to wasp or mock exposure teachers of the opposite species.
For cohabitation experiments where the two species did not have visual cues, the two species were incubated in bottles with 100 female and 20 males of each species per bottle in complete darkness. The only difference between this method and other training sessions was the lack of light—meaning flies were subject to 25°C with 40% humidity. Every two days, flies were placed into new bottles prepared in the identical manner. Flies were exposed to light for less than 30 seconds, during which they were placed into a new bottle, and immediately returned to the dark. Following the weeklong dark-cohabitation, flies were anesthetized and the two species were separated. The flies were then used as students to wasp or mock exposure teachers of the opposite species.
For cohabitation experiments under monochromatic light settings, batches of 3 bottles with 100 female and 20 males of each species were placed into 27.9cm x 16.8cm x 13.7cm plastic boxes (Sterilite 1962 Medium Clip Box with Blue Aquarium Latches sold by Flikis). These boxes were externally wrapped with colored cellophane wrap, allowing only a certain wavelength of light to be transmitted into the boxes. Red and blue cellophane wraps were purchased from Amscam (Amscan Party Supplies for Any Occasion Functional Cellophane Wrap, 16' x 30", Rose Red and Spanish Blue). Cellophane wrapped boxes with bottles containing flies were subject to 25°C with 40% humidity under the same light intensity as previous experiments. Light intensity within the red wrapped box was 112 and within the blue wrapped box was 115 measured using the Sekonic L-308DC light meter. Every two days, flies were placed into new bottles prepared in the manner described previously. Flies were exposed to broad-spectrum light for less than 30 seconds, during which they were placed into a new bottle, and immediately returned to monochromatic light. Following the weeklong monochromatic-light-cohabitation, flies were anesthetized and the two species were separated. The flies were then used as students to wasp or mock exposure teachers of the opposite species.
For the one-day cohabitation experiments, batches of 3 bottles with 100 females and 20 males of each species were placed at 25°C with 40% humidity for 24 hours. Following the 24-hour cohabitation, flies were anesthetized and the two species were separated. The flies were then used as students to wasp or mock exposure teachers of the opposite species.
In order to ask whether socialization is needed for learning ability between D. melanogaster, we performed isolation experiments (Fig 3). In order to acquire isolated flies, we performed a 24-hour egg lay using approximately 100 females and 20 males of 3-5-day old Canton S at 25°C with 40% humidity on grape juice agar plates. Grape juice plates were made in aliquots of 30 plates, containing a total of 100 mL. We mixed: Dextrose (5.8 g), Sucrose (3.0 g), Agar (2.2 g), and Yeast (2.2 g). We added 86 mL distilled water and 12 mL grape juice concentrate (welches brand) to these solids. This solution was brought to a boil in a microwave, and allowed to pour and solidify. Plates were used immediately upon cooling.
Following the 24-hour egg lay, flies were removed and the egg lay plate was placed at 25°C with 40% humidity with a 12:12 light: dark cycle for a second 24-hour period, after which, L1 larvae were collected and placed into a Falcon round-bottom polypropylene tube (catalog number 352063) containing 1 mL standard Drosophila media. Larvae were allowed to pupate and eclose in isolation. Each tube was kept separate such that no visual information could be transferred between tubes. Following eclosion, 3–5 day old flies were used as students. 1 female and 1 male isolated Canton S were used as students, paired with 1 female, 1 male Canton S raised under typical socialized conditions. Social conditions were achieved by performing the same egg lay protocol as above, but 100 L1 larvae were transferred to standard Drosophila bottles (Genesee catalog number 32–130) containing 50 mL standard Drosophila media and allowed to pupate and eclose at 25°C with 40% humidity.
RU486 (Mifepristone) was used from Sigma (Lot number SLBG0210V) as the ligand for Gene-Switch experiments. Dialect training bottles were prepared by directly pipetting an RU486 solution onto the 3 Kimwipes in the bottle. The solution was prepared by dissolving 3.575 mg of RU486 in 800μL methanol (Fisher Scientific Lot number 141313). This solution was added to 15.2 mL of distilled water. The total solution (16 mL) was thoroughly mixed and 4000 μL was pipetted onto the Kimwipe in each bottle. For bottles containing no RU486 (methanol only) 800μL methanol was mixed with 15.2 mL of distilled water. The total solution (16 mL) was thoroughly mixed and 4000 μL were pipetted onto the Kimwipe in each bottle. Flies were shifted to new bottles prepared in the exact same manner every two days. Flies were cohabitated for approximately 7 days. Following cohabitation, flies were anesthetized and the two species were separated. The flies were then used as students to wasp or mock exposure teachers of the opposite species.
Ovaries were collected from flies that were placed in vials along with female wasps for experimental or no wasps for control settings. Flies were placed in batches into standard vials (Genesee catalog number 32–116) of 20 females, 2 males along with 20 female wasps for exposed vials, or simple placing 20 female and 2 male flies in vials for the unexposed treatments. Three vials were prepared to produce three replicates to account for batch effects. We observed no batch effects so each of the 12 ovaries imaged from each treatment were then counted as a replicate, thus providing an n of 36. Ovaries that were prepared for immunofluorescence were fixed in 4% methanol-free formaldehyde in PBS with 0.001% Triton-X for approximately five minutes. The samples were then washed in PBS with 0.1% Triton-X, and blocked with 2% normal goat serum (NGS) for two hours. The primary antibody, cleaved Drosophila Dcp-1 (Asp216) (Cell Signaling number 9578) at a concentration of 1:100, was used to incubate the ovaries overnight at 4° C in 2% normal goat serum (NGS). The secondary antibody used was Fluorescein isothiocyanate (FITC) conjugated (Jackson Immunoresearch), and used at a concentration of 1:200 for a two-hour incubation at room temperature. This was followed by a 10-minute nuclear stain with 4', 6-diamidino-2-phenylindole (DAPI). For confocal imaging of D. melanogaster ovaries, wheat germ agglutinin (WGA) was also used as a membrane marker (Fig 1F and 1J, S2 Fig). All egg chambers were counted to acquire total egg chamber number and egg chambers showing Dcp-1 signal were counted as positive for Dcp-1. All ovary quantifications were performed in a blinded manner such that the counter did not know the condition (exposed v unexposed) or species of the Drosophila ovaries being counted.
A Nikon A1R SI Confocal microscope was used for imaging activated Dcp-1 caspase staining in D. melanogaster (Fig 1D–1K, S2 Fig). Image averaging of 4x during image capture was used for all images. A Nikon E800 Epifluorescence microscope with Olympus DP software was used to image Dcp-1 caspase staining on all other Drosophila species tested (S3–S16 Figs). This microscope was also used to quantify egg chambers with Dcp-1 signal and total number of egg chambers in all species tested (S17 and S18 Figs).
Statistical tests on exposed v unexposed/teacher v student interactions were performed in Microsoft Excel. Welch’s two-tailed t-tests were performed for data. P-values reported were calculated for comparisons between paired treatment-group and unexposed and are included in S1 File.
Categorization assignments were made based on the criteria of mean value and statistical significance compared to unexposed. ‘No communication’ is assigned in instances where there was not a statistically significant decrease of the exposed group. ‘Partial communication’, requires a statistically significant decrease of the exposed group, with an exposed mean above 50%. ‘Full communication’, criteria are a statistically significant decrease of the exposed group, along with a mean below 50%. Direct comparisons between partial and full communication groups would require analysis of data collected at different times and between genotypes, rendering any such p-values invalid. However, to satisfy the desire for p-values associated with the partial/full threshold, one sample one tailed t tests were performed on exposed samples that were statistically less than unexposed (S2 File). Corresponding p-values asses if the exposed group is statistically less than 50%. Statistical comparisons were performed in R (version 3.0.2 “Frisbee Sailing”).
|
10.1371/journal.pgen.1007607 | Identification of expression quantitative trait loci associated with schizophrenia and affective disorders in normal brain tissue | Schizophrenia and the affective disorders, here comprising bipolar disorder and major depressive disorder, are psychiatric illnesses that lead to significant morbidity and mortality worldwide. Whilst understanding of their pathobiology remains limited, large case-control studies have recently identified single nucleotide polymorphisms (SNPs) associated with these disorders. However, discerning the functional effects of these SNPs has been difficult as the associated causal genes are unknown. Here we evaluated whether schizophrenia and affective disorder associated-SNPs are correlated with gene expression within human brain tissue. Specifically, to identify expression quantitative trait loci (eQTLs), we leveraged disorder-associated SNPs identified from 11 genome-wide association studies with gene expression levels in post-mortem, neurologically-normal tissue from two independent human brain tissue expression datasets (UK Brain Expression Consortium (UKBEC) and Genotype-Tissue Expression (GTEx)). Utilizing stringent multi-region meta-analyses, we identified 2,224 cis-eQTLs associated with expression of 40 genes, including 11 non-coding RNAs. One cis-eQTL, rs16969968, results in a functionally disruptive missense mutation in CHRNA5, a schizophrenia-implicated gene. Importantly, comparing across tissues, we find that blood eQTLs capture < 10% of brain cis-eQTLs. Contrastingly, > 30% of brain-associated eQTLs are significant in tibial nerve. This study identifies putatively causal genes whose expression in region-specific tissue may contribute to the risk of schizophrenia and affective disorders.
| An estimated 21 million people live worldwide with schizophrenia, 60 million with bipolar disorder, and 400 million with major depressive disorder. Recent genome-wide association studies have shed light on the genetic variants linked to these disorders, and increasing evidence suggests that their genetic architectures may overlap. However, understanding the roles of these variants in disease biology remains limited. Here we questioned whether genetic variation associated with these disorders is correlated with the expression of genes that are proximally located within the genome. Importantly, we evaluate this in two large and independent human brain tissue datasets. We subsequently identify, with high confidence, >2,200 disease-associated variants as putative regulators of expression for nearby genes. The identification of these regulated genes provides new insights into disease biology and will help prioritise associations for future mechanistic follow-up studies.
| Schizophrenia and affective disorders, comprising bipolar disorder and major depressive disorder, constitute a significant global burden of disease. Worldwide it is estimated that more than 21 million individuals are living with schizophrenia, 60 million with bipolar disorder and over 400 million with major depressive disorder [1]. The consequences are staggering as evidenced by the fact these three diseases, which usually emerge in early-adulthood, accounted for over 90 million disability-adjusted life years in 2010 [2].
Ineffective management, which contributes to the enormous disease burden, is largely due to our lack of understanding about the pathobiology underlying these disorders. Family and twin studies have estimated heritability to be between 70–80% for schizophrenia and bipolar disorder and up to 40% for major depressive disorder [3]. This has prompted the establishment of genome-wide association studies (GWAS) to identify genetic variants associated with these disorders. Recent and large-scale GWAS such as those organized the Psychiatric Genomics Consortium (PGC), CONVERGE Consortium and 23andMe, have been published for schizophrenia [4–7], bipolar disorder [8–10], major depressive disorder [11–13] as well as for a multiple-disorder analysis [14]. Results from these studies have suggested that schizophrenia, bipolar disorder and major depressive disorder may share common genetic architecture [15, 16].
While GWAS have identified numerous loci associated with human diseases [17, 18], understanding their roles in disease biology remains limited. Expression quantitative trait loci (eQTLs) are genetic variants that affect gene expression levels and may offer insights into mechanisms contributing to health and disease [19–21]. An eQTL can act in cis, meaning that the variant is associated with expression of a gene within 1Mb, or in trans, with the variant located outside of this window. Studies have leveraged GWAS data, particularly that of single nucleotide polymorphisms (SNPs), to identify eQTLs [22, 23]. Moreover, it is becoming evident that disease-associated variants are enriched for eQTLs [14, 24]. Hence, eQTL studies based on disease-GWAS may implicate important molecular processes underlying pathobiology.
In this study, we therefore sought to identify eQTLs associated with schizophrenia and affective disorders in neurologically-normal post-mortem brain tissue. By leveraging gene expression data from UK Brain Expression Consortium (UKBEC) [25] and NIH Genotype-Tissue Expression (GTEx) [26] and performing multi-region analyses [27], we identified cis-eQTLs that were pervasive across various brain regions and determined the extent to which these overlapped with those detected in more accessible tissue including blood and peripheral nervous system. Results reported here may help prioritize future studies of GWAS SNPs associated with these disorders.
Publicly available GWAS data from the 11 most-recent and largest studies (those with >10,000 cases and controls) related to schizophrenia, bipolar disorder and major depressive disorder were selected for analysis (Fig 1). SNPs were collated from the following studies: PGC-SCZ1 [4], PGC-SCZ1+Sweden [5], PGC-SCZ2 [6] and SCZ-Chinese [7] for schizophrenia; PGC-BIP [8], PGC-MooDs [9] and 40K_BPD [10] for bipolar disorder; PGC-MDD [11], CONVERGE [12] and 23andMe [13] for major depressive disorder, and PGC-Cross Disorder Analysis [14] for multiple disorders (schizophrenia, bipolar disorder, major depressive disorder, autism spectrum disorder and attention-deficit hyperactivity disorder). Of note, many of these studies included samples that were analysed in previous studies (S1 Table). We included overlapping studies to maximize the number of disorder-associated SNPs in our analysis as there may be loci identified in one study but not in another.
Guided by the NHGRI-EBI GWAS catalogue [17, 28], we included disease-associated SNPs with a study p-value < 5 x 10−5 in our analysis in order to also capture those SNPs with suggestive associations. For SNPs from all studies except PGC-SCZ1+Sweden, PGC-SCZ2 and 23andMe, we also obtained SNPs that are in moderate-high linkage disequilibrium (LD, R2 ≥ 0.5) with the study-SNPs [29] using the web-based tool rAggr (http://raggr.usc.edu/). LD-analysis settings were as follows: CEU population from 1000 Genomes Phase 3 October 2014 release, build hg19, minimum minor allele frequency = 0.001, maximum distance = 500kb, maximum Mendelian errors = 1, cut-off for Hardy-Weinberg p-value = 0, and minimum genotype = 75%. Study-SNPs from PGC-BIP, PGC-MDD and PGC-Cross Disorder Analysis were lifted from hg18 to hg19 prior to LD-analysis. SNPs in LD with study-SNPs from PGC-SCZ1+Sweden, PGC-SCZ2 and 23andMe were not obtained as these studies had >5,000 study-SNPs and were not LD-pruned. A total of 106,397 analysis-SNPs (from autosomal and X chromosomes) resulted from collation of the 11 studies (Fig 1, S1 Table).
Genotype and gene expression data were obtained from UKBEC [25]. These data contained samples from 134 neurologically-normal individuals from the following brain regions: cerebellum, frontal cortex, hippocampus and putamen (see S2 Table for sample sizes per region). UKBEC study brain regions not overlapping with GTEx brain regions were not included in this study. Genotype processing is previously described [25]. Briefly, these data included ~5.88 million imputed (1000 Genome, March 2012 release) and typed SNPs. Raw expression data from Affymetrix Human Exon 1.0 ST microarrays were processed as described previously with minor modifications. Specifically, all ~5 million probes were initially re-mapped to Ensembl v75 annotations using BioMart. Only probe sets containing three or more probes free of the ‘polymorphism-in-probe’ problem were used for subsequent analysis [30]. Following robust multi-array average (RMA) normalisation and background filtering, we calculated gene-level estimates by taking the winsorised mean of all probe sets for a given gene. Prior to eQTL detection, these were covariate-corrected for sex, cause of death, post-mortem interval, RNA integrity number (RIN) and study group using a linear regression model.
Data from GTEx [26] (dbGaP accession phs000424.v6.p1) were also included in our study. The following brain regions were analysed: cerebellum, frontal cortex (BA9), hippocampus and putamen (see S2 Table for sample sizes per region). GTEx study brain regions not overlapping with UKBEC brain regions were not included in this study. Whole blood and tibial nerve served as comparisons to brain tissues. Genotype and expression processing is described at http://gtexportal.org. Briefly, the dataset contained ~11.55 million imputed (1000 Genomes, August 2012 release) and typed autosomal variants and ~26,000 transcripts. To make this dataset conform with UKBEC, we excluded GTEx individuals identified as having neurological diseases based on description of their comorbidities/cause of death or inclusion of certain variables (see S2 Table for variable list). Prior to eQTL detection, gene expression was covariate-corrected for first three genotyping principal components (PCs), array platform, sex and probabilistic estimation of expression residuals (PEER) factors using a linear regression model.
The R package Matrix eQTL [31] was used to identify eQTLs. Cis-eQTLs were defined as SNPs within 1Mb of the transcription start site; those SNPs outside this region, i.e. trans-eQTLs, were not considered in this analysis. Study genotype data were limited to those that were present in the set of analysis-SNPs described above. To minimize eQTL signals being called as significant when only present as a strong effect in a few individuals, analysis-SNPs that had a minor allele frequency < 5% within the analysis population were excluded. Gene expression input was based on expression residuals from linear regression of gene expression and genotype. cis-eQTL analysis was performed independently for each region per study (S5 and S6 Figs). eQTL (beta) effect sizes are given as standardised expression units (EU) per allele and each eQTL is a unique SNP-gene pair. We used the stringent Bonferroni correction method to set the critical p-value threshold per brain region and gene expression database combination analysed; hence each combination had its own unique p-value threshold (S3 Table). For completeness, in S3 Table, we also included the number of eQTLs with a False Discovery Rate (FDR) < 0.05.
To determine which eQTLs were present in multiple tissues, we performed two separate approaches that were then intersected. First, and using single-region eQTL results, we performed a meta-analysis of test statistics (eQTL effect size estimates and standard errors calculated by Matrix eQTL) using the metagen function in the R package meta with a random-effects model. FDR was calculated using R p.adjust with the Benjamini-Hochberg method. Three groups of meta-analyses were performed (Fig 1): A) four regions in UKBEC (p-value denoted as pukbec); B) four regions in GTEx (pgtex); and C) of eight regions between UKBEC and GTEx (pmeta). The first ten principal components of the UKBEC-GTEx meta-analysis for each of the overlapping regions are shown in S1–S4 Figs. As above, we used the stringent Bonferroni correction method to set the critical p-value threshold per brain region and gene expression database combination being analysed (S4 Table). For completeness, in S4 Table, we also include the number of eQTLs with an FDR < 0.05.
Second, we utilized a hierarchical Bayesian model for multi-region eQTL analysis (MT-eQTL) [27]. This model incorporates variation patterns of the presence/absence of eQTLs as well as their effect size heterogeneity across tissues. MT-eQTL indicates in which tissue(s) a gene-SNP pair is expected to be an eQTL (isEQTL variable). We similarly applied this approach across three groups (Fig 1): four regions in UKBEC; B) four regions in GTEx; and C) of eight regions between UKBEC and GTEx. An eQTL was deemed probable in each analysis if isEQTL = 1 in all tissues tested (ie all four tissues in UKBEC and GTEx, separately and all eight tissues in UKBEC and GTEx combined, S4 Table).
Through intersection of both analyses, we defined an eQTL as high confidence if (1), the meta-analysis p-value was less than the p-value threshold set by the Bonferroni correction (S4 Table), and (2), it was determined to be an eQTL in all tissues assessed using MT-eQTL (S4–S7 Tables, S7 Fig). To further validate this approach, permutation analysis was used to test the robustness of eQTLs identified in the multi-region meta-analysis of overlapping regions in UKBEC and GTEx datasets (Fig 1, analysis C). 10,000 iterations were performed for each region-study combination per high-confidence eQTL by shuffling sample IDs in the gene expression file prior to eQTL analysis. Meta-analysis of each permuted iteration per combination yielded a meta p-value. Permuted p-value was calculated as the number of meta p-values equal to or less than the nominal p-value (pmeta), divided by the number of iterations (10,000). Those eQTLs with permuted p-values less than 5 x 10−6 (0.05/10000) surpassed permutation testing, thus revealing that all high-confidence eQTLs reached this stringent threshold. The aim of this rigorous and stringent thresholding was to identify disease-associated eQTLs across multiple brain regions with high confidence.
Cis-eQTLs identified in the multi-region meta-analysis above (Fig 1, analysis C) were assessed for overlap with cis-eQTLs from GTEx samples of whole blood and tibial nerve tissue (calculated using Matrix eQTL as described above, S2 Table), as well as those downloaded from the Blood eQTL Browser (http://genenetwork.nl/bloodeqtlbrowser/) [32]. eQTL beta effect sizes for the GTEx samples and meta-analysis were converted to z-scores using the R scale with center = TRUE and scale = TRUE for effect size comparison; for eQTLs from the Blood eQTL Browser, the “OverallZscores” variable was used.
We also accessed for overlap with two recently published eQTL studies involving human dorsolateral prefrontal cortex tissue. The Religious Orders Study and Memory and Ageing Project (ROSMAP) study [33] analysed QTL data from 411 older individuals and resulting cis-eQTL associations were downloaded from the Brain xQTLServe (http://mostafavilab.stat.ubc.ca/xQTLServe/). cis-eQTLs with FDR < 0.05 from the CommonMind Consortium (CMC) case-control study [34] of over 250 individuals with schizophrenia were obtained from the CMC Knowledge Portal (https://www.synapse.org/#!Synapse:syn2759792/wiki/69613).
Replication rates of this study’s brain tissues eQTL analysis with other tissues and databases were assessed using the π1 statistic [35], which estimates the proportion of non-null hypotheses. For eQTL overlap between multi-region meta-analysis (Fig 1, analysis C) and GTEx whole blood, GTEx tibial nerve and ROSMAP [33], the R package qvalue was utilised. As Blood eQTL [32] and CMC [34] data only reported limited eQTLs, available FDR values were used to estimate π1.
Bayesian colocalisation analysis was conducted to assess the extent of overlap between eQTL (related to the PGC-SCZ2 [6] study) and GWAS signals using the R package Coloc [36]. Summary statistics of all SNPs (regardless of GWAS p-value) within 200kb of the lead GWAS SNP and common in both the GWAS and eQTL studies were inputted into Coloc, which was run with default parameter settings [36]. Regions showing evidence of colocalisation between the GWAS and eQTL signals were identified utilizing pre-defined thresholds [37]: PP3 (posterior probability that there exist two distinct causal variants, one for each trait) + PP4 (posterior probability that these exists a single causal variant common to both traits) ≥ 0.80 and PP4/PP3 ≥ 3.
We sought to determine if SNPs associated with schizophrenia and affective disorders also served as eQTLs in brain tissue. Study-SNPs identified in 11 GWAS [4–14] as well as SNPs in moderate to high LD [29] with these SNPs (see Methods) were included in this analysis (Fig 1). Combined, this yielded 106,397 analysis-SNPs across the 11 GWAS for eQTL interrogation. GWAS-independent and neurologically-disease free genotype and gene expression data were obtained from both UKBEC [25] and GTEx [26]. Of the 106,397 analysis-SNPs, 84,786 and 84,308 were present in UKBEC and GTEx, respectively. Expression data originated from four UKBEC and GTEx brain regions (cerebellum, frontal cortex, hippocampus and putamen).
To identify high confidence cis-eQTLs across regions in each study, we initially utilised an additive linear model with Matrix eQTL [31] in each of four regions within UKBEC and GTEx datasets, separately (S5 and S6 Figs). The number of detected cis-eQTLs using a stringent Bonferroni correction threshold varied considerably between studies and regions (S3 Table).
We next leveraged these results across multiple regions within same datasets to increase discovery power and identify gene-SNP pairs that are cis-eQTLs across the four brain regions in each study [38, 39]. To do so, we first meta-analysed the single region cis-eQTLs across the four regions in UKBEC or GTEx separately (Fig 1, analyses A and B, S8 Fig, S5 and S6 Tables). Using a stringent Bonferroni correction threshold (S4 Table), 2,672 and 18,462 unique cis-eQTLs had pukbec/gtex < pbonferroni in the UKBEC and GTEx data, respectively, associated with expression of 37 and 168 genes. Separately, we applied an alternative hierarchical bayesian MT-eQTL model that allows heterogeneity in both the distribution of eQTLs and their effect sizes across multiple tissues, and which additionally controls for correlated measurements of gene expression that can be apparent when sampling multiple tissues from the same donors [27, 40]. This model identified that 6,131 of 6,266 (97.8%) single-tissue eQTLs were highly likely in all four regions in UKBEC. In GTEx, 28,608 of 31,194 (91.7%) single-tissue eQTLs were highly likely in all four regions (S7a and S7b Fig). These were respectively associated with 74 and 282 genes. Finally, by intersecting both meta-analysis and MT-eQTL analysis, we identified 2,672 UKBEC and 18,458 GTEx cis-eQTLs present in all four tissues and with a pukbec/gtex < pbonferroni (Fig 2B), associated with 37 and 167 eGenes, respectively (Fig 2C).
The aforementioned analyses were applied independently to both UKBEC and GTEx datasets. In order to generate a high-confidence list of eQTLs shared between datasets, as well as to discover additional ones, we next sought to determine if eQTLs identified in either UKBEC or GTEx data were significant when assessed across both studies. Following identical strategy to multi-region anaylsis on the separate study datasets, we first applied a meta-analysis of all cis-eQTLs present across eight regions (four UKBEC and four GTEx, Fig 1, analysis C). This identified 2,346 eQTLs with pmeta < pbonferroni (S4 Table) associated with expression of 43 genes. Next, we applied the MT-eQTL model to these same eight regions (four UKBEC and four GTEx), and revealed that 18,340 of 20,347 (90.1%, S4 Table, S7c Fig) eQTLs were present in all tissues analysed and associated with 274 genes.
Intersection of both meta-analysis and MT-eQTL analysis identified 2,224 cis-eQTLs present in all four tissues across both studies that were associated with 40 eGenes (Fig 2A–2C, Table 1, S7 Table, S9 Fig). Of these, 1,070 cis-eQTLs, associated with 20 genes, were significant in the UKBEC multi-region intersection, whilst 1,100 cis-eQTLs, associated with 15 genes were also significant in the GTEx intersection (Fig 2B and 2C). Demonstrating reproducibility between the two studies, 697 of the 2,224 cis-eQTLs were also significant in UKBEC and GTEx (Fig 2B). Therefore, an additional 751 cis-eQTLs were detected when both UKBEC and GTEx datasets were leveraged in combination, thus demonstrating the benefit of a combined analysis. Indeed, this yielded discovery of 14 regulated genes not detected by either dataset alone (Fig 2C). Collectively we consider these 2,224 cis-eQTLs and 40 eGenes as the set of high confidence candidates to take forward. Importantly, a 10,000-fold permutation analysis validated these cis-eQTLs.
Of the 2,224 cis-eQTLs, 58% (1,292) were located with 1MB upstream of the eGene TSS (transcription start site), 11% were located within the eGene itself while the remaining 31% were downstream of the TES (transcription end site). Positional analysis relative to the associated eGenes showed that cis-eQTLs were most enriched within 250kb of the TSS and TES, respectively (Fig 2D). With respect to eQTL variant positions, 45.5% were localised to introns (Fig 2E), as classified by Ensembl Variant Effect Predictor. This high proportion of introns likely includes overlap with introns of neighbouring transcripts. Both observations are consistent with other studies [33, 34]. No clear gene ontology was associated with the eGenes.
Mean and median of the (absolute value of) effect sizes of candidate eQTLs were 0.26 and 0.25 EU per allele, respectively (S10 Fig). eQTLs positioned within 250kb of the TSS and TES, respectively, had the greatest absolute effect size (S11 Fig), similar to what was seen with respect to relative TSS position and variant frequency (Fig 2D). The largest effect size was observed with rs9461434 (-0.56 ± 0.09 EU per allele for AA relative to CC), associated with the expression of the pseudogene ZNF603P (45kb upstream of the TSS, pmeta = 3.7 x 10−9, FDR = 2.9 x 10−6, S12 Fig). The most significant cis-eQTL was rs12438181 on chromosome 15q25.1, 46kb upstream and correlated with the expression of CHRNA5 (pmeta = 2.4 x 10−27, FDR = 1.8 x 10−21), which encodes the α5 subunit of the nicotinic cholinergic receptor. This eQTL had an effect size of 0.28 ± 0.03 EU per allele for AA relative to GG (Fig 3A, 3B and 3D, S9 Fig).
Interestingly, 28 cis-eQTLs (S8 Table) were classified as leading to missense mutations. Three eQTLs resulted in missense mutations within genes whose expression they were then also additionally associated. The most significant of these was rs16969968 (pmeta = 2.3 x 10−21), which is located within exon 5 of CHRNA5, a highly conserved region (Fig 3C, 3E and 3F, S9 Fig). The minor allele A encodes for an amino acid change to asparagine from aspartic acid (major allele G) at position 398, which may affect receptor function (see Discussion). The effect size of GG, relative to AA, is 0.21 ± 0.02 EU per allele, thus implying that both expression changes and receptor activity could be contributing to the effect of this variant.
To assess replicability with other eQTL studies using brain tissue, we evaluated our findings relative to similar studies. Previously, 27 brain eQTLs were identified in a meta-analysis of GWAS SNPs associated with five neuropsychiatric disorders (including three studied here [14]) using cortical expression data from five separate studies [41]. We identified one of these overlapping with our study: rs4523957 associated with expression of SRR on chromosome 17p13.3. Though this was not the most significant eQTL for this gene, thus suggesting that the lead eQTL and the GWAS signals might be different (see below), the association still surpassed the Bonferroni threshold with pmeta = 8.1 x 10−11, FDR = 1.3 x 10−7. SRR encodes for an enzyme that converts L- to D-serine, which has be found to be lower in the CSF of patients with schizophrenia [42], supporting the role of glutamatergic neurotransmission in the biology of schizophrenia and affective disorders [43, 44]. Indeed, it is also a candidate drug target for schizophrenia, again highlighting the potential for eQTL studies to identify pathobiology that might be targeted pharmacologically. ZSCAN31, also known as ZNF323, is another cis-eQTL associated gene that has been previously identified as significantly associated with schizophrenia, bipolar disorder and psychosis in both a GWAS [45] and an eQTL study using 193 human prefrontal cortex samples [46]. In our study, multiple cis-eQTLs associated with this gene were significant (pmeta = 5.0 x 10−15 for the top eQTL). Likewise, two genes that were prioritized as putatively causal from integrative analysis of PGC-SCZ2 GWAS with both whole blood and UKBEC data averaged across 10 brain regions [47], SNX19 and NMRAL1, were associated with cis-eQTLs with some evidence of significance, pmeta = 1.6 x 10−4 and 6.5 x 10−6, respectively (both genes had probable eQTLs in all eight brain regions analysed across UKBEC and GTEx).
Next, we sought to test replication of our high-confidence candidate eQTLs in a disease-associated eQTL study. Here, the CommonMind Consortium (CMC) study [34] recently analysed gene expression of dorsolateral prefrontal cortex in over 500 schizophrenia cases and controls. In comparing the two analyses, of the 2,224 eQTLs surpassing the pmeta Bonferroni threshold, 1,969 (88.5%) had a FDRCMC < 0.05, accounting for 30 out of 40 eGenes identified in our study (S13A Fig). Importantly, 7,555 of the 10,286 eQTLs with FDRmeta < 0.05 were also significant in the CMC study, suggesting strong overlap between the two analyses despite our use of healthy brain tissue alone (π1 = 0.73, S9 Table). The most significant overlapping eQTL was rs139708473 associated with the expression of LINC00499, a brain and testis expressed non-coding RNA on chromosome 4q28.3 with unclear function (pmeta = 2.7 x 10−19, FDR = 3.7 x 10−15, effect size = -0.51 ± 0.06 EU per AA, compared to GG, S13B Fig).
A key consideration in eQTL analyses is the source of tissue used for gene expression data. While identification of eQTLs is increased by studying associations in multiple tissues [48, 49], other studies have found that analyses in disease-related tissues are enriched for disease-associated eQTLs [50, 51]. Recent availability of datasets detailing gene expression in various regions of human brain have now allowed for eQTL analyses in nervous tissue [25, 26, 52]. Previous eQTL analyses for schizophrenia and affective disorders have relied on transcriptome data collected from more readily available biological specimens such as whole blood [53–55] or more disease-related specimens such as single brain regions [55]. An important and clinically relevant question is whether disease-related eQTLs can be detected in samples that are more accessible and related from living patients. If so, this may facilitate larger study designs in the future where high quality biological material is more readily selected.
To test whether our significant cis-eQTLs detected in brain regions can be identified in more clinically-accessible tissue, we assessed the associations of the 2,224 cis-eQTLs in whole blood tissue. Of these, only 50 were detected in the largest meta-analysis of peripheral whole blood to date (Westra Blood, n = 5,311) [32], with 36 (1.6%) reaching significance in the Westra Blood samples (Fig 4A and 4D, S14D and S14G Fig). These eQTLs were associated with four genes (BTN3A2, HIST1H4H, CHRNA5 and GSTO2) (S15B Fig). Overall, replication of our data in the Westra Blood samples was very low (π1 = 0.05). Of note, eQTLs associated with CHRNA5 expression had a minimum pblood-Westra > 7 x 10−4.
Similar results were observed within the GTEx database, where only 1,146 of the 2,224 cis-eQTLs were detected in GTEx-whole blood. However, only 139 (6.3%) reached the stringent Bonferroni correction threshold for GTEx-whole blood (Fig 4B and 4E, S14E and S14H Fig, π1 = 0.32). Looking in more detail, of the 40 genes associated with brain cis-eQTLs from the multi-region meta-analysis, thirty had detectable expression in the GTEx whole blood samples. However, only BTN3A2, SRD5A3 and DYNC1I2 were associated with blood cis-eQTLs having pblood-GTEx < pbonferroni (S14B and S15A Figs). The most significant brain cis-eQTL in whole blood was rs72841536 (chromosome 6p22.2), which correlated with BTN3A2 expression (pblood-GTEx = 2.7 x 10−46 and FDR = 5.1 x 10−43). Of note, the direction of the effect was the same as that seen in the brain eQTL; the effect size was 0.86 ± 0.05 EU per allele in blood compared to an effect size of 0.44 ± 0.08 EU per allele in brain (S15C and S15D Fig). Contrastingly, the top brain cis-eQTL, rs12438181, associated with CHRNA5 expression, only had a pblood-GTEx = 0.05 and FDR = 0.40 (effect direction was the same). Thus, whilst brain-related cis-eQTLs can be detected in whole blood in principle, nearly 50% were not detected and over 90% did not reach significance in whole blood analyses.
Although it has been demonstrated that gene expression profiles of the central nervous system (CNS) are different than those of the peripheral nervous system (PNS) [56, 57], data suggest that the transcriptomes of the CNS and PNS are more similar than that between CNS and blood [26]. These findings, along with increased tissue accessibility of the PNS over the CNS, motivated us to ascertain whether significant cis-eQTLs captured in the multi-region meta-analysis were also significant in eQTL analysis of tibial nerve tissue, a component of the PNS. Of the 2,224 cis-eQTLs, 2,220 were detected in GTEx tibial nerve tissue and more importantly, 695 (31.2%) reached the Bonferroni correction threshold for GTEx tibial nerve (Fig 4C and 4F, S14F and S14I Fig, S10 Table, π1 = 0.66), five-fold more than that detected in whole blood. Moreover, 38 of 40 significantly associated eGenes had detectable expression in the GTEx tibial nerve samples (LINC00499 and AF131216.5 were not detected in tibial nerve samples). Importantly, seven genes (ZNF603P, DDHD2, ALMS1P, RSRC1, AC068039.4, CD46, LMAN2L, S10 Table) were associated with overlapping significant eQTLs (pmeta and pnerve-GTEx both < pbonferroni, S16 Fig). This suggests that while the tibial nerve does not fully capture the eQTLs identified in brain, it may be a better proxy than whole blood. Taken together, these findings strongly suggest that eQTL analyses should either be performed in disease-relevant tissue wherever possible or that more extensive studies targeting minimally invasive tissues, such as those from the PNS, are necessary to identify suitable brain correlates.
In this multi-region meta-analysis, we report the presence of eQTLs in brain tissue for SNPs that are associated with a risk of developing schizophrenia and affective disorders. Even with the requirement of a stringent Bonferroni correction p-value threshold and being an eQTL in all tissues assessed, we identified 2,224 cis-eQTLs that were correlated with expression of 40 genes. These associations held across four brain regions from two independent studies (UKBEC and GTEx) of neurologically-normal individuals as well as through permutation analyses of 10,000 iterations, supporting the robustness of these findings.
Of the 2,224 cis-eQTLs detected, nearly two-thirds were located within introns or intergenic regions, consistent with previous studies [55, 58]. Meanwhile, 1.4% were classified as causing missense mutations in the proteins encoded from genes harbouring the eQTL. The most significant eQTL that results in a missense mutation within the gene whose expression it is associated with, rs16969968, leads to an amino acid change at position 398 in CHRNA5, a subunit in nicotinic acetylcholine receptors. Interestingly, functional in vitro studies have demonstrated that receptors containing this missense mutation are less responsive to a nicotinic agonist than ones with the more common variant [59, 60]. This leads to reduced cell-depolarization and cholinergic signalling and is consistent with the reported hypofunction of cholinergic signalling in schizophrenia [61]. Moreover, pharmaceutical modulation of cholinergic signalling may improve outcomes for schizophrenia [62].
There is modest evidence that this and other eQTLs affecting CHRNA5 expression are associated with schizophrenia and affective disorders [6, 63, 64]. However, rs16969968 has also been shown to be associated with increased tobacco use [59, 65] and incidence of lung cancer [66, 67]. Therefore, given the disproportionate percentage of individuals with mental illness that smoke [68], further studies are needed to ensure that these eQTLs are not associated with a confounding behaviour seen in such individuals. Nonetheless, these cis-eQTLs for CHRNA5 demonstrate how genetic analyses can identify variants that may increase disease-risk while concurrently being potential therapeutic targets.
Another clinically interesting eQTL is that of rs12491598, associated with the expression of RSRC1 on chromosome 3 (S17 Fig). While this SNP was initially identified in a GWAS of MDD [13], the eGene from our eQTL-analysis has been associated with schizophrenia in a combined case-control imaging genetics study utilising left dorsal lateral prefrontal cortex activation as an intermediate phenotype [69, 70]. RSRC1 is involved in pre-mRNA splicing [71] as well as a marker of subventricular zone progenitor cells within the foetal and postnatal forebrain [72], supporting the hypothesis of a developmental aetiology for schizophrenia. However, SNPs cis to this gene were not significantly associated with schizophrenia [6], with a minimum p-value of 1.4 x 10−3. Moreover, RSRC1 has also been associated with the extreme ranges of height [73]. These findings necessitate further investigation of the role of this gene in health and disease.
More generally, comparative analysis found that our approach replicated several cis-eQTLs found in other studies. This includes the well-powered report of the CommonMind Consortium [34], which utilised both healthy and diseased brain samples to discover eQTLs in the dorsolateral prefrontal cortex. Indeed, 1,969 (88.5%) had a FDRCMC < 0.05, accounting for 30 out of 40 eGenes identified in our study (π1 = 0.73). This supports our approach of leveraging information across eQTL studies to discover new regulatory elements that aids interpretation of GWAS results. Indeed, these findings also support the suitability of non-disease state tissue to investigate disease pathobiology, as has been demonstrated with the use of normal prostate tissue to study prostate cancer [74].
Despite the highlighted overlaps with previous studies, several reported associations were not significant here. CACNA1C and ZNF804A are two genes that have been implicated in schizophrenia and affective disorders through multiple studies, including GWAS and case/control brain expression analyses [75–79]. While both UKBEC and GTEx data contain expression data regarding these genes and the set of analysis-SNPs contain variants that are in cis, we did not find any significant eQTLs in our meta-analysis of UKBEC and GTEx across the four overlapping brain regions. For CACNA1C (minimum pmeta = 1.3 x 10−3), the effect size estimates of the cis-eQTLs had different signs across the various tissues and studies (effect size of 0.06 ± 0.02 EU per allele, S11 Table). Similarly, for ZNF804A, the effect size estimates also demonstrated different signs across the tissues and studies (pmeta = 3.1 x 10−1, effect size of -0.02 ± 0.02 EU per allele for the most significant meta-eQTL, S12 Table). This suggests that some disease-implicated cis-eQTLs may have varying effects in different brain regions, warranting further investigation. Case inclusion may also be a confounding issue when comparing to some studies. Suggestive of this, overlap of our eQTLs was less apparent in the longitudinal ROSMAP study [33] that had reported eQTLs from the dorsolateral prefrontal cortex (π1 = 0.40). We note that despite the 500 study individuals that were healthy at the time of enrolment in ROSMAP, over half developed Alzheimer’s disease by the time brain tissue was donated for analysis. Further investigation will therefore be required to determine if sample size and/or confounding disease genetics could be the reason for limited intersection of these two eQTL studies, especially given the intense research into the proportion of disease- and tissue-specific eQTLs [34, 80, 81].
It is important to determine which specific isoforms are the target of the cis-eQTLs identified in this study [34], and which are the causal variants associated with GWAS signal. Fine-mapping tools such as SHERLOCK [82], RTC [83], Coloc [36] and eCAVIAR [84] have been developed to achieve the latter, and have so far been used to prioritize certain genes in schizophrenia and affective disorders [34, 47, 85]. A Coloc analysis of our high-confidence eQTLs with the PGC-SCZ2 schizophrenia GWAS supported colocalisation between the GWAS and lead eQTL signals for seven of the thirty-three eGenes (S13 Table). This ratio is comparable to other studies which demonstrate that less than a fifth of GWAS loci for schizophrenia have had genes prioritized in this way. However, it is also important to appreciate that current estimates suggest 5–25% of credible intervals identified with such tools may not actually contain the causal variants [86]. Moreover, many cis-eQTLs are driven by multiple independent SNPs [58, 87, 88]. As Coloc assumes at most one causal variant per region [36], it may be the case that many of the signals that we have captured as high-confidence eQTLs are secondary or tertiary. This remains an area we are further investigating computationally. Further, this strongly encourages follow up experimental validation to help elucidate causal SNPs and genes [34]. Meanwhile, emerging research showing integrated epigenome and transcriptome QTL analysis can complement fine-mapping approaches for variant prioritization is an exciting new avenue that merits further exploration [33, 89].
With reference to future QTL studies, progress in understanding the biology governing schizophrenia and affective disorders has been hampered by difficulty in accessing disease-relevant tissue. Therefore, peripheral blood, as it is more accessible, has previously been used as a proxy for studying eQTLs of complex diseases [21, 32, 47]. However, it is unclear as to how robust findings from blood samples are with respect to studying diseases outside of the hematopoietic system. By comparing cis-eQTL results from brain and blood, we find that a minority of the brain-eQTLs were detected in blood and their significance (as marked by p-value) was greatly diminished. Substantially more overlap (over >99% of brain-eQTLs) was demonstrated in GTEx tibial nerve samples, with 31.2% having significant association in both analyses. The extent to which this overlap is driven by power versus tissue specificity is unclear. Therefore, studying eQTLs in disease-related tissues (i.e. central or peripheral nervous tissue in this case) is warranted and may prioritize eQTLs for further validation and mechanistic studies. Indeed, this has been demonstrated through the identification of eQTLs associated with prostate cancer with the use of non-diseased prostate tissue [74].
In summary, we have identified robust cis-eQTLs associated with schizophrenia and affective disorders in human brain tissue. Of the eQTL-associated genes, many have been implicated previously, such as CHRNA5 and RSRC1, while others are novel associations (i.e. ZNF603P) that now merit further analyses. We also demonstrate that eQTL analysis in disease-related tissues allows for prioritization of associations for follow-up studies in diseased-tissue. These results provide insight into putative mechanisms related to development of schizophrenia and affective disorders, thereby identifying potentially new therapeutic targets.
|
10.1371/journal.pcbi.1007024 | Predicting three-dimensional genome organization with chromatin states | We introduce a computational model to simulate chromatin structure and dynamics. Starting from one-dimensional genomics and epigenomics data that are available for hundreds of cell types, this model enables de novo prediction of chromatin structures at five-kilo-base resolution. Simulated chromatin structures recapitulate known features of genome organization, including the formation of chromatin loops, topologically associating domains (TADs) and compartments, and are in quantitative agreement with chromosome conformation capture experiments and super-resolution microscopy measurements. Detailed characterization of the predicted structural ensemble reveals the dynamical flexibility of chromatin loops and the presence of cross-talk among neighboring TADs. Analysis of the model’s energy function uncovers distinct mechanisms for chromatin folding at various length scales and suggests a need to go beyond simple A/B compartment types to predict specific contacts between regulatory elements using polymer simulations.
| Three-dimensional genome organization is expected to play crucial roles in regulating gene expression and establishing cell fate, and has inspired the development of numerous innovative experimental techniques for its characterization. Though significant progress has been made, it remains challenging to construct chromosome structures at high resolution. Following the maximum entropy approach pioneered by Zhang and Wolynes, we developed a predictive model and parameterized a force field to study chromatin structure and dynamics using genome-wide chromosome conformation capture data (Hi-C). Starting from one-dimensional sequence information that includes histone modification profiles and CTCF binding sites, this model predicts chromosome structure at a 5kb resolution, thus establishing a sequence-structure relationship for the genome. A significant advantage of this model over comparable approaches is its ability to study long-range specific contacts between promoters and enhancers, in addition to building high-resolution structures for loops, TADs and compartments. Furthermore, the model is shown to be transferable across chromosomes and cell types, thus opens up the opportunity to carry out de novo prediction of genome organization for hundreds of cell types with available epigenomics but not Hi-C data.
| The human genome contains about 2 meters of DNA that is packaged as chromatin inside a nucleus of only 10 micrometers in diameter [1]. The way in which chromatin is organized in the three-dimensional space, i.e., the chromatin structure, has been shown to play important roles for all DNA-templated processes, including gene transcription, gene regulation, DNA replication, etc [2–4]. A detailed characterization of chromatin structure and the physical principles that lead to its establishment will thus greatly improve our understanding of these molecular processes.
The importance of chromatin organization has inspired the development of a variety of experimental techniques for its characterization. For example, using a combination of nuclear proximity ligation and high-throughput sequencing, chromosome conformation capture and related methods quantify the interaction frequency in three-dimensional space between pairs of genomic loci [5,6], and have revealed many conserved features of chromatin organization. A consistent picture that is emerging from these experiments is the formation of chromatin loops and topologically associating domains (TADs) at the intermediate scale of kilobases to megabases, and the compartmentalization of chromatin domains that are millions of base pairs apart in sequence [7–11]. Many of the findings from these cross-linking experiments are now being validated and confirmed with microscopy imaging studies that directly probe spatial contacts [12–20].
Polymer modeling has played a critical role in our understanding of the genome organization and in interpreting features of Hi-C contact maps [21]. In particular, due to its deviation from the value of an equilibrium globule [6], the power-law exponent of the contact probability between pairs of genomic segments as a function of the genomic separation has attracted the attention of numerous research groups [22–28]. Of the many mechanisms that have been proposed, the non-equilibrium extrusion model [29–31], which assumes that cohesin molecules function as active enzymes to inch along the DNA and fold the chromatin until encountering bound CTCF molecules, has gained wide popularity [32]. Notably, this model succeeds in explaining the flanking of CCCTC-binding factor (CTCF) and cohesin binding sites at the boundaries of chromatin loops and TADs [7,9–11,33]. On the other hand, phase separation, which is emerging as the key mechanism for organizing numerous membraneless organelles [34–36], has been suggested as the driving force for chromosome compartmentalization [37–39]. Since polymer molecules that differ in chemical compositions are known not to intermix [40], micro-phase separation can contribute to the formation and compartmentalization of chromatin domains with distinct histone modification profiles. Finally, besides these mechanism-based modeling strategies, data-driven approaches have also been quite successful in reconstructing chromosome structures directly from Hi-C data and revealing structural features of both interphase and metaphase chromosomes [41–45].
In parallel, bioinformatics studies have provided powerful tools in addressing potential biases in Hi-C data [46–48], and offered numerous insights in our understanding of genome organization. In particular, correlating one-dimensional genomics and epigenomics data with 3D contacts has been rather informative and has led to many proposals on the molecular mechanism of chromatin folding [4,49–54]. Furthermore, using advanced machine learning techniques, numerous groups have developed predictive models to identify specific contacts between regulatory elements [55–58]. Though not able to construct the whole contact map and 3D chromosome structures, these machine learning approaches have achieved the level of resolution and specificity needed to study functionally important contacts within a TAD. On the other hand, it remains challenging to quantitatively study such functionally important contacts using polymer modeling approaches, though significant progress towards that direction is being made [39,59–63]
The difficulty in predicting contacts between specific regulatory elements using polymer models is at least twofold. First, existing phase separation models based on A/B compartments or six subcompartments are inadequate for such purposes, despite their success in recapitulating the long-range block-wise patterns observed in Hi-C. As chromosome compartments are defined based on contact patterns revealed by Hi-C at a coarse resolution from 50kb to 1 Mb, they tend to group many regulatory elements together as one “active” type and fail to capture the distinction among them [6,7,47]. The ambiguity of these compartments significantly limits the accuracy of polymer models built upon them. To study enhancer-promoter interactions, one must introduce new chromatin types at a higher resolution to achieve the required specificity. How to define these types and how many types are needed remain unclear. Secondly, even with our current understanding of chromatin folding mechanisms, developing a quantitative polymer model to predict contact probability between pairs of genomic loci is still a non-trivial task. In particular, robust and efficient schemes are needed to derive parameters of polymer models to ensure their accuracy.
In this paper, we report the development of a predictive and transferable polymer model to simulate the structure and dynamics of chromosomes at five kilo base resolution. This model takes combinatorial patterns of epigenetic marks and genomic location and orientation of CTCF binding sites as input, and can be parameterized from Hi-C data with a robust and efficient maximum entropy approach [64,65]. A key innovation of this model is its use of chromatin states to capture the wide variety of regulatory elements and to probe their interactions. Computer simulations of this model provide a high-resolution structural characterization of chromatin loops, TADs, and compartments, and succeed in quantitatively reproducing contact probabilities and power-law scaling of 3D contacts as measured in Hi-C and super-resolution imaging experiments. Many significant enhancer-promoter contacts can be captured in simulated contact maps as well. As the model incorporates ingredients from both the extrusion and the phase separation mechanism, its success in quantitative predictions of genome organization provides strong support for such mechanisms. In the meantime, detailed analysis of the model parameters further reveals a significant difference between the interactions that stabilize TAD and those that drive compartmentalization, providing additional insight into chromatin folding not appreciated in existing modeling efforts. Finally, we demonstrate that the model is transferable across chromosomes and cell types, setting the stage for de novo prediction of the structural ensemble for any given chromatin segment using only one-dimensional sequencing data that is available for hundreds of cell types.
We introduce a predictive model to study cell-type specific 3D chromatin folding. This model takes a sequence of chromatin states derived from genome-wide histone modification profiles and a list of CTCF binding sites as input. We selected these genomic features due to their known roles in organizing the chromatin at various length scales (Fig 1A). At the core of this model is an energy function—a force field—that is sequence specific and ranks the stability of different chromatin conformations. Starting from the input for a given chromatin segment, we use molecular dynamics simulations to explore chromatin conformations dictated by the energy function and to predict an ensemble of high-resolution structures. These structures can be compared directly with super-resolution imaging experiments or converted into contact probability maps for validation against genome-wide chromosome conformation capture (Hi-C) experiments.
As shown in Fig 1B, a continuous genomic segment is represented as beads on a string in this model. Each bead accounts for five-kilo bases in sequence length and is assigned with a chromatin state derived from the underlying combinatorial patterns of 12 key histone marks. Chromatin states are known to be highly correlated with Hi-C compartment types [39,54,66] and, therefore, will help model large-scale chromosome compartmentalization. In the meantime, chromatin states can go beyond traditional A/B compartments or subcompartments to provide polymer models with the specificity needed for studying interactions between regulatory elements. We define a total of 15 chromatin states, identified using a hidden Markov model [67], to distinguish promoters, enhancers, heterochromatin, quiescent chromatin, etc (see Methods). Detailed histone modification patterns for these chromatin states are shown in Fig 1C. We note that 15 is large enough to capture the diversity of epigenetic modifications while still being small enough to ensure a sufficient population of each state for a robust inference of interaction parameters between them (Figure A1 in S1 Supporting Information). We further studied a hidden Markov model with 20 states, and found that further increasing the number of states does not lead to a discovery of additional epigenetic classes with significant populations (Figure A2 in S1 Supporting Information). A polymer bead is further labeled as a CTCF site to mark chromatin loop boundaries if both CTCF and cohesin molecules are found to be present in the corresponding genomic region. We define the orientation of these CTCF sites by analyzing the underlying CTCF motif and the relative position of CTCF molecules with respect to cohesin. Details for the definition of CTCF binding sites are provided in Methods.
The potential energy for a given chromatin configuration r is a sum of three components, and UChrom(r) = U(r) + UCS(r) + UCTCF(r). U(r) is a generic polymer potential that is included to ensure the continuity of the chromatin, and to enforce excluded volume effect among genomic loci. UCS(r) is a key innovation of the chromatin model, and is crucial to capture the formation of TADs and compartments. It quantifies the chromatin state specific interaction energies between pairs of loci. As detailed in Section: Physical principles of chromatin organization and Methods, we used a general form for UCS(r) to capture its dependence on genomic separation. UCTCF(r) is inspired by the loop extrusion model [29–31], and facilitates the formation of loop domains enclosed by pairs of CTCF binding sites in convergent orientation (Fig 1A). Both UCS(r) and UCTCF(r) contain adjustable parameters that can be derived from Hi-C data following the optimization procedure developed by one of the authors [64,65]. Segments of chromosomes 1, 10, 19 and 21 from GM12878 cells were used for parameterization to ensure a sufficient coverage of all chromatin states (see Figure A1 in S1 Supporting Information). Detailed expressions for the potential energy, and the parameterization procedure are provided in Methods and in the S1 Supporting Information.
Using the parameterized energy function, we simulated the ensemble of chromatin structures and determined the corresponding contact probability map for a 20 Mb region of chromosome 1 from GM12878 cells. As shown in Fig 2A, the simulated contact map is in good agreement with the one measured by Hi-C experiments from Ref. [7] and reproduces the overall block-wise checkerboard pattern that corresponds to the compartmentalization of chromatin domains. A zoomed-in view along the diagonal of the contact map provided in Fig 2B and 2C further suggests that chromatin TADs and loops are also well reproduced. Similar comparisons for other chromosomes used in parameterizing the model are provided in Figure B in S1 Supporting Information. We note that the length 20 Mb was chosen for computational efficiency, but the model can be easily generalized to longer chromatin segments (see Figure C in S1 Supporting Information).
To go beyond the visual inspection and quantify the correlation between simulated (GM-Sim) and experimental (GM-Exp) contact maps, we calculated the Pearson correlation coefficient (PCC) between the two for chromosome 1 and found that it exceeds 0.96. Importantly, this number is higher than the PCC (0.94) between GM-Sim and Hi-C data from IMR90 cells (IMR-Exp). Breaking down the PCC at different genomic separations also supports that GM-Sim is more correlated with GM-Exp at all ranges than with IMR-Exp (Figure D in S1 Supporting Information). In addition, we also determined the stratum-adjusted correlation coefficient (SCC) that takes into account the distance-dependence effect of contact maps by stratifying them according to the genomic distance [68], and obtained 0.7 for GM-Sim/GM-Exp, and 0.66 for GM-Sim/IMR-Exp. Therefore, SCC analysis also validates our model’s ability in reproducing Hi-C contact maps and in capturing the distinction between cell types. We note that the magnitude of SCC can be sensitive to the smoothing parameter used in its calculation and should be interpreted with caution (Figure E in S1 Supporting Information).
We further examined the agreement between simulated and experimental contact maps using multiple feature-specific metrics. First, we define the contact enhancement for a pair of genomic loci as the ratio of their contact probabilities over the mean contacts averaged over a locally selected background region (see Figure F1 in S1 Supporting Information). The contact enhancement for chromatin loops from chromosome 1 is always larger than one, indicating a strong enhancement of spatial colocalization between loop anchors. Furthermore, over 74% of the loop pairs exhibit a contact enhancement that is larger than the 90th percentile of the distribution for random genomic pairs. These random pairs are selected regardless of CTCF occupancy but with comparable sequence separations as those found in chromatin loops. Therefore, if we use the 90th percentile of the random distribution as a threshold (1.16) and predict every convergent CTCF pairs as loops, the prediction will have a false negative rate of 26%, and a false positive rate less than 10%. The false positive value is an upper bound since most of the random pairs are not flanked with convergent CTCF. The sensitivity of chromatin loop predictions on the threshold is shown in Figure F2 in S1 Supporting Information. It is worth pointing out that the contact enhancement for chromatin loops calculated using Hi-C data is in general larger than simulated values and separates better from that for random pairs (Figure F3 in S1 Supporting Information). The overlap between the two distributions in our simulation is due to that random pairs include a significant fraction of convergent CTCF pairs whose contacts are enhanced as a result of the potential UCTCF(r). Many of these pairs, however, are not recognized as loops in Hi-C, and more advanced algorithms than simple binding site orientations are probably needed to identify loop forming CTCF pairs [69].
To go beyond CTCF mediated contacts and evaluate our model’s ability in reproducing strong interactions between genomic loci, we selected statistically significant contact pairs from simulated and experimental contact maps for chromosome 1 using the software Fit-Hi-C [48] (Figure G in S1 Supporting Information). As a quantitative metric, we define the matching score as the percent of experimental pairs that can be found in the list extracted from simulation. The reverse matching score can be similarly defined as the percent of simulated pairs found in the experimental list. The matching score for the top 1000 chromatin contacts is determined to be 46% and 52% for the reverse matching. To examine specific interactions between regulatory elements, we performed a similar analysis by selecting the top 100 enhancer (state: EnhW1)-promoter (state: PromD1) pairs with highest contact probabilities based on simulated and experimental contact maps. We find that over 70% of experimental pairs are captured in our simulation for chromosome 1. These results suggest that our model based on chromatin states and CTCF mediate interactions is able to reproduce a large fraction of significant contacts detected in Hi-C experiments. Further improving the model’s ability in predicting functionally important pairs would potentially require considering the effect of other proteins, such as YY1 that are known to mediate chromatin interactions [70], and will be an interesting future direction.
We next determined the correlation coefficients between the top five eigenvectors for simulated and experimental contact matrices. As shown in Figure H in S1 Supporting Information, the contact maps reconstructed using only these eigenvectors recapitulate the formation of TADs and compartments observed in the original maps. The high correlation between simulated and experimental eigenvectors (with PCC at approximately 0.8) supports that the corresponding features are well captured by the computational model, and confirms the qualitative observations from Fig 2 and Figure B in S1 Supporting Information.
To more closely examine the quality of simulated TADs, we calculated the insulation profile by sliding a uniform 500kb × 500kb square along the diagonal of the contact matrix and averaging over all contacts within the square. The minima of this profile can be used to identify TAD boundaries as inter-TAD contacts are sparser compared to intra-TAD contacts, resulting in a drop in the insulation score profile as the sliding window crosses TAD boundaries [71]. The PCC between experimental and simulated insulation profiles for chromosome 1 is 0.7. We find that the matching score for TAD boundaries is 80% and 100% for the reverse matching. As another independent validation, we determined TAD boundaries using the software TADbit [43], and found that the simulated results again match well with experimental ones (see Figure I in S1 Supporting Information).
To demonstrate the transferability of the computational model across chromosomes and cell types, we performed additional simulations for chromosomes from GM12878, K562, and Hela cells, whose Hi-C data were not included during the parameterization procedure. As shown in Fig 3 and Figure J in S1 Supporting Information, these de novo predictions are in good agreement with experimental results as measured by PCC (Fig 3B) and SCC (Fig 3C) between experimental and simulated contact maps, matching score between TAD boundaries detected from the insulation profile (Fig 3D) and from TADbit (Figure K1A in S1 Supporting Information), PCC between experimental and simulated insulation profiles (Figure K1D in S1 Supporting Information), matching score between significant contacts detected using Fit-Hi-C (Fig 3E), matching score between interacting enhancer-promoter pairs (Figure K2C in S1 Supporting Information), correlation coefficients of the top five eigenvectors (Fig 3F and Figure H in S1 Supporting Information), and false negative rate of loop predictions (Fig 3F). Furthermore, the model succeeds in revealing the cell-type specificity of Hi-C contact maps, and the simulated contact maps are always more correlated with the corresponding experimental data from the same cell type than with those from IMR90 cells (light colors in Fig 3B and 3C). The matching scores between experimental and simulation results are also significantly higher than those calculated between experimental and control data (light colors in Fig 3D and 3E), which were obtained by randomly shuffling the size of loops/enhancer-promoter pairs/TADs along the chromosome while keeping their total number unchanged. The success of these de novo predictions supports that the chromatin-state-based model introduced here provides a consistent description of the 3D genome organization across cell types.
We next analyze the simulated 3D structural ensembles to gain additional insights on chromatin organization. Consistent with previous experimental and theoretical studies [37,72,73], our model reproduces the clustering of active chromatin state and their preferred location at the exterior of chromosomes (Figure L in S1 Supporting Information).
Super-resolution imaging experiments probe chromatin organization in 3D space to quantify spatial distances between genomic segments. These 3D measurements can be compared directly with simulated chromatin structures, and thus provide a crucial validation of the computational model parameterized from Hi-C experiments with independent datasets. To understand the overall compactness of various chromatin types, we selected a set of active, repressive and inactive chromatins and determined their radiuses of gyration from the ensemble of simulated structures. These different chromatin types are identified using two key histone marks H3K4me2 and H3K27me3 (Fig 4A). The complete list of chromatin domains with their genomic locations is provided in the Extended Data Sheet. As shown in Fig 4B, the radius of gyration increases at larger genomic separation following a power law behavior in all cases with exponents of 0.34, 0.31 and 0.23 for the three chromatin types respectively. These scaling exponents are in quantitative agreement with imaging measurements performed for Drosophila chromosomes [12] and support the notion that active chromatins adopt less condensed conformations to promote gene activity. Consistent with the imaging study performed on chromosome 21 from IMR90 cells [13,20], we also observe a strong correlation between Hi-C contact probabilities and spatial distances for pairs of genomic loci (Fig 4C).
One of the most striking features revealed by high-resolution Hi-C experiments is the formation of chromatin loops anchored at pairs of convergent CTCF sites [7,10,74,75]. Microscopy studies that directly visualizes 3D distances using fluorescence in situ hybridization (FISH) methods further find that these loops are dynamic, and despite their high contact frequencies, loop anchors are not in close contact in every cell [16,41,76]. Consistent with their dynamic nature, chromatin loops in our simulation adopt flexible conformations as well. As shown in Fig 5A, for the loop formed between chr1:39.56–39.73 Mb, we observe a large variance in the probability distribution of its end-to-end distances. Additional results for other loop pairs are provided in Figure M in S1 Supporting Information. Two example configurations of the loop domain with distance at 0.08 and 0.24 μm are shown in the inset. A systematic characterization of all the loops identified in Ref. [7] for the simulated chromatin segment shows that the conformational flexibility is indeed general, though there is a trend in decreasing variance for loops with larger contact probabilities (Fig 5B). We also emphasize that though higher contact probabilities, in general, corresponds to smaller end-to-end distances, their relationship is not strictly monotonic. The opposite correlation can be seen in numerous cases in Fig 5B. Such seemingly paradoxical observations have indeed been found in previous experimental studies that compare 3C with FISH experiment [16,77], and can naturally arise as a result of dynamical looping or loop extrusion [78].
Compared to chromatin loops, TADs are longer and are stabilized by a complex set of interactions [79]. The analysis of their structural ensemble is less straightforward, and the end-to-end distance may not be sufficient for a faithful description of their conformational fluctuation [80]. It is desirable to analyze TAD structures using reaction coordinates that not only help to distinguish different clusters of chromatin conformations, but can also provide insight into the mechanism of TAD folding and formation. Borrowing ideas from protein folding studies, we approximate these reaction coordinates using collective variables with slowest relaxation timescales as determined following the diffusion map analysis [81,82]. Progression along these variables approximates well the most likely transition between two sets of structures and can, therefore, shed light on the pathway for conformational rearrangements. Diffusion map analysis has been successfully applied to a variety of systems to provide mechanistic insights on the conformational dynamics involved in protein folding, ligand diffusion, etc. [83,84].
We applied the diffusion map technique to the predicted structural ensemble of the genomic region chr1:34–38 Mb from GM12878 cells that consists of three visible TADs. As shown in Fig 6, several basins are observed in the probability distribution of chromatin conformations projected onto the first two reaction coordinates, suggesting the presence of multiple stable TAD structures, rather than a unique one. Conformational heterogeneity in TADs has indeed been observed in a recent super-resolution imaging study that characterizes single cell chromatin structures [20]. To gain physical intuition on the reaction coordinates and insight on the transition between TAD structures, we calculated the corresponding contact maps at various values of these coordinates. As shown in the top panel, reaction coordinate one captures the formation of contacts between TAD1 and TAD3 while the structures for all three TADs remain relatively intact. On the other hand, progression along reaction coordinate two (left panel) leads to significant overlaps between TAD1 and TAD2. Interaction between TAD2 and TAD3 can also be observed along a third coordinate as shown in Figure N in S1 Supporting Information. Example structures for the three TADs in various regions are also provided on the right panel. These results are consistent with the notion that TADs are stable structural units for genome organization [79], but also suggest the presence of significant cross-talk among neighboring TADs [85].
Though the exact molecular mechanism and driving force for chromatin folding remain elusive, it is becoming increasingly clear that different molecular players are involved in organizing the chromatin at various length scales [49,60,86,87]. For example, transcription factors and architectural proteins are critical in stabilizing the formation of chromatin loops and TADs [4,33,79]. On the other hand, nuclear compartments, such as the nucleolus and the nuclear envelope, contribute to chromatin compartmentalization and mediate contacts among chromatin domains separated by tens of Mb in sequence [50,88]. We expect that these different molecular mechanisms will give rise to distinct interaction energies at various genomic length scales. For example, for the same pair of chromatin states, as the genomic separation between them is varied, the interaction energy that stabilizes their contact should vary. In the following, we examine the dependence of inferred contact energies on genomic separation to reveal the principles of genome organization.
Fig 7A presents the derived contact energies among chromatin states UCS(r) at various genomic separations (500kb, 1.5 Mb, 4 Mb and 10 Mb from left to right), with blue and red for attractive and repulsive interactions respectively. A notable feature for all four length scales is the clear partition of chromatin states into at least two groups that correspond to well-known active and repressive chromatins respectively. For example, attractive interactions are observed among the top half chromatin states that include promoters (PromD1, PromU), enhancers (TxEnh5, Enhw1) and gene body (Tx), and for the bottom half that includes inactive chromatin (Quies), polycomb repressed domain (ReprPC) and heterochromatin (Het). The unfavorable interactions among active and repressive chromatins will drive their phase separation shown in Fig 2D and Figure L in S1 Supporting Information. Partitioning of chromatin states into active and inactive groups is also evident from the dendrogram shown in Fig 7B, and the eigenvectors for the largest in magnitude eigenvalue of the interaction matrices shown in Fig 7C.
Despite their overall similarities, the interaction energies at various genomic separations differ from each other. To quantify their differences, we determined the pairwise Pearson correlation coefficients between the interaction matrices. As shown in Fig 7C, the interactions that are responsible for TAD formation (~ 1 Mb) indeed differ significantly from those that lead to chromatin compartmentalization (~ 10 Mb), as evidenced by the low correlation among them. Strikingly, the correlation coefficient between interaction matrices at 4 Mb and 10 Mb exceeds 0.9, indicating the convergence of chromatin interactions at large genomic separation.
We further compared the complexity of the interaction matrices by calculating the ratio of the first n eigenvalues over the sum of all eigenvalues. Fig 7D plots this complexity measure as a function of n, and absolute values of the eigenvalues were used to calculate the measure. For all three matrices with genomic separation larger than 1 Mb, we find the top first six eigenvectors can explain a large fraction of their complexity (over 80%). This observation is consistent with the success of our previous effort in modeling chromatin organization with six compartment types [37]. However, more eigenvectors are needed, especially for short range in sequence interactions, to capture the full matrix complexity. These results together highlight the presence of distinct mechanisms that fold the chromatin at various genomic separations, and argues the importance of using sequence length dependent contact energies.
We introduced a novel computational model for studying 3D genome organization by integrating bioinformatics analysis with polymer modeling. This integration brings together the best of both worlds and results in a powerful predictive tool. Similar to bioinformatics approaches, our model succeeds in identifying cell-type specific interactions between regulatory elements. As in polymer modeling, the availability of 3D chromosome conformations makes it possible to characterize contacts between any genomic segments and construct the whole contact map, to study global properties of the genome organization that involve many-body interactions, and to explore the physical mechanism and driving force of genome folding.
This predictive model presents a significant improvement from our previous effort in simulating chromatin structures [37] by switching the input from compartment types to chromatin states. In particular, unlike compartment types that are results from clustering Hi-C contact matrices [7], chromatin states are defined as combinational patterns of histone modification profiles. Uncoupling the input from Hi-C data is critical to ensure that the model is genuinely predictive. Furthermore, chromatin states allow us to model chromatin structures at a much higher resolution (5kb) to provide a detailed structural characterization of chromatin loops and TADs, and to resolve long-range specific contacts between promoters and enhancers. On the other hand, chromatin models based on compartment types are inherently limited to 50kb [37,39], a resolution at which compartment types can be robustly derived from Hi-C data [7]. Finally, as shown in Fig 7, the novel sequence-separation dependent contact potential developed here enables a rigorous assessment of the number of “types” needed for modeling chromatin structures, and suggests that the six compartment types are insufficient for an accurate description of TAD formation. Since the data required to define chromatin states are available for hundreds of cell types via the epigenome roadmap project [89], we anticipate a straightforward application of the model developed here to characterize the differences of chromatin structures across cell types and to understand the role of 3D genome organization in cell differentiation and cell fate establishment.
Histone modifications have long been recognized as crucial for the genome’s function [90]. The “histone code” hypothesis was proposed to rationalize the presence of numerous types of histone marks and the importance of their combinatorial roles [91]. However, a mechanistic understanding of the relationship between these chemical modifications and the functional outcome remains lacking [92]. The success of the computational model introduced here in predicting chromatin structures argues for the importance of histone modifications in organizing the genome. It is tantalizing to hypothesize that the histone code can be understood from a structural perspective. Epigenome engineering experiments that perturb histone modifications at specific genomic locations will be helpful to elucidate further whether the relationship between 1D histone modifications and 3D genome organization is causal.
The energy function of the chromosome model, which can be rigorously derived following the maximum entropy principle [64,65], adopts the following form
UChrom(r)=U(r)+∑I,J∑i∈I∑j∈JαIJ(|j-i|)f(rij)+∑K,L∑K≤k<l≤L[αCh,Ch+αC,Ch+αC,C)f(rkl].
U(r) defines the generic topology of the chromosome as a confined polymer with excluded volume effect. The second term incorporates the sequence length dependent contact energies αIJ (|j − i|) between pairs of loci i, j characterized with chromatin states I, J respectively. As discussed in the main text, the dependence of contact energies on sequence length separation is crucial to reproduce the hierarchical genome organization, and to detect independent mechanisms of chromatin folding at different length scales. f(rij) measures the contact probability between a pair of loci i and j separated by a distance rij, and is defined as follows
f(r)={12[1+tanh(σ(rc-r))],ifr≤rc12(rcr)4,forr>rc#
Where rc = 1.76 and σ = 3.72. As shown in Figure O in S1 Supporting Information, compared to a simple hyperbolic tangent function used in previous studies [64,65], the new expression decays to zero for large distances r at a slower rate. This new form is motivated by the power law relationship between spatial distances and Hi-C contact probabilities observed in Ref. [13].
Finally, the last term, inspired by the recently proposed extrusion model [29–31], is included to model the formation of chromatin loops. In particular, the genomic segment enclosed by a pair of convergent CTCF binding sites experiences a condensing potential due to the binding of cohesin molecules. We limit this potential to convergent CTCF pairs that are separated by no more than 4 CTCF binding sites with 5’– 3’ orientation or 4 CTCF binding sites with 3’– 5’ orientation to mimic the finite processivity of cohesin molecules [30]. For generality, three different potentials are used for CTCF-CTCF interaction (αC,C), CTCF-chromatin interaction (αC,Ch) and chromatin-chromatin interaction (αCh,Ch).
The explicit mathematical expression for UChrom(r) is provided in the SI. It contains a total of 1883 parameters. This seemingly large number is a result of our use of chromatin states and the dependence of their interaction energies, αIJ (|j − i|), on genomic separation. Both of these two features are innovations of our model to predict specific contacts between enhancers and promoters, and to capture the different biological mechanisms for TAD formation and chromosome compartmentalization. We emphasize that since a specific experimental constraint can be defined for each one of these parameters, their values can be derived robustly and efficiently using the iterative maximum entropy algorithm introduced by Zhang and Wolynes [64]. As proven before, the value of these parameters in principle is unique [76]. Numerical values of the parameters are provided in the Extended Data Sheet.
After a careful analysis of the interaction energies shown in Fig 7, however, we believe that the number of parameters could potentially be significantly reduced without sacrificing the model accuracy. In particular, the number of chromatin states used here is probably “too many” since the complexity of the interaction energy matrices can be well explained with the top 10 eigenvectors. Furthermore, the interaction energies also converge at larger genomic separation, making its dependence on |j − i| unnecessary. These insights will prove useful for future chromatin modeling efforts.
We carried out constant temperature simulations to predict chromatin structures consistent with the energy function UChrom(r) using the molecular dynamics software package LAMMPS [93]. For each contact map presented in the manuscript, a total of eight independent 40-million-timestep long simulations were performed to ensure sufficient statistics. On an Intel Xeon E5-2690 v4 2.6GHz node with 14 cores, each one of such simulations takes approximately 30 hours to finish. More details on the simulation are provided in the supporting information.
To enable a quantitative comparison between simulated chromatin structures with microscopy imaging data, we estimate a 5kb long genomic segment with a width of 30 nm and a length of 60 nm based on a high-resolution chromatin structure characterized by cryogenic electron microscopy (Cryo-EM) technique [94].
Experimental contact maps at 5kb resolution from Ref. were downloaded using the Gene Expression Omnibus (GEO) accession number GSE63525 (see Extended Data Sheet). We used the combined contact matrices constructed from all read pairs that map to the genome with a MAPQ> = 30. The raw matrices were then normalized with the KR method using the normalization vector provided in the same dataset. To convert the contact matrices into probabilities, we further divided each matrix element with the diagonal value Cii = 1035 obtained from averaging over all chromosomes. With this probability conversion, all the genomic segments that are within in 5kb along the sequence will on average have a contact probability of 1. Since in the computational model, a 5kb segment has a diameter of σ = 30 nm, this probability conversion is equivalent of specifying the contact probability as 1 for genomic loci that are within a spatial distance of 30 nm. Such a probability definition is indeed consistent with the contact function f(r) defined in Eq. [3] and plotted in Figure O in S1 Supporting Information.
A key input of the computational model is the sequence of chromatin states that captures the variation of epigenetic modifications along the genome sequence. Following Ref. [67], we defined chromatin states as the set of unique combinatorial patterns of histone marks. Using a multivariate hidden Markov model that maximizes the posterior probability of assigning a hidden state to each genomic segment given the sequence of observed histone modifications [95], we derived 15 chromatin states from genome-wide profiles of 12 key histone marks collected from six cell types that include GM12878, K562, HeLa, H1hesc, Huvec and Hepg2. A single set of chromatin states is crucial to ensure the transferability of the parameterized force field across cell types. The dataset used for chromatin state inference is listed in the Extended Data Sheet. Detailed histone modification patterns for these chromatin states are shown in Fig 1C. With the set of chromatin states specified, every five-kilo-base long segment can then be assigned to a chromatin state based on its histone modification profiles, and a sequence of chromatin states for the entire chromatin segment can be defined as the simulation input.
To capture the formation of chromatin loops, we compiled a list of CTCF-binding sites along the chromatin of interest using cell-type specific ChIP-Seq data.
Starting from the peak profile downloaded from ENCODE (see Extended data sheet), we identified the center of binding for each peak of both CTCF and cohesin subunit Rad21. As both CTCF and cohesin molecules are found at the boundaries of most chromatin loops, we selected loop forming CTCF binding sites as those that have at least one Rad21 molecule located within 50bp of their genomic locations.
We then determined the orientation of each CTCF-binding site as follows. We first attempted to align the binding sites to the set of CTCF motifs compiled in Refs. [7] and [96] (see Extended data sheet). If the alignment succeeds and a motif is found within 100bp of the binding site, the orientation of the binding site was then assigned based on the DNA sequence of that motif. If no motif can be aligned, the orientation of the CTCF-binding site is determined using the genomic location of its binding center relative to that of the nearest binding center of Rad21. For example, we assign the orientation as 5’– 3’ if the nearest Rad21 binding center is in the downstream of the CTCF binding site; otherwise, the orientation is assigned as 3’– 5’.
The above procedure will result in a list of oriented CTCF sites at single base resolution. From this list, we defined a 5kb-long bead in the computational model as a CTCF site if there is at least one CTCF binding site falls into the genomic region enclosed by that bead. If all the CTCF sites within the 5kb region have the 5’– 3’ orientation, then the bead is assigned with the 5’– 3’ orientation; similarly, if all the CTCF sites within the 5kb region have the 3’– 5’ orientation, then the bead is assigned with the 3’– 5’ orientation. If CTCF sites with both orientations are present, then the bead is assigned with dual orientation as well.
For molecular systems that exhibit a separation of timescales, it is often desirable to approximate their dynamics at long time limit with a handful of slow variables. The time evolution of these slow variables should be Markovian and independent of the fine details of the high dimensional system to capture the dynamical behavior of the system on a coarsened timescale. Mathematically it has been proven that an optimal choice of these slow variables is the first few eigenfunctions of the backward Fokker–Planck diffusion operator [81]. Diffusion map is a data-driven approach that approximates these eigenfunctions and therefore the slow variables by defining a random walk process on the simulation data [97].
In particular, for N chromatin configurations selected from the simulated structural ensemble, we first constructed a transition probability matrix K for the random walk by defining its elements as
Kij=exp(-dij2ϵiϵj).
The eigenfunctions of the above transition matrix can be shown to converge to that of the Fokker–Planck operator in large N limit. The distance between two configurations dij was calculated as the mean difference of their corresponding contact probability maps. We followed the algorithm proposed in Ref. [82] to normalize the matrix and to estimate ϵi. From the normalized transition matrix, we then determined its eigenfunctions and used the top two with the smallest non-zero eigenvalues as the reaction coordinates shown in Fig 6 (see Figure N in S1 Supporting Information for eigenvalues).
|
10.1371/journal.pgen.1005426 | mTORC1 Prevents Preosteoblast Differentiation through the Notch Signaling Pathway | The mechanistic target of rapamycin (mTOR) integrates both intracellular and extracellular signals to regulate cell growth and metabolism. However, the role of mTOR signaling in osteoblast differentiation and bone formation is undefined, and the underlying mechanisms have not been elucidated. Here, we report that activation of mTOR complex 1 (mTORC1) is required for preosteoblast proliferation; however, inactivation of mTORC1 is essential for their differentiation and maturation. Inhibition of mTORC1 prevented preosteoblast proliferation, but enhanced their differentiation in vitro and in mice. Activation of mTORC1 by deletion of tuberous sclerosis 1 (Tsc1) in preosteoblasts produced immature woven bone in mice due to excess proliferation but impaired differentiation and maturation of the cells. The mTORC1-specific inhibitor, rapamycin, restored these in vitro and in vivo phenotypic changes. Mechanistically, mTORC1 prevented osteoblast maturation through activation of the STAT3/p63/Jagged/Notch pathway and downregulation of Runx2. Preosteoblasts with hyperactive mTORC1 reacquired the capacity to fully differentiate and maturate when subjected to inhibition of the Notch pathway. Together, these findings identified the role of mTORC1 in osteoblast formation and established that mTORC1 prevents preosteoblast differentiation and maturation through activation of the Notch pathway.
| The coordinated activities of osteoblasts and osteoclasts in bone deposition and resorption form the internal structure of bone. Disruption of the balance between bone formation and resorption results in loss of bone mass and causes bone diseases such as osteoporosis. Current therapies for osteoporosis are limited to anti-resorptive agents, while bone diseases due to reduced osteoblast activity, such as senile osteoporosis, urgently require targeted treatment and novel strategies to promote bone formation. mTORC1 has emerged as a critical regulator of bone formation and is therefore a potential target in the development of novel bone-promoting therapeutics. Identifying the detailed function of mTORC1 in bone formation and clarifying the underlying mechanisms may uncover useful therapeutic targets. In this study, we reveal the role of mTORC1 in osteoblast formation. mTORC1 stimulated preosteoblast proliferation but prevented their differentiation and attenuated bone formation via activation of the Notch pathway. Pharmaceutical coordination of the pathways and agents in preosteoblasts may be beneficial in bone formation.
| The skeleton is a highly specialized and dynamic structure undergoing constant remodeling [1]. The remodeling process is executed by temporary cellular structures that comprise teams of coupled osteoblasts and osteoclasts. The rate of genesis as well as death of these two cell types is vital for the maintenance of bone homeostasis [2], and common metabolic bone disorders such as osteoporosis are largely caused by a derangement in the proliferation, differentiation or apoptosis of these cells [3].
Osteoblasts, which are the chief bone-making cells, differentiate and produce bone matrix during skeletal development [4]. The differentiation process of osteoblasts is often divided into stages of mesenchymal progenitors, preosteoblasts and osteoblasts (often called mature osteoblasts) [5]. Osteoblasts are often characterized by the expression of osteocalcin, while preosteoblasts are usually considered to express the transcription factor Runx2 or both Runx2 and osterix (Osx). Preosteoblasts have been shown to actively divide in vitro and are multipotent in differentiating, thus proliferative expansion and osteoblastic differentiation of preosteoblasts are essential for bone formation. Understanding the intracellular signaling pathways that control preosteoblast proliferation and differentiation is critical for preventing bone loss-related disease caused by impaired bone formation such as senile osteoporosis.
The mechanistic target of rapamycin (mTOR) is a conserved Ser/Thr kinase nucleating at least two distinct multi-protein complexes, mTOR complex 1 (mTORC1) and mTOR complex 2 (mTORC2) [6]. mTORC1 uniquely contains raptor and is the sensitive target of rapamycin, it integrates both intracellular and extracellular signals, including growth factors, nutrients, energy levels, and cellular stress [7]. The tuberous sclerosis 1 (TSC1)-TSC2-TBC1D7 (TSC-TBC) complex is the major upstream inhibitory regulator of mTORC1 [8], and loss of this complex causes cells and tissues to display constitutive mTORC1 activation. TSC-TBC accelerates the intrinsic rate of GTP hydrolysis of Rheb, converting Rheb from the GTP-bound (active) to the GDP-bound (inactive) form. The active GTP-bound form of Rheb directly interacts with mTORC1 to stimulate its activity. mTORC1 phosphorylates the translational regulators, eukaryotic initiation factor 4E-binding protein-1 (4E-BP1) and S6 kinase 1 (S6K1), to regulate biosynthesis of proteins and modulates autophagy, biosynthesis of lipids and organelles and mitochondrial metabolism as well, through which mTORC1 exerts an essential role in regulating cell metabolism, survival, growth and proliferation [6,9].
mTORC1 signaling has emerged as a critical regulator of bone formation. Patients with tuberous sclerosis due to mutation in TSC1/2 present with sclerotic and lytic bone changes [10,11,12,13,14]. Moreover, mTOR has recently been identified among genes and pathways that are connected with human skeletal growth [15]. However, results from in vitro and in vivo studies on the function of mTORC1 in osteoblast lineage are inconsistent. The mTORC1 inhibitor, rapamycin, showed controversial capacity to influence the differentiation of various osteoblastic lineage cell lines in vitro [16,17,18,19,20,21,22] and demonstrated inconsistent potential for bone formation in vivo [23,24,25]. In addition to an undefined function in the osteoblast lineage, the mechanisms through which mTORC1 modulates osteoblast differentiation and bone formation are unknown. Using conditional Tsc1 knockout cell and mouse models, we demonstrated that mTORC1 activation is crucial for preosteoblast proliferation, but prevents their differentiation and maturation.
Notch signaling mediates communication between neighboring cells to control cell fate decision [26, 27]. Notch has been reported to be positively regulated by mTORC1 in various cell lines and be responsible for the impaired cell differentiation by mTORC1 [28, 29, 30]. We determined here that it is true in osteoblast lineage cells as well. Our mechanistic studies revealed that Notch signaling is strongly activated by mTORC1 to inhibit osteoblastic transcription factor Runx2 and prevent preosteoblast maturation and bone formation.
mTORC1 is known to promote cell proliferation in many types of cells. To examine the relationship between preosteoblast proliferation and mTORC1 activity, we analyzed the level of mTORC1 during the proliferative expansion of MC3T3-E1 cells, a murine preosteoblast cell line [31], and subsequent cessation of growth. We counted cells each day to monitor their growth and observed that the cells reached confluence after 6–7 days, when proliferation ceased due to contact inhibition (Fig 1A). Western blot analysis reflected this growth inhibition as a decrease in the levels of the cell cycle markers, cyclin D1 and proliferative cell nuclear antigen (PCNA). As expected, the high level of mTORC1 activity (indicated by P-S6K (T389) and P-S6 (S235/236)) decreased when proliferation of MC3T3-E1 cells slowed down and eventually ceased (Fig 1B). Cells did not differentiate during this growth period, as the level of osterix (Osx, a marker of early osteoblasts) was unchanged throughout. These data indicate that the level of mTORC1 activity was positively correlated with the rate of preosteoblast proliferation.
We then induced osteoblastic differentiation in confluent MC3T3-E1 cells. The induced cells showed a decreased mTORC1 activity during osteoblast differentiation in parallel with increases in markers of osteoblast differentiation (i.e. osterix (Osx) and osteocalcin (Ocn)) (Fig 1C). A similar pattern of mTORC1 activity was observed in osteoblastic induced fetal rat calvarial cells (Fig 1D), which implicated that mTORC1 activity is not required for osteoblastic differentiation of preosteoblasts.
We next investigated the role of reduced mTORC1 activity caused by rapamycin in the proliferation and differentiation of preosteoblasts. As shown in Fig 1A, the growth of cells treated with 0.1 nM rapamycin significantly lagged behind control cells, and the decreased level of proliferative markers (cyclin D1 and PCNA) in rapamycin-treated cells revealed the underlying mechanism (Fig 2A). We next determined the role of reduced mTORC1 activity in the differentiation of preosteoblasts. As seen in Fig 2B, a low concentration of rapamycin (0.1nM) increased the expression of osterix and osteocalcin. Separate sets of cells were tested for mineralization capacity, a terminal differentiation parameter for osteoblasts, by staining with alizarin red, and the results confirmed enhanced mineralization of the extracellular matrix (ECM) in MC3T3-E1 cells with impaired mTORC1 activity (Fig 2C). Increased mineralization of ECM was also observed in fetal rat calvarial cells treated with 0.1nM rapamycin (S1 Fig).
To test the results in vivo, we subcutaneously injected newborn wild-type C57BL/6 mice with rapamycin daily (0.1 mg/kg body weight/day) for 10 weeks. mTORC1 activity was efficiently down-regulated in the primary spongiosa as observed by the reduction in immunohistochemistry staining for S6 phosphorylation (Ser235/236) in rapamycin-treated mice (Fig 2D). Micro-CT analysis of the distal femur showed a marked decrease in bone mass in the rapamycin-treated mice (Fig 2E), as demonstrated by a significant decrease in BV/TV, trabecular number or trabecular thickness, coupled with an increase in trabecular separation (Table 1). In line with the reduced trabecular bone mass, mice treated with rapamycin showed decreased cortical thickness as well as smaller outer and inner femoral mid-shaft bone perimeters when compared with controls (Table 1). The decrease in cortical thickness was due to deficient periosteal apposition, as evidenced by a decrease in the outer perimeter of the mid-shaft, while the inner perimeter was also decreased.
We then analyzed the cellular basis of decreased bone mass in rapamycin-treated mice. Tartrate-resistant acid phosphatase (TRAP) staining on femoral sections revealed a reduction in osteoclasts within the trabecular bone region in rapamycin-treated mice (S2 Fig). Thus, decreased bone mass in the rapamycin-treated mice was not caused by an overall decrease in bone resorption. We next investigated bone formation parameters by first examining the total number of osteoblasts. ALP staining of the distal femoral section from rapamycin-treated mice showed a marked reduction in osteoblastic lineage cells when normalized to the bone perimeter (Fig 2F and 2G). Immunohistochemistry staining for BrdU revealed that proliferation of osteoblastic lineage cells was attenuated in rapamycin-treated mice (Fig 2H and 2I). To determine if differentiation of osteoblasts was affected, we next measured their numbers at different stages of differentiation by immunostaining femur sections from rapamycin-treated mice and controls. The number of osterix-positive preosteoblasts (Fig 2J and 2K) and osteocalcin-positive mature osteoblasts (Fig 2L and 2M) on the bone surfaces of rapamycin-treated mice were less than those in the controls. However, the mean density of osterix and osteocalcin-positive cells was increased in mice treated with rapamycin (Fig 2K and 2M), indicating that expression of osterix and osteocalcin in single osteoblasts was increased in rapamycin-treated mice. Thus, in agreement with the in vitro results, rapamycin attenuated proliferation of preosteoblasts but promoted their differentiation in vivo.
To characterize the role of mTORC1 activation in the regulation of proliferation and differentiation of osteogenic progenitors, we generated mice in which mTORC1 were selectively activated in osteoprogenitor cells committed to the osteoblast lineage. To achieve such cell type-specific knockout, we crossed floxed Tsc1 mice with Osx-GFP::Cre mice (which express a GFP-Cre fusion protein under the direction of the Osx1 promoter) to generate conditional Tsc1 knockout mice. We mated Osx-GFP::CreTG/+;Tsc1flox/+ mice and selected female mice with the genotype Osx-GFP::CreTG/+;Tsc1flox/flox (hereafter, referred to as ΔTsc1) for detailed analysis. Female Osx-GFP::CreTG/+;Tsc1+/+ littermates served as controls. ΔTsc1 mice were born at the expected Mendelian frequency, and recombination of Tsc1 alleles only occurred in skeletal tissues (i.e., legs and skull) as demonstrated by allele specific PCR (Fig 3A). Immunohistochemical staining of distal femur sections showed a dramatic increase in S6 phosphorylation (Ser235/236) in ΔTsc1 mice (Fig 3B), indicating that mTORC1 was activated by Tsc1 disruption.
At the gross level, 4-week-old ΔTsc1 mice demonstrated square skulls and prominent dwarfism (Fig 3C and 3D). The body weight of the Tsc1-deficient mice was significantly lower than that of their control littermates, indicating retardation of growth in the mutant mice (Fig 3E). Whole body X-ray analysis showed increased bone mass in the skull, vertebrae and long bones (Fig 3D). Because of thickened cortical bone, pale bones were observed in ΔTsc1 mice (S3 Fig). Micro-CT analysis of the distal regions of the femur confirmed a marked increase in cancellous bone mass in ΔTsc1 mice (Fig 3F), as reflected in a 240%, 140% and 170% increase in BV/TV, trabecular number and trabecular thickness, respectively, coupled with a 50% decrease in trabecular separation. Analysis of femoral mid-shaft revealed a similar increase in cortical bone mass in ΔTsc1 mice, as reflected in a 230% and 130% increase in cortical thickness and outer perimeter, respectively, and a 90% decrease in inner perimeter (Table 2). Although bone mass was increased in ΔTsc1 mice, micro-CT radiograms showed more porous areas in the bones of mutant mice (Fig 3F), which resulted in a uniform 90% decrease in bone mineral density (BMD) in cancellous and cortical bone (Table 2). The appearance of more porous areas in mutant mice was the result of increased areas of hypomineralization. As shown by Goldner’s Masson trichrome staining, ΔTsc1 mice had 35% more osteoid/hypomineralized areas (stained red) in bone (Fig 3G and 3H). Together, these data suggest that mTORC1 activation in osteogenic progenitors stimulates these progenitors to produce increased amounts of immature woven bone.
We then analyzed the cellular basis for the increased amounts of immature woven bone in ΔTsc1 mice. ALP staining of 10-week-old ΔTsc1 mouse femurs indicated an increased number of osteoblast lineage cells (Fig 4A and 4B). To determine the underlying molecular mechanism of the increase in osteoblasts in transgenic bone, we cultured postnatal day 3 calvarial osteoblasts and found a significant increase in BrdU-positive cells, indicating increased cellular proliferation in ΔTsc1 calvarial osteoblasts (S4 Fig). In addition, the percentage of proliferative osteoblasts on bone surface was increased in ΔTsc1 mice (Fig 4C and 4D).
We next determined whether the differentiation of preosteoblasts was influenced in ΔTsc1 mice. Osteoblasts in ΔTsc1 mice lost normal morphology and appeared immature as shown by scanning electron microscopy (SEM) analyses (S5 Fig), indicating a faulty maturation process from precursors to osteoblasts. Immunohistochemical staining of distal femur sections showed decreased expression of osterix (Fig 4E and 4F) and osteocalcin (Fig 4G and 4H), indicating impaired differentiation of preosteoblasts in mutant mice. To detect osteoblast activity, we performed double fluorochrome labeling analyses. Incorporation of the two fluorochromes was evident in the control mice bone, while mutants’ bone displayed diffuse fluorochrome labeling, a characteristic feature of immature woven bone. Although mineralizing surface was dramatically increased abnormally in cortical bone, distance between calcin-labeled mineralization fronts at endosteum of the midshaft of femur was smaller in ΔTsc1 mice than that in the controls (Fig 4I). Histomorphometric measurements showed that the endosteum mineral apposition rate (MAR) of ΔTsc1 mice was lower than that of controls (Fig 4J), suggesting impaired performance of individual osteoblasts in ΔTsc1 mice.
To further examine the impact of mTORC1 activation on osteoblastic precursor differentiation, we used primary calvarial preosteoblast cultures in which Tsc1 was eliminated prior to the induction of osteoblast differentiation due to suppression of Osx-Cre by doxycycline to specifically delete Tsc1. On the day the cells reached confluence (day -3), mTORC1 was not activated (as indicated by unchanged phosphorylation of S6K and S6) and markers of differentiated osteoblasts (Osx and Ocn) remained normal in ΔTsc1 cells compared with the controls. Doxycycline was then discontinued to activate Osx-Cre and delete Tsc1 and osteoblastic differentiation was induced for another 14 days. Total proteins were extracted for western blot on different days after induction (day 0, 7, 14). Osteoblasts lacking Tsc1 exhibited the expected activation of mTORC1 and reduced expression of osterix and osteocalcin (Fig 4K). Alizarin red staining confirmed that ΔTsc1 mice exhibited a marked decreased capacity to form mineralized nodules (Fig 4L).
Notably, osteoclast numbers on the bone surface (OC.N/B.Pm) were significantly decreased in mutant mice (S6 Fig).These data suggest that mTORC1 activation in preosteoblasts promotes their proliferation but prevents their differentiation and decreases osteoclast number by an undefined mechanism.
To identify whether the increased amount of immature woven bone was mTORC1-dependent, ΔTsc1 mice were administered with rapamycin prenatally from E13.5 (the approximate day of Osx-Cre expression) and then after birth until 10 weeks old. After treatment with 0.1 mg/kg/day rapamycin, S6 phosphorylation (Ser235/236) was significantly decreased in the primary spongiosa (Fig 5A), accelerated proliferation was reduced to normal level (Fig 4C and 4D), increased osteoblastic lineage cells with ALP (Fig 4A and 4B) were reduced to normal, expression of osterix (Fig 4E and 4F) and osteocalcin (Fig 4G and 4H) was increased, diffuse and narrowed distance between fluorochrome labeling was distinct and broadened (Fig 4I and 4J), high bone mass was notably reduced (Fig 5B), narrowed bone marrow cavity was expanded (Fig 5C), areas of hypomineralization were decreased to normal (Fig 5D and 5E) and bone architecture was partially normalized (Table 3). These results indicate that preosteoblasts in ΔTsc1 mice re-acquired the ability to differentiate, and form normal bone. The results of rapamycin treatment in differentiating calvarial cells confirmed these findings. Following treatment with a low concentration (0.1 nM) of rapamycin for 14 days, decreased expression of osterix and osteocalcin was elevated (Fig 5F) and mineralized nodules were formed increasingly (Fig 5G) in ΔTsc1 cells.
We next investigated the mechanism by which mTORC1 regulates osteoblast differentiation. Runx2, a Runt domain-containing transcription factor, is required for commitment of mesenchymal osteochondroprogenitors to the osteoblastic lineage, differentiation into mature osteoblasts and terminal differentiation into osteocytes [32]. ΔTsc1 calvarial cells exhibited a reduced Runx2 expression level, while rapamycin treatment led to an increase in Runx2 above baseline (Fig 6A). The regulation of Runx2 by mTORC1 was further confirmed in ΔTsc1 mice (Fig 6B and 6C). Thus, mTORC1 attenuated the expression of Runx2 in vitro and in vivo, and the delay in osteoblast differentiation was probably due to repression of Runx2 in ΔTsc1 mice.
In consideration of the same negative role of Notch signaling in osteoblast differentiation [32,33,34,35] as mTORC1 signaling and the positive correlation between Notch and mTORC1[28,29,30], we next investigated the potential role of the Notch pathway in the mechanism underlying the impaired differentiation potential of mTORC1-activated cells. Jagged1 (a Notch ligand) protein expression was significantly enhanced in primary ΔTsc1 calvarial cells, and rapamycin reduced this expression to normal (Fig 6D) and dose-dependently decreased its expression in MC3T3-E1 cells (Fig 6E). In addition, a similar pattern of expression of the Notch transactivator NICD domain and Hes1 (a direct target of Notch) was also observed in these cells (Fig 6D and 6E). These results demonstrated that mTORC1 activated Notch signaling in preosteoblasts.
As the Notch pathway has been reported to reduce Runx2 activity [32], we next determined whether Notch signaling impaired differentiation of preosteoblasts upstream of Runx2. We found increased Runx2 expression in differentiating MC3T3-E1 and primary cavarial cells following suppression of the Notch pathway using the Notch inhibitor N-[N-(3, 5-difluorophenacetyl-L-alanyl)]-S-phenylglycinet-butylester (DAPT) (Fig 6F and 6H) and Notch siRNA (Fig 6G and 6I). We conclude that mTORC1 inhibited Runx2 expression by activating Notch signaling in preosteoblasts.
To define how activated mTORC1 influenced the Jagged1/Notch/Hes1 pathway, we determined the expression of STAT3 and p63, as STAT3 is a transcriptional activator of p63 [36] as well as the downstream target of mTOR [37], and p63 is a positive regulator of Jagged expression and Notch activity [38,39,40] which is induced by PI3K [41]. Cavarial cells with disruption of Tsc1 showed increased phosphorylation of STAT3 on Ser727 and expression of TP63 (total p63) (Fig 7A), and rapamycin reduced these parameters to normal levels. Knockdown of STAT3 by siRNA led to reduction in p63, Jagged1 and Hes1 without affecting the activity of mTORC1 (Fig 7B), and downregulation of p63 resulted in reduced Jagged1 and Hes1 but no changes of STAT3 phosphorylation and mTORC1 activity (Fig 7C). These findings suggested that the STAT3/p63 cascade is positively regulated by mTORC1 and serves as a conjunction between mTORC1 and the Notch pathway.
We next sought to define how mTORC1 regulates STAT3/p63 cascade in osteoblasts. As has been reported that mTORC1 directly phosphorylates STAT3 at Ser727 [37], we firstly examined whether mTORC1 could phosphorylate STAT3 at Ser727 in osteoblasts. We immunoprecipitated mTOR complex from control and ΔTsc1 calvarial cells and subjected the immunoprecipitates to in vitro kinase assays using a recombinant GST-tagged full-length STAT3 peptide. As shown in Fig 7D, mTOR immunoprecipitates could phosphorylate STAT3 at Ser727 in vitro and constitutive activated mTORC1 from ΔTsc1 calvarial cells enhanced the phosphorylation. Thus the elevated phosphorylation of STAT3 (S727) in ΔTsc1 osteoblasts is due to increased mTOR kinase activity. Next, we examined the role of phosphorylated STAT3 (S727) in the regulation of the p63 gene transcription by electrophoretic mobility shift assay (EMSA). Nuclear protein of osteoblasts bound specifically to a double-strand probe containing a consensus STAT3-specific binding sequence in the promoter region of ΔNp63 [36], and anti-pSTAT3 (S727) antibody significantly reduced the binding of STAT3 to p63 promoter (Fig 7E). Accordingly, nuclear protein from ΔTsc1 osteoblasts showed an increased binding of pSTAT3 (S727) to the probe (Fig 7E). Taken together, these results suggest that phosphorylated STAT3 (S727) may bind to p63 promoter to regulate its transcription in osteoblast, and mTORC1 activates the Notch pathway through STAT3/p63/Jagged cascade.
As mTOR is a positive regulator of the Jagged1/Notch/Hes1 pathway, we next determined whether hyperactive Notch signaling was responsible for the impaired differentiation of preosteoblasts by mTORC1 activation. Firstly, we examined the correlation between the level of Notch activity and differentiation of preosteoblasts. Low activity of Notch signaling is required during differentiation of preosteoblasts, as the expression of Jagged1, NICD and Hes1 was decreased in parallel with an increase in the markers of osteoblast differentiation (Fig 8A), and inhibition of the Notch pathway by DAPT or si-Notch1 promoted osteocalcin expression (S7A Fig and Fig 8B) and mineralized nodules formation in MC3T3-E1 cells (S7B Fig and Fig 8C). Importantly, a reduction in Notch by DAPT (S7C and S7D Fig) or siRNA (Fig 8D and 8E) potentiated the differentiation of ΔTsc1 preosteoblasts.
To test these results in vivo, we administered 6-week old mice with DAPT or equivalent volume of DMSO for 4 weeks. Micro-CT analysis of the distal femur showed a marked normalizing of bone architecture in DAPT-treated ΔTsc1 mice (Fig 8F and 8G and Table 4) when compared to their DMSO-treated control. Inhibition of Notch pathway by DAPT elevated expression of Runx2 and subsequent osterix and osteocalcin in osteoblasts (Fig 8H–8M), and significantly reversed the bone phenotypes of ΔTsc1 mice (Table 4). Therefore, Notch acts as an inhibitor of osteoblast differentiation, while deregulated mTORC1 activation impairs preosteoblast differentiation through activation of the Notch signaling pathway.
Together, these in vitro and in vivo loss- and gain-of-function studies lend support for a central role of mTORC1 signaling in regulating both proliferation and differentiation of preosteoblast during bone homeostasis (S8 Fig).
In this study, we defined the essential role of mTORC1 signaling in osteoblastgenesis. We observed a decline in mTORC1 activity during differentiation of preosteoblasts and enhancement of osteoblastic differentiation following inhibition of mTORC1 in vitro and in vivo. Using conditional knockout cell and mouse models, we further revealed that activation of mTORC1 prevented preosteoblast differentiation through activation of the Notch signaling pathway.
mTORC1 is ubiquitously expressed in all types of cells to regulate growth and metabolism. Proliferation in many types of cells is promoted by mTORC1, thus it is not surprising that the activity of mTORC1 positively correlated with the proliferative rate of preosteoblasts. However, it is interesting that mTORC1 activity resulted in a decline in the differentiation of preosteoblasts. We speculate that the decline in mTORC1 activity is due to differentiated osteoblasts. As osteoblasts differentiate, calcification of the extracellular matrix gradually increases, which generates hyperosmolarity and cellular stress. Coincidentally, this type of stimulus has been shown to inhibit mTORC1 [42].
The role of mTORC1 in bone formation was originally identified in various cell lines and animals treated with rapamycin, however, the results were controversial. Depending on the cell type, rapamycin either stimulates or inhibits osteoblastic differentiation. In rat osteosarcoma (ROS 17/2.8) cells, rapamycin inhibited proliferation, but promoted osteogenic differentiation [16]. In C2C12 cells, rapamycin potentiated the effect of BMP-2 inducing late markers of osteoblast differentiation [17]. In Human Embryonic Stem Cells, rapamycin induced the up-regulation of early osteogenic markers and further promoted the expression of late osteoblastic marker mRNA and/or proteins and mineralized bone nodule formation following induction for 2–3 weeks [43]. On the other hand, rapamycin has been shown to block osteogenic protein-1 induction of alkaline phosphatase activity in fetal rat calvarial cells [19] and reduce alkaline phosphatase activity, osteocalcin expression and the calcium content in mesenchymal stem cells [20]. In MC3T3-E1 subclone 4 (MC-4) cells and primary mouse bone marrow stromal cells (BMSCs), rapamycin inhibited osteoblast differentiation by targeting osteoblast proliferation and the early stage of osteoblast differentiation [21]. Although the reasons for the discrepant results are not readily apparent, one possible explanation is that the cells used in these reports were in various stages of osteoblastic differentiation and mTORC1 activity is stage-specifically required during the differentiation process. The results of the present study demonstrated that inhibition of mTORC1 is essential for preosteoblast differentiation. Another possible explanation for these contradictory results could be related to the treatment conditions including concentration and duration. Rapamycin at a concentration as low as 0.1 nM is effective in significantly suppressing mTORC1 activity [21], while higher concentrations and long-term treatment may produce a decrease in cell viability and growth, and inhibition of mTORC2. Thus, we used a low dose of rapamycin (0.1 nM) as a precaution against non-specific effects, and obtained reliable results that low mTORC1 activity is crucial for osteogenesis.
Previous results in animals administered rapamycin have also been inconsistent. Administration of rapamycin for 14 days resulted in no change in cancellous bone volume in rats [23], and after a longer duration of treatment (28 days), rats still lacked an osteopenic phenotype despite increased bone turnover [24]. However, mice treated with a similar concentration and duration of treatment showed a phenotype with less new bone formation and lower trabecular bone mass [25], due to a reduced number of osterix and osteocalcin positive cells coupled with unaffected osteoclasts. In the present study, rapamycin-treated mice had a similar phenotype of bone loss as well as decreased numbers of osterix and osteocalcin positive cells, except that the mean density of these cells was enhanced and the number of osteoclasts was reduced by rapamycin. Although we could not rule out that the systemic effect of rapamycin acted on other cell types to affect bone mass indirectly, the results of rapamycin administration in vivo are in agreement with those from rapamycin treated cells in vitro, which consolidates our conclusions that hyperactive mTORC1 is not required for differentiation of preosteoblasts and rapamycin promotes preosteoblast differentiation by blocking mTORC1.
Clinically, rapamycin was initially identified as a macrocyclic antifungal agent [44] and is used for immunosuppression in transplantation. The experimental and clinical trials have showed that rapamycin reduced brain, kidney, and skin lesions of tuberous sclerosis complex (TSC) [45, 46]. Some TSC children are being started on rapamycin at a young age (< 5 years), and it is sometimes continued for many (>5) years [47, 48]. In this situation, bone densitometry and morphological measurements must be advised to monitor the possible side effects of rapamycin on bone. Rapamycin is likely to impair bone microarchitecture and quality of those children as indicated by the current study, although it presented no effect on growth in weight and height of them [48].
Recent mouse genetic studies also highlighted the critical role of mTORC1 in skeletal development, whereas the function of mTORC1 in osteoblast formation is uncertain and the underlying mechanisms are not defined. Mice with Pten [49] and Tsc2 [50] disrupted by OC (osteocalcin)-cre recombinase (ΔTsc2) in mature osteoblasts shared an elevated mTORC1 activity, and exhibited different phenotypes. Although both mouse models had uniform high bone mass, osteoblasts with deletion of Pten showed increased osteoblastic differentiation, while Tsc2 disruption led to impaired differentiation of osteoblasts and mineralized nodule formation. Because inhibition of mTORC1 by rapamycin restored the differentiation defects in osteoblasts due to disruption of Tsc2, we reason that impairing osteoblastic differentiation is the function of hyperactive mTORC1 in mature osteoblasts, while an mTORC1-independent mechanism may account for the distinct phenotype in mice with disruption of Pten. Akt, which has been reported to promote osteoblast differentiation and bone development [51], is highly activated following Pten deletion and significantly inhibited following Tsc2 disruption. Acceleration of osteoblast differentiation by Akt may exceed the inhibition due to mTORC1 activation and cause increased differentiation in osteoblasts with Pten ablation. ΔTsc2 mice presented a similar phenotype and same differentiation defect in osteoblasts as ΔTsc1 mice in the current study. However, in contrast to ΔTsc1 mice, ΔTsc2 mice exhibited a decreased proliferation rate of osteoblast, which indicates that mTORC1 may play different role in regulating proliferation of immature and mature osteoblasts, since OC (osteocalcin) is expressed exclusively in mature osteoblasts. Moreover, though mTORC1 activation resulted in a same defect of differentiation in mature osteoblasts and preosteoblasts, the underlying mechanisms may be different. In this sense, our work revealed the role of mTORC1 in preosteoblast differentiation under physiological and pathological conditions and explored a different mechanism (STAT3/p63/Jagged/Notch pathway) responsible. More recently, another similar phenotype of immature woven bone as that of ΔTsc1 mice has been reported in mice with disruption of Lkb1 in preosteoblasts (ΔLkb1) [52], but the increased trabecular bone mass was more severe in ΔLkb1 mice than in ΔTsc1 mice, as BV/TV increased 7-fold versus 2.4-fold. In contrast, cortical thickness, which showed a 2.3-fold increase in ΔTsc1 mice, was decreased 3-fold in ΔLkb1 mice and osteoclasts were increased inversely in ΔLkb1 mice when compared with ΔTsc1 mice. These discrepancies may also be attributed to the activation of mTORC1-independent pathways in ΔLkb1 mice, as AMP kinase (AMPK), the target of Lkb1, regulates many signaling pathways besides mTORC1. In addition, there is no evidence to show that rapamycin restores bone phenotypes in ΔLkb1 mice. Unlike Lkb1, mTORC1 is a well-established target of TSC-TBC, and the phenotype of immature woven bone in ΔTsc1 mice was reversed by rapamycin, which further confirmed that bone changes in ΔTsc1 mice were the result of mTORC1 activation. Together, these findings demonstrate that differentiation of both preosteoblasts and osteoblasts is attenuated by mTORC1 activation.
Our data suggest that the impaired differentiation of preosteoblasts following mTORC1 activation is due to over-activated Notch signaling downstream. Notch signaling mediates communication between neighboring cells and decides their fate [32,33,34,35,53]. Specifically, gain of Notch function in preosteoblasts under the control of the type I collagen (Col1a1) promoter results in a phenotype similar to ΔTsc1 mice in the present study [32]. By inhibiting mTORC1 and the Notch pathway via rapamycin and DAPT/Si-Notch1, respectively, we showed that Notch acts downstream of mTORC1 and mediates the attenuation of osteoblastic differentiation by mTORC1. These findings are supported by a report which revealed that mTOR positively regulates Notch signaling in mouse and human cells through induction of the STAT3/p63/Jagged signaling cascade [28]. More recently, Wang et al. reported that mTORC1 activates Notch3 to accelerate the development of hypoxia-induced pulmonary hypertension [30] and Karbowniczek M et al. found that mTORC1 activates Notch in tuberous sclerosis complex and Drosophila external sensory organ development [54]. In contrast, Notch has also been shown to promote mTORC1 signaling by increasing Raptor protein expression in rat hepatoma cells and primary mouse hepatocytes [55]. We speculate that the interaction between mTORC1 and Notch may depend on cell type. Our data also confirmed that Runx2 is negatively regulated by mTORC1 and identified the Notch pathway as the responsible mechanism. Notch was identified as a major mediator of mTORC1 signaling in the impairment of preosteoblast differentiation in the present study.
In summary, this study clarified the potential role of mTORC1 signaling in the regulation of preosteoblast proliferation and differentiation and identified Notch signaling and Runx2 as critical downstream mediators. Pharmaceutical coordination of the pathways and agents in preosteoblasts may be beneficial in bone formation.
The preosteoblast cell line, MC3T3-E1, was maintained in alpha-MEM (Gibco) supplemented with 10% FBS (Gibco), 100 U/ml penicillin, and 100 mg/ml streptomycin sulfate, at 37°C with 5% CO2. For growth curve analysis, MC3T3-E1 cells were plated in six-well plates at a density of 8×104 cells/well and cultured until confluent (8th day). Growth rate was assessed by cell counting. Primary osteoblastic cells were prepared from the calvaria of 21-day-old fetal rats or newborn mice as described previously [56, 57, 58] and cultured using the same method as for MC3T3-E1 cells.
For osteogenic induction, 100 μg/ ml ascorbic acid (Sigma-Aldrich) and 10 mM β-glycerol phosphate (Sigma-Aldrich) were added to confluent cells. Rapamycin (Sigma-Aldrich) was added as stated in the Results section. Alizarin red staining was carried out according to standard techniques.
The Southern Medical University Animal Care and Use Committee approved all procedures involving mice. Mice importing, transporting, housing and breeding were conducted according to the recommendations of "The use of non-human primates in research." Newborn C57BL/6 mice were purchased from the Laboratory Animal Centre of Southern Medical University. Tsc1flox/flox [59] and Osx-GFP::Cre [60] mice were both purchased from The Jackson Laboratory. The background of Tsc1flox/flox mice is 129S4/SvJae, and Osx-GFP::Cre mice were backcrossed on to a 129S4/SvJae background for 8 generations prior to use. We performed genotyping using genomic DNA isolated from tail biopsies, and the primers used are shown in S1 Table. The specificity of recombination was examined by PCR using primers flanking the floxed allele (S1 Table). Mice were sacrificed by cervical dislocation to ameliorate suffering.
Femur tissues dissected from the mice were fixed using 4% paraformaldehyde in PBS at 4°C for 24 hours and then decalcified in 15% EDTA (pH 7.4) at 4°C for 14 days. The tissues were embedded in paraffin or optimal cutting temperature (OCT) compound (Sakura Finetek), and 2–5 μm sagittal-oriented sections were prepared for histological analyses. H&E and Toluidine blue staining was performed as previously described [61]. Tartrate-resistant acid phosphatase (TRAP) or alkaline phosphatase (ALP) staining was performed using a standard protocol (Sigma-Aldrich). For IHC, we incubated primary antibodies which recognized mouse phospho-S6 ribosomal protein (Ser235/236) (Cell Signaling, 1:100, #2211), proliferating cell nuclear antigen (PCNA) (Cell Signaling 1:200, #13110), osterix (Abcam, 1:500, ab22552), osteocalcin (Abcam, 1:500, ab93876), and Runx2 (Cell Signaling, 1:100) overnight at 4°C. All sections were observed and photographed on Olympus BX51 microscope. Immunohistochemical staining was evaluated by cell number counting and computerized optical density (OD) measurements. In proliferation analysis, osteoblast proliferation fraction was calculated as BrdU+ osteoblasts per total osteoblasts on bone surface. Osteoblasts on bone surface were discerned by morphology and calculated by two independent observers blinded to the groups. In immunohistochemistry assays, cells per bone perimeter (B.Pm) was used to calculate the number of positive cells, and integrated optical density per area of positive cells (IOD/area, mean density) was used to quantify the staining intensity by detecting in 6 different images taken at 100x magnification with Image Pro Plus 6.0 software (Media Cybernetics, MD,USA)[62]. Briefly, positively stained regions of the image were selected by HSI (hue, saturation and lightness) with S from 0–255, I from 0–210 and H from 0–25, and then the brown color was converted into grayscale signal. The grayscale signal was measured as mean optical intensity of staining (mean density) within the tissue masks. At least three mice per group were examined. Three equidistant sections spaced at 200 μm apart throughout the midsagittal section of femur were evaluated.
The narcotized mice were analyzed using X-ray radiography. Quantitative analysis was performed in mice femora at 12 μm resolution on a micro-CT Scanner (Viva CT40; Scanco Medical AG, Bassersdorf, Switzerland) [63]. Briefly, scanning was performed at the lower growth plate in the femora and extended proximally for 300 slices. We started morphometric analysis with the first slice in which the femoral condyles were fully merged and extended for 100 slices proximally. Using a contouring tool, we segmented the trabecular bone from the cortical shell manually on key slices, and morphed the contours automatically to segment the trabecular bone on all slices. The three-dimensional structure and morphometry were constructed and analyzed for BV/TV (%), BMD (mg HA/mm3), Tb.N. (mm–1), Tb.Th. (mm) and Tb.Sp (mm). We also performed micro CT imaging in the mid-diaphysis of the femur and performed mid-shaft evaluation of 100 slices to quantify the cortical thickness, bone mineral density and outer/inner perimeter of the mid-shaft.
Cells and tissues were lysated by 2% sodium dodecyl sulfate with 2 M urea, 10% glycerol, 10 mM Tris-HCl (pH 6.8), 10 mM dithiothreitol and 1 mM phenylmethylsulfonyl fluoride. The lysates were centrifuged and the supernatants were separated by SDS-PAGE and blotted onto a nitrocellulose (NC) membrane (Bio-Rad Laboratories). The membrane was then analyzed using specific antibodies and visualized by enhanced chemiluminescence (ECL Kit, Amersham Biosciences).
To label the mineralization fronts, 10-week-old mice were subcutaneously injected with calcin (Sigma, 15 mg/kg body weight) in 2% sodium bicarbonate solution 10 days and 3 days before death [25]. After dissection, the femurs were fixed in 4% paraformaldehyde for 24 hours. They were then dehydrated through a graded series of ethanol (70–100%) and xylene before being embedded in methylmethacrylate (MMA) without prior decalcification [64]. 5 μm-thick sections were prepared for Goldner’s-Masson trichrome [65], and 10 μm-thick sections were prepared for double-labeling fluorescent analysis.
After initial culture for 48 hours, primary calvarial cells were replated and expanded for an additional 24 hours. The cells were then treated with BrdU labeling reagent (Invitrogen) for 6 hours according to the manufacturer’s instructions and washed with PBS. The cells were fixed with 70% ethanol for 25 min at 4°C, and then stained for immunocytochemical analysis. Three to five areas for each group (n = 3 slides) were counted by two independent observers blinded to the groups. BrdU-positive cells were scored visually. For in vivo proliferation analysis, mice were injected with BrdU (1ml/100g body weight) 2 hours before sacrifice.
The surface of the MMA embedded femurs were polished and acid-etched with 37% phosphoric acid for 2–10s. After washing for 5 min with 5% sodium hypochlorite they were coated with gold and palladium before examining with SEM (S-3700N, Hitachi, Japan).
To prevent the Osx promoter from driving Cre expression, pregnant mice were exposed to 200 μg/ml doxycycline (Sigma-Aldrich) in drinking water until their progeny had been processed for calvarial cell isolation. Doxycycline 100 μg/ml was added sequentially to the isolated cells until they reached confluence.
We transiently transfected cells with siRNA using Lipofectamine RNAi MAX (Invitrogen, Carlsbad, CA, USA) in Opti-MEM medium (Invitrogen), according to the manufacturer’s instructions. The efficiency of transfection was measured by western blot. The sequences of siRNA used in this study were as follows: Notch1: sense: 5’-CCAAGAAGUUCCGGUUUGATT-3’, and anti-sense: 5’-UCAAACCGGAACUUCUUGGTT-3’; STAT3: 5’-CTGGATAACTTCATTAGCA-3’; p63 (all isoforms): 5’-CACAGACCACGCACAGAAUdTdT-3’, 5’-UCCAGAUGACUUCCAUCAAdTdT-3’ (1:1 ratio mixture). Non-specific siRNA sequences were used as negative controls: sense: 5’-UUCUCCGAACGUGUCACGUTT-3’, and anti-sense: 5’-ACGUGACACGUUCGGAGAATT-3’. (GenePharma, Shanghai, China).
Primary calvarial cells were lysed in ice-cold buffer (40mM HEPES (pH 7.4), 2mM EDTA, 10mM pyrophosphate, 10mM glycerophosphate, 0.3% CHAPS, and one tablet of EDTA-free protease inhibitors (Roche, Basel, Switzerland) per 25 ml). Supernatants were incubated with anti-mTOR antibody for 2h at 4°C, followed by addition of 30 μl of 50% slurry of protein G Sepharose beads for another 1h. Beads were then washed four times with lysis buffer and once kinase buffer (25mM HEPES (pH7.4), 50mM KCl, 10mM MgCl2, 250 μM ATP). 0.4μg of recombinant GST-tagged full-length STAT3 peptide (Creative BioMart, #1496H) was added to 30μl kinase buffer. Kinase assays were performed for 30 min at 30°C, and terminated by the addition of the 2×SDS sample buffer followed by boiling for 5 min.
A total of 2μg of nuclear protein extracted from calvarial cells was incubated with biotin-labeled STAT3 binding-site DNA probe in binding buffer (EMSA kit; Thermo scientific) for 30 minutes at room temperature. The probe used for the reaction contains the STAT3 binding site of the ΔNp63 promoter with a sequence of 5'-GGATTCCTATTTCCCGTACATAATATGGAT-3'. After incubation, the samples were separated on a 6% polyacrylamide gel in Tris-borate ethylenediaminetetraacetic acid, transferred onto a nylon membrane, and fixed on the membrane by ultraviolet cross-linking. The biotin-labeled probe was detected with streptavidin-horseradish peroxidase (EMSA kit; Thermo scientific). A probe lacking nuclear extracts was used as a negative control. The specificity of the identified STAT3-DNA binding activity was confirmed by using a 200-fold excess of unlabeled probe containing a same sequence. For supershift analysis, 1 μg monoclonal anti-phosphorylated STAT3 (S727) (Cell Signaling Technology) was incubated with nuclear extracts for 30 minutes before the addition of the biotin-labeled DNA probe.
All results are presented as the mean ± S.D. Curve analysis was determined using Prism (GraphPad). The data in each group were analyzed using unpaired, two-tailed Student’s t-test. The level of significance was set at P < 0.05.
|
10.1371/journal.pgen.1006306 | RNA-Binding Protein FXR1 Regulates p21 and TERC RNA to Bypass p53-Mediated Cellular Senescence in OSCC | RNA-binding proteins (RBP) regulate numerous aspects of co- and post-transcriptional gene expression in cancer cells. Here, we demonstrate that RBP, fragile X-related protein 1 (FXR1), plays an essential role in cellular senescence by utilizing mRNA turnover pathway. We report that overexpressed FXR1 in head and neck squamous cell carcinoma targets (G-quadruplex (G4) RNA structure within) both mRNA encoding p21 (Cyclin-Dependent Kinase Inhibitor 1A (CDKN1A, Cip1) and the non-coding RNA Telomerase RNA Component (TERC), and regulates their turnover to avoid senescence. Silencing of FXR1 in cancer cells triggers the activation of Cyclin-Dependent Kinase Inhibitors, p53, increases DNA damage, and ultimately, cellular senescence. Overexpressed FXR1 binds and destabilizes p21 mRNA, subsequently reduces p21 protein expression in oral cancer cells. In addition, FXR1 also binds and stabilizes TERC RNA and suppresses the cellular senescence possibly through telomerase activity. Finally, we report that FXR1-regulated senescence is irreversible and FXR1-depleted cells fail to form colonies to re-enter cellular proliferation. Collectively, FXR1 displays a novel mechanism of controlling the expression of p21 through p53-dependent manner to bypass cellular senescence in oral cancer cells.
| Understanding the mechanisms underlying evasion of cellular senescence in tumor cells is expected to provide better treatment outcomes. Here, we identify RNA-binding proteins FXR1 (Fragile X-Related protein 1), that is overexpressed in oral cancer tissues and cells bypasses cellular senescence through p53/p21-dependent manner. Once FXR1 is amplified in oral cancer cells, protein p21 is suppressed and non-coding RNA TERC expression is aided, resulting in reduction of cellular senescence and promotion of cancer growth. Here, we demonstrate the importance of FXR1 in antagonizing tumor cell senescence using human head and neck tumor tissues and multiple oral cancer cells including the cells expressing p53 wild-type and mutants. This finding is important as FXR1/TERC overexpression is associated with proliferation of HNSCC and poor prognosis, pointing to possible stratification of HNSCC patients for therapies.
| Cellular senescence is a critical biological process occurring in normal and aging cells either due to developmentally programmed or DNA damage-induced causes. Cancer cells escape senescence by utilizing either transcriptional and/or co-transcriptional gene regulatory processes to control gene expression. For example, transcriptional activators including p53 [1,2] promote senescence by activating subset of genes and also get affected by upstream stress responses such as the DNA damage response (DDR). A majority of the transcriptionally activated genes such as p21 (CIP1/CDKN1A), p27 (CDKN1B), p16 (CDKN2A), and PTEN (Phosphatase and tensin homolog) are well-characterized for promoting cellular senescence through either activating p53 or p16-mediated senescence pathways [3]. Although changes in transcription play a major role in cellular senescence, the post-transcriptional changes associated with cellular senescence has not been well studied.
The post-transcriptional gene regulation is often controlled by RBPs in conjunction with noncoding RNAs [4]. Most importantly, aberrant expression of RBPs can alter the gene expression patterns and, subsequently, involve in carcinogenesis in multiple cancers including HNSCC [5]. A very few RBPs are known to be associated with senescence pathway by controlling mRNA processing, transport, stability, and translation of proteins responsible for senescence in mammalian cells. For example, RBPs like HuR, AUF1 and TTP can directly or indirectly control turnover and translation of mRNAs encoding senescence proteins [6,7,8]. In addition, the involvement of RBPs in DDR is rapidly growing and now they are considered as the major players in the prevention of genome instability [9]. RBPs prevent harmful RNA/DNA hybrids and are involved in DDR, and many different cell survival decisions. For example, in response to DNA damage, p53 induces RNPC1 expression and PCBP4 [poly(rC)-binding protein 4], which in turn represses translation of the mRNA encoding p53 and stability of the mRNA encoding p21, respectively [10,11]. Thus, RBPs are known to contribute to the cell fate decisions such as apoptosis and/or permanent cell cycle arrest to induce cellular senescence. A pro-senescence approach to cancer therapy is an attractive alternative approach to chemotherapeutic strategies [12]. However, abundant reports indicate that cellular senescence occurs in the pre-malignant stage of oral squamous cell carcinoma (OSCC) but is lost once malignant transformation has occurred [13,14,15,16,17]. In contrast, stress or oncogene-induced senescence (OIS) also reported in OSCC and indicated that OIS and its markers could play a role in OSCC tumor progression [18,19,20]. Furthermore, OSCC cells expressing high-risk p53 mutations are sensitized to cisplatin therapy by the selective wee-1 kinase inhibitor, MK-1775, which subsequently promoted mitotic arrest and cellular senescence [21]. Thus, understanding the molecular mechanisms that underpin RBP-mediated senescence may yield invaluable data for the management of OSCC.
FXR1 belongs to the Fragile X-Related (FXR) family of RBPs, which also includes Fragile X Mental Retardation 1 (FMRP) and Fragile X-Related 2 (FXR2). FXR1 is frequently amplified in chromosome 3q26-27 in lung squamous cell carcinomas [22]. A recent observation indicates that FXR1 is a key regulator of tumor progression and is critical for growth of non-small cell lung cancer cell (NSCLC), and head and neck squamous cell carcinoma (HNSCC) [23]. Similar to the functions of other RBPs, FXR1 is involved in mRNA transport, translational control, and mRNA binding via AU-rich elements (ARE) or G4 RNA structures [24,25]. FXR1 is shown to bind to G4-RNA structure at the 3’-UTR of p21 and reduce its half-life in mouse C2C12 cells [26]. The G4 RNA structure containing human telomerase reverse transcriptase (hTERT) and its RNA component TERC RNA [27], are suppressors of cellular senescence in a variety of cells as deregulation of their function leads to the progressive shortening of telomere [28]. The regulatory mechanisms controlled by RNA G4 structures involve the binding of protein factors that modulate G4 RNAs turnover and serve as a bridge to recruit additional protein regulators. For example, G4 structure forming sequences protect TERC from degradation and interact with RNA helicase associated with AU-rich element (RHAU), a DEAH-box RNA helicase that exhibits G quadruplex-RNA binding and resolving activity [29]. In this report, we have identified that FXR1, overexpressed in oral squamous cell carcinoma, binds and destabilizes G4 containing RNA p21 and in turn reduces its protein expression in oral cancer cells. Thus controls cell cycle at G0/G1 phase, maintains cancer cell proliferation and evades cellular senescence. In addition, FXR1 associates and stabilizes non-coding RNA TERC resulting in suppression of cellular senescence and increased cancer growth. Thus, FXR1 is an important protein that regulates RNAs such as p21 and TERC to promote cancer progression.
To determine whether expression of RBPs is different in HNSCC, we utilized the cbioportal cancer genomics database (www.cbioportal.org) to examine the Cancer Gene Atlas (TCGA) in Head and Neck cancer study. We first analyzed copy number variation of the predicted and well-conserved 424 RBPs [30]. The TCGA HNSCC dataset contains tumor samples from 516 patients, of which 302 were analyzed for copy number alterations and mutation status with 5% cut off (Fig 1A, S1 Table). Next, to test the mRNA expression pattern of 424 RBPs, we set an mRNA expression onset of two standard deviations above or below the mean z-score, to identify patients with highly altered RBP mRNA levels. Using an arbitrary cut-off (RBP EXP > = 1.5), we identified 123 RBPs which were altered in 10% (51/516) or more of HNSCC patients (Fig 1B, S2 Table). Among those shown significant alteration, FXR1 showed a 31% alteration in a combined DNA copy number and mRNA in 279 tumor samples. Moreover, patients without FXR1 mRNA alteration showed a significant (p<0.01) overall survival rate compared to the ones with mRNA alteration, by Kaplan- Meier estimates (S1A Fig). DNA copy number status of FXR1 was independently verified by fluorescence in situ hybridization (FISH) analysis in a HNSCC tissue microarray (TMA). As shown in Fig 1C (S3 Table), FXR1 is highly expressed in tumor compared to normal adjacent tissues. Next, we used UCSC Cancer Genome Browser to compare the mRNA expression levels of tumor and normal tissues in a large cohort of patients (Normal = 43 and Tumor = 521). FXR1 mRNA levels along with two other FXR1 families of proteins FMR1 and FXR2 were determined in 521 tumor samples. As shown in Fig 1D (S4 Table), FXR1 and FMR1 mRNAs are significantly (p<0.0001) expressed in tumor samples compared to normal. However, we did not observe a significant difference in expression for FXR2 in normal adjacent and HNSCC tumor tissues. Next, we tested the mRNA levels of FXR1 in eight matched tumor vs normal adjacent samples (obtained from HNSCC patients from MUSC biorepository) by qRT-PCR (clinical parameters are tabulated in S5 Table). As shown in Fig 1E, FXR1 mRNA is overexpressed in tumor compared to normal adjacent tissues, whereas FMR1 and FXR2 mRNA levels are comparable to their normal mRNA expression. All the values were normalized to 1 corresponding to their normal adjacent tissues. To confirm the above observations, the levels of FXR1, FMR1, and FXR2 proteins from eight representative matched HNSCC tumor and normal adjacent samples were analyzed. As expected the level of FXR1 is highly amplified in cancer tissues compared to normal adjacent tissues and we do not see differential expression of FMR1 and FXR2 (Fig 1F). Next, we tested the mRNA levels of FXR1, FXR2 and FMR1 in eight HNSCC cell lines compared to the normal primary human oral keratinocytes (HOK, value was taken as 1). As shown in Fig 1G, FXR1 mRNA is significantly (p<0.05) overexpressed in HNSCC cells compared to HOK cells, whereas FMR1 and FXR2 mRNAs are not significantly overexpressed. Finally, the protein levels for FXR1, FXR2 and FMR1 were determined in HNSCC cell lines compared to HOK. FXR1 protein expression is high in all the cell lines tested compared to HOK (Fig 1H). Unlike the mRNA expression pattern, FXR2 does not uniformly express in cell lines (Fig 1H). The FMR1 protein is not detected in these cell lines. Collectively, our data show that FXR1 DNA, mRNA, and protein is amplified and expressed at high levels in HNSCC tumor tissues and cell lines.
We have utilized two different shRNA constructs and both shRNAs are able to silence FXR1 protein in HNSCC cells (S1B Fig). We selected shRNA-2 (TRCN0000159153) for further gene silencing experiments in this report. However, in order to delineate the off target effects of shRNAs, we used additional shRNA-1 (TRCN0000158932) to ascertain the important experiments (S1C, S1D and S1E Fig). First, to test whether knock down (KD) of FXR1 has any effect on cell viability, we used MTT (3-(4, 5-dimethylthiazol-2-yl)-2, 5-diphenyltetrazolium bromide) colorimetric assay. As shown in S1F Fig, a significant decrease in number of viable cells is observed in both UMSCC74A (p<0.05) and UMSCC74B (p<0.01) cells with FXR1 KD compared to control, indicating FXR1 plays a key role in oral cancer cell viability and/or survival. To investigate the underlying mechanism of reduction in number of viable cells following FXR1 KD, we performed cell cycle analysis of the FXR1 KD and control cells by flow cytometry. As shown in Fig 2A, FXR1 KD induces cell cycle arrest in G0/G1 providing evidence that FXR1 might regulate cell division in oral cancer. Interestingly, FXR1 KD does not induce apoptosis in UMSCC74A and UMSCC74B cells, as demonstrated by the absence of cleaved PARP and Caspase-3 (S1G Fig). Cell cycle arrest is one of the key features of cellular senescence [31]. Hence, to test whether FXR1 KD-induced cell cycle arrest is associated with induction of cellular senescence, we assessed the expression of senescence-associated β-galactosidase (SA-β-gal) activity. As shown in Fig 2B, FXR1 KD UMSCC74A and UMSCC74B exhibit increased SA-β-gal staining compared to control cells. In addition, silencing of FXR1 in multiple oral cancer cells also promotes senescence by positive SA-β-gal staining (S2A Fig). Furthermore, to quantify the β-gal enzyme activity, we measured MUG (4-Methylumbelliferyl β-D-Galactopyranoside) conversion by senescence associated β-galactosidase to 4-MU as described by [32]. As shown in Fig 2C, both FXR1 KD UMSCC74A and UMSCC74B cells, the 4MU fluorescence per microgram is significantly high (p<0.05) compared to control shRNA treated cells. To confirm that cells underwent DNA damage associated cellular senescence, we used immunofluorescence to study DNA damage. An important step during the cellular response to DNA double stand break (DSB) is the phosphorylation of histone H2AX at the break site, giving rise to discrete nuclear foci, termed γ-H2AX foci and also phospho-ATM (pATM) foci, accumulation at the break sites and can also be visualized as distinct foci [33]. Here, we silenced FXR1 and estimated the level of γ-H2AX and pATM foci in both oral cancer cell lines. As shown in Fig 2D, the appearance of spontaneous γ-H2AX and pATM foci in FXR1 depleted cells demonstrate that DSBs occur after silencing of FXR1 compared to control shRNA treated cells. Representative quantitative information of the foci formation is shown under each cell line. Cells containing two or more foci are counted as positive [34]. Next, to confirm that depletion of FXR1 promotes cellular senescence; we have used FXR1 knockout mouse embryonic fibroblasts (MEFs) for SA-β-gal qualitative assay. As expected, FXR1 KO MEFs showed bright staining of SA-β-gal compared to WT MEFs (S2B Fig). Collectively, our analyses indicate that silencing of FXR1 in oral cancer cells led to DNA damage, cell cycle arrest and cellular senescence.
To test if FXR1 regulated cellular senescence occurs through alteration of post-transcriptional gene expression; we analyzed gene expression of senescence markers. As shown in Fig 3A, qRT-PCR analysis of p21, p27, p53, and PTEN revealed that their mRNA levels have increased in the absence of FXR1 in both UMSCC74A and UMSCC74B cells. Intriguingly, the level of TERC RNA has decreased significantly (p<0.05) in the absence of FXR1. Consistent with this analysis the protein levels of p53, PTEN, p21, and p27 (CDKN1B) increased upon FXR1 KD in UMSCC74A and UMSCC74B cells (Fig 3B). In addition, reduced levels of pAkt (Ser-473) is observed in FXR1KD UMSCC74A and UMSCC74B cells compared with control and total Akt (Fig 3B), indicates that FXR1 regulated cellular senescence is possibly aided through inactivation of phosphatidylinositol 3 kinase/Akt signaling pathway. As we observe an increase in p53 protein upon FXR1 KD (Fig 3B), next we tested whether p53 plays a role in promoting senescence in the absence of FXR1. Many human tumors are not entirely lacking p53, but instead they have “hot spot” p53 mutations which serve as tumor suppressors. Here, we used two isogenic oral cancer cell lines in the background of PCI13 where they express wild type (Wt) and mutant p53 (C238F), respectively, as described [35]. Fig 3C illustrates the levels of p21 and p53 mRNAs upon FXR1 KD. Both mRNAs exhibit a significant increase in p53Wt cells in comparison with p53 mutant cells. To test the mRNA changes alter the expression of proteins, we tested the protein expression patterns of p53 and p21 in FXR1 KD cells. Interestingly, upregulation of p53 and p21 protein levels are observed in the absence of FXR1 in WT p53 cell line. Moreover, silencing of FXR1 in WT p53 cells exhibit SA-β-gal staining compared to the mutant (Fig 3E). This observation demonstrates that FXR1-regulated senescence utilizes p53 activated senescence pathway in oral cancer cells. Next, to corroborate our observation in Fig 2D that silencing of FXR1 promotes DNA damage; we determined the H2AX expression (phosphorylation of H2AX at Ser 139 (γ-H2AX) correlates well with each double stranded DNA break) under FXR1 KD condition. As shown in Fig 3F, FXR1 KD UMSCC74A and 74B cells express γ-H2AX compared to control indicating that DNA damage occurs in the absence of FXR1. Next, to determine whether the senescent protein coding mRNAs are associated with FXR1 to exert their function, we employed RNP IP assay as described [36]. As shown in Fig 3G, both p21 and TERC are associated with FXR1 in comparison with p27 and p53. FXR1 IP data in the figure is added to show that FXR1 efficiently binds to the beads and elutes out with the bound mRNAs under RNP IP conditions. Interestingly, both the 3’-UTR of p21 and full-length TERC RNAs contain G4 RNA sequences (S2D Fig), identified by using QGRS mapper software [37]. Agarose gel analyses of PCR amplified RNAs obtained from FXR1 RNA IP samples (Fig 3G) show that p21 and TERC bind to FXR1 protein (Fig 3H). Altogether, these studies indicate that FXR1 coordinates the expression of several senescence-associated genes; most importantly it binds and regulates the expression of both p21 and TERC in oral cancer cells.
First, to estimate the levels of p21 at the DNA copy number, we used FISH analysis in HNSCC TMA. As shown in Fig 4A, compared to normal (left panel), p21 is not amplified in HNSCC TMA tested in this study (S3 Table). Next, the copy number changes of p21 was verified at the mRNA level with the data obtained from Cancer Genome Browser, and as shown in Fig 4B p21 levels are highly comparable with normal tissue samples. The expression of this mRNA is not significant in cancer genome browser. This observation is well correlated with the TCGA data, where p21 (DNA copy number and mRNA combined) expression is only 3% from 279 patients. As shown in Fig 4C, p21 levels are downregulated, though not significantly, in eight representatives matched normal adjacent and HNSCC tumor tissue samples. Furthermore, the levels of p21 protein were tested from the same eight representative matched HNSCC tumor and normal adjacent samples. As shown in Fig 4D, p21 is mainly expressed in the eight normal adjacent samples compared to tumor tissues. We did observe a discrepancy in p21 mRNA and protein expression patterns in normal and tumor tissues. In addition to mRNA changes, there is a possibility that p21 protein is altered at the post-translational level. And p21 protein analyses of the tumor samples correlated with the mRNA data shown in Fig 4C. Thus, p21 is downregulated in HNSCC compared to normal adjacent tissues tested here.
Second, based on our data in Fig 3G that FXR1 binds to p21 as well as previous report showing association of FXR1 with G4 RNA structure in the 3'–UTR region p21 [26], we therefore probed the relationship between FXR1 and p21 in HNSCC. Based on our RNA IP data (Fig 3G), p21 is associated with FXR1. Hence, we planned to determine whether silencing of FXR1 influences the expression of p21 mRNA. As shown in Fig 3A, FXR1 KD has induced the expression of p21 in both the oral cancer cell lines. Next, to test if FXR1 KD was directly correlated with p21 mRNA expression, we used a time course assay. As shown in Fig 4E, 0hrs of post transduction with shFXR1 exhibits a steady and significant (p<0.05) increase in p21 mRNA as tested until 72 hrs in UMSCC74B cells. Next, to test the changes in p21 exerted by FXR1 in translation, we tested p21 protein levels at different time points. FXR1 KD promotes p21 protein expression over the course of 0 to 72hrs (Fig 4F). Altogether, these data indicate that silencing of FXR1 promotes the expression of p21 in oral cancer cells. To estimate the post-transcriptional changes caused by FXR1 to control cellular p21 stability, we treated shcontrol- or shFXR1-transduced UMSCC74B cells with the transcription inhibitor actinomycin D (ActD), and measured the half-life of p21 by qRT-PCR. Linear regression on semi-log values of p21 mRNA decay rate in shcontrol-transfected cells provided an estimated half-life of approximately 2:30±0:54 hrs compared to shFXR1 exhibited a statistically significant (p<0.05, n = 3) increased half-life of 4:30±0:48 hrs (Fig 4G). Next, in order to determine the specific G4 sequences in human p21 and their association with FXR1, first, we measured the G-scores within the RNAs. Based on high and low G-scores, we cloned G4 regions of p21 3’UTR in the 3’UTR of Renilla luciferase plasmid (S2C and S2E Fig). The segment 1 (seg1) of p21 3’UTR has highest G-score compared to seg2. 3’UTR of human p21 shows different G4 containing regions with high and low G-scores compared to mouse p21 3’UTR where the G4 region is enriched in one location [26]. As shown in Fig 4H, the luciferase activity of seg1 in control cells is lower than seg2 because FXR1 binds and destabilizes seg1 compared to seg2. However, in the absence of FXR1 seg1 exhibits increased luciferase activity compared to seg2 which corroborates with total mRNA levels as shown in Fig 3A. Furthermore, seg2 in control cells exhibits increased luciferase activity indicating that FXR1 preferentially binds to seg1 compared to seg2 and promote destabilization of the RNA. Next, we wanted to show a direct binding of FXR1 to these luciferase constructs containing seg1 and seg2 of p21 3’UTR by RNP-IP assay. UMSCC74B cells are transfected independently with 3’UTR-empty or 3’UTR-p21seg1 or 3’UTR-p21seg2. 48hrs of post-transfection, cells are collected for RNP-IP analyses. As shown in Fig 4I, qRT-PCR analysis for luciferase RNA shows that seg1 binds to FXR1 more efficiently (p<0.05, n = 2) compared to seg2 with a low G4 sequences. As shown in S2D Fig, qRT-PCR for luciferase in these input samples shows that the transfection efficiency is comparable in all samples.
Thus, silencing of FXR1 in oral cancer cells facilitates an increase in steady-state level of p21 and subsequently promotes its protein expression, indicates that overexpressed FXR1 in HNSCC destabilizes p21 mRNA and reduces its protein expression.
First, we determined the TERC DNA copy number (chromosomal locus is 3q26.2) status which is independently verified by FISH analysis in a HNSCC TMA. As shown in Fig 5A, compared to normal (left panel), TERC DNA is amplified in multiple loci of the tumor tissue samples (Fig 5A and S3 Table). Next, we tested the RNA levels of TERC in eight tumor tissue samples compared to normal adjacent tissues. Our analyses indicate that TERC is overexpressed in HNSCC tumor tissues compared to normal adjacent tissues (Fig 5B). A similar trend is also observed in TCGA where the combined DNA copy number and RNA expression is amplified in 25% of 279 tissue samples. Collectively, these observations indicate that TERC is overexpressed in HNSCC tissues.
Second, our initial analysis presented above demonstrates that FXR1 depletion correlates with downregulation of TERC. Given that TERC associates with FXR1, we sought out to determine whether FXR1 directly regulates TERC accumulation in HNSCC cells by utilizing RNA turnover pathway. To understand the correlation between FXR1 and TERC levels over the course of the time, we tested both FXR1 and TERC simultaneously in FXR1 KD UMSCC74B cells. As shown in Fig 5C, an equally steady decrease of both TERC and FXR1 is observed under FXR1 depleted oral cancer cells. To estimate the post-transcriptional changes caused by FXR1 to control endogenous TERC stability, we treated shcontrol- or shFXR1-transduced UMSCC74B cells with the transcription inhibitor actinomycin D (ActD), and measured the half-life of TERC by qRT-PCR. Linear regression on semi-log values of TERC mRNA decay rate in shFXR1-transfected cells provided an estimated half-life of approximately 15.3±0:58 min compared to shControl exhibited a statistically significant (p<0.05, n = 3) half-life of 26.3±2.1 min (Fig 5D). Taken together, these observations possibly suggest that the observed upregulation of TERC in HNSCC primarily at the post-transcriptional level through increased RNA stability that is mediated by FXR1. Next, in order to determine the specific G4 sequences and their association with FXR1, first, we measured the G-scores within the TERC RNA. Based on high and low G-scores, we cloned full-length and mutant TERC at 3’UTR of Renilla luciferase plasmid (S2C and S2E Fig). When full-length and truncated TERC (28 base deletion at 5’-end, TERCmut) (S2C and S2E Fig) are used for luciferase assay, TERC exhibits increased luciferase activity in control cells compared to TERCmut (Fig 5E). However, in FXR1 KD cells, TERC exhibits a decreased luciferase activity compared to TERCmut indicating that FXR1 binds to specific G-rich sequences in these RNAs and possibly controls their turnover. Next, we wanted to establish a direct binding of FXR1 to these luciferase constructs containing TERC and TERCmut by RNP-IP assay. UMSCC74B cells are transfected independently with 3’UTR-empty or 3’UTR- TERC or 3’UTR-TERCmut. 48hrs of post-transfection, cells are collected for RNP-IP analyses. As shown in Fig 5F, qRT-PCR analysis for luciferase RNA shows that full-length TERC binds to FXR1 more efficiently (p<0.05, n = 2) compared to TERCmut with low G4 sequences. As shown in S2D Fig, qRT-PCR for luciferase in these input samples shows that the transfection efficiency is comparable in all samples. As TERC deregulation is often associated with telomere length [38], down regulation of TERC in HNSCC cells prompted us to determine the telomerase activity by using TRAPeze® Telomerase Detection Kit (Millipore, USA). FXR1 KD cells showed an appearance of the internal control (36bp) band compared to control which is correlated with a reduced telomerase activity (Fig 5G and 5H). We quantified both the internal control (36bp) and the first telomerase activity (*) bands in shControl and shFXR1 lanes. The data was normalized against the internal control bands from the two heat inactivated sample lanes for each experiment set, respectively (Fig 5F). Thus, silencing of FXR1 reduces the level of TERC and subsequently interferes with the telomerase activity. Altogether, these data indicate that overexpression of FXR1 in HNSCC play a key role in stabilizing TERC to bypass cellular senescence.
Based on the two independent experiments described above, FXR1 concurrently destabilizes p21 (Fig 4E–4G) and stabilizes TERC (Fig 5C and 5D) to repress cellular senescence in HNSCC. Upregulation of p21 induces cell cycle arrest during replicative senescence in cell culture [39,40]. In addition, TERC is an essential RNA component of the telomerase enzyme complex that has been directly implicated in the maintenance of telomere length and in the prevention of premature senescence and aging. In support of this function, TERC-deficient mice displayed short telomeres, chromosomal instability, and premature aging [41]. To test whether these two coordinated events such as down-regulation of p21 and up-regulation of TERC in cancer cells, independently or in combination to control senescence, we overexpressed p21 and silenced TERC by individual and combinatorial transfections in HNSCC cells. Overexpression of p21 by plasmid transfection and silencing of TERC by siRNA were verified by qRT-PCR analyses. As shown in Fig 6A, p21 is overexpressed more than 18-fold (p<0.05) and TERC is significantly (p<0.05) downregulated in oral cancer cells. We see changes in FXR1 levels after these transfections but the protein levels did not change. In addition, silencing TERC by siRNA does not alter the expression of p21 (Fig 6A). We further confirmed the levels of FXR1 and p21 protein levels in UMSCC74B cells (Fig 6B, additional cell line UMSCC74A- S3A and S3B Fig); and the protein expression is also quantified as a mean of two separate western blots (Fig 6C). The data indicate that ectopic expression of p21 increases the expression of p21 protein significantly without altering the levels of FXR1 or TERC. Next, to study whether these changes alter cellular senescence, we stained the cells with SA-β-gal. The data indicate that, independent overexpression of p21 did not show SA-β-gal staining, however, silencing of TERC showed weak staining (Fig 6D) (additional cell line UMSCC74A- S3C Fig). Quantitation of MUG conversion to 4-MU by senescence associated β-galactosidase for Fig 6D is shown in Fig 6E. To further confirm that FXR1 is a key regulator of senescence in these oral cancer cell lines and it does so by modulation p21 and TERC RNA, we set up the following experiment. As shown in Fig 6F, we treated UMSCC74B cells with shControl and shFXR1. And at the same time, the treated cells are also treated with shp21 and/or transfected with TERC overexpression plasmid. 72h of post-treatment, the treated cells were stained with SA-β-gal. Strong SA-β-gal staining was observed again in shFXR1 only treated cells. Light staining was also observed in cells treated with both shFXR1and shp21. Here, the light staining with SA-β-gal is similar to our observation in Fig 6D where the use of siTERC showed light staining in UMSCC74B cells. TERC RNA expression was estimated in all the plasmid transfected and/or shRNA transduced cells (Fig 6G). As shown in Fig 6H, FXR1 and p21 protein levels were determined in the treated cells by western blot analyses. Nevertheless, concurrent overexpression of p21 and silencing of TERC significantly promoted senescence evidenced by staining of SA-β-gal (Fig 6D). Furthermore, the 4MU fluorescence per microgram are significantly high (UMSCC74A, p<0.05 and UMSCC74B, p<0.005) in those with concurrent overexpression of p21 and silencing of TERC compared to independent changes (Fig 6E and S3D Fig). Taken together, these data indicate that FXR1-regulated repression of senescence involves both down-regulation of p21 and up-regulation of TERC in oral cancer cells.
Senescence has been shown to be either reversible or irreversible depending on cell type and exposure to cytotoxic agents [42,43]. To test whether FXR1-dependent senescence is reversible or irreversible, we used inducible clones of shcontrol and shFXR1 to silence FXR1 under Isopropyl β-D-1-thiogalactopyranoside (IPTG, 0.5mM) promoter. First, the stable UMSCC74B cells, transfected with shcontrol and shFXR1 (MISSION 3X LacO Inducible Non/Target shRNA), were treated with 1mM IPTG for 6 days to test the phenotype. Second, IPTG was removed from the medium to induce FXR1 expression back in the cells. The cells were then grown for additional 6 days to test senescence. Under both the conditions, the levels of FXR1, p21, and TERC are quantified by qRT-PCR. As shown in Fig 7A, IPTG-shFXR1 treated cells indicate that both FXR1 mRNA and TERC are significantly (p<0.05) downregulated and p21 mRNA is significantly (p<0.05) upregulated compared to non-induced control cells. However, when IPTG is removed, the RNA levels return back to comparable levels with the non-induced control cells (Fig 7A). Next, the protein levels were measured to ensure that the FXR1 shRNA inducible cells respond to the treatment of IPTG. Protein p21 is increased as FXR1 protein is silenced under the treatment of IPTG compared to control and IPTG removed cells (Fig 7B). Next, we have stained the cells with SA-β-gal to estimate the senescence. The SA-β-gal staining shows an appearance of bluish green stain in both IPTG incubated and removed FXR1 KD cells compared with induced control (Fig 7C). To note, the IPTG removed cells express FXR1 protein comparable to induced control cells (Fig 7B), hence, it was interesting to test whether the cells are reverting back from senescence by estimating their proliferation ability and survival through colony formation assays. As expected, the IPTG incubated FXR1 KD cells fail to form colonies after 18 days (longer time is given to form adequate number of colonies). Surprisingly, IPTG removed FXR1 expressing cells also fail to form colonies in comparison with non-induced control cells which form large colonies (Fig 7D). The number of colonies are reduced from 337 to 118 (35% decrease) upon FXR1 KD. Moreover, IPTG removed FXR1 expressing cells also exhibit reduced colony number of 123 (36% decrease) in UMSCC74B cells (Fig 7E). Taken together, these data indicate that FXR1-regulated senescence is not reversible in oral cancer cells, even after expressing FXR1 back into the cultured cells.
Post-transcriptional control of gene expression is gaining much attention due to the fact that RBPs are critical regulators of genes involved in DDR and genome instability [9]. Till date, there are few RBPs that have been implicated in DDR and functionally involved in promoting or suppressing cellular senescence [44]. In this report, we show that knockdown of RBP FXR1 is a major factor involved in cell cycle arrest and promoting cellular senescence through turnover of two distinct RNA targets in oral cancer cells. These findings indicate that in addition to changing the transcriptional landscape of RNAs through DDR, overexpression of certain RBPs are critically involved in increasing the life span of cancer cells. As RBPs are widely recognized for their extensive roles in several cancer biological processes such as survival, apoptosis and metastasis [45,46], a better understanding of RBPs role in cellular senescence will provide a new insight into the post-transcriptional gene regulation.
Firstly, cellular senescence, one of the hallmarks of cancer and aging, can be induced by telomere dysfunction that specific DNA damage, chromatin instability, and oncogene activation [47]. Our findings indicate, specific shRNA mediated knockdown of FXR1 induces the expression of mRNAs of p53, p27 and p21 and their encoding proteins (Fig 3A and 3B). An increased stabilization of p21, in particular, is a marker of cancer cell senescence [48]. An increased p21 protein levels is also associated with reduced cell growth in cancer [49]. Interestingly, it has been noted that in mouse C2C12 myoblasts a reduced level of FXR1 promotes p21 expression by association with G4-RNA structures present in the 3' UTR of p21 [26]. Taken together it suggests that the overexpression of FXR1 protein in cancer may aid in an important mechanism for evasion of cellular senescence through reduced mRNA and protein levels of p21. Although recent work that examined FXR1 in human cancers showed silencing of FXR1 exhibited reduced cancer cell growth in vitro and in vivo [23], the precise molecular mechanism of FXR1-regulated cancer cell growth was not addressed. Our proposed model (Fig 7F) demonstrates that overexpression of FXR1 post-transcriptionally facilitates p21 mRNA destabilization and reduces its expression in HNSCC, possibly promoting cancer cell proliferation.
Secondly, cancer cell senescence is shown to be associated with DDR and correlated with p53-mediated gene expression patterns and also telomere shortening in multiple cancer models [44,50,51,52]. Here, we show that, loss of FXR1 results in DNA damage (Figs 2D, 2E and 3F), induces p53 mRNA and protein (Fig 3A and 3B), and ultimately resulting in senescence in oral cancer cells. Increase in p53 and p21 are correlated with DDR-induced senescence [3,31], and our data demonstrate that silencing of FXR1 promotes double stranded DNA breaks by yH2AX foci formation (Figs 2D, 2E and 3F). In addition, upregulation of p21 by the p53 tumor suppressor gene has been well documented [39,49,53]. Here, we show that FXR1-mediated downregulation of p21 plays a role in repressing p53-dependent cellular senescence. Furthermore, our data indicates that p53 appears to play a critical role in p21 induction in FXR1 depleted cells (Fig 3C–3E). In addition, we have observed destabilization of TERC and slightly reduced telomerase activity (Fig 5C–5I) which indicates that TERC degradation could be an important biological process that promotes senescence in part with p21. Treatment with ActD shows a significant decrease in TERC RNA half-life in FXR1 KD cells (Fig 5D). This data refers that FXR1 plays a role in the stability of TERC which is consistent with previous work identifying RBP Dyskerin as a key regulator of TERC that in turn controls the mammalian telomerase activity [54,55,56]. Interestingly, Dyskerin is the only well-characterized RBP that has shown to bind to TERC, but our study has now identified FXR1 as another controller of TERC which binds and stabilizes the RNA. Additional work is needed to warrant the recruitment of TERC by FXR1 and how this process controls telomere length in cancer.
Finally, studies demonstrate a link between p21 and TERC, increased telomere erosion, and DDR. For example, knockout of TERC causes progressive telomere shortening that persistently activates DDR and leads to numerous abnormalities in stem cell function and accelerated aging [57,58]. At the cellular level, DDR promotes a permanent cell-cycle arrest and initiates senescence [31,59]. Interestingly, loss of p21 in TERC -null mice with dysfunctional telomeres leads to improved stem cell function and increased lifespan without accelerating tumor formation [60]. In this regard, our findings provide a physiological basis by which FXR1 prevents cellular senescence through loss of p21 and upregulation of TERC. Our data also indicate that, overexpression of p21 alone is not sufficient for promoting senescence in oral cancer cells (Fig 6D and 6E); it also requires down-regulation of TERC (Fig 6D and 6E). This is also clear from Fig 6F where senescence is observed when only FXR1 alone is knocked down. Correlative with Fig 6D and 6E, we also see some senescent cells where both FXR1 and p21 are knockdown suggesting down-regulation of TERC can still bring about senescence but at a lower degree. RBPs play a major role in genome instability and DDR to control gene expression patterns [45,61,62]. The action of DDR-mediated activation of p53/p21 activated senescence requires down-regulation of TERC by FXR1 which also provides a basis for RBPs’ role in DDR and genome instability. Studies strongly support that G-4 RNA structures present in telomere RNAs [63,64], play a critical role in promoter activity as well as its expression. Our observation on G4 RNA structures, which are strongly enriched in 3' UTR sequences including that of p21 and full length TERC (more the G-score stronger the G4 structure, Figs 4H and 4I, 5F and 5G and S2E Fig), provides an opportunity to shed light on their importance in senescence and aging. Our data clearly implicate that both G4 structure containing p21 and TERC bind to FXR1 (Figs 4H and 4I, 5F and 5G and S2E Fig), providing a basis for post-transcriptional control of two distinct mechanisms (Fig 7F). Cellular senescence is considered irreversible in the sense that known physiological stimuli cannot force senescent cells to re-enter the cell cycle [59]. Our data supports that FXR1-regulated elucidation of senescence is irreversible based on the colony formation assay (Fig 7D and 7E). Furthermore, FXR1 promotes cellular senescence in WT p53 expressing cells compare to p53 mutant stable HNSCC cells (Fig 3C–3E). Thus, FXR1 possibly plays a role in suppressing p53 for checkpoint control. Therefore, our data highlight an important unifying role of FXR1 towards the p53/p21-TERC-pathway in dictating cell growth over cellular senescence in HNSCC (Fig 7F).
Human tissues were obtained from Hollings Cancer Center (HCC) biorepository with written informed consent and local MUSC Internal Review Board approval (Pro00009235 (CT)#101547). Frozen tumor tissues are micro-dissected to assure that > 80% of tumor contained HNSCC. The HNSCC and normal adjacent tissues contained tissue microarrays that are used for the evaluation of FXR1 and p21, and TERC DNA expression using FISH. Samples are subjected to protein and RNA extraction for immunoblotting and qPCR analyses, respectively.
HNSCC cell lines UMSCC-11A, -11B, -74A and -74B were obtained from University of Michigan and SCC4, SCC9, SCC25 and CAL27 were obtained from ATCC. Cell lines were routinely grown in Dulbecco’s modified Eagle medium (DMEM-Hyclone) containing 10% fetal bovine serum (FBS) with 100U/ml penicillin and 100 μg/ml streptomycin. HOK cells (Science Cell) were grown in keratinocyte serum-free medium supplemented with BPE and EGF (Gibco, BRL). SCC cell lines were grown in DMEM: F12 (1:1) containing 400 ng/ml hydrocortisone, 10% FBS, and 100U/ml penicillin and 100 μg/ml streptomycin. Different shRNA constructs for FXR1 (TRCN0000158932 and TRCN0000159153) and p21 (TRCN0000287021) were obtained from Sigma Mission. FXR1 inducible shRNA clone with TRCN0000159153 was obtained from Sigma Mission. p21, pCEP-WAF1 was a gift from Bert Vogelstein (Addgene plasmid # 16450) and TERC, pBABEpuro U3-hTR-500 was a gift from Kathleen Collins (Addgene plasmid # 27666), over-expression plasmids are obtained from Addgene. siTERC RNAs (100uM; GGGCGUAGGCGCCGUGCUU and CCCACUGCCACCGCGAAGA) were purchased from Sigma Mission whereas control siRNA (20nM; GTTCAATTGTCTACAGCTA) was from Dharmacon RNAi Technologies. Regular transfection was done with lipofectamine-2000 (life technologies). siRNA transfections were done with HiPerfect (QIAGEN) transfection reagent, following the manufacturer’s protocol. FXR1 antibodies were obtained from Cell Signaling Technology (CST) and EMD Millipore, FMR1 was from Abgent and FXR2 was from Bethyl Laboratories. p21 and p27 antibodies were from BD Pharminogen. P53 was from Santa Cruz. AKT, p-AKT (S473), PTEN, γ-H2AX, and GAPDH were from CST. β-Actin was purchased from Sigma. Alexa Fluor 488 was bought from life technologies. SA-β-gal assay kit was purchased from CST. MUG was purchased from Sigma-Aldrich. Horseradish peroxidase-conjugated anti-mouse and anti-rabbit immunoglobulinG were procured from GE Healthcare Biosciences (Uppsala, Sweden). Protein A/G beads were purchased from Santa Cruz Biotechnology. Fugene HD transfection reagent and LightSwitch Luciferase Assay kit (LS010) were purchased from Switchgear Genomics.
FISH assays are performed on unstained TMA sections. BAC clones (FXR1: BAC clones: RP11-314I4 (green) RP11-480B15 (red); p21: RP11-265F6 (green) RP11-624F22 (red); TERC RNA: RP11-990E14 (green) RP11-480B15 (red) are selected from the UCSC genome browser and purchased through BACPAC resources (Children's Hospital, Oakland, CA). Following colony purification DNA is prepared using QiagenTips-100 (Qiagen, Valencia, CA). DNA is labeled by nick translation method with biotin-16-dUTP and digoxigenin-11-dUTP for 3' and 5' probes and locus and control probes respectively (Roche, USA). Probe DNA is precipitated and dissolved in hybridization mixture containing 50% formamide, 2X SSC, 10% dextran sulphate, and 1% Denhardt's solution. Approximately 200ng of labeled probe is hybridized to normal human chromosomes to confirm the map position of each BAC clone. FISH signals are obtained using anti-digoxigenin-fluorescein and AlexaFluor-594 conjugate to obtain green and red colors, respectively. Fluorescence images are captured using a high resolution CCD camera controlled by ISIS image processing software (Metasystems, Germany).
SA-β-gal activity is measured according to the manufacturer’s instructions (Cell Signaling Technology, Beverly, MA, USA). SA-β-Gal activity is detected using X-gal (5-bromo-4-chloro-3-indolyl β-D-galactoside) staining at pH 6.0 at 72 hours post-transduction with shRNAs unless otherwise mentioned. Using light microscope, three representative fields are captured under white light for three independent experiments. The senescence associated β-gal activity in UMSCC74A and UMSCC74B is quantified by a method as described elsewhere [32]. Briefly, SA-β-gal is measured by the rate of conversion of 4-methylumbelliferyl-α-D-galactopyranoside (MUG) to a fluorescent hydrolysis product 4-methylumbelliferone (4-MU) at pH 6.0. Treated UMSCC74A and UMSCC74B cells grown in 60-mm plates are washed three times with Hank’s balanced salt solution. Cells are then lysed by 200μl of lysis buffer, scraped, transferred to a 1.5-ml tube, vortexed, and centrifuged at 12,000g for 5min. The clear supernatant is then used for the assay after measuring the total protein by Biorad spectrophotometer. Reaction buffer at 2X strength is mixed with 1.7mM of MUG added immediately prior to use from a 34mM stock in dimethyl sulfoxide. For final reaction the 2X reaction buffer (150μl) is mixed with 150μl of clarified lysate (100μl of lysate diluted with 50μl of lysis solution) and carried out at 37°C water bath for 0, 1, 2, and 3 hours. At the end of each time points the reaction is stopped with 400mM sodium carbonate. The stopped reaction mixture is read by using150μl per well in a 96-well plate using a plate reader with excitation at 385 nm, emission at 465nm, and gain held constant at 460. Normalized SA-β-gal activity is expressed as observed fluorescence divided by micrograms of total protein in the assay.
FXR1 RNP IP is performed as previously described [36] with some modifications. Briefly, cell lysates are prepared from exponentially growing UMSCC74B cells. Equal amounts of protein are used (750–1000μg). FXR1 monoclonal antibody (Millipore) or isotype control IgG (Santa Cruz) are pre-coated onto protein A/G Sepharose beads (PAS) and extensively washed using NT2 buffer [50mM Tris–HCl, 150mM NaCl, 1mM MgCl2, 0.05% Nonidet P-40 (NP-40), pH 7.4]. Individual pull-down assays are performed at 4°C for 1–2 h to minimize potential reabsorbing of mRNAs. For RNA analysis, the beads are incubated with 1ml NT2 buffer containing 20 U RNase-free DNase I (15 min, 30°C), washed twice with 1ml NT2 buffer and further incubated in 1 ml NT2 buffer containing 0.1% SDS and 0.5mg/ml proteinase K (15 min, 55°C) to digest the proteins bound to the beads. RNA is extracted using phenol and chloroform, and precipitated in the presence of glycogen. For analysis of individual mRNAs, the RNA isolated from the IP is subjected to reverse transcription (RT) using random hexamers and SuperScriptII reverse transcriptase (Biorad).
Total RNA is prepared from oral cancer tissues and HNSCC cell lines using Trizol (Ambion) or RNeasy mini kit (QIAGEN) by following manufacturer’s protocol. qRT-PCR for all m/RNA targets is performed using an Applied Biosystems StepOne Plus system with the SYBR green master mix RT-PCR kit (SA Biosciences). Primer sequences are provided in S6 Table.
The cell cycle analyses of UMSCC74A and 74B-FXR1 knockdown cells was performed by flow cytometry using propidium iodide. After 72hrs of shRNA treatment, a total of 50,000 FXR1 KD and control cells were fixed and stained with propidium Iodide (PI), and analyzed by fluorescence-activated cell sorting (BD Fortessa X-20 Analytic Flow Cytometer) to evaluate the number of cells in different stages of cell cycles. Cell cycle analysis was done using the ModFit LT software.
UMSCC74A and 74B-FXR1 knockdown and control cells were washed and fixed with 4% paraformaldehyde. Fixed cells were blocked with 10% normal donkey serum followed by incubation with yH2AX and pATM primary antibodies for 1hr. Finally, the cells were washed and incubated with Alexa Fluor 488 secondary antibody and 4, 6-diamidino-2-phenylindole. Stained cells were subjected to analysis by using Olympus BX61 Microscope with Green and DAPI-FITC filter.
The shRNA mediated 74B-FXR1 knockdown and control cells were used for luciferase assay. Different segments of human 3′UTRs of p21, TERC, and GAPDH were systematically identified (S2C and S2D Fig) based upon the G-scores (QGRS mapper tool) and cloned into a luciferase reporter vector system, pLightswitch-3’UTR from Switchgear Genomics. The segments were cloned between Nhe1 and Xho1 sites to express chimeric m/RNAs spanning two of the luciferase p21 3′UTR segments and TERC RNAs based upon high and low G-scores. The luciferase GAPDH 3′UTR and 3’UTR empty vector negative controls were included in all assays. Each construct was transfected in triplicates separately with either 74B-FXR1 knockdown and control cells with Fugene HD transfection reagent. Plates were incubated at 37°C for 48 h post-transfection before being removed. 100 μl of luciferase assay (buffer + substrate) reagent (LightSwitch Luciferase Assay) was added to each well of 96 well solid bottom white plates, and was incubated at room temperature for 30 min. Luminescence was measured by using a VICTOR3 1420 Multilabel Counter (PerkinElmer) and the data obtained was normalized using Lightswitch normalization protocol using GAPDH 3′UTR and 3’UTR empty vector controls.
Cell viability rate upon FXR1 KD in UMSCC74A and UMSCC74B cells are determined using MTT cell proliferation assays (Invitrogen). Briefly, post-shRNA transfected 5×103 cells were inoculated into each well of a 96-well plate (well area = 0.32cm2). Right after plating (considered as 0hr) and after that every 24h, medium was replaced with experimental medium (100μl). MTT solution was prepared fresh (5 mg/ml in H2O), filtered through a 0.22-μm filter, and kept for 5 min at 37°C. The MTT solution (10μl) was added to each well, and plates were incubated in the dark for 2 h at 37°C. Then reaction was stopped using MTT solution (10% SDS in 1N HCl) and further incubated overnight at 37°C. Next morning the absorbance was measured at A570 nm using a plate reader (Bio-Rad).
UMSCC74B inducible control and FXR1 KD (two wells) cells were counted and 1,000 cells were plated on a 6-well dish with 1mM IPTG for 9 days. After 9 days IPTG was removed from one well containing FXR1 KD cells and left for another 9 days. At the end the colonies were fixed and stained with crystal violet (0.5%w/v) in 20% methanol for 30min, plate was washed, and counted using a microscope.
Data are expressed as the mean ± the standard deviation. Two-sample t-tests with equal variances are used to assess differences between means. Results with p values less than 0.05 is considered significant.
|
10.1371/journal.pntd.0006569 | The small non-coding RNA response to virus infection in the Leishmania vector Lutzomyia longipalpis | Sandflies are well known vectors for Leishmania but also transmit a number of arthropod-borne viruses (arboviruses). Few studies have addressed the interaction between sandflies and arboviruses. RNA interference (RNAi) mechanisms utilize small non-coding RNAs to regulate different aspects of host-pathogen interactions. The small interfering RNA (siRNA) pathway is a broad antiviral mechanism in insects. In addition, at least in mosquitoes, another RNAi mechanism mediated by PIWI interacting RNAs (piRNAs) is activated by viral infection. Finally, endogenous microRNAs (miRNA) may also regulate host immune responses. Here, we analyzed the small non-coding RNA response to Vesicular stomatitis virus (VSV) infection in the sandfly Lutzoymia longipalpis. We detected abundant production of virus-derived siRNAs after VSV infection in adult sandflies. However, there was no production of virus-derived piRNAs and only mild changes in the expression of vector miRNAs in response to infection. We also observed abundant production of virus-derived siRNAs against two other viruses in Lutzomyia Lulo cells. Together, our results suggest that the siRNA but not the piRNA pathway mediates an antiviral response in sandflies. In agreement with this hypothesis, pre-treatment of cells with dsRNA against VSV was able to inhibit viral replication while knock-down of the central siRNA component, Argonaute-2, led to increased virus levels. Our work begins to elucidate the role of RNAi mechanisms in the interaction between L. longipalpis and viruses and should also open the way for studies with other sandfly-borne pathogens.
| Sandflies are important insect vectors that transmit many species of Leishmania, bacteria and viruses. We know very little about how this insect vector responds to viral infection. RNA interference (RNAi) utilizes small non-coding RNAs to regulate different aspects of animal physiology, including immune responses. Small interfering RNAs (siRNAs) mediate a major antiviral response in insects. Virus-derived PIWI-interacting RNAs (piRNAs) can also be generated during infection, at least in some insects. Finally, endogenous microRNAs (miRNA) can regulate the host response to infection. Here we show that virus infection triggers activation of the siRNA pathway but not production of piRNAs in the sandfly Lutzomyia longipalpis. Furthermore, activation or inhibition of the siRNA pathway had a direct effect on viral replication. We also show that virus infection caused mild changes to the expression of endogenous miRNAs. Our work describes for the first time a model to study virus infection in sandflies and highlights the importance of the siRNA pathway for the control of virus infection in L. longipalpis. The framework described here can be used to explore other aspects of the vector-pathogen interactions.
| Phlebotomine sandflies (Diptera: Pshychodidae) are important vectors for a wide range of pathogens [1]. Protozoans of the Leishmania genus are the most studied of sandfly-borne pathogens but these insects can also transmit bacteria and viruses. Arthropod-borne viruses (arboviruses) transmitted by sandflies are associated with several human and animal diseases, mostly characterized by flu-like symptoms but also some severe cases of encephalitis [2–4]. Sandfly-borne viruses belong to several genera including Vesiculovirus (family Rhabdoviridae), Phlebovirus (family Bunyaviridae) and Orbivirus (family Reoviridae) [5–7]. Vesiculovirus and Orbivirus are mainly restricted to the Americas while Phlebovirus are of important concern in Southern Europe and Turkey [7–10]. The Vesiculovirus genus includes several human and animal pathogens such as Vesicular stomatitis virus (VSV) that causes outbreaks in cattle and horses [11]. Despite the importance of sandfly borne viruses, viral infections remain poorly studied in this vector. Current laboratory models to dissect sandfly-pathogen interactions are mostly restricted to Leishmania [12–14] while most information on the transmission of viruses is based on field studies [6, 8, 15, 16].
Arboviruses need to overcome physical and immunological barriers to replicate in the insect vector and be transmitted to the vertebrate host. One of the most important antiviral immune responses in insects is mediated by RNA interference (RNAi). RNAi refers to different mechanisms of regulation of gene expression mediated by small non-coding RNAs [17]. The small interfering RNA (siRNA) is considered the main antiviral response in insects [18, 19]. This pathway is activated by double stranded RNA produced during viral replication that is processed by the RNaseIII enzyme Dicer-2 into 21 nucleotide (nt) siRNAs [20]. These siRNAs are then loaded by Dicer-2 and the small dsRNA binding protein r2d2 onto the nuclease Argonaute- 2 (AGO2) to form the RNA-induced silencing complex (RISC) that destroys complementary targets [21]. The PIWI-interacting RNA (piRNA) pathway can also be activated during viral infection although it is restricted to certain viruses and insect hosts [22–24]. In this case, the biogenesis of piRNAs does not require dsRNA specific nucleases such as Dicers. Rather, single stranded piRNA precursors are processed by different nucleases such as Zucchini nuclease and PIWI proteins to generate 24–29 nt small RNAs [25]. In addition to RNAi mechanisms that can directly target the virus, host small RNAs may also participate in the immune responses by regulating the expression of endogenous genes. MicroRNAs (miRNA) are endogenous 22 nt small RNAs that control the expression of hundreds of genes [26, 27]. A large number of immune genes are regulated by host miRNA that can be affected by infection. The majority of our knowledge about the small RNA response to infection in insects comes from studies in D. melanogaster and mosquitoes [22, 24, 28–36]. Currently, there is only one study about small RNAs in sandflies by our group, which suggested that the siRNA pathway was not activated by viruses in L. longipalpis [23].
Here, we develop a laboratory model to dissect virus infection in L. longipalpis utilizing VSV that is transmitted by sandflies in the wild [2]. Using this model, we analyzed the global small non-coding RNA response to virus infection. We found that VSV strongly induced the production of siRNAs but not piRNAs in adult sandflies. We further demonstrate that this pattern is conserved in response to two novel viruses found in Lulo, a cell line derived from L. longipalpis. Interestingly, VSV did not induce significant changes the profile of vector miRNAs, which is in agreement with the mild effects observed in infected cells and adult insects. Our data are the first to characterize the small RNA response to viruses and the antiviral role of RNAi in sandflies using a relevant model of VSV infection. This study should open the way for more in depth studies of the role of RNAi in regulating immune responses in sandflies as well as the tripartite interaction between vector, viruses and other pathogens such as Leishmania.
Previous studies have suggested that sandflies play a role in maintenance and transmission of VSV in nature [7, 16, 37–40]. In order to explore the interaction between VSV and sandflies, we initially characterized viral replication in cultured LL5 cells derived from L. longipalpis. Dose-response experiments demonstrated that using different multiplicities of infection (MOI) 0.4, 2 and 10 plaque-forming units (PFU)/cell, VSV achieved the same levels of viral RNA at 24 hours post infection (hpi) (S1 Fig). Using 10 PFU/cell, we observed that VSV RNA levels increased early after infection and reached a plateau at 6 hpi that remained similar until 24 hpi (Fig 1A). We also observed that infectious viral particles and viral RNA were detected in supernatant of LL5 cells during early hours of infection (Fig 1B and S1 Fig). Despite clear viral replication, we observed no obvious cytopathic effects (CPE) in LL5 cells after VSV infection (Fig 1C). In contrast, mammalian cell lines (Vero) were completely destroyed by VSV infection (Fig 1C). Mosquito C6/36 cells also displayed CPE such as cell rounding after VSV infection (Fig 1C). Although the kinetics of viral replication was different, LL5 cells generated a similar number of viral particles compared to Vero and C636 cells (Fig 1D). These experiments suggested that VSV productively replicates in L. longipalpis cells without causing much damage. Similar observations have been reported elsewhere [41].
We next performed a dose-response experiment to analyze the threshold of VSV infection in vivo. Adult female L. longipalpis were fed with blood containing different concentrations of VSV and the infection was monitored at 2 and 4 days post feeding (dpf). These two time points allowed us to analyze the amount of virus in the inoculum before full blood digestion at 2 dpf from productive infection at 4 dpf. In these experiments, we observed that the minimum concentration required to productively infect sandflies was 106 PFU/mL (Fig 2A). At this concentration we observed that about 25% of individuals were infected at 4 dpf, after blood digestion. At higher viral concentrations, 107 and 108 PFU/mL in the blood meal, over 40 and 60% of adult sandflies were infected at 4 dpf, respectively (Fig 2A). At the highest concentration of VSV in the blood meal, 108 PFU/mL, levels of the viral RNA increased continually from 1 to 6 dpf. Infected sandflies at 6 dpf had 100 fold more VSV RNA than the inoculum detected at 1 dpf (Fig 2B). Infectious viral particles were also detected in sandflies throughout the kinetics of infection in vivo (Fig 2C). VSV RNA was not detected in mock infected sandflies, which were used as controls (Fig 2B). We observed no differences in the mortality of sandflies infected by VSV compared to the control group. These results indicate that our laboratory model reproduces the susceptibility of sandflies to VSV observed in field collected insects [7, 37]. The benign effect of VSV in cells and adult L. longipalpis indicates that the virus is maintained at tolerable levels by the host. These observations are consistent with sandflies working as vectors for this virus in the wild.
The siRNA pathway is a major antiviral defense mechanism in insects and could be involved in controlling VSV replication in L. longipalpis. In order to analyze the antiviral response mediated by the siRNA pathway in L. longipalpis, small RNA libraries were constructed using total RNA from VSV infected individuals at 2, 4 and 6 dpf (indicated in Fig 2B). Mock infected individuals were used as controls. Raw sequencing results from libraries prepared from infected and control individuals were similar (S1 Table).
We next analyzed the presence of VSV-derived small RNAs in libraries from infected and control sandflies. Control individuals did not show significant accumulation of virus derived small RNAs (Fig 3A). Infected individuals accumulated increasing amounts of VSV-derived small RNAs from 2 to 6 dpf proportionally to the viral load observed at each time point (Fig 3B). In infected individuals, the profile of VSV-derived small RNAs showed a clear peak at 21 nt in size, a ratio of ~1 between sense/antisense sequences without a clear 5’ base preference (Fig 3B). In addition, the small RNAs mapped across the entire length of the viral genome on both positive and negative strands of the genome (Fig 3C). The profile of VSV-derived small RNAs in L. longipalpis was very similar to the antiviral siRNA response observed in Drosophila melanogaster [32, 36]. These results suggest that VSV triggers an antiviral response mediated by the siRNA pathway in sandflies.
In order to investigate whether this was unique to VSV, we sequenced small RNAs from cell lines derived from sandflies and used a metagenomic strategy to identify viruses [23]. In one L. longipalpis cell line, Lulo, our strategy identified two novel viruses causing persistent infections. Phylogenetic analysis suggested that they belong to two separate genera of RNA viruses, Luteovirus and Alphapermutotetravirus, and were named Lulo virus 1 and 2, respectively (LV1, LV2) (S2 Fig). In Lulo cells, both viruses generated small RNAs with a peak of 21 nt in size symmetrically derived from both positive and negative strands of the genome and no 5’ base preferences (Fig 3D and 3E). The small RNA profiles observed for LV1 and LV2 are consistent with the activation of siRNA pathway. These results suggest that the siRNA pathway is broadly activated by viral infection in L. longipalpis. It is important to point out that the Lulo and LL5 cell lines were obtained from embryonic tissues [42, 43] and therefore may not mimic a bona fide RNAi response produced by a differentiated cell.
Immune genes are often inducible upon infection in order to optimize the defense response [44, 45]. In a few cases, increased expression of RNAi genes has been observed in insects and other organisms upon viral infection or exposure to dsRNA [46–48]. In order to analyze the expression of siRNA genes in sandflies, we first identified in the genome of L. longipalpis genes encoding orthologs to core components of the siRNA pathway, Dicer-2, AGO2 and r2d2. We next measured RNA levels of these genes in LL5 cells and adult female L. longipalpis after VSV infection. There were no significant changes in the expression of Dicer-2, AGO2 and r2d2 in cells between 1 and 48 hpi and sandflies with VSV at 1, 2, 4 and 6 dpf (Fig 4A and 4B). These results indicate that similar to most insects, the activity of the siRNA pathway is not transcriptionally regulated during infections in sandflies.
The generation of VSV-derived siRNAs in L. longipalpis suggested that the siRNA pathway mediates an antiviral response against this virus as observed in D. melanogaster [32, 36]. In order to investigate whether activation of the siRNA pathway has an antiviral effect in L. longipalpis, we first transfected LL5 cells with dsRNA complementary to VSV prior to virus infection. Pre-engagement of the siRNA pathway with virus-specific dsRNA led to a significant reduction in viral replication at 24 and 48 hpi compared to control cells treated with a non-related dsRNA targeting the Firefly Luciferase gene (Fig 5A). We next utilized dsRNA treatment to silence AGO2, the central component of the siRNA pathway, in LL5 cells. dsRNA treatment lead to a significant reduction in AGO2 levels (Fig 5B). As a result, AGO2 silenced cells had significantly increased viral RNA levels compared to controls cells at 24 and 48 hpi (Fig 5A). These results indicate that the siRNA pathway has an important role in controlling viral infection in L. longipalpis.
In addition to the activation of the siRNA pathway, production of virus-derived piRNAs has been observed during viral infection in dipterans. In vivo, Aedes mosquitoes infected with Chikungunya virus (CHIKV) and Phasi Charoen like-virus induced clear production of piRNAs [22, 23]. Virus-derived piRNAs have been detected in cell lines derived from mosquitoes and also other dipterans such as culicoides and D. melanogaster [22, 49–51]. However, it remains unclear whether other dipterans generate virus-derived piRNAs in vivo since in vitro results lack the context of whole animals. For example, despite numerous efforts, virus-derived piRNAs have not been detected in adult Drosophila [24]. In this regard, sandflies are closely related to mosquitoes and fruit flies and could help understand whether virus-derived piRNAs appeared only in the mosquito lineage or were lost in Drosophila. Thereby we separately analyzed virus-small RNAs in the size range of piRNAs, 24–30 nt, detected both in adult L. longipalpis and cell lines. We observed accumulation of longer virus-derived small RNAs in adult sandflies infected with VSV and Lulo cells infected with LV1 and LV2 (Fig 6A). These small RNAs did not exhibit 1U or 10A enrichment or 10 nt overlap between sense and antisense strands that are considered canonical piRNA characteristics (Fig 6B and 6C).
In order to further investigate the absence of virus-derived piRNAs, we re-analyzed previously published small RNA datasets from L. longipapis infected with three other viruses, LPRV1, LPRV2 and LPNV [23]. These viruses had a broad size distribution of smalls RNAs but none showed characteristics of canonical piRNAs (S3 Fig). Together, our results suggest that the piRNA pathway is not engaged by infection with VSV, LV1 and LV2 in L. longipapis.
In addition to the production of virus-derived small RNAs, modulation of host small RNAs, especially miRNAs, may also occur during viral infection in insects [52]. Thus, we first analyzed the profile of endogenous small RNAs derived from L. longipalpis in infected and control individuals. We observed no differences in profile of small RNAs between infected and control sandflies. In both cases, the size distribution of small RNAs showed two clear peaks between 21–23 and 24–30 nt representing the typical length of siRNAs/miRNAs and piRNAs, respectively (S4 Fig).
In order to analyze changes in specific small RNAs during VSV infection, we first identified and annotated miRNA genes in the L. longipalpis genome (data from the sandfly genome consortium). In this effort, we were able to identify 206 miRNAs, most of them conserved in other dipterans such as D. melanogaster and A. aegypti. Using this reference, we analyzed differential expression of miRNAs in sandflies infected with VSV compared to control individuals at 2, 4 and 6 dpf. At 2 and 6 dpf, we observed no significant differences in the expression of Lutzomyia miRNAs between infected and control individuals. At 4 dpf, we detected 5 miRNAs that were significantly modulated by VSV infection (Fig 7). llo-miR-11-5p and llo-miR-263a-5p were up-regulated in VSV infected individuals while llo-miR-new3-3p, llo-miR-929-5p and llo-miR-79-3p were down-regulated. However, it is noteworthy that the magnitude of changes in miRNA expression was smaller than 2-fold between infected and control individuals. Together, these results show that VSV infection causes mild changes in host miRNAs and other endogenous small RNAs. However, it is possible that more pronounced changes might be observed if specific tissues are analyzed.
Here, we successfully developed a laboratory model to dissect the interaction between sandflies and viruses. Sandflies are well recognized as vectors for species of Leishmania but we lack information about host-pathogen interactions in the context of viruses. This work is, to the best of our knowledge, the first to analyze antiviral responses against arboviruses in sandflies. Here, we focused in the characterization of RNA interference mechanisms since they mediate powerful antiviral responses in insects such as fruit flies and mosquitoes [22, 24, 28–30, 32, 34–36]. In this regard, our work in sandflies is also an important contribution since most of our understanding about insect small RNA responses to virus infection is based on fruit flies and mosquitoes.
Our results indicate that the siRNA pathway has a conserved antiviral role in sandflies despite previous result from our own group that suggested the absence of virus-derived siRNAs in Lutzomyia [23]. Here, we observed that virus-derived siRNAs are produced in vivo and in vitro and pre-engagement of the siRNA pathway had a direct effect on VSV replication. Furthermore, silencing of AGO2 in Lutzomyia cells leads to increased viral replication, which suggests they fail to control infection. This antiviral role of the siRNA pathway in sandflies is in agreement with numerous observations in insects and other invertebrates [32, 53, 54]. The broad size profile of small RNAs derived from the viruses we described before suggest that they had the ability to inhibit the siRNA pathway in L. longipalpis rather than the absence of such response [55]. We note that most studies on the natural antiviral role of RNAi in vector insects have focused on arboviruses that have positive-stranded RNA genomes [22, 34, 35, 56, 57]. VSV is a negative stranded RNA virus that is controlled by the siRNA pathway during artificial infections in Drosophila [32, 36]. Our data show that this is also the case in sandflies that are natural vectors for this arbovirus in the wild.
The conspicuous absence of virus-derived piRNAs in sandflies suggests that viruses do not engage the piRNA pathway in these insects. Previous results have indicated that viruses activate the piRNA pathway in mosquitoes but not fruit flies, at least in vivo [22, 24]. Sandflies provide an important evolutionary perspective since they are closely related to mosquitoes and fruit flies. Therefore, activation of the piRNA pathway by virus infection seems to have been an acquired characteristic of the mosquito lineage although its role in antiviral defense remains unclear. However, it is still possible that activation of the piRNA pathway is limited to certain viruses or restricted to sandfly tissues that were not analyzed in our model.
In addition to activation of siRNA and piRNA pathways targeting the virus, we investigated whether host miRNAs changed in response to infection. Modulation of specific host miRNAs during infection has been observed in insects, which may be part of a coordinated host response to infection or a consequence of viral replication [58]. However, we observed that VSV induced little changes in the expression of host miRNAs, which is consistent with this infection being well tolerated by sandflies.
Finally, our model can be further used to dissect more complex interactions involving multiple sandfly-borne pathogens and understand other aspects of immune responses. Notably, epidemiologic data showed strong association between transmission of Leishmania and Phlebovirus in Southern France [59]. This result suggests that viruses may affect sandfly vectorial capacity for Leishmania and vice-versa. Viruses may decrease host fitness and affect the ability of sandflies to transmit other pathogens. Good vectors need to tolerate infections to prevent loss of fitness and maintain pathogen levels to allow transmission to a vertebrate host. Our work should allow further studies to dissect the possible effects of tripartite vector-virus-Leishmania interactions in sandflies. It is also important to point out that our work is also the first description of small RNAs in sandflies. This opens the doors for future studies on RNAi mechanisms in other contexts such as the role of specific miRNAs during Leishmania infection.
Vero cells obtained by ATCC (Maryland, USA) were grown in Dulbecco's Modified Eagle's Medium (DMEM) (Life Technologies) supplemented with 5% FBS and penicillin/ streptomycin (final concentration 100 units/mL, 100 μg/ mL, respectively) at 37°C/ 5% CO2. Lutzomyia longipalpis embryonic LL5 [42] and Lulo [43] cells were a kind gift from Dr. André Pitaluga (Fiocruz–Rio Janeiro). Both cells and Aedes albopictus C6/36 cells were grown in L-15 medium (SIGMA—Aldrich) supplemented with 10% fetal bovine serum (FBS) (GIBCO) and 1% penicillin/ streptomycin (final concentration 100 units/mL, 100 μg/mL respectively, Life Technologies), at 28°C.
VSV Indiana virus expressing green fluorescent protein (GFP) were a kind gift from Dr. Curt Horvath (Northwestern University). Stocks were prepared in Vero cells and virus titrations were done by plaque assay.
Six-well plates containing Vero monolayers with approximately 90% confluence were inoculated with virus and incubated at 37°C for 1h in an atmosphere supplemented with 5% of CO2. DMEM supplemented with 2% FBS was added to each well in a volume sufficient to maintain cell monolayers during the subsequent incubation period of 72h at 37°C in atmosphere supplemented with 5% of CO2. Vero monolayers were fixed with formalin at 10% (MERCK MILLIPORE, USA) and stained with crystal violet solution at 1% (SYNTH, Brazil) allowing naked eye observation of cytopathic effects. All samples were tested in triplicate.
LL5 cells (3x106 per well) were seeded in plaque one day prior to infection and infected with VSV at a multiplicity of infection (MOI) of 10 by directly adding the virus to the culture medium. One hour post infection the culture medium was removed, the cells washed with PBS followed by new culture medium addition. The cells were maintained at 25°C and the cells and supernatants were harvested for RNA isolation and plaque assay.
dsRNA targeting L. longipalpis AGO2, VSV and Firefly luciferase were synthesized using the T7 and SP6 megascript kit. 7 μg of dsRNA was transfected into LL5 cells using Cellfectin (Invitrogen) according to the manufacturer’s protocol. 3 days after transfection, cells were infected with a MOI of 1 PFU/cell of VSV.
The sandflies used in this study were obtained from a colony reared in the laboratory that was originally started from individuals from Teresina, Brazil, and maintained at the Laboratory of Physiology of Hematophagous Insects (Department of Parasitology, Instituto de Ciências Biológicas, Universidade Federal de Minas Gerais). Adult sandflies were fed a 70% sucrose solution (w/v) and blood-fed with anesthetized (ketamine, 200 mg/kg) hamsters to trigger egg development. After oviposition, eggs were collected and reared to preserve the colony. All larval instars were fed a crushed mixture of rabbit feces, rabbit chow, and garden soil. Third and fourth instars were supplemented with a mixture of white soy protein and cereal flakes (1:1). The insects were maintained at 27°C, humidity of 80–95%, and a photoperiod schedule of 12 h light/12 h dark.
Three-day-old female sand flies and private of sugar solution food for a period of 24h were used in this study. The insects were fed with heparinized blood through chick skin membrane attached in an artificial feeder (Hemotek). The artificial infection was performed with blood containing VSV or a control without virus at 37°C. Fully engorged females were selected and harvested individually at 2, 4 and 6 dpf. Insects were anesthetized with carbon dioxide and directly ground in TRIzol (Invitrogen) using glass beads as previously described [23].
All procedures involving animals were approved by the ethical review committee of the Universidade Federal de Minas Gerais (CEUA 33/2016 to M.R.V.S).
Total RNA was extracted from individual insects and cells using Trizol reagent according to the manufacturer’s protocol (Invitrogen). 1 μg of total RNA was reverse transcribed using 250 ng of random primers specific primers per reaction. The resulting cDNA was used as template for qPCR reaction containing SYBR Green (Invitrogen) and primers specific for the amplification of the genes of interest. The relative amount of the indicated RNAs normalized to an internal control (Rpl32) was calculated using the delta Ct method. Reverse transcription reactions were performed in the absence of primers or enzyme as negative controls for qPCR to ensure the identity of the products. The endogenous gene Rpl32 qPCR was used as normalization standard. Oligonucleotides designed in this study are described in S2 Table.
For libraries construction, individual insects were first tested for the presence and levels of VSV by RT-qPCR. RNA samples were then pooled according to the viral load in each insect. Total RNA extracted from adult insects or cell lines were used from the construction of small RNA libraries. Small RNAs were selected by size (~18–30 nt) on a denaturing PAGE before being used for construction of libraries as described [36]. All sequencing runs were performed at the IGBMC Microarray and Sequencing platform, a member of the ‘France Génomique’ consortium (ANR-10-INBS-0009), as 1 × 50 base pairs using a HiSeq 2500 instrument. Accession numbers of all libraries described in this study are in S1 Table.
Reference genome of L. longipalpis (Jacobina strain, version J1.2) was downloaded from the VectorBase website (www.vectorbase.com). The accession number KU721836 was the reference genome of VSV.
Nucleotide sequences of core siRNA genes from D. melanogaster (Dicer-2, AGO2, and r2d2) were used as query to identify putative homologs in the genomes of L. longipalpis using SoftBerry software [60]. Putative orthologs were analyzed for domain conservation in predicted protein by using InterProScan [61]. Gene IDs are shown in S3 Table.
miRNA genes in L. longipalpis were identified using mirDeep [62]. miRNA genes were annotated as part of sandfly genome consortium. Small RNA libraries were mapped against miRNA genes annotated in the L. longipalpis genome. Bowtie [63] was used for mapping allowing 0 mismatches in the seed and 1 mismatch in the rest of sequence. miRNA counts were normalized and used to evaluate differential expression among samples using R with package DESeq2 considering p<0.05 [64].
Pattern-based analysis, small RNA size profile, 5’ base enrichment, density of coverage and additional data analysis were evaluated using in-house Python, Perl and R scripts. Statistics of 5’ base enrichment was calculated as described [32]. All of the experiments were analyzed for statistical significance using the software GraphPad Prism. The Shapiro-Wilk normality test was first applied. Kruskall Wallis test was used for multiple comparisons. p<0.05 was considered statistically significant.
|
10.1371/journal.pgen.1006429 | Proteomic Landscape of Tissue-Specific Cyclin E Functions in Vivo | E-type cyclins (cyclins E1 and E2) are components of the cell cycle machinery that has been conserved from yeast to humans. The major function of E-type cyclins is to drive cell division. It is unknown whether in addition to their ‘core’ cell cycle functions, E-type cyclins also perform unique tissue-specific roles. Here, we applied high-throughput mass spectrometric analyses of mouse organs to define the repertoire of cyclin E protein partners in vivo. We found that cyclin E interacts with distinct sets of proteins in different compartments. These cyclin E interactors are highly enriched for phosphorylation targets of cyclin E and its catalytic partner, the cyclin-dependent kinase 2 (Cdk2). Among cyclin E interactors we identified several novel tissue-specific substrates of cyclin E-Cdk2 kinase. In proliferating compartments, cyclin E-Cdk2 phosphorylates Lin proteins within the DREAM complex. In the testes, cyclin E-Cdk2 phosphorylates Mybl1 and Dmrtc2, two meiotic transcription factors that represent key regulators of spermatogenesis. In embryonic and adult brains cyclin E interacts with proteins involved in neurogenesis, while in adult brains also with proteins regulating microtubule-based processes and microtubule cytoskeleton. We also used quantitative proteomics to demonstrate re-wiring of the cyclin E interactome upon ablation of Cdk2. This approach can be used to study how protein interactome changes during development or in any pathological state such as aging or cancer.
| The proliferation of mammalian cells is driven by proteins called cyclins, which bind and activate their catalytic partners, the cyclin-dependent kinases (Cdks). Cyclin-Cdk complexes drive cell proliferation by phosphorylating several cellular proteins. This study focuses on the E-type cyclins (cyclins E1 and E2). These proteins were shown to play important roles in cell cycle progression, and are often overexpressed in human cancers. It remained unclear, however, whether in addition to their well-established cell cycle roles E-type cyclins perform distinct tissue-specific functions in vivo, in the living mouse. To address this question we developed a system, which combines gene targeting in embryonic stem cells with high-throughput mass spectrometry. This approach allowed us to determine the identity of cyclin E protein partners, as well as phosphorylation substrates of cyclin E and its associated kinase, Cdk2, in different mouse organs. These analyses identified several novel, previously unanticipated interactors and phosphorylation substrates of cyclin E-Cdk2, and revealed that cyclin E plays distinct molecular functions in different compartments in vivo. This approach can be used to study the in vivo function of essentially any protein, in any model organism, and in any physiological and pathological state.
| E-type cyclins (cyclins E1 and E2, collectively referred to as ‘cyclin E’) represent components of the core cell cycle machinery. The two E-cyclins are encoded by separate genes, but they show substantial amino acid sequence similarity. In proliferating cells, E-cyclins become upregulated during the late G1 phase. Once induced, E-cyclins bind and activate their catalytic partner, the cyclin-dependent kinase 2 (Cdk2). Cyclin E-Cdk2 complexes phosphorylate proteins involved in cell cycle progression (the retinoblastoma protein pRB, p107, p130, p27Kip1), centrosome duplication (NPM1, CP110), histone biosynthesis (p220NPAT) and DNA replication (Cdc6, MCMs), thereby driving cell proliferation [1,2]. Consistent with growth-promoting roles for E-cyclins, amplification of the cyclin E1 and/or E2 genes and pathological overexpression of cyclin E proteins were documented in a wide range of human cancer types [1].
While the E-type cyclins have been extensively studied using biochemical approaches and in in vitro cultured cells, much less is known about the molecular functions of these proteins in different cell types within a living organism. In particular, it is not known whether cyclin E plays distinct molecular functions in different compartments or at different stages of development. Analyses of mice lacking E-cyclins revealed that both cyclin E1-null and E2-null mice are viable and develop relatively normally [3,4]. The only phenotype observed in cyclin E2-deficient mice was a defect in spermatogenesis leading to decreased male fertility. This phenotype was further exacerbated in mice with reduced dosage of cyclin E1 (E1+/-E2-/-), and was most pronounced upon conditional ablation of both E-cyclins in the male germline (E1Δ/ΔE2-/-), indicating that the two E-cyclins perform redundant functions in spermatogenesis [3,5]. Strikingly, the testicular phenotype of cyclin E-deficient mice closely mimics abnormalities seen in knockout mice devoid of cyclin E catalytic partner, Cdk2 [6,7]. These observations strongly suggest that cyclin E-Cdk2 kinase plays an important function during the spermatogenic process. However, the molecular role of cyclin E-Cdk2 in the male germline is largely unknown.
Whereas genetic ablation of individual cyclins yielded viable mice, a combined ubiquitous deletion of both E-type cyclins resulted in an early embryonic lethality [3,4]. This has been taken as an indication that the two E-cyclins perform overlapping functions in normal development. However, the lethality of cyclin E-deficient animals hampered analyses of cyclin E function at later stages of development and in adult organs.
To overcome these limitations, and to investigate the molecular functions of cyclin E in different compartments of the living organism, we developed a system that involves high-throughput proteomic analyses of organs derived from genetically engineered mice. Using this system, we provide insights into tissue-specific molecular roles for cyclin E and its associated kinase, Cdk2 in vivo.
In order to elucidate the in vivo functions of cyclin E, we decided to generate knock-in mouse strains expressing tandemly (Flag- and hemagglutinin, HA-) tagged versions of cyclin E1 in place of wild-type cyclin E1. We reasoned that these mice would allow us to use tandem immunoaffinity purifications with anti-Flag and -HA antibodies, followed by repeated rounds of high-throughput mass spectrometry, to determine the repertoire of cyclin E1-associated proteins in essentially any tissue or cell type, and at any stage of development.
We inserted DNA sequences encoding Flag and HA tags into the amino terminus of cyclin E1, immediately downstream of the start codon, using gene-targeting in embryonic stem (ES) cells (Fig 1A). Subsequently, homozygous cyclin E1Ntag/Ntag mice were generated using standard procedures. Since a tag at a particular end of cyclin E1 molecules might destabilize the protein, or render it non-functional in vivo, we also generated the second knock-in strain, in which we inserted DNA sequences encoding these two tags into cyclin E1 carboxy terminus, immediately upstream of the stop codon, yielding cyclin E1Ctag/Ctag animals (Fig 1B). We verified that in the tissues of knock-in mice the tagged cyclin E1 alleles were expressed at the same levels as wild-type cyclin E1 in the corresponding tissues of control animals (Fig 1C and 1D). We also verified that both amino- and carboxy-terminally tagged cyclin E1 retained the ability to bind and to activate cyclin E catalytic partner, Cdk2 (Fig 1C and 1D). Like wild-type cyclin E1, tagged cyclin E1 was expressed at high levels in several organs of adult mice, as well as in embryonic brains (Fig 1E and 1F).
In our proteomic analyses we decided to focus on five compartments: embryonic brains, adult brains, spleens, thymuses and testes, as these organs expressed particularly high levels of cyclin E1 (Fig 1E and 1F and S1 Fig). First, we analyzed protein lysates from mouse organs using size exclusion chromatography, to determine the molecular weight of cyclin E1-containing complexes. We found that cyclin E1 was present in a wide range of protein fractions (56–840 kD), suggesting that it forms a multitude of distinct protein complexes in vivo (Fig 1G). We next purified cyclin E1-containing protein complexes from each of the five organs using sequential immunoaffinity purifications with anti-Flag and -HA antibodies (Fig 1H), and identified cyclin E1-associated proteins using repeated rounds of liquid chromatography–tandem mass spectrometry (6–10 independent purifications/mass spectrometry runs). In parallel, we performed the same number of purifications/mass spectrometry analyses from control wild-type animals, which do not express tagged cyclin E1, and the identified proteins were subtracted as a background (see S1 Appendix).
These procedures allowed us to determine the identity of cyclin E1-associated proteins (‘E1-interactomes’) in embryonic brains, adult brains, spleens, thymuses and testes. In total, we detected 117 high-confidence cyclin E1 interactors (Fig 2, S2 Fig and S1 Table). Thirty-seven of these were detected in at least two different compartments (Fig 3A and S1 Table). These shared interactors contained essentially all well-established cyclin E-binding proteins, including Cdk2, ‘pocket proteins’ p107 and p130, as well as cell cycle inhibitors p27Kip1, p57Kip2 and p21Cip1, and were highly enriched for cell cycle proteins (Fig 2 and S1 Table).
In addition to these shared interactors, 80 proteins were found to associate with cyclin E1 only in one organ (Fig 3A). Cyclin E1-interactomes in embryonic and adult brains contained the highest fraction of unique interactors (57.5% and 56.0%, respectively), followed by spleens (34.7%), testes (21.4%) and thymuses (5.3%) (Fig 3B). Surprisingly, the fraction of shared proteins between embryonic and adult brain interactomes was very low (8.9%), suggesting distinct molecular roles for cyclin E in embryonic versus in adult brains. In contrast, testes shared nearly half of their interactors (42.4%) with thymuses and 32.8% with spleens, while the overlap between thymic and splenic interactomes was 28.3% (Fig 3C).
We next assigned Gene Ontology functions to interactors found in different organs. In spleens, thymuses and embryonic brains the most frequent function was cell cycle, while transcription represented the second most frequent category. In addition, over 10% of cyclin E1 interactors in spleens and in thymuses belonged to proteins involved in protein ubiquitination and in metabolism, respectively (Fig 3D and S2 Table). In the testicular interactome, transcription and cell cycle constituted two equally most frequent functions. In the adult brains, in addition to cell cycle proteins, a significant fraction of cyclin E1-interactors represented proteins involved in neuronal function (20%), and 16% belonged to proteins that play roles in regulation of microtubules and cytoskeleton (Fig 3D and S2 Table).
We constructed a biological process enrichment heat map of cyclin E1 interactors (Fig 3E and S3 Table). As expected, cell cycle control category was enriched across all organs. In addition, we observed that apoptosis/cell death and chromatin modification categories were enriched in multiple compartments. Indeed, cyclin E was postulated to play roles in both processes [9–11]. Strikingly, embryonic brain and adult brain interactors shared functions involved in neuron development, neurogenesis and regulation of synaptic plasticity, while only interactors from adult brains showed enrichment for regulation of microtubule-based processes and microtubule cytoskeleton (Fig 3E). These findings indicate that cyclin E plays previously unanticipated functions in neuronal differentiation as well as in regulating microtubules and cytoskeleton in terminally differentiated neurons.
Cyclin E was previously shown to interact with Cdk2 and, to a lesser extent, with Cdk1 [12]. Indeed, we detected Cdk2 bound to cyclin E1 in all organs, whereas we observed Cdk1 in some (embryonic brains, spleens, testes) but not in other compartments (thymuses, adult brains) (Fig 2A and S1 Table).
Unexpectedly, we observed association of cyclin E with Cdk5 in all five organs studied (Fig 2A and S1 Table). Cdk5, together with its regulatory partners p35 and p39 represents neuronal-specific kinase that plays key roles in neuronal differentiation [13]. Cyclin E1 was shown to interact with Cdk5 in terminally differentiated neurons, where it negatively regulates phosphorylation of synaptic proteins [8]. Our unexpected finding that cyclin E1 interacts with Cdk5 in all compartments studied here indicates that cyclin E-Cdk5 complexes likely play much wider physiological roles than previously anticipated, by acting outside the nervous system.
We also observed, and confirmed by immunoprecipitation (IP)–western blotting, an unexpected association of cyclin E1 with another cell cycle kinase, Cdc7 (Fig 2A, S2A Fig and S1 Table). This protein is known to interact with its regulatory subunits Dbf4 and Drf1, and to play a rate-limiting role in firing mammalian DNA replication origins [14,15]. Importantly, we did not detect Dbf4 or Drf1 as cyclin E1-interacting proteins, suggesting that cyclin E1 forms distinct complexes with Cdc7. Of note, cyclin E-Cdk2 and Dbf4-Cdc7 complexes were postulated to play distinct molecular functions during firing of mammalian DNA replication origins [16]. Our findings raise a possibility that cyclin E may contribute to this process through cyclin E-Cdk2 as well as cyclin E-Cdc7 complexes.
Collectively, these results revealed that cyclin E interacts in vivo with a much wider range of cell cycle-related kinases than previously appreciated.
We reasoned that application of quantitative proteomic methods would allow us to visualize how the cyclin E1 interaction network is perturbed in different pathological states. As a proof of principle, we decided to analyze the changes in cyclin E1 interactome upon ablation of Cdk2, the major catalytic partner of cyclin E. To address this, we bred cyclin E1Ntag/Ntag mice with Cdk2+/- animals and generated Cdk2-/-/cyclin E1Ntag/Ntag mice. We then used spleens and thymuses from Cdk2-/-/cyclin E1Ntag/Ntag and from cyclin E1Ntag/Ntag mice (control, Cdk2+/+) to isolate cyclin E1-associated proteins using sequential immunoaffinity purifications, as above. We then compared the abundance of cyclin E1 interactors between these two genotypes using isobaric tags for relative and absolute quantification (iTRAQ) approach. In this method, protein purification products from different samples are labeled with isobaric tags containing different reporter isotopes. After labeling, the samples are mixed and subjected to quantitative mass spectrometry analysis. Analogous peptides derived from each sample are distinguished due to the mass differences of the isotope reporter ions, and the ratio of the reporter ion peak intensities reflects the relative abundance of the peptides in each sample [17].
We found that ablation of Cdk2 led to a strong increase in association of cyclin E1 with several other cyclin-dependent kinases, such as Cdk1, Cdk4 and Cdk5 (Fig 4A and S4 Table). While association of cyclin E1 with Cdk1 in Cdk2-null cells has been reported before [12], increased binding to Cdk4 and Cdk5 was unexpected. We verified using IP–western blotting that in the wild-type setting, cyclin E1 bound predominantly to Cdk2. However, ablation of Cdk2 led to a dramatic increase of cyclin E1-Cdk1, E1-Cdk4 and E1-Cdk5 interactions, both in Cdk2-null spleens and thymuses (Fig 4B and S3A Fig). Importantly, binding of Cdk1 and Cdk4 to their physiological cyclin partners (cyclins B1 and D1, respectively) remained unchanged in Cdk2-/- organs (S3B Fig), arguing against re-distribution of Cdk1 and Cdk4 away from its normal partners to cyclin E, and suggesting that the “free” pool of Cdks bound to cyclin E in the absence of Cdk2. Interestingly, binding of cyclin A2 to Cdk1 was increased in Cdk2-/- cells (S3B Fig). Like cyclin E, cyclin A2 normally interacts with both Cdk2 and Cdk1, and ablation of Cdk2 increases cyclin A2 binding to the Cdk1 subunit.
We also tested the effect of chemical inhibition of Cdk2 kinase activity on the composition of cyclin E-containing complexes. To do so, we used CVT-313, a reversible ATP-competitive inhibitor of Cdk2 (IC50 = 0.5 μM), which can also inhibit cyclin Cdk1 with lower potency (IC50 = 4.2 μM) [18]. In contrast to genetic Cdk2 ablation, inhibition of Cdk2 activity did not increase the interaction of cyclin E with other Cdks (Fig 4C), indicating that the loss of Cdk2 protein, rather than inhibition of Cdk2 kinase triggered the observed re-wiring.
In addition to changes in cyclin E-Cdk complex composition, ablation of Cdk2 led to the loss of cyclin E1 binding to p107 (Fig 4A and 4B). Hence, Cdk2 is required for cyclin E1-p107 interaction, and other Cdks cannot replace Cdk2 in this process.
To the best of our knowledge, our quantitative proteomic analysis of Cdk2-/- organs described above represents the first unbiased, proteome-wide study of how cells re-wire protein-protein interactions upon a particular genetic insult.
Since cyclins are known to physically interact with cyclin-Cdk phosphorylation substrates [19], we hypothesized that mining of the cyclin E1 interactome might allow us to identify new, possibly tissue-specific substrates of cyclin E-Cdk2 kinase. To test this prediction, we first intersected the cyclin E1 interactome with the list of Cdk2 substrates at PhosphoSitePlus (Cell Signaling Technology). Among cyclin E1 interactors we observed several well-established Cdk2 phosphorylation substrates, including p107, p130, p27Kip1, Mcm proteins (Mcm3 and Mcm4), Cdc6 and Brca2 (Fig 2A and S1 Table). In total, fifteen out of 117 interactors (12.8%) represented known Cdk2 substrates, compared with 1.95% in the whole proteome (p = 8.9 x 10−9) (Table 1).
We next screened the amino acid sequence of all 117 cyclin E1 interactors with a Scansite 3.0 program [20]. This software allows one to identify proteins containing a Cdk phosphorylation motif, which hence might represent bona fide Cdk2 phosphorylation substrates. We found that 39/117 (33.3%) proteins contained high-confidence and 70/117 (59.8%) medium-confidence Cdk phosphorylation motifs, again a very strong enrichment as compared to the whole proteome (p = 2.4 x 10−15 for high-confidence and p < 2.2 x 10−16 for medium-confidence substrates) (Table 2 and S2 Table). We concluded that the cyclin E1 interactome likely contains several novel Cdk2 substrates. Gene Ontology analysis of predicted Cdk2 substrates revealed strong enrichment for cell cycle (p = 1.5 x 10−17) and transcriptional functions (p = 5.0 x 10−4, S2 Table), indicating that cyclin E-Cdk2 kinase influences these functions in vivo.
Analyses of individual organ-specific interactomes revealed that in all organs except for testes, approximately one third of proteins represented predicted high confidence Cdk2 phosphorylation substrates, whereas approximately half corresponded to medium confidence substrates. In contrast, in testes this proportion was significantly higher (43.3% and 66.7%, respectively, Table 3), raising a possibility that cyclin E-Cdk2 kinase plays particularly important functions in this compartment. We further explored this possibility in our mechanistic studies of testicular interactome described below.
Our analyses of cyclin E1 interactomes from different organs revealed that cyclin E1 associates with components of the DREAM complex in several proliferating tissues (Figs 2 and 5A and S1 Table). The DREAM complex represents a group of proteins that have been conserved between C. Elegans, Drosophila and humans [21,22]. In mammalian cells, the DREAM complex consists of a ‘pocket’ protein (p107 or p130), transcription factors E2f4 or E2f5 and Dp1 or Dp2, and five proteins homologous to products of the C. Elegans synMuvB group of genes: Lin9, Lin37, Lin52, Rbbp4 (Lin53) and Lin54 [22]. The mammalian DREAM complex physically binds to E2f responsive promoters in quiescent cells and represses their transcription. During G0 → S phase progression, the Lin subunits dissociate from the p130-E2f-Dp module through an unknown mechanism, and they form a new DNA-binding complex with Myb transcription factor [22].
We first verified the interaction of cyclin E1 with Lin proteins and other components of the DREAM complex in mouse spleens by IP-western blotting (Fig 5B). Although we did not detect Lin37 as cyclin E1 interactor in our mass spectrometry analysis, IP-western blot analysis confirmed that also this DREAM complex subunit associated with cyclin E1 (Fig 5B, S4A and S4B Fig).
Next, we performed immunoprecipitation followed by re-IP and western blotting to verify that cyclin E1, p130 and Lin proteins are present within the same complex (Fig 5C). We also determined that cyclin E1 associates with Lin proteins and with other components of the DREAM complex in mouse embryonic fibroblasts (MEFs) (S4A Fig) and in human glioblastoma T98G cells (S4B Fig). Importantly, when we used for immunoprecipitation an anti-Lin52 antibody which recognizes only ‘free’ Lin proteins, but fails to bring down Lin52 associated with p130, E2f and Dp because its epitope overlaps with the p130-binding site in Lin52 [23,24], we found that this pool of ‘free’ Lin proteins does not associate with cyclin E1 (S4B Fig). These findings indicate that cyclin E1 binds to the DREAM complex through the p130-E2f-Dp module. Consistent with this notion, interaction of cyclin E1 with the DREAM complex was abrogated in knockout MEFs lacking p107, p130 and pRB (S4C Fig).
To determine when during cell cycle progression cyclin E1-DREAM interaction takes place, we synchronized T98G cells by serum starvation and then forced them to re-enter the cell cycle by addition of serum. We found that cyclin E1 associates with Lin37 starting at approximately 10 hrs post-stimulation; the onset of cyclin E1-Lin37 binding coincided with decreased interaction of Lin37 with p130 and preceded association of Lin37 with B-Myb (Fig 5D). This timing suggested that cyclin E1 might play a role in disrupting the DREAM complex.
We determined that cyclin E1 kinase partner Cdk2 also associated with the DREAM proteins (Fig 5C), with kinetics essentially identical to cyclin E1-DREAM interaction (Fig 5D) suggesting that cyclin E-Cdk2 might phosphorylate components of this complex and trigger its disassembly. To test this prediction, we synchronized T98G cells as above and released them in the presence of a CDK2 inhibitor, CVT-313. Indeed, inhibition of CDK2 kinase prevented dissociation of Lin37 from p130 (Fig 5D). Strikingly, CVT313-treatment also prevented binding of cyclin E1 and Cdk2 to Lin37, indicating that an active Cdk2 kinase is needed for cyclin E1/Cdk2-DREAM interaction (Fig 5D).
Next, we ectopically expressed cyclin E1 and Cdk2 together with p130 in T98G cells. We found that expression of cyclin E-Cdk2 led to disruption of the p130-Lin complex (S4D Fig). Significantly, cyclin E1-Cdk2 was also able to decrease binding of Lin subunits to phosphorylation-deficient p130 mutant in which all Cdk phosphorylation sites have been mutated to alanines [25], suggesting that phosphorylation of Lin components by cyclin E1-Cdk2 might contribute to disruption of the p130-Lin complex (S4D Fig). Consistent with this hypothesis, analyses of the cyclin E1 interactome with Scansite 3.0 (S2 Table) indicated that, in addition to a well established Cdk2 substrate p130, four additional components of the DREAM complex, Lin9, Lin37, Lin52 and Lin54 represented predicted Cdk phosphorylation substrates.
To test whether these proteins are indeed phosphorylated by cyclin E-Cdk2, we incubated purified recombinant Lin9, Lin37, Lin52 and Lin54 with cyclin E1-Cdk2 kinase in the presence of radioactive ATP. We found that cyclin E1-Cdk2 readily phosphorylated these proteins in vitro (Fig 5E). Subsequently, we used mass spectrometry to map residues on Lin proteins that are phosphorylated by cyclin E1-Cdk2 (Fig 5F and S5 Fig). We extended these observations in vivo using cells engineered by us to ectopically express ‘analog-sensitive’ Cdk2 together with cyclin E1 and Flag-tagged Lin37. Unlike wild-type kinases, analog-sensitive kinases can use ‘bulky’ N6-substituted ATP-analogs such as N6-furfuryl-ATP (6-Fu-ATP) to phosphorylate their substrates. Therefore, by supplementing cells expressing analog-sensitive kinases with thio-containing 6-Fu-ATP (6-Fu-ATPγS), one can label their substrates with thiophosphate moieties [26,27]. IP with anti-Flag antibody followed by immunoblotting with anti-thio-phosphate antibody revealed that cyclin E-Cdk2 kinase indeed phosphorylates Lin37 in vivo (Fig 5G).
Collectively, these findings suggest that cyclin E-Cdk2 kinase may play an important role in disrupting the DREAM complex during G0 → S phase progression, likely through phosphorylation of multiple DREAM subunits.
We next focused our attention on the cyclin E1 testicular interactome. As mentioned above, genetic ablation of the E-type cyclins led to severe defects in spermatogenesis [3,5]. Strikingly, mice lacking cyclin E kinase partner, Cdk2, displayed a very similar phenotype with spermatocytes arrested at the pachytene stage [6,7]. The molecular basis of this phenotype remains unknown. We hypothesized that cyclin E plays a Cdk2-dependent function in testicular development, through phosphorylation of key proteins that drive the spermatogenic process.
Our analyses of the testicular cyclin E1 interactome revealed the presence of several proteins that have been shown to play important roles in spermatogenesis, notably Mybl1, Dmrtc2, Cdc7 and cyclin B3, in addition to Cdk2 (Fig 6A and S1 Table). In particular, Mybl1 is thought to represent a ‘master regulator’ of male meiosis [28]. Intriguingly, mice lacking Mybl1 show defective spermatogenesis, which resembles the phenotype of Cdk2- or cyclin E-deficient mice [28,29]. Dmrtc2 is also essential for male spermatogenesis, and Dmrtc2-deficient mice show a similar testicular phenotype with spermatogenic arrest at the pachytene stage [30,31]. Importantly, our Scansite 3.0 analyses of the testicular interactome identified both Mybl1 and Dmrtc2 as predicted Cdk phosphorylation substrates (S2 Table). These findings suggested that Mybl1 and Dmrtc2 might represent essential downstream targets of cyclin E-Cdk2 kinase in testes. Consistent with this possibility, we confirmed the physical interaction between the endogenous Mybl1 and Dmrtc2 proteins and cyclin E and Cdk2 in the mouse testes, by IP–western blotting (Fig 6B).
To test whether Mybl1 and Dmrtc2 can be phosphorylated by cyclin E-Cdk2, we performed in vitro kinase assays using recombinant proteins. Indeed, we detected phosphorylation of both proteins by cyclin E-Cdk2 (Fig 6C). We next examined in vivo phosphorylation of Mybl1 and Dmrtc2 by cyclin E-Cdk2 using cells engineered by us to express Flag-tagged Mybl1 or Dmrtc2 together with cyclin E and analog-sensitive Cdk2. Immunoprecipitation with anti-Flag antibody followed by western blot analysis with anti-thio-phosphate antibody revealed that cyclin E-Cdk2 kinase indeed phosphorylates Mybl1 and Dmrtc2 in vivo (Fig 6D). We next identified residues of Mybl1 and Dmrtc2 that are phosphorylated by cyclin E-Cdk2 using mass spectrometry. A total of ten cyclin E-Cdk2 phosphorylation sites in Mybl1 and three sites in Dmrtc2 were detected (Fig 6E and S6 Fig).
The essential role of Mybl1 in spermatogenesis is thought to be mediated by the ability of this transcription factor to regulate expression of crucial meiotic genes such as Miwi and Morc2b. Indeed, disruption of these Mybl1 downstream targets can also cause abnormal spermatogenesis [28]. During the course of normal spermatogenesis, Mybl1 expression increases at the pachytene stage, leading to increased transcription of Mybl1 downstream targets such as Miwi (Fig 7A). Strikingly, in Cdk2-/- mice the expression of Mybl1 remained low (Fig 7B and 7C). Consequently, Cdk2-null testes failed to express normal levels of Miwi, a rate-limiting transcriptional target of Mybl1, as revealed by reverse transcription–quantitative PCR (RT-qPCR), western blotting and immunohistochemistry (Fig 7B–7E). These results suggest that phosphorylation of Mybl1 by cyclin E-Cdk2 is required to stabilize and activate Mybl1 at the pachytene stage.
In addition, cyclin B3 also represents a direct transcriptional target of Mybl1. Unlike the majority of Mybl1 target genes, where Mybl1 serves as an activator of transcription, expression of cyclin B3 is repressed by Mybl1 [28]. Cyclin B3 is highly expressed in pre-pachytene spermatocytes and downregulated when cells enter the pachytene stage and start expressing Mybl1 [32] (Fig 7A). Aberrant expression of cyclin B3 beyond the pachytene stage causes spermatogenic defects [33]. Consistent with reduced Mybl1 levels observed by us in Cdk2-null testes, we found that expression of cyclin B3 was elevated in the absence of Cdk2 (Fig 7F). These results suggest that cyclin E-Cdk2 negatively regulates expression of cyclin B3 through Mybl1, and that deregulated cyclin B3 expression also likely contributes to the testicular phenotype seen in Cdk2-null testes.
DNA-binding protein Dmrtc2 represented another testicular-specific substrate of cyclin E1-Cdk2 kinase identified in our analyses (Fig 6B–6D). Dmrtc2 is an essential regulator of spermatogenesis, and mice lacking this protein present impairment in spermatogenesis characterized by developmental arrest at the pachytene stage [30,31]. However, the molecular function of Dmrtc2 in spermatogenesis is not yet well understood. Western blotting of Cdk2-null testes revealed that ablation of Cdk2 led to a reduction of Dmrtc2 protein levels (Fig 7B and 7C). We propose that cyclin E-Cdk2 kinase may also affect spermatogenesis via Dmrtc2, by regulating the levels and function of Dmrtc2, through direct phosphorylation.
Of the two cyclin E proteins, cyclin E2 is more abundant than cyclin E1 in mouse testes [5,34]. Correspondingly, cyclin E2-null mice manifest more evident spermatogenic defects and decreased fertility, although the abnormalities are not as pronounced as those seen in Cdk2-null mice, due to compensation from cyclin E1 in E2-null cells [3,5]. Despite these differences in phenotypic severity, cyclin E2-/- testes showed very similar aberrant expression of Mybl1, Miwi and Dmrtc2 to that seen Cdk2-null mice (Fig 7B and 7C). Low expression of Miwi that persists in cyclin E2-null spermatocytes indicates low residual activity of Mybl1, which is likely caused by phosphorylation of Mybl1 by cyclin E1-Cdk2, and may explain why cyclin E2-/- mice have milder testicular phenotype than Cdk2-null animals.
Taken together, these results strongly suggest that cyclin E-Cdk2 kinase plays a critical role in the male germline by phosphorylating and controlling the activity of key regulators of spermatogenesis Mybl1 and Dmrtc2.
In this study we used knock-in mice expressing tandemly tagged cyclin E1 in place of the wild-type protein to delineate the proteomic landscape of cyclin E1 interactions in vivo. When combined with large-scale mass spectrometry, tagged knock-in mouse system allows one to determine the set of cyclin E1 interacting proteins (E1-interactome) in essentially any organ or cell type. We also demonstrate that this system can be used to identify novel, tissue-specific substrates of cyclin E-Cdk2 kinase. Moreover, by using quantitative proteomic approaches, one can visualize how the cyclin E1 interactome changes under different physiological or pathological conditions. As a proof of principle, we bred our tagged knock-in mice into the Cdk2-null background, and demonstrated re-wiring of the cyclin E1 interactome upon ablation of the major catalytic partner of cyclin E1. In the future, the same approach can be used to determine how the set of cyclin E1-interating proteins or cyclin E-Cdk2 phosphorylation substrates changes in any genetic background and upon any genetic insult. Furthermore, our system allows to visualize and to quantify how the cyclin E1-interactome changes at different stages of normal development or in different pathological states (for example, by comparing young versus aging stem cell compartments). It can also be used to study cyclin E function at different stages of the neoplastic process. Given the well-documented role of cyclin E1 in oncogenesis [35], it will be of interest to compare cyclin E1-interactomes and cyclin E1-Cdk2 phosphorylation targets in pre-malignant lesions, during tumor initiation, progression, and in the metastatic spread.
Our study illustrated that upon ablation of Cdk2, cyclin E1 binds several cell cycle kinases. Such re-wiring of interactomes likely takes place in several mouse knockout strains and compensates for the loss of a given protein. These observations underscore the fact that the experiments using knockout mice should not be over-interpreted, and absence of phenotypes does not rule out a physiological role for a given protein.
Our proteomic analyses revealed that regulation of cell cycle and transcription represents the major biological functions regulated by cyclin E1 (and E1-Cdk2 kinase) in vivo. The role of cyclin E in cell cycle progression has been extremely well documented. Several studies also implicated cyclin E1 in transcription [36–42]. Our analyses indicate that cyclin E plays a rate-limiting role in regulating transcription in vivo in the male germline, where it controls expression of key meiotic genes via Mybl1 (and likely Dmrtc2).
We found that cyclin E-Cdk2 kinase phosphorylates a master regulator of spermatogenesis, Mybl1, and that the levels of Mybl1 are decreased in Cdk2- and cyclin E2-null testes. Moreover, the expression of Mybl1 transcriptional targets (such as Miwi and cyclin B3) was deregulated, consistent with the loss of Mybl1 activity. Intriguingly, mice lacking cyclin E, Cdk2, Mybl1, or Miwi share similar spermatogenic defects [3,5–7,28,29]. We propose that cyclin E-Cdk2 kinase serves to maintain transcriptional activity of Mybl1 in the male germline. It remains to be seen how exactly cyclin E-Cdk2 regulates Mybl1 function. The absence of faithful in vitro systems to study meiotic cells precluded us from performing mechanistic analyses to address this point.
Dmrtc2 represents another critical protein implicated in mouse spermatogenesis, and spermatocytes lacking Dmrtc2 exhibit pachytene arrest, similar to testicular phenotype seen in Cdk2- or cyclin E-deficient animals [30,31]. We found that, like Mybl1, Dmrtc2 also represents a direct cyclin E-Cdk2 phosphorylation target, and that the levels of Dmrtc2 are reduced in the testes of mice lacking Cdk2 or cyclin E. Collectively, these results suggest that cyclin E-Cdk2 represents a crucial upstream regulator of the transcriptional cascade in the male germline, by acting through Mybl1 and Dmrtc2.
Analyses of cyclin E interactomes in embryonic and adult brains suggest novel, previously unanticipated functions for cyclin E in neurogenesis and in regulation of microtubule-based processes and microtubule cytoskeleton. These findings will now allow one to design hypothesis-driven studies to elucidate cyclin E functions in these processes, based on the interactors identified in our study. Cyclin E is highly expressed in adult mouse brains, where it was shown to regulate synaptic plasticity by inhibiting phosphorylation of synaptic Cdk5 substrates [8]. However, very little is known about the function of cyclin E in regulating neuronal cytoskeleton and in neurogenesis. Previous work in Drosophila revealed a role for cyclin E in fate determination in the central nervous system. Loss of cyclin E function in the developing neuroblast lineage was shown to result in generation of only glial cells, while ectopic expression of cyclin E led to generation of neuronal sublineage, in addition to the glial cells [43]. This role of cyclin E was attributed to regulation of localization and function of a homeobox protein Prospero [44]. Given results of our proteomic analyses, it seems likely that cyclin E affects neuronal differentiation also in mammalian cells via currently unknown mechanism(s).
The study described here focused on a key component of mammalian core cell cycle machinery, which has been conserved from yeast to humans. Our results indicate that while preserving their ‘core’ cell cycle functions, in the process of evolution these proteins acquired novel, tissue-specific roles. In the future, the same approach can be applied to study the function of any protein in any model organism. By combining and overlaying interactomes of various interacting proteins (for example kinases and all their regulatory and accessory partners) one will be able to visualize complex biological networks that control, in a cell type-specific fashion, different cellular functions. Unbiased, biocomputational analyses of these networks will help to understand the biological interplay between different proteins, and to elucidate how perturbations of components of these networks contribute to various pathological states.
Detailed procedures to generate knock-in mice carrying a Flag-HA tag at the N-terminus of the cyclin E1 gene (cyclin E1Ntag/Ntag) have been described previously [8]. A targeting vector to knock-in a Flag-HA tag at the C-terminus of the cyclin E1 gene was constructed by replacing a stop codon in the last exon with DNA sequences encoding a Flag-HA tag followed by a termination codon, and by inserting a loxP-flanked hygromycin (Hyg) resistance cassette into ScaI site in the intron 11 (Fig 1B). The construct spanned 10 kb KpnI–HpaI fragment of the cyclin E1 gene. The targeting vector was electroporated into embryonic stem (ES) cells and homozygous cyclin E1Ctag/Ctag animals were obtained using standard procedures [45]. Cyclin E1Ntag/Ntag mice were crossed with Cdk2+/- animals (kindly provided by Dr. Philipp Kaldis). All experiments conformed to the relevant regulatory standards, and were approved by the Institutional Animal Care and Users Committee.
Spleens, testes, thymuses, and brains were dissected from 1-month-old cyclin E1Ntag/Ntag or cyclin E1Ctag/Ctag mice. Embryonic brains (heads) were collected from E14.5–15.5 embryos. We used pooled 20 to 30 adult organs or 20 embryonic heads for a single purification. To maximize the capture of interactors, we used approximately 1:1 mixture of organs derived from amino- and carboxy-tagged mice. After homogenizing tissues, cyclin E1 and its associated proteins were immunoprecipitated using anti-Flag M2 agarose (Sigma), eluted twice with Flag peptide (Sigma), then immunoprecipitated again with anti-HA antibody (12CA5 ascites fluid, Covance) coupled to protein A sepharose beads (Amersham). Complexes were then eluted with 0.1 M glycine (pH 2.5). For mass spectrometry, purified protein complexes containing at least 200–300 ng of cyclin E1 were used for a single run. For each organ, we performed 6–10 purifications (each yielding 200–300 ng of cyclin E1), followed by 6–10 independent mass spectrometry runs. In parallel, we performed 6–11 ‘mock’ purifications from the same number of organs from wild-type mice, followed by 6–11 mass spectrometry runs. Detailed procedures for LC-MS/MS and iTRAQ have been described previously [8].
Please see S1 Appendix. LC-MS analyses of phosphorylated peptides were performed as previously [46].
Organs or cells were homogenized in lysis buffer (50 mM Tris-HCl pH 7.4, 150 mM NaCl, 0.5% NP-40, 10 mM NaF, and protease inhibitor cocktail). Proteins were separated on SDS-PAGE gels and transferred to Immobilon-P membranes (Millipore). Membranes were blocked with blocking buffer (TBST, 5% non-fat skim milk) before immunoblotting. Immunoblots were visualized by ECL (Pierce) or Odyssey imaging system (LI-COR). Quantification was performed with ImageJ. For immunoprecipitation, organs or cells were homogenized in lysis buffer containing 100 mM Tris-HCl pH 8.0, 100 mM KCl, 0.1% NP-40, 0.1% Tween 20, 10 mM NaF, and protease inhibitor cocktail.
Testes were collected from wild-type and Cdk2-/- mice at P15, fixed in 4% paraformaldehyde, and embedded in paraffin. Section (5 μm-thick) were deparaffinized through xylene and graded ethanol dilutions, followed by antigen retrieval by microwaving in PBS buffer (5 mM Tris pH 8.0, 1 mM EDTA). After blocking with 5% normal goat serum (NGS; Sigma) and 0.2% Triton X-100 in PBS for 1 hr at RT, sections were incubated with primary antibodies for 2 hrs in PBS containing 5% NGS and 0.2% Triton-X100 at RT, washed with PBS, then incubated with secondary antibodies (Alexa 568 and Alexa 488; Invitrogen) for 1 hr at RT in PBS with 5% NGS. After rinsing with PBS, sections were mounted with Vectashield mounting medium containing DAPI (Vector Laboratories) and analyzed on a fluorescent microscope (Nikon E600).
For immunoblotting, immunoprecipitation and immunostaining, we used antibodies against cyclin E1 (Santa Cruz, BioLegend or Millipore), cyclin D1, cyclin D3, cyclin B1, Cdk1, Cdk2, Cdk4, Cdk5, p107, Pcna, B-Myb, Fig 4, Rbbp9 (Santa Cruz), p130 (Santa Cruz or BD Transduction Labs or Bethyl Laboratories), HA (Covance), actin, Flag, Mybl1, cyclin A2 (Sigma), Gapdh (Invitrogen or Millipore), Lin9, Lin37, Lin52, Lin54, Rbbp4, Rbbp7 (Bethyl Laboratories), E2f4, Dp1, Cdc7 (Lab Vision), thiophosphate ester, Dmrtc2 (Abcam or Sigma), Mapk15 (Abcam), RalA, mTOR/Frap1, cyclin B3, phosphor-Rb, Tubulin and Miwi (Cell Signaling). Purified rabbit IgG was from either Bethyl Laboratories or Santa Cruz and mouse IgG from Santa Cruz. For immunoprecipitation, we also used anti-Flag M2 agarose (Sigma) and HA beads that were prepared by conjugating Protein A Sepharose (Amersham) with anti-HA antibody in 12CA5 ascites fluid (Covance).
To construct GST-Mybl1 and GST-Dmrtc2, cDNA fragments were amplified by PCR using human Mybl1 cDNA and mouse Dmrtc2 cDNA (from Dr. D. Zarkower) as templates and subcloned into pGEX-5X-3 vector (GE Healthcare). GST-Mybl1-N and GST-Mybl1-C contain DNA segments encoding N-terminal (aa 1–376) and C-terminal (aa 376–752) fragments of Mybl1, respectively. GST-Dmrtc2-N contains a DNA segment encoding N-terminal fragment (aa 1–201) of Dmrtc2. To construct GST-Lin9, -Lin37, -Lin52 and -Lin54, corresponding full-length human cDNAs were subcloned into pGEX-6P-3 vector (GE Healthcare). The GST-containing constructs were expressed in E. Coli BL21, and proteins were purified using glutathione sepharose (GE Healthcare).
To construct pCMV-Flag-Mybl1 and pCMV-Flag-Dmrtc2, cDNA fragments were amplified by PCR using human Mybl1 cDNA and mouse Dmrtc2 cDNA (from Dr. D. Zarkower) as templates and subcloned into p3XFLAG-CMV expression vector (Sigma). Wild-type and AS mouse Cdk2 were cloned into pCMV vector, and pCMV-cyclin E was provided by Dr. B. Clurman.
Protein lysates from spleens collected from wild-type mice were incubated with 20 μM CVT-313 (Santa Cruz Biotechnology) for 30 min at room temperature. Subsequently, lysates were used for immunoprecipitation in the presence of 20 μM CVT-313 inhibitor. Washing buffer also contained 20 μM CVT-313.
GST-Lin9, GST-Lin37, GST-Lin52, GST-Lin54, GST-Mybl1 and GST-Dmrtc2, were constructed as described above. For kinase assays, GST proteins (1 μg) or histone H1 (1 μg; Roche) were incubated with recombinant cyclin E-CDK2 (1 μg; Millipore) at 30°C for 30 min in 20 mM Tris-HCl pH 8.0, 1 mM EGTA, 10 mM MgCl2, 1 mM dithiothreitol, 25 μM cold ATP and 10 μCi [γ32P]-ATP. For inhibition of Cdk2 kinase activity, 20 μM CVT-313 was added to both reaction and washing buffer.
To generate an analog-sensitive (AS) version of Cdk2, we introduced a mutation in mouse Cdk2 cDNA that changes phenylalanine 80 to a glycine as described previously [47]. Human embryonic kidney 293T cells in 6-well plates were co-transfected using lipofectamine 2000 (Invitrogen) with plasmids encoding Flag-tagged Lin37, Mybl1 or Dmrtc2 and AS Cdk2 or wild-type Cdk2 with cyclin E1. After two days, cells were washed with PBS and incubated in the wells for 20 min at room temperature with 200 μl of a kinase reaction buffer [20 mM HEPES pH 7.5, 100 mM KOAc, 5 mM NaOAc, 2 mM MgOAc2, 1 mM EGTA, 10 mM MgCl2, 0.5 mM DTT, 30 μg/ml digitonin, 5 mM GTP, 0.1 mM ATP, 0.1 mM N6-(phenethyl) ATPγS (Biolog), 1X phosphatase inhibitor cocktail I and II (Sigma), and 1X complete protease inhibitors, EDTA-Free (Roche)], as described previously [26]. After the labeling step, 200 μl of 2x RIPA buffer (100 mM Tris pH 8.0, 300 mM NaCl, 2% NP-40, 0.2% SDS, 20 mM EDTA) with 2.5 mM p-nitrobenzyl mesylate (PNBM; Abcam) was added, and samples were incubated for 30 min at room temperature (RT). Mybl1 and Dmrtc2 were immunoprecipitated using anti-Flag antibody-coupled resin (Sigma) and the phosphorylation was detected by western blotting with an anti-thiophosphate ester antibody (Abcam).
Testes dissected from wild-type or Cdk2-/- mice were homogenized in Trizol (Invitrogen), and total RNA was extracted. Reverse transcription and qPCR were performed using SuperScript III SYBR Green One-Step kit (with ROX; Invitrogen). Data were normalized to Gapdh levels, and calculations were made based on the ΔΔCT method.
All cells were cultured in DMEM with 10% fetal bovine serum (Sigma). Early passage wild type MEFs and TKO (Rb-/-p107-/-p130-/-) MEFs [48] (a gift from H. te Riele) were grown until 80% confluent, harvested and analyzed by immunoprecipitation and immunoblotting.
Protein lysates were separated by size exclusion chromatography using a Superdex 200 10/300 GL (GE Healthcare). Approximately 250 μL of samples were loaded onto the Superdex size exclusion column in buffer (50 mM Tris-HCl, pH 8.0).
All statistical analyses were performed by using R or MATLAB.
|
10.1371/journal.ppat.1000300 | Species-Specific Activity of HIV-1 Vpu and Positive Selection of Tetherin Transmembrane Domain Variants | Tetherin/BST-2/CD317 is a recently identified antiviral protein that blocks the release of nascent retrovirus, and other virus, particles from infected cells. An HIV-1 accessory protein, Vpu, acts as an antagonist of tetherin. Here, we show that positive selection is evident in primate tetherin sequences and that HIV-1 Vpu appears to have specifically adapted to antagonize variants of tetherin found in humans and chimpanzees. Tetherin variants found in rhesus macaques (rh), African green monkeys (agm) and mice were able to inhibit HIV-1 particle release, but were resistant to antagonism by HIV-1 Vpu. Notably, reciprocal exchange of transmembrane domains between human and monkey tetherins conferred sensitivity and resistance to Vpu, identifying this protein domain as a critical determinant of Vpu function. Indeed, differences between hu-tetherin and rh-tetherin at several positions in the transmembrane domain affected sensitivity to antagonism by Vpu. Two alterations in the hu-tetherin transmembrane domain, that correspond to differences found in rh- and agm-tetherin proteins, were sufficient to render hu-tetherin completely resistant to HIV-1 Vpu. Interestingly, transmembrane and cytoplasmic domain sequences in primate tetherins exhibit variation at numerous codons that is likely the result of positive selection, and some of these changes coincide with determinants of HIV-1 Vpu sensitivity. Overall, these data indicate that tetherin could impose a barrier to viral zoonosis as a consequence of positive selection that has been driven by ancient viral antagonists, and that the HIV-1 Vpu protein has specialized to target the transmembrane domains found in human/chimpanzee tetherin proteins.
| Tetherin is a cell surface protein that acts as an antiviral defense. It functions by tethering newly assembled HIV-1 particles to the surface of the infected cell, such that the viral particle is unable to depart and disseminate to other, uninfected cells. HIV-1 possesses an antagonist of tetherin, termed Vpu, that abolishes tetherin function. We found that HIV-1 is an effective antagonist of human and chimpanzee variants of tetherin but is unable to antagonize tetherins from two monkey species. Additionally, we found that sequence differences in a portion of the protein that is embedded in cell membranes determined whether or not it could be antagonized by Vpu. Since the Vpu protein is alsi a membrane embedded protein, this result suggests that Vpu and tetherin interact within cell membranes. We also show that tetherin has been evolving rapidly, and has likely been placed under selective pressure to change sequence. Notably, portions of tetherin that appear to have been placed under selective pressure coincide with positions that influence Vpu antagonism. Therefore, the evolutionary history of primates determines the effectiveness of HIV-1 Vpu in modern species. Thus, tetherin could impose a barrier to cross species transmission of retroviruses.
| Eukaryotic cells can constitutively or inducibly express a variety of molecules that inhibit the replication of viruses. Among these antiviral defenses are components of the type-I interferon (IFN) -induced innate immune system [1],[2]. In turn, viruses have evolved to express proteins that either limit IFN-induced gene expression or directly antagonize the function of antiviral proteins.
We and others recently identified an IFN-induced antiviral protein, termed tetherin, that functions by a novel mechanism. Specifically, tetherin blocks the release of nascent virions from HIV-1 infected cells [3]–[5]. Tetherin is an integral membrane protein with a unique topology. In particular, it encodes a transmembrane anchor towards its N-terminus, as well as a putative glycophosphatidyl-inositol lipid anchor at its C-terminus [6]. These two membrane anchors are linked by an extracellular domain that is predicted to form a coiled-coil. Ectopic expression of tetherin in cells that do not ordinarily express it results in the formation of protease-sensitive tethers that causes retention of retrovirus particles on the surface of infected cells, from where they can be internalized [4],[5],[7],[8]. This pronounced ability to retain and internalize HIV-1 particles is present constitutively in cells that normally express tetherin, but is suppressed when tetherin is depleted. Tetherin colocalizes with Gag and appears to act by inducing adherence of virion and cell membranes. Thus, virions that are retained by tetherin are fully formed and mature, and have lipid bilayers that are discontinuous with cell membranes [4],[7].
Notably, an HIV-1 accessory transmembrane protein, Vpu, acts as a viral antagonist of tetherin [4],[5]. Indeed tetherin dramatically inhibits the release of Vpu-defective HIV-1 virions, but has only modest effects on wild-type Vpu-expressing HIV-1. Moreover, Vpu colocalizes with tetherin and prevents the localization of tetherin to nascent virions, perhaps through its ability to reduce the amount of tetherin at the cell surface [4],[5]. Thus, the existence of tetherin explains the previously observed requirement for Vpu during HIV-1 particle release from certain cells, particularly those that have been exposed to type-I IFN [3], [7], [9]–[12].
The wide expression of tetherin upon exposure of cells to IFN-alpha [4],[13] and the wide range of retroviruses and filoviruses that are inhibited by tetherin [8] suggests that it might be a general component of an innate immune defense against many enveloped viruses. As such, tetherin could provide an impetus for the evolution of antagonists in viruses other than HIV-1. Indeed, the Kaposi's sarcoma herpesvirus (KSHV) also encodes a likely antagonist of tetherin, since expression of the KSHV K5 protein decreases the steady state level of tetherin protein [14]. Additionally, certain retroviral envelope proteins, in particular the HIV-2 Env, have Vpu-like activity [15],[16]. Thus, it seems likely that tetherin antagonists other than Vpu exist.
Here, we show that tetherin proteins from different species exhibit marked differences in sensitivity to antagonism by HIV-1 Vpu. Specifically, tetherin proteins from two old world monkey species as well as from mouse, are effective inhibitors of HIV-1 particle release, but are resistant to Vpu. Moreover, we show that the transmembrane domain of tetherin contains determinants of sensitivity to Vpu, and that two mutations in the human tetherin sequence are sufficient to generate a protein that is entirely resistant to antagonism by Vpu. Interestingly, tetherin sequences that are predicted to be accessible to cytoplasmic or integral membrane antagonists, such as HIV-1 Vpu, exhibit evidence of positive selection at numerous codons, including some that we demonstrate determine the effectiveness of Vpu antagonism. Thus, past selective pressures imposed on tetherin by viral antagonists likely provides a barrier to the establishment of zoonotic infections by modern primate lentiviruses, and potentially enveloped viruses from other species, whose spread depends on antagonism of tetherin function.
Inspection of sequence databases revealed the presence of putative tetherin proteins in various mammalian species. For functional analyses, we amplified tetherin sequences from cDNAs derived from rhesus macaque (rh), African green monkey (agm), chimpanzee (cpz) and mouse (mo). Coexpression of hu-tetherin, cpz-tetherin, rh-tetherin, agm-tetherin, mo-tetherin, with HIV-1 (delVpu) caused a marked decease in the yield of viral particles, as measured using infectivity or western blot assays (Fig. 1 A,B,C). The magnitude of the reduction in virus yield varied somewhat, depending which tetherin protein was coexpressed. In each case, inhibition of virion release by tetherin did not lead to a dramatic accumulation of cell associated viral proteins. This finding suggest that the virions that are retained by tetherin, are destroyed at rate that exceeds, or is not greatly different, to their synthesis. This would likely be through endocytosis, followed by lysosomal degradation. Moreover, it may simply be the case that only a fraction of the viral protein that is synthesized by infected cells is actually released as particles. Thus the amount of viral protein that is observed in cell lysates would be determined by its intrinsic turnover rate, rather than particle release versus retention.
The differences in potency that were observed among the various tetherin proteins were only partly explained by variation in the levels at which each tetherin was expressed (Fig. 1, Fig. S1A), A finding which suggests that natural variation in potency may exist among mammalian tetherin proteins, and that hu-tetherin is particularly potent. However, a caveat to be attached to this conclusion is that transiently expressed tetherin exisits as a variety of species, presumably reflecting heterogeneous glycosylation. At present it is not clear whether all of the various tetherin species are active, and it is possible that the amount of active tetherin is not uniformly represented in analyses of total tetherin protein levels. Nonetheless, each tetherin protein was clearly capable of reducing infectious HIV-1(delVpu) yield, by 20-fold or more. Strikingly, however, and in contrast to hu-tetherin, the rh-, agm- and mo-tetherin proteins were equally efficient in reducing HIV-1(WT) and HIV-1(delVpu) virion yield (Fig. 1A,B,C). Conversely, as with hu-tetherin, HIV-1(WT) was substantially resistant to inhibition by cpz-tetherin, while the release of HIV-1(delVpu) particles was strongly inhbited (Fig. 1 D, E). Thus, the non-hominid tetherin proteins were apparently insensitive to antagonism by HIV-1 Vpu. while both hu-tetherin and cpz-tetherin were Vpu-sensitive.
Earlier work indicated that expression of an intact transmembrane (TM) segment of HIV-1 Vpu was most important, and perhaps sufficient, to enhance HIV-1 particle release [17]. This finding, combined with the notion that only the N-terminal portion of tetherin should be physically accessible to Vpu, suggested the possibility that Vpu might target the transmembrane domain of tetherin. Thus, we generated chimeric proteins, termed hu(agmTM) and hu(rhTM), in which the transmembrane domain of hu-tetherin was replaced with corresponding sequences from rh- or agm-tetherin. These chimeric proteins differed in the magnitude with which they inhibited HIV-1 virion release, concordant with differences in expression level (Fig. S1B), but both were entirely resistant to antagonism by Vpu (Fig. 2A, C). Indeed, the hu-tetherin protein containing the agm TM domain inhibited both HIV-1(delVpu) and HIV-1(WT) particle release with similar or greater potency with which the intact hu-tetherin protein selectively blocked HIV-1(delVpu) release (Fig 2A, C).
In a reciprocal experiment, rh-tetherin and agm-tetherin proteins encoding a hu-tetherin transmembrane domain were generated. The rh(huTM) protein was expressed at a slightly higher level and was a more effective inhibitor of HIV-1 particle release than was the agm(huTM) protein (Fig. 2B,C, Fig. S1B). However, both proteins selectively inhibited HIV-1(delVpu) particle release, indicating that they were sensitive to antagonism by HIV-1 Vpu. Thus, the exchange of the TM domain between tetherin proteins from different species transferred sensitivity and resistance to antagonism by HIV-1 Vpu.
Inspection of tetherin TM domain protein sequences revealed several differences between the monkey and human proteins, distributed along the length of the TM domain (Fig. 3A). To determine which of these were responsible for the Vpu resistance of the monkey tetherin proteins to antagonism by Vpu, we generated mutant forms of hu-tetherin, each bearing an individual change in the TM domain that is found at the corresponding position in rh-tetherin. (The exceptions to this scheme were the L23V,L24I mutant in which two contiguous amino acids were changed to their rh-tetherin counterparts and the delGI change, where a two amino acid deletion that is present in the rh-tetherin sequence was introduced). This panel of mutant tetherin proteins varied in the potency with which they reduced HIV-1(delVpu) virion yield, and in their ability to be antagonized by Vpu (Fig. 3B,C). Variation in the potency of antiviral activity among the mutant panel correlated with expression level in most cases (Fig. S1C). Notably, none of the individual hu-tetherin mutants recapitulated the phenotype of the rh-tetherin or hu(rhTM)-tetherin proteins that appeared completely resistant to Vpu antagonism. Rather, several of the individual mutants appeared partly resistant to antagonism by Vpu, in that Vpu was not able enhance virion release as effectively in their presence as it did in the presence of unmanipulated hu-tetherin (Fig. 3B,C). Because these analyses were slightly confounded by the variation in the potency and expression of the individual tetherin mutants, we measured infectious HIV-1(WT) and HIV-1(delVpu) virion yield in the presence of varying levels of the mutant tetherin proteins (Fig. 4). Overall, these analyses identified a single amino acid difference (P40L) as contributing substantially to the Vpu sensitivity of hu-tetherin and the Vpu-resistance of rh-tetherin. Other changes in the hu-tetherin TM had more modest effects, or no effect on Vpu sensitivity, when present in isolation, or substantially affected tetherin potency (Fig. 3, Fig. 4).
Because our initial analysis revealed that no single change in the TM domain of hu-tetherin could abolish sensitivity to Vpu, we next tested whether mutations that individually had minor or partial effects on Vpu antagonism could exert more dramatic effects when present in combination. In particular, several mutations were combined with the most obvious difference between human and monkey tetherin TM domains, namely the delGI change. Notably, the delGI,T45I double mutant strongly inhibited HIV-1 particle release and was completely resistant to antagonism by Vpu (Fig. 5A,B). Additionally, the I33V,I36L combination mutation which would be predicted to target proximal residues on the face of a TM alpha helix (Fig. 3A) appeared to confer at least partial resistance to antagonism by Vpu. However, this double mutation generated a tetherin protein with only modest activity. Nonetheless, when combined with a mutation at a third proximal residue (generating V30G,I33V,I36L) this combined mutation conferred partial resistance to Vpu antagonism, in the context of a protein with potent antiviral activity (Fig. 5A, B). Moreover a hu-tetherin bearing combined delGI,I33V,I36V mutations was almost completely resistant to antagonism by Vpu, and exhibited substantial antiviral activity (Fig. 5A, B). Finally, combining the delGI mutation with subsitiutions at contiguous residues that corresponded to rh-tetherin residues (delGI, L23V,L24I) resulted in a protein with only weak inhibitory activity, whereas combining the delGI mutation with contiguous L23A,L24V mutations (corresponding to agm-tetherin residues) generated a protein that was potent and partly Vpu resistant (Fig. 5A, B). Notably, the differences in activity and Vpu sensitivity among the various combination-mutant tetherin proteins was not explained by differences in expression level (Fig. S1D).
To determine whether mutations that conferred resistance to antagonism by Vpu in virion release assays also conferred resistance to the previously described phenomenon of Vpu-induced downregulation of hu-tetherin from the cell surface [5], we generated cell lines stably expressing either wild type hu-tetherin-HA protein or the Vpu-resistant delGI/T45I hu-tetherin mutant (Fig. 6). Upon infection with HIV-1 (WT), hu-tetherin was efficiently depleted from the surface of the vast majority of infected cells (Fig. 6A, B). Conversely, infection with HIV-1 (delVpu) resulted in little or no hu-tetherin downregulation from the surface of infected cells. Strikingly, and unlike the WT hu-tetherin protein, the delGI/T45I mutant hu-tetherin was not removed from the cell surface upon infection with HIV-1 (WT) (Fig. 6A, B) and, thus, was resistant to surface downregulation by Vpu.
Overall, these experiments revealed that no single difference between the hu-tetherin and rh-tetherin proteins accounted for their respective sensitivity and resistance to antagonism by HIV-1 Vpu. Rather, they indicated that the particular combination of residues in the tetherin TM domain can affect antiviral potency, and that multiple differences between human and monkey proteins, including the delGI indel, and the I33V, I36L, P40L and T45I differences, influence the differential sensitivity of tetherins to antagonism by Vpu.
Inspection of a larger collection of mammalian tetherin sequences amplified from various old world primates, or retrieved from sequence databases, revealed some striking features. Firstly, among the nonprimate mammalian tetherin sequences, the N-terminal cytoplasmic domain was hypervariable, both in length and sequence (data not shown). Because of these properties, it proved impossible to unambiguously align non-primate and primate tetherin sequences in order to perform tests for positive selection. Therefore, we confined further analyses to primate tetherin sequences, which could be aligned unambiguously throughout the entire length of the coding sequence. Even within primates there was considerable sequence divergence between species, ranging from 0.5% to 40.0% at the nucleotide level. Sampling within three old world monkey species (rhesus macaques, pig-tailed macaques and sooty mangabeys) also revealed the presence of significant polymorphism within species (Fig. S2 and data not shown); it remains to be seen whether nonsynonymous polymorphism extends to the tetherin locus of other primate lineages. Positive selection was tested using the REL(HyPhy) [18] and CODEML (PAML) [19] methods and these analyses revealed that codons exhibiting high dN/dS ratios, and therefore likely to have been subjected to positive selection, were enriched in the N-terminal cytoplasmic and TM domains in primate tetherins (Fig. 7). Tetherin evolution in primates was also evaluated under several standard models of sequence evolution as implemented in the CODEML program. These comprise three nested pairs of models (M0 and M3; M1a and M2a; M7 and M8) in which the second model of each pair is derived from the first by allowing sites to evolve under positive selection. Nested models were compared using the likelihood ratio test, and in each case allowing individual sites to evolve under positive selection (M3, M2a, M8) gave a significantly better fit to the primate sequence data than the corresponding model without positive selection (M0, M1a and M7, respectively) (Table 1). The M3, M2a and M8 models identified a largely overlapping set of sites in the tetherin coding sequence with dN/dS>1, consistent with an evolutionary history characterized by frequent episodes of positive selection. Notably, some codons that exhibited a high probability of having evolved under positive selection coincided with residues that determined the effectiveness of Vpu antagonism (Fig. 7). However, there were numerous additional codons, particularly in the tetherin cytoplasmic domain, that also exhibited high dN/dS ratios, suggesting that antagonists other than Vpu have also imposed selective pressure on primate tetherin sequences.
Previously, we reported that Vpu could reverse the inhibitory effect of IFN-alpha on HIV-1 particle release from human cells, but that Vpu failed to reverse such IFN-alpha induced inhibition in African green monkey cells [3]. The subsequent discovery that tetherin is an IFN-induced inhibitor that is antagonized by Vpu [4],[5] leads to the prediction that agm-tetherin should be resistant to Vpu. Here we show that agm-tetherin, as well as rh-tetherin and mo-tetherin are indeed resistant to antagonism by Vpu, in contrast to tetherin variants found in species (human and chimpanzee) that are permissive hosts for HIV-1. Moreover, these studies identify the TM domain of hu-tetherin as a major determinant of the effectiveness of Vpu antagonism. That the TM domain of tetherin harbors critical determinants of Vpu sensitivity is concordant with previous observations indicating that the TM domain of Vpu is critical for its virus release activity [17]. Thus, these observations suggest a model in which Vpu and tetherin interact via their TM domains. Such an interaction could be direct, but further work will be required to resolve whether this is indeed the case, and precisely how Vpu tetherin antagonism and/or downregulation from the cell surface is achieved is not known. In this regard, the recent report suggests that the host cell protein CAML is important for Vpu activity [20], but how tetherin, Vpu and CAML interact in a functional sense is unclear at present.
Importantly, no single change in the hu-tetherin TM domain abolished Vpu sensitivity, which is consistent with a model in which Vpu, or a bridging factor, makes multiple contacts with the TM domain of tetherin. Indeed, the hu-tetherin mutant, delGI,T45I, that had the most striking phenotype, in that it retained full activity but was completely resistant to antagonism by Vpu, harbored mutations at positions close to the opposing ends of the TM domain. Additionally, a different combination of mutations (delGI, I33V, I36L) also conferred complete Vpu resistance, again consistent with the notion that multiple contacts with the tetherin TM domain are made during its antagonism by Vpu. Concordant with these findings, the chimpanzee and human tetherin proteins were both sensitive to HIV-1 Vpu, and differed from each other at only a single position in their TM domains. Thus, it is likely only minor, if any, adaptation in the SIVCPZ Vpu protein that is immediately ancestral to the HIV-1 Vpu proteins would have been required in order for it to target human tetherin.
Primate tetherin TM domain sequences, that should only be accessible to integral membrane antagonists, exhibit clear evidence of positive selection at several codons, including some that determine the effectiveness of Vpu antagonism. Thus, it is likely that antagonists encoded by pathogenic viruses have driven the selection of the tetherin variants that exist in modern primates. Such antagonists could include Vpu itself, since several primate lentiviruses encode Vpu proteins (http://www.hiv.lanl.gov/). However, other viral antagonists, including the KSHV K5 protein [14], or homolgues of it, are also reasonable candidates for factors that have imposed selective pressure on tetherin sequences. In addition to the TM, several sites with the highest dN/dS ratios mapped to the N-terminal cytoplasmic domain. Such sites may define an accessible target exploited by other virally encoded inhibitors of tetherin. It is noteworthy in this regard that many primate lentiviruses do not encode a Vpu like protein, and may have evolved alternative strategies targeting this or other regions of the protein. Additionally, tetherin sequences might also have evolved to better target specific types of viral particles in some host species, as a consequence of varying viral challenges. Such evolution might also include adaptations in domains of tetherin that, for example, modify its trafficking within cells (likely including the cytoplasmic and TM domains). Overall, there are several potential sources of evolutionary pressure that could give rise to positive selection and diversification of tetherin genes. We note that the sequence of the TM domain of tetherin is obviously constrained by the need to retain the biochemical characteristics of a TM domain, which might mitigate against the detection of positive selection in this protein domain.
HIV-1 is well adapted to replicate in human cells, but fails to replicate in many nonhuman primate cells. This is in large part because it is unable to evade or antagonize the species-specific variants of antiviral genes, such as TRIM5 and APOBEC3, which show evidence of positive selection that is assumed to have resulted from past retroviral epidemics [21]–[23]. Tetherin represents a third example of an antiviral gene in monkeys that exhibits activity against intact HIV-1 as a consequence of positive selection in the primate lineage. Thus an array of antiviral molecules limit the replication of primate lentiviruses in non-natural host cells, creating barriers to zoonosis, and revealing potential opportunities to mobilize intrinsic antiretroviral defenses by therapeutic inhibition of the activity of their viral antagonists.
A hu-tetherin cDNA, cloned into pCR3.1 vector, and its N-terminally HA tagged counterpart have been described previously [4]. An internally HA-tagged tetherin expression construct, pCR3.1/hu-tetherin-HA was derived from this by inserting an NheI restriction site at nucleotide position 463 of the tetherin gene. Thereafter, complimentary oligonucleotides encoding an HA epitope tag were inserted into the NheI site. Similarly, the tetherin coding sequence from rhesus macaque, African green monkey chimpanzee and mouse was amplified using cDNA generated from IFN-alpha treated 221, COS-7, chimpanzee fibroblasts, and NIH3T3 cells respectively. For the monkey tetherins, the HA epitope was inserted at a position orthologous to nucleotide 463 as described above, while a N-terminally tagged murine tetherin construct was used. Thereafter, overlap-extension PCR approaches were used to exchange TM domain segments between human and monkey proteins, or to introduce point mutations into the hu-tetherin sequence.
293T cells were maintained in DMEM media supplemented with 10% fetal calf serum and gentamycin, as were HeLa-TZM cells which express CD4 and CCR5 and contain a lacZ reporter gene under the control of an HIV-1 LTR. To measure tetherin and Vpu activity, 293T cells were seeded in a 24 well plate at a concentration of 1.5×105 cells/well and transfected the following day using polyethylenimine (PolySciences) with 500 ng of an unmanipulated HIV-1 proviral plasmid NL4-3(WT) or a Vpu-defective counterpart NL4-3(delVpu). Additionally, 50 ng of a tetherin expression plasmid and 50 ng pCR3.1/cherry fluorescent protein (to monitor transfection efficiency) were included in the transfection. In experiments where the level of tetherin was varied, the tetherin expression plasmids were serially diluted from 200 ng to 12.5 ng per transfection and pCR3.1 was used as a DNA filler. Stable WT and mutant hu-tetherin-HA expressing 293T-derived cell lines were generated by retroviral transduction, as previously described [8].
Transfected 293T cells were place in fresh medium at 20 hrs post transfection and virion containing cell supernatants were harvested and filtered (0.2 µm) at 40 hrs post transfection. Infectious virus release was determined by inoculating, in triplicate, sub-confluent monolayers of HeLa-TZM cells seeded in 48 well plates at 2.5×104 cells/well with 50 µl of serially diluted supernatants. At 48 hrs post infection, ß-galactosidase activity was determined using GalactoStar reagent as per the manufacturer's instructions. The remainder of the virion containing supernatant (450 µl) was layered onto 800 µl of 20% sucrose in PBS and centrifuged at 20,000 g for 90 minutes at 4°C and virion yield determined by western blot assays
Pelleted virions and the corresponding cell lysates were resuspended in SDS-PAGE loading buffer and separated on NuPAGE Novex 4–12% Bis-Tris Mini Gels (Invitrogen). Proteins were blotted onto nitrocellulose membranes. Thereafter, HIV-1 Gag or capsid proteins, as well as tagged tetherin proteins were revealed using anti-capsid and anti-HA antibodies and chemiluminescent detection reagents, as described previously.
293T cells stably expressing either WT or mutant (delGI/T45I) tetherin-HA were plated on poly-D-lysine coated dishes (Mattek). The following day the cells we infected with VSV-G pseudotyped HIV-1(WT) or HIV-1(delVpu) variants that carried Cerulean-FP (CFP) embedded in the stalk region of the matrix domain of Gag. The virus dose was chosen so that approximately 40% of the cells were infected. At 48 h after infection, cells were fixed but not permeabilized in order to confine tetherin-HA staining to surface expressed protein. Fixed cells were sequentially incubated with an anti-HA monoclonal antibody (Covance) followed by an anti mouse IgG Alexafluor 594 conjugate. The cells were imaged using a Deltavision microscopy suite and infected cells (identified by the presence of CFP fluorescence) were scored for the presence of intense tetherin-HA staining on the cell surface.
The aforementioned human, chimpanzee, rhesus monkey, African green monkey and mouse sequences were included in an analysis for positive selection. In addition, tetherin genes were amplified from lymphocyte RNA from several rhesus macaques (n = 8) pigtail macaques (n = 1), crab eating macaques (n = 6) and sooty mangabeys (n = 6). These were cloned using a kit TOPO-TA kit (Invitrogen) and the sequences of multiple clones determined. Representative alleles were included in the analysis described below. Additional tetherin sequences from gorilla, gibbon, and marmoset were retrieved from the raw data archives of ongoing genome sequencing projects by TraceBLAST (http://www.ncbi.nlm.nih.gov/BLAST/Blast.cgi), using each of the hu-tetherin coding exons as a separate query. Sequences were aligned using Macvector, and adjusted manually. Codon-based nucleotide alignments were used in conjunction with phylogenetic trees generated using the DNAPARS program (PHYLIP) as input for the random effects likelihood (REL) program (HyPhy) to detect positive selection. Input files for analysis using CODEML in the PAML suite (version 3.14) were generated by first aligning amino-acid sequences using the CLUSTAL-W algorithm, converting the alignment back to nucleotides, and adjusting manually where necessary using MEGALIGN (DNASTAR, Madison, WI). Tree files were generated by Neighbor-Joining, and sites with dN/dS>1 were identified using the resulting tree or a tree constrained to accept the known major branches of primate evolution as input, with similar results. The F3X4 model of codon frequencies was used for all analyses in CODEML. Paired, nested models of sequence evolution implemented in CODEML (M0, M3; M1, M2; M7, M8) were also compared using the likelihood ratio test. Evaluation with the chi-square test assumed either 4 degrees of freedom (M0, M3) or 2 degrees of freedom (M1,M2; M7, M8).
|
10.1371/journal.pcbi.0030233 | Small Regulatory RNAs May Sharpen Spatial Expression Patterns | The precise establishment of gene expression patterns is a crucial step in development. Formation of a sharp boundary between high and low spatial expression domains requires a genetic mechanism that exhibits sensitivity, yet is robust to fluctuations, a demand that may not be easily achieved by morphogens alone. Recently, it has been demonstrated that small RNAs (and, in particular, microRNAs) play many roles in embryonic development. Whereas some RNAs are essential for embryogenesis, others are limited to fine-tuning a predetermined gene expression pattern. Here, we explore the possibility that small RNAs participate in sharpening a gene expression profile that was crudely established by a morphogen. To this end, we study a model in which small RNAs interact with a target gene and diffusively move from cell to cell. Though diffusion generally smoothens spatial expression patterns, we find that intercellular mobility of small RNAs is actually critical in sharpening the interface between target expression domains in a robust manner. This sharpening occurs as small RNAs diffuse into regions of low mRNA expression and eliminate target molecules therein, but cannot affect regions of high mRNA levels. We discuss the applicability of our results, as examples, to the case of leaf polarity establishment in maize and Hox patterning in the early Drosophila embryo. Our findings point out the functional significance of some mechanistic properties, such as mobility of small RNAs and the irreversibility of their interactions. These properties are yet to be established directly for most classes of small RNAs. An indirect yet simple experimental test of the proposed mechanism is suggested in some detail.
| Early embryonic development depends on robust patterning along the axes of the embryo. At the cellular level, neighboring segments are often identified via the concentrations of several gene products: the expression of such a gene may, for example, be high in the cells of one segment, and negligible in those of another. Recently, it has been suggested that small RNA molecules, such as microRNAs, may play a role in establishing a sharp boundary between two neighboring segments, but are not required for the overall patterning. Here, we investigate this possibility using a mathematical model, which assumes that small RNAs diffuse in the tissue. Surprisingly, we find that mobility of the small RNAs may generate a sharp interface in the expression profile of its target gene. We analyze the properties of the interaction between the two molecules that are required to achieve this function. An experimentally testable prediction is detailed, and two possible realizations in the fruit fly and in maize are discussed.
| Morphogenesis proceeds by sequential divisions of a developing embryo into domains, each expressing a distinct set of genes. Each combination of genes is associated with a particular cell identity. At advanced stages of development, most genes that define cell identity are either highly expressed (“on”) or strongly inhibited (“off”) in a given cell. For example, two adjacent domains may be differentiated by high expression of some genes in one, and low expression in the other. In such cases, it is important that cells of the two populations do not intermix. Moreover, the number of cells that show intermediate levels of expression, typically found at the interface between the two sets, should be kept to a minimum. These demands are necessary in order to unambiguously define the identity of each cell. A spatial gene expression pattern that obeys these demands is said to exhibit a sharp interface.
A crucial step in setting the interfaces of gene expression patterns is often the establishment of a concentration gradient of molecules called morphogens. Some morphogens are transcription factors that regulate gene expression directly [1,2]. Others are ligands that bind cell-surface receptors signaling the activation of target expression [3]. Since morphogens act in a concentration-dependent manner, a morphogen gradient may be transformed into a gradient of its target messenger RNA (mRNA).
In principle, a single morphogen interacting cooperatively with its target enhancer can generate a sharp interface in the target transcription profile, by modulating the rate of its mRNA transcription as a function of the nuclear spatial coordinate [4]. This may be done, e.g., by cooperative binding to a receptor or to a promoter [5] or by zero-order ultrasensitivity [6]. As an example, in Drosophila early embryonic development, Hunchback transcription depends on the cooperative binding of about five Bicoid molecules [7]. An obvious limitation in this mechanism is the need for large cooperativity factors or cascades of reactions, which make it prone to fluctuations and slow to adapt [7–10]. Recently, a role for small regulatory RNAs in establishing developmental patterning has been documented in plants [11–13] and animals [14]. In particular, it has been suggested that microRNAs (miRNAs) confer accuracy to developmental gene expression programs [15]. This raises the possibility that small RNAs aid morphogen gradients in establishing sharp interfaces between “on” and “off” target-gene expression.
In this study, we formulate a mathematical model in which small regulatory RNAs help morphogens to determine cell identity by sharpening morphogen-induced expression patterns. For specificity, we assume here that the small RNA belongs to the miRNA family, and consider another class of small RNA in the Discussion. miRNAs constitute a major class of gene regulators that silence their targets by binding to target mRNAs. In metazoans, primary miRNA transcripts are transcribed and then processed both inside and outside of the nucleus to form mature transcripts approximately 21 nucleotides (nt) in length that are then loaded into the RNA-induced silencing complex (RISC) [16]. They are found in plants [17] and animals [18], including human [19], and are predicted to target a large fraction of all animal protein-coding genes [19–21]. In plants, miRNAs are known to affect morphology [11,12,22], implying that they play an important role in determining cell identity. This is underscored by the fact that the spatiotemporal accumulation of miRNAs is under tight control in plants [23], fly [14,24], and zebrafish [25].
Our model is constructed in one spatial dimension, namely along one spatial axis. Domains of gene expression are laid out along this axis, and we assume no significant variance along other, perpendicular axes. Two key ingredients of the model are a strong interaction between miRNA and mRNA, and intercellular mobility of the miRNA. Within this framework, miRNAs generate a sharp interface between those cells expressing high levels of the target mRNA and those expressing negligible levels of mRNA. We use physical arguments to understand the range of parameters in which this sharpening occurs. Our model predicts that the spatial position of the interface is precisely determined: mobile miRNAs spatially average individual cellular fluctuations without compromising the interface sharpness. We use computer simulations to show that this is also true even with low numbers of molecules. A consequence of our model is that a local change to the transcription profiles can induce a nonlocal effect on the mRNA concentration profile; we outline an experiment to detect this nonlocal property. Finally, we consider possible applications of these ideas in plants and fruit fly.
Our theory comprises three central elements. First, miRNAs and their targets are taken to be transcribed in a space-dependent manner. Second, we assume that the interaction between miRNA and target irreversibly disable the target mRNA from being translated into proteins; this, for example, may be done by promoting the degradation of the target. Furthermore, the miRNA molecule itself may be consumed during this interaction. Last, we allow for the possibility that miRNAs move between cells. Before defining the model, let us review the available data regarding each of these processes.
miRNAs and their targets are often expressed in a coordinated manner [26]. Often, the regulatory network is designed to express the miRNA and its targets in a mutually exclusive fashion. For example, the expression patterns of the miRNA miR-196 and its target Hoxb8 are largely nonoverlapping in mouse [27] and chick [28]. Similarly, the nascent transcripts of ubx (i.e., ubx transcripts still attached to the DNA) are expressed in a stripe near the center of the early embryo, whereas nascent transcripts of its regulator, iab-4, are simultaneously observed in nuclei posterior to this domain [29]. A recent large-scale study in Drosophila showed that miRNAs and their target genes are preferentially expressed in neighboring tissues [15]. Likewise, in mouse [30] and in human [31], predicted miRNA targets were found at lower levels in tissue expressing the cognate miRNA than in other tissues.
Our model assumes that the synthesis rate of the miRNA and its target are smoothly varying along a spatial axis, x. This, for example, may be the result of a common morphogen regulating (either directly or indirectly) the two species. The transcription profiles αμ(x) and αm(x) of the miRNA and its target are assumed to be largely anticorrelated.
The detailed interaction between miRNAs and their targets is currently a topic of intense investigation [32,33]. miRNAs induce the formation of a ribonucleoprotein complex (RISC). Targeting of a specific mRNA by a RISC is done via (often imperfect) base-pair complementarity to the miRNA [18]. Upon binding, protein synthesis is suppressed by either translational repression or mRNA destabilization [32,33]. Although it is likely that miRNA can go through a few cycles of mRNA binding [34], the increased endonucleolytic activity conferred by the miRNA makes it plausible that the miRNA is sometimes degraded in the process. In addition, evidence suggests that mRNAs that are translationally repressed by miRNA may be colocalized to cytoplasmic foci such as stress granules [35] or processing bodies [34–36]. Stress granules are cytoplasmic aggregates that appear under stress and sequester untranslated RNA, perhaps to protect these molecules or to regulate translation [37]. Processing bodies, which are enriched with endonucleases, are believed also to be places of mRNA degradation [38]. The two types of RNA granules are also known to interact, possibly an indication that stored RNA in stress granules may be targeted for degradation [37]. In both cases, RNA granules may sequester or degrade not only the mRNA, but also its bound miRNA. Taken together, these facts make it improbable that miRNAs act in a fully catalytic manner.
A pair of mRNA–miRNA reactions that describe a spectrum of plausible scenarios is
where m represents the mRNA concentration and μ represents that of the miRNA. Here, θ is the average number of targets degraded by a given miRNA before it is itself lost in the process. These reactions may be realized in different ways. For example, the two species may reversibly form a complex that is then subject to degradation. Another possibility is that the two species irreversibly associate to form an inert complex. Furthermore, the reaction between the species may be reversible, as long as the typical dissociation time is much longer than the relevant biological timescale. One way in which the cell may control the dissociation time is by regulating exit of the RNA pairs from processing bodies [39].
Can miRNAs move from cell to cell? Short interfering RNA (siRNAs), another important class of small RNAs, are known to elicit non–cell-autonomous RNA silencing in plants, worms, fly, and possibly mouse (reviewed in [40]). This may also be the case for trans-acting siRNA [13]. Evidence in favor of intercellular mobility of miRNA is found in pumpkin [41]. There, miRNAs have been found in the phloem sap that is transported throughout the plant by phloem tissue. In animals, many small RNAs, including many miRNAs, were found in exosomes from mouse and human mast cell lines, which can be delivered between cells [42].
In our model, we consider the possibility that miRNAs migrate from cell to cell. Mobility of the miRNA species is likely to rely on active export from the cell followed by import to neighboring cells, or perhaps on transport between neighboring cells, e.g., via gap junctions. On the tissue scale, these transport processes are expected to result in effective diffusion. We therefore ignore the small-scale transport processes, and model miRNA mobility as pure diffusion. Finally, we combine these processes into a steady-state mean-field model given by
The β terms describe independent degradation (i.e., by processes independent of the other RNA species) and the k term describes coupled degradation of both RNA species. Note that the case θ > 0 (where miRNA may go through multiple rounds of interactions with target mRNAs) can also be brought into this form by rescaling Equation 2a [43]. Mobility of the miRNA is described by an effective diffusion constant D. The spatial coordinate x measures distance along one dimension of a tissue. All our numerical results shall be presented in units of the tissue size, i.e., 0 ≤ x ≤ 1.
Equation 2a and Equation 2b cannot be solved analytically. In what follows, we solve these equations numerically, imposing zero-flux boundary conditions. These exact numerical solutions can be used to draw the steady-state expression profiles of both RNA species for a particular set of parameters. To gain further insight, we also develop an approximate analytical solution.
As described above, a desired target protein profile comprises a domain of cells that express this protein abundantly, adjacent to a domain of cells where this protein does not accumulate. Furthermore, one requires that the number of cells with intermediate expression levels lying in between the two domains be minimized—this is our definition of a sharp interface. In this section, we discuss one scenario in which the mutual consumption of a diffusive miRNA and its target leads to such a sharp interface in the mRNA profile.
We assume that some morphogen controls the transcription rate of the target. The transcription profile—namely, the transcription rate as a function of the spatial coordinate of the nucleus—is laid down as a smooth gradient, falling from one end of the developing tissue (which, for convenience, will be called “left”) toward the other end (“right”). Motivated by a recent study that showed that miRNA and their targets are preferentially expressed in neighboring tissues [15], we focus on the scenario in which the miRNA transcription is controlled in a fashion opposite to that of the mRNA: miRNA transcription is peaked at one end of the tissue (where the mRNA transcription rate is minimal), and decreases toward the other end (where the mRNA transcription rate is maximal). The kind of “mutually exclusive” transcription profiles we have in mind is depicted in Figure 1A. In this figure, and hereafter, we denote the mRNA transcription profile by αm(x) and the miRNA transcription profile by αμ(x), explicitly noting their dependence on the spatial coordinate x. We note that, in the absence of miRNA–target interaction and of miRNA diffusion, the concentration profiles of mRNA and miRNA simply follow their transcription profiles (Figure 1A and 1D). Each is rather smooth and overlaps the other near the center of the tissue.
Before studying the full model, it is instructive to consider first the interacting system in the absence of miRNA diffusion. In the context of mutually exclusive transcription profiles, we expect that each cell would be dominated by one RNA species (either the miRNA or the target mRNA), which we will call the majority species, and be depleted of the other, the minority species. In other words, we are making the critical assumption that the decay of the minority species in each cell is governed by the interaction with the other species (rather than by its independent degradation). This assumption can be made quantitative in terms of the model parameters; see Equation S1 in Text S1. Under this assumption, which we refer to as the strong interaction limit, it is straightforward to show (Equation S3 in Text S1) that the density of the majority species in each cell is proportional to the difference between the two transcription rates in that cell, whereas the minority species is essentially absent. Consequently, in the context of mutually exclusive transcription profiles, the mRNA level becomes vanishingly small in any cell for which αm(x) < αμ(x), namely every cell to the right of the point where the two transcription profiles are equal. The concentration profiles of the two RNA species are shown in Figure 1B and 1E, in which one can see that the mRNA and miRNA spatial expression domains are now complementary and more sharply defined.
The threshold response that arises when both RNA species do not diffuse from cell to cell provides insurance against the possibility that the mRNA transcription profile is not as step-like as is required for unambiguous cell differentiation. In other words, miRNA regulation acts as a failsafe mechanism whereby incorrectly transcribed low-abundance transcripts in the region αm(x) < αμ(x) are silenced, while correctly transcribed high-abundance transcripts in the region αm(x) > αμ(x) are only mildly affected [15,26]. This threshold response in the target profile has been observed in the context of small RNAs—another class of posttranscriptional regulators—in bacteria [43].
We now return to our full model, which allows for diffusion of the miRNA. To simplify the analysis, let us keep the strong-interaction limit, described above (and in Equation S1 in Text S1). In general, one expects that diffusion makes the miRNA profile more homogeneous, and this is confirmed by exact numerical solution of the model, as shown in Figure 1C. Surprisingly, however, the mRNA profile does not become smoother. In fact, Figures 1C and 1F show that this profile actually develops a sharper drop from high to low mRNA levels than there was in the absence of diffusion. More specifically, miRNA diffusion creates an interface between high and negligible target expression. Increasing diffusion moves the interface deeper into the mRNA-rich region and thereby accentuates the drop in mRNA levels across the interface. Although some miRNA diffusion is required to establish a sharp interface in the mRNA profile, the diffusion constant cannot be too large. As Figure 1G demonstrates, increasing the diffusion constant may result in smoothing the interface. A corresponding increase in the interaction strength, k, can compensate for the increased diffusion, regaining the interface sharpness (Figure 1H). We will quantify these observations below.
Diffusing miRNAs can find themselves in one of two very different regions. In the miRNA-rich region (including the region to the right of the point where the transcription profiles are equal), miRNA decay occurs mainly via processes independent of their interaction with the target. In this region, our model boils down to a simple diffusion process accompanied by linear decay. Such processes are characterized by a length scale, denoted by λ, which essentially measures how far a miRNA can travel (due to diffusion) before being consumed (by independent degradation). It is thus an increasing function of the diffusion constant D, but a decreasing function of the independent decay rate βμ. On the other hand, in the mRNA-rich region, a miRNA decays mainly via co-degradation with its target. In this region, miRNAs decay faster, and one expects them to be able to travel over much shorter distances than in the miRNA-rich region. In fact, diffusion in this region is characterized by another, smaller, length scale, denoted by ℓ, which again increases with D, but is now a decreasing function of the interaction strength, k. Explicit expressions for the two length scales are given in Text S1 (Equations S5 and S6 in Text S1).
To obtain a sharp interface in the mRNA profile, miRNAs should be able to travel from the miRNA-rich zone into the mRNA-rich zone. This means that the first length scale, λ, should be of the same order as the tissue length. This, for example, can be achieved if the diffusion constant D is large enough. On the other hand, the vicinity of the interface is governed by the other length scale, ℓ. This length scale is what determines the “width” of the interface, namely the number of cells that exhibit intermediate levels of mRNA expression (see blue box in Figure 1C). A sharp interface, therefore, means a small value of ℓ, and one way to achieve a small value of ℓ is to make the diffusion constant D small enough. These two contradicting requirements on D suggest that there might be an intermediate range of values for the diffusion constant that allows for a sharp interface, but also raises the suspicion that this range may be very small and requires some fine-tuning. This, however, is not the case: the fact that λ is strongly dependent on βμ (whereas ℓ does not depend on βμ at all), and that ℓ strongly depends on k (whereas λ does not) means that the range of allowed values of D can be set as large as needed.
In Text S1, we develop an approximate analytical expression for the mRNA profile in terms of the various parameters and the “input” profiles αm(x) and αμ(x) (Figure S1). There are two lessons to be learned from this exercise. First, the interface established by the mRNA–miRNA interaction is effectively impermeable to miRNA diffusion in the strong-interaction limit. The system thus separates into two parts which—in steady state—do not exchange molecules between them. This property allows one to calculate the position of the sharp interface in the mRNA profile.
The second lesson comes from the resulting equation for the interface position. This equation takes the form of a weighted spatial average of the difference between the two transcription profiles (Equation S11 in Text S1). Before interpreting the full result, it is instructive to consider the limiting case in which miRNAs cannot be degraded independently (βμ = 0). In this case, our result (Equation S12 in Text S1) implies that the interface is positioned such that total synthesis rates of mRNA and miRNA to its right are equal. Thus, it is the total production rates in that part of the system that determine the interface position, and not any particular cell by itself.
In the more general case (βμ > 0), the contribution of each cell is weighted by some nontrivial function. Still, in order to determine the interface position, one needs to perform a sum over many nuclei, each contributing the difference between the local transcription rates of the two RNA species. Clearly, these rates may be influenced by many factors, and in a description that is somewhat closer to reality, one would expect this difference to be fluctuating around αμ(x) − αm(x). However, the interface position is a sum of these fluctuating objects, and one might hope that the sum of these fluctuations—which are uncorrelated—would be close to zero. In this case, the interface position would be robust to fluctuations of this type. Indeed, a stochastic simulation of the model shows no change in the interface position (or structure), as compared with the deterministic model discussed so far (Figure S3; see Text S1 for details of our simulations).
In a multicellular tissue, mRNA are typically not expected to be transferred from cell to cell. Therefore, most of the work presented here does not consider the possibility that mRNA can also be mobile. Nevertheless, mRNA mobility should be considered in some cases. For example, the early Drosophila embryo is a syncytial blastoderm, in which nuclei multiply in a common cytoplasmic space. In the absence of cell membranes, mRNA is likely to be mobile, although probably with a small diffusion constant [44]. In Text S1, we generalize our model to include mRNA mobility. We find that a sharp interface can be achieved as long as the typical distance traveled by target mRNAs, even in the absence of miRNAs, is small compared with any other length scale (such as the interface width). Denoting the mRNA diffusion constant by Dm, this condition can be written as
. We note that this condition does not contradict any of the conditions mentioned before; see Figure S2.
In passing, let us note that the conditions required so far—namely, strong interaction between the miRNA and its target, and small ℓ—may be reached by making the mRNA completely stable (βm → 0). However, our analysis shows that in this case, the system would never relax to a steady state, since target mRNAs would accumulate at the left end of the tissue without limit. Our analysis here is, therefore, only applicable if the mRNA molecules undergo independent degradation, in addition to the miRNA-dependent degradation.
Using the insight gained in the previous section, we briefly show how a stripe is formed when the miRNA transcription profile αμ(x) is similar to αm(x) but displaced from it (Figure 2A). Suppose, for example, that the synthesis of an miRNA and its target are activated by the same transcription factor. In any given nucleus, the two promoters experience the same concentration of this transcription factor. However, they need not react in the same way: if the binding affinity of one promoter is stronger than that of the other, there will be intermediate concentrations of the transcription factor such that the first promoter will be activated while the other will not. Such a scenario is depicted in Figure 2A, where a common transcription factor, which exhibits a spatial gradient, activates the target gene as well as the miRNA gene. In this case, the target promoter has higher affinity to the transcription factor than the miRNA promoter. Thus, some cells in the middle of the developing tissue express the target mRNA, but not the miRNA.
Unlike the case studied in the previous section, in which the transcription profiles crossed at one point, here the transcription profiles cross at two points. Let us retrace our steps in the previous section by first considering the case of no diffusion. For low values of the interaction rate k, the miRNA and mRNA profiles are qualitatively similar to their transcription profiles. As k is increased however, miRNA deplete mRNA levels at any position where αμ > αm and thus confine mRNA expression to a stripe between the two crossing points of the transcription profiles (green curve in Figure 2B). Can diffusion make this profile sharper, as in the previous case? Indeed, diffusing miRNAs that survive annihilation on the left and right diffuse into the interval between the two crossing points, and establish sharp interfaces in the mRNA concentration profile. The resulting stripe resides within this interval, but is narrower (blue curve in Figure 2B). It is therefore important that parameters allow for sharp interfaces, without making the stripe too narrow (or even disappear). Therefore, to sustain a well-defined stripe of gene expression, the interface width must be much smaller than the distance between the two crossing points of the transcription profiles.
One can use the same analytic method mentioned earlier to calculate the new positions of the stripe boundaries (see Figure S4 and Text S1). This exemplifies how the method can be used to analyze geometries of increasing complexity.
The sharp interface that we predict can be detected directly in an imaging experiment, provided the light intensity varies linearly with mRNA concentration and the spatial resolution is high enough. However, experiments often do not supply quantitative data that are faithful to the underlying concentration profile. This, for example, is the case if an experimental setup is designed to identify the presence/absence of a molecular species. The application of nonlinear filters, such as photomultipliers, may result in spurious sharp boundaries. In contrast, low spatial resolution may make a sharp interface appear smoother than it really is.
Here, we address the task of making predictions that are based on quantitative analysis, yet can be tested using qualitative data. To this end, we consider a worst-case (if somewhat artificial) scenario in which the apparatus' readout is binary: concentrations below an apparatus-dependent threshold are not detected, whereas concentrations larger than this elicit a concentration-independent fluorescence intensity. In such scenarios, it is impossible to tell a smooth and sharp concentration profile apart as both yield a sharp interface in the binary readout (Figure 3A and 3B).
Fortunately, our model of miRNA-mediated morphogenic regulation possesses another signature that is visible at such coarse experimental resolution. To detect this signature, one needs to overexpress the miRNA in a small patch of cells (hereafter denoted the “patch”). Our model then predicts that this patch has a qualitatively different effect depending on which side of the interface it occurs.
The technique one uses to generate the patch may differ according to the stage of development under consideration. In the early blastoderm stages of Drosophila development, for example, a Gal4 driver may be used to drive expression of the miRNA in those cells in which an endogenous gene is expressed [45]. Many endogenous genes are expressed in stripes along the anterior–posterior axis during these stages and some have dedicated enhancers for single stripes [4,46]. As an example, the yeast FLP-FRT recombination system has been used to misexpress the gap gene knirps in a stripe by placing it under the control of the eve stripe 2 enhancer [47].
In later stages of Drosophila development, e.g., imaginal discs, one technique is the random generation of a mosaic of mutant clones (patches) by mitotic recombination [45,48]. The patches are generated at a low rate, and one then screens for those embryos containing a single patch.
We model the localized overexpression of miRNA by an effective local increase of the transcription rate (by an amount αc = 5). This increase in transcription rate occurs in a small number of cells, which in our model is about 5% of the tissue length. More specifically, we choose a position xc for the center of the patch, and for every point x that resides within a distance w/2 of xc, we change the transcription rate from αμ(x) to αμ(x) + αc. Here w is the “width” of the patch, which takes a value w ≃ 5% of the tissue length.
Consider positioning the patch first in the miRNA-rich region of the developing tissue (Figure 3C). One sees that, even if positioned at a distance from the expression domain of the target, the effect of the additional miRNA is to push the interface toward the left. The localized patch of cells therefore has a nonlocal effect.
As mentioned earlier, the position of the interface is determined by a global comparison of the mRNA and miRNA transcription rates to the right of the interface. Ectopic expression of miRNA to the right of the interface changes this balance and displaces the boundary. Note that this displacement can only be achieved if the patch is positioned to the right of the interface, since the interface position is not influenced by transcription balance to its left. This effect is quantified in Text S1.
This experiment should be contrasted with one in which the overexpressing patch is positioned in the mRNA-rich region, as shown in Figure 3D. Such ectopic miRNA expression has a local effect, with excess miRNA creating a hole in the mRNA expression domain. The hole edges constitute two additional interfaces in the system, the sharpness of each again determined by ℓ.
One can go further and make a quantitative prediction, relating the number of patches in a mosaic of patches with the lateral shift in the interface position. To first approximation, one needs to count the number of patches in the miRNA-rich region, and disregard completely the patches in the mRNA-rich region. The displacement of the interface position is then linearly proportional to this number; see Text S1 for details. Simulated experimental results that would verify this prediction are shown in Figure 4.
The distinct nonlocal effect described above does not occur when the miRNA are unable to move between cells. Also, we have checked (for the parameters used in this study) that it does not occur when the miRNA acts purely catalytically (Figure S5). Rather, both miRNA mobility and a strong interaction between miRNA and target are required. The presence or absence of the nonlocal effect would therefore confirm or falsify the hypotheses that miRNA are mobile and that they interact stoichiometrically with mRNA while in this mobile state.
In this study, we have analyzed a model in which miRNAs sharpen target-gene expression patterns by generating an interface between high and low target expression. This effect is due only to the strong noncatalytic interaction between the miRNA and its target, and requires no additional interactions or feedbacks.
A necessary condition for a sharp interface to occur is that the miRNA and target are co-degraded; a miRNA–mRNA interaction in which miRNAs promote mRNA degradation, but in which miRNAs themselves are unaffected, is insufficient. One can, in principle, test the existence of coupled miRNA–mRNA degradation by inhibiting the transcription of mRNA and monitoring its decay rate, which in this case would be time-dependent,
(t) = βm + k μ(t). We note also that a stoichiometric interaction may complicate the interpretation of sensor transgene experiments [27], as the transgene would then sequester miRNA and thereby alter the original expression patterns.
In principle, any interaction between a pair of molecules that obeys the rules of our model, such as an irreversible noncatalytic interaction, can set up a sharp expression interface. For example, suppose that a transcription factor is deactivated by binding irreversibly to an inhibitor protein. In this case, the concentration profile of active transcription factors can exhibit a sharp interface via the mechanism described above. In common with classical reaction–diffusion models for developmental patterning [49,50], an essential property here is that the inhibitor diffuses much faster than its target.
The interface between low and high mRNA levels is characterized by low molecule numbers of both RNA species. In such cases, fluctuations in the molecule number of either species may have macroscopic effects. For example, a small RNA-target pair in bacteria shows enhanced fluctuations when their transcription rates become comparable (E. Levine, M. Huang, Y. Huang, T. Kuhlman, Z. Zhang, and T. Hwa, unpublished data) [51]. These fluctuations can in turn give rise to noise-induced bistability, which manifests itself experimentally as diversity in a population of cells (E. Levine, M. Huang, Y. Huang, T. Kuhlman, Z. Zhang, and T. Hwa, unpublished data). We performed Monte Carlo simulations of our model, but found that fluctuations have no macroscopic effect, even near the transition point where molecule numbers of both species are low. This in-built robustness to fluctuations arises because the interface position is determined by an integrated transcriptional flux which averages out individual cellular fluxes. Thus spatial averaging results in high spatial precision without smoothing out the interface.
Strong cooperative activation, as often occurs in morphogenetic regulation at the transcriptional level (e.g., Bicoid has about five binding sites in target promoters of Drosophila), would seem to make pattern formation by morphogens inherently susceptible to temperature variations [52,53]. Nevertheless, embryonic patterning appears to be quite robust to temperature variations, as has been documented for Hunchback [52] and for Eve [54] in Drosophila. The only cooperative reaction required in the model presented in this work is coupled degradation of miRNA and mRNA, suggesting the possibility that miRNAs filter fluctuations arising from temperature variations.
Candidate systems in which to test the ideas put forth in this study include the establishment of dorsoventral (adaxial/abaxial) leaf polarity in plants, as well as the segmentation of the early Drosophila embryo. We now discuss these two systems in some detail.
Leaf polarity in plants is established shortly after the emergence of the leaf primordium from the meristem. Specification of leaf polarity depends on the Sussex signal [55], a meristem-borne signal that specifies adaxial cell fates. Members of class III of the homeodomain-leucine zipper (HD-ZIPIII) proteins specify adaxial fate [11,56]. In Arabidopsis and in maize, the polar expression pattern of these genes results from their inhibition by two miRNAs, miR165/166, which exhibit a complementary expression pattern [11,12]. Recently, it has been shown that in maize, restriction of miR165/166 to the abaxial side of the developing leaf depends on the polarized expression of LBL-1, a protein involved in the biosynthesis of trans-acting RNAs, ta-siR2141/2 [13]. Possible targets of ta-siR2141/2 include members of the arf3 gene family (a transcription factor that is expressed abaxially), as well as members of the miR166 family [13]. Although miRNAs in plants are thought to act mainly cell autonomously [57], DCL4-dependent siRNAs, such as ta-siR2141/2, may exhibit cell-to-cell movement [40].
The following model is consistent with these data (Figure 5A). The RNA transcript TAS3 is cleaved to produce ta-siR2141/2 in the meristem. These small RNAs then propagate (diffuse) into the adaxial side of the leaf, inhibiting the expression of miR166 either directly or through the ARF3 transcription factor. The target (either miR166 or ARF3) is transcribed uniformly throughout the leaf, and is localized to the cell where it is synthesized. If one further assumes that the interaction between ta-siR2141/2 and its target is noncatalytic, then this model belongs to the class of models studied in this work, and can therefore exhibit a sharp interface between the abaxial domain of high target expression and the adaxial domain of no expression; see Figure 5. In agreement with this model is the low abundance of ta-siR2141/2 in Arabidopsis [58,59], despite their distinct phenotypic role.
Early embryonic development in Drosophila proceeds via a cascade of gene activities that progressively refine expression patterns along the anterior–posterior axis of the embryo. A recent study of the expression patterns of nascent miRNA transcripts suggests that a number of miRNAs play a role in this process. The miRNAs miR-309clus, miR-10, and iab-4 (which all reside between annotated mRNA genes on the genome), and miR-11, miR-274, and miR-281clus (which all reside within introns of annotated genes) are all expressed in a graded fashion along the anterior–posterior axis of the blastoderm embryo [14,60].
The complementary transcription profiles of iab-4 and its target ubx at stage 5 of development make this miRNA-target system a candidate for the sharpening mechanism proposed in this study. The early ubx transcript pattern is broadly distributed over the posterior half of the embryo, becoming localized to a stripe at the center of the embryo by the completion of cell formation [29,61], probably as a result of transcriptional repression by Hunchback in the anterior and posterior regions of the embryo [62]. The nascent transcript profile of its regulator, iab-4, is broadly distributed posterior to this stripe [29].
It may be the case that iab-4 is also expressed before cell formation and that the absence of cell membranes makes iab-4 mobile. The much larger ubx mRNA, on the other hand, may be effectively stationary on the timescales of interest [7]. Furthermore, the transcription profiles of iab-4 and ubx at stage 5 do not seem to overlap [29], suggesting that iab-4 intercellular mobility is crucial to allow it to interact with ubx at this stage of development. Assuming then that only the miRNA iab-4 is mobile, the complementary expression patterns of iab-4 and its target, ubx, measured in [29] is consistent with our model of miRNA-induced sharpening. Sample profiles predicted by the model are shown in Figure 6.
A possible difficulty with regard to applying our model to ubx/iab-4 is that the system may not have reached steady state before stage 6, when cells begin to migrate. In particular, no ubx protein was detected at stage 5, possibly because of the time needed to transcribe the large ubx locus [29].
Like iab-4, the miRNA miR-10 is also expressed at stage 5 in a broad posterior region along the anterior–posterior axis [14]. The homeotic gene Scr is a predicted target of miR-10 [63] and is also expressed in the blastoderm at stage 5 [64]. The miR-10 site in the Scr 3′ UTR is likely to be functional because the pairing is well conserved in all drosophilid genomes and because the miRNA site is conserved in the Scr genes in mosquito, the flour beetle, and the silk moth [20]. Unlike ubx, the protein of Scr is detected at this stage of development in a stripe of ectodermal cells about four cells wide in the parasegment-2 region, though it may not be functional at this time as the protein (a transcription factor) was not localized to the nucleus [64]. This spatial expression pattern is proximal to the anterior limit of miR-10 expression [14,64]. Hence the interaction of miR-10 with Scr at stage 5 of Drosophila development is also a candidate for the sharpening mechanism.
The sharpening mechanism is most effective when the spatial transcription profiles of miRNA and target are regulated in such a way as to be mutually exclusive. The genomic locations of the miRNAs iab-4 and miR-10 are proximal to their targets, which is certainly consistent with the possibility of coordinated regulation [65].
To obtain the concentration profiles for the mRNA and miRNA in the different scenarios considered in this paper, we integrated numerically Equation 2. To do this, one needs to specify the transcription profiles, αm(x) and αμ(x), and the values of the parameters βm, βμ, k, and D. Unless mentioned otherwise in the text, we chose βm = βμ = D = 0.01 and k = 1 throughout.
The transcription profiles of Figures 1 and 3 were
where Am = 2, Aμ = 1, xtsx = 0.5, and λtsx = 0.2. In the stripe geometry (Figure 2), the transcription profile for m was as above, with xtsx = 0.7. The transcription profile of the miRNA was given by
with Aμ = 2, Aμ0 = 0.6, and xtsx = 0.3.
In the Discussion, we outline possible applications in two systems: leaf polarity in maize and segmentation in the early Drosophila embryo. Here, we did not aim to estimate parameters from experimental data (which, in most cases, is not quantitative enough to allow for parameter inference). Instead, parameters were chosen arbitrarily to allow clear demonstration of possible results. In the case of leaf polarity (Figure 5), we chose αm(x) = Am, αs(x) = Aμθ(x − xtsx) with Am = 1,Aμ = 50, and xtsx = 0.99. Here, θ(x) is the unit step function. In the Drosophila embryo (Figure 6), the transcription profile of the iab-4 miRNA was the same as in Equation 3, whereas the ubx mRNA transcription profile was given by
with xtsx = 0.1, λtsx = 0.05, and Am = 2.
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank/) GeneIDs for the genes discussed in this paper are arf3 (817014), eve (36039), hb (41032), hoxb8 (15416), iab-4 (3772110), lbl1 (100037819), miR-10 (3772568), miR-11 (3771987), miR-196 (387191), miR-274 (3771876), miR-281–1 (3772402), miR-281–2 (3772497), miR-309 (3772613), scr (40833), tas3 (3768766), and ubx (42034). |
10.1371/journal.pgen.1008129 | Regulation of ectopic heterochromatin-mediated epigenetic diversification by the JmjC family protein Epe1 | H3K9 methylation (H3K9me) is a conserved marker of heterochromatin, a transcriptionally silent chromatin structure. Knowledge of the mechanisms for regulating heterochromatin distribution is limited. The fission yeast JmjC domain-containing protein Epe1 localizes to heterochromatin mainly through its interaction with Swi6, a homologue of heterochromatin protein 1 (HP1), and directs JmjC-mediated H3K9me demethylation in vivo. Here, we found that loss of epe1 (epe1Δ) induced a red-white variegated phenotype in a red-pigment accumulation background that generated uniform red colonies. Analysis of isolated red and white colonies revealed that silencing of genes involved in pigment accumulation by stochastic ectopic heterochromatin formation led to white colony formation. In addition, genome-wide analysis of red- and white-isolated clones revealed that epe1Δ resulted in a heterogeneous heterochromatin distribution among clones. We found that Epe1 had an N-terminal domain distinct from its JmjC domain, which activated transcription in both fission and budding yeasts. The N-terminal transcriptional activation (NTA) domain was involved in suppression of ectopic heterochromatin-mediated red-white variegation. We introduced a single copy of Epe1 into epe1Δ clones harboring ectopic heterochromatin, and found that Epe1 could reduce H3K9me from ectopic heterochromatin but some of the heterochromatin persisted. This persistence was due to a latent H3K9me source embedded in ectopic heterochromatin. Epe1H297A, a canonical JmjC mutant, suppressed red-white variegation, but entirely failed to remove already-established ectopic heterochromatin, suggesting that Epe1 prevented stochastic de novo deposition of ectopic H3K9me in an NTA-dependent but JmjC-independent manner, while its JmjC domain mediated removal of H3K9me from established ectopic heterochromatin. Our results suggest that Epe1 not only limits the distribution of heterochromatin but also controls the balance between suppression and retention of heterochromatin-mediated epigenetic diversification.
| Suppression of unscheduled epigenetic alterations is important for maintenance of homogeneity among clones, while emergence of epigenetic differences is also important for adaptation or differentiation. The mechanisms that balance both processes warrant further investigation. Epe1, a fission yeast JmjC domain-containing protein, is thought to be an H3K9me demethylase that targets ectopic heterochromatin via its JmjC-dependent demethylation function. Here we found that loss of epe1 induced stochastic ectopic heterochromatin formation genome-wide, suggesting that the fission yeast genome had multiple potential heterochromatin formation sites, which were protected by Epe1. We found that Epe1 prevented deposition of ectopic H3K9me independently of its JmjC-mediated demethylation before heterochromatin establishment. By contrast, Epe1 could attack already-established ectopic heterochromatin via its JmjC domain, but demethylation was not 100% effective, which provided a basis for epigenetic variation. Together, our findings indicate that Epe1 is involved in both maintenance and alteration of heterochromatin distribution, and shed light on the mechanisms controlling individual-specific epigenome profiles.
| Heterochromatin is a silent chromatin structure characterized by methylation of histone H3 at lysine 9 (H3K9me), to which heterochromatin protein 1 (HP1) binds and recruits various effectors including silencing factors. Euchromatin, another well-defined chromatin structure, is generally open and accessible to the transcriptional machinery. Protecting the genome from improper heterochromatin formation in euchromatin regions is important for constitutive gene expression. On the other hand, heterogeneous heterochromatin distribution, which leads to perturbation of gene expression, can contribute to adaptation to specific conditions.
The fission yeast Schizosaccharomyces pombe is a well-established model organism to analyze heterochromatin because of its conserved but simplified heterochromatin assembly system. H3K9 methylation is mediated by the sole H3K9 methyltransferase Clr4 [1, 2]. Constitutive heterochromatin in fission yeast is limited to centromeric repeats, subtelomeric regions, and the mating-type locus, while the other genomic regions consist almost entirely of euchromatin [3]. Centromeric heterochromatin formation depends mainly on the RNAi pathway that requires Dicer and Argonaute proteins encoded by dcr1 and ago1, respectively [4]. Subtelomeric heterochromatin assembly relies redundantly on RNAi and the telomere DNA-binding protein Taz1 [5], but this assembly scheme might not be applicable to subtelomeric heterochromatin of chromosome III. Chromosome III subtelomeres contain ribosomal DNA (rDNA) repeats, which lie adjacent to telomeres and are subject to heterochromatin silencing. However, the assembly mechanism of heterochromatin at rDNA repeats is largely unknown [3, 6–8]. Heterochromatin formation at the mating-type locus depends redundantly on RNAi and ATF/CREB family proteins Atf1/Pcr1 [9].
The JmjC domain is conserved from yeast to human, and demethylates methyl-lysines of histones using Fe2+ and 2-oxoglutarate as cofactors [10]. A JmjC domain-containing protein, S. pombe Epe1, has been identified as an anti-silencing factor at heterochromatin-euchromatin boundaries [11–14]. The Cul4-Ddb1Cdt2 ubiquitin ligase complex moderately degrades Epe1 located inside the boundaries [15, 16]. Overexpression of Epe1 reduces centromeric di-methylated H3K9 (H3K9me2) in vivo [12, 17] and facilitates transcription in heterochromatin and at specific euchromatin loci [12, 13, 17]. In addition, Epe1 promotes histone turnover in heterochromatin [18]. Epe1 physically interacts with the HP1 homologs Swi6 and Chp2 and localizes to heterochromatin in a largely Swi6-dependent manner [12, 15, 19, 20]. Loss of epe1 partially destabilizes centromeric heterochromatin, whereas combined loss of epe1 and dcr1 or ago1 retains an epe1Δ-like silenced state at centromeres, indicating that epe1Δ bypasses the requirement for RNAi [12, 13, 21]. The mechanisms underlying these phenotypes remain unknown. Despite accumulating in vivo evidence, Epe1 does not display demethylation activity in vitro [22]. Epe1 may therefore not be a canonical H3K9 demethylase.
Recent studies suggest genome-wide functions of Epe1. Twenty-one small H3K9me peaks, designated heterochromatin islands (Is), exist in euchromatic regions and Epe1 represses expansion of the islands as well as emergence of additional islands [23]. Combined loss of Epe1 and the histone acetyltransferase Mst2 induces strong ectopic heterochromatin on many euchromatic loci including essential genes, resulting in a severe growth defect, which is eliminated by heterochromatin formation on heterochromatin assembly genes [24]. An artificial Clr4-tethering system establishes extensive heterochromatin on euchromatic regions in the presence of Epe1, and after release of tethered Clr4, the heterochromatin is disrupted by Epe1 in a JmjC-dependent manner [25, 26]. These studies suggest that Epe1 targets ectopic heterochromatin and antagonizes heterochromatin formation in euchromatic regions via its JmjC-dependent function, although whether Epe1 actually removes H3K9me in spontaneously established ectopic heterochromatin is not known.
Here, we found that loss of Epe1 resulted in the repression of a ribonucleotide synthesis gene and a change in cell phenotype by mediating ectopic heterochromatin formation on this gene. Further analyses showed that ectopic heterochromatin was stochastically established, thereby producing an epigenetically diversified population of epe1Δ clone cells. Surprisingly, Epe1 prevented ectopic H3K9me deposition independently of both its JmjC-mediated demethylation and heterochromatin association ability. We identified the N-terminal transcriptional activation (NTA) domain of Epe1 and it contributed to the prevention function. Epe1 removed H3K9me from already-established ectopic heterochromatin, but the removal was not complete. The persistent ectopic heterochromatin harbored an H3K9me supply source that did not deposit H3K9me in the presence of Epe1 (designated as latent H3K9me source). Collectively, the results show that Epe1 has two distinct functions: the protection of euchromatic regions from stochastic de novo ectopic heterochromatin formation via a mechanism involving its N-terminal transcriptional activation (NTA) domain, and incomplete disruption of already-established ectopic heterochromatin via its JmjC domain. Thus, Epe1 could be a key regulator in the formation of individual-specific H3K9me landscapes.
An ade6+ marker inserted in a centromeric heterochromatin region (otr1R::ade6+; Fig 1A) is conventionally used to monitor the state of heterochromatin [27, 28]. On adenine-limited medium, silenced otr1R::ade6+ confers a red color on cells whereas its partial or strong desilencing confers a pink or white color, respectively, since ade6+ encodes an enzyme metabolizing 5-aminoimidazole ribotide (AIR), which develops into a red pigment through multiple steps (Fig 1B). epe1Δ induces heterogeneous expression of otr1R::ade6+ to generate a mixture of red and white colonies. Replating red colonies induces epe1Δ-like variegation, whereas replating white colonies frequently produces white colonies with some red/pink ones [13]. In this study, sequential white-colony isolation of an otr1R::ade6+ epe1Δ strain established several white clones (W70, W164–166) whose color was different from the light pink color of the clr4Δ strain, a canonical heterochromatin-defective mutant (Fig 1C, S1A and S1B Fig). W70 cells stably produced completely white colonies with a frequency of 89%.
ChIP analysis revealed that, unlike clr4Δ cells, epe1Δ W70 cells retained a substantial amount of H3K9me on otr1R::ade6+ (Fig 1D). epe1Δ W70 cells also showed low level of expression of otr1R::ade6+, which was comparable to that of parental epe1Δ cells and red-isolated epe1Δ R69 cells (Fig 1E). By contrast, H3K9me levels on otr1R::ade6+ in W164–166 cells were less than a quarter of those in the WT (S1C Fig), which probably explained the expression of otr1R::ade6+ and the white phenotypes. These results suggested that the white phenotype of W70 was not linked to otr1R::ade6+. The re-appearance of red colonies from epe1Δ W70 cells (Fig 1C) suggested that the white phenotype was due to epigenetic rather than genetic alterations.
To explore the cause of the white phenotype of W70, we performed microarray transcriptome analysis in W70 and R69 cells, which uncovered a cluster of genes silenced only in W70 (Fig 1F, S1D Fig). This cluster encompassed a 23 kb euchromatic region neighboring the right subtelomere of chromosome III (subtel3R); this region contained the ade5 gene, whose product acts upstream of Ade6 in IMP biosynthesis (Fig 1B). Since loss of Epe1 increases H3K9me levels at subtel1L and 2L [23, 24], we hypothesized that the ade5 gene was silenced by ectopically deposited H3K9me, which arrested red pigment formation. ChIP-qPCR analysis in W70 cells revealed the presence of strong ectopic heterochromatin on ade5 with H3K9me levels comparable to those on centromeric ade6+ in WT cells, but such ectopic heterochromatin was not detected in parental epe1Δ, epe1Δ R69 or W164–166 cells (Fig 1D, S1C Fig). Consistently, qRT-PCR analysis revealed that W70 cells had lower ade5 expression than WT, parental epe1Δ, and R69 cells (Fig 1E). In W70 cells, ade5-like accumulation of H3K9me was observed at neighboring puf6 and SPCC569.03 but not at nsa2, which lay outside the cluster identified by transcriptome analysis (S1E Fig). Accordingly, we detected accumulation of Swi6 at these subtel3R genes but not at nsa2 in W70 cells (Fig 1G, S1F Fig). These results suggest that silencing of the ade5 gene by strong ectopic heterochromatin formation caused the white phenotype of W70.
To confirm the cause of the white phenotype, we performed genetic complementation of the W70 strain. We supplied single copies of epe1+ and/or ade5+ using a diploid complementation system (Fig 1H). Control diploid strains displayed expected phenotypes: epe1+/+ and epe1Δ/+, red; epe1Δ/Δ, red-white variegation; and ade5Δ/Δ, completely white (S1G Fig). We designated the ectopic heterochromatin-containing allele in W70 as ade5*. When the W70 strain (epe1Δ ade5*) was mated with an epe1Δ ade5Δ strain to generate epe1Δ/Δ ade5*/Δ, the white phenotype of epe1Δ ade5* was retained, indicating that no alleles except epe1Δ and ade5* of W70 caused the white phenotype (Fig 1H). By contrast, introduction of ade5+ complemented the white phenotype (epe1Δ/Δ ade5*/+), showing that ade5* was responsible for the white phenotype. Similarly, provision of epe1+ complemented the white phenotype of epe1Δ ade5* (epe1Δ/+ ade5*/Δ). ChIP analysis of these diploid cells showed the loss of H3K9me at the ade5 region, indicating that re-introduction of Epe1 promoted demethylation of H3K9me at ade5* (Fig 1I). Thus, we concluded that ectopic heterochromatin-mediated repression of ade5 caused the white phenotype of the epe1Δ W70 strain.
The features of W70 mentioned above prompted us to speculate that loss of Epe1 induces the white phenotype independently of otr1R::ade6+ via ectopic heterochromatin formation at the ade5 locus, although a large part of the red-white variegation in epe1Δ cells depended on variegated expression of otr1R::ade6+. To test this conjecture, we used a strain harboring ade6-m210, which has no functional allele of the ade6 gene. The strain showed a uniform red phenotype on adenine-limited medium. Introduction of epe1Δ induced a red-white variegated phenotype with a lower proportion of pink and white colonies than that of the epe1Δ otr1R::ade6+ strain (Fig 2A and 2B).
We hypothesized that ectopic heterochromatin formation was responsible for the white phenotype in the epe1Δ ade6-m210 strain and that its stochastic formation resulted in red-white variegation. To investigate this hypothesis, we introduced mutations of genes required for heterochromatin formation and analyzed their effects on the variegation phenotype (Fig 2A and 2B). Loss of Clr4 abolished red-white variegation in the epe1Δ ade6-m210 background, indicating a requirement for the H3K9 methyltransferase Clr4. clr3 and sir2, which encode histone deacetylases, are required for self-propagation of heterochromatin [30–34]. The introduction of clr3Δ or sir2Δ into the epe1Δ background also induced a uniform red phenotype. Similarly, loss of Swi6 suppressed variegation. These results suggest that red-white variegation relies on heterochromatin assembly.
We next examined the requirement for Ago1 and Taz1 for epe1Δ-induced variegation, because both factors are involved in subtelomeric constitutive heterochromatin formation [5]. epe1Δ ago1Δ, epe1Δ taz1Δ, and epe1Δ ago1Δ taz1Δ strains displayed red-white variegated phenotypes (S2A and S2B Fig), indicating that neither RNAi nor Taz1 was essential for epe1Δ-induced variegation.
To confirm the relationship between the red-white variegation and stochastic ectopic heterochromatin formation, we established red- and white-isolated clones from ade6-m210 strains harboring epe1Δ, epe1Δ ago1Δ, and epe1Δ taz1Δ (Fig 2C). Using ChIP-sequencing, we analyzed the whole genome distribution of H3K9me of these strains. The results revealed formation of strong ectopic heterochromatin on ade5 in epe1Δ W1-1, W2-1, and W9-1, which was associated with decreased ade5 expression (Fig 2D–2F). This suggests that ade5 ectopic heterochromatin caused the appearance of white colonies of epe1Δ ade6-m210 cells. epe1Δ ago1Δ W2-1 and epe1Δ taz1Δ W7-2 also harbored ade5 ectopic heterochromatin, indicating that neither RNAi nor Taz1 was required for ectopic heterochromatin formation at the ade5 region (Fig 2D, S2C Fig).
The epe1Δ W5-1 strain displayed a pink phenotype without increased H3K9me on ade5, but displayed ectopic heterochromatin formation on ade1 and its neighboring region, and a decreased level of ade1 expression (Fig 2G and 2H, S2D Fig). Since, like Ade5, Ade1 functions upstream of Ade6 in de novo IMP biosynthesis (Fig 1B), the pink phenotype was probably due to ectopic silencing of ade1. Accordingly, red isolates showed no stable pink/white phenotype or ectopic heterochromatin formation on pigmentation-associated genes. These results suggest that white colonies are generated when ectopic heterochromatin is formed at pigmentation-associated genes.
We also performed H3K9me ChIP-seq analysis in previously isolated W70 and W164 strains (otr1R::ade6+ background). Despite having the same parental strain, epe1Δ W70 and W164 clones did not share ectopic heterochromatin at right subtelomeres (Fig 3A). W70 had large ectopic heterochromatin at subtel1R and 3R, while W164 had it at subtel2R. The result suggested that these ectopic heterochromatin domains had been differentially established in the epe1Δ cell mixture. We additionally obtained a W70-like white-isolated clone from the otr1R::ade6+ epe1Δ ago1Δ strain, designated as W173 (Fig 3B, S3A Fig). In contrast to epe1Δ W70 and W164, epe1Δ ago1Δ W173 had ectopic heterochromatin at subtel1R, 2R, and 3R (Fig 3A, 3C and 3D, S3B Fig). We noticed that the gal1 gene was embedded in subtel2R ectopic heterochromatin formed in epe1Δ W164 and epe1Δ ago1Δ W173 clones (Fig 3A and 3E). Swi6 levels on gal1 were higher and gal1 transcript levels were substantially lower in these mutants than in the WT (Fig 3C and 3D), and cells harboring gal1 ectopic heterochromatin displayed defective growth on galactose-containing medium (Fig 3F), indicating that gal1 was silenced by ectopic heterochromatin. This implies that ectopic heterochromatin can affect phenotypes other than red pigment formation. Interestingly, the parental strain of epe1Δ ago1Δ W173 had gal1 ectopic heterochromatin (Fig 3E), indicating that subtel2R ectopic heterochromatin already existed in the parental strain and that ectopic heterochromatin at subtel2R and 3R had been established at different times.
Loss of Epe1 increases H3K9me levels at heterochromatin islands [17, 23, 24, 35, 36]. We observed heterogeneous development of islands among clones harboring the ade6-m210 background (S3C Fig). Note that some of the islands were hardly detected in the WT strains and a large amount of H3K9me stochastically accumulated among isolated epe1Δ clones (for example, Is 3, 8, 19 in S3C Fig). These results suggest that some islands represent an H3K9me source that is inconsistently active.
The red-isolated epe1Δ R2-1 and white-isolated epe1Δ ago1Δ W2-1 clones harbored ectopic heterochromatin on pigmentation-unrelated genes, pdi4 and can1, respectively, both of which did not correspond to known euchromatic H3K9me peaks [23, 24, 37–42], and had no H3K9me in their original WT cells (Fig 3G and 3H, S3D Fig). This suggests that unknown potential H3K9me sources still existed in the S. pombe genome and strong ectopic heterochromatin can be established at these source-positive sites. In addition, epe1Δ taz1Δ W7-2 cells harbored ectopic heterochromatin at lys1 (cen1L) and clr2 loci (S3C Fig). Other ectopic H3K9me peaks in clones are shown in S2 and S3 Tables. These results showed that each epe1-null cell had a distinct epigenotype.
It is noteworthy that the subtel1R heterochromatin landscape was different between the WT (epe1+) strains (Fig 3A and 3I): the WT ade6-m210 strain harbored extended subtelomeric heterochromatin, which was not detected in the WT otr1R::ade6+ strain. The result indicates that the epigenetic profile varies between the WT strains even though they have functional Epe1 and similar genetic backgrounds.
Since Epe1 promotes demethylation of artificially deposited H3K9me in vivo in a JmjC domain-dependent mechanism [25, 26], we asked whether Epe1 demethylation plays a role in suppression of the red-white variegation observed in Fig 2A. To examine this, we generated strains expressing N-terminal tagged wild-type Epe1 (3FLAG-Epe1) and Epe1H297A, which harbors an alanine substitution at the first Fe2+-binding residue [10, 13]. H297A is a canonical catalytic-dead mutant that lacks in vivo demethylation activity on artificially deposited H3K9me [25, 26]. Both proteins were expressed from the endogenous epe1 promoter and displayed almost the same protein expression levels (S4A Fig). Unlike loss of Epe1, the H297A mutation generated few pink/white colonies in the ade6-m210 background (Fig 4A); indeed, 96.2% of Epe1H297A cells formed WT-like red colonies, while 61.7% of epe1Δ cells did. This result indicates that Epe1 almost fully suppressed ectopic heterochromatin-mediated variegation in a JmjC domain-independent manner.
We have already shown that introduction of epe1+ into W70 cells led to a red phenotype via the removal of H3K9me at ade5* (Fig 1H and 1I); however, Epe1 suppressed variegation independently of JmjC activity. Thus, we next analyzed whether the JmjC domain is required for the removal of H3K9me at ade5*, using a diploid complementation assay (Fig 4B–4D). We constructed a carrier strain (epe1Δ W-t1) bearing ade5 ectopic heterochromatin (ade5*) transferred from the epe1Δ ago1Δ W173 strain, and confirmed the retention of subtel3R ectopic heterochromatin including ade5* after sexual reproduction (S4B and S4C Fig). Introduction of epe1H297A as well as epe1Δ into epe1Δ W-t1 resulted in diploid cells that retained the white phenotype and H3K9me at ade5*, while introduction of epe1+ complemented the white phenotype and almost depleted H3K9me of ade5* (Fig 4C and 4D). Furthermore, since the anti-H3K9me antibody recognizes mono-, di-, and tri-methylated H3K9, we tested whether ade5 ectopic heterochromatin contained di- and tri-methylation marks. We detected ectopic H3K9me2 and me3 signals at ade5 and found that their levels were substantially reduced in Epe1-introduced diploid cells (S4D Fig). These results indicate that Epe1 promoted demethylation of methyl-H3K9 including H3K9me2 and me3 at ectopic heterochromatin in a JmjC-dependent manner. However, despite of the loss of its demethylation function, Epe1H297A suppressed ectopic heterochromatin-mediated variegation (Fig 4A). Therefore, Epe1 has a JmjC domain-independent function that suppresses ectopic heterochromatin formation before accumulation of large amounts of H3K9me.
Epe1H297A cells generated a few pink colonies that were not generated by wild-type cells (Fig 4A). Thus, we suspected that the JmjC domain-independent function cannot completely suppress ectopic heterochromatin formation and that JmjC-dependent demethylation contributes to full suppression to some extent. We isolated a white clone (Epe1H297A W2-1) by replating minor pink colonies generated from Epe1H297A cells (Fig 4E). ChIP-seq analysis revealed that Epe1H297A W2-1 cells had ectopic heterochromatin at the subtel3R region (Fig 4F). This suggested that ectopic heterochromatin occurred at a low frequency by escaping JmjC-independent prevention and that JmjC-dependent demethylation removed this ectopic heterochromatin.
In the course of the analysis of Epe1H297A, we examined whether the H297A mutation affected Epe1 localization at heterochromatin. FLAG ChIP analysis revealed that H297A reduced appreciably Epe1 enrichment on centromeric dg repeats and IRC3 (Fig 4G), a centromeric boundary sequence where Epe1 accumulates to a high level [12, 14]. By contrast, H3K9me was maintained at IRC3 and dg (S4E Fig). Although the Swi6 level was significantly reduced, it was still present at both regions (S4F Fig), suggesting that the reduced Swi6 level may not have been the primary cause of the drastic reduction in Epe1H297A enrichment on heterochromatin. We next validated the physical interaction between the Epe1H297A mutant and Swi6 by yeast two-hybrid analysis. Because budding yeast has no endogenous H3K9me or HP1, neither of these factors could affect the analysis. The H297A substitution slightly impaired the interaction (S4G Fig), which was confirmed by the results of a bait-prey exchange experiment. We performed co-immunoprecipitation analysis of Swi6 with Epe1H297A. Consistent with the results of yeast two-hybrid assay, the Epe1H297A mutant interacted with Swi6 with a slightly lower efficiency than wild-type Epe1 (Fig 4H). Together, these results showed that mutation of the Fe2+-binding residue in the JmjC domain slightly impaired the physical interaction between Epe1 and Swi6, but largely reduced targeting of Epe1 to heterochromatin. In addition, these results showed that heterochromatin association activity of Epe1 was not important for the JmjC-independent prevention of ectopic heterochromatin formation.
However, how Epe1 finds target sites to prevent ectopic heterochromatin formation is unknown. Since Epe1 physically interacts with the bromodomain protein Bdf2, which is required for heterochromatin-euchromatin boundary formation [14], we predicted that Bdf2 would recruit Epe1 to the target sites. However, bdf2Δ cells showed an almost uniform red phenotype in the ade6-m210 background (S4H Fig), suggesting that Bdf2 was not related to suppression of variegation and ectopic heterochromatin formation.
In the yeast two-hybrid system, the Epe1 protein expressed as bait activates transcription of reporter genes without prey [15] (Fig 4I). We found that deletion of the N-terminal 171 amino acids (Epe1ΔN) abolished transcriptional activation by Epe1 and the N-terminal 208 amino acids (Epe1N208) activated transcription of the HIS3 reporter independently of JmjC (Fig 4I), suggesting that the N-terminal 171 amino acids region is required for the transcriptional activation activity.
To examine whether the N-terminal 1–171 amino acids region has a transcriptional activation function in fission yeast, we established a plasmid-based reporter system (Fig 4J). We constructed a reporter plasmid containing the coding sequence of ade6-m210 transcribed from 154 bases of the nmt1 promoter combined with three copies of Gal4 binding sites instead of its thiamine regulatory element [43]. We also constructed expression plasmids for expressing Epe1 or its mutants fused to the Gal4 DNA binding domain (Gal4DBD). These plasmids were introduced into a fission yeast strain harboring the ade6-M216 allele. Since ade6-m210 and ade6-M216 alleles show intragenic complementation, cells expressing ade6-m210 would grow on medium without adenine, like ade6+ cells. Indeed, cells expressing the VP16 transactivation domain (TAD), a well-characterized transcriptional activation domain, fused to Gal4DBD grew on medium without adenine, while cells expressing Gal4DBD alone did not (Fig 4J). Cells expressing Epe1, Epe1H297A, and NTA fused to Gal4DBD, but not Epe1ΔN, showed Ade+ phenotypes (Fig 4J). This indicated that the NTA domain (1–171 amino acids region) harbored transcriptional activation activity in fission yeast. Note that expression of Gal4DBD-VP16TAD or -NTA fusion protein induced growth defect on the control plate, which is likely caused by off-target effects of Gal4DBD fusion proteins with a strong activation domain.
We introduced Epe1ΔN into ade6-m210 cells to examine the effect of the ΔN mutation on the suppression of ectopic heterochromatin formation. Epe1ΔN cells formed pink/white colonies with a slightly lower frequency than epe1Δ cells (Fig 4K), indicating that the NTA domain contributed to the suppression of ectopic heterochromatin-mediated variegation. We performed co-immunoprecipitation analysis of Swi6 with Epe1ΔN and found that Epe1ΔN interacted with Swi6 with much lower efficiency than wild-type Epe1 (Fig 4H). However, consistent with the previous report [13], the C-terminal half of Epe1 (487–948 amino acids region) interacted with Swi6 in the yeast two-hybrid system, but the N-terminal half (1–486) did not (S4I Fig). These results suggest that the C-terminal half contains a Swi6 binding site and the NTA domain indirectly contributes to Epe1–Swi6 interaction. In summary, our results suggest that the NTA domain in Epe1, rather than its JmjC domain or heterochromatin association, is required for the prevention of ectopic H3K9me deposition. By contrast, the intact JmjC domain of Epe1 is required for the removal of already-established ectopic heterochromatin.
As shown in Fig 4D, introduction of a single copy of epe1+ almost completely removed H3K9me on ade5*. However, H3K9me as well as specific H3K9me2 and H3K9me3 marks on a subtel3R gene, SPCC569.03, were maintained after introduction of epe1+ (Fig 5A, S4D Fig). Importantly, no H3K9me accumulation was observed on SPCC569.03 before deletion of epe1 (S1E Fig), indicating that ectopic heterochromatin formation at subtel3R was a partially irreversible epigenetic alteration and that Epe1 selects the position for demethylation.
We next asked whether an increase in the level of Epe1 would reduce persistent ectopic H3K9me. Introduction of an Epe1 overexpression (Epe1OE) allele achieved complete removal of H3K9me on SPCC569.03, suggesting that the amount of Epe1 is critical for the removal of residual H3K9me (Fig 5A).
After re-introduction of Epe1 into epe1Δ ade5* cells, we observed a difference in persistent H3K9me levels between ade5 and SPCC569.03, both of which had not undergone H3K9me deposition in the presence of Epe1 (Figs 4D and 5A). Since SPCC569.03 is close to subtelomeric constitutive heterochromatin, we speculated that Epe1 would not remove already-established heterochromatin neighboring a constitutive supply source of H3K9me. To investigate this, we used a mild H3K9me source that does not deposit H3K9me in the presence of Epe1 but does in the absence of Epe1. Taz1, a subunit of the telomere protection complex Shelterin, binds to telomeric repeats [44, 45]. Five heterochromatin islands harbor 2–5 copies of telomere repeat units coupled with a late replication origin, which recruit Clr4 via Shelterin to deposit H3K9me [36, 46]. Thus, we predicted that low numbers of telomere repeats lacking the late origin would induce little or no H3K9me accumulation when Epe1 was present, but initiate H3K9me deposition when Epe1 was absent. We constructed a Taz1-binding cassette (LEU2-4TBS), which consists of the coding sequence (CDS) of S. cerevisiae LEU2 and four copies of a unit of telomere repeats [44, 47], and introduced it into the ade6-m210 background strain, where it was inserted after the SPCC569.06 promoter, a location distant from ade5 (Fig 5B). S. cerevisiae LEU2 functions in place of S. pombe leu1. The inserted LEU2 complemented the leu1− allele harbored by the ade6-m210 strains, resulting in growth on −Leu medium (Fig 5C). The epe1Δ LEU2-4TBS strain displayed marginal growth retardation on −Leu medium, indicating that 4TBS did not strongly silence LEU2 even in the absence of Epe1. This strain also showed red-white variegation. White colony isolation constantly generated clones that displayed Leu− and irreversible white phenotypes. One of the clones, epe1Δ LEU2-4TBS W1-1, harbored higher H3K9me deposition on ade5, LEU2, and surrounding genes (Fig 5C and 5D, S5A Fig). We designated the silenced ade5 and LEU2 alleles as ade5* and LEU2*, respectively. Diploid-based complementation showed that introduction of wild-type Epe1 resulted in a red colony and Leu− phenotype (Fig 5E and 5F). Consistently, H3K9me on ade5 was almost completely removed, whereas that on LEU2 and SPCC569.03 was not reduced by the provision of a single copy of Epe1 (Fig 5G, S5B Fig), suggesting that Epe1 failed to antagonize heterochromatin neighboring the H3K9me source but greatly decreased H3K9me levels at non-contiguous locations. By contrast, introduction of Epe1OE resulted in a red colony and Leu+ phenotype and the complete removal of subtel3R ectopic H3K9me, indicating that Epe1OE overcame 4TBS-induced self-retaining heterochromatin. Constitutive heterochromatin harboring a strong maintenance system such as RNAi, in contrast to weak systems such as 4TBS, withstands Epe1 overexpression to some extent [12, 17]. Our results suggest that the balance between the level of Epe1 and intensity of the self-maintenance system in each domain determines the extent of heterochromatin retention. Importantly, LEU2-4TBS deposited no H3K9me on LEU2 or ade5 in the epe1+ background, whereas accumulation was obvious in epe1Δ (Fig 5D). This strongly suggests the existence of a mechanism for the retention of altered chromatin structure, in which Epe1 suppresses de novo ectopic heterochromatin establishment, but after establishment via escape from suppression, the ectopic heterochromatin is retained by the latent H3K9me source despite the presence of Epe1. In summary, we reveal that Epe1 allows retention of robust ectopic heterochromatin, in which the level of persistent ectopic H3K9me is determined by the amount of Epe1, the distance between target H3K9me and the H3K9me source, and the intensity of the source.
In this study, we demonstrated that Epe1 suppressed ectopic heterochromatin formation via two mechanisms as shown in Fig 6. First, Epe1 prevents the initial deposition of H3K9me at potential ectopic heterochromatin formation sites via a mechanism that does not require the function of its JmjC domain. Second, Epe1 promotes demethylation of H3K9me from established ectopic heterochromatin via a mechanism that requires its JmjC domain. These two distinct mechanisms cooperate to suppress ectopic heterochromatin formation. Loss of Epe1 induced stochastic accumulation of H3K9me, resulting in heterogeneous ectopic heterochromatin formation among clonal cells. Because the demethylation function of Epe1 is antagonized by a constitutive supply of H3K9me, some of the ectopic heterochromatin established by escape from prevention can be retained and provide a basis for epigenetic differences.
The JmjC mutant Epe1H297A almost completely suppressed the red-white variegation induced by ectopic heterochromatin assembly, while it failed to remove already-established ectopic heterochromatin (Fig 4A and 4D). Therefore, Epe1 prevents H3K9me deposition before the establishment of ectopic heterochromatin. Although epe1H297A cells largely maintained heterochromatin at IRC3 and dg (S4E and S4F Fig), Epe1H297A largely lost its heterochromatin localization (Fig 4G). Nevertheless, Epe1H297A prevented ectopic heterochromatin formation. This suggests that Epe1 recognizes potential ectopic heterochromatin formation sites in a manner distinct from heterochromatin targeting. However, the targeting mechanism is unclear at this stage.
Although no domain other than JmjC had been found in Epe1, the newly identified NTA domain showed transcriptional activation ability in both budding and fission yeasts (Fig 4I and 4J). These results show that the NTA domain functions as a conserved transcriptional activation domain like VP16 TAD to recruit RNA polymerase II. Accordingly, Epe1 interacts with a histone acetyltransferase complex, SAGA, which is involved in transcriptional activation [48]. Importantly, loss of the NTA domain induced a variegation phenotype (Fig 4K). These results raise the possibility that the transcriptional activation activity of Epe1 is involved in the prevention of H3K9me deposition. Recent reports show that loss of Leo1, a component of the transcription elongation complex Paf1C, causes ectopic heterochromatin formation [37, 49, 50]. Similarly, loss of Mst2, a histone acetyltransferase, induces ectopic heterochromatin formation [24]. Importantly, loss of each of these proteins causes a decrease in histone turnover [24, 50]. Since histone turnover is associated with transcription [50, 51] and Epe1 promotes histone turnover in heterochromatin [18], Epe1 might exclude H3K9me-containing nucleosomes by activating histone turnover coupled with transcriptional activation at the potential ectopic heterochromatin formation sites. It is still possible that the NTA domain has a function, other than transcriptional activation, which contributes to the prevention of ectopic heterochromatin formation. We found that lack of the NTA domain impaired the interaction of Epe1 with Swi6 (Fig 4H). Heterochromatin association is not required for the prevention of ectopic H3K9me deposition as described above, but Epe1–Swi6 interaction might contribute to the prevention through a mechanism independent of heterochromatin association. Further studies such as determination of important amino acids for the transcriptional activation and interaction with Swi6 will clarify the mechanism of JmjC-independent prevention.
The conserved histidine at residue 297 in the JmjC domain, which is important for Fe2+ binding, is essential for the promotion of H3K9 demethylation by Epe1 in vivo (Fig 4D) [25, 26]. We found that histidine 297 was required for localization to heterochromatin (Fig 4G). We also found that the H297A mutation slightly impaired the interaction of Epe1 with Swi6, although this may not fully explain the reduced heterochromatin localization of Epe1 (Fig 4H, S4G Fig). In contrast, Swi6 interacted with the C-terminal half of Epe1, which lacks the JmjC domain (S4I Fig) [13]. Swi6 is shown to interact with a JmjC mutant, Epe1Y307A, which retains the metal-binding residues [12]. Thus, we speculate that conformational changes in the JmjC domain induced by perturbations in Fe2+ binding result in a slight alteration of the interaction surface for Swi6 binding, while severely disrupting the structure of a region essential for heterochromatin association. Heterochromatin association of Epe1 might require another mechanism in addition to its binding to Swi6. Combined with the result of Epe1ΔN–Swi6 interaction analysis (Fig 4H), it is possible that the N-terminal half region of Epe1 contributes to proper conformation for Epe1–Swi6 interaction and heterochromatin association.
Analysis of the H3K9me distribution of isolated epe1Δ clones by ChIP-seq revealed that each clone had a unique H3K9me landscape. Thus, ectopic heterochromatin is formed stochastically and maintained stably, but its landscape occasionally shifts to another state. Therefore, we could detect several metastable heterochromatin landscapes among epe1Δ isolates.
Ectopic heterochromatin domains formed in epe1Δ cells seemed to have preferred positions on the genome. In addition to heterochromatin islands, we found novel ectopic heterochromatin domains such as clr2, pdi4, ade1, and can1 (Figs 2G and 3G, S3C Fig), which appeared to have no identifiable H3K9me source, except for convergent genes [23, 24, 36, 38–40]. In addition, many ectopic heterochromatin domains identified in several mutant backgrounds have no known H3K9me source [23, 24, 37, 42], while polymerase pausing is an emerging heterochromatin-inducible mechanism [41]. These facts suggest that many potential H3K9me sources for heterochromatin formation exist in the genome and are incidentally activated to form ectopic heterochromatin.
Loss of H3K9 demethylase in other organisms commonly alters traits by perturbing the epigenomic state: in male mice, stochastic formation of testes, ovaries, and testis-ovary hybrids is induced by perturbations in the levels of histone modifications on the testis-determining gene Sry [52, 53]; and in Arabidopsis, a variety of developmental defects and genome-wide deposition of H3K9me and 5mC are observed [54, 55]. We further assume that mutations in H3K9 demethylases accelerate intratumoral heterogeneity, which interferes with chemotherapy by generating drug-tolerant subpopulations, because tumor evolution involves increases in repressive and decreases in active histone marks [56, 57], and co-dependency between genetic and epigenetic mutations possibly enhances the heterogeneity [58].
We found that re-introduction of single copy Epe1 did not erase ectopic heterochromatin when an H3K9me source existed nearby, while Epe1 overexpression completely erased it. On the other hand, increased levels of Epe1 impair constitutive heterochromatin [12, 15, 17]. The expression level of endogenous Epe1 appears to be appropriately regulated to allow ectopic heterochromatin retention while keeping constitutive heterochromatin intact. Since Epe1 is degraded in S phase by the Cul4-Ddb1Cdt2 complex [15] and phosphorylation of Swi6 affects localization of Epe1 to heterochromatin [20], we assume that transient inactivation, loss of expression, or delocalization of Epe1 provides an opportunity to change the heterochromatin landscape. Differences in the H3K9me landscape between wild-type strains might reflect transient changes in Epe1 activity (Fig 3A and 3I). Recently, Gallagher et al. reported that low temperature induces the formation of additional heterochromatin islands including SPCC569.03 [59], where we found robust ectopic heterochromatin that tolerated the re-introduction of a single copy of epe1+ (Fig 5A). This emergence of islands might be caused by impaired Epe1 function at low temperatures. Such regulatory mechanisms appear to be widely applicable to organisms that have demethylases that erase repressive histone methyl marks. Although the genome contains numerous potentially H3K9me-inducible sequences [60–62], not all of them exhibit H3K9me accumulation. Ectopic heterochromatin formation at these sequences, inducible under specific conditions, might result in unprogrammed epigenetic differences.
Variation of the H3K9me landscape could produce adaptive subpopulations, because the variation could switch metabolic pathways or induce changes in growth to those suited for survival in a particular environment [63]. Indeed, ectopic heterochromatin affects ribonucleotide synthesis (ade5 and ade1) and carbon source metabolism (gal1). This study provides insights into the mechanisms of epigenetic diversification and maintenance, which underlie cellular homeostasis and heterogeneous evolution.
The S. pombe strains used in this study are listed in S5 Table. The media recipes used were previously described [64, 65]. The isolated clones were obtained by two rounds of single-colony isolation, in which “R” or “W” in the strain name means sequential red- or white-colony isolation. The DNA fragments for gene deletion or tagging were constructed using the polymerase chain reaction (PCR)-based method as previously described [66]. For gene deletion, target genes were replaced by the drug-resistant cassettes kanMX6, hphMX6, and natMX6 that confer resistance to G418, hygromycin B, and nourseothricin, respectively. The cassettes for expressing 3FLAG-tagged Epe1 and its mutants were constructed on plasmids. The cassettes were then cut out with SmaI and introduced into S. pombe cells. The 3FLAG tag was placed at the N-terminus of Epe1; C-terminal tagging was avoided because of its effect on Epe1 functions [17, 50]. The four copies of a Taz1 binding sequence (4TBS) were represented by 5’-GGGTTACAGGGGTTACAGGGGTTACAGGGGTTACAG-3’, composed of four GGTTACAG sequences combined with guanine stretches [44, 47]. For overexpression of 3FLAG-tagged Epe1 (Epe1OE), the urg1 promoter and 3FLAG sequence were inserted between the promoter and the CDS of epe1. All integrations were confirmed by PCR. The haploid h+ and h−strains composing diploid cells harbor kanMX6 and natMX6 at the epe1 locus, respectively. Diploid cell formation was confirmed by dark magenta color on medium containing 5 mg/L Phloxine B (PB; Nacalai Tesque) and resistance to both 100 mg/L G418 sulfate (Wako) and ClonNAT (nourseothricin dihydrogen sulfate; WERNER BioAgents).
Saturated cells were adjusted to 1 × 108 cells/mL in sterilized water. For preculture of diploid cells, medium containing 100 mg/L G418 sulfate and ClonNAT was used. Diploid cells were saturated without the antibiotics. The suspended cells were serially 10-fold diluted up to 1 × 103 cells/mL. The suspension (6 or 8 μL) was spotted on YE-based complete media or PMG-based synthetic media. To complement genetic mutations, supplements mix (225 mg/L adenine, uracil, histidine, leucine, and lysine, as final concentration) was added (YES and PMGS). Silencing assays were performed on YE media with adenine-dropout supplements mix (Low Ade) and PMG media with leucine-dropout supplements mix (–Leu). The galactose-containing medium (YEGal) was made by replacement of the major carbon source: 30 g/L of galactose, instead of glucose, was added to YES. For haploid cells, plates were incubated at 30°C for 4 days. For diploid cells, plates were incubated at 30°C for 3 days. No assay medium contained antibiotics.
Cells were cultured in 20 mL of YES to 1 (within ±0.1) × 107 cells/mL. The harvested cells were washed with PBS and stored at −80°C. The cell pellet was suspended in AE buffer (50 mM sodium acetate (pH 5.2) and 10 mM EDTA) containing 1% SDS, and then the equivalent volume of acid phenol was added to the suspension. Total RNA was extracted by a freeze-thaw treatment made up of five cycles of rapid freezing in liquid nitrogen followed by incubation in a water bath at 65°C with vortexing. The RNA was subjected to another acid phenol treatment followed by acid phenol/chloroform and chloroform treatments. RNA was recovered by ethanol precipitation and treated with 5 U of recombinant DNase I (Takara Bio) at 37°C for 30 min. DNase I was removed by acid phenol/chloroform treatment. Using Oligo (dT)15 primer, 1 μg of total RNA was reverse transcribed into cDNA with PrimeScript Reverse Transcriptase (Takara Bio) at 37°C for 30 min. Quantitative PCR (qPCR) was performed with SYBR Green I dye on a Thermal Cycler Dice Real Time System TP-850 (Takara Bio). The RT− samples (pseudo-experiments without reverse transcriptase) were also subjected to qPCR. The primer sets are listed in S6 Table. The standard curve for each primer set was created from serially 1-to-1000-fold diluted cDNA samples of clr4Δ cells. The signals of RT− samples were low and seldom fell within the standard curve, and consequently no RT− sample was adequately analyzed. Relative concentrations of cDNA based on the standard curve were divided by the concentration of act1 to determine the transcript levels relative to act1. The error bars represent the standard deviation of the mean of three independent experiments (n = 3). Each experiment was independently performed from cell culture to qPCR. Note that ade6-DN/N harbors a 153 bp of deletion between NcoI sites [28]. The primer set for ade6 specifically detects the centromere-derived transcripts, avoiding amplification of the truncated allele.
The microarray analysis was performed as described previously [67]. Based on the expression ratio, genes with a fold change >1.5 (upregulated) or <1.5 (downregulated) are highlighted in S1D Fig. This experiment was not repeated.
Cells grown to 1 (within ±0.1) × 107 cells/mL in 50 mL of YES were fixed with 1% formaldehyde (Nacalai Tesque) for 20 min at 25°C. Diploid cells were precultured in YES containing 100 mg/L G418 sulfate and ClonNAT and then grown to 5 (within ±0.5) × 106 cells/mL in 50 mL of YES without antibiotics followed by the same formaldehyde treatment as for haploid cells. Quenching of the fixative was performed with 150 mM glycine. The cells were harvested by centrifugation and stored at −80°C. Note that the diploid cells were stored for no more than 1 day. The cells were resuspended in Buffer 1 (50 mM HEPES-KOH (pH 7.5), 140 mM NaCl, 1mM EDTA, 1% Triton X-100 (Nacalai Tesque), and 0.1% sodium deoxycholate (Merck Millipore)) containing a protease inhibitor cocktail, and then homogenized with 30–40 cycles of bead beating for 60 s at 4°C to render them refraction-negative under a light microscope. The cell extracts were centrifuged for 60 min at 21,880 × g at 4°C. After discarding the supernatant, the pellets were resuspended in Buffer 1 containing a protease inhibitor cocktail and sonicated with resonant metallic bars for 360 s with Bioruptor UCW-310 (Cosmo Bio) set at 310 W (High level) and cooled to around 4°C. The sonicated cell extracts were centrifuged for 15 min at 21,880 × g at 4°C, and the resultant supernatant was recovered. Before IP, Dynabeads M-280 Sheep anti-Mouse or anti-Rabbit IgG (Invitrogen) were washed once with Buffer 1 and incubated with 1 μg of anti-H3K9me (mouse monoclonal, m5.1.1, a kind gift from T. Urano, Shimane University), anti-FLAG (mouse monoclonal, M2, Sigma-Aldrich), anti-Swi6 (rabbit polyclonal, made by S.T.), anti-H3K9me2 (mouse monoclonal, 6D11, a kind gift from H. Kimura, Tokyo Institute of Technology), or anti-H3K9me3 (mouse monoclonal, 2F3, a kind gift from H. Kimura, Tokyo Institute of Technology) [68] antibody for 2 h with mild rotation at 4°C followed by washing with Buffer 1. Note that the anti-H3K9me antibody detects mono-, di-, and tri-methylated H3K9 but not unmethylated H3K9. The beads were incubated with the supernatant for 2 h with mild rotation at 4°C. After IP, the beads were washed twice each with Buffer 1, Buffer 1’ (50 mM HEPES-KOH (pH 7.5), 500 mM NaCl, 1 mM EDTA, 1% Triton X-100, and 0.1% sodium deoxycholate), and Buffer 2 (10 mM Tris-HCl (pH 8.0), 250 mM LiCl, 0.5% NP-40 (Roche), and 0.5% sodium deoxycholate) followed by another wash with Buffer 1. The beads were resuspended in Buffer 1 containing RNase A and incubated for 15 min at 37°C and then incubated in 0.25 mg/mL Proteinase K and 0.25% SDS for 2 h at 37°C to obtain IP samples. For input samples, one-fifth volume of the supernatant applied to IP was equally subjected to RNase A and Proteinase K treatments. IP and input samples were incubated for 12–16 h at 65°C for reverse crosslinking. DNA was extracted by neutral phenol/chloroform treatment and recovered by ethanol precipitation. qPCR was performed with SYBR Green I dye on Thermal Cycler Dice Real Time System TP-850 (Takara Bio). The primer sets are listed in S6 Table. The standard curve for each primer set was created from serially diluted input samples of WT cells. Relative concentrations of IP samples based on the standard curve were divided by those of input samples to determine the IP efficiency (IP/input). For FLAG ChIP analysis, IP efficiency of FLAG-tagged cells was divided by that of no tag (WT) cells to determine fold enrichment (fold over no tag). For Swi6 ChIP analysis, incubation at 18°C for 2 h before fixation, which has been usually done to increase signal intensity in previous studies, was not carried out because low temperature induces ectopic heterochromatin formation [59, 69]. This might have increased background levels. The error bars represent the standard deviation of the mean of three independent experiments (n = 3). Each experiment was independently performed from cell culture to qPCR.
Cells were grown to 1 (within ±0.1) × 107 cells/mL in 500 mL of YES. The following procedure was identical to that of ChIP-qPCR analysis except for qPCR. In the ChIP-seq analysis of ade6-m210 cells, DNA was recovered using the spin column-based QIAquick PCR Purification Kit (Qiagen) instead of ethanol precipitation. The ChIP libraries for the Illumina platform were prepared according to the manufacturer’s instructions. The libraries from ade6-m210 and otr1R::ade6+ strains were sequenced on the Illumina HiSeq 1500 system (single-end, 51 bp) and the HiSeq 2500 system (single-end, 101 bp), respectively. The sequenced reads were mapped onto the S. pombe genome (972) using BWA (version 0.7.17), and then processed using SAMtools (version 1.6) and MACS (version 2.1.1). MACS extended each read to the expected fragment length of 200 bp using the option—extsize 200. The processed ChIP-seq data were loaded into IGV. Blue graphs indicate normalized piled-up fragment counts at single base-pair resolution: the scale of the vertical axis is represented as fragment counts per five million mapped reads. The number of both raw and mapped reads is listed in S4 Table. The experiments were not repeated.
Cells grown to 1 × 107 cells/mL in 15 mL of YES were harvested, washed, resuspended in water, and heated at 95°C for 5 min. An equal volume of buffer containing 8 M Urea, 4% SDS, 0.12 M Tris-HCl (pH 6.8), 20% glycerol, and 0.6 M β-mercaptoethanol was added to cell suspensions, and the cells were homogenized by bead beating. The cell extracts were heated at 95°C for 5 min and spun down, and then the supernatants were recovered. The supernatant samples were separated by polyacrylamide gel electrophoresis, and the proteins were blotted onto nitrocellulose membranes. The membranes were first probed with the anti-FLAG (M2, Sigma-Aldrich) and anti-α-tubulin (B-5-1-2, Sigma-Aldrich) antibodies, and then with horseradish peroxidase-conjugated anti-Mouse IgG (GE Healthcare).
Matchmaker GAL4 Two-Hybrid System 3 (Clontech) was used for the Epe1-Swi6 interaction analysis. pGBKT7 (containing TRP1) was used for bait expression. pGADT7 (containing LEU2) was used for prey expression. pGBKT7 and pGADT7 plasmids were introduced into the AH109 host strain by a polyethylene glycol/lithium acetate (PEG/LiAc)-mediated method. AH109 harbors a HIS3 reporter gene. Yeast strains were cultured on proper minimal synthetic dropout (SD) media according to the user manual (Clontech). 3-AT (15 mM; 3-amino-1,2,4-triazole), an inhibitor of the HIS3 product, was added for precise analysis of the Epe1-Swi6 interaction. Glucose was applied to SD media as the carbon source. The experimental procedure for serial dilution assays of S. cerevisiae strains was the same as for S. pombe strains.
Six hundred cells were spread onto a plate with a 90 mm diameter containing adenine-limited media, which generated 200–500 colonies depending on the mutation. For cell spreading, sterile glass beads were used. The cells were incubated at 30°C for 4 days on adenine-limited media, and photographed for assessment. No image processing software was applied to the assessment. Colony color on adenine-limited medium was assessed in randomized photographs and the sample names were masked. The color was grouped into five types: red or dark red; pale red; reddish pink or pink; light pink; and completely white color. The group “completely white” only included the white color observed in ade5Δ cells. Colonies too small for color assessment were grouped into “too small” and excluded from the percentage graph. Although the color of small colonies appeared pale, colony color was not adjusted for colony size in the assessment. When samples displayed a uniform color (at least >99%), the assessment area was reduced to half of the plate. The number of colonies of a particular color type is shown in S1 Table.
Co-immunoprecipitation analysis was performed as described previously with some modifications [12, 14]. Cells grown to 1 (within ±0.1) × 107 cells/mL in 50 mL of YES were harvested by centrifugation, washed with 2 × HC buffer (200 mM HEPES-KOH (pH 7.5), 300 mM KCl, 2 mM EDTA, and 20% glycerol), frozen in liquid nitrogen, and stored at −80°C. Cells were resuspended in 2 × HC buffer containing a protease inhibitor cocktail and 2 mM DTT, and then homogenized with 16 cycles of bead beating for 15 s at 4°C to render 90% of them refraction-negative under a light microscope. The cell lysate was centrifuged for 10 min at 14,000 × g at 4°C, and the resultant supernatant was recovered. Twenty microliter of the supernatant was mixed with an equal volume of 2 × Laemmli buffer (8 M Urea, 2% SDS, 0.12 M Tris-HCl (pH 6.8), 20% glycerol, and 0.6 M β-mercaptoethanol) and heated at 95ºC for 5 min. Before IP, Dynabeads M-280 Sheep anti-Mouse IgG (Invitrogen) were washed once with 1 × HC buffer and incubated with 1 μg of anti-FLAG (mouse monoclonal, M2, Sigma-Aldrich) antibody for 2 h with mild rotation at 4°C followed by washing with 1 × HC buffer. The beads were suspended in the cell lysate mixed with an equal volume of 200 mM KCl containing a protease inhibitor cocktail, and incubated for 4 h with mild rotation at 4°C, where IP reaction was actually performed in 1 × HC buffer containing 250 mM KCl and 1 mM DTT. After IP, the beads were washed eight times with 1 × HC buffer containing 250 mM KCl, suspended in 20 μl of 1 × HC buffer followed by addition of an equal volume of 2 × Laemmli buffer, heated at 95ºC for 5 min, and spun down. The denatured samples were separated by polyacrylamide gel electrophoresis, and the proteins were blotted onto nitrocellulose membranes. The membranes were first probed with the anti-FLAG and anti-Swi6 antibodies, and then with horseradish peroxidase-conjugated anti-Mouse and anti-Rabbit IgG (GE Healthcare), respectively.
To construct the reporter plasmid, a sequence including three Gal4-binding sites of the GAL10 promoter was amplified with the primers 5’-CTTGCATGCGTGAAGACGAGGACGCAC-3’ and 5’-CTCATTGCTATATTGAAGTACGG-3’ from the S. cerevisiae W303 genome; the 154-bp region of the nmt1 promoter, which contained core promoter sequence but lacked the thiamine regulatory element [43], was amplified with the primers 5’-CCGTACTTCAATATAGCAATGAGGCAGCGAAACTAAAAACCG-3’ and 5’-GTCGACATGATTTAACAAAGCGAC-3’; the coding sequence of ade6-m210 was amplified with the primers 5’-CTTTGTTAAATCATGTCGACGAGCGAAAAACAGGTTGTAG-3’ and 5’-TTTACCCGGGCTATGCAGAATAATTTTTCCAACC-3’ from the S. pombe FY648 genome. These DNA fragments were fused by the polymerase chain reaction (PCR)-based method as previously described [66]. The fused fragment was cut with SphI and XmaI and ligated into the pSLF173 plasmid [70] digested with the same enzymes for removal of the nmt1 promoter, and the resultant plasmid was named pGNP154-Am. To construct expression plasmids, the 3FLAG-Gal4DBD fragment was made by PCR fusion and ligated into the pREP41 plasmid with NdeI and BamHI. The resulting plasmid was used as an empty (control) plasmid, named pNFD41. An SV40 nuclear localization signal (NLS) sequence and a multi-cloning site (MCS) were added to the 5’ end of the forward and reverse primers, respectively, to amplify the Gal4DBD sequence (5’-GACAAGGGTGGTGGCTCCCCAAAAAAGAAGAGAAAGGTCGAAGACGCAATGAAGCTACTGTCTTCTATCG-3’ and 5’-TTCTGGATCCGTCGACGCGGCCGCCATGGAACCTCCTCCCGATACAGTCAACTGTCTTTG-3’). The extended Gal4DBD fragment was then fused to the 3FLAG fragment by another PCR to make the 3FLAG-Gal4DBD fragment. Target sequences, epe1 mutants or the transactivation domain (TAD) of VP16 derived from human herpesvirus 1, were ligated into the empty plasmid with NcoI and SalI. Not all primers used for construction of expression plasmids are shown because of the complexity. The plasmids are listed in S7 Table. The TP4-1D strain, harboring ade6-M216, was transformed with reporter and expression plasmids, which contain ura4 and LEU2 markers, respectively. The transformants were cultured on PMG medium lacking uracil and leucine with 15 μM of thiamine. Serial dilution assay was performed using PMG medium lacking uracil, leucine, and adenine without thiamine. Control and assay plates were incubated at 30°C for 8 days.
The sequences of the probes and the original data from the microarray experiments were deposited in GEO (http://www.ncbi.nlm.nih.gov/geo) under accession number GSE108448. The sequencing data of ChIP-seq analyses including input and IP were deposited in DDBJ (https://ddbj.nig.ac.jp/DRASearch) under accession number DRA006424 for the ade6-m210 strains and DRA006425 for the otr1R::ade6+-harboring strains.
|
10.1371/journal.pcbi.1002738 | A Structural Systems Biology Approach for Quantifying the Systemic Consequences of Missense Mutations in Proteins | Gauging the systemic effects of non-synonymous single nucleotide polymorphisms (nsSNPs) is an important topic in the pursuit of personalized medicine. However, it is a non-trivial task to understand how a change at the protein structure level eventually affects a cell's behavior. This is because complex information at both the protein and pathway level has to be integrated. Given that the idea of integrating both protein and pathway dynamics to estimate the systemic impact of missense mutations in proteins remains predominantly unexplored, we investigate the practicality of such an approach by formulating mathematical models and comparing them with experimental data to study missense mutations. We present two case studies: (1) interpreting systemic perturbation for mutations within the cell cycle control mechanisms (G2 to mitosis transition) for yeast; (2) phenotypic classification of neuron-related human diseases associated with mutations within the mitogen-activated protein kinase (MAPK) pathway. We show that the application of simplified mathematical models is feasible for understanding the effects of small sequence changes on cellular behavior. Furthermore, we show that the systemic impact of missense mutations can be effectively quantified as a combination of protein stability change and pathway perturbation.
| Small changes in protein sequences, such as missense mutations resulting from genetic variations in the genome, can have a large impact on cellular behavior. Consequently, numerous studies have been carried out to evaluate the disease susceptibility of missense mutations by directly analyzing their structural or functional impact on proteins. Such an approach has been shown to be useful for inferring the likelihood of a mutation to be disease-associated. However, there are still many unexplored avenues for improving disease-association studies, due to the fact that the dynamics of biological pathways are rarely considered. We therefore explore the practicality of a structural systems biology approach, combining pathway dynamics with protein structural information, for projecting the physiological outcomes of missense mutations. We show that stability changes of proteins due to missense mutations and the sensitivity of a protein in terms of regulating pathway dynamics are useful measures for this purpose. Furthermore, we demonstrate that complicated mathematical models are not a prerequisite for mapping protein stabilities to network perturbation. Thus it may be more feasible to study the systemic impact of missense mutations associated with complex pathways.
| How one links genetic information to physiological outcomes is an important issue in the current ‘post-GWAS’ (genome-wide association studies) era [1]. One specific topic regarding this problem is the functional annotation of non-synonymous single nucleotide polymorphisms (nsSNPs) that cause amino acid changes in proteins. However, the difficulty of annotating nsSNPs has slowed down the pace of investigating their molecular consequences. Therefore, as the speed of identifying new SNPs is high, there is now a distinct sense of urgency to resolve this problem – an immediate focus is the 1000 Genomes Project (http://www.1000genomes.org/) that has identified approximately 100,000 nsSNPs in need of further analyses. Indeed, the urgent requirement for SNP annotation has also motivated the CAGI experiment (Critical Assessment of Genome Interpretation; http://genomeinterpretation.org/) that encourages community-wide efforts in predicting the phenotypic impacts of genome variation.
Interpreting the physiological effect on cells due to missense mutations in proteins is not a simple task. This is partly achievable through analyzing the increasing number of protein structures deposited in the Protein Data Bank (http://www.rcsb.org/) and through functional annotation of proteins [2]. Investigating protein structures allows for a qualitative view of pathway dynamics; a more quantitative approach is to use mathematical modeling. Indeed, our understanding of cellular behavior during the last two decades has been significantly improved through the application of mathematical modeling methods such as ordinary differential equations (ODE) and rule-based simulations [3], [4], [5], [6], [7], [8]. However, the idea of integrating the dynamical aspects of proteins and their associated pathways to investigate the systemic impact of missense mutations is still in an early stage of development. In 2007 Stein et al. proposed the idea of integrating structural and pathway information for estimating key kinetic constants associated with biochemical pathways [9]. More recently, Kiel and Serrano [10] studied how missense mutations in the Ras-binding domain of c-Raf (RafRBD) affect the expression of the downstream protein Erk by investigating the structure of RafRBD and constructing an ODE model describing Erk signaling pathway [10].
The work of Kiel and Serrano suggests that integrating protein structural analysis with pathway modeling can be a useful method to facilitate the physiological annotation of missense mutations in proteins. However, the effectiveness of this approach at quantifying missense mutations located in different proteins remains unexplored. Also unexplored is the utility of this approach with simpler mathematical models, considering only the dynamics of key proteins while the remaining proteins in the pathway are omitted – this is potentially a more practical approach for achieving an improved inference of the parameter space, thereby increasing the reliability of the analysis (current ODE models describing biological pathways often contain tens or hundreds of parameters that can neither be easily measured nor calibrated experimentally). Furthermore, extensive investigation is required to determine how the approach performs when annotating missense mutations whose physiological outcomes can be clinically defined and examined.
These issues are discussed in this work by gauging the systemic impacts of missense mutations through integrating protein and pathway behavior via reduced ODE models. Here we present and discuss the measurement of a ‘systemic impact factor’ (SIF), defined as a function of free energy change (ΔΔG) and systemic control (CSpi, see Methods section ‘Control coefficient’), as a practical approach for evaluating the relative effects of missense mutations in a specific system. For mutations appearing in proteins whose complexed and uncomplexed states are both considered in the model, we calculate their maximum SIFs by taking the maximum ΔΔG between the two states. This is because the average score of the two protein states does not necessarily have a clear biophysical meaning in terms of describing the overall stability change of a mutation. Although summing the ΔΔGs calculated in the two protein states may have biophysical meaning, complications will be incurred when comparing the SIFs to other proteins that only have one conformational state analyzed in the model (either complexed or uncomplexed). Therefore, by using the maximum ΔΔGs we do not compromize the biophysical meaning of SIF and at the same time make the SIF scores more comparable across different proteins that may or may not have two states.
The benchmark includes two biological systems: (1) the fission yeast G2 to Mitosis (G2-M) transition and (2) the human MAPK signaling pathway. The first system is a well-defined system for studying the genotype-phenotype relationship as the systemic perturbation of missense mutations can be directly benchmarked to the length change of yeast cells. We use the temperature-sensitive yeast strains as experimental models, each of them containing a single missense mutation in protein Cdk1 or Cyclin B (CycB), and we measure their cell lengths at septation (septation is immediately followed by mitosis). Finally, the practicality of the SIF score in quantifying the systemic effect of missense mutations is evaluated by the correlation between the calculated SIF scores and in vivo cell lengths. The second benchmark system represents a more complex example in which the target mutations are spread within four different proteins (H-Ras, Raf-1, Braf and Me) and lead to clinical symptoms (in this case the neuro-cardio-facial-cutaneous syndrome) that have different prognoses and risk of complication. To determine whether or not a simple ODE model can be used to infer the systemic perturbation of missense mutations, we construct a reduced ODE model that includes only 12 parameters for the calculation of SIF values. We then place the mutations into subgroups according to their predicted SIF scores, and record whether our classification reveals the underlying difference between disease mechanisms.
The G2-M transition controls when a cell enters mitosis and determines the size of a cell at the point of division into two daughter cells. In fission yeast, Schizosaccharomyces pombe, this involves Cdk1, CycB, Wee1 and Cdc25. In the G2 phase, Cdk1 and CycB form a complex known as the mitosis promoting factor (MPF), which brings about the G2-M transition [11]. The activity of MPF is regulated by the protein kinase Wee1 [12] and the protein phosphatase Cdc25 [13], [14]: Wee1 inhibits the activity of MPF by phosphorylating Cdk1, and Cdk1 also exerts negative feedback on Wee1 by phosphorylating it. In addition, Cdc25 activates MPF by dephosphorylating Cdk1 and vice versa [15]. The Wee1-MPF-Cdc25 control system increases the ratio of active MPF over its inactive state and eventually promotes a cell into mitosis (Figure 1A).
The model we present here (Table 1) is based on the first realistic model of MPF activation published by Novak and Tyson [16]. Two basic assumptions of our model are 1) the total amount of Cdk1 (Cdk1T) present in the system is constant and in excess (far greater than the initial concentration of CycB) [17], and 2) all CycB forms a complex with Cdk1 immediately after it is synthesized since Cdk1 binds to CycB strongly and is in excess of CycB: that is, CycBT = MPF (active form of MPF)+preMPF (inactive form of MPF).
Investigation of the parameter space through the replica exchange Monte Carlo algorithm (see Methods section ‘Replica exchange Monte Carlo method’) shows that the parameters in our in silico model are confined to a small range (Figure S1A) and parameter variations do not change the general trend of the relation between the various rate constants (Figure S1B).
Here we consider each missense mutation as a perturbation to the wild-type status as described in the in silico model mentioned above, and the systemic impact of each mutation is projected as the extent that a mutation is likely to deviate from the wild-type state. In our model, entry into mitosis occurs when CycB reaches a concentration (dimensionless) equivalent to an active MPF concentration of 2.0. Assuming cells grow continuously and linearly in time during interphase, the systemic impact of perturbing each rate constant can be gauged through the change of CycB concentration when active MPF = 2.0: the higher (lower) the CycB concentration, the longer (shorter) the cell size at mitosis (Figure 1C). Mathematically, this is implemented by calculating the control coefficients CSpi that indicate the change of CycB concentration under a consistent amount of perturbation to each rate constant (Methods section ‘Control coefficient’). The sign of CSpi shows the direction of CycB concentration change: positive CSpi values correspond to an increase of CycB concentration, whereas a negative CSpi indicates the opposite. An overall view of CSpi calculated for the G2-M mechanism shows that perturbing the Cdc25-related rate constants has a larger impact on the shift of the MPF curve compared to perturbations to the Wee1-related rate constants (Figure S2). This suggests an asymmetric relationship between the positive and negative feedback loops on MPF activation controlled by Cdc25 and Wee1 respectively, which is in agreement with the recent paper by Domingo-Sananes and Novak [18]. Although the unequal impact between Wee1 and Cdc25 has not been confirmed, previous experimental evidence in Xenopus egg extracts [19], [20] suggests this may be the case.
The systemic perturbation of the G2-M transition (Table 2) is studied by examining the effect of four temperature-sensitive mutations in Cdk1 (all mutations except C67Y and G183E) and two temperature-sensitive mutations in CycB that attenuate the function of MPF. The effect of these mutations on protein stability or function is more pronounced when the temperature increases (as proteins are allowed a greater degree of movement). Phenotypically, these mutations allow cells to divide at a greater length than their wild type states when the temperature increases. The cell does divide with the non-temperature sensitive mutant in Cdk1 (C67Y and G183E), but at a smaller cell size.
The modeled structure of MPF shows that mutation G43E in Cdk1 is located at the interface of MPF subunits and thus is likely to have a significant effect on the stability of the MPF complex (see Methods section ‘Homology modeling of Cdk1, CycB and MPF structures’ regarding structural modeling). Mutations A177T, G183E and P208S in Cdk1 are located at or close to the active site and hence are likely to cause functional effects; C67Y and G227C in Cdk1 and W395R in CycB are at the periphery of the proteins and thus are mainly structurally related. Mutation C379Y in CycB is within a hydrophobic core and is likely to have a considerable impact on the MPF complex by destabilizing the structure of CycB (Figure 1B).
The link between SIF and systemic perturbation (SP) can be statistically established through regression:where ΔΔG is the free energy change caused by a mutation to a target protein (here FoldX [21] is applied to approximate the ΔΔG of the mutations studied), which approximates the change in a specific rate constant of the target ODEs (see Text S2 for further information regarding the application of ΔΔG as an evaluation for systemic perturbations); CSpi is the control coefficient (Methods section ‘Control coefficient’) that reflects how sensitive the concentration change of the reporter protein (in this case protein CycB) is to the specific parameter.
Hence for the G2-M model the magnitude of each SIF value indicates the degree of impact a mutation can have on the quantity of CycB, which determines when a cell enters mitosis and therefore the length of the cells.
The fundamental concept of our approach is to build a wild-type model that faithfully reflects in vivo cellular behavior and then considers each missense mutation as a perturbation to the wild-type status. We do not intend to formulate a model that describes the mutant-type states; we only project the extent that a mutation is likely to deviate from the wild-type state. The procedure of calculating SIF scores is shown in Figure 1. Firstly, a target biological system (in this case the G2-M transition in the cell cycle) is chosen and a reporter protein, whose expression profile can be used to gauge the systemic behavior, is identified. The reporter protein used here is the MPF protein complex. The mutations are then mapped onto three-dimensional protein structures and linked to the associated parameters in the ODEs. For each mutation, its ΔΔG is approximated as the size of perturbation introduced to the associated ODE parameter. To improve the estimation of ΔΔG for each mutation, we applied molecular dynamic (MD) simulations to sample the movement of the flexible regions in the modeled Cdk1 and CycB structures, and then calculated an average ΔΔG based on the sampled conformations (Methods section ‘Molecular dynamic simulations and free energy calculations’). Next, the sensitivity of the expression profile of the reporter protein to each ODE parameter is explored by calculating the CSpi. Finally, the systemic consequence of each mutation is inferred by calculating its SIF score based on ΔΔG and CSpi. In the case of the G2-M model, a larger SIF reflects a greater delay for a cell to enter mitosis. Hence a longer cell length should be observed.
For the eight missense mutation studies presented here, their SIF values are calculated (Table 3) and the length of their host yeast cells are measured at septation (Methods and Material section ‘Yeast strains and cell length measurement’). As shown in Figure 2, the in silico SIF score generally reflects the in vivo cell length well: at the semi-restrictive temperature (30°C) a medium-to-strong correlation R2 = 0.69 (p value = 0.04; all the p values shown in this study are based on the two-tailed model) is obtained.
To validate the function of our temperature-sensitive yeast strains, their lengths are also measured at the permissive temperature of 25°C: a condition that allows all the mutants and wild-type cells to grow normally, so the effect of mutation on cell length should be minimal. As shown in Figure 2, there is indeed a much smaller effect of the mutations on cell length at division and a weak correlation (R2 = 0.29) between SIFs and in vivo cell lengths was observed.
The MAPK pathway plays an essential role in cell survival, proliferation, differentiation and development (Figure 3A). Its three-tier MAPK cascade, i.e. Raf-Mek-Erk, is a highly conserved systemic structure that regulates the switch-like behavior of the pathway's signal transduction mechanism [22]. The important features of this cascade manifest themselves as representatives to evaluate the behavior of the parental pathway, and previous studies of the human MAPK pathway have shown analytical results which support this [23], [24]. To explore the effectiveness of a model that focuses on the dynamics of the three-tier structure, a reduced model is constructed here based on previous work that simulated the signaling cascade from the epidermal growth factor (EGF) receptor to the Erk reporter protein [25]. By omitting redundant terms whose removal has little effect on the expression curve of the downstream protein Erk, a set of succinct ODEs is derived as shown in Table 4 (the derivation is presented in Text S3).
To benchmark the behavior of both the reduced and original model, a sensitivity analysis is performed over three target quantities of the reporter protein Erk (Methods section ‘Quantifying the change of expression curves’): the amplitude (maximum activation), duration (time until signal drops down to 50% of its maximum activation) and peak time (time of maximum activation). For the test, the initial concentration of the key proteins in both models is varied and their effects on controlling the target quantities of Erk is compared (the key proteins include ShcGS (Shc: Src homology and collagen domain protein), GS, Grb2 (growth factor receptor binding protein 2), SOS (son of sevenless homologous protein), Ras, Raf, Mek and Erk). As a result, the control coefficients in both models demonstrate a similar pattern across all three-target quantities (Figure 4A,B), which indicate that the reduced model does not sacrifice the overall dynamics of the original model to achieve its simpler structure.
Here 40 mutations associated with neuro-cardio-facial-cutaneous syndrome are collected and studied (Table S1). As shown in Figure 3B, all the mutations can be mapped to crystal structures of H-Ras, Raf-1, B-Raf and Mek, and each mutation is classified as mainly functionally or mainly structurally important according to its location in the target protein.
Unlike missense mutations in the yeast G2-M model, there are no quantitative measurements of the physiological outcomes for the mutations in the MAPK pathway that can be used to calculate the correlation with SIF scores. Hence, as an indirect way to evaluate the relationship between mutations and clinical symptoms, each mutation is represented by three SIF scores calculated according to the systemic impact on the wild-type Erk expression curve: measured as amplitude, duration and peak time differences. The trajectory of the SIFs corresponding to each mutation as a function of these three target quantities shows that mutations in Raf1, B-Raf and Mek are more likely to be overlapped in a similar region, whereas mutations in H-Ras tend to distribute in a very different trajectory to the direction of the other mutations (Figure 5A). To determine if the different distribution of H-Ras mutations is a robust feature, a different set of initial concentrations that were measured experimentally in HeLa cells by Fujioka et al. [26] is used to derive two new parameter sets: one produces expression curves similar to those of the original model, whilst the other one produces curves fitted to the in vivo FRET data measured by Fujioka et al [26] (the parameters of both models are available in Text S1). As shown in Figure 5B and 5C, both parameter sets distribute H-Ras mutations in a trajectory different from other mutations, which suggests that the separation of H-Ras is not sensitive to variations to initial concentrations and parameter space. As a benchmark, the three dimensional SIF scores from the original model are also presented (Figure 5D). Consistently, H-Ras mutations are distributed into a distinctly different group.
It has been demonstrated that using ensembles of simulated protein structures, rather than a single conformation as represented by a crystal or modeled structure, can improve the estimation of free energy change [27]. In order to determine if the use of structural ensembles affects the SIF distribution, molecular dynamic (MD) simulations are also applied to sample the movement of flexible regions in the key kinases. Eventually an average ΔΔG, and therefore an average SIF score, was calculated for each mutation based on the alternative structures sampled by the MD simulations (Methods section ‘Molecular dynamic simulations and free energy calculations’). By using the average SIF scores calculated over the conformation ensemble, a less narrow distribution for B-Raf and Raf-1 mutations is observed in the reduced model with parameters fitted to the experimental FRET data while the distribution of mutations in other models remained largely unchanged (Figure S4). Moreover, the overall distribution of the SIF scores in all of the models is in agreement with the results using only the crystal structures. This suggests that the SIF scores are not overly sensitive to movements away from the experimentally determined atomic positions.
A closer examination of mutant SIF scores reveals that H-Ras mutations perturb the MAPK pathway in a distinctly different manner from that of the mutations in Raf-1, B-Raf and Mek (Figure 5A–D): H-Ras mutations tend to dominantly affect the duration of the Erk expression profile whereas the other mutations mainly affect the amplitude of the expression profile, followed by a smaller impact on peak time and an even smaller effect on the duration of the Erk activation. Physiologically, this indicates that the cellular response to H-Ras mutations is different to the other mutations. Indeed, the duration of Erk activation is known to be a critical factor for determining cell fate: in PC12 cells, it has been shown experimentally that prolonged activation of Erk is sufficient for cell differentiation whereas transient activation is associated more closely with cell proliferation; in fibroblast, a reverse relationship between duration and cell fate is observed [28]. Although the amplitude of Erk activation has also been experimentally shown to be a determinant of cell fate, its effect is more complicated: high level of Erk activation usually promotes cell-cycle progression but sometimes it leads to cell-cycle arrest as well [28]. Also, the mutations in Raf-1, B-Raf and Mek mainly reduce the amplitude of Erk expression and hence it is likely that they have less effect on cell growth than H-Ras mutations, which mainly increase the duration of Erk expression.
In this work we presented the SIF function as an effective measure for the systemic impact of missense mutations. SIF values reflect in a simple manner the fact that proteins are functional units in the cell whose interaction networks regulate cellular behavior. It is of particular interest to see that SIF scores reflect the in vivo phenotype in the yeast G2-M model when there is no additional parameter introduced to distinguish functionally and structurally important mutations. This suggests that, although they change protein behavior in different ways, functional and structural mutations can perturb a pathway to a similar extent.
A potential way to improve the current correlation between SIF and systemic outcome is to consider an additional parameter λ that describes the amount of parameter perturbation caused by free energy change. Now the SIF function becomes:By assigning different λ constants for functional and structural mutations in the G2-M model, we found that using a larger λ for functional mutations consistently provides smaller correlations (less than 0.68). This suggests that ΔΔG in this case over-estimates the systemic impact of functional mutations and thus should be scaled down by a smaller λ when used for analyzing mutations at functional sites. This also indicates that functional mutations may be better annotated by considering other protein-protein interactions besides protein stability. However, this would make it much more difficult to quantify the impact of protein interactions. Using a smaller λ for functional mutations may be suitable for the mutations studied here; nevertheless further investigation is required to determine if this is a general criterion that could be applied for mutations in other biological systems.
When we considered only structural mutations in the G2-M model, the correlation between SIF and cell length increases from 0.69 to 0.73 (p value = 0.026). This suggests that the current SIF formula may perform much better in annotating the systemic effect of mutations whose role is more structural than functional. This could be due to the way we approximate the functional impact of a missense mutation through Michaelis constants and link its perturbation to ΔΔG as an approximation of Kd (Text S2).
Although the current SIF function correlates linearly with in vivo measurements, the data cannot rule out an exponential relationship between SIF and phenotypic outcome. As described in Text S2, if we broadly approximate the amount of perturbation in each rate constant to be , SIF can be formulated as ∼To approximate the direct use of ΔΔG, we may transform the SIF function to . Following this formula, the correlation between ln(cell length) and ln(SIF) is reasonable: 0.62. Further studies will be required to explore the optimal correlation between SIF and systemic effects.
A very intriguing result of this study is that systemic impacts can be reasonably gauged through simple or reduced ODEs. This indicates that it is possible to study the systemic perturbation of a pathway when there is incomplete information about its components – an important observation, given the fact that the majority of biological pathways have missing components waiting to be discovered or confirmed. Another import aspect of this work is that, for the purpose of studying systemic perturbation, it is feasible to study the missense mutations through “fuzzy” parameters – that is, the systemic impact of a mutation can be extrapolated through rate constants that account for general protein-protein interactions rather than detailed enzyme catalytic reactions. Finally, the advantage of using a simpler model is also reflected in facilitating a lower chance of associating multiple parameters with a perturbation, which means the difficulty of discussing the impact of a missense mutation can be reduced.
The simplicity of the G2-M model lies in two aspects. First, it has only four major component proteins (Cdk1, CycB, Wee1 and Cdc25) used to simulate cell growth, and the model can be considered to be linear, terminating when MPF reaches a certain critical concentration. The second aspect is that, rather than capturing their time-course data, the model reflects the relationship between the component proteins. Normally this raises the difficulty of parameter optimization, as it increases the chance of converging to multiple parameter sets that all give simulation curves satisfying a particular phenotypic outcome. Fortunately, parameter inference is not a concern in this case, since the general trend of the Cspi relation between parameters is conserved, regardless of parameter variations (Figure S1B), i.e. the correlation between the SIF values and in vivo cell lengths is not sensitive to parameter variation.
In preserving the overall dynamics of the original model (Figure S5), our reduced MAPK model is efficient in terms of parameterization; it has only 12 rate constants, compared to the original 27. The simplified ODEs allow us to conduct a straightforward analysis on missense mutations, which may not be the case in a more complicated model. For example, a mutation in the functional site of Ras can affect two downstream interactions in the original model (see Figure S5: one is between Ras and Gap; the other is between Ras and Raf), whereas it can only affect the interaction between Ras and Raf in the reduced model. Furthermore, without reduction we would not be able to implement the robustness test on the SIF projections shown in Figure 5, since it is most unlikely that one could obtain robust parameters given the expression data profile from Fujioka et al [26]. The practicality of a simpler model suggest that the idea of using ODEs to model the dynamics of a pathway can be more feasible than previously thought, as long as we can reduce a complex pathway to smaller modules that account for the functional core of a pathway.
It is generally non-trivial to infer cellular phenotypes from studying pathway dynamics since many cellular functions have complex underlying mechanisms. However, the medium-to-strong correlation between the SIF values and in vivo yeast cell lengths in the G2-M model shows that it is possible to estimate effectively the phenotypic effect of missense mutations through gauging systemic impacts. This is due to two essential factors underlying our G2-M model. Firstly, cell length at septation (cell division) is a faithful indicator for identifying cells at the beginning of mitosis. This is because fission yeast grows only in length and thus it can be positioned in its cell cycle simply by its length and does not grow between entry into mitosis and septation. Secondly, the chosen reporter protein, MPF, is closely linked to the initiation of mitosis. A strong support for this is a recent discovery that MPF is a necessary and non-redundant factor for triggering mitosis [29].
The SIF values simulated from the MAPK model, on the other hand, reflect a more complex relationship with phenotype. We expected that most of the mutations studied here should be projected into similar regions, as they are associated with overlapping symptoms under a broad term ‘neuro-cardio-facial-cutaneous syndrome’. However, H-Ras mutations are projected into distinctly different trajectories from the other mutations with respect to their effects on the ERK expression profile. This suggests that H-Ras mutations are likely to have different characters in terms of the disease prognosis and risk of complications depending more upon the genotype than on the phenotype. Given the clinical symptoms of patients from which the missense mutations studied here were identified (as shown in Figure 5E, all the H-Ras mutations are associated with Costello syndrome (CS); most of the Raf-1 mutations are associated with Noonan syndrome (NS); most of the B-Raf and all of the Mek mutations are associated with cardio-facio-cutaneous syndrome (CFCS)), the result in Figure 5A–5D suggests that NS and CFCS may share some degree of similarity in terms of disease development. Indeed, it is often difficult to distinguish an infant with CFCS from NS, although the phenotype becomes more distinctive with time [30]. Interestingly, current knowledge of the genotype-phenotype correlations suggests that the presence of mutations in the H-Ras gene is associated with a much higher tendency of cancer compared to the other mutations [31], indicating a potentially different system dynamic, as indeed demonstrated in this study. As a whole, the MAPK model serves as a good example to show how qualitative annotation of mutations (the classification of mutations) can contribute to the understanding of disease mechanisms. This is practically useful as it is often hard to clinically quantify various disease phenotypes that lead to differences in prognosis and drug response.
The two systems in our study show that SIF can reflect phenotype or the underlying mechanism of missense mutations in proteins. In general, we may confidently interpret systemic impacts as an indicator for phenotype only if a reporter protein is strongly and non-redundantly linked to a target phenotype; otherwise a more reserved view would be appropriate.
One confounding factor associated with the performance of SIF is the relationship between ΔΔG of a mutation and its actual phenotypic effect. This is because different proteins may have different stability states and hence they may respond differently to the same amount of ΔΔG caused by missense mutations. The issue of benchmarking the effect of ΔΔG on different proteins has been an active topic in annotating nsSNPs. Previous studies show that proteins belonging to different structural families can respond differently to the same amount of ΔΔG, but in general a small margin of ΔΔG (1–3 kcal/mol) can be approximately used to define missense mutations that may not cause an immediate effect on protein fitness [32], [33], [34]. On the other hand, for proteins that share Immunoglobulin-like folds, a clearer phenotypic threshold of ΔΔG (2 kcal/mol) can be used to define missense mutations that generally result in severe disease phenotypes [35], [36]. Hence, taking a more stringent view, this implies that proteins sharing similar structures are more likely to react similarly to mutations that cause the same amount of ΔΔG.
For the proteins studied in this work, the concern of comparing the effect of ΔΔG across different proteins is likely to be alleviated due to the above reasoning. In the G2-M model, CycB and Cdk1 form a complex and hence the uncertainty of comparing ΔΔG in two different proteins is reduced. In the MAPK model, all the key proteins are kinases that share the same well-structured fold.
Another factor that may affect the performance of SIF is the complication of assigning the role of a mutation as mainly functional or structural. This issue is especially hard to deal with if a missense mutation is likely to cause long-range structural effects on its host proteins - for example, a mutation can exist far away from a functional site (and thus is considered as a structural mutation) but still affect the function of its host protein by inducing long-range conformational changes. Hence additional attention should be paid to calculating SIF for mutations located in proteins that are not well studied or have versatile conformations. For the cases studied in this work, the problem of assigning functional and structural mutations is not significant because most of the key proteins are kinases that have well-defined functional sites (see Text S2 for further information).
One other factor that is associated with SIF performance is the accuracy of calculating ΔΔG. So far most of the methods for predicting ΔΔG do not show a good correlation with the experimental ΔΔG; however, they do perform well when used to estimate the average effect of mutations on protein stability [37]. This is likely to support the good correlation between SIFs and the in vivo cellular phenotypic outcomes measured in our study, since we calculated an average ΔΔG for each missense mutation based on the simulated structures and used it to correlate with the experimental data.
Finally, it is worth mentioning that the performance of SIF can be considerably compromised by mutations with large ΔΔG values. These mutations can be too extreme to be considered a perturbation to a target system, and hence the ODE model describing the wild-type condition is not applicable. On the other hand, large ΔΔG values can also be the result of Van der Waals clash that are often heavily penalized in ΔΔG calculations (as likely the case for mutations G43E and C379Y in the yeast G2-M model). All in all, in the cases where ΔΔG is large, caution should be taken when applying the SIF function.
Our study as a whole suggests that it is beneficial to combine multi-level knowledge to investigate the effects of missense mutations on cellular behavior. The advance in protein structure prediction techniques will particularly make the calculation of SIF more feasible, since it requires the structural information of proteins that host the target missense mutations. Overall, there is sufficient reason for us to be confident that future studies on integrating protein and pathway dynamics will become increasingly viable, as there are constant efforts across the scientific community in solving protein structures and identifying new components in biological pathways.
Simulating pathway dynamics through ODEs, as demonstrated here, provides a convenient platform for utilizing the information on protein structures. However, the application of ODEs implies two major limitations. One is in the availability of time-course data of protein expression in public resources; at the moment this is relatively low and sparse compared to that of gene expression data. This will be alleviated as more high-throughput time course data becomes available. The other limit is in our knowledge of the biological pathways – a majority of them have only been partially uncovered. A feasible way to circumvent the problem is to develop a simpler model by considering only key proteins that are essential for preserving pathway behaviors, as we have demonstrated in the case of MAPK pathway and G2-M transition.
The SIF function in its current form gives a good approximation of systemic perturbation resulting from the missense mutations in the G2-M and MAPK models. With further development on a larger dataset, especially with the inclusion of more parameters to further characterize protein function and structure, we are likely to obtain better correlations with quantitative phenotypes. The process of refining the SIF equation will tell us more about the relationships between protein function and structure, and pathway dynamics, which is one of the most important questions considered by structural biologists.
The advance of high throughput technology has enabled us to identify mutations in a large number of inter-connected pathways. It is becoming apparent that performing experiments to check the impact of individual mutations on the pathway level will be extremely time-consuming and costly, let alone monitoring all the possible cross-interactions and combinatorial effect of multiple mutations. From this perspective, multi-level mathematical modeling, such as that described here, will provide an efficient mechanism for pre-screening systemic impact in a cost-effective way. This is particularly useful for studying the etiology of complex diseases that are usually the result of accumulating multiple mutations.
Yeast strains used in this study are listed in Table 3.
All the strains except strain 4932 were generated following our protocol previously published by Nurse et al. [38]. Strain 4932 was generated and identified as described in the work of Fong et al. [39] with the following changes: Genomic DNA from a cdc13hph tagged strain was used as the starting template. TaKaRa LA-Taq polymerase (Takara Bio) was used for the first round of PCR and Z-Taq (Takara Bio) for the mutagenic PCR reaction that was supplemented with 10XdGTP. Mutation positions were identified using Big Dye (Applied Biosystems) terminator cycle sequencing.
Cells were grown to mid exponential growth (∼5×106 cells/ml) in rich media at 25°C and 30°C [40] and photographed using a Zeiss Axioplan microscope. Cell lengths upon mitosis, by unbiased sampling of 30 septated cells, were measured using ImageJ.
Here we applied the replica exchange Monte Carlo method (REM) – also known as parallel tempting (PT) – to implement parameter inference. For a non-linear system, as represented by the G2-M and MAPK model, the energy surface is normally rugged and it is hard to ensure unbiased sampling along the uneven energy space. Nevertheless REM has been shown to be very useful for this purpose, especially at low temperatures, and has been used extensively for finite-temperature simulation of biomolecules [41], [42]. The general idea of REM is to simulate a number of subsystems {X(m)} with different inverse temperatures βm (replicas) in parallel. At particular intervals, the sampling trajectory is exchanged from one subsystem to the others (usually adjacent replicas) with the following probability specified in [43]:where Δβ = βm+1−βm is the difference between the inverse temperature of the two replicas and ΔE = E (X(m+1))−E (X(m)) is the energy difference between them (in our case the deviation of the protein expression time course).
Practically, the exchange of replicas with different temperatures effectively generates repeated heating and annealing cycles, which avoids the parameter search from becoming trapped in a local energy minimum.
For sampling the trajectories, PEPP used the Metropolis algorithm [44] with modifications that allow uneven sizes of sampling steps. This echoes the idea that the coexistence of large and small changes in phase space is essential for sampling unstable structures [45]. To determine the size of each sampling step (Δx), PEPP adopted the method introduced in [45]:where d and e are random integers uniformly distributed in d[1,9] and e[, ] and s is a binary random number that is either 1 or −1 with probability of 0.5. and determine the logarithmic scales of the smallest and largest step move, respectively. As a result, the overall sampling density is a mixture of uniform distributions with different scales; it has a sharp peak near zero and very long tails.
The iteration of the Metropolis algorithm in our model is as following:
|
10.1371/journal.pcbi.1002546 | Prosthetic Avian Vocal Organ Controlled by a Freely Behaving Bird Based on a Low Dimensional Model of the Biomechanical Periphery | Because of the parallels found with human language production and acquisition, birdsong is an ideal animal model to study general mechanisms underlying complex, learned motor behavior. The rich and diverse vocalizations of songbirds emerge as a result of the interaction between a pattern generator in the brain and a highly nontrivial nonlinear periphery. Much of the complexity of this vocal behavior has been understood by studying the physics of the avian vocal organ, particularly the syrinx. A mathematical model describing the complex periphery as a nonlinear dynamical system leads to the conclusion that nontrivial behavior emerges even when the organ is commanded by simple motor instructions: smooth paths in a low dimensional parameter space. An analysis of the model provides insight into which parameters are responsible for generating a rich variety of diverse vocalizations, and what the physiological meaning of these parameters is. By recording the physiological motor instructions elicited by a spontaneously singing muted bird and computing the model on a Digital Signal Processor in real-time, we produce realistic synthetic vocalizations that replace the bird's own auditory feedback. In this way, we build a bio-prosthetic avian vocal organ driven by a freely behaving bird via its physiologically coded motor commands. Since it is based on a low-dimensional nonlinear mathematical model of the peripheral effector, the emulation of the motor behavior requires light computation, in such a way that our bio-prosthetic device can be implemented on a portable platform.
| Brain Machine Interfaces (BMIs) decode motor instructions from neuro-physiological recordings and feed them to bio-mimetic effectors. Many applications achieve high accuracy on a limited number of tasks by applying statistical methods to these data to extract features corresponding to certain motor instructions. We built a bio-prosthetic avian vocal organ. The device is based on a low-dimensional mathematical model that accounts for the dynamics of the bird's vocal organ and robustly relates smooth paths in a physiologically meaningful parameter space to complex sequences of vocalizations. The two physiological motor gestures (sub-syringeal pressure and ventral syringeal muscular activity), are reconstructed from the bird's song, and the model is implemented on a portable Digital Signal Processor to produce synthetic birdsong when driven by a freely behaving bird via the sub-syringeal pressure gesture. This exemplifies the plausibility of a type of synthetic interfacing between the brain and a complex behavior. In this type of devices, the understanding of the bio-mechanics of the periphery is key to identifying a low dimensional physiological signal coding the motor instructions, therefore enabling real-time implementation at a low computational cost.
| The complex motor behavior originating the rich vocalizations of adult oscine birds results from the interaction between a central pattern generator (the brain) and a nonlinear biomechanical periphery (the bird's vocal organ) [1], [2]. The fact that this complex behavior is learned, together with the parallels between the physical mechanisms of birdsong and human speech production, make birdsong an ideal model to study how a complex motor behavior is acquired, produced and maintained [3].
In an effort to understand what gives rise to complexity in this behavior, a part of the birdsong community has set focus on the capabilities of the periphery to produce vocalizations owning a diverse set of nontrivial acoustic features [2]. The avian vocal organ, comprised mainly by the respiratory system, the syrinx and the vocal tract, is a highly nonlinear biomechanical device. The complexity of its dynamics leaves traces in the sounds that can be produced in it. In this way, several acoustic features found in vocalizations can be related to nonlinear phenomena occurring in the syrinx [4], [5] or introduced by acoustic interactions between the syrinx and the tract [6]–[9]. In all these cases, the complexity of the behavior does not require a complex motor pattern to drive the vocal organ, but rather simple, smooth gestures.
Through a combination of experimental observations and theoretical analysis, low-dimensional mathematical models have been proposed that account for the physical mechanisms of sound production in the avian vocal organ [10], [11]. In particular, a model based on Titze's proposed flapping mechanism for oscillations in human vocal folds [12] was recently used to synthesize the song of the Zebra finch (Taeniopygia guttata) [13], [14]. This model captures the nonlinear dynamics of the folds oscillating to produce sound, in a way that a variety of complex vocalizations are generated by the tuning of parameters related to physiologically observable motor gestures elicited by the bird.
Part of the appeal of counting with this model is the prospect of applying it to the construction of a bio-prosthetic device. In this scenario computation is relatively inexpensive because of the low dimension of the mathematical model. In addition, the physical description of the peripheral effectors led to the identification of a set of smoothly varying parameters that determine the behavior [13]. By recording the physiological activity related to the parameters and feeding it to a device that solves the equations of the model in real-time, vocal behavior can be emulated by a prosthesis controlled by a subject via its motor instructions.
The usual strategy of BCIs and BMIs (Brain Computer Interfaces and Brain Machine Interfaces) is to decode motor commands from recordings of physiological activity in the brain and use this activity to control bio-mimetic devices [15]–[17]. In [17], for instance, multi-electrode recordings of tens to hundreds of neurons in different cortical areas of primates are used to drive a robotic arm. In recent work, Cichocki et. al. discuss the perspectives of using electroencephalographic (EEG) recordings to generate noninvasive BCI solutions [16]. In these examples (as well as in many other BCI implementations), the crucial problem is the classification of the features of the large data set which correspond to a determined set of motor tasks. The feature extraction is performed by different techniques which include linear decomposition in a diversity of vector spaces and machine learning algorithms [15], [16], [18]. In this way, accurate control of bio-mimetic effectors is achieved for a finite number of specific tasks, such as grasping or cursor moving.
Our current understanding of the biophysics of the avian vocal organ, particularly our capacity to identify the dynamical mechanisms by which complex behavior occurs when the peripheral systems are driven by low dimensional, smooth instructions, allows us to propose an example of a different kind of bio-prosthetic solution. The model predicts a diversity of qualitatively different solutions to the system for continuous paths in a parameter space. Not only is this parameter space suggested by the model, but it is also physiologically pertinent.
We present a device that is driven by a freely behaving Zebra finch to produce realistic, synthetic vocalizations in real-time. The device is based on the real time integration of the mathematical model of the vocal organ on a Digital Signal Processor (DSP). It is controlled by the bird's subsyringeal air sac pressure gesture, which is transduced, digitized and fed to the DSP to provide the model with the appropriate path in parameter space.
The work is organized as follows. In the Methods section we describe the Zebra finch vocal organ and the mathematical model that accounts for its dynamics. We also discuss its applicability to the construction of a model-based bio-prosthetic device, and introduce the physiological motor gestures that relate to parameters of the model. We present the steps leading to the real-time implementation of the model on a device controlled by a spontaneously singing bird. In the Results section we show the example of a successful case. Finally, we summarize the results and discuss the impact of this device as an example of a kind of bio-prosthetic device enabled by the low-dimensional dynamical model of the peripheral effector.
All experiments were conducted in accordance with the Institutional Animal Care and Use Committee of the University of Utah.
One of the most studied species of songbirds is the Zebra finch. Its song presents a set of diverse acoustic features, which can be accounted for by the dynamics displayed by the mathematical model of its vocal organ [5], [11]. This low-dimensional mathematical model, when driven by the appropriate gesture in parameter space, is capable of producing realistic, synthetic birdsong. By implementing this model on a Digital Signal Processor, we are able to construct a bio-prosthetic vocal organ.
The mathematical model for the vocal organ, the reduction of the system ruling the dynamics of the sound source, and the identification of the pertinent parameters accounting for its motor control, they all make way for the construction of a bio-prosthetic device. The parameters determining acoustic properties of vocalizations in the normal form (5) are physiologically meaningful and the set of differential equations is easy enough to compute in a portable platform such as a Digital Signal Processor. By fitting the parameters and integrating the system in real-time, synthetic song can be produced in a device controlled by the motor instructions elicited by a freely behaving bird.
In many bio-prosthetic solutions, the physiological motor gestures used to drive the device are degraded respect to those recorded in the intact subjects [15]. One application of this device is the performance of altered auditory feedback experiments [32]–[34]. In experiments enabled by this device, the bird's own auditory feedback can be replaced by synthetic birdsong computed in real-time. Since the synthetic feedback is produced by the integration of the model when fed with actual physiological motor gestures elicited by a freely behaving bird, alterations of the feedback are possible that are consistent with alterations in the motor gestures intended to produce them. Here, the prosthetic vocal organ is driven by a bird that is muted via the insertion of a cannula through its inter-clavicular air sac. Phonation is prevented as the airflow is bypassed away from the syrinx. As a side effect, the pressure pattern registered on the muted bird differs from that recorded in the intact animal.
The device reconstructs the intended motor gesture from this degraded pressure gesture to trigger integration of the model. When the pressure pattern corresponding to the syllables comprising a motif are identified, the mathematical model for the vocal organ is computed with the appropriate paths in parameter space, to produce the corresponding synthetic output.
The device succeeds in synthesizing song online when driven by the pressure gesture of a muted bird. From the altered motor gesture, the algorithm infers the segment of a motif intended by the bird and computes the model to produce the vocalizations. An example is illustrated in Fig. 5.
The upper panels of the figure display the recorded subsyringeal pressure and sonogram of a segment of a bout with its preceding introductory call. A song bout of this bird (B06) is composed of a number of introductory notes (O) and the repetition of a simple motif containing two syllables (A and B), indicated by different colors and opacity of shading in Fig. 5. An initial segment of syllable A (marked with clearer shading in the figure) is used to detect the intention to elicit a motif in the pressure gesture of the muted bird. The part that is produced upon triggering of the synthesis appears in a darker shade of the same color. These recordings were used to fit the parameters of the model to produce synthetic vocalizations showing a match in fundamental frequency and spectral content.
The bird is then muted and placed in the setup to drive the electronic vocal organ with its pressure gesture. In the lower panel of Fig. 5 we show the pressure gesture of the muted bird and the sonogram of the sound recorded by a microphone placed close (about ) to the bird and the speaker. It can be noted in the sonogram that no sound is produced during the bird's attempts to phonate an introductory note. When the pressure motor gesture corresponds to the first syllable in the bout (syllable A), the instruction is recognized and the corresponding song is synthesized and played through the speaker. In the first bout, the bird only elicits the gesture corresponding to the first syllable, and then stops. In this case the algorithm detects the interruption and turns off the integration. In the second bout, the bird continues with the second syllable and drives the electronic syrinx to the end of the motif.
This example illustrates how this device works, and shows that it is successful in synthesizing the song motif as the bird drives it. We evaluate its success by counting the times the motif was properly detected and synthesized, and how many times a false trigger occurred. During a session of hours (the second day after the muting took place), a muted bird elicited about calls, out of which less than generated false triggering of synthetic song. In most cases the false trigger event was recognized by the algorithm and silenced after less than . The rate of success in detecting the beginning of a bout was of about in attempted bouts elicited by the bird.
Despite the variability of the altered pressure gesture in the subsequent days ( days after the muting), a brief calibration before each daily session allowed the rates of success and false triggers to be maintained. To do this calibration, the pressure gesture was recorded during and these data were used to re-set the cross-correlation thresholds, while keeping the test segment of the intact pressure gesture previously selected.
We show here that realistic vocal behavior is synthesized in real-time by our device, as it is controlled by the spontaneous behavior of a muted bird, a physiological signal (its air sac pressure) that is degraded in respect to the one recorded in the intact bird. The computing platform is a low cost, portable processor, and the initial rate of success is high. This is an encouraging example of the plausibility of a kind of interface between the central motor pattern generator and the synthetic, bio-mimetic behavior. DSP technology is being implemented in a variety of biologically inspired problems, and together with Field Programmable Gate Array technology (FPGA) is likely to become a standard solution for a variety of bio-mimetic applications [15], [35].
Brain computer and brain machine interfaces (BCI and BMI) typically read physiological data and attempt to decode motor instructions that drive peripheral devices in order to produce synthetic behavior [15], [17], [35], [36]. Because we have a physical model of the peripheral effector, the origins of the complexity of the behavior were linked to smooth paths in a low dimensional parameter space. Following the identification of the pertinent parameters and their physiological link to the pattern generator (activity of the ventral syringeal muscle and sub-syringeal air sac pressure), a further simplification of the system was carried out, on dynamical grounds, by eliminating irrelevant nonlinear terms (performing a reduction of the model to its normal form). This led to the possibility of implementing our bio-prosthetic device on a programmable electronic platform. Since the computing capabilities of the platform are greatly enhanced by the low costs of our implementation, technological advances in this front will have great impact on the complexity of the peripheral biomechanical system that can be emulated.
We have built a device that emulates complex motor behavior when driven by a subject by its actual (yet degraded) physiological motor gestures. It successfully reproduces the result of the stereotyped motor gesture that leads to the behavior, i.e., the diverse and complex set of sounds comprising the bird's song bout. The realistic synthetic vocalizations are produced in real-time, by computing a mathematical model of the vocal organ on a portable Digital Signal Processor.
The relative computational and technological simplicity of the device relies on the current level of understanding of the peripheral biomechanical effector [1], [2], [12], [24]. We have been able to construct a physical model of the syrinx that presents the adequate level of description and converges to a low dimensional (as low as two dimensions) dynamical system. The deterministic model of the vocal organ on which our device is based is not a statistical attempt to capture causal relationships between motor commands and behavior; instead, it is a hierarchization of the interactions within the biomechanical periphery and with the pattern generator. It aims to identify the dynamical mechanisms by which the behavior is produced.
Furthermore, exploration of the model leads to the finding that much of the diversity and complexity of the behavior can be explained in terms of the dynamical features of this nonlinear system [5], [28], requiring only simple instructions of the nervous system to produce a rich variety of vocalizations. Just as it is identified that a low dimensional system reproduces the main features of the complexity of the vocal organ, it can also be concluded that the control parameters are few and their behavior is simple (i.e., the physiological motor gestures linked to the paths in space are smooth). In this way, the parameter space of the model not only suggests the pertinent physiological instructions determining the main properties of the output but also how they are expected to behave. Paths in parameter space reconstructed in order for the model to produce vocalizations matching experimental recordings are indeed effective in predicting the physiological motor gestures [14].
In addition, knowledge of nonlinear dynamics allows us to find the simplest system with equivalent oscillatory behavior. The reduction of the low dimensional mathematical model for the syrinx to its normal form reduces the computational requirements and makes way for the implementation on a real-time computing solution, such as a DSP.
Realistic vocal behavior is synthesized online, controlled by the motor gesture of a freely behaving muted bird, which is a physiological signal that is degraded respect to the one recorded in the intact bird. This was achieved by computing in real time a mathematical model describing the mechanisms of sound production in the interface between the motor pattern generator and the behavior, the highly nonlinear vocal organ. The computing platform is a low cost, portable processor. This successful avian vocal prosthesis is an encouraging example of the plausibility of a kind of interface between the central motor pattern generator and the synthetic, bio-mimetic behavior. An advance towards models in which certain complex features of the motor behavior are understood in terms of the underlying nonlinear mechanisms of the peripheral effectors has the potential to enhance solutions of brain-bio-mimetic effector interfaces in many ways.
|
10.1371/journal.pcbi.1004517 | A Generative Statistical Algorithm for Automatic Detection of Complex Postures | This paper presents a method for automated detection of complex (non-self-avoiding) postures of the nematode Caenorhabditis elegans and its application to analyses of locomotion defects. Our approach is based on progressively detailed statistical models that enable detection of the head and the body even in cases of severe coilers, where data from traditional trackers is limited. We restrict the input available to the algorithm to a single digitized frame, such that manual initialization is not required and the detection problem becomes embarrassingly parallel. Consequently, the proposed algorithm does not propagate detection errors and naturally integrates in a “big data” workflow used for large-scale analyses. Using this framework, we analyzed the dynamics of postures and locomotion of wild-type animals and mutants that exhibit severe coiling phenotypes. Our approach can readily be extended to additional automated tracking tasks such as tracking pairs of animals (e.g., for mating assays) or different species.
| The roundworm Caenorhabditis elegans is a widely used model organism. Its locomotion, for instance, enables the study of genetic and cellular mechanisms that underlie behavior and may be broadly conserved. Characterizing C. elegans locomotion requires identifying its body posture and tracking how posture changes with time. Existing machine vision approaches have greatly aided this effort. However, they have been limited in cases where the body of the animal curved strongly such that one part of the animal rested or slid against another part. We present a method for automated detection of such complex body postures and its application to the analysis of locomotion. At the core of our method are progressively detailed statistical models of the shape of the animal. These models enable us to assess the probability that a given image contains a suggested posture. Our approach does not require manual initialization and can be readily parallelized for large-scale applications. We used our approach to analyze locomotion in mutants that severely exaggerate their body bends, called coilers. This approach can readily be extended to additional automated tracking tasks such as pairs of interacting roundworms or different organisms.
| The nematode Caenorhabditis elegans is a simple animal model system, widely used to study the genetic foundations of behavior. Among its key advantages are its tractable genetics, short life cycle, relatively simple anatomy and behavior patterns, and evolutionary conserved pathways [1–3]. The locomotion patterns of C. elegans have been extensively studied. Historically, this was largely done relying on visual phenotyping.
In recent years, several machine vision tools have been developed for automated posture analysis, collectively referred to as “trackers”, and spanning a range of capabilities [4–10]. Accurate identification of head and tail and reconstruction of the midline of the body are important steps in automated analyses of C. elegans postures. Typically, the topological genus of images of wild-type animals is zero, i.e., the body image only rarely forms closed loops. However, loops are observed in coiling mutants and more rarely in wild-type or other mutants.
Existing trackers were rarely used to automate the study of severe coiler phenotypes, plausibly because in such cases they either require frequent manual intervention or may misidentify the posture [11–14]. Therefore, we refer to such non-self-avoiding postures as complex.
Generative statistical models describe the expected images given a particular posture. This expectation is formulated in terms of a probability distribution, referred to as the data or likelihood term [15]. Knowledge about the data, such as expected body length or smoothness, is accounted for by specifying an a priori distribution of postures. The algorithm then optimizes the posterior probability of the posture, i.e., the product of the likelihood and the prior term. This framework can enable the identification of complex postures.
Here, we present a posture detection method based on generative statistical models and a coarse-to-fine strategy. Our approach allows a computationally efficient implementation and yields reliable detections of many complex postures. First, a small set of characteristic features for the head and body regions is defined as functions of oriented edges in the image. Next, we formulate a statistical model describing the likely configurations of these features given a hypothesized posture. At run time, we search for the maximum a posteriori posture of the worm given the observed image. This calculation yields a coarse outline of the animal.
To refine the outline, a second statistical model is employed which directly uses the edge information in the image and hence enables more precise identification. The advantage of this coarse-to-fine technique is computational efficiency. The coarse search runs over a grid which is much smaller than the original image grid. The fine search is then initialized using the result of the coarse detection and is only required to explore a small subset of possibilities. Searching for the posture at the fine scale without considering the information obtained from a coarse search would have been computationally intractable. To demonstrate the utility of our method, we assayed wild-type animals and several mutant strains that were previously associated with a coiler phenotype.
A coarse statistical model was defined to identify approximate positions of the head, tail and midline of the animal. The model is based on prominent features of the head and the body that are identified on a coarse-grained grid H, in which every point corresponds to a block of pixels in the original image grid G. Through trial and error, we found that a coarse grid unit length of one quarter of the width of the worm offered a good tradeoff between efficiency and accuracy. The key components of this model and the resulting detection are descried below.
The estimated point sequence θ→ on the coarse grid facilitates finding the boundaries of the animal body at the resolution of the original grid, from which a refined midline can be derived. The key components of this model and the resulting detection are described below.
To test the proposed algorithm we assayed wild-type animals and mutants that were previously reported to exhibit coiling. Of particular interest were three strains with severe locomotion defects. The Gq protein alpha subunit ortholog, encoded by egl-30, was shown to affect locomotion, viability, egg laying, synaptic transmission, and pharyngeal pumping [16–28]. The voltage-insensitive cation leak channel, a subunit of which is encoded by unc-77/nca-1, assists transmission of presynaptic activation from the cell body to the synapses [29,30]. The unc-8 gene encodes a putative mechanosensory channel [31]. A gain-of-function (gf) allele of either of these genes result in exaggerated body bends and coiling [27,30,31].
Examples of successfully detected complex postures for six mutants that display a coiler phenotype are shown in Fig 3E and S2 Movie. In this work, anterior coils were defined as periods when the head was in close proximity to any point along the body (within 5% of the midline). Posterior coils were similarly defined for the tail. We note that these definitions were not mutually exclusive (Fig 4A). The rate of detections using the generative statistical algorithm was compared to that of a previously described image analysis tool [9], which uses a standard morphological approach for single-frame detection that solely relies on the contrast between object and background (see also the Discussion section). The statistical approach yielded midlines of appropriate length for >90% of the images and these midlines were very similar to those obtained using the morphological approach, when the latter was available. For coiler mutants, the differences between the two algorithms mirrored the abundance of coils, indicating that detecting complex postures was key to the observed improvements (Fig 4B).
As a coarse measure of the severity of different coiling defects, the durations and frequencies of coiling were measured for each of the strains tested. Typical timescales were obtained by fitting the data to a Weibull distribution [32] (Fig 5A) and the full distributions for a severe coiler, egl-30(gf), are depicted in S2 Fig. The proposed algorithm thus enabled us to obtain a nearly continuous record of posture dynamics, uninterrupted by coiling.
To examine the relation between coiling and locomotion, we derived the propagation of dorsoventral body-bends from the time-series of postures as previously described (see Materials and Methods section and [9]). Coiling events in wild-type animals were rare, their durations were short, and they rarely interrupted directed locomotion (Fig 5A and 5B). In contrast, the majority of bouts of directed locomotion were interrupted by coiling in egl-30(gf), and unc-8(gf) mutants and coils were longer and more frequent than wild-type in these mutants (Fig 5A and 5B).
During continuous periods in which the posture of C. elegans is non-self-avoiding, the directional propagation of body bends can be disrupted to varying degrees. Therefore, an alternative measure of the impact of coiling can be obtained by asking how it affects directed locomotion. To address this, we measured the propensities for directed locomotion during coiling events. In each case, these propensities were compared to their baseline values, i.e., when coiling was absent (Fig 5C). Wild-type animals mostly progressed forward during a coil. This was the case since wild-type coiling was largely caused by Ω-turns which facilitate turning and are not detrimental to directed locomotion (see below). In egl-30(gf) and unc-77/nca-1(gf) mutants, propensities for forward or backward locomotion exceeded their baseline levels during anterior or posterior coiling, respectively (Fig 5C). Thus, in these mutants directed propagation of dorsoventral bends could be sustained despite coiling and, as further shown below, locomotion and coiling were likely linked. Taken together, these results suggest that the proposed statistical approach can be used to characterize coiler phenotypes.
Principal component analysis (PCA) was proposed as an unbiased and efficient approach for describing C. elegans behavior [7]. It has been used to characterize the dimensionality and dynamics of locomotion, as well as behavioral motifs [7,8,33,34]. When complex postures are largely inaccessible, two of the leading modes describe sinuous oscillations associated with directed locomotion and a third is associated with turning [7,34].
We constructed an ensemble of complex postures by equally sampling the coiled postures of the mutants we assayed. The resulting two leading modes were associated predominantly with anterior and posterior curvature. Thus, the severity and direction (dorsal/ventral) of anterior and posterior coils corresponded to the amplitudes and signs of these modes, respectively (Fig 6A and 6B). The third mode contributed opposite curvatures to the edges and the mid-body. Additional modes introduce higher order corrections and more than 95% of the variance in the data was accounted for by the leading six modes (Fig 6A and 6B).
Identifying typical coiling postures is ambiguous due to their broad distributions. Nevertheless, a heuristic definition can highlight prominent features and provide a useful starting point. We used k-means clustering to sub-divide the dataset of amplitudes of the modes composing coiled postures. Scree plots [35] would typically lead to dividing a coiler dataset to k = 6–8 similarly sized clusters. However, we found that generating a larger number of smaller clusters was useful: the centroids of the most populated small clusters resembled postures that were frequently observed in the raw data. Representative examples of cluster centroids and the postures that were reconstructed from them were projected onto the plane of the two leading modes and depicted in Fig 6B.
The dynamics of the amplitudes during continuous periods of coiling could vary between different animals and different types of coils. At the two extremes, the duration of a coil could be spent in a static posture that is not easily released or in a continuous sequence of exaggerated body-bends (also referred to as loopy motion). The amplitudes of the leading modes demonstrate that posterior coils of egl-30(gf) and unc-77/nca-1(gf) mutants were highly dynamic, such that ventral (positive amplitudes) and dorsal (negative amplitudes) coiling were averaged out while the animal exhibited a continuous succession of non-self-avoiding postures. In contrast, anterior coils of unc-8(gf) and unc-3 mutants were characterized by locking into a static coiled posture (Fig 6C).
More detailed information can be obtained from focusing on specific families of coils. We defined a spool as any posture for which the product of the two leading amplitudes, a1·a2, was larger than unity, i.e., anterior and posterior curvatures were sufficiently high and in the same direction. The centroids of the 10 most populated clusters of postures that satisfied this condition spanned the observed range of loops typical of Ω-turns to compact spirals. The shaded areas in the left panel of Fig 6D represent the convex hulls of these centroids for wild-type animals and four coiler mutants. The surrounding postures were reconstructed from 26 of these centroids. The center of the panel, where a1·a2 was small, was populated by loops that could typically be observed during Ω-turns. As a1·a2 grew larger, we preferentially observed spirals in which the head was at the center in unc-8(gf) and unc-122 mutants. In contrast, we observed a significant fraction of spirals in which the tail was at the center in egl-30(gf) and unc-77/nca-1(gf) mutants. Taken together with the dynamics of the amplitude, these data suggested that the two pairs of mutants preferentially formed spirals differently. Exaggerating an anterior bend prevented dorsoventral undulations and developed into a static head-centered spiral, perhaps through proprioceptive coupling, i.e., the trigger that compels body regions to bend in the same direction as their anterior neighboring region after a short time delay during forward locomotion [36]. Reversing into an exaggerated posterior bend formed a tail-centered spiral that did not suppress dorsoventral bending and was more rapidly released. Assaying the propagation of body-bends concurrently with coiling (described below) supported this interpretation.
PCA analysis of spools can also be used to assess the severity of a defect. Wild-type animals rarely exhibit postures for which a1·a2 >4, but coilers do (Fig 6D, right panel) and this trend was not sensitive to the exact value of the threshold. Projecting the sub group of spools onto its own low-dimensional space could facilitate testing more detailed hypotheses. Intuitively, the resulting three leading principal components corresponded to nearly uniform curvature, tightening of anterior bending, and tightening of posterior bending (S3 Fig). Similarly applying the condition a1·a2 < −4 would result in charactering number-8-like coils where anterior and posterior curvature have opposite signs. These results demonstrate that complex posture recognition can be integrated with existing analysis methods for large scale and unbiased studies of severe locomotion defects.
Is the initiation of directed locomotion particularly favorable for coiling in certain mutant backgrounds? To address this question, we assayed the temporal dynamics of locomotion upon entering and exiting a coiling event (Fig 7). In large part, wild-type coiling resulted from omega turns: acute turns composed of a reversal, an Ω-like posture, and forward locomotion in the new direction [37–41]. As a result, a rise in the propensity to reverse was observed shortly prior to coiling and high levels of forward locomotion were observed immediately following the coil (Fig 7A). In the cases of egl-30(gf) and unc-77/nca-1(gf) mutants, the signature of a reversal-to-forward switch was detected immediately prior to entering an anterior coil (Fig 7B and 7C, middle panels). Upon posterior coiling, these mutants exhibited the opposite behavioral switch (Fig 7B and 7C, right panels). However, similar trends were not observed in other coiler mutants (Fig 7D and 7E).
As a complementary measure of the association of coiling with locomotion transitions, we measured the fraction of coiling events that occurred within 5 sec from the initiation of directed locomotion. The signature of wild-type Ω-turns could be clearly detected: a large fraction of all coils promptly followed the initiation of forward locomotion after a reversal (Fig 8A). In coiler mutants, an exaggerated posterior body-bend upon a forward-to-reversal switch could increase the likelihood of coiling shortly following the initiation of the reversal. This trend (and the opposite one for anterior coiling) was displayed by egl-30(gf) and unc-77/nca-1(gf) mutants but not by other coilers (Fig 8A).
Our analysis typically identified brief periods of dwelling during transitions between forward and backward locomotion. Therefore, to visualize selected behavioral trends at the termination and initiation of directed locomotion, we aligned the data at the initiation and termination of short bouts of dwelling. Locomotion was then compared between two sub-categories of the full dataset: events in which, shortly following the onset of dwelling, anterior coiling was identified or no coiling was detected (Fig 8B; additional examples shown in S4 Fig). Coiling upon switching by egl-30 mutants manifested as an exaggerated reversals peak prior to dwelling and elevated forward propensities following dwelling. Thus, in egl-30(gf) and unc-77/nca-1(gf) mutants, the dorsoventral bends that initiate directed locomotion may be more likely to exaggerate and result in coiling than those that follow.
Are coiler phenotypes asymmetric with respect to the dorsoventral axis? The deep head bend of an Ω- turn is known to be ventral (Fig 9A) [37–41]. However, ectopic deep bends could potentially arise from the misregulation of bending in either direction. Interestingly, posterior coiling of unc-77/nca-1(gf) mutants was more likely when the tail bent dorsally (Fig 9B). The asymmetry in the bending direction of the tail also manifested as higher dorsal (as compared to ventral) posterior curvature in the period leading to a coil (Fig 9C).
The NCA-1 leak channel was recently implicated in persistent motor circuit activity required for sustaining locomotion [42]. Curiously, the gain-of-function of UNC-77/NCA-1 was shown to eliminate some of the spontaneous activity in muscles (miniature postsynaptic currents) [30]. The asymmetric behavior of unc-77/nca-1(gf) mutants can lead to hypotheses regarding the structure and function of the backward motor circuit. For instance, given the expression of unc-77/nca-1 in AVA premotor interneurons, AVA may be capable of asymmetrically activating dorsal and ventral motoneurons. Alternatively, unc-77/nca-1 may be asymmetrically expressed in motoneurons [43] or not expressed in AS neurons which innervate only dorsal muscles [2,30,44]. In the latter case, AS may play a role in maintaining dorsoventral balance.
We note that the initiation of coiling is not generally restricted to the initiation of directed locomotion. To demonstrate this we examined animals carrying a gain-of-function mutation in the unc-8 gene, encoding a putative mechanosensory channel [31] or a loss-of-function mutation of unc-122, affecting postsynaptic neuromuscular signaling [45]. Neither of these mutants exhibited the signature peaks associated with switching before coiling (Figs 7D, 7E, 8D and 8E). Curiously, unc-8(gf) was the only mutant we examined that exhibited significant anterior coiling while reversing, as evident by the unique rise of reversal probability prior to anterior coiling (Fig 7D left and middle panels). These data indicate that the proposed statistical model can be used for testing detailed hypotheses regarding cellular and molecular locomotion mechanisms.
The standard approach for identifying C. elegans in a digitized image applies simple morphological operations and/or heuristically motivated processing steps [5]. Typically, a background subtraction step is followed by thresholding to obtain a binary image. The largest connected component in the binary image is identified as the animal. Next, morphological closing (dilations followed by erosions) or morphological hole filling is applied and a skeletonization algorithm computes the midline of the body. The head is distinguished from the tail either by manual inspection or by comparing the regions in vicinity of the end points of the midline. Typically, when imaging in “artificial dirt” chambers, the brighter region is associated with the head.
Alternatively, the boundaries of the body in the binary image are determined by subtracting an eroded version of the image (or an equivalent edge detection method). A spline can then be fitted to all boundary points and the end with the higher peak curvature is associated with the tail [10]. If visual inspection is feasible and the duration of the measurement is limited, manual detection of the head and information about the motion of the center of mass can be used to resolve situations where parts of the worm overlap [11–14]. Such approaches are limited in their ability to reliably detect complex (non-self-avoiding) postures based on a single frame.
The approaches described in [11–13] have been applied to non-self-avoiding postures (including some cases of self-crossing midlines) and implemented commercially. They are based upon a geometric model for postures and a motion model for deformations of postures during locomotion. The posture in a given frame is assumed to be a small deformation of the posture in the preceding frame. Given this assumption, complex postures are resolved by tracing them back to simpler ones. These approaches require an initially resolved simple posture that sufficiently resembles the complex one. The simple posture is either provided manually [11] or assumed to be automatically attainable [12,13]. Once such an algorithm loses track of an animal it cannot autonomously recover, but may resume tracking given manual input [11].
These and similar approaches were not specifically designed to address severe phenotypes such as the prolonged continuous periods of coiling exhibited by egl-30(gf) mutants. Correspondingly, in published datasets, they were strictly applied to short video sequences in which bouts of coiling were brief. Our formulation of the object recognition problem is qualitatively different: we introduce sparse visual features that enable single-frame detection as opposed to solely relying on the differences in brightness between the imaged animal and its background. In addition to minimizing error propagation and manual intervention, single-frame detection can be parallelized easily and applied efficiently to large datasets.
An enhancement of the standard morphological methods is described in [46], where the skeleton could be determined for omega or spiral shaped postures. In this work, sophisticated heuristics were used to locate and dissect instances of self-touching for certain coiled body configurations. However, this approach is limited to specific postures and cannot be easily generalized. In addition, a laterally coupled snake model was developed for accurate contour detection of coiled animals [47]. When faced with complex postures, this method requires initializations that are close to the correct posture and therefore cannot be used in high-throughput, automated, applications.
Importantly, existing methods do not provide a measure of quality of detection, as they lack a cost function that allows comparison of candidate solutions in a meaningful way. A key advantage of a global generative statistical model is that it is principled: it enables to quantitatively assess the plausibility of the detected posture and can be naturally adapted to different experimental circumstances. An additional advantage of the proposed approach is its scalability. Analyzing a single frame at a time (rather than relying on neighboring frames) is an embarrassingly parallel problem, i.e., one that requires no dependency or communication between the parallel tasks, and eliminates propagation of errors. We implemented our algorithm using open source, freely available tools and libraries that are virtually guaranteed to be available at any research-computing environment. Therefore, our implementation can seamlessly be incorporate in a “big data” workflow for the timely analysis of large volumes of data.
In order to apply our approach to a different species it would be necessary to identify distinct visual features analogous to the edge formations described here. For this work it was sufficient to represent postures as a sequence of instantiation points but more sophisticated representations can be used instead.
To summarize, we presented a computationally efficient method, which correctly detects the posture of C. elegans in a variety of complex cases where standard morphological operations are inadequate. If higher precision is required, our fine detection method can be extended to a more computationally expensive procedure, e.g., an additional stage of further refinement. The presented analysis of coiler mutants demonstrates the flexibility and usability of this method for generating and testing detailed hypotheses.
C. elegans strains were maintained and grown according to standard protocols [1]. The following strains were used: wild-type strain N2, CG21 egl-30(tg26); him-5(e1490), DR1089 unc-77(e625), CB15 unc-8(e15), CB4870 unc-122(e2520), CB151 unc-3(e151), CB719 unc-1(e719).
Animals were grown at 20°C on standard NGM plates seeded with E. coli OP50 bacteria. Mid to late L4 individuals were sealed into individual “artificial dirt” chambers filled with an overnight OP50 culture concentrated tenfold and resuspended in NGM medium [48]. Animals were imaged at 10 frames per second at a 4.2x magnification for posture-based analysis using a CCD camera (Prosilica GC2450, Allied Vision Technologies, Stadtroda, Germany). Motion and quiescence were identified using previously described methods [49].
The proposed algorithm for identifying body features was implemented in Python (performance-critical parts were programmed in Cython) and integrated with our previously described suite of image analysis tools, called PyCelegans [9,49]. In brief, once we identified the body midline, head, and tail in each frame, each midline was divided into 20 equal intervals and the relative angles of all 18 next-nearest neighbor interval pairs (corresponding to the curvature of the body) were calculated. The dynamics of these 18 relative angles were used to identify quiescence and directed locomotion states. The propagation of body bends from anterior to posterior or vice versa was defined as forward or backward locomotion, respectively. Complete lack of motion was defined as quiescence. All other states were defined as dwelling. Although directional propagation of body bends corresponded well to centroid motion, directed locomotion could be scored even if the animal was slipping and the centroid was not propagating in the laboratory frame of reference. Data analysis was performed using custom Matlab scripts (Mathworks Inc., Natick MA). Our source code and documentation are publicly available at https://github.com/david-biron/pycelegans-2.0.
Data analysis was performed using custom Matlab scripts. For comparisons in summary statistics panels, significance was calculated using a one-way ANOVA test. Post-hoc correction for multiple comparisons was performed using the Bonferroni adjustment. For the purpose of performing principal component analysis (PCA), the posture of the animal was represented using the same 18 relative angles between next-nearest neighbor intervals that were used for the analysis of locomotion. For the purpose of calculating the principal modes, approximately 5,000 anterior coiled frames and 5,000 posterior coiled frames (or spooled, for S3 Fig) were randomly picked from the full dataset of each of the six coiler mutants assayed. Thus, PCA modes were calculated using a total of 60,000 frames. K-means clustering was performed with a redundancy of 5 using k = 50 clusters for Fig 6 and k = 25 for the more restricted set of spools (S3 Fig). Results were not sensitive to an exact choice of the number of clusters, k.
|
10.1371/journal.pgen.1001193 | A Coastal Cline in Sodium Accumulation in Arabidopsis thaliana Is Driven by Natural Variation of the Sodium Transporter AtHKT1;1 | The genetic model plant Arabidopsis thaliana, like many plant species, experiences a range of edaphic conditions across its natural habitat. Such heterogeneity may drive local adaptation, though the molecular genetic basis remains elusive. Here, we describe a study in which we used genome-wide association mapping, genetic complementation, and gene expression studies to identify cis-regulatory expression level polymorphisms at the AtHKT1;1 locus, encoding a known sodium (Na+) transporter, as being a major factor controlling natural variation in leaf Na+ accumulation capacity across the global A. thaliana population. A weak allele of AtHKT1;1 that drives elevated leaf Na+ in this population has been previously linked to elevated salinity tolerance. Inspection of the geographical distribution of this allele revealed its significant enrichment in populations associated with the coast and saline soils in Europe. The fixation of this weak AtHKT1;1 allele in these populations is genetic evidence supporting local adaptation to these potentially saline impacted environments.
| The unusual geographical distribution of certain animal and plant species has provided puzzling questions to the scientific community regarding the interrelationship of evolutionary and geographic histories for generations. With DNA sequencing, such puzzles have now extended to the geographical distribution of genetic variation within a species. Here, we explain one such puzzle in the European population of Arabidopsis thaliana, where we find that a version of a gene encoding for a sodium-transporter with reduced function is almost uniquely found in populations of this plant growing close to the coast or on known saline soils. This version of the gene has previously been linked with elevated salinity tolerance, and its unusual distribution in populations of plants growing in coastal regions and on saline soils suggests that it is playing a role in adapting these plants to the elevated salinity of their local environment.
| Uncovering the genetic polymorphisms that underlie adaptation to environmental gradients is a critical goal in evolutionary biology, and will lead to a better understanding of both the types of genetic changes and the gene functions involved. Such understanding will not only provide insight into how organisms may respond to future global climate change, but will also provide tools for the development of agricultural systems and ecological services that are more resilient to such changes.
Patterns of phenotypic diversity across environmental gradients can be indicative of adaptive responses to selection, and evaluation of these patterns has the potential to lead to the identification of the genetic polymorphisms underlying these adaptive responses. Numerous studies in animals and plants have identified phenotypic clines in various life history traits, but only a few have determined the genetic changes driving such traits. In Arabidopsis thaliana, plasticity in seasonally regulated flowering appears to be modulated by a network of gene interactions responsive to both vernalization and photoperiod signals [1]. Adaptive clines in resistance to oxidative stress and chilling [2], and wing size [3] in Drosophila melanogaster are modulated by the Insulin-like Receptor (InR) and Drosophila cold acclimation (Dca) genes, respectively. While adaptation to high altitude in Peromyscus maniculatus (Deer mice) is associated with enhanced pulmonary O2 loading driven by alterations in α-globin and β-globin genes [4]. These genetic changes are all associated with adaptation to variation in environmental factors that vary with latitude or altitude. Such systematic variation has greatly facilitated the discovery of these loci and their adaptive significance. Clines in various life history traits have also been identified in plants growing on serpentine [5], saline [6], [7], and mine impacted soils [8]. Progress has been made in outlining the genetic architecture of these adaptive traits [5], [8]–[10], though a molecular genetic understanding is still needed.
A. thaliana is broadly distributed in its native Europe and central Asia, where it experiences a wide range of altitudinal, climatic, and edaphic conditions, leading to a range of selective pressures [11]. Whether the wide variety of natural phenotypic and genetic variation observed in A. thaliana [12] contributes to its local adaptation is an important unresolved question that is currently attracting a significant amount of attention [13].
Because of its relevance to crop production, salinity tolerance in plants has been studied intensively [14], and natural plant populations adapted to such conditions have provided an excellent system for studying the evolutionary mechanisms of adaptation and speciation in coastal [6], [10] and salt marsh [7], [9], [15]–[18] environments. The primary effects of excess Na+ on plants are water deficit resulting from a water potential gradient between the soil solution and plant cells, and cytotoxicity due of intracellular Na+ accumulation [14]. To overcome these effects plants must both accumulate solutes for osmotic regulation, and detoxify intracellular Na+ either by limiting its accumulation, or by compartmentalizing Na+ into the vacuole. In addition, Na+ compartmentalization facilitates vacuolar osmotic adjustment that is necessary to compensate for the osmotic effects of salinity by maintaining turgor pressure for cell expansion and growth. Plants therefore need to strike a balance between the accumulation of Na+ to maintain turgor, and the need to avoid Na+ chemical toxicity, and this balance will depend in part on soil salinity levels. Given the critical role Na+ accumulation plays in salinity tolerance, we used this life history trait to probe the global A. thaliana population for signals of adaptive selection for growth in saline impacted environments.
We grew 349 accessions of A. thaliana in a controlled common garden in non-saline soil, and analyzed leaf Na+ accumulation. We observed a wide range of leaf Na+ accumulation across the accessions (330–4,848 mg kg−1 dry weight). If this natural variation in leaf Na+ accumulation capacity is related to adaptation to growth in saline soils we would expect to find evidence of an adaptive cline, or a gradient of leaf Na+ accumulation that correlates with the geographical distribution of variation in soil salinity. Salinity impacted soils are expected to occur in coastal regions due to air born deposition of sea spray which can occur many tens of km inland [19]–[22], but can also occur in areas distant from the coast through high Na+ in the soil or ground water. Elevated soil salinity can also be caused by inappropriate irrigation practices such as irrigation with saline water or poor drainage.
To test for the existence of an adaptive cline in leaf Na+ accumulation capacity and soil salinity we related leaf Na+ accumulation capacity to the distance of the collection site for each accession to the coast, or to the nearest known saline soil, whichever is the shortest. We focused on European accessions since a good soil salinity map exists for this region [23], which left 300 accessions. Regressing the distance to the coast, or nearest known saline soil, on leaf Na+ for all 300 accessions revealed a significant relationship (p-value<2e-12), establishing that accessions with elevated leaf Na+ are more likely to grow in potentially saline impacted soils (Figure 1A and 1B).
To investigate the genetic architecture underlying this cline in leaf Na+ accumulation capacity we performed a genome-wide association (GWA) study (previously described for a smaller data set [24]) to identify regions of the genome at which genetic variation is associated with leaf Na+ accumulation capacity. The 337 A. thaliana accessions used in our GWA study, which are a subset of the 349 accessions phenotyped for leaf Na+, were genotyped using the Affymetrix SNP-tilling array Atsnptile1a which can interrogate 248,584 SNPs. To assess evidence of association between SNPs and leaf Na+ accumulation we used a mixed-model approach [25] to correct for population structure, as previously described [24]. In the current analysis we identified a single strong peak of SNPs associated with leaf Na+, with the peak centered on AtHKT1;1 (Figure 2), a gene known to encode a Na+-transporter [26]. Accessions with a thymine (T) at the SNP most significantly associated with leaf Na+ at position 6392276 bp on chromosome 4 (Chr4:6392276) have significantly higher leaf Na+ than accessions with a cytosine (C) at this same position (2,325 vs. 955 mg Na+ kg−1 dry weight, p-value<2e-16). This SNP explains 32% (without accounting for population structure) of the total variation in leaf Na+ accumulation observed.
Previously, in independent test crosses between the high leaf Na+ accessions Ts-1 and Tsu-1 (both containing a T at Chr4:6392276) and the low leaf Na+ accession Col-0 (containing a C at Chr4:6392276) QTLs for leaf Na+ centered on AtHKT1;1 were identified in both F2 populations [27]. Such genetic evidence provides independent support that the peak of SNPs associated with leaf Na+ observed in our GWA analysis, centered at AtHKT1;1 (Figure 2), represents a true positive association and not a false positive driven by the high degree of population structure known to exist in A. thaliana [24]. Reduced expression of AtHKT1;1 in Ts-1 and Tsu-1 was concluded to drive the elevated leaf Na+ observed in these two accessions [27]. Here, we expand on this observation by establishing the strength of the AtHKT1;1 alleles in four further high Na+ accumulating accessions (Bur-1, Duk, PHW-20 and UKNW06-386) that all contain a T at Chr4:6392276, along with a low leaf Na+ accession (Nd-1) with a C at Chr4:6392276. By examining the leaf Na+ accumulation in F1 plants from crosses of each of these accessions to Col-0hkt1-1 and Col-0HKT1, we were able to establish a significant correlation between leaf Na+ accumulation and the strength of the AtHKT1;1 alleles (Figure 3A). These crosses confirmed that all accessions tested with elevated leaf Na+, and that contain a T at Chr4:6392276, have hypofunctional alleles of AtHKT1;1 relative to the Col-0 allele. Furthermore, analysis of the expression of AtHKT1;1 in the same set of accessions revealed that allelic variation in AtHKT1;1 strength is modulated at the level of gene expression (Figure 3B), consistent with what was previously observed for Ts-1 and Tsu-1 [27]. Though the SNP most significantly associated with leaf Na+ (Chr4:6392276) is unlikely to be causal for these AtHKT1;1 expression level polymorphisms, this SNP can be used as a linked genetic marker to determine the type of AtHKT1;1 allele present, with a T at this SNP being associated with weak AtHKT1;1 alleles.
Using the SNP at Chr4:6392276 as a genetic marker for the type of AtHKT1;1 allele (strong or weak) allowed us to test the hypothesis that the leaf Na+ soil salinity cline we observe in European populations of A. thaliana (Figure 1A and 1B) is associated with weak alleles of AtHKT1;1. By comparing the means of distances to the coast, or known saline soil, for the collection site of all 300 accessions with and without a T at Chr4:6392276, we determined that a significant association (parametric test p-value = 0.0001; non-parametric Wilcoxon rank-sum test p-value = 0.0062) exists between A. thaliana growing on potentially saline impacted soils and the presence of a weak allele of AtHKT1;1 (Figure 1A and 1B). Such a strong correlation between the presence of allelic variation at AtHKT1;1 known to drive elevated leaf Na+, and the observed cline in leaf Na+ and saline soils, is evidence for the involvement of AtHKT1;1 in determining this geographical distribution. Furthermore, using 13 SNPs within a 20kb region centered on HKT1;1 to define the HKT1;1 haplotype, we identify 7 haplotypes (6 if you combine haplotypes with only 1 SNP different) in accessions with high leaf Na+ (>2,500 ppm), suggesting that weak alleles of HKT1;1 have arisen independently multiple times.
However, to be credible it is also important to provide evidence that selection for growth on saline soils could be acting on the phenotype driven by allelic variation at AtHKT1;1; in this case elevated leaf Na+. Such evidence is provided by the previous observation that the weak allele of AtHKT1;1 in the coastal Tsu-1 A. thaliana accession not only causes elevated leaf Na+ but is also genetically linked to the elevated salinity tolerance of this accession [27].
In A. thaliana AtHKT1;1 functions to unload Na+ from xylem vessels in the root, controlling translocation and accumulation of Na+ in the shoots [26], [28]. Therefore, modulation of its function would allow the balancing of Na+ accumulation in the shoot with soil salinity. We note here that the hkt1-1 null mutation in the Col-0 background causes plants to exhibit dramatic leaf Na+ hyperaccumulation and increased NaCl sensitivity [29], [30]. We interpret this to mean that expression of AtHKT1;1 in the hkt1-1 null mutant is reduced to such an extent that leaf Na+ accumulation saturates the capacity for cellular detoxification of Na+ by vacuolar compartmentalization. We propose that the naturally occurring weak alleles of AtHKT1;1, that we show are associated with populations growing in potentially saline impacted environments, allow sufficient Na+ to accumulate in leaves for osmotic adjustment, conferring elevated Na+ tolerance. However, these weak, but not complete loss-of-function AtHKT1;1 alleles, do not saturate the mechanism whereby the accessions avoid Na+ cytotoxicity. The basis of this Na+ detoxification mechanism remains to be determined, though an active leaf vacuolar Na+ compartmentalization mechanism driven by AtNHX1 is one likely candidate.
In conclusion, here we provide evidence supporting the involvement of specific cis-regulatory polymorphisms at AtHKT1;1 in the potentially adaptive cline in leaf Na+ accumulation capacity we observe in A. thaliana populations to saline impacted environments. We have identified a strong association between the AtHKT1;1 allele frequency in A. thaliana populations and their growth on potentially saline impacted soils (Figure 1A and 1B). Further, we have confirmed by GWA mapping, experimental complementation crosses, and gene expression studies, that this allelic variation directly causes changes in the clinally varying leaf Na+ accumulation phenotype via cis-regulatory polymorphisms (Figure 2 and Figure 3). And, finally, we have previously established that the weak AtHKT1;1 alleles we show to be associated with potentially saline soils, are also linked to elevated salinity tolerance [27], providing a plausible mechanistic link between selection for growth on saline soils and variation in AtHKT1;1 allele frequency. Such discoveries provide tantalizing evidence that points to selection acting at AtHKT1;1 in natural populations of A. thaliana in adaptation to growth in saline environments.
Plants were grown in a controlled environment with 10 h light/14 h dark (90 µmol m−2s−1 photosynthetically active light) and 19 to 22°C, as previously described [31]. Briefly, seeds were sown onto moist soil (Promix; Premier Horticulture) in 10.5″×21″ 20 row trays with various elements added to the soil at subtoxic concentrations (As, Cd, Co, Li, Ni, Rb, and Se [31]) and the tray placed at 4°C for 3 days to stratify the seeds and help synchronize germination. Each tray contained 108 plants, six plants each from 18 accessions, with three plants of each accession planted in two different parts of the tray. Each tray contained four common accessions (Col-0, Cvi-0, Fab-2 and Ts-1) used as controls, and 14 test accessions. Trays were bottom-watered twice per week with 0.25-strength Hoagland solution in which Fe was replaced with 10 µM Fe-HBED[N,N′-di(2-hydroxybenzyl)ethylenediamine-N,N′-diacetic acid monohydrochloride hydrate; Strem Chemicals, Inc.). After 5 weeks plants were non-destructively sampled by removing one or two leaves and the elemental composition of the tissue analyzed by Inductively Couple Plasma Mass Spectroscopy (ICP-MS). The plant material was rinsed with 18 MΩ water and placed into Pyrex digestion tubes. For complementation experiments plants were crossed to Col-0 or Col-0hkt1-1 and approximately 12 F1 plants were grown in the conditions described above.
A set of 360 A. thaliana accessions were selected from 5,810 worldwide accessions to minimizing redundancy and close family relatedness, based on the genotypes at 149 SNPs developed in a previous study [32]. Figure S1 and Table S1 show the genetic variation in the core set of 360 accessions vs. a random set of 360 accessions chosen from the genotyped 5,810 accessions. From the selected core set of 360 accessions a subset of 349 were phenotyped using ICP-MS, and of these 337 were genotyped using the Affymetrix SNP-tilling array Atsnptile1 which contains probe sets for 248,584 SNPs. Details of the SNP-tilling array and methods for array hybridization and SNP-calling are the same as previously described [24]. In brief, approximately 250 ng of genomic DNA was labeled using the BioPrime DNA labeling system (Invitrogen) and 16 µg of the labelled product hybridized to each array. SNPs were called using the Oligo package after slight modifications. Quality control (QC) of the genotypes, and imputation of the missing SNPs were performed following the procedure previously described [24], except that a 15% mismatch rate was used to filter out low quality arrays. After QC and imputation, the 337 accessions had genotypes for at least 213,497 SNPs. The core set of 360 accessions selected are all available from the Arabidopsis Biological Resource Center (http://abrc.osu.edu/), and the SNP genotypes for the 337 accessions used for the GWA study are available from http://borevitzlab.uchicago.edu/resources/genetic/hapmap/BaxterCore/.
Samples were analyzed as described by Lahner et al. [31]. Tissue samples were dried at 92°C for 20 h in Pyrex tubes (16×100 mm) to yield approximately 2–4 mg of tissue for elemental analysis. After cooling, seven of the 108 samples from each sample set were weighed. All samples were digested with 0.7 ml of concentrated nitric acid (OmniTrace; VWR Scientific Products), and diluted to 6.0 ml with 18 MΩ water. Elemental analysis was performed with an ICP-MS (Elan DRCe; PerkinElmer) for Li, B, Na, Mg, P, S, K, Ca, Mn, Fe, Co, Ni, Cu, Zn, As, Se, Rb, Mo, and Cd. A liquid reference material composed of pooled samples of A. thaliana leaves was run every 9th sample to correct for ICP-MS run to run variation and within-run drift. All samples were normalized to the calculated weights, as determined with an iterative algorithm using the best-measured elements, the weights of the seven weighed samples, and the solution concentrations, implemented in the Purdue Ionomics Information Management System (PiiMS) [33] (for a full description see www.ionomicshub.org). Data for all elements is available for viewing and download at www.ionomicshub.org in trays 1478–1504.
To quantify the levels of AtHKT1;1 mRNA in roots of the various accessions studied, we used a protocol similar to that of Rus et al. [27]. Roots from plants grown under identical conditions to those used for ICP-MS analysis were separated from the shoots and rinsed thoroughly with deionized water to remove any soil contamination. The samples were frozen in liquid nitrogen and stored at −80°C until extraction. Total RNA was extracted, and DNase digestion was performed during the extraction, using the Invitrogen PureLink RNA Mini Kit. Two micrograms of total RNA were used as a template to synthesize first-strand cDNA with random hexamers, using SuperScript II Reverse Transcriptase (Invitrogen Life Technologies). Quantitative real-time PCR (qRT-PCR) was performed with first strand cDNA as a template on four technical replicates from three independent biological samples for each accession, using a sequence detector system (StepOne Plus, Applied Biosystems). For normalization across samples within a qRT-PCR run the expression of the Actin 1 gene (At2g37620) was used with the following primers: CPRD66, 5′-TGG AAC TGG AAT GGT TAA GGC TG-3′ and CPRD67, 5′-TCT CCA GAG TCG AGC ACA ATA C-3′. For quantification of AtHKT1;1 the following primers were used: HKT-RTF, 5′-TGG GAT CTT ATA ATT CGG ACA GTT C-3′ and HKT-RTR, 5′-GAT AAG ACC CTC GCG ATA ATC AGT-3′. The fold induction relative to AtHKT1;1 expression in Col-0 roots was calculated following the method of Livak and Schmittgen [34]. CT values were determined based on efficiency of amplification. The mean CT values were normalized against the corresponding Actin 1 gene and ΔCT values calculated as CTAtHKT1;1–CTActin 1. The expression of AtHKT1;1 was calculated using the 2∧(ΔCT) method [34]. To normalize between samples analyzed in separate qRT-PCR runs, we divided the ΔCT for each line by the ΔCT of Col-0 roots in that run.
ICP-MS measurements below zero and extreme outliers (those values that were greater than the 90th percentile + percentile) within each tray were removed. To account for variation in the growth environment, the four control accessions included in each tray were used to create a tray specific normalization factor. Briefly, for each element, each control accession in a given tray was compared to the overall average for that accession across all trays to obtain an element×line×tray specific normalization factor. The four element×line×tray factors in a give tray were then averaged to create a tray×element normalization factor for the tray. Every value for the element in the tray was then multiplied by the normalization factor. See Figure S2 for data of control accessions before and after the normalization. The mean of each accession was then used for all subsequent analysis. Normalized Na+ values and their frequency distribution can be found in Dataset S1 and Figure S3.
Genotype calls for all 349 accessions were obtained using the methods previously described [24]. GWA analysis was done with correction for confounding using a mixed-model that uses a genetic random effect with a fixed covariance structure to account for population structure [25] implemented in the program EMMA [24]. The contribution of the best performing SNP (C or T at Chr4:6392276 = isT) was checked using un-normalized Na+ data and the linear model:(1)using the lm and anova functions from R v2.9.1. The control accessions were excluded from this analysis. The output of the statistical model can be found in Text S1. Although the samples were nested in trays, Figure S4 indicates that the best performing SNP is essentially evenly distributed across all trays.
The geographical location of each accession was obtained from TAIR (www.arabidopsis.org). When processing the original data, we found an inconsistency for one of the high-Na accessions, CS28373 (also known as Jm-1). The listed latitude and longitude (49, 15) of the accession do not match the location name “Jamolice” from where this accession was collected. The town Jamolice is located at 49.0721283 latitude and 16.2532139 longitude (http://www.gpsvisualizer.com/geocode). In the interests of consistency, we used the original coordinates, although altering the location did not materially change the analysis. The distance to the coast or saline/sodic areas was calculated by obtaining the longitudes and latitudes of the shoreline/coast from the National Oceanic and Atmospheric Administration's National Geophysical Data Center (NOAAs NGDC http://www.ngdc.noaa.gov/ngdc.html) and the saline and sodic soils data from the European Soils database [23]. The pointDistance function in R 2.10.0 and the package raster were used to calculate the Great-circle distance to the shoreline or saline/sodic areas. We created a variable (toSeaSal) representing the shortest distance from the target accession to the shoreline/coast or saline/sodic area. The accession coordinates, distance to sea, distance to saline environment and SNP genotype at Chr4:6392276 can all be found in Dataset S1.
The method used to collect accessions and assemble the population might introduce unintended confounding effects that violate the assumption of independent locations used by our models. To determine whether the locations of the accessions were spatially dependent we performed a Mantel test [35] on the distances from the 300 accessions to the coast or known saline/sodic areas. The simulated p-values of 50 permutations tests with 999 repeatedly simulated samples are 0.996, indicating that an assumption of independency for the response variable toSeaSal is acceptable.
To test for associations between leaf Na+ (Na), genotype at the highest scoring SNP (C or T at Chr4:6392276 = isT), and the distance to the nearest coast or saline/sodic area (toSeaSal), we used the package lm in R 2.10.0 to fit linear models, with the weights determined by the following approach. First, to quantify the strength of the relationship between toSeaSal and the leaf sodium Na, we fit the data into a linear model and regressed toSeaSal on Na.(2)Second, we applied a regression approach to single-factor analysis [36] between toSeaSal and isT and tested if the average distance to coast or saline/sodic areas of samples having the high Na T allele is significantly different from the average of samples having the C allele.(3)Finally, we regressed toSeaSal on the interaction between Na and isT to inspect how the two predictors jointly affect the distance to sea or saline/sodic.(4)To perform the significance tests on the linear coefficients, Na should be centered at the mean [36]. The extent of variation of distances to saline environments changes with both leaf Na+ concentrations and genotypes (Figure S5). Therefore, all three models account for this heterogeneity of variation, and parameters of the models are fitted using weighted least squares. The variances of the error terms in equation 2, 3, and 4 are not constant, and are related to the predictors according to the diagnosis on the model residuals. The models were fit using iterative weighted least squares [36]. In addition to the parametric test (model 3), we performed a non-parametric test (Wilcoxon rank-sum test or Wilcoxon-Mann-Whitney test [37]) using the wilcox.test function in R package stats, to assess whether toSeaSal is higher in the lines with the T allele than those with the C allele at Chr4:6392276. The p-value of the Wilcoxon rank-sum test is 0.006224 indicating that both the parametric and non-parametric approaches reach the same conclusion. The statistical output of all models can be found in Text S1.
|
10.1371/journal.ppat.1003284 | HPV16 E7 Protein and hTERT Proteins Defective for Telomere Maintenance Cooperate to Immortalize Human Keratinocytes | Previous studies have shown that wild-type human telomerase reverse transcriptase (hTERT) protein can functionally replace the human papillomavirus type 16 (HPV-16) E6 protein, which cooperates with the viral E7 protein in the immortalization of primary keratinocytes. In the current study, we made the surprising finding that catalytically inactive hTERT (hTERT-D868A), elongation-defective hTERT (hTERT-HA), and telomere recruitment-defective hTERT (hTERT N+T) also cooperate with E7 in mediating bypass of the senescence blockade and effecting cell immortalization. This suggests that hTERT has activities independent of its telomere maintenance functions that mediate transit across this restriction point. Since hTERT has been shown to have a role in gene activation, we performed microarray studies and discovered that E6, hTERT and mutant hTERT proteins altered the expression of highly overlapping sets of cellular genes. Most important, the E6 and hTERT proteins induced mRNA and protein levels of Bmi1, the core subunit of the Polycomb Group (PcG) complex 1. We show further that Bmi1 substitutes for E6 or hTERT in cell immortalization. Finally, tissue array studies demonstrated that expression of Bmi1 increased with the severity of cervical dysplasia, suggesting a potential role in the progression of cervical cancer. Together, these data demonstrate that hTERT has extra-telomeric activities that facilitate cell immortalization and that its induction of Bmi1 is one potential mechanism for mediating this activity.
| The human papillomaviruses (HPVs) are critical elements in the etiology of cervical cancer, as well as several other human cancers. The E6 protein, in combination with the E7 protein of these viruses, immortalizes epithelial cells and increases the expression of the hTERT protein. In the current study we show that the enzymatic activity of hTERT is not required for cooperating in cell immortalization. We further demonstrate that hTERT proteins increase the expression of the Bmi1 protein, which is also capable of cooperating in cell immortalization. We anticipate that these findings will stimulate new studies of telomerase in HPV biology, cancer etiology, and stem cell reprogramming.
| Cell immortality is a hallmark of cancer cells [1] and the high-risk oncogenic HPVs encode two major transforming genes, E6 and E7, which are required for the immortalization of human primary genital keratinocytes [2], [3]. These two oncogenes are uniformly retained and expressed in cervical cancers and their continued expression is required for the cells to retain the tumorigenic phenotype [4], [5], [6], [7], [8]. The E6 and E7 proteins were initially identified as targeting the p53 and Rb tumor suppressor pathways in host cells, thereby disrupting cell cycle controls [5], [6], [7], [8]. E7 stimulates the cell cycle via its ability to bind and inactivate the cellular Rb protein while E6 binds to p53, leading to its degradation via the proteosomal pathway [5], [6], [7], [8].
In addition to p53 degradation, E6 induces telomerase activity in epithelial cells [6], [9], [10]. Telomerase is a specialized reverse transcriptase that synthesizes the telomeric repeat DNA sequences at the ends of chromosomes [11]. The absence of telomerase activity in most normal human cells results in the progressive shortening of telomeres with each cell division [12], [13], ultimately leading to chromosomal instability and cellular replicative senescence [12], [14]. For this reason, telomere shortening is thought to represent the “mitotic clock” that determines cellular lifespan. In contrast to most human somatic cells, approximately 90% of immortalized and cancer cells express telomerase activity and consequently maintain minimal, stable telomeres and indefinite proliferative potential [15]. Therefore, telomerase activation is considered a critical event in the process of cell immortalization. Recent studies indicate that telomerase may assist in bypassing two separate events which block the continuous replication of primary human cells: mortality stage 1 (M1, replicative senescence) followed by mortality stage 2 (M2, crisis) [16]. In some cells, especially those with decreased function of the p16/Rb pathway, telomerase activity is sufficient to bypass both M1 and M2 blockades and to stabilize and elongate telomeres [17], [18], [19], [20].
Studies have demonstrated that activation of telomerase by E6 is critical for cell immortalization by HPV [17], [21]. E6 executes this increase in telomerase activity by multiple mechanisms [8], [22], [23], [24], [25], [26]. While increased hTERT is required for viral-mediated cell immortalization [8], [17], [21], our previous studies demonstrated that telomeres erode in HPV-expressing keratinocytes similar to normal keratinocytes [10], suggesting that the role of hTERT overexpression in cell immortalization might involve functions additional to those in telomere elongation.
Evidence is accumulating that hTERT has important non-canonical functions. For example, mTERT has been ascribed roles in altering apoptotic responses [27], [28], tumor formation in mice [29], [30], stem cell migration and renewal [30] and chromatin remodeling [31]. The Artandi laboratory has shown that mTERT can not only augment breast cancer development in mice, but also can regulate the transcription of genes responsive to the Wnt/β-catenin pathway [30]. Smith et al demonstrated that in human mammary epithelial cells (hMECs) telomerase modulates expression of growth-controlling genes, including epidermal growth factor receptor (EGFR) [32]. Vascular endothelial growth factor (VEGF) and fibroblast growth factor (FGF) also appear to be induced by hTERT in fibroblasts, along with many other targets [33], [34]. Interestingly, the majority of these data have been recapitulated with an hTERT mutant that is catalytically-inactive, suggesting that these non-canonical roles of hTERT are independent of the reverse transcriptase function.
Although hTERT has been shown to be a key player in cellular immortalization, in many cases it does not immortalize alone [17]. Interestingly, the Bmi1 protein has been shown to cooperate with hTERT in immortalization and to induce hTERT mRNA [35], [36]. Bmi1 is the central protein in polycomb repressive complex 1 (PRC1). The Polycomb group (PcG) complex of proteins act through remodeling chromatin to silence hundreds of genes and have been implicated in controlling cell fate, development, and cancer [37], [38].
In the current study, we used quantitative assays to measure telomerase activity and telomere length following transduction of foreskin keratinocytes by E6/E7, hTERT/E7 and mutant hTERT/E7. These activities were correlated with the ability of the various hTERT mutant proteins to immortalize cells. Our studies indicate that a telomerase-independent activity of hTERT collaborates with E7 in the immortalization of primary human cells. To elucidate the underlying mechanism, whole genome expression profiling was performed in keratinocytes expressing E6, hTERT, or a catalytically inactive hTERT mutant (hTERT-D868A). Increased expression of Bmi1 mRNA was observed in this screening and follow-up experiments indicate that Bmi1 appears to be a functional component of hTERT- or E6-mediated cell immortalization and that its expression further increases during cancer progression.
Immortalized cells generally do not display the phenotypic properties of cancer cells (e.g. growth in soft agar or tumor formation in nude mice) without further gain of genetic changes or transduction of additional genes [5], [8], [39]. To determine if changes in telomerase activity might contribute to the differences between immortalized and tumorigenic cervical cells, we compared the levels of telomerase activity for cells immortalized by the HPV-16 E6/E7 genes to those found in 3 cervical cancer cell lines (SiHa, HPV16 positive; HeLa, HPV-18 positive; C33A, HPV negative). The E6/E7 immortalized cell lines (HFK E6/E7 at population doubling (PD) 90 and HEC E6/E7 at PD 98) exhibited similar levels of telomerase activity as the three cervical cancer lines, indicating that further increases in telomerase are not required for progression to malignancy (Figure 1A).
While an early hypothesis for HPV-mediated cell immortalization suggested that telomerase induction by E6 maintained telomere length [8], [9], [17], our previous studies showed that E6/E7 expressing cells continued to shorten telomeres even in the presence of induced telomerase activity [10]. To quantify these changes in telomere length, we used a PCR-based assay to screen E6/E7 immortalized cells and cervical cancer cell lines. At PD 2, HFKs had long average telomere lengths with a T/S ratio of 1.0 (Figure 1B). Since the approximate length of telomeres in early passage HFKs is 10 kb, the T/S ratio can be converted into telomere length (where 1.0 T/S ratio equals 10 kb telomere length). At PD 18, HFKs had a T/S ratio of 0.6 or 6 kb size. Immortalized, PD 96 HFK E6/E7 cells, which have bypassed crisis, had amongst the shortest telomeres (Figure 1B; T/S ratio of 0.2, or 2 kb length). These short telomeres were also seen in all three cervical cancer cell lines, including the HPV-negative cancer cell line, C33A. Our data suggest that E6/E7-immortalized cells continue to degrade their telomeres until they reach a length of 2 kb, at which point they become stabilized and equivalent in length to telomeres in cervical cancer cell lines.
The kinetics of passage-dependent shortening or degradation of telomeres during cell immortalization by E6/E7 were also studied (Figure 1C). By PD 74, HFK cells expressing E6/E7 achieved their shortest telomere length, after which telomere length became stable.
An hTERT protein that was epitope-tagged at its C-terminus (hTERT-HA) retained telomerase activity, but alone was unable to elongate telomeres or immortalize HFFs or SV40 transformed epithelial cells [40], [41], [42]. To test the functions of wild-type hTERT and hTERT-HA in human keratinocytes, HFKs were co-transduced with vectors expressing E7 and the hTERT proteins. As expected, wild-type hTERT cooperated with E7 to immortalize HFKs, while hTERT or E7 alone were unable to immortalize HFKs. Surprisingly, hTERT-HA also immortalized HFKs in collaboration with E7 [26], indicating that telomere maintenance is not critical for hTERT/E7 immortalization. The functionality of E7 was verified by demonstrating that the Rb protein level was significantly decreased in all E7 expressing cells (Figure S1).
To verify that the hTERT-HA mutant generated telomerase activity in HFKs, Telomeric Repeat Amplification Protocol (TRAP) assays were performed. HFKs transduced with hTERT or hTERT-HA alone, or in combination with E7, exhibited similar levels of telomerase activity (Figure 2A). HFKs with E7 alone did not exhibit significant telomerase activity. Consistent with our earlier results, Figure 2B illustrates that telomeres lengthened during immortalization of the hTERT/E7 cells, but shortened in the hTERT-HA/E7 cells (Figure 2B).
The preceding experiments indicate that cell immortalization is independent of telomere lengthening and raise the possibility that other telomere-related functions of hTERT were involved in this process. To evaluate this possibility, we therefore examined the immortalizing activity of additional hTERT mutants that were known to be catalytically inactive (hTERT-D868A) [43] or had impaired recruitment to telomeres (hTERT N+T) [42], [44]. Similar to hTERT-HA, both the hTERT-D868A and hTERT N+T mutants were able to immortalize keratinocytes in conjunction with E7 (Figure 3A). Cells immortalized by the hTERT N+T mutant exhibited similar levels of telomerase activity as cells immortalized by E6/E7, so decreased telomerase activity could be ruled out as a potential mechanism (Figure 3B). The results with hTERT-D868A were even more significant. Cells immortalized by this defective mutant were as efficient at cell immortalization as wild-type hTERT (Figure 3A), despite the complete lack of telomerase activity in early passage keratinocytes (Figure 3B). Immunofluorescence studies demonstrated that hTERT-D868A exhibited a similar expression level and localization as wild-type hTERT (Figure S2). Thus, the catalytic activity of hTERT and its ability to elongate telomeres is dispensable for the immortalization of keratinocytes with E7.
Another unexpected finding was that cells immortalized by the telomerase-defective hTERT-D868A mutant exhibited high telomerase activity at late passages (Figure 3C) in contrast to the lack of telomerase at early passages (Figure 3B). This led us to ask whether hTERT proteins, including hTERT-D868A activate the endogenous hTERT promoter. To test this, we transfected wild type or mutant hTERT proteins along with an hTERT core promoter construct into HFK. The data from luciferase reporter assays demonstrated that neither wild type hTERT nor mutant hTERT-D868A activated the hTERT promoter (Figure 4A). This is consistent with the lack of endogenous telomerase activity in early-passage cells transduced with the hTERT-D868A mutant (Figure 3B). HPV E6 was used as positive control in this experiment (Figure 4A). We also confirmed that both wild type and mutant hTERT proteins were biologically active in this assay and able to activate the cyclin D promoter (Figure 4B) as described previously [30]. Together, these findings suggest that inactive hTERT proteins (in collaboration with E7) can mediate transit through crisis (the M2 phase of cell), but that continued cell proliferation correlates with increased endogenous hTERT expression, identical to what is observed in cells immortalized by the E6 protein. The implications of these findings are considered in the Discussion.
Since previous studies have defined extra-telomeric functions of hTERT, we attempted to identify a potential telomere-independent mechanism to explain the ability of the inactive hTERT to immortalize cells, with a specific focus on cellular gene expression. Given the conflicting reports of hTERT on gene expression in various model systems [32], [33], [34], [45], [46], [47], [48], it was important that we examined hTERT effects in primary keratinocytes. We therefore stably expressed E6, wild-type hTERT (hTERTwt), or hTERT-D868A in primary HFKs and conducted array-based whole genome expression analysis (Figure S3, and Dataset S1). Because E6 is a known activator of the hTERT protein [8], expression changes shared by hTERT and E6 could represent hTERT-dependent E6 targets. As expected, significant changes in mRNA expression in E6 cells were also seen in cells with wt hTERT (1379 of 6991, or 20% of E6 changes with fold change >1.33 and p-value <0.01) (Figure 5A). More than half of the wt hTERT changes (58%, 1379/2359) were also seen in E6 cells, suggesting that changes seen in hTERT-expressing cells are also altered by E6-expressing cells, possibly through an hTERT-dependent pathway.
To pursue whether the mRNA expression changes seen in wt hTERT cells were dependent on changes in telomere biology, we also expressed the catalytically inactive mutant hTERT-D868A in primary HFKs. A total of 2359 mRNA probe sets were altered in wt hTERT HFKs compared to 5467 changes in hTERT-D868A HFKs (Figure 5B). Interestingly, 2077 of the 2359 (88%) of the RNA probes altered by wt hTERT were also altered by hTERT-D868A. Thus, the gene expression alterations seen following hTERT expression are largely independent of reverse transcriptase function.
1258 changes in mRNA expression were shared by wt hTERT, hTERT-D868A, and E6 (Figure 5C, Dataset S1). Thus, additional considerations were required to focus our study (Figure 5C). The wt hTERT/hTERT-D868A overlapping gene set was submitted for analysis using Database for Annotation, Visualization and Integrated Discovery (DAVID) [49]. Hierarchical clustering of the 2077 catalytically-independent changes was used to identify 408 gene clusters that were visualized as a heatmap [50] (Figure S4). Based on enrichment scores, genes associated with “Chromosome Organization”, “Chromatin Organization”, and “Chromatin Modification” were grouped with the SP-PIR Keyword “Chromatin Regulator” (Dataset S2). Of particular interest in the enriched chromatin regulation cluster was the gene Bmi1, which exhibited significant increases in two overlapping probe sets (Figure 5D, probe set IDs L13689 and NM_005180). We verified by RT-PCR that E6, wt hTERT and hTERT-D868A expression led to 5–7 fold increases in Bmi1 transcript levels (Figure 5E). More important, these three genes also increased Bmi1 protein expression (Figure 6A, B).
To further validate the RT-independent ability of hTERT to induce Bmi1, we analyzed two additional mutated hTERT constructs that lacked telomerase activity. Both of these telomerase mutants increased Bmi1 transcript levels, similar to hTERT-D868A mutant (Figure S5), further substantiating the telomerase-independent activity involved in Bmi1 induction.
Since Bmi1 expression is increased acutely by hTERT, we investigated whether Bmi1 levels remained increased in late-passage HFKs immortalized by hTERT/E7. We doubly transduced HFKs with hTERT and E7 and propagated the cells beyond the time when they would normally enter crisis. Bmi1 protein was shown to be increased in early passage (28 population doublings, PD 28) hTERT/E7 immortalized HFKs and remained high after serial passaging (PD 204) (Figure 6C). We also confirmed increased Bmi1 protein expression in hTERT/E7 immortalized HFKs by immunohistochemistry (Figure 6D). Compared to control HFKs, Bmi1 was significantly increased in the nuclei of early passage (PD42) and late passage (PD220) hTERT/E7 immortalized HFKs (Figure 6D). Together, these data indicate that Bmi1 protein is not only increased acutely by hTERT in primary keratinocytes but that its increased expression is maintained in late passage immortalized cells.
Given the correlative data between hTERT and Bmi1 expression and a previous study showing that Bmi1 immortalizes mammary and oral epithelial cells [51], we speculated that Bmi1 might substitute for hTERT and immortalize human keratinocytes. To test this, we transduced HFKs with Bmi1 and E7 together or separately, as well as with empty vector. Cells were passaged to determine the growth rate and lifespan of these cell populations (Figure 7A). As expected, the HFKs infected with empty vector alone failed to reach 25 population doublings. Introducing E7 or Bmi1 alone extended lifespan by approximately 15 population doublings, as previously described for E7 [8]. However, cell immortalization (>50 population doublings) was observed only when Bmi1 and E7 were co-expressed (Figure 7B). It is noteworthy that the Bmi1/E7 cells also exhibited shortened telomere length, similar to E6/E7 cells. Thus, these data indicate that Bmi1, like hTERT and E6, is able to cooperate with E7 in the immortalization of HFKs.
Although Bmi1 was increased in immortalized, non-tumorigenic HFKs, we queried whether its expression might be altered in cervical cells and during the progression to cervical cancer. We therefore first examined Bmi1 expression in immortalized and tumorgenic cervical cell lines. Consistent with our data in HFKs, expression of E6/E7 in primary human ectocervical cells (HECs) leads to immortalization, enhances endogenous hTERT expression, and increases in both Bmi1 mRNA and protein (Figure 8A, Figure S6). Bmi1 protein levels also showed increases in the tumor-derived, telomerase-positive HeLa cancer cell line (Figure 8A, 1A). These data indicated that Bmi1 is increased in HPV-immortalized and tumorigenic cervical cells. Furthermore, immunohistochemical (IHC) staining of cervical cancer tissue specimens demonstrated increased Bmi1 levels in invasive lesions (Figure 8B).
Given the evidence demonstrating increased expression of Bmi1 in immortalized cells and tumorigenic cervical cancer cells and the observation that even invasive cervical tumors overexpress Bmi1, we examined whether Bmi1 expression correlated with the severity of cervical cancer progression. Bmi1 levels were evaluated by IHC in 21 cervical tissues identified by pathological review as either normal, precancerous, or cancerous lesions. Bmi1 was observed in modest amounts in the nucleus of cells in the basal and immediate suprabasal epithelium in normal cervical tissue (Figure 8D ii), CIN1, and CIN2 (Figure 8D iv, vi). However, a striking increase in staining was observed in the epithelial layers of CIN3 and invasive carcinoma (Figure 8D viii, x). For cases of invasive cervical carcinoma, Bmi1 staining was specific to cancerous lesions (Figure 8D x). To quantify Bmi1 expression, intensity and positivity scores were determined (Figure 8C) with the mean and standard deviation shown (Figure 8E). CIN3 and invasive carcinoma show combined scores that are significantly higher than both CIN1 and CIN2 (CIN3 vs. CIN1, p = 0.0083; CIN3 vs. CIN2, p = 0.0372; invasive vs. CIN1, p = 0.0012; invasive vs. CIN2, p = 0.0130; p values as calculated by student t-test). These data indicate that Bmi1 expression does indeed correlate with the degree of cervical dysplasia and with progression to cancer.
Overall our findings demonstrate that a novel extra-telomeric and non-catalytic function(s) of hTERT contributes to cell immortalization by hTERT and E7 in human keratinocytes. These findings not only define new properties of hTERT that contribute to cell immortalization, but they potentially modify our concept of the mechanism by which E6 is mediating cell immortalization. Using our current and published data, we have constructed a summary table listing the properties of HFK expressing the various hTERT mutants and HPV oncogenes (Table S1).
While we have shown that E6 and E7 are required for the efficient cell immortalization of primary cells, one study has shown that E7 immortalizes human keratinocytes at a very low efficiency when cultured in serum-free synthetic medium [8], [52] and another study has shown that wild type HPV16 E6 as well as an natural mutant E6 are able to immortalize human keratinocytes [53]. However, it is important to note that when E6 and E7 are used alone, an obvious “crisis” period or “flat” phase of cell growth is observed, indicating that cell immortalization is infrequent and arises from a small subpopulation of cells. Most likely additional genetic or epigenetic changes are required for escape from “crisis”. Although we have also noted that E7 is highly efficient for immortalizing keratinocytes without a “crisis” period of time when co-cultured with feeders or conditioned medium (Liu, X., et al. unpublished data), this is best explained by the ability of feeder cells or conditioned medium to induce telomerase [54].
A basic tenet of cell immortalization is that hTERT reverse transcriptase activity is essential for maintaining or elongating telomeres, thus allowing for continued cell replication [11]. However, our early studies showed that E6-induced telomerase activity could be dissociated from telomere maintenance [10]. Supporting this hypothesis, the current study clearly indicates that immortalized cells exhibit similar levels of telomerase activity, yet telomeres shorten during cell passaging and stabilize at late passages (Figure 1B, 1C). We have also demonstrated the same pattern of telomere length in keratinocytes immortalized by a Rho kinase (ROCK) inhibitor [55], [56], [57]. These data suggest that extra-telomeric functions of telomerase or hTERT play a role during cell immortalization independent of telomere maintenance.
Several reports indicate that hTERT-HA fails to elongate telomeres and immortalize human fibroblasts [41], [42] or HA1 cells [40] despite a high level of telomerase activity. Surprisingly, we found that hTERT-HA reproducibly and efficiently immortalized HFK cells in cooperation with HPV E7 [26]. Even more interesting, these immortalized HFK cells had short telomeres (Figure 2B). We also observed the same results with the hTERT N+T mutant, which is positive for telomerase activity, but negative for telomere recruitment and elongation (Figure 3). Here, we demonstrate in keratinocytes that telomerase activity is not required, as the catalytically-defective hTERT-D868A retains its ability to immortalize HFKs in cooperation with E7.
We have also shown that immortalization is independent of the telomere lengthening function of hTERT. Telomere lengthening is predominantly carried out by telomerase, but can also occur via the alternative telomere lengthening (ALT) pathway [58]. However, the exact mechanisms of telomere maintenance or elongation remain elusive. Studies have suggested that many mechanisms, including enzymatic activity, telomere-capping, and recombination, may play roles in the final stabilization of telomeres in immortalized and human cancer cells [42], [58], [59], [60], [61], [62], [63].
A number of non-canonical functions for hTERT have been reported in literature, and the list is increasing rapidly [27], [28], [29], [30], [31], [33], [34], [47], [48], [64]. This led us to pursue whole genome expression studies to probe altered signaling pathways in primary cells expressing hTERT. Surprisingly, our data have shown that 88% of the genes altered by wild-type hTERT are also altered in the same direction by hTERT-D868A (Figure 5B). Somewhat surprisingly, hTERT-D868A regulates about twice as many genes as wild-type hTERT, suggesting that elimination of the catalytic function of hTERT actually augments non-canonical functions.
It is critical to note that bypass of the Hayflick limit by enzymatic-defective hTERT mutants is accompanied by the global induction of many cellular genes, including endogenous hTERT (Figure 3C). This induction of endogenous hTERT is not due to the direct, acute transactivation of endogenous hTERT by mutant hTERT (Figure 3A and 4A). Rather the endogenous hTERT activation is part of the larger number of gene sets that are increased during transit through M1/M2 restriction points.
A very recent study [47] demonstrates the existence of a splice-variant of hTERT (Δ4-13) containing an in-frame deletion of exons 4 through 13 that encode the catalytic domain of telomerase. This variant is expressed in telomerase negative normal cells and tissues as well as in transformed telomerase positive cell lines and in cells that employ an alternative method to maintain telomere length. The overexpression of the Δ4-13 significantly elevated the proliferation rate of several cell types without enhancing telomerase activity, while decreasing the endogenous expression of this variant using siRNA technology reduced cell proliferation. The expression of the Δ4-13 variant stimulates Wnt signaling. This is the first report that a naturally occurring hTERT splice variant that lacks telomerase activity exhibits an ability to stimulate cell proliferation, supporting our conclusions that non-canonical hTERT functions contribute cell immortalization.
We have also used real-time RT-PCR to validate more than 20 genes that were altered greater than two-fold as determined by microarray analysis. The RT-PCR results were virtually identical to the microarray data, and several of these genes are critical regulators of keratinocyte growth, apoptosis, and differentiation. Bmi1 was one such target of hTERT. Our microarray data also demonstrated that E6 and hTERT increased the expression of RB and ROCK1 mRNA in HFKs (Dataset S2) and this increased expression was confirmed by real time RT-PCR. This induction of RB is presumably counteracted by the activity of the E7 protein. There are similar parallel events in the cooperation of the E6 and E7 proteins in cell immortalization and transformation [6], [7], [8]. For example, these two genes seem to have evolved both complementary and opposing functions that are necessary to prevent senescence and/or apoptosis. For example, while E7 stabilizes p53 protein, E6 degrades this tumor suppressor protein. Similarly, while E6 stabilizes RB protein, E7 inactivates and destabilizes it. The yin-yang regulation of E6/E7 functions and telomerase and RB/16 pathways may be critical for fine tuning the growth and differentiation of keratinocytes as well as for regulating the viral replication cycle.
Bmi1 has been identified as a marker of cancer progression in a number of carcinomas, including those derived from the nasopharynx, breast, pancreas, and other sites [65], [66], [67]. Equally important, hTERT has been identified a potential universal cancer target, since it is up-regulated in most cancers [15], [68]. Beyond identification of Bmi1 in our genetic screen, hTERT and Bmi1 have been linked in several previous reports. PcG components Bmi1 and SIRT1 have been shown to be altered in hTERT-expressing urothelial cells [45], and Bmi1 has been shown to induce endogenous telomerase in human mammary epithelial cells [35]. Besides Bmi1, other chromatin remodeling complex members have also been associated with the non-telomere effects of hTERT, including transcriptional regulation of Wnt targets by binding BRG1, a Trithorax group protein (TrxG) [30]. Our data links hTERT expression to changes in Bmi1, suggesting that there is an hTERT-Bmi1 signaling pathway. In this study, we have shown that Bmi1 overexpression occurs prior to full transformation since both hTERT-expressing HFKs and multiple types of telomerase-positive immortalized cells (Figure 5E, Figure 6, & Figure 8A) overexpress Bmi1.
The above Bmi1 findings may have clinical relevance. Our in vivo studies reveal a differential expression of Bmi1 in carcinoma in situ and invasive carcinomas compared to preneoplastic lesions (Figure 8B–E). Indeed, Bmi1 mRNA expression is increased in cervical cancer compared to corresponding noncancerous tissues [69]. Additionally, Bmi1 overexpression has been significantly correlated with tumor size, clinical stage, and regional lymph node metastases in cancers of the cervix [70]. Another PcG protein, EZH2, was also recently shown to be up-regulated in high grade squamous cervical intraepithelial lesions (HSILs) compared to normal cervical epithelium [71], further implicating chromatin remodeling changes in tumor initiation and progression.
While Bmi1 appears to be a significant contributor to cell immortalization, it is also obvious that other genes detected in the mRNA expression screen may also contribute to this process. Indeed, the known anti-apoptotic activity of hTERT might also be expected to assist in the bypass of cellular senescence.
pLXSN vector and pLXSN-16E6, pLXSN-16E7, pLXSN-16E6E7 were as described previously [10], [21], [22], [72]. pBABE-puro-hTERT, pBABE-puro-hTERT-N+T [42], pBABE-puro-hTERT-D868A [43] were gifts from Dr. Elizabeth Blackburn, pBABE-puro-hTERT-HA [40] from Dr. Robert Weinberg, and pCLMSCV-puro-Bmi1 [36] from Dr. Tohru Kiyono. Other hTERT mutants were made using the QuikChange XL Site-Directed Mutagenesis Kit (Stratagene, La Jolla, CA). An N-terminal double Flag epitope tag was added to hTERT using a PCR insertion method. SD3443 retrovirus packaging cells were transfected with pLXSN vectors or pBABE-puro vectors described above using LipofectAmine 2000 (Invitrogen) as instructed. Culture supernatants containing retrovirus were collected 48 hours after transfection.
Primary human foreskin keratinocytes (HFKs) and human foreskin fibroblasts (HFFs) were isolated and cultured from neonatal foreskins as described58. Primary human ectocervical keratinocytes (HECs) were derived from fresh cervical tissue similarly and obtained after hysterectomy for benign uterine diseases. Standard trypsinization procedures were used to isolate the keratinocytes, which were cultured in serum-free keratinocyte medium supplemented with 50 µg/ml of bovine pituitary extract and 25 ng/ml of recombinant epidermal growth factor (Invitrogen). The cells were cultured in serum-free keratinocyte growth media (Invitrogen) supplemented with gentamycin (50 µg/ml). Primary HFKs, HFFs, and HECs were transduced with amphotropic pLXSN retroviruses expressing HPV-16E6, E7, or both E6 and E7 and/or pBABE-puro retroviruses expressing hTERT or its mutants (see above). Retrovirus-transduced cells were selected in G418 (100 µg/ml) for 5 days and/or puromycin (2 ug/ml) for 3 days. Resistant colonies were pooled and passaged every 3–4 days (1∶4 ratio for HFKs and HECs, 1∶8 ratio for HFFs). HeLa, C33A, SiHa cells were maintained in complete DMEM medium. All cells were cultured on plastic tissue culture dishes or flasks.
To prepare cells for immunocytochemistry, cells were pelleted and then fixed with 4% paraformaldehyde solution overnight and resuspended in HistoGel (Richard-Allan Scientific) at a ratio of 1∶1 per volume. The gel matrix was processed through graduated alcohols and Clear-Rite 3 (Richard-Allan Scientific) for paraffin embedding using the Leica ASP300 system (Leica Microsystems, Wetzlar, Germany). Paraffin sections were cut at 5 µm and mounted on Superfrost Plus slides (Fisher Scientific).
Patient samples were acquired through the Histopathology and Tissue Shared Resource at the Lombardi Comprehensive Cancer Center (Washington, DC). Twenty one cervical tissues were acquired which represented different pathological stages–one normal tissue core and five tissue cores for each of the following pathological stages: Cervical Intraepithelial Neoplasia Stage 1 (CIN1), CIN2, CIN3 or carcinoma in situ, and invasive cervical carcinoma.
Human keratinocytes and fibroblasts were lysed and analyzed by Q-TRAP [72] with SYBR Green Supermixture (Bio-Rad). A standard curve was produced for the real-time Q-TRAP assay using serially diluted HeLa cell extracts. All samples were run in triplicate.
Genomic DNA was extracted from cells using Qiagen DNeasy Blood & Tissue Kit. Average telomere length was assessed by a modified method of the real-time PCR–based telomere assay [73]. Briefly, the telomere repeat copy number to single gene copy number (T/S) ratio was determined using the Bio-Rad IQ5 thermocycler in a 96-well format. Five nanograms of genomic DNA was subjected to PCR reactions with Bio-Rad SYBR Green Super mixture. The primers for telomere length and HBG1 (a single copy gene) were as below:
Tel-1 5′ CGGTTTGTTTGGGTTTGGGTTTGGGTTTGGGTTTGGGTT-3,
Tel-2 5′-GGCTTGCCTTACCCTTACCCTTACCCTTACCCTTACCCT-3′;
HBG1 5′- TGTGCTGGCCCATCACTTTG,
HBG2 5′- ACCAGCCACCACTTTCTGATAGG-3′.
The reactions proceeded for 1 cycle at 95°C for 5 min, followed by 41 cycles at 95°C for 15 s, 60°C for 45 s. All samples for both the telomere and HBG1 reactions were done in triplicate. In addition to the samples, each 96-well plate contained a six-point standard curve from 0.0, 0.2, 1.0, 5.0, 25.0, 125.0 ng using genomic DNA (telomere length 10.4 kb) from Roche Telo-kit. The T/S ratio (dCt) for each sample was calculated by normalizing the average HBG1 Ct value from the average telomere Ct value.
1×105 telomerase-negative HFKs were seeded onto 24-well plates and grown overnight. Transient transfections were performed using LipofectAmine 2000 reagent (Invitrogen) according to the protocol provided by the manufacturer. Cotransfections were performed using 0.5 ug of a core hTERT or Cyclin D1 promoter reporter plasmids and 50 ng of each expression vector as indicated (HPV16E6, hTERTwt or hTERT-D868A) or empty vectors as control for basal promoter activity. Cells also were cotransfected with 2 ng of the pRL-CMV plasmid (Promega), which contains the Renilla reniformis luciferase gene as a transfection control. Firefly and Renilla luciferase activities were measured 24 hr after transfection using the Dual luciferase reporter assay system (Promega).
hTERT, hTERT-D868A, HPV E6 or the pBP vector was stably expressed in primary HFKs. Cells were grown on 100 mm tissue culture dishes (BD Falcon) to confluency before harvesting RNA with 1 mL TRIzol Reagent according to manufacturer's protocol. DNAse treatment was performed (Ambion, Austin, TX). RNA was sent to MOGene, LC (St. Louis, MO) for microarray analysis. E6, hTERT, and hTERT-D8686A were run separately against the pBP on a two-color Agilent whole human genome slide with a 4 x 44K format. A total of six comparative arrays were run- hTERT, hTERT-D868A, or E6 vs. empty vector and run with a duplicate for dye swap. RNA was amplified using the Agilent Low Input Linear Amplification kit (Agilent Technologies, Santa Clara, CA), and then labeled with either cyanine-5 or cyanine-3 using the ULS RNA Fluorescent Labeling Kit (Kreatech Biotechnology, Amsterdam, The Netherlands). 825 ng each of labeled c-DNA was hybridized overnight at 65°C in an ozone-free room to protect the label. All washes and hybridization conditions followed were consistent with the Agilent processing manual (protocol version 4.0). Arrays were scanned using an Agilent C scanner and extracted using the Agilent Feature Extraction software 10.7.1 (Agilent Technologies, Santa Clara, CA). Initial data analysis was performed by MOGene using the Rosetta Luminator software (Agilent). Expression arrays were submitted to DAVID Bioinformatics Resources 6.7 (NIAID, NIH) for Functional Annotation Clustering [49]. Using the MultiExperiment Viewer v4.8 (TM4 Microarray Software Suite, Rockville, MD) and data from the four hTERT comparative arrays, a heat map was constructed with the cluster of interest [50].
SuperScript III Reverse Transcriptase kit (Invitrogen) was used to perform reverse transcription PCR (RT-PCR), as previously described [72]. Reactions were annealed and analyzed using a Bio-Rad iCycler and accompanying software (Bio-Rad Laboratories). Primer sets used include the following:
Bmi1-F: 5′ TGCCCAGCAGCAATGACTGT3′
Bmi1-R: 5′ GTCCATCTCTCTGGTGACTGATCTTC3′
GAPDH-F: 5′ TCTCCTCTGACTTCAACAGC3′
GAPDH-R: 5′ GAAATGAGCTTGACAAAGTG3′
Stable cell lines were lysed in 2X SDS gel electrophoresis sample buffer. Proteins were separated on a 4–20% Tris-glycine gradient gel (Invitrogen) and electrophoretically transferred to an Immobilon-P PVDF membrane (Millipore). The membranes were blocked in 5% dry milk-PBST and incubated with pRb antibody (1∶1000, Cell Signaling), hTERT (1∶1000, Y182, Epitomics), Bmi1 (1∶200, F6, Millipore), P53 (1∶1000, Pab 1801, Santa Cruz), and HPV16-E7 (1∶1000, ED17, Santa Cruz); and a secondary antibody with HRP conjugation and detected by chemiluminescence (anti-rabbit IgG or anti-mouse IgG; Santa Cruz Biotechnology). Equal protein sample loading was monitored using an anti-β-actin (1∶5000, Sigma) or anti-GAPDH (1∶2000, FL-335, Santa Cruz). The membranes were visualized by using Western Blotting Chemiluminescence Luminol Reagent (Santa Cruz).
Immunocytochemistry of HFK cell pellets and immunohistochemistry of cervical tissue was performed for Bmi1. Five micron sections from formalin fixed, paraffin embedded tissues were de-paraffinized with xylenes and rehydrated through a graded alcohol series. Heat induced epitope retrieval (HIER) was performed by immersing the tissue sections at 98°C for 20 minutes in 10 mM citrate buffer (pH 6.0) with 0.05% Tween. Immunohistochemical staining was performed using the VectaStain Kit from Vector Labs according to manufacturer's instructions. Briefly, slides were treated with 3% hydrogen peroxide for 10 minutes. Endogenous biotin was blocked using an avidin/biotin blocking kit from Invitrogen. The slides were then treated with 10% normal goat serum and exposed to primary antibodies for Bmi1 (1∶200, F6, Millipore) for 1 hour at 22°C. Slides were exposed to appropriate biotin-conjugated secondary antibodies (Vector Labs), Vectastain ABC reagent and DAB chromagen (Dako). Slides were counterstained with Hematoxylin (Fisher, Harris Modified Hematoxylin) at a 1∶17 dilution for 2 minutes at RT, blued in 1% ammonium hydroxide for 1 minute at 22°C, dehydrated, and mounted with Acrymount. Consecutive sections with the omitted primary antibody were used as negative controls.
To quantify expression of immunohistochemical staining, slides were subjected to a randomized, blinded scoring performed by a board-certified clinical pathologist. Combined scores were calculated by adding the intensity score and positivity scores. Mean and standard deviation of combined scores were calculated. A subset of slides was scored multiple times to demonstrate reproducibility. Each case received an intensity score from 0–3 (0 = negative, 1 = weak, 2 = moderate, 3 = intense) and the percentage of positive cells was recorded, which was converted to a tiered positivity score (0 = less than 10%, 1 = 11–49%, 2 = 50–74%, 3 = 75–100%).
HFKs were transfected with either wild-type hTERT or an hTERT mutant, and then grown on sterile glass cover slips, fixed in 4% (wt/vol) paraformaldehyde, and labeled with the primary and secondary antibodies. The following primary antibodies were used: anti-hTERT (Rockland 1∶500 dilution) and anti-hTERT serum from rabbit immunized with KLH conjugated Ac-CSRKLPGTTLTALEAAANPAL-amide (aa1104-1123). The secondary antibodies, AlexaFluor 488 donkey anti-mouse IgG and AlexaFluor 555 donkey anti-rabbit IgG (Invitrogen) were used at a concentration of 5 µg/mL. A Zeiss Axioskop microscope and a HAMAMATSU ORCA-ER Digital Camera were used for visualization and microphotography.
The HFK cells were prepared from human neonatal foreskins at Georgetown University Hospital, normally these tissues are de-identified and discarded. The cervical tissue samples from the Histopathology and Tissue Shared Resource at the Lombardi Comprehensive Cancer Center were anonymized. These protocols (2002-021 and 1992-048) have been approved by the Georgetown University Institutional Review Board.
BMI1 (NM_005180), BRG1 (NM_003072), EGFR (NM_005228), EZH2 (NM_004456), FGF2 (NM_002006), HPV16 E6 (NP_041325), HPV16 E7 (NP_041326), hTERT (NM_198253), P16 (NM_058195), RB (NM_000321), SIRT1 (NM_012238), TP53 (NM_000546), VEGF (NM_003376).
|
10.1371/journal.pbio.1002319 | Prolonged Intracellular Na+ Dynamics Govern Electrical Activity in Accessory Olfactory Bulb Mitral Cells | Persistent activity has been reported in many brain areas and is hypothesized to mediate working memory and emotional brain states and to rely upon network or biophysical feedback. Here, we demonstrate a novel mechanism by which persistent neuronal activity can be generated without feedback, relying instead on the slow removal of Na+ from neurons following bursts of activity. We show that mitral cells in the accessory olfactory bulb (AOB), which plays a major role in mammalian social behavior, may respond to a brief sensory stimulation with persistent firing. By combining electrical recordings, Ca2+ and Na+ imaging, and realistic computational modeling, we explored the mechanisms underlying the persistent activity in AOB mitral cells. We found that the exceptionally slow inward current that underlies this activity is governed by prolonged dynamics of intracellular Na+ ([Na+]i), which affects neuronal electrical activity via several pathways. Specifically, elevated dendritic [Na+]i reverses the Na+-Ca2+ exchanger activity, thus modifying the [Ca2+]i set-point. This process, which relies on ubiquitous membrane mechanisms, is likely to play a role in other neuronal types in various brain regions.
| The accessory olfactory system is essential for chemical communication in animals during social interactions. During this process, the principle cells of the accessory olfactory bulb (AOB) may respond to transient stimulation with prolonged activity, sometimes lasting for minutes—a property known as persistent activity. This property, which has been observed in other brain areas, is usually attributed to positive feedback mechanisms either at the cellular or the network level. Here, we show how persistent activity can emerge without feedback, relying on slow changes in internal ionic concentrations, which keep a record of past neuronal activity for long periods of time. We used a combined computational and experimental approach to show that the complex interaction between various ions, their extrusion mechanisms, and the membrane potential leads to stimulus-dependent persistent activity in the AOB. The same mechanism may apply to other neuronal types in various brain regions.
| The accessory olfactory system, also known as the vomeronasal system, mediates chemical communication between conspecifics of most mammalian and reptilian species during social interactions [1]. Inputs to this chemosensory system originate from the sensory neurons of the vomeronasal organ (VNO) that synapse on the mitral cells of the accessory olfactory bulb (AOB), which provide the output of the bulb [2]. Previously, we have shown that AOB mitral cells in vitro respond to brief afferent nerve stimulation with persistent firing activity lasting several minutes [3].
Persistent activity, defined as the ability of neurons to remain active in the absence of external inputs, was documented in many brain areas. Such activity enables the brain to maintain an internal state without continuous external input. It has been suggested that persistent activity is a neuronal correlate of working memory [4], and that it can mediate neuronal integration over long time scales [5].
The time scale of persistent activity (>1 min) is much longer than that of most biophysical mechanisms (typically 0.5–100 ms). Most attempts to explain how the extremely prolonged time scales of persistent activity emerge from such rapid biophysical processes have involved feedback mechanisms [6]. Such feedback can be implemented with recurrent excitation at the network level [7–9], or alternatively, by biochemical pathways at the cellular level. An example of the latter is the mechanism proposed to underlie persistent activity in the entorhinal cortex [10,11] and hippocampal CA1 pyramidal neurons [12,13]. The mechanism involves an interaction between Ca2+ influx during spiking and a calcium-activated non-selective (CAN) cation conductance that depolarizes the cell. However, theoretical models of prolonged spiking based on feedback mechanisms are hard to construct in a way that is robust to small parameter changes, immune to noise and continuously graded [10,14–16].
Persistent activity in AOB mitral cells was shown to depend upon Ca2+ influx and CAN conductance. However, this intrinsic cellular mechanism does not depend on a feedback cycle involving ongoing neural activity, as persistent firing readily resumes after a temporal firing cessation [3].
In the present study we combined electrophysiological, imaging, and computational approaches to explore the mechanisms underlying persistent firing in AOB mitral cells. We describe a novel mechanism involving interplay between homeostatic processes controlling intracellular Na+ and Ca2+ concentrations. This novel mechanism, which does not rely upon feedback, is both resistant to noise and allows multiple stable firing states.
Prolonged firing activity of AOB mitral cells was demonstrated in behaving mice during social investigation of conspecifics [17]. It has remained unclear whether this sustained activity reflects the continuous detection of the stimulus or network properties. In order to explore this issue, we examined AOB responses in anesthetized mice following well-controlled chemosensory stimulus application to the VNO (Fig 1A and 1B, S1 Fig) [18].
While response dynamics often matched those attributed to the vomeronasal pump [18,19], in other cases, elevated firing rates remained high well beyond this time scale, sometimes even after the stimulus was flushed from the nasal cavity and the VNO. Under a highly strict statistical criterion (see Data Analysis in Materials and Methods), reliable cases of persistent activity were found in about one percent (n = 7) of the recorded units, and were associated with a particular stimulus, while other stimuli elicited only transient response in the same cells (Fig 1A and 1B). This stimulus selectivity is consistent with a requirement for a high level of activation to trigger the prolonged firing (see below).
Similarly, prolonged single unit AOB spiking activity could be readily elicited in anesthetized mice by direct stimulation of the vomeronasal nerve with a metal electrode (Fig 1C), further confirming that the sustained responses are independent of VNO dynamics. Finally, In agreement with our previous study [3], persistent firing could be elicited in AOB mitral cells in brain slices. An example is shown in Fig 1D (top), where a 4 s train of action potentials is followed by a prolonged period of persistent spiking at a rate of 1–3 Hz lasting for over a minute. The reproducibility of this firing epoch is demonstrated by the mean rate response for the three recorded cells (Fig 1D, bottom). Altogether, these results and the results of our previous studies [3,20] prove that AOB mitral cells are capable of persistent firing responses, both in vitro and in vivo to either electrical or chemical sensory stimulation.
Conducting the in vitro protocol described above while shifting the membrane potential to −60 mV (Fig 2A) blocked the persistent firing and unmasked a prolonged depolarization with a similar time course as the firing activity (compare to Fig 1D). To analyze the currents underlying the prolonged depolarization, the hybrid-clamp methodology was used (see Materials and Methods). Cells were voltage-clamped to −80 mV and trains of action potentials at various frequencies were delivered during a 4 s long current-clamp period. The evoked inward current (Fig 2B) comprised an initial, rapidly decaying phase (transient phase, enlarged in Fig 2C), followed by a second, prolonged phase (persistent phase), that peaked after >10 s (Fig 2B, arrows) and slowly decayed with a more prolonged time course (>30 s). Notably, the charge transfer during each of the phases monotonically increased with the stimulus frequency (Fig 2D). Thus, the prolonged inward current underlying persistent firing in AOB mitral cells seems to involve transient and persistent components that are proportional to the firing frequency during the stimulation.
The complex dynamics of the prolonged inward current suggest that multiple biophysical mechanisms are involved. To isolate the participating processes, we abolished the inward current, previously shown to be mediated by Ca2+-dependent, CAN conductance [3]. Removal of Ca2+ from the extracellular solution, as well as blocking the increase in [Ca2+]i by adding 5 mM BAPTA to the pipette solution (S2 Fig), abolished the prolonged inward current. Under these conditions, the stimulating train was followed by an outward current of 18±3 pA that monotonically decayed with a single time constant (τ = 5±1 s, n = 11 cells, Fig 2E, green trace). Subtracting the outward current from the control condition current (Fig 2E, blue trace) yields a net inward current (Fig 2E, black trace), which is likely due to the CAN conductance. Similar results were previously obtained by blocking N/R type voltage-sensitive Ca2+ channels [3].
As shown in Fig 2F, the outward current measured in the absence of Ca2+ ions was independent of membrane potential, suggesting that it is not mediated by ionic conductance. The most likely candidate for a voltage-insensitive outward current is an ionic pump current, such as the one produced by the plasma membrane Na+-K+ pump (Na+-K+ ATPase) [21]. As shown in Fig 2G (green), blocking the Na+-K+ pump using ouabain unmasks a strong net inward current peaking immediately after the spike train. The difference between the currents before and after ouabain application is a net outward current resembling the one measured in the absence of Ca2+ ions (Fig 2G, black). Thus, the outward current is most likely mediated by the Na+-K+ pump.
Overall, these data suggest that the complex dynamics of the prolonged inward current reflect the sum of two opposing currents—a voltage-independent outward current (Na+-K+ pump) decaying over a few seconds and a prolonged Ca2+-dependent inward current (ICAN) that remains active for minutes.
To study the spatio-temporal relationship between [Ca2+]i and ICAN, we correlated [Ca2+]i indicator fluorescence in various cellular compartments with the simultaneously recorded somatic inward current. To that end, AOB mitral cells were filled with a Ca2+ indicator using the patch pipette (inset in Fig 3). Then, tuft fluorescence (Fig 3A) and the corresponding inward currents (Fig 3B) were simultaneously monitored as trains of action potentials at various frequencies were delivered via the patch pipette to activate the neurons. The transient increase in [Ca2+]i in the dendritic tuft was followed by an extremely prolonged decay lasting longer than the interval between stimuli, resulting in summation of [Ca2+]i levels over consecutive trains (Fig 3A). Similarly, the prolonged inward current also persisted longer than the interval between stimuli, resulting in progressive increase in inward current as well (Fig 3B).
To analyze the relationship between dendritic [Ca2+]i and the inward current, the current amplitude at each time point was plotted against the simultaneously measured fluorescence level. Fig 3C shows this analysis, applied to the data shown in Fig 3A and 3B. As apparent, the inward current shows a clear sigmoidal dependence on the fluorescence signal, suggesting that the [Ca2+]i in the dendritic tuft tightly correlates with the slow dynamics of inward current (see S3E and S3F Fig for more examples). In contrast to tuft fluorescence, somatic fluorescence does not correlate with the magnitude of the inward current (Fig 3D, S3A and S3D Fig). The close relationship between tuft [Ca2+]i and the magnitude of the inward current suggests that the prolonged current reflects the extended elevation of tuft [Ca2+]i. Indeed, close examination of tuft fluorescence levels (Fig 3E) revealed that the decay of [Ca2+]i in the tuft followed two distinct time scales: fast initial decay (τ = 1.9 s, mean of three cells) followed by very slow decay (τ = 47.0 s). This slow process suggests that [Ca2+]i is in a quasi-stable state, the level of which is determined by the stimulus frequency. Consistent with this, increasing the stimulation frequency from 15 Hz to 30 Hz almost doubled the quasi-stable state level (Fig 3E, blue and green traces).
Overall, these results suggest that the inward current underlying persistent firing of AOB mitral cells is mediated by dendritic Ca2+-dependent ionic conductance (CAN) and that its slow dynamics likely reflect a complex interaction between several ionic extrusion mechanisms.
The result described above, in which the tuft [Ca2+]i decays to a quasi-stable state determined by the stimulation frequency, suggests that the quasi-stable state is generated by slowly changing, activity-dependent quantity. One such quantity may be the tuft [Na+]i, which affects the Ca2+ dynamics by interacting with ionic transport mechanisms such as the Na+-Ca2+ exchanger. This exchanger, which is the major mechanism for control of large excess Ca2+ [22], uses the Na+ electrochemical gradient to extrude Ca2+. Thus, increase in [Na+]i which leads to a decreased Na+ gradient, reduce or even reverse the Ca2+ flux through the exchanger [23]. We examined this possibility in a simple abstract dynamical model, with a minimal number of parameters (Fig 4A; see S1 Text for a description of the model equations). In this model, [Ca2+]i and [Na+]i increase at a rate proportional to an abstract “voltage” quantity, given that the “voltage” is above a certain threshold. [Na+]i decays exponentially to zero over time, while [Ca2+]i decays to a level linearly determined by [Na+]i (the quasi-stable state). The “voltage” is a sum of three components: externally applied current, inward Ca2+-dependent current, and outward (negative) Na+-dependent “pump” current.
Fig 4B and 4D shows the results of running this model with a pulse of externally applied current (black bar). As apparent, the “voltage” (Fig 4B) behavior qualitatively resembles the experimental observations (compare to Figs 1D and 2A). This voltage trajectory is due to the changes in [Na+]i and [Ca2+]i (Fig 4C) and the corresponding currents (Fig 4D). Thus, the feasibility of the mechanism suggested above is confirmed by this simple abstract model. In order to further test this hypothetical mechanism and produce quantitative predictions, we incorporated the principles of this mechanism into a realistic conductance-based model.
A realistic conductance-based model (see Materials and Methods), was constructed using the detailed morphology of a single typical mitral cell (Fig 4E and S4 Fig) for which the electrophysiological properties were characterized. The model assumes that active conductances reside in the apical dendrites and dendritic tufts, as well as in the soma and axon initial segment [24], so that [Na+]i increase in these compartments following firing. A novel feature of our model is the incorporation of compartmental [Na+]i as state variables along with longitudinal ionic diffusion. Accordingly, [Na+]i not only sets the local Na+ reversal potential but also affects localized ionic extrusion mechanisms (Na+-K+ pumps, Na+-Ca2+ exchangers). The Ca2+ influx, buffering and extrusion mechanisms (including a simulated Ca2+ indicator), as well as a Ca2+ dependent conductance, were introduced in the dendritic tufts. The spatial distribution of membranal mechanisms in the model is shown in Fig 4E (See Materials and Methods for a link to the model source code and S1 Text for a full description of the model equations and parameters). Using such a model, one can calculate the temporal dynamics of [Na+]i and [Ca2+]I in various cellular compartments.
An evolutionary multi-objective algorithm [25,26] was used to find the biophysical parameters that best fit our electrophysiological observations. An initial evolutionary process was employed to find the best fit to the following measured parameters: the response to a hyperpolarizing current pulse (Fig 5A), the shape of the action potential (Fig 5B), the modulation of the spike amplitude during a strong (350 pA) current injection (Fig 5G), and the I-f curve (Fig 5H). As shown, the model accurately reproduces the behavior of the real cell with respect to these objectives. Notably, the spike amplitude modulation during depolarizing current injection simulated by the model precisely fitted the experimental observations, despite the fact that only a 350 pA current injection was used as an objective (Fig 5E–5G). Importantly, the spike amplitude modulation in the model was the result of [Na+]i accumulation during the spike train and would not be reproduced when [Na+]i accumulation was prevented (S5 Fig).
The goal of the next evolutionary process was to find the parameters that reproduce the tuft Ca2+ indicator fluorescence and the prolonged somatic inward current (see Materials and Methods). Trains of action potentials with frequencies of 1, 15, and 30 Hz were used to activate the model neuron. The simulated dendritic tuft fluorescence and the accompanying prolonged inward current were then compared to the experimental observations. At a frequency of 1 Hz (Fig 6A), the simulation (red line) perfectly reproduced the observed fluorescence signal (blue line). At higher frequencies (15 Hz and 30 Hz, Fig 6B), both measured (blue and green lines) and simulated (orange and red lines) fluorescence levels, rapidly increased to saturation levels during the stimulation (Fig 6B, top green bar). The rapid increase was followed by an equally rapid decline to a low quasi-stable level that strongly depended on the stimulation frequency (Fig 6B). Moreover, the simulated prolonged inward current also closely fit the experimentally measured current (Fig 6C).
In order to assess the sensitivity of the model to changes in its parameters, we created a population of 1,200 model neurons. In each model, each parameter (except the channels' half-activation voltage parameters) was randomly selected from a uniform distribution that spanned between -10% and +10% relative to the original value. We then examined, in each of the models, the predicted prolonged inward currents evoked by 30Hz spike train. The properties of the resulting currents distribute normally (see example histograms for the maximum current in S6A Fig and the residual current after 1 min for the train end in S6B Fig). As apparent from the 80% bounds of the distribution (S6C Fig), this change in parameters did not cause a large deviation from the fit of the model to the experimental data.
A critical validation of the model is its ability to reproduce the persistent firing recorded in AOB mitral cells. Indeed, a train of simulated spikes evoked long lasting persistent activity (Fig 6D) which resembled the experimental observations (Fig 1D). Adding Gaussian current noise introduced variability to the responses that upon averaging reproduced the PSTH observed in vitro (compare Figs 6E to 1D).
Another critical validation is the ability of the model to predict the time course of dendritic [Na+]i, and particularly its rise during stimulation and very slow subsequent decay that maintains a quasi-stable state for the tuft [Ca2+]i. To that end, we used two fluorescent Na+ indicators, sodium-binding benzofuran isophthalate (SBFI) and Sodium Green, to image the dendritic Na+ dynamics following a stimulus train [27]. As shown in Fig 6F, the averaged observed dynamics (thick green line) indeed match the predicted dynamics (red line) (For dF/F signal not normalized by standard deviation and similar results using the Sodium Green indicator, see S7 Fig) As shown in the presence of tetrodotoxin (TTX,blue lines), the signal does relate to opening of voltage-gated Na+ channels. Notably, similar dynamics were previously observed experimentally in cortical pyramidal neurons [27].
Thus, we conclude that our mitral cell model adequately reproduces the experimental observations.
We used this model to examine possible mechanisms underlying the prolonged current responses of AOB mitral cells. We first examined the time course of [Na+]i in two cellular compartments: the axon initial segment (AIS) and the dendritic tuft, while the model neuron was activated by a 4 s train of 15 or 30 Hz (Fig 7A, orange and red traces, respectively). The model predicts that during the stimulus, [Na+]i at the AIS (dashed lines) would reach a very high level (50 and 63 mM for 15 and 30 Hz stimulation, respectively), followed by a relatively rapid decline (τ = 3.85 s) to baseline levels. The large increase in AIS [Na+]i is due to the high density of voltage-gated Na+ channels and the limited volume of the compartment. The relatively fast recovery results from the activity of the Na+-K+ pump and the fast diffusion to the compartments adjacent to the AIS (axon and soma) that serve as diffusion sinks.
In contrast to the AIS, tuft [Na+]i increased only to a moderate level (15 and 17 mM for 15 and 30 Hz stimulation—solid orange and red traces in Fig 7A, respectively), followed by extremely slow exponential recovery over a time course of minutes (τ = 130 s and 115 s for 15 and 30 Hz stimulation, respectively). This slow time course, also observed in real mitral cells using Na+ fluorescent indicators (Fig 6F), stems from the slow diffusion in the thin dendritic process and the low density of Na+-K+ pumps in this compartment.
The prolonged elevated [Na+]i in the dendritic tuft is bound to affect the Na+-Ca2+ exchanger, hence to determine the quasi-stable state of [Ca2+]i. We examined this prediction using the model by calculating the stable-state [Ca2+]i for various fixed [Na+]i. values. As shown in Fig 7B (inset), an increase in [Na+]i elevated the stable-state [Ca2+]i non-linearly. This relationship was used to calculate the quasi-stable state of [Ca2+]i during and after the stimulation train, based on the instantaneous values of [Na+]i. As shown in Fig 7B for stimulation frequencies of 15 Hz and 30 Hz, the simulated tuft [Ca2+]i (solid lines) quickly dropped to its quasi-stable state level (dashed lines), and then closely followed the slow decrease of the quasi-stable state. The proposed mechanism that maintains [Ca2+]i in a quasi-stable state level is demonstrated in Fig 7C and 7D, where the Ca2+ currents of the exchanger (dashed green line) and the Ca2+ pump (dashed orange line) are shown along with schematic diagrams depicting each state. Immediately following the stimulus train, both currents are positive (outward, Fig 7D, middle), but then the exchanger current becomes negative (inward) while the pump current remains positive. As a result, the net current of the Ca2+ regulatory mechanisms reaches a near-zero value (Fig 7C, blue line; Fig 7D, right). The inward current mediated by the exchanger reflects the condition of high [Na+]i level, that causes the exchanger to operate in a "reverse mode" (Ca2+ influx). In this state [Ca2+]i is determined by the slow change of the quasi-stable state, which is due to the slow return of tuft [Na+]i back to baseline levels (Fig 7A).
As described above, the prolonged quasi-stable state of [Ca2+]i is the result of the opposing actions of the pump Ca2+ efflux and Ca2+ influx due to the reverse-mode of the Na+-Ca2+ exchanger (Fig 7C and 7D). Therefore, blockade of the exchanger should result in acceleration of the recovery of [Ca2+]i. We examined this prediction using the model by calculating the decay of the tuft fluorescence under control conditions (Fig 8A, orange line) and after blocking the exchanger (Fig 8A, red line). Indeed, a much faster return to baseline levels was obtained in the absence of exchanger activity. Experimentally, such blockade can be realized by substituting Na+ ions in the extracellular solution with Li+, which cannot be transported by the exchanger [28–30]. This manipulation was simulated by modeling the effect of Li+ on the exchanger and the Na+-K+ pump [31,32]. Fig 8B shows that following substitution of Na+ by Li+ in the model, the slow inward current (orange line) is replaced by a fast, transient inward current (red line) that rapidly declines to baseline.
Similar results were obtained experimentally (Fig 8C and 8D) by substituting Na+ with Li+ in the bath solution. In the presence of Li+ (green) the decay of the fluorescence signal in the tuft (Fig 8C) followed a single time constant (τ = 11 s, dashed line, compare to inset in Fig 8A). Thus, in the absence of exchanger activity the elevated quasi-stable state was blocked. In accordance with the fluorescence measurements, the inward current (Fig 8D) decayed faster in the presence of Li+ (compare green to blue lines) and the initial “bump” (arrow in Fig 8D) created by the Na+-K+ pump outward current was absent (compare to the effect of ouabain application, Fig 2G). These experimental observations strongly support the proposed model in which a quasi-stable state of [Ca2+]i is generated by the reversed action of the Na+-Ca2+ exchanger.
We examined the response of our detailed model to VNO inputs using a simple feed-forward network simulation (see Materials and Methods). The input stage of the network represents the responses of vomeronasal sensory neurons (VSNs) to natural stimuli, which follow a simple ligand-receptor interaction [33,34]. A simple model of VSN firing was established by first calculating the predicted time course (Fig 9A, red line) of a response to a brief stimulus application (red bar), and then using it to generate a random spike train (vertical lines). For the purpose of the model, we assumed that each tuft is innervated by 13 VSNs (Fig 9B; within the lower range of reported convergence [35]).
The unitary synaptic response of an AOB mitral cell to sensory fiber stimulation (Fig 9C, blue line) was measured by stimulating the dendritic tuft of a mitral cell (see Materials and Methods, micrograph in Fig 9C). A model of synaptic conductance was fitted to the averaged response (Fig 9C, orange line) and assigned to the dendritic tufts of the aforementioned mitral cell model. The simulated response was then tested for two ligands presented simultaneously to two groups of 13 VSNs, converging on two different dendritic tufts of the model mitral cell (Fig 9B). This simulation was run using a range of stimulus durations for both ligands, and a range of concentrations for one of them (the concentration of the second ligand was kept constant). Random white noise was also injected to the model mitral cell. Using these parameters, we encountered two possible outcomes: one was a transient firing response (Fig 9D) while in the other firing persisted for ~200 s (Fig 9E).
Fig 9F summarizes the average firing durations for different combinations of stimulus duration and concentration. As shown, longer stimulus durations and higher concentrations led to persistent firing responses. The bimodal distribution of the response duration is shown in Fig 9G. As apparent, the response was either transient, following the stimulus application, or persistent peaking at 200 s, as previously demonstrated for AOB mitral cells [3]. Thus, our model cell embedded in a realistic small network simulation reproduces the transition of the responses of AOB mitral cells between transient and persistent modes as a function of stimulation strength and duration, as observed both in vitro and in vivo.
Here, we combined in vivo and in vitro electrophysiological recordings, Ca2+ and Na+ imaging, and computational modeling to investigate how persistent firing responses are generated in AOB mitral cells. We first demonstrated that these cells are capable of responding in vivo with persistent firing to natural stimuli applied to the VNO, in a stimulus-specific manner. We then used AOB slices to explore the ohmic and non-ohmic currents that underlie the persistent firing responses. The dynamics and inter-relationships of these current sources were analyzed by pharmacological and ionic manipulations. By combining patch-clamp recordings with Ca2+ and Na+ imaging we showed that the dynamics of the currents underlying persistent firing rely on slow changes in internal ionic concentrations within specific cellular compartments. Using first abstract and then realistic computational models of AOB mitral cell we demonstrated that these changes are governed by passive diffusion, as well as by local active processes involving ionic pumps and exchangers. The model results accurately predict the persistent firing responses, the dynamics of Na+ in the dendrites and the results of substituting Na+ by Li+. Finally, we employed our model in a local network simulation and established that the model cell indeed shifts between transient and persistent firing responses to stimuli detected by VNO neurons, as a function of the duration and strength of stimulation.
We have previously hypothesized [3] that the accessory olfactory system reports the social context to an animal by inducing specific brain states based on the persistent activity of its mitral cells. These states can then change the processing of sensory information in other brain areas [36–38]. This hypothesis is supported by the current study where we showed, both in vivo and in vitro that AOB mitral cells shift from transient to persistent firing responses to stimuli arriving from the VNO. In agreement with our previous reports [3,20], this transition is relatively sharp and depends on sufficient stimulus strength and duration. This may correspond, for example, to the presence of a rich source of social chemosensory cues, e.g., a conspecific animal, which would elicit persistent firing in AOB mitral cells, thereby conveying to higher brain centers the presence of a social partner.
Persistent activity has been reported in several brain areas, including mitral cells of the main olfactory bulb in vivo [39], and was suggested to mediate working memory or prolonged brain states [4]. To date, most of the mechanisms purposed for persistent activity include either network or biophysical feedback loops [5,10,16,39,40]. We propose a novel mechanism for persistent neuronal activity in AOB mitral cells, in which a train of action potentials that back-propagate to the dendritic tuft [24] elevates the tuft [Na+]i. This increase in the tuft [Na+]i shifts the [Ca2+]i stable state upwards [29], thus creating an elevated quasi-stable state for [Ca2+]i. The slow decay of [Ca2+]i is dictated by the slow removal of Na+ ions from the tuft (Fig 7). The novelty of our model is in the absence of biochemical, biophysical, or network feedback mechanisms or hysteresis.
Two alternative models were examined in an attempt to reproduce the observed experimental results. One model incorporated two distinct CAN conductances—a fast one and a slow one (compare with [41]), without activity-dependent outward current. Each of these currents produces a different phase of the observed inward current (Fig 2B and 2C). While this model reproduces the results used as objectives in the model training phase (Figs 5 and 6A–6C), it failed to produce the persistent activity (Fig 6D and 6E), emphasizing the importance of this validation. The long time constant kept the slow Ca2+-dependent current away from its stable-state value in the hybrid clamp simulations, but this current would greatly intensify and cause a runaway effect in a current clamp simulation (S8A Fig).
In another model, a single Ca2+-dependent non-specific cation conductance was used, along with a Ca2+-dependent K+ conductance (SK/BK) that account for the outward current observed immediately after the stimulating train of action potentials. Although this model reproduces the results used as objectives in the model training phase (Figs 5 and 6A–6C), it failed to explain the outward current recorded in the absence of [Ca2+]o (Fig 2E).
The ultimate rejection of these models demonstrates the importance of the model validation step.
According to our results, [Na+]i plays a key role in both transient and prolonged biophysical processes. First, increase in AIS [Na+]i alters the Na+ Nernst potential, thus lowering the spike amplitude (Fig 5E–5I and S5 Fig). Second, elevation in [Na+]i increases the Na+-K+ pump-mediated outward current that terminates bursts of activity by hyperpolarizing the cell (Fig 2E). Third, increase in [Na+]i decreases Ca2+ efflux by modulating the Na+-Ca2+ exchanger (Fig 8). The latter endows [Na+]I with the ability to dictate [Ca2+]i and thus the inward ICAN.
In thin and slightly active processes like the dendritic tuft, [Na+]i is only moderately increased by neuronal activity and its extrusion is exceptionally slow (Fig 6F,[27]). Thus, [Na+]i can attain a range of quasi-stable states—a property that renders it an ideal candidate to integrate epochs of high neural activity. This is in contrast to [Ca2+]i, which is highly dynamic as a function of neuronal activity on one hand, and rapid extrusion on the other.
Notably, changes in [Na+]i are rarely tracked in conductance-based models (important exceptions in [42–45]), although in thin active processes, such as axons and dendrites, they may substantially affect neuronal activity. Indeed, the effects of [Na+]i on a variety of Ca2+-dependent processes were previously demonstrated: reducing Na+ extrusion or inhibiting the Na+-Ca2+ exchanger were shown to extend [Ca2+]i transients, and thus to facilitate Ca2+-dependent mechanisms, such as synaptic plasticity and learning [29,46,47].
The activity-dependent Na+-K+ pump-mediated outward current is an important factor protecting neurons from a runaway positive feedback loop (S8B and S8C Fig). Without this current, the high [Ca2+]i during and immediately after stimulation would result in a strong inward current (as demonstrated using ouabain—Fig 2G, red line) that would in turn evoke high frequency spiking, and thus a further increase in [Ca2+]i.
Since the mechanism described here contains mostly elements which are ubiquitous in neurons, we argue that it is relevant (wholly or partially) to other brain areas. The essential non-trivial building blocks required to produce persistent activity by this mechanism are: a) changes in [Na+]i and [Ca2+]i in thin processes (axon or dendrites), attained by (back)-propagation of action potentials and/or excitatory synaptic activity; b) Co-localization of an excitatory Ca2+-dependent conductance at the site of [Na+]i changes; c) Low density of [Na+]i active extrusion mechanisms at the site of [Na+]i changes.
Thus, the proposed mechanism is very likely to explain persistent activity in other brain areas, most likely with some variation of the time scale and of the factors necessary to evoke it. For example, [Na+]i dynamics may take alternative forms in other neuronal types, as a function of the spatial distribution of Na+ channels and Na+-K+ pumps as well as cell morphology. Moreover, the persistent activity may be either superthreshold, i.e. persistent firing, or long term integrative changes in membrane potential and excitability. It may be assumed that very slow changes in resting potential or firing rate, such as those produced by the mechanism presented here, may be under-reported because they are frequently filtered out and require prolonged recording sessions.
Furthermore, the importance of the mechanism proposed by us is beyond the scope of persistent activity. We show that the long-term behavior of intracellular Ca2+ depends upon past activity—an idea previously presented in [29], though in a much shorter time scale. Given that Ca2+ is an important cellular signaling molecule, long-term changes in its concentration may broaden the time scale of processes such as synaptic plasticity [47], activity-dependent gene expression, and more. Similarly to membrane potential changes, reports about [Ca2+]i changes may also suffer from "time-scale bias," wherein long-term changes in [Ca2+]i seems to be regularly filtered out or not recorded due to photo damage and dye-bleaching concerns.
An even broader consequence is the ability of Na+ to integrate activity (discussed above), which at the very least causes changes to Na+ reversal potential. Thus, persistent activity is only one of several end results of the mechanism we describe. Our results at least demonstrate that changes in [Na+]i should not be regarded as negligible in conceptual and computational models of neuronal activity.
C57BL/6J and BalbC male mice were maintained in the SPF mouse facility of the Hebrew University of Jerusalem under veterinary supervision, according to National Institutes of Health standards, with food and water ad libitum and lights on from 7:00 A.M. to 7:00 P.M. Eight- to twenty-week-old mice (25–35 g) were held in groups of 5–10 mice per cage. All experiments were approved by the Animal Care and Use Committee of the Hebrew University (permit number: NS-12-13310-4). Mice were anesthetized for in vivo experiments (ketamine, medatomidine). For in vitro experiments, mice were anesthetized (pentobarbitone) and killed by cervical dislocation.
Secretions for in vivo recordings were collected from C57BL/6J and BalbC females (housed in the animal facility of the Hebrew University). Samples were pooled and immediately frozen in liquid nitrogen and stored at −80°C until use. For urine collection, mice were gently held over a plastic sheet until they urinated. Vaginal secretions were collected by flushing the vagina with 30μl of ringer’s solution repeatedly. 20 μl were stored. For saliva collection, isoproterenol hydrochloride (0.2 mg/100 g) and pilocarpine (0.05 mg/100 g) were injected i.p. to increase salivation [48]. Following a delay of 5 min, saliva was collected from the mouth using a micropipette. Stimuli were diluted in Ringer’s solution.
In vivo multi-unit activity followed by electrical stimuli was recorded in anesthetized mice (ketamine, 10 mg/kg, medatomidine, Pfizer, 1 mg/kg). A recording electrode (glass micropipette filled with 1 M potassium acetate) was placed in the AOB external plexiform layer using a micromanipulator (Luigs and Neumann). A stimulating coaxial bipolar electrode was inserted through the medial frontal lobe, to the point where contact was made with the vomeronasal nerve and field potentials appeared in the AOB in response to brief stimuli. A train of brief shocks (0.1 ms, 1–100 V), given at 2 Hz for 2.5 s was applied to the vomeronasal nerve at intervals of 60 s via an isolated stimulator.
Electrophysiological recordings of AOB neurons followed by natural stimuli were performed as previously described in detail [18]. Briefly, BalbC mice were anesthetized with 100 mg/kg ketamine and 10 mg/kg xylazine. A tracheotomy was made using a polyethylene tube to allow breathing during flushing; a cuff electrode was placed on the sympathetic nerve trunk with the carotid serving as a scaffold. Incisions were closed and the mouse was placed in a custom-built stereotaxic apparatus where anesthesia was maintained throughout the entire experiment with 0.5–1% isoflurane in O2. A craniotomy was opened immediately rostral to the rhinal sinus, the dura was removed around the penetration site, and electrophysiological probes were advanced into the AOB using an electronic micromanipulator (MP-285; Sutter instruments). All recordings were made with 32 channel probes (NeuroNexus Technologies). During each trial, 2 μl of stimulus solution was placed directly in the nostril (“stimulus application”) and after 20 s, a square-wave stimulation train (duration: 1.6 s, current: ±120 μA, frequency: 30 Hz) was applied through the sympathetic nerve cuff electrode to induce VNO pumping and, accordingly, stimulus entry to the VNO lumen (“sympathetic stimulation”). A pump was turned on 40 s after each stimulus presentation, followed (after 10 s) by application of Ringer’s solution (1–2 ml) to the nostril that was flushed through the nasopalatine duct to cleanse the nasal cavity. 20 s after Ringer's application sympathetic stimulation was performed to ensure the VNO lumen cleansing (the second stimulation).
Using an RZ2 processor, PZ2 preamplifier, and two RA16CH head-stage amplifiers (Tucker-Davis Technologies), neuronal activity was sampled at 25 kHz and band-pass filtered at 0.3–5 kHz. Custom MATLAB (Mathworks) programs were used to extract spike waveforms. Spikes were sorted automatically according to their projections on two principle components using KlustaKwik [49] and then manually verified and adjusted using the Klusters program [50].
Mice were anesthetized (pentobarbitone; 60 mg/kg) and killed by cervical dislocation. Olfactory bulbs were dissected into a physiological solution containing the following (mM): 125 NaCl, 25 NaHCO3, 5 glucose, 3 KCl, 2 CaCl2, 1.3 NaH2PO4, and 1 MgCl2, oxygenated by bubbling through a 95% O2 and 5% CO2 mixture, pH 7.4, 36°C. Parasagittal olfactory bulb slices, 300–400 μm thick, were prepared and equilibrated for 0.5–3h in the same solution at physiological temperature [51].
For electrophysiological recordings, slices were submerged in oxygenated physiological solution (identical to above) at room temperature in a recording chamber and perfused at a constant rate of 5–7 ml/min. To test the effect of substituting Na+ by Li+, equimolar amount of LiCl was used instead of NaCl. To test the effect of Ca2+ removal, equimolar amount of MgCl2 was used instead of CaCl2. Where indicated, picrotoxin (100 μM) was added to the bath solution to block GABAA receptors, or ouabain (Tocris Bioscience) was added to the bath solution in excess (10–100 μM) to block the Na+-K+ pump.
For electrophysiological recordings, we used an Olympus BX61WIF microscope equipped with a motorized stage and manipulators (Luigs and Neumann), pulse generator (Master8, A.M.P.I.), isolated stimulator (ISOFlex, A.M.P.I.), and a MultiClamp 700B amplifier (Molecular Devices).
Mitral cells were visualized using infrared differential interference contrast (DIC) video microscopy via a 40x or a 60x water-immersion objective. Mitral neurons were identified by the location of the cell body on the ventral side of the external plexiform layer of the AOB. Whole-cell recordings were performed using borosilicate pipettes filled with standard intracellular recording solution containing the following (mM): 130 K-gluconate, 10 Na-gluconate, 10 HEPES, 10 phosphocreatine, 4 MgATP, 0.3 NaGTP, and 4 NaCl (pH = 7.25 with KOH, 5–12 MΩ). When BAPTA was used, BAPTA—tetrapotassium (Sigma) was dissolved in this solution to a final concentration of 5 mM. Seal resistance was at least 2 GΩ and typically 5–10 GΩ.
In most experiments, a 4-s-long spike train was evoked by injecting a series of depolarizing pulses (rate, 1–30 Hz; amplitude, 1–2 nA; pulse duration, 10 ms). In the hybrid-clamp procedure, membrane potential was clamped to −80 or −70mV throughout the experiment, excluding 4 s periods during which the amplifier was switched to current-clamp mode to deliver the train of current pulses. It should be noted that the hybrid-clamp methodology has been proven useful for investigating the firing activity of neurons, since by preventing most of the feedback and interference that ongoing firing activity may elicit upon itself, it enables a relatively clean examination of the underlying currents.
All amplified signals were digitized at 2–20 kHz using a National Instruments board and homemade software written in LabVIEW (National Instruments).
A unitary measurement of EPSC was done by filling a cell with Alexa 488 (Life Technologies) and visually positioning a bipolar theta electrode filled with physiological solution and Alexa 488 close to one of the dendritic tufts (Fig 8C).
For calcium imaging experiments, Oregon Green BAPTA-1 (OGB-1, Life Technologies, 50 μM) was added to the pipette solution. Fluorescence signals were recorded during the hybrid clamp protocol using a high speed camera (MiCAM Ultima, Brainvisions) and converted to (F−Fmin)/Fminratio after subtracting the ongoing background signal. In order to perform imaging of the dendritic tuft, a 60x water-immersion objective was used, and the dye was allowed to fill the cell for >20 min before recording was started. Evoked spikes were used to increase fluorescence and facilitate the visual search for a dendritic tuft.
For sodium imaging experiments, SBFI salt (TEFLabs, 2mM) and Sodium Green salt (Life Technologies, 500uM) was added to a modified pipette solution containing (mM): 130 K-gluconate, 10 HEPES, 5 phosphocreatine, 4 MgATP, 0.3 NaGTP, 20 KCl, and 0.2 EGTA (pH = 7.2 with KOH, 5–12 MΩ). Fluorescence signals were recorded from the apical dendrite as in the case of calcium imaging. A standard Fura-2 filter set (Ex. 380 nm; Em. 510 nm, Chroma Technology) was used for SBFI imaging. In order to cancel the substantial dye bleaching, trials without stimulus were subtracted from stimulus trials.
Unless otherwise noted, recorded current or voltage traces were averaged for each recorded cell, and the presented result is the mean of the cell population.
Value error range reported is SEM unless otherwise noted.
For stimulus-induced in vivo AOB recordings, units from various sets of experiments were considered. Six hundred and sixty-three single units for which at least one stimulus was presented at least five times were considered. The procedure for identifying persistent responses was as follows: For each single trace (one unit, one presentation of one stimulus), the responses in consecutive 2 s bins were defined as significant if the rate within it was larger than the mean baseline rate by at least five times the SEM of the baseline rate. The bins spanned a period of 160 s, which includes the baseline period, stimulus delivery, and the VNO flushing period, extending 20 s after the sympathetic stimulation during flushing (second stimulation, see above). The response duration was defined as a period beginning and ending with a significant bin, and in which at least 85% of the bins were significant. If this response lasted more than 10 s following the second sympathetic stimulation, it was designated as a persistent response. Finally, units for which at least half of the responses were persistent for a given stimulus were defined as persistent.
The seven persistent firing units came from three sets of experiments differing in stimulus sets used. For three units the stimuli were undiluted saliva, urine, and vaginal secretions. In three other units the stimuli were saliva, 100F diluted urine, and vaginal secretions. For the remaining units, the stimuli were urine at 1F, 10F, and 30F dilutions.
The abstract dynamical model was constructed in MATLAB/SIMULINK (Mathworks). See S1 Text for details and equations.
We constructed the conductance-based model using the NEURON simulation environment with Python [52,53]. The model was based on experimental measurements and morphological reconstruction of an AOB mitral cell [54], and included influx, diffusion, and extrusion of Na+ and Ca2+. It assumed a presence of active Na+ channels in the apical dendrites and tufts [24], as well as non-uniform channel properties across different compartments (Fig 4E, [55]). Evolutionary multi objective optimization algorithm [25,26] was used to find the model parameters, based on recorded electrophysiological and imaging data. Some membranal mechanisms were based upon published models hosted by ModelDB [56–60]. See S1 Text for additional information. The model code is available online at:
https://senselab.med.yale.edu/ModelDB/ShowModel.cshtml?model=185332
In order to test the response of the model mitral cell to natural stimuli, a simple two-layer network model was constructed. The firing response of the VNO sensory neurons was modeled using a simple ligand-receptor interaction [33] that triggers a semi-random spike train (Fig 8A). Thirteen of such sensory neurons converged on each of the mitral cell's dendritic tufts (Fig 8B) [35], where the synaptic current was modeled using a double exponential fit to the unitary response measured experimentally using a theta electrode (a bipolar micropipette pulled from a tubing with a θ-like cross-section) positioned next to a dye-filled mitral cell's dendritic tuft (Fig 8C).
|
10.1371/journal.pcbi.1002189 | Identification of Potent EGFR Inhibitors from TCM Database@Taiwan | Overexpression of epidermal growth factor receptor (EGFR) has been associated with cancer. Targeted inhibition of the EGFR pathway has been shown to limit proliferation of cancerous cells. Hence, we employed Traditional Chinese Medicine Database (TCM Database@Taiwan) (http://tcm.cmu.edu.tw) to identify potential EGFR inhibitor. Multiple Linear Regression (MLR), Support Vector Machine (SVM), Comparative Molecular Field Analysis (CoMFA), and Comparative Molecular Similarities Indices Analysis (CoMSIA) models were generated using a training set of EGFR ligands of known inhibitory activities. The top four TCM candidates based on DockScore were 2-O-caffeoyl tartaric acid, Emitine, Rosmaricine, and 2-O-feruloyl tartaric acid, and all had higher binding affinities than the control Iressa®. The TCM candidates had interactions with Asp855, Lys716, and Lys728, all which are residues of the protein kinase binding site. Validated MLR (r2 = 0.7858) and SVM (r2 = 0.8754) models predicted good bioactivity for the TCM candidates. In addition, the TCM candidates contoured well to the 3D-Quantitative Structure-Activity Relationship (3D-QSAR) map derived from the CoMFA (q2 = 0.721, r2 = 0.986) and CoMSIA (q2 = 0.662, r2 = 0.988) models. The steric field, hydrophobic field, and H-bond of the 3D-QSAR map were well matched by each TCM candidate. Molecular docking indicated that all TCM candidates formed H-bonds within the EGFR protein kinase domain. Based on the different structures, H-bonds were formed at either Asp855 or Lys716/Lys728. The compounds remained stable throughout molecular dynamics (MD) simulation. Based on the results of this study, 2-O-caffeoyl tartaric acid, Emitine, Rosmaricine, and 2-O-feruloyl tartaric acid are suggested to be potential EGFR inhibitors.
| Tumor growth is associated with overexpression of epidermal growth factors receptors. Targeted control of EGFR by EGFR inhibitors is an attractive therapy alternative to conventional cancer treatment that offers specificity and reduced adverse effects. The purpose of this study was to identify natural compounds from traditional Chinese medicine that may be used as EGFR inhibitors. The top four TCM compounds with the highest binding affinity to EGFR were selected and their suitability as EGFR inhibitors confirmed with different statistical prediction models. The candidate compounds had higher bioactivity than Iressa®, the drug that is clinically used. The TCM compounds also met key structural components that were characteristic among known inhibitors. In addition, the binding between TCM compounds and EGFR were stable which is a fundamental requirement for any targeting drug. Results from bioactivity prediction, structural component matching, and binding stability all point to the possibility of these TCM compounds as suitable EGFR inhibitor candidates.
| Target-specific therapies have generated much attention in addition to conventional cancer treatments [1]–[3]. By targeting key molecules essential for cellular function, replication, or tumorigenesis, such therapies may exert cytostatic or cytotoxic effects on tumors while minimizing nonspecific toxicities associated with chemotherapy or irradiation [4].
The epidermal growth factor receptor (EGFR) signaling pathway is one of the most important pathways in mammalian cells [5]. Specific ligands, such as epidermal growth factor (EGF) and transforming growth factor alpha (TGFα), bind and activate EGFR, triggering autophosphorylation of the intracytoplasmic EGFR tyrosine kinase domain [6], [7]. The phosphorylated tyrosine kinase residues serve as binding sites for signal transducers and activators of intracellular substrates, which then stimulate intracellular signal transduction cascades that upregulate biological processes such as gene expression, proliferation, angiogenesis, and inhibition of apoptosis [8]. EGFR overexpression has been shown to activate downstream signaling pathways, resulting in cells that have aggressive growth and invasive characteristics [9]. Tumor cell motility, adhesion, metastasis, and angiogenesis have also been associated with stimulated EGFR pathways [10]–[12]. Since EGFR over-expression often differentiates tumor cells from normal cells, it is possible for EGFR inhibitory molecules to act on tumor cells and attenuate their proliferation rates [4].
Several tyrosine kinase inhibitors were approved for clinical use. Iressa® (gefitinib) is highly selective for EGFR tyrosine kinase and is commonly used for treating lung cancer [13]. EGFR downstream signaling is competitively inhibited by Iressa® at its ATP binding site [14]. Other therapeutic agents with inhibitory mechanisms similar to Iressa® include Erlotinib (Tarceva®) against non-small cell lung cancer (NSCLC) and pancreatic cancer [15], [16], and Vandetanib (Zactima®) against late stage medullary thyroid cancer [17]. Lapatinib (Tykerb®) is a dual inhibitor of EGFR and HER2 tyrosine kinases approved for metastatic breast cancer [18], [19]. Though the effect of Iressa® on lung cancer has been well established, severe side effects has also been reported [20]. Adverse reactions listed under Iressa® product information include diarrhea, skin rash and dryness, nausea, vomiting, haemorrhage, anorexia, asthenia, and in some cases, interstitial lung disease with fatal outcomes [21]. The adverse effects of available treatments necessitate continuous search efforts for alternatives with less toxicity.
Computational predictions in biology and biomedicine are of significant importance for generating useful data which otherwise be time-consuming and costly through experiments alone [3], [22]–[27]. Computational predictions, combined with information derived from structural bioinformatics analysis, can provide useful insights and timely information for both basic research and drug development [28], [29]. Much cutting-edge cancer drug development has been conducted through the use of computational bioinformatics and modeling [30]–[37]. The powerful ability of modern computational prediction and bioinformatics were adopted in this research to search for novel EGFR inhibitors.
Traditional Chinese medicines (TCM) are natural substances with therapeutic effects on many diseases [38]–[40]. The vast number of TCM represents a rich resource that can be explored for drug development. We had investigated kinase inhibitor candidates from TCM targeting HER2 and HSP90 receptors before [28], [41]–[42]. Though EGFR kinase inhibitors have been investigated through different screening and modeling scenarios [43]–[47], none from TCM compounds has been reported to date. This study utilized the world’s largest TCM Database@Taiwan [48] to screen for potential EGFR inhibitors from TCM compounds and applied structure- and ligand-based methods to evaluate the suitability of candidates as EGFR inhibitors.
A useful predictor for a biological system should include the following steps [49]: (i) selection of a valid dataset to train and test the predictor; (ii) formulate samples with an effective mathematical expression that reflects intrinsic correlation with the attribute to be predicted; (iii) develop a powerful algorithm to operate the prediction; (iv) objectively evaluate accuracy of the predictor through cross-validation tests. The experimental design of the current study follows these guidelines and details are explained in the following sections.
The EGFR protein sequence (EGFR_HUMAN, P00533) used in this study was obtained from Swiss-Prot [50], and the 3D structure (PDB: 2ITY) [51] used for analyses was downloaded from Protein Data Bank. The tyrosine kinase was encoded from Phe712-Leu979, and the ATP binding site was located at Leu718–Val726.
The Traditional Chinese Medicine (TCM Database@Taiwan) database (http://tcm.cmu.edu.tw) was used to screen for potential EGFR inhibitors from more than 20,000 compounds within the database. All compounds were operated using the Prepare Ligands module with Lipinski’s rule of five using Discovery Studio 2.5 (DS 2.5; Accelrys Inc., San Diego, CA). Iressa® was selected as the control. The LigandFit program (DS 2.5) was used to locate the best docking pose for different confirmations under the Chemistry at HARvard Macromolecular Mechanics (CHARMm) force field [52]. Results for the docking studies were ranked according to Dock Score.
Twenty ligands with demonstrated inhibition against EGFR were used in this study (Table S1) [53]. Descriptors for each ligand were identified using the Calculate Molecular Properties program in DS 2.5. Predictive models containing five optimum descriptors were generated using the Genetic Function Approximation (GFA) algorithm. Descriptors in the model with the highest r2 value were used to construct ligand activity prediction models.
A MLR model using the descriptors from the top GFA algorithm was constructed using Matlab Statistics Toolbox (MathWorks, Natick, MA) and validated using MLR Leave-One-Out validation [54]. The MLR model was trained with 17 randomly selected ligands with EGFR inhibitory activity (Table S1) and used to predict the activity (pIC50) of the control and TCM candidates.
The identical descriptors were normalized to the range of [−1,+1] and plugged into the libSVM program to generate a SVM prediction model[55]. Following model training with the 17-ligand training set, the SVM model was used to predict the activity of the control and TCM candidates.
Ligands used in the previous sections were also used for 3D-QSAR analysis. The 2-dimensional (2D) and 3-dimensional (3D) ligand structures were drawn with ChemBioOffice 2008 (PerkinElmer Inc., Cambridge, MA) under a Molecular Mechanics 2 (MM2) force field. Following ligand alignment, Comparative Molecular Field Analysis (CoMFA) and Comparative Molecular Similarities Indices Analysis (CoMSIA) models were constructed using partial least squares statistical method (PLS). Cross-Validated (CV) correlation coefficient (q2) and non-cross-validation correlation coefficient (r2) were used to validate the models. Biological activities of Iressa® and TCM candidate compounds were predicted using the generated 3D-QSAR contour map.
Molecular dynamics (MD) of Iressa® and the TCM candidates were simulated using DS2.5 Standard Dynamics Cascade and Dynamics package. Sample preparation was conducted under the following parameters: [minimization] steepest descent and conjugate gradient: each with maximum steps of 500, [heating time] 50 ps, [equilibration time] 200 ps. The simulations were produced with a total production time of 20 ns with NVT, constant temperature dynamics of Berendsen weak coupling method, a temperature decay time of 0.4 ps, and a target temperature of 310K. Root mean square deviations (RMSD) of protein-ligand complex and individual ligands, total energy of protein-ligand complex, hydrogen bond (H-bond), and H-bond distance were analyzed using the Analyze Trajectory function following MD simulation.
The top four TCM candidates with the highest Dock Score were 2-O-caffeoyl tartaric acid, Emitine, Rosmaricine, and 2-O-feruloyl tartaric acid (Table 1). Corresponding scaffolds of the top TCM candidates are illustrated in Figure 1. Iressa®, Emitine, and Rosmaricine had amine groups available for H bonding whereas 2-O-Caffeoyl tartaric acid and 2-O-feruloyl tartaric acid had carbonyl groups. The different residues available for H bonding resulted in different binding poses (Figure 2). Binding of Iressa® (Figure 2a), Emitine (Figure 2c), and Rosmaricine (Figure 2e) to tyrosine kinase were located within the pocket, with H-bonds formed between the amine group of the ligand compounds and the carboxyl group of Asp855. 2-O-Caffeoyl tartaric acid (Figure 2b) and 2-O-feruloyl tartaric acid (Figure 2e) docked outside the tyrosine kinase pocket and formed multiple H-bonds through their carboxyl groups with Lys716 and Lys728. The binding location of 2-O-caffeoyl tartaric acid and 2-O-feruloyl tartaric acid directly blocks the ATP binding site of tyrosine kinase located from Leu718–Val726. Dock scores of each TCM candidate is given in Table 1. All candidates have higher dock scores than Iressa®, indicating higher binding affinities to the tyrosine kinase receptor than Iressa®.
Representative descriptors from the top GFA algorithm include: Num_H_Acceptors_Lipinski (equivalent of N+O count), Molecular_SurfaceArea (the total surface area for each molecule using a 2D approximation), Kappa_1 (Kappa Shape Indices), PMI_Y (principle moment of inertia Y-component), and Shadow Xlength (length of molecule in the X dimension). The descriptors were validated using Leave-One-Out method which is the most objective of all available cross-validation methods [56]. The MLR model established with the determined descriptors was:
The SVM model was also established with the five identified descriptors using libSVM.
Correlation between the predicted and observed pIC50 activities on EGFR ligands of known activity using the constructed MLR and SVM models were illustrated in Figure 3a and 3b, respectively. Correlation coefficients based on the training set were 0.7858 for the MLR model and 0.8754 for the SVM model. Activity predictions of Iressa® and the TCM candidates using MLR and SVM were summarized in Table 1. Both models indicate that Iressa and the TCM candidates are compounds with acceptable predicted activities. Predicted activities (pIC50) of Iressa by the trained MLR and SVM models were 6.715 and 5.110, respectively. The Iressa activity predicted by SVM was closer to experimentally determined Iressa activities (pIC50) between 4.76–5.96 [57], thus SVM values may be more accurate predictions of the actual activity.
The results of CoMFA and CoMSIA model generation are detailed in Table 2. Steric field was the sole factor in the CoMFA model since the electrostatic field value was zero. Cross-validated (q2) and non-cross-validated (r2) correlation coefficient values of 0.721 and 0.986, respectively, indicated a high level of confidence for this model. The small standard error of estimates (SEE) and large F-test value further supported the reliability of this model. In contrast, CoMSIA models were influenced by multiple factors including steric field, hydrophobic region, and hydrogen bond donor/acceptors. Among all generated versions of the CoMSIA model, CoMSIA_SHD had the highest r2 (0.988), lowest SEE (0.133), and highest F value (134.272), thus was selected as the optimum CoMSIA model for use in this study. The pIC50 of 20 ligands predicted by the constructed CoMFA and CoMSIA models were compared with observed pIC50 reported by Fidanze et al. [53] in Table 3. In general, both models gave similar predicted values and were close to the experimentally determined activities. Correlations between predicted and observed pIC50 using CoMFA and CoMSIA models are summarized in Figure 4a and 4b, respectively. High correlation coefficients validated the reliability of the constructed CoMFA (r2 = 0.9860) and CoMSIA(r2 = 0.9877) models.
Ligand activities of Iressa® and the TCM candidates can be predicted based on structural conformation to the 3D-QSAR feature map, including features in steric field, hydrophobic field, and H-bond donor/acceptor characteristics. As illustrated in Figure 5, Iressa and the TCM candidates were able to match the generated 3D-QSAR model features. The benzene in Iressa® favored steric and hydrophobic fields, and H-bond was favored between its amine group and Asp855. In 2-O-Caffeoyl tartaric acid, the benzene structure favored steric and hydrophobic fields, and the carboxyl group favored H-bond formation with Lys716 and Lys728. The carbon chain structure in Emetine contoured to the steric and hydrophobic fields, and the amine group favored H-bond formation with Asp855. Rosmaricine had benzene and isopropyl structures that favored steric and hydrophobic fields, and an amine group that favored H-bond with Asp 855. The benzene structure in 2-O-feruloyl tartaric acid favored steric fields and the carboxyl group favored H-bond formations with Lys716 and Lys728. Iressa® and the TCM candidates have structural components that contour to the features of the 3D-QSAR model, thus were likely to be biologically active.
Binding stability of the control and TCM candidates was validated using MD simulation. RMSDs of protein-ligand complex (Figure 6a) and individual ligand (Figure 6b) stabilized after 10 ns. The RMSDs of the protein-ligand complexes stabilized at approximately 1.6Å. With regard to individual ligands, the RMSDs of Iressa and 2-O-caffeoyl tartaric acid was 2.0 and 1.6Å, respectively. All other compounds registered RMSD values of approximately 1.0Å. The lower RMSD values of the TCM candidates suggest more stability within the receptor compared to Iressa. The energy trajectory of each compound is shown in Figure 6c. Complexes formed by Rosmaricine and 2-O-feruloyl tartaric acid had the lowest total energy (<−14,800 kcal/mol), followed by Iressa® and Emetine (approximately −14,700 kcal/mol), and 2-O-caffeoyl tartaric acid (−14,600 kcal/mol). Stabilization of total energy in ligand-protein complexes was achieved after 12 ns.
H-bond distance profiles in the EGFR receptor were summarized in Figure 7. A single H-bond between the amine group on Iressa® and the carboxyl group on Asp855 was formed after 9.74 ns and stabilized after 20 ns (Figure 7a). Two H-bonds were formed between the carboxyl group of 2-O-caffeoyl tartaric acid and Lys716 and Lys728 of the EGFR receptor (Figure 7b). The formation of two H-bonds contributed to a higher stability between 2-O-caffeoyl tartaric acid and the EGFR receptor. However, an increase in H-bond distance was observed towards the end of the 20 ns simulation period, suggesting a weakening of the H-bond at Lys728. Emetine formed a total of four H-bonds with the receptor, two with Asp722 and two with Ala855 (Figure 7c). Bond distances stabilized after 10 ns for Ala722 and 4 ns for Asp855. Rosmaricine formed three H-bonds each at Asp841 and Arg855 (Figure 7d). The multiple H-bonds enabled Rosmaricine to remain in a stable state within the protein. 2-O-Feruloyl tartaric acid also formed multiple H-bond at Lys716 and Lys728, enhancing its stability within the receptor site (Figure 7e). However, similar to 2-O-caffeoyl tartaric acid, an increase in H-bond distance was also observed at Lys728 for 2-O-feruloyl tartaric acid. These observations suggest that the bond at Lys728 weakens throughout the MD simulation process, and that the H-bond at Lys716 may be the primary bond for 2-O-caffeoyl tartaric acid and 2-O-feruloyl tartaric acid. In addition, periodic fluctuations in H-bond distances were observed in 2-O-caffeoyl tartaric acid, Rosmaricine, and 2-O-feruloyl tartaric acid. These phenomena can be attributed to the rotation of the amine group where the H-bond is formed. These MD results support our docking findings which identify Asp855, Lys716, and Lys 728 as key residues for docking.
As determined in the CoMSIA model, hydrophobic interactions were key factors contributing to ligand bioactivity. Toward the final 20 ns of analysis, hydrophobic amino acids surrounding the docking region were Leu718, Val726, Ala743, Cys775, Phe795, Cys797, and Leu844. The hydrophobic subgroups of Iressa®, Emetine, and Rosmaricine were surrounded by Val726, Cys797, and Leu844 (Figure 8a). Hydrophobic groups of 2-O-caffeoyl tartaric acid were also surrounded Val726, Cys797, and Leu844 (Figure 8b). The hydrophobic region of 2-O-feruloyl tartaric acid was attracted to the Phe795 on EGFR (Figure 8b). The significance of matching the hydrophobic region of the ligand to that of the receptor may be to increase stability of the ligand-protein complex, and contribute to the bioactivity of the activated ligand. Our results indicate that Iressa® and the TCM candidates remained stable within the EFGR hydrophobic area following MD simulations.
Structural and ligand based methods supported 2-O-caffeoyl tartaric acid, Emetine, Rosmaricine, and 2-O-feruloyl tartaric acid as potential EGFR inhibitors. Structurally, the TCM candidates were capable of forming H-bonds with key residues Asp855, Lys716, and Lys728 and matched hydrophobic regions of the receptor. Bioactivity of the candidates were evaluated using validated MLR, SVM, CoMFA, and CoMSIA models. All models indicated that the TCM candidates have good predicted bioactivity. Molecular simulation results further supported the high potential for the TCM candidates in drug development. Iressa®, the drug currently used clinically, bound to the ERGF receptor through a single H-bond at Asp855. In comparison, multiple H-bonds formed at Asp855 and additional H-bonds formed at Ala722 and Arg841 increase the stability of Emetine and Rosmaricine, respectively. The ability of carboxyl groups in 2-O-caffeoyl tartaric acid and 2-O-feruloyl tartaric acid to form multiple H-bond networks that directly blocked the ATP binding site was also a unique characteristic worthwhile of further investigation. Contour to hydrophobic regions of the TCM candidates within the receptor site provides additional support for the stability of the protein-ligand complex. In summary, using different simulation and validation methods, we have identified four TCM compounds that may have potential as novel EGFR inhibitors. As the four TCM compounds have two distinctive types of binding locations and bond formation within the EGFR binding site, we suggest exploring the possibility of connecting Emetine/Rosmaricine with 2-O-caffeoyl tartaric acid/2-O-feruloyl tartaric acid through a spacer. The connection could allow more of points of attachment, which in turn would contribute to more stable binding within the tyrosine kinase site.
|
10.1371/journal.pcbi.1006864 | Identification of pathways associated with chemosensitivity through network embedding | Basal gene expression levels have been shown to be predictive of cellular response to cytotoxic treatments. However, such analyses do not fully reveal complex genotype- phenotype relationships, which are partly encoded in highly interconnected molecular networks. Biological pathways provide a complementary way of understanding drug response variation among individuals. In this study, we integrate chemosensitivity data from a large-scale pharmacogenomics study with basal gene expression data from the CCLE project and prior knowledge of molecular networks to identify specific pathways mediating chemical response. We first develop a computational method called PACER, which ranks pathways for enrichment in a given set of genes using a novel network embedding method. It examines a molecular network that encodes known gene-gene as well as gene-pathway relationships, and determines a vector representation of each gene and pathway in the same low-dimensional vector space. The relevance of a pathway to the given gene set is then captured by the similarity between the pathway vector and gene vectors. To apply this approach to chemosensitivity data, we identify genes whose basal expression levels in a panel of cell lines are correlated with cytotoxic response to a compound, and then rank pathways for relevance to these response-correlated genes using PACER. Extensive evaluation of this approach on benchmarks constructed from databases of compound target genes and large collections of drug response signatures demonstrates its advantages in identifying compound-pathway associations compared to existing statistical methods of pathway enrichment analysis. The associations identified by PACER can serve as testable hypotheses on chemosensitivity pathways and help further study the mechanisms of action of specific cytotoxic drugs. More broadly, PACER represents a novel technique of identifying enriched properties of any gene set of interest while also taking into account networks of known gene-gene relationships and interactions.
| Gene expression levels have been used to study the cellular response to drug treatments. However, analysis of gene expression without considering gene interactions cannot fully reveal complex genotype-phenotype relationships. Biological pathways reveal the interactions among genes, thus providing a complementary way of understanding the drug response variation among individuals. In this paper, we aim to identify pathways that mediate the chemical response of each drug. We used the recently generated CTRP pharmacogenomics data and CCLE basal expression data to identify these pathways. We showed that using the prior knowledge encoded in molecular networks substantially improves pathway identification. In particular, we integrate genes and pathways into a large heterogeneous network in which links are protein-protein interactions and gene-pathway affiliations. We then project this heterogeneous network onto a low-dimensional space, which enables more precise similarity measurements between pathways and drug-response-correlated genes. Extensive experiments on two benchmarks show that our method substantially improved the pathway identification performance by using the molecular networks. More importantly, our method represents a novel technique of identifying enriched properties of any gene set of interest while also taking into account networks of known gene-gene relationships and interactions.
| Large-scale cancer genomics projects, such as the Cancer Genome Atlas [1], the Cancer Genome project [2], and the Cancer Cell Line Encyclopedia project [3], and cancer pharmacology projects, such as the Genomics of Drug Sensitivity in Cancer project [2], have generated a large volume of genomics and pharmacological profiling data. As a result, there is an unprecedented opportunity to link pharmacological and genomic data to identify therapeutic biomarkers [4–6]. In pursuit of this vision, significant efforts have been invested in identifying the genetic basis of drug response variation among individual patients. For instance, a recent study performed a comprehensive survey of genes with basal expression levels in cancer cell lines that correlate with drug sensitivity, revealing potential gene candidates for explaining mechanisms of action of various drugs [7].
While significant efforts have focused on specific genes that interact with compounds and confer observed cellular phenotypes, there has been relatively little progress in studying the synergistic effects of genes. These effects are key factors in comprehensively deciphering the mechanisms of action of compounds and understanding complex phenotypes [8]. Similarly, pathways, which comprise a set of interacting genes, have emerged as a useful construct for gaining insights into cellular response to compounds. Analysis at the pathway level not only reduces the analytic complexity from tens of thousands of genes to just hundreds of pathways, but also contains more explanatory power than a simple list of differentially expressed genes [9]. Consequently, an important yet unsolved problem is the effective identification of pathways mediating drug response variation. Although the associated pathways for certain drugs have been studied experimentally [10–12], in vitro pathway analysis is costly and inherently difficult, making it hard to scale to hundreds of compounds.
Fortunately, a growing compendium of genomic, proteomic, and pharmacologic data allows us to develop scalable computational approaches to help solve this problem. Although statistical significance tests and enrichment analyses can be naturally applied to compound-pathway association identification (e.g., by testing the overlap between pathway members and differentially expressed genes), these approaches fail to leverage well-established biological relationships among genes [13–16]. Even when analyzing individual genes, molecular networks such as protein-protein interaction networks have been shown to play crucial roles in understanding cellular drug response [8, 17–20]. Therefore, we propose to combine molecular networks with gene expression and drug response data for pathway identification. However, integrating these heterogeneous data sources is statistically challenging. Moreover, networks are high-dimensional, incomplete, and noisy. Thus, our algorithm needs to accurately and comprehensively identify pathways while exploiting suboptimal networks.
Here, we present PACER, a novel, network-assisted algorithm that identifies pathway associations for any gene set of interest. PACER first constructs a heterogeneous network that includes pathways and genes, pathway membership information, and gene-gene relationships such as protein-protein physical interaction. It then applies a novel dimensionality reduction algorithm to this heterogeneous network to obtain compact, low-dimensional vectors for pathways and genes in the network. Pathways that are topologically close to the given set of genes (e.g., drug response-related genes) in the network are co-localized with those genes in this low-dimensional vector space. Hence, PACER ranks each pathway based on its vector’s proximity to vectors representing the given genes. We used the proposed algorithm to discover chemosensitivity-related pathways, by applying it to genes whose basal expression level correlates with drug sensitivity. We evaluated PACER’s ability to identify compound-pathway associations with two “ground truth” sets built from compound target data [7] and LINCS differential expression data [21]. When comparing PACER to state-of-the-art methods that ignore prior knowledge of interactions among genes, we observed substantial improvement of the concordance with the chosen benchmarks. Even though we developed PACER and tested its ability to identify compound-pathway associations, the algorithm is applicable to any scenario in which one seeks to discover pathways related to a pre-specified gene set of interest, while utilizing a given gene network.
We obtained a large-scale compound response screening dataset from Rees et al. [7], which spans 481 chemical compounds and 842 human cancer cell lines encompassing 25 lineages. These 481 compounds were collected from different sources including clinical candidates, FDA-approved drugs and previous chemosensitivity profiling experiments. Area under the drug response curve (AUC) was used by the authors of that study to measure cellular response to individual compounds. We also obtained gene expression profiles for these cell lines from the Cancer Cell Line Encyclopedia (CCLE) project [22], profiled using the GeneChip Human Genome U133 Plus 2.0 Array. Since these expression measurements were done in each cell line without any drug treatment, they are referred to as “basal” expression levels. In contrast, the expression profiling of a cell line was performed after treatment with a drug in certain studies such as LINCS L1000 [21] and CMAP [23]. We obtained the SMILE specification of each drug from PubChem [24].
We obtained a collection of six human molecular networks from the STRING database v9.1 [25]. These six networks include experimentally derived protein-protein interactions, manually curated protein-protein interactions, protein-protein interactions transferred from model organism based on orthology, and interactions computed from genomic features such as fusion-fusion events, functional similarity, and co-expression data. There are 16,662 genes in the network. We used all of the STRING channels except “text-mining” and used the Bayesian integration method provided by STRING. Since our approach can deal with different edge weights, we did not set a threshold to remove low-confidence edges. We referred to this integrated network as the “STRING-based molecular network”.
To test whether genes that are highly correlated with many compounds tend to have higher degrees in the network, we formed two groups of genes. One group contained genes that are correlated with over 100 compounds, and the other group contained the remaining genes. We then used the Wilcoxon signed-rank test to test whether the degrees of genes in these two groups were from the same distribution.
We obtained a collection of 223 cancer-related pathways from the National Cancer Institute (NCI) pathway database [26]. These manually curated pathways include human signaling and regulatory pathways as well as key cellular processes.
PACER integrates pathway information with the STRING-based molecular network described above by constructing a heterogeneous network of genes and pathways. An edge exists between two genes if they are connected in the network. An edge exists between a pathway and a gene if the gene belongs to the pathway. There are no direct pathway-pathway edges in the heterogeneous network.
Formally, let A denote the weighted adjacency matrix of the STRING-based molecular network with n genes (or proteins). Let B ∈ {0, 1}n×m denote the gene pathway association matrix, where Bij = 1 if gene i is in pathway j. The heterogeneous network H ∈ R ( n + m ) × ( n + m ) is then defined as:
H i j = { A i j , i ≤ n , j ≤ n B i - n , j T , i > n , j ≤ n B i , j - n , i ≤ n , j > n 0 , i > n , j > n (1)
PACER adopts diffusion component analysis (DCA), a recently developed network representation algorithm to learn a low-dimensional vector for each node in the network [27]. Because of its ability to handle noisy and missing edges in the biological network, DCA has achieved state-of-the-art results in several computational biology tasks [27, 28]. DCA takes H as input. It outputs the d-dimensional vectors V ∈ R ( n + m ) × d for each node in H. According to the definition of H, the first n columns of H are the embedding vectors for genes. The remaining columns of H are the embedding vectors for pathways.
Since compounds are not nodes in the constructed heterogeneous network, only genes and pathways are projected onto the low-dimensional space. After learning the low-dimensional representations of all nodes (genes and pathways), PACER ranks pathways based on the cosine similarities between the low-dimensional representations of the pathway and a set of genes most correlated with response to a compound. Formally, the PACER score sij between pathway i and compound j is defined as:
s i j = ∑ k ∈ RCG ( j ) w k · cos ( V k , V i + n ) , (2)
Here, wk is the weight for gene k. PACER can take input gene weights to weight these cosine similarities. In this paper, we weight the cosine similarities by using the Pearson correlation between the gene expression vector and the chemosensitivity vector. We further calculate an empirical p-value for each compound-pathway association. For a given drug with n response-correlated genes, we use a new, randomly generated set of n genes and compute its pathway association scores using PACER. This is repeated k = 10, 000 times. With m pathways, we then have a total of km PACER scores. The empirical p-value of each original drug-pathway PACER score is its (fractional) rank in this set of PACER scores from random gene sets.
LINCS is a data repository of over 1.3 million genome-wide expression profiles of human cell lines subjected to a variety of perturbation conditions, which include treatments with more than 20 thousand unique compounds at various concentrations. Each perturbation experiment is represented by a list of differentially expressed genes that are ranked based on z-scores of perturbation expression relative to basal expression. For each gene, we first took the difference between its expression in a perturbation condition and its expression in a control condition (i.e., treatment with pure DMSO solvent). We then considered the differential expression of the gene in multiple perturbation experiments involving that compound (i.e., different concentrations, time points, and cell lines). We used the maximum differential expression to represent the compound’s effect on that gene’s expression. All genes were then ranked by their differential expression on treatment with the compound, and the top 250 genes were treated as differentially expressed genes (DEGs) of the compound, provided their z-score has an absolute value greater than 2.
We implemented the method of Huang et al. [13] ourselves using the exact same input (i.e., chemosensitivity and gene expression data) as PACER. We first computed a gene’s correlation to a drug by calculating the Pearson correlation coefficient between the gene’s expression values and the drug response values across cell lines. Let the set of genes in pathway p be denoted by Gp, and their correlation values to a drug d by C(Gp, d). Conversely, the set of genes not in pathway p is denoted as G p ¯, and their correlation values to d as C ( G p ¯ , d ). We then performed the Kruskal-Wallis H test, following Huang et al., to test if the medians of C(Gp, d) and C ( G p ¯ , d ) were significantly different. We used the resulting p-value to rank pathways for each drug.
Following the work of Rees et al. [7], we first examined correlations between the compound sensitivity and basal gene expression profiles across hundreds of cell lines. We calculated Pearson correlation coefficients between each gene’s expression and the cellular response to each compound (measured as AUC, see Methods), across different cell lines (Fig 1A). In contrast to IC50 and EC50 scores, AUC simultaneously captures the efficacy and potency of a drug. Of the 8.7 million pairs of genes and compounds tested, we found 294,789 to be significantly correlated (p-value < 0.0001 after Bonferroni correction, corresponding to a Pearson correlation coefficient of 0.215.) Since the Rees et al. dataset comprises measurements on 842 cell lines, each correlation was computed over 842 pairs of values (drug response, gene expression pairs). This is why even a modest-looking Pearson correlation of 0.215 was deemed highly statistically significant. The key observation from this initial analysis, also noted by Rees et al., is that basal gene expression levels are highly correlated with cytotoxic response for large numbers of compound-gene pairs. Within these significantly correlated pairs, 26 genes were correlated with over 250 compounds (Fig 1B, S1 Table). We note that these key genes tend to be high-degree nodes in the STRING-based molecular network (Wilcoxon rank-sum test p-value < 9.6e-14, see Methods). We also found that some (10 of 481) compounds were significantly correlated (Pearson correlation p-value < 0.0001 after Bonferroni correction) with more than 3,200 genes (Fig 1C). Five of these ten compounds are chemotherapeutic agents (S2 Table). In contrast, about 100 compounds were not significantly correlated with any genes; these compounds are mostly probes that either lack FDA approval or are not clinically used. The large disparity among the examined compounds in terms of the number of correlated genes reflects the diversity of these 481 small molecules. While many of them are chemotherapeutic, which can affect the expression of a large number of genes, some compounds may be targeting specific mutations, post-translational modifications, or protein expression. A closer examination revealed that the compounds with the highest AUC had the fewest gene correlations (i.e., fewest genes whose expression correlates with cytotoxic response) (Fig 1 in S1 Text). This suggests that the strategy of identifying compound-associated genes by correlating basal gene expression profiles with cytotoxicity is likely to be more effective for more potent compounds, for which average response is stronger. Note that the gene expression profiles used here are basal and not in response to treatment with compound, hence it was not clear a priori that more effective compounds would have larger numbers of gene correlates. In summary, examination of individual genes’ correlations with chemical response confirmed previous reports [2, 7, 29] that basal gene expression is significantly correlated with cytotoxicity across cell lines, especially for effective cytotoxic drugs. For each compound, we refer to the top 250 genes whose expression are most significantly correlated with chemosensitivity (Pearson correlation p-value < 0.0001 after Bonferroni correction) as “response-correlated genes” (RCGs) for this compound.
The above evidence for correlations between basal gene expression and chemical response raised the possibility that one might discover important biological pathways associated with the response by a systems-level analysis of gene expression data. To explore this, we considered a collection of 223 cancer-related pathways from the National Cancer Institute (NCI) pathway database [26] and used Fisher’s exact test to quantify the overlap between the set of genes in a given pathway and RCGs. A significantly large overlap between the two sets indicates an association between the pathway and the compound. We performed a multiple hypothesis correction on all pathway association tests for each compound, using FDR = 0.05. The results of this baseline method for predicting pathway associations are shown in Fig 1D (distribution of the number of compounds that are significantly associated with each pathway) and Fig 1E (distribution of the number of pathways significantly associated with each compound). Both distributions revealed a long tail. For instance, while each pathway was associated with an average of 18 compounds (of the 481 tested), there were 10 pathways that were associated with over 150 compounds (S3 Table). Likewise, while each compound was associated with an average of eight pathways, there were 12 compounds associated with over 25 pathways (S4 Table). We show the details of these long tails in Fig 2 in S1 Text.
We observed above that key RCGs (i.e., those correlated with many compounds) tend to be enriched in high degree nodes in the STRING-based molecular network. This suggests that an analysis combining this network with pathway enrichment tests might provide additional insights. We therefore developed a novel network-based method, called PACER, for scoring compound-pathway associations. PACER (Fig 2A) first constructs a heterogeneous network consisting of genes and pathways as nodes. In this network, gene-pathway edges denote pathway memberships based on a compendium of pathways and gene-gene edges from the STRING-based molecular network introduced above (also see Methods). PACER then creates a low-dimensional vector representation for each gene and pathway node in the heterogeneous network, reflecting the node’s position in this heterogeneous network. This is done by the Diffusion Component Analysis (DCA) approach reported in previous work [27, 28]. Nodes (i.e., pathways or genes) will have similar vector representations if they are near each other in the network. For instance, two pathway nodes will have similar vector representations if the pathways share genes and/or their genes are related in the STRING-based molecular network. In a similar vein, two genes will have similar representations if they belong to the same pathway(s) and/or possess the same network neighbors. A gene and a pathway can also be compared in the low-dimensional space, and will be deemed similar if the gene is in the pathway and/or the gene is related in the network to other genes of the pathway. Using the low-dimensional vectors calculated by DCA, PACER next scores a pathway based on the average cosine similarity between the vector representation of the pathway and those of the RCGs. A pathway can thus be found to be associated with a compound if, in the network, the pathway genes are closely related to the compound’s RCGs; this association can be discovered even if the pathway does not actually include the RCGs. We note that scores assigned by PACER are not statistical significance scores and are meant only to rank pathways for association with a given compound. Also, a negative score assigned to a compound-pathway pair does not imply a negative correlation between expression levels of pathway genes and chemosensitivity. Rather, it only implies a lack of evidence for an association between the compound-pathway pair. Since pathway association analysis is likely to be meaningless for compounds with very few RCGs, we limited the following analysis to the 330 compounds for which more than 5 RCGs were identified.
The PACER association scores for all combinations of 330 compounds and 223 NCI signaling pathways are shown in Fig 2B. Since PACER scores are not easily assigned statistical significance levels, we chose to examine, for each compound, the n pathways with the highest PACER scores, where n is the number of statistically significant pathway associations (FDR < 0.05) found by the baseline method above for the same compound. (This choice also allows a fair comparison between the two methods in subsequent sections.) We found literature support for several of these associations. For example, PACER analysis associates ruxolitinib, a JAK/STAT inhibitor, with integrin-linked kinase signaling pathway. In a previous study, it was shown that beta 4 integrin enhances activation of the transcription factor STAT3, which is a target of ruxolitinib [30]. Fig 2B reveals that the pathways cluster into many distinct groups, each with different compound association profiles. In some cases, we noted functionally related pathways being grouped together. For example, one group consists of pathways describing various integrin cell surface interactions including “integrin family cell surface interactions”, “alpha E beta 7 integrin cell surface interactions”, “alpha 6 beta 4 integrin-ligand interactions”, and “beta 5 beta 6 beta 7 and beta 8 integrin cell surface interactions” (marked as blue rectangle in Fig 2B). These pathways are known to play crucial roles in communications among cells in response to small molecules [31]. Notably, the integrin-mediated pathways promote invasiveness and oncogenic survival, and contribute to cancer cell survival and resistance to chemotherapy [32, 33]. Another group consists of different interleukin signaling pathways including “IL4-mediated signaling events”, “IL8- and CXCR1-mediated signaling events”, “IL3-mediated signaling events”, and “IL2 signaling events mediated by PI3K” (marked as green rectangle in Fig 2B). Our analysis found that this group of pathways is associated with decitabine. A recent study shows that decitabine’s effect of PD-1 blockade-based immunotherapy is enhanced in colorectal cancer through upregulation of many immune-related genes [34].
Fig 2B also shows compounds clustered into different groups based on their associations with pathways. We noted examples where many compounds with similar structures were grouped together. For example, teniposide and etoposide had a Tanimoto similarity score of 0.94 between their SMILE specifications, which was substantially higher than the average Tanimoto similarity score of 0.3716 for all pairs of compounds. They were clustered together in the same group (marked as black rectangle in Fig 2B), which had seven compounds. Among the pathways that are associated with this group, we found a set of similar pathways, including “p53 pathway”, “direct p53 effectors”, “signaling mediated by p38-alpha and p38-beta”, and “signaling mediated by p38-gamma and p38-delta”. We found support in the literature in favor of some of these associations. For example, a previous study reported that etoposide activates p38MAPK and can be used as a combined treatment approach when used with p38MAPK inhibitor SB203580 [35]. As another example, temsirolimus and tacrolimus, which are both epipodophyllotoxins and inhibit topoisomerase II, have a Tanimoto similarity score of 0.82, and are grouped closely in Fig 2B (marked as pink rectangle in Fig 2B).
We noted a substantial degree of complementarity between the top predictions of PACER and those of the baseline method that uses Fisher’s exact test between RCGs and pathway genes (see S5 Table). For instance, PACER found that bexarotene is associated with the “IGF1 pathway”. A recent study showed that treating rats with high doses of bexarotene substantially decreased serum IGF1 levels [36]. The baseline approach did not find this association to be significant. Similarly, PACER reported that the “ATM pathway” is associated with simvastatin, while the baseline method did not. Simvastatin has been reported to activate ATM when it is used to treat chronic lymphocytic leukemia patients [37].
For a more systematic comparison between the two methods, we evaluated PACER based on a database of known compound targets. We performed the evaluation under the assumption that a pathway containing at least one known target is an associated pathway. Huang et al. suggested and used this approach [13]. We used it here to evaluate PACER, the baseline method, as well as a third method presented by Huang et al. [13] Although this third method was proposed to detect associations between pathways and drug clades, it can directly detect pathway-compound associations. We implemented the method ourselves (see Methods) and included it in our evaluations. We obtained the known targets for our compound set from Rees et al. [7] and STITCH database [38]. We then computed the AUROC of pathway predictions made by PACER for each compound, and plotted this information alongside analogous information for the baseline method and the method of Huang et al. [13] As shown in Fig 2C, PACER identified pathways with higher AUROC compared to the other two methods. For example, PACER identified pathways with an AUROC greater than 0.75 for 23 different compounds, while the baseline method achieved this level of AUROC for only 5 compounds. Table 1 shows the 10 compounds for which PACER achieved highest AUROC (Fig 4-7 in S1 Text).
We further compared the associations predicted by the three methods to those identified from an external data set. We mined the Library of Integrated Network-Based Cellular Signatures (LINCS) L1000 data [21], which reports genes differentially expressed upon treatment of various cell lines with a compound. For each compound in our analysis that is also included as a perturbagen in the L1000 compendium, we established a LINCS-based benchmark of significantly associated pathways. This was based on a Fisher’s exact test (p-value < 0.05) between pathway genes and the most differentially expressed genes from treatments with the same compound (see Methods). We required this criterion to be met in at least one of the cell lines for which data was available from LINCS. We then assessed the concordance between this set of LINCS-based compound-pathway associations and those predicted by either method presented above. We recognize that this is not an ideal benchmark: LINCS data points to genes (and, indirectly, to pathways) that are differentially expressed in response to treatment, while PACER and the compared methods base their pathway predictions on genes that have basal expression levels across cell lines that correlate with chemical response. At the same time, we expect the pathways affected by chemical treatment to also be, to an extent, involved in interpersonal variation of chemosensitivity, making this a suitable evaluation procedure. This was inspired by similar observations in cancer biology: genes and pathways disrupted in cancer tissues overlap with genes and pathways whose mutation status in germline non-tumor samples is informative about disease susceptibility and progression.
To test whether the significant pathways identified from LINCS data agree with the pathways predicted by one of the methods being evaluated (based on chemical response variation in CCLE cell lines), we counted the compounds for which the two sets of predicted pathways overlapped significantly (Fisher’s exact test p-value < 0.05). As shown in Fig 2D, the PACER approach predicts pathways concordant with the corresponding LINCS-based benchmark for more compounds, compared to the baseline method and that of Huang et al. [13] For instance, when the baseline method used an FDR threshold of 10% to designate significant pathway associations for each compound, and the PACER method predicted the same number of pathways, the latter’s predictions were concordant with the LINCS-based benchmark for 118, a nearly two-fold improvement over the baseline method’s predictions. Our evaluations actually provide evidence for the above-mentioned possibility that pathways predictive of drug sensitivity overlap with genes that mediate drug response. In fact, we found 113 compounds for which the pathways identified from basal expression correlations and the pathways identified from LINCS signatures overlap with FDR < 5%.
After observing the substantial improvement of PACER, we then investigated whether the performance of PACER is stable when using only experimental derived protein-protein interactions as input. We found that this is indeed the case, as per the two evaluation strategies presented above (Fig 8-9 in S1 Text). We further demonstrated that the result of our method is robust to different numbers of top response-correlated genes used in PACER, as shown in Fig 10-11 in S1 Text. We compared different values for ‘k’ in the ‘top k’ genes chosen by PACER. We found that results were comparable when using k = 100, 150, 200, 250 and 300. This demonstrates the stability of the algorithm’s performance to different but reasonable values of k in its choice of top k response-correlated genes.
We have shown that embedding prior knowledge in a gene network can more accurately identify compound-pathway associations. Our new method, called PACER, identified many compound-pathway associations that are supported by known compound targets as well as literature evidence. Due to its unique ability to incorporate any suitable compendium of gene interactions, our approach may provide complementary insights into drug mechanisms of action.
Historically, pathways associated with a particular gene set are identified by using popular statistical methods such as Gene Set Enrichment Analysis [39], Fisher’s exact test, or the Binomial test (Reactome [40]). These tools test the overlap between differentially expressed genes and pathway members. They may also be applied to the set of drug-response-correlated genes (RCGs) analyzed here. Ingenuity Pathway Analysis [41] is another related tool, which utilizes information about causal interactions between pathway members. Our study is similar to the above tools in that PACER also seeks to find pathways implicated by a gene set. However, our approach differs from these existing tools in that known molecular interactions (e.g., PPI) among different genes are taken into consideration. Thus, a gene set, be it the RCGs of a compound or the members of a pathway, is not treated merely as the sum of its parts, but also includes the relationships among those parts. Since the dominant theme in existing approaches is assessment of overlaps between two gene sets (MSigDB, DAVID, and Reactome adopt variations on this theme), our extensive comparisons between PACER and the baseline method of Fisher’s exact test shed light on the relative merits of the new approach. A related line of work aims to identify differentially expressed subnetworks in a given interaction network, e.g., KeyPathwayMiner [42], but these studies are only superficially relevant to our work since we aim to prioritize existing pathways instead of finding new pathways.
We consider two potential reasons for the strong performance of PACER. First, it is widely appreciated that a chemical compound not only affects individual genes, but also combinations of genes in molecular networks corresponding to core processes, such as cell proliferation and apoptosis. Our method postulates that even if the RCGs and a pathway may only have a few genes in common, they may be close to each other in the network. Although current compound pathway maps are incomplete, much relevant information is available in public databases of human molecular networks. While traditional pathway enrichment analysis methods like Fisher’s exact test identify pathways according to the number of shared genes, PACER prioritizes pathways based on their proximities to RCGs in molecular networks. Second, manually curated pathways may have arbitrary boundaries due to the need to capture knowledge at different levels of detail. Consequently, identifying drug-related pathways might be hindered by pathway boundaries. By leveraging the prior knowledge in molecular networks, PACER is more robust to the noise in pathway boundaries, thus improving the sensitivity of detecting compound-pathway associations.
We see many opportunities to improve upon the basic concept of PACER in future work. First, although the current PACER framework was developed in an unsupervised fashion, the scores assigned to each pathway for the given gene set can be used as the feature and plugged into off-the-shelf machine learning classifiers for compound-pathway association identification. Second, although this study focused on chemosensitivity response, the PACER method is broadly applicable to testing the association between two sets of genes according to their proximity in the network. Finally, although we use gene expression data as the molecular profile of each cell line, it might be interesting to test our method based on other molecular data such as somatic mutations and copy number alterations.
|
10.1371/journal.pntd.0005260 | The Fleas (Siphonaptera) in Iran: Diversity, Host Range, and Medical Importance | Flea-borne diseases have a wide distribution in the world. Studies on the identity, abundance, distribution and seasonality of the potential vectors of pathogenic agents (e.g. Yersinia pestis, Francisella tularensis, and Rickettsia felis) are necessary tools for controlling and preventing such diseases outbreaks. The improvements of diagnostic tools are partly responsible for an easier detection of otherwise unnoticed agents in the ectoparasitic fauna and as such a good taxonomical knowledge of the potential vectors is crucial. The aims of this study were to make an exhaustive inventory of the literature on the fleas (Siphonaptera) and range of associated hosts in Iran, present their known distribution, and discuss their medical importance.
The data were obtained by an extensive literature review related to medically significant fleas in Iran published before 31st August 2016. The flea-host specificity was then determined using a family and subfamily-oriented criteria to further realize and quantify the shared and exclusive vertebrate hosts of fleas among Iran fleas. The locations sampled and reported in the literature were primarily from human habitation, livestock farms, poultry, and rodents’ burrows of the 31 provinces of the country. The flea fauna were dominated by seven families, namely the Ceratophyllidae, Leptopsyllidae, Pulicidae, Ctenophthalmidae, Coptopsyllidae, Ischnopsyllidae and Vermipsyllidae. The hosts associated with Iran fleas ranged from the small and large mammals to the birds. Pulicidae were associated with 73% (56/77) of identified host species. Flea-host association analysis indicates that rodents are the common hosts of 5 flea families but some sampling bias results in the reduced number of bird host sampled. Analyses of flea-host relationships at the subfamily level showed that most vertebrates hosted fleas belgonging to 3 subfamilies namely Xenopsyllinae (n = 43), Ctenophthalminae (n = 20) and Amphipsyllinae (n = 17). Meriones persicus was infested by 11 flea subfamilies in the arid, rocky, mountainous regions and Xenopsyllinae were hosted by at least 43 mammal species. These findings place the Persian jird (M. persicus) and the Xenopsyllinae as the major vertebrate and vector hosts of flea-borne diseases in Iran including Yersinia pestis, the etiological agent of plague. We found records of at least seven vector-borne pathogenic agents that can potentially be transmitted by the 117 flea species (or subspecies) of Iran.
Herein, we performed a thorough inventary of the flea species and their associated hosts, their medical importance and geographic distribution throughout Iran. This exercise allowed assessing the diversity of flea species with the potential flea-borne agents transmission risk in the country by arranging published data on flea-host associations. This information is a first step for issuing public health policies and rodent-flea control campaigns in Iran as well as those interested in the ecology/epidemiology of flea-borne disease.
| The data about flea-borne emerging or re-emerging infections throughout Iran are limited. This paper showed that the flea fauna of Iran were dominated by seven families. Moreover flea-host association analysis indicates that rodents are common hosts of flea families and most vertebrates hosted fleas belonging to the subfamilies Xenopsyllinae, Ctenophthalminae and Amphipsyllinae. We showed that the Persian jird (Merions persicus Blanford, 1875) and the Xenopsyllinae are respectively the major vertebrate and potential vectors of flea-borne diseases in Iran. Further efforts are needed to inventorize and screen molecularly wild and domestic mammals flea fauna (>3kg) in order to monitor the risk of and control flea-borne infections in Iran, especially in the ecoregions with high diversity of flea and host species and in the old endemic plague foci of the country.
| Vector-borne diseases (VBDs) are globally responsible for more than 17% of all infectious diseases [1]. There are a large number of viral, rickettsial, bacterial and parasitic diseases that are transmitted by insect vectors [2]. In the last two decades, many zoonotic VBDs have emerged in areas where they previously did not occur, and the incidence of these diseases both in endemic areas and outside their known range has increased [3]. In recent years, most studies on zoonotic diseases have focused on tick- and mosquito-borne diseases, less attention has been given to flea-borne diseases[4].
Fleas (Siphonaptera) are small, bloodsucking or hematophagous ectoparasites that may transmit pathogens through several possible mechanisms, including: contaminated feces (e.g. R. typhi, B. henselae), soiled mouthparts (e.g. Y. pestis, viral pathogens), regurgitation of gut contents (e.g. Y. pestis), and infectious saliva (e.g. R. felis in salivary glands)[4].
Over 2500 flea species belonging to 16 families and 238 genera have been described worldwide [5]. Fleas are mainly ectoparasites of mammals while birds are infested by only 6% of the known species. This is partly due to reduced collection efforts and sampling bias as only few bird fleas are in close contact with humans [6]. Fleas are one of the most common insect groups that can serve as vector and intermediate host of pathogenic zoonotic agents between vertebrate hosts, including humans [4, 7–8]. Fleas can have a direct pathogenic effect by causing allergic dermatitis [9–10] or paralysis subsequent to the injection of saliva into their hosts skin or blood [11]. Notorious human pathogens such as Yersinia pestis (plague), Rickettsia typhi (murine typhus), Francisella tularensis (tularemia) and Bartonella henselae (cat scratch disease) are transmitted by fleas [12–15].
Some fleas tend to be host specific (restricted or specialist), but others have a wide host range (permissive, opportunistic). The permissive species group are more significant than the restricted ones, because they can spread infectious agents among and within their multiple hosts and across a diverse series of habitats [6]. In order to prevent or control the occurrence and spread of flea-borne diseases, it is thus necessary to establish a taxonomical inventory of the flea fauna and their specific distribution range.
Climate changes, due to global warming and human intervention, have led to changes in the biological parameters and distribution ranges of vectors and hence of VBDs [16]. On the bases of vulnerability assessments and models, it is predicted that climate change will result in raised incidence of communicable diseases embracing VBDs; however the short and long term effects will be mitigated and will be linked to vector life cycles (e.g.: developments of preimaginal stages) and geographic area [17]. Reasonable proofs tend to suggest that changes in climatic factors may affect VBDs incidence especially acting on the off-host developmental life stages of arthropods and hence disease transmission dynamics. Insects as poikilotherm organisms have no internal control of their body temperatures, and as such depend on their host(s)—the imago as a transient habitat -, and abiotic conditions for survival, which both condition their vector capacity, as well as their reproduction rate[18]. Moreover, vector capacity is linked to the nature of the pathogen transmitted, survival rate inside its vector host—which may or may not affect vector fitness—and incubation or turnover rate that is inversely proportional to temperature[19]. Moreover, climate and human behavior changes increase human exposures to vectors and the pathogenic agents they transmit [20]. Studies of plague transmission in the U.S.A, China and Kazakhstan have found that the patterns of human or rodent plague are shifting as temperatures warms up or link to climatic oscillations (such as El Niño) and precipitation pattern [20].
Iranian physicians were familiar with the human plague for a long time. Although there are little information about the situation of plague from earlier centuries, more documented evidence are available from the 19th and 20th centuries. As a matter of fact, faunistic studies of Iranian fleas have been carried out mainly about 60 years ago in a context of plague research and most species described at the time were collected and described off plague hosts [21]. When plague research stopped, flea inventories did so too and there are no current updates on the flea fauna of Iran. However, a recent study detected antibodies against Y. pestis in dogs—known to be a good sentinels for plague surveillance- while human plague hasn’t been reported for 50 years [22]. This finding triggered some concern about the possible plague reemergence in the countryside, in the old plague foci and called for an update on the state-of-knowledge of the flea diversity in the country. The aims of the present study were to update by reviewing the current state of knowledge of the Iranian Siphonaptera diversity, their host range and especially the medically important species.
This review was based on a search of the online scientific databases (Scientific Information Database) PubMed and Google Scholar from 1952 through 31st August 2016. Keywords—submitted in English, French, Turkish and Russian—for the search were “flea AND fauna AND Iran”; “Iran AND puce”, “Iran AND siphonaptera”; “Iran AND ectoparasite”. Searches were conducted in the titles, abstracts, keywords and full text. The majority of our knowledge on the Siphonaptera of Iran is derived from plague studies[23], the concept of “telluric plague” is coeval with these researches[24] and studies of two flea specialists, the Iranian Farhang-Azad and the French J.M. Klein.
In each case the flea species, its host, and location of sampling were extracted from the published papers. The flea distribution maps were prepared using ArcGIS (ArcGIS version 9.3, ESRI). An online software were used to further classify and quantify the shared and exclusive vertebrate hosts of fleas with the “family or subfamily” filtering criteria[25].
The data for this study were extracted from about 100 relevant papers in English, French, Istanbul Turkish or Russian. Faunistic reviews of the medically significant fleas showed the presence of fleas through 31 Iranian provinces (Fig 1). In the old classification of Iran provinces used by Farhang-Azad (1972b), the Khorasan province, which was the largest province of Iran in the plague research era, is currently divided in three provinces namely Razavi Khorasan, North Khorasan, and South Khorasan. This means that the spatial scale of the flea range resolution is less accurate in the old literature as it covers a larger area where the flea and their host are not homogenously found. Based on the information in the studied papers, the sampling locations mainly were human houses, animal husbandry premises, poultry farms, and rodents’ burrows.
According to the literature, about 117 species or subspecies of fleas belonging to 7 families and 35 genera have been described in Iran. Most flea species reported in the studied literature belonged to the Ceratophyllidae (n = 33), Leptopsyllidae (n = 24), Pulicidae (n = 21), Ctenophthalmidae (n = 20) and Coptopsyllidae (n = 9) families. The flea species of the Ischnopsyllidae (bat-fleas) and Vermipsyllidae (carnivore-fleas) families consisted of only 6 and 4 species of the whole collection respectively (Tables 1 and 2).
The Ceratophyllidae, the more represented flea family, consisted of 33 species belonging to 6 genera, comprising Callopsylla, Ceratophyllus, Citellophilus, Myoxopsylla, Nosopsyllus and Paraceras.
The Leptopsyllidae, bird and rodent fleas, consisted of 24 species consisting of 10 genera including Amphipsylla, Caenopsylla, Ctenophyllus, Frontopsylla, Leptopsylla, Mesopsylla, Ophthalmopsylla, Paradoxopsyllus, Peromyscopsylla and Phaenopsylla.
The Ctenophthalmidae consisted of 20 species belonging to 7 genera comprising Ctenophthalmus, Doratopsylla, Neopsylla, Palaeopsylla, Rhadinopsylla, Stenoponia and Wagnerina.
The Pulicidae, a cosmopolitan family of the most notorious plague vectors (genus Xenopsylla), included 21 species distributed in 7 genera comprising Archaeopsylla, Ctenocephalides, Echidnophaga, Pulex, Synosternus, Parapulex, and Xenopsylla.
The Coptopsyllidae was limited to 9 species in the genus Coptopsylla.
In the above-mentioned five families, the most commonly reported fleas belong to the genera Nosopsyllus (Ceratophyllinae), Xenopsylla (Xenopsyllinae), Ctenophthalmus (Ctenophthalminae) Coptopsylla (Coptopsyllidae) Amphipsylla (Amphipsyllinae), Leptopsylla (Leptopsyllinae), and Mesopsylla (Mesopsyllinae). Detailed information is presented in Table 1.
The hosts associated with Iran fleas ranged from the small mammals (Rodentia, Chiroptera, Lagomorpha, Insectivora) to the large mammals (Ungulata, Carnivora, Primates, Artiodactyla) and birds as well. On the whole, 166 vertebrate host species were reported infested by fleas in Iran in the literature including Pulicidae (n = 56), Ceratophyllidae (n = 38), Ctenophthalmidae (n = 29), Leptopsyllidae (n = 22), Coptopsyllidae (n = 11), Ischnopsyllidae (n = 7) and Vermipsyllidae (n = 3). By filtering the compiled data, we recognized 77 vertebrate host species among all seven flea families.
Eight potential mammals were hosted by ≤7 flea (sub-) family respectively; these were: Calomyscus bailwardi (7), Meles meles (7), Mus musculus (7), Meriones vinogradovi (8), Vulpes vulpes (8), Cricetulus migratorius (9), Meriones libycus (9) and Meriones persicus (11). Actually flea (sub-) families can infest ≥10 vertebrate hosts were Xenopsyllinae (n = 43), Ceratophyllinae (n = 37), Archaeopsyllinae (n = 20), Ctenophthalminae (n = 20), Pulicinae (n = 19), Amphipsyllinae (n = 17), Stenoponiinae (n = 12) and Coptopsyllidae (n = 11). Detailed information is presented in Table 3.
At least 23, 6, 5, 5 and 1 host species are exclusively infested by Pulicidae, Ischnopsyllidae, Ceratophyllidae, Ctenophthalmidae and Leptopsyllidae respectively. However restricted host species was not found in the Coptopsyllidae and Vermipsyllidae (Table 4).
A total of 53 vertebrate species were reported infested by six subfamilies of Ctenophthalmidae including Ctenophthalminae (n = 20), Stenoponiinae (n = 12), Rhadinopsyllinae (n = 9), Anomiopsyllinae (n = 6), Doratopsyllinae (n = 3) and Neopsyllinae (n = 3). By filtering the compiled data, 29 vertebrate host species were distinguished among all six subfamilies. Correspondingly 8, 6 and 1 host species are exclusively included in the Ctenophthalminae, Stenoponiinae and Doratopsyllinae. However there were not found any restricted vertebrate host species in the Anomiopsyllinae, Neopsyllinae and Rhadinopsyllinae (Table 5).
A total of 33 vertebrate species were reported infested by three subfamilies of Leptopsyllidae including Amphipsyllinae (n = 17), Mesopsyllinae (n = 9) and Leptopsyllinae (n = 7). By filtering the compiled data, 22 vertebrate host species were distinguished among three subfamilies. Investigation on the flea-host associations in subfamilies of the Leptopsyllidae showed that there were no common host species shared by the three subfamilies. However 6, 3 and 2 host species are exclusively included in the Amphipsyllinae, Leptopsyllinae and Mesopsyllinae respectively (Table 6).
A total of 83 vertebrate species were reported infested by three subfamilies of Pulicidae including Xenopsyllinae (n = 43), Pulicinae (n = 20) and Archaeopsyllinae (n = 20). By filtering the compiled data, 56 vertebrate host species were distinguished among three subfamilies. Exploration of flea-host associations in Pulicidae pointed out that there are eight common hosts including Capra hircus (Linnaeus, 1758), Hemiechinus auritus (Gmelin, 1770), Herpestes auropunctatus (Hodgson, 1836), Hyaena hyaena (Linnaeus, 1758), Meles meles (Linnaeus, 1758), Ovis aries (Linnaeus, 1758), Rattus rattus (Linnaeus, 1758) and Vulpes vulpes (Linnaeus, 1758) among three subfamilies. Although a number of 27, 5 and 5 host species are exclusively included in the Xenopsyllinae, Pulicinae and Archaeopsyllinae respectively (Table 7).
The literature inventory of the fleas of Iran showed that there are seven Siphonaptera families in this country namely Ceratophyllidae, Leptopsyllidae, Pulicidae, Ctenophthalmidae, Coptopsyllidae, Ischnopsyllidae and Vermipsyllidae. These flea families are distributed in all parts of the country where sampling occurred and where data were available. According to the literature reviewed, the distribution range of those families extends in Hamadan and Kurdistan (West Iran) provinces rather than in Ardabil (northwest), Northern Khorasan (northeast), Bushehr (south), Mazandaran, Golestan and Gilan provinces (north). This fact is partly due to a collection bias in plague foci during the sixties (1963–1975 Baltazard, Klein, Farhang-Azad and Mollaret)[65–70]. The distribution maps of the studied fleas showed that further sampling, especially from provinces with poor faunistical studies, is necessary, especially in a context of vector-borne disease epidemiology where known mammalian hosts of pathogenic agents are also present.
Most fleas of medical or veterinary importance belong to the Ceratophyllidae, Leptopsyllidae, Pulicidae, Ctenophthalmidae and Vermipsyllidae families [12]. Pulicidae, a family including most cosmopolitan flea species of medical importance and in particular the Xenopsylla genus, was by far the most reported family in Iran [8, 29–30, 32, 35, 53–55, 57–60]. Analysis of common mammal hosts and their flea diversity revealed that M. persicus was infested by 11 flea subfamilies and Xenopsyllinae were hosted by at least 43 mammal species.
The Persian Jird, M. persicus, is distributed from Eastern Anatolia to Afghanistan and western Pakistan. Iran is the most extensive geographical region in the distribution range of the Persian Jird; indeed five of the six subspecies are found in the country [71].
At the first, the research team of Baltazard (1952) and then Golvan & Rioux (1963) and Poland and Dennis (1999) offered initial illustrations of the role of resistant or silent enzootic reservoirs in the maintenance of Y. pestis and human plague outbreaks in the Kurdistan focus. They showed that M. vinogradovi and M. tristrami were extremely sensitive to Y. pestis while M. libycus and M. persicus were highly resistant. Tatera indica has also been associated with transmission of Y. pestis in the country. Flea densities were reported to be high on M. persicus [23, 72–73]. In that era flea species including Pulex irritans, Xenopsylla cheopis, X. astia, X. buxtoni, X. conformis, Nosopsyllus fasciatus N. iranus iranus, and Stenoponia tripectinata were listed as favorite candidate Y. pestis vectors within and among vertebrates including man [74–79].
In 1980, Karimi et al. surveyed the Sarab focus in East Azarbaijan province where fourteen samples of Y. pestis were isolated from M. persicus, M. vinogradovi, and Mesocricetus auratus and from their fleas; Xenopsylla conformis and Nosopsylla iranus iranus [80]. The Y. pestis strains isolated from the M. persicus in the Trans-Arax focus in Armenia were characterized by higher virulence than those that are isolated from voles in the Transcaucasus Mountainous focus[81].
In a recent serological survey carried out by Esmaeili et al., in Western Iran antibodies against Y. pestis F1 capsular antigen were detected in a M. persicus [22]. Whether Y. pestis strains lacking the F1 antigen naturally occur in Iran is not known but could lead to an underestimation of the current seroprevalence.
Meriones species notably M. persicus were reported to be main reservoir host for pathogens rather than bacterium Y. pestis. In the parasitological studies sandfly-borne Leishmania spp. including L. major [82], L. infantum [83] and L. donovani [84] were isolated from M. persicus specimens. Meriones species rather than M. persicus (M. libycus and M. hurrianae) have been reported as the major reservoir host of zoonotic cutaneous leishmaniasis in several endemic areas of Iran [85–89]. The endoparasites ranging from Acanthocephala to Cestoda and Nematoda were identified in M. persicus as well [90]. These findings place the Persian jird and the Xenopsyllinae as the major vertebrate and vector hosts of flea-borne diseases in Iran including Y. pestis, the etiological agent of plague.
Indeed, Xenopsylla spp. were collected from 18 provinces with a wide array of climatic conditions ranging from cold mountainous areas to warm and dry sandy plains and deserts (Table 1).
Most species of the Pulicidae family are notorious vectors of disease agents causing plague, murine typhus, and tularemia but also transmit helminths. Several species of the Xenopsylla genus play an important role in the transmission of Y. pestis, the etiological agent of plague, from rodents to human [91]; the most classical and significant vector being X. cheopis [92].
Indeed, X. cheopis accounts for 80% of the fleas collected off rodent hosts in the natural endemic plague foci of Iran [93]. X. cheopis is also the vector of various human pathogenic Bartonella species [6, 94]. The cat scratch disease, caused by B. henselae, has been considered as an emerging zoonotic bacterial pathogen in veterinary and human medicine. Cats are the basic source of the bacteria. Bacteria are transferred from cat to cat by the flea Ctenocephalides felis, another cosmopolitan flea, which have been reported in the Iranian cat population [95]. Murine typhus or endemic typhus caused primarily by Rickettsia typhiis another rodent-borne disease that is transmitted to humans by the flea X. cheopis [96]. Pulex irritans and Nosopsyllus fasciatus are secondary vectors of murine typhus Rickettsia [97] that is endemic through coastal regions of the Caspian Sea and the Persian Gulf [98].
Rickettsia felis is the cause of another flea-borne “spotted fever group” rickettsiosis. R. felis is transmitted by the bite or faeces of several flea species, and transovarially in Ctenocephalides felis felis (and the African subspecies C. f. strongylus) but also in C. orientis present in Iran, so that they are considered as vectors and reservoir hosts of this pathogen [99].
Ctenocephalides felis, C. canis- that have been collected from the studied areas extensively (Table 1)—and P. irritans are the intermediate hosts of flatworms such as Dipylidium caninum, or nematodes as the filaria Acanthocheilonema reconditum. Hence dog, cat and rarely human infection occurs following ingestion of infected fleas [100–101]. Typically, a human is bitten more often by a cat flea (C. felis) than a dog flea (C. canis) which is very or even monospecific. Cosmopolitan fleas as helminths vector have less medical than veterinary importance, since the helminth species they transmit rarely infest humans and are virtually harmless.
Nosopsyllus fasciatus, a Ceratophyllidae and Coptopsylla lamellifer, a Coptopsyllidae, were collected in 14 different regions of Iran. They play a role in enzootic plague cycles, that is in circulating the plague bacterium Y. pestis between rodents but since they do not readily bite humans in a natural setting, are only accidental vectors of Y. pestis to humans exposed [38, 41, 102–103].
Fleas are also considered vectors of F. tularensis the etiological agent of tularemia [104]. Vulnerable animals such as hares and rodents frequently die in large numbers during epizootics. Human infections take place through several routes, including insect bites and direct contact to an infected animal. It can affect the skin, eyes, lymph nodes, lungs and, less often, other internal organs. According to recent studies (which have shown the presence of this disease in western and eastern regions of Iran) and the previous studies (which have shown the presence of this disease in the east and north-west of the country [105]), the possibility of transmission of this agent by fleas should be considered in all parts of the country [106].
Most leptopsyllids parasitize rodents and a few birds. Species of Frontopsylla, Leptopsylla, Mesopsylla, Ophthalmopsylla and Paradoxopsyllus are known as main or suspected vectors of plague, murine typhus, erysipeloid, listeriosis and salmonellosis in the Central Asia[107]. In an experiment it was showed that L. segnis is more successful in transmitting R. typhi to rats than X. cheopis [64]. Leptopsylla aethiopica aethiopica which transmits plague in Africa recently have been reported from Semnan province [50]; however its presence and identity in the region is very questionable.
People who travel to rural areas should consciously avoid flea bites especially in populations camping outside (herders, travelers, nomads) and avoid exposure to wild rodents and their fleas. In domestic areas, in order to prevent bites and thus disease transmission to humans, the floors and walls, as well as the rodents’ burrows around settlements, should theoretically be sprayed with insecticides. A few days later the application of rodenticides is necessary.
There were virtually no records of some flea species in a few provinces like North Khorasan (Fig 1). This is mainly due to inadequate inventories, especially in remote areas, or minorly due to the changing of geographical boundaries where the number of provinces in old classification has increased from 10 to 31 provinces.
In this paper we highlighted the geographical gaps on the Siphonaptera fauna of Iran. Generally, it shows that extensive fundamental and systematic research is still needed to determine the impact of off-host abiotic conditions and host identity (either mammal or bird) on host specificity, and on the potential for flea-borne diseases spread and transmission risk.
Co-evolution partly explains host-flea relationships which are translated into various degrees of host specificity (as shown in Tables 4–7) and morphological adaptations of the parasite [108]. Host specificity is important from the perspective of transmission of disease agents. It is more probable that, vertebrate hosts with related taxonomy or similar ecologies will have flea species that share similar pathogens. Depending on the level of infestation, flea species do not cause major problem to their hosts [108]. While some fleas species, virtually exclusively females, (Echidnophaga spp., Vermipsylla spp., Dorcadia spp., Tunga spp), spend much of their adult lives embedded or fixed in the host skin, this is far from being the rule. Indeed, most species jump on a host to feed intermittently before returning to the host dwelling place, usually a nest or burrow [6].
Den/nest making hosts (mammals or birds) display a more specific flea fauna than non roosting species [6]. It has been shown that fleas possibly appeared with mammals and speciated with rodents which still have the most speciose extant fauna (74%)[109].
Since rodent-borne, bat-borne and vector-borne diseases are the major rising concerns to health authorities, and threats to public health making inventories of the host and their ectopoarasitic fauna has become as never before a priority. Although most flea-borne diseases are not classified in the 17 neglected tropical diseases (NTDs) list made by the World Health Organization, this doesn’t mean those are unimportant or not causing an underestimated morbidity burden worldwide. The lack of recognition by major stakeholders, and the local lack of diagnostic tools and awareness are impeding improvements into flea-borne disease research. However, with about seven human or zoonotic highly pathogenic agents circulating among -possibly- the 117 flea species throughout Iran, there is an urgent need to organize and fund flea-host-pathogen ecological surveys in the face of rapid environmental and human behavioral changes.
The first step in identifying the risk linked to flea exposure is to make a list of the species before any public health measures can be taken. Flea-borne diseases are caused by emerging and re-emerging infectious agents which distribution, prevalence and incidence are currently increasing. However, the data about fleas and their medical significance in different geographical regions of Iran is limited. We took the first step in this paper but supplementary studies are required to i) complete the list, especially in areas where there are no reportsor poor faunistic studies and ii) perform molecular screening of flea pools in order to detect specific pathogen circulation in domestic fauna and wildlife in order to prevent future epidemics.
|
10.1371/journal.pntd.0004780 | Discovery of Point Mutations in the Voltage-Gated Sodium Channel from African Aedes aegypti Populations: Potential Phylogenetic Reasons for Gene Introgression | Yellow fever is endemic in some countries in Africa, and Aedes aegpyti is one of the most important vectors implicated in the outbreak. The mapping of the nation-wide distribution and the detection of insecticide resistance of vector mosquitoes will provide the beneficial information for forecasting of dengue and yellow fever outbreaks and effective control measures.
High resistance to DDT was observed in all mosquito colonies collected in Ghana. The resistance and the possible existence of resistance or tolerance to permethrin were suspected in some colonies. High frequencies of point mutations at the voltage-gated sodium channel (F1534C) and one heterozygote of the other mutation (V1016I) were detected, and this is the first detection on the African continent. The frequency of F1534C allele and the ratio of F1534C homozygotes in Ae. aegypti aegypti (Aaa) were significantly higher than those in Ae. aegypti formosus (Aaf). We could detect the two types of introns between exon 20 and 21, and the F1534C mutations were strongly linked with one type of intron, which was commonly found in South East Asian and South and Central American countries, suggesting the possibility that this mutation was introduced from other continents or convergently selected after the introgression of Aaa genes from the above area.
The worldwide eradication programs in 1940s and 1950s might have caused high selection pressure on the mosquito populations and expanded the distribution of insecticide-resistant Ae. aegypti populations. Selection of the F1534C point mutation could be hypothesized to have taken place during this period. The selection of the resistant population of Ae. aegypti with the point mutation of F1534C, and the worldwide transportation of vector mosquitoes in accordance with human activity such as trading of used tires, might result in the widespread distribution of F1534C point mutation in tropical countries.
| Aedes aegpyti is one of the most important vectors of yellow fever and dengue fever. Pyrethroid insecticides are emerging as the predominant insecticides for vector control, and resistance of vector mosquitoes to pyrethroid is a major problem for the vector control program. Several mutations in the voltage-gated sodium channel were reported to play important roles in pyrethroid resistance of Aedes aegypti. Recently, a novel F1534C mutation was reported to be strongly correlated with resistance to DDT and pyrethroid. We observed a high resistance to DDT and moderate resistance to permethrin in both Ae. aegypti aegypti (Aaa) and Ae. aegypti formosus (Aaf) colonies collected in Ghana. Concurrently, high frequencies of F1534C mutations were found in the above mosquito colonies, and this was its first detection on the African continent. We found a strong linkage of F1534C mutation and the introns between exon 20 and 21 commonly found in South East Asian and South and Central American countries. The DDT and pyrethroid resistance in Ghanaian Ae. aegypti population was suggested to be caused by the introgression of Aaa genes from the above area.
| Aedes aegypti (L.) is found throughout West Africa from sea-level to at least 1,220 m in Nigeria, and from the coastal swamp zone to the northern Guinea savannas. Various types of breeding sites have been reported for this species, including crab burrows, holes in trees, fallen leaves, rock pools, anthropogenic containers, etc. Transportation and urbanization of new areas are major causes of the spread of Ae. aegypti [1].
Yellow fever is endemic in Ghana and major outbreaks, which involved 319 cases and 79 deaths, occurred in 1969–1970 in the northern part of the country. In December 2011, the Ministry of Health of Ghana declared a yellow fever outbreak. Cases were recorded in three districts located in the midwestern part of the country. A total of three laboratory-confirmed cases and seven deaths were reported [2]. Aedes aegpyti is one of the most important yellow fever vectors implicated in the Ghana outbreaks [3]. Although there have been no reports of dengue fever outbreaks in Ghana, it has been detected in the adjacent countries of Côte d’Ivoire and Burkina Faso, both of which share borders with Ghana [4]. Increasing migration of people across the borders of these countries and the absence of organized mosquito control in Ghana might lead to dengue fever transmission in Ghana in the future [4]. A recent seroprevalence survey in Ghana revealed the presence of IgM and IgG dengue antibodies in 3.2% and 21.6% of the children, respectively, with confirmed malaria. This indicated the possible co-infection of dengue fever and malaria, and previous exposure of the children to dengue virus [5]. Although no flavivirus was detected in Aedes mosquitoes from the study sites, larval densities and adult biting rates of Aedes mosquito in study areas were thought to be sufficient to promote outbreaks of dengue fevers [4].
Pyrethroid insecticides are emerging as the predominant insecticides for vector control. Pyrethroid resistance of vector mosquitoes may become a major problem for vector control programs because there are currently no substitutes for pyrethroids [6]. Although there are some alternative chemicals to pyrethroids, no chemical seems to surpass pyrethroids in the toxicological and economical point of view. The kdr-type resistance has been observed in several mosquitoes, including Anopheles gambiae Giles [7], Anopheles stephensi Liston [8], Culex quinquefasciatus Say [9], and Ae. aegypti [10]. Several mutations in segment 6 of domain II of the voltage-gated sodium channel were reported to play important roles in pyrethroid resistance of Ae. aegypti (I1011M, I1011V, V1016G, and V1016I) [10–12]. Recently, a novel F1534C mutation in segment 6 of domain III in DDT/permethrin-resistant Ae. aegypti was reported [13,14] and this point mutation was confirmed to be strongly correlated with resistance to DDT and pyrethroid [15]. The S989P mutation in domain II, which occurs in deltamethrin-resistant Ae. aegypti, is another principal kdr mutation that works synergistically with the V1016G mutation [16].
The mapping of the nation-wide distribution and the detection of insecticide resistance of vector mosquitoes in Ghana will provide beneficial information for forecasting dengue and yellow fever outbreaks and developing effective control measures. Differences in the insecticide susceptibilities associated with seasonal or regional differences in the distribution of the subspecies Ae. aegypti aegypti (Aaa) and Ae. aegypti formosus (Aaf) is also of interest. Aaf which originated from African forest area is believed to be the ancestral species of Ae. aegypti s. l. Aaa is predominantly anthropophilic and adapted to the human environment, while Aaf is more associated with a forest environment [17]. Adults of Aaa prefer an indoor environments and use artificial water containers for oviposition, while Aaf prefer an outdoor environment and the forest edge and breed in natural containers such as tree holes, rock pools and plant axils. Aaa is highly susceptible to dengue and yellow fever virus, and is considered to be a more efficient virus vector than Aaf [18].
In the present paper, insecticide-susceptibility of Ae. aegypti s. l. populations (mixed populations of Aaa and Aaf) collected from used tires located in several locations in Ghana was examined. The presence of mutations in the voltage-gated sodium channel gene, S989, I1011, L1014, and V1016 and a recently identified amino acid replacement at F1534 were examined. Possible causes of insecticide resistance in Ghanaian Ae. aegypti s. l. populations were discussed based on phylogenetic analysis.
Ethical approval for the Ghanaian field study was reviewed by Noguchi Memorial Institute for Medical Research IRB (DF22). Ethical approval for the Kenyan study was reviewed by KEMRI Ethic (CSS No. 2126).
We drove along the main roads in three cities (Accra, Kumasi, and Tamale) and 2 towns (Abuakwa/Suhum and Kintampo) in Ghana, from December 4–10, 2013 (Accra, Abuakwa/Suhum, Kumasi, Kintampo, and Tamale in the beginning of the dry season) and from September 2–10, 2014 (Accra, Abuakwa, Kumasi, and Kintampo in the late rainy season) (Fig 1). Accra is the capital city located in southern Ghana and faces the Atlantic Ocean. It experiences high humidity (monthly average in 2012 and 2013, 77.4% relative humidity, Ghana Meteorological Agency, Legon-Accra, Ghana), but relatively low precipitation (monthly average 2012–2013, 46.7 mm). Kumasi, the 2nd biggest city, is located in a tropical rainforest area with high precipitation (monthly average 2012–2013, 119.6 mm) and little sunshine (monthly average 2012–2013, 4.9 hrs/day). Abuakwa/Suhum is also located in a tropical rainforest area and experiences high precipitation (annually 1270–1650 mm). Kintampo and Tamale are located in a tropical savanna climate with relatively lower humidity (monthly average 2013–2014, 60.6% for Tamale) and high temperatures (monthly average 2012–2013, 29.1°C for Tamale).
Used tires were found primarily along the periphery of repair shops, and mosquito larvae were collected from used tires with nets and dippers. We recorded the geographical location of the collection site using a global positioning system (GPS). Mosquito collection points were plotted on a shape file map available from DIVA-GIS (http://www.diva-gis.org/gdata) using ArcGIS 10.2 (ESRI Japan Corp., Tokyo, Japan). In 2013, there were 14, 11, nine, seven, and seven collection points in Accra, Abuakwa/Suhum, Kumasi, Kintampo, and Tamale, respectively. There were eight, seven, seven, and eight collection points in Accra, Abuakwa, Kumasi, and Kintampo, respectively, during 2014. Mosquito larvae collected from separate collection points in a town or city were mixed into 1 batch and reared in dechlorinated tap water at room temperature until adult emergence.
Tests of adult susceptibility to insecticides were performed using World Health Organization (WHO) test tube kits for the field-collected Ae. aegypti colonies. Procedures were carried out according to WHO instructions (WHO/CDS/CPC/MAL/98.12). Although the WHO recommended discriminating concentration for permethrin of 0.25% and discriminating time of contact for DDT as 30 min for Ae. aegypti, we used 1 h exposure to a higher concentration of permethrin (0.75%) and longer time (1 h) exposure for DDT (4%). Mixed adult mosquito colonies (F0) consisting of Ae. aegypti aegypti (Aaa) and Ae. aegypti formosus (Aaf) emerged from the field-collected larvae were used for the insecticide susceptibility test in the collection of 2013 (Total 138 female adults). The F1 adult mosquitoes that emerged from the eggs of F0 colonies were used for the tests in 2014 (Total 600 female adults). One- to 5-day-old unfed female mosquitoes were released into WHO test tubes and were exposed to an insecticide-impregnated paper. Basically, 3 to 4 replicates using 20 female adults per a replicate were made in the test. In the test of 2013, however, 1 to 3 replicates using the smaller number of mosquitoes were done because we could not get enough number of mosquitoes. Control tests using the papers without insecticides were done in each replication. Time to knockdown was recorded. Insects were then transferred to a clean tube and fed via cotton soaked with a 5% glucose solution, and mortality was recorded after 1 day. The time required for 50% knockdown (KT50) was determined, and average mortality was calculated.
Adult Ae. aegypti specimens were observed microscopically and identified using keys by Huang [19] into two subspecies, Aaa and Aaf. Specimens with a large, median patch of pale scales on abdominal tergite 1were identified as Aaa, and those without the median patch of pale scales were identified as Aaf (Fig 2).
Direct DNA sequencing was conducted to test for the presence of point mutations at S989, I1011, L1014, V1016, and F1534 for different adult individuals from those used in the insecticide susceptibility test. One or two legs from each specimen were placed in a 1.5-mL PCR reaction tube. The sample was homogenized in a mixture of extraction solution (20 μL) plus tissue-preparation solution (5 μL) (REDExtract-N-Amp Tissue PCR Kit; Sigma, St. Louis, MO) for extraction of DNA. The solution was heated at 95°C for 3 min and neutralized with the neutralization solution. Initial amplification was carried out using the primers AaSCF1 (AGACAATGTGGATCGCTTCC) and AaSCR4 (GGACGCAATCTGGCTTGTTA) for S989P, I1011M (or V), L1014F, and V1016G (or I); or AaSCF7 (GAGAACTCGCCGATGAACTT) and AaSCR7 (GACGACGAAATCGAACAGGT) for F1534C. The PCR mixture contained 4 μL of REDExtract-N-Amp ReadyMix (Sigma), 0.5 μM of each primer, and 1 μL of the DNA template in a total volume of 10 μL. PCR was performed under the following conditions: initial denaturation at 94°C for 3 min, 35 cycles each of 94°C for 15 s, 55°C for 30 s, and 72°C for 30 s, followed by a final elongation step at 72°C for 10 min. The amplified fragments of the expected size were purified with ExoSAP-IT (USB Corporation, Cleveland, OH) at 37°C for 30 min, and then 80°C for 15 min. DNA sequencing was carried out using the primers AaSCF3 (GTGGAACTTCACCGACTTCA) and AaSCR6 (CGACTTGATCCAGTTGGAGA) for S989P, I1011M (or V), L1014F; and V1016G (or I), or AaSCR8 (TAGCTTTCAGCGGCTTCTTC) for F1534C. A BigDye Terminator v 3.1 Cycle Sequencing Kit (Applied Biosystems Japan Ltd., Tokyo, Japan) was used for DNA sequencing, according to the manufacturer’s instructions. Two micromoles of each primer were added to a tube, making total mixture volume 10 μL. PCR was performed under the following conditions: initial denaturation at 96°C for 1 min. 25 cycles each of 96°C for 10 s, 50°C for 5 s, and 60°C for 2 min. Direct DNA sequencing was performed on the 3730 DNA Analyzer (Applied Biosystems Japan Ltd.). The electropherogram of the targeted amino acid replacement was analyzed with MEGA 6.0 public domain software (http://www.megasoftware.net/). The unique DNA haplotype sequences were deposited in GenBank.
The genetic diversities in the introns between 1015V and 1016V (the sequences produced by the direct sequence described as above) located in the domain II area of the voltage-gated sodium channel in the field-collected Ae. aegypti specimens from Africa, Asia, and South and Central America, along with other genetic information for this species in GenBank were analyzed to determine the genetic affinity of Ghanaian Ae. aegypti populations (423 Aaa and 336 Aaf) to other populations. Newly determined sequences (Ghana, Malawi, Zambia, Zimbabwe, Kenya, Philippines, Singapore, Vietnam, and El Salvador) and the sequences obtained from GenBank (Brazil, India, Indonesia, and Myanmar) (S1 Table) were aligned initially using MEGA version 6 [20], and subsequently modified manually if needed. The alignment was performed for a 263 bp of fragment with gaps (total fragment lengths were from 228 to 250 bp). Thus, two datasets were prepared. In the first dataset, all indels were completely removed from the fragment (final length was 207 bp). In the second, those gaps were treated as a 5th variable in maximum parsimony analysis. A single gap was assumed to have evolved once, whereas a longer indel was assumed to have been caused by either one- or two-time events by comparing same sites of other sequences.
KT50 (time to cause 50% knockdown) was calculated using the Bliss' probit method [21]. Chi-square tests were used for the comparison of subspecies composition of Aaa and Aaf between the collection in 2013 and 2014, and the comparison of insecticide susceptibility between the two subspecies. For the two phylogenetic datasets, a total of four phylogenetic analyses were conducted: 1) maximum parsimony analysis (MP), 2) maximum likelihood analysis (ML), and 3) neighbor joining analysis (NJ) were all conducted using the 1st data, and 4) MP was also conducted using the 2nd data. Based on the model selection program of MEGA, the Tamura 3-parameter model was the evolution model used. The first three analyses were conducted by MEGA, whereas the MP tree for the 2nd dataset was constructed using PHYLIP3.69 (http://evolution.genetics.washington.edu/phylip.html). For the all constructed trees, Bootstrap replication was operated for 1,000 times to calculate how strongly the branches were supported.
Subspecies composition of Aaa and Aaf based on the identification criteria by Huang [20] is shown in Fig 3. In the first collection in November and December 2013, Aaa appeared dominant in Accra (82.7%), Kintampo (87.0%), and Tamale (79.3%), whereas the composition rate was relatively lower in Kumasi (65.2%). Conversely, in the 2nd collection performed in September 2014, the composition rates of Aaa in Accra (46.4%), Abuakwa (25.5%), and Kintampo (60.0%) were lower than those in 2013 collection, while the composition rate in Kumasi was in the same range as in 2013 (63.7%). The composition rates of Aaa were significantly lower in Accra (χ2 = 16.8, df = 1, P < 0.0001) and Kintampo (χ2 = 4.66, df = 1, P = 0.031) as compared to those in 2013, whereas no such significant change in composition was observed in Kumasi (χ2 = 0.077, df = 1, P = 0.78).
Totally, there was no difference in the susceptibility to permethrin (χ2 = 0.010, df = 1, P = 0.92 in 2013 collection; χ2 = 1.46, df = 1, P = 0.23 in 2014 collection) and DDT (χ2 = 1.26, df = 1, P = 0.26 in 2013 collection; χ2 = 0.021, df = 1, P = 0.88 in 2014 collection) between Aaa and Aaf used for the susceptibility test. The susceptibilities to the insecticides were, therefore, compared with mixed colonies of both subspecies (Fig 4). High resistance to DDT (less than 70% mortality at 1 h contact) was observed in all mosquito colonies. KT50s with DDT for these colonies were >60 min, except for Tamale (2013) and Kumasi (2013) colonies (KT50 was 56.4 and 58.2 min, respectively), indicating low knockdown ability of DDT against these colonies, as well as low killing rates. Susceptibilities to permethrin (0.75%) were relatively higher in all colonies as compared to those for DDT. Resistance to permethrin was, however, suspected in the Accra (<90% mortality) and Abuakwa colony (81.3% mortality). For the Kumasi and Kintampo colonies collected in 2014, mortalities with permethrin were higher than the other colonies, but were less than 100% (98.8% and 95.0% mortality, respectively), indicating the possible existence of resistance or tolerance to permethrin.
Total 759 specimens (262 in 2013 collection and 497 in 2014 collection) were sequenced. No mutation at S989, I1011, or L1014 was detected among 707, 756, and 734 mosquitoes sequenced, respectively. Conversely, F1534C mutations were detected at high frequency: 294 homozygous and 259 heterozygous mutations among 759 mosquitoes sequenced (accession No. LC050217, LC050218). Table 1 shows the homozygous percentages and allelic frequencies of point mutations at 1534F in Aaa mosquitoes collected from five different places in Ghana. Allelic frequencies of F1534C mutations were higher in Accra (68.4%), Kumasi (64.6%), and Kintampo (58.3%) than other places. The allelic frequencies of F1534C mutations in Aaf were also higher in Accra (52.6%), Kumasi (60.0%), and Kintampo (45.2%) than other places (Table 2). The frequency of F1534C allele and the ratio of F1534C homozygous mosquitoes in Aaa was significantly higher than that in Aaf in the mixed populations from all collection places in 2013 and 2014 (Table 3). Additionally, one heterozygote point mutation (V1016I, accession No. LC050223) was found in Accra among 732 mosquitoes (Table 4). Homozygous F1534C was concurrently found in this individual (Aaa). No V1016I mutation was found in Aaf (Table 5).
We detected two types of introns between exon 20 and 21 in the Ghanaian Ae. aegypti populations (183 specimens of Aaa and Aaf): Ghana 001 (250 bp, Accession No. LC036551) and Ghana 257 (234 bp, Accession No. LC036552). The point mutations at 1534F (F1534C) on exon 31 were found to be strongly linked with the intron of the former group (Group A in Table 6). When the two types of the intron were treated as two alleles, strong linkage disequilibrium was observed between mutation at 1534F and the two types of intron (using Genepop, G-test with 100 repeats of 10000 iteration per batch, P < 0.001). All phylogenetic trees showed similar topology (Fig 5 and S1–S5 Figs). Thus, we show one of the MP trees constructed using the 1st dataset (no indel) in Fig 5. The sequences from each geographic area were not in the same clade, and were distributed paraphyletically. Two large clades were observed: Clade 1 consisted of the sequences from southeastern Asia and the South and Central America with two from Kenya and one from Ghana, and Clade 2 consisted of the remaining African samples and strongly supported Asian or American branches. Ghana 001 (Group A in Table 6) shared the same sequence with most of other Asian and South-American sequences within Clade 1. These 2 clades were strongly supported by a consensus tree of four parsimonious trees using the 1st dataset (S2 Fig). Most African sequences were placed in Clade 2, and Asian and American sequences were distributed in three strongly supported monophyletic clades within Clade 2. When the consensus tree of MP analysis using the 1st dataset was compared to the consensus tree using the 2nd dataset (indel was treated as 5th variable), only the branch position of clade 1 was changed (S1 and S5 Figs). ML and NJ trees also showed similar topology, with a clade consisting of two Kenyan and one Ghanaian sequences and another clade consisting of the remaining sequences (S3 and S4 Figs).
The rainy season in Ghana starts in March and lasts until the end of October. The collection period in our survey was November and December 2013 and September 2014. Therefore, our collection dates corresponded to the beginning of the dry season and the late rainy season, respectively. Accordingly, the proportion of the Ae. aegypti aegypti (Aaa) collected in samples from used tires in Accra and Kintampo were higher in the dry than rainy season, although no such difference was observed in Kumasi (Fig 3). The same seasonal shift in subspecies abundance was reported for Ae. aegypti s. l. in tree holes and fruit husks in southeastern Senegal where most of the Ae. aegypti s. l. in the wet season were subspecies formosus [22]. It is interesting that there was no such seasonal difference in subspecies composition in Kumasi. This appeared attributable to the relatively consistent precipitation in Kumasi throughout the year (monthly average precipitation in 2012 and 2013 was 162.4 mm in the rainy season and 88.9 mm in the dry season; Ghana Meteorological Agency, Legon-Accra, Ghana) as compared to Accra (68.2 mm and 29.6 mm for the wet and dry season, respectively) and Kintampo (75.1 mm and 46.6 mm for the wet and dry season, respectively, in adjacent Tamale).
Trpis and Hausermann collected larvae of Ae. aegypti s. l. in three principal habitats (domestic, peridomestic, and feral) in the Rabai area in eastern Kenya [23]. The mosquitoes from the domestic habitat were represented by the domestic form, Aaa, and the feral mosquitoes from tree holes were represented by the feral subspecies, Aaf. A hybridization experiment showed that house-entering behavior was genetic, and the percentage entering houses was highest in the domestic Aaa populations and lowest in the feral Aaf populations. The authors suggested that the populations from the peridomestic habitat may represent hybrids between the domestic and feral forms. The involvement of a gene expression related the odorant receptor of human-specific odor component (sulcatone) was also suggested to explain the behavioral difference between the two subspecies [24]. Sylla et al. found that both Aaa and Aaf may survive the tropical dry season in natural habitats, such as tree holes and husks, in West Africa [22]. Our finding, that both of the subspecies were found in an artificial habitat (used tires) provides another contrasting trend to that of previous studies in East Africa that reported household containers were the exclusive larval habitat for Aaa and tree holes the predominant habitat for Aaf [17, 23].
Source reduction and use of insecticides, such as organophosphates, carbamates, and pyrethroids, were recommended by the WHO as preventive control measures for the vector mosquitoes of yellow fever. Use of organochlorine compounds, however, is not recommended because of widespread resistance of Ae. aegypti to these compounds in the 1980s [25]. Thermal fogging, mist blowers, Ultra low volume (ULV) spray, and indoor residual spraying (IRS) with the above insecticides (organophosphates, carbamates, and pyrethroids) were recommended by the WHO as emergency control measures for Ae. aegypti. Resistance of Ae. aegypti to hexachlorocyclohexane (HCH) in Navrongo, Kassena and dieldrin resistance in upper region of Lawra were reported in Ghana in 1971 [25]. The present study is perhaps the first to report the resistance of Ae. aegypti s. l. to DDT and permethrin in Ghana, although DDT resistance in Ae. aegypti was common in countries adjacent to Ghana, such as Côte d'Ivoire (1968), Togo (1969), and Benin (1968) [25]. High resistance to DDT seems to be widespread and resistance to pyrethroids is also suspected to be common in Ghana. Although the concentration of DDT and pyrethroids used and how they were applied for the control of Ae. aegypti in Ghana is unknown, it is clear that these insecticides were one of the causative factors in the resistance of Ae. aegypti s. l. populations in Ghana. Organochlorine pesticides were most popular and extensively used by farmers in Ghana with lindane commonly used for pest control on cocoa, vegetables, and maize, and endosulfan on cotton, vegetables, and coffee. DDT and lindane were once employed to control ectoparasites of farm animals and pets in Ghana [26]. Lambda-cyhalothrin and cypermethrin are used by vegetable growers on tomato, pepper, okra, eggplant, cabbage, and lettuce farms [26]. The contamination caused by the aforementioned pesticides to the breeding area might have served as selection pressure for the development of the resistance in Ae. aegypti populations.
Additionally, indirect effects of long lasting insecticidal nets (LLINs), IRS, and other insecticide treatment for malaria control have contributed to the development of DDT and pyrethroid resistance as previously reported in Ae. aegypti populations in Vietnam [27–30]. The malaria vector Anopheles gambiae s. l. in southwestern Ghana has developed a high resistance to DDT and pyrethroid insecticides in an area where the species was susceptible to these chemicals just a decade ago [31, 32]. The use of insecticides, such as LLINs and IRS, in the Ghana National Malaria Control Program is believed to be the major cause for the cross-resistance between DDT and pyrethroids. This was mainly attributable to the kdr gene [31] as reported in the adjacent countries of Mali [33] and Burkina Faso [34]. Pyrethroid treatment for malaria vector control appears to have been intensively conducted in the interior and along the periphery of human habitation areas, where the breeding and resting sites of Ae. aegypti are located. This likely contributed to the strong selection pressure toward Ae. aegypti (especially Aaa) because this species is domestic and endophagic. Extensive use of DDT for malaria control before it was banned may have also contributed to the development of pyrethroid resistance in Ae. aegypti because the target site (i.e., the voltage-gated sodium channel) is common to both DDT and pyrethroids.
F1534C mutations were reported worldwide (i.e., South Asian, South East Asian, South and Central American countries and Macaronesian islands). After the first description of the F1534C point mutation in Ae. aegypti collected in Thailand [13, 14], the same mutation was reported in succession in Vietnam [27], Grand Cayman Island [15], Madeira Island [35], Brazil [36], Myanmar [37], Venezuela [38], India [39], and Malaysia [40]. The mutations at 1016V were also reported worldwide. Two different types of the mutation at the same locus are distributed independently. Valine to glycine replacements (V1016G) are commonly distributed in South East Asia [37, 40–43], whereas valine to isoleucine replacements (V1016I) are common in South and Central America [11, 36, 38, 44–46].
The present study provides the first description of F1534C and V1016I mutations found in African Ae. aegypti s. l. populations. Accra and Kumasi are the two largest cities in Ghana, each home to one to two million people. Tamale is the 3rd largest city with a population of approximately 400,000 people. Kintampo and Abuakwa, both of which have populations of approximately 40,000, are much less populated compared to the other sampled cities. Allelic frequencies of F1534C and percentage of homozygous individuals of the same point mutation were higher in the two large cities, Accra and Kumasi, than the other collection locations. Interestingly, both the allelic frequency and homozygous percentage of F1534C in Aaa was significantly higher than that in Aaf, though we could not observe the significant difference in the susceptibility to DDT and permethrin between the two subspecies. Above discrepancy might suggest that the F1534C mutation is not a single resistance mechanism but is combined with the other unknown mechanisms such as metabolic factors etc. Recently, some reports called attention to the role of glutathione-S-transferases (GST) in the cross resistance between DDT and pyrethroids in mosquitoes [47, 48]. Riveron et al. demonstrated that the single amino acid change in GST gene (L119F) confers high level of metabolic resistance to DDT in Anopheles funestus [47]. The authors also showed that this mutation strongly related to the metabolism of permethrin. Several Epsilon GST genes were reported to play a role in pyrethroid resistance in Ae. aegypti [48]. The above new findings, as well as the other metabolic factors, should be taken into consideration for further study.
Martins et al. reported two types of haplotype group A (250 pb) and B (234 pb) in the intron between exons 20 and 21 on domain II of the voltage-gated sodium channel with pronounced differences in both sequence and size in Brazilian Ae. aegypti [49]. The introns in the sequence of accession No. FJ479611 referred in S1 Table and Fig 5 correspond to the haplotype group A and those of FJ479609, FJ479610, and FJ479613 correspond to the haplotype group B. The authors also noted point mutations at 1011I (I1011M) on exon 20 appeared in half of the group A sequences, whereas no such mutation occurred in group B sequences. The same kind of the evidence of linkage equilibrium was reported by Saavedra-Rodriguez et al. The authors found the same intron as reported above (group A) strongly linked with V1016I mutation and hypothesized that a genetic sweep of the V1016I allele and its proximate intron sequences has occurred through DDT and subsequent pyrethroid selection [11]. In the present study, we could detect the same two types of intron in the Ghanaian Ae. aegypti populations: Ghana 001 for group A and Ghana 257 for group B. Interestingly, the point mutations at 1534F (F1534C) on exon 31 were found to be strongly linked with the intron of group A. Furthermore, phylogenetic analysis using this intron in the present study clearly showed that the two Ghanaian haplotypes belonged to two haplotype groups (Clades 1 and 2). Given Ae. aegypti was originally from Africa, Clade 2 is thought to be the ancestral clade of Clade 1 because Clade 2 contained most African haplotypes. Interestingly, one of the two Ghanaian haplotypes (Ghana 001) and two Kenyan haplotypes were placed in Clade 1, apparently suggesting those haplotypes were introduced from other continents, such as Asia or South or Central America.
Aedes aegypti is thought to have originated on the African continent. The sub-Saharan part of the continent still contains Aaf, which is believed to be the ancestral species of Ae. aegypti s. l. The subspecies that has been domesticated and adapted to anthropomorphic environments (Aaa) expanded its habitat around human domiciles and has been dispersed by human movement. Aedes aegypti aegypti spread to the western hemisphere in the 17th centuries, to the Mediterranean coastal area in the 18th centuries, and to tropical Asian and Pacific islands in the 19th to 20th centuries. This subspecies was eradicated from the Mediterranean area in 1950s and from south America from 1950 to 1960. However, it has reinfested most of the countries from which it was eradicated [50]. The Ae. aegypti eradication program, initiated by the Pan American Health Organization (PAHO) in the 1940s and 1950s to prevent urban epidemics of yellow fever, was successful in most of the countries in South and Central America, resulting in a dramatic decrease in the distribution of mosquito populations. However, the discontinuation of the eradication program in 1970s led the reinfestation of the mosquitoes and Ae. aegypti regained a similar distribution to that of the 1940s by 1995 [51]. The worldwide eradication programs, presumably with organochlorine insecticides, in 1940s and 1950s might have caused high selection pressure on the mosquito populations and expanded the distribution of insecticide-resistant Ae. aegypti populations [25]. Selection of the F1534C point mutation could be hypothesized to have taken place during this period. DDT resistance in Ae. aegypti was first reported in the Caribbean countries in the 1950s and the resistance persists in almost all regions that had achieved Ae. aegypti eradication, despite the fact that DDT is no longer used. DDT resistance, as well as the F1534C point mutation might have been maintained in the Ae. aegypti populations by selection pressures from pyrethroid insecticides, such as permethrin, as both insecticides have the same target site [15].
Used and discarded tires are one of the most important breeding sites for Ae. aegypti. They provide a habitat for the larvae and are capable of supporting larval development immediately after they are discarded. Accumulation of microorganisms in time improves the breeding environment [52]. It is noteworthy that there has been worldwide focus on the dispersal of containers breeding mosquitoes in the used tires for the past three decades [53]. The mosquito species that has played the leading part in the above event are Ae. albopictus (Skuse), which together with the other four Japanese mosquito species was thought to have arrived at the western coastal ports of the United States in used tires by 1983 and were widely distributed in the United States and Brazil by 1986 [54]. Since 1986, tire shipments infested with Ae. albopictus have been found in the South and Central America, South and West Africa, Oceania, and European countries [54]. Used tire trading is worldwide with complicated commercial networks, including those in African countries, such as South Africa, Kenya, Uganda, Niger, Nigeria, and Ghana [53, 54]. The history of the worldwide invasion by Ae. aegypti in association with human activity might be longer than that of Ae. albopictus.
The worldwide transportation of vector mosquitoes in accordance with human activity such as trading of used tires, might result in the introgression of Asian or Central or South American type haplotype into Ghanaian Ae. aegypti population. It is not known whether Ghanaian F1534C mutations "hitched a ride" with the above haplotypes or they were selected convergently after the above introgression. However, strong linkage disequilibrium between F1534C mutation and intron haplotypes may support introgression of the mutation. Phylogeographic analyses with other loci (e.g. mtDNA, ITS) and detailed population genetic analyses on intron (e.g. mismatch analysis) could provide more evidence supporting the hypothesis. Discovery of the V1016I mutation, although it was only 1 heterozygote, and co-occurrence of this mutation with F1534C might be noteworthy. Recently, Linss et al. detected the widespread co-occurrence of V1016I and F1534C point mutations in Ae. aegypti populations in Brazil [36]. The same co-occurrence of the two point mutations were reported in Grand Cayman Island [15]. The high frequency of F1534C and the co-occurrence of V1016I with this mutation, therefore, might explain one of the possible history of introduction of these mutations into the Ghanaian Ae. aegypti population from South and Central America. The importance of the sequential evolution of F1534C and V1016I was advocated in Mexican Ae. aegypti population [55]. V1016I mutation was unlikely to have evolved independently because of low fitness, while F1534C mutation evolved first but conferred only a low level resistance. V1016I mutation then rapidly evolved from 1016V/F1534C haplotype under the high pressure of pyrethroids since these double mutants confer higher pyrethroid resistance. The authors suggested that knowledge of the frequencies of mutations in both S6 in domains II and III are important to predict the potential of a population to evolve kdr. They also sounded an alarm that susceptible populations with high 1016V/F1534C frequencies are at high risk for kdr evolution [55]. We, therefore, have to pay great attention to the genomic status in Ghanaian Ae. aegypti populations for predicting the evolution of pyrethroid resistance.
|
10.1371/journal.pgen.1001303 | Genome-Wide Transcript Profiling of Endosperm without Paternal Contribution Identifies Parent-of-Origin–Dependent Regulation of AGAMOUS-LIKE36 | Seed development in angiosperms is dependent on the interplay among different transcriptional programs operating in the embryo, the endosperm, and the maternally-derived seed coat. In angiosperms, the embryo and the endosperm are products of double fertilization during which the two pollen sperm cells fuse with the egg cell and the central cell of the female gametophyte. In Arabidopsis, analyses of mutants in the cell-cycle regulator CYCLIN DEPENDENT KINASE A;1 (CKDA;1) have revealed the importance of a paternal genome for the effective development of the endosperm and ultimately the seed. Here we have exploited cdka;1 fertilization as a novel tool for the identification of seed regulators and factors involved in parent-of-origin–specific regulation during seed development. We have generated genome-wide transcription profiles of cdka;1 fertilized seeds and identified approximately 600 genes that are downregulated in the absence of a paternal genome. Among those, AGAMOUS-LIKE (AGL) genes encoding Type-I MADS-box transcription factors were significantly overrepresented. Here, AGL36 was chosen for an in-depth study and shown to be imprinted. We demonstrate that AGL36 parent-of-origin–dependent expression is controlled by the activity of METHYLTRANSFERASE1 (MET1) maintenance DNA methyltransferase and DEMETER (DME) DNA glycosylase. Interestingly, our data also show that the active maternal allele of AGL36 is regulated throughout endosperm development by components of the FIS Polycomb Repressive Complex 2 (PRC2), revealing a new type of dual epigenetic regulation in seeds.
| Seeds of flowering plants consist of three different organisms that develop in parallel. In contrast to animals, a double fertilization event takes place in plants, producing two fertilization products, the embryo and the endosperm. Imprinting, the parent-of-origin–specific expression of genes, typically takes place in the mammalian placenta and in the plant endosperm. A prevailing hypothesis predicts that a parental tug-of-war on the allocation of available recourses to the developing progeny has led to the evolution of imprinting systems where genes expressed from the mother dampen growth whereas genes expressed from the father are growth enhancers. The number of imprinted genes identified in plants is low compared to mammals, and this precludes the elucidation of the epigenetic mechanisms responsible for this specialized expression system. Here, we have used genome-wide transcript profiling of endosperm without paternal contribution to identify seed regulators and, among these, imprinted genes. We identified a cluster of downregulated MADS-box transcription factors, including AGL36, that was subsequently shown to be imprinted by an epigenetic mechanism involving the DNA methylase MET1 and the glycosylase DME. In addition, the expression of the active AGL36 allele was dampened by the FIS Polycomb Repressive Complex, identifying a novel mode of regulation of imprinted genes.
| Seed development is a tightly regulated process that is controlled, both before and after fertilization and requires tight coordination of parental gene expression [1]. A paradigm for the importance of balanced parental contribution is the observation that certain genes in the developing offspring of flowering plants are exclusively or preferentially expressed from only one of the two parental genomes, a phenomenon called genomic imprinting that has also been observed in mammals [2], [3]. The relevance of parent-of-origin effects was first found in interploidy crosses [4]. Typically, an increase in the paternal genome results in larger seeds, while the opposite is observed if the maternal gene dosage is higher than normal [5]. This is in agreement with the parental conflict theory, which implies that fathers direct maximal amount of maternal resources to their own offspring and thereby promote growth. Mothers on the other hand would seek to distribute the resources equally among all their offspring, and balance their resource between themselves and their offspring. Thus, maternal factors are thought to dampen growth [6].
In mammals, imprinted genes are often involved in growth control [7]–[10]. In Arabidopsis, the endosperm is the major tissue regulating the flow of nutrients to the embryo, and is therefore a likely site for parent-of-origin dependent gene expression.
Imprinting results from differences in epigenetic marks, involving DNA methylation and post-translational modifications of histones on the parental alleles [11], [12]. Trimethylation of lysine 27 on histone H3 (H3K27me3) leading to repression of gene expression, has been found to be a particularly important imprinting mechanism in plants. In Arabidopsis seeds, H3K27me3 mark is set by the FIS Polycomb Repressive Complex 2 (PRC2), which consists of at least four components; the histone methyltransferase MEDEA (MEA), FERTILIZATION INDEPENDENT SEED 2 (FIS2), FERTILIZATION INDEPENDENT ENDOSPERM (FIE), and MULTICOPY SUPPRESSOR OF IRA 1 (MSI1). The corresponding genes were identified in screens for autonomous endosperm development, indicating that the FIS complex acts as a repressor of endosperm development prior to fertilization [13]–[17].
An equally important regulatory mechanism in imprinting is DNA methylation resulting from the activity of several different methyltransferase enzymes, where each has specificity for cytosine (C) in certain sequence contexts. So far, imprinting has been shown to be under the influence of MET1, the major Arabidopsis maintenance DNA methyltransferase involved in CG-methylation [11], [18]–[20]. DNA demethylation can be achieved either by a passive process i.e. the repression of MET1 expression [21], [22], or by an active mechanism involving DNA glycosylase enzymes such as DME [23]. Several lines of evidence show that DME, which is expressed in the central cell of the female gametophyte, is necessary for maternal-specific gene expression in the endosperm [11], [18], [19], [24].
So far, only about a dozen genes in Arabidopsis have been identified to have parental-specific gene expression, and they illustrate different modes of imprinting [3]. MEA, ARABIDOPSIS FORMIN HOMOLOGUE 5 (AtFH5) and PHERES 1 (PHE1) are imprinted by the action of FIS PRC2, where only the latter is paternally expressed [13], [25]–[31]. FIS2, FLOWERING WAGENINGEN (FWA) and MATERNALLY EXPRESSED PAB C-TERMINAL (MPC) are all maternally expressed and regulated by the dual action of MET1 and DME [11], [19], [24], [32]–[34]. Recently, five novel imprinted genes, HOMEODOMAIN GLABROUS 3 (HDG3), HDG8, HDG9, At5g62110 and ATMYB3R2 were identified by differential DNA methylation in embryo and endosperm [35].
In comparison to Arabidopsis, more than 100 genes have been shown to have a uniparental or preferential parental expression pattern in mammals [36]–[39]. This suggests that additional genes in Arabidopsis are imprinted. Furthermore, the low number of known imprinted genes in plants precludes the identification of general principles in this kind of gene expression control and thus, the identification of further imprinted genes is pivotal. Moreover, the targets of imprinted genes, as well as genomic pathways and regulatory modules influenced by imprinted genes are largely unknown.
Here, we have designed a microarray strategy for the identification of seed regulators by exploiting the cdka;1 mutation. Using this approach, we have identified a cluster of previously uncharacterized AGAMOUS-LIKE (AGL) Type-I MADS-box transcription factors that are downregulated in endosperm with no paternal contribution. Here, we report that AGL36 is imprinted by the dual action of MET1 and DME. In addition, AGL36 is regulated throughout endosperm development in its maternal expression cycle by the Polycomb FIS-complex, thereby identifying a novel mode of regulation for imprinted genes.
Here we have used cdka;1 as a tool to identify factors sensitive to the vital parental gene balance in the endosperm. In heterozygous cdka;1 mutants, the second pollen mitosis is either missing or is severely delayed. However, mutant pollen can successfully fertilize the egg cell while leaving the central cell unfertilized [40], [41]. A detailed analysis by Aw and colleagues has revealed that a second sperm cell is delivered to the central cell, but that karyogamy does not take place [42]. Although not properly fertilized, the majority of the central cells in cdka;1 fertilized ovules (70–90%) are triggered to initiate endosperm proliferation [40], [42], [43]. Thus, fertilization by cdka;1 sperm cells creates a unique situation where endosperm initially develops without any paternal contribution (in the following also referred to as cdka;1P). The endosperm, however remains under-developed, and ultimately the seed aborts, further demonstrating the importance of the paternal contribution to the endosperm for proper seed development. Since activation of maternal alleles by loss of maternal FIS PRC2 could rescue seed lethality [43], we hypothesized that the disturbance of parental gene balance in the endosperm is the main cause leading to developmental arrest of cdka;1P at 3–4 days after pollination (DAP).
To identify factors and mechanisms sensitive to such an imbalance in gene dosage in the endosperm and with that likely key regulators of seed development, we performed microarray transcript profiling of cdka;1 fertilized seeds at 3 DAP (Figure S1A). Due to the heterozygous nature of the cdka;1 mutant line used, a transcript that is absent in cdka;1p seeds will lead to a reduction of maximal 50% in the genome profiling experiment. For example, genes that are only expressed from the paternal genome would show such reduced expression levels (Figure S1B). Likewise, maternally expressed genes that require activation by a paternally expressed gene(s) would be downregulated (Figure S1C), whereas genes that are acted upon by paternally expressed repressors were expected to be upregulated in the microarray screen (Figure S1D).
When we compared the transcriptional profiles of Ler x cdka;1 versus Ler x Col seeds 3 DAP, we detected 17223 nuclear genes that were expressed in all biological replicates of both mutant (cdka;1 set) and wild-type (WT set) seed profiles. Our result is in good agreement with a set of genes identified by Goldberg & Harada laboratories (GH) in globular stage seeds of Arabidopsis Ws-0 plants as 68% of our genes were also identified by GH, and our gene set included >90% of the GH globular seed gene set (Figure 1A; http://seedgenenetwork.net, [44]).
To further validate the quality of our dataset, we examined the expression pattern of genes known to be preferentially expressed from the paternal allele. To date, only three genes have been identified that show a predominant paternal expression pattern; PHE1, HDG3 and At5g62110, where all three genes were found to be downregulated in our arrays (Figure S1E), supporting our working hypothesis that paternally expressed genes can be detected amongst downregulated genes. In addition, out of seven imprinted maternally expressed genes present in our microarray sets, four were also detected as downregulated (Figure S1E). This could reflect required activation by paternal factors (Figure S1C), or be a result of more complex deregulation in response to change in gene dosage. To exclude array artifacts we tested all down-regulated genes by means of real-time PCR and could confirm their deregulation (Figure 1B).
Due to the background noise in the microarray experiment, modest but reproducible downregulation of arithmetic ratios (ar) ranging from 0.5 to 1.0 will produce False Discovery Rates (FDR, see materials and methods) with insignificant q values. Since the absence of paternally expressed genes was the simplest hypothesis to account for downregulation, we defined a functional limit for screening purposes that allowed us to detect two out of three known paternally expressed genes in the array. Both PHE1 and HDG3 are detected at q values of 0.35 and a downregulation cutoff of 0.8 (ar). Consequently these values were chosen and used to filter the microarray data.
Using these criteria, a set of 602 genes was extracted (q≤0.35 and ar ≤0.8), subsequently called Down 0.8. For upregulation, we worked with two gene sets. For the first set, Up 1.2, we used parameters equivalent to the downregulated set (q≤0.35 and ar ≥1.2), which resulted in a set of 1030 genes. For the second set, Up 1.5, resulting in 323 genes, we chose ar ≥1.5, a threshold for deregulation commonly used in genome-wide expression studies (Table S3).
To test whether the deregulated genes could preferentially be attributed to a certain seed structure, we compared our data to gene sets expressed in different seed regions and compartments of globular stage seeds using data generated by Goldberg & Harada (GH) laboratories available at http://seedgenenetwork.net [44]. The overlap between the upregulated gene sets and the GH embryo, seed coat and endosperm was significantly lower than expected for independent sets of genes, indicating that among the upregulated genes we preferentially find those that are below the detection limit of the GH analyses. However looking at the downregulated genes, the picture was different. While we found slightly less overlap than expected by chance for the GH embryo set, the overlap was clearly larger than expected by chance for GH seed-coat (1.2<2.7e−07) and even more significant for the GH endosperm (rf = 1.3, p<2.0e−13, Figure S2A, S2B).
In order to functionally classify the deregulated gene sets according to their molecular functions we used the GO Slim classification system (Figure 1C). Only for the GO Slim term “Transcription factor activity” we find a higher percentage and significant over-representation of both up- and down-regulated groups when compared to all genes on the array/all genes expressed. Since key regulators of seed development are likely to be transcription factors (TF), we analyzed this class in detail.
When comparing the fraction of deregulated genes among the different TF families, the Mγ MADS-box transcription factors clearly stood out with more than 60% of the seed expressed members being downregulated in Ler x cdka;1 arrays (Figure S3A, S3B). We therefore focused on this MADS Type-I class for further analysis. Searches in publically available expression databases (www.genevestigator.com, Figure S4) revealed that all identified genes were exclusively expressed in the seed and predominantly in the endosperm. From the identified Type-I Mγ MADS-box genes, we selected AGL36 for further in depth analysis (Figure S4). AGL36 was the previously undescribed Mγ candidate that interacted with the highest number of described AGLs in a Y2H screen performed by de Folter et al [45]. Both AGL36 and PHE1 have been shown to interact with AGL62, which plays a major role in endosperm development [45], [46]. Within the Mγ class, AGL36 clusters together with AGL34 and AGL90 [47], which are both also detected as downregulated in our microarray experiment (Figure S4). AGL36 shares 85.7% and 84% nucleotide identity with AGL34 and AGL90, respectively (Figure S8). On the amino acid level this results in of 80.2% similarity of AGL36 with AGL34 and 83.9% similarity with AGL90.
Real-time PCR measurement of AGL36 relative expression level three days after pollination (3 DAP) in Ler ovules fertilized with either Col or cdka;1 pollen confirmed that AGL36 expression was reduced in cdka;1 fertilized seeds, (27% when normalized towards ACT11, and 36% when normalized towards GAPA) compared to wild-type seeds (Figure 2A).
To determine whether AGL36 has parental-specific expression, we took advantage of an AGL36 Single Nucleotide Polymorphism (SNP) existing between the Col and Ler ecotypes. This SNP allows the PCR product of Col cDNA to be digested by AlwNI, leaving the Ler cDNA PCR product intact (Figure 2B). We performed reciprocal crosses between Col and Ler ecotypes, and analyzed the digested RT-PCR fragments on an Agilent Bioanalyzer Lab-on-a-Chip, allowing accurate measurement of fragment sizes and their concentrations. When Colmaternal is crossed with Lerpaternal, we only detected the Col bands (165 bp+234 bp) after AlwNI digestion, indicating only maternal expression (Figure 2C). Similarly, in the reciprocal cross when Lermaternal is fertilized with Colpaternal pollen, the cDNA PCR digest resulted only in an undigested band (399 bp) originating from Ler, indicative of maternal expression (Figure 2C). This testified that AGL36 was only expressed from the maternal genome after fertilization and thus identified as a novel imprinted gene.
AGL36 expression level in wild-type seeds (Ler x Col) at different stages of seed development was monitored over a period of 12 days after pollination. Initially, a low expression level was detected (1 DAP), followed by a rapid increase and subsequent peak in AGL36 expression at 4 DAP, when the embryo is at the late globular stage of development, before declining (Figure 3A). At the embryo heart stage, corresponding to 6 DAP, AGL36 expression had decreased to similar levels as 1 DAP. To address whether AGL36 imprinting is maintained throughout its expression cycle, we performed a SNP analysis of the RT-PCR product obtained from Ler x Col crosses harvested during 1 to 12 DAP (Figure 3B). We found that AGL36 expression is originating from the maternal genome (Ler) throughout the experiment. By plotting the molarities of the maternal band obtained by Agilent Bioanalyzer, an expression profile closely identical to the pattern obtained in the real-time PCR analysis was found (Figure 3C).
To rule out that the observed maternal expression is due to expression of AGL36 in the ovule integument, which is a maternal tissue, we generated a reporter construct consisting of 1752 bp of the AGL36 promoter region fused to a GUS reporter (pAGL36::GUS) (Figure 4A). Single-copy lines carrying this construct were used in reciprocal crosses with wild-type Ler and Col plants to examine GUS expression at 3 and 6 DAP. When inherited maternally, pAGL36::GUS expression in the seed was indeed found to be restricted only to the fertilization product (Figure 4B, Figure S7D). In the reciprocal cross, when pAGL36::GUS was inherited from the paternal genome, no GUS expression was detected, (Figure 4C, Figure S7E). Consistent with the SNP analysis, this demonstrated that AGL36 was imprinted and only maternally active throughout its expression cycle. Furthermore, the 1.7 Kb promoter fragment used in this analysis appears to be sufficient to confer parent-of-origin specific expression of the reporter.
To further investigate the biological function of AGL36, we screened the Koncz T-DNA collection for insertions [48]. We identified a mutant line, agl36-1, harboring a single T-DNA insertion 16 bp upstream of the AGL36 ATG start codon (Figure S5A). The agl36-1 line showed Mendelian segregation of the T-DNA insertion, as 75% of the plants were resistant to Hygromycin (N = 1025, χ2 = 0,83, Table S1).
To test the transmission through the male and female gametes directly, reciprocal crosses of both hemizygous and homozygous agl36-1 mutant plants with wild-type plants were performed (Table S1). In a reciprocal cross, a hemizygous mutant will segregate 50% of the T-DNA resistance marker if the disrupted gene is not vital for gametophyte transmission or function. Thus, gametophyte requirement can be scored directly as reduced frequency of resistant plants [49]. In reciprocal crosses with agl36-1, no transmission distortion through female or male gametophytes could be observed (N = 661, χ2 = 0,13 and N = 1015, χ2 = 0,00 respectively, Table S1).
The position of the T-DNA insertion in agl36-1 predicts AGL36 expression failure, and indeed real-time PCR analyses of 3 DAP seeds of homozygous agl36-1−/− plants compared to Col wild-type indicate a 1000-fold AGL36 downregulation in the mutant seeds (Figure S5B). In line with an imprinted and maternal-only expression of AGL36, close to 50% reduction of the transcript level was observed in 3 DAP hemizygous agl36-1+/− seeds (Figure S5B). We thereby concluded that agl36-1 represents a loss-of-function allele of AGL36.
Although depletion of AGL36 did not interfere with the fitness of the mutant allele in our experimental system, we have shown that AGL36 is specifically expressed from the maternal allele in the fertilization product, in a time frame between 2 and 6 DAP. To investigate whether this was reflected morphologically or developmentally in the developing seed, we compared embryo and endosperm development in wild-type and homozygous agl36-1−/− seeds within the AGL36 expression time frame.
After fertilization of the egg and the central cell, the endosperm in Arabidopsis undergoes three syncytial rounds of nuclear divisions before the first asymmetric division of the zygote that creates the apical embryo proper and the basal suspensor that connects the embryo proper and the maternal tissue (Figure S5C). At the 2 DAP stage, no obvious difference could be observed between wild-type and agl36-1−/− seeds, both typically harboring a 1–2 cell embryo proper and a 16–32 nucleated endosperm (Figure S5C, left section). The embryo continues to divide through radial, longitudinal and transverse divisions to produce the so-called globular stage at 4 DAP (Figure S5C, middle section). The endosperm also undergoes 3–4 syncytial nuclear divisions and remains uncellularized as cell proliferation at the upper half of the embryo forms the cotyledon primordia at the so-called heart stage at 6 DAP (Figure S5C, right section). Although the main AGL36 expression peak occurs during this time frame, no obvious deviation between wild-type and agl36-1−/− could be observed at these stages. Similarly, using an endosperm specific pFIS2::GUS reporter [33], a wild-type endosperm division pattern was observed in agl36-1+/− seeds (Figure S5D).
The majority of imprinted, maternally expressed genes identified in Arabidopsis so far have been shown to be paternally silenced by mechanisms involving symmetric CG methylation, maintained by MET1 [11], [18], [19]. Although not directly linked to imprinting, methylation can also be directed by CHROMOMETHYLASE 3 (CMT3) that has specificity for CNG, and members of the DOMAINS REARRANGED METHYLTRANSFERASE (DRM) family; DRM1 and DRM2, that are mainly responsible for asymmetric CHH methylation [50]. In order to address the involvement of DNA methylation in the regulation of paternal AGL36 expression, we performed SNP analyses of 3 DAP ovules from reciprocal crosses with mutants that have been shown to be involved in DNA methylation. In the SNP RT-PCR analysis of mutant pollen crossed to wild-type, paternal AGL36 expression is expected if the tested mutants are involved in AGL36 imprinting.
CMT3 DNA methylation has been reported to be guided to specific sites by KRYPTONITE (KYP) H3K9 methylation [51]. When mutant cmt3-7 and kyp-2 pollen were crossed to Col wild-type plants, no difference in AGL36 expression was observed (Figure 5A). In the reciprocal cross with cmt3-7 also no difference could be detected compared to wild-type expression (Figure S6).
DRM1 and DRM2 are mainly responsible for asymmetric DNA CHH methylation [50] and rely on small interfering RNAs, processed by ARGONAUTE4 (AGO4), for target template guidance [52]. In our assays, fertilization by pollen lacking DRM1;DRM2 and pollen lacking AGO4 had no effect on the AGL36 expression pattern (Figure 5A). Likewise, AGL36 expression in the reciprocal cross was identical to wild-type (Figure S6).
DECREASE IN DNA METHYLATION1 (DDM1) is involved in maintenance of DNA methylation [53]. In our SNP RT-PCR analyses where mutant ddm1-2 pollen was used to fertilize wild-type ovules, paternal AGL36 expression was not activated (Figure 5A). In summary, CMT3, KYP, DRM1;DRM2, AGO4 and DDM1 appear not to be involved in the establishment nor maintenance of AGL36 imprinting (Figure 5A, Figure S6).
However, paternal AGL36 expression was detected when plants hemizygous for the met1-4 mutation were used as pollen donor in crosses with wild-type Ler (Figure 5B). In the reciprocal cross, using met1+/− as the maternal partner, no AGL36 expression from the paternal genome could be observed (Figure 5B). Furthermore, we performed crosses using pollen from homozygous met1-4 parents. When first generation homozygous met1 plants were used as pollen donor on wild-type plants, prominent AGL36 expression from the paternal Col genome could be observed (Figure 5B). This strongly suggests that the repression of the paternal copy of AGL36 is lifted due to the met1-4 mutation, and that MET1 is required for maintaining paternal inactivation of AGL36. In the reciprocal crosses, only expression from the maternal genome could be detected, both in the heterozygous and the homozygous met1-4 situation, further substantiating the requirement of MET1 in the male germline in order to maintain AGL36 imprinting (Figure 5B). Maternal AGL36 expression levels using homozygous met1-4 as the maternal cross partner appeared to be equal to maternal levels in the reciprocal crosses (Figure 5B). This opens for the interpretation that DNA methylation is not required for the regulation of maternal AGL36 expression.
In public expression databases, AGL36 is reported to be expressed in the seed and more precisely in the endosperm [54] (Figure S4). In order to monitor AGL36 expression in vegetative tissues and its dependence on DNA methylation, we performed a real-time PCR experiment on vegetative tissues from reciprocal Ler x Col crosses and homozygous met1-4 tissues. In biological replicates of progenies from both reciprocal crosses, weak AGL36 expression ranging from 1–6% of the seed expression level could be detected in seedlings, leaves and flowers (Figure 6A). This showed that AGL36 was expressed throughout the plant life cycle, although at very low levels. In the same experiment, we monitored expression in met1-4 tissues. AGL36 expression levels were 50–90-fold higher in met1-4 leaves compared to seed expression levels (Figure 6A). In a direct comparison, expression levels were elevated 2000-fold in homozygous met1-4 leaves compared to wild-type Col x Ler leaves (Figure 6B). In flowers, the upregulation was more than 20-fold in met1-4 compared to wild-type Col x Ler flowers (Figure 6C). In conclusion, these data showed that silencing of AGL36 in vegetative tissues involves MET1, suggesting that the absence of maintenance DNA methylation elevates vegetative AGL36 expression beyond the maternal expression levels found in seeds.
In order to investigate the parental expression pattern of AGL36 in vegetative tissues, we performed SNP analyses of flowers from F1 hybrids of Ler and Col reciprocal crosses. In both reciprocal crosses, AGL36 appeared to be expressed equally from the parental Ler and Col genomes, indicating biparental expression in flowers (Figure 6D). This indicates that parental-specific expression, i.e. imprinting of AGL36, as expected, only takes place in the seed and that a low basal biparental expression is present throughout the plant life cycle. Interestingly, biallelic expression in flowers suggests that further silencing of AGL36 takes place in the male germline before uniparental expression in the seed (Figure 6D).
According to our data, the action of MET1 suppresses AGL36 expression throughout the vegetative phase and this suppression is maintained in the fertilization product through the male germline. AGL36 imprinting thus requires specific activation of the maternal allele. DNA demethylation by DME has previously been shown to mediate maternal-specific gene expression in the endosperm [11], [18], [19], [24], and we therefore investigated AGL36 expression in dme-6 mutant plants. Since dme cannot be maintained in a homozygous state, we harvested siliques of dme-6+/− heterozygous plants pollinated with Col pollen at 3 and 6 DAP. We monitored the relative expression by means of real-time PCR using FWA and FIS2 as controls. At 3 DAP, both controls were downregulated by 69±0.09% and 53±0.30% respectively (Figure 6E), in line with a lack of functional DME in 50% of the seeds in heterozygous dme-6+/− plants. AGL36 was downregulated in a similar manner as FIS2 (41±0.20%), suggesting that DME is indeed involved in early activation of the maternal AGL36 allele.
We also tested the expression of FWA and FIS2 in 6 DAP samples and found that their downregulation were sustained as predicted (Figure 6E). However, to our surprise AGL36 expression in dme-6+/− seeds was elevated more than 50-fold (Figure 6E). This result was unexpected, and implicated a more intricate regulation of AGL36.
DME is required for the activation of MEA, the core histone H3K27 methyltransferase (HMTase) of the PRC2 FIS-complex [46], [55], [56]. To determine whether PRC2 FIS is involved in the regulation of AGL36, we analyzed the relative expression of AGL36 over time (1 to 12 DAP) in mea mutant seeds compared to wild-type (Figure 7A). While AGL36 expression in wild-type seeds was at its maximum at 4 DAP, we observed that AGL36 expression in mea seeds surpassed the maximum levels of wild-type at 4 DAP, and reached its highest levels at around 6 DAP. At this point, the AGL36 relative expression in mea mutant seeds was approximately 40-fold higher than wild-type expression at the same stage, and 7-fold higher than the maximum AGL36 level found in wild-type seeds at 4 DAP (Figure 7A). Our data thus indicate that the FIS-complex is indeed a repressor of AGL36 expression, and could also explain the elevated AGL36 expression level in 3 DAP dme-6+/− seeds (Figure 6E). In line with these findings, we found highly elevated AGL36 relative expression levels in mutant seeds from three different mutant alleles of mea (Figure 7C). Similar results were also obtained with mutants of other components of the FIS PRC2 complex (FIS2, FIE and MSI1, data not shown).
To investigate whether FIS activity was exerted on the maternal and/or paternal allele of AGL36, we performed SNP analyses on the RT-PCR product of AGL36 obtained from mea mutant plants (in Ler background) pollinated with Col wild-type pollen. We found that AGL36 is expressed only from its maternal allele in the mea background throughout the duration of our experiment (Figure 7B). In comparison to the expression pattern in wild-type (Figure 3B), strong ectopic maternal expression was also observed at 9 and 12 DAP stages. No paternal expression could be observed in these stages. By plotting the molarities of the maternal band detected by the Agilent Bioanalyzer, an expression profile for the maternal allele could be generated (Figure 7B, lower panel). This demonstrated that in the absence of MEA, AGL36 expression continues to increase after 4 DAP, and although the intensity decreases from 6 DAP, high level of AGL36 is maintained at 12 DAP. Hence, the FIS-complex represses the maternal allele of AGL36 after the 3 DAP stage.
To further substantiate that maternal AGL36 expression is regulated by the maternal action of MEA, we crossed mea mutant plants with pollen expressing the pAGL36::GUS reporter line. Here, no obvious activation of the paternal transgene could be observed at 3 DAP (Figure S7A). Surprisingly, at 6 DAP, corresponding to embryo heart stage, weak expression of the paternal copy in the embryo could be found (Figure S7A). In addition, we performed reciprocal crosses with the pAGL36::GUS reporter line in mutant mea background. When the transgene was contributed from the female side in mea background, a GUS signal was found in 3 DAP stages that increased drastically up to 6 DAP (Figure S7B). In the reciprocal cross however, no expression could be observed (Figure S7C).
The E(z) class of H3K27 histone methyltransferases (HMTases) in Arabidopsis consists of MEA, SWINGER (SWN) and CURLY LEAF (CLF) that participate in different PRC2 complexes. To test whether AGL36 repression is a specific function of FISMEA PRC2, we analyzed AGL36 expression in homozygous swn-4 and clf-2 seeds. For mutants of both HMTases values similar to the wild-type situation were found, and in conclusion AGL36 appear to be specifically regulated by FISMEA PRC2 (Figure 7C).
In summary, maternal AGL36 expression appears to be repressed specifically by the maternal action of FIS PRC2.
For all genes known to be imprinted by PRC2, the FIS-complex is involved in the repression of the silenced allele [25]-[27], [30], [56]. Our data suggest that silencing of the paternal AGL36 allele requires MET1 whereas the maternal allele is activated by DME. Modulation of female AGL36 expression by PRC2 thus represents a novel mechanism in this type of gene expression system, and adds an additional level of parent-of-origin specific gene expression to the scheme. In order to investigate if this regulation applies to other genes imprinted by the dual action of MET1/DME [11], [18], [19], we analyzed the relative expression levels of FWA, FIS2, AGL36 and MPC in a mea mutant. At 3 DAP expression levels were unchanged or slightly downregulated (0.40–0.99) for all genes tested (Figure 7D). However, while the expression of FWA and FIS2 remained stable at 6 DAP, AGL36 and MPC levels were elevated up to 80-fold (Figure 7D). Thus, genes imprinted by means of MET1/DME can be divided in two classes based on their dependence of FIS PRC2 for additional regulation of the expressed allele. Whereas one class appears not to be regulated by FIS PRC2, the other class depends on the action of the FIS-complex for developmental regulation of its expression.
We have performed genome-wide microarray transcript profiling of seeds with only maternal endosperm as a screening method to identify novel regulators of seed development. Previous experiments have shown that a paternal genomic contribution is essential in wild-type Arabidopsis plants for successful seed development. Thus, our working hypothesis was that in the absence of the paternal genome in the endosperm, key regulators of seed development are not present or not effectively transcribed.
Using selection criteria that allowed for the identification of known paternally expressed genes, we extracted a set of downregulated genes that significantly overlapped with a set of endosperm expressed genes identified by Goldberg & Harada laboratories. The GO-Slim term Transcription factor activity was overrepresented in both down- and up-regulated gene sets, and a closer analysis revealed a striking overrepresentation of the Type-I Mγ MADS-box class among the downregulated transcription factors. With the selection criteria used, each detected gene could be a false positive at a probability of 0.35 at the highest, and thus a thorough examination of candidate genes, as performed in this report for AGL36, will be required.
MADS-box transcription factors play important roles in developmental control and signal transduction pathways in most if not all eukaryotes [57]. They are divided into two groups: the very well studied Type-II group (46 genes) including the MIKC class with important regulators such as AGAMOUS, and the Type-I group (61 genes), on which there is very limited information related to function [54], [58], [59]. Emerging data suggest that Type-I MADS-box genes differ from Type-II genes by being involved in female gametophyte and seed development [46], [60]–[62]. In addition they were found to be only weakly expressed, and most members of this group contain no introns [63].
A comprehensive interaction study with members of the Arabidopsis MADS-box protein family by de Folter and colleagues indicated a complex network of interactions between these proteins (Figure S4). It revealed for instance that PHE1 interacts with AGL62, which in turn interacts with both AGL36 and AGL80. AGL62 itself is regulated by the FIS-complex, and functions as a suppressor of endosperm cellularization [46], [59]. PHE1 and AGL36 on the other hand both interact with AGL28. In addition, mutant analysis has shown that AGL80 function is required for the expression of DME in the central cell, and is therefore an upstream regulator of FIS PRC2 [60]. Moreover, AGL61 is required for central cell development, and there is evidence that a heterodimerization between AGL61 and AGL80 is necessary for AGL61 translocation to the nucleus [59], [62]. PHE1 expression is upregulated in A. thaliana (At) x A. arenosa (Aa) incompatible hybrids due to loss of maternal PHE1 silencing, and introgression of phe1 could improve seed viability in semi-compatible 4xAt x 2xAa crosses [64]. In A. thaliana, expression of a PHE1 antisense construct (MEApromoter::asPHE1) could partially restore the seed abortion phenotype in mea mutants [29]. Peculiarly, PHE1 loss-of-function has no phenotypic effect in A. thaliana [56]. However, given the high sequence similarity within the Mγ class of Type-I MADS-box factors, it is possible that MEApromoter::asPHE1 silenced not only PHE1 but also many other Mγ class genes. Taken together, it seems likely that additional Type-I MADS-box factors are upregulated in mea mutants and a collective downregulation by antisense PHE1 would thus restore some of the defects in mea.
In the cluster of Type-I AGL proteins identified in our screen we also found a large overlap with genes recently shown to be upregulated in incompatibly balanced At x Aa crosses compared to semi-compatible At x Aa maternal excess crosses (AGL35, AGL36, PHE1, PHE2, AGL62, AGL90) [65]. In accordance, mutations of both AGL62 and AGL90 partially restore seed lethality in incompatibly balanced At x Aa crosses, accompanied with selective transmission of the mutant alleles. This array of genes was also found to be upregulated in a PRC2 fis2 mutant [65]. In addition, AGL36, AGL62, AGL90 and PHE1 were commonly upregulated in transcriptional profiles of At paternal excess crosses using tetraploid or unreduced jason (jas) pollen [66].
Together with these recent findings, the network of interactions with AGL62 (AGL36, PHE1, PHE2, AGL90) and PHE1 (AGL40, AGL62) and interactors of these proteins (AGL40, AGL45 and AGL90) strongly suggest that the here identified cluster of Type-I AGL proteins plays key roles in parent-of-origin dependent regulation of seed development. An in-depth study of different members of this group will therefore be of great value in understanding this process, and aid the identification of novel imprinted genes.
Here, we report that AGL36 is a novel imprinted gene that is only expressed from its maternal allele in the endosperm. Silencing of the paternal allele requires the action of MET1, as paternal expression is restored in met1 mutants.
In public high-density DNA methylation maps prepared from wild-type seedlings (http://signal.salk.edu), both the AGL36 transcribed region and the 5′and 3′regulatory regions are decorated by CG methylation. In line with this, AGL36 was expressed at very low levels in vegetative tissues. Transcript levels however, were highly elevated in the absence of MET1, in accordance with the virtual absence of CG methylation in met1 mutants (http://signal.salk.edu) [67].
AGL36 is expressed from both parental alleles at low levels in vegetative tissues, which show that AGL36 imprinting occurs specifically in the endosperm. Other imprinted genes in Arabidopsis have been shown to have biallelic expression in the embryo and other vegetative tissues [11], [34], [68]. However, for most imprinted genes this issue is not clarified [3]. Since paternal AGL36 expression is absent in the seed, it suggests that further silencing of AGL36 takes place by entry into the male germline. Moreover, silencing in the female germline must be lifted to allow AGL36 expression in the seed. Alternatively, maintenance methylation and further silencing do not take place on the AGL36 gene in the female gametophyte. The majority of previously described imprinted genes are regulated by a dual switch of methylation and demethylation involving MET1 and DME [11], [18]–[20], [35]. Here we have shown that AGL36 expression is reduced in a dme mutant, indicating that DME has an activating function towards AGL36. In accordance with this, mutants of CMT3, KYP, AGO4, DDM1 and DRM1/2 had no effect on paternal AGL36 expression suggesting that maintenance and repression by MET1 and activation by DME is sufficient for AGL36 imprinting.
In our SNP analyses, a weak paternal signal was observed only at the 2DAP stage. This was interpreted as an artifact since the signal was absent both before and after this stage. If this is a real paternal signal, it suggests an alternative hypothesis where silencing is achieved in the endosperm post fertilization. Further analyses are however required to support this.
In two recent studies, the genome-wide methylation profile of the seed was dissected by comparing cytosine methylation in wild-type embryos to wild-type and dme endosperm. This showed that endosperm development, and hence the activity of endosperm-specific genes, is marked by an extensive demethylation of the maternal genome, especially at specific transposon sequences [35], [69]. According to the Zilberman Lab Genome Browser (http://dzlab.pmb.berkeley.edu/browser/), such demethylation indeed takes place in the 5′and 3′regulatory regions of AGL36. Methylation patterns are regained in the dme mutant, supporting our data that AGL36 is maternally activated through the action of DME.
In an elegant approach by Gehring and colleagues, novel imprinted genes have recently been identified by the prediction of Differentially Methylated Regions (DMRs) between embryo and endosperm. In support of our findings, significant DMRs were also mapped to 5′and 3′region regions of AGL36 [35].
Imprinting could be demonstrated in transgenic pAGL36::GUS seeds, thus indicating that the 1752 bp promoter fragment used is sufficient for parent-of-origin specific expression. The genomic environment of imprinted genes is highly correlated with transposable elements (TE), and imprinting has been postulated to be an evolutionary byproduct of silencing of invading transposons [23], [69], [70]. For instance, methylation of a SINE-related tandem repeat structure in the 5′-region correlates with FWA expression [32], [71], and DMRs in MEA, PHE1, HDG3 and HDG9 map to TE [35]. In line with this, a variety of remnants of TE reside in both the 5′and 3′ regulatory regions of AGL36 (Figure 4A). The 1752 bp pAGL36::GUS promoter fragment harbors remnants of Helitrons and parts of an Arnoldy TE. An 800 bp DMR maps immediately (78 bp) upstream of the AGL36 transcriptional start site overlapping the Helitron TEs ([35], Mary Gehring, personal communication). Clearly, the 1752 bp 5′region is sufficient for basal AGL36 imprinting, and similar to the abovementioned examples, AGL36 DMRs map to TE. Further investigations will be needed to elucidate the role and the mechanisms of additional 5′and 3′DMRs as well as the involvement of small RNAs in AGL36 imprinting [72].
Distinct from the expression pattern of AGL36 that subsides at the time of cellularization in wild-type seeds, AGL36 maternal expression in mea mutant seeds was highly elevated and sustained throughout seed development. Recently, Walia et al. also reported AGL36 upregulation obtained in five days old seeds from selfed fis2+/− plants [65]. Our results show that FIS-complex mediated repression acts exclusively on the expression of the maternal allele of AGL36. The paternal allele was efficiently silenced throughout endosperm development.
Surprisingly, weak paternal pAGL36::GUS expression could be observed in 6 DAP early heart stage embryos when the mother was homozygous for mea. MEA has been shown to have biallelic expression in the embryo [28], and thus the observed paternal expression in hemizygous mea embryos is not caused by the lack of functional MEA. This could hint to dosage-dependent regulation of paternal AGL36 expression by MEA, directly or indirectly, but in lack of further experiments this remains speculation.
Different PRC2 complexes can regulate common genes [30]. However, in mutants of CLF and SWN, the paralogues of MEA, no significant effect on AGL36 expression levels was found, indicating that AGL36 regulation is specific to PRC2FIS. H3K27 trimethylation mediates PRC2s repressive function, and in a whole-genome assay for H3K27 methylation more than 4400 target genes were detected [73] (Daniel Bouyer, personal communication). AGL36 was however not part of this set of genes. Since this material was obtained from seedlings and may not reflect the situation in the seed, it is not known whether AGL36 is a direct target of H3K27 trimethylation.
Repression of the maternal AGL36 allele identifies a novel means of dual epigenetic regulation of imprinted genes. In this scenario, the expressed maternal AGL36 allele is antagonistically activated by DME and repressed by PRC2FIS. To our knowledge, this is the first report of an imprinted gene where the expressed allele is concurrently regulated by a repressive epigenetic mark.
We asked whether this type of regulation was specific for AGL36 by investigating the fis mutant for expression of three other imprinted genes that are activated by DME. We found that these genes fall into two distinct groups; FWA and FIS2 which were largely unaffected by the lack of FIS, and MPC along with AGL36 which showed strong upregulation. This suggests that additional PRC2 regulation of DME-activated alleles defines a common mechanism that applies to a subset of imprinted genes.
In Arabidopsis, three imprinted genes, MEA, PHE1 and AtFH5 are known to have their silenced allele repressed by PRC2FIS, and two of these genes, MEA and PHE1 are additionally regulated by DNA methylation [55]. In these cases however, the repressed allele is silenced by PRC2 whereas the active allele is regulated by DNA methylation [74]. Here, we show that AGL36 defines a novel type of regulation where the same allele is activated by DME and repressed by PRC2FIS in a sequential fashion. This suggests that maternal AGL36 expression after DME activation needs to be dampened and developmentally regulated by PRC2FIS, in accordance with the strong AGL36 expression observed in hypomethylated met1−/− plants. Interestingly, DME is required to activate both PRC2MEA and AGL36, and is thus a key player in developmental tuning of parent-of-origin specific AGL36 expression.
AGL36 was identified in our transcript profiling as a downregulated gene when the paternal contribution to the endosperm was absent. A simple hypothesis to account for this regulation would be that AGL36 is under the control of one or more paternally expressed factor(s) that activate the maternal allele of AGL36. The identity of such factor(s) remain unknown, and was not approached in this work, but a simple prediction from this hypothesis is that AGL36 would be upregulated in paternal excess interploidy crosses. In a recent report, AGL36 is indeed upregulated in such crosses, as well as in crosses with unreduced diploid jas pollen [66]. Such parental cross-talk is however likely to involve complex genetic and epigenetic regulatory mechanisms, and the mechanism that cause the observed transcriptional response of AGL36 and other previously described imprinted genes in cdka;1p seeds remains to be clarified.
In our study, we have shown that AGL36 is only maternally expressed. Our current model suggests that the paternal allele is silenced by the action of MET1 and the maternal allele activated by DME (Figure 8). In addition, we have also shown that PRC2FIS regulates the expression of the maternal AGL36 allele at the transition between proliferation and cellularization (Figure 8).
Although AGL36 is identified as a novel target of the imprinting machinery in Arabidopsis, we have limited knowledge about its function during plant and seed development. Since expression of AGL36 and its interacting partners coincide with the transition of endosperm from proliferation to differentiation, we speculate that it plays an important role in this process. This is in agreement with recent findings [65], showing that suppression of an AGL cluster including AGL36 is critical for successful transition of endosperm from syncytial to cellularized stage.
In this work we have identified a novel imprinted gene that is controlled by a novel type of dual epigenetic regulation in the seed. This underscores the importance of further investigations to identify imprinted genes in order to unravel the complex network of epigenetic regulation of parent-of-origin effects in seed development.
All plant lines used in these experiments were obtained from the Nottingham Arabidopsis Stock Centre (NASC) unless otherwise stated. The mutant lines cdka;1-1 (SALK_106809; [40], [41]), ddm1-2 (a kind gift from E. Richards; [53]), dme-6 (GK-252E03-014577; Figure S9), mea-8 (SAIL_55_C04; [75]), mea-9 (SAIL_724_E07; Figure S9), met1-4; (SAIL_809_E03; [76] and swn-4 (SALK_109121; [77]) were in the Col accession. The mutant lines ago4-1 (N6364; [78]), clf-2 (N290; [79]), cmt3-7 (N6365; [80]), fis1 (a kind gift from A. Chaudhury; [14]) and kyp-2 (N6367; [51] were in the Ler accession. The drm1;drm2 (N6366; [81]) line was in the Ws-2 accession. Mutants used in this study were genotyped using gene-specific and T-DNA specific primers as described in Table S2. The ddm1-2 mutant line was genotyped by an allele-specific PCR test using dCAPS primers DDM1f and ddm1-2Rsa, as described by [68], allowing digestion of the PCR fragment of the ddme1-2 allele with RsaI restriction endonuclease, generating a ∼130 bp band.
We obtained the agl36-1 allele from the Koncz collection [48]. Allele-specific PCR, using the primers HOOK1 (left border T-DNA primer) and AGL36-AS2-KONCZ (genomic AGL36 primer), was carried out to verify the T-DNA insertion, followed by sequencing analysis of the PCR product using the HOOK1 primer. The left border of the insertion was verified to be 16 bp upstream of the ATG start codon of AGL36. In addition, there is an 11 bp long DNA filler located between the genomic sequence and the T-DNA sequence.
Arabidopsis seeds were surface-sterilized using EtOH, bleach and Tween20 prior to plating out on MS-2 plates [82] supplemented with 2% Sucrose, containing the correct selection when necessary. Seeds on the MS-2 plates were stratified at 4°C O.N before they were incubated for 14 days at 18°C to germinate. The seedlings were then put on soil and grown in long day conditions (16 hr light) at 18°C.
To increase tissue specificity, siliques were cut open and seeds were isolated directly in tubes containing pre-chilled ceramic beads (Roche MagNA Lyser Green Beads). Isolated tissues were stored at −80°C. Homogenization was performed by the addition of Lysis buffer containing β-ME (Sigma Spectrum Plant Total RNA Kit) directly to the samples, followed by 3×15 second intervals of homogenization using a MagNA Lyser Instrument (Roche). To prevent RNA degradation, samples were chilled on ice two minutes between each homogenization interval. After the last homogenization step, the samples were centrifuged at 4°C for one minute prior to the transfer of the lysate to a new 1.5 ml tube. RNA extraction was performed according to the Sigma Plant Total RNA Kit protocol, except that all centrifugation steps were done at 4°C and not at room temperature as indicated in the protocol. RNA was eluted in 50 µl volume. cDNA was synthesized by first preparing the RNA for real-time PCR by treatment with DNase I (Sigma) followed by Reverse Transcription using Oligo(dT) and SuperScript III Reverse Transcriptase (Invitrogen) according to the manufacturer's protocols. The synthesized cDNA was purified utilizing QIAquick PCR Purification kit (Qiagen) and eluted in 30 µl volume prior to measurement of cDNA concentration using a NanoDrop 1000 Spectrophotometer.
Plants were grown and seeds isolated as described above. Total RNA was isolated as described above. For microarray analysis, three biological replicas were generated, each consisting of approximately 35 hand-pollinated siliques from ten different plants. The microarray experiment was conducted by the NARC Microarray Service in Trondheim. Microarray slides were printed by the Norwegian Microarray Consortium (Trondheim, Norway). A custom made Arabidopsis chip with 32567 unique 70-mer oligo probes was used in the experiments. Total RNA (15 mg) and Super-Script III reverse transcriptase (Invitrogen) were used in a reverse transcription reaction. A 3DNA Array 350 kit with Cy3- and Cy5-labelled dendrimers (Genisphere Inc.) was used for labeling. Hybridizations were performed in a Slide Booster Hybridization Station (Advalytix), and the slides were washed according to the manufacturers' descriptions (Genisphere and Advalytix). The slides were scanned at 10 mm resolution on a G2505B Agilent DNA microarray scanner (Agilent Technologies). The resulting image files were processed using GenePix 5.1 software (Axon Instruments). Spots identified as not found or manually flagged out as bad were filtered out. Spots with more than 50% saturated pixels were also excluded. The data sets were log-transformed and normalized using the print-tip Loess approach [83]. Within-array replicated measurements for the same gene were merged by taking the average between the replicates. The data were then scaled so that all array data sets had the same median absolute deviation. The differentially expressed genes were identified using the Limma software package [84]. The resulting set of p-values were used to compute the q-values as described [85].
The microarray data generated in this publication have been deposited in NCBI's Gene Expression Omnibus and are accessible through GEO Series accession number GSE24809 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE24809).
We defined the following sub-sets for our microarray data (see Table S3): All expressed = all genes having a present call (17223 genes); Down 0.8 = in Ler x cdka;1 downregulated genes with q≤0.35 and arithmetic ratio (ar) ≤0.8 (602 genes); Up 1.5 = in Ler x cdka;1 upregulated genes with q≤0.35 and ar ≥1.5 (323 genes); Up 1.2 = in Ler x cdka;1 upregulated genes with q≤0.35 and ar ≥1.2 (1030 genes). The q-value is the false discovery rate (FDR) of the p-value, and was adjusted with Storey's q-procedure [85]. The threshold for analysis was set to q≤0.35 since this value detected paternally expressed genes at an arithmetic ratio (ar) ≤0.8. A functional classification was done at http://www.arabidopsis.org/tools/bulk/go/ using the GO-Slim Molecular Function classification system. For the detailed transcription factor analysis we used the Transcription Factor (TF) classification from the Arabidopsis transcription factor database (AtTFDB) hosted on the Arabidopsis Gene Regulatory Information Server (AGRIS, http://arabidopsis.med.ohio-state.edu/AtTFDB). The MADS TFs were sub-grouped as in de Folter et al [45]. We compared our microarray data with seed expression data generated by the Goldberg & Harada laboratories, available at http://seedgenenetwork.net/analyze?project=Arabidopsis.
For data comparison a reference set of genes was used that contained all genes covered by the Operon chip used in our study (Arabidopsis thaliana 34K NARC serie 8; GEO Platform GPL11051GPL) and the Affimetrix chip used by Goldberg & Harada (Ath1, GEO Platform GPL198). For the Ath1 chip we used the annotation provided by Goldberg & Harada available at http://seedgenenetwork.net/media/Arab_Final_Annotations_09-07-07_completed.txt. For the operon chip we used the current TAIR 9.0 annotation. From these annotations all AGIs for nuclear genes were extracted and the overlap was calculated. This reference set contained 22130 genes.
We used the reference set overlap of the following Goldberg/Harada datasets for comparison: GH seed = call all present and experiment in Arabidopsis ATH1 Array/Arabidopsis/Globular Stage/Seed; GH seed coat = call all present and experiment in Arabidopsis ATH1 Array/Arabidopsis/Globular Stage/Chalazal Seed Coat or Arabidopsis ATH1 Array/Arabidopsis/Globular Stage/General Seed Coat; GH endosperm = call all present and experiment in Arabidopsis ATH1 Array/Arabidopsis/Globular Stage/Chalazal Endosperm or Arabidopsis ATH1 Array/Arabidopsis/Globular Stage/Micropylar Endosperm or Arabidopsis ATH1 Array/Arabidopsis/Globular Stage/Peripheral Endosperm; GH embryo = call all present and experiment in Arabidopsis ATH1 Array/Arabidopsis/Globular Stage/Embryo Proper.
Venn diagrams were generated using the VENN diagram generator designed by Tim Hulsen at http://www.cmbi.ru.nl/cdd/biovenn/ [86]. The test for statistical significance of the overlap between two groups of genes was calculated by using software provided by Jim Lund accessible at http://elegans.uky.edu/MA/progs/overlap_stats.html.
To generate the pAGL36::GUS construct we utilized the Gateway cloning technology (Gateway; Invitrogen). The promoter region (÷1740–12) spanning the ATG start codon was amplified using the attB sequence containing primers attB1-pAGL36-AS7 and attB2-pAGL36-S4 (Table S2), and cloned into the pMDC162 GUS-vector [87]. The resulting construct, after checking the DNA sequence, was introduced to Col ecotype by Agrobacterium tumefaciens mediated transformation using the floral-dip method [88].
Histochemical assays were performed after a modified protocol from Grini et al. (2002) by incubating the tissues in staining buffer (2 mM X-Gluc; 50 mM NaPO4, pH 7.2; 2 mM K4Fe(CN)6 x 3H2O; 2 mM K3Fe(CN)6; 0.1% Triton) overnight at 37°C before the reaction was terminated using 50% EtOH. The tissues were cleared and mounted on slides according to Grini et al. (2002), and inspected using an Axioplan 2 Carl Zeiss Microscope. Images were acquired with an AxioCam HRc Carl Zeiss camera and processed with AxioVs40 V 4.5.0.0 software.
Real-time PCR was performed using a Light-cycler LC480 instrument (Roche) according to the manufacturer's protocol. To ensure high PCR efficiency and to avoid undesired primer dimers, all oligonucleotide pairs were initially tested by melting point analysis using SYBR Premix Ex Taq (TaKaRa). To obtain higher level of gene specificity, probe-based real-time PCR with confirmed primers were performed using Universal Probe Library (UPL) hydrolysis probes (Roche) in combination with Premix Ex Taq (TaKaRa).
For AGL36 real-time PCR, we used primers AGL36-160-LP and AGL36-160-RP, which gave a 60 bp amplicon (Table S2). Comparison of the sequences of the coding region and the 3′UTR of AGL36 with AGL34 and AGL90, revealed more than 85% and 84% sequence similarity respectively between these genes (Figure S8). To ensure that the abovementioned primers are only amplifying AGL36, we cloned the obtained amplicon from four independent reactions into the pCR2.1 vector (Invitrogen), and subsequently sequenced two clones of each construct with M13-Forward and M13-Reverse primers. Sequence results revealed exclusive and specific AGL36 amplification.
ACTIN11 (ACT11), a housekeeping gene that is strongly expressed in the developing ovules [89], was shown in a preliminary analysis not to be affected by our experimental conditions (data not shown), and was therefore selected as a reference gene. GLYCERALDEHYDE-3-P DEHYDROGENASE A-SUBUNIT (GAPA) was used as an additional reference gene. The oligo sequences, their amplicons and appropriate UPL probes are shown in Table S2.
Real-time PCR of all samples and reference controls were performed in two independent biological replicates and repeated at least two times (technical replicas) unless otherwise stated. The PCR efficiency was determined independently for all replicates (biological and technical) by series of dilutions (100 ng, 50 ng, 20 ng, 5 ng template/rxn) for each experiment. This allowed us to obtain the efficiency for each single reaction. Calculations of relative expression ratios were performed according to a model described by Pfaffl [90] with minor exceptions. Since we had efficiency for all reactions (four values for each calculation corresponding to Etarget-sample, Etarget-standard, Ereference-sample and Ereference-standard), we calculated the average Etarget and Ereference values from the standards and the samples, ending up with two E-values that we could use in the formula described by Pfaffl.
RNA was isolated and cDNA synthesized and purified as described above. Polymorphisms between various ecotypes were identified using TAIR Genome Browser (www.arabidopsis.org) and/or the Arabidopsis SNP Sequence Viewer tool provided by the Salk Institute Genomic Analysis Laboratory (http://natural.salk.edu/cgi-bin/snp.cgi). A selected region spanning the SNP of interest was amplified by PCR using TaKaRa Ex Taq DNA polymerase applying 100 ng template per reaction, and the following PCR parameters in a 50 µl reaction: 94°C-3 min, 35×(94°C-1 min, 58°C-30 sec, 72°C-1 min/kb), 72°C-5 min, 4°C-∞. Parental-specific expression based on SNP was determined by setting up an appropriate restriction digest. For AGL36 SNP analysis, 20 µl of the SNP PCR reaction was digested with 15 U of AlwNI at 37°C for a duration of 2.5 hrs, followed by a 20 min inactivation at 65°C. For the FWA control SNP, due to the absence of a restriction site in the SNP region in both Col and Ler ecotypes, dCAPS primers were used, generating a NheI restriction site in the Col ecotype. The obtained amplicons for both ecotypes were digested with NheI [11].
In cases where the detected SNP did not result in digestion in neither ecotype, a primer sequence was designed to introduce a base exchange adjacent to the SNP, leading to restriction digestion of one of the ecotypes. The obtained amplicon for both ecotypes were then treated with the appropriate restriction enzyme. In all experiments either genomic DNA or cDNA from wild-type plants from both ecotypes used in the study was used as controls for the presence or absence of digestion. The digested samples were analyzed using DNA-1000-LabOnChip and 2100 Bioanalyzer (Agilent Technologies).
To rule out that the primers used for AGL36 SNP PCR (AGL36-SP7-SNP and AGL36-ASP6-SNP) (Table S2) would amplify the highly similar AGL90, we oriented the AGL36-SP7-SNP primer such that it was located in a region that was annotated as intron in AGL90 but not in AGL36 (Figure S8). First, the presence of the intron in AGL90 was confirmed by amplifying the intron-flanking region (AGL90-SP1-subcloning and AGL90-ASP2-subcloning primers (Table S2)), and comparing the size differences obtained between the genomic PCR and cDNA PCR. Due to high sequence similarity, we suspected to amplify both AGL36 and AGL90 in these PCR reactions. To distinguish between these two amplicons, we took advantage of the presence of two unique restriction sites (MslI and BspBI) in the amplified region of AGL36 that are absent in AGL90.
Sequence comparison between the abovementioned AGL36-SNP primers and AGL34 showed that there was approximately 70% and 91% sequence similarity between the primers and the AGL34 gene. However, if these primers were functional in amplifying AGL34, they would result in a smaller amplicon than AGL36 amplicon (373 bp versus 399 bp respectively). This difference could easily be detected using a DNA-1000-LabOnChip. Our SNP data only showed the expected 399 bp band, verifying that AGL34 was not amplified using the above primers. The paternally imprinted FWA gene was used as a positive control by utilizing primers FWA-RTf and FWA-dNheI (Table S2) for PCR amplification followed by NheI restriction digest.
|
10.1371/journal.ppat.1000919 | The Haemophilus influenzae HMW1C Protein Is a Glycosyltransferase That Transfers Hexose Residues to Asparagine Sites in the HMW1 Adhesin | The Haemophilus influenzae HMW1 adhesin is a high-molecular weight protein that is secreted by the bacterial two-partner secretion pathway and mediates adherence to respiratory epithelium, an essential early step in the pathogenesis of H. influenzae disease. In recent work, we discovered that HMW1 is a glycoprotein and undergoes N-linked glycosylation at multiple asparagine residues with simple hexose units rather than N-acetylated hexose units, revealing an unusual N-glycosidic linkage and suggesting a new glycosyltransferase activity. Glycosylation protects HMW1 against premature degradation during the process of secretion and facilitates HMW1 tethering to the bacterial surface, a prerequisite for HMW1-mediated adherence. In the current study, we establish that the enzyme responsible for glycosylation of HMW1 is a protein called HMW1C, which is encoded by the hmw1 gene cluster and shares homology with a group of bacterial proteins that are generally associated with two-partner secretion systems. In addition, we demonstrate that HMW1C is capable of transferring glucose and galactose to HMW1 and is also able to generate hexose-hexose bonds. Our results define a new family of bacterial glycosyltransferases.
| Decoration of proteins with carbohydrates has an important impact on protein function throughout biology and has been recognized increasingly in pathogenic bacteria. Haemophilus influenzae is a common cause of both bacterial respiratory tract disease and bacterial invasive disease and initiates infection by colonizing the upper respiratory tract. The Haemophilus HMW1 adhesin is a large protein that resides on the bacterial surface and mediates bacterial attachment to respiratory epithelial cells, an essential step in the process of colonization. In recent work, we discovered that HMW1 is decorated at multiple sites with short carbohydrate units that serve to prevent degradation and to stabilize association with the bacterial surface. In the current study we identify the enzyme responsible for adding carbohydrate units at specific sites of HMW1. In addition, we demonstrate that this enzyme is capable of creating both carbohydrate-protein and carbohydrate-carbohydrate bonds. The amino acid sequence of this enzyme is similar to the sequences of proteins in several other bacteria, suggesting a new family of bacterial enzymes capable of creating carbohydrate-protein and carbohydrate-carbohydrate bonds.
| Glycosylation of proteins is an essential process that plays an important role in protein structure and function and represents a strategy to fine tune cell-cell recognition and signaling. For a long period of time, glycosylation of proteins was believed to be restricted to eukaryotes. However, in recent years glycoproteins have been identified increasingly in prokaryotes as well, including pathogenic bacteria such as Pseudomonas aeruginosa, Campylobacter spp., Neisseria spp., and E. coli, among others [1]–[10].
Nonencapsulated (nontypable) Haemophilus influenzae is a human specific pathogen that is a common cause of localized respiratory tract and invasive disease and initiates infection by colonizing the upper respiratory tract [11], [12]. Approximately 75–80% of isolates express two related high-molecular weight proteins called HMW1 and HMW2 that mediate high-level adherence to respiratory epithelial cells and facilitate the process of colonization [13], [14]. The HMW1 and HMW2 adhesins are encoded by homologous chromosomal loci that appear to represent a gene duplication event and contain 3 genes, designated hmw1A, hmw1B, and hmw1C and hmw2A, hmw2B, and hmw2C, respectively [15], [16].
HMW1 and HMW2 are synthesized as pre-pro-proteins (Figure 1A) and are secreted by the two-partner secretion system [17]–[19]. Amino acids 1–68 represent an atypical signal peptide and direct the pre-pro-proteins to the Sec apparatus, where they are cleaved by signal peptidase I [18]. The resulting pro-proteins are targeted to the HMW1B and HMW2B outer membrane translocators and undergo cleavage between amino acids 441 and 442, removing the pro-pieces and generating mature species that are 125 kDa and 120 kDa, respectively [18]–[21] (Figure 1A). Following translocation across the outer membrane, mature HMW1 and HMW2 remain non-covalently associated with the bacterial surface [18], [19].
In recent work, we demonstrated that HMW1 is a glycoprotein and undergoes glycosylation in the cytoplasm in a process that is dependent upon HMW1C [22]. Functional analyses revealed that glycosylation of HMW1 protects against premature degradation, analogous to some eukaryotic proteins [22]. In addition, glycosylation appears to influence HMW1 tethering to the bacterial surface, a prerequisite for HMW1-mediated adherence [22]. Based on carbohydrate composition analysis of purified HMW1 using gas chromatography and combined gas chromatography-mass spectrometry, the modifying sugars include glucose, galactose, and possibly small amounts of mannose [22]. Analysis of HMW1 proteolytic fragments by mass spectrometry identified 31 sites of modification [23]. All of the modified sites were asparagine residues, in all except one case within the conventional sequence motif for eukaryotic N-linked glycosylation, namely NX(S/T) where X is any residue except for proline [23]. LC-MS/MS analysis, accurate mass measurement, and deuterium replacement studies established that the modifying glycan structures were mono-hexose or di-hexose units rather than N-acetylated hexosamine units that comprise the di-N-diacetyl chitobiose core of eukaryotic and many bacterial asparagine-linked glycans. These results suggested a novel N-linked carbohydrate-peptide transferase activity that does not require assembly of the monosaccharide units onto a lipid-linked intermediate [23].
In the present study, we studied the enzymatic mechanism responsible for the glycosylation of asparagine residues in HMW1. We found that the HMW1C protein encoded in the hmw1 gene cluster is capable of transferring glucose and galactose to the HMW1 adhesin. In addition, HMW1C is capable of generating hexose-hexose linkages.
In earlier work we found that insertional inactivation of hmw1C in H. influenzae strain Rd-HMW1 resulted in a loss of glycosylation of HMW1 [22], suggesting that HMW1C participates in the process of glycosylation. Further analysis revealed that amino acids 386–439 in HMW1C share 40–41% identity and 51–65% similarity with a domain conserved in a family of eukaryotic O-GlcNAc transferases, including human O-GlcNAc transferase, rat O-GlcNAc transferase, and a plant protein called Spy [22], raising the possibility that HMW1C is a glycosyltransferase.
To address the possibility that HMW1C is the glycosyltransferase responsible for N-linked glycosylation of HMW1, we purified HAT-tagged HMW1C and Strep-tagged HMW1802–1406 (Figure 1B). HMW1802–1406 corresponds to just over half of mature HMW1 (HMW1442–1536), contains 18 documented N-linked glycosylation sites, and was more amenable to purification than mature HMW1 (Figure 1A). Subsequently, we incubated approximately equimolar quantities of HAT-HMW1C and Strep-HMW1802–1406 with both UDP-α-D-glucose and UDP-α-D-galactose at room temperature for 60 minutes, then examined the reaction mixture for reactivity with the DIG-glycan reagents. As shown in Figure 2A, we observed efficient glycosylation of HMW1802–1406 that was dependent on both HMW1C and the UDP-hexoses. To extend this result, we performed the same experiment with UDP-α-D-glucose by itself, UDP-α-D-galactose by itself, GDP-α-D-mannose by itself, UDP-α-D-N-Acetylglucosamine by itself, and UPD-α-D-N-Acetylgalactosamine by itself. As shown in Figure 2B, we observed glycosylation with UDP-α-D-glucose alone and UDP-α-D-galactose alone but not with GDP-α-D-mannose, UDP-α-D-N-Acetylglucosamine, or UPD-α-D-N-Acetylgalactosamine alone. To determine whether smaller amounts of HMW1C are associated with appreciable glycosylation of HMW1802–1406, we repeated assays with a fixed amount of HMW1802–1406, fixed amounts of UDP-α-D-glucose and UDP-α-D-galactose, and dilutions of HMW1C. Based on analysis using DIG-glycan reagents, we observed efficient glycosylation with molar quantities of HMW1C that were less than one-tenth the molar quantity of HMW1802–1406 (data not shown).
To address whether the glycosylation of HMW1802–1406 in in vitro reactions mimicked glycosylation of native HMW1 in whole bacteria and to gain further insight into which sugars modify which sites, we repeated reactions with purified Strep-tagged HMW1802–1406, purified HAT tagged HMW1C, and UDP-α-D-glucose alone, UDP-α-D-galactose alone, GDP-α-D-mannose alone, or UDP-α-D-glucose plus UDP-α-D-galactose plus GDP-α-D-mannose and then examined the reaction mixtures by LC-MS/MS. As a positive control we examined purified HMW1802–1406 recovered from DH5α/pASK-HMW1802–1406 + pHMW1C, and as a negative control we examined HMW1802–1406 recovered from DH5α/pASK-HMW1802–1406 (lacking pHMW1C).
As summarized in Table 1, we detected 10 of the 18 predicted sites of glycosylation and 11 distinct glycopeptides in HMW1802–1406, including 10 glycopeptides with a single site of glycosylation and one glycopeptide with two sites of glycosylation (KNITFEGGNITFGSR). Interestingly, of the 10 sites of glycosylation, all were modified in the in vitro reactions with UDP-α-D-glucose alone and with UDP-α-D-glucose plus UDP-α-D-galactose plus GDP-α-D-mannose. In contrast, only 6 of the 10 sites of glycosylation were modified in the in vitro reactions with UDP-α-D-galactose alone. Consistent with our observations using DIG-Glycan reagents, no sites were glycosylated in the in vitro reactions with GDP-α-D-mannose alone.
As demonstrated by the collision-induced fragmentation spectra shown in Figure 3 and Figure S1, the glycopeptide NLSITTNSSSTY (HMW1 amino acids 946–958, with glycosylation at N952) and the glycopeptide AITNFTFNVGGLFDNK (HMW1 amino acids 909–924, with glycosylation at N912) were present in two forms, including one with a mono-hexose at the predicted site of glycosylation and the other with a di-hexose at the predicted site of glycosylation. The forms containing a mono-hexose were detected in the in vitro reactions with UDP-α-D-glucose alone, UDP-α-D-galactose alone, and UDP-α-D-glucose plus UDP-α-D-galactose plus GDP-α-D-mannose, while the forms containing a di-hexose were detected only in the in vitro reactions with UDP-α-D-glucose alone and with UDP-α-D-glucose plus UDP-α-D-galactose plus GDP-α-D-mannose, suggesting that glucose must be the first hexose linked to asparagine in the glycopeptides containing di-hexose modification.
Together these findings demonstrate that the HMW1C protein is a glycosyltransferase and has a novel activity capable of transferring glucose and galactose to asparagine residues in HMW1 and creating hexose-hexose bonds. In addition, they demonstrate that the di-hexosylated sites at N951 and N912 are initially modified with a glucose monosaccharide.
To extend our understanding of glycosylation of HMW1 and confirm our observation that HMW1 is modified with glucose and galactose in in vitro glycosylation assays, we examined the effect of insertional inactivation of galU (open reading frame HI0812 in strain Rd) on glycosylation of HMW1 in strain Rd-HMW1. The galU gene encodes glucose-1-phosphate uridyl transferase, which converts glucose-1-phosphate to UDP-glucose (Figure S2). UDP-glucose in turn can be converted directly to UDP-galactose by GalE (UDP Gal-4-epimerase) or can serve as the donor of UDP for conversion of galactose-1-phosphate to UDP-galactose. In assessing the effect of inactivation of galU, we incubated Rd-HMW1/galU in supplemented brain heart infusion broth [24], which contains glucose as the primary carbon source. Interestingly, inactivation of galU mimicked the effect of inactivation of hmw1C described in our earlier work [22], eliminating HMW1 glycosylation as assessed by DIG-glycan blots (Figure 4A), virtually eliminating HMW1 tethering to the bacterial surface (Figure 4B), and abolishing HMW1-mediated adherence (Figure 4C).
Consistent with our in vitro glycosyltransferase assays with purified HMW1802–1406 and HMW1C, these results indicate that UDP-glucose is required for glycosylation of HMW1 in H. influenzae under standard growth conditions in supplemented brain heart infusion broth.
In this study, we found that the H. influenzae HMW1C protein encoded in the hmw1 gene cluster is a glycosyltransferase and is capable of transferring glucose and galactose to asparagine residues in the HMW1 adhesin, providing the first example of a glycosyltransferase that transfers hexose units rather than N-acetylated amino sugars to asparagine residues in protein targets. Further analysis revealed that HMW1C is capable of creating both hexose-asparagine and hexose-hexose linkages, suggesting multi-functionality as a glycosyltransferase.
All previously reported carbohydrate modification of asparagine residues in proteins in Eukarya and Bacteria involve the en bloc transfer of oligosaccharides from a lipid-linked intermediate by an oligosaccharyltransferase complex [25]. In Archaea, the mechanisms of N-glycosylation are less well understood. Glycosylation of asparagine residues with a trisaccharide moiety in the flagellin and S-layer proteins of Methanococcus voltae has been proposed to proceed via a lipid-linked intermediate [26]. More recently it has been shown that hexose units are attached directly to asparagine residues in an S layer glycoprotein of Haloferax volcanii [27]. A pentasaccharide with the structure Hex-X-hexuronic acid-HexA-HexA-Hex-peptide was identified at two glycosylation sites. Interestingly, these two sites were different from the conventional N-glycosylation sequence motif observed in eukaryotes and in HMW1. It is currently unclear whether the H. volcanii Hex-Asn linkage is formed from a lipid-linked intermediate or via activated monosaccharides as we have found with HMW1 and HMW1C.
In earlier work, we performed carbohydrate composition analysis on purified HMW1 and detected glucose, galactose, and small amounts of mannose [22]. Given the potential for contaminating sugars to be detected in this analysis, we were uncertain as to whether mannose was truly present as a modifying sugar in HMW1, especially given that it accounted for only 2.5–3% of the total carbohydrate [22]. Our analysis in the current study argues that mannose is not present in HMW1. In particular, in in vitro glycosyltransferase assays using purified HMW1802–1406, HMW1C, and GDP-α-D-mannose, we were unable to detect modification of HMW1802–1406 using either DIG-Glycan reagents or LC-MS/MS.
Based on assessment of the 10 glycopeptides that we detected in our in vitro glycosylation assays with HMW1802–1406, which corresponds to just over half of mature HMW1, we observed that HMW1C transfers glucose to all glyscosylated asparagines and transfers galactose to only a subset of glyscosylated asparagines. All of these glycosylation sites correspond to the conventional sequence motif of N-linked glycans, namely NX(S/T), with X being any amino acid except proline. Examination of the primary amino acid sequence of the sites that are modified only with glucose and the sites that are modified with either glucose or galactose in in vitro assays reveals no apparent distinction, suggesting that factors beyond the amino acid sequence influence the specificity or potentially the efficiency of glycosylation. This observation is consistent with the fact that only a fraction of conventional sequences motifs are glycosylated in HMW1 purified from H. influenzae [23].
Further analysis of the glycopeptides detected after in vitro glycosylation revealed two peptides that were modified with a di-hexose. Interestingly, in both cases the glycopeptides were detected only in the reactions performed with UDP-α-D-glucose alone and with UDP-α-D-glucose plus UDP-α-D-galactose plus GDP-α-D-mannose, indicating modification with UDP-α-D-glucose. In contrast, the corresponding glycopeptides containing a single hexose at the asparagines in question were detected in the reactions performed with UDP-α-D-glucose alone, with UDP-α-D-galactose alone, and with UDP-α-D-glucose plus UDP-α-D-galactose plus GDP-α-D-mannose, indicating modification with either glucose or galactose. Considered together, these results suggest that glucose must be linked to asparagine in the glycopeptides containing di-hexose modification. At this point, it is unclear whether the di-hexose is generated prior to modification of the acceptor asparagine residue or whether instead a single hexose is linked to the target asparagine and then a second hexose is linked to the first hexose, although the conventional interpretation is that the hexose is added to the protein and then the chain is extended. In either event, it appears that HMW1C is responsible for creating the hexose-hexose bond.
Interestingly, homology analysis reveals 42–68% identity and 58–83% similarity between the full-length HMW1C sequence and proteins in a number of other gram-negative bacterial pathogens, including the enterotoxigenic E. coli (ETEC) EtpC protein and predicted proteins in Yersinia pseudotuberculosis, Y. enterocolitica, Y. pestis, H. ducreyi, Actinobacillus pleuropneumoniae, Mannheimia spp., Xanthomonas spp., and Burkholderia spp, among others (Table S1). In ETEC, Y. pseudotuberculosis, Y. enterocolitica, and Y. pestis, these homologs are encoded by genes that are adjacent to known or predicted two-partner secretion loci. The H. ducreyi, Mannheimia succiniciproducens, and Burkholderia xenovorans genomes contain genes that encode predicted two-partner secretion proteins as potential targets for the HMW1C homologs, although these genes are in unlinked locations. The ETEC EtpC protein is encoded by a two-partner secretion locus called etpBAC and has been shown to be required for glycosylation of the EtpA adhesin, a high-molecular weight protein that has a predicted molecular mass of ∼177 kDa and promotes adherence to intestinal epithelial cells and colonization of the intestine in mice [28], [29]. These observations suggest that that there is a family of bacterial HMW1C-like proteins with glycosyltransferase activity.
To summarize, in eukaryotes N-linked glycosylation occurs in the endoplasmic reticulum and involves an oligosaccharyltransferase that catalyzes the transfer of the oligosaccharide from the lipid donor dolichylpyrophosphate to the acceptor protein. Similarly, in bacteria, N-glycosylation generally occurs in the periplasm and involves an oligosaccharyltransferase that transfers the glycan structure from a lipid donor to the acceptor protein. In contrast, in the case of the H. influenzae HMW1 adhesin, N-linked glycosylation occurs in the cytoplasm and involves direct transfer of hexose units to the acceptor protein by HMW1C, with no requirement for a lipid donor. In this study, we have established that the H. influenzae HMW1C protein is a multi-functional enzyme that is capable of transferring glucose and galactose to asparagine residues in selected conventional N-linked sequence motifs in HMW1 and is also capable of creating hexose-hexose linkages. Based on homology analysis, it is likely that a variety of other bacteria possess HMW1C-like proteins with similar enzymatic activity. In future work, we will examine whether these HMW1C-like proteins are identical to HMW1C in terms of the glycan units that they transfer and the acceptor protein sequence motifs that they recognize.
The strains and plasmids used in this study are listed in Table 2. H. influenzae strain Rd-HMW1 is a derivative of strain Rd that contains the intact hmw1 locus and expresses fully functional HMW1 [22]. H. influenzae strain Rd-HMW1/hmw1C is a derivative of strain Rd-HMW1 that contains an insertionally inactivated hmw1C gene [22]. The H. influenzae Rd-HMW1 derivative harboring a kanamycin cassette in galU was constructed by transforming competent Rd-HMW1 with genomic DNA recovered from RdgalU and selecting for kanamycin resistance [30].
In order to overexpress HMW1802–1406 with a Strep tag at the N terminus, the fragment encoding HMW1802–1406 was amplified by PCR from pHMW1-14 using a 5′ primer that incorporated a BamHI site and a 3′ primer that incorporated a SalI site. The PCR amplicon was digested with BamI and SalI and then ligated into BamHI-SalI-digested pASK-IBA12 (IBA, BioTAGnology), creating pASK-HMW1802–1406.
In order to overexpress the HMW1C protein with a HAT epitope at the N terminus, the hmw1C gene was amplified by PCR from pHMW1-14 using a 5′ primer that incorporated a BamHI site and a 3′ primer that incorporated an EcoRI site. The PCR amplicon was digested with BamHI and EcoRI and then ligated into BamHI-EcoRI-digested pHAT10 (Clontech), creating pHAT-HMW1C.
Plasmids were introduced into E. coli by chemical transformation [31]. DNA was introduced into H. influenzae using the MIV method of transformation described by Herriott et al. [32]. Transformants were selected by plating on agar containing kanamycin, and mutations were confirmed by PCR analysis using primers that anneal to regions flanking the target gene.
To purify HMW1802–1406, E. coli strain DH5α/pASK-HMW1802–1406 was grown at 37°C to an OD600 of 0.7, then induced for 2 hrs with the addition of 100 µg/ml of anhydro-tetracycline (Sigma). Cells were harvested, resuspended in 100 mM Tris pH 8.0, 150 mM NaCl with Complete Mini protease inhibitor (Roche), and lysed by sonication. Insoluble material was removed by centrifugation at 12,500 × g for 30 min. The supernatant was loaded onto a Strep-Tactin Superflow cartridge and eluted according to the manufacturer's instructions (IBA, BioTAGnology). Eluted fractions were analyzed for purity by SDS-PAGE and were pooled. To purify HMW1C, E. coli strain DH5α/pHAT-HMW1C was grown at 37°C overnight. Cells were recovered, resuspended in 50 mM sodium phosphate buffer pH 7.0, 300 mM NaCl (bufferA), and lysed by sonication. Insoluble material was removed by centrifugation at 12,500 × g for 30 min. The supernatant was loaded onto a 1 ml Talon column (Clontech) and eluted with a gradient of 0 to 300 mM imidazole in Buffer A. Fractions were analyzed for purity by SDS-PAGE and were pooled.
In standard in vitro glycosyltransferase assays, 1.5 µg (23 pmole) of purified HMW1802–1406 was combined with a mixture containing 20 µl of 50 mM UDP-α-D-glucose, 50 mM UDP-α-D-galactose, 50 mM GDP-α-D-mannose, 50 mM UDP-α-D-N-Acetylglucosamine, or 50 mM UDP-α-D-N-acetylgalactosamine (Calbiochem) either as individual sugars or as mixtures. The reactions were initiated with addition of 1.5 µg (21 pmole) of purified HMW1C in a final volume of 150 µl in 25 mM Tris pH 7.2, 150 mM NaCl. Samples were incubated for 60 minutes at room temperature and then further incubated at 4°C overnight.
To detect protein glycosylation, DIG Glycan reagents (Roche) were employed. Use of these reagents is based on the oxidation of hydroxyl groups in carbohydrates to aldehydes either in solution or bound to nitrocellulose membranes. Digoxigenin is then covalently linked to the aldehyde groups, and an anti-digoxigenin alkaline-phosphatase conjugated agent is used for detection of labeled carbohydrates.
FACS analysis was performed by the Duke University Medical Center Cancer Research Center Flow Cytometry Shared Resource Center using a Becton Dickinson FACS Calibur instrument at a wavelength of 488 nm. Bacterial suspensions were fixed with 1% formaldehyde in PBS at room temperature for 30 min. After washing once with Tris buffered saline (TBS), bacteria were resuspended in 1 ml of TBS, 50 mM EDTA, 0.1% bovine serum albumin, and a 1∶1000 dilution of guinea pig antiserum GP85 directed against HMW1 [33] and were incubated with gentle rocking at room temperature for 1 hr. Samples were then centrifuged, washed twice with PBS, and resuspended in 200 µl of PBS, 0.1% bovine serum albumin, and a 1∶200 dilution of Alexa Fluor488 anti-guinea pig antibody (Molecular Probes). Samples were incubated with gentle rocking at room temperature for 1 hr. After two additional washes with PBS, bacterial pellets were re-suspended in 1 ml of PBS and were then analyzed. Data were analyzed with CELLQUEST software (Becton Dickinson). To quantify histograms, markers were drawn on plots, and positive events within the markers were determined as a percentage of the positive control (set at 100%).
Adherence assays were performed with Chang epithelial cells (human conjunctiva; ATCC CCL 20.2) (Wong-Kilbourne derivative clone 1-5c-4) as described previously [16]. Percent adherence was calculated by dividing the number of adherent colony-forming units by the number of inoculated colony-forming units. All strains were examined in triplicate, and each assay was repeated at least two times.
Whole cell sonicates were prepared by suspending bacterial pellets in 10 mM HEPES, pH 7.4 and sonicating to clarity. Proteins were resolved by SDS-PAGE using 10% polyacrylamide gels. Western blots were performed using guinea pig antiserum GP85 against the HMW1 protein [30].
Samples were precipitated using the 2D protein clean up kit (GE Healthcare) according to the manufacturer's instructions. Bovine serum albumin (100 ng) was added to each sample as an internal standard. Pellets were dissolved in 40 µl 9 M urea and aliquoted into 0.5 ml microfuge tubes. Samples (20 µl in 9 M urea) were reduced with 5 mM TCEP at pH 8.0 at room temperature for 30 min and were alkylated with 10 mM iodoacetamide (Bio-Rad) in the dark at room temperature for 30 min. TCEP and iodoacetamide were quenched with 5 mM dithiothreitol (DTT) at room temperature for 10 min. The reduced and alkylated proteins were digested with 1 µg of endoproteinase Lys-C (Roche) at 37°C overnight. Samples were diluted with 64 µl H2O to reduce the concentration of urea to 2 M and were then digested with 4 µg trypsin (Sigma) at 37°C overnight. Peptides were acidified with 5.5 µl formic acid (Sigma) and extracted 6 times with 10–200 µl NuTip porous graphite carbon wedge tips (Glygen) according to the manufacturer's directions and were then eluted into 1.5 ml autosampler vials with 60% acetonitrile (Burdick & Jackson) in 0.1% formic acid. The peptide digests were evaluated for quality and detergent contaminants using MALDI-TOF/TOF [34] prior to LC-MS analysis. For MALDI-TOF/TOF analysis, the peptide sample (0.5 µl) was mixed with an equal volume of MALDI matrix solution (Agilent Technologies) prior to spotting. For nano-LC-FTICR-MS analysis, the peptide sample was dried and immediately dissolved in 10 µl aqueous acetonitrile/formic acid (1%/1%).
The complex mixtures of peptides and glycopeptides from HMW1802–1406 were analyzed using high-resolution nano-LC-MS on a hybrid mass spectrometer consisting of a linear quadrupole ion-trap and an Orbitrap (LTQ-Orbitrap XL, Thermo-Fisher). The liquid chromatographs were nanoflow HPLC systems (NanoLC-1Dplus™ and NanoLC-Ultra™) that were interfaced to the mass spectrometer with a nanospray source (PicoView PV550; New Objective). The in-house packed LC column (Jupiter C12 Proteo, 4 µm particle size, 90 Å pore size [Phenomenex]) was equilibrated in 98% solvent A (aqueous 0.1% formic acid) and 2% solvent B (acetonitrile containing 0.1% formic acid). The samples (10 µL) were injected from autosampler vials using the LC-systems autosamplers at a flow rate of 1.0 µL/min and were eluted using a segmented linear gradient (250 nL/min) with solvent B: isocratic at 2% B, 0–2 min; 2% B to 40% B, 2–65 min; 40% B to 80% B, 65–70 min; isocratic at 80% B, 70–72 min; 80% B to 2% B, 72–77 min; and isocratic at 2% B, 77–82 min. The survey scans (m/z 350–2000) (MS1) were acquired at high resolution (60,000 at m/z = 400) in the Orbitrap, and the MS/MS spectra (MS2) were acquired in the linear ion trap at low resolution, both in profile mode. The maximum injection times for the MS1 scan in the Orbitrap and the LTQ were 50 ms and 100 ms, respectively. The automatic gain control targets for the Orbitrap and the LTQ were 2×105 and 3×104, respectively. The MS1 scans were followed by six MS2 events in the linear ion trap with wideband collision activation in the ion trap (parent threshold = 1000; isolation width = 2.0 Da; normalized collision energy = 30%; activation Q = 0.250; activation time = 30 ms). Dynamic exclusion was used to remove selected precursor ions (−0.25/+1.5 Da) after MS2 acquisition with a repeat count of 2, a repeat duration of 30 s, and a maximum exclusion list size of 200. The following ion source parameters were used: capillary temperature 200 °C, source voltage 2.5 kV, source current 100 µA, and the tube lens at 79 V. The data were acquired using Xcalibur, version 2.0.7 (Thermo-Fisher).
The MS2 spectra were analyzed both by searching a customized protein database that contained the sequences of HMW802–1406 and by expert manual interpretation. The exact masses of the glycopeptides and fragmentation ions were calculated using the Molecular Weight Calculator, version 6.45 (http://ncrr.pnl.gov/software/). For database searches, the LC-MS files were processed using MASCOT Distiller (Matrix Science, version 2.3.0.0) with the settings previously described [35]. The resulting MS2 centroided files were used for database searching with MASCOT, version 2.1.6, and the following parameters: enzyme, trypsin; MS tolerance = 10 ppm; MS/MS tolerance = 0.8 Da with a fixed carbamidomethylation of Cys residues and the following variable modifications: Methionine, oxidation; Pyro-glu (N-term); Maximum Missed Cleavages = 5; and 1+, 2+, and 3+ charge states.
|
10.1371/journal.pcbi.1000593 | The GAAS Metagenomic Tool and Its Estimations of Viral and Microbial Average Genome Size in Four Major Biomes | Metagenomic studies characterize both the composition and diversity of uncultured viral and microbial communities. BLAST-based comparisons have typically been used for such analyses; however, sampling biases, high percentages of unknown sequences, and the use of arbitrary thresholds to find significant similarities can decrease the accuracy and validity of estimates. Here, we present Genome relative Abundance and Average Size (GAAS), a complete software package that provides improved estimates of community composition and average genome length for metagenomes in both textual and graphical formats. GAAS implements a novel methodology to control for sampling bias via length normalization, to adjust for multiple BLAST similarities by similarity weighting, and to select significant similarities using relative alignment lengths. In benchmark tests, the GAAS method was robust to both high percentages of unknown sequences and to variations in metagenomic sequence read lengths. Re-analysis of the Sargasso Sea virome using GAAS indicated that standard methodologies for metagenomic analysis may dramatically underestimate the abundance and importance of organisms with small genomes in environmental systems. Using GAAS, we conducted a meta-analysis of microbial and viral average genome lengths in over 150 metagenomes from four biomes to determine whether genome lengths vary consistently between and within biomes, and between microbial and viral communities from the same environment. Significant differences between biomes and within aquatic sub-biomes (oceans, hypersaline systems, freshwater, and microbialites) suggested that average genome length is a fundamental property of environments driven by factors at the sub-biome level. The behavior of paired viral and microbial metagenomes from the same environment indicated that microbial and viral average genome sizes are independent of each other, but indicative of community responses to stressors and environmental conditions.
| Metagenomics uses DNA or RNA sequences isolated directly from the environment to determine what viruses or microorganisms exist in natural communities and what metabolic activities they encode. Typically, metagenomic sequences are compared to annotated sequences in public databases using the BLAST search tool. Our methods, implemented in the Genome relative Abundance and Average Size (GAAS) software, improve the way BLAST searches are processed to estimate the taxonomic composition of communities and their average genome length. GAAS provides a more accurate picture of community composition by correcting for a systematic sampling bias towards larger genomes, and is useful in situations where organisms with small genomes are abundant, such as disease outbreaks caused by small RNA viruses. Microbial average genome length relates to environmental complexity and the distribution of genome lengths describes community diversity. A study of the average genome length of viruses and microorganisms in four different biomes using GAAS on 169 metagenomes showed significantly different average genome sizes between biomes, and large variability within biomes as well. This also revealed that microbial and viral average genome sizes in the same environment are independent of each other, which reflects the different ways that microorganisms and viruses respond to stress and environmental conditions.
| Metagenomic approaches to the study of microbial and viral communities have revealed previously undiscovered diversity on a tremendous scale [1],[2]. Metagenomic sequences are typically compared to sequences from known genomes using BLAST to estimate the taxonomic and functional composition of the original environmental community [3]. Many software tools designed to estimate community composition (e.g. MEGAN) annotate sequences using only the best similarity [4]. However, the best similarity is often not from the most closely related organism [5]. In addition, most metagenomes contain a large percentage of sequences from novel organisms which cannot be identified by BLAST similarities, further complicating analysis [1],[6],[7].
Mathematical methods based on contig assembly have been developed to estimate viral diversity and community structure from metagenomic sequences regardless of whether they are similar to known sequences [8]. These similarity-independent methods require the input of the average genome length of viruses from a given sample [8]. Having an accurate value of this average is important because it takes a potentially large range spanning 3 orders of magnitude, and has a large influence on the diversity estimates. Average genome length for an environmental community can be determined using Pulsed Field Gel Electrophoresis (PFGE) [9],[10]. PFGE gives a spectrum of genome lengths in a microbial or viral consortium, indicated by electrophoretic bands on an agarose gel, which can be used to calculate an average genome length. Due to the large variability of dsDNA virus genome length, PFGE can discriminate and identify dominant viral populations [11]. However, PFGE is limited because the bands are not independent and a single band can contain different DNA sequences [12],[13].
Average genome length in environmental samples has also been used as a metric to describe community diversity and complexity [9], [14]–[17]. In PFGE, both a larger size range and a greater number of bands indicate a wider variety of genomes and hence, a more diverse community [9],[14],[16],[17]. The average genome length of a microbial community has been shown to serve as a proxy for the complexity of an ecosystem [15]. Longer average genome lengths indicate higher complexity [15], since larger bacterial genomes can encode more genes and access more resources [18].
Here we introduce Genome relative Abundance and Average Size (GAAS), the first bioinformatic software package that simultaneously estimates both genome relative abundance and average genome length from metagenomic sequences. GAAS is implemented in Perl and is freely available at http://sourceforge.net/projects/gaas/. Unlike methods that rely on microbial marker genes to estimate genome length, the GAAS method can be applied to viruses, which lack a universally common genetic element [19]. GAAS determines community composition and average genome length using a novel BLAST-based approach that maintains all similarities with significant relative alignment lengths, assigns them statistical weights, and normalizes by target genome length to calculate accurate relative abundances. Using GAAS, the community composition and average genome length for over 150 viral and microbial metagenomes was derived from four different biomes, including the Sargasso Sea virome previously described in Angly et al. [1]. The average genome lengths were used in a meta-analysis to determine how genome length varies at three levels: between biomes (e.g. terrestrial versus aquatic), between related sub-biomes (e.g. ocean versus freshwater), and between microbial and viral communities sampled from the same environment.
GAAS provided more accurate estimates of average genome length and community composition than standard BLAST searches (i.e. no length normalization, no relative alignment length filtering, top BLAST similarity only) (Figure 1). The accuracy of GAAS estimates was benchmarked using artificial viral metagenomes. To simulate environmental metagenomes, 80% of species were treated as unknowns and viral communities were created with either power law or uniform rank-abundance structures. The error for power law metagenomes was consistently higher than for the uniform case (data not shown). Significance of BLAST similarities was determined using relative alignment length and percentage of similarity in addition to an E-value cutoff. The accuracy of GAAS was dramatically increased by normalizing for genome length; average errors decreased significantly for community composition (p<0.001, Mann-Whitney U test), as well as genome length (p<0.001, Mann-Whitney U test) (Figure 1 A, B). Metagenomes consist of sequence fragments derived from the available genomes in an environment [20]. Even if two genomes are present in equal abundances, a larger genome has a higher probability of being sampled because it will produce more fragments of a given size per genome (Figure S1). Length normalization in GAAS corrected for this sampling bias inherent to the construction of random shotgun libraries such as metagenomes. Using all similarities weighted proportionally to their E-values further reduced errors in composition. This reduction was significant in comparison to average error when only the top BLAST similarity was used (p<0.001, Mann-Whitney U test) (Figure 1 C). When no species were treated as unknown, the error on the GAAS estimates decreased dramatically (Figure S2). GAAS performed well in benchmarks using artificial microbial metagenomes obtained from JGI (Figure S3). Figure S4 shows that it is harder to distinguish between closely related strains than unrelated species using local similarities: the error on the relative abundance estimates is higher than for more distantly related microorganisms (Figure S3). However, GAAS improves both estimates of relative abundance and average genome length, from ∼2% relative error for the average genome size when keeping only the top similarity to ∼0.2% using all similarities and weighting them (Figure S4).
Variations in metagenomic read lengths did not affect the accuracy of GAAS relative genome length estimates (Figure 2, Figure S5, Figure S6). GAAS was benchmarked on simulated viral metagenomes containing 50, 100, 200, 400, or 800 base pair sequences. Read length had no effect on the accuracy of average genome length estimates (p = 0.408, Kruskal-Wallis test). Average errors in composition increased significantly (p<0.001, Kruskal-Wallis test) with increasing read length, but there was only a very weak positive correlation between increased errors and longer reads (tau = 0.07, p<0.001). The accuracy of GAAS estimates was thus not very susceptible to changes in read length on average. This contrasts with a report on the inappropriateness of short reads for characterizing environmental communities, mainly on the basis that they miss more distant homologies than longer sequences [21]. In addition, the longest reads tested here (800 bp) achieved both the lowest and highest error on the relative abundance estimates (Figure S5). This indicates that the choice of appropriate filtering parameters is more important for longer sequences than for short sequences. In summary, GAAS can be used to accurately and effectively estimate both composition and average genome length for sequences from a variety of available technologies: very short (∼50 bp) sequences obtained by reversible chain termination sequencing (e.g. Solexa), mid-size sequences produced by Roche 454 pyrosequencing (∼100–400 bp), and long 700+ bp reads sequenced by synthetic chain-terminator chemistry (Sanger).
Re-analysis of the Sargasso Sea virome using GAAS revealed that small ssDNA phages were more important than previously assessed, representing ∼80% of the viral community (Figure 3). Community composition and average genome size for the Sargasso Sea virome were calculated using both the GAAS method and the standard method (no length normalization, top similarities only) for comparison. Both the pie charts and length spectra in Figure 3 were generated directly by GAAS. Using the standard method, the Sargasso Sea viral community was dominated by Prochlorococcus phages (64%), with lesser abundances of Chlamydia phages (15%), Synechococcus phages (12%), Bdellovibrio phages (3%) and Acanthocystis chlorella viruses (2%). In contrast, using GAAS, Chlamydia phages were the most abundant organism (79%), whereas Prochlorococcus phages only comprised 16% of the community. The presence of Chlamydia phages in the Sargasso Sea was previously verified experimentally using molecular methods [1]. In contrast to the standard method, the GAAS method also indicated very low relative abundances (<1%) of Synechococcus phages and Chlorella viruses, which have larger genomes.
Most of the variations in community composition estimates were explained by differences in viral genome lengths (Figure 3, right panel). The corrected relative abundance estimates provided by GAAS indicated that species with larger genomes were less abundant than previously thought, and that normalizing by genome length was essential for accurate estimation of community composition (as shown in benchmark tests, Figure 1). A lack of normalization could lead to poor and possibly misleading community composition estimates, as our results have shown, since relative abundance does not equal percentage of similarities.
Phages with small genomes (20–40 kb) are believed to be the most abundant oceanic viruses [11]. In the re-analysis of the Sargasso Sea metagenome, GAAS estimated that 80% of the viral particles were Microviridae (mainly Chlamydia phages), viruses with a genome size smaller than 10 kb. Multiple Displacement Amplification (MDA) was used during the preparation of the Sargasso Sea virome and could have led to over-representation of this viral family. Despite this potential bias, the Chlamydia phage content of this virome was still higher than in all viromes prepared with MDA (except for the stromatolite viromes [6]) (data not shown). In addition, diverse marine circovirus-like genomes, with a length of less than 3 kb, have also been reported in the Sargasso Sea [22], suggesting that small single-stranded viruses play important roles in this marine habitat.
Both microbial and viral average genome lengths calculated by GAAS were significantly different between marine, terrestrial, and host-associated biomes (Figure 4A, Table S1, Table S2). Of the 169 metagenomes analyzed, 146 had a sufficient number of similarities for estimation of average genome length. The average for genome length across all aquatic viral metagenomes was consistent with the previous estimate of 50 kb for marine systems using PFGE by Steward et al. [9]. Host-associated and aquatic viromes had average genome lengths spanning a wide range, from 4.4 to 51.2 kb and from 4.6 to 267.9 kb respectively. Viral average genome lengths were significantly smaller in host-associated metagenomes than in aquatic systems (p = 0.002, Mann-Whitney U test). Estimates of microbial average genome length for aquatic and terrestrial biomes were similar to those predicted using the Effective Genome Size (EGS) method [15], a computational technique based on finding conserved bacterial and archaeal markers in metagenomic sequences. Aquatic microbiomes also showed large variation in average genome sizes, ranging from 1.5 to 5.5 Mb for Bacteria and Archaea and from 0.7 to 25.7 Mb for protists. Microbial average genome lengths in the terrestrial biome were significantly higher than in the host-associated and aquatic biomes (p<0.0001, Mann-Whitney U test). Genome lengths of Bacteria and Archaea from soil environments have previously been shown to be larger than those observed in other biomes [15]. A larger genome is characteristic of the copiotroph lifestyle [23] as it provides microbes a selective advantage in the complex soil environment where scarce but diverse resources are available [24].
Microbial and viral average genome lengths were also significantly different between aquatic sub-biomes. Aquatic metagenomes were grouped into five categories (ocean, freshwater, hypersaline, microbialites, and hot springs) to determine if the variation in average genome lengths could be accounted for by the influence of distinct sub-biomes (Figure 4B, Table S1, Table S2). Other biomes did not include enough metagenomes from different sub-biomes to allow for meaningful classification and analysis. While average genome lengths still varied over a range of values in sub-biomes, the variability was much lower than in the aquatic biome as a whole (Table S1). The average genome sizes in oceanic viromes varied from 20 to 163 kb, well within the range described in [17]. In hypersaline metagenomes, the average genome length varied from 51 to 263 kb, which is comparable to viral genome sizes detected in ponds of similar salinities [16]. A number of average genome lengths were significantly different between sub-biomes for both viruses and microbes (Figure 4B). The stromatolite metagenomes had an average genome length which was significantly different from the oceanic and hypersaline sub-biomes (p<0.05, Mann-Whitney U test), but not from freshwater systems. Oceanic and hypersaline environments were not significantly different. In comparison with the biome level (Figure 4A), the range of average genome lengths at the sub-biome level was reduced (Figure 4B). This suggests that differences in average genome lengths may be driven by environmental factors at a more specific level (e.g. the sub-biome) than what can be encompassed by general biome classifications. Previous work has demonstrated that both metabolic profiles and dinucleotide composition vary at the sub-biome level, and significant differences between both composition and metabolic functions have been reported for marine (ocean), hypersaline, microbialite, and freshwater environments [7],[25].
Microbial and viral average genome lengths varied independently of each other across biomes and aquatic sub-biomes, and reflected differences in the way microbial and viral consortia react to stressors and environmental conditions (Figure 5). Using GAAS estimates for average genome lengths, we compared 25 pairs of viral and microbial metagenomes sampled from the same environment at the same time point. Viral and microbial community compositions have been shown previously to co-vary [26], however, there was no consistent trend between microbial and viral average genome length across all biomes (Kendall's tau = −0.21, p = 0.10).
Most viromes in this analysis were obtained by the collection of viral particles small enough to pass through 0.22 µm pore size filters. The four viral metagenomes collected using 0.45 µm filters [27] had a larger viral average genome length (in light blue in Figure 5). These data show that large viruses may be omitted when sampling with 0.22 µm filters and the capsid size of DNA viruses is likely positively correlated with their genome length. Sampling biases, however, do not account for the independence of viral and microbial length reported here.
Paired metagenomes from oceanic and hypersaline aquatic sub-biomes were characterized by small fluctuations in viral genome lengths coupled with large variations in microbial genome lengths. The four paired ocean metagenomes (Figure 5, light blue squares) were taken from waters surrounding coral atolls in the Northern Line Islands [27]. Microbial communities changed dramatically along a gradient of human disturbance, with populations of pathogens and heterotrophic microbes increasing with human activity [27], which could have resulted in large differences in average microbial genome lengths between atolls. Across all four atolls, viral communities were dynamic but dominated in general by Synechococcus and Prochlorococcus phage, according to both the original [27] and the GAAS analysis (not shown). The large genome of these widespread phages resulted in a less variable viral average genome length. In hypersaline metagenomes (Figure 5, blue diamonds), a similar trend of low variation in viral genome lengths coupled with larger ranges of microbial genome lengths was observed. This corresponded to known differences in the ranges of genome lengths of dominant halophilic viruses and microbes. The most abundant viruses in hypersaline systems have genome lengths between 32 and 63 kb, while predominant Halobacteria have genome lengths varying across a larger range, from 2.6 to 4.3 Mb [28],[29].
The relationship between viral and microbial average genome lengths in manipulated coral metagenomes reflected differences in how viral and microbial consortia reacted to stress (Figure 5, yellow triangles). Five of the six manipulated metagenome pairs used in this analysis were metagenomes from Porites compressa corals subjected to a variety of stressors [30],[31]. Nutrient, DOC, temperature, and pH stress all resulted in an increased abundance of large herpes-like viruses over the control, which could lead to increased average viral genome lengths overall [30]. However, shifts in the microbial consortia (consisting of Bacteria, Archaea, and eukaryotes) were more variable depending on which stressor was applied [31]. For example, temperature stressed corals showed a dramatic increase in fungal taxa, which could be driving the larger average microbial genome length seen here.
The GAAS software package implements a novel methodology to accurately estimate community composition and average genome length from metagenomes with statistical confidence. GAAS provides the user with both textual and graphical outputs, including genome length spectra, relative abundance pie charts, and relative abundances mapped to phylogenetic trees. GAAS can easily be applied to any database of complete sequences to perform taxonomic or functional annotations, and provides filtering by relative alignment length as a standard for selecting significant similarities regardless of which database is used. Since GAAS controls for sampling bias towards larger genomes and considers all significant BLAST similarities, it has the potential to identify key players in ecosystems that may be ignored by other analyses. For example, the re-analysis of the Sargasso Sea virome indicated that small ssDNA phage were very abundant and may play a previously overlooked role in the oceanic ecosystem. GAAS could also be applied in metagenomic studies of disease outbreaks and epidemics. Many emerging and highly virulent human pathogens are ssRNA viruses with small genomes, which could be missed by standard analysis methods, which do not normalize for genome length. Meta-analysis using GAAS provided insight into how environmental factors may affect average genome lengths in microbial and viral communities and the relationships between them. The lack of covariance between microbial and viral average genome lengths indicates that natural and applied stressors have different effects on microbes and viruses from the same environment.
NCBI RefSeq (ftp://ftp.ncbi.nih.gov/refseq/release) (Release 32, August 31, 2008) was used as the target database for the estimation of taxonomic composition and average genome size. Three databases containing exclusively complete genomic sequences were created from the viral, microbial, and eukaryotic RefSeq files. All incomplete sequences were identified as having descriptions containing words such as “shotgun”, “contig”, “partial”, “end” and “part”, and were removed from the database.
A taxonomy file containing only the taxonomic ID of the sequences in these three databases was produced using the NCBI Taxonomy classification. Sequences with a description matching the following words were excluded from that file unless the chromosomal sequences were also available for the same organism: “plasmid”, “transposon”, “chloroplast”, “plastid”, “mitochondrion”, “apicoplast”, “macronuclear”, “cyanelle” and “kinetoplast”. The complete viral, microbial, and eukaryal sequence files with accompanying taxonomic IDs are available at http://biome.sdsu.edu/gaas/data/.
Similarly to the Interactive Tree Of Life (ITOL) [40] and MetaMapper (http://scums.sdsu.edu/Mapper), GAAS is able to graph the relative abundance of viral, microbial or eukaryotic species on phylogenetic trees such as the Viral Proteomic Tree (VPT) or Tree Of Life (http://itol.embl.de). The Viral Proteomic Tree was constructed using the approach introduced in the Phage Proteomic Tree and extending it to the >3,000 viral sequences present in the NCBI RefSeq viral collection (Edwards, R. A.; unpublished data, 2009).
Simulated metagenomes were created to test the validity and accuracy of the GAAS approach using the free software program Grinder (http://sourceforge.net/projects/biogrinder), which was developed in conjunction with GAAS. Grinder creates metagenomes from genomes present in a user-supplied FASTA file. Users can simulate realistic metagenomes by setting Grinder options such as community structure, read length and sequencing error rate. Over 9,500 simulated metagenomes based on the NCBI RefSeq virus collection were generated using Grinder. The viral database was chosen since its large amount of mosaicism and horizontal gene transfer represents a worst-case scenario. Therefore, benchmark results using the viral database are expected to be valid for higher-order organisms such as Bacteria, Archaea and eukaryotes. The parameters used were a coverage of 0.5 fold, and a sequencing error rate of 1% (0.9% substitutions, 0.1% indels). Half of the simulated metagenomes had a uniform rank-abundance distribution, while the other half followed a power law with model parameter 1.2. Sequence length in the artificial metagenomes was varied from 50 to 800 bp for the analysis of read length effects on GAAS estimates.
For each simulated viral metagenome, GAAS was run repeatedly with different parameter sets (relative alignment length and percentage of identity). The maximum E-value was fixed to 0.001 in order to remove similarities due to chance alone. Each set of variable parameters was tested on a minimum of 1,200 different Grinder-generated metagenomes. All computations were run on an 8-node Intel dual-core Linux cluster.
Due to the limited number of whole genome sequences available, a great majority of the sampled organisms in a metagenome cannot be assigned to a taxonomy. To evaluate the effect of sequences from novel organisms on GAAS estimates, the taxonomy of 80% randomly chosen organisms in the database was made inaccessible to GAAS rendering them “unknown”. A control simulation with 100% known organisms was run for comparison (Figure S2).
The accuracy of GAAS estimates was evaluated by comparing GAAS results to actual community composition and average genome size of the simulated metagenomes. The relative error for average genome size was calculated as , where x and xe are the true and estimated values respectively. For the composition, the cumulative error was calculated as , where ri is the relative error on the relative abundance of the target genome i and n is the total number of sequences in the database.
Because the benchmark results were not normal, non-parametric statistical tests were used for all pairwise (Mann-Whitney U test) and multi-factor comparisons (Friedman test) of average errors. Non-parametric correlations were calculated using Kendall's tau.
GAAS was also tested on the three simulated metagenomes available at IMG/m (http://fames.jgi-psf.org). Parameter setting and data processing were conducted as in viral benchmark experiments. Points on the IMG/m microbial benchmark graphs represent the average of 58 repetitions.
Microbial strains typically have a largely identical genome, with a fraction coding for additional genes and accounting for differences in genome length. An additional simulation was performed to investigate how the presence of closely related genomes influences the accuracy of the GAAS estimates. The 15 Escherichia coli strains present in the NCBI RefSeq database, ranging from 4.64 to 5.57 Mb in genome size, were used to produce ∼4,500 shotgun libraries with Grinder. The parameters used were the same as for the simulated viral metagenomes, but with a coverage of 0.0014 fold (>1,000 sequences). Half of the simulated metagenomes were treated as in the viral benchmark, using the GAAS approach and assuming no unknown species. The other half were treated similarly but taking only the top similarity. Points on the graph of the microbial strain benchmark represent the average of >2,200 repetitions.
The composition and average genome size for 169 metagenomes were calculated using GAAS. Most of these metagenomes were publicly available from the CAMERA [41], NCBI [42], or MG-RAST [43] (Table S2), and a few dozens were viromes and microbiomes newly collected from solar saltern ponds, chicken guts, different soils and an oceanic oxygen minimum zone (Protocol S1). The metagenomes used here therefore represent viral, bacterial, archaeal, and protist communities sampled from a diverse array of biomes and were categorized as one of the following: “aquatic”, “terrestrial”, “sediment”, “host-associated”, and “manipulated / perturbed”. The large number of aquatic metagenomes was further subdivided into: “ocean”, “hypersaline”, “freshwater”, “hot spring” and “microbialites”. Sampling, filtering, processing and sequencing methods differed among compiled metagenomes. Table 1 provides a summary of the number of metagenomes from each biome (a list of the complete dataset is presented in detail in Table S2).
For all metagenomes, GAAS was run using a threshold E-value of 0.001, and an alignment relative length of 60%. In addition, for bacterial, archaeal and eukaryotic metagenomes, similarities were calculated using BLASTN with an alignment similarity of 80%. Due to the low number of similarities in viral metagenomes using BLASTN, TBLASTX was used for viruses, with a threshold alignment similarity of 75%. All average genome length estimates produced from less than 100 similarities were discarded to keep results as accurate as possible. Manipulated metagenomes were ultimately not used in the meta-analysis because they do not accurately represent environmental conditions. Statistical pairwise differences between average genome lengths across biomes were assessed using Mann-Whitney U rank-sum tests.
The average genome length and relative abundance results obtained for all metagenomes with our GAAS method were compared to the “standard” analytical approach where: 1) only the top similarity for each metagenomic sequence is kept, 2) there is no filtering by alignment similarity or relative length, and 3) no normalization by genome length is carried out. The virome from the Sargasso Sea was chosen to illustrate in detail the difference between the results obtained with the two methods (Figure 3).
Average genome lengths were calculated for 25 pairs of microbial and viral metagenomes sampled from the same location at the same time. The statistical relationship between viral and microbial average genome length in paired metagenomes was evaluated using Kendall's tau, since lengths were not normally distributed. Regression analysis was performed with Generalized Linear Models (GLM). Interactions between genome lengths and biome classifications were not significant and were not included in final models.
All statistical analyses of the GAAS benchmark results, environmental genome length and genome length correlations described above were performed using the free statistical software package R (http://www.R-project.org/) [44].
|
10.1371/journal.ppat.1004746 | Early Virological and Immunological Events in Asymptomatic Epstein-Barr Virus Infection in African Children | Epstein-Barr virus (EBV) infection often occurs in early childhood and is asymptomatic. However, if delayed until adolescence, primary infection may manifest as acute infectious mononucleosis (AIM), a febrile illness characterised by global CD8+ T-cell lymphocytosis, much of it reflecting a huge expansion of activated EBV-specific CD8+ T-cells. While the events of AIM have been intensely studied, little is known about how these relate to asymptomatic primary infection. Here Gambian children (14–18 months old, an age at which many acquire the virus) were followed for the ensuing six months, monitoring circulating EBV loads, antibody status against virus capsid antigen (VCA) and both total and virus-specific CD8+ T-cell numbers. Many children were IgG anti-VCA-positive and, though no longer IgM-positive, still retained high virus loads comparable to AIM patients and had detectable EBV-specific T-cells, some still expressing activation markers. Virus loads and the frequency/activation status of specific T-cells decreased over time, consistent with resolution of a relatively recent primary infection. Six children with similarly high EBV loads were IgM anti-VCA-positive, indicating very recent infection. In three of these donors with HLA types allowing MHC-tetramer analysis, highly activated EBV-specific T-cells were detectable in the blood with one individual epitope response reaching 15% of all CD8+ T-cells. That response was culled and the cells lost activation markers over time, just as seen in AIM. However, unlike AIM, these events occurred without marked expansion of total CD8+ numbers. Thus asymptomatic EBV infection in children elicits a virus-specific CD8+ T-cell response that can control the infection without over-expansion; conversely, in AIM it appears the CD8 over-expansion, rather than virus load per se, is the cause of disease symptoms.
| Primary infection with EBV, a common human herpesvirus, is typically asymptomatic in childhood but, if occurring in adolescence or later, often presents as AIM. This febrile illness is characterised by high virus loads in the blood and an exaggerated EBV-specific CD8+ T-cell response that pushes total CD8+ T-cell numbers well above normal levels. By contrast, very little is known about the events of asymptomatic primary infection. We therefore studied young Gambian children at an age at which many acquire EBV, monitoring them over six months for evidence of EBV infection by virus load in the blood, virus-specific IgM and IgG antibody status, and virus-specific CD8+ T-cell responses. Focusing on IgM-positive children with very recent EBV infection but no history of symptoms, we found that they carried a virus load equivalent to that seen in AIM patients and also mounted a classical virus-specific CD8+ T-cell response. However, that response, though it could occupy at least 15% of the circulating CD8+ T-cell pool, occurred without the huge global expansion of CD8 numbers seen in AIM. This work reinforces the idea that the host’s exaggerated CD8+ T-cell response, rather than the virus load per se, leads to the symptoms of AIM.
| Epstein-Barr Virus (EBV) is a ubiquitous gamma herpesvirus associated with occasional severe primary infections, several malignancies and significant pathology in immunosuppressed hosts. It does not, however, cause significant morbidity in the majority of those infected. In The Gambia most children are infected during childhood, in contrast to most developed countries where the majority of primary infections occur at a later age, often in adolescence [1,2]. It is estimated that between a quarter and up to three quarters of those infected in adolescence will develop a sometimes-severe disease, AIM [3,4]. Paradoxically, those infected during childhood tend to have minor self-limiting illnesses that often go undetected [5]. It is not fully understood why individuals that contract EBV during childhood are usually asymptomatic and do not develop AIM. Of note, most of the published literature regarding the immunopathogenesis of primary EBV infection is derived from studies of AIM, rather than asymptomatic infections.
Many studies in adults have characterised cellular immune responses ex vivo during AIM, both among CD8+ and to a lesser extent CD4+ T-cell subsets [6–13]. The EBV-specific CD8+ T-cell response is hugely amplified, such that total CD8+ T-cell numbers in the blood may reach five to ten-fold higher than usual. Indeed individual lytic antigen reactivities (typically against epitopes within the immediate early (IE) and some early (E) proteins) can account for up to 40% of the highly expanded CD8+ T-cell population, and individual latent antigen reactivities (typically against epitopes from the EBV nuclear antigen 3A, 3B, 3C family) occupying up to 5%. These CD8+ T-cells display a phenotype consistent with recent antigen stimulation, being perforin-positive with direct ex vivo cytotoxic function [14–16] and express the activation marker CD38 and cell cycling marker Ki-67 [8,10,14,17]. What drives these expansions in AIM is unclear, but factors such as an initial lack of natural killer cell control [18], cross-reactive recognition by clonotypes in pre-existing CD8+ T-cell memory [19], and genetic factors [20–22], including polymorphism of the IL-10 promoter [23], have all been proposed.
Whether the cellular response to early EBV infection in asymptomatic children shows features of disruption similar to those described in AIM has been difficult to investigate, mostly because donors without clinical symptoms can’t be readily identified. However, an understanding of how EBV infection is controlled with minimal immunopathology, i.e. without the development of AIM, is important, as AIM is associated with an increased risk of EBV-associated diseases such as EBV-positive Hodgkin lymphoma and multiple sclerosis [24,25]. Of the few published studies on asymptomatic primary EBV infections, Silins et al identified four adult patients undergoing silent seroconversion within a vaccine trial [26]. Interestingly some of these had high EBV loads yet did not have massive T-cell expansions and, where studied, most did not have the distorted T-cell receptor (TcR) repertoires usually seen in AIM; however, their EBV-specific T-cell response was not studied. As to infections during childhood, an early report suggested that this occurs with a serological picture distinct from AIM and without obvious lymphocytosis [27], while another small study of children aged 20–35 months detected EBNA3A, B and C-specific CD8+ T-cell responses in the blood without addressing issues of viral load or hyper-expansion [28].
A detailed study of the EBV-specific immune response and EBV dynamics in asymptomatically infected children utilising modern immunological and virological tools is lacking. The present work follows a cohort of 114 Gambian children longitudinally over six months, using serology and viral load to determine EBV status. It describes the EBV-specific CD8+ T cell responses in those that have had a relatively recent primary EBV infection without any obvious clinical history, and additionally captures six children undergoing silent seroconversion.
Gambian children of an age, 14–18 months, likely to be undergoing primary EBV infection were recruited for study from an infant vaccination clinic. Blood samples were collected from the children at baseline (visit one), they were vaccinated one week later against Diptheria, Tetanus, Pertussis, Hepatitis B and Haemophilus influenzae B (visit two). A five-millilitre blood sample was collected a week later (visit three) primarily to monitor vaccine responses and a further sample at six months (visit four). Of 120 children screened, six were ineligible due to concurrent illnesses or malnutrition at screening whereas 114 were enrolled, of which 99 remained in the study until completion at six months. The study dropouts did not significantly differ in age, sex, weight, haemoglobin, leucocyte and lymphocyte counts, compared to those who continued to participate (S1 Table).
Initially children were tested for their EBV-VCA antibody status and categorised into one of three groups: non-infected (IgM−IgG−), established infection (IgM−IgG+) or very recently infected (IgM+IgG+/−). At visit one, 71 children had established EBV infection as judged by the presence of VCA-specific antibodies of IgG but not IgM class (Fig. 1). Another four children showed evidence of recent infection, with three having IgM VCA-specific antibodies only and one having both IgM and IgG VCA-specific antibodies. The remaining 39 children appeared to be non-infected, having no detectable VCA-specific antibodies or viral genomes in their peripheral blood mononuclear cells (PBMCs). At visit four (six months later) 17 of these 39 initially EBV non-infected donors had become VCA IgM−IgG+, another two had become VCA IgM+IgG+, 13 remained VCA IgM−IgG−, while seven dropped out of the study. The four initially IgM+ children had now become IgM- and had developed VCA-specific IgG. All children, including those with VCA-specific IgM antibodies, were asymptomatic for classical symptoms of AIM (fever, lymphadenopathy, malaise) prior to recruitment and at subsequent visits, based on maternal history and clinical evaluation. Overall, 62% of children showed serological evidence of being EBV infected at baseline, rising to 86% among those remaining in the study six months later. An analysis of VCA IgG titre in a subset of 25 pairs of samples from children at visit one and four showed no significant difference in titre (p = 0.774, S1A Fig).
Since acquisition of EBV in similar African cohorts begins between six and twelve months after birth [29,30], it is likely that at least some children who were IgG reactive to VCA within the cohort were infected with EBV within the last six months prior to recruitment. To examine for evidence of recent infection among these donors, EBV genome loads in PBMCs were measured by qPCR analysis. Fig. 2 shows viral genome load data from PBMCs collected from 70 IgM−IgG+ donors at baseline, 58 of these donors six months later, and six very recently infected IgM VCA reactive donors (some of whom also had VCA-specific IgG antibodies). Genome load data from Caucasian adolescent patients undergoing primary symptomatic EBV infection, AIM, assessed using the same qPCR assay are also included for comparison.
Almost all of the children who had IgG antibodies to VCA at baseline had high EBV genome levels in their PBMCs, ranging up to two million genomes per million PBMCs. Indeed, for the whole cohort of IgM− IgG+ children sampled at visit 1, the median load of 3000 genomes per million PBMCs (IQR 900 to 8000 genomes per million PBMCs) was not significantly different to that observed in AIM patients. When PBMCs from these same children were assessed for genome loads six months later, a narrower range of values was found and the median load was eight to tenfold lower than at baseline. The decreasing virus loads observed over these two time points suggests that these donors were establishing their virus host balance following recent EBV infection. Comparing virus load to VCA IgG titre in a subset of 25 children at the two time points showed no correlation between load and titre (S1B and C Fig). When the samples from the six IgM+ donors were analysed, these also showed high EBV genome loads with a median value of 8000 genomes per million PBMC, slightly higher than but not significantly different from loads in IgM-IgG+ positive children measured at baseline and again similar to that seen in AIM patients. Such data are consistent with these IgM+ children having been very recently EBV-infected.
Primary symptomatic infection with EBV is associated with dramatic expansions in the frequency and absolute number of lymphocyte subsets, especially CD8+ lymphocytes [6]. Evidence of disruptions to the lymphocyte compartments of the three groups of children (IgM− IgG−, IgM− IgG+ or IgM+ IgG+/−) were studied by determining absolute numbers of lymphocytes within the CD3, CD4, CD8 and CD19 subsets. Fig. 3 shows results of absolute cell counts and, for comparison, counts of equivalent subsets from six Caucasian AIM patients. Dramatic expansions of the CD3+ and CD8+ (but not CD4+) T-cell numbers and a contraction of B cell numbers were seen in AIM patient samples. However, no obvious or significant expansion of lymphocyte subsets was observed when comparing uninfected children with the two EBV-infected groups. Furthermore, no significant changes in the CD4:CD8 ratios were observed in PBMCs from a subset of 14 children over the six month study period (p = 0.76, S2 Fig). This indicated that there was little disruption to peripheral lymphocyte subsets in children at these different stages of asymptomatic EBV infection.
To further understand the dynamics of CD8+ T-cell responses and virus loads at early stages of infection, MHC-class I tetramers were used to assess the frequency of EBV-specific T-cells in the PBMCs of children with VCA-specific IgM− IgG+ antibodies at baseline and six months later and compared to their virus loads at these time points. Children with relevant EBV-specific responses were identified by screening PBMCs from visit one for responses by ELISpot to pools of peptides containing peptide-epitopes known to be presented by HLA types frequent within the Gambian population. This identified 14 children with responses which could be assessed with HLA-A*0201, HLA-B*0801 or HLA-B*3501MHC class I tetramers. Virus loads from this subset of donors were representative of the overall population shown in Fig. 2 at the two time points and their loads were significantly decreased at the second time point (Fig. 4A).
Frequencies of EBV-specific responses in each of the children’s paired visit one and visit four PBMCs samples were obtained by staining with an appropriate tetramer which would identify an immunodominant response. These tetramers presented epitopes derived from the immediate early lytic cycle protein BZLF1 protein, either the HLA-B*0801 presented RAKFKQLL peptide or the B*3501 presented EPLPQGQLTAY peptide, or from the early lytic cycle protein BMLF1, the HLA-A*0201 presented GLCTLVAML peptide. Analysis of the frequency of tetramer-specific CD8+ T-cells in PBMCs from this group is shown in Fig. 4B. This shows that at the first time point EBV-specific responses are made, but their frequency of up to 2.5% of CD8+ T cells is not obviously increased as is seen in AIM patients, in whom previous reports have documented up to 40% of the total CD8+ T-cell population as being EBV-specific (Callan et al. 1998). Analysis of the visit four time point, collected six months later, showed that responses were still present, however they were on average significantly lower than those detected at visit one.
This reduction in EBV-specific T-cell numbers with time, coupled with a falling virus load illustrates a pattern consistent with recent EBV infection and establishment of a long term carrier state in these children. In all but one of these 14 children there were sufficient cell numbers to examine the changing phenotype of the T-cell response over time. Both total CD8+ T-cells and tetramer-positive CD8+ T-cells were examined for markers known to be expressed by CD8+ T-cells responding to acute EBV infection; namely activation status as defined by co-expression of CD38 and HLA-DR, cycling status as defined by the expression of Ki-67, and apoptosis sensitivity as indicated by loss of Bcl-2, which is down regulated in activated EBV-specific cells in AIM patients. Of note CD38 and HLA-DR co-staining for activation status was used as lymphocytes from young children may constitutively express CD38 which is progressively lost with age [31]. The graphs on the left hand side of Fig. 5 summarises the results of this analysis while the flow cytometry plots show representative analyses for each stain at the two time points; note that the top flow plots combine data from all CD8+ T cells in black, with the tetramer-positive population in red, while the middle and bottom flow plots show the Ki-67/tetramer and Bcl-2/tetramer profiles. At the first time point there were significantly more HLADR+CD38+ (p = 0.002) and Ki-67+ (p = 0.01) EBV-specific CD8+ T-cells compared to the total CD8+ T-cell population, while tetramer positive CD8+ T-cells expressed lower levels of the anti-apoptotic marker Bcl-2 (p <0.0001). The percentage of HLADR+CD38+ EBV-specific CD8+ T-cells in the IgM−IgG+ children declined significantly over time (p = 0.003), whereas cellular expression of Bcl-2 significantly increased (p = 0.013), with a non-significant decline seen in frequency of cells expressing Ki-67. Again, the phenotypic analysis of these VCA antibody IgG+ donors suggests recent EBV infection and the establishment of a long term carrier state.
To get the clearest understanding of the early events in EBV-specific CD8+ T-cell responses in children undergoing asymptomatic infection, the EBV tetramer-specific responses and phenotype of these cells were examined in children with IgM VCA-specific antibodies. Of the six IgM+ children found in this study, three were suitable for study with the tetramer panel for frequency and phenotype analysis at the different time points. However the three other donors did not show expanded numbers of CD8+ T cells (Fig. 3) and showed, at most, small changes in the frequencies of CD8+ T cells from 40.9%, 36.8% and 50% at visit one when they were IgM+, to 27.3%, 32.5% and 50.1% at visit four respectively when they were IgG+.
Of the samples from children which could be analysed with tetramers, 082 and 007 were seronegative at the first time point, but had developed IgM+ VCA-specific antibodies at six months. Although there was no significant increase in the absolute numbers of CD8+ T-cells before and after EBV acquisition, both donors 082 and 007 showed increases in the frequencies of CD3+ CD8+ lymphocytes from baseline values of 18.2% and 19.8% to values of 50.3% and 40.8% respectively six months later. As shown in Fig. 6, tetramer analysis of samples at baseline when the donors 082 and 007 were VCA seronegative showed no tetramer-specific staining. However at six months, when the children were VCA IgM+ IgG+, the HLA-B*0801 donor 082 made a substantial response with 6.9% of their CD8+ T-cells being specific to the RAK-epitope while the HLA*0201 donor 007 made a small response of 0.43% CD8+ T-cells to the GLC-epitope. In both cases, tetramer positive cells were highly activated with the majority of EBV-specific cells co-expressing HLA-DR and CD38. In both of these donors a substantial frequency of EBV-specific cells were in cycle; interestingly in the case of child 007 a large proportion of non-tetramer-specific CD8+ T-cells were also in cycle, likely representing other EBV-specificities, possible bystander activation or coincident infection with another pathogen. In the EBV-specific CD8+ T-cells from these children there was little if any expression of Bcl-2 (Fig. 6).
The third child studied, the HLA-B*0801 donor 061, had VCA-specific IgM+ antibodies at the first time point and made a substantial response to the RAK epitope which allowed serial monitoring of this response (Fig. 7). The frequency of PBMC CD8+ T cells did not change over time with levels being 38% at visit one, 37% at three two weeks later, and 37% at visit four with no significant expansion of absolute numbers (see Fig. 1). However, as shown in Fig. 7, tetramer analysis at the first time point detected a large RAK-specific response of 15.9% of CD8+ T-cells, with 32.2% of these being activated, 14.8% in cycle and few expressing Bcl-2. After two weeks the tetramer-specific frequency had decreased to 5.4% of CD8+ T-cells with no associated decrease in activation marker and Ki-67 expression, nor increase in Bcl-2 expression by the tetramer-specific cells. By six months, the frequency of RAK-specific cells had decreased to 1.75% of CD8+ T-cells. Here the RAK-specific cells phenotypically resembled the EBV-specific cells from the VCA-IgG+ donors at visit 4, with only 5% expressing activation markers, 5% of cells in cycle and 15% expressing Bcl-2 (Fig. 7). Overall these findings suggest that in asymptomatic primary EBV infection, the frequency of activated EBV-specific cells in the CD8 population can be substantial but this occurs without significant expansion of the CD8 compartment as a whole.
Current understanding of the immunological changes seen during primary EBV infection is almost exclusively derived from studies of AIM in adolescents and adults. These may not apply to the situation in asymptomatic primary EBV in early childhood, when a large proportion of infections occur. In this study, virological and immunological parameters of African children, studied during a time period when they undergo primary asymptomatic EBV infection, were examined to better understand the pathogenesis of primary asymptomatic EBV infection. Children who had EBV infection established for at least some months, as judged by the presence of VCA IgG antibodies but not IgM, have high virus loads, comparable to AIM patients, and their EBV-specific CD8+ T-cells show evidence of recent activation. These loads dropped significantly when tested six months later but were still elevated when compared to loads in other populations such as UK carriers, but were similar to those detected in healthy older children from The Gambia [32]. This suggests that in this situation, the virus set point is established several months after primary infection and that it is higher than in EBV-infected carriers in the UK.
What determines the high virus set point detected in these children is unclear. One factor known to increase EBV loads is exposure to malaria, with children living in malaria endemic regions have higher EBV loads compared to those living in areas of sporadic transmission [32,33]. How EBV loads in children in the present study relate to loads in other African childhood cohorts is difficult to determine due to differences in assays and sample sources used to quantify EBV loads [33,34]. Malaria infection is thought to increase EBV loads either through promoting B cell proliferation [35] or altering T cell responses [36]. However the children in this study showed no evidence of acute malaria and data at the time of the study demonstrated a low prevalence of malaria in The Gambia [37,38]. Others have suggested virus load may be related to the age at infection, with African infants infected shortly after loss of maternal antibodies having higher virus loads than those infected later [34]. In this context, comparing virus loads to western populations should be interpreted with some caution as the timing of infection in these latter populations is relatively delayed and is particularly dependent on the ethnic group studied [39].
A clearer picture of the early immune response dynamics comes from studying children with IgM+ VCA-specific antibodies, who are likely to have had very recent EBV infection. Although these children had high virus loads in their PBMCs, akin to those seen in studies of symptomatic AIM [2,3,40], they showed no obvious physical signs of infection. There was no significant expansion of lymphocytes or the CD3+ or CD8+ T-cell compartment; a finding consistent with those of others who have studied lymphocyte compartments in primary asymptomatic EBV infections [1,26,27]. Nevertheless, there was an ongoing virus-specific CD8+ T cell response in the IgM+ children with, in one case, greater than 15% of CD8+ T cells directed against a single dominant EBV epitope. This increased frequency of EBV-specific T cells found in the periphery in the presence of high virus loads may call in to question their role in control of the infection. However we have previously found that in AIM patients EBV-specific T cells lack expression of lymphoid homing receptors such as CCR7 and CD62L which allow access to tissues such as the tonsil. Virus replication and transformation of B lymphocytes, in AIM patients at least, occurs in this tissue and appears poorly controlled by the inefficient recruitment of EBV-specific T cells to this site [10]. As activated EBV-specific T cells express low levels of CCR7 and CD62L, we propose that the activated EBV-specific T cells in the children are similarly inefficiently recruited to the tonsil and virus replication and transformation poorly controlled at this site, allowing higher loads of virus to be detected in the presence of these strong responses.
The phenotype of the antigen-specific CD8+ T-cells from the IgM+ VCA antibody children was consistent with what has been described in AIM donors, being highly activated (HLADR+CD38+), in cycle (Ki-67+) and pro-apoptotic (low Bcl-2 expression) [8,10,14,17,41,42]. Why these children do not develop the expanded numbers of CD8+ T-cells observed in AIM patients after EBV infection is not clear. Possible reasons for the AIM associated hyper-expansion have included the development of heterologous immunity where an existing response to an epitope coded by a previously encountered pathogen cross reacts with another from EBV, amplifying the pool of T-cells responsive to EBV challenge and potentially inducing an exaggerated response. Children with less antigenic exposure would have a more limited repertoire of T-cells capable of responding in this heterologous manner [19]. Alternatively two recent studies comparing the incidence of AIM in monozygotic compared to dizygotic twins, and first-, second- or third-degree relatives have shown concordance in the development of AIM, suggesting that there may be a genetic component underlying disease development [20,22]. Perhaps more importantly, an emerging concept in the control of early EBV infection comes from studies using immunodeficient mice reconstituted with human haematopoietic cells, which repopulate human NK and T-cell repertoires. Depletion of NK cells in this model followed by challenge with EBV recapitulates an AIM-like response including splenomegaly, increased plasma levels of the pro-inflammatory cytokine IFN-γ and increased CD8+ T-cell numbers and frequencies [18]. Currently there is a lack of clarity in the literature as to the dynamics and role of NK cell responses in AIM and so the immediate relevance of these experimental findings to natural infection remain to be resolved. Some studies have indicated that there are inverse relationships between NK cell numbers and both severity of symptoms and virus load [43], while others have shown a positive correlation between NK cell numbers both with virus load and severity of symptoms in AIM patients [3]. However NK cells are a heterogeneous population and so a key question arising from these studies is whether there is a difference in subsets of NK cells in terms of numbers or function between individuals who go on to develop AIM compared to children or others who don’t develop this disease.
It is important to recognise that the immune system of children is different in comparison to adolescents where AIM is typically seen. Neonates are born with high levels of the immunosuppressive cytokine IL-10, high levels of plasma immunosupressive factors such as adenosine, and have Th2 and Th17 skewed immunity, all of which decline by 1–2 years of age to near adult levels [44]. This “infant-adapted” immune profile evolves in the first few years of life, with a gradual increase in Th1 capacity, maturing of B cell and antibody responses, and development of T and B cell immunological memory [44]. By the time infants were recruited into the study at 14–18 months of age, they would have a considerably matured immune system compared to birth but some differences would persist compared to adolescents. Such immunological differences could have contributed to the asymptomatic primary infection that is seen in the children compared to adolescents.
Throughout this study we have used the VCA serological status as a guide to when infection occurred, consequently the precise timing of the primary EBV infection cannot be determined. However, from previous data on infants at the same study site, only 18% of infants were found to be EBV infected at nine months of age [29] and one can, therefore, assume that the majority of children would have been infected six months to one year prior to recruitment. Furthermore, the IgM+IgG± donors were likely infected within the last 120 days and may be at a different stage of infection. Secondly, ensuring children are truly asymptomatic in this setting can be challenging as reliance on maternal perception may not be reliable and careful clinical studies of infants or children have described AIM like symptoms in some instances [45]. To combat this, the children all underwent a health screen by the study clinician, including baseline clinical observations such as weight, height, temperature and heart rate and where indicated a rapid malaria test. Recent work by Balfour et al. has suggested that 89% of primary EBV infections in a cohort of University students display some symptoms, although they may not fulfil all classic criteria for AIM, which is different to our observations in Gambian children [3].
In summary, this study supports the notion that AIM is an immunopathological disease and that symptoms are caused by the significantly expanded CD8+ T-cell responses to the virus. It provides clear evidence that during primary asymptomatic infection EBV-specific responses are indeed activated and can occupy a significant percentage of the circulating CD8+ T cell pool. However these responses appear able to contain the infection without the massive expansion that characterises AIM. Conversely then, the symptoms of AIM appear to derive from the absolute CD8 expansion rather than from the virus infection per se.
The Gambian Government/ MRC Laboratories Joint Ethics Committee approved this study. Participants were enrolled after individual written informed consent was obtained from the participant's parent/guardian. SCC 1206.
This study was conducted in a peri-urban Medical Research Council (MRC) UK clinic, Sukuta, situated within the Government Sukuta Health Centre, serving a low-income population living in crowded conditions. A cohort of 120 children aged between 14 and 18 months were screened when they attended the local government health centre for their routine booster vaccination of diphtheria, tetanus, whole cell pertussis (DTwP) combined vaccine. All children were screened by enquiring about a maternal history of recent illness (e.g. fever) and a clinician examination for signs and symptoms of infectious mononucleosis, including weight and baseline observations (temperature, heart rate, weight and length). Any child found to be unwell (observations outside normal clinical range for age or maternal report of recent illness) or had a weight below that specified on the local Infant Welfare Card Growth Chart were not recruited into the study. (n = 6). Of the children not recruited, none of these showed clinical features suggestive of infectious mononucleosis. Five millilitres of blood was collected from each child into vacutainers containing EDTA (BD). A 500μl aliquot was removed and used to obtain a full blood count on each child using a M-series M16/M20 Haematology Analyser (Medonic, Sweden). A further 250μl aliquot was removed and whole blood flow cytometric staining performed. The remaining blood was layered on to 4mls of Lymphoprep (Axis-Shield, UK) in 15ml Leucosep tubes (Greiner Bio-One, UK). Following centrifugation, the plasma layer was removed, and cryopreserved in 2ml aliquots and stored at −70°C for downstream serology. The lymphocyte interphase was harvested and washed. Cells were counted and re-suspended in freezing medium (FCS (Sigma-Aldrich) supplemented with 10% (v/v) dimethyl sulfoxide (DMSO)) at approximately 5 x 106/ml. Children were brought back one week later to receive the Pentavalent vaccination (DTwP, Hep B, Hib) (Easy Five Panacea Biotec). They were subsequently invited to return a week after vaccination and again at six months to undergo further blood sampling.
AIM patients were recruited from a cohort of young adults (18–25 years old) collected at the University of Birmingham, UK. All patients gave written informed consent to donate samples and experiments were approved by the South Birmingham Local Research Ethics Committee (reference number 07/Q2702/24). Patients were defined as AIM by having tonsillitis/sore throat, high lymphocyte counts and being heterophile antibody positive. Mononuclear cells were harvested from blood specimens and stored as described above.
EBV genome loads were assayed by quantitative real-time PCR, as described elsewhere [46]. DNA extraction was performed from 1x106 PBMCs using QIAmp DNA Blood Mini kit (Qiagen).
IgG and IgM reactivity to EBV Viral Capsid Antigen (VCA) were measured using a previously described in-house immunofluorescence assay at the Institute for Cancer Studies, Birmingham [47,48] and the MRC-University of Glasgow Centre for Virus Research, University of Glasgow [49,50] respectively.
For children a 100 μl volume of whole blood for each donor was stained with antibodies to the following surface markers: CD3 PE, CD4 PerCP, (BD Biosciences), CD8 efluor450 and CD27 APCalexafluor750 (Ebioscience) for 30 min at 4°C. Red blood cells were then lysed using 1:10 FACS Lysing Solution (BD Biosciences) and incubated for 10 min at room temperature. Cells were then washed twice in FACS buffer (PBS, 5% BSA, 5% EDTA) and re-suspended in Cytofix (BD, Biosciences). Samples were acquired on a Cyan ADP flow cytometer using Summit software (Beckman Coulter) at MRC Gambia.
Lymphocyte subsets from AIM patients were identified by staining with antibodies specific to: CD19 FITC, CD4 PE (Biolegend), CD27 APC elfluor 780, CD3 efluor 450 (eBioscience) and CD8 qDot 655 (Invitrogen). Samples were stained for 30 min on ice, washed and analysed immediately on an LSR-II flow cytometer (BD Biosciences). Data was analysed using Flow-Jo software (Treestar Inc).
Tetramers were used to identify and analyse the surface marker phenotype of epitope-specific CD8+ T-cells. From the aforementioned IFN-γ ELISPOT data we selected the following epitopes, B*0801 RAKFKQLL, B*3501 EPLPQGQLTAY and A*0201 GLCTLVAML for tetramer manufacture, as they were frequent targets of the immune response. Markers of activation (CD38 & HLDR), proliferation (Ki-67) and the anti-apoptotic marker, Bcl-2 were assessed. Tetramers were validated for specificity against HLA-matched and mismatched seropositive and seronegative donors. Tetramer staining was performed on cryopreserved PBMCs as described elsewhere. Cells were thawed and stained with LIVE/DEAD fixable Aqua Dead Cell Stain for 30 min at 4°C, washed and stained with 1μg of tetramer-PE for 15min at 37°C. Following two further washes, surface staining with CD3-Qdot655, CD4-Qdot605, CD8-Qdot705, CD14-V500, CD19-V500, CD38-APC and HLA-DR-Alexafluor700 were performed. Following fixing and permeabilisation as described above, intracellular staining with Ki67-Alexafluor488 and Bcl-2 (B-cell lymphoma 2)-V450 was performed. Fluorescence minus one samples were included to aid gating during subsequent flow cytometric analysis. A comparison of expression of the above phenotypic markers on EBV-specific and the total CD8+ T-cell populations were performed.
Compensation for fluorescence ‘spill-over’ was performed using the BD CompBead Anti-Mouse Ig set (BD Biosciences) and the antibodies described above. Briefly, antibodies were added to separate tubes containing one drop each of Anti-Mouse Ig beads and the negative control beads (which do not bind κ light chain-bearing immunoglobulin). Following a 30 min incubation at 4°C, beads were washed and re-suspended in FACS buffer.
All statistical analyses were performed using Graphpad Prism version 5.0 for Macintosh (GraphPad Software, San Diego California, USA, www.graphpad.com). Comparisons between variables were performed using the Mann-Whitney U test (two-tailed) and for non-parametrically distributed data, the Wilcoxon matched pairs test (for comparisons between total CD8+ and virus-specific CD8+ T-cells made within individuals) was used. Correlations between non-normally distributed data were made using the Spearman’s rank correlation coefficient.
|
10.1371/journal.pgen.1003767 | Reversible and Rapid Transfer-RNA Deactivation as a Mechanism of Translational Repression in Stress | Stress-induced changes of gene expression are crucial for survival of eukaryotic cells. Regulation at the level of translation provides the necessary plasticity for immediate changes of cellular activities and protein levels. In this study, we demonstrate that exposure to oxidative stress results in a quick repression of translation by deactivation of the aminoacyl-ends of all transfer-RNA (tRNA). An oxidative-stress activated nuclease, angiogenin, cleaves first within the conserved single-stranded 3′-CCA termini of all tRNAs, thereby blocking their use in translation. This CCA deactivation is reversible and quickly repairable by the CCA-adding enzyme [ATP(CTP):tRNA nucleotidyltransferase]. Through this mechanism the eukaryotic cell dynamically represses and reactivates translation at low metabolic costs.
| Adequate reprogramming of metabolic activities by environmental stress or suboptimal growth conditions is crucial for cell survival. Cells employ a remarkable diversity of processes to maintain its homeostasis at all levels of gene expression, including chromatin remodeling, mRNA expression and degradation, translation and protein degradation. Each of these processes shapes cell response at different time scales. In this study, we analyzed the cellular response to oxidative stress at the level of translation. Translation, as one of the most downstream processes in gene expression, provides the necessary plasticity for immediate changes of cellular activities. Using high-sensitive approaches to probe the structural integrity of cellular tRNAs, we show that upon exposure to oxidative stress tRNAs are rapidly deactivated by a cleavage within their ubiquitous, single-stranded 3′-CCA termini by oxidative stress-activated nuclease, angiogenin. The CCA-ends deactivation is reversible and quickly repairable by a ubiquitous enzyme, CCA-adding enzyme, whose natural function is to attach post-transcriptionally the CCA overhang to the 3′-termini of all tRNAs in an mRNA template-independent manner. We propose that this is a mechanism to dynamically repress and reactivate translation at low metabolic costs.
| Environmental stress or suboptimal growth conditions reduce cell viability and put cells at risk. Cells maintain their internal homeostasis by adequate reprogramming of metabolic activities at all levels of gene expression, including chromatin remodeling, mRNA expression and degradation, translation and protein degradation. Given the considerable time needed to activate new genes and/or de novo synthesize mRNA, the translation of existing mRNAs provides the necessary plasticity for the cell to selectively and rapidly respond to stress [1], [2]. Translation is divided into three distinct phases: initiation, elongation and termination. Translation initiation, as a rate-limiting process, is a major point to reprogram translation in response to stress [3], [4]. A key mechanism to repress translation initiation is the phoshorylation of the alpha-subunit of translation initiation factor 2 (eIF2) by stress-activated kinases [5], [6]. However, a sizeable set of cellular mRNAs are initiated in an eIF2-independent manner, which allows for escaping the global kinase-dependent inhibition of translation initiation [3], [4]. It remains elusive, which alternative mechanisms the cell employs to regulate translation during adverse environmental stress.
Transfer RNAs (tRNAs) enter ribosome-mediated protein biosynthesis in a translationally competent state, which includes post-transcriptional modifications at various positions, including the anticodon loop, and the presence of an intact single-stranded CCA-sequence at the 3′-terminus that is required for amino acid attachment by the corresponding aminoacyl-tRNA-synthetase [7]. The CCA ends are generated and maintained by the CCA-adding enzyme [8]. Some bacteria carry tRNA genes encoding CCA termini, thus the CCA-adding enzyme is primarily involved in repairing damaged CCA ends in these organisms [9]. In contrast, all eukaryotic tRNA genes lack the CCA ends and the role of the CCA-adding enzyme is to attach post-transcriptionally the CCA overhang to the 3′-termini of all tRNAs [8], [10], [11]. The functional repertoire of the CCA-adding enzyme has been expanded by its recently discovered role in the quality control of hypo-modified tRNAs [12].
Mature, translationally competent tRNAs are very stable under normal growth conditions, with a half-life of approximately one to several hours [13]. However, environmental changes dynamically modulate the concentration of the tRNA pool. Some tRNAs are cleaved in the anticodon loop in response to various environmental stress factors (e.g., oxidative stress, heat shock or ultraviolet irradiation) [14], [15], [16], [17]. The endonucleolytic tRNA cleavage is a conserved feature in higher eukaryotes. Thereby, two tRNA-halves (designated 5′- and 3′-tiRNAs) are generated by a ubiquitously expressed enzyme, angiogenin [18]. This cleavage, however, does not significantly reduce the level of mature tRNAs, which implies that tiRNAs may rather act as a signal transducer to modulate translation of specific mRNAs, than to globally repress translation [18]. Furthermore, in response to external stimuli, retrograde translocation of mature tRNAs to the nucleus[19] or selective charging of different tRNA isoacceptors [20] transiently alter the pool of translationally active tRNAs in the cytoplasm. Consequently, these stress-induced alterations in the tRNA concentration will decrease the amount of ternary complex (that is, the complex of charged tRNA with the GTP-loaded elongation factor). However, the primary mechanism, that triggers a general inhibition of translation elongation during stress, remains surprisingly elusive.
Here, using high-sensitive approaches to probe the structural integrity of cellular tRNAs, we show that upon exposure to oxidative stress all tRNAs are rapidly deactivated by a cleavage within their 3′-CCA termini by oxidative stress-activated nuclease, angiogenin. Since 3′-CCA ends are ubiquitous for all tRNAs, angiogenin-induced deactivation of tRNAs provides a means for global repression of translation at the level of elongation. On a much slower scale, at longer times of exposure to stress, some tRNAs are also cut in their anticodon. The CCA ends deactivation is reversible and quickly repairable by the CCA-adding enzyme. We propose that this is a mechanism to dynamically repress and reset translation at low metabolic costs.
Angiogenin, the nuclease that endonucleolytically cleaves tRNAs during oxidative stress, is constitutively expressed, but kept inactive through an inhibitor RNH1 [18]. Oxidative stress dissociates the inhibitor and activates angiogenin [21]. To investigate the susceptibility of the cellular tRNAs to angiogenin-mediated cleavage, we exposed confluent HeLa cells to arsenite which elicits oxidative stress and activates angiogenin. A small amount of tiRNAs was generated, but only at prolonged exposure to arsenite (>30 min) (Figure 1A). Next, we used tRNA microarrays [22] with immobilized oligonucleotide probes complementary to the full-length tRNA sequences to determine the susceptibility of each tRNA species to angiogenin. Only a subset of all tRNAs bearing a CA sequence in the anticodon loop was predominantly cleaved into tiRNAs; a minor fraction with UA or GC motifs in the anticodon loop was also cleaved (Figure S1A). This cleavage pattern mirrors the substrate specificity of angiogenin: it targets single-stranded ribonucleic acid sequences with 10–30-fold higher preference for CA over UA [23] and 3-fold higher for CA over CG [24]. While there is a large variability in the composition of the anticodon loops of all tRNAs and only a fraction of tRNAs possesses a CA-motif in the anticodon loop (Figure S1B), we realized that the ubiquitous, single-stranded 3′-CCA sequence post-transcriptionally attached to all eukaryotic tRNAs [8], bears the strongest angiogenin recognition motif, the CA motif. To investigate whether the 3′-CCA ends can be targeted by angiogenin upon exposure to oxidative agents, we used enzymatic ligation of a fluorescent stem-loop oligonucleotide that complementary pairs only to the intact 3′-CCA end of tRNA (Figure 1B, schematic inset). The fluorescent signal, which is proportional to the amount of tRNAs with intact 3′-CCA ends, decreased noticeably in conditions of severe oxidative stress while the total tRNA amount remained relatively constant (500 µM arsenite; Figure 1B), consistent with the idea that the 3′-CCA ends of tRNAs are primary substrates of angiogenin. Oxidative stress altered the structural integrity of the 3′-CCA termini of tRNAs in a dose-sensitive manner; at lower dose of stress (100 µM arsenite) much smaller fraction of tRNAs than at high stress dose (500 µM arsenite) was unable to ligate the fluorescent oligonucelotide (Figure 1B). To our surprise, the removal of the 3′-CCA ends of the tRNAs occurred on a much faster time scale (Figure 1B) compared to the appearance of the tiRNA fragments (Figure 1A).
To confirm that angiogenin cleaves the 3′-CCA termini of tRNAs upon exposure to arsenite, we upregulated the level of aniogenin and analyzed the integrity of the 3′-CCA ends using the fluorescent oligonucleotide-ligation approach. A small increase in the cellular level of angiogenin led to a noticeable enrichment of tRNAs with cleaved 3′-CCA termini, confirming its role in the oxidative stress-mediated cleavage of the 3′-CCA ends (Figure 1C). Note that angiogenin can be only moderately upregulated for short expression times (Figure S2A); longer expression perturbs the vitality of the cell. The ectopically enhanced levels of angiogenin, a fraction of it might be additionally deactivated by the excess of the RNH1 inhibitor in the cell, is most likely far below the concentration of stress-activated angiogenin, thus the effect of arsenite-induced 3′-CCA end cleavage (+arsenite) was much stronger (Figure 1C).
Intrigued by the different time scales of the stress-induced alterations of cellular tRNAs, we next analyzed the kinetics of 3′-CCA end cleavage and tiRNA generation in vitro. Total tRNA was isolated from confluent, non-stressed HeLa cells, radioactively labeled at their 5′- or the 3′-end and subsequently subjected to angiogenin treatment. Strikingly, while 5′-labeled tRNAs are still visible at 120 min of incubation with angiogenin, the 3′-labeled tRNAs completely disappeared after 30 min (Figure 2A). The fast decay of the signal of the 3′-labeled full length tRNAs than the 5′-labeled tRNAs (Figure 2A) is consistent with a preferred and much faster cleavage in the 3′-CCA termini of the tRNAs. All tRNAs were equally sensitive to angiogenin cleavage at the 3′-CCA termini; the signal for all 3′-radioactively labeled tRNAs decayed almost simultaneously during the angiogenin treatment (Figure 2B and S3). In the fluorescent oligonucleotide-ligation approach (Figure 1B, schematic inset), we observed a clear progressive decrease of the yield of the oligonucleotide ligated to the 3′-CCA end of the tRNAs upon angiogenin treatment (Figure 2C). Notably, fragments migrating at the height of the tiRNAs appeared much later (Figure 2C), suggesting that angiogenin degraded the 3′-CCA ends more rapidly than it cleaved tRNAs in the anticodon loop, thus recapitulating the observations in HeLa cells (Figure 1). To define the exact cleavage site in the 3′-CCA end, we used internally radioactively labeled variants of tRNAPhe(GAA) with intact 3′-CCA and truncated 3′-CC end. tRNAPhe(GAA) lacks the CA motif in the anticodon loop and hence, only the 3′-CCA terminus is susceptible to angiogenin cleavage. Angiogenin cleaved endonucleolytically within the 3′-CCA motif between the C and A nucleotide and removed exclusively the adenosine residue (Figure S4), implying a high CA-dependent endonucleolytic activity of angiogenin.
While in vitro all tRNAs lost their 3′-ends after 10 min, as evidenced by almost complete signal extinction of the 3′-labeled full-length tRNAs (Figure 2A), in vivo the signal plateaued at about 60% of the initial signal intensity (Figure 1B). Translationally competent tRNAs are aminoacylated and complexed with elongation factor EF1α, which may protect tRNAs from angiogenin cleavage. Aminoacylation per se did not influence angiogenin cleavage (Figure S5). The crystal structure of the ternary complex from Thermus aquaticus indicates that EF1α contacts only the phosphate groups of tRNA bases 73–75 [25]. [Note, C75A76 is endonucleolytically targeted by angiogenin]. Thus, the elongation factor EF1α may marginally interfere with the angiogenin binding and partly protect the aminoacyl-tRNA.
In cells, the CCA-adding enzyme repairs the partially degraded 3′-CCA ends of tRNAs without a nucleic acid template and highly discriminates between adding cytidine at position 75 and adenosine at position 76 [8], [10], [11]. Thus, we hypothesized that the lower amount of tRNAs with deactivated CCA termini in HeLa cells (Figure 1B), compared to the in vitro angiogenin treatment (Figure 2A), might represent a steady-state equilibrium between the angiogenin-mediated cleavage and simultaneous repair by the CCA-adding enzyme whose activity remained unchanged upon the arsenite treatment (Figure S4C). To determine the effect of these two opposing processes, total HeLa tRNAs were successively treated with angiogenin and human CCA-adding enzyme. Indeed, the CCA-adding enzyme repaired the CCA termini (Figure 3A) by adding the cleaved adenosine (Figure 3B), implying that stress-damaged 3′-CCA ends of tRNAs can be easily repaired and tRNAs are converted back to translationally-competent species. The angiogenin-catalyzed endonucleolytic cleavage of the CA motif results in a 3′-terminal 2′,3′-cyclic phosphate at the cytosine residue. Thus, prior to treatment with CCA-adding enzyme HeLa tRNAs (Figure 3A) or single tRNAPheCC were treated with T4 polynukleotide kinase (PNK). In the cell, the 3′-end cyclization products are quickly hydrolyzed to 3′OH by 2′,3′-cyclic 3′-phosphodiesterases [26], [27].
An attempt to reduce the cellular concentration of the CCA-adding enzyme was unsuccessful: even though de novo synthesis of the enzyme was significantly inhibited by targeting its mRNA with specific siRNA probe (Figure S2B), the concentration of the mature CCA-adding enzyme remained unchanged (Figure S2C). As the CCA-adding activity is essential for cell viability, an intrinsic robustness of this enzyme has the advantage of maintaining a constant function and permitting a prompt stress response.
The 3′-CCA ends are indispensable for tRNA aminoacylation and subsequently for translation. What is the effect of angiogenin-induced deactivation of the 3′-CCA termini of cellular tRNAs on protein translation? Exposure of HeLa cells to acute oxidative stress (500 µM arsenite) altered the polysomal profile and shut down translation (Figure 4A). Importantly, at low arsenite concentration (100 µM) the cells retained some translation activity, detectable as a considerable polysomal fraction (Figure 4A). The most potent inhibition of translation is mediated by eIF2α phosphorylation upon oxidative stress via haem-regulated inhibitor kinase (HRI), which represses translation of mRNAs with scanning- or cap-dependent translation initiation [3]. By contrast, a sizeable subset of genes are translated through a cap-independent mechanism: internal ribosome-entry sites (IRES) direct translation initiation without the aid of canonical initiation factors and initiator Met-tRNA [28]. We hypothesized that cap-dependent translation will be influenced at much lower arsenite concentrations compared to mRNAs with scanning-independent initiation; the combined effect of oxidative stress on eIF2α phosphorylation and the deactivation of the 3′-CCA ends of all tRNAs will have much higher impact on mRNAs initiated in a scanning-dependent manner. In contrast, in the IRES-initiated translation, as only the 3′-CCA-end inactivation should play a role the effect should be less pronounced. We therefore tested the effect of two arsenite concentrations, representing severe (500 µM) and moderate (100 µM) oxidative stress using bicistronic mRNA encoding renilla luciferase (Rluc), initiated in a cap-controlled manner, and firefly luciferase (Fluc), initiated via cricket paralysis virus IRES (CrPV-IRES) (Figure 4B), an IRES sequence described to confer translation independent of any initiation factor [29]. At a low arsenite concentration (100 µM), the Fluc activity remained at >80%, while Rluc activity progressively decreased, indicating much potent inhibition of cap-dependent initiation compared to IRES-dependent initiation (Figure 4C,D). At a high arsenite concentration (500 µM), however, a similar decrease for both Rluc and Fluc activity was observed, implying that both IRES-dependent and scanning-controlled initiation were equally inhibited (Figure 4C,D). This cannot be attributed to the decrease of mRNA levels, since the mRNA expression levels of the bicistronic construct remained similar upon stress exposure (Figure S6). Variations in the transfection efficiency are not likely; transfection efficiency was equal in all experiments as assessed with fluorescent reporter. This suggests that under acute oxidative stress translation of all mRNAs is globally repressed, while moderate oxidative stress affects more strongly the cap-dependent than the IRES-controlled initiation due to the combined effect on eIF2α phosphorylation and the tRNAs deactivation.
Here, we analyze the effect of oxidative stress on the structural integrity of the cellular tRNAs and define the mechanisms of oxidative stress-induced global repression of translation at the level of elongation. Our observations clearly suggest a sequential order of tRNA deactivation upon exposure to oxidative stress: the 3′-terminal CCA sequence is first targeted, while the deactivation into tRNA halves occurs much later. The first event, the CCA cleavage, is not restricted to specific tRNAs; the CCA ends of all tRNAs can be targeted by angiogenin. At severe oxidative stress (500 µM) all tRNAs are rapidly deactivated which leads to a global repression of translational elongation of both mRNAs with scanning and non-scanning (IRES) controlled initiation. The deactivation of the 3′-CCA ends is a mechanism to reversibly repress translation at very low metabolic costs; the 3′-CCA tRNA ends are quickly repaired by the CCA-adding enzyme [8], [10], [30] and translation is reset. In contrast, cleavage in the anticodon loop proceeds on a much slower timescale and is specific for only a subset of tRNAs, so that tiRNAs with specific primary sequences can be generated. This mirrors the reported specificity of the tiRNAs to either selectively target translation of a defined fraction of mRNAs [14] or trigger formation of stress granules [31]. The role of some tiRNAs to silence specific functions [14] indirectly suggests a separation of the tRNA halves upon a cleavage in the anticodon loop. Regeneration of such tRNAs is more metabolically demanding for the cell, as the cleaved tRNAs can be replaced only through a new transcription cycle. A sizeable fraction of tRNAs with cleaved anticodons may not dissociate into tiRNA halves and undergo a repair by tRNA ligases [32]. In vertebrates, the tRNA-ligase activity is coupled to tRNA splicing and is mainly localized in the nucleus [33], [34]; a cytoplasmic localization, although conceivable given the observation for cytoplasmic mitochondrial surface in yeast and plants [35], [36], has not yet been described. This, in turn, would require a translocation of the cleaved tRNAs into the nucleus. Stress-induced retrograde translocation to the nucleus has been shown for mature tRNAs [19]. If upon cleavage within the anticodon the tRNA structure is maintained nearly to the native one, the retrograde transport into the nucleus would be possible and tRNAs can be repaired by the tRNA ligases.
Arsenite derived oxidative stress has also been shown to induce elevated tRNA misacylation with methionine [37]. However, Met-misacylation was obtained upon 1 µM arsenite which is much lower than the concentrations in the experiments presented here (100 or 500 µM). Furthermore, the duration of the treatment is much longer (4 hours) [37], which exceeds the time of the first response towards oxidative stress – the 5′-CCA-ends cleavage. Met-misacylation has been proposed to potentially serve as a protective mechanism for cell's own proteins against oxidative inactivation [37]. This mechanism is distinct from the 3′-CCA cleavage which is useful to regulate global translation activity.
Our observation for selective translation of transcripts with scanning-independent initiation under moderate oxidative stress (100 µM arsenite) adds another layer to selectively reprogram protein translation under stress at the level of elongation. The inhibition of translation initiation through eIF2α phosphorylation upon oxidative stress is a potent mechanism to repress translation of mRNAs with scanning-controlled initiation [3], [16]. At low doses of arsenite (100 µM arsenite) translation of mRNAs with cap-dependent translation initiation is compromised, while the non-scanning, IRES-dependent translation continues to function to a certain level (Figure 4) as in the cell only a small fraction of the total tRNA pool is with deactivated 3′-CCA ends (Figure 1B). Finally, as many transcripts involved in proliferation, differentiation and apoptosis [16] are initiated in a cap-independent manner, the observed differential inactivation of protein synthesis which allows these mRNAs to bypass the global translational repression and activate the selective stress response [3], [16], [38].
HeLa cells were usually cultured in DMEM with 10% fetal bovine serum and L-Glu (2 mM) to 80–90% confluency. Oxidative stress was exerted by adding 100 or 500 µM sodium arsenite (Fluka) for indicated times. Human angiogenin was cloned in pCDNA3 plasmid (Invitrogen) and transfected in sub-confluent HeLa cells using polyethylenimine (PEI, Polysciences Europe GmbH). After 8 h angiogenin expression was detected with polyclonal antibodies (1∶1000, Santa Cruz Biotechnology). To decrease the expression of CCA-adding enzyme, pSuper plasmid (Oligoengine) bearing shRNA (5′-CCGGCGCAGAGATCTCACTATAAATCTCGAGATTTATAGTGAGATCTCTGCGTTTTTG-3′) that targeted the CCA-adding enzyme mRNA was transfected using polyethylenimine and expressed for 12 h. shRNA with the same, but randomly scrambled sequence was used as a control. Prior to harvesting, an aliquot of cells was additionally exposed to 500 µM sodium arsenite for 1 h. mRNA was quantified by real-time qRT-PCR and the protein level with polyclonal antibodies (1∶200, Santa Cruz Biotechnology) against human CCA-adding enzyme.
Statistical analyses were performed with Fisher's exact test. Differences were considered statistically significant when p<0.05.
A bicistronic construct was created by cloning renilla luciferase (Rluc) gene under the CMV promoter and downstream of it a firefly luciferase (Fluc) gene under the cricket paralysis virus IRES (CrPV-IRES) into pECFP-C1 (Clontech); note CFP was deleted from pECFP-C1 prior to cloning. HeLa cells were transfected with this bicistronic reporter construct using polyethylenimine (PEI, Polysciences Europe GmbH) and expressed for 8 h in total. Prior to harvesting cells were exposed to 100 and 500 µM sodium arsenite for various times and harvested. Luciferase activities were measured using Dual-Luciferase® Reporter Assay System (Promega).
For isolation of non-charged tRNAs, HeLa cells were harvested by mechanical scrapping and total RNA was isolated with TriReagent (Sigma-Aldrich) according to the manufacturer's protocol. For isolation of charged tRNAs, total RNA was isolated under acidic conditions. Briefly, HeLa cells were re-suspended in 0.3 M NaOAc 10 mM EDTA pH 4.5 and extracted two times with acidic phenol. The aqueous phase, containing RNA, was precipitated with one volume isopropanol and washed with 80% ethanol. The total uncharged or charged tRNAs were separated on 10% PAGE gels at 4°C. Bands corresponding to the tRNAs were visualized by UV-shadowing, cut and eluted from the gel overnight at 4°C, for uncharged tRNAs in elution buffer (50 mM potassium acetate, 200 mM potassium chloride, pH 7.0) or for charged tRNAs in acidic elution buffer (0.3 M NaOAc, 10 mM EDTA pH 4.5).
1.5 Mio. HeLa cells were treated for 30 min with 100 or 500 µM arsenite. Ten minutes prior to harvesting, cycloheximide (CHX) to a final concentration of 100 µg/ml was added to the medium. Cells were trypsinized (trypsin solution was also supplemented with CHX) and collected by centrifugation at 232×g for 5 min. The cell pellet was resuspended in 320 µl of ice-cold lysis buffer (10 mM Tris-HCl pH 7.4, 5 mM MgCl2, 100 mM KCl, 1% Triton-X, 100 µg/ml, 2 mM DTT) and cells were sheared with a 26-gauge syringe. After pelleting of the debris at 5000×g for 8 min at 4°C, the supernatant was layered onto 15 to 50% (w/v) sucrose gradient (20 mM HEPES-KOH pH 7.4, 5 mM MgCl2, 100 mM KCl, 100 µg/ml CHX, 2 mM DTT) and centrifuged for 1.5 h at 35,000 rpm in SW 55Ti rotor (Beckman) at 4°C. The gradient was slowly pumped out from the bottom of the tubes and A254 nm was recorded via a flow-through UV spectrophotometer cell (Pharmacia LKB-UV-M II).
One µg of the total mRNA from HeLa cells was treated with DNase I (Fermentas), the cDNA was synthesized with reverse transcriptase using oligo-dT primer (both Fermentas) and quantified using the 2× Fast SYBR® Green Master Mix (Applied Biosystems) and the 7500 Fast Real-Time PCR system (Applied Biosystems). The following primers were used for amplification of the bicistronic Fluc-Rluc construct: forward (5′-GCTGTTTCTGAGGAGCCTTC-3′) and reverse (5′-GCACTCTGATTGACAAATACGATT-3′), and for CCA-adding enzyme: forward (5′-GATCGCAAAAGAGGAGAAAAAC-3′) and reverse (5′-GCATCAGGTTCCCTAGAATC-3′). mRNA expression was normalized to β-actin.
The degree of aminoacylation of isolated HeLa tRNAs was tested using the periodate protection assay described previously [39]. Total tRNA sample was treated with 50 mM sodium periodate, which oxidizes the 3′-ends of uncharged tRNAs, which prevents the ligation of the fluorescent stem-loop DNA/RNA oligonucleotide. tRNAs are resolved on denaturing 10% PAGE and the ligation efficiency serves as a measure for the levels of charged tRNAs.
tRNA probes covering the full-length sequence of 42 cytosplasmic tRNA species with sequences described previously [22] were spotted onto amino-coated slides. The probes for each tRNA are arranged in clusters of six replicates. Radioactively labeled tRNA samples were mixed with 0.17 mg/ml salmon sperm DNA (Invitrogen), 0.17 mg/ml polyA (Sigma-Aldrich) in hybridization buffer (Sigma-Aldrich) and hybridized on the microarrays for 16 h at 60°C. Subsequently, the microarrays were washed three times in 6×SSC at 35°C and once in 2×SSC and 0.2×SSC at 30°C. The composition of the 20×SSC buffer was as follow: 3 M sodium chloride, 300 mM sodium citrate, 0.1% SDS. Radioactivity was detected on a FUJI BAS scanner.
Yeast tRNAPhe with intact CCA ends or tRNAPheCC were generated according to the procedure described in [40]. Radioactive, internally labeled transcripts were synthesized in the presence of 3 µCi 32P-α-ATP.
Total HeLa tRNA was dephosphorylated with calf intestine alkaline phosphatase (Roche) for 30 min at 37°C. The enzyme was removed by phenol/chloroform extraction and the tRNA was precipitated. Radioactive phosphate was incorporated by T4 polynukleotide kinase (USB) and 32P-γ-ATP for 30 min at 37°C. Radioactively labeled RNA was separated on a denaturing 10% PAGE gel, tRNA bands were cut and eluted in the elution buffer (4 h, 25°C).
Total HeLa tRNAs were deacylated for 45 min at 37°C in 0.1 M TrisHCl, pH 9.0 and dephosphorylated with T4 polynucleotide kinase (USB) in the absence of ATP. 3′-CMP was phosphorylated with 32P-γ-ATP (30 min, 37°C) using PNK (Fermentas) and ligated to the RNA with T4 RNA ligase (NEB) by incubation over night at 16°C. Radioactively labeled RNAs were separated on a denaturing 10% PAGE gel, tRNA bands were cut and eluted in the elution buffer (4 h, 25°C).
To prepare the tRNA for subsequent digestions, isolated tRNA was heated at 90°C for 2 min and cooled down at room temperature in 30 mM HEPES 30 mM sodium chloride, pH 7.0 for 3 min. MgCl2 and BSA were added to final concentrations of 2 mM and 0.01%, respectively, and further incubated for 5 min at 37°C. Recombinant human angiogenin (R&D systems) was added to a final concentration of 0.2 or 1 µM to the total HeLa or yeast tRNAPhe and incubated at 37°C for the indicated times. In the radioactive experiments, total non-labeled HeLa tRNA was spiked with radioactive 5′- or 3′-labeled tRNA. The reactions were stopped by extraction with phenol/chloroform or adding gel loading buffer (95% formamide, 0.025% SDS, 0.5 mM EDTA, 0.25% (w/v) bromophenolblue, 0.25% (w/v) xylene cyanol) and shock freezing in liquid nitrogen. To test the integrity of the 3′-CCA ends, a fluorescent stem-loop RNA/DNA oligonucleotide, with a sequence described previously [22], was ligated over night at 16°C with T4 DNA ligase (NEB). Full-length tRNA and tiRNAs were separated on a denaturing 10% PAGE gel.
Human CCA-adding enzyme was purified as described [41]. Total HeLa tRNA and 3′-radioactive labeled yeast tRNAPhe (0.5 µM) were treated with 0.2 µM angiogenin at 37°C for 4 h, dephosphorylated with PNK to convert the 2′,3′-cyclophosphate generated by angiogenin to 3′-OH [40] and subsequently treated with 50 nM human CCA-adding enzyme at 30°C for 30 min in 20 mM HEPES pH 7.6, containing 20 mM KCl, 6 mM MgCl2, 2 mM DTT and 1 mM NTPs.
Oxidative stress was exerted by adding 500 µM sodium arsenite (Fluka) to confluent HeLa cells for indicated times. RNA was isolated using mirVana miRNA Isolation kit (Ambion) and subsequently deacylated in 0.1 M Tris.HCl buffer, pH 9.0 at 37°C for 30 min. Fluorescent stem-loop RNA/DNA oligonucleotide [22] (Figure 1B, schematic inset) was ligated over night at 16°C with T4 DNA ligase (NEB). Ligation efficiency was analyzed by resolving the samples on denaturing 10% PAGE and detected by fluorescence (Fujifilm LAS-4000) or SYBR Green (Invitrogen) staining.
|
10.1371/journal.ppat.1008004 | Fas-associated factor 1 mediates NADPH oxidase-induced reactive oxygen species production and proinflammatory responses in macrophages against Listeria infection | Fas-associated factor 1 is a death-promoting protein that induces apoptosis by interacting with the Fas receptor. Until now, FAF1 was reported to interact potentially with diverse proteins and to function as a negative and/or positive regulator of several cellular possesses. However, the role of FAF1 in defense against bacterial infection remains unclear. Here, we show that FAF1 plays a pivotal role in activating NADPH oxidase in macrophages during Listeria monocytogenes infection. Upon infection by L. monocytogenes, FAF1 interacts with p67phox (an activator of the NADPH oxidase complex), thereby facilitating its stabilization and increasing the activity of NADPH oxidase. Consequently, knockdown or ectopic expression of FAF1 had a marked effect on production of ROS, proinflammatory cytokines, and antibacterial activity, in macrophages upon stimulation of TLR2 or after infection with L. monocytogenes. Consistent with this, FAF1gt/gt mice, which are knocked down in FAF1, showed weaker inflammatory responses than wild-type mice; these weaker responses led to increased replication of L. monocytogenes. Collectively, these findings suggest that FAF1 positively regulates NADPH oxidase-mediated ROS production and antibacterial defenses.
| Phagocytic NADPH oxidase plays a pivotal role in generating reactive oxygen species (ROS) and in defense against bacterial infections such as L. monocytogenes. ROS eliminate phagocytosed bacteria directly and are implicated in transduction of signals that mediate inflammatory responses. Here, we show that the apoptotic protein FAF1 regulates ROS production in macrophages by regulating phagocytic NADPH oxidase activity upon infection by L. monocytogenes. FAF1 interacts directly with and stabilizes p67phox, a regulatory protein of the phagocytic NADPH oxidase complex, to induce ROS production during L. monocytogenes infection. Production of ROS leads to release of proinflammatory cytokines, chemokines and, ultimately, to bacterial clearance. Interestingly, FAF1gt/gt mice deficient in FAF1 expression exhibit weakened inflammatory responses and are thus more vulnerable to bacterial infection than FAF1+/+ mice. This study reveals that FAF1 is a crucial regulator that induces inflammatory responses to bacterial infection via ROS production.
| Innate immune cells are the first barrier encountered by invading microbial pathogens. Among these cells, phagocytes such as macrophages and neutrophils play key roles in host protection against bacterial infection. Upon recognition and phagocytosis of bacteria, phagocytes produce reactive oxygen species (ROS) that kill and inactivate bacteria directly. This mechanism is known as the respiratory burst. NADPH oxidase, one of several ROS sources, is critical for this process [1, 2]. The redox center of phagocytic NADPH oxidase is a heterodimer comprising transmembrane-associated protein subunits p22phox and gp91phox (Nox2). This heterodimer, also known as flavocytochrome b588, forms a phagocytic NADPH oxidase complex together with the cytosolic regulatory subunits p67phox, p47phox, p40phox, and the small GTPase Rac [1, 3, 4]. The ROS generated by the NADPH oxidase complex are not only toxic to the cell but also participate in host defense responses such as NF-kB activation and release of proinflammatory cytokines [5]. A life-threatening genetic disorder called chronic granulomatous disease (CGD), in which the phagocytic NADPH oxidase is dysfunctional, leads to life-threatening bacterial and fungal infections. CGD is caused by mutations in any one of the genes that encode subunits of the phagocytic NADPH oxidase complex [6–8].
Upon phagocytosis of bacteria, toll-like receptors (TLRs), which are transmembrane receptors that play a critical role in innate immune recognition of pathogens, act as a first line of host defense [9, 10]. The TLR family in humans and mice includes more than ten different members, all of which have been studied extensively with respect to infection. Phagocytes express all TLR members, stimulation of which induces diverse biological processes, including inflammation, antigen presentation, and direct bactericidal effects [10, 11]. The interplay between these TLR recognition and activation of NADPH oxidase during phagocytosis of bacteria is well characterized. Especially, interaction between Nox2 and TLR2 is required for ROS production and inflammatory responses during mycobacteria infection [12–14]. Moreover, TLR2 mediates expression of Nox2 in microglia during peripheral nerve injury [15]. Nox4 is also required for TLR4-mediated ROS production in response to lipopolysaccharide (LPS) [16]. Except for bacterial infection, Nox4 is necessary for generation of macrophage migration inhibitory factor during parasite infection [17]. Recent studies show that the TLR4-Nox1 redox signaling axis plays a role in metastasis of colon cancer and lung cancer cells [18, 19].
Fas-associated factor 1 (FAF1) was identified initially in a yeast two-hybrid assay using the cytoplasmic domain of Fas protein as bait [20]. FAF1, which contains a Fas-interacting domain (FID), a death effector domain-interacting domain (DEDID), and a C-terminal domain [21] potentiates Fas-mediated apoptosis as a member of the death-inducing signaling complex [22, 23]. FAF1 interacts with different molecules and is involved in a variety of biological processes; it plays a role in regulating cell death and/or tumor progression, ubiquitination-mediated proteosomal degradation, chaperones, NF-kB signaling, and interferon signaling [24–29].
To better understand the biological role of FAF1, we examined the relationship between NADPH oxidase and FAF1 in host defense against bacterial infection. We show that FAF1 is a crucial positive regulator of the phagocytic NADPH oxidase complex, which promotes ROS production by macrophages in response to L. monocytogenes infection. FAF1 controls phagocytic NADPH oxidase-mediated inflammatory responses upon L. monocytogenes infection by interacting with p67phox, thereby inhibiting bacterial growth.
Based on a previous study of the role of FAF1 in antiviral responses against infection by RNA virus [27], we asked whether FAF1 is also involved in responses to bacterial infection. To examine in vivo host responses to infection by L. monocytogenes, FAF1+/+ and FAF1gt/gt mice were infected intraperitoneally with L. monocytogenes (5 × 105 CFU per mouse) and serum cytokine levels, bacterial load, and proinflammatory gene expression in the spleen and liver were measured at 24 h post-infection (hpi). The bacterial load in the spleen and liver of FAF1gt/gt mice was approximately 10-fold and 3-fold higher, respectively, than that in FAF1+/+ mice (Fig 1, panel A). Serum cytokine levels were also reduced markedly in FAF1gt/gt mice (Fig 1, panel B). Expression of proinflammatory genes in the spleen and liver of FAF1gt/gt mice was also lower than that in FAF1+/+ mice (Fig 1, panels C-D). To determine whether these effects were mediated by peritoneal macrophages in response to L. monocytogenes infection, FAF1+/+ and FAF1gt/gt mice were infected with L. monocytogenes intraperitoneally and peritoneal macrophages (PMs) were isolated at 24 hpi. Expression of proinflammatory cytokines and chemokines by these cells was measured by quantitative real-time PCR. Expression of mRNA encoding IL-6, CXCL10, and RANTES was significantly lower in PMs from FAF1gt/gt mice than in those from FAF1+/+ mice (S1 Fig). These data suggest that FAF1 plays an important role in host defense against L. monocytogenes infection in vivo. Next, to determine whether FAF1 is involved in inflammatory responses, we examined induction of FAF1 in macrophages exposed to L. monocytogenes. Importantly, mRNA expression of FAF1 was induced with 2–3 folds at 15 or 60 m post-infection (mpi) with L. monocytogenes in Raw264.7 cells, suggesting that FAF1 responds to bacterial infection at early time (S2 Fig, panel A). Also, FAF1 responded to high MOI of bacterial infection in mouse bone marrow-derived macrophages (BMDMs) and Raw264.7 cells (S2 Fig, panels B-C). At early time points after L. monocytogenes infection, host defense is regulated by secretion of several cytokines and chemokines, including IL-6, IL-12, and RANTES [30, 31]. Therefore, we examined the effect of FAF1 on proinflammatory cytokine secretion by BMDMs and resident PMs against L. monocytogenes infection or TLR2 ligands. First, expression of FAF1 was confirmed by immunoblotting of extracts from BMDMs or PMs isolated from FAF1+/+ and FAF1gt/gt mice (S3 Fig, panels A-B). Next, cells were infected with L. monocytogenes or treated with zymosan or bacterial lipoprotein (BLP). The supernatants were harvested at 12 or 24 hpi to measure IL-6 and IL-12 by ELISA. As results, FAF1 knockdown reduced cytokine production by both BMDMs (Fig 1, panels E-F) and PMs (S3 Fig, panels C-D). Next, we generated control and FAF1-knockdown murine macrophage cells (Raw264.7) using lentiviruses harboring non-specific or FAF1-specific small hairpin RNA (shRNA). Expression of FAF1 was then determined by immunoblot analysis (S4 Fig, panel A). Next, cells were infected with L. monocytogenes or treated with zymosan or BLP and supernatants harvested at 12 or 24 hpi to measure cytokine secretion. Consistent with the results obtained from primary cells, shRNA-mediated knockdown of FAF1 led to a marked reduction in secretion of proinflammatory cytokines IL-6 and IL-12 upon stimulation via TLR2 (Fig 1, panels G-H). Additionally, production of chemokines regulated upon activation (i.e., normal T cell expressed and secreted (RANTES) and monocyte chemoattractant protein (MCP)-1) was lower in FAF1-knockdown cells than in control cells (S4 Fig, panels B-C). Collectively, these results suggest that FAF1 expression has a marked effect on host defense responses against L. monocytogenes.
To determine whether FAF1 activates proinflammatory signaling pathways, we performed immunoblot analysis to examine expression of activated forms of molecules related to the NF-κB and MAPK signaling pathways in BMDMs isolated from FAF1+/+ and FAF1gt/gt mice infected with L. monocytogenes. The result showed that knockdown of FAF1 leads to a marked reduction in phosphorylation of p65 (NF-κB), IκBα, and SAPK/JNK, but had no effect on activation of p38 MAPK or Erk1/2 (Fig 2, panel A). In agreement with this, knockdown of FAF1 in Raw264.7 cells suppressed activation of p65 (NF-κB), IκBα, and SAPK/JNK, but not p38 or Erk1/2 MAPK (Fig 2, panel B). Furthermore, to measure expression of proinflammatory genes, BMDMs and resident PMs isolated from FAF1+/+ and FAF1gt/gt mice were infected with L. monocytogenes and subjected to real-time PCR at 12 hpi to detect Il6, Nos2 (iNOS), Ptgs2 (COX-2), Cxcl10, and RANTES. The expression of these genes in FAF1-knockdown BMDMs cells were lower than those in wild-type cells (Fig 2, panel C) or resident PMs (S5 Fig). Consistent with this, similar results were obtained from FAF1-knockdown Raw264.7 cells (Fig 2, panel D). Taken together, these data suggest that FAF1 plays a role in inflammatory responses against L. monocytogenes infection.
To identify the target protein of FAF1, large-scale cultured HEK293 cells were used for immunoprecipitation with an anti-FAF1 antibody, followed by mass spectrometry analysis. The result identified NADPH oxidase activator 1 (NoxA1) (S6 Fig). NoxA1 regulates activation of Nox1, which can generate ROS and is expressed at high levels in colon cancer cells [3]. As a homolog of NoxA1, p67phox acts mainly as an activator of Nox2 in phagocytes. Moreover, there is a high degree of domain homology between NoxA1 and p67phox, although the proteins show only 28% amino acid identity [3, 32, 33]. Based on these reports, we asked whether FAF1 interacts with p67phox to regulate ROS production in phagocytes. Mock- or L. monocytogenes-infected Raw264.7 cells were harvested at various time points and cell lysates were immunoprecipitated with an anti-FAF1 antibody, followed by immunoblotting with antibodies against components of the phagocytic NADPH oxidase complex (Fig 3, panel A). The result indicated that FAF1 transiently interacts with p67phox, p47phox, and p40phox at 30 mpi with L. monocytogenes (Fig 3, panel A). To test whether FAF1 expression affects those interaction, efficiency of FAF1-specific siRNA was priorly determined in Raw264.7 cells or BMDMs for further experiments (S7 Fig, panels A-B). FAF1 knockdown-BMDMs showed a weak interaction between both molecules as well as lower expression of p67phox compared with control BMDMs upon L. monocytogenes infection (S8 Fig, panel A). Similar result was obtained following siRNA-mediated knockdown of FAF1 in Raw264.7 (S8 Fig, panel B). Moreover, confocal microscopy analysis exhibited that FAF1 is translocated to phagosomal membranes upon zymosan treatment in BMDMs where it co-localizes with p67phox (Fig 3, panel B). Additionally, immunoprecipitation with an anti-V5 antibody using FAF1-overexpressing Raw264.7 cells showed strong interaction between ectopic FAF1 and endogenous p67phox without stimulation, suggesting that FAF1 might present high affinity to p67phox (Fig 3, panel C). Next, we performed CFU assay to examine the growth rate of intracellular L. monocytogenes in BMDMs following siRNA-mediated knockdown of FAF1. Knockdown of FAF1 showed a significant increase of bacterial growth compared to control BMDMs (Fig 3, panel D). This result encouraged us to verify that reduced ROS production and inflammation in FAF1-knockdown cells might result in a more favorable environment for L. monocytogenes replication. To investigate whether FAF1 affects ROS production upon L. monocytogenes infection, we used fluorescence absorbance to measure ROS production following siRNA-mediated knockdown of FAF1 in BMDMs. As expected, H2O2 and O2-produced by FAF1-knockdown BMDMs were significantly lower than those produced by control BMDMs in response to L. monocytogenes infection (Fig 3, panel E). Similar results were obtained from Raw264.7 cells following siRNA-mediated knockdown of FAF1 (Fig 3, panel F). Moreover, knockdown of FAF1 exhibited lower ROS production upon stimulation with zymosan in BMDMs, Raw264.7, and PMs compared to control cells. (S9 Fig, panels A-C). ROS acts as a signal transduction mediator in response to diverse stimuli [5]. To determine further whether proinflammatory cytokine production in response to L. monocytogenes infection or TLR2 signaling correlated directly with ROS generation in the presence or absence of FAF1, we measured NO and IL-6 levels in supernatants from BMDMs (Fig 3, panel G) and Raw264.7 cells (Fig 3, panel H) stimulated with L. monocytogenes or zymosan in the presence/absence of ROS inhibitors [N-acetyl-L-cysteine and diphenyleneiodonium]. As expected, treatment with ROS inhibitors reduced NO production toward nearly basal level despite of stimulation with L. monocytogenes or zymosan. Secretion of IL-6 by these cells also fell markedly, regardless of FAF1 expression. In other words, impaired ROS production resulted in no significant difference in cytokine secretion by wild-type and FAF1-knockdown macrophages. Knockdown of FAF1 attenuates IκBα degradation and NF-κB activation, but not phosphorylation of IKKs, suggesting that FAF1 indirectly regulates the inflammatory responses via ROS production on TLR2 signaling (S10 Fig). It was also supported by evidences that FAF1 augments inflammatory responses depending on NADPH oxidase complex (S11 Fig, panels A-B). Likewise, knockdown of FAF1 led to increased bacterial growth than control in BMDMs. However, there was no differences under DPI or NAC treatment, which suggests that FAF1 inhibits bacterial growth by mediating ROS production (Fig 3, panel I). Taken together, these findings demonstrate that FAF1 enhances inflammatory responses and intracellular bacterial clearance via ROS generation by interacting with p67phox upon infection by L. monocytogenes.
FAF1 contains three well characterized domains, the FID, DEDID, and C-terminal domains [21, 24, 29]. To define which domain of FAF1 interacts with p67phox, we generated GST-tagged domain constructs and performed GST pull-down assays in HEK293T cells. p67phox bound strongly to both the DEDID and C-terminal domains of FAF1 (Fig 4, panel A). The region of FAF1 responsible for interaction with p67phox was narrowed down to amino acids 330–489 (Fig 4, panel B). In the reverse experiment, GST-tagged domain constructs of p67phox were generated, and GST pull-down assays were performed in HEK293T cells to determine which domain of p67phox interacts with FAF1. FAF1 bound to the 4X tetratricopeptide repeat (TPR) domain of p67phox (Fig 4, panel C). A diagram of the domains mediating binding between FAF1 and p67phox is shown in panel D (Fig 4, panel D). Next, a mutant construct of FAF1 with a deletion in amino acids 330–489 was generated (Δ330–489). This mutant showed weak binding to endogenous p67phox and p47phox compared with wild-type FAF1 by immunoprecipitation with an anti-V5 antibody in Raw264.7 cells (Fig 4, panel E). Furthermore, ectopically expressed p67phox co-localized with wild-type FAF1 but not FAF1 Δ330–489 in HEK293T cells (Fig 4, panel F). Taken together, these data indicate that amino acids 330–489 of FAF1 comprise the critical region that mediates interaction with the 4X TPR domain of p67phox.
As described above, we identified the binding site through which FAF1 interacts with p67phox and found that a deletion mutant of this region (FAF1 Δ330–489) was unable to bind to p67phox. To determine whether FAF1 Δ330–489 lost the ability to promote inflammatory responses due to the weak interaction with p67phox, we generated Raw264.7 cells stably overexpressing a control vector, wild-type FAF1, or FAF1 Δ330–489. The expression level of FAF1 in each stable cell line was determined by western blotting (S12 Fig, panel A). The cells were then stimulated with TLR2 ligands, including L. monocytogenes, zymosan, and BLP, and levels of proinflammatory cytokines and chemokines in the supernatants were measured. As expected, cells expressing FAF1 Δ330–489 showed reduced secretion of proinflammatory cytokines and chemokines in response to TLR2 signaling compared with cells expressing wild-type FAF1 (Fig 5, panels A-B and S12 Fig, panels B-C). We next evaluated the activation of molecules involved in NF-κB and MAPK signaling in L. monocytogenes-infected Raw264.7 cells. Cells expressing wild-type FAF1 showed enhanced activation of the NF-κB and SAPK/JNK pathways upon L. monocytogenes infection, as previously observed (Fig 2, panels A-B), while activation of these signaling molecules was not altered in cells overexpressing FAF1 Δ330–489 compared with control cells (Fig 5, panel C). In addition, IL-6, O2, and NO production were impaired in cells overexpressing FAF1 Δ330–489 compared with cells overexpressing wild-type FAF1. Cytokine secretion was not induced in the presence of ROS inhibitors, with no significant difference between individual cell lines (Fig 5, panels D-E). While the growth of intracellular bacteria in cells overexpressing wild-type FAF1 was lower than in control cells, cells overexpressing FAF1 Δ330–489 exhibited a similar level of bacterial growth as control cells (Fig 5, panel F). These findings indicate that the effects of FAF1 on inflammatory responses, and intracellular bacterial clearance through ROS generation are dependent on its interaction with p67phox.
Among NADPH oxidase regulatory proteins, p67phox has a critical role in the activation of NADPH oxidase [4, 34]. To determine whether FAF1 affects phagocytic NADPH oxidase activity, Raw264.7 cells overexpressing control vector, wild-type FAF1, or FAF1 Δ330–489 were infected with L. monocytogenes, and a chemiluminescence assay was performed to measure NADPH oxidase activity. The result showed that overexpression of wild-type FAF1 augments the phagocytic NADPH oxidase activity, whereas overexpression of FAF1 Δ330–489 doesn`t (Fig 6, panel A). This finding suggested that amino acids 330–489 is the critical region for the interaction with p67phox, which results in increased ROS production in response to L. monocytogenes infection. We also found that expression of FAF1 leads to higher and persistent expression of p67phox in a binding-specific manner (Fig 6, panel B). Overexpression of FAF1 induces a bit more mRNA expression of p67phox and p47phox without stimulation (S13 Fig). Additionally, Raw264.7 cells overexpressing control vector, wild-type FAF1 were treated with mock or zymosan/cycloheximide and used for immunoblot analysis for expression levels of p67phox and p47phox over time. This result showed that overexpression of FAF1 increases the stability of p67phox and p47phox upon zymosan treatment (Fig 6, panel C). The intensity of the protein bands on the blot shown in Fig 6, panel D is quantitated (Fig 6, panel D). Furthermore, to check the localization of p67phox around phagosomes depending on expression level of FAF1, BMDMs were treated with zymosan particles for 30 min following siRNA-mediated knockdown of FAF1, then followed by confocal microscopy using anti-p67phox antibody (Fig 6, panel E). As correlated with prior data, knockdown of FAF1 exhibited considerable decrease of p67phox localized to phagosomal membranes compared to control, which supports evidently that FAF1 potentiates the stability of p67phox in phagosomes. Collectively, these results suggest that FAF1 in macrophages effectively augments p67phox stability in response to infection with L. monocytogenes, resulting in increased phagocytosis-mediated NADPH oxidase activity.
FAF1 (Fas-associated factor 1), a member of the Fas death-inducing signal complex, modulates a variety of biological processes by interacting with diverse molecules [24–26]. However, the role of FAF1 in host defense against bacterial infection remains unclear. Here, we report that FAF1 is a positive regulator that increases activity of the phagocytic NADPH oxidase complex, resulting in production of ROS and in activation of NF-κB signaling, inflammatory responses, and antibacterial activity upon L. monocytogenes infection. First, FAF1gt/gt mice exhibited reduced serum cytokine levels, reduced inflammatory gene expression, and increased bacterial burden during L. monocytogenes infection. Second, primary macrophages (BMDMs and PMs) isolated from FAF1gt/gt mice showed decreased ROS production and inflammatory responses as well as bacterial clearance than macrophages from FAF1+/+ mice upon L. monocytogenes infection or TLR2 stimulation. Consistent with these data, knockdown of FAF1 in Raw264.7 cells also showed significantly reduced ROS production, NF-κB activation and inflammatory responses, upon L. monocytogenes infection. Third, FAF1 transiently interacted strongly with the p67phox-p47phox-p40phox complex at early time points after L. monocytogenes infection in macrophages, and FAF1 region comprising amino acids 330–489 was responsible for interaction with the TPR domain of p67phox. Finally, interaction between FAF1 and p67phox stabilized p67phox and increased activity of phagocytic NADPH oxidase upon L. monocytogenes infection. Collectively, these findings strongly suggest that FAF1 plays a crucial role in promoting antibacterial responses by interacting with p67phox in macrophages during intracellular microbial infection.
NADPH oxidase and dual oxidase induce ROS production by various cells and tissues in response to growth factors, cytokines, and pathogen-mediated signals [1]. Among these, the phagocyte NADPH oxidase is a multi-component complex in which the membrane glycoprotein gp91phox (known as NOX2) is tightly associated with p22phox; this complex is activated via association with cytosolic regulatory proteins such as p67phox, p47phox, p40phox, and the small GTPase Rac, resulting in ROS generation [1, 3, 4]. Based on homology to gp91phox (Nox2), the Nox family of NADPH enzymes comprises seven members: Nox1 through Nox5, plus Duox1 and Duox2 [35]. All gp91phox-related enzymes (except gp91phox (Nox2)) belonging to the Nox family are non-phagocytic enzymes expressed in epithelial or endothelial cells within diverse tissues and organs [1]. ROS generated by these NADPH oxidases take part in biological processes such as cell signaling, hormone biosynthesis, and host innate immune responses [5, 36]. In particular, ROS play essential roles in phagocyte-mediated defense against bacterial infection; ROS kill engulfed pathogens directly, or indirectly by activating intracellular signaling pathways related to inflammatory responses, which then protect the host. ROS are necessary to eliminate intracellular bacteria such as mycobacteria, Listeria, and Salmonella, efficiently [37].
Recognition of pathogens by TLRs is the first line of host innate defense; indeed, interaction between TLRs and NADPH oxidase in phagocytes has been well studied. For example, Yang et.al., report that Nox2 is essential for TLR2-dependent inflammatory responses and for intracellular control during mycobacterial infection. They showed that Nox2 and TLR2 interact directly during mycobacterial infection [13, 14]. Moreover, the interplay between TLR4 and NADPH oxidases such as Nox4 or Nox1 was studied in non-phagocytic cells [16, 38]. However, lipopolysaccharide (LPS; a TLR4 agonist) also activates NADPH oxidase in phagocytes indirectly by increasing association of gp91phox (Nox2) with regulatory proteins in plasma membrane [39]. Here, we demonstrated that FAF1 is a crucial regulator of the phagocytic NADPH oxidase (Nox2) complex required for ROS production by macrophages in response to L. monocytogenes infection as well as TLR2 stimulation. We also examined whether FAF1 regulates inflammatory responses in macrophages stimulated by TLR4. As results, TLR4-mediated cytokine secretion by BMDMs isolated from FAF1gt/gt mice was lower than that by cells from FAF1+/+ mice (S14 Fig, panels A-B). These results suggest that FAF1 also activates signaling pathways related to TLR4 stimulation, as proposed by DeLeo et.al. However, further studies demonstrating a detailed mechanism of how FAF1 controls TLR4 signaling pathway are needed.
Nox2 in resting phagocytes is inactive; however, it is activated by phagocytosis of invading bacteria, leading to ROS production and their subsequent effects on host defense (e.g., killing bacteria and regulating intracellular signaling). Once bacteria are recognized by host TLRs, NADPH oxidase Nox2 (gp91phox) heterodimerizes with p22phox at the phagocyte membrane and is rapidly activated by cytosolic regulatory proteins [1, 3, 4]. These p22phox and cytosolic regulatory proteins (p67phox, p47phox, p40phox, and the small GTPase Rac) are indispensable for regulation of NADPH oxidase; indeed, increasing evidence suggests that it is important to control NADPH oxidase subunits to ensure appropriate ROS generation. For example, RUBICON interacts with p22phox, thereby increasing ROS production in response to infection by Gram-positive bacteria [12]. Moreover, Nox-dependent ROS production occurs in Parkinson’s disease (autosomal recessive, early onset) 7 (Park7)-p47phox [40]. Our findings suggest that FAF1 is a key molecule that regulates activation of NADPH oxidase via strong interaction with p67phox.
As noted above, mass spectrometry analysis identified the NoxA1 protein as an interacting protein of FAF1. NoxA1 is expressed at high levels by colon epithelial cells and is responsible for activation of Nox1 (and thereby for subsequent ROS production) [1, 3, 33]. p67phox is a homolog of NoxA1 expressed by phagocytes. Thus, we hypothesized that FAF1 interacts with p67phox to regulate ROS generation in macrophages. In this study, we found that FAF1 interacts with p67phox at an early time point after L. monocytogenes infection. Moreover, amino acids 330–489 of FAF1 are required for interaction with p67phox. Conversely, FAF1 interacted with the TPR domain (amino acids 1–155) of p67phox. This domain, which comprises four 34 amino acid-long TPR motifs, is involved in a variety of protein-protein interactions [4, 32, 41, 42]. In macrophage cell lines stably overexpressing wild-type FAF1 or FAF1 Δ330–489, we found that overexpression of wild-type FAF1 but not FAF1 Δ330–489 increased NADPH oxidase activity, ROS production, proinflammatory cytokine production, and antimicrobial activity. Ultimately, interaction between FAF1 and p67phox facilitated stabilization of p67phox and increased activity of NADPH oxidase upon L. monocytogenes infection. These results suggest that regulation of phagocytic NADPH oxidase by FAF1 is dependent on binding to p67phox. Consequently, we demonstrated that FAF1 positively regulates the NADPH oxidase 2 complex via stabilization of p67phox.
FAF1 interacts potentially with many different proteins and functions as a negative and/or positive regulator in a variety of biological possesses [25, 26, 43]. Previous studies report that FAF1 homologs suppress antibacterial immunity in Drosophila and Locusta migratoria [44, 45]. In addition, Park et al. showed that FAF1 suppresses IKK activation and nuclear translocation of NF-κB in fibroblasts [29, 46]. However, we clearly demonstrate that FAF1 acts as a positive regulator of antibacterial responses by regulating ROS production. In particular, we identified the physiological role of FAF1 in defense responses against L. monocytogenes infection in FAF1+/+ and FAF1gt/gt mice. FAF1gt/gt mice attenuates bacterial clearance due to reduced inflammatory responses. We also have focused on the role of FAF1 in phagocytes such as macrophages in vitro. Raw264.7 cells in which FAF1 was knocked down, as well as BMDMs and resident PMs isolated from FAF1gt/gt mice, showed lower ROS production, proinflammatory responses, and bacterial killing activity than FAF1+/+ mice upon L. monocytogenes infection. In addition, activation of molecules involved in NF-κB signaling was markedly reduced in FAF1-knockdown cells when exposed to L. monocytogenes. However, the upstream molecules responsible for FAF1 activation upon L. monocytogenes infection or TLR2 stimulation remain unclear. To assess this, further studies are necessary to examine the detailed molecular mechanisms (e.g., phosphorylation of FAF1) that operate in macrophages under infectious conditions.
In summary, we demonstrated that FAF1 is a critical positive regulator of the phagocytic NADPH oxidase (Nox2) complex responsible for ROS generation by phagocytes upon L. monocytogenes infection or TLR2 stimulation. FAF1 interacts directly with p67phox and stabilizes p67phox, thereby triggering NADPH oxidase-mediated ROS production, release of proinflammatory cytokines, and bacterial clearance in response to infection by L. monocytogenes. Taken together, the results suggest a plausible mechanism involving interaction between p67phox and FAF1 and increases our understanding of molecules that control ROS signaling and antibacterial defense responses against L. monocytogenes infection or TLR2 stimulation.
All animal experiments were managed in strict accordance with the Guide for the Care and Use of Laboratory Animals (National Research Council, 2011) and performed in BSL-2 and BSL-3 laboratory facilities with the approval of the Institutional Animal Care and Use Committee of Bioleaders Corporation (Reference No., BLS-ABSL-16-002) and Chungnam National University (Reference No., CNU-00763).
Zymosan (tlrl-zyn), Pam3CSK4 (tlrl-pms) were purchased from Invivogen. DPI(D2926), NAC(A9165), Cycloheximide(C7698), DHE(D7008), Amplex red and peroxidase were purchased from Sigma. DCFH-DA and pHrodo Red Zymosan A Bioparticles conjugate (P35364) were obtained from Molecular probe. For western blot analysis, specific antibodies for p-IKKα/β (2697), IKKβ (8943), p-JNK (4668), JNK (9258), p-NF-κB p65 (3033), NF-κB p65 (4764), p-Erk1/2 (9101), Erk1/2 (9102), p-p38 (4631), p38 (9212), p-IκBα (2859), IκBα (4812) and p67phox (3923) were purchased from Cell Signaling Technology. Antibodies for FAF1 (sc-393965), p47phox (sc-14015), IKKα (sc-7606), β-actin (sc-47778), GST (sc-138) were purchased from Santa Cruz Biotechnology. Anti-p67phox antibody (ab109366) was purchased from abcam. Anti-V5 antibody (46–0705) was purchased from Invitrogen life technology. The anti-FAF1 monoclonal antibody was kindly provided by Dr. Eunhee Kim (Chungnam National University, Daejeon, Korea).
L.monocytogenes (KVCC-BA0000087) was from Korean Veterinary Culture Collection (KVCC), and grown at 37°C in Brain Heart Infusion (BHI) broth medium (BD 237500). Log phase bacteria (O.D. value, 0.6–0.8) were used for all assay. Cultures were aliquoted and stored at -80°C. To determine bacterial titer (Colony-forming unit, CFU), bacteria thawed was diluted 10-fold. Each diluent was plated on BHI agar (BD 241830) and incubated at 37°C for one day.
FAF1+/+ and FAF1gt/gt mice on a C57BL/6 background were kindly provided by Dr. Eunhee Kim (Chungnam National University, Daejeon, Korea). A hypomorphic allele, designated FAF1gt/gt, was generated by a gene-trap insertion in intron 8 [47]. All mice were bred in pathogen-free condition. Offspring were genotyped by PCR as previously described [27].
Sex-matched mice (six-week-old) were intraperitoneally infected with L. monocytogenes (5 × 105 CFU/mouse). Liver, spleen, peritoneal macrophage and serum were collected to determine the bacterial load, cytokines or chemokines levels or mRNA levels of cytokines or chemokines as described below.
FAF1 plasmid was kindly provided by Dr. Eunhee Kim (Chungnam National University, Daejeon, Korea). To generate plasmid constructs with GST expressing vector, FAF1 and its mutants or p67phox and its mutants were amplified by PCR and inserted into pEBG vector. For construction of V5-tagged expression plasmid, FAF1, FAF1 Δ330–489 and p67phox were amplified by PCR and inserted into pIRES-V5.
Raw264.7 cells, HEK293T cells, BMDMs and PMs were maintained in Dulbecco’s Modified Eagle’s medium (Gibco) supplemented with 10% FBS (Thermo-Hyclone) and antibiotic-antimycotic (Gibco) at 37°C in 5% CO2. To establish stable expressing cell line, Raw264.7 cells were transfected with empty vector or V5-tagged FAF1 wild-type or V5-tagged FAF1 Δ330–489 using Lipofectamine 2000 (Invitrogen), then selected with 2μg/ml puromycin (Gibco) containing culture media for 2 weeks. To generate FAF1 knockdown cell lines, Raw264.7 were infected with lentivirus containing non-specific shRNA or FAF1-specific shRNA in the presence of 8μg/ml polybrene (Sigma AL118), and then selected with 2μg/ml puromycin for 2 weeks as previously described [27]. Resident peritoneal macrophages were obtained by flushing peritoneal cavity of FAF1+/+ and FAF1gt/gt mice with HBSS w/o phenol red as previously described [48]. BMDMs were isolated from FAF1+/+ or FAF1gt/gt mice and cultured for 6 days in medium containing 20 ng/ml recombinant mGM-CSF (Creagen) as previously described [27].
For silencing of FAF1 gene expression, the pGIPZ lentiviral vector, which contains FAF1-specific shRNA sequences was purchased from Open Biosystems. (http://www.openbiosystems.com). Lentiviruses were produced as previously described [27]. In brief, HEK293T cells were transiently transfected with packaging plasmids (psPAX2 and pMD2.VSV-G) and pGIPZ containing non-specific shRNA or FAF1-specific shRNA sequences using Lipofectamine 2000. At 48–72 hr post-transfection, virus-containing media was collected and filtrated (0.45 μm filter, Millipore).
The sequence of the mouse FAF1-specific siRNA #1 (duplex) were as follow, 5`-CCG CCU UCA UCA UCC AGC C-3`and 5`-GGC UGG AUG AUG AAG GCG G-3`. The mouse FAF1-specific siRNA duplex #2–4 (14084–1, 14084–2, and 14084–3) were purchased from Bioneer Corp. The mouse p47phox-specific siRNA (sc-36157) was purchased from Santa Cruz Biotechnology. A non-targeting siRNA was used as a control. Cells were transfected with duplex siRNA using Lipofectamine RNAiMAX (Invitrogen), according to the manufacturer`s protocol, then incubated for 36 hrs before stimulation.
For immunoprecipitation, Raw264.7 cells or BMDMs were lysed with RIPA buffer containing protease inhibitor cocktail (Roche). Lysates were incubated with a primary antibody overnight at 4°C, followed by incubation with protein A/G agarose (Santa Cruz Biotechnology) for 3hrs at 4°C. Then immunoprecipitates were washed with 1% nonidet P-40 for 3 times. For GST-pull down assay, HEK293T cells were transiently transfected with the indicated plasmids, and at 36hr post-transfection, were lysed with RIPA buffer containing protease inhibitor cocktail (Roche). Lysates were pre-cleared by sepharose 6B (GE Healthcare) for 3hrs at 4°C, followed by incubation with glutathione (GSH) sepharose 4B (GE Healthcare) overnight at 4°C. Precipitated beads were washed with 1% nonidet P-40 for 3 times.
Whole cell lysates (WCLs) were prepared using Radio-immunoprecipitation assay (RIPA) lysis buffer (50mM Tris-HCl, 150mM NaCl, 0.5% sodium deoxycholate, 1% IGEPAL, 1mM NaF, 1mM Na3VO4, proteinase inhibitor cocktail). Samples were separated by SDS-PAGE and transferred onto a PVDF membrane (Bio-Rad) using Trans-Blot semi dry transfer cell (Bio-Rad). Membranes were blocked for 1hr in 5% bovine-serum albumin with TBST or 5% skim milk within PBST, followed by incubation with primary antibody overnight at 4°C. Next day, membranes were washed with TBST or PBST, and incubated at room temperature with HRP-conjugated secondary antibody. Then, membranes were washed 3 times with TBST or PBST, followed by developing with Western blotting detection reagents (GE healthcare, ECL select Western Blotting Detection Reagent).
HEK293 cells were lysed with RIPA buffer containing protease inhibitor cocktail (Roche). Lysates were incubated with anti-FAF1 antibody overnight at 4°C, followed by incubation with protein A/G agarose (Santa Cruz Biotechnology) for 3hrs at 4°C. Then immunoprecipitates were washed with 1% nonidet P-40 for 3 times and separated by PAGE gels, followed by Coomassie blue staining. Protein bands were excised from the gel and identified by Q-TOF mass spectrometer.
ELISA was performed to detect the secreted cytokines and chemokines in sera or cell culture supernatants. Mouse IL-6 (BD biosciences, 555240) and mouse IL-12 (BD biosciences, 555165), RANTES (Invitrogen), MCP-1 (Invitrogen) were used for analysis according to the manufacturer’s protocols.
The oxidative fluorescent dye 20μM CM-H2DCFDA (Molecular probe), 20uM Amplex red (Sigma) with 0.1 unit/ml peroxidase (Sigma) and 20μM DHE (Sigma) were used to detect intracellular total ROS (Excitation 485nm, Emission 530nm), H2O2 production (Excitation 530nm, Emission 620nm) and O2- (Excitation 485nm, Emission 620nm), respectively, using fluorescence module of GloMax-Multi Microplate Reader (Promega, E7031). NO detection in culture media was performed using Griess reagent (G4410, Sigma) at 540 nm. Lucigenin (M8010) and NADPH (N1630) were used to determine NADPH oxidase activity as previously described [12].
Total RNA was isolated from cells and murine tissues using the RNeasy Mini Kit (Qiagen). cDNA synthesis was performed using reverse transcriptase (TOYOBO), cDNA was quantified by real-time polymerase chain reaction (PCR) using QuantiTect SYBR Green PCR kit (TOYOBO), according to the manufacturer`s instructions on a Rotorgene (Qiagen). The primer sequences were as follow: mFAF1, 5'-GGT GAC TGC CAT CCT GTA TTT T-3' (forward) and 5'-TGC TCT GTT GGT GTC CTT TG-3`(reverse), mGAPDH, 5`-TGA CCA CAG TCC ATG CCA T-3`(forward) and 5`- GAC GGA CAC ATT GGG GGT AG-3`(reverse); mIL-6, 5`- GAC AAC TTT GGC ATT GTG G-3`(forward) and 5`- ATG CAG GGA TGA TGT TCT G-3`(reverse); mIL-12p40, 5'-CAG AAG CTA ACC ATC TCC TGG TTT G-3' (forward) and 5'-TCC GGA GTA ATT TGG TGC TTC ACA C-3' (reverse); mIL-1β, 5`-TTG TGG CTG TGG AGA AGC TGT-3`(forward) and 5`-AAC GTC ACA CAC CAG CAG GTT-3`(reverse); mCOX-2, 5`-TGA GTA CCG CAA ACG CTT CT-3`(forward) and 5`-CTC CCC AAA GAT AGC ATC TGG-3`(reverse); miNOS, 5`-TGG GAA TGG AGA CTG TCC CAG-3`(forward) and 5`-GGG ATC TGA ATG TGA TGT TTG-3`(reverse); CXCL10, 5`-GCC GTC ATT TTC TGC CTC A-3`(forward) and 5`-CGT CCT TGC GAG AGG GAT C-3`(reverse); RANTES, 5`- CCA GAG AAG AAG TGG GTT CAA G-3`(forward) and 5`- AAG CTG GCT AGG ACT AGA GCA A-3`(reverse); mp67phox, 5´-CAG ACC CAA AAC CCC AGA AA-3´ (forward) and 5´-AGGGTGAATCCGAAGCTCAA-3´ (reverse); mp47phox, 5´-GTC CCT GCA TCC TAT CTG GA-3´ (forward) and 5´-TAT CTC CTC CCC AGC CTT CT-3´ (reverse); mgp91phox, 5´-TCG CTG GAA ACC CTC CTA TG-3´ (forward) and 5´-GGA TAC CTT GGG GCA CTT GA-3´ (reverse) (Bioneer, Daejeon, Republic of Korea).
BMDMs or HEK293T cells were seeded into eight-chamber slides, followed by treatment or transfection desired. The cells were fixed in 4% paraformaldehyde at RT for 20min, then permeabilized by incubation for 20min with 100% Methanol at -20°C. Then, the fixed cells were incubated with 2% FBS for 1hr to block non-specific binding of antibodies. The appropriate primary antibodies were incubated overnight at 4°C. The cells were washed three times with PBST, and incubated with the appropriate secondary antibodies (Invitrogen) for 1hr at RT without light exposure. Then, the cells were washed three times with PBST, and stained with DAPI (ratio, 1:100,000), washed with PBS, and mounted in mounting solution (VECTOR). Images were acquired under a Nikon laser scanning confocal microscope (C2plus) and analyzed using NIS-Elements software.
BMDMs or Raw264.7 cells were infected with L. monocytogenes (MOI = 0.1) for 1hr and washed 3 times with sterile PBS, followed by incubation with DMEM containing gentamycin (100μg/ml) for the indicated times. Finally, cells were harvested and lysed with 0.1% Triton X-100 (Sigma) to release the intracellular bacteria, and cell lysates were then diluted 10-fold in BHI broth (BD). Each sample were plated on BHI agar (BD) and incubated at 37°C for one day. Colony-forming unit (CFU) was utilized to ensure quantification of intracellular bacteria.
Prism 6 software (GraphPad Software) was used for charts and statistical analyses. The significance of results was analyzed by an unpaired two-tailed Student’s t-test and Mann-Whitney test with a cutoff P value of 0.05. Error bars and P-values are indicated in the figure legends.
|
10.1371/journal.pntd.0005795 | Animal-related factors associated with moderate-to-severe diarrhea in children younger than five years in western Kenya: A matched case-control study | Diarrheal disease remains among the leading causes of global mortality in children younger than 5 years. Exposure to domestic animals may be a risk factor for diarrheal disease. The objectives of this study were to identify animal-related exposures associated with cases of moderate-to-severe diarrhea (MSD) in children in rural western Kenya, and to identify the major zoonotic enteric pathogens present in domestic animals residing in the homesteads of case and control children.
We characterized animal-related exposures in a subset of case and control children (n = 73 pairs matched on age, sex and location) with reported animal presence at home enrolled in the Global Enteric Multicenter Study in western Kenya, and analysed these for an association with MSD. We identified potentially zoonotic enteric pathogens in pooled fecal specimens collected from domestic animals resident at children’s homesteads. Variables that were associated with decreased risk of MSD were washing hands after animal contact (matched odds ratio [MOR] = 0.2; 95% CI 0.08–0.7), and presence of adult sheep that were not confined in a pen overnight (MOR = 0.1; 0.02–0.5). Variables that were associated with increased risk of MSD were increasing number of sheep owned (MOR = 1.2; 1.0–1.5), frequent observation of fresh rodent excreta (feces/urine) outside the house (MOR = 7.5; 1.5–37.2), and participation of the child in providing water to chickens (MOR = 3.8; 1.2–12.2). Of 691 pooled specimens collected from 2,174 domestic animals, 159 pools (23%) tested positive for one or more potentially zoonotic enteric pathogens (Campylobacter jejuni, C. coli, non-typhoidal Salmonella, diarrheagenic E. coli, Giardia, Cryptosporidium, or rotavirus). We did not find any association between the presence of particular pathogens in household animals, and MSD in children.
Public health agencies should continue to promote frequent hand washing, including after animal contact, to reduce the risk of MSD. Future studies should address specific causal relations of MSD with sheep and chicken husbandry practices, and with the presence of rodents.
| Diarrheal disease is one of the leading causes of death worldwide in children younger than 5 years. Exposure to animals in homes may be a risk factor for diarrhea in children. To test this, we studied a subset of children in the Global Enteric Multicenter Study (GEMS) in rural western Kenya, whose caretakers reported the presence of animals in the children’s homesteads. In GEMS, children with moderate-to-severe diarrhea (MSD) were matched with children without MSD, who were of the same sex, similar age and who lived in the same area. We asked questions about the presence and management of animals in the children’s homesteads. We also collected fecal specimens from domestic animals present at homesteads and tested these for microbes that could cause diarrheal disease in children. We found that children who reportedly washed their hands after animal contact, and who lived in a homestead with adult sheep that were not confined to a pen overnight, had a lower risk of MSD. Children who lived in homesteads that owned more adult sheep, or in which fresh rodent droppings were observed frequently, had a higher risk of MSD, as did children who reportedly participated in providing water to chickens in the homestead. We did not find any association between the presence of particular pathogens in household animals, and MSD in children.
| Diarrheal disease remains among the leading causes of global mortality in children younger than 5 years [1, 2]. Although the mortality rate due to diarrheal disease in this age group in Africa has decreased by nearly 4% per year since 2000, it remains unacceptably high: it is estimated that 12% of deaths in children younger than five years in Africa are due to diarrhea, amounting to almost half a million childhood deaths annually [2]. While mortality rates have decreased, the incidence of diarrheal disease in young children in low- and middle-income countries has shown little change, from 3.4 episodes/child year in 1990 to 2.9 episodes/child year in 2010 [3]. Persistently high incidence rates in these countries are concerning because early childhood diarrhea may have long-term effects on child growth and development [4, 5]. Data characterising risk factors and etiologies of diarrheal disease in children in these settings are important for focusing interventions to decrease associated morbidity and mortality rates.
Many viral, bacterial and protozoal pathogens have been demonstrated as causes of diarrheal disease in children younger than 5 years in developing countries [6]. Contact with domestic animals, including livestock, poultry and companion animals, has been shown to play a role in the epidemiology and transmission to people of a number of these pathogens [7, 8] including Campylobacter spp. [9–11], non-typhoidal Salmonella [11, 12], diarrheagenic Escherichia coli strains [12, 13], Cryptosporidium spp. [12–14] and Giardia duodenalis [15]. In addition, some reports implicate dogs as a possible source of human infections with unusual strains of rotavirus [16, 17]. Livestock and poultry play a significant role in rural livelihoods in developing countries, providing a variety of benefits to poor households, such as animal-source food (energy-dense food with high biological-value protein, rich in micronutrients), draft power for ploughing and transport, nutrient recycling through manure, income through sale of animals or their products, as well as a form of savings and insurance [18]; however, animal husbandry may also have negative impacts on households, including the transmission of zoonotic and foodborne diseases. In a meta-analysis of demographic health survey data from 30 sub-Saharan African countries examining associations between child health outcomes and household ownership of livestock, Kaur et al [19] found a negative association between livestock and stunting (an indicator of chronic malnutrition), a positive association between livestock and all-cause mortality in children, and no association between livestock and diarrheal illness. In a systematic review and meta-analysis of human diarrhea infections associated with domestic exposure to food-producing animals, Zambrano et al. [20] found consistent evidence of a positive association between exposure and diarrheal illness in people, across a range of animal species and enteric pathogens. Close contact with domestic animals (such as animals sleeping in the house or room) is also associated with impaired growth in children [21, 22]. Considering the potential positive benefits of animal husbandry to rural livelihoods in resource-poor settings, there is a need to identify specific husbandry-related practices associated with diarrheal illness. Such evidence can serve as bases for interventions to reduce transmission of enteric pathogens to household members, especially to children, who are particularly vulnerable to mortality, sequelae and developmental consequences of diarrheal disease. Identifying etiologies of diarrheal illness in household members and concurrent infections in domestic animals may provide further utility for these efforts [23–25].
The Global Enteric Multicenter Study (GEMS), a large-scale case-control study designed to identify the etiology and population-based burden of diarrheal disease in children younger than 5 years in developing countries [6, 26], provided an opportunity to study the association between animal-related exposures and diarrheal illness in household children at a rural site in western Kenya. GEMS was a 3-year, prospective, age-stratified, matched case-control study of moderate-to-severe diarrheal illness in children aged 0–59 months, residing in populations under demographic surveillance at four sites in sub-Saharan Africa and three sites in south Asia. The methodology [26–28] and main findings [29] of GEMS have been published. The GEMS Zoonotic Enteric Diseases (GEMS-ZED) sub-study was conducted among a subset of case children and their matched controls enrolled at one of the six GEMS sentinel health centers in rural western Kenya. The objectives of the GEMS-ZED study were to identify animal-related exposures associated with cases of moderate-to-severe diarrhea (MSD) in children, and to identify the major zoonotic enteric pathogens present in the domestic animals residing in the homesteads of case and control children.
The GEMS sentinel health center for this study was St Elizabeth Mission Hospital in Lwak (henceforth referred to as Lwak Hospital), located in Rarieda sub-county, Siaya County (formerly Nyanza Province) in western Kenya. Lwak Hospital is the designated referral facility for population-based infectious disease surveillance (PBIDS) conducted in the surrounding 33 villages by the Kenya Medical Research Institute (KEMRI) and the U.S. Centers for Disease Control and Prevention (CDC) [30]. The area also falls within the KEMRI/CDC health and demographic surveillance system (HDSS) site in western Kenya [31]. The HDSS provides general demographic and health information including population age-structure, migration, fertility rates, birth and death rates, verbal autopsy, access and utilization of health care for approximately 220,000 inhabitants in 55,000 households. The primary economic livelihood is subsistence farming and fishing, and an estimated 70% of the population lived below the poverty line in 2003 [32]. The area is culturally homogeneous, with 95% of people being ethnically Luo [33]. Households in the PBIDS villages are clustered into compounds composed of related family units, with most compounds having between one and five family units [33]. Animal husbandry is common: 89% of compounds own at least one species of livestock or poultry, with 86% owning poultry (median flock size: 10), 49% cattle (median herd size: 4), 48% goats (median herd size: 4) and 18% sheep (median herd size: 3) (KEMRI/CDC HDSS data for 2008). Among compounds that own livestock, approximately one-half also own cats and/or dogs (International Emerging Infections Program–Zoonoses Project data for 2009). Rodents, including black rats (Rattus rattus), are also commonly present in and around houses in the PBIDS site [34].
From January 31, 2008 through January 29, 2011, children 0–59 months old who sought care at selected sentinel health centers (including Lwak Hospital) and belonged to the HDSS population were screened for diarrhea. To be eligible for inclusion in GEMS, the diarrhea episode had to meet the case definition for MSD [29], which was three or more loose stools within the previous 24 h, with onset within the previous 7 days after a period of at least 7 diarrhea-free days, with one or more of the following: sunken eyes; loss of skin turgor; intravenous rehydration administered or prescribed; dysentery; or hospitalized with diarrhea or dysentery. Each GEMS site restricted enrollment to the first nine eligible cases per age stratum per fortnight. Three age strata were targeted: infants (0–11 months), toddlers (12–23 months), and children (24–59 months). For every enrolled case, one to three children without diarrhea were enrolled as controls. Controls were matched to individual cases by age (within 2 months of age for patients aged 0–23 months, and within 4 months of age for patients aged 24–59 months), sex, and residence (same or nearby village as patient). Potential controls were randomly selected from the KEMRI/CDC HDSS database and enrolled during a home visit within 14 days of the matched case. Potential controls who had diarrhea in the previous 7 days were ineligible. At enrollment, primary caregivers (parent or other caretaker) of cases and controls were interviewed to obtain demographic, epidemiological and clinical information. In addition, each case and control provided at least 3 g of fresh stool, which was submitted to the laboratory for identification of enteric pathogens using standard methods as described by Panchalingam et al. [28].
The GEMS-ZED substudy collected and analysed additional data on animal-related factors from a subset of GEMS case and matched control children with reported animal presence at home. From November 4, 2009 through February 4, 2011, all cases enrolled into GEMS at Lwak Hospital were screened for inclusion in the GEMS-ZED study. (Enrollment into GEMS continued for a short period after the official end date of January 29, 2011, during which time 3 case-control pairs were enrolled into GEMS-ZED. Data from the GEMS study [laboratory test results and wealth index] are not available for these 3 pairs.) Between zero and six cases per fortnight (median of two) were enrolled into GEMS at Lwak Hospital during the GEMS-ZED study period. Only cases and controls whose primary caregiver reported presence of animals (domestic animals as well as peridomestic wild rodents) at the child’s compound during the GEMS enrollment interview were considered eligible. For each eligible case, the first eligible GEMS-enrolled matched control was identified, resulting in one-to-one matching in the GEMS-ZED dataset. If no eligible child could be identified among the GEMS set of one to three matched controls, then the case was not enrolled into GEMS-ZED. Caregivers of eligible cases and controls were approached for enrollment into the GEMS-ZED study during a separate home visit that took place within 2 weeks of their enrollment into the GEMS study. Written informed consent for participation in the study was sought from the primary caregiver, as well as from the head of the compound of residence of each eligible child; only compounds in which both individuals provided consent were enrolled. Compounds were excluded if the child participating in GEMS had died subsequent to enrollment, or if no domestic animals were found to be resident (for example, if animals had died or were sold subsequent to GEMS enrollment).
Following enrollment, both the head of the compound and the child’s caregiver were interviewed using a standard questionnaire. The questionnaire consisted of two parts: the first part dealt with residence and husbandry of domestic animals in the compound (livestock, poultry, dogs and cats), as well as observations relating to the presence of rodents in and around the compound, and was asked of the person in the compound responsible for the management of animals (typically the head of the compound). The second part dealt with information specific to the participating child, relating to exposures to animals and their environment, and was asked of the child’s caregiver. A summary of the items included in the questionnaire is presented in S1 Table.
At the enrollment visit, fecal specimens were collected from a convenience sample of domestic animals resident at the compound. Specimens from a single species and age category (young, unweaned animals vs. older animals) were pooled together, with specimens from a maximum of five animals collected in a single pool, and a maximum of two pools per species and age category combination (i.e. a maximum of ten animals per species and age category combination were sampled from a compound). A previous study showed good agreement of bacterial culture results between individual and pooled fecal samples of five individuals per pool [35]. Between 3 and 10 g of feces were collected directly from the rectum of larger animals (cattle, sheep, goats and adult dogs). For smaller animals (cats and young dogs), three moistened cotton-tipped swabs were used to collect samples from the animal’s rectum and placed directly into transport media (two in modified Cary Blair and one in buffered glycerol saline); whole feces were not routinely collected from smaller animals.
For poultry, groups of birds of a single species (chickens or ducks) were confined overnight on a sheet of thick plastic. Owners were asked to confine approximately five birds per group, and not more than two groups of birds per species. Fecal specimens from a single pool of animals were mixed in a stool cup. Following thorough mixing of the pooled feces, two cotton-tipped swabs were inserted into the feces and then placed in a vial containing modified Cary Blair transport medium. A third swab was placed in a vial containing buffered glycerol saline. All specimen containers were clearly labelled and placed in a sealed bag in a coolbox with icepacks for transport to the laboratory.
Identification of potentially zoonotic enteric pathogens in animal specimens (Campylobacter jejuni, Campylobacter coli, non-typhoidal Salmonella, diarrheagenic E. coli, Cryptosporidium, Giardia, and rotavirus) was carried out using an identical protocol to that described for the human stool specimens tested in GEMS [28]. Briefly, bacterial agents were isolated and identified using conventional culture techniques. Three putative Escherichia coli colonies of different morphology types were pooled and analysed by multiplex PCR that detect targets for enterotoxigenic (ETEC), enteroaggregative (EAEC), enteropathogenic (EPEC), and enterohaemorrhagic E. coli (EHEC). The following gene targets defined each E. coli pathotype: ETEC (either eltB for heat-labile toxin [LT], estA for heat-stable toxin [ST], or both), ST-ETEC (either eltB and estA, or estA only), typical EPEC (bfpA with or without eae), atypical EPEC (eae without either bfpA, stx1, or stx2), EAEC (aatA, aaiC, or both), and EHEC (eae with stx1, stx2, or both, and without bfpA). Commercial immunoassays were used to detect rotavirus (ProSpecT Rotavirus kit, Oxoid, Basingstoke, UK), Giardia and Cryptosporidium spp. (TechLab, Inc., Blacksburg, VA, USA). Immunoassays were only performed on whole fecal specimens of adequate volume (≥ 3 g), and were therefore not completed for the majority of cat specimens, because volumes from this species were often inadequate.
To better understand the zoonotic potential, we genotyped Cryptosporidium parasites from immunoassay-positive animal fecal specimens. DNA was extracted from 0.5 ml of fecal specimens using a FastDNA SPIN Kit for Soil (MP Biomedicals, Santa Ana, CA). Cryptosporidium species present were differentiated by PCR-restriction fragment length polymorphism (RFLP) analysis of the small subunit (SSU) rRNA gene, and confirmed by DNA sequencing of the PCR products [36].
Data were analysed using R statistical software version 3.1.3 [37]. We used conditional logistic regression (clogit function applying the exact method in R package ‘survival’ [38]) with one-to-one matching to identify animal-related exposures that were significantly associated with MSD.
Exposure variables were screened for inclusion in the multivariable model using univariable conditional logistic regression. As part of the screening process, each exposure variable was evaluated for potential recoding. Husbandry-related variables for which values were conditional upon residence of the species in question were evaluated and recoded if this made biological sense. For example, the question “Do adult sheep enter the cooking area?’” was conditional on residence of adult sheep in the compound. If no adult sheep were resident, the response was recoded as “No–no adult sheep present” rather than a missing value, and compared against “No–adult sheep present but do not enter cooking area” and “Yes–adult sheep present and enter cooking area”. For these variables, the null state (species not resident) was taken as the reference level. Variables related to exposures of children to animals and their environments were kept as binary variables. For example, the question “Does the child play in an area of the compound where adult sheep defecate?” had one of two responses: ‘no’ if no adult sheep were resident in the compound or adult sheep were resident but the child did not play in the area where they defecated, and ‘yes’ if there were adult sheep resident and the child played in the area where they defecated. For categorical variables with four or more categories, we created new binary variables by combining categories based on frequencies. For example, the original four levels for frequency of observation of rodents or their excreta (never, seldom, often or daily) were dichotomised to never/seldom vs. often/daily. Both the original and new variables were tested in the univariable analysis. Continuous variables (e.g. number of chickens owned) were categorised into three categories [category 1: zero values; category 2: values greater than zero and less than or equal to the median value (excluding zeros); category 3: values greater than the median value (excluding zeros)]; both the original continuous variable and the new categorical variable were assessed in the univariable analysis. Variables with a significant number of missing values (>10% of observations) were discarded. Variables with a Wald test p-value greater than 0.2 on univariable analysis were excluded from further analyses. If both the original and recoded variable had a p-value below the threshold of 0.2, the one with the smaller p-value was retained.
After the univariable screening, we assessed collinearity between the selected exposure variables using condition indices (colldiag function in R package ‘perturb’ [39]). A condition index is a number ranging from 1 to infinity that is computed from data on a set of exposure variables–the higher the condition index, the greater the amount of collinearity [40]. The condition indices were investigated by calculating the variance decomposition proportion (VDP) for each condition index over 30, beginning with the largest. Exposure variables with a VDP >0.5 were considered potentially collinear. In cases where it made biological sense to do so, collinear variables were combined to create a new categorical variable. For example, the collinear variables “Chicken manure used in farm” and “Chicken manure used in the compound” were combined to create a variable “Chicken manure used”. When this did not make biological sense, or when the new variable still exhibited collinearity, the collinear variable with the higher univariable p-value was excluded. Remaining variables were taken forward for consideration in the multivariable conditional logistic regression model.
We compared main effects models using Akaike’s information criterion (AIC), whereby models with a smaller AIC are considered more optimal. We used a forward stepwise regression process to select exposure variables to retain in the model. Missing values were handled through multiple imputation (R package ‘mice’ [41]). Building of the main effects model was stopped when the addition of a variable resulted in an increase in the AIC. We assessed interactions between variables in the final main effects model by adding two-way interaction terms to the model and evaluating their effect on the AIC.
For evaluation of the final model, we identified outliers and influential pairs, using the transformation method described in [42] and applying a Bonferroni outlier test. We computed leverage values and delta β statistics to identify influential pairs (in R package ‘car’ [43]). To determine if these pairs were having an undue effect on the model, we refit the model with them omitted.
In GEMS, a wealth index quintile for households was generated by principle component analysis of thirteen household assets [26, 44]. The wealth index quintile was forced into the final model as an ordinal variable to evaluate the potential confounding effect of wealth.
The GEMS protocol was approved by the KEMRI Scientific and Ethical Review Committee (protocol no. 1155) and the Institutional Review Board at the University of Maryland, School of Medicine, Baltimore, MD, USA. The Centers for Disease Control and Prevention, Atlanta, GA, USA, formally deferred to the IRB at the University of Maryland for review (protocol no. 5038). Written informed consent was obtained from the parent or primary caretaker of each participant before initiation of study activities. The GEMS-ZED study protocol was approved by the KEMRI Scientific and Ethics Review Unit (protocol no. 1572) and the CDC Institutional Review Board (protocol no. 5683). Written informed consent for participation in the study was provided by the parent or primary caretaker of each participant, as well as from the head of the compound of residence of each participant. Protocols for animal involvement were reviewed and approved by the KEMRI and CDC Institutional Animal Care and Use Committees (protocol no. SSC 1572 and 2088OREMULX, respectively). CDC IACUC protocols comply with the Animal Welfare Act (AWA) regulations promulgated by the United States Department of Agriculture (USDA) under Title 9, Code of Federal Regulations, Parts 1–3 as well as the Public Health Service Policy on Humane Care and Use of Laboratory Animals (PHS Policy) administered by the National Institutes of Health (NIH), Office of Laboratory Animal Welfare (OLAW). In Kenya, all vertebrates are protected under Cap 360 (the Prevention of Cruelty to Animals Act) (1963, revised 1983).
A flow diagram showing the enrollment of children into the GEMS-ZED study is shown in Fig 1. Of the 90 children with MSD enrolled at Lwak Hospital from November 4th, 2009 through February 4th, 2011, 73 of their households participated in GEMS-ZED, along with 73 control households matched on age, sex and location of the case and control children. The median time between enrollment into GEMS and enrollment into GEMS-ZED was 4 days (range: 0–13 days).
Residence (presence/absence) of particular animal species did not differ significantly between case and control compounds based on the exact McNemar’s test values (Table 1). The wealth index quintile distribution also did not differ between case and control compounds (p = 0.4).
During the screening process, 497 exposure variables were evaluated (including recoded variables). Of these, 100 variables were discarded because they were not applicable or had >10% missing observations. Of the remaining 397 variables, 45 were selected after screening using univariable conditional logistic regression (Wald test p-value ≤ 0.2). Results of the univariable analysis for these variables are presented in S2 Table. After assessment of these variables for collinearity, and combination or exclusion of collinear variables, 37 variables were available for inclusion in the multivariable model (S3 Table). Results of the final model are shown in Table 2. All two-way interactions between variables in the final model were assessed; none resulted in a decrease in the AIC. We also tested for two-way interactions between age group and the main effects in the final model. No interaction terms were significant, meaning that the association between the main effects and MSD did not vary significantly by age group.
Variables that were associated with decreased risk of MSD were washing hands after animal contact, and presence of adult sheep that were not confined in a pen overnight. Variables that were associated with increased risk of MSD were increasing number of sheep owned, frequent observation of fresh rodent excreta (feces/urine) outside the house, and participation of the child in providing water to chickens. Inclusion of the wealth index did not result in a substantial change in the log odds ratio of the variables in the final model (<20% change).
In the evaluation of the final model, three pairs were detected as outliers or influential. When we refit the model with these pairs omitted, the same variables as in Table 2 remained in the final model, with the exception that the variable “Adult sheep sleeping in the pen” was replaced by the variable “Distance of sleeping area between child and adult sheep”. Compared with the reference level of no adult sheep, the matched adjusted odds ratio was 0.01 (95% CI 0–0.2) for a distance of 30m or more, and 0.05 (95% CI 0.01–0.04) for a distance of less than 30m.
We collected fecal specimens of acceptable quality for diagnostic testing from 2,174 domestic animals of eight species, resulting in a total of 691 pools (median of 5 and range of 1 to 10 pools per compound). Of these, 159 pools (23%) tested positive for one or more potentially zoonotic enteric pathogens (Campylobacter jejuni, C. coli, non-typhoidal Salmonella, diarrheagenic E. coli, Giardia, Cryptosporidium, or rotavirus). Test results for particular pathogens by host species and age group are given in Table 3. Species with the highest proportion of positive pools for particular pathogens were chickens for C. jejuni [18/231 (7.8%)] and non-typhoidal Salmonella [26/231 (11.3%)]; goats for C. coli [6/106 (5.7%)]; donkeys for diarrheagenic E. coli [1/12 (8.3%)]; dogs for Giardia [19/69 (27.5%)] and Cryptosporidium [4/69 (5.8%)]; and cattle for rotavirus [4/153 (2.6%)].
Domestic animals from 45/73 (61%) compounds at which a child with MSD resided tested positive to one or more pathogens, compared with 44/73 (60%) compounds with a control child. There were no significant associations on univariable conditional logistic regression between the presence of particular pathogens in domestic animals residing in compounds, and MSD in the participating child from the compound (Table 4). When considering the children’s GEMS laboratory results, we found 21 instances in which the pathogen identified in the child was also identified in one or more species of domestic animals residing in the compound (Table 5).
Nineteen pooled specimens positive for Cryptosporidium spp. by immunoassay were analysed by PCR, including 14 pooled specimens from chickens, 4 from dogs, and 1 from calves. Among them, 7 chicken specimens and the bovine specimen generated the expected PCR products. RFLP analysis indicated the presence of C. meleagridis in 6 chicken specimens, C. bovis in one chicken specimen, and C. parvum in one bovine specimen. None of the canine specimens analysed were positive by PCR.
We identified several animal-related factors associated with MSD in children younger than 5 years from compounds in rural western Kenya in which one or more species of domestic animals were resident. Children who reportedly washed their hands after contact with animals had significantly lower odds of MSD. Water, sanitation, and hygiene (WASH) interventions, including hand washing promotion, are shown to significantly reduce the risks of diarrheal illness in less developed countries [45, 46], but their effectiveness in reducing pathogen exposure specifically from domestic animals in these settings has not been explored. While the protective effect of hand washing has been demonstrated in outbreaks of enteric diseases associated with exposure to domestic animals in public settings [12, 13, 47], in their review Zambrano et al. [20] could find no studies that focused on WASH as a means of limiting disease transmission following domestic exposure to food-producing animals. Our study may be the first to report evidence of a protective effect of hand washing following exposure to household domestic animals in a developing country context. Hand washing after contact with animals may be a reflection of an overall higher frequency of hand washing in these children, and thus the protective effect may extend beyond (or be unrelated to) the risk of diarrheal illness after animal exposure. We recognise that a limitation of our study is reliance on self-reporting of behaviour, including hand washing.
Children from compounds that reported frequent observation of fresh rodent excreta outside the house had significantly higher odds of MSD. In a previous study in the area, a number of rodents were trapped in compounds, including a high proportion of black rats [34]. Rodents, and particularly rats, can be infected with pathogens that cause diarrheal illness in humans [48], including Salmonella Typhimurium [49, 50], Shiga-toxin producing E. coli [51] and Cryptosporidium parvum [52, 53]. Fresh rodent feces in areas of the compound may therefore be a source of exposure of children to these pathogens. Absence of rodent excreta could also be a reflection of better sanitation in these compounds, which may be associated with decreased risk of MSD independent of rodents.
Ownership and husbandry of sheep was found to be associated with MSD, but the nature of their role is not clear, with increasing numbers of sheep associated with increased odds, and not confining adult sheep in a pen overnight associated with decreased odds. Distance between children’s sleeping areas and where sheep are kept overnight may also play a role. Sheep are not a common livestock species in the study area, with only 18% of compounds owning sheep (compared with 49% owning cattle and 48% owning goats). Evidence from the literature of a specific role for sheep as risk factors for diarrheal illness in children is scant [54–57]. Consumption of mutton was found to be a risk factor for gastrointestinal illness in children and young adults in Isiolo, eastern Kenya [58]. In our study, we found a low prevalence of potentially zoonotic enteric pathogens in sheep feces (0% - 5%), with the exception of Giardia (21%). Giardia infection in children was not associated with MSD in GEMS [29].
Participation of the child in providing water to chickens was identified as a risk factor for MSD. In our study, a relatively high proportion of chicken fecal pools were positive for non-typhoidal Salmonella (11.3%), Campylobacter jejuni (7.8%) and diarrheagenic E. coli (7.6%). In their meta-analysis of six studies, Zambrano et al. [20] showed that poultry exposure more than doubled the odds of Campylobacter spp. infections in humans. Limiting exposure to household poultry, by for example corralling poultry, should therefore reduce the incidence of Campylobacter enteritis in children; however, in a randomized study to test this, Oberhelman et al. [59] found that rates of Campylobacter-related diarrhea were in fact significantly higher in children from households in which chickens were corralled, compared to those from households in which chickens were not confined. They speculated that this was due to the effect that corralling had on concentrating infected feces in one area, which would increase the risk of exposure to high doses of Campylobacter in children who entered corrals. Similarly, in our study we speculate that provision of water to chickens will be carried out mainly in situations where chickens are confined rather than free-ranging, increasing exposure of any accompanying children to enteric pathogens in the accumulated feces; however, we lack more detailed information on the nature of the reported exposure to substantiate this supposition. Active ingestion of chicken feces by infants has been observed in a rural African setting [60], highlighting the risk of zoonotic transmission of enteric pathogens.
In general, the prevalence of potentially zoonotic enteric pathogens in chicken feces in our study was lower than those reported in other studies in comparable settings [9, 24, 59, 61, 62]. Prevalence of zoonotic enteric pathogens in ruminants in our study was also lower when compared with other studies [24, 25, 61–65]. While this may be a reflection of differences in the diagnostic methods used, it could also be due to the extensive, subsistence nature of animal husbandry in our study site and the very small herd/flock sizes. We found no evidence of any association between the presence of particular pathogens in domestic animals and MSD in children, or of infection of children with the same pathogen species, although we note this was a pilot study with a small sample size, which may have limited our ability to detect associations. Enteric pathogens are often shed intermittently in the feces of carrier animals, so it is possible that carrier animals may not have been identified at the time of the specimen collection. The sensitivity of the microbiological methods used in children and in animals is low, as shown by a recent reanalysis of GEMS specimens using quantitative molecular diagnostic methods [66]. Even when the same pathogen species are found in children and in domestic animals in close contact, further characterization often shows genotypic differences between human and animal strains [24, 67, 68], although in some instances further subtyping provides support for zoonotic transmission [69]. In our study, most Cryptosporidium species identified from chickens and calves are pathogenic in humans, but further subtyping of species in child and animal specimens is needed to better understand the role of zoonotic transmission in cryptosporidiosis epidemiology.
We tested a large number of animal-related variables for an association with MSD in children. We recognise that with this many variables, significant associations may arise by chance, although the use of AIC in model selection should mitigate this. Furthermore, we do not infer a causal relation from the observed associations. We recommend that our results be used to generate hypotheses of causal links that can be tested in specific studies that address causal relations. These could include the role of sheep, chickens and rodents as risk factors for childhood diarrhea, and the application of WASH interventions to reduce risk. These studies should include established predictors of diarrhea in infants and young children, including breastfeeding and HIV status, in their causal models [70]. Future studies might further examine animal-related factors associated with environmental enteric dysfunction, as a number of zoonotic enteric pathogens have been found to be associated with this condition [71]. The use of quantitative molecular diagnostic methods in well-designed case-control and cohort studies of linked human and animal populations will also be important to understand the role of animals in domestic environments as reservoirs of human enteric pathogens.
|
10.1371/journal.pcbi.1000577 | Analyzing Short-Term Noise Dependencies of Spike-Counts in Macaque Prefrontal Cortex Using Copulas and the Flashlight Transformation | Simultaneous spike-counts of neural populations are typically modeled by a Gaussian distribution. On short time scales, however, this distribution is too restrictive to describe and analyze multivariate distributions of discrete spike-counts. We present an alternative that is based on copulas and can account for arbitrary marginal distributions, including Poisson and negative binomial distributions as well as second and higher-order interactions. We describe maximum likelihood-based procedures for fitting copula-based models to spike-count data, and we derive a so-called flashlight transformation which makes it possible to move the tail dependence of an arbitrary copula into an arbitrary orthant of the multivariate probability distribution. Mixtures of copulas that combine different dependence structures and thereby model different driving processes simultaneously are also introduced. First, we apply copula-based models to populations of integrate-and-fire neurons receiving partially correlated input and show that the best fitting copulas provide information about the functional connectivity of coupled neurons which can be extracted using the flashlight transformation. We then apply the new method to data which were recorded from macaque prefrontal cortex using a multi-tetrode array. We find that copula-based distributions with negative binomial marginals provide an appropriate stochastic model for the multivariate spike-count distributions rather than the multivariate Poisson latent variables distribution and the often used multivariate normal distribution. The dependence structure of these distributions provides evidence for common inhibitory input to all recorded stimulus encoding neurons. Finally, we show that copula-based models can be successfully used to evaluate neural codes, e.g., to characterize stimulus-dependent spike-count distributions with information measures. This demonstrates that copula-based models are not only a versatile class of models for multivariate distributions of spike-counts, but that those models can be exploited to understand functional dependencies.
| The brain has an enormous number of neurons that do not work alone but in an ensemble. Yet, mostly individual neurons were measured in the past and therefore models were restricted to independent neurons. With the advent of new multi-electrode techniques, however, it becomes possible to measure a great number of neurons simultaneously. As a result, models of how populations of neurons co-vary are becoming increasingly important. Here, we describe such a framework based on so-called copulas. Copulas allow to separate the neural variation structure of the population from the variability of the individual neurons. Contrary to standard models, versatile dependence structures can be described using this approach. We explore what additional information is provided by the detailed dependence. For simulated neurons, we show that the variation structure of the population allows inference of the underlying connectivity structure of the neurons. The power of the approach is demonstrated on a memory experiment in macaque monkey. We show that our framework describes the measurements better than the standard models and identify possible network connections of the measured neurons.
| So far, it is still unknown which statistics are crucial for analysis in order to understand the neural code. One approach is to analyze simultaneous spike-counts of neural populations. Responses from populations of sensory neurons vary even when the same stimulus is presented repeatedly, and the variations between the simultaneous spike-counts are usually correlated (noise correlations) at least for neighboring neurons. These noise correlations have been subject of a substantial number of studies (see [1] for a review). For computational reasons, however, these studies typically assume Gaussian noise. Thus, correlated spike rates are generally modeled by multivariate normal distributions with a specific covariance matrix that describes all pairwise linear correlations.
For long time intervals or high firing rates, the average number of spikes is sufficiently large for the central limit theorem to apply and the normal distribution is a good approximation for the spike-count distributions. Several experimental findings, however, suggest that processing of sensory information can take place on shorter time scales, involving only tens to hundreds of milliseconds [2],[3]. In this regime the normal distribution is no longer a valid approximation:
Though not widespread for modeling spike-counts, alternative models have been proposed in previous studies that have Poisson distributed marginals and separate parameters for higher order correlations, e.g. the multiple interaction process model [6] and the compound Poisson model [7]. Both models are point processes. In terms of their induced spike-count distributions these models are special cases of the multivariate Poisson latent variables distribution first introduced by Kawamura [8] and presented in a compact matrix notation by Karlis and Meligkotsidou [9]. Similar to the multivariate normal distribution this model has also a couple of shortcomings for spike-count modeling: (1) Only Poisson-marginals can be modeled. (2) Negative correlations cannot be represented. (3) The dependence structure is inflexible: features like tail dependence cannot be modeled.
We use and extend a versatile class of models for multivariate discrete distributions that overcome the shortcomings of the afore-mentioned models [10],[11]. These models are based on the concept of copulas [12], which allow to combine arbitrary marginal distributions using a rich set of dependence structures. In neuroscience they were also applied to model the distribution of continuous first-spike-latencies [13].
Figure 1 illustrates the copula concept using spike-count data from two real neurons. Figure 1A shows the bivariate empirical distribution and its two marginals. The distribution of the counts depends on the length of the time bin that is used to count the spikes, here . In the case considered, the correlation at low counts is higher than at high counts. This is called lower tail dependence [12]. Figure 1B shows the discretized and rectified multivariate normal distribution. On the other hand, the spike-count probabilities for a copula-based distribution (Figure 1C) correspond well to the empirical distribution in Figure 1A.
The paper is organized as follows. The next Section “Materials and Methods” contains a description of methodological details regarding the multivariate normal distribution, the multivariate Poisson latent variables distribution, the copula approach for spike-counts and the model fitting procedures. In this section we will also introduce a novel transformation for copula families. The method is innovative and yields a novel result. We will then describe the computational model used to generate synthetic data and the experimental recording and analysis procedures. In the Section “Results” copula-based models will be applied to artificial data generated by integrate-and-fire models of coupled neural populations and to data recorded from macaque prefrontal cortex (PFC) during a visual memory task. The appropriateness of the models is also investigated. The paper concludes with a discussion of the strengths and weaknesses of the copula approach for spike-counts.
All procedures were approved by the local authorities (Regierungspräsidium) and are in full compliance with the guidelines of the European Community (EUVD 86/609/EEC) for the care and use of laboratory animals.
The multivariate normal (MVN) distribution is characterized by a probability density over continuous variables and its cumulative distribution function (CDF) with mean and covariance matrix is given by
In order to apply it to spike-count distributions (which are discrete and non-negative) it is discretized and rectified (probability for negative values is set to zero). Its CDF is given bywhere denotes the floor operation for the discretization. The probability mass function will have peaks at the zero count rows, due to the rectification of the CDF. It would be desirable to distribute the cut off mass equally to the complete domain. However, this is infeasible for more than three dimensions, because the necessary normalization term is computationally too time-consuming. Note that is no longer the mean of the distribution corresponding to , because the mean is shifted to larger values as is rectified. This shift grows with the dimension .
The Poisson latent variables distribution is characterized by a probability mass function (PMF) over non-negative integer variables [9]. A random variable with this distribution is composed of latent variables . These latent variables are independent univariate Poisson distributed with rates takes the form , where is a mixture matrix. The PMF of is then given by
When we set to we can vary all pairwise and higher order interactions separately using the rates of the latent variables. However, only non-negative correlations can be modeled, because the rates of the latent variables are non-negative. Furthermore, the are marginally Poisson distributed.
A copula is a cumulative distribution function (CDF) which is defined on the unit hypercube and which has uniform marginals [12]. Formally, a copula is defined as follows:
Our goal is to construct multivariate distributions for simultaneously recorded spike-counts that can model a wide range of dependence structures. Copulas make it possible to model multivariate distributions based on two distinct parts: the distributions of the individual elements and the dependence structure. Let us now assume that represents the spike-count of neuron within a given interval. According to Theorem 1 we can then describe the joint cumulative distribution function of the spike counts by choosing a copula from a particular family, and by setting and . are the models of the marginal distributions, i.e. the cumulative distributions of spike-counts of the individual neurons. Often, the Poisson distribution is a good approximation to spike-count variations of single neurons [16], hence the CDFs of the marginals take the form is the mean spike-count of neuron for a given bin size. A more flexible marginal is the negative binomial distribution,which allows to model spike-count distributions showing overdispersion. Here is the gamma function, is again the mean spike-count of neuron , and is a positive parameter, which controls the degree of overdispersion. The smaller the value of , the greater is the Fano factor, and as approaches infinity, the negative binomial distribution converges to the Poisson distribution.
The second part of the model is the copula family. Many different families have been discussed in the literature in the past. Families differ by the number of free parameters and by the class of dependence structures they can represent. The most simplistic copula is the product copula defined as for which independence is attained. We selected a number of useful copula families (see Table 1). Figure 2 shows their bivariate probability density functions (PDFs).
The Clayton family has a so-called lower tail dependence: the correlation between its elements is higher for low values than for high values (see Figure 2A). The scalar parameter controls the strength of dependence. Note that does not only control the strength of pairwise interactions but also the degree of higher order interactions. We define .
The Gumbel-Hougaard (short Gumbel) family has an upper tail dependence. Here, the region of high correlation is in the upper right corner of the density. Hence, the correlation between its elements is higher for high values than for low values (see Figure 2B). The scalar parameter controls the strength of dependence.
The Frank family has no tail dependence. There is no difference between the correlation for low and for high values (see Figure 2C). Again, the scalar parameter controls the strength of dependence and we define .
The Ali-Mikhail-Haq (AMH) family models are positively ordered, i.e. for it holds for all (see Figure 2D). Again we define .
The Farlie-Gumbel-Morgenstern (FGM) family has parameters that individually determine the pairwise and higher order interactions. It has parameters less than the Poisson latent variables distribution because the rates of the neurons can be parametrized by the marginals. Non-zero values of the parameter indicate the presence of order interaction. For order interactions are absent. If, for example all for , the corresponding probability distribution includes only parameters of second order, similar to the multivariate normal distribution. The constraints on the parameters , however, constrain the corresponding correlation to be small in terms of their absolute value.
We now introduce a novel extension of standard copula models, which is particularly useful for modeling distributions of spike-counts. It is based on the orthant dependence concept. Here, an orthant refers to one of the hypercubes of equal size in the unit hypercube, i.e. a “corner” of the copula distribution. Let us consider a distribution with a so-called lower tail dependence (see Figure 3A), i.e. a distribution, for which the correlation between spike-counts of two neurons is higher for low values than for high values. We now introduce the flashlight transformation which allows to shift the region of high correlation to an arbitrary orthant (see Figure 3B–D). The whole dependence structure between spike-counts is rotated accordingly, but remains unchanged otherwise. The transformation is a function that operates on CDFs. Yet, it rotates the corresponding PDF, not the CDF.
The flashlight transformation is specified in the following theorem (see Text S1 in the supplementary material):
Once a family of marginal distributions and a family of copulas for describing the dependence structure has been selected, model parameters have to be estimated from the data, i.e. from the empirical distribution. Here we suggest a method which is similar to maximum likelihood estimation.
Theorem 1 provides a method to construct multivariate CDFs based on copulas. Therefore, the approach yields a CDF of a multivariate distribution. In order to calculate the likelihood we have to transform the CDF to a probability mass function (PMF).
For this purpose we define the sets and , . The probability of a particular set of spike-counts can then be expressed using only the CDF , making use of the so-called inclusion-exclusion principle of Poincaré and Sylvester [19]:(2)
Letdenote the sum of log likelihoods of the marginal distribution , where are the parameters of the chosen family of marginals. Furthermore, letbe the log likelihood of the joint probability mass function, where denotes the parameter of the chosen copula family. The so-called inference for margins (IFM) method [20] now proceeds in two steps. First, the marginal likelihoods are maximized separately:
Then, the full likelihood is maximized given the estimated marginal parameters:
It was shown that the IFM estimator is asymptotically efficient [20]. The estimator is computationally more convenient than the maximum likelihood estimator, because parameter optimization in low dimensional parameter spaces needs less computation time.
Depending on whether the copula parameters are constrained, either the Nelder-Mead simplex method for unconstrained nonlinear optimization [21] or the line-search algorithm for constrained nonlinear optimization [22] can be applied to estimate the copula parameters using Eqn 2 as the objective function.
For mixtures of copulas, where the values of the latent variables have to be estimated in addition, we suggest to use the expectation-maximization algorithm [23],[24]. In the expectation step, the weights are updated usingwhere is the PMF of the model based on the copula . In the maximization step the copula parameters are determined for fixed values of by applying the IFM method. Both steps are repeated until parameter values converge.
The leaky integrate-and-fire neuron is a simple neuron model that models only subthreshold membrane potentials. The equation for the membrane potential is given bywhere denotes the resting membrane potential, is the total membrane resistance, is the synaptic input current, and is the time constant. The model is completed by a rule which states that whenever reaches a threshold , an action potential is fired and is reset to [25]. In all of our simulations we used , , , , and initialized with . These are typical values that can be found in [25].
Current-based synaptic input for an isolated presynaptic release that occurs at time can be modeled by the so-called -function [25]:
The function reaches its peak at time and then decays with time constant . We can model an excitatory synapse by a positive and an inhibitory synapse by a negative . We used for excitatory synapses, for inhibitory synapses, and .
Neural activity was recorded from the lateral prefrontal cortex within an area of located on the ventral bank of the principal sulcus of an adult female rhesus monkey (macaca mulatta). Recordings were performed simultaneously from up to adjacent sites with an array of individually movable fiber micro-tetrodes (manufactured by Thomas Recording) with an inter-tetrode distance of . Data were sampled at and bandpass filtered between and . Recording positions of individual electrodes were chosen to maximize the recorded activity and the signal quality. The recorded data were processed by a principal component analysis-based spike sorting method. Automatic cluster cutting was manually corrected by subsequent cluster merging if indicated by quantitative criteria such as the ISI-histograms or amplitude stability.
Activity was recorded while the monkey performed a visual working memory task. One out of visual stimuli (fruits and vegetables) were presented for approximately . After a delay of , during which the monkey had to memorize the sample, a test stimulus (“test”) was presented and the monkey had to decide by differential button press whether both stimuli were the same or not. Correct responses were rewarded. Match and non-match trials were randomly presented with equal probability ().
The mutual information between spike-counts and stimuli is given by(3)where is the set of stimuli, is the probability distribution over the stimuli, and is the likelihood of a neural response given a stimulus . For higher dimensions the sum over prohibits an exact computation of , since the number of terms of the sum grows exponentially with . The evaluation of this sum is therefore practically infeasible unless the number of neurons is very small. However, we can estimate the mutual information using Monte Carlo sampling. For each of the stimuli , we can estimate the second sum by drawing samples with probability . The termwill then converge to the second sum in Eqn 3, as approaches infinity [26].
Typically the number of samples that can be obtained in electro-physiological experiments is small. Thus, it might appear to be hopeless to estimate a multidimensional model with a detailed dependence structure. However, since our marginal distributions are discrete the copula matters only at a small number of points. In the following, we will demonstrate that it is not always necessary to obtain a great number of samples for a reliable model estimation. For this purpose we selected the Clayton-copula model with negative binomial marginals as a ground truth model which was used to draw samples. We calculated the deviation of the log likelihood of the estimated model from the log likelihood of the ground truth model in percent of the ground truth log likelihood. The correlation strength of the ground truth model was varied by the Clayton parameter. The results are shown in Figure 4 for three different Clayton parameters of the ground truth model. For moderate dependence strengths (as are typically found in the data) samples were sufficient for estimations of the log likelihood with an error of less than .
One cause for dependence between spike-counts of different neurons are common input populations. Therefore, we investigated network models with different types of common input. We set up two current based leaky integrate-and-fire neurons (see Section “Materials and Methods”) and three input populations modeled as Poisson spike generators. The left input population projected only to neuron 1 and the right input population projected only to neuron 2. The center input population was the common input population, projecting to both neuron 1 and neuron 2. We investigated all four combinations of excitatory (E) and inhibitory (I) projections from the common population to the two neurons (see Figure 5A1–A4).
In this network model a lower tail dependence should arise if the projections from the common input projection are mostly inhibitory: each time the common population is active the firing rates of both neurons will decrease simultaneously. Therefore, only low spike-counts should be strongly correlated and the Clayton family should provide a good fit to the responses of such a network. Similarly, two excitatory projections should result in an upper tail dependence and other combinations should become apparent as dependence blobs in other corners of the probability density function of the copula. The flashlight transformation shifts the dependence blob of a given copula with orthant dependence into other orthants of the probability density function and is thus capable of modeling different types of common input populations in a stochastic manner. For two neurons, the lower left corner models an inhibitory input population, the upper right corner models an excitatory input population, and the other corners model a combination of excitatory and inhibitory input populations.
The spike trains of the two neurons were binned into intervals. We applied copula-based models with negative binomial marginals to fit the generated data from the four models using the IFM method (see Section “Model Fitting”). Four different copula families were applied: the unmodified bivariate Clayton family and the three remaining flashlight transformations of the Clayton family (Figure 3). Figure 5C1–C4 shows the log likelihoods of the fits for the corresponding networks as shown in Figure 5A1–A4. The respective model performed best for the combination of projection types of the common input population it was supposed to model, i.e. Clayton for I-I, Clayton survival for E-E, etc. Hence, by determining the best fitting transformation the most likely combination of input types could be identified. Each of the transformations could be associated with a distinct combination of projection types.
To investigate whether the results of the reconstruction depend on the strengths of the synapses we varied between and for excitatory synapses and between and for inhibitory synapses (data not shown). While the relation of the best fitting copula families was constant across all strengths the differences between the curves decreased for decreasing strengths. For it was hard to distinguish between the likelihoods of lower and upper tail dependencies. Therefore, tail dependencies were less pronounced in the spike-counts. In the multi-tetrode data, however, we found significant differences between the likelihoods of the copula families (see Section “Application of Copula-Based Models to Multi-Tetrode Data”).
To investigate the impact of the bin size on the reconstruction performance we also binned the data into smaller and larger intervals (data not shown). When the bin size was too small or too large ( and ) the reconstruction did not succeed. In the intermediate range (, ), however, the connection types could be reconstructed. This can be explained by the asymptotic distributions of the multivariate spike-counts. According to the central limit theorem the multivariate normal distribution provides a good approximation when the bin size is sufficiently large. Hence, tail dependencies will vanish. On the contrary, when the bin size becomes too small the marginal distributions are essentially Bernoulli distributed and the tail dependencies will vanish as well. Of course, the range of the intermediate bin size depends on the rates of the neurons. The larger the rates the smaller the bin sizes in the intermediate range. For the simulated data the rates were comparable to the data recorded from the prefrontal cortex (see Section “Multi-Tetrode Recordings”).
Our copula-based models are capable of modeling different dependence structures with marginals that are tailored to single neuron spike-count distributions. Thus, we expected that the copula-based models would provide a much better fit to data recorded from real neurons than the multivariate normal distribution or the multivariate Poisson latent variables distribution. To test this, we applied copula-based models from different families and with different marginal distributions to data, which has been recorded from macaque prefrontal cortex for each of the twenty presented stimuli and each of the four phases (pre-stimulus presentation, stimulus presentation, delay, presentation of the test stimulus) of the visual working memory task. We compared the results to models of the discretized multivariate normal and the Poisson latent variables distribution (see Section “Materials and Methods”)
We randomly selected count vectors for each task phase and each stimulus as the validation set. We then estimated the model parameters on the remaining count vectors (training set) and used the validation set for obtaining an unbiased estimate of the likelihoods of the selected models.
We used the IFM-estimator for the copula-based models and the maximum likelihood estimator for the Poisson latent variables distribution. The parameters and of the discretized MVN distribution were estimated by the sample mean and the sample covariance matrix of the spike-counts. This procedure does not correspond to the maximum likelihood estimate of the discretized distribution. We used it, because the maximum likelihood estimator was too expensive to compute for six neurons. The high computational costs come from the estimation of the CDF of the MVN.
The rate parameter for the Poisson distribution and negative binomial distribution were estimated via the sample mean. The maximum likelihood estimates for the overdispersion parameter were computed iteratively by Newton's method.
Figure 6A summarizes the results for the discretized MVN, the Poisson latent variables distribution, and two copula-based distributions with different marginals, the Poisson distribution, and the negative binomial distribution. The negative binomial distribution provided for all four task phases a significantly better fit than the Poisson distribution, the MVN distribution, and the Poisson latent variables distribution. The likelihood for the copula-based models was significantly greater than for the discretized MVN model (, paired-sample Student's t test over stimuli) and the Poisson latent variables model (). Moreover, the likelihood for the negative binomial marginals was even greater than that for the Poisson marginals (). Thus, the copula-based approach provided models that were indeed superior for the data at hand. Moreover, the additional flexibility of the negative binomial marginals improved the fit significantly.
We applied different copula families to examine the importance of the dependence structure for the model fit. Figure 6B shows an evaluation of the different copula families with different dependence structures for the best fitting marginal, which was the negative binomial distribution. The model based on the Clayton copula family provided the best fit. The fit was significantly better than for the second best fitting copula family (), the Gumbel family. In spite of having more parameters, the FGM copulas performed worse. However, the FGM model with third order interactions fitted the data significantly better than the model that included only pairwise interactions ().
The best fitting copula-based model, the Clayton copula, is characterized by a lower tail dependence. Apart of the Gumbel family, the other families that we applied so far do not model orthant dependencies. To check whether other orthant dependencies would improve the fit, we applied the flashlight transformation and we transformed the Clayton copula tail towards all corners of the six dimensional hyper cube. The results are shown in Figure 7. The standard Clayton copula with lower tail dependence had the significantly highest value of the log likelihood on the validation set indicating that the empirical spike-count distribution has indeed a lower tail dependence. The second highest peak was reached by the Clayton survival copula. The central peak corresponded to those transformations that were close to the Clayton and the Clayton survival copulas: sectors and ( and decimal). Thus, a common lower tail dependence was prominent in the data.
We applied mixtures of copulas as described in Section “The Flashlight Transformation and Mixtures of Copulas” to check whether there was indeed a prominent common upper tail dependence beside the lower tail dependence in the data. Therefore, we fixed the Clayton copula (which models a lower tail dependence) as the first mixture component and varied the sector of the flashlight transformed Clayton copula for the second mixture component. Figure 7C shows the mean log likelihoods of the mixture models with negative binomial marginals on the same data set used for Figure 7B. All of the mixture models exhibit similar performance. Therefore, the upper tail dependence that we observed for the unmixed model appears to be an artifact of the lower tail dependence.
In summary, we could show that the copula-based approach provided a significant improvement in the goodness of fit compared to the discretized and rectified multivariate normal distribution and the Poisson latent variables distribution. Moreover, the dependence structure alone has a significant impact as well.
Our model consists of two parts: 1) the copula and 2) the marginals. We already analyzed the effect of the copula. In this section we describe the investigation of the marginals. In particular, we are interested in understanding how the goodness of fit is influenced by the marginals. For this purpose we compared the log likelihoods of the Clayton-copula model with Poisson, negative binomial, and empirical marginals fitted to the training set of the sample stimulus presentation phase. The model with empirical marginals was a so-called semiparametric distribution consisting of a parametric dependence structure (the copula family) and nonparametric marginals. We drew samples from these distributions in order to learn whether the training and validation sets were typical samples from the fitted distributions. For a complex model we expect the likelihood of training samples to be close to the mode of the histogram, while we expect the validation samples to have a much smaller likelihood. Contrary, for a model with small complexity we expect the likelihood of the training samples to be close to the likelihood of the validation samples. When the complexity is too small we expect the likelihoods of the training and the validation samples to be much smaller than the mode of the histogram.
In our setting the most complex model is the one with empirical marginals. Histograms of the log likelihoods for copula models with the three different marginals are shown in Figure 8. For Poisson marginals, the log likelihoods of both the training set and the validation set were much smaller than the log likelihoods of the samples drawn from the fitted distribution. Thus, the Poisson marginals seem to be too simple for a good fit to the data, whereas the negative binomial marginals generalized well in spite of their increased complexity. On the training set the model with the empirical marginals performed best. However, there was a huge discrepancy to the likelihood of the model with empirical marginals on the validation set, whereas the likelihoods of the other two models did not change much. This result can be explained by overfitting. The empirical marginals matched the marginals of the training set perfectly. The empirical marginals of the training set, however, were noisy representations of the true marginals, because of the limited sample size. Hence, a perfect fit is not beneficial when it comes to novel data. In contrast to that, the likelihoods of the models with Poisson and negative binomial marginals were almost equal to the respective likelihoods on the training set. Thus, these models did not suffer from overfitting.
In order to relate these findings to the number of samples in our training set we can compare the number of samples to the estimated number of required samples for the toy example in Section “Model Fitting”. Figure 6 shows that the log likelihood for the Clayton-copula model deviated from the second best family by . In Section “Model Fitting” we showed that for this model samples were sufficient for good estimations of the log likelihood. For the delay phase and for the test stimulus phase, the number of samples varied between and per stimulus. Therefore, the number of samples was sufficient for these phases. Taken together with the histogram analysis, we found that the model complexity was appropriate for the available amount of data at hand.
We will now show that the copula-based models can be used to measure the short-term information about a stimulus that is encoded by the spike-count dependence structure of the recorded neurons. The first step is to estimate the total information of the spike-count responses. We applied the best fitting copula model, the Clayton-copula model with negative binomial marginals, to estimate the mutual information between stimuli and responses via Monte Carlo sampling (see Section “Materials and Methods”). Figure 9A shows the estimated mutual information for each of the four task phases. The mutual information was greater during the sample stimulus interval and the test stimulus interval than during the delay interval. Therefore, a stimulus presentation evoked a spike-count response which instantly encoded information about the stimulus. In the test stimulus phase the dotted line is above the dashed line, so the spike-counts coded more information about the sample stimulus that was previously presented than about the test stimulus.
Figure 9B shows the information estimate , normalized to the mutual information that is shown in Figure 9A. The dependence structure carried between and of the mutual information. During the test stimulus interval the dependence structure encoded almost twice as much information about the test stimulus as about the sample stimulus that was previously presented.
We developed a framework for analyzing the noise dependence of spike-counts and used synthetic data from a model of leaky integrate-and-fire neurons to derive interpretations for different dependence structures. Applying the framework to our data from the macaque prefrontal cortex we found that: (1) copula-based models with negative binomial marginals rather than the multivariate normal distribution or the Poisson latent variables distribution are appropriate models of spike-count data for short time intervals; (2) the dependence structure encodes between and of the mutual information about the presented stimuli; (3) the amount of data required for a good likelihood estimation is present in our data set; and (4) a lower tail dependence between all neurons is present in the data and can be explained by common inhibitory input.
The copula approach has many advantages compared to previous models. Recently, the Ising model gained a lot of attention in neuroscience [4],[27]. This model is a maximum entropy model of binary variables called spins that have only pairwise interactions [28]. The model is applied to the neuroscience setting by binning spike trains into very short time intervals such that at most one spike falls into each bin. The spin for that bin then indicates whether or not a spike was present. Using this model pairwise interactions between simultaneously recorded neurons can be modeled [4]. The Ising model is a special case of a more general class of nested maximum entropy models [29]. Other models in this class can be used to model higher order interactions between neurons. Nevertheless, an independence assumption for subsequent bins is necessary due to the limited number of samples present in typical neuroscience settings. Therefore, the marginal spike-counts of individual neurons will be binomial distributed. The variance of this distribution is always smaller than its mean which is a severe disadvantage of this model class. The copula approach on the other hand can model arbitrary marginals.
Another class of models are doubly stochastic models where some parameters of the data distribution are themselves random variables. The doubly stochastic Poisson point process presented by Krumin and Shoham belongs to this class [30]. For such models the marginal distributions change whenever the dependence is modified. It is thus very hard to disentangle the effects of the dependence structure from the effects of the marginals.
In contrast to the multivariate normal distribution and the multivariate Poisson latent variables distribution the copula approach can be used to model arbitrary marginal distributions that are appropriate for the data at hand. The marginal distributions can therefore be discrete without any mass on the negative axis and with variance greater than the mean. We compared the fits of negative binomial marginals to Poisson and empirical marginals and found that only the negative binomial marginals provided a reasonable fit to the data. Contrary to the Poisson marginals, the negative binomial marginals were complex enough such that likelihoods of samples from the model were consistent with the likelihood of the data. Moreover, the negative binomial marginals did not suffer from overfitting as did the empirical marginals. We conclude that the negative binomial marginals are appropriate to describe the spike-counts recorded from the prefrontal cortex.
The dependence structure of the copula approach is flexible. Higher order interactions can be parametrized separately if desired. Furthermore, in contrast to the multivariate Poisson latent variables distribution, negative correlations can be modeled as well. Another advantage of the copula approach is that it is modular in the sense that the copula family used for the data analysis can be easily exchanged by another family. Many different copula families exist, each representing and parameterizing different properties of the dependence structures. Thus, it is easy to test for different properties of a distribution. Specific examples are the Clayton and Gumbel families. These families have lower and upper tail dependencies, respectively. Lower and upper tail dependencies can arise from common input populations with inhibitory and excitatory projections, respectively. By deriving the flashlight transformation we could construct additional families that account for combinations of inhibition and excitation.
When applying the flashlight transformation to the data from the prefrontal cortex, we found that the unmodified Clayton family provided the best fit to the test data. Therefore, a common lower tail dependence to all neurons is present in the data. One explanation is a common input population whose projections are mostly inhibitory to all the analyzed neurons. Two types of common inhibitory sources are possible: (1) A local source of inhibitory input such as common interneurons. (2) Another area projecting to the prefrontal cortex. It was found that interneurons have a reach of no more than a few hundred micrometers whereas the inter-tetrode distance was . Thus, it is unlikely that a population of common interneurons inhibits all the stimulus specific neurons that we recorded from. Another area, therefore, is more likely to be the source of the common inhibitory input. One possibility could be the ventral tegmental area (VTA). In the rat cortex it was found that the VTA exerts a direct inhibitory influence on the PFC. In a study of recorded PFC neurons were inhibited as a result of VTA stimulation [31]. Moreover, the VTA is thought to be a central component of the reward system [32] which is essential for a memory task. Our analysis provides evidence for such an influence based on the spike-count statistics.
The second best fit was achieved by the Clayton survival family. One explanation for this result is provided by an upper tail dependence between all neurons in addition to the stronger lower tail dependence. We applied mixtures of copulas to elucidate this issue and found that a mixture of the Clayton and Clayton survival family did not provide the best fit out of all mixtures of the Clayton family with a Clayton flashlight transformation. At first sight it is puzzling that the upper tail dependence seems to disappear when mixed with the lower tail dependence. However, the Clayton copula and the Clayton survival copula have their dependence along the same line in the six dimensional space that is spanned by the neuronal spike-counts, though predominantly at different ends of this line. Hence, the Clayton survival family can capture some of the dependence that is inherent to the Clayton family. We conclude that the prominence of the upper tail dependence that was observed for the unmixed model is an artifact of the lower tail dependence component.
The results show that important properties of dependence structures such as tail dependencies arise very naturally in simple input scenarios, and that the copula approach can be used to construct generative models that are capable of capturing these aspects of this underlying connectivity. In principle, copula-based models can be used to guide reconstructions of functional connectivity, but this topic is outside the scope of this study. If the reader is interested in detailed reconstruction of functional connectivity we recommend the studies in [33]–[35] as a starting point.
We could show that there is important information represented in the dependence structure which has been ignored in studies reporting only the correlation coefficient. Based on the flashlight transformation we could derive novel copula families with interesting interpretations for neuroscience: the statistical dependence gives insight into possible connections of the underlying network. Other copula families might be applicable to investigate different properties of the network.
We could also show that the Gaussian distribution is not an appropriate approximation of the spike-count distribution of short time intervals. Yet, many studies applied this approximation in their investigations. Therefore, these studies should be reassessed with respect to their validity for short-term coding.
We also compared the copula-based approach to the multivariate Poisson latent variables distribution. In terms of spike-counts this model corresponds to previous point process models that account for higher order correlations. The copula-based approach overcomes a number of shortcomings of this distribution, namely the Poisson marginals, the restriction to non-negative correlations and the inflexible dependence structure. We could show that the improvement in the goodness-of-fit is significant.
Taken together, the copula-based approach allows us to model and analyze spike-count dependencies in much more detail than previously applied models. A drawback is the small number of neurons to which the approach can be applied so far. The approach is computationally too demanding for higher numbers of neurons because the model fitting complexity is exponential in the number of neurons. Approximate inference methods might provide a solution to the computational problem. However, another problem is the number of samples available in typical electro-physiological experiments. We could show that samples are sufficient for six dimensional data with moderate dependence strengths. Nevertheless, the amount of required data increases dramatically for increasing dimensions, i.e. for the number of neurons. A combination with dimensionality reduction techniques might provide a solution to this problem.
|
10.1371/journal.pgen.1004813 | Genetic Analysis of the Cardiac Methylome at Single Nucleotide Resolution in a Model of Human Cardiovascular Disease | Epigenetic marks such as cytosine methylation are important determinants of cellular and whole-body phenotypes. However, the extent of, and reasons for inter-individual differences in cytosine methylation, and their association with phenotypic variation are poorly characterised. Here we present the first genome-wide study of cytosine methylation at single-nucleotide resolution in an animal model of human disease. We used whole-genome bisulfite sequencing in the spontaneously hypertensive rat (SHR), a model of cardiovascular disease, and the Brown Norway (BN) control strain, to define the genetic architecture of cytosine methylation in the mammalian heart and to test for association between methylation and pathophysiological phenotypes. Analysis of 10.6 million CpG dinucleotides identified 77,088 CpGs that were differentially methylated between the strains. In F1 hybrids we found 38,152 CpGs showing allele-specific methylation and 145 regions with parent-of-origin effects on methylation. Cis-linkage explained almost 60% of inter-strain variation in methylation at a subset of loci tested for linkage in a panel of recombinant inbred (RI) strains. Methylation analysis in isolated cardiomyocytes showed that in the majority of cases methylation differences in cardiomyocytes and non-cardiomyocytes were strain-dependent, confirming a strong genetic component for cytosine methylation. We observed preferential nucleotide usage associated with increased and decreased methylation that is remarkably conserved across species, suggesting a common mechanism for germline control of inter-individual variation in CpG methylation. In the RI strain panel, we found significant correlation of CpG methylation and levels of serum chromogranin B (CgB), a proposed biomarker of heart failure, which is evidence for a link between germline DNA sequence variation, CpG methylation differences and pathophysiological phenotypes in the SHR strain. Together, these results will stimulate further investigation of the molecular basis of locally regulated variation in CpG methylation and provide a starting point for understanding the relationship between the genetic control of CpG methylation and disease phenotypes.
| Epigenetic marks provide information that is not encoded in the primary DNA sequence itself but in modifications of genomic DNA and of the associated proteins. Methylation of genomic DNA at cytosine residues is an important epigenetic modification that is associated with developmental processes, carcinogenesis and other diseases. Genome-wide extent of, and reasons for inter-individual differences in cytosine methylation, and their association with phenotypic variation are poorly characterised. To address these questions we have determined and compared the genome-wide methylation patterns in heart tissue of two inbred rat strains, the spontaneously hypertensive rat, an animal model of human disease and a control rat strain. Comparison of methylation differences between genetically identical animals from the same strain and differences between animals from different strains allowed us to quantify association of epigenetic and genetic differences. We show that differences in an individual's germline DNA sequence are important determinants of the variability in methylation between individuals. Comparison with previous reports implicates common mechanisms for regulation of cytosine methylation that are highly conserved across species. Finally, we find correlation between a proposed blood biomarker for heart failure and variation in DNA methylation, suggesting a link between germline DNA sequence variation, methylation and a disease-related phenotype.
| Cytosine methylation at CpG dinucleotides is a key epigenetic mark with an essential role in regulating gene expression and other cellular and whole body phenotypes. While the molecular mechanisms for de novo and maintenance methylation of CpG cytosines are well established [1]–[3], allele-specific influences on CpG methylation have been documented [4]–[12], and association of genotype and epigenotype has been shown in plants recently [13], the extent of, and reasons for inter-individual differences in cytosine methylation, and their association with phenotypic variation in mammals are poorly characterised.
Isogenic inbred lines provide a powerful platform for establishing relationships between the germline genome and downstream phenotypes. The ability to breed experimental crosses, to study multiple genetically identical animals, to minimise environmental influences and to sample tissues as required are significant advantages over comparable studies in humans. We have generated extensive genetic, genomic and physiological resources in our studies of the SHR, including expression datasets [14], [15] and the SHR genome sequence [16], [17] that have led to identification of several genes underlying pathophysiological traits in the SHR strain [18]–[22].
Given these resources and the relevance of SHR cardiac phenotypes to related human disorders [18], [19], we sought to investigate cytosine methylation in the SHR heart. We applied whole-genome bisulfite sequencing (WGBS) to multiple isogenic animals from the SHR and BN strains, to test the hypothesis that inter-individual variation in CpG dinucleotide methylation is regulated by genomic DNA sequence and may be important in the development of other genetically determined SHR phenotypes.
We generated WGBS datasets from the left ventricles of four BN and four SHR rats (European Nucleotide Archive accession number ERP002215), producing a total of 1.3 and 1.5 billion mapped reads in each strain. This was equivalent to 40–50× mean strand-specific read coverage per strain, and >28× mean coverage after quality filtering and removal of clonal and non-uniquely mapped reads (Table S1). Of the ∼40 million CpG dinucleotides covered by mapped reads on either strand in either strain, we focus here on the 10,614,445 CpG dinucleotides that were sequenced at a coverage depth of at least 5× strand-specific reads across a minimum of three animals per strain (Figure S1).
Hierarchical clustering of the methylation profiles showed that CpG dinucleotide methylation was more variable between strains than within strains. The average distance between methylation profiles of animals within each strain (average distance BN = 5,255, average distance SHR = 5,430) was ∼5 times lower than the average distance between animals across strains (average distance BN-SHR = 24,281) (Figure 1A). To quantify the contribution of strain specific differences to the variation in methylation measurements we carried out a principal component analysis which showed that the first principal component which separates the methylation profiles by strain explains 28.13% of the total variance (Figure 1B). No differences in global methylation levels between the strains were detected by pyrosequencing (Figure S2). These data suggested that differences in CpG dinucleotide methylation may be at least in part dependent on genomic sequence variation. To test this hypothesis we investigated the relationship between genome-wide inter-strain methylation differences and inter-strain differences in genomic DNA sequence.
First, we defined a set of CpG dinucleotides that were differentially methylated between animals from the SHR and BN strains. Of the 10,614,445 analysed CpG dinucleotides, 77,088 (0.7%) were significantly differentially methylated between the two strains (false discovery rate (FDR) <5%). Of the 77,088 differentially methylated CpGs, 47,775 cluster into 12,128 differentially methylated regions containing between 2 and 213 CpG dinucleotides, while 29,313 are single CpG dinucleotides that are at least 500 bp away from the next differentially methylated CpG dinucleotide (Figure S3). In more than 96% of differentially methylated regions, methylation differences between strains at individual CpGs were in the same direction. In line with previous observations, the large majority of analysed CpG dinucleotides were highly methylated (≥80% methylation) (Figure S4) but a significant fraction of CpG dinucleotides (2.9%) had very low methylation (≤10% methylation). Within CpG islands (CGIs), differentially methylated CpGs were found equally across different genomic features whilst outside CGIs, the proportion of differentially methylated CpGs was increased around the transcription start site (TSS) (Figure S5).
To determine whether differential methylation could be associated with discrete molecular or cellular functions important to the biological differences between SHR and BN rats, we carried out a Gene Ontology analysis of genes overlapping with or in close proximity (within 5000 bp) of differentially methylated regions containing five or more differentially methylated CpGs. Based on these criteria 2,525 differentially methylated regions were associated with 1,283 genes. This set of genes showed strong enrichment for gene products localised to the plasma membrane (p = 1.2×10−14) or involved in neuron differentiation (p = 2.9×10−4) or cell communication (p = 2.9×10−4) (Table S2).
To validate WGBS methylation calls and to define a set of differentially methylated CpG dinucleotides for genetic analysis, we carried out Sanger sequencing and multiplexed Illumina sequencing of PCR products generated from bisulfite converted genomic DNA. We selected a set of CpG dinucleotides (Table S3) in 40 regions (out of the 12,128 DMRs detected in the parental strains) that showed between 1% and 93% differential methylation between BN and SHR, were located at varying distances from TSSs (the majority were within 3 kb of the TSS) and were within a variety of genomic elements. By Sanger sequencing in BN and SHR, there were 81 scoreable CpG dinucleotides for which 70 were covered in the WGBS data set. By multiplexed Fluidigm amplification and Illumina sequencing, there were 131 scoreable CpG dinucleotides, all of which were covered in the WGBS dataset (Table S3). CpG methylation measured from bisulfite-converted PCR-amplified products was highly concordant with WGBS methylation data, both for the Sanger and for the Fluidigm/Illumina-sequenced PCR products (rSanger = 0.86, p<10−4, Figure S6A; rFluidigm = 0.93, p<10−4, Figure S6B).
We next sought to map the genetic determinants of CpG dinucleotide methylation of the 212 CpG dinucleotides residing within the 40 amplicons in the BN- and SHR-derived BXH/HXB panel of recombinant inbred (RI) strains [23], [24], using methylation percentage as a quantitative trait. Mean cytosine methylation for 32 of the 40 amplicons mapped in cis (LOD scores 4.8–23.3) and two in trans (LOD 4.9 and 6.1) (Table 1, Figure 2, Figure S7). The remaining six amplicons showed no linkage for mean CpG methylation although individual CpG dinucleotides in three of these amplicons (Epha2 [ENSRNOG00000009222], Ppp1r13b [ENSRNOG00000012653], Tbc1d30 [ENSRNOG00000023951]) showed strong cis linkage (LOD 5.1–9.9; Fig. 2C, Figure S7, Table S3). Individual CpG dinucleotides within all three of these amplicons showed opposing allelic effects (Table S3), which was validated by amplicon sequencing in the parentals (Figure S8), explaining the lack of linkage to mean CpG methylation across these amplicons. Individual CpGs within other amplicons generally showed similar quantitative differences and directional change (Table S3). The strength of linkage to methylation at individual CpG cytosines was highly dependent on the difference in methylation between the parental strains (Figure 3). We assessed the heritability of cytosine methylation by segregation in the RI strains as previously described [15] for the 32 amplicons for which cytosine methylation was regulated in cis and found that the heritability of mean amplicon methylation ranged from 12% to 99%, with mean heritability being 62% (SD 22%). We also assessed heritability by calculating the proportion of the linkage signals at these 32 loci that was explained by cis linkage and found that, on average 58% (SD 23%) of the total phenotypic variability in mean methylation in the RI strains at these loci was explained by cis linkage. Taken together, these data imply that methylation percentage at most differentially methylated loci are regulated in cis. In addition, the observed large sizes of the cis-QTL effects suggest that these methylation phenotypes are essentially under the control of a single locus, defined here as the extent of the linkage region, and are strongly regulated in cis.
The differentially methylated CpG cluster at Odfp2 [ENSRNOG00000014584] on chromosome 3 was regulated in trans by a locus on chromosome 2. Four of the seven protein-coding genes in the 2-LOD linkage confidence interval – or trans-regulated methylation quantitative trait locus (meth-QTL) - contain non-synonymous coding sequence variants between BN and SHR and two of the six non-protein coding genes contain sequence variants (Table S4). In an RNA-seq data set that we generated from SHR and BN left ventricle (Table S5, ArrayExpress accession number E-MTAB-1702), we detected expression for 8 of the 13 genes in the 2-LOD confidence interval, of which three showed evidence of differential expression or alternative splicing between BN and SHR (Table S4). None of the 13 genes in the 2-LOD confidence interval nor their mouse or human orthologues had Gene Ontology annotations related to DNA methylation, chromatin status or DNA binding, suggesting a potentially novel mechanism for trans-regulated methylation caused by sequence variation or altered expression of one of the protein-coding or non-protein-coding genes in this region. The trans-regulated CpG cluster at Asap2 [ENSRNOG00000006056] on chromosome 6 was regulated by a locus on chromosome 1. The region of linkage contained more than 180 genes, many of which showed either differential expression or coding sequence variation between SHR and BN.
Because our studies of differential methylation were carried out on intact heart tissue, it is possible that inter-strain differences in methylation were due to differences in cellular composition of the tissue samples rather than inter-strain differences in methylation for a particular cell type. We therefore studied CpG methylation at differentially methylated sites in cardiomyocytes and non-cardiomyocytes isolated from intact SHR and BN heart. Whilst a small proportion of differentially methylated loci showed methylation differences that were in part dependent on cell type, in 69% of cases methylation differences in cardiomyocytes and non-cardiomyocytes were at least in part dependent on strain, and for 77% of CpGs methylation differences were independent of cell type (Figure S9, Table S6). A principal component analysis (PCA) of methylation measurements (Figure 4) confirmed that the variation in methylation is in large part explained by the strain background. The first principal component (PC) which retains 43% of the variance separates samples into BN and SHR while the second PC which explains 20% of the variance distinguishes samples by tissue (Figure 4A). Separate PCA of methylation measurements done in each cell population, showed similar distribution of loadings for the 1st PC (strain difference) in both cell populations (Figure 4B), suggesting similar patterns of strain dependant methylation in the two populations.
We carried out a quantitative trait methylation (QTM) analysis to test for associations between variation in methylation and phenotypic variation. We searched for correlation between average methylation levels at the 40 loci previously tested for genetic linkage (Table S3) and 241 physiological quantitative traits measured across the RI strain panel. We found significant negative correlation (Pearson's r = −0.75) of CpG methylation on chr5:143,573,578–143,573,701 and serum chromogranin B (CgB) levels (P = 1.27×10−2, FDR = 5%). Subsequent QTL mapping for the serum chromogranin B trait located the peak of linkage to chr5:143,426,944 (LOD score = 3.92, P = 0.007) which was identical to the peak of linkage of the methylation QTL that correlated with the phenotype.
To identify genomic variants that may underlie cis-regulatory control of CpG dinucleotide methylation, we analysed the distance to the nearest SNP for all differentially methylated CpG dinucleotides. We found an increased frequency (5-fold enrichment) of CpG dinucleotides with closely adjacent SNPs in the set of differentially methylated CpGs compared to the analysed set of CpG dinucleotides and to all CpG dinucleotides in the genome (Figure 5A). SNP enrichment in DNA adjacent to differentially methylated CpG dinucleotides was not due to SNPs that disrupt adjacent CpG dinucleotides because there was no enrichment of CpG-disrupting SNPs at these sites (Figure 5A). There was an increase in SNP frequency starting at around 250 bp from differentially methylated CpGs with a further sharp increase occurring within 5 bp of the differentially methylated CpG dinucleotide (Figure 5B): 58.8% of differentially methylated CpGs were no further then 250 bp away from a SNP, 5.3% had a SNP within 5 bp. SNP proximity was, however, not associated with the extent of differential methylation at the affected CpG dinucleotide (Figure S10).
To investigate further the relationship between local SNP density and differential methylation, we sought to determine whether there was an association between nucleotide preference within 5 bp of differentially methylated CpG dinucleotides and either increased or decreased methylation. We observed preferential nucleotide usage for increased and decreased CpG dinucleotide methylation within this 12 bp window with particularly strong association in the 1–3 bp immediately adjacent to differentially methylated CpGs (Figure 5C). The nucleotide signatures that we identified were very similar to those recently found to be associated with increased and decreased methylation in mouse brain [12]. While the consensus sequence of the nucleotide usage at hypermethylated CpG dinucleotides is identical the consensus sequence at hypomethylated CpGs differs only at a single position (Table S7). To quantify the observed similarity we compared our position frequency matrices with those obtained by Xie et al (personal communication) using the TOMTOM comparison tool [25]. Both the nucleotide usage patterns around hyper- as well as hypomethylated CpGs are highly similar between rat heart and mouse brain (Phypo = 2.76×10−10, Phyper = 1.06×10−10, Table S7).
After categorising SNP differences between BN and SHR 1 bp up- and 1 bp downstream of the differentially methylated CpG we found that SNPs which changed adenines and thymines into guanines and cytosines both upstream and downstream of the differentially methylated CpG more often resulted in increased methylation in SHR than in BN (Figure S11). The nucleotide bias associated in this study with increased and decreased CpG methylation showed remarkable similarity to that found previously in mouse brain [12] and in human blood [26].
To study further the relationship between allelic sequence variation and CpG dinucleotide methylation and to search for possible parent-of-origin (PO) effects, we carried out WGBS of left ventricle DNA from F1 animals from reciprocal crosses between SHR and BN parents (Table S8). After filtering, we retained and analysed approximately 25 million CpG dinucleotides (Figure S12, Figure S13). This number was greater than in the parental strains because of increased coverage in the F1 crosses compared to the parental data sets. The absence of cross specific clusters in the hierarchical clustering analysis (HCA) (Figure 6A) and the lack of a clear separation of profiles by the 1st PC (Figure 6B), which captured strain specific variation in the parentals, showed that the variation in methylation profiles across animals within and between the two crosses was small. Consistent with the small variation in methylation we detected only 2,627 differentially methylated CpG dinucleotides between the reciprocal crosses (<0.01% of all analysed CpGs). This was less than one tenth of the variability seen between the two parental strains, despite the higher coverage in the F1 animals, reflecting the genetic differences between the parental strains, but the identical genomes (apart from the sex chromosomes) of all the F1 animals.
In order to detect allele-specific methylation (ASM) in the F1 animals, SNP-containing reads were phased by parental genotype. Methylation profiles derived from the phased read data did cluster by parental genotype (Figure 6C) and were separated by genotype when projected on the 1st PC (Figure 6D). The absence of cross specific sub-clusters is most likely due to the subtlety of methylation differences between the cross which is smaller than the technical variation of the method. Of the 2,170,074 CpG dinucleotides testable for ASM differences after phasing, 38,152 showed significant ASM differences (FDR<5%). Because cis-regulated differences in cytosine methylation should be detectable as both differential methylation in the parental strains and as ASM in the F1s, we looked for the intersection between the differentially methylated CpG dinucleotides in the parental and the F1 datasets. Of the 1,705,718 CpG dinucleotides analysed in both the parental and phased F1 data (Figure 7A, intersection), there were 8,340 autosomal CpG dinucleotides differentially methylated in an allele-specific manner in the F1s that were also differentially methylated between BN and SHR parental strains (Figure 7B, intersection). This number was 15 fold greater than expected by chance (X2 = 137040.2, p<4.9×10−324). We observed the highest enrichment for proximal sequence variation amongst the subset of differentially methylated CpGs in the F1 dataset that were also found to be differentially methylated in the parental strains (Figure 7C). The increased frequency of local SNPs that we observed in this set is further evidence of enrichment for putative cis-regulatory sequence variants adjacent to differentially methylated CpG dinucleotides. The striking concordance of differential methylation in the parental animals and allelically-determined methylation in the F1 animals strongly suggests that the methylation level at a large number (many thousands at least) of CpG cytosines is regulated in cis, partly by local sequence variation. The enrichment for SNPs in the 250 bp surrounding differentially methylated CpGs, and particularly in the 10 bp window around these CpGs (Figure 7C) indicates that sequence variation in the immediate vicinity of differentially methylated CpGs underlies at least part of these differences in cytosine CpG methylation.
In addition to the finding of ASM differences in F1 animals, we also identified 723 CpG dinucleotides within 145 regions that were significantly differentially methylated (FDR<5%) in a PO-specific manner (Table S9). In 35 regions we detected a mean methylation difference between the paternal and maternal allele of more than 50%. Twenty-five of these regions have been reported previously as imprinted loci in human or mouse while 10 regions were previously undescribed (Table 2). To validate the parent-of-origin specific methylation differences we measured methylation by Fluidigm amplification and Illumina sequencing at 15 loci (four known imprinted and eleven novel) at which we detected PO-specific methylation by WGBS. The methylation differences detected by WGBS at the selected loci ranged from 20–97%. Within these 15 loci, we were able to confirm parent-of-origin dependent methylation differences for the four known imprinted loci and five of the novel loci (F1LUE4_RAT-Sgk1, 5SrRNA-Pmfbp1, Col9a2, Mthfd2l, Region ID #97; Table S10). Validation for one of the loci was only achieved on the BN parental chromosome, most likely due to insertion deletion polymorphisms on the SHR chromosome which led to the absence of sequencing reads from the SHR chromosome at this locus.
Table 3 and Figure 8 summarise methylation differences we detected in this study by source of variation. Table 3 shows the proportion of CpGs that were differentially methylated in the analysis of i) inter-strain differences in methylation, ii) differential methylation between reciprocal crosses, iii) allele specific methylation and iv) parent-of-origin dependent methylation. Figure 8 shows the distribution of methylation differences in the respective analyses. CpGs methylated in a parent-of-origin dependent manner exhibited the highest median difference in methylation followed by CpGs that showed allele-specific methylation. The small number of differentially methylated CpGs detected between the F1 reciprocal crosses had a median methylation difference similar to that of CpGs differentially methylated between the parental strains albeit with a broader range of differences.
Cytosine methylation at CpG dinucleotides is amongst the most studied epigenetic marks, yet at present there is little information about the extent to which inter-individual differences in CpG methylation in adult tissues are genetically determined by germline DNA sequence variation and how inter-individual variation in methylation relate to whole-body phenotypes. Previous studies in humans and experimental animals have shown evidence of individual or allele-specific differences in methylation at CpG cytosines [4]–[12]. However, these have either assayed methylation at a small fraction of genomic loci, used low resolution techniques, not taken account of SNPs that disrupt CpG sites, studied single or small numbers of individuals, or only indirectly inferred the difference between cis and trans regulation of CpG methylation.
We applied WGBS to assay CpG methylation at single-nucleotide resolution across the genome in multiple isogenic animals, and carried out segregation studies in F1s and linkage studies in RI lines to define the extent and mode of inheritance of inter-individual variation in CpG cytosines in the adult rat heart. While methylation profiling has been carried out previously in the rat [27] this is the first genome wide study at single nucleotide resolution methylation to look at the association of genotype, epigenotype and phenotype in an animal model of human disease. Methylation studies in human heart tissue have used Illumina BeadChip arrays [28] and MeDIP-seq [29] to examine cytosine methylation in cardiomyopathy (CM). However, these studies did not account for inter-individual germline differences in the genome sequence which as we show here is an important determinant of inter-individual variation in DNA methylation and may in part explain the methylation differences observed in CM patients.
We measured cytosine methylation at over 10 million CpG sites across the genome, showed that inter-strain variability in cytosine methylation is markedly greater than intra-strain variability and detected differential methylation between the strains at over 75,000 CpG cytosines, amounting to 0.7% of all analysed CpG sites. Our studies in animals from parental strains, F1s and RI lines show that inter-strain differences in CpG methylation are substantially cis-regulated, essentially under control of sequence variation at a single locus in cis defined as the extent of the linkage region and the resolution of the genetic map which has an average distance between genetic markers of ∼2.7 MB. Furthermore, we were able to demonstrate that the detected methylation differences are largely independent of differences in tissue composition between the two strains.
Our Gene Ontology analysis showed that genes that were closest to differentially methylated regions, defined as a minimum of five adjacent differentially methylated CpGs, were strongly enriched for gene products localising to the plasma membrane, or involved with neuron differentiation or cell communication. Interestingly, a study of genetic variation in the SHR rat found that genes showing major variation in their coding region between SHR and BN also showed enrichment for genes encoding plasma membrane components and neurological processes [16]. These data suggest that, in addition to direct cis effects on methylation, differential methylation near genes associated with the enriched GO categories could be the result of genetic variation in genes with similar biological function that act indirectly through feedback mechanisms at the molecular or organismal level. Although potentially of interest, such observations would require more detailed investigation to confirm and understand these associations. Furthermore, we find significant correlation between CpG methylation and serum CgB levels. Serum CgB levels are a correlate of sympathetic nervous system overactivity and have been proposed as a biomarker for heart failure [30]. These findings suggest a link between genetic, epigenetic and disease-associated phenotypes the causal relationship of which remains to be determined.
Analysis of genomic sequence showed that the 250 bp surrounding differentially methylated CpGs are enriched for sequence variants between the SHR and BN strains. We found biases in nucleotide usage in the 5 bp immediately adjacent to differentially methylated CpGs that are associated with increased or decreased methylation. Remarkably, the nucleotide signatures that we identified are almost identical to those recently identified as being associated with increased and decreased methylation in mouse brain [12] indicating near complete conservation across tissues and species. Nearly identical signatures were also associated with increased and decreased methylation in a separate analysis of CpG methylation associated with ageing in human blood [26]. This extent of conservation across tissues and species suggests a common mechanism contributing to regulation of CpG methylation that merits investigation in future studies.
Further to our findings of differential methylation in the parental strains, we identified over 38,000 CpG dinucleotides that showed allele-specific methylation in F1 animals. This finding of allele-specific methylation in F1 animals is most likely due to cis-regulatory mechanisms and is consistent with previous studies in humans and mice reporting cis-regulation of CpG methylation [4]–[10], [12], [31]. Our data define a genome-wide set of CpGs that are susceptible to allele-specific regulation of CpG methylation in the mammalian heart. Since in our study allele-specific methylation could only be detected in F1 animals when sequence reads could be phased by parental genotype, we could only seek the presence of allele-specific methylation in 2.1 million of the total of more than 25 million CpG dinucleotides in the rat genome. Accordingly it is likely that the number of CpG dinucleotides influenced by cis-regulated methylation is up to 10-fold greater than the 38,000 detected in this study in the genomes of these two strains.
Our studies in F1 animals also detected parent-of-origin (PO) effects on DNA methylation with 723 CpG dinucleotides clustered in 145 regions showing evidence of PO-type regulation of methylation. To our knowledge, this is the first empirically-based prediction of imprinted genes in the rat. Thirty-five of these regions showed CpG methylation differences of more than 50% between paternally and maternally-inherited alleles. Of these, 25 have been previously reported as imprinted loci in human and mouse, including 19 identified recently in brain [12] while 10 regions were previously undescribed.
The detected methylation differences ranged from a ∼20% median difference in the parental strains to ∼50% median difference for parent-of-origin specific methylation. The continuous distribution of methylation differences and the practical absence of mono-allelic methylation (i.e. 100% methylation on one allele 0% methylation on the other) is probably explained by i) the methylation measurements being composite signals coming from a mix of different cell types present in the left ventricle and ii) the stochastic nature of methylation reaction.
Our mapping studies in the BXH/HXB RI strains showed that, treated as a quantitative trait, methylation at 151 of 212 CpGs, residing within 36 amplicons, showed cis linkage with LOD scores of 3.9–32.0. Most of these showed complete segregation of CpG methylation with genotype indicating essentially single-locus control in cis. Methylation of six CpGs in the Odfp2 amplicon mapped in trans (LOD scores 3.6–7.6), with the trans-regulating meth-QTL locus residing on a different chromosome to the regulated CpGs. The fact that over 95% of the meth-QTLs for individual CpGs that showed linkage in the RI strains were cis-regulated is in keeping with our nucleotide preference data in the parental strains, our cis/trans tests in the F1s, and with previous data [11], [12] that correlate local sequence variation with CpG methylation status. The explanation for the two loci that showed trans-regulated methylation is less clear. For Odfp2, there were no candidate genes in the 2-LOD confidence interval of the trans meth-QTL that had GO annotations related to DNA methylation, chromatin status or DNA binding, suggesting either a novel mechanism for regulation of DNA methylation, or possibly a role for one of the several non-coding RNAs in this interval. For Asap2, the large number of genes in the QTL interval precludes assessment of candidates without further investigation. The finding that methylation is predominantly regulated in cis complements the analysis of SNP frequency adjacent to differentially methylated CpGs. Both results point to the importance of cis regulation although the underlying mechanisms and relative contribution of proximal and distal sequence effects require further investigation. Whilst a previous analysis has suggested that most cis-acting allelic influences on CpG methylation may be up to 149 kb [9], our study and that of Gibbs et al [6] detect allelic associations over much shorter distances of 1–45 bp from differentially methylated CpGs. The observed co-segregation of genotype and differential methylation is consistent with the results of a recent study of inbred lines in plants [13], suggesting that cis-regulated control of inter-individual variation in cytosine methylation is strongly conserved across species and between animals and plants.
Many previous studies of CpG methylation have been carried out in cell lines or homogeneous cell types isolated in primary cultures, while we used whole cardiac tissue for these studies. The spontaneously hypertensive rat is a model of cardiovascular disease that has been extensively characterised genetically and phenotypically for hypertension, cardiac hypertrophy and failure and insulin resistance [18]–[20], [22], [32]–[34] but the role of the cardiac methylome in these phenotypes remains unexplored. It might be expected that the study of whole tissue harvested ex vivo might reduce our ability to detect genetic effects because of cell type heterogeneity, possible paracrine effects or secondary effects of whole-body phenotypes such as hypertension. Despite this, we find consistent genetic effects on CpG methylation in cardiac tissue from parental strains, F1 animals and RI lines derived from the two parental strains. In particular, the finding that the CpG sites showing allele-specific methylation in F1 animals were enriched 15-fold (p<10−324) for CpG sites also showing differences in CpG methylation in the parental strains is evidence of the stability and robustness of these genetic data.
Previous studies of gene expression showed that the heritability of gene expression in heterogeneous compared to homogeneous cell types is dominated by cis-regulated gene expression [35], and also that expression differences that are found in single cell populations may not be found in heterogeneous tissues [36]. By analogy, it is possible that genetic analysis of DNA methylation in more homogeneous cell populations than were studied here would detect stronger cis effects and more extensive trans effects on CpG methylation than were detected in our studies of intact heart tissue. Notwithstanding, our study was able to detect strong allelic effects on CpG cytosine methylation and, in addition, our analysis of methylation differences in isolated cardiomyocytes and non-cardiomyocytes showed that for the large majority (>75%) of CpGs analysed in this study, differential CpG methylation in the heart tissue was independent of cell type. Taken together, these studies define a minimum extent of genetic regulation of CpG methylation across these two rat strains in the rat heart.
Rats were housed in an air-conditioned animal facility and allowed free access to standard laboratory chow and water. All experiments were done in agreement with the Animal Protection Law of the Czech Republic and were approved by the Ethics Committee of the Institute of Physiology, Czech Academy of Sciences, Prague. We collected left ventricular tissue from unfasted males from six week old male BN-Lx/Cub (referred to in this study as BN), SHR/Olalpcv (referred to as SHR) (n = 4 per strain); and from (BNxSHR)F1 and (SHRxBN)F1 (n = 4 per cross); and from male rats from 29 BXH/HXB recombinant inbred strains (n = 2 per strain) [23] between 9:00am and 10:00am. Left ventricles were snap frozen in liquid nitrogen and stored at −80°C. The BN-Lx/Cub and SHR/Olalpcv lines have been inbred over more than 80 generations [17]. For tissue heterogeneity studies, cardiomyocytes and non-cardiomyocytes were isolated as previously described [37] from SHR/Ncrl and BN.NCrl rats (four biological replicates each) and studied alongside whole heart tissue from the same strains, as well as left ventricle tissue from SHR/Olalpcv and BN-Lx/Cub rats (three biological replicates each). No differences were found between the SHR/Olalpcv and BN-Lx/Cub and the respective NCrl strains which are also referred to as SHR or BN in the analysis of tissue heterogeneity.
Frozen tissues from BN rats, SHR rats and the reciprocal F1 animals were processed without pooling to generate methylation profiles on the HiSeq 2000 platform (Illumina) as described in [38].
A luminometric-based assay for global DNA methylation was performed according to the protocol described in [39]. Briefly, 250 ng of genomic DNA extracted from liver (n = 3), kidney (n = 3) and left ventricle (n = 4) of BN and SHR animals was subjected to double-digestion with EcoRI and methylation-sensitive HpaII and methylation insensitive MspI restriction endonucleases and global methylation between strains quantified by pyrosequencing.
RNA was extracted from 25 mg of crushed left ventricular tissue from four BN and four SHR animals without pooling, using Trizol (Invitrogen) according to manufacturer's instructions. 4 µg of total RNA was used to generate RNA-seq libraries using TruSeq RNA kit (Illumina) according to manufacturer's instructions. Libraries were multiplexed in pairs and run on a single lane of the HiSeq 2000 platform (Illumina) to generate 100 bp paired-end reads.
500 ng of genomic DNA was bisulfite converted using the MethylCode Bisulfite Conversion Kit (Invitrogen) according to manufacturer's instructions. Diluted by 1∶8, bisulfite-converted DNA was amplified using PFU Turbo Cx polymerase with primers designed with the Meth primer program (http://www.urogene.org/methprimer/index1.html). PCR products were purified using the MultiScreen PCRμ96 plate (Millipore). Sanger-sequenced CpG dinucleotides were scored in the Sequencher 5.0 analysis software (Gene Codes Corporation). A secondary peak was called if it was ≥5% of the primary peak height. The peak height ratio was used to calculate percentage methylation at CpG sites. Primer sequences are available on request. To ensure that we mapped CpG dinucleotide methylation rather than methylation of CpG disrupting SNPs (e.g. CG>TG), we carried out genomic DNA sequencing in BN and SHR for all 14 regions containing the differentially methylated CpG dinucleotides and only mapped CpG dinucleotide methylation where the CpG dinucleotide was present in both strains.
500 ng of genomic DNA was bisulfite converted using the MethylCode Bisulfite Conversion Kit (Invitrogen) according to manufacturer's instructions. 50 ng of bisulfite-converted DNA and target regions were PCR amplified utilising the Fluidigm 48.48 Access Array. Primers were designed with the Sequenom primer program (http://www.epidesigner.com) and FastStart High Fidelity PCR system (Roche) was used for amplification. PCR amplification was performed for 40 cycles with an annealing temperature of 57°C. Primer sequences are available on request. PCR products from each sample were indexed, pooled and purified using the Agencourt AMPure XP system (Beckman Coulter). Sequencing libraries were prepared from the pooled samples and sequenced on the Illumina MiSeq platform. Demultiplexed sequencing reads were mapped and methylation profiles generated as described in [38]. Differentially methylated CpG dinucleotides were analysed for the presence of a disrupting SNP or INDEL and only CpG dinucleotide methylation where the CpG dinucleotide was present in both strains was mapped.
The rat genome assembly RGSC3.4 (rn4) was used as the reference sequence. Reference sequence files in FASTA format were obtained from the UCSC genome browser website (ftp://hgdownload.cse.ucsc.edu/goldenPath/rn4/chromosomes).
Transcription start site, exon and intron annotations for the rat genome assembly RGSC3.4 were obtained from the Ensembl database (http://www.ensembl.org) version 62. CpG island coordinates (cpgIslandGgfAndyMasked) were obtained from the UCSC genome browser (http://genome.ucsc.edu).
We used the BN-Lx/Cub and SHR/OlaIpcv genome sequence variation data previously reported [17], using the SNV and indel caller described in [17].
WGBS read mapping and processing was carried out as described in detail in [38]. In brief, WGBS reads were aligned against the rat genome reference sequence with BWA [40] version 0.5.8a. Prior to alignment, reads were pre-processed as follows: i) the conversion state of read pairs was masked by converting all C base calls of read1 to Ts and all G base calls of read2 to As ii) the first read base was clipped to avoid false negative methylation calls from unmethylated cytosines introduced at the fragment 5′ end during the end-repair step of library preparation iii) reads were quality trimmed at the 3′ end with the -q option of BWA using a Q score cutoff of 20.
To map reads originating from the bisulfite converted forward strand of the genomic DNA, reads were aligned to an in silico bisulfite-converted version of the reference sequence with all Cs converted to Ts; to map reads originating from the bisulfite converted reverse strand reads were aligned to the G-to-A converted reference sequence. To reduce allele-specific mapping bias, reads generated in SHR samples were mapped to a reference sequence that was converted to the SHR allele at BN-SHR SNP positions prior to C-to-T/G-to-A conversion. Reads generated in F1 animals and thus of unknown haplotype at BN-SHR SNP positions were mapped against a reference sequence with all SNP positions masked by replacing them with N. Spiked-in unmethylated lambda phage control DNA was mapped to the lambda reference sequence (Genbank accession NC_001416.1) which was in silico bisulfite converted in the same way as the rat reference genome sequence.
After alignment, the 3′ ends of overlapping read pairs were clipped retaining the higher quality end in order to avoid duplicate methylation calls in the overlap region. Subsequently read mappings were filtered removing i) clonal reads ii) reads with a mapping quality <20 iii) read pairs mapping to both the in silico converted forward and the in silico converted reverse strand iv) reads with invalid mapping orientation.
Depth of coverage statistics for the filtered WGBS mappings were calculated with the DepthOfCoverage command of the Genome Analysis Toolkit [41] version 1.1.23.
Mapped, processed and filtered reads were piled up with the samtools [42] version 0.1.16 mpileup command and the number of cytosine and thymine base calls counted at each cytosine position. The relative methylation level at each cytosine position was calculated as the percentage of cytosine base calls of the total number of cytosine and thymine base calls. Methylation calls were corrected for incomplete bisulfite conversion by subtracting the average number of expected unconverted cytosines at the depth of coverage at the respective position given the bisulfite conversion rate for the respective sample prior to calculating relative methylation levels. For WGBS libraries generated from F1 samples the bisulfite conversion rate was calculated from the frequency of unconverted cytosines in the unmethylated lambda control spike-ins. For WGBS libraries generated from BN and SHR samples, the bisulfite conversion rate was estimated by regression analysis of non-CpG conversion rate and lambda conversion rate in the F1 samples. Bisulfite conversion rates for BN and SHR samples were then predicted from the non-CpG conversion rate in these samples. Bisulfite conversion rates were >97% for all samples.
Differential methylation at cytosines was tested for by Fisher's exact test on a 2×2 contingency table testing for independence of strain/cross/allele and the frequency of unconverted and converted cytosines across all replicates. P-values were adjusted for multiple testing using the false discovery rate (FDR) method by Benjamini and Hochberg [43].
Methylation profiles were filtered prior to multiple testing correction i) Retaining only cytosines with a minimum combined read coverage of 5× across a minimum of three replicates. ii) Removing CpN dinucleotides affected by BN-SHR sequence variation (SNP and indels) to exclude methylation differences resulting from the disruption or deletion of methylation sites.
Methylation profiles of CpG dinucleotides with at least 5× coverage in each of the replicates/phased data sets were clustered based on the pairwise euclidean distance between the vectors of methylation levels for each animal scaled down to 5× coverage. CpG dinucleotides affected by BN-SHR sequence variation (SNPs and indels) were removed from the data set prior to the analysis. Profiles were subsequently clustered using Ward's minimum variance method [44]. Distance calculations and clustering were carried out with the statistical software package R [45] using the dist and hclust functions, respectively. Principal component analysis was carried out with the pca function of the pcaMethods R package.
From Sanger and Illumina sequencing of PCR products, we derived mean methylation values from the two biological replicates for each RI strain. We carried out genome-wide linkage analysis in the 29 BXH/HXB RI strains for all 212 CpG dinucleotides and the average methylation percentage of all CpG within each of the 40 PCR amplicons. Linkage analysis was carried out as previously described [14] using QTL Reaper (K.F. Manly; University of Tennessee Health Science Center, Memphis, Tennessee) except that a denser SNP map of 1384 markers was used [46]. To account for non-normally distributed data, the empirical significance of the meth-QTLs was assessed by 1 million permutations [47]. The same genome wide correction was applied for testing both cis and trans linkage. We defined regions as regulated in cis or in trans by defining cis when the peak of linkage marker lay within 5 Mbp of the genomic location of the amplicon, with other linkages being defined as trans. Candidate genes and Gene Ontologies for the Odfp2 and Asap2 meth-QTL interval were taken from Ensembl release 66.
Principal component analysis of methylation measurements in cardiomyocytes and non-cardiomyocyte cell populations isolated from BN and SHR cardiac tissue was carried out using the pca function of the pcaMethods R package with unit variance scaling.
Measurements of 241 phenotypic traits measured across the RI strain panel (Table S11) were provided by Michal Pravenec. Outliers were removed from the raw measurements within each strain using boxplot analysis, grubb test and Nalimov test. After outlier removal mean and standard error where calculated for each trait and strain. Association of average locus methylation in the RI strains and physiological phenotypes was then assessed by Pearson's correlation test. P-values were adjusted for multiple testing using the false discovery rate (FDR) method by Benjamini and Hochberg [43].
RNA-seq reads were aligned to the rat genome reference sequence assembly RGSC3.4 (rn4) with TopHat [40] version 1.3.0. To reduce allele-specific mapping bias, reads generated in SHR samples were mapped to a reference sequence that was converted to the SHR allele at BN-SHR SNP positions.
Exon expression was quantified by counting read-pairs mapped to exon locations annotated in Ensembl version 62 using the NxtGenUtils (http://code.google.com/p/nxtgen-utils) version 0.12 CountReads command. Differential expression analysis was carried out with DESeq [48] version 1.6.1 on the raw exon fragment counts. The analysis was run with the default parameters except for the estimateDispersions function which was run with the ‘pooled’ option set to true. An FDR-adjusted P-value<0.05 was chosen as the significance threshold for differential expression.
SNP allele bias for alleles at differentially methylated CpGs was examined by calculating the information content [49] for the observed SNP allele frequencies at the five base pairs up- and downstream of CpGs showing increased or decreased methylation in the SHR and BN strains. Calculations and visualisation were carried out with the Bioconductor seqLogo package implemented in R. The seqLogo code was modified to take into account nucleotide usage in the rat genome (29% A, 29% T, 21% C, 21% G). To compare nucleotide usage around hyper- and hypomethylated CpGs in rat and mouse nucleotide usage information was obtained as personal communication from the authors of [Xie et al., 2012]. Position frequency matrices derived from rat and mouse nucleotide usage data were compared with the TOMTOM tool of the MEME suite version 4.9 [25].
WGBS read pairs generated in the reciprocal F1 crosses that mapped to BN-SHR SNP positions were phased by parental genotype by determining the SNP allele in the read sequences using the NxtGenUtils (http://code.google.com/p/nxtgen-utils) version 0.12 PhaseRead command. C/T SNP positions were not used for phasing because of the difficulty of distinguishing allelic variation from bisulfite conversion at these positions in WGBS reads. Read pairs with ambiguous allele patterns (equal frequency of BN and SHR SNP alleles) were discarded.
Allele and parent-of-origin specific differences in methylation were detected by carrying out differential methylation analysis on the phased F1 WGBS read data as described above for the parental and unphased F1 data.
Allele-specific methylation differences were determined separately for maternally and paternally derived chromosomes, i.e. methylation on the (maternally-derived) BN allele in the (BNxSHR)F1 cross was compared to methylation on the (maternally derived) SHR allele in the (SHRxBN)F1 cross. Methylation on the (paternally-derived) BN allele in the (SHRxBN)F1 cross was compared to methylation on the (paternally-derived) SHR allele in the (BNxSHR)F1 cross. A CpG was reported as allele-specifically methylated if it showed a statistically significant difference (as above) in methylation in either or both the comparisons.
By analogy, parent-of-origin specific methylation differences were determined by comparing methylation on the maternal and paternal allele separately for the BN and SHR derived chromosomes. Only differentially methylated regions that show a consistent direction of methylation differences in all comparisons were reported.
Differentially methylated cytosines were clustered into differentially methylated regions by grouping differentially methylated cytosines not further away than 500 bp from each other irrespective of the direction of the methylation difference using BEDTools [50].
Genes whose gene body (exons and introns) or the region 5000 bp upstream or downstream overlapped with differentially methylated regions containing five or more differentially methylated CpGs showing the same direction of methylation differences were tested for enrichment of Gene Ontology terms. As background the set of all genes whose gene body +/−5000 bp overlapped with CpGs tested for differential methylation was used. The Gene Ontology analysis was carried out with DAVID 6.7 [51], [52] using default settings. Enriched Gene Ontology terms were subsequently clustered by similarity with REVIGO [53] setting the output option for ‘list size’ to ‘small’. Otherwise default settings were used.
|
10.1371/journal.pcbi.1002513 | Measuring Granger Causality between Cortical Regions from Voxelwise fMRI BOLD Signals with LASSO | Functional brain network studies using the Blood Oxygen-Level Dependent (BOLD) signal from functional Magnetic Resonance Imaging (fMRI) are becoming increasingly prevalent in research on the neural basis of human cognition. An important problem in functional brain network analysis is to understand directed functional interactions between brain regions during cognitive performance. This problem has important implications for understanding top-down influences from frontal and parietal control regions to visual occipital cortex in visuospatial attention, the goal motivating the present study. A common approach to measuring directed functional interactions between two brain regions is to first create nodal signals by averaging the BOLD signals of all the voxels in each region, and to then measure directed functional interactions between the nodal signals. Another approach, that avoids averaging, is to measure directed functional interactions between all pairwise combinations of voxels in the two regions. Here we employ an alternative approach that avoids the drawbacks of both averaging and pairwise voxel measures. In this approach, we first use the Least Absolute Shrinkage Selection Operator (LASSO) to pre-select voxels for analysis, then compute a Multivariate Vector AutoRegressive (MVAR) model from the time series of the selected voxels, and finally compute summary Granger Causality (GC) statistics from the model to represent directed interregional interactions. We demonstrate the effectiveness of this approach on both simulated and empirical fMRI data. We also show that averaging regional BOLD activity to create a nodal signal may lead to biased GC estimation of directed interregional interactions. The approach presented here makes it feasible to compute GC between brain regions without the need for averaging. Our results suggest that in the analysis of functional brain networks, careful consideration must be given to the way that network nodes and edges are defined because those definitions may have important implications for the validity of the analysis.
| Modern cognitive neuroscience views cognition in terms of brain network function. A network is a physical system of nodes connected to each other by edges. From the network perspective, cognitive function depends on activity patterns involving the nodes and edges of functional brain networks. It is important then, to appropriately define the nodes and edges of functional brain networks in order to understand cognition. In this study we consider the nodes of functional brain networks to be brain regions, and demonstrate a method that effectively measures the edge pattern between regions with a technique called Granger Causality. Our method is made possible by the utilization of recent advances from the field of statistics. Our approach is generally applicable to functional brain network analysis and contributes to the understanding of network properties of the brain.
| The modern understanding of human cognition relies heavily on the concept of large-scale functional brain networks, and large-scale functional network analysis of Blood-Oxygenation-Level-Dependent (BOLD) signals from functional Magnetic Resonance Imaging (fMRI) is playing an increasingly important role in cognitive neuroscience [1]. From this perspective, knowledge of cognition may be obtained from BOLD signals by identification of the nodes and edges of large-scale functional brain networks. An important unresolved question remaining in the field, however, is how best to define the nodes and edges of large-scale functional brain networks.
A node is typically represented in brain network studies of fMRI BOLD activity as a lumped Region Of Interest (ROI), formed by averaging the BOLD signals of all the ROI's voxels [2]–[6]. This collapse of the ROI by averaging has the benefit of reducing the dimensionality of analysis, but rests on the twin assumptions: (1) that the BOLD activity of an ROI is homogeneous over all its voxels; and (2) that the functional interactions (connectivity) between the voxels of an ROI with those in other ROIs is also homogeneous. If these homogeneity assumptions are not true, edge measurements computed from ROI-averaged BOLD signals may be erroneous since averaging may distort the time series information.
Here we present a novel procedure for the analysis of directed interregional functional interactions that is based on the BOLD activity of the individual voxels of ROIs and the Granger Causality (GC) measure of directed interaction between voxels. GC tests whether the prediction of the present value of one time series by its own past values can be significantly improved by including past values of another time series in the prediction. If so, the second time series is said to Granger cause the first, and the degree of significance of the improvement may be taken as the strength of GC [7]. The GC measure is typically implemented by AutoRegressive (AR) modeling [8] and has been shown to be a powerful and flexible tool for measuring the predictability of one neural time series from another [9]–[14]. It has advantages as an edge measure over the typically utilized correlation: first, it provides the strength of functional interaction between time series in both directions, as opposed to a single non-directional strength; second, its grounding in prediction allows stronger statements to be made about functional interactions than does simple correlation. The use of GC to measure directed interactions in the brain from fMRI BOLD data has received intense scrutiny in recent years, with some arguing in its favor [15]–[18] and others opposed to it [19]–[23]. In the present work, we focus on improving the application of GC analysis to fMRI BOLD data in order to better understand the role of top-down influences in visuospatial attention.
Previous evidence from GC analysis of fMRI BOLD data argues against the assumption of homogeneous interregional functional interactions, and thus suggests that averaging BOLD signals prior to edge measurement may not be appropriate. Bressler et al. [24] found that GC between ROIs varies considerably across voxel pairs, with the distribution of GC values being highly skewed and only a small fraction of values in the tail of the distribution being significantly different from zero. These results indicate that GC is heterogeneous across voxel pairs, suggesting that the investigation of functional interactions between ROIs should take into account the interactions of all the voxels within the ROIs.
An approach to the problem of heterogeneous functional interaction between ROIs is to compute the distribution of GC values using a bivariate AR model for each pairwise combination of voxels in the ROIs. This pairwise-GC approach, followed by Bressler et al. [24], not only avoids the possible pitfalls of averaging, but also makes feasible the separate measurement of GC density and strength between ROIs, two factors that are conflated by averaging. Thus, deriving a summary GC statistic between ROIs from the distribution of GC values across all voxel-voxel pairs may be statistically more informative than simply setting it to the GC between across-voxel averages. There is a further problem, however, with the pairwise-GC measure: some GC values may be identified as being significant when actually they are not. This problem arises, for example, if one voxel (x) ‘drives’ a second voxel (y), while voxel y ‘drives’ a third voxel (z), without there being a ‘drive’ from voxel x to voxel z (Figure 1A). In this case, the GC from voxel x to voxel z may be spuriously identified as being significant. As another example, the problem also occurs if voxel x ‘drives’ both voxels y and z with different delays, without there being a ‘drive’ from y to z (Figure 1B). In this case, the GC from y to z may be spuriously identified as being significant. Since x, y, and z may be in the same or different ROIs, one point these examples make clear is that the GC within ROIs should be taken into account in order to reduce the possibility of spuriously identifying GCs as being significant. A second point is that improvement in the GC computation may be possible by using a method that can mitigate the problem of spurious GC significance.
Our approach to the problem of spurious GC significance rests on the concept of conditional GC [25]–[27]. Conditional GC analysis tests for a significant GC from one time series to a second with the effect of a third time series removed. By this procedure, it is possible to determine whether a significant GC measured between two time series is attributable to the third time series. In this paper, we utilize the conditional GC concept for ROI-level analysis in an approach that essentially measures the GC between any pair of voxels in two ROIs conditional on all the other voxels in the ROIs. This is accomplished by constructing a single Multivariate Vector AutoRegressive (MVAR) model from the voxel time series, as opposed to the pairwise-GC method, in which a separate bivariate AR model is constructed for each voxel pair. Use of the MVAR model offers the promise of reducing or eliminating the problem of spurious significant GC identification in the assessment of directed functional interactions from fMRI BOLD signals.
To make use of the MVAR model for ROI-level GC analysis necessitates overcoming one further problem that often occurs in model estimation: the number of available observations (data points) limits the number of parameters (model coefficients) that can accurately be estimated. This problem commonly arises in neurobehavioral studies where the number of data points that can realistically be acquired limits the size of the MVAR model that can be estimated. This limitation can be mitigated, however, if it is assumed that the voxel-voxel functional interactions between ROIs are sparse (i.e., have a low connectivity density) [28]. Under the assumption of sparseness (low connectivity density), the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm [29] is used to pre-select variables for inclusion in the MVAR model, and thus to overcome the problem of a limited number of data points. The LASSO algorithm has previously been tested on numerical experiments [30], gene-network data [31] and simulated and experimental fMRI BOLD data [28], [31]–[33].
This paper presents a novel application of the MVAR model to study voxel-based region-to-region interactions in the brain, particularly long-range, top-down interregional interactions in visuospatial attention. We demonstrate that the LASSO algorithm can be effectively used to pre-select model variables, thereby enabling estimation of the coefficients of a voxel-based MVAR model of two predefined ROIs. The originality of our methods derives from: (1) estimation of the MVAR model for fMRI voxel-level BOLD time series from two ROIs; (2) use of the LASSO algorithm for variable pre-selection prior to MVAR model estimation; (3) use of the General Cross-Validation criterion for determining optimal predictors in the MVAR model from the LASSO algorithm; and (4) creation of two types of summary statistics at the ROI level that represent the separate measurement of density and strength of GC between ROIs.
We report the results of both MVAR model simulations and the application of MVAR model estimation to an empirical fMRI BOLD dataset obtained during a visuospatial attention task [34]. The simulation results demonstrate that voxel-based approaches can better capture the GC between two ROIs than the averaging approach. When LASSO is used to pre-select variables for inclusion in the MVAR model estimation, voxel-based GC summary statistics are more sensitive to coefficient changes in the model than GC values computed from averaged signals. LASSO pre-selection allows MVAR model estimation to fit the simulated data accurately as long as the GC functional connectivity is sparse, i.e. has relatively low density. We also report that construction of the voxel-based GC distribution by pairwise bivariate AR model estimation, instead of by MVAR model estimation with LASSO pre-selection, may yield spuriously significant GC values. The empirical results show that the assumption of sparse GC functional connectivity is realistic, and that LASSO variable pre-selection followed by MVAR model estimation is thus effective, for empirical fMRI BOLD data. Also, the low GC connectivity density observed for this dataset suggests that directed interregional functional interaction in the brain is heterogeneous and that averaging the voxels of an ROI prior to GC connectivity analysis is inappropriate. Furthermore, the observed directional asymmetry, as measured by the GC strength summary statistic, is consistent with current theory on top-down modulation in visuospatial attention.
We conclude that LASSO variable pre-selection and MVAR model estimation can be effectively used to measure Granger Causality between cortical regions from voxelwise fMRI BOLD signals. Through the MVAR model, it is beneficial to analyze all the voxels in an ROI, instead of taking an average over the ROI. In this way, directed interregional functional interactions are captured with less distortion of the information carried in the BOLD time series.
Simulation MVAR models were created based on Equation 3 (see Methods), and iterated to generate simulated fMRI BOLD time series data for pseudo-voxels in two pseudo-ROIs having fixed sizes (30 pseudo-voxels in X and 50 pseudo-voxels in Y). The innovation process for the simulation model was created by iterative random sampling of a zero-mean normal distribution with 0.1 standard deviation. The predictors were initialized with random values also taken from a zero-mean normal distribution with 0.1 standard deviation. The four coefficient submatrices (Bxx, Byx, Bxy and Byy) were constructed separately. For each submatrix, some coefficients (bij) were randomly set to zero and the rest were randomly drawn from a normal distribution with zero-mean and a specific standard deviation (0.08 for Bxx and Byy, 0.2 for Byx, and 0.1 for Bxy). For each simulation, 200-point-long time series for each pseudo-voxel were created by model iteration. A total of 56 simulation models were created. The density of model connectivity was systematically increased with increasing model identification number by augmenting the number of voxel pairs connected by non-zero b values (Table 1).
To verify model validity, we determined that the correlations of the model residuals were low for all 56 models. A representative residuals correlation matrix from one of the simulations is displayed in the Figure S6, showing randomly distributed weak correlation across simulated voxel pairs. We considered the models to be valid based on these observations.
We then considered the effect of averaging the BOLD activity of all voxels in an ROI on the measurement of interregional GC. The GC between two ROIs, each of which is represented by an averaged time series, was measured by a single t-score in each direction (see Methods). For it to properly portray the connectivity between ROIs, the t-score was expected to follow the change of parameters across the simulation models shown in Table 1. We tested this prediction by measuring the correlation of the t-scores from the averaged time series with two summary statistics (the fraction of significant connections, f, and the average connectivity strength, W) (see Methods). These summary statistics were computed directly from the simulation models, and thus followed the change of parameters across the simulation models. The t-scores derived from averaged time series were not significantly correlated (p<0.05) with either summary statistic (Figure 2). That the t-scores did not follow the change of parameters across simulation models indicates that computing GC from averaged voxel time series does not accurately capture inter-ROI connectivity patterns.
We next examined how well voxel-based methods recovered the actual GC patterns of the four submatrices across the simulation models shown in Table 1. The analysis for each simulation model consisted first of estimating the full B matrix from the simulated data generated by that model using two methods: (1) pairwise-GC estimation; and (2) LASSO-GC estimation, i.e. LASSO pre-selection of variables for inclusion in an MVAR model, followed by GC estimation from the model. Then, the results from each method were compared with the actual values in the model. Each of these two methods is voxel-based. The pairwise-GC method constructs the B matrix by estimating a separate bivariate AR model for each voxel pair, whereas the LASSO-GC method computes the B matrix by estimating an MVAR model whose variables are pre-selected by LASSO. Unlike the approach of averaging across voxels, both methods compute a t-score for each b coefficient in the B matrix, testing whether the value of that coefficient significantly deviates from zero. A significant non-zero b value is equivalent to a significant GC value when the model order is one. Figure 3 illustrates the results from a simulation in which the LASSO-GC method (Figure 3B) closely estimated the pattern of b values of the model (Figure 3A), whereas the pairwise-GC method (Figure 3C) yielded a large number of spurious non-zero values.
To determine how typical were the results seen in Figure 3 across all simulation models, summary statistics from pairwise-GC and LASSO-GC estimations were compared with those computed directly from the models. First to be used was the f summary statistic, which measures the fraction of significant b values. Figure 4 compares how well the pairwise-GC and LASSO-GC methods recovered the actual f summary statistic computed directly from the simulation models. It reveals that in most simulations the f summary statistic from pairwise-GC estimation was greater than the actual simulation model value, whereas that from LASSO-GC estimation closely matched the actual simulation model value. We defined the distance between estimated and model f values by their absolute difference, and compared the distances resulting from the LASSO-GC method with that from the pairwise-GC method. Paired t-tests showed highly significantly (p<0.01) smaller distances with the LASSO-GC method for all four submatrices.
The W summary statistic, which reflects the average strength of significant GC from voxels in one ROI to voxels in another, was used next to compare the pairwise-GC and LASSO methods. As with the f statistic, the W statistic from the LASSO method matched the actual W statistic computed from the simulation model more closely than that from the pairwise-GC method (Figure 5). Also as with the f statistic, the distances between estimated and model W values for the two methods were compared. Paired t-tests showed highly significantly (p<0.01) smaller distances with the LASSO-GC method for all four submatrices.
The comparison of LASSO-GC versus pairwise-GC across 56 runs can be considered as 56 repeated tests of the two methods for their efficiency in estimating model parameters. The fact that LASSO-GC yielded more accurate estimations than pairwise-GC over a range of different parameter settings demonstrates LASSO-GC's robustness. To further validate this conclusion, we repeated each 56-run test on 20 separate iterations, each iteration using an independently generated dataset with the parameters from Table 1. The resulting GC patterns across the 20 iterations are consistent with those shown in Figures 2, 4 and 5 (as demonstrated in Figures S1, S2 and S3). To summarize the results up to this point, the LASSO-GC method was found to outperform the pairwise-GC method and the average-signal based method in recovering simulation model connectivity. We next applied the LASSO-GC method to explore functional connectivity in an empirical fMRI BOLD dataset.
An fMRI BOLD dataset from a slow event-related visuospatial attention task paradigm was analyzed with the LASSO-GC method. Details about the experimental design and the fMRI recording are available in [24] and [34]. Within each of 6 subjects, bilateral areas V1v, V2v, VP, V3A and V4 were in the Visual Occipital Cortex (VOC), and bilateral areas Frontal Eye Field (FEF) and anterior and posterior IntraParietal Sulcus (aIPS and pIPS) were in the Dorsal Attention Network (DAN). MVAR models of order-one were estimated from the time series of all voxels from each pair of VOC and DAN ROIs by the LASSO-GC method. The largest MVAR model contained approximately 150 voxels. Repeated trials (average number ∼70) at each time point were used as observations. For each ROI pair, a full B matrix was first estimated, and the f and W statistics were then computed for each of the four submatrices.
As with the simulation results, the correlations of the MVAR model residuals were found to be low, indicating that the models were valid. Because of the large data dimension, not all ROI pairs could be examined. Instead, 10 ROI pairs were randomly selected from each subject for examination: the residuals correlation matrix for one representative ROI pair is displayed in Figure S7. Since most of the correlation scores were weak (near or below 0.5), the MVAR models were considered to be valid representations of the fMRI BOLD data, and we thus proceeded to explore the connectivity patterns.
The results of functional connectivity analysis between the VOC and DAN are presented in Figure 6 for a representative ROI pair in one subject. GC connectivity diagrams are shown between the right VP region (having 25 voxels) in VOC and the right FEF region (having 56 voxels) in the DAN (Figure 6A). The four diagrams represent GC connectivity within right VP (VP → VP), from right FEF to right VP (FEF → VP), from right VP to right FEF (VP → FEF), and within right FEF (FEF → FEF). GC connectivity is sparse both within and between ROIs, meaning that a low fraction of t-scores is significant at p<0.05 (fVP−>VP = 0.08, fFEF−>VP = 0.04, fVP−>FEF = 0.06, fFEF−>FEF = 0.03). Both significantly positive (orange arrows) and significantly negative (blue arrows) GCs are present both within and between ROIs. A positive GC indicates that increased activity of the “sending” voxel predicts increased activity of the “receiving” voxel, whereas a negative GC signifies that increased activity of the “sending” voxel predicts decreased activity of the “receiving” voxel.
Connectivity based on the correlation measure is also considered. For correlations measured directly on the fMRI BOLD time series, a larger fraction of connections is significant at p<0.05 (fVP−VP = 0.67, fVP−FEF = 0.30, fFEF−FEF = 0.45) for the same ROI pair and subject (Figure 6B), suggesting that a large portion of the voxels are correlated. More sparse connectivity from LASSO-GC than from correlations is clearly seen in the f and W summary statistics (Figure 6C). This more sparse connectivity found with LASSO-GC than with correlation might have been a simple effect of LASSO pre-selection. However, this was found not to be the case since the correlations still showed much more dense patterns than the LASSO-GC results even when computed after LASSO pre-selection (Figures S4, S5). That the average correlation connectivity strength (W in Figure 6C) is greater than 2, both within and between ROIs, indicates that each voxel receives, on average, connections from more than 2 other voxels. This observation of relatively high correlation density suggests that the LASSO-GC method is needed to reduce correlation-induced spurious GC estimates.
To extend the functional connectivity analysis to the full fMRI BOLD dataset, we applied the LASSO-GC method to all 60 VOC-DAN ROI pairs in each of the 6 subjects. The f and W summary statistics were then averaged across ROI pairs and subjects, yielding mean f and W summary statistics for VOC-to-VOC connectivity, DAN-to-DAN connectivity, DAN-to-VOC connectivity, and VOC-to-DAN connectivity (Figure 7). These four connectivity types correspond to the four coefficient submatrices of the estimated B matrix in LASSO-GC analysis (see Methods): VOC-to-VOC and DAN-to-DAN connectivity refers to connectivity within a single region of VOC or DAN, not to connectivity between different VOC or DAN regions. The mean f summary statistic is below 0.1 for all submatrices, indicating overall sparse within- and between-ROI GC connectivity.
Paired-sample t-tests with subjects as repeated measures (df = 5 for all comparisons) were performed on both f and W to compare: (1) top-down (DAN-to-VOC) with bottom-up (VOC-to-DAN) connectivity; (2) within-VOC with within-DAN connectivity; (3) top-down with within-VOC connectivity; and (4) bottom-up with within-DAN connectivity. The comparison of top-down with within-DAN connectivity and the comparison of bottom-up with within-VOC connectivity were not performed because these comparisons are ambiguous, i.e., they are based on GC to voxels in a sending region, whereas the W summary statistic is based on voxels in a receiving region (see Methods). The results show that within-VOC (VOC→VOC) connectivity was significantly greater than top-down (DAN→VOC) connectivity for both the f (t = 4.60, p<0.05) and W (t = 2.85, p<0.05) summary statistics, indicating that the local GC between voxels within VOC is both more dense and stronger than the long-range, top-down GC from the DAN. Connectivity within DAN (DAN→DAN) was also significantly greater than that in the bottom-up direction (VOC-to-DAN) for the W summary statistic (t = 4.07, p<0.05), but not for the f summary statistic, indicating that the local GC between voxels within DAN is stronger, but not more dense, than the long-range GC from VOC. Finally, connectivity in the top-down direction (DAN→VOC) was significantly greater than that in the bottom-up direction (VOC→DAN) for the W summary statistic (t = 3.93, p<0.05) but not for the f summary statistic, indicating a long-range directional strength asymmetry between DAN and VOC, with stronger top-down connectivity.
We have shown that Granger Causality (GC) computed from voxel-level BOLD signals better reflects the pattern of directed functional interaction between ROIs than that computed from voxel-averaged signals. We conclude that brain regions are not unitary elements, that network structure exists at the voxel level, and that ROI-level GC connectivity is best measured by summary scores computed over voxel-level connectivity patterns.
We emphasize that our conclusions apply specifically to GC between pre-defined ROIs, and do not necessarily extend to the computation of maps showing GC between a “seed” signal, averaged over the voxels in one cortical region, and voxels throughout the rest of the cortex [35]. In fact, an interesting extension to the mapping approach has recently been proposed by Garg et al. [33]. Their technique, called Full-brain AutoRegressive Modeling (FARM), also adopts LASSO for voxel-level analysis. Although their use of LASSO to make MVAR modeling feasible for large numbers of voxels is similar to ours, their problem of GC mapping for the entire brain is different from the interregional interaction question that we have investigated. It is possible that full-brain mapping and interregional analysis will prove complementary to each other. To explore the relationship of a particular region to the remainder of the cortex, the mapping method would appear to be more appropriate since it yields a global functional interaction pattern. However, to test specific theories in cognitive neuroscience that involve particular cortical networks, might require the use of regions that are pre-defined from previous clinical or experimental evidence. In that case, one would be interested in examining the details of functional interaction between regions, and a full-brain LASSO algorithm could be insufficient because its tuning parameter might be too severe, making the model overly sparse: even if the global connectivity pattern were preserved, the details of interregional connectivity might still be lost. With these concerns in mind, one might use full-brain mapping as a first step to establish global functional interaction patterns, and then a more detailed exploration could be performed using voxel-based interregional analysis.
In addition to mapping, another common analytic method in the literature examines region-to-region correlations based on averaged signals and identifies topological properties from large-scale networks that involve hundreds of ROIs [36]. Our findings do not necessarily negate this approach: since GC and correlation are different measures, inhomogeneity in GC does not imply inhomogeneity in correlation. It is possible that correlation-based connectivity with averaged signals may be effective even though GC analysis requires a voxel-based approach.
We have demonstrated that the LASSO-GC method can better identify GC connectivity between ROIs in simulated fMRI BOLD data than the pairwise-GC method by more accurately estimating the connectivity density and strength. The pairwise-GC method can yield spuriously significant coefficients if correlated predictors are present in the MVAR model. The close fit of the LASSO-GC results to the actual results from the simulation models demonstrates that the LASSO-GC method is better able to avoid false positives, and also shows the sensitivity of this method in detecting model changes. By contrast, GC values computed from averaged data do not systematically follow changes in simulated ROI models, suggesting that summary statistics computed from voxel-to-voxel GCs are better able to represent ROI-level connectivity than single region-to-region GCs computed after averaging over ROI voxels.
The estimated f summary statistics from the LASSO-GC method matched the actual f statistics from the simulation models better when the B matrices were more sparse. Although the LASSO algorithm could potentially fail for high connectivity densities, we were not able to observe such a failure because the simulated voxel activity at high connectivity density becomes unstable. Nonetheless, it is unlikely that the low f values observed for the empirical BOLD data are artifactual because if the B matrices were ill-estimated, then the directional asymmetry found with the W statistic would not display the high degree of consistency across subjects that was observed. The fact that the range of f found for the empirical BOLD data fell within the range of f in the simulations further suggests the suitability of the LASSO-GC technique for application to BOLD data. Moreover, the low values of the f summary statistic from the empirical BOLD data indicate that GC connectivity between cortical ROIs is sparse. Given evidence from anatomical studies that axonal connectivity of the cortex is generally sparse [37], [38], it is more likely that the sparse GC connectivity reflects actual functional interaction patterns than that it is a mere statistical byproduct.
Directional asymmetry in GC connectivity between the Dorsal Attention Network (DAN) and Visual Occipital Cortex (VOC) was reported in our previous work [24] using the pairwise-GC method for computing GC and f as the summary statistic. Using the LASSO-GC method, we report here that the directional asymmetry is found in the W, but not the f, summary statistic (Figure 5). The difference in results from the pairwise-GC and LASSO-GC methods may be understood by examining the properties of the W summary statistic. The finding that W values in the top-down DAN-to-VOC direction are significantly greater than in the bottom-up VOC-to-DAN direction suggests that VOC voxels are modulated more strongly by DAN voxels than DAN voxels are by VOC voxels, despite there being similar fractions of voxels being modulated in both directions. The greater top-down modulation strength may have introduced a bias in the pairwise-GC results from our previous work, yielding an apparently greater fraction of significant top-down GC values. Relatively high correlation density (Figure 6) may have contributed to such a bias.
The problem of bias actually has multiple facets. It is known from theory that the LASSO method may be biased if predictors are highly correlated. There are two main problems caused by correlated predictors. First, some predictors in a system may not be included in the model of the system. This is the case when estimation of multiple bivariate AR models is employed in place of MVAR model estimation: the estimation may be biased by undetected influences from the excluded predictors. The use of LASSO helps to mitigate this problem by allowing estimation of a full MVAR model. Second, even when all the predictors are taken into account, correlation among predictors may still bias model estimation, a situation often referred to as the collinearity problem for multiple regressions. An example of such bias would be the case of a group of strongly correlated predictors, where LASSO tended to select only one predictor from the group. Extensions of LASSO have been proposed to mitigate this problem by selecting the entire group instead of a single predictor. Such extensions include fused LASSO [39], the elastic net [40] and the group LASSO [41]. A thorough review of this issue is available in [42]. Whether such selection bias becomes a problem in brain network analysis depends on the specific research question being considered. On the one hand, if the exact relationship among predictors is of central interest, as when large-scale cortical network structure is explored using ROIs as predictors [28], one may consider the use of extended LASSO algorithms to avoid losing important correlated ROIs. On the other hand, in our voxel-based investigation of pre-selected ROIs, the voxels are used as multiple representations of the corresponding ROIs, and omitting some of the correlated voxels in a group is not expected to radically change the collective functional connectivity at the ROI level. In the present study, we are more interested in the summary statistics over the entire connectivity matrix than in the details of the connectivity patterns within the matrix. Thus, in our study, although the collinearity problem may exist, and a complete solution is not currently available from theory, our results are nonetheless not invalidated. We found that the f and W summary statistics effectively recovered the modeled connectivity values from simulated data, even though those data had significant correlations between most voxel pairs. Furthermore, in empirical BOLD data analysis, it is often desirable to compare summary statistics across different conditions rather than to precisely identify their values. For such comparison, any possible bias introduced by voxel-voxel correlations would exist in both conditions and thus would not alter the comparison.
Although the MVAR models used in this paper were implemented with order one, models having higher order (p>1 in Equation 3) can be implemented within the same framework. For model orders greater than one, multiple b coefficients at different time lags (t-k) contribute to the GC from one voxel to another, and it is not sufficient simply to test the significance of a single b coefficient. In that case, testing for significant between-voxel GC would need to be performed differently, and the summary statistics would accordingly be defined differently. For example, a criterion for between-voxel GC to be significant might be that at least one of the b coefficients from different lags must be significant. A summary statistic equivalent to f might then be defined as the fraction of significant between-voxel GCs rather than the fraction of significant b values. Similarly, a summary statistic equivalent to W could base the average strength of significant GC on all significant b values for a voxel pair instead of a single b value. A straightforward way to do this would be to sum the significant b values from different lags over all inputs to receiving voxels. In this way the W statistic would be sensitive to three factors: the magnitude of all significant b values, their corresponding time lags, and the total number of converging significant inputs to receiving voxels. However, to compare the W statistic between models of different order, the time-lag factor would need to be removed, possibly by averaging b values over time lags, to avoid bias due to different total numbers of b values.
In conclusion, our work suggests that the LASSO algorithm can be effectively employed for pre-selection of voxels that are then used in an MVAR model to measure functional connectivity between ROIs, using voxel-based fMRI BOLD signals. It indicates that the f and W summary statistics reveal different aspects of directed influence between ROIs. Used in tandem, these statistics may provide consistent information about influences between brain regions that is richer than that from either one alone. Additional summary statistics will likely be found in the future that will further our understanding of directed influences between brain regions.
We first consider an fMRI BOLD dataset from m voxels in ROI X and n voxels in ROI Y. The dataset consists of time series of t points recorded from every voxel in X and Y, and can be written in matrix form as:(1)(2)
The relationship between X and Y can be expressed in the form of a Multivariate Vector AutoRegressive (MVAR) model. A general matrix representation of the model is:(3)where Zt is the dependent variable in vector form, representing the BOLD data values at arbitrary time t of all voxels in X and Y; Zt-k represents the values of the Z vector at arbitrary earlier time point t-k; lag k ranges from 1 to p, the model order; Bk is the corresponding coefficient matrix at lag k; and Et is the residual vector.
When expanded, the product term in Equation 3 becomes:(4)
Each element of the Zt-kth vector is a predictor, and each element (bkij) of the Bk matrix is a coefficient representing the degree of prediction of the ith element of Zt by the jth predictor. If a value of bkij significantly differs from zero, then a significant GC is said to exist from voxel j to voxel i. The magnitude (strength) of that GC may be assessed by the magnitude of the statistic (e.g. t-statistic) used to measure the difference of the b value from zero. The sum of product terms over all lags is the total prediction of Zt by the model.
The model order (p) was set to one in this paper, based on our prior experience with the analysis of fMRI BOLD data [24]. The MVAR model in Equation 3, with model order one, was used here for both simulation and GC analysis. For simulation, the residual vector represented an innovation process that generates random values, the B matrix was known, and the X and Y time series data were simulated. For GC analysis, the X and Y time series data were known, the B matrix was estimated in order to determine GC, and the residual vector represented prediction errors.
We also employed pairwise-GC analysis for comparison with MVAR analysis. In the pairwise-GC approach, coefficients are estimated (and the significance of GC determined) by constructing a separate bivariate model for each pair of voxels, one in X and one in Y:(5)
In pairwise-GC analysis, the assumption is made that the predictors are independent of one another. Under this assumption, the GC between X and Y can be assessed solely from the bivariate models in Equation 5, and it is not necessary to estimate the coefficients representing GC within X or Y. In fact, however, the predictors may be correlated for BOLD time series, making the pairwise-GC approach problematic. If the predictors are correlated, estimation by separate bivariate (or partial) models may be biased, and all of the coefficients in the B matrix should be estimated simultaneously [43]. Nonetheless, simultaneous estimation may be impossible in the analysis of data from neurobehavioral studies, in which the number of observations is often limited.
The Least Absolute Shrinkage and Selection Operator (LASSO) technique is a method that makes model estimation feasible when only a limited number of observations is available. Under the assumption that the B matrix is sparse (i.e., many coefficients are zero), the LASSO algorithm effectively determines which b values are actually zero. Our goal in using LASSO is to identify non-zero coefficients and then estimate them simultaneously, thus avoiding bias due to partial regression with correlated predictors. The pre-selection process in LASSO involves determining an optimal set of predictors.
In the MVAR model, pre-selection is carried out in a row-wise manner. LASSO adds a constraint on each row equation of Equation 3 that restricts the total absolute values of the coefficients. The constraint is expressed as:(6)where c is a tuning parameter.
Regression of the ith row of Equation 3 under the constraint provided by Equation 6 is equivalent to the regression of:(7)
Finding a least-squares solution of Equation 7 requires a subset of the b values to be set to zero. To achieve this goal we use the Least Angle RegreSsion (LARS) algorithm developed by Efron et al. [44], which starts with all b's equal to zero and then iteratively adjusts their values to fit the model. Some of the b's remain zero after the adjustment, resulting in the identification of an optimal set of non-zero b values for a particular c value (corresponding to λ in Equation 7).
The next step in model estimation is to tune the parameter c to achieve a best fit of Equation 7. The minimum value that c can take is zero, corresponding to the extreme case where all b's are zero. The upper boundary is reached when LARS does not penalize any b to zero, making c equal to the sum of the absolute values of all b's. Within this interval, a number (approximately 100 in our case) of c values is chosen to compute the subsets of b's. For each c value, after an optimal subset has been found, the corresponding Residual Sum of Squares (RSS) is used to calculate a General Cross-Validation (GCV) statistic [28]:(8)where n is the number of independent observations and df is the estimated degrees of freedom from the LARS algorithm. From all the solutions, a GCV curve is plotted. The minimum GCV value determines the single most optimal set of predictors over all c values.
A subsequent Ordinary Least Squares (OLS) procedure is then applied to the new row equation with the selected predictors. To avoid using the same data to estimate the LASSO and OLS models, we randomly sort the data trials into two sets. One set is used to estimate the LASSO coefficients for model selection, and the other is then used to estimate OLS coefficients for the new row equation. In the second step, if the model order is one, as in our application, there is only one coefficient for each predictor. Either an F-test or a t-test is performed for each coefficient to determine whether its value is significantly different from zero. The resulting F-score or t-score characterizes the prediction by a predictor on the RHS of Equation 3 of the dependent variable on the LHS, and corresponds to the GC strength from that predictor to the dependent variable. Here we used the t-score to measure GC because it has a signed value, and thus indicates whether the GC is enhancing or reducing, in addition to indicating GC strength.
The full B matrix may be estimated by following the above procedures for every row equation in Equation 3. It consists of four submatrices (Bxy, Byx, Bxx, and Byy), where the first subscripted index represents the predictor and the second represents the dependent variable. Thus, Bxy represents connectivity from X to Y, Byx represents connectivity from Y to X, and Bxx and Byy represent connectivity within X and within Y, respectively. In order to measure GC from one ROI to another (i.e., X→Y or Y→X), one or more statistics are needed to summarize the voxel-to-voxel GCs represented by significant coefficients in Bxy or Byx.
The first summary statistic that we used was the fraction (f) of significant b values in the B matrix or one of its submatrices (representing the fraction of significant GCs) (Figure 8). The fraction of b values found to be significantly different from zero at p<0.05 was corrected for multiple-comparisons by the False Discovery Rate (FDR). This summary statistic is a measure of density of the ROI-level connectivity. Because each b value represents a potential functional “connection”, the f summary statistic summarizes the fraction of all possible voxel-to-voxel connections from one ROI to another by which the two ROIs are actually connected.
The second summary statistic used was the average strength of significant GC from voxels in one ROI to voxels in another (Figure 8). Consider, for example, the GC from (“sending”) ROI Y to (“receiving”) ROI X. Significant voxel-to-voxel GCs are represented by significant coefficients in Byx. For any given voxel x in ROI X having at least one significant (p<0.05) t-score (indicating a significant GC) from ROI Y, we first summed the t-scores of all the GCs to x. This sum represents the total significant “input” to the “receiving” voxel x from all “sending” voxels in Y. Because the t-scores can be positive or negative, signifying that changes of activity in the “sending” voxel contribute to a change of activity in the “receiving” voxel either in the same or opposite direction, the sum of t-scores takes into account the balancing effect of positive and negative inputs to the same receiving voxel. We then computed the average strength of significant input over all receiving voxels in ROI X as the W summary statistic. The significance threshold for determining the number of significant inputs was corrected for multiple-comparisons by the FDR. The same procedure was also followed to assess the average strength of significant GC in the other direction, i.e. from ROI X to ROI Y using Bxy. Although not the focus of this paper, W could also be computed to assess the average strength of significant GC within ROI X using Bxx, or within ROI Y using Byy. W measures the average strength of GC from ROI Y to ROI X, but is not simply a weighted version of the f summary statistic. For example, a high W value from Y to X depends on a combination of the following: 1) many voxels in Y have high GC values to voxels in X; and 2) single voxels in X have significant GC values from multiple voxels in Y.
Simulation models were constructed using the R statistical computing package. For the purpose of comparing the GC strength of a simulation model with its estimated values, we computed the simulation W statistic directly from the b values of the simulation model. To make the estimated and simulation W measures comparable, we normalized the t and b values to z-scores (i.e. subtracted the mean and then divided by the standard deviation). Computer code used in this study will be freely provided for legitimate research purposes upon request from the first author.
|
10.1371/journal.pntd.0001971 | Mesenchymal Bone Marrow Cell Therapy in a Mouse Model of Chagas Disease. Where Do the Cells Go? | Chagas disease, resulting from infection with the parasite Trypanosoma cruzi (T. cruzi), is a major cause of cardiomyopathy in Latin America. Drug therapy for acute and chronic disease is limited. Stem cell therapy with bone marrow mesenchymal cells (MSCs) has emerged as a novel therapeutic option for cell death-related heart diseases, but efficacy of MSC has not been tested in Chagas disease.
We now report the use of cell-tracking strategies with nanoparticle labeled MSC to investigate migration of transplanted MSC in a murine model of Chagas disease, and correlate MSC biodistribution with glucose metabolism and morphology of heart in chagasic mice by small animal positron emission tomography (microPET). Mice were infected intraperitoneally with trypomastigotes of the Brazil strain of T. cruzi and treated by tail vein injection with MSC one month after infection. MSCs were labeled with near infrared fluorescent nanoparticles and tracked by an in vivo imaging system (IVIS). Our IVIS results two days after transplant revealed that a small, but significant, number of cells migrated to chagasic hearts when compared with control animals, whereas the vast majority of labeled MSC migrated to liver, lungs and spleen. Additionally, the microPET technique demonstrated that therapy with MSC reduced right ventricular dilation, a phenotype of the chagasic mouse model.
We conclude that the beneficial effects of MSC therapy in chagasic mice arise from an indirect action of the cells in the heart rather than a direct action due to incorporation of large numbers of transplanted MSC into working myocardium.
| Chagas disease, resulting from infection with the parasite Trypanosoma cruzi, is a major cause of heart disease in Latin America. Treatment options are limited to a small number of drugs that were developed more than four decades ago and which have various drawbacks. Stem cell therapy with bone marrow mesenchymal cells (MSCs) has emerged as a novel therapeutic option for cell death-related heart diseases, but efficacy of MSCs has not been tested in Chagas disease therapy. Due to the established role of the immune system in the physiopathology of Chagas disease and the immune modulatory properties of MSC we hypothesized that MSC could be an optimal cell type for therapy in chagasic cardiomyopathy. Therefore, in this study we have used cell tracking strategies following labeling of MSCs with nanoparticles to investigate migration of transplanted MSCs in a murine model of Chagas disease, and have correlated MSCs migration with cardiac function in chagasic animals by small animal positron emission tomography imaging technique.
| Chagas disease is a serious public health problem in all Latin American countries [1], where it is estimated that 15–16 million people are infected with the its causative agent, the parasite Trypanosoma cruzi (T. cruzi) [2]. Although T. cruzi is endemic in Latin America, thousands of people are infected in Europe, United States, Canada, among other countries, due to migration of infected people [3], [4].
Approximately one-third of individuals with Chagas disease develop a symptomatic chronic phase decades after the infection, of which 90% develop heart disease and the other 10% are affected by gastrointestinal diseases [5]. Chronic Chagas heart disease is a progressive, fibrotic inflammatory cardiomyopathy that results in permanent heart damage [6]. This heart damage leads to dilation and cardiac arrhythmia, and ultimately to congestive heart failure, which is the primary cause of death in chronic Chagas heart disease patients [7], [8]. For more than 40 years, the only treatment option for Chagas disease in the acute phase has been the anti-parasitic drugs nifurtimox and benznidazole. However, these drugs have side effects and lead to parasite resistance [9]. In the chronic phase, when congestive heart failure ensues, heart transplantation is often the only therapeutic option, which is also fraught with many problems.
In this complex scenario, where an estimated 20,000 people die of chronic Chagas heart disease each year [1], cell therapies appear as an alternative solution. In a mouse model of chronic chagasic cardiomyopathy (CCC) we have previously shown that mononuclear cells from the bone marrow decrease inflammation and fibrosis, reduce or reverse right ventricular dilation and significantly restore gene expression pattern to that of control, non-infected hearts [10]–[12]. However, given the established role of the immune system in the physiopathology of Chagas disease [13] and the immune modulatory properties of bone marrow mesenchymal cells (MSC) [14] we hypothesized that MSC could be an optimal cell type for therapy in chagasic cardiomyopathy. In addition, preliminary studies with mononuclear cells from chronic chagasic patients have revealed a diminished colony forming capacity (unpublished data), which can compromise autologous therapy. Due to the immune privileged characteristics of MSC, these cells can be used as an allogenic product [15]. Furthermore, previous studies with cellular therapy have focused primarily on the chronic phase of the disease and data about the effect of cellular therapy at early stages, such as 1 month after infection, was not previously evaluated. Thus, we wanted to examine the hypothesis that cell therapy is effective at earlier stage of the disease.
Therefore, in this study we describe the use of cell tracking strategies following labeling of MSC with nanoparticles to investigate migration of intravenously transplanted cells in an acute murine model of T. cruzi infection. Furthermore, we correlated MSC migration with glucose metabolism and morphology of heart in chagasic mice by small animal positron emission tomography (microPET).
All experiments were performed on adult male CD-1 mice in accordance with the U.S. National Institutes of Health Guide for the Care and Use of Laboratory Animals (NIH Publication No. 80-23), approved by the Institutional Animal Care and Use Committee of the Albert Einstein College of Medicine.
To obtain bone marrow cells, tibias and femurs of approximately 8 week old mice were isolated, the epiphyses were removed, the bones were individually inserted in 1 mL automatic pipette polypropylene tips and then put in 15 mL tubes. The bones were centrifuged at 300× g for 1 min and the pellets suspended in Dulbecco's modified Eagle's high glucose medium (DMEM; Invitrogen Inc., Carlsbad,CA), supplemented with 10% fetal bovine serum (FBS; Invitrogen Inc.), 2 mM l-glutamine (Invitrogen Inc.), 100 U/mL penicillin (Sigma-Aldrich Co., St. Louis, MO), and 100 µg/mL streptomycin (Sigma-Aldrich). The cells were plated in 100 mm culture dishes with supplemented DMEM and maintained in 5% CO2 atmosphere at 37°C. The medium was replaced 48–72 hrs after initial culture to remove non adherent cells and the adherent cells were grown to confluence before each passage. Medium was replaced three times a week. All experiments were performed on second or third passage cells.
In the present study we used fluorescent nanoparticles called X-Sight nanospheres (Carestream Health Inc., Rochester, NY): X-Sight 761 (761 nm excitation and 789 nm emission) and X-Sight 549 (549 nm excitation and 569 nm emission). We incubated MSC with a solution of 0.3 mg/mL X-Sight in supplemented DMEM in 5% CO2 atmosphere at 37°C for 4 hours.
The labeled cells with X-Sight were then washed three times with phosphate-buffered saline (PBS), trypsinized and centrifuged at 300× g for 5 min. Subsequently, the labeled cells were used for in vitro experiments or for tracking after transplant.
The Brazil strain of T. cruzi was maintained by serial passage in C3H mice (Jackson Laboratories, Bar harbor, ME). Eight to 10 week old male CD-1 mice (Charles River) were infected by intraperitoneal injection of 5×104 trypomastigotes in saline solution. One month after infection (1MAI) these mice received a single dose of 3×106 MSC in 100 µL of PBS, or 100 µL of PBS via tail vein. For cell tracking, both control and chagasic mice received single doses of 3×106 labeled MSC via tail vein.
The X-Sight 761-labeled MSC were visualized by the in vivo imaging system (IVIS) Kodak Image Station 4000MM PRO (Carestream Health) equipped with a CCD camera. For the fluorescence imaging, the machine was configured for 760 nm excitation, 830 nm emission, 3 min exposure, 2×2 binning and f-stop 2.5. The acquired images were analyzed with the Carestream MI Application 5.0.2.30 software (Carestream Health).
For in vitro visualization of labeled cells, the MSC were grown on glass coverslips coated with 0.2% gelatin, incubated with X-Sight 549 for 4 hours, washed with PBS and fixed for 20 min in 4% paraformaldehyde. The cells were then observed by confocal microscopy to ascertain intracellular incorporation of the particles.
Besides the IVIS technique, we tracked the labeled cells in the heart by microscopy. The same hearts used for IVIS tracking were fixed overnight in 4% paraformaldehyde, incubated in optical cutting temperature resin (Sakura Finetek USA, Inc., Torrance, CA) and sliced in 5 µm frozen sections. The photomicrographs shown in this study were obtained using a Zeiss LSM 510 Duo confocal microscope.
Mice were administered 300–400 µCi (12–15 MBq) of [18F] fluoro-2-deoxyglucose (18F -FDG) in 100 µL saline via tail vein and imaging was started 1 hour after injection. Imaging was performed on an Inveon Multimodality scanner (Siemens Healthcare, Erlagen, Germany) using its PET module. The mice were anesthetized with isoflurane inhalation anesthesia (2% in 100% oxygen) administered via a nose cone. PET imaging was performed using the PET gantry which provides 12.7 cm axial and 10 cm transaxial active field of view. The PET scanner has no septa and acquisitions are performed in 3-D list mode. A reconstructed full-width-half-maximum (FWHM) resolution of <1.4 mm is achievable in the center of the axial field of view. After each acquisition (approximately 3 minutes), data were sorted into 3D sinograms, and images were reconstructed using a 2D-Ordered Subset Expectation Maximization algorithm. Data were corrected for dead-time counting losses, scatter, random coincidences and the measured non-uniformity of detector response (i.e., normalized) but not for attenuation. Analysis was performed using ASIPRO and IRW (Siemens Healthcare) dedicated software.
The images were acquired from animals at 1 month after infection (before treatment) and 15 and 30 days after treatment with PBS or cells. Control group imaging was performed at every time point and for statistical analysis they were combined, thus, the technical number of control animals was 12. In addition, as the collected data at 15 days after treatment was very similar to the collected data at 30 days, we combined these time points to increase the sample number in a group called 15–30d, thus, the technical number of animals were 8 and 9 for the PBS and MSC treated groups, respectively.
Statistical significance was evaluated using one-way ANOVA with Newman-Keuls post-test for comparison among multiple groups. All calculations were done using GraphPad Prism 5 for Windows (GraphPad Software, San Diego, CA) and p<0.05 was considered as statistically significant. The data are presented as mean and the error bars represent the standard error of the mean.
By microscopy, we observed that all of the cells were labeled with X-Sight nanoparticles in vitro after 4 hours of incubation (Figure 1A-A″) and because of that we considered that it was not necessary to quantify the number of labeled cells. By confocal microscopy, we confirmed that the nanoparticles were incorporated into the cell cytoplasm (Figure 1B). We did not observe a cytotoxic effect of the X-Sight on cellular proliferation, evaluated by ki67 antibody, or on viability, evaluated by trypan blue staining (data not shown). We analyzed the retention time of nanoparticles in vitro for up to 4 weeks using different cell plating densities, and we observed a substantial loss of fluorescence intensity over time (Figure 1C), likely due to cellular proliferation as previously described by us [16]. The wells plated at 5×105 cells could not be monitored beyond 2 days because of high confluence and, consequently, cellular death. However lower plating densities allowed us to detect signals for up to 1–4 weeks. At 2 days after initial exposure to X-Sight for 4 hours a direct relation between cell number and fluorescence intensity was observed and a small number of cells, as low as 5×103 was detected (Figure 1D and E).
Control and chagasic mice at 1 month post infection received X-Sight 761-labeled MSC via tail vein. The images were acquired by IVIS, 2 or 15 days after cell transplantation. A weak signal from the labeled cells was observed in whole body images (Figure 2A) and a better signal was detected in ex vivo organs (Figure 2B and H). Despite the filters being set to near infrared excitation and emission, a basal level of fluorescence was detected in control mice (Figure 2A-CTRL), which did not receive cells, and similar fluorescence intensity was found in the infected mice that received unlabeled cells, indicating that neither the disease nor the unlabeled cells affected the basal fluorescence (data not shown).
From all analyzed organs, including heart, bladder, lung, liver, spleen and kidney we observed that approximately 70% of the fluorescence was localized in the lung, liver and spleen of control and chagasic mice (Figure 2C–H). We also harvested some tissue samples, including leg muscle, white and brown adipose tissue but we did not observe cell migration to these tissues (data not shown). When we compared the fluorescence intensity in the organs, 2 and 15 days after transplantation, a decrease of approximately 60% in total intensity was seen (Figure 2C–H). Based on the possibility that nanoparticles might be released and secondarily label other cells, such as macrophages, we injected free X-Sight in the animals. In contrast to the distribution of labeled cells, free X-Sight was distributed more widely in whole body and about 60% of signal was found exclusively in the liver (data not shown).
It was interesting to note that despite the fluorescence signal in lung and spleen was stronger in control animals than in chagasic animal 2 days after therapy (Figure 2E and H), in the heart we noticed the opposite. The quantification of fluorescence intensity showed that signal from chagasic hearts was statistically higher when compared to hearts obtained from control mice when ex vivo images of heart were compared 2 days after transplantation (Figure 2C), suggesting the homing of cells to the, most affected tissue by the disease (Figure 2C). In Figure 3A and B it is possible to observe with more detail the ex vivo images and graph of the hearts evaluated. In histological examination of the heart there were rare X-Sight-labeled cells in this tissue by confocal microscopy (Figure 3C-C″), which corroborates our data shown in Figure 2B–H where it is possible to note that only few cells migrate to this organ when compared to lung, liver and spleen.
All analyzed organs in this experiment were weighed and wet weights were found not to be affected by the infection, except the spleen. The spleens of the chagasic animals were heavier than control, independent of PBS or MSC treatment for 1 month (84.35±2.5; 236.4±23.7 and 229.9±30.9 mg, for control or chagasic mice treated with PBS or MSC, respectively).
MicroPET was performed with two main goals: to evaluate the glucose metabolism of the heart using the radioactive tracer 18F-FDG, and to measure the right ventricle (RV) dilation which is typical of the murine Chagas disease model [12], [17]. Figure 4A represents a whole body image, in a horizontal plane, from an animal that received 18F-FDG. In high magnification, in a transversal plane, it was possible to visualize the heart of control (Figure 4B), mice infected for 1 month and treated with PBS (Figure 4C) or MSC (Figure 4D). Note the size of the RV in Figure 4C, from an untreated mouse. A high glucose activity was observed 1 month after infection (without treatment) in the LV (Figure 4E) as well as in the RV (Figure 4F). However, 15–30 days after PBS treatment, a decrease in the glucose activity was observed in both ventricles what was not observed in the MSC treated groups. These data indicate that MSC can increase the glucose metabolism in infected hearts. Besides the heart, we observed a high glucose uptake in the brain, but we found no difference in uptake among the different experimental groups (data not shown). When the RV area was measured, the dilation observed in the RV of mice 1 month post infection was significantly reduced by cell therapy 15–30 days after cell transplantation (Figure 4G).
There are two types of cell therapy approaches that have been applied to animal models of Chagas disease: bone marrow mononuclear cells in mice [10]–[12] and co-cultured skeletal myoblasts with MSC in rats [18]. In human clinical studies of patients with end-stage heart failure due to Chagas disease the administration of autologous bone marrow mononuclear cells did not improve cardiac function [19]. The transplant of skeletal myoblast was associated with cardiac arrhythmias due to their inability to form electric coupling with cardiac myocytes [20]. Thus, these cells are not recommended for a disease with high incidence of arrhythmias, such as Chagas disease. On the other hand, MSC electrically couple to host cardiac myocytes and they have been suggested as a better cell type for cardiac therapy than other cell types, such as skeletal myoblasts [15].
MSC is an immune privileged cell type which can interact with cells of both the innate and adaptive immune systems and release trophic factors [14], [21]. Hence, MSC might modulate the inflammation and reverse the tissue damage caused by T. cruzi infection. However there is no previous report of therapy for Chagas disease using only MSC. The present study thus pioneers the analysis of MSC therapy and biodistribution of these cells in infected mice.
In contrast to myocardial infarction which causes a regional damage, Chagas disease affects the heart globally. Therefore, systemic delivery of MSC has been chosen for small animals infected with T. cruzi since multiple local injections into several heart areas would be expected to generate tissue damage [22]. Thus, it is very important to identify the preferential sites of MSC migration and correlate with their effects in cardiac function. In the present study an efficient visualization of X-Sight-labeled MSC was obtained and a small cell number as low as 5×103 could be detected in vitro by IVIS. However, there was a rapid decrease in fluorescence intensity over time. Based on our previous study with MSC labeled with superparamagnetic oxide iron nanoparticles [16], cellular proliferation seems largely responsible for the signal decrease observed in vitro. Regarding cell homing to the site of infection, we observed that migration was significantly higher to the hearts of infected mice when compared to controls. MSC migration to the damaged tissue has been reported by several authors in different models such as tumors [23], arthritic joints [24], middle cerebral artery occlusion [25] and myocardial infarction [26]. Although MSCs reside in specialized niches their perivascular location allows global access to the tissues and when they migrate to an injured region they may secrete large amounts of immune regulatory and trophic bioactive factors [21]. The exact mechanisms and molecules involved in migration of MSCs to areas of inflammation are unknown. It is assumed that the process of MSC migration is similar to that of leukocytes [27]. This process comprises different types of molecules such as chemokines and their receptors, adhesion molecules and proteases [28]. The increase in chemokine concentrations at the site of inflammation is crucial for the MSC migration to injury site. The stromal cell-derived factor-1 (SDF-1) is a member of the inflammatory chemokine family and stimulates the migration of various progenitor cells to injury site, including hematopoietic stem cells and MSCs, due to the CXC chemokine receptor type 4 (CXCR4) [29].
Although some cells migrate to the heart, as visualized by IVIS and confocal microscopy, the quantity is negligible when compared to liver, lungs and spleen, where about 70% of the fluorescence intensity was found. Our findings are consistent with other studies where the majority of intravenous injected MSC were found in the liver, lung and spleen, including in dogs with myocardial infarction [26] and patients with cirrhosis [30]. The fluorescence intensity 15 days after transplantation was greatly reduced, which is likely due to nanoparticle exocytosis, cellular proliferation and/or death [31]. It has been shown that most MSC die within days or weeks of transplantation, yet their beneficial effects can be seen over a much longer term, suggesting a critical time window for MSC action [32].
The radioactive tracer 18F-FDG has been used to analyze the area of infarcted myocardium in mouse [33] and humans [34] and seems to be helpful in the diagnosis of infection and inflammation [35]. Here we used the 18F-FDG technique to evaluate glucose metabolism and morphology of the hearts by microPET. 18F-FDG uptake was increased in chagasic animals 15–30 days after infection and decreased to control levels 45–60 after infection (1MAI + PBS 15–30d group). Although these data differ from those of a previous study from our group, which showed the increase of 18F-FDG uptake in all time points studied [17], in both studies there is a peak of uptake at 15–30 days after infection that corresponds to the peak of parasitemia, 25–30 days after infection in this model [36]. Although the incorporation of 18F-FDG is related to the general glucose metabolism, several authors suggest that the incorporation increase in tissues may also be related to inflammation and infection events [35], [37]. It was interesting to note that MSC therapy increased 18F-FDG uptake in the heart, since the number of MSC present in this organ is very low, as revealed in the tracking experiment we suggest that the MSC induced increase in 18F-FDG uptake is due to the known effects of MSC in damaged tissues, such as enhanced angiogenesis, stimulation of mitosis in stem and progenitor cells, recruitment of circulating stem cells, inhibition of apoptosis and/or change in extracellular matrix composition [14], [38]. We performed western blot analyses of heart tissue which reveled that MSC did not modify inflammatory proteins, such as interferon-γ (INF- γ) and interleukin 1β (IL-1β) and 10 (IL-10) in chagasic animals at a time point 1 month after infection with 1 month of therapy (total of 2 months of infection). Evaluating another time point of treatment (at 2 months after infection) we did not observe alterations in these proteins at 1 month after treatment (total of 3 months of infection) either; however, we did note differences in INF- γ and IL-10 due to MSC therapy after 2 months of treatment (total of 4 months of disease). Thus, despite the fact that we did not observe an immunomodulation after 1 month of therapy, we have obtained evidence that MSCs are able to immunomodulate after a longer term (data not shown). The remodeling of the right ventricle (RV) has been shown to be a characteristic phenotype of the chagasic mouse model used by our group [12], [17], [36], [39]. RV dysfunction was described as a predictor of mortality in patients with chagasic cardiomyopathy [40]. Regarding the heart morphology evaluation by microPET, we observed that the RV dilation caused by T. cruzi infection was reduced after cellular therapy. This result demonstrates that MSC therapy is able to reduce the RV dilation in Chagas disease model and corroborates with another study from our group, in which magnetic nuclear resonance was used to show that RV dilation was reduced after bone marrow mononuclear cell transplantation [12].
To summarize, this study is the first to use MSC for therapy and cell tracking in chagasic mice. It was interesting to note that despite a very small number of cells migrating to the heart when compared to the other organs, a statistically significant preferential migration to the damage heart was observed. Since the vast majority of the intravenous injected cells migrated to lung, liver and spleen we suggest that the beneficial effect observed by MSC cell therapy in chagasic mice is due to an indirect action of the cells in the heart rather than a direct action by incorporation of large numbers of MSC into the working myocardium.
|
10.1371/journal.pcbi.1006052 | Epigenetic regulation of cell fate reprogramming in aging and disease: A predictive computational model | Understanding the control of epigenetic regulation is key to explain and modify the aging process. Because histone-modifying enzymes are sensitive to shifts in availability of cofactors (e.g. metabolites), cellular epigenetic states may be tied to changing conditions associated with cofactor variability. The aim of this study is to analyse the relationships between cofactor fluctuations, epigenetic landscapes, and cell state transitions. Using Approximate Bayesian Computation, we generate an ensemble of epigenetic regulation (ER) systems whose heterogeneity reflects variability in cofactor pools used by histone modifiers. The heterogeneity of epigenetic metabolites, which operates as regulator of the kinetic parameters promoting/preventing histone modifications, stochastically drives phenotypic variability. The ensemble of ER configurations reveals the occurrence of distinct epi-states within the ensemble. Whereas resilient states maintain large epigenetic barriers refractory to reprogramming cellular identity, plastic states lower these barriers, and increase the sensitivity to reprogramming. Moreover, fine-tuning of cofactor levels redirects plastic epigenetic states to re-enter epigenetic resilience, and vice versa. Our ensemble model agrees with a model of metabolism-responsive loss of epigenetic resilience as a cellular aging mechanism. Our findings support the notion that cellular aging, and its reversal, might result from stochastic translation of metabolic inputs into resilient/plastic cell states via ER systems.
| Cell reprogramming, a process that allows differentiated cells to re-acquire stem-like properties, is increasingly considered a critical phenomenon in tissue regeneration, aging and cancer. In light of the importance of metabolism in controlling cell fate, we designed a computational model capable of predicting the likelihood of cell reprogramming in response to changes in aging-related metabolites. Our predictive mathematical model improves our understanding of how pathological processes that involve changes in cell plasticity, such as cancer, might be accelerated or attenuated by means of metabolic reprogramming.
| Aging is associated with profound changes in the epigenome involving large disturbances of the epigenetic landscape and genome architecture [1, 2]. Studies in model organisms have not only revealed the complex changes occurring in chromatin structure and functioning during aging, but also the remarkable plasticity of age-associated epigenetic marks [3–5]. Thus, whereas epigenetic alterations in DNA methylation, post-translational modification (PTM) of histones and chromatin remodelling are considered highly conserved hallmarks of aging [4, 6], the ability of cellular reprogramming-driven epigenetic remodelling to ameliorate age-associated phenotypes has been described recently. This finding unequivocally supports the causative role of epigenetic dysregulation as a driver of aging [7]. The reversible nature of epigenetic regulation of aging is receiving increasing attention as it might offer a revolutionary strategy to simultaneously delay or reverse a spectrum of diseases, including cancer, clustered in older individuals [8, 9]. A mechanistic understanding of the dependence and inter-relationship between aging and the functional status of specific epigenetic modifiers, for example histone demethylases (HDMs) and histone deacetylases (HDACs), is largely lacking.
There is an increasing awareness of the relationship between epigenetic modifiers and metabolism. Common metabolites of intermediary metabolism, such as acetyl-CoA, NAD+, α-ketoglutarate, succinate, FAD, ATP or S-adenosylmethionine, drive epigenetic processes by directly regulating epigenetic modifiers. The usage of these intermediates as substrates and regulators of chromatin-modifying enzymes provides a direct link between the metabolic state of the cell and epigenetics [10–17]. However, it remains intriguing how aging-related changes in cellular metabolism (e.g., loss of NAD homeostasis [18–20]) might control the layers of epigenetic instructions that influence cell fate without involving changes in the DNA sequence.
The capacity of the chromatin structure to affect cellular identity and cellular state transitions can differ as a function of metabolic conditions that change during aging. However, the possibility that cellular aging might result from the stochastic translation of metabolic signals into cellular epigenetic states has not been formally evaluated.
In this paper, we explore the causative relationship between cofactor (e.g. metabolite) variability and chromatin modification state underpinning the aging-associated loss of epigenetic resilience, which leads to a gain of more plastic cell and tissue features. This fact might predispose aging tissues to cancer [21, 22]. To this end, we generated an ensemble of epigenetic regulation (ER) systems by means of Approximate Bayesian Computation (ABC) whose heterogeneity reflects the inhomogeneous abundance of cofactors used by epigenetic modifiers. By analysing the robustness of ER systems in response to the regulation of HDM and HDAC activity, we present a model of ER capable of formulating strategies aimed at modifying the aging process and the aging-dependency of cancer, based on the control of epigenetic resilience and plasticity.
Recent advances in experimental determination of the mechanisms of ER have triggered an interest in developing mathematical models capable of reducing their intrinsic complexity to essential components such as ER of gene expression [17, 23–27] and epigenetic memory [24, 25, 27–32]. For comprehensive reviews, we refer the readers to [25, 27]. In order to put our model into context, we briefly summarise the current state of the art in ER modelling.
Models of ER were originally formulated in order to shed light onto the mechanisms of epigenetic memory; since DNA during cell cycle is duplicated and, therefore, the epigenetic marks diluted, early ER models were aimed at explaining how epigenetic-regulatory states remain stable upon cell division and transmitted to daughter cells. Such models must satisfy two essential properties, namely, they must be bistable, i.e., each steady state corresponding to an alternative epigenetic state, and the basin of attraction of such states must allow that large perturbations of the ER systems undergoing DNA replication should not change the epigenetic state thus allowing mitotic heritability [29].
Dodd et al. [28] developed the first of such ER models. The authors considered a region of DNA consisting of N nucleosomes, each assumed to be in either of three states, namely unmodified (U), methylated (M), and acetylated (A). Because modifying and de-modifying enzymes carry out nucleosome modifications and removal of marks, a crucial ingredient of the model by Dodd et al. [28] is that histone-modifying enzymes are recruited by modified nucleosomes, thereby providing the necessary positive feed-back for the system to be bistable. However, recruitment based on next-neighbours interactions is not enough to produce robust bistability. Long-range correlations are necessary.
The model by Dodd et al. [28] has been modified and extended in several ways [31]. Sneppen and Dodd have successfully applied the same ideas [32] to modelling the patterns of epigenetic regulation in CpG islands [33]. Another interesting feature of the model developed by Sneppen and Dodd [31] is that medium-length correlations are provided by the size of nucleosomes, which allows relaxing the requirement for recruited demethylation. Angel et al. [30] have proposed an ER model to explain quantitative epigenetic control associated with the phenomenon of vernalisation, i.e. the perception and epigenetic memory of a period of cold temperatures to initiate flowering later. This model is capable of reproducing both the patterns of flowering locus C (FLC) and the quantitative dependence with respect to the duration of the exposition to low temperatures.
Besides the issue of maintaining stable epigenetic memory, recent efforts have been dedicated to the study of the regulation of epigenetic modifications by transcription factors [23, 26]. Based on the experimental observation that transcription factors (TFs) can recruit histone-modifying enzymes, Sneppen et al. [23] proposed a model where transcription factors are coupled to ER. A similar approach, although with rather significant differences, has been recently proposed by Berry et al. [26]. An essential feature of this model is the proposed feedback between transcription and epigenetic chromatin modification: activation of transcription depends on the balance between positive and negative modifications, and, in turn, each passage of RNA polymerase II, which is modelled as a discrete event, causes demethylation (see [26] for details). An important feature that distinguishes this model from its predecessors is the assumption of next-neighbour recruitment as exclusively opposed to long-distance recruitment.
Bintu et al. [24] have recently proposed a more phenomenological ER model capable of explaining experimental data obtained by using a reporter gene that expresses a fluorescent protein with induced recruitment of a number of epigenetic-modifying enzymes. The model by Bintu et al. [24] considers active, reversible silent, and irreversible silent states and is able to predict the rates of transition between states.
In this Section, we provide an account of our stochastic model of epigenetic regulation of gene expression which extends our previous work [17]. Our model belongs to a family of models which consider that single unmodified (U) loci can be modified so as to acquire positive (A) or negative (M) marks. A positive feedback mechanism is introduced whereby M marks help to both add more M marks and remove A marks from neighbouring loci. The positive marks are assumed to be under the effects of a similar positive reinforcement mechanism [27, 28].
The stochastic model of epigenetic regulation is formulated in terms of the associated Chemical Master Equation (CME), which, in general, is given by:
∂ P ( X , t ) ∂ t = ∑ i ( W i ( X - r i ) P ( X - r i , t ) - W i ( X ) P ( X , t ) ) (1)
where X = (X1, …, Xn) is the vector containing the number of molecules of each molecular species at time t, Wi(X) is the transition rate corresponding to reaction channel i and ri is a vector whose entries denote the change in the number of molecules of each molecular species when reaction channel i fires up, i.e. P(X(t + Δt) = X(t) + ri|X(t)) = Wi(X)Δt. Our model (see Table 1) is based on the stochastic models by Dodd et al. [28] and Menéndez et al. [34].
Dodd et al. [28] consider that direct transitions between M and A are very unlikely. Instead, they assume that transitions occur in a linear sequence given by M ⇌ U ⇌ A. They further put forward the hypothesis that such nucleosome modifications are of two types, namely, recruited and unrecruited. Mathematically, recruited modifications are represented by non-linear dependence on the number of M-nucleosomes and A-nucleosomes of the corresponding transition rates (see Table 1).
Specifically, the reactions involved in our model are:
All these reactions can be both recruited or unrecruited. The associated reactions rates are reported in Table 1.
We consider the scenario where both hyper-(hypo-)abundance of A (M) marks allows for genes to be expressed, insofar the associated transcription factors are present [10]. On the contrary, we associate hypo-(hyper-)abundance of A (M) marks with silent states where genes are not expressed even in the presence of the appropriate transcription factors. We here focus on the conditions for bistability to arise and the robustness of the associated open and closed states particularly in connection with the abundance or activity of HDMs and HDACs. Our aim is to analyse the effects of varying the concentration of these enzymes as well as possible synergies between them.
In more detail, we focus our analysis on plastic behaviour of the epigenetic regulatory states when the activity of histone-modifying enzymes (HMEs) is down-regulated against the background of heterogeneity due to variability in the pool of cofactors for chromatin-modifying enzymes. We proceed by first defining a base-line scenario (which we categorise as normal cell) in which the associated epigenetic regulatory system is such that, for average values of HDM and HDAC activities, the differentiation-promoting gene ER is open and the pluripotency-promoting gene ER is closed. We then proceed to generate an ensemble of ER systems that satisfy the requirements imposed by this base-line scenario; the necessary variability to generate this ensemble is provided by heterogeneity in abundance of epigenetic cofactors. Analysis of this ensemble reveals that the requirements of the base line scenario restrict the values of a few parameters only, leaving ample flexibility to fix the rest of them. This behaviour is typical of the so-called sloppy models [35], where available data constrains a limited number of parameters (or parameter combinations), the system being robust to the choice of a large number of model parameters. In our case, this feature is absolutely essential since, nested within this heterogeneous ensemble of ER systems, there exists a sub-ensemble of plastic ER systems.
In order to gain some insight into the behaviour of the stochastic ER model, we analyse its mean-field limit regarding time scale separation and the quasi-steady state approximation. For a full account of the technicalities we refer the reader to our previous work [36, 37].
The mean-field equations, which describe the time evolution of the ensemble average of the variables Xi, associated to the stochastic system with rates given in Table 1 are:
d Q i d t = ∑ j = 1 16 r j , i W j ( Q ) (2)
where Q is a vector whose entries, Qi, are Qi ≡ 〈Xi〉. In order to proceed further, we assume that the variables describing the system are divided into two groups according to their characteristic scales. More specifically, we consider the situation where the subset of chemical species Xi, with i = 1, 2, 3, scale as Xi = Sxi, where xi = O(1), whilst the remaining species are such that Xi, with i = 4, 5, 6, 7, scale as Xi = Exi, where xi = O(1). Key to our approach is the further assumption that S and E must be such that ϵ = E S ⪡ 1. The averaged variables, Qi, are similarly divided into two groups: slow variables, i.e. Qi = Sqi (i = 1, 2, 3), and fast variables, i.e. Qi = Eqi (i = 4, 5, 6, 7).
Under this rescaling, we define the following scale transformation for the transition rates in Table 1: Wj(Q) = k4S2Eωj(q). We further rescale the time variable so that a dimensionless variable, τ, is defined as τ = k4SEt. It is now straightforward to verify that, upon rescaling, the mean-field equations become:
d q i d τ = ∑ j = 1 16 r j , i ω j ( q ) , i = 1 , 2 , 3 , (3) ϵ d q i d τ = ∑ j = 1 16 r j , i ω j ( q ) , i = 4 , 5 , 6 , 7 . (4)
with ϵ = E/S.
If ϵ = E/S ≪ 1 holds, Eqs (3) and (4) naturally display multiple scales structure, which we will exploit to simplify our analysis by means of a quasi-steady state approximation (QSSA) [38], which is given by:
d q 1 d τ = e H D M ( κ 1 + q 3 ) ( κ 3 + κ 6 q 3 ) q 2 ( κ 2 + κ 3 ) + ( κ 1 + q 3 ) q 2 + ( κ 5 + κ 6 ) q 3 + e H D A C ( κ 9 + κ 12 q 2 ) ( κ 11 + κ 14 q 2 ) q 3 ( κ 10 + κ 11 ) + ( κ 9 + κ 12 q 2 ) q 3 + ( κ 13 + κ 14 ) q 2 - ( κ 8 q 2 + κ 7 + κ 16 q 3 + κ 15 ) q 1 (5)
d q 2 d τ = - e H D M ( κ 1 + q 3 ) ( κ 3 + κ 6 q 3 ) q 2 ( κ 2 + κ 3 ) + ( κ 1 + q 3 ) q 2 + ( κ 5 + κ 6 ) q 3 + ( κ 8 q 2 + κ 7 ) q 1 (6)
d q 3 d τ = - e H D A C ( κ 9 + κ 12 q 2 ) ( κ 11 + κ 14 q 2 ) q 3 ( κ 10 + κ 11 ) + ( κ 9 + κ 12 q 2 ) q 3 + ( κ 13 + κ 14 ) q 2 + ( κ 16 q 3 + κ 15 ) q 1 (7)
q 4 = e H D M κ 2 + κ 3 + ( κ 5 + κ 6 ) q 3 ( κ 2 + κ 3 ) + ( κ 1 + q 3 ) q 2 + ( κ 5 + κ 6 ) q 3 (8)
q 5 = e H D M ( κ 1 + q 3 ) q 2 ( κ 2 + κ 3 ) + ( κ 1 + q 3 ) q 2 + ( κ 5 + κ 6 ) q 3 (9)
q 6 = e H D A C κ 10 + κ 11 + ( κ 13 + κ 14 ) q 2 ( κ 10 + κ 11 ) + ( κ 9 + κ 12 q 2 ) q 3 + ( κ 13 + κ 14 ) q 2 (10)
q 7 = e H D A C ( κ 9 + κ 12 q 2 ) q 3 ( κ 10 + κ 11 ) + ( κ 9 + κ 12 q 2 ) q 3 + ( κ 13 + κ 14 ) q 2 (11)
where the re-scaled parameters κj are defined in Table 2, and the conservation laws q4(τ) + q5(τ) = eHDM and q6(τ) + q7(τ) = eHDAC hold. These conservation laws account for the fact that the total number of enzyme molecules, i.e. the enzyme molecules in their free form and those forming a complex must be constant. Hence, the quantities eHDM and eHDAC are defined as e H D M = z 0 E and e H D A C = v 0 E, respectively, where z0 and v0 are the numbers of HDM and HDAC enzyme molecules, respectively. E is the characteristic scale (i.e. average) of abundance of the histone-modifying enzymes which, for simplicity, has been taken to have the same value for both HDMs and HDACs. This result opens interesting avenues to investigate, since both oncometabolic transformation and aging appear to reduce the number of both types of enzymes. Our theory thus allows us in a natural manner to explore the effects of these anomalies on the stability of epigenetic regulatory states.
We now proceed to explore the behaviour of our system as the number of HDMs and HDACs vary relative to their average abundance against the background of variability provided by our ABC-ensemble approach.
We first focus on a bifurcation analysis of the mean-field QSSA Eqs (5)–(11), to investigate the qualitative behaviour of the ER system as the relative abundances of HDMs and HDACs are varied. Results are shown in Fig 3(a) and 3(b). In particular, the phase space of both ER systems obtained by varying the parameters eHDM and eHDAC. Both these diagrams display three differentiated regions: one in which the only stable steady-state is the one associated with a silenced gene, another one in which the only stable steady-state is the corresponding to an open gene, and a third one where the system is bistable. Fig 3(a) is associated with the differentiation-promoting gene, and Fig 3(b) corresponds to the pluripotency-promoting gene (parameters as per Table A, Table B in S1 File, respectively). In order to clarify the three regions (open, closed and bistable) displayed in Fig 3(a), a 3D plot is shown in Fig 4(a), where the vertical axis shows the level of positive marks (q3). This plot shows that the system dysplays bistable behaviour: depending on the parameter values eHDM and eHDAC, the system may be both in the open state (high levels of q3, top of the plot), or in the closed state. Fig 4(b) displays the projection on the xy-plane of the plot shown in Fig 4(a), where we can clearly identify the three regions described in Fig 3(a).
A more detailed picture of the situation illustrated in Figs 3(a) and 4 is given in Fig 3(c), which shows the bifurcation diagram where eHDM, i.e. HDM concentration, is taken as the control parameter, whilst keeping eHDAC constant. In particular we show the steady state value of q3, i.e. the variable with positive marks, as a function of HDM concentration. This allows to distinguish the three regions displayed in Fig 3(a). We observe, that a decrease in HDM makes the corresponding gene inaccessible to the transcription machinery (corresponding to the closed region, Fig 3(a)). As HDM concentration recovers, the system enters a bistable regime where both the active and silent states coexist (region marked as bistable in Fig 3(a)). Further increase of the demethylase concentration drives the system through a saddle-node bifurcation, beyond which the only stable steady-state is the active state (region labelled as open in Fig 3(a)). It is noteworthy that these results are in agreement with the oncometabolic transformation scenario associated with IDH mutations proposed by Thompson and co-workers [10, 42] in which downregulation of HDM activity locks differentiation genes into a silenced state which favours reprogramming of the differentiated state of somatic cells into a pluripotent phenotype [17]. The association between IDH mutations and cancer progression has been well established in the case of glioblastomas and acute myelogenous leukaemia [43–46].
In Fig 3(e), we show the bifurcation diagram associated with fixing eHDM and varying eHDAC. Within the scenario we are considering, i.e. the epigenetic regulation of a differentiation-regulating gene, reduced HDAC concentration recovers the base-line state where the epigenetic regulatory machinery is set to the open state. As HDAC concentration recovers, the system enters a bistable regime in which both the active and silent states coexist. Further increase in HDAC activity locks the system into the close chromatin state so that the gene is silenced. This implies that reduced HDAC activity may help to rescue differentiation-regulating genes from the effects of IDH mutation.
Numerical results which verify the predictions of the bifurcation analysis are presented and discussed in Section I in S1 File.
We now proceed to analyse in more detail the implications of the bifurcation analysis, regarding robustness of the epigenetic regulatory state. In Fig 3(d), which shows the phase diagram of both modes of epigenetic regulation (differentiation- and pluripotency-promoting) in the same phase space, the region between the solid red line and the dashed blue line represents the part of the phase space where the differentiation genes are open and the pluripotency genes are closed (region marked as Normal Cell in Fig 3(d)). This sub-space is therefore associated with normal, differentiated somatic cells. As we have previously shown [17], efficient reprogramming requires both closed differentiation genes and open pluripotency genes. Such situation is not viable under the scenario shown in Fig 3(d) because these two conditions cannot hold simultaneously, which we therefore dubb as the refractory scenario.
By contrast, Fig 3(f) corresponds to a plastic scenario, where, under appropriate conditions, cells become poised for reprogramming. The main difference with the refractory scenario is the intersection between the bistability regions of both the differentiation regulator and the pluripotency gene. In Fig 3(f), the regime where both bistability regions overlap is the one between the red solid line and the blue dashed line (region marked as Rep. in Fig 3(f)). Within this region, since both genes are in the bistable epigenetic regulatory regime, it is possible to find the differentiation gene in its closed state and the pluripotency gene in the open state. Such situation makes reprogramming much more likely to occur [17] and therefore we identify this feature of the phase space with plastic behaviour. By driving the ER system into this region by means of down-regulation of both HDM and HDAC activity, cells become epigenetically poised to undergo reprogramming. This is consistent with evidence according to which both oncometabolic transformation (e.g. IDH mutation leading to down-regulation of JHDM activity [10, 42]) and aging (e.g. down-regulation of SIRT6 [5, 19, 47]) induce loss of HDM and HDAC activity thus facilitating reprogramming.
In order to study the robustness of the refractory and plastic scenarios with respect to variations of the model parameters, kj (see Table 1), we first generate an ensemble of parameter sets θ = (kj, j = 1, …, 16) compatible with simulated data for the epigenetic regulation systems. Such ensemble is generated using Approximate Bayesian Computation [48] (for further details see Section III in S1 File). Our approach is as follows. For each mode of epigenetic regulation, we have generated simulated data (denoted as “raw data” in Fig 2) using the stochastic simulation algorithm on the model defined by the transition rates Table 1. This simulated data will play the role of the experimental data, x0, to which we wish to fit our model. We consider two different data sets x 0 d and x 0 p, corresponding to the differentiation gene (reaction rates from Table A in S1 File) and the pluripotency gene (reaction rates from Table B in S1 File), respectively. Each data set consists of 10 realisations and 25 time points per realisation. For each time point, ti, we consider two summary statistics: the mean over realisations, x ¯ ( t i ), and the associated standard deviation, σ(ti). We then run the ABC rejection sampler method until we reach an ensemble of 10000 parameter sets which fit the simulated data, x0, within the prescribed tolerances for the mean and standard deviation. Fig 2(a) & 2(b) shows results comparing the reference (raw simulated) data to a sub-ensemble average (full posterior distributions are shown in Fig. C in S1 File, differentiation-promoting gene, and Fig. D in S1 File, pluripotency-promoting gene).
The above procedure provides us with an ensemble of parameter sets that are compatible with our raw data, i.e. such that they fit the data within the prescribed tolerances. The heterogeneity associated with the variability within this ensemble has a clear biological origin. The rates kj are associated with the activity of the different enzymes that carry out the epigenetic-regulatory modifications (HDMs, HDACs, as well as, histone methylases (HMs) and histone acetylases (HACs)), so that variation in these parameters can be traced back to heterogeneity in the availability of cofactors, many of them of metabolic origin such as NAD+, which are necessary for these enzymes to perform their function (as illustrated in Fig 1).
We first consider the differentiation ER system. In particular, we focus on the sub-ensemble of the 400 parameter sets that best fit the raw data. Within such sub-ensemble, we proceed to evaluate the robustness of the different scenarios we study. We consider that a particular scenario is sensitive to a specific parameter, kj, if its distribution is significantly different from the uniform distribution [49]. We first analyse the base-line scenario for the epigenetic regulation of a differentiation-regulated gene, namely, (i) when eHDM = eHDAC = 1, the regulatory system is mono-stable (only the open chromatin state is stable), and (ii) for eHDM < 1, eHDAC < 1 there exists a region of bistability. Out of all the parameter sets of the considered sub-ensemble, only 94 fulfill these requirements. We refer to these as the viable set. The remaining 307 are bistable at eHDM = eHDAC = 1, and they will be referred to as the non-viable set. In Fig 5, we present the cumulative frequency distributions (CFDs) of each kj within both sets. The rationale for looking into this is that the requirements upon system behaviour associated with both sets should reflect themselves on the corresponding CFDs.
Regarding the viable set, we seek to assess which kinetic constants have distributions which deviate in a statistically significant manner from the uniform distribution [49]. Such parameters are deemed to be the essential ones for the ER system to exhibit the behaviour associated with the viable set. We perform this analysis by means of the Kolmogorov-Smirnov (KS) test [50, 51], which we use to compare our samples with the uniform distribution. According to such analysis, the kinetic constants k1, k3, k6, k7, k12, k14, and k16 are not uniformly distributed (p-values are reported in Table E in S1 File).
Nested within the viable set, there are parameter sets which exhibit plastic behaviour, as characterised by a phase diagram as per Fig 3(f). We thus continue by studying the plastic subset regarding both its frequency within the viable subset and further restrictions imposed on parameter variability. We first check the number of the plastic parameter sets within the viable set relative to the pluripotency-gene ER system defined by Table D in S1 File. Somehow unexpectedly, the plastic scenario is rare, but not exceptional: amongst the 94 parameter sets that we have identified as viable, 10 exhibit plasticity (see Fig 5 for their CFDs).
Further restrictions on parametric heterogeneity imposed by the plastic scenario are analysed regarding the variation of the CFDs of kinetic constants when compared to those associated with the whole viable subset. The results of KS analysis performed on the data shown in Fig 5 show that only the distributions of k1 (associated with recruited demethylation), k9 (unrecruited deacetylation), and k14 (recruited deacetylation) are significantly modified by the plasticity requirement (p-values reported in Table G in S1 File).
From a more mechanistic perspective, we observe that, within the plastic set, the mass of the CFDs of k1, k9 and k14 is displaced towards the large-value end of their intervals with respect to their behaviour within the full viable set. In other words, k1, k9 and k14 tend to be larger for plastic ER systems than for non-plastic, viable ER systems. In essence, we observe that ER systems exhibiting plastic behaviour tend to have increased activity in the enzymes performing histone deacetylation. This is consistent with recent evidence that aging decreases histone acetylation and promotes reprograming [5, 19, 47].
The same analysis has been conducted regarding the ensemble of parameter values generated using ABC for the pluripotency gene ER system (full posterior distribution in Fig. D in S1 File). The results of this analysis are shown in Fig 6. Detailed analysis using the KS test of the ensemble viable pluripotency ER systems shows that k3, k8, k12, k14, k15, and k16 are significantly constrained by the requirements of such scenario (i.e. their CDF departs significantly from the uniform distribution, as shown by the p-values from Table F in S1 File). We then move on to investigate further restrictions within the plastic set. We observe that only the CDFs associated with k2 and k6 are significantly different (p-values reported in Table H in S1 File). In both cases, values of k2 and k6 associated with plasticity are larger than in the general viable population. Both parameters are associated with demethylation activity.
Our ensemble analysis thus provides a rationale for the coupling between variations in the size of the pool of epigenetic cofactors and increased reprogramming in a heterogeneous cell population. A notable case in point is provided by metabolic changes during aging: those cells where key metabolites such as acetyl-CoA and NAD+ are less abundant lose acetylation capability (in our model, this is reflected through the dependence of histone-modifying enzyme activity on the concentration of these cofactors), leading to cells poised for reprogramming.
This analysis provides a rationale for a strategy to interfere with the epigenetic regulatory system, regarding the ability to either drive the system away from plastic behaviour or to drive it to the plasticity scenario, while keeping it functional (i.e. within the restrictions of the base-line scenario). An example illustrating the effectiveness of this strategy is shown in Fig 7. Consider the viable set of the ER differentiation-promoting gene, Fig 5, which is neutral with respect to the value of k9: k9 remains uniformly distributed within the viable subset. By contrast, when plasticity is required, the admissible values of k9 accumulate mostly towards the large-value end. This suggests that decreasing the value of k9 might be a viable strategy to restore resilience. To check this, we consider the parameter set, θ = kj/k4, j = 1, …, 16, that gives rise to the plastic behaviour depicted in Fig 3(f) (Table C in S1 File, for the differentiation-promoting gene). We then analyse the effect of modifying the value of k9 for the differentiation-promoting gene on system behaviour. The new parameter set, θ ′ = k j ′ / k 4 , j = 1 , … , 16, is such that k 9 ′ = k 9 / 4 and k j ′ = k j for all j ≠ 9 (kj values as per Table C in S1 File). Parameter values for the pluripotency gene remain unchanged (as per Table D in S1 File). The corresponding phase space is shown in Fig 7(a). We observe that by reducing deacetylase activity in this fashion, the ER system reverts to resilient behaviour. This suggests that, by regulating the abundance of cofactors associated with (de)acetylation, we can drive the system off the plastic regime into the base-line behaviour.
Similarly, we can seek for complex, combined strategies to increase the robustness of plastic behaviour. An example of such strategy is shown in Fig 7(b). Based on the results of the KS test for the differentiation-promoting gene, we observe that deacetylation-related rates k9 and k14 are significantly increased in plastic scenarios. Taking parameter sets from a resilient scenario (Tables A & D in S1 File, which lead to a combined phase diagram qualititatively similar to that shown in Fig 3(d)) and modifying k9 and k14 for the differentiation-promoting gene so that k 9 ′ = 3 k 9 and k 14 ′ = 3 k 14 while keeping all the others at the same value, the resulting ER system corresponds to a plastic system. Futhermore, this combined strategy results in more robust plasticity (as compared to e.g. the case shown in Fig 3(f)), as measured by the area of the phase space region where reprogramming is feasible. This indicates that by combining the strategies suggested by the statistical analysis of the plastic sub-ensemble, we can find conditions for optimal conditions to achieve robust reprogramming. This, in turn, highlights the importance of cofactor levels, since as it has been shown in Fig 7, depending on its availability, the same ER system can be driven to the plastic or resilient state.
These strategies require close attention to be payed to the correlations between parameters. Parameters in complex systems biology models exhibit strong correlations which confer the system with essential properties such as sloppiness, which refers to the property exhibited by many multi-parameter systems biology models, whereby the system’s behaviour is insensitive to changes in parameter values except along a small number of parameter combinations [35]. In order to quantify such correlations, we have used hierarchical clustering. The results are shown in Fig. E(a) & E(b) in S1 File for the base-line and the plastic scenarios of the differentiation-regulating ER system, respectively. Not unexpectedly, we observe that, with respect to the base-line scenario, correlations substantially change when the plastic scenario is considered. Although the strategies illustrated in the results shown in Fig 7 changed one or two parameters alone independently of all the others, more general situations will require to closely monitor these correlations to understand which combinations of parameters are relevant to control the system’s behaviour [35].
We here provide computational evidence for the role of stochastic translation of epigenetic cofactors into resilient/plastic cell states via ER systems as a mechanistic facilitator of cellular aging, and its reversal. When changes in levels of such cofactors operate as regulators of the kinetic parameters associated with chromatin-modifying enzymes such as HDMs and HDACs, the ensemble of ER configurations reveals the occurrence of cell-to-cell phenotypic variability in terms of different epi-states (see Fig 8). This model provides a rationale for the responsiveness of cellular phenotypes to metabolic signals, as metabolic pools serve as epigenetic cofactors. The metabolic control of epigenetic landscapes and cell state transitions might therefore operate as a common hub capable of facilitating the pathogenesis of aging-related diseases including cancer.
Several layers of molecular communication exist between cell metabolism and chromatin remodelling [16, 52–56]. A first layer of metabolo-epigenetic regulation includes metabolites/nutrient-responsive TF-dependent transcriptional regulation of chromatin regulators (HMT, HAT, DNMTS, etc), which can lead to global changes on chromatin structure. Second, metabolites can modulate chromatin modifications at specific genomic loci by affecting the activity/localisation of proteins that recruit or regulate chromatin-modifying enzymes during, for example, transcriptional activation phenomena. Third, chromatin-modifying enzymes employ many metabolites as donor substrates and cofactors, and changes in levels of these bona fide epigenetic metabolites can in turn lead to changes not only in the global status of chromatin modifications but also to gene specific regulation under different metabolic conditions.
Our mathematical model only incorporates the third such layer through cofactor-induced heterogeneity. Because any metabolic input has the potential to affect various chromatin marks via its effects on transcription, our model ignored metabolic regulation of TF activity. In contrast to other metabolically-regulated enzymatic activities such as phosphorylation in which the substrate (ATP) is present in cellular concentrations far greater than the enzyme Km values, i.e., the concentration of metabolite at half maximum velocity of enzyme-mediated reaction, the physiological cellular concentrations of donors and cofactors that are employed by histone-modifying enzymes (e.g., organic ketoacids such as the demethylase cofactor α-ketoglutarate for HDMs or the NAD+ deacetylase cofactor for HDACs) are close to HDM and HDAC Km values [16, 57]; consequently, based solely on the intrinsic biochemical characteristics of chromatin-modifying enzymes such as HDMs and HDACs, small fluctuations in the concentrations of such metabolites could significantly alter HDM and HDAC activities, either increasing or decreasing their respective histone-modifying activities. This layer of metabolo-epigenetic regulation is commonly viewed as a direct link from cell metabolism to chromatin-modification status, which could be mathematically modelled and tested as has been confirmed in our current computational model (see Fig 8).
Evidence accumulates demonstrating that differing metabolomes can be found in distinct cell states, thereby suggesting how changes in metabolism can impact and probably specify cell fate via alteration of the chromatin landscape [58–63]. Yet, there is a scarcity of examples showing that metabolic changes can restructure the epigenetic landscape and lead to different cell states regardless of other global changes in cell physiology occurring in response to this variation in metabolite levels. Our findings support the notion that changes in the abundances of certain metabolites would alter specific chromatin marks, thereby determining both the stability of cell types and the probability of transitioning from one epi-state to another [64]. Our model infers that such a change in metabolite level would be sufficient to either impede or allow cell epi-state transitions by regulating the height of the phenotypic barriers in the context of Waddington’s landscape (Fig 8). However, we should acknowledge that the necessary involvement of cellular metabolism on the structure of the epigenetic landscape will require the experimental coupling of defined metabolic conditions with epigenome editing systems (e.g., CRISPR-Cas9) capable of targeting specific histone PTMs playing important roles in chromatin structure [65].
Our ensemble approach provides mechanistic support to the notion that emergence of the cellular and molecular hallmarks of aging including cancer might result from a metabolically driven loss of epigenetic resilience. Flavahan et al. [57] have recently proposed that non-genetic stimuli including aging and metabolic insults can induce either overly restrictive chromatin states, which can block tumor-suppression and/or differentiation programs, or overly permissive/plastic chromatin states, which might allow normal and cancer cells to stochastically activate oncogenic programs and/or nonphysiologic cell fate transitions. Our ensemble approach provides a framework that supports heterogeneity of epigenetic states as an engine that facilitates cancer hallmarks and other aging diseases. On the one hand, the ability of resilient states to maintain large epigenetic barriers refractory to non-physiologic cell fate transitions might explain why the NAD+-dependent HDAC/sirtuin pathway is one of the few mechanisms described to mediate the correction or resetting of the abnormal chromatin state of aging cells induced by calorie restriction, the most robust life span-extending and cancer preventing regimen [2, 66–68]. On the other hand, the ability of plastic states to lower epigenetic barriers, and increase the sensitivity of primed cells to undergo reprogramming-like events leading to loss of cell identity is consistent with the ability of certain metabolites to promote oncogenesis by epigenetically blocking the HDM-regulated acquisition of differentiation markers [17, 69–71].
The traditional view of cancer formation (i.e., the Knudson model [72]) exclusively involves the binary acquisition and accumulation of genetic alterations as the principal driver mechanism for the age-dependency of multistage cancer development. Our ensemble approach suggests an alternative, namely, that oncogenic chromatin aberrations might also occur via purely epigenetic stimuli. Our model shows that, nested within the ensemble of ER systems, those that prime cells for reprogramming exhibit properties associated with age-induced epigenetic dis-regulation [73, 74]. Aging-responsive ER reprogramming might thus operate in a more progressive and graded manner to increase cancer susceptibility without the need to induce genetic mutations. Our ensemble model is mechanistically consistent with the fact that those cancers in which the sole presence of epigenetic metabolites (e.g., oncometabolites) suffices to stabilise undifferentiated cellular states by preventing demethylation of genes implicated in differentiation have accelerated models of oncogenesis [44, 75–82]. Whereas the epigenetic signature of adult somatic cells must be partially and acutely erased to adopt a more plastic epigenome, such cellular plasticity, which might occur via metabolically driven epigenetic activation of promoter regions of pluripotency genes, could impose a chronic, locked gain of stem cell-like states disabled for reparative differentiation.
The existence of metabolism-permissive resilient and plastic epigenetic landscapes might have predictive power on the susceptibility of a cell to lose its normal cellular identity through reprogramming-like resetting phenomena. The beneficial or deleterious decision paths during the maintenance of cell and tissue homeostasis might be closely related to the ability of epigenetic landscapes to modulate the intrinsic responsiveness to reprogramming cellular identity. The incapability of finishing cellular reprogramming, or at least to increase cellular epigenetic plasticity, might impede tissue self-repair in response to injury, stress, and disease, thus driving the observed aging phenotypes. Accordingly, the infliction of chronic injury and the aging phenotype have been shown to render tissues highly permissive to in vivo reprogramming [47] while the cyclic, transient expression of reprogramming factors has recently been shown to increase lifespan in a murine model of premature aging via remodeling of the chromatin landscape [7]. Because our model suggests that the fine-tuning of metabolic epigenetic cofactors might direct plastic epigenetic states to re-enter into epigenetic resilience, and vice versa, it would be relevant to experimentally evaluate whether specific metabolic interventions might either mimic transient reprogramming and revert some age-associated features without promoting complete undifferentiation, or prevent the occurrence of unrestricted/uncontrolled plasticity in chronically injured tissues such as those occurring in aging and cancer.
In summary, by integrating the ability of chromatin epigenetic modifiers to function as sensors of cellular metabolism, our ensemble model provides computational support to the notion that a metabolism-responsive loss of epigenetic resilience might mechanistically facilitate cellular aging. The stochastic translation of metabolic signals into resilient/plastic cell states via ER systems might be viewed as a metabolo-epigenetic dimension that not only facilitates cellular aging, but that also offers new therapeutic and behavioural avenues for its reversal. Our findings strongly suggest that the development of predictive mathematical models and computational simulation platforms capable of operatively integrate the metabolic control of epigenetic resilience and plasticity and its combination with confirmatory lab-based testing might accelerate the discovery of new strategies for metabolically correcting the aberrant chromatin structure that affects cellular identity and epi-state transitions in aging and aging-related diseases.
|
10.1371/journal.pcbi.1002427 | A Common Model for Cytokine Receptor Activation: Combined Scissor-Like Rotation and Self-Rotation of Receptor Dimer Induced by Class I Cytokine | The precise mechanism by which the binding of a class I cytokine to the extracellular domain of its corresponding receptor transmits a signal through the cell membrane remains unclear. Receptor activation involves a cytokine-receptor complex with a 1∶2 stoichiometry. Previously we used our transient-complex theory to calculate the rate constant of the initial cytokine-receptor binding to form a 1∶1 complex. Here we computed the binding pathway leading to the 1∶2 activation complex. Three cytokine systems (growth hormone, erythropoietin, and prolactin) were studied, and the focus was on the binding of the extracellular domain of the second receptor molecule after forming the 1∶1 complex. According to the transient-complex theory, translational and rotation diffusion of the binding entities bring them together to form a transient complex, which has near-native relative separation and orientation but not the short-range specific native interactions. Subsequently conformational rearrangement leads to the formation of the native complex. We found that the changes in relative orientations between the two receptor molecules from the transient complex to the 1∶2 native complex are similar for the three cytokine-receptor systems. We thus propose a common model for receptor activation by class I cytokines, involving combined scissor-like rotation and self-rotation of the two receptor molecules. Both types of rotations seem essential: the scissor-like rotation separates the intracellular domains of the two receptor molecules to make room for the associated Janus kinase molecules, while the self-rotation allows them to orient properly for transphosphorylation. This activation model explains a host of experimental observations. The transient-complex based approach presented here may provide a strategy for designing antagonists and prove useful for elucidating activation mechanisms of other receptors.
| Class I cytokines activate their receptors via a 1∶2 complex, but the conformational rearrangements leading to receptor activation remain unclear. To elucidate the activation mechanism, here we calculated the transient complex, an on-pathway intermediate close to the 1∶2 complex. Similar rotational motions were found for three cytokine (growth hormone, erythropoietin, and prolactin) receptors on going from the transient complex to the 1∶2 complex. They involve both scissor-like rotation between the extracellular domains of two receptor molecules and self-rotation of the molecules. Based on these results, we propose a common model for receptor activation by class I cytokines. The model explains a number of experimental observations, including differences in receptor orientations between erythropoietin and its antagonistic and partially agonistic mimetics. Transient complexes present a novel type of targets for designing antagonists. The detailed activation model developed here and our transient-complex based approach will be useful for studying the activation mechanisms of other receptors.
| Cytokines are a large family of small proteins that bind to specific cell surface receptors to initiate signals critical for cell proliferation, differentiation, and apoptosis. Among the best characterized cytokines are class I helical cytokines, including growth hormone (GH), erythropoietin (EPO), and prolactin (PRL). Each of these cytokines has two receptor binding sites, referred to as site 1 and site 2, with high and low affinities, respectively. Each cytokine receptor consists of an extracellular domain (ECD) and an intracellular domain (ICD), connected by a single transmembrane helix (TMH). The ECD in turn is composed of two β-sandwich subdomains linked by a short hinge [1]. It is well known that the binding of two receptor molecules, to site 1 and site 2 on the cytokine, results in receptor activation, leading to transphosphorylation of two Janus kinase 2 (JAK2) molecules, each associated with a receptor ICD at a proline-rich region (box 1). Once phosphorylated, the JAK2 molecules initiate downstream signaling [2]–[5].
The structures of the 1∶2 complexes of GH, EPO, and PRL with the ECDs of the corresponding receptors have been determined [1], [6], [7] (Figure 1). The structures are overall similar, but differ in many details. Each cytokine contacts both ECD subdomains of each receptor molecule around the hinge. The two C-terminal subdomains are nearly parallel to each other (and presumably to the normal of the cell membrane), while the two N-terminal domains lie on a plane parallel to the membrane, at 130°–160° angles. These structures have been very valuable, but they do not reveal the rearrangement of the two ECDs induced by the cytokine binding. Since the structures lack the TMHs and the ICDs, there is also no information on the ICDs' rearrangement, which initiates downstream signaling. The aim of the present study is to compute the cytokine-induced rearrangement of the ECDs and develop a detailed model for receptor activation.
In the early model proposed by Fuh et al. [8] for GH receptor activation, GH first binds to one receptor molecule via site 1, and then recruits the second receptor molecules via site 2. This sequential receptor-dimerization model was based on three important observations. First, site 1 has much higher affinity than site 2. Second, a G120R mutation disrupting site 2 did not affect receptor binding to site 1 but abolished GH-induced cell proliferation. Third, the dose response curve of cell proliferation was bell-shaped, suggesting that engagement of each receptor molecule by a separate GH molecule (via site 1) interferes with receptor dimerization and signaling.
It is now clear that receptors likely exist as preformed dimers in the absence of the cytokines [9]–[11]. For both GH receptor (GHR) and EPO receptor (EPOR), the TMHs are implicated in dimer formation [10], [12], [13]. However, dimerization alone is insufficient for activation. For example, two EPO mimetic peptides (EMP1 and EMP33) bind to EPOR to form 1∶2 complexes, but in each of these complexes the ECDs (and their subdomains) have an orientational arrangement that is different from that in the EPO:(EPOR)2 complex [6], [14], [15]. (EMP1 and EMP33 each are present as dimers in the complexes with two EPORs. We treat these dimers as a single ligand and refer to the stoichiometry of the complexes as 1∶2.) In signaling EMP1 acted as a partial agonist but EMP33 as an antagonist. Seubert et al. [16] engineered EPOR dimers by replacing the ECDs with a dimeric coiled coil. Through deletions of up to 6 residues, they explored the full range of relative orientation of the two TMHs in the EPOR dimers, and found one of them to be constitutively active in cell proliferation.
For GHR, Rowlinson et al. [17] found monoclonal antibodies that competed against GH for GHR binding but failed to act as agonists, again indicating that dimerization is insufficient for activation. Brown et al. [10] demonstrated constitutive dimer formation of GHR by FRET experiments, and after inserting alanine residues in the TMH or in the sequence immediately before box 1, observed constitutive activity. Interestingly, constitutive activity required different numbers of inserted alanine residues in the TMH and before box 1.
The deletion and insertion results of Seubert et al. [16] and Brown et al. [10] suggest that rotation of the TMH is involved in receptor activation. However, the orientational rearrangement of the ECDs that is induced by cytokine binding and triggers the TMH rotation remains unclear.
Even in binding to a preformed dimer, it is still believed that engagement of site 1 precedes engagement of site 2 [5], [10], [18]. The initial step, i.e., the binding of a cytokine to the first receptor molecule (R1) via site 1, leads to a 1∶1 complex. The 1∶1 complex is very likely an on-pathway intermediate since the structures of the 1∶1 complexes formed by GH and GHR ECD [19], [20] and by PRL and PRL receptor (PRLR) ECD [21] are very similar to those in the corresponding 1∶2 complexes [1], [7], [22]. The 1∶1 complexes were obtained by introducing the site-2 disrupting mutation G120R to GH and a corresponding mutation, G129R, to PRL.
Recently we calculated the rate constants for forming the 1∶1 complexes of PRL, GH, and EPO [23], using our transient-complex theory [24]. These rate constants differ by 5000-fold, mostly arising from differing levels of charge complementarity across the site-1 interface. Moreover, the rate constants of the initial binding apparently anti-correlate with the circulation concentrations of the cytokines, such that the pseudo-first order receptor binding rate constants are close to the limits set by the half-lives of the receptors, ensuring their participation in cytokine binding before internalization and degradation.
The transient complex in a binding process refers to an intermediate that has near-native relative separation and orientation but not the short-range specific interactions of the native complex, and is formed by translational and rotational diffusion of the subunits. The transient complex is located at the rim of the energy well of the native complex, and is therefore a late on-pathway intermediate. Structural differences between the transient complex and the native complex reveal the orientational rearrangement of the subunits at the late stage of the binding process. This stage starts after some of the native contacts are already in proximity, but before the precise fit of all the native contacts. As such it is at a critical juncture of the binding process. Yet its characterization enjoys certain technical advantages. First, because we focus on the late stage, we completely avoid any issues concerning how the subunits reach the transient complex, such as 2-dimensional diffusion of the membrane-bound receptors. Second, because the transient complex is formed before the formation of the stereospecific native contacts, we also avoid the necessity of accurately treating the native contacts. Instead, the transient-complex ensemble is largely dictated by the shape of the binding interface.
Here we applied the transient-complex theory to study the binding of a second receptor molecule (R2) to a 1∶1 complex, to form the 1∶2 activation complex. By calculating the transient complex for this step, we identified the orientational rearrangement between the ECDs of R1 and R2 leading to receptor activation. Similar rotational motions were found for three cytokine-receptor systems (GH, EPO, and PRL with their receptors). At the start of the late-stage orientational rearrangement, R2 is loosely bound to the 1∶1 complex around site 2 of the cytokine, with the C-terminal subdomains of R1 and R2 far apart. R1 and R2 then rotate like a scissor, around an axis along the N-terminal subdomain of R2, to close up the membrane-proximal ends of the two C-terminal subdomains. In addition, R1 and R2 both self-rotate but to different extents, such that the angle between the two N-terminal subdomains is reduced. We propose that the scissor-like rotation separates the intracellular domains of the two receptor molecules to make room for the associated Janus kinase molecules, while the self-rotation allows them to orient properly for transphosphorylation. This common model for receptor activation explains a host of experimental observations on the three cytokine-receptor systems.
The focus of the present study is the late-stage orientational rearrangement between the two receptor molecules in forming the 1∶2 complex. The start of the late stage is the transient complex, in which R2 is loosely bound to the 1∶1 complex around site 2 of the cytokine. The transient complex is identified by mapping the energy landscape over the native-complex energy well and the surrounding region, using the structure of the native complex as input [24], [25]. Within the native-complex well, the rotational freedom of the subunits is severely restricted. As the two subunits separate, there is a sudden increase in the rotational freedom. The transient complex is identified with the midpoint of this transition, which is largely dictated by the shape of the binding interface.
Receptor activation occurs at cell membranes, where receptors likely exist as preformed dimers. However, the rate constants for binding to the 1∶1 complex by R2 ECD coming from the bulk solution, rather than from a preformed receptor dimer, have been measured for the GH-GHR and PRL-PRLR systems [22], [26]. Our transient-complex theory can make accurate predictions for the rate constants of protein association in bulk solution, as demonstrated by results spanning five orders of magnitude for 49 protein complexes [25]. We carried out rate constant calculations for the 1∶2 complexes of the three cytokines with the corresponding receptor ECDs. The results were within the range, 104 to 106 M−1 s−1, of the in vitro measurements (see Supporting Text S1 for details and implication for R2 binding to the 1∶1 complex in the cellular environments).
Each transient complex was an ensemble of configurations located at the rim of the native-complex energy well. It was generated from the structure of the 1∶2 complex and would be a late on-pathway intermediate, even if R2 came from a preformed receptor dimer.
As noted above, the transient complex was identified by mapping the energy landscape over the native-complex energy well and the surrounding region. The internal conformations of R2 and the 1∶1 complex (referred to as two subunits) were fixed at those in the 1∶2 native complex. This is justified since the available structures of the isolated 1∶1 complexes of the GH and PRL systems [19]–[21] are very similar to those in the corresponding 1∶2 complexes [1], [7], [22] (with Cα RMSDs of ∼1.2 Å); similarly the structures of apo GHR [10] and of apo EPOR [9] as well as EPORs in EMP1:(EPOR)2 and EMP33:(EPOR)2 [14], [15] are similar to the R2 structures in the respective 1∶2 complexes for GH and EPO (with Cα RMSDs of ∼1.3 Å). In particular, there is no evidence for significant change in the relative orientation between the N-terminal and C-terminal subdomains of either ECD upon forming any 1∶2 complex. (Calculations using some of these alternative structures as well as those taken from molecular dynamics simulations of the 1∶2 complexes produced similar results.) There were then only six remaining degrees of freedom in mapping the inter-subunit energy landscape: three for relative separation and three for relative rotation.
To facilitate describing the orientational rearrangement on going from the transient complex to the native complex, we refer to the N-terminal and C-terminal subdomains of the R1 ECD as N1 and C1, and analogously N2 and C2 for the subdomains of R2. We present orientational changes as rotations of R2 relative to R1. To that end, we define a coordinate system in which the z axis is the long axis of C1 (directed upward), the y axis is perpendicular to the long axes of C1 and N2, and consequently the x axis is in the plane defined by the two long axes and roughly parallel to the N2 long axis (Figure 2A). We refer to the view into the z axis as top view, and the view into the x axis as side view. Figure 2B–D presents the configurations of the receptor molecules in the 1∶2 native complexes of the three systems in these two viewing directions.
In Figure 3 we display 5 representative transient-complex configurations each for the GH-GHR, EPO-EPOR, and PRL-PRLR systems. The top view shows that, for each of the three systems, R2 undergoes clockwise rotation around the z axis on going from the transient complex to the 1∶2 native complex. This “self-rotation” is most prominent for N2 and less so for C2, since the latter is roughly parallel to the rotation axis (i.e., z axis). Meanwhile the side view shows that, again for each of the three systems, R2 undergoes counterclockwise rotation around the x axis on going from the transient complex to the 1∶2 native complex. This “scissor-like rotation” brings together the membrane-proximal ends of C1 and C2.
To quantitatively characterize the orientational rearrangement, we define two angles: γ for the angle between the projections of the N1 and N2 long axes on the x-y plane; and φ for the angle between the projections of the C1 and C2 long axes on the y-z plane. The values of these angles in the native complexes of are: γ = 163° and φ = −7° in GH:(GHR)2; γ = 132° and φ = 0° in EPO:(EPOR)2; and γ = 157° and φ = 20° in PRL:(PRLR)2 (Figure 2B–D). From the transient complex to the native complex, clockwise self-rotation can be recognized as a decrease in γ, and scissor-like rotation can be recognized as a decrease in φ. The distributions of γ and φ in the transient complexes of the three systems are shown in Figures S1, S2, and S3. The distributions are asymmetric with respect to the γ and φ values in the native complexes, with higher values more favored in the transient complexes, supporting the self-rotation and scissor-like rotation illustrated in Figure 3 on going from the transient complex to the native complex.
Mark and co-workers [27], [28] carried out molecular dynamics simulations of (GHR)2 after removing GH from its 2∶1 complex and of (PRLR)2 after removing PRL from its 2∶1 complex. In the former simulations they found prominent self-rotation corresponding to that depicted in the top view of Figure 3A. In the latter simulations they found prominent scissor-like rotation corresponding to that depicted in the side view of Figure 3C. The simulation results thus accord well with our transient-complex calculations.
Examination of the structures of the three 1∶2 native complexes revealed that the asymmetry in φ can be attributed to the wrapping of a C1 loop (between strands A and B) around C2 (Figure S4). A C2 configuration with φ lower than the native value tends to encounter steric clash with the C1 loop. In contrast, C1 presents a relatively flat surface on the side of the native C2 where φ is higher than the native value, allowing the sampling of the high φ values. In the cases of GH:(GHR)2 and PRL:(PRLR)2, the extended N-terminal tail of the cytokine enforces the asymmetry in γ by providing an additional interaction surface for N2 configurations with γ higher than the native value. Recent experimental results of Jomain et al. [21] have implicated a role of the PRL N-terminal tail in receptor activation. The dictation of the transient-complex ensemble by the interface shape is reminiscent of observations on the binding of a ribotoxin to an RNA loop on the ribosome [29]; there ribosomal proteins around the binding interface were found to shift the positioning of the transient-complex ensemble.
Our transient-complex calculations revealed the ECD orientational rearrangements of the three receptor dimers induced by the binding of the corresponding cytokines. These orientational rearrangements are similar, involving both self-rotation and scissor-like rotation, and are largely dictated by the shape of binding interface.
The orientational rearrangement of the ECDs has to be transmitted via the TMHs to the ICDs, to properly position and orient the associated JAK2 molecules for transphosphorylation. Based on our previous study [23] and the present results on the three cytokine-receptor systems, we propose a common model for receptor activation illustrated in Figure 4 (see also Supplementary Video S1). First a cytokine binds to an unoccupied receptor R1 via site 1 to forms a 1∶1 complex. Then R2 in the preformed dimer approaches site 2. Initially the ECD N-terminal subdomains of R1 and R2 are separated at ∼180° and the membrane-proximal ends of the two ECD C-terminal subdomains are apart. Subsequently the two ECDs undergo scissor-like rotation to bring together the membrane-proximal ends of the two C-terminal subdomains, and simultaneously self-rotation to reduce the angle between the N-terminal subdomains. As a result of the scissor-like rotation, the ECD-TMH linkers and the N-terminals of the TMHs move closer, while the C-terminals of the TMHs and the box-1 regions of the ICDs are separated, making room for the associated JAK2 molecules. Meanwhile the self-rotation allows the JAK2 molecules to orient properly for transphosphorylation.
Our calculations were based on the structures of the 1∶2 complexes of the three cytokines with the corresponding receptor ECDs. These structures are likely preserved in the 1∶2 complexes involving the full-length receptors bound to cell membranes, for the following reasons. First, structures of the receptor ECDs in apo form and in 1∶1 and 1∶2 complexes with their cytokines have been determined by different groups. As noted above, the multiple structures for each system are all very similar, attesting to their stability. Second, the ECD of each receptor is separated from the TMH by a linker of ∼10 residues, suggesting minimal perturbation of the ECD by the TMH in the full-length receptor. While separating the ECDs from the TMHs, the linkers play the important role of relaying the rotational motions of the ECDs to the TMHs. (A similar role was identified for an inter-domain linker in the activated of a ligand-gated ion channel [30].)
The ECD orientational rearrangements of the receptor dimers determined here occur after the two receptor molecules are loosely bound, and thus the fact that the molecules reach this state via diffusion in the 2-dimensional membrane has no bearing. The resulting motions of the TMHs and box-1 regions are speculated, but seem to be supported by a host of experimental observations, as we detail below.
Our transient-complex calculations identified a common rotational pathway that receptor dimers are likely to follow upon ligand binding. If the rotations induced are incomplete, then the ligand will likely act as a partial agonist or antagonist. This conclusion is supported by the EPOR partial agonist EMP1 and antagonist EMP33. In EMP1:(EPOR)2, γ = 168° and φ = 39° (Figure 2C). Both values are higher than the counterparts in EPO:(EPOR)2, just like those in the transient complex of EPO:(EPOR)2 (Figure S2). That is, in terms of receptor orientational arrangement, EMP1:(EPOR)2 and the transient complex of EPO:(EPOR)2 deviate from EPO:(EPOR)2 from the same direction. The receptor configuration induced by EMP1 can thus be viewed as an intermediate along the way to the fully activated configuration as found in EPO:(EPOR)2, explaining why EMP1 is only a partial agonist. In EMP33:(EPOR)2, γ = 182° and φ = 38° (Figure 2C), the former angle deviating even more than that in EMP1:(EPOR)2 from the counterpart in EPO:(EPOR)2. The receptor configuration induced by EMP33 is thus an earlier intermediate compared to that induced by EMP1, and hence EMP33 is an antagonist. The fact that EMP1 is a partial agonist but EMP33 is an antagonist despite the similar φ angles of EMP1:(EPOR)2 and EMP33:(EPOR)2 directly supports our contention that both scissor-like rotation and self-rotation are required for receptor activation (see below for further discussion).
We also calculated the transient complexes formed by EMP1 and EMP33 with EPOR, and found that they too followed the common rotational pathway of the GH-GHR, EPO-EPOR, and PRL-PRLR systems. The distributions of γ and φ for the EMP1 and EMP33 transient-complex ensembles are shown in Figure S2. Figure S5 displays 5 representative configurations each for the EMP1 and EMP33 transient complexes. Clockwise self-rotation (top view) and scissor-like rotation (side view) similar to those shown in Figure 3 are also seen in approaching the native complexes here.
From the distributions of γ and φ in Figure S2, it can seen that the EMP33 transient complex is comprised of configurations closely clustered around the EMP33 native complex, and they all fall inside the configurational space of the EMP1 transient complex. It appears that EMP33 locks the receptor dimer in the configurations found in the EMP1 transient complex and prevents it from further orientational rearrangement toward more active configurations. EMP33 differs from EMP1 by two additional bromine atoms on Tyr4 residues (located in site 1 and site 2) of the dimeric ligand. The additional contacts seem key to the locking action of EMP33.
Our analysis on the complexes of EMP1 and EMP33 with EPOR suggests a strategy for designing antagonists based on transient-complex calculations. One first uses the configurations constituting the transient complex of a full agonist as targets; ligands (like EMP1) that stabilize these transient-complex configurations may be candidates for partial agonists. In the next iteration, configurations constituting the transient complex of a thus designed partial agonist become targets; ligands (like EMP33) that stabilize the new generation of transient-complex configurations may be candidates for antagonists. This process may be further iterated.
Constitutively active receptors obtained by Seubert et al. [16] and Brown et al. [10] through deletion or insertion mutations on TMHs demonstrate the involvement of self-rotation in receptor activation. Insertions and deletions move residues on the C-terminal side of the point of mutation along the helical wheel. This has the same effect as self-rotation on the associations JAK2s. Each deleted (inserted) residue in the TMH corresponds to a 103° counterclockwise (clockwise) rotation (top view). Starting with the state in which R2 is loosely bound to the 1∶1 complex (Figure 4C), we find that, after either deleting three residues or inserting four residues on the TMHs, the associated JAK2s are oriented in proximity (Figure S6), similar to that brought about by the receptor self-rotation in our activation model (Figure 4D). These are precisely the numbers of deleted and inserted residues that Seubert et al. [16] and Brown et al. [10] found to result in constitutive activity. We emphasize, however, both self-rotation and scissor-like rotation are required in our model of receptor activation. We note that the dimeric coiled coil replacing the ECDs in the constitutively active EPOR mutant engineered by Seubert et al. [16] would likely bring the N-terminals of the TMHs together, thus achieving the same effect as cytokine-induced scissor-like rotation.
Other experimental observations also support the proposed role of scissor-like rotation in receptor activation. Zhang et al. [31] found that a disulfide linkage between Cys241 residues, located in the middle of the ECD-TMH linkers (Figure 4), occurred only after forming the GH:(GHR)2 complex. This observation suggests that the ECD-TMH linkers are apart before GH binding and come into contact in the 1∶2 complex. This movement of the linkers is just what is brought about by the scissor-like rotation of R1 and R2 (Figure 4).
Brooks et al. [32] using FRET observed that GHR ICDs moved part by ∼9 Å in an active receptor dimer relative to an inactive dimer. They concluded that reorientation (akin to our self-rotation) is critical but insufficient for full activation. Their observation and conclusion are in line with our model of receptor activation.
Recently Liu and Brooks [33] replicated the alanine-insertion of Brown et al. [10] on PRLR. In contrast to the results of Brown et al. for GHR, Liu and Brooks did not find any constitutively active dimer after inserting up to four alanines. Since it takes seven residues to cover all positions on a helix wheel, insertions of five and six alanines would be required to complete the full range of relative orientation of the two TMHs. It is possible that the five- or six-alanine insertion mutant would be constitutively active. It is also possible that none of these alanine-insertion PRLR mutants has sufficient scissor-like rotation for activation.
Other experiments can be designed to further test our model of cytokine receptor activation. For example, inter-receptor distances at different positions along the z axis could be obtained by double cross-linking with bifunctional reagents, which bridge between two receptor molecules and can be used as molecular rulers [34]. The distances, before and after cytokine binding, between residues in the ECD-TMH linkers and between residues in the box-1 regions will be particularly useful for validating and refining our model. It will then even be worthwhile to start building structural models for receptor constructs that are truncated only after the box-1 region, as either preformed dimer or in an activated complex.
Orientational rearrangements such as self-rotation have been implicated in the activation of thrombopoietin receptor and many tyrosine kinase receptors [35]–[37]. The detailed activation model presented here for three cytokine receptors and our approach based on transient-complex calculations will be useful for elucidating the activation mechanisms of a wide range of receptors.
In conclusion, our calculations suggest that R2 undergoes a combined scissor-like rotation and self-rotation to reach the activated state upon binding to the cytokine-R1 complex. The similar observations in all the three cytokine-receptor systems allow us to propose a common model for class I cytokine receptor activation. Both the scissor-like and self-rotation are required for the activation
The implementation of our transient-complex theory used the structures of native complexes as input. Here native complex referred to a 1∶2 complex comprised of one cytokine molecule and two receptor molecules. The structures of the 1∶2 complexes of the GH-GHR, EPO-EPOR, PRL-PRLR, EMP1-EPOR, and EMP33-EPOR systems were from Protein Data Bank entries 3HHR [1], 1EER [6], 3NPZ [7], 1EBP [14], and 1EBA [15], respectively. In the complex containing either EMP1 or EMP33, the EPO mimetic peptide was present as a dimer. All hydrogen atoms were added and energy minimized by the AMBER program.
The N-terminal tail of GH (residues 1 to 5) changes orientation on going from the 1∶1 complex to the 1∶2 complex, from extending sideways to wrapping around R2. We used the orientation of the N-terminal tail of GH in the 1∶1 complex, but counted those N-terminal residues in touch with R2 in the 1∶2 complex when calculating contacts for determining the transient complex (see below).
The N-terminal tail of PRL (residues 1 to 10) is disordered in both the 1∶1 and 1∶2 complexes, and shows an ensemble of conformations in the NMR structure of the unbound state (Protein Data Bank entry 1RW5) [38]. Jomain et al. [21] implicated a role of the N-terminal tail in receptor activation. We thus chose to build the N-terminal tail by Modeller (version 9v8) [39], in an orientation wrapping R2 and similar to that in one of the NMR models for the unbound PRL. To further mimic the situation with the GH-GHR system, we pulled the N-terminal tail so that it extended sideways. The subsequent treatment of this N-terminal tail when determining the transient complex was the same as described for the GH-GHR system.
The implementation of our transient-complex theory for protein-protein association has been described previously [23]–[25]. Briefly, while fixing the 1∶1 complex in space, R2 was translated and rotated around the native-complex configuration. The three translational degrees of freedom were represented by the displacement vector r between the centers of the binding surfaces on the two subunits. The binding surfaces were defined by heavy atoms making <5 Å cross-interface contacts in the native complex. Of the three rotational degrees of freedom, two were a unit vector e attached to the mobile R2 and the remaining one was the rotational angle χ around the unit vector. The unit vector was perpendicular to the least-squares plane of the interface heavy atoms.
To sample the native-complex energy well and the transition region to the unbound state, the six translational and rotational coordinates (r, e, χ) were randomly generated, with the magnitude, r, of r restricted: r≤rcut. The value of rcut was automatically determined to ensure that the clash-free fraction of the randomly generated configurations was ≥10−4 [25]. The resulting rcut values were 6, 6, 12, 6, and 7 Å for the GH-GHR, EPO-EPOR, PRL-PRLR, EMP1-EPOR, and EMP33-EPOR systems, respectively. Clash between the 1∶1 complex and R2 was detected exhaustively over all inter-subunit atom pairs.
For each clash-free configuration, the total number, Nc, of contacts, either native or nonnative, made by a list of “interaction-locus” atoms across the binding interface was calculated as a surrogate of short-range interaction energy. The interaction-locus atoms were selected from the interface atoms as follows. Native pairs of the interface heavy atoms were sorted in ascending order of interatomic distances; each pair was then evaluated against preceding pairs for possible elimination. Specifically, a pair was eliminated if it was within 3.5 Å of a preceding pair on either side of the interface. The final remaining list constituted the interaction-locus atoms. The purpose of the selection process was twofold: to increase the chance that retained native pairs were distinct from each other; and to decrease the chance of nonnative contacts so that there was a proper balance between native and nonnative contacts. The value of Nc in a randomly generated configuration was calculated by counting the number of native contacts and nonnative contact. The upper limit in distance for forming a native contact was the native distance plus 3.5 Å. To count nonnative contacts, the native distance of each native pair was split in half to define the contact radii of the two atoms. A nonnative contact was considered formed when the interatomic distance was less than the sum of their contact radii plus 2.5 Å.
The Nc values of the GH-GHR, EPO-EPOR, PRL-PRLR, EMP1-EPOR, and EMP33-EPOR native complexes were 56, 32, 81, 31 and 44, respectively. As the two subunits moved apart, Nc decreased gradually and the range of allowed rotation angles, as indicated by the standard deviation in χ of the clash-free configurations, increased sharply. The midpoint of this sharp transition (where Nc≡Nc*) defined the transient complex (Figure S7) [25]. From 8×106 clash-free configurations, the values of Nc* were determined to be 12, 15, 16, 19, and 13, respectively, for the five systems, and the 9,114, 48,078, 2,276, 19,407, and 13,361 configurations with these respective Nc values constituted the transient-complex ensembles.
By calculating the basal rate constant to reach the transient complex and the electrostatic interaction energy within the transient complex, the transient-complex theory further predicts the protein association rate constant in solution. Details of these two components and the calculated rate constants are given in Supporting Text S1.
|
10.1371/journal.pmed.1002573 | Future cost-effectiveness and equity of the NHS Health Check cardiovascular disease prevention programme: Microsimulation modelling using data from Liverpool, UK | Aiming to contribute to prevention of cardiovascular disease (CVD), the National Health Service (NHS) Health Check programme has been implemented across England since 2009. The programme involves cardiovascular risk stratification—at 5-year intervals—of all adults between the ages of 40 and 74 years, excluding any with preexisting vascular conditions (including CVD, diabetes mellitus, and hypertension, among others), and offers treatment to those at high risk. However, the cost-effectiveness and equity of population CVD screening is contested. This study aimed to determine whether the NHS Health Check programme is cost-effective and equitable in a city with high levels of deprivation and CVD.
IMPACTNCD is a dynamic stochastic microsimulation policy model, calibrated to Liverpool demographics, risk factor exposure, and CVD epidemiology. Using local and national data, as well as drawing on health and social care disease costs and health-state utilities, we modelled 5 scenarios from 2017 to 2040:
Scenario (A): continuing current implementation of NHS Health Check;
Scenario (B): implementation ‘targeted’ toward areas in the most deprived quintile with increased coverage and uptake;
Scenario (C): ‘optimal’ implementation assuming optimal coverage, uptake, treatment, and lifestyle change;
Scenario (D): scenario A combined with structural population-wide interventions targeting unhealthy diet and smoking;
Scenario (E): scenario B combined with the structural interventions as above.
Scenario (A): continuing current implementation of NHS Health Check;
Scenario (B): implementation ‘targeted’ toward areas in the most deprived quintile with increased coverage and uptake;
Scenario (C): ‘optimal’ implementation assuming optimal coverage, uptake, treatment, and lifestyle change;
Scenario (D): scenario A combined with structural population-wide interventions targeting unhealthy diet and smoking;
Scenario (E): scenario B combined with the structural interventions as above.
We compared all scenarios with a counterfactual of no-NHS Health Check.
Compared with no-NHS Health Check, the model estimated cumulative incremental cost-effectiveness ratio (ICER) (discounted £/quality-adjusted life year [QALY]) to be 11,000 (95% uncertainty interval [UI] −270,000 to 320,000) for scenario A, 1,500 (−91,000 to 100,000) for scenario B, −2,400 (−6,500 to 5,700) for scenario C, −5,100 (−7,400 to −3,200) for scenario D, and −5,000 (−7,400 to −3,100) for scenario E. Overall, scenario A is unlikely to become cost-effective or equitable, and scenario B is likely to become cost-effective by 2040 and equitable by 2039. Scenario C is likely to become cost-effective by 2030 and cost-saving by 2040. Scenarios D and E are likely to be cost-saving by 2021 and 2023, respectively, and equitable by 2025. The main limitation of the analysis is that we explicitly modelled CVD and diabetes mellitus only.
According to our analysis of the situation in Liverpool, current NHS Health Check implementation appears neither equitable nor cost-effective. Optimal implementation is likely to be cost-saving but not equitable, while targeted implementation is likely to be both. Adding structural policies targeting cardiovascular risk factors could substantially improve equity and generate cost savings.
| Previous evidence for population cardiovascular screening (e.g., National Health Service [NHS] Health Check in England) has failed to conclusively provide answers around effectiveness, cost-effectiveness, and equity.
A recent systematic review found some evidence of cost-effectiveness, but some of the studies included were methodologically flawed.
The researchers simulated the life courses of synthetic individuals under different scenarios, within a close-to-reality synthetic population of Liverpool.
This study provides detailed evidence on effectiveness, cost-effectiveness, and equity of health checks while also uniquely providing estimations of the speed at which these programmes will reach these thresholds.
The results show that, while different implementation structures for health checks may improve cost-effectiveness and equity, these approaches are dominated by the addition of population-wide, structural interventions to improve diet and reduce smoking (e.g., mandatory salt reformulation of processed foods or increased taxation on tobacco products).
Liverpool is a city with a high concentration of cardiovascular risk factors and high burden of cardiovascular disease (CVD), and therefore the likelihood of programme effectiveness increases.
The findings highlight the importance of a carefully considered approach to the implementation of health checks while also considering other approaches for CVD prevention.
The addition of structural interventions will lead to primary CVD prevention becoming equitable and cost-effective earlier.
| Cardiovascular disease (CVD) is a leading cause of morbidity and mortality globally [1]. In 2015, almost 500,000 deaths across the United Kingdom were attributed to CVD, more than any other cause of death [2]. As part of the prevention of CVD, the National Health Service (NHS) Health Check programme was implemented across England in 2009. Local-authority commissioning of NHS Health Check has been a statutory requirement since 2013 as part of the Health and Social Care Act [3]. The programme involves CVD risk stratification—at 5-year intervals—of all adults between the ages of 40 and 74 years, excluding any with known preexisting vascular conditions (including CVD, diabetes mellitus, and hypertension, among others) [4]. Those identified as high-risk are offered appropriate treatment, including pharmacological and behavioural interventions. There are limited data available on the national programme. There is little consistency in commissioning arrangements; however, most areas use local General Practices (GPs).
There is conflicting evidence regarding the effectiveness—and cost-effectiveness—of health checks. In the UK, cost-effectiveness for public health interventions is often compared to a threshold of £20,000 per quality-adjusted life year (QALY) gained [5]. The programme was modelled by the UK Department of Health in 2008 [6], suggesting that health checks were cost-effective. However, concern was raised that assumptions about the effectiveness of lifestyle interventions might have been overestimated. A more recent study looking only at the weight-loss impact of NHS Health Check found them cost-effective, through attributing the weight-loss effect exclusively to health checks [7]. A Cochrane review found that, while health check programmes increased diagnosis, they had no significant effect on survival; however, the majority of studies in the review were from before 1980, prior to the introduction of many pharmacological interventions (such as statins for hypercholesterolaemia) [8]. A systematic review of economic evaluations for health checks reported that some modelling and observational studies found programmes to be cost-effective. However, some of these were methodologically flawed (i.e., pre–post designs without a control group) and were based on assumptions regarding uptake and treatment effectiveness that are not supported by empirical evidence [9]. A recent study has suggested that targeting health checks to those most at risk might increase cost-effectiveness [10].
There are also concerns about the potential effect of health checks on health inequalities. An analysis of the first 4 years of the NHS Health Check programme suggested low, but improving, overall uptake and higher uptake in the most deprived communities. However, this did not incorporate a complete 5-year cycle; therefore, it may reflect that, in some areas, the invitation strategy prioritised people with lower socioeconomic circumstances [11]. A rapid systematic review for the Expert Scientific and Clinical Advisory Panel identified that there was wide variability in uptake across England, and the evidence regarding differential uptake by sex, ethnicity, or deprivation was inconclusive [12]. Public Health England (PHE) recently suggested that the programme could potentially be successful in reducing inequalities, but this would require high and equitable uptake in high-risk groups [13]. Local authorities have the flexibility to focus more resources on high-risk communities, making a targeted approach on top of universal coverage possible through an approach of proportionate universalism [14]. A microsimulation study suggests that this could increase the equity of the intervention [15]. Yet there is continuing concern that interventions that target high-risk individuals such as the Health Check programme might exacerbate health inequalities given that they have the potential to favour populations with higher levels of resources [16,17].
On the other hand, population-wide policy approaches to reducing CVD can be effective, cost-effective, and improve health inequalities [16]. These approaches typically focus on tobacco and alcohol control as well as dietary interventions, such as mandatory salt reformulation of processed food or a sugar-sweetened beverage (SSB) levy [18–23]. Therefore, the optimal combination of individual and population-level strategies to reduce the unequal burden of CVD is still not well-defined.
The current study focusses on Liverpool, a city in northwest England with high levels of socioeconomic deprivation. Liverpool is ranked the fourth most deprived local authority in England when considering the proportion of neighbourhoods in the most deprived quintile [24]. Furthermore, there are high levels of inequality within the city, with notable clustering of CVD risk factors that might favour targeted, high-risk approaches to prevention. Health checks were implemented in Liverpool in 2010, but this remains suboptimal; an audit of the programme in 2015 and 2016 found large variations in practice [25]. Barely two-thirds of eligible residents were invited, of which less than one-third completed a health check. These figures are below the national average of about 86% and 49%, respectively [26]. Patients from more deprived areas were significantly less likely to attend, and men less so than women.
The aim of this study was therefore to quantify the cost-effectiveness and equity impact of the NHS Health Check programme in Liverpool, comparing current implementation, targeted implementation, or their combination with structural interventions.
The IMPACTNCD model, a discrete-time stochastic microsimulation, was used to counterfactually simulate individual life courses in the scenarios. The model structure and validation is described in detail elsewhere [20,27,28]. We present, in detail, model inputs, model outputs, scenario specification, and additional results and validation in S1 Text. In brief, we set the simulation horizon to 2040 to enable this preventative intervention to have time to become effective, and we present the results for ages 30 to 84.
We inputted data on demographics and projections (by age, sex, and English Index of Multiple Deprivation quintile group [QIMD]) for Liverpool into the IMPACTNCD model to create a synthetic dynamic population of 200 million adults aged 30 to 84 at baseline. QIMD is a measure of relative area deprivation based on the Index of Multiple Deprivation [29]. According to this system, all Lower Super Output Areas in England (average population of 1,500) are ranked in order of increasing deprivation, based on 7 domains of deprivation: income; employment; health deprivation and disability; education, skills, and training; barriers to housing and services; crime and disorder; and living environment. Then, the QIMD is formed from the quintiles of the above index. Population exposure to 7 known CVD risks was extracted from the Health Survey for England (HSE) using a subsample of northwest England residents. The 7 risks were inadequate fruit and vegetable consumption, physical inactivity, smoking, high body mass index, hypertension, hypercholesterolaemia, and diabetes mellitus. Trends in these risks between 2002 and 2014 were projected to 2040, stratified by demographics, to estimate future exposure. We used the published relative risks from high-quality meta-analyses and cohort studies to link risk factor exposure with disease outcomes. Table 1 summarises the input sources for IMPACTNCD, and Table 2 presents the key modelling assumptions.
The following list summarises model outputs:
Second-order Monte Carlo simulation was used to estimate 95% uncertainty intervals (UIs), propagating estimated uncertainty of inputs to the outputs. Many sources of uncertainty are shared between scenarios; therefore, between-scenario results are not statistically independent and covary to an extent. Therefore, a crossover between UIs for scenarios should not be seen as evidence against statistical significance (please refer to S1 Text for between-scenario comparisons). We present the probability estimates of each scenario being cost-effective (net cost of less than £20,000 per QALY gained), cost-saving (negative net cost), and equitable (reduces both absolute and relative socioeconomic inequalities in health). We defined a probability threshold of 80% to determine when—or whether—a scenario becomes cost-effective (or cost-saving, or equitable).
We used local audit data to determine costs [25], while anonymised aggregated data from the local Clinical Commissioning Group were used to estimate prescription rates. The modelled costs were based on the payment that Liverpool City Council makes to GPs to provide the NHS Health Check programme [25]. This is £5.11 for an invitation and £13 to £19 per participant who undergoes a Health Check. The range of participation cost reflects that Liverpool City Council incentivises GP practices that achieve high uptake. Disease costs were drawn from NICE economic modelling and were separated by the first year of diagnosis, subsequent years, and fatal CVD events [60–63]. Deprivation weighting was used to match the deprivation profile of Liverpool, as there is evidence that there is a social gradient of costs [64]. Estimates assume that costs are equal for all ages and sexes, while myocardial infarction is used as a proxy for coronary heart disease (CHD). Non-CVD complications of diabetes are not included in cost estimates. Costs are inflated using the UK Treasury GDP deflator, November 2016 [65].
We searched for disease utility index score multipliers that used the Euroqol 5-dimension scale (EQ-5D), which is seen as the ‘gold standard’ for health technology assessment in the UK and is recommended by NICE [66]. The evidence base was used to determine baseline utility scores by age [67]. A multiplier of 0.778 was used for CHD [68], 0.629 for stroke [69], and 0.901 for diabetes [70]. Hypertension was not given a multiplier because there was no consistent evidence, especially considering the link between hypertension and other morbidities.
Productivity losses were estimated separately for CHD and stroke, accounting for both working years lost due to early mortality and sickness absence [71,72]. Prices were inflated using ONS data, then weighted based on Liverpool median earnings (95% of the median earnings for the UK). Indirect costs (such as informal care) were not included, and productivity losses were not included in the main cost estimations, which were from a health and social care perspective.
An annual discount rate of 3.5% was applied from 2016 to net costs and QALYs. Results from prior to 2016 were inversely discounted. This rate was selected based on guidance from the UK Treasury [73].
In total, 20 scenarios were progressively developed through an iterative process with public health practitioners in Liverpool. Those scenarios included isolated improvements in coverage, uptake, prescription, referrals to lifestyle services, and some of their combinations. None were shown to be substantially better than the current implementation except the 5 scenarios we present in this study (for a detailed description and justification, please refer to S1 Text).
(A) Current implementation. In this scenario, we modelled the impact and costs assuming that the NHS Health Check programme continued unchanged in Liverpool throughout the modelled period to 2040. This assumed that annual coverage would remain at 13.8%, annual uptake would be 32.3%, and prescription rates would be 9.1% for low-risk, 25.8% for medium-risk, and 41.7% for high-risk participants. The costs were estimated at £5.11 per invitation and £13.28 per successful participation in NHS Health Check.
(B) Current plus targeted. In this scenario depicting a proportionate universalism approach, we modelled the impact and costs if the NHS Health Check programme was targeted toward individuals in the most deprived quintile, with increased coverage and uptake in this population while coverage and uptake in all other groups remain as the current implementation scenario. The coverage in the most deprived quintile is assumed to be 20%, with uptake at 66%. Crucially, in this scenario, we assume that the hypothetical recruitment strategy in the most deprived areas manages to attract participants with a higher cardiovascular risk profile than in scenario A. Prescription rates are unchanged from the current implementation scenario. The costs were estimated at £5.11 per invitation and £15.00 per successful participation in NHS Health Check. The increase in participation cost is in line with current Liverpool City Council practice to monetarily incentivise GP practices that achieve high uptake.
(C) Optimal uptake and prescription rates. In this scenario, we modelled the impact and costs if the NHS Health Check programme met the PHE requirements to invite 20% of the eligible population, and uptake was 66% [74]. Prescription rates in this scenario were set to 9.1% for low-risk, 80% for medium-risk, and 80% for high-risk participants to better reflect current prescription guidelines. In addition, we assumed highly effective lifestyle services. The costs were estimated at £5.11 per invitation and £15.00 per participant.
(D) Current plus structural interventions. This scenario modelled the impact and costs of the current implementation of NHS Health Check (scenario A) with the addition of structural interventions. The structural interventions were stricter tobacco control, an increase of fruit and vegetable consumption by a portion per day in 50% of individuals in the population, mandatory salt reformulation of processed foods, and an SSB tax of 20%. Their effectiveness was informed by published, mostly modelling studies [20,21,75–77].
(E) Current plus targeted plus structural interventions. This scenario modelled the impact and costs of the targeted implementation of NHS Health Check (scenario B) with the addition of structural interventions. The structural interventions are as in the previous scenario.
All results were reported following CHEERS guidelines (S1 CHEERS Checklist).
The model estimated that about 94% (95% UI: 93% to 96%) of CHD incidence can be attributed to the modelled risk factors in the least-deprived quintile group. This gradually increased with deprivation to reach 96% (95% UI: 94% to 97%) for the most-deprived quintile group. Similarly, for stroke, the proportions were 86% (95% UI: 81% to 90%) and 92% (95% UI: 89% to 94%) for the least- and most-deprived quintile groups, respectively.
Table 3 compares scenarios considering the number of CVD cases prevented and number of QALYs gained. The ‘current implementation’ was the worst-performing scenario for both outcomes and by both years (2030 and 2040). The ‘optimal implementation’ of Health Check performed better (estimated to prevent or postpone approximately 750 and 2,000 CVD cases by 2030 and 2040, respectively). However, that performance was dwarfed when structural interventions were added. The ‘targeted Health Check plus structural interventions’ was the best-performing scenario, estimated to prevent or postpone approximately 1,800 and 3,800 CVD cases by 2030 and 2040, respectively. Notably, even ‘current implementation plus structural interventions’ outperforms all programme-only scenarios.
Table 4 and Fig 1 show cost-effectiveness analysis. The current implementation was unlikely to become cost-effective before 2040, while the targeted implementation would only pass the 80% threshold by 2040. The optimal implementation would pass the threshold by 2032. Comparatively, the addition of structural interventions means that scenarios D and E might become cost-effective by 2023, almost a decade earlier. Fig 2 shows that current and targeted implementations are unlikely to be cost-saving by 2040, while optimal implementation might reach the threshold by 2040. The current implementation with structural interventions could reach cost-saving fastest—by 2024—while targeted and structural approaches are likely to achieve cost-saving by 2026.
Table 5 and Fig 3 display findings concerning health inequalities. All 3 programme-only scenarios showed a decrease in absolute inequality—modest for the current implementation and substantial for the optimal implementation. When considering relative inequalities, only the targeted implementation was associated with an improvement in equity, crossing the 80% threshold by 2039. By contrast, the structural intervention scenarios both resulted in much larger improvements in both absolute and relative inequalities. Both would cross the 80% threshold before 2025.
All 5 scenarios would produce productivity savings by 2040 (Table 6). These savings would be largest for the targeted Health Check plus structural interventions (scenario E), with savings of approximately £23 million above the current implementation. This was over 5 times greater savings than for targeted implementation (scenario B) alone and 4 times greater than the savings increase for optimal implementation.
Our study suggests that simply continuing the current implementation of the Health Check programme in Liverpool is unlikely to be cost-effective or equitable, even when modelled 2 decades into the future. If implementation of the programme met optimal PHE recommendations for uptake and prescribing [74], then cost-effectiveness might be achieved by 2030 and cost-savings achieved by 2040 for the Liverpool population. However, it would not improve relative health inequalities. Conversely, a targeted approach towards the most deprived (those at the highest risk) would likely improve equity by 2039 but only reach cost-effectiveness by 2040.
Moreover, our findings may have wider implications nationally and beyond. NHS Health Check is an intervention that targets high-risk individuals. Theoretically, these interventions are more effective when the risk is concentrated in some subgroups of the population [78]. Historically, Liverpool is a city with concentrated deprivation and concentrated CVD risk that translates consistently to worse CVD burden and outcomes compared to the national average [79]. Therefore, we expect the effectiveness and cost-effectiveness of NHS Health Check—even considering optimal implementation—to be worse elsewhere than in Liverpool.
These findings add to the results of a systematic review by Lee and colleagues in 2017 [9]. Analysing the evidence for the cost-effectiveness of population-wide CVD screening, the review found that there was a lack of robust evidence to support the implementation of such screening. As a result, the authors recommended that further evidence is required to identify the cost-effectiveness of such screening, and different delivery models should be examined, such as targeted implementation or population-wide interventions. This modelling study examines these issues directly and provides the evidence recommended.
The addition of structural interventions addressing smoking, diet, and an SSB tax provides a stark contrast. These scenarios would almost certainly be cost-effective, cost-saving, and equitable and reach these 80% thresholds much more quickly. This is unsurprising when remembering Rose’s paradigm about sick individuals and sick populations [80].
The UK Department of Health estimated that the ICER for NHS Health Check would be under £3,000 per QALY gained [6]. However, our findings suggest the ICER would be substantially higher than this by 2030, regardless of implementation. Furthermore, substantial improvements to the delivery and implementation of NHS Health Check would be required to achieve an ICER under £3,000 per QALY gained by 2040. Our estimates are comparable to those from Crossan and colleagues suggesting that optimal implementation would reduce ICERs, perhaps to less than £10,000 (versus £23,000 or more for typical implementation) [10]. The different estimates of these studies are due to differences in the modelled NHS Health Check implementation—and thus effectiveness—and wider differences in the modelling assumptions overall (i.e., assumptions regarding the future burden of CVD or use of different simulation horizons).
These results from our IMPACTNCD model may be of particular interest to local commissioners. Public health interventions have generally been found to be cost-saving [81], thus NHS Health Check appears to represent a comparatively expensive approach to prevention [82]. However, the NHS Health Check programme is currently a statutory requirement for local governments in England. Given this and the major pressures on local-authority budgets [83], the need to improve the cost-effectiveness of prevention programmes with the addition of cost-saving interventions is becoming increasingly clear.
The UK government introduced a sweetened drinks industry levy (SSB tax) in April 2018 [84]. However, other population strategies to reduce CVD could be added. Rather than wait for further action at the national level, we would advocate that local governments take the initiative in implementing population health-improvement strategies. Liverpool did that in the past. In 2004, the local government took the initiative to make Liverpool a smoke-free city. This local initiative was instrumental in the passage of the relevant national law 2 years later [85].
A number of cities across the United States, such as Boston and Oklahoma [86,87], have engaged in collective approaches to encourage healthier diets and weight loss. New York City implemented a number of approaches to improve fruit and vegetable consumption, particularly among low-income individuals [88]. These approaches included financial support to reduce cost barriers, incentives to improve access, support to improve supply and demand for healthy food, improving food quality in different organisations (including schools and childcare centres), and individual support programmes. This combined approach was associated with small but significant improvements in fruit and vegetable consumption. The legal and political feasibility of similar local public health policies for English cities is largely unexplored. Moreover, devolution deals such as those in Liverpool and Manchester may give local governments additional power to implement more structural preventive policies.
To our knowledge, IMPACTNCD is the first dynamic stochastic microsimulation model to estimate cost-effectiveness and equity of the NHS Health Check programme at a city level. Focusing on the city level allowed us to use real-world data on disease rates, costs, and programme success that are not easily available at the national level. In addition, the choice of Liverpool—a city in which the concentration of CVD risk theoretically would favour NHS Health Check—allows the derivation of useful analogies for other areas.
There are, however, a number of limitations. First, health-related quality of life decrements were assumed equal across all socioeconomic groups. However, when considering CVD, people from lower socioeconomic backgrounds have significantly reduced quality of life compared to higher socioeconomic backgrounds. This is likely to underestimate our cost-effectiveness and equity estimates for scenarios B and E [89]. Second, the costs for NHS Health Check only consider the prices paid by the Liverpool City Council. We have assumed that these costs are consistent and have not considered the potential effects of saturation or diminishing returns. The costs of medications for CHD, stroke, diabetes, and hypertension are included in the broad healthcare unit cost estimates. Only the costs of statins for people with no other diseases are not explicitly included, but these costs are small, at around £20 to £40 per patient per annum. Furthermore, we did not include opportunity costs, resulting in the potential for the Health Check to displace more effective health interventions [90]. Third, while most of the structural policies we modelled are incremental improvements on existing policies, their political feasibility is unclear in the current environment. Fourth, our model only considers CVD and diabetes mellitus. Both NHS Health Check and the structural interventions we modelled can potentially reduce the burden of other noncommunicable diseases, too.
Our results suggest that current NHS Health Check implementation appears neither equitable nor cost-effective for local authorities. Optimal administration and implementation might result in better value for money in the next few decades. However, the addition of structural interventions could substantially reduce CVD risks, improve equity, and generate cost savings within short time-scales.
|
10.1371/journal.pcbi.1005970 | The exclusive effects of chaperonin on the behavior of proteins with 52 knot | The folding of proteins with a complex knot is still an unresolved question. Based on representative members of Ubiquitin C-terminal Hydrolases (UCHs) that contain the 52 knot in the native state, we explain how UCHs are able to unfold and refold in vitro reversibly within the structure-based model. In particular, we identify two, topologically different folding/unfolding pathways and corroborate our results with experiment, recreating the chevron plot. We show that confinement effect of chaperonin or weak crowding greatly facilitates folding, simultaneously slowing down the unfolding process of UCHs, compared with bulk conditions. Finally, we analyze the existence of knots in the denaturated state of UCHs. The results of the work show that the crowded environment of the cell should have a positive effect on the kinetics of complex knotted proteins, especially when proteins with deeper knots are found in this family.
| Self-tying of knotted proteins remains a challenge both for theoreticians and experimentalist. In this work, we study the proteins with complex, the 52 knot, in a bulk and confined within a chaperonin box. We show that in our model we recreate the experimental results, identify two topologically distinct folding pathways and explain the beneficial role of confinement for complex knotted proteins. Encapsulation provides a possibility to fold via alternative pathway—folding via trefoil intermediate knot (N-terminal pathway) from entropic reason while folding via the C-terminal (direct tying) appears with the same probability. The results of this work show, how crowded environment in the real cell may enhance self-tying of proteins. The results are also the first step to the identification of possible oligomerization-prone forms of UCHs, which may cause neurodegenerative diseases.
| The role of knots in protein structures is still not fully understood. The topological complexity induces stability to the structure [1, 2] and enforces local motifs favorable for active sites of enzymes [3]. The latter fact may explain, why over 80% of known knotted proteins are enzymes with the active site located at the entangled region [4]. Nevertheless, folding process of knotted proteins is a fundamental and still not solved problem.
One of the families of knotted proteins is Ubiquitin C-terminal Hydrolase (UCH) of which the characteristic feature is the presence of a complex topological fingerprint 523131 [4] as shown in the Fig 1. This means that the entire protein forms a 52 knot as a whole, but some of its subchains form two trefoil knots (see Table 1). Each entry in the matrix indicates the knot type, formed by one continuous subchain, by one particular color; e.g. the unknot is denoted in white. Each such subchain starts with the N-terminal amino acid at position x and ends with the C-terminal amino acid at position y, and the corresponding colored entry in the matrix is shown in position (x, y) (along respectively horizontal and vertical axes). Specifically, one can trace what is the topology of the subchain with one end in N-terminus. Pictorially, this is represented as the traveling down of the left-most vertical line in the matrix in Fig 1). In the beginning, successive subchains are unknotted, however reaching at least Ile163 the subchain becomes trefoil knotted (first green patch in the matrix). The subchain Met1-Tyr173 is still knotted, however then the chain winds back forming a slipknot loop and when the end of the subchain is in-between Glu174 and Pro180 (parts of C-terminal β-strands), such subchain is unknotted (the break between the green patches). Next, the subchain starting in N-terminus and ending in-between residues Pro182 and Ala216 is again trefoil knotted (bottom green patch) and finally the whole chain is 52 knotted (blue patch). The 523131 fingerprint is unique and conserved in all UCH members, which are separated by billion years of evolution and exhibit a very low sequence similarity (below 30%) [5]. Notably, the formation of the larger trefoil results in the formation of the inner-most (dipper) trefoil knot. Therefore, in subsequent analysis by “formation of 31 knot”, we mean the formation of the larger (and hence both) trefoil knot.
The UCH superfamily is a group of deubiquitinating enzymes (DUBs). Their exact substrates have not yet been determined, however it seems that the role of UCHs is to detach the ubiquitin from small nucleophiles. Four of the UCH family members exist in humans: UCH-L1, UCH-L3, human UCH-L5 (UCH37) and BAP-1. They share a high degree of homology in their catalytic domains [6], surrounded by the deepest 31 knot. Moreover, UCHs have a tissue-specific expression in complex organisms such as humans and their activity is crucial from the therapeutical point of view. For example, UCH-L3 has been shown to be upregulated in breast cancer tissues [7], and a high expression of UCH-L5 is significantly associated with poor prognosis in human epithelial ovarian cancer [8]. On the other hand, UCH-L1 is one of the most common proteins in human brain (composing up to 1-2% of the brain the total protein content [9]), and it is highly expressed in pancreatic [10], esophageal [11], prostate [12], medullary thyroid [13], colorectal carcinomas [14] and HPV16-transformed cells [15]. Its misfolded forms were connected with neuronal disorders such as Parkinson’s, Huntington’s and Alzheimer’s diseases [16], which justifies the importance of studying the UCH folding process.
In general, it is expected that folding of knotted proteins is governed mainly by the depth of the knot and the complexity of the topological fingerprint [17, 18]. Self-tying was observed theoretically for the smallest knotted proteins with DNA binding motif and a rather shallow 31 knot [19, 20]. They mainly fold by a slipknot conformation [17, 19]. Similarly, it was shown that proteins with a deep trefoil knot, such as YibK and YbeA, can self-tie [21]. The theoretical results obtained for these proteins with a structure based model additionally revealed that a knotting event is a rate-limiting step [22] and the folding efficiency can be controlled by non-native contacts [23] or consideration of cotranslational on-ribosome folding [24]. For the protein with a 61 knot, DehI [25], there were only a few successful folding pathways observed theoretically. Surprisingly, this protein folds via a simple mechanism: a large twisted loop formed on the backbone flips over another protein fragment previously arranged in a twisted loop, and in consequence, the six-fold knot is created in a single movement. Even though this protein is prone to aggregate, the experimental data support this mechanism [26]. These results suggest that bulk structure-based models can be used to investigate knotted proteins.
On the other hand, experimental data show that knotting process of trefoil-knotted YibK and YbeA bacterial proteins can be specifically and significantly accelerated by the GroEL-GroES chaperonin complex [21] encapsulating the folding protein. This agrees with the theoretical investigation showing that knotting probability of polymers increases in confinement [27]. Only due to the encapsulation (following [28]), successful reversible folding was observed for members of knotted proteins with DNA binding motif VirC and DndE [29]. It is then natural to expect, that chaperonins encapsulating proteins may also facilitate folding and self-knotting of eucaryotic UCHs, although no experimental result in this topic is available yet. Nevertheless, one has to bear in mind, that encapsulation is the simplest possible model of chaperonin, lacking many “biological features”—specific binding to the cage, chaperonin conformational changes, etc.
However, any theoretical results have to be confronted with the experimental data concerning UCHs folding. It has been shown already that UCHs can fold and refold reversibly in two parallel pathways, each consisting of one slow and one fast phase, as determined from chevron plot [30, 31]. Despite a common mechanism, folding processes of different UCHs is characterized by different kinetic parameters. Such differences can stem from a various depth of knots in UCH family [4], ranging from rather shallow (from the N-terminus) to the deep knot, which we just found, as shown in Table 1. The most attention-drawing are the S18Y and I93M mutations, which were found to modify (either decrease or increase) the risk of Parkinson’s disease [32], and the intermediates on folding pathway, as these are especially prone to oligomerization [33]. However, the results concerning these mutations are variable and differ in different studies. Despite the successful assignment of the majority of signals in NMR spectrum of the UCH-L1 [34], and in spite of studies of its tryptophan variants [33], the exact conformation of intermediates is still unclear. This may be due to a broad structural plasticity around the intermediate states [30]. The self-tying was postulated via the direct knotting event in accordance with the theoretical study of the on-lattice model of a designed by hand heteropolymer chain with 52 knot [35]. However, the optical tweezers stretching experiment showed that the threading significantly decelerates the folding [36]. Still, because the topology cannot be detected in the in vitro experiment, the mechanism of knot tying remains unresolved.
In this study, we asked following questions: What is the difference between the two experimentally observed, parallel folding pathways? What is the influence of a chaperonin cage (confinement) on the folding and self-tying of UCHs? And more generally, what are the dynamical properties of UCH in a bulk and in a confinement? To answer these questions, we performed a comprehensive study of representative UCH members in a model of chaperonin cage (mimicked by repulsive cavity) and in the bulk, using structure-based model simulations. To ensure robustness of the results, we investigated proteins from different organisms, with low sequence similarity, and different depth of the knot. The results show that the structure-based model was sufficient to knot and unknot each of the studied proteins with and without the presence of confinement. However, only in the confinement, the simulations in transition temperature were accessible. We performed a comprehensive analysis of knot occurrence during simulations, resulting in the identification of two topologically distinct pathways. To relate our results to the experiment, we reproduce the chevron plot for a representative protein member of UCH family, revealing the existence of fast and slow phases. Next, we studied short-lived knots on the folding and the unfolding pathway and revealed for the first time the existence of random knots in the unfolded protein chain. To our knowledge, this is the first theoretical study with a direct investigation of the influence of the excluded volume on proteins containing complex knots.
To obtain robust results, three sequentially different members of UCH family denoted with their PDB codes 3IRT (UCH-L1), 2LEN (UCH-L1), and 4I6N (UCH-L5) were studied. Additionally, to check an influence of the length of knot tails, we constructed in silico mutant of 4I6N, denoted as 4I6N-m, obtained by removing 7 residues from the N-terminus. All investigated structures feature a left-handed 52 knot and complex topological fingerprint 523131. Alignment of the 3-dimensional structures of the studies proteins is shown in the Fig 1 and their most important topological and structural information is summarized in Table 1. Further structural and sequential comparison of chosen structures is presented in S1 Appendix, part 1. Not determined region (amino acids 142–152) in the structure of 4I6N was repaired using Modeller software [37], where the model with the lowest DOPE potential (Discrete Optimized Protein Energy) was chosen. The DOPE potential is one of the quantities assessing the structure correctness [38].
The dynamics of investigated structures was studied in structure based Cα model [39, 40] with standard parameters as proposed by the SMOG server [41]. The model included bonded interactions (bonds, planar and dihedral angles), bead excluded volume (Lennard-Jones repelling part) and non-bonded interactions described with a 10-12 Lennard-Jones potential. The non-bonded attraction was applied between residues forming contacts in the native structure, as defined in [42]. The number of the native contacts for all considered proteins is presented in Table 1. The folding/unfolding transitions were studied through constant temperature molecular dynamics simulations with the Nose-Hoover thermostat (coupling constant eq. 0.025) using Gromacs v4.5.4 package [43]. Temperature T ˜ in Gromacs is defined by the equation T ˜ = ( k b T ) / ( ϵ k˜b), where k˜ b= 0 . 00831451. Through the text, the temperature is denoted simply as T and the Boltzman constant k ˜b as kb. There were performed 200 simulations for each structure and temperature. The number of steps was in the range 107 − 1.6⋅109 steps depending on the condition.
The confinement is represented by a cylinder (Fig 1) with a diameter equal to its height equal 6.0 nm [44], introduced into the system as in [29]. The interactions between the inner wall of the cylinder and protein are purely repulsive (only a confinement effect). Such model was previously used to study confinement or crowding effect on protein folding [44–46].
The data concerning the position of the knotted core and the length of the knot tails (Table 1) are taken from KnotProt server [4]. The knot type of each of the subchains of the protein is determined using the implementation of the HOMFLY-PT polynomial [47–49] and the chain closing method as in [5, 50]. The same algorithm was used to detect the entanglement along the protein backbone during simulations. The knot was regarded as present in the simulation if it was detected for at least 5 consecutive frames.
The similarity to the native state was measured by the fraction of native contacts, Q. At given conformation, each native contact was regarded as present, if the distance between a pair of Cα atoms was less than 1.2 times their native distance. The untied structure was regarded as unfolded, if Q < 0.2. By unfolding pathway, we mean the shortest part of trajectory connecting knotted structure with Q > 0.9 and unfolded structure. By folding pathway, we mean the shortest part of the trajectory connecting unknotted structure with Q < 0.4 and knotted structure with Q > 0.9. Folding trajectories start from one of the previously generated 100 unknotted conformations with Q < 0.2. All initial structures belong to separate clusters with 0.1 nm cutoff, to remove any possible bias.
The structures were visualized using UCSF Chimera [51].
The landscape of successful folding pathways that leads to the correctly knotted native conformation is observed for each of the proteins, in both conditions through our simulations. As the folding of all UCHs is similar [31], we concentrated our study on the most commonly analyzed protein—UCH-L1 (PDB code 3IRT), comparing results with other UCHs when needed. Our results naturally split into three parts: description of folding/unfolding landscape, kinetics of the process and an analysis of the random and short-lived knots.
The UCHs are known to fold and refold along two parallel pathways, each featuring one slow and one fast phase [30, 31, 52]. Therefore, to correlate our model with experimental results we recreated the chevron plot for UCH-L1, with the temperature as a denaturant. In conventional chevron plots, the folding constants are calculated based on the time dependence of e.g. fluorescence. The fluorescence of a protein’s tryptophan depends on its neighborhood. Hence, the fluorescence trace can be understood as a measure of similarity of the tryptophan neighborhood to the native structure. In our simulations, such a measure is given by the fraction of the native contacts—Q. Therefore, we calculated the average Q as a function of time (representative trace in Fig 5A). The average was taken over all simulations in a given T and in each condition (bulk/confinement). Next, we fitted the smoothed Qaver(t) with the sum of exponential functions. In particular, we fitted the trace with the sum of the highest number of exponents, for which the fitting errors were lower than 5% of the value. The details of the plot along with the values of obtained constants and errors are presented in S1 Appendix, part 6. Although Qaver(t) is only a rough equivalent for the fluorescence, in almost all cases we were able to decompose the trace as a sum of 2-4 exponentials, i.e. to find up to 2 fast and 2 slow phases. These data are shown in Fig 5B (bulk) and 5C (confinement) in the form of chevron plot with the inverse of temperature (precisely −ϵ/kbT) mimicking the denaturant concentration [53]. For the consistency with conventional chevron plot, we plot the logarithm of k = 1/τ where τ is a characteristic time of a given phase. The obtained values create the trends characteristic for chevron plots, therefore they were connected by dashed lines. Note that in some cases the connection of values is arbitrary. For most cases, the fitting error was an order of magnitude smaller than the value obtained (Fig 5), and those, for which the error was higher (e.g. the point for confinement, −ϵ/kbT = −1.10), still corresponds to reasonable values.
The presence of the fast and slow phase during folding/unfolding process shows that our results are clearly consistent with previous experimental observations [30, 52]. However, in most cases, we were not able to determine four individual phases (as in the experiment). This may be due to similar characteristic times of separate phases, the model imperfection, or because of a much more complicated folding pathway. Indeed, the curvature of the limbs of the chevron plot indicates more complicated mechanism in each phase, again consistently with the experiment [33]. The detailed analysis of folding/unfolding pathway should be the next step in investigating of these proteins.
To determine the influence of the confinement, we compared the “most complete” kinetic trace for the slow and the fast phases for bulk and confinement (Fig 5D). The slowest phases can be fitted to an equation describing chevron plot, which yields an approximate Tf equal 114 (−ϵ/kbT = −1.05) for bulk and 120 (−ϵ/kbT = −0.99) for the confinement (for details see S1 Appendix, part 6). This indicates that confinement stabilizes UCH as it was observed for proteins with trivial topology [44, 54]. Moreover, the chevron plot indicates that the confinement significantly accelerates the folding process, especially the slow phase. In particular, the simulations in the Tf for bulk were not accessible computationally due to very slow rates, while they were accessible in Tf for the confinement. This enables us to calculate near-equilibrium F(Q) dependence for confinement, which in principle could give additional information on UCHs folding. However in this case, due to the complexity of the folding landscape, the standard ways of its representation do not reveal any new information (S1 Appendix, part 7).
Regardless of the conditions, a collapse of the protein (the first phase of folding) occurs relatively fast, which stays in accordance with the experimental results, that the knotting (occurring in our model for Q > 0.7) should be the rate-limiting step [36]. Therefore in our case, the fast phase corresponds to arriving at collapsed, non-knotted form (first part of folding) and the slow phase should correspond to knotting and subsequential reaching of the native structure. The impact of the confinement on the slow phase indicates that the confinement facilitates knotting by restricting the conformational space of the termini. On the other hand, the confinement slows down the unfolding process by slowing down the unknotting—note the change of order in the curves in Fig 5D. The lower unfolding rates in the confinement may be also a result of the retying during unfolding (discussed in the next section). As it turns out that the once unfolded knot has a higher probability to retie in the confinement which results in higher knot stability and slower unfolding. Slow unfolding is again in agreement with intuition and experimental observation made for proteins with trivial topology [55].
To additionally investigate the influence of the confinement on both unfolded and folded state, we determined the average asphericity [56] parameter for bulk and confinement in both states. Again, the asphericity of folded state was comparable in both conditions, indicating, that the confinement does not influence the near-folded structures significantly. On the other hand, the asphericity of unfolded state was different in the confinement than in bulk, showing the influence of confinement on the unfolded basin (S1 Appendix, part 8).
The probability of knot presence in a polymer chain increases rapidly with its length. As a result, it is highly probable that the sufficiently long polymer will spontaneously form a knot. However, the fraction of knotted proteins is far lower than in the case of equally long polymers [57]. Moreover, the spontaneous self-tying of protein chains in the denaturated state was not reported so far even in the natively deeply-knotted structures [22, 58, 59] or in the case of small knotted proteins in confinement [29]. However, in the case of UCHs we observe a significant fraction of (in most cases short-lived) knots, appearing during folding/unfolding pathway, or in the denaturated state (with Q < 0.2).
We found that confinement leads to faster and more efficient folding of UCH proteins for two reasons. First, encapsulation provides the possibility to fold via an alternative pathway. More precisely, the confinement facilitates folding via the trefoil knot (the FC pathway) for entropic reasons, while it does not affect folding via the FN pathway (direct tying, the N-terminus folds the last). This surprising behavior is supported by the experimental observation of uneven influence of chaperonin [63] on the substrate protein rhodanese, which decelerates the folding of the C-terminal domain, but leaves the folding rate of the N-terminal domain unaffected [64]. Second, at the same confinement stabilizes native interactions and destabilizes non-native ones in comparison to bulk, and thus it reduces the height of the free energy barrier and accelerates the folding rate as it was observed for a protein with trivial topology [55, 65, 66]. The two pathways clearly distinguishable in our analysis are in accordance with two pathways identified in experiment [30]. However, there is still no technique, which could determine the topology of the protein during folding in vitro, which prevents from direct validation of our results. Some insights can however be given by the study of the tryptophan variants of UCH [33]. In particular, it was shown that the pathways differ in the structure of intermediates for which highly stable central β-sheet core and flanking α-helices and loop regions are formed differently. This is in accordance with our results, however detailed analysis of folding pathway, with special emphasis put on the location of tryptophan mutations, is required to precisely compare the experimental and theoretical results.
On the other hand, our results show that the confinement introduced by the chaperone-like cage decelerates the unfolding of UCH proteins. Firstly due to a decrease in the effective mobility of the protein backbone upon encapsulation and under topological constrains, which reduce the rate at which new configuration can be explored (especially in twisted loops). This argument without topological contribution is used to explain lower folding rate for a protein with the trivial topology [64]. Secondly, the decelerated unfolding in the confinement is caused by the retying phenomena. It is worth pointing out that the retying phenomena can be used by other knotted proteins with a rather shallow knot, e.g. carbonic anhydrase, to stabilize the structure in a crowded environment (moreover carbonic anhydrase structures with deeper knots also start to be crystallized).
Under the confinement, a significant number of short living knots is observed in the denatured state and in folding and unfolding routes of UCH proteins, what has not been reported for other knotted proteins. More and more complicated knots are more common to occur upon encapsulation, which is in the agreement with the polymer theory [61, 62]. These knots seem to have only positive effect, i.e. their formation accelerates folding. In principle, deeply knotted structures could lead to misfolding, but contrary to the situation in the bulk, they are not formed due to the constrained configurational space.
In summary, we took advantages of structure based model and knot theory, and made the step forward in characterizing folding/unfolding routes for UCH proteins identified experimentally in [30]. We identified possible oligomerization-prone forms of UCHs, which may cause neurodegenerative diseases. We found that weak confinement smooths the rough and not continuous free energy landscape of UCH proteins in a subtle way, e.g. enhancing an indirect tying route. However, at low temperature or strong confinement slower folding should be again observed due to restriction on indirect tying. The deceleration under strong confinement was suggested for a protein with the trivial topology in [63].
|
10.1371/journal.pcbi.1004658 | NINJA-OPS: Fast Accurate Marker Gene Alignment Using Concatenated Ribosomes | The explosion of bioinformatics technologies in the form of next generation sequencing (NGS) has facilitated a massive influx of genomics data in the form of short reads. Short read mapping is therefore a fundamental component of next generation sequencing pipelines which routinely match these short reads against reference genomes for contig assembly. However, such techniques have seldom been applied to microbial marker gene sequencing studies, which have mostly relied on novel heuristic approaches. We propose NINJA Is Not Just Another OTU-Picking Solution (NINJA-OPS, or NINJA for short), a fast and highly accurate novel method enabling reference-based marker gene matching (picking Operational Taxonomic Units, or OTUs). NINJA takes advantage of the Burrows-Wheeler (BW) alignment using an artificial reference chromosome composed of concatenated reference sequences, the “concatesome,” as the BW input. Other features include automatic support for paired-end reads with arbitrary insert sizes. NINJA is also free and open source and implements several pre-filtering methods that elicit substantial speedup when coupled with existing tools. We applied NINJA to several published microbiome studies, obtaining accuracy similar to or better than previous reference-based OTU-picking methods while achieving an order of magnitude or more speedup and using a fraction of the memory footprint. NINJA is a complete pipeline that takes a FASTA-formatted input file and outputs a QIIME-formatted taxonomy-annotated BIOM file for an entire MiSeq run of human gut microbiome 16S genes in under 10 minutes on a dual-core laptop.
| The analysis of the microbial communities in and around us is a growing field of study, partly because of its major implications for human health, and partly because high-throughput DNA sequencing technology has only recently emerged to enable us to quantitatively study them. One of the most fundamental steps in analyzing these microbial communities is matching the microbial marker genes in environmental samples with existing databases to determine which microbes are present. The current techniques for doing this analysis are either slow or closed-source. We present an alternative technique that takes advantage of a high-speed Burrows-Wheeler alignment procedure combined with rapid filtering and parsing of the data to remove bottlenecks in the pipeline. We achieve an order-of-magnitude speedup over conventional techniques without sacrificing accuracy or memory use, and in some cases improving both significantly. Thus our method allows more biologists to process their own sequencing data without specialized computing resources, and it obtains more accurate and even optimal taxonomic annotation for their marker gene sequencing data.
| The advent of next-generation sequencing technologies, combined with major advances in molecular and bioinformatics techniques, have enabled rapid growth in the culture-independent sequencing of amplified marker genes (amplicons) from environmental microbial communities. The major benefit of amplicon sequencing is that it allows reasonable resolution of taxonomic composition in these communities at a fraction of the cost of deep metagenomic sequencing. Once these sequences are generated, a common analysis approach is to bin them by sequence identity into operational taxonomic units (OTUs)[1–4]. For environments containing a large fraction of novel taxa, one must rely on unsupervised (“de novo”) clustering of amplicons to convert the raw reads to features representing organisms belonging to distinct evolutionary clades. On the other hand, in habitats with mostly well-characterized microbes, we have the option of matching the generated amplicon sequences to reference databases containing example marker genes from known taxa [5]. A hybrid approach may also be used, where sequences are first compared to a reference database, with subsequent de novo clustering of those that failed to match. As the number of published culture-independent amplicon-based surveys of microbial communities continues to grow, our ability to rely on reference sequences also increases. However, although the crucial analysis step of mapping generated amplicons to reference marker genes has received much attention from the microbial bioinformatics field, with a variety of solutions proposed [6–10], there is much room for improvement in terms of speed, accuracy, memory footprint, and openness of code. NINJA-OPS, our portable, open-source OTU picking pipeline, realizes these goals.
Originally conceived as a means to make data more compressible, the Burrows-Wheeler transform (BWT) [11] is a lossless, reversible transformation that effectively positions series of like characters close to each other in a way that can easily be undone to recover the original data. It involves creating a circular suffix array, sorting the final column lexicographically, and storing that column as the transformed data for later compression. This algorithm also has the interesting property of enabling rapid substring search, with O(1) order of growth in finding exact string matches. As long as there is an efficient indexing scheme that stores the indices of the transformed bases into the original string, the BWT can be used for fast database substring search amounting to binary searching (or looking up via rank matrix) the transformed reference string representation and mapping back to the original, and has hence been employed in a number of commonly used DNA alignment tools [11–14]. Although these tools are approximate methods due to the high additional computational cost of performing optimal local or global alignment search when mismatches occur, they are generally fast and widely used in the genome-enabled research community (http://bowtie-bio.sourceforge.net/bowtie2/other_tools.shtml).
Here we demonstrate that BWT-enabled DNA alignment can be effectively used for accurate and fast assignment of marker-gene sequences to a reference database. We present the NINJA-OPS pipeline utilizing several novel contributions to achieve an order of magnitude speedup and higher accuracy when compared to commonly used approaches (or up to two orders of magnitude when combined with denoising). To test the accuracy and efficacy of our approach, we perform closed-reference OTU-picking on a wide range of biological data sets from varied environments.
Accuracy was evaluated using an optimal aligner which produces a BLAST-style %ID for each query sequence against the reference sequence chosen by the OTU picking method. Speed was assessed as the elapsed time from parsing the correctly-formatted input FASTA file, which is accelerated by NINJA’s fast C parser and simple format requirements, until the alignment (against a pre-generated database) has terminated. However, it may be useful to note that NINJA also significantly speeds up the subsequent steps of tallying reads, incorporating taxonomic annotations, and producing an OTU table in sparse BIOM 1.0 format, as well as other steps prior to the alignment such as reverse complementing and trimming reads. Hence, the NINJA pipeline accelerates many stages of the OTU-picking pipeline in addition to the alignment step.
The pipeline follows three stages: filtering, aligning, and parsing. After forming the concatenated reference string, called the “concatesome,” from the individual references, NINJA applies a powerful filtering step which uses a 3-way radix quicksort on string pointers to rapidly de-duplicate millions of reads, construct a sample dictionary, and output a reduced-size filtered FASTA file and sample dictionary (Fig 1). The program implements this lossless filtering approach as well as a lossy variant, making use of singleton filtering as well as statistical profiling over the entire set of reads to exclude reads with a user-defined number of duplicates or rare segments (k-mers) appearing below a user-defined threshold of prevalence. The lossy filtering, which is not enabled by default, is intended to identify reads with probable read error independently, and speeds up the resulting alignment by excluding such reads from the BWT aligner. This adds an additional speedup because BWT string matching spends a disproportionate amount of search effort to align erroneous or low-identity reads. Because choice of k (from 8 to 14) and prevalence threshold are highly domain- and dataset-specific, it is difficult to issue a general recommendation for this setting. Although we have found k = 8 to k = 14 at a threshold of 0.05%-0.01% to be a safe minimum for 16S data, user experimentation is recommended.
The NINJA filter step also performs reverse-complementing and sequence length trimming at the same time as the other filtering steps. Because of this simultaneous multi-step filtering, no intermediate files are created prior to the alignment stage, and all filtering steps are performed rapidly in optimized C code on data structures already in memory. This takes a fraction of the time used by other filtering pipelines which perform sequential operations often written in general-purpose scripting languages and generate numerous large intermediate files after each step. Using the base NINJA filter parameters, the entire filtering process itself takes approximately 10–20% of the time it takes to align the resulting filtered file when using all optional filtering steps.
Next, the filtered reads are aligned against a reference database containing the (the concatesome) via any BWT-derived short read aligner such as BWA[11], Bowtie[13]/Bowtie2[12], hpg-aligner [15], SOAP2 [16]—or, more broadly, any read aligner whatsoever capable of outputting to headerless SAM format[17] and suppressing unmatched input reads. Utilizing SAM is much faster than BAM (binary compressed SAM) after deduplication, as the alignment step is not I/O bound and the overhead of BAM’s additional compression/decompression step can be significant. We have chosen to standardize NINJA around Bowtie2 for our tests and publish the command line options for Bowtie2 as we have found it to be suitable for the purposes of BLAST-identity-based OTU picking. Following alignment, the resulting SAM file is fed to the NINJA parsing step, which takes in the sample dictionary metadata as well as an optional taxonomy map to rapidly re-assign each de-duplicated read to the biological sample(s) in which it originally occurred, add taxonomy annotation to each picked OTU, bin all reads by their matched OTU into a sample-by-OTU matrix (OTU table), and output the result in sparse BIOM 1.0 format or a tab-delimited human-readable legacy QIIME format. This can also be incorporated into an open-reference OTU-picking pipeline.
The BWT has received a lot of attention in the alignment of short reads to a reference genome, and now enjoys routine use in clinical and other settings as a well-vetted technique for mapping short DNA reads to a longer reference, where it is known as Burrows-Wheeler alignment. The BWT is based on the principle that a long string of text can be reversibly transformed to reduce the complexity of substring queries to effectively two binary searches into the transformed representation of the original string, which is then converted back to indices into the original reference string with a short walk-back (the BW Last-First, or LF walk) or lookup. The efficacy of this approach in matching short reads to a reference database of numerous short reference marker genes has remained largely unexplored [1].
Benchmarks for this article were performed on a 2013 MacBook Air (MacBook Air 6,2) with a dual-core Core i7 CPU and 256GB SSD.
The runtime performance of the database generation is significantly longer than is practical to perform on the fly. This step only needs to be performed once for each reference database. Although ninja_prep performs the concatenation of references rapidly (it is I/O bound on the Macbook’s SSD), the BWT program may spend a long time generating the BW index. For bowtie2 on our test machine, this takes over half an hour (with a maximum of one thread) on the Greengenes 97% OTU representative sequence database. For this reason, it is best to store and use pre-compiled databases for all subsequent alignments, and NINJA-OPS is distributed with a number of pre-compiled databases for commonly performed 16S bacterial marker-gene OTU matching.
NINJA filtering takes approximately 10–15% of the alignment time. For our 1.6 million read 175bp test data, without additional processing, filtering runs in 3.5 seconds and outputs a de-duplicated FASTA file approximately 1/5 the size of the original.
Bowtie2 with the settings mentioned in methods aligns the entire test dataset of 1.6 million 175bp reads in under 40 seconds on a single thread of the test laptop. Performance for the default and maximum-fidelity (“max”) NINJA presets were measured (Fig 2). The “max” preset not only demonstrates higher accuracy than either the default preset or USEARCH, but also retains significantly more reads. The “fast” preset displays similar accuracy characteristics to “default,” but misses about 2% of the alignments detected by the latter, usually of the lowest identity. Total pipeline runtime on the same dataset decreases to less than 10 seconds when using the recommended singleton-based denoising option (parameter “D 2”), in combination with default preset. In addition, the speedup versus USEARCH 8 was measured using the default preset without denoising (“D 0”) across different datasets (S2 Fig). RAM usage during alignment was 205MB in all cases, while that of USEARCH 8 was 720MB. Using multiple threads during alignment decreases the running time further, but speedup is sublinear, having somewhat more advantage in datasets with longer reads or higher error rates (and hence more difficult alignments).
Parsing with ninja_parse takes roughly 0.2–3 seconds on datasets in the size range included here (0.5–2 million sequences). Outputting to legacy tab-delimited format instead of BIOM increases the runtime by a second or two. A Python-based convenience wrapper distributed with NINJA adds additional overhead if the user requests a fasta file containing the sequences that failed to match the database.
To assess the accuracy of the alignments found by NINJA, and to compare them to existing tools, we calculated the optimal alignment, using a semi-global version of the Smith-Waterman algorithm, of each query sequence with the reference sequence assigned by a given tool. We found that NINJA (default preset) generally finds higher-accuracy matches than USEARCH 8 (Mann-Whitney U test p < 2.2e-16) (Figs 3 and 4). In a published dataset containing healthy subjects and patients with Crohn’s disease the two methods produced the same list of differentiated genera across disease conditions with occasional disagreements about the direction of the association (S3 Fig). NINJA produced a comparable percentage of matches with default preset to USEARCH (S4 Fig), and generally comparable taxonomic assignments despite some interesting differences (S5 Fig).
Our tool leverages a combination of several novel approaches to accomplish an order of magnitude speedup over existing methods without compromising accuracy, and in many cases NINJA-OPS is more accurate than popular existing tools. In combination with a recommended denoising step, the pipeline achieves up to two orders of magnitude speedup over USEARCH 8. The key innovation of this tool is our use of a single long reference genome, or concatesome, composed of concatenated marker genes. This approach allows NINJA to leverage the benefits of the Burrows-Wheeler transform (Fig 5). The code is available at http://ninja-ops.ninja (or https://github.com/GabeAl/NINJA-OPS).
Optimizations within NINJA-OPS include tweaks to the parsing and filtering programs to increase the throughput of the processes leading up to the alignment. Deduplication is a viable strategy in marker-gene sequencing contexts because samples usually consist of fewer taxa than there are reads, and in fact are often dominated by a few highly abundant species. This results in a large number of identical reads which can be filtered out to reduce the alignment time. In human gut datasets which are quality-trimmed (or where the marker gene reads are of approximately equal length), this may result in losslessly discarding 80% of the reads as duplicates, depending on the microbial community sampled, which can speed up the downstream alignment step substantially (S6 Fig). A sequence-to-sample(s) dictionary keeps track of the abundance of each sequence in each sample to ensure that each original sequence is properly accounted for wherever it was originally found. By default, NINJA-filter also performs read compaction (parameter “-d 1"), which normalizes for variation in read lengths within a dataset by treating reads which are subsets of longer reads as copies of the longer reads. This increases consistency of OTU calling as well as decreasing runtime. This behavior can easily be disabled (parameter “-d 0”).
An optional beneficial feature during the filter step is the ability to perform lossy denoising. NINJA performs this in two ways. The first and most straightforward for amplicon reads is to discard singleton reads (parameter “-d 2”); that is, reads that have no identical match in the entire list of queries, or which are not perfectly contained in a longer read. This can be extended as the user desires from singletons to doublets and so on (parameter “-d 3”, “-d 4”, etc.). The second form of denoising is discarding reads judged to be erroneous by breaking each read into its component overlapping k-mers and comparing each of these k-mers to the counts of that k-mer in an empirical distribution of all k-mers in the body of input reads. Reads with k-mers that fail to meet user-defined criteria for support (appearing under a certain % in the dataset) are discarded completely from subsequent analysis. The resulting speedup for the downstream alignment is often much greater than the proportion of reads discarded, because Burrows-Wheeler alignment programs expend a disproportionately large amount of effort attempting to align erroneous reads that will not match the database compared to non-erroneous reads which will often find perfect (or near-perfect) matches in a well-populated database. The BW substring search is designed for perfect substring searches, so it performs most efficiently in aligning reads that have few to no mismatches with a subsequence of the database. This is also why NINJA and BWT tools perform most effectively when the alignment identity is high (%ID in the mid-to upper-90’s, with taxonomic resolutions at the level of genus or finer). Performance of BWT-based tools is expected to increase as the diversity of available reference sequences increases, because the probability of finding a perfect match likewise increases.
One early concern as we were considering how to most effectively construct the concatesome was that some reads would align by chance to the boundary between two concatenated marker genes, which would produce a meaningless mapping. However, in practice, such an occurrence is exceedingly unlikely to occur in end-to-end marker gene alignments at genus-level or greater resolution due to the high identity expected over the entire length of the input read. This is even more true of marker gene alignment, where reads are much more similar to each other than in shotgun data, and the possible sites of alignment seeding are likewise similar, with significantly less randomness than would produce alignments with the boundary region by chance. The prevalence of such reads in our 16S test data is accordingly less than 1 in 1,000,000 reads aligned. Furthermore, in the unlikely event that such an alignment does occur, it is trivial to discard it in the final parsing step by testing whether the index at the end of the alignment is equal to or greater than the starting index of the subsequent marker gene. NINJA-OPS automatically discards reads that map to junctions between concatenated marker genes.
An interesting finding that corroborates past findings [21] is that the commonly used bowtie, bowtie2, and BWA alignment tools do not scale linearly with increasing read length. However, due to the ability to substitute alternative BWT-based alignment programs for the alignment step, it is possible to use alternatively optimized variants such as HPG Aligner, which uses uncompressed suffix arrays instead of “traditional” BWT but shares many of the same characteristics with the added benefit of better scaling for longer reads. GPU-accelerated variants of the original algorithm are also available [22]. Additionally, NINJA-OPS is not restricted to the domain of 16S OTU picking, although it is distributed with a pre-built 16S database. Marker genes such as ITS for fungal identification [24], bacterial rpoB [25], and the recently proposed Cpn60 universal bacterial barcode [26] are easily incorporated into NINJA-OPS simply by compiling the included “ninja_prep.c” and running it on an appropriately-formatted FASTA file containing the desired marker sequences, followed by the BWT-based aligner’s database generation step. Further, NINJA-OPS can be incorporated as a preliminary step in another pipeline; for instance, NINJA-OPS can be used to group reads prior to de novo assembly [27]. This flexibility of the pipeline in allowing substitution of the aligner itself, as well as the marker gene database used, makes NINJA-OPS applicable for situations and optimizations beyond what were envisioned at the time of writing, and enable the pipeline to keep pace with emerging technologies in the sequencing and computing spheres alike.
|
10.1371/journal.pntd.0005318 | Identification of Functional Determinants in the Chikungunya Virus E2 Protein | Chikungunya virus (CHIKV) is a mosquito-transmitted alphavirus that causes high fever, rash, and recurrent arthritis in humans. It has efficiently adapted to Aedes albopictus, which also inhabits temperate regions, including Europe and the United States of America. In the past, CHIKV has mainly affected developing countries, but has recently caused large outbreaks in the Caribbean and Latin America. No treatment or licensed CHIKV vaccine exists.
Here, we have identified determinants in the CHIKV cell-attachment protein E2 that facilitate cell binding. The extracellular part of the E2 gene is subdivided into the three domains, A, B, and C. These domains were expressed in E. coli and as Fc-fusion proteins generated from HEK293T cells and used for cell-binding assays. Domains A and B bound to all cells tested, independently of their permissiveness to CHIKV infection. Domain C did not bind to cells at all. Furthermore, CHIKV cell entry was promoted by cell-surface glycosaminoglycans (GAGs) and domain B interacted exclusively with GAG-expressing cells. Domain A also bound, although only moderately, to GAG-deficient cells. Soluble GAGs were able to inhibit CHIKV infection up to 90%; however, they enhanced the transduction rate of CHIKV Env pseudotyped vectors in GAG-negative cells.
These data imply that CHIKV uses at least two mechanisms to enter cells, one GAG-dependent, via initial attachment through domain B, and the other GAG-independent, via attachment of domain A. These data give indications that CHIKV uses multiple mechanisms to enter cells and shows the potential of GAGs as lead structures for developing antiviral drugs.
| The chikungunya virus (CHIKV) glycoprotein E2 mediates cell attachment and consists of three domains A, B and C. Since the cell entry process of CHIKV is not understood in detail, we analyzed the binding properties of the three E2 domains with proteins expressed in E. coli or as Fc-fusion proteins and the role of glycosaminoglycans (GAGs) on E2 cell binding and CHIKV entry. The two surface-exposed E2 domains, A and B, both bound to cells and domain B bound only to cells expressing GAGs. Domain A bound additionally to GAG-deficient cells and domain C did not bind to cells. CHIKV-pseudotyped lentiviral vector and CHIKV entry were enhanced in cells expressing GAGs. Our results suggest that CHIKV uses at least two entry mechanisms, one GAG-dependent, via attachment through E2 domain B, and the other GAG-independent, via binding of domain A. These data give indications that CHIKV uses multiple mechanisms to enter cells and shows the potential of GAGs as lead structures for developing antiviral drugs. In addition, it shows that domain A and B might constitute good targets for vaccine development.
| The Chikungunya virus (CHIKV) is a mosquito-transmitted alphavirus that causes high fever, rash, and recurrent arthritis in humans. The majority of symptoms disappear after about one week. However, in about 30% of cases, arthritis can last for months or even years, which may cause substantial economic losses [1], [2]. The virus has been endemic in Sub-Saharan Africa, the Indian Ocean islands, India, and Southeast Asia. However, the virus spread to the Caribbean in late 2013 and is now responsible for a large, still-ongoing outbreak there and in Latin America with 1.9 million suspected cases as of December 2016 (www.paho.org/hy/). The mortality rate is very low (0.1%), but the infection rates are high (sometimes 30%) and asymptomatic cases are rare (about 15%). Due to climate change, globalization, and vector switching, the virus will most likely continue to cause new, worldwide outbreaks. Additionally, more temperate regions of the world like Europe or the USA, which have recently reported their first cases, will likely become targets [3], [4]. Alarmingly, no specific treatment or vaccination against CHIKV is available so far.
CHIKV is a (+) single-stranded RNA virus. Like other alphaviruses, it enters cells by receptor-mediated endocytosis and a subsequent pH-dependent fusion step. CHIKV has two surface proteins that mediate cell entry: the transmembrane glycoproteins E2 and E1. E2 mediates cell attachment and E1 is a class II viral fusion protein [5], [6]. E2 and E1 associate as trimers of heterodimers (E2–E1) on the particle surface [7], [8], [9]. The E2 protein contains two N-glycosylation sites at position 263 and 345. The E2 envelope protein consists of domain C, located close to the viral membrane, domain A, in the center of the protein, and domain B, at the distal end, prominently exposed on the viral surface [7], [8]. These domains are promising sites of interaction with the target cell.
Potential interaction partners of viruses on the cell surface are glycosaminoglycans (GAGs), which are ubiquitously present on the surfaces of all animal cells and are an essential part of the extracellular matrix (ECM) [10],[11], [12]. They consist of long linear chains of disaccharide units (30–60 per chain). These disaccharides are sulfated to different degrees and are thus negatively charged. GAGs that are covalently linked to a core protein are called proteoglycans (PGs). They differ depending on the carbohydrates that form the disaccharide units. The best characterized GAGs linked to core proteins on human cells are heparan sulfate (HS), chondroitin sulfate (CS), and dermatan sulfate (DS) [12]. Since GAGs are ubiquitously present on the cell surface, many pathogens exploit them to cross the cell membrane barrier and use them for initial cell attachment or as entry receptors. These pathogens include several bacteria, parasites, and viruses [10], [13]. Cell surface HS, the most extensively studied GAG, promotes attachment and/or entry of herpes simplex virus type 1 (HSV-1), human immunodeficiency virus (HIV), hepatitis C virus (HCV), vaccinia virus (VACV), dengue virus (DENV), and adeno-associated virus isolate 2 (AAV-2) into cells [13]. Binding of an alphavirus, the eastern equine encephalitis virus (EEEV), to cell surface HS, thereby enhancing its neurovirulence, has also been reported [14]. So far, the role of GAGs in CHIKV replication has only been studied in the context of viral attenuation. Point mutations within domain A of the E2 protein (e.g., E79K or G82R) have been found in attenuated vaccine strains that were cell culture adapted and showed enhanced GAG dependency but reduced in vivo replication [15], [16], [17], [18]. Additionally, it was reported that cell-surface PGs promote replication of some CHIKV strains, but this replication was not inhibited by the presence of soluble GAGs [15]. Furthermore, one CHIKV strain has been shown to be not influenced at all by the presence or absence of PGs [17]. The E2 domains A and B have been suggested before to be putative receptor binding sites [7], [8].
Since the cell entry process of CHIKV is not understood in detail [19], we examined the role of GAGs in CHIKV cell entry and analyzed the binding properties of the E2 domains A, B and C, and their dependency on GAGs. The two surface-exposed E2 domains, A and B, both bound to cells expressing GAGs. Domain C was not involved in cell binding at all. We could show that CHIKV entry is enhanced in cells expressing GAGs and that domain B binds exclusively to GAGs. Domain A also bound to cells that do not express GAGs. Our results suggest that CHIKV uses at least two entry mechanisms, one GAG-dependent, via attachment through E2 domain B, and the other GAG-independent, via domain A.
All cells used for this study were cultured at 37°C under 5% CO2. HEK 293T (ATCC: CRL-1573) cells were incubated in Dulbecco’s modified Eagle’s medium (DMEM; Lonza, Verviers, Belgium). Jurkat (ATCC: TIB-152) and BHK-21 (CCL-10) cells were grown in Roswell Park Memorial Institute medium (RPMI; Biowest, Nuaille, France). CHO-K1 and pgsA-745 cells were grown in Ham’s F-12 medium (Life technologies, Darmstadt, Germany) and 2 mM glutamine. Media were supplemented with 10% FBS (v/v; PAA, Pasching, Austria) and 5% L-glutamine (200 mM; Lonza, Verviers, Belgium). The glycosaminoglycans chondroitin sulfate, dermatan sulfate, heparan sulfate, heparin and dextran sulfate were purchased from Sigma-Aldrich (Taufkirchen, Germany).
The codon-optimized CHIKV E3-E1 gene (based on isolate “S27”) was synthesized by GeneArt (Life Technologies, Darmstadt) and cloned into the plasmid pIRES2-eGFP (Clontech/Takara, Saint-Germain-en-Laye, France) as described previously [20].
The E2 domain A (including the β-ribbon connector), B (aa 172–231), and C (aa 271–341) genes were cloned into the bacterial expression vector pET-15b. The same was done with the entire extracellular part of the E2 protein (E2ex). Cloning was achieved by adding the restriction sites NdeI and BamHI via primers by PCR (template DNA: pIRES2-EGFP-CHIKV E3-E1), cutting the DNA products and the vector, and subsequent ligation. For domain A, two fragments were derived via PCR. One fragment contained domain A itself and the first part of the β-ribbon connector (C-terminus of domain A (aa 1–171)). The other fragment contained the second half of the connector C-terminus of domain B (aa 231–270). The fragments were cloned into the pET-15b vector by a triple ligation (via NdeI and BamHI). The two fragments were linked via a shared SmaI restriction site (at the C-terminal part of fragment one and the N-terminal part of fragment two, respectively). By this procedure, E2 domain B was bypassed and replaced with the sequence G4PG5. The constructs also contained an N-terminal poly-histidine-tag for purification.
The primers used for cloning were:
The Fc-fusions were constructed by cloning E2 fragments via Apa I, Nhe I sites introduced by PCR, into the vector pCMV2.5-hIgG1Fc-XP (kind gift of Stephan Dübel, TU Braunschweig). This generated expression vectors that express E2-Fc fusion proteins without any tags. The following primers were used to amplify the E2 fragments from the E.coli expression vectors as template: Domain A fw 5’ AAAAGGGCCCAGCACCAAGGACAACTTCAAC, rev 5’AAAAGCTAGCCTTGGGCACCATGCAGGTC; domain B fw 5’ AAAAGGGCCCCCCGACACCCCCGATAGAA, rev 5’AAAAGCTAGCGGTCACGGCGGCGTGGC; domain C fw 5’ AAAAGGGCCCGCCCGGAACCCTACCGTG, rev 5’AAAAGCTAGCCTGGGGCCAGTACTTGTAGG; domain A-ß fw 5’ AAAAGGGCCCAGCACCAAGGACAACTTCAAC, rev 5’AAAAGCTAGCGGGGTCGTGGTGGAAGGG. All mutations were introduced by site directed mutagenesis.
Proteins were expressed in BL21-CodonPlus (DE3)-RIPL competent cells (Agilent Technologies, Böblingen, Germany) transformed with the pET-15b plasmid containing construct A, B, C, or E2ex. Bacteria were inoculated into 100 ml of LB medium containing ampicillin (0.1 mg/ml) and grown overnight (37°C, 220 rpm). After 16 hrs, 2 l of LB medium were inoculated with the 100 ml overnight culture. The bacteria were grown to an OD600 of 0.5–0.7, and then protein expression was induced by the addition of 1 mM IPTG. After another 2.5 hrs of incubation, cells were harvested and the pellets were frozen at –20°C.
The recombinant proteins were purified from the bacterial pellets under native (B, C) or denaturing (A, E2) conditions using HisTrap FF Crude columns (GE Healthcare, Freiburg, Germany) and the ÄKTA system (GE Healthcare, Freiburg, Germany) as described by [21]. For A and E2, ion-exchange chromatography was additionally performed to remove contaminating bacterial proteins. After purification, proteins were dialyzed against PBS using Slide-A-Lyzer Dialysis Cassettes 3.5K MWCO (Pierce, Thermo Scientific, Bonn, Germany) and concentrated with Ultra-4 3 kDa Centrifugal Filter Units (Merck Millipore, Schwalbach, Germany). The protein concentration was determined by SDS-PAGE with marker proteins, following staining with Coomassie (Bio-Rad, Munich, Germany). Proteins were then quick-frozen with liquid nitrogen and stored at –80°C. For experiments, proteins were thawed in a 37°C water bath.
The empty Fc control protein and the E2 domain-Fc-fusion proteins were produced in HEK293T cells by transient transfections. Transiently transfected HEK293T cells were grown in DMEM containing 10% FCS. After 48 h, supernatants of the transfected cells were harvested two times at 24 h intervals. The secreted Fc-fusion proteins were purified by affinity chromatography with protein A-agarose, eluted at a pH 2.5, neutralized with 1M Tris, pH 9.0 and dialyzed against PBS pH 7 and stored at -80°C.
The plasmid pCHIKV-mCherry-490 [22] was in vitro-transcribed with T7 RNA polymerase after NotI linearization. The mRNA was transfected into BHK-21 cells using Lipofectamine 2000 (according to the manufacturer’s protocol; Life Technologies). Virus-containing supernatants were harvested 48 hrs later and used to reinfect fresh BHK-21 cells for virus amplification. Infected cells showed a clear red fluorescence. Supernatants were collected and stored at –80°C or used to determine the viral titer.
For CHIKV infections, 293T cells were seeded onto a 24 well plate. Cells were incubated at 37°C for 16–24 hrs and counted (about 3 × 105 cells). Subsequently, the mCherry-tagged CHIKV (CHIKV-mCherry-490) [22] was added at a multiplicity of infection (MOI) of 1. After 6 hrs, the cells were collected in medium, washed and resuspended in 2% paraformaldehyde in PBS, and analyzed by flow cytometry. At least 10,000 events were acquired with an LSRII instrument (BD Biosciences) and analyzed using FACS Diva software.
Lentiviral vector particle production was performed as described previously [20]. Briefly, 293T cells were seeded in 10 cm dishes in 10 ml DMEM. Cells were cotransfected 16 hrs post seeding with the plasmids pRRLsinhCMV-GFP-pre (a lentiviral vector genome encoding GFP) or pCSII-Luc (a lentiviral vector genome encoding luciferase), pMDLg/pRRE, pRSVrev, and pHIT-G or pIRES2-eGFP-CHIKV E3-E1 using Lipofectamine 2000 (according to the manufacturer’s protocol; Life Technologies). After 24 hrs incubation, the medium was discarded and replaced with 5 ml of fresh DMEM. Another 24 hrs later, the supernatant containing the vector particles was harvested, sterile filtered with 0.45 μm filters (Sartorius, Göttingen, Germany), and frozen at -80°C.
Cells were transduced with pseudotyped lentiviral vector particles in 384-well plates as described previously [20]. Briefly, 6000 293T cells per well were seeded (using a MultiFlo Microplate Dispenser; BioTek, Bad Friedrichshall, Germany) in 20 μl DMEM in white CELLSTAR 384-well microtiter plates (Greiner Bio-One, Frickenhausen, Germany) and incubated for 16–24 hrs at 37°C. The same was done for CHO-K1 and pgsA-745 cells using Ham’s F-12 medium and 3000 cells per well were seeded.
Soluble glycosaminoglycans (GAGs) were serially diluted (three-fold) four times in DMEM containing vector particles in 96-U-well plates (Thermo Scientific, Rockford, IL, USA), and incubated at 4°C for 1 h. This resulted in equal amounts of vector and serially diluted compound. The vector particle mixtures were then added to the cells using a Matrix Multichannel Equalizer Electronic Pipette (Thermo Scientific, Rockford, IL, USA), transferring 20 μl each to three wells of the 384-well plate out of one well of the 96-well plate. This resulted in a final concentration of GAGs/dextran sulfate ranging from 500 to 6.2 μg/ml. Cells were incubated with the vector particle mixtures for another 16–24 hrs. Afterwards, 20 μl of BriteLite substrate (PerkinElmer, Rodgau, Germany) was added. After 5 minutes incubation at room temperature, the luciferase signal was detected using the PHERAstar FS microplate reader (BMG LABTECH, Ortenberg, Germany).
Statistical analyses were done using the GraphPad Prism 5.04 software (La Jolla, CA, USA). The p-values were determined by the unpaired two-tailed t-test.
The recombinant proteins purified from E. coli or E2-Fc-fusion proteins were incubated for 30 min at 4°C with cells in PBS/2% FCS. The cells were then washed and bound protein was detected via an anti His-tag antibody (Dianova, Hamburg, Germany) and an anti-mouse IgG-FITC antibody or an FITC coupled anti-human IgG antibody (Fc-fusion proteins) followed by flow cytometry. Binding was detected as the mean change in fluorescence. In addition, the CHIKV E2-derived protein sA (containing the surface exposed regions of domain A connected by linkers) [23] was used as a negative control for the binding of E.coli derived proteins and the Fc-protein served as negative control for Fc-fusion proteins and the values are given as fold increase in binding compared to Fc.
First it was analyzed if GAGs play a role in viral entry. For this investigation, 293T cells, CHO-K1 cells, and the CHO-K1 derived cell line pgsA-745 which, due to an enzymatic defect, is not able to produce GAGs, were transduced with CHIKV Env- or VSV-G-pseudotyped lentiviral vectors encoding GFP. Transduction was determined as the number of GFP-positive cells and was standardized as percentage of GFP-positive cells obtained after transduction with VSV-G-pseudotyped vectors. The results displayed in Fig 1A show that CHIKV cell entry into 293T and CHO-K1 cells was almost equally efficient. However, the transduction rate of pgsA-745 cells was significantly reduced by more than 50% in comparison to the parental cell line. Thus, cell entry of CHIKV Env-pseudotyped vectors into GAG-deficient cells is strongly and significantly reduced in comparison to those carrying cell-surface GAGs. However, cell entry was only reduced by about 50%, indicating the existence of at least one more entry pathway.
To confirm the relevance of the above experiments, the dependency of CHIKV infections on cell-surface GAGs was studied. For this, CHO-K1 and pgsA-745 cells were both infected with the recombinant CHIKV-mCherry-490 using an MOI of 1. This virus contains an mCherry gene within the nsP3 gene of CHIKV and has growth characteristics similar to the wild-type virus [22], [24]. Viral replication was determined at 6 and 24 hrs post-infection by flow cytometry. The infection rate at 6 hrs post-infection was significantly reduced (9.9-fold) in pgsA-745 cells compared to CHO-K1 cells (Fig 1B). At 24 hrs post-infection, the difference decreased to a 1.5-fold higher, yet still significantly different, infection rate in CHO-K1 cells compared to pgsA-745 cells. Thus, both the cell entry of CHIKV Env-pseudotyped vectors and the replication of CHIKV were reduced on pgsA-745 cells lacking cell-surface GAGs, but not fully inhibited, indicating that GAGs enhance, but are not essential for entry.
CHIKV cell entry into GAG-deficient cells was reduced; accordingly, the presence of soluble GAGs might inhibit CHIKV entry into GAG-expressing cells or influence entry into GAG-free cells. Therefore, 293T, CHO-K1, and pgsA-745 cells were transduced with CHIKV Env-pseudotyped vectors in the presence of different amounts of soluble GAGs. Dextran sulfate (DX), which consists of long chains of highly sulfated glucose units, was used as a control for the soluble GAGs, in addition to HP. It has a similar charge to HP and the other GAGs, but its structural background is built up of entirely different carbohydrates. The experiment was carried out in a 384-well plate format with vectors encoding firefly luciferase [20].
Transducing 293T and CHO-K1 cells with CHIKV Env-pseudotyped vectors transferring a luciferase gene in the presence of GAGs generally resulted in dose-dependently reduced transduction efficiencies compared to the control without GAGs (Fig 2, top). DX and HP were the most potent inhibitors on both cell lines. On 293T cells, the cell entry could be inhibited to about 10–20% of the untreated control at a concentration of 500 μg/μl GAGs. On CHO-K1 cells, transduction was reduced to about 20–30% of the GAG-free control maximally (Fig 2, top). Cell entry of CHIKV Env-pseudotyped vectors into pgsA-745 cells was, on the contrary, dose-dependently enhanced by the addition of rising GAG concentrations (Fig 2, bottom). The highest values reached, at 500 μg/μl, were between 121 (DX) and 177% (HP) of the untreated controls. Substantial inhibition of transduction with values lower than 100% was only observed for DX at concentrations lower than 500 μg/μl. In conclusion, the data indicate that cell entry of CHIKV Env-pseudotyped vectors is enhanced by GAGs, since low level transduction is still possible in pgsA-745 cells and in the presence of soluble GAGs. In addition, the entry into GAG-deficient pgsA-745 cells could be enhanced with increasing amounts of GAGs, indicating an activation of CHIKV Env-pseudotyped vectors.
As shown above, CHIKV replication in GAG-deficient cells was diminished. Accordingly, this raised the question of whether infection of cells is inhibited in the presence of soluble GAGs. 293T cells were infected with CHIKV-mCherry (MOI 1) in the presence of 500 μg/ml of the respective soluble GAGs (but 500 U/ml HP). Six hours later, cells were analyzed by flow cytometry. Fig 3A shows that all GAGs reduced viral replication significantly by at least 76.3% (HS). HP was most effective, reducing replication by 90.7% compared to the GAG-free control. To determine whether the inhibition of CHIKV replication by GAGs occurs at the attachment/entry step of the viral life cycle, CHIKV and the different GAGs were incubated with target cells at 4°C to allow viral attachment to the cells but to avoid the subsequent endocytosis step. After 30 minutes, unbound virus and the GAGs were washed away and the cells were incubated for 6 hrs at 37°C in fresh medium. Infection rates were measured via flow cytometry. Fig 3B shows, similar to the previous experiment, a significant reduction of cell attachment/entry by 62.1 (HS) to 82.4% (HP).
In conclusion, replication of recombinant CHIKV-mCherry-490 in 293T cells is significantly inhibited by the addition of soluble GAGs. Attachment and/or endocytosis are the critical steps in the viral life cycle where this inhibition occurs.
The E2 protein is the cell-binding moiety of CHIKV and has been frequently described to contain determinants recognized by neutralizing antibodies [7], [8], [25]. Therefore, one could speculate that parts of E2 are components that bind the unidentified cellular receptor of CHIKV. To analyze which domains of the E2 protein are involved in cell binding, the sequences encoding domain A (including the β-ribbon connector, aa 1–171 and 231–270), domain B (aa 172–231) and domain C (aa 271–341) were cloned into the bacterial expression vector pET-15b. The same was done for the entire extracellular part of the E2 protein (E2ex), which served as a positive control. The constructs also contained an N-terminal poly-histidine-tag for purification. The proteins were expressed in E. coli, and partially purified by Ni2+ affinity chromatography under native (E2 domains B and C) and denaturing (E2 domain A and E2ex) conditions. For domain A and E2ex additional ion-exchange chromatography was performed to remove bacterial protein contaminants. Fig 4 shows a Coomassie-stained SDS-PAGE separation of the purified proteins. Domain A has a molecular mass of 26.5 kDa, B of 8.5 kDa, C of 10.7 kDa, and E2ex of 40.4 kDa. The purified proteins migrated at the expected size.
Their cell binding was analyzed with CHO-K1, the GAG deficient pgsA-745, 293T and Jurkat cells.293T cells show good transduction efficiencies with CHIKV Env-pseudotyped vectors. In contrast, Jurkat cells have revealed very low transduction efficiencies [20]. The recombinant proteins were incubated with cells at 4°C. Bound protein was detected by flow cytometry using an anti-His-tag antibody and an anti-mouse IgG-FITC antibody. Data are presented as fold increase in the mean FITC values of the sample in comparison to those of the control (cells only treated with staining antibodies). In addition, the protein sA, which contains only surface-exposed domains of A and did not induce neutralizing anti-E2 antibodies upon vaccination [23] and was used as a negative control for the binding studies. Fig 5 reveals that domains A and B, and protein E2ex showed a significant difference in the mean FITC signal compared to the control sample, indicating binding of these proteins to 293T and CHO-K1 cells. In contrast, there was no significant cell binding detectable for the negative control, sA, and domain C. An identical pattern was obtained for Jurkat cells, although at a generally lower binding level. Conducting the experiment with pgsA-745 cells resulted in weak, yet still significant, binding of domain A and E2ex, and no binding of domain B or C (Fig 5). Accordingly, these data indicate that binding of domain B to cells is enabled by cell-surface GAGs, while that of domain A is only partly dependent on these molecules.
To further prove that cell binding of the E2 protein domains is GAG dependent, inhibition of their cell binding by addition of soluble GAGs was analyzed. The recombinant E2 proteins and different cell lines were again incubated at 4°C for binding; however, this time in the presence of 500 μg/ml soluble GAGs. Heparin (HP) was used as a control to ascertain the role of charge in the cell binding. HP is structurally derived from heparan sulfate (HS), but is more heavily sulfated and thus has a higher negative charge density. Fig 6 shows the results of soluble GAG-induced protein-binding inhibition and indicates that binding of domain A to 293T cells was not reduced in the presence of any of the soluble GAGs, except HP (about 2.5-fold). In contrast, the presence of all GAGs reduced the cell binding of domain B up to 3.0-fold. Again, HP showed the most efficient inhibition (4.5-fold). Inhibition of protein E2ex binding was in between that of domains A and B. A similar picture was observed for CHO-K1 cells (Fig 6, middle). Here, in contrast to 293T cells, HP did not strongly inhibit domain A binding. Minor inhibition of domain A binding was detectable after addition of HS and chondroitin sulfate (CS). The same experiment was performed with pgsA-745 cells (Fig 6, bottom). Here, domain C was used as an additional control. However, the binding of the recombinant proteins to pgsA-745 cells was not inhibited or enhanced in the presence of soluble GAGs. The viability of 293T cells incubated with 500 μg/ml of the different GAGs (the maximum concentration used in all experiments) was tested using MTT assays. None of the GAGs were significantly cytotoxic at this concentration (data not shown).
In summary, binding of E2 domain A to 293T and CHO-K1 cells was not or weakly inhibited by the addition of soluble GAGs. Only the strong negatively charged HP inhibited the interaction of domain A with 293T cells. In contrast, binding of domain B to both cell lines was massively decreased in the presence of HP, HS, CS, and dermatan sulfate (DS). The addition of GAGs to domains A, B, or C had no influence on the binding properties of these proteins towards GAG-deficient pgsA-745 cells.
There was a serious concern that the E2 domains expressed in E. coli might not be correctly folded or have incorrect disulfide bond patterns. Therefore the E2 domains were expressed as Fc-fusion proteins in a soluble form in eukaryotic cells by transient transfections of 293T cells. Supernatants containing these proteins were partially purified by protein-A Sepharose chromatography and used for binding assays (Fig 7A). Western blot analysis of the proteins under native conditions confirmed that the proteins were Fc-dimers, as expected for antibodies (Fig 7B). This implies that disulfide bonds were formed and might be present not only in the Fc part, but also in the E2 fragments.
First the three subdomains A, B and C were tested for their cell binding ability as Fc-fusion proteins towards CHO-K1 and pgsA-745 cells. The Fc protein was used as negative control. Two concentrations were tested and “low” represents one sixth of the original sample. Again, domain A and B bound to CHO-K1 cells (Fig 7C). Domain C bound only marginally to CHO-K1 cells. This distribution is similar to the one determined with proteins expressed by E.coli (Fig 5), although domain A expressed from eukaryotic cells had a higher binding affinity compared to domain B, than domain A derived from E.coli. Analysis of cell binding to pgsA-745 cells revealed that domain A shows residual cell binding in the absence of GAGs. Domain C did not bind to pgsA-745 cells.
Point mutations in domain A have been described before to increase CHIKV binding to GAGs [15], [16], [18], therefore we generated two mutants of domain A and expressed them as Fc-fusion proteins, domain A with a mutation at position 79 (Fc-CHIKV-E2-A-E79K) and position 166 (Fc-CHIKV-E2-A-166K). In addition, we generated a construct containing domain A without the ß-ribbon connector (A-ß) (Fig 8A). Again, the proteins run in SDS-PAGE under native conditions as dimers (Fig 8B). The two point mutants in domain A showed a higher binding affinity to CHO-K1 cells compared to domain A (Fig 8C) and their binding to pgsA-745 cells was only weakly increased. This indicates that introducing the point mutations in domain A increased its binding affinity to GAGs. Deletion of the ß-ribbon connector decreased domain A binding to CHO-K1 cells, however binding to pgsA-745 cells remained unaltered compared to domain A containing the ß-ribbon connector. This indicates that the ß-ribbon connector partially contributes to GAG binding and that the GAG independent binding site is located in domain A.
The alphavirus CHIKV enters cells by receptor-mediated endocytosis and a subsequent pH-dependent fusion step [26]. Host factors that are required for CHIKV entry are still only poorly understood. First hints have emerged from genome-wide RNAi screens, where downregulation of archain1, fuzzy homologue or TSPAN9 inhibited CHIKV infection but also that of alphaviruses in general [27]. On the viral side, the CHIKV E2 glycoprotein mediates cell attachment; however, the detailed mechanism has not been studied well so far [5], [6].
The structures of the CHIKV envelope proteins have been solved by X-ray crystallography [7], [8]. The extracellular part of E2 has three immunoglobulin (Ig)-like extracellular domains called A, B, and C [7]. Here, we expressed the extracellular part of E2 (E2ex) and its three domains A, B and C in E. coli. Cell-binding experiments revealed a significant binding of domains A and B, and E2ex, but not domain C, to all cells tested. Domain C is found close to the viral membrane and is followed by the stem region, which is a linker to the transmembrane region composed of hydrophobic amino acids [7]. Since domain C is not surface accessible and does not contain epitopes for neutralizing antibodies [7], it was not surprising that this protein lacked cell-binding activity. Domain B is located at the tip of the protein protruding from the viral surface and is linked to domain A via a long β-ribbon connector (containing an acid-sensitive region [ASR]). Domains A and B are prominently exposed on the viral membrane and it has previously been speculated that the cellular receptor mainly interacts with domain A [7,8,28]. The recombinant proteins containing domains A and B bound to cells independently of the cells’ ability to allow cell entry of CHIKV [20]. Binding to non-permissive Jurkat cells was reduced in comparison to binding to highly permissive 293T cells, which indicates that the binding moiety might be present on all cells but at different densities.
Further analysis of the possible cellular attachment factor was performed with the help of CHO-K1 and pgsA-745 cells. The pgsA-745 cells are derived from CHO-K1 cells and lack glycosaminoglycans (GAGs) on the cell surface. Binding assays revealed that domains A and B bound to CHO-K1, but only domain A bound to pgsA-745 cells, however markedly reduced but still significant. Furthermore binding of domain B to GAG-expressing cells in the presence of soluble GAGs was highly reduced. This suggests that domain B binds GAGs and thereby facilitates CHIKV cell binding. In contrast, domain A binding to GAG-deficient pgsA-745 cells and competition with soluble GAGs only marginally decreased domain A cell binding, indicating GAG-independence.
These data were confirmed with Fc-fusion proteins containing the E2 domains A, B and C, suggesting that the E.coli derived proteins have a native structure. In addition, point mutations in domain A showed the previously observed increase in GAG binding. Deletion of the ß-ribbon connector decreased the GAG binding of domain A, indicating that the ß-ribbon connector is partially responsible for GAG binding. A GAG binding consensus sequence is located in the ß-ribbon connector (-DRKGK- amino acid 251–255) [29]. GAG independent binding was not affected and might be mediated by domain A.
These data were substantiated by transduction of pgsA-745 cells with CHIKV Env-pseudotyped vector particles, which was significantly reduced by over 50% in comparison to transduction of the parental CHO-K1 cells, and soluble GAGs inhibited transduction of GAG-expressing cells. Remarkably, the transduction rates could not be reduced to less than 10% of the soluble GAG-free control, indicating that at least one GAG-independent entry pathway must exist. The presence of soluble GAGs did not decrease, but rather enhance transduction of pgsA-745 cells with CHIKV Env-pseudotyped vectors. This might be simply due to a local acidic environment generated by soluble GAGs, which may induce a fusion competent conformation of the glycoprotein. Alternatively, a pre-activation of the CHIKV envelope proteins through binding to the soluble GAGs might occur. Soluble GAGs did not enhance the binding of domains A or B to pgsA-745 cells; therefore, it is tempting to speculate that GAGs might induce conformational changes within the envelope proteins that allow them to bind more effectively to structures on the cell surface, which then might promote viral uptake. Such an activation of the virus has been described for the human papillomavirus type 16 (HPV-16) in the presence of HP, which allowed HPV-16 infection in the absence of cell-surface GAGs [30]. Cell-surface HS is an important attachment factor for HPV-16 [31]. For AAV-2 particles, slight structural rearrangements on the viral surface have been described upon HP binding [32]. Furthermore, it has been proposed that initial structural rearrangements on the alphavirus surface occur directly after binding to the cell surface [9], based on the observation that transitional epitopes of the Sindbis virus (SINV) became accessible to antibodies upon cell binding [33,34, 35].
Replication of CHIKV was also significantly reduced in pgsA-745 cells in comparison to CHO-K1 cells at 6 hrs post infection. However, this effect was less pronounced 24 hrs post infection, although still significant. CHIKV replication in 293T cells was significantly reduced in the presence of soluble GAGs. The lack of GAG-dependency at a later time point during infection might be explained by potential direct cell-to-cell transmission, as it has been observed for CHIKV in cell culture [36], or could just be due to the saturation of infection. For example, the majority of cells could already be infected 18 hrs post-infection and the viral titer might already be at a plateau, regardless of whether the cell entry is less efficient or not. However, an enhancement of infection by soluble GAGs was not observed. This might indicate that higher concentrations of GAGs are needed, which are possibly within their toxic range.
So far, the role of GAGs in CHIKV replication has mainly been studied with regard to viral attenuation. Point mutations within the E2 domain A (e.g., E79K, G82R or E166K) have been found in attenuated vaccine strains that were cell-culture adapted and showed enhanced GAG-dependency but reduced in vivo replication [15], [16], [18]. These mutations increase the positive charge in domain A [16], and as shown here for the first time, directly affect its binding affinity. In addition, recently it has been shown that CHIKV binding to GAG receptors on mammalian cells enhances replication in those cells and cell binding was influenced by the N-glycosylation pattern of the viral envelope proteins [37]. There are two glycosylation sites in E2, one in the ß-ribbon connector (N-263) and another one in the stem region, outside the region that was analyzed here. Deletion of the ß-ribbon connector and expression of domain A in E. coli decreased domain A’s cell binding, indicating that glycosylation of N-263 may partially contribute to GAG binding. In contrast to domain A, domain B has mostly been described to be associated with covering the E1 protein and thus antibodies binding to domain B mainly inhibit the movement of domain B during fusion, but not cell attachment [25]. Our findings reveal a unique role for cell-surface GAGs during CHIKV infection, in which they are not absolutely necessary for CHIKV replication, but undoubtedly promote viral entry and replication. There is at least one GAG-independent entry pathway, as CHIKV entry into GAG-deficient cells is still possible, and soluble GAGs cannot fully block CHIKV cell entry. This additional pathway(s) consequently include different cell surface receptor(s). The proposed entry pathways are most likely mediated by different binding sites on the E2 protein. The GAG-dependent pathway would be characterized by binding of the prominently exposed domain B most likely in combination with domain A to cell-surface GAGs, inducing a conformational change of the CHIKV Env molecules which might allosterically spread to the other CHIKV Env molecules on the particle surface [9]. This activation might enable the envelope molecules to bind to a second molecule on the cell surface, possibly via domain A. Following this binding, the virus is taken up by receptor-mediated endocytosis and the further pH-induced conformational changes occur. The proposed key role for domain B is supported by experiments with chimeric viruses which revealed that domain B is the critical target of neutralizing antibodies in humans and mice [38]. The second, GAG-independent pathway in this model possibly involves direct binding of domain A to another cell-surface molecule which might not play a role in the GAG-dependent pathway.
GAGs on the cell surface are thus not absolutely required for CHIKV cell entry, but they are part of a strategy CHIKV employs in order to enter cells. They might be comparable with T-cell immunoglobulin and mucin (TIM) membrane proteins that promote, although are not absolutely required for, the cell entry of a number of viruses, including CHIKV [39]. The fact that CHIKV is able to infect a wide range of, and evolutionary only distantly related, species, and to infect many different cell types and organs within one organism [6], makes the scenario plausible, in which the virus can utilize the ubiquitously expressed GAGs for entry [40], but additionally exploits other opportunities and receptors to get into the host cell. Finally, GAG mimetics might thus be promising antiviral candidates for the treatment of CHIKV infections however they will not inhibit the GAG-independent entry pathway [41,42]. Because the E2 domains B and A contain cell binding moieties, they might be promising targets for vaccine development [23].
|
10.1371/journal.ppat.1003483 | Mechanism of HIV-1 Virion Entrapment by Tetherin | Tetherin, an interferon-inducible membrane protein, inhibits the release of nascent enveloped viral particles from the surface of infected cells. However, the mechanisms underlying virion retention have not yet been fully delineated. Here, we employ biochemical assays and engineered tetherin proteins to demonstrate conclusively that virion tethers are composed of the tetherin protein itself, and to elucidate the configuration and topology that tetherin adopts during virion entrapment. We demonstrate that tetherin dimers adopt an “axial” configuration, in which pairs of transmembrane domains or pairs of glycosylphosphatidyl inositol anchors are inserted into assembling virion particles, while the remaining pair of membrane anchors remains embedded in the infected cell membrane. We use quantitative western blotting to determine that a few dozen tetherin dimers are used to tether each virion particle, and that there is ∼3- to 5-fold preference for the insertion of glycosylphosphatidyl inositol anchors rather than transmembrane domains into tethered virions. Cumulatively, these results demonstrate that axially configured tetherin homodimers are directly responsible for trapping virions at the cell surface. We suggest that insertion of glycosylphosphatidyl inositol anchors may be preferred so that effector functions that require exposure of the tetherin N-terminus to the cytoplasm of infected cells are retained.
| The cellular restriction factor, tetherin, prevents HIV-1 and other enveloped virus particles from being disseminated into the extracellular milieu by infiltrating their envelopes and by physically crosslinking them to the cell surface. It is known that tetherin consists of pairs of membrane anchors, situated at either end of a rod-shaped molecule, but how tetherin causes virion tethering has been difficult to unambiguously determine. In this work, we develop genetic and biochemical approaches to probe tetherin molecules that have infiltrated tethered virions. We show that tetherin adopts an “axial” configuration in its functional state, with a pair of membrane anchors situated at one end of the rod-like structure inserted into a tethered virion. While either end of the rod can be inserted into a virion, there is a preference for the insertion of its lipid (glycosylphosphatidyl inositol) modified carboxyl-terminus into virion envelopes. These studies demonstrate unequivocally that the tetherin molecule itself is directly responsible for trapping virions, and dissect the molecular mechanism underpinning its antiviral activity.
| Cells have evolved numerous defense measures to inhibit the replication of infectious agents. In animal cells, sensing of viruses by pattern recognition receptors leads to interferon production and signaling, which induces the expression of hundreds of interferon-stimulated genes (ISGs) in infected and bystander cells [1]–[3]. Among these are several classes of autonomously acting proteins (the APOBEC3 proteins, TRIM5 proteins, tetherin and SAMHD1). These proteins are popularly termed “restriction factors”, and are considered to comprise an intrinsic immune system [4] or a specialized arm of conventional innate immunity. Recent efforts have revealed that these proteins directly inhibit the replication of viruses via remarkably divergent and elegant mechanisms of action [5], [6].
Tetherin (also known as BST-2, CD317, or HM1.24) is a type II membrane glycoprotein whose expression is strongly upregulated by type I interferon in most cell types. Tetherin expression causes the physical entrapment of nascent mature enveloped virions at the cell surface [7]–[11]. Structurally, tetherin comprises of a short N-terminal cytosolic tail, a single pass transmembrane helix, an extracellular domain that is predominantly alpha helical [12]–[15], and has three extracellular cysteine residues stabilizing parallel homodimer formation via disulphide bridges. Tetherin is also modified at its C-terminus by a glycosylphosphatidylinositol (GPI) membrane anchor [16], [17].
A few pieces of evidence suggest that tetherin acts directly and autonomously to trap virions at the cell surface. First, trapped virions can be liberated from the cell surface by treatment with the protease subtilisin A, indicating that protein is an essential component of the tethers [18]. In such experiments, tetherin fragments can be found in subtilisin-liberated virions [19]. Second, inactive tetherin proteins in which one of the two membrane anchors is removed are efficiently incorporated into virions [19]. Third, fluorescent and electron microscopic analyses demonstrate that tetherin is localized at sites of virion entrapment [19]–[22]. Fourth, an artificial tetherin protein assembled from heterologous protein domains that have similar configuration but no primary sequence homology to tetherin, recapitulates tetherin function [19]. Taken together these findings suggest that (i) the biological activity of tetherin can be ascribed to its overall configuration rather than its primary sequence and (ii) tetherin does not require specific cofactors or the recognition of specific viral components to cause virion entrapment. These findings are difficult to reconcile with complex models in which tetherin might act as a virion sensor to induce other factors that have tethering activity. Rather, they are more easily explained by the idea that tetherin acts autonomously and directly to trap virions, simply as a consequence of being incorporated into the lipid envelope of virions as they bud through cell membranes. Consistent with these arguments, tetherin exhibits antiviral activity against a broad spectrum of enveloped virions whose proteins have essentially no sequence homology [23]–[30].
Another argument in favor of the notion that tetherin acts rather nonspecifically to trap enveloped virions arises from the mechanisms that viruses have evolved to evade tetherin action. Rather than acquiring viral protein sequence changes that might enable escape from interaction with tetherin, viral proteins have instead adapted to gain interaction with, and thereby antagonize, tetherin. For example, the HIV-1 accessory protein Vpu interacts with the tetherin transmembrane domain [31]–[36], and employs surface downregulation [37]–[41] and degradation [32], [42]–[45] to antagonize tetherin. Additionally, the SIV Nef proteins [46]–[49], the KSHV K5 protein [27], [50], and the HIV-2 Env [38], [51], SIVMAC Env [52] and Ebola Env [53], [54] proteins have adapted to counteract tetherin proteins in their hosts by targeting different portions of the tetherin cytoplasmic tail or ectodomain.
One question that remains incompletely addressed is the precise molecular mechanisms by which tetherin exerts its antiviral activity. As discussed above, a preponderance of the evidence support a direct tethering mechanism, wherein tetherin dimers infiltrate the lipid envelope of assembling particles [19]–[21]. However, while previous biochemical analyses [19] and structural studies [12]–[15] indicate that tetherin forms a rod-like structure with membrane anchors at either end, the configuration adopted by the tetherin protein during entrapment is unknown. For example, because the membrane anchors are spatially separated from each other, it is possible that one pair of anchors partitions into the lipid envelope of assembling particles, while the other pair remains rooted in the plasma membrane of the infected cell (axial configuration, Figure 1A). In this configuration, each tetherin dimer could potentially link viral and cell membranes in either “polarity”, i.e. with N-termini inserted into either the infected cell or the assembling particle. Other obvious possibilities by which entrapment might be achieved would be via the separate partitioning of dimerized tetherin molecules into virion and cell membranes (equatorial configuration, Figure 1B) or the non-covalent oligomerization of tetherin dimers that have both pairs of anchors embedded in either virion envelopes or cell membranes (Figure 1C).
Because a protected, β-mercaptoethanol-sensitive, dimeric amino-terminal tetherin fragment can be recovered from virions that have been liberated by protease treatment, it appears that at least some trapped virions are infiltrated by both N-termini of a parallel tetherin homodimer, favoring the models shown in Figure 1A or 1C [19]. Moreover, a tetherin variant that lacks a GPI anchor preferentially localizes to sites of viral budding, suggesting that the tetherin N-terminus provides the dominant driving force for infiltration into budding virions [19]. These results have also been supported by other studies involving super-resolution microscopy [22]. Nevertheless, it remains a challenge to establish if any of the aforementioned configurations are adopted during the retention of virions, or whether the contribution of any one outweighs that of the others.
Herein, we have employed quantitative biochemical experiments and engineered tetherin proteins to demonstrate conclusively that tetherin acts directly to trap virions and to elucidate the mechanisms of virion entrapment. Specifically, we placed epitope tags and cleavage sites for the site-specific protease Factor Xa at strategic positions in the tetherin molecule. Virions that were tethered at the cell surface by these modified tetherin proteins were liberated upon specific protease treatment and analyzed. Our results demonstrate that tetherin dimers trap virions by adopting the axial configuration (Figure 1A), with either transmembrane domains or GPI anchors capable of infiltration into assembling particles. Quantitative analyses suggested that, on an average, a few dozen tetherin dimers are involved in trapping each virion and that there is a ∼3–5 fold preference for a tetherin orientation in which the GPI anchored C-terminus rather than the transmembrane domain is inserted into a tethered particle. Taken together, our biochemical experiments constitute the most compelling evidence to date that tetherin is directly responsible for trapping virions at the cell surface and that this is achieved using axially positioned tetherin homodimers, that are primarily configured with their GPI anchored C-termini inserted into virions.
In this study, we endeavored to develop biochemical assays to unequivocally determine whether tetherin acts as a direct tether in trapping virions, and to determine the configuration of tetherin dimers that are engaged in virion entrapment. A variety of approaches, including hydropathy analyses, fusion with reporter enzymes, or the insertion of target sites for proteases, antibodies and chemical modifiers, have been used to deduce membrane protein topology [55]. For example, the insertion of Factor Xa cleavage sites into hydrophilic loops has proven to be useful in such analyses [56], [57]. We adapted these approaches by engineering modified human tetherin proteins that carried (i) single cleavage sites for Factor Xa and (ii) epitopes such as hemagglutinin (HA) and FLAG tags positioned either N- or C-terminal to the Factor Xa site (Figure 1D). Previous experiences with modified tetherin proteins led to the expectation that these alterations should have no or modest effects on antiviral activity [7], [19].
Initial experiments in which Factor Xa sites alone were incorporated into Tetherin resulted in proteins that were somewhat refractory to proteolysis (unpublished observations). Hence, we reasoned that the introduction of flexible linkers into its primary sequence might facilitate access to the cleavage site, and increase the efficiency of proteolysis. Therefore, we generated a panel of proteins in which we inserted five and eight GGGGS peptide linker units into the extracellular domain of tetherin, either N-terminal (at amino acid 50) or C-terminal (at amino acid 157) to the predicted coiled-coil domain. The GGGGS peptide is predicted to be unstructured because the glycine residues impart flexibility, and the polar serine residue permits hydrogen bonding to the solvent [58], [59].
Among the panel of linker modified tetherin proteins, we determined that the proteins with eight linker units C-terminal to the coiled-coil (C8, Figure 1D) and five linker units N-terminal to the coiled-coil (N5, Figure 1D) were expressed at comparable levels to WT tetherin (Figure 2A). Note that Tetherin is heterogeneously glycosylated, and because the cells were lysed in non-reducing buffer, the tetherin proteins migrated primarily as a smear of dimeric species [19] (Figure 2A). To examine the antiviral activity of the linker-modified tetherin proteins, we co-expressed an HIV-1 proviral plasmid (HIV-1(WT)) or its Vpu-deficient counterpart (HIV-1(ΔVpu)) along with varying amounts of plasmids expressing WT tetherin or one of the modified tetherin proteins. Hereafter, WT tetherin (Figure 1D) refers to a previously described construct that harbors an HA epitope tag at amino acid 155 in the extracellular domain, but retains the antiviral activity of the untagged, endogenous protein [7]. As expected, WT tetherin potently inhibited the release of HIV-1 (ΔVpu) in a dose-dependent manner, while only marginally affecting the release of HIV-1 (WT) (Figure 2B, C). Importantly, the C8 and N5 tetherin proteins were only modestly impaired in their antiviral activity compared to WT tetherin as determined by infectious virion yield and extracellular particulate CA protein measurements (Figure 2B, C) and the levels of cell-associated Gag protein were unaffected by the expression of the tetherin proteins (Figure 2C). Thus, the insertion of linker sequences into tetherin was well tolerated with little effect on antiviral activity.
We next programmed the C8 and N5 tetherin proteins with single Factor Xa cleavage sites. The rationale was that these proteins (referred to hereafter as C8Fac and N5Fac respectively, Figure 1D) would differ in the relative ordering of the HA epitope tag and the protease site. Thus, the epitope tag is positioned N-terminal to the protease site in the C8Fac protein, whereas it is positioned C-terminal to the protease site in the N5Fac protein. In addition to the C8Fac and N5Fac proteins that carried only one epitope tag, we also appended the N-terminus of the N5Fac construct with three tandem FLAG epitope tags. This manipulation results in FLAG and HA epitope tags flanking the protease site (Flag N5Fac, Figure 1D). The use of three FLAG tags in tandem reportedly enhances signal intensity by ∼10–20-fold [60].
Analysis of the antiviral activity of the Factor Xa site-modified tetherin proteins revealed that the C8Fac and N5Fac proteins were only slightly impaired in activity relative to WT tetherin, while the Flag N5Fac protein was nearly indistinguishable in antiviral activity to WT tetherin (Figure 2B, C). The C8Fac and N5Fac proteins were expressed at slightly lower levels than the C8 and N5 proteins respectively (Figure 2A), and were proportionately impaired in antiviral activity (Figure 2B, C). Interestingly, despite harboring more tags as compared to any of the other modified tetherin proteins, the Flag N5Fac protein was virtually as potent as WT tetherin, and expressed at levels indistinguishable from WT tetherin. Vpu antagonized all modified tetherin proteins and restored the yield of extracellular virions (Figure 2B, C). Thus, all modified tetherin proteins mimicked the biological activity and Vpu sensitivity of WT tetherin.
We next generated a panel of 293T cells that stably expressed the epitope-tagged Factor Xa-cleavable tetherin proteins. The levels of cell surface tetherin in these stable cell lines was assessed by flow cytometry using a monoclonal antibody that recognizes the extracellular region of human tetherin. Importantly, the surface expression levels of the WT and modified tetherin proteins were quite similar to each other, varying over a 2.5-fold range (mean fluorescent intensities were 6200, 15000, 8800, and 12000 for WT, C8Fac, N5Fac and Flag N5Fac tetherin proteins, respectively) and were only 1.5 to 3-fold greater than that of the endogenous protein in HeLa cells, a prototype tetherin-positive cell line (mean fluorescent intensity = 5000, Figure 3A). Additionally, we verified that the engineered tetherins exhibited antiviral activity in the stable cell lines using single-cycle HIV-1 replication assays (Figure 3B). As expected, both the WT and the modified tetherin proteins inhibited the release of virions from infected cells, but did not affect cell associated Gag protein expression (Figure 3B). Also, the expression of Vpu reversed the inhibitory effect of the modified tetherin proteins (Figure 3B).
Our previous studies have employed a protease “stripping” assay [18], [61] in which a relatively nonspecific protease (subtilisin A) was used to demonstrate that tetherin causes virions to become entrapped on cell surfaces by a protein based tether. The logic underpinning the assay described herein was that if the tetherin protein itself functions as the direct tether, then treatment of cell surfaces with a specific protease (Factor Xa) would trigger the release of virions, only when tetherin was programmed with a Factor Xa cleavable site (Figure 4A). Moreover, cleavage should result in partitioning of the epitope-tagged proteolytic fragments either into the liberated virions or the infected cells. Because the epitope tags were strategically positioned relative to the protease site, topological information could be deduced about tetherin in its functional state (Figure 4B). However, because we expect that only a minority of the tetherin molecules on the cell surface would actually be involved in tethering virions, only fragments that are found in virions should be regarded as informative with respect to tetherin topology during virion entrapment.
Note that if tetherin adopts the equatorial configurations depicted in Figure 1B or 1C then we would not expect Factor Xa cleavage to result in virion release, because the cleavage sites are positioned outside the region of tetherin-tetherin interaction, in the rod like portion of the molecule. Indeed, the cleavage sites are positioned in artificially introduced sequences whose insertion did not perturb tetherin function (Figure 2, 3). Conversely, if tetherin adopts the axial configuration in virion tethers, then Factor Xa cleavage should result in virion release. Moreover, if as depicted in Figure 4B, the HA-tagged proteolytic fragments partition with virions that are liberated from Factor Xa-treated, C8Fac-expressing cells, it would suggest that tetherin dimers exist with their N-termini inserted into the interior of the virion. Conversely, if HA-tagged proteolytic fragments partition with virions that are liberated from the N5Fac cell line, we would deduce that tetherin dimers exist with their GPI anchors embedded in the virion membrane. If, however, tetherin dimers adopt both polarities, then HA-tagged proteolytic fragments would be observed in virions liberated by Factor Xa from both C8Fac and N5Fac expressing cell lines.
We first investigated the utility of this approach using cell lines expressing the single epitope tagged C8Fac and N5Fac tetherin proteins. Cells were infected with single-cycle, Vpu-deficient HIV-1, and constitutively released particles were harvested from culture supernatants. Thereafter, the monolayer of cells was treated with Factor Xa, and then the cell lysates and any liberated virions were also harvested. As before, infected tetherin-negative control cells constitutively released comparatively high levels of virions into the culture supernatant, while virion yield from cells expressing WT, C8Fac or N5Fac tetherin proteins was substantially reduced (Figure 4C). The levels of HIV-1 Gag expression in cell lysates were uniform (Figure 4C).
Incubation in Factor Xa cleavage buffer alone resulted in the release of only low levels of pelletable CA from tetherin-deficient cells. This may have represented virion particles that were constitutively released during incubation, or virions that were loosely adhered to the cell surface (Figure 4C). Even lower levels of particles were released from cells expressing the WT, C8Fac or N5Fac tetherin proteins that were incubated in Factor Xa cleavage buffer alone. Strikingly however, Factor Xa treatment of the C8Fac and N5Fac resulted in the release of substantial amounts of particulate CA (Figure 4C). Crucially, Factor Xa treatment of tetherin-negative or WT tetherin expressing cells did not increase particle release over the low background levels that were observed in the absence of protease, underscoring the strict requirement for a Factor Xa-cleavable tetherin in Factor Xa-induced virion release (Figure 4C). Notably, proteolytic fragments of tetherin were observed in virions released by Factor Xa from both C8Fac and N5Fac expressing cells and these virion-associated fragments were consistent with the incorporation of tetherin dimers therein. These dimers were the only tetherin species that were detectable on non-reducing SDS PAGE gels (Figure 4C).
Because tetherin is intrinsically heterogeneous, due to variable glycosylation as well as dimer formation, it was difficult to assess the extent of Factor Xa cleavage in cell lysates (Figure 4C, center panel), or to unambiguously demonstrate that only cleaved tetherin fragments were present in Factor Xa liberated virions (Figure 4C, bottom panel). Therefore we treated cell and virion lysates with PNGase-F and repeated the western blot analyses under reducing conditions. We observed that the HA-tagged proteolytic fragments (predicted molecular weights of ∼20.8 kDa and ∼17 kDa for C8Fac and N5Fac respectively) could be resolved from the full-length molecules (∼24.7 kDa and ∼23.8 kDa for C8Fac and N5Fac respectively) (Figure 4D). This analysis revealed that about half of the cell-associated C8Fac and N5Fac protein was cleaved by Factor Xa that was applied to the cell surface. The incomplete cleavage may have been due to the intracellular localization of a fraction of the tetherin protein. As expected, no proteolysis of the WT tetherin protein was observed (Figure 4D). Notably, only the cleaved tetherin protein was found in PNGase-F-digested virion lysates, consistent with the notion that tetherin cleavage by Factor Xa was necessary for virion release in this assay. To assess the efficiency of tetherin cleavage and virion release by Factor Xa, we compared the levels of virion released from C8Fac and N5Fac expressing cell lines following treatment with Factor Xa or with subtilisin A (Figure 4E). Similar amounts of virions were released by the site-specific and non-specific proteases. This result suggested that tetherin cleavage and virion release caused by Factor Xa was quite efficient. It also suggested that it was unlikely that a significant fraction of virions are retained using alternative configurations of tetherin (Figure 1) in which virion release might be resistant to Factor Xa treatment.
Overall, these results strongly suggested that tetherin traps virions by adopting the axial configurations depicted in Figure 1A. Moreover, because HA-tagged proteolytic fragments from both C8Fac and N5Fac tetherin proteins partitioned with virions these data suggested that both polarities depicted in Figure 1A are adopted by tetherin during virion entrapment.
To estimate the number of tetherin dimers that were involved in the entrapment of a single virion, we used a quantitative western blotting approach and PNGase-F-digested virion lysates to measure the relative number of CA and HA epitopes associated with virions that had been tethered by the C8Fac and N5Fac proteins, and then released by Factor Xa cleavage. First, we generated an appropriate internal standard protein to enable relative quantitation. This standard consisted of a fusion protein that comprised the HIV-1 p24CA protein, appended at its C-terminus with three tandem FLAG tags and an HA epitope tag. Thus, this single protein included each of the epitopes that we planned to probe, at a stoichiometric ratio of 1∶1∶1 and could be used as a standard to compare the relative numbers of HA and CA epitopes in tethered virions liberated from C8Fac and N5Fac expressing cells. Specifically, serial dilutions of cell lysates expressing the HA-Flag-CA protein were run on SDS-PAGE gels, blotted onto membranes and probed with antibodies against CA and HA. The band intensities were analyzed using a LiCOR Odyssey scanner (Figures 5A, B), and regression analysis was performed over the linear range of signal intensities (Figures 5A, B). Dilutions of the PNGase-F-treated virion lysates recovered from C8Fac and N5Fac expressing cells that also yielded band intensities in the linear range of the assay were resolved on the same gel as the standard, and the relative amounts of CA and HA epitope in each samples were deduced by interpolation using the standard curves (Figures 5A, B).
HIV-1 virions have been reported to contain between 1000–5000 copies of the Gag protein, of which only a fraction contribute to core formation [62]–[66]. We calculated our estimates of tetherin dimers per virion based on the extremities of this range (Table 1). Thus, if each virion contains 1000 CA protein molecules, we estimate that 16±5 dimers of the N-terminus of C8Fac and 71±26 dimers of the C-terminus of N5Fac tetherin dimers were associated with a single tethered virion (Table 1). Conversely, if a single virion contains 5000 CA epitopes, then we estimate that 80±25 dimers of the N-terminus of C8Fac and 355±130 dimers of the C-terminus of N5Fac tetherin dimers were associated with a single tethered virion. Thus these numbers suggested a preference (∼4 to 5-fold) for the insertion of the GPI-anchored tetherin C-terminus, rather than the N-terminal transmembrane domain into virions. Note that the larger number of HA tags associated with virions in the case of N5Fac cannot be explained by differences in tetherin expression levels. In fact, there were lower levels of N5Fac on cell surfaces (MFI = 8800, Figure 3A) as compared to the C8Fac protein (MFI = 15000, Figure 3A).
The aforementioned experiments indicated that tetherin directly tethers HIV-1 particles in an axial configuration (Figure 1A) and suggested that both polarities, with either N- or C- termini inserted into virions contribute to antiviral activity. However, it was possible that the two different estimates for the numbers of tetherin molecules inserted into virions with each polarity might reflect intrinsic differences in the properties of the two different tetherin molecules used (C8Fac and N5Fac). Therefore, we quantitated tetherin insertion into virions in a second set of experiments employing a single tetherin species with two different epitope tags on either side of the Factor Xa cleavage site (Flag N5Fac, Figure 1D, Figure 6A). Additionally, we have previously found that virions that accumulate on the surface of cells as a result of tetherin action can sometimes be tethered to each other as well as to the cell surface. This scenario could be the result of virion assembly at sites on the cell surface already occupied by trapped virions and would result in both ends of a tetherin molecule being associated with virions. These events would tend to reduce any indication that tetherin N-or C-termini are preferentially inserted into virion envelopes. Because the accumulation of virions should exacerbate this effect over time, we treated the surface of cells expressing Flag N5Fac with Factor Xa at predetermined time intervals following infection with HIV-1ΔVpu, and quantified HA- and FLAG-tagged proteolytic fragments in liberated virions.
The HIV-1 Gag protein became detectable in infected Flag N5Fac-expressing cell lysates at ∼24 h after infection and levels progressively increased with time thereafter (Figure 6B). Treatment of these infected cells with Factor Xa resulted in a time dependent increase in the amount of recovered virions (Figure 6B). The Factor Xa site is positioned N-terminal to the sites of N-linked glycosylation as well as to the extracellular cysteines in the Flag N5Fac molecule (Figure 1D, Figure 6A) and so the Factor Xa cleavage of the 65–70 kDa dimeric, glycosylated Flag-N5Fac protein yields a cell associated dimeric, glycosylated ∼50–55 kDa αHA reactive species as well as a cell associated monomeric, nonglycosylated 10 kDa α-FLAG reactive species (Figure 6B). Notably, both the dimeric glycosylated ∼50–55 kDa αHA reactive species and the 10kDa α-FLAG reactive species were observed in virions, and their levels in the virion fraction increased with time, approximately in parallel with the increasing yield of Factor Xa liberated virions (Figure 6B). Notably, the N-terminal FLAG tagged fragment of Flag N5Fac was also found in virions in a form that was consistent with the formation of dimers. We hypothesize that this is because the tetherin cytoplasmic tail contains two cysteines that can form disulphide bonds in the interior of virions. Consistent with this idea, only the smaller of the two Flag tagged species was observed when virion lysates were subjected to SDS PAGE gel electrophoresis under reducing conditions (Figure 6C). Additionally, the dimeric glycosylated ∼50–55 kDa αHA reactive species collapsed to a single ∼17 kDa band when samples were deglycosylated with PNGase and reduced (Figure 6C).
We used quantitative western blot analyses of PNGase-F-digested virion lysates to estimate the number of copies of HA- and FLAG-tagged proteolytic fragments per trapped virion (Figure 6C, D). Again we used the FLAG-HA-CA protein as a standard to determine the relative numbers of HA, FLAG and CA epitopes in the virions liberated from Flag N5Fac expressing cells. Although tethered virions could be recovered from the surface of Flag N5Fac expressing cells beginning at 24 h after infection, we could not make reliable estimates of the HA and FLAG fragments at this time point, as they were present at levels that were close to the limit of detection. However, we could make reasonably robust estimates of the levels of incorporation of HA- and FLAG-tagged fragments into virions beginning at 32 h after infection.
Importantly, the number of FLAG-tagged dimers that were estimated to be present in virions (assuming 1000 CA molecules per virion) tethered by Flag N5Fac (11±3 [at 32 h] to 16±6 [at 48 h], Figure 6C, D, Table 2) correlated quite well with the number of HA-tagged dimers present in virions tethered by C8Fac (16±5 [at 48 h], Table 1). Similarly, the number of HA-tagged dimers in tethered virions recovered from the Flag N5Fac expressing cells (34±18 [at 32 h] to 55±28 [at 48 h]) (Figure 5D, Table 2) correlated quite well with the number of copies of HA-tagged dimers in tethered virions recovered from the N5Fac expressing cells (71±26 at 48 h) (Table 1). Notably, we estimated that the virions liberated from Flag N5Fac expressing cells carried ∼3 to 4-fold more HA tags than FLAG tags, again suggesting that axially configured tetherin dimers infiltrate assembling particles, with a tendency to embed their C-termini rather than their N-termini in tethered virions. Also noticeable was a marginal trend for the appearance of increasing numbers of tetherin molecules per virion over time. This trend was not statistically significant and could be due to some unknown bias in the measurements. However, it is also possible that virions with smaller numbers of tetherin molecules are more readily released, leading to the preferential accumulation of virions with greater numbers of tetherin molecules on the surface of cells. Finally, to confirm that alternative Factor Xa-resistant configurations of Flag N5Fac tetherin were not responsible for retaining a significant fraction of virion particles, we compared the levels of virions released from Flag N5Fac expressing cells by Factor Xa or by subtilisin A treatment. Similar amount of particles were released by the two proteases, suggesting that axially configured Flag N5Fac tetherin molecules were the major form responsible for virion retention (Figure 6E).
We devised a biochemical approach to probe tetherin molecules that have infiltrated virions at the cell surface, with the goal of elucidating the configuration adopted by tetherin during virion entrapment. This approach was based on two previous findings. First, a non-specific protease, subtilisin, could be used to liberate tethered particles from the infected cell's surface [18], [61]. Second, the primary sequence of tetherin can be drastically altered while retaining biological activity [19]. Thus, we employed the site-specific protease Factor Xa to liberate virions trapped by tetherin molecules that were engineered to include its cleavage site. This manipulation gave the approach tight specificity and enabled the unequivocal demonstration that the tetherin protein itself is an essential component of virion tethers. Moreover, the use of a site specific protease to release tethered virions from cell surfaces enabled the preservation of epitope tags inserted into the tetherin ectodomain, allowing us to infer the organization of tetherin molecules in virion tethers. We could use a double epitope-tagged version of tetherin, as well as single epitope-tagged versions to analyze the incorporation of both N- and C-terminal proteolytic fragments into virions, and thereby determine tetherin configuration. Additionally, we constructed a protein standard and performed quantitative western blotting to estimate the numbers of tetherin dimers in each orientation that are associated with trapped virions.
Because virions were efficiently liberated by Factor Xa treatment of N5Fac or C8Fac expressing cells, our data effectively exclude the “equatorial” configuration shown in Figure 1B, as cleavage of the tetherin peptide backbone in this context would leave intact the majority of the bonds holding the virion on the cell surface. Moreover, the fact that tetherin fragments found in virions liberated by Factor Xa were exclusively disulphide linked homodimers also constitutes strong evidence disfavoring this model. While our data do not completely discount the possibility that tetherin multimers adopt the equatorial configuration, with virions becoming trapped via hypothetical noncovalent dimer-dimer interactions (Figure 1C) this scenario appears unlikely for two reasons. First, such a configuration would not be expected to result in virion release upon Factor Xa cleavage, because dimer-dimer interactions would not be expected to be perturbed, particularly since the Factor Xa cleavage site is placed within a foreign spacer sequence whose insertion does not itself perturb tetherin function. Second, the scenario envisaged in Figure 1C would result in precisely equal numbers of tetherin N- and C termini being placed in tethered virions. We found that there were modestly, but clearly, more tetherin C-termini than N-termini in virions, arguing that tetherin N- and C-termini partition separately into virion and cell membranes. Overall our experiments indicated that tetherin homodimers adopt an axial configuration in their functional state, with a preference for the insertion of their GPI-anchored C-termini into virions during their entrapment at the surface of infected cells. Quantitative analysis indicated that an average of ∼80 to 400 tetherin dimers (depending on how many CA molecules are assumed to be present in each virion) associated with each tethered particle. Our findings do not discount the discount the possibility that higher order tetherin multimers, e.g. tetramers, might contribute to tethering, but if such complexes do exist, then they must involve non-covalent interactions between axially configured tetherin molecules and be arranged in such a way that all N-termini and in one membrane (be it virion envelope or cell membrane) and all C-termini are in the opposing membrane.
Previous studies have not resolved the configuration adopted by tetherin during virion entrapment. For example, conflicting results have been obtained in studies where the release of virions was attempted by cleavage of the tetherin GPI anchor using phosphatidyl-inositol-specific phospholipase C (PI-PLC). In one study, the efficiency of virion release induced by PI-PLC treatment was poor (∼20% compared to subtilisin) [22], while other studies indicated that PI-PLC treatment fails to liberate any virions [20], [67]. Second, the failure of reducing agents to release virions would tend to suggest that the equatorial model shown in Figure 1B is incorrect [20]. However, this argument is somewhat confounded by the fact that tetherin molecules are twisted around each other in a dimer, and so breaking the disulphide bonds in an already-formed tether would not necessarily be expected to cause virion release.
One caveat of our assay is that some tetherin dimers might infiltrate particles and yet be uninvolved in restriction. Thus, it is possible that the number of tetherin molecules that we measured to be associated with a virion might be greater than the number of molecules actually involved in virion entrapment. Indeed, previous studies have shown that low levels of complete tetherin molecules can be found in the small number of virions that are released from tetherin-positive cells [20]. However, to be uninvolved in restriction would require that both tetherin N- and C-termini were embedded in virions. If the numbers of tetherin dimers that were inserted into virions in this way was in excess of the numbers of tetherin dimers involved in tethering, with N- and C-termini partitioned separately into virion and cell membranes, then there would be little or no difference in the number of tetherin N- and C-termini found in virions. The fact that we do indeed observe a 3-to 5-fold excess C-termini in tethered virions, argues strongly that most of the tetherin molecules (at least 65–80%) that are tethered-virion associated, have their N- and C-termini separately partitioned into virion and cell membranes. Thus most tetherin molecules must be in the axial configuration with only their C-termini embedded in virions.
Our estimates of the number of tetherin molecules that are associated with tethered virions are several-fold higher than those obtained using super-resolution microscopy approaches (i.e. 4–7 dimers per virion) [22]. At least three factors could account for this discrepancy. First, the microscopy studies use a tetherin-mEosFP fusion protein, that includes a bulky 230 amino acid (∼26 kDa) protein at its N-terminus, appended to the otherwise short (21 amino acid) native tetherin cytoplasmic tail. This could very easily reduce the numbers of tetherin molecules that associate with virions. Second, the estimates made in the microscopy studies correspond to groups of tetherin molecules present at the same location as clusters of Gag molecules that may not represent completely assembled virions. Thus, microscopy studies cannot determine whether the imaged tetherin molecules are in the act of restriction. Conversely, our estimates are based on bona fide tethered virions that are recovered from cells by specific cleavage of the tether. Finally, the cell lines that we used to derived our estimates modestly overexpressed tetherin (1.5- to 3- fold) as compared to HeLa cells, which might have slightly elevated the numbers of tetherin molecules that were associated with virions. In previous studies [22], transfected HeLa cells were used, and the levels of tetherin-mEosFP relative to preexisting endogenous tetherin, or the total (endogenous plus exogenous) levels of tetherin expression were not determined, which could lead to underestimates or overestimates of tetherin association with tethered virions.
Given that virions are trapped not only at the cell surface, but are also linked to each other, it should be expected that both tetherin N- and C-termini would be found in virions. Most likely, the appearance of virions tethered to each other results from the assembly of a virion at a location on the plasma membrane already occupied by a tethered particle. This being so, our finding of a 3- to 5-fold preference for the insertion of C-termini rather than N-termini into virion membranes may represent an underestimate of the true preference. If this is the case, then one might expect that the apparent preference for the insertion of C-termini into virions would become less apparent over time as virion accumulate at the cell surface and the likelihood of a virion assembly at a site already occupied by a tethered virion increased. However, we did not observe such a trend, and thus it remains unclear whether the 3- to 5-fold preference for C-terminus insertion into virions is an accurate number, or an underestimate resulting from virion accumulation.
The biophysical mechanism underpinning the apparent preference for the insertion of GPI-anchored C-termini over TM domain anchored N-termini into virions is unclear at present. Although it is not the predominant scenario, the tetherin N-terminal domain is clearly capable of being incorporated into virions. Indeed, a tetherin molecule lacking the GPI anchor is efficiently incorporated into released virions [19]. Moreover, it is the N- terminus that is targeted by Vpu to block tetherin incorporation into virions [19], [33]. Perhaps the tetherin N-terminal domain acts as a sensor of membrane curvature, driving localization to assembly sites, but the GPI anchor diffuses more freely into virion membranes. Consistent with this idea, recent work indeed indicates that tetherin colocalizes better with HIV-1 Gag proteins that cause membrane curvature than those which do not [68].
There is potential biological utility in preferentially inserting GPI anchored tetherin C-termini rather than N-termini into virions. In such a scenario, the tetherin N-terminus remains available to the cytoplasm of the infected cell, from where it may execute important functions. For instance, virions trapped at the cell surface are internalized and degraded in lysosomes [61], [69]. Moreover, human tetherin appears capable of initiating signaling cascades, particularly when it is engaged in tethering, and in some respects may act as a virion sensor [70]–[72]. Thus, the need to interact with the endocytic machinery and/or initiate signaling might favor a scenario in which tetherin dimers are oriented with their N-termini in the infected cell and their C-termini in the virion membrane.
Tetherin was transiently expressed using pCR3.1 (Invitrogen) based plasmids or stably expressed using pLHCX (Clontech) based retroviral vectors. A human tetherin protein internally tagged with an HA epitope at amino acid 155 and, expressed using pCR3.1 or LHCX vectors, has been described previously [7]. Eight copies of a peptide linker sequence, each comprising the amino acid sequence GGGGS, were inserted immediately C-terminal to the HA tag, to generate the C8 modified tetherin protein (Figure 1D). Similarly, five GGGGS linker units were inserted immediately C-terminal to the tetherin transmembrane domain at amino acid position 50, to generate the N5 modified tetherin protein. Because the BamHI recognition site (GGATCC) encodes a glycine and serine, we incorporated its sequence into the fourth and third linker units for the C8 and N5 proteins respectively. We then used these BamHI sites for the subsequent insertion of a Factor Xa cleavage site (IEGR) to generate the C8Fac and N5Fac proteins (Figure 1D). Thereafter the Flag N5Fac protein was generated by inserting three copies of a FLAG epitope tag at the N-terminus of the N5Fac protein (Figure 1D).
The protein standard used for quantitative western blotting was generated by appending the C-terminus of HIV-1 p24 CA protein with three FLAG epitope tags and an HA epitope tag. Specifically, the p24 CA coding sequence was amplified from the proviral plasmid pNL4-3 using oligonucleotides that encoded the epitope tags, and inserted as an EcoRI-NotI fragment into the multiple-cloning site of pCRV-1, a previously described hybrid expression vector [73] that is derived from pCR3.1 and from a highly modified HIV-1 provirus (V1B). All mutagenesis was accomplished by using overlap-extension PCR.
Human embryonic kidney (HEK) 293T cells and HeLa-TZM cells expressing CD4/CCR5 and a LacZ reporter gene under control of the HIV-1 LTR were maintained in Dulbecco's Modified Eagle Medium (DMEM) supplemented with 10% FBS and gentamycin (2 µg/ml, Gibco). HEK293T cells were transduced using pLHCX based retroviral vectors expressing genes of interest and selected with hygromycin (50 µg/ml) (MediaTech, Inc) to generate cell lines expressing either the empty vector or epitope-tagged WT or modified tetherin proteins.
The 293T cells stably expressing the modified tetherin proteins and HeLa cells were harvested in PBS plus 5mM EDTA, washed in FACS buffer (PBS plus 2% BSA), and stained with PE anti-human CD317 (tetherin) antibody (Biolegend). Dead cells were excluded by DAPI staining. All data were acquired on an LSR II flow cytometer (Becton Dickinson), and data were analyzed with FlowJo software (Tree Star).
A HIV-1 proviral plasmid that expresses green fluorescent protein (GFP) in place of Nef has been described previously [74]. 293T cells were seeded in 10 cm plates at a concentration of 3×106 cells/plate and were cotransfected the following day using polyethylenimine (PolySciences) with 10 µg of wild-type (HIV-1(WT)) or Vpu-deficient (HIV-1(ΔVpu)) GFP reporter plasmids, along with 1 µg of a VSV-G expression plasmid. The culture medium was replaced the following day. At 48 hours post transfection, the culture supernatants were harvested, clarified by centrifugation at 3000 rpm, and filtered through a 0.2 µm PVDF membrane (Millipore). The viruses were stored at -80°C. Infectious virus titers were determined by inoculating sub-confluent monolayers of 293T cells that were seeded in 96 well plates at 30,000 cells/well with 100 µl of serially diluted supernatants. At 48 hours post infection, the cells were dispersed with trypsin, fixed in 4% paraformaldehyde and analyzed by flow cytometry.
293T cells were seeded in 24-well plates at a concentration of 2×105 cells/well and were cotransfected the following day using polyethylenimine (PolySciences) with 350 ng of wild-type (HIV-1(WT)) or Vpu-deficient (HIV-1(ΔVpu)) proviral plasmids along with varying amounts of a Tetherin expression plasmid (25 ng to 100 ng) and a plasmid expressing YFP (75 ng), to monitor transfection efficiency. In all transfection experiments, the total amount of DNA was held constant by supplementing the transfection with an empty expression vector. The culture medium was replaced the following day. At 48 hours post transfection, the culture supernatants were harvested, clarified by centrifugation at 3000 rpm, and filtered through a 0.2 µm PVDF membrane (Millipore). Infectious virus yield was determined by inoculating sub-confluent monolayers of HeLa-TZM cells that were seeded in 96 well plates at 10,000 cells/well with 100 µl of serially diluted supernatants. At 48 hours post infection, β-galactosidase activity was determined using GalactoStar reagent, in accordance with the manufacturer's instructions (Applied Biosystems). Physical particle yield was determined by layering 700 µl of the virion containing supernatant onto 1 ml of 20% sucrose in PBS followed by centrifugation at 20,000×g for 90 minutes at 4°C. Virion pellets were then analyzed by Western blotting.
Cells (HEK293T) stably expressing WT or engineered tetherin proteins were infected with VSV-G-pseudotyped HIV-1(WT) or HIV-1(ΔVpu) GFP at 1 infectious unit per cell in 10 cm dishes. The inoculum was removed 6 h later. At 48 hours post transfection, the culture supernatants were harvested, clarified by centrifugation at 3000 rpm, and filtered through a 0.2 µm PVDF membrane (Millipore). Physical particle yield was determined as outlined above. Simultaneously, the cells were washed with Factor Xa reaction buffer (20 mM Tris·Cl, pH 6.5; 50 mM NaCl; 1 mM CaCl2) and incubated with 50 µg of Factor Xa in 5 ml of Factor Xa reaction buffer for 2 hours at 37°C. Alternatively, the cells were washed with with subtilisin A buffer (10 mM Tris ,pH 8.0; 1 mM CaCl2; 150 mM NaCl), and treated with 5 ml of 1 µg/ml of subtilisin A (Sigma) for 3 min at room temperature. Subtilisin treatment was stopped using DMEM containing 10% FCS, 5 mM PMSF, and 20 mM EGTA. Thereafter, the supernatants were centrifuged, filtered and virions pelleted as described above, and the cells were lysed for analysis of viral protein expression by Western blotting.
Lysates of cell and liberated virions were denatured with 0.5% SDS at 100°C for 10 minutes and then treated with 1% NP-40 to neutralize the SDS. The lysates were incubated with (or without) 500 U of peptide-N-glycosidase-F (New England Biosciences) at 37°C for 3 hours. Thereafter, the reactions were quenched with SDS-PAGE loading buffer and the samples were analyzed with western blotting.
Pelleted virions and cell lysates were resuspended in SDS-PAGE loading buffer, in the presence or absence of β-mercaptoethanol, and resolved on NuPAGE Novex 4–12% Bis-Tris Mini Gels (Invitrogen) in MOPS running buffer. Proteins were blotted onto nitrocellulose membranes (HyBond, GE-Healthcare) in transfer buffer (25 mM Tris, 192 mM glycine). The blots were then blocked with Odyssey blocking buffer and probed with mouse anti-HIV-1 capsid (NIH), rabbit anti-HA (Rockland), and mouse anti-FLAG (Sigma) primary antibodies. For quantitative western blotting, the bound primary antibodies were detected using fluorescently labeled secondary antibodies (IRDye 800CW Goat Anti-Mouse Secondary Antibody, IRDye 680LT Goat Anti-Rabbit Secondary Antibody and IRDye 680LT Goat Anti-Mouse Secondary Antibody; LI-COR Biosciences). Fluorescent signals were detected using a LI-COR Odyssey scanner and quantitated with Odyssey software (LI-COR Biosciences).
|
10.1371/journal.pgen.1005243 | The Centrosomal Linker and Microtubules Provide Dual Levels of Spatial Coordination of Centrosomes | The centrosome is the principal microtubule organizing center in most animal cells. It consists of a pair of centrioles surrounded by pericentriolar material. The centrosome, like DNA, duplicates exactly once per cell cycle. During interphase duplicated centrosomes remain closely linked by a proteinaceous linker. This centrosomal linker is composed of rootletin filaments that are anchored to the centrioles via the protein C-Nap1. At the onset of mitosis the linker is dissolved by Nek2A kinase to support the formation of the bipolar mitotic spindle. The importance of the centrosomal linker for cell function during interphase awaits characterization. Here we assessed the phenotype of human RPE1 C-Nap1 knockout (KO) cells. The absence of the linker led to a modest increase in the average centrosome separation from 1 to 2.5 μm. This small impact on the degree of separation is indicative of a second level of spatial organization of centrosomes. Microtubule depolymerisation or stabilization in C-Nap1 KO cells dramatically increased the inter-centrosomal separation (> 8 μm). Thus, microtubules position centrosomes relatively close to one another in the absence of linker function. C-Nap1 KO cells had a Golgi organization defect with a two-fold expansion of the area occupied by the Golgi. When the centrosomes of C-Nap1 KO cells showed considerable separation, two spatially distinct Golgi stacks could be observed. Furthermore, migration of C-Nap1 KO cells was slower than their wild type RPE1 counterparts. These data show that the spatial organization of centrosomes is modulated by a combination of centrosomal cohesion and microtubule forces. Furthermore a modest increase in centrosome separation has major impact on Golgi organization and cell migration.
| During most of interphase, the two centrosomes of a cell are kept together by a proteinaceous linker, called the centrosomal linker. It is clear that the linker has to be dissolved by Nek2 kinase and other mechanisms before mitosis in order to assemble a functional bipolar mitotic spindle. Yet the relevance of the centrosome linker for cell function during interphase is not understood. Here we describe for the first time the analysis of a knockout (KO) cell line that lacks an essential component of the centrosome linker, C-Nap1. We observed that centrosomes in these cells are devoid of linker proteins and Nek2 kinase whereas other centrosomal proteins localize to centrosomes as in wild type cells. On average the centrosome distance is moderately increased in C-Nap1 KO cells from 1 to 2.5 μm. We further show that the centrosomal linker is only one element that positions centrosomes close to each other in interphase cells. In linker deficient cells, microtubules spatially organize centrosomes. This resolves a long discussed issue on the role of microtubules in centrosome cohesion. Moreover, we observed that linker deficient cells mis-organize the Golgi. Furthermore, migration of C-Nap1 KO cells was slower than their wild type RPE1 counterparts.
| The centrosome is the principal microtubule organizing center (MTOC) in most animal cells. By nucleating and anchoring microtubules, the centrosome influences microtubule directed processes including shape, polarity, organelle transport, adhesion, motility and division of cells [1]. Centrosomes consist of the centrioles and the pericentriolar material (PCM) that has microtubule nucleation activity [2].
In telophase/G1 the two perpendicularly joined centrioles become separated by the activities of polo kinase and separase [3,4]. Simultaneously, a proteinaceous linker, called the centrosomal linker, assembles at the proximal end of the two centrioles and keeps them connected [5]. In G1/S phase, each of the two linked centrioles initiate the process of duplication at the end of which the cell has two centrosomes each with two centrioles. The two centrosomes remain connected by the centrosomal linker [6] until the onset of mitosis when the centrosomal linker is dissolved [7–9]. This enables the two centrosomes to organize the poles of the mitotic spindle and to segregate the chromosomes. Since the two centrosomes are closely connected in interphase by the centrosomal linker, it was suggested that they function as a single MTOC [7].
At the molecular level, several proteins have been shown to play a role in the assembly and disassembly of the centrosomal linker. C-Nap1 acts as a docking site for all linker proteins at the proximal end of centrioles [7,10–14]. The protein rootletin forms filaments that physically connect the two centrosomes [14,15]. Recently, Cep68, LRRC45 and centlein were identified as structural components of the centrosomal linker [11–13]. At the onset of mitosis, enhanced activity of polo kinase Plk1, a major mitotic kinase, activates Nek2A through the Ste20-like kinase Mst2 that directs Nek2A to centrosomes [16,17]. Epidermal growth factor (EGF) also recruits Nek2A to centrosomes and so regulates linker dissolution in a mode of control that is linked to external cues [18]. In addition, cyclin B2 overexpression and p53 transcriptional activity split centrosomes prematurely by activating the Plk1-Mst2-Nek2A pathway [19].
At centrosomes, Nek2A phosphorylates C-Nap1, rootletin and other linker components [7,11,15]. This phosphorylation leads to linker disassembly without degradation of its components. In contrast, phosphorylation of Cep68 by Plk1 in prometaphase triggers proteolytic degradation of Cep68 by the E3 enzyme βTrCP and the proteasome [20]. Cep68 degradation seems to be mainly important for removal of the protein CEP215/CDK5RAP2 from the PCM but not for linker dissolution. After the linker is dissolved, the centrosomes migrate away from one another as a consequence of anti-parallel microtubule sliding forces of the kinesin-5 motor protein Eg5/Kif11 and HKlp2/Kif15 [21–23].
Most studies to date focused on identifying components that are important for centrosomal linker structure and regulation, yet the function of the centrosomal linker remains elusive. Here, we have constructed RPE1 linker deficient cells by inserting a premature stop codon together with the neomycin resistance gene as a selection marker into exon 15 of CEP250 gene (coding for C-Nap1). These C-Nap1 knockout (KO) cells were used to assess the functional importance of the centrosomal linker at the cellular level. C-Nap1 KO cells exhibited normal chromosome segregation and surprisingly the average distance between centrosomes was only modestly increased, suggesting an additional level of centrosomal organization. Microtubules but not dynein or actin were an additional organizing component of inter-centrosome positioning. Nevertheless, the moderate increase in centrosome distance in C-Nap1 KO cells was sufficient to affect Golgi organization and reduce the speed of cell migration.
The importance of the centrosomal linker that connects the two centrosomes during interphase is not understood. To gain insights into the cellular function of this connection, we disrupted both copies of the CEP250 gene (encoding for C-Nap1) in hTERT-immortalized retinal pigment epithelial cells (RPE1). C-Nap1 was chosen because all known linker proteins are dependent on this protein for centriole binding [7,10–14]. We employed a zinc finger nuclease (ZFN)-induced homologous recombination strategy in combination with a neomycin resistance donor construct to insert a premature stop codon into the C-Nap1 open reading frame (Fig 1A). The ZFN strategy produced rare random integrants (Fig 1B, clone 19), but mostly single (clone 22 and 23) and double CEP250 neomycin integrants that targeted both copies of exon 15 (named C-Nap1 KO cells) (Fig 1B; clones 7, 17 and 18 are independent clonal cell lines). RT-PCR analysis confirmed the absence of wild type (wt) C-Nap1 mRNA in clones 7, 17, and 18 (Fig 1C). At the protein level, we analyzed C-Nap1 KO clones by immunobloting (Fig 1D). We could observe that the single allele knockout clones 22 and 23 expressed full length C-Nap1 at lower levels compared to the RPE1 wt cell line. Furthermore, full length C-Nap1 was undetectable in the C-Nap1 KO clones 7, 17, and 18. However, the C-Nap1 antibody that is directed against the N-terminus of the protein detected the predicted N-terminal C-Nap1 fragment of 65 kDa in the single (Fig 1D; clones 22 and 23) and double allele knockout cell lines (clones 7, 17, and 18). Using the same C-Nap1 antibody, the full-length protein was detected by indirect immunofluorescence at centrosomes of RPE1 wt cells. Yet we did not observe a centrosomal or another defined signal in the C-Nap1 KO clones 7, 17, and 18 (Fig 1E). Thus, the N-terminal C-Nap1 fragment of C-Nap1 KO cells was unable to provide centrosomal linker function and was most likely dispersed in the cytoplasm. This analysis confirmed that our strategy has generated cell lines that lack functional C-Nap1 protein. Clones 7, 17, and 18 were used in further experiments to address the function of C-Nap1 and the overall function of the centrosomal linker.
siRNA depletion of C-Nap1 impairs localization of all other linker proteins [12]. Consistent with these data, rootletin and Cep68 were no longer associated with centrosomes in two independent C-Nap1 KO cell lines (Fig 2A, S1A and S1B Fig). However, we noticed cytoplasmic filament-like assemblies of rootletin and Cep68 in 20–30% of C-Nap1 KO cells that were not connected to centrosomes and did not contain the N-terminal C-Nap1 fragment (Fig 2A and S1B Fig). This suggests that C-Nap1 is not needed as an organizer for rootletin/Cep68 filaments per se, rather for the anchorage of these filaments to centrosomes.
LRRC45 that associates with the linker and with appendages of the mother centriole [12], lost linker localization in C-Nap1 KO cells but still bound to appendages (Fig 2A and S1B Fig). Nek2A no longer associated with the centrosomes of C-Nap1 KO cells indicating that the majority of this kinase binds to centrosomes via linker proteins (S1A and S1B Fig) [24]. Interestingly, Nek2A did not colocalize with the rootletin or Cep68 filaments that were observed in the proximity of centrosomes in C-Nap1 KO cells (S1A and S1B Fig). Transient transfection of C-Nap1 KO cells with a C-Nap1 expression construct restored localization of centrosomal linker proteins and centrosomal linker function (Fig 2B–2D).
Analysis of other centrosomal proteins revealed no change in the distribution of centrin [25], the centriole duplication proteins Cep135 and Sas-6 [26,27], the PCM proteins pericentrin and γ-tubulin [28,29], and the distal appendage protein Cep164 [30] (S1A and S1B Fig). Thus, although centrosomes from C-Nap1 KO cells lack centrosomal linker proteins, the localization of other centrosomal proteins was as in wt cells. Consistent with this conclusion, analysis of C-Nap1 KO cells by electron microscopy did not reveal obvious changes in centriole structure when compared to RPE1 wt cells (S2 Fig).
The term pericentriolar satellite defines electron-dense granules around the centrosome. These granules recruit centrosomal proteins and have functions in cilia formation [31,32]. In C-Nap1 KO cells the pericentriolar satellite protein PCM-1 [33] was preferentially positioned towards the Cep164-marked mother centriole (S1A–S1C Fig). Because of the close linkage of both centrioles in RPE1 wt cells, it was impossible to say whether this was the normal behaviour of pericentriolar satellites. The relevance of this asymmetric localization of PCM-1 remains unclear but it did not lead to a cilia formation defect in RPE1 C-Nap1 KO cells (S3A and S3B Fig), as this has been reported for siRNA depletion of pericentriolar satellite proteins [32,34]. Cilia formation in C-Nap1 KO cells is consistent with the published assembly of cilia upon siRNA depletion of C-Nap1 [30].
One of the main functions of centrosomes is the formation of the mitotic spindle. Analysis of mitotic C-Nap1 KO cells did not reveal any striking mitotic defects such as lagging or mis-segregated chromosomes (S4 Fig). The bipolar mitotic spindles of C-Nap1 KO cells (N = 20) were normal in appearance, suggesting that the main cellular function of the centrosomal linker is not in mitotic spindle formation.
Interfering with linker proteins increases the distance between both centrosomes during interphase [10–13]. We also observed an increase in average inter-centrosome distance from 1 μm in RPE1 wild type cells to 2.5 μm in three independent C-Nap1 KO cell lines (Fig 2C). When we categorized cells with a centrosome distance >2 μm as separated, only ~35% of C-Nap1 KO cells had separated centrosomes compared to the 5% recorded for wt cells (Fig 2D). These findings suggest that in the absence of the centrosomal linker another mechanism keeps the two centrosomes near to one another.
Early studies have shown that treatment of cells with nocodazole promotes centrosome separation [35,36]. This observation has been interpreted in different ways (see Discussion), however, the authors of these studies were not in a position to analyze the role of microtubules in centrosome positioning in the absence of centrosomal linker function. To this end, we perturbed microtubule and actin function with nocodazole, taxol and cytochalasin D, respectively. Nocodazole at 5 μM completely depolymerised microtubules (Fig 3A). In C-Nap1 KO cells, nocodazole drastically increased the average distance between centrosomes from 2.5 μm to ~8 μm (Fig 3B). Centrosome separation increased from 35% to ~80% upon drug treatment (Fig 3C).
In contrast, in RPE1 wt cells 5 μM nocodazole only moderately increased average centrosome distance from 1 to 2.5 μm (Fig 3B). It is important to note that these cells fell into two phenotypic groups. 70% of RPE1 wt cells maintained the short centrosome distance of 1 μm. In 30% of cells, centrosome distance increased to 4–15 μm, indicative of a failure of centrosomal linker function (Fig 3B). In order to understand why short treatment with nocodazole increases centrosome separation only in some interphase cells, we analyzed the state of the centrosomal linker of RPE1 wt cells in the presence and absence of nocodazole. RPE1 centrosomes associated with the linker proteins C-Nap1 and rootletin independent of their distance and whether cells were treated with nocodazole (Fig 3D). However, while centrosomes with a distance <2 μm were connected by rootletin fibres (Fig 3D, cells on the left), centrosomes with a distance >2 μm carried two unconnected rootletin bundles (Fig 3D, cells on the right). This was observed invariantly of nocodazole treatment. Thus, separated interphase centrosomes of RPE1 wt cells carry linker proteins but these are disconnected and therefore non-functional.
We next asked whether microtubule dynamics is important for the centrosome position in linker deficient cells. Low concentrations of nocodazole (100 nM), had only a modest impact on the integrity of the microtubule cytoskeleton (Fig 3A). At this concentration nocodazole mainly affects microtubule dynamics rather than overall architecture [37]. Despite the presence of a microtubule network, 100 nM nocodazole was as efficient in inducing centrosome separation in RPE1 C-Nap1 KO cells as complete microtubule depolymerization (Fig 3B and 3C). Interestingly, the effect of microtubule depolymerisation on centrosome positioning in C-Nap1 KO cells was reversible. Nocodazole wash out restored the relatively close juxtaposition of centrosomes within 2 h (Fig 3E). Thus, changes in microtubule dynamics are probably sufficient in disturbing centrosome positioning. Consistent with this notion, we observed that the microtubule stabilizer taxol also increased the distance between unlinked centrosomes, while having only a minor impact on centrosomes separation in RPE1 wt cells (Fig 3B and 3C).
Cells treated with cytochalasin D showed a complete collapse of the actin cytoskeleton (S5 Fig). However, actin depolymerization did not induce centrosome separation of C-Nap1 KO cells (Fig 3A–3C). Thus, we concluded that microtubules, not actin, maintain the proximity of unlinked centrosomes.
The microtubule motor protein dynein positions centrosomes to defined cellular locations in a number of cell types [38–40]. We therefore asked whether dynein is important for centrosome coordination in linker deficient C-Nap1 KO cells. Disruption of dynein motor activity using the dynein inhibitor ciliobrevin D [41] did not increase the distance between centrosomes in C-Nap1 clones (S6A and S6B Fig). The same was observed in overexpression experiments with the dynein inhibitor construct p50/dynamitin [42] (S6C and S6D Fig). Disorganization of the Golgi network was used as control for dynein inhibition in both experiments (S6A and S6C Fig). Thus, the spatial coordination of centrosomes in C-Nap1 KO cells is not dependent on dynein and likely requires the activity of other microtubule and/or cell cortex associated proteins.
The data above do not give insights into the dynamics of centrosomes in the absence of linker proteins. We addressed this point by transfecting RPE1 wt and C-Nap1 KO cells with a construct expressing mNeonGreen-PACT. The PACT domain targets the fluorophore to the centrioles [43]. In RPE1 wt mNeonGreen-PACT cells, the distance between the two centrosomes remained relatively constant at 0.5–1.0 μm (Fig 4A and 4B). Time lapse analysis indicated that the two centrosomes of C-Nap1 KO cells moved back and forth but kept an average distance of 2 μm and rarely separated further apart than 3.5 μm (Fig 4A and 4B). In confirmation of the data from fixed cells, 100 nM nocodazole treatment uncoupled the two centrosomes of C-Nap1-KO cells (Fig 4C and 4D). Shortly after nocodazole addition, the two centrosomes separated by up to 10–12 μm and maintained an average distance of 8–9 μm (Fig 4D). Live cell analysis therefore supported the view that in the absence of the centrosomal linker, microtubule dependent forces coordinate the closed spatial positioning of centrosomes.
Based on these experiments, we conclude that in RPE1 cells the centrosomal linker maintains the stable association of centrosomes. In the absence of the linker, microtubule forces keep the two centrosomes relatively close together.
The data above showed that centrosomal linker function is not essential for cell viability and spatial centrosome organization in RPE1 cells. In order to provide a coherent picture on centrosome linkage and to understand the contribution of the centrosomal linker and microtubules on centrosome positioning in different cell lines, we analyzed RPE1, Human Bone Osteosarcoma Epithelial Cells (U2OS) and HeLa cells in combination with siRNA depletion of C-Nap1 and nocodazole treatment (S7 Fig). Identical data as for RPE1 C-Nap1 KO cells were obtained upon siRNA depletion of C-Nap1 in RPE1 cells (S7A–S7C Fig) demonstrating that the RPE1 C-Nap1 KO phenotypes were not affected by adaptation or expression of the N-terminal C-Nap1 fragment. Consistent with published data [7], U2OS cells had a robust centrosomal linker and microtubule forces kept centrosomes relatively close together in the absence of linker function (S7D–S7F Fig).
The high genetic instability makes HeLa cells heterogeneous. We therefore analyzed two HeLa cell lines of different origin. Interestingly, the average centrosomal distance in HeLa-ATCC and HeLa-B cells was 2.5 to 4-fold higher than in RPE1 and U2OS cells (S7B, S7E, S7H, and S7K Fig). The basal level of centrosome separation was 35% for HeLa-ATCC and 60% for HeLa-B cells (S7I and S7L Fig). Interestingly, HeLa-ATCC cells fell into two groups. 70% had a centrosome distance of 1–2 μm, the distance of the others was >2 μm. This variation indicates heterogeneity in centrosomal linker function within this cell population. A substantial portion of HeLa-ATCC cells had centrosomal linker function as indicated by the C-Nap1 siRNA depletion induced centrosome separation from 35 to 60% (S7I Fig). Nocodazole treatment and C-Nap1 depletion had a synergistic effect on centrosome separation (S7H and S7I Fig).
HeLa-B cells had very little, if any, centrosomal linker function since C-Nap1 depletion hardly increased the already high level of centrosome separation (S7K and S7L Fig). However, microtubule depolymerization increased the centrosome distance from 4 to 6 μm (S7K Fig) indicating that microtubules provided some centrosome coordination in these cells.
These differences in centrosome behaviour were reflected in the morphology of the linker (S8 Fig). Most HeLa-B cells did not have a connecting centrosomal linker independent of the centrosome distance although linker proteins were associated with centrosomes (S8C Fig). U2OS and HeLa-ATCC cells had a functional linker when the centrosome distance was <2 μm (S8A and S8B Fig). Taken together, centrosomal linker function is variably established in human cell lines.
Recent data suggest a linkage between centrosomes and Golgi function [44,45]. We therefore analyzed distribution of the Golgi marker GM130 in RPE1 wt and C-Nap KO cells (Fig 5A and 5B). The area occupied by the Golgi increased at least 2-fold in C-Nap1 KO cells. In addition, we observed correlation between the centrosome distance and the Golgi area (Fig 5C) suggesting that the increase in centrosome distance is to some extent causing the Golgi organization defect. In cells with a centrosome distance of >8 μm, two well-separated Golgi stacks could even be observed (Fig 5A, cell on the right, white asterisks). Such observations suggest that both of the centrosomes are capable of independently organizing the Golgi stack.
An increase in centrosome number affects the migration behaviour of cells [46,47]. We therefore tested whether the increase in centrosome distance in RPE1 C-Nap1 KO cells, also has consequences on cell migration speed. Time-lapse analysis measured a medium speed of 28 μm/sec of RPE1 wt cells. Strikingly, the two independent C-Nap1 KO cells moved at a markedly reduced rate of only 15 and 18 μm/sec, respectively (Fig 6A and 6B). Using the random migration data, we analyzed if the directionality of movement was also altered in C-Nap1 KO cells. However, no significant difference was observed between the C-Nap1 KO and wt RPE1 cells as indicated by the very similar directionality index which is calculated as a ratio between Euclidean and accumulated distance.
A wound healing assay was performed in order to test if directed cell migration was also affected. RPE1 C-Nap1 KO cells also had a reduced migration speed in this assay compared to wt cells, but to a lesser extent than in the random migration assay (Fig 6D and 6E). Both migration assays confirm that unlinked centrosomes are reducing the speed of cell migration.
For the majority of interphase, a proteinaceous linker, called the centrosomal linker, physically connects the two centrosomes of a cell close together. The linker needs to be dissolved by Nek2A kinase and other mechanisms at the onset of mitosis in order to allow the assembly of a bipolar mitotic spindle [7,9,16,18,19]. However, the relevance of centrosome linkage for cell function during interphase is not understood, largely due to the lack of a model cell line in which the linker function is impaired.
C-Nap1 is the central anchor point of linker proteins at the proximal end of the two connected centrioles and its absence impairs binding of all other linker proteins [7,10–15]. RPE1 cells, in which both CEP250 copies were disrupted and hence lack the centrosomal linker, (Figs 1, 2, and S1 Fig) provide an excellent model system for the functional analysis of centrosomal linker function. As predicted from observations of the consequences of interfering with the C-Nap1 function [10], centrosome distance increased from 1 to 2.5 μm in C-Nap1 KO cells. However, to our surprise, some level of spatial organization persisted in the centrosomes of linker deficient interphase cells (Figs 3 and 4). We have determined that the microtubule cytoskeleton plays a role in keeping unlinked centrosomes close together. Low concentration of nocodazole, that mainly affect microtubule dynamics but not microtubule integrity, or the microtubule stabilizer taxol, impaired the positioning of the two unlinked centrosomes and increased the average distance from 2.5 to 8 μm (Fig 3B). We therefore suggest that forces that are dependent on the dynamic properties of microtubules position the two centrosomes in relatively close proximity when the centrosomal linker is missing. Dynein positions centrosomes in other model systems like C. elegans and Drosophia [38,39] but did not affect the distance of centrosomes in linker deficient C-Nap1 KO cells (S6 Fig). Identifying the players that keep centrosomes together in the absence of the linker is an important task for future studies, both from the perspective of centrosome biology in its own right and because this mechanism may contribute to the clustering of over-amplified centrosomes that supports the viability of cancer cells [48].
Previous studies have reported an impact of microtubules on interphase centrosome separation [35,36]. These results were first interpreted as microtubule tension forces that target the two unlinked centrosomes to the same cellular location [35]. With the discovery of the centrosomal linker, the nocodazole induced centrosomal separation was interpreted as a shift in balance between kinases and phosphatases that regulate linker proteins [36]. If the latter model was correct, we should not have observed an increase in centrosome separation in the C-Nap1 KO cell line upon microtubule depolymerization. Furthermore, upon nocodazole induced centrosome separation in RPE1 wt cells, we observed that the centrosomal linker was still at the centrosomes. We propose that both the centrosomal linker and microtubule dependent forces cooperate to keep the centrosomes in close proximity.
Our data suggest that 95% of RPE1 cells have a robust linker that keeps the centrosomes together (Fig 2). FRAP data on centrosomal linker proteins C-Nap1 and rootletin are consistent with this notion of a relatively stable linker [24]. Puzzlingly, despite this linkage, treatment for 1 h with either nocodazole or taxol increased the number of RPE1 cells with disengaged centrosomes from 5 to 20–30% (Fig 3C). We propose that the linker can temporarily loose connection between the two centrosomes with low frequency for example through the loss of rootletin-rootletin interactions within the rootletin polymer [15]. This model is supported by the split linker morphology of RPE1 wt cells with separated centrosomes (Fig 3D). In the absence of coordinating microtubules, centrosomes are able to separate beyond a recoverable distance and the breakage of the linkage becomes permanent. In the presence of microtubules the centrosomes are kept in close proximity to one another and the linker reforms between the two centrosomes to restore centrosome cohesion. Alternatively, cell cycle differences regarding the centrosomal linker may explain the nocodazole induced centrosome splitting. Only cells in a particular cell cycle phase could be more sensitive to centrosomes splitting in response to microtubule depolymerisation while the centrosomal linker of cells in other cell cycle phases remains unaffected.
Currently we do not know the exact mechanism by which microtubules keep unlinked centrosomes in relative close proximity. Here we discuss three models that are not mutually exclusive. The first model includes a possibility that there is another centrosomal linker that is C-Nap1 independent. We do not favour this view since we have shown using live cell imaging that unlinked centrosomes are moving back and forth and do not show the same coordinated movement like their linked counterparts in RPE1 wt cells (Fig 4). In the second model, the two unlinked centrosomes are close together because of similar spatial coordination. It has been suggested that microtubules originating from the two centrosomes interact with dynein at the cell cortex and equal pulling forces position both centrosomes in the cell center [40]. Such a hypothesis would argue that the relative proximity of unlinked centrosome would be dynein dependent. In fact, we have shown that dynein is not responsible for the proximity of unlinked centrosomes (S6 Fig). However, motor proteins other than dynein may position centrosomes in this way in RPE1 cells.
A third possibility is that the microtubules organized by the two centrosomes overlap and are used to establish the close inter-centrosome distance. Motors at centrosomes or associated within the anti-parallel microtubule overlap could position the two centrosomes close to each other. Such principle was shown to be important in yeast karyogamy where the minus end directed motor protein Kar3 is positioned at both spindle pole bodies (SPBs). Kar3 at one SPB pulls on microtubules organized by the other SPB. In this way the two nuclei are moved together until they finally fuse [49]. Additional experimental work is necessary to discriminate between these models.
Using C-Nap1 siRNA depletion and nocodazole treatment, we have analyzed RPE1, U2OS and HeLa cells to obtain a coherent picture of the spatial organization of centrosomes. In RPE1 and U2OS cells, we observed robust linker formation (S7 and S8 Figs). In both cell lines, C-Nap1 depletion and nocodazole treatment showed a clear synergistic effect on centrosome separation indicating that both organizing principles were active in these cells. Efficient centrosomal linker formation in U2OS and RPE1 cells is consistent with published data [7,15] and our RPE1 C-Nap1 KO analysis. Centrosomes of HeLa-ATCC cells responded to C-Nap1 depletion and microtubule depolymerization, indicating that both mechanisms were active. However, the relative high basal level of centrosome separation of 30% suggests that centrosome cohesion was suboptimal in these cells (S7I Fig). In HeLa-B cell the basal level of centrosome separation was very high at 60% (S7L Fig). Two lines of evidence suggest that the linker is not functional in HeLa-B cells. First, in most HeLa-B cells the centrosomal linker was not clearly connecting the two centrosomes even when they were close together (S8C Fig). Second, the high basal level of centrosome separation with an average distance of 4 μm indicates linker defects (S7K Fig). The reason for this failure in linker establishment in HeLa-B cells is presently unclear. High EGF signalling, overexpression of cyclin B2 or the misbalanced expression of linker proteins may cause this defect [18,19]. Furthermore, differences in microtubule dynamics could play a role in explaining the differences between different cell lines.
In any case, our observation of a non-functional linker in some cell lines may help to explain previous experimental findings. For example, siRNA depletion of CDK5RAP2, a microtubule organizing protein that is not associated with the linker [50], increased centrosome distance in U2OS and HeLa cells [13,51]. The penetrance of this phenotype appeared to be stronger in HeLa cells than in U2OS cells. This centrosome separation phenotype may be caused by altered microtubule properties in the cells. In addition, HeLa cells were reported to have highly mobile centrosomes with fluctuating inter-centrosomal distance [52,53]. Having in mind the data presented in this manuscript, a possible interpretation would be that these HeLa cells did not have a functional centrosomal linker allowing centrosome separation.
Recent data suggest a linkage between centrosomes and Golgi function [44,45]. These reports motivated us to analyze Golgi structure in RPE1 C-Nap1 KO cells. The Golgi of C-Nap1 KO RPE1 cells is spread over a wider area than in wild type cells (Fig 5B). A linkage of the Golgi stacks with centrosomes via microtubules is well established [44,45,54]. Microtubules organized by centrosomes interact with Golgi stacks [44]. The Golgi apparatus is also a MTOC that recruits γ-tubulin complexes through the proteins GM130, AKAP450, CDK5Rap2 and myomegalin [44,45,55]. The correlation between centrosome distance and the area covered by the Golgi (Fig 5C) suggests that the increase in centrosome distance is causing the Golgi organization defect. We propose that the increase in centrosome distance in C-Nap1 KO cells leads to greater spreading of the microtubules that are organised by the centrosomes to distribute the Golgi stacks over a larger area. Supporting this notion, when centrosomes of C-Nap1 KO cells were separated by >8 μm, two spatially organized Golgi organelles could be observed (Fig 5B). Such observation suggests that the two unlinked centrosomes can independently organize Golgi stacks. Moreover, cells with supernumerary centrosomes that are unlinked and scattered over a larger area also generate a Golgi organization defect [47].
Cells with extra centrosomes were recently reported to have disrupted cell migration due to the presence of multiple scattered microtubule organizing centers [47]. Since it was suggested that the centrosomal linker effectively joins the two centrosomes into one functional MTOC [7], we investigate whether unlinked centrosomes would have the same effect on cell migration as supernumerary centrosomes. Random migration of RPE1 C-Nap1 KO cells was compared to RPE1 wt cells and revealed a 35–46% decrease in the velocity of C-Nap1 KO cells (Fig 6A and 6B). A decrease in cell migration speed of C-Nap1 KO cells was also observed in the directional migration using the wound healing assay. Thus, centrosomal linker deficient RPE1 cells migrate with reduced speed (Fig 6D and 6E). In contrast, the directionality of the random movement was unaffected by the absence of the centrosomal linker (Fig 6C). The exact nature and molecular connection of the reduced cell migration in C-Nap1 KO cells is presently unclear.
Taken together, the centrosomal linker is important for joining the two centrosomes into one functional MTOC unit during interphase. A relatively small increase in centrosomal distance already affects the organization of the Golgi and has consequences on cell migration. It will be interesting to see how this relatively modest defect in centrosomal linker deficient cells impacts upon the development and function of an organism.
HeLa and U2OS cells were cultured in Dulbecco's Modified Eagle's Medium (DMEM) Glutamax (Gibco) supplemented with heat inactivated 10% (v/v) FBS and 2 mM L-glutamine. hTERT-RPE1 cells were cultured in Dulbecco's Modified Eagle's Medium F-12 Nutrient mixture (DMEM/F-12, Gibco) supplemented with 10% FBS and 2 mM L-glutamine. All cell lines were cultured at 37°C in a humidified atmosphere with 5% CO2. Plasmid transfection was performed using Lipofectamine LTX according to manufacturers (Life Technologies) instructions. For siRNA-based experiments, Lipofectamine RNAi MAX was used according to manufacturers instructions. The following siRNA oligos were used: Non targeting siRNA (Human Dharmacon ON-Target plus, Nr. 1, Thermo Scientific, sequence: 5´-UGUUUACAUGUCGACUAA-3´) and C-Nap1 siRNA (Human CEP250 Dharmacon ON-TARGET plus, Nr. 1 and 3, Thermo Scientific, sequence: 5’-GAGCAGAGCUACAGCGAAU-3’ and 5´- AAGCUGACGUGGUGAAUAA-3). Microtubule depolymerization was performed using nocodazole at 100 nM or 5 μM for 1 h. Microtubule stabilization was performed with 10 μM taxol for 1 h. In order to depolymerize actin filaments, 1 μM Cytochalasin D was used for 1 h. Dynein was inhibited with 0.125 mg/ml Ciliobrevin D for 1 h.
Sigma Aldrich designed a ZFN with a specific cut site in exon 15 of the CEP250 gene (coding for C-Nap1). The donor vector was constructed as reported [56] by PCR amplification of the genomic locus 800 bp upstream and downstream of the ZFN cut site using primers with the following sequences FW1: 5´-GCTGAGGCAGGAGAATCTCTTG-3´ and REV1: 5´-GGGCCAGCTGT CTGGCTGC-3´. The PCR product was subsequently subcloned in pJet 1.2 vector. The ZFN cut site was mutagenized in the donor vector by PCR mutagenesis. The 3xSTOP codon in every frame-Neomycin resistance cassette was inserted at the cut site. RPE1 wt cells were co-transfected with the ZFN plasmids and the donor vector using electroporation (Invitrogen, Neon transfection system). After transfection cells were cultured for 2 d at 37°C and then for 2 d at 30°C in order to increase ZFN efficiency. 4 days after the transfection, cells were seeded in 96-well plates (100 cells per well) in the presence of 0.5 mg/ml G418 for selection. 2–3 weeks after the selection onset, colonies were picked and screened for the presence of the Neomycin resistance gene at the correct genomic locus with primers FW2: 5´-TGCCTGTAATCCCAACTACTCG-3 and REV2 5´-TGTGCGAGG CCAGAGGCC-3´. The wt genomic locus was amplified using FW1 and REV1 primers. Absence of wt mRNA was confirmed by using FW3: 5´-CTGTGTGCAGCAGAATGGAGGCC-3 and REV3: 5´- CCTCTAGAGCCGCTTTCTCTCG-3´ primers that would amplify from wt exon 14 to exon 15, but could not amplify the mutated exon 15 probably because of its larger size.
For indirect immunofluorescence, cells were fixed with ice-cold methanol for 5 min at -20°C or with 4% PFA for 15 min at room temperature. Cells were permeabilized with 0.1% Triton X-100 for 10 min, blocked with 10% (v/v) fetal calf serum (FCS) for 30 min and stained with antibodies in 3% (w/v) BSA (bovine serum albumin) in PBS. DNA was stained with Hoechst 33342 (0.2 g/ ml, Calbiochem). The following antibodies were used in immunofluorescence microscopy experiments: anti-C-Nap1 (BD Biosciences, recognizes N-terminus of C-Nap1), anti-C-Nap1 [16], anti-rootletin [16], anti-Cep68 (kind gift from E. Nigg [13]), anti-γ-tubulin (Abcam, TU-30), anti-Nek2 [16], anti-pericentrin (Abcam, ab-4448), anti-LRRC45 [12], anti-α-tubulin (Sigma, T-9026), anti-GM130 (Cell Signalling, D681), anti-MM491 (human centrosome auto-immune serum), anti-Cep164 (kind gift from G. Pereira [57], anti-Sas6 (Santa Cruz, sc-81431), anti-Cep135 (raised against a recombinant Cep135 fragment of amino acids 1–658 in rabbit by EuroGentec), anti-PCM-1 (kind gift from O. Gruss [34]), anti-acetylated-tubulin (C3B9 monoclonal antibody; kind gift of G. Pereira). Secondary antibodies were donkey anti-rabbit IgG coupled to Alexa Fluor 488, Alexa Fluor 594 or Alexa Fluor 647, donkey anti-mouse IgG coupled to Alexa Fluor 555 or Alexa Fluor 488, and donkey anti-goat IgG coupled to Alexa Fluor 555 (all 1:500; Invitrogen) and donkey anti-human coupled to Alexa Fluor 488 (used 1:200).
Imaging was performed on a DeltaVision RT system (Applied Precision) with an Olympus IX71 microscope equipped with FITC (fluorescein isothiocyanate), TRITC (tetramethyl rhodamine isothiocyanate) and Cy5 filters (Chroma Technology), a plan-Apo ×100 NA 1.4 and ×60 NA 1.4 oil immersion objective (Olympus), a CoolSNAP HQ camera (Photometrics), a temperature controller (Precision Control) and Softworx software (Applied Precision). Confocal imaging was performed using Zeiss LSM780 microscope with standard equipment.
Centrosome distance was manually calculated in 3 dimensions using the formula (3D distance)2 = (2D distance)2+(z stack distance)2.
Live cell imaging was performed using the Nikon Biostation microscope IM-Q (cell migration assays) or DeltaVision RT system (fluorescence live cell imaging). Both microscopes were used at 37°C in a humidified atmosphere with 5% CO2. All live cell imaging experiments were performed with Ibidi glassware. Directionality index was calculated as a ratio between Euclidean distance and accumulated distance using the following formula: D = (Euc.dist.)/(acc.dist.). Wound healing assay migration speed was calculated as the average speed of the cell front.
Cells were collected by trypsinization. After washing with PBS, the cells were lysed in 10 mM Tris-Cl pH 7.5, 150 mM NaCl, 5 mM EDTA, 0.1% SDS, 1% Triton X-100, 1% deoxycholate supplemented with 1 mM PMSF (Sigma) and protease inhibitor cocktail (Roche) for 30 min cell. Lysates were centrifuged and the supernatant was boiled with Laemmli buffer. SDS-PAGE was performed as previously described [58]. Transfer to membrane was done using a BioRad Mini-Transblot Electrophoretic Transfer System. The membranes were subsequently blocked in 5% non-fat milk in TBS-T. The following primary antibodies were used: anti-C-Nap1 (BD Biosciences), anti-GAPDH (Cell Signalling Technology), anti-α-tubulin (Sigma, T-9026); with appropriate secondary antibodies: donkey HRP-coupled anti-mouse and donkey HRP-coupled anti- rabbit antibodies (from Jackson laboratories).
Cells were grown on coverslips and fixed using 2.5% glutaraldehyde in 0.1 M Na cacodylate buffer, pH 7.2, at room temperature for 30 min. The cells were subsequently washed with 0.1 M Na cacodylate buffer and postfixed with 2% osmium tetroxide in Na cacodylate buffer for 1 h on ice. The samples were washed and contrasted in 0.5% uranyl acetate over night. The samples were subsequently washed and gradually dehydrated by immersing them in a graded ethanol solution from 50, 70, to 90% and finally two times in 100% ethanol. Dehydrated cells were embedded in Epoxy medium using Epoxy Embedding kit (Fluka) and serial sections were generated using Reichert Ultracut S Microtome (Leica Instruments). Sections were post-stained with 2% uranyl acetate (in 70% methanol) and lead citrate. Finally, serial sections were viewed using a CM120 electron microscope (Phillips Electronics), operated at 120 kV, and images obtained by a Keen view CCD camera (Soft imaging systems).
ImageJ software was used for image analysis [59]. Centrosomes were counted as separated if the distance between them exceeded 2 μm. Mitotic cells were excluded from this analysis. For fluorescence intensity quantification of PCM-1, a square 3x3μm was used either having the centrosomal pair (in the case of wt) or having the mother or daughter centrosome in the center (separately, in the case of C-Nap1 KO clones). Average of background intensities were subtracted from each measurement in each channel. All statistical analyses were performed using GraphPad PRISM software.
|
10.1371/journal.ppat.1003907 | Competitive and Cooperative Interactions Mediate RNA Transfer from Herpesvirus Saimiri ORF57 to the Mammalian Export Adaptor ALYREF | The essential herpesvirus adaptor protein HVS ORF57, which has homologs in all other herpesviruses, promotes viral mRNA export by utilizing the cellular mRNA export machinery. ORF57 protein specifically recognizes viral mRNA transcripts, and binds to proteins of the cellular transcription-export (TREX) complex, in particular ALYREF. This interaction introduces viral mRNA to the NXF1 pathway, subsequently directing it to the nuclear pore for export to the cytoplasm. Here we have used a range of techniques to reveal the sites for direct contact between RNA and ORF57 in the absence and presence of ALYREF. A binding site within ORF57 was characterized which recognizes specific viral mRNA motifs. When ALYREF is present, part of this ORF57 RNA binding site, composed of an α-helix, binds preferentially to ALYREF. This competitively displaces viral RNA from the α-helix, but contact with RNA is still maintained by a flanking region. At the same time, the flexible N-terminal domain of ALYREF comes into contact with the viral RNA, which becomes engaged in an extensive network of synergistic interactions with both ALYREF and ORF57. Transfer of RNA to ALYREF in the ternary complex, and involvement of individual ORF57 residues in RNA recognition, were confirmed by UV cross-linking and mutagenesis. The atomic-resolution structure of the ORF57-ALYREF interface was determined, which noticeably differed from the homologous ICP27-ALYREF structure. Together, the data provides the first site-specific description of how viral mRNA is locked by a herpes viral adaptor protein in complex with cellular ALYREF, giving herpesvirus access to the cellular mRNA export machinery. The NMR strategy used may be more generally applicable to the study of fuzzy protein-protein-RNA complexes which involve flexible polypeptide regions.
| Herpes viruses invade cells, hijacking cellular components to sustain their lifecycle and replicate. A critical step of infection is the export of viral mRNA from the nucleus to the cytoplasm, where the molecular machinery to produce proteins is located. To provide a link between their mRNA and cellular components of the mRNA export pathway, all herpesviruses use special adaptor proteins. These adaptor proteins specifically select viral mRNAs from the mixture present in the nucleus, and introduce them to cellular mRNA export factors, such as ALYREF. How these viral adaptors manage to trick ALYREF to accept foreign genetic material has not been understood on a molecular level. In this study we reveal how a typical viral adaptor protein ORF57 recognizes specific viral RNA motifs, and also how it binds to the cellular protein ALYREF. We uncover details of how ORF57 transfers the viral RNA to ALYREF, locking it in the cooperative ternary complex. We also describe the atomic-resolution structure of ORF57-ALYREF interaction interface. Together the data provides the first molecular insight of how viral mRNA is transferred between viral and cellular proteins, thus helping virus to hijack a cell.
| Mammalian gene expression is coupled with mRNA maturation, where nascent transcripts undergo a continuous series of splicing and processing events finally leading to nuclear export to the cytoplasm [1]. This process is tightly regulated and orchestrated, ensuring that only mature and fully-processed cellular mRNA is exported from the nucleus, to be correctly translated into proteins in the cytoplasm. The recruitment of protein markers acquired during this maturation process, such as UAP56, UIF and ALYREF (otherwise known as Aly, REF, Aly/REF, REF/Aly, BEF, Thoc4 in metazoan and Yra1 in yeast), is essential for the export of cellular mRNA via the NXF1 pathway (otherwise known as TAP) [2]–[4]. These markers are part of the multicomponent TREX complex which associates with the 5′ end of cellular mRNAs during splicing [3]. TREX recruits NXF1 to mRNA and TREX triggers a conformational change in NXF1, such that it binds mRNA with high affinity [5], [6]. The cellular protein ALYREF functions as an export adaptor, binding mRNA as part of TREX, and also interacting with NXF1 [7], [8]. The structure of ALYREF has been characterized: it consists of central folded RRM domain [9] flanked by two largely flexible multifunctional N- and C-terminal domains [10]. ALYREF primarily uses its N-terminal flexible arginine-rich region for interaction with NXF1; this region closely overlaps with the RNA binding site [10]. The arginines within this region become methylated, which reduces its RNA binding activity and may serve as a control mechanism for RNA displacement from ALYREF to NXF1 [11]. ALYREF and Thoc5 binding remodels NXF1, increasing its binding affinity for mRNA, ensuring transfer of mRNA to NXF1 [5], [6]. NXF1 then introduces the mRNA to nucleoporins, committing it to exit from the nucleus through the nuclear pore [12]–[14].
Herpesviridae possess an intriguing ability to circumvent the sophisticated cellular controls which ensure that only mature spliced mRNA can be exported from the nucleus. Viral mRNA is generally unspliced, therefore it cannot acquire the normal protein markers during splicing, which would signal that mRNA is ready for export to the cytoplasm. However, all herpesviruses express an essential multi-functional adaptor protein which specifically recognizes viral mRNA, and bridges its interaction with TREX complex via binding to cellular mRNA export factors such as ALYREF and UIF [15]–[19], for subsequent export via the NXF1 pathway [20]–[23]. It was also recently suggested that ALYREF may be recruited by viral adaptors to stabilize the viral nuclear RNAs independently of their export [24]. The infected cell protein 27 (ICP27) from Herpes Simplex Virus type 1 (HSV-1) is probably one of the most well-studied examples of the viral multifunctional adaptors [25]. In Herpesvirus Saimiri (HVS), which is the prototype γ-2 herpesvirus with close similarity to human Kaposi's Sarcoma-associated herpesvirus (KSHV), a similar function is carried out by the protein ORF57 [21], [23]. Homologs of these adaptor proteins are also known as ORF57 in KSHV [26], [27], EB2 in Epstein-Barr virus (EBV) [28], and UL69 in the human cytomegalovirus [29]. All these viral adaptor proteins contain long intrinsically-unstructured but functionally-important regions, with relatively poor sequence homology. Although these proteins appear to have a very similar function in promoting viral mRNA export via the cellular NXF1 pathway, the location and appearance of their RNA-binding regions vary, and the precise location of ALYREF binding sites cannot be inferred from their amino acid sequences. How exactly they perform their viral mRNA export function, and introduce viral mRNA to cellular proteins such as ALYREF, has not been described in detail yet.
Recently, the structure of the interaction interface between HSV-1 adaptor protein ICP27 and cellular mRNA export factor ALYREF was determined [30]. (It should be noted that while in our previous study [30] ALYREF protein was referred to as REF, due to recent recommended changes by the HUGO Gene Nomenclature Committee [31] here we will be referring to the same protein as ALYREF). In this structure, interaction with the RRM domain of ALYREF is achieved via a very short peptide fragment of the flexible N-terminal region of ICP27 [30]. Additionally, a ALYREF-interacting region aa103–120 was mapped on HVS ORF57 protein [30]. The mostly unstructured ORF57 region aa8–120 [30] mediates specific recognition of HVS mRNA via the viral RNA sequence motif GAAGRG [32]. Although this ORF57 fragment contains an arginine-rich region, it lacks any canonical RNA-binding sequence features such as an RGG box, which is present in ICP27 [33], [34]. Therefore, the exact location of the RNA binding site remained unknown, along with the mechanism of RNA transfer from ORF57 to ALYREF. Which protein sites are involved at different stages of such a transfer? What is the structure of the ternary ORF57-RNA-ALYREF complex? Answers to these questions would enable further functional and mutagenesis studies, to reveal how the assembly and disassembly of complexes involved in RNA recognition, transfer and export are achieved at the molecular level [35], [36].
In this study we used solution state NMR to reveal molecular details of the ternary complex assembly of functional fragments of HVS ORF57, HVS RNA and ALYREF, and suggest a model for the mechanism of RNA transfer between protein molecules in this system. The mapping experiments show a clear difference between binding of non-specific random-sequence RNA oligos, and RNA oligos containing HVS-specific sequence motifs, to a flexible arginine-rich region of ORF57. We reveal that for the ORF57 protein, its ALYREF binding site also forms part of the specific viral RNA recognition region, with adjacent arginine-rich sequences also contributing to RNA binding. We present the atomic-resolution structure of the ORF57-ALYREF binding interface, which somewhat differs from that of ICP27-ALYREF identified earlier [30]. Using a new strategy based on principles of saturation-transfer (ST) between molecules [37] and isotopically-discriminated NMR [38], we followed the changes in RNA binding sites which accompany transfer of RNA from one protein molecule to another. In the ternary ORF57-RNA-ALYREF complex, RNA is partially displaced from its binding site on ORF57 by ALYREF, but is retained in the complex by the synergistic action of flanking flexible regions of both ALYREF and ORF57. The detailed model obtained based on NMR data was supported by mutagenesis studies, and cooperativity in ternary complex assembly was additionally characterized by fluorescence measurements.
Previously the recognition of specific short viral mRNA sequences was attributed to Herpesvirus saimiri ORF57 protein region aa8–120 [32], however the precise binding site within this fairly long region was unknown, and no obvious sequence patterns (such as RGG boxes) indicative of RNA-binding sites could be identified. To locate the RNA-binding regions experimentally within ORF578–120, and study specific vs non-specific binding, we used NMR spectroscopy. Sequence specific signal assignments of all amides of ORF578–120 allowed the mapping of interaction sites to a residue-level resolution. The unlabelled RNA oligonucleotides were added to 15N-labelled protein samples and residue-specific signal changes were monitored. The effect of non-specific RNAs of different lengths (oligonucleotides named 7merN and 15merN, see Materials section) on signal position and shape were compared with ORF57-specific RNA oligos (7merS and 14merS) containing the previously identified HVS motif GAAGAG [32] (Fig. 1A and Fig. S1A). For non-specific 7merN and 15merN (even at two-fold excess) the amide signal changes in ORF578–120 were small and scattered across the entire sequence (Fig. S1A), suggesting only transient non-specific binding. Similarly, addition of a two-fold excess of non-specific 7merN caused no significant change in local mobility of the ORF57 polypeptide chain, as evidenced by 15N{1H}-NOE both for the ORF578–120 and ORF5756–140 constructs (Fig. S1B). (The latter construct was used as a control to ensure that the absence of binding with 7merN is not an artifact of C-terminal truncation of a potential binding site). In contrast, the ORF57-specific oligos caused substantial signal broadening in all signals corresponding to the region aa64–120 (Fig. 1A). The severity of signal perturbations was also dramatically dependent on the length of the oligo used, reflecting differences in the apparent affinity of RNA binding. For specific 7merS, all signals within the aa64–120 region were broadened beyond detection once a 1∶1 stoichiometry was reached, whereas for 14merS, equivalent signal loss occurred at 0.2∶1 RNA∶protein ratio. Notably, the ALYREF-binding region aa103–120 [30] was affected most severely by the addition of RNA, suggesting that the RNA and ALYREF-binding sites partially overlap. To determine if this ALYREF binding region is sufficient for specific RNA binding, the short ORF57103–120 peptide was titrated with 7merS, however no NMR signal broadening occurred and only small signal perturbations (under 0.04 ppm) were observed even with a 3-fold excess of RNA (Fig. S1C). Signal perturbation mapping therefore suggested a specific RNA binding site encompassing aa64–120 within ORF578–120, whilst also showing the ALYREF-binding region aa103–120 located within this site is not sufficient for the recognition of specific viral RNA. Fluorescence measurements were also used to estimate the Kd for 14merS binding as 7.57±0.06 µM, compared to 38.8±0.6 µM for 7merN (Fig. S7A,B).
Unexpectedly, the addition of both specific and non-specific RNA oligos caused small NMR signal shifts in the acidic region of ORF578–120 (aa10–40). We could not observe intermolecular NOEs between RNA and protein for the definitive binding epitope mapping, therefore, to separate possible indirect effects of conformational changes on signal shifts brought about by RNA binding and identify direct points of contact, we used RNA→ORF57 cross saturation transfer (ST) experiments [37], [39] (see Fig. S1D). These experiments report directly on the spatial proximity of RNA moieties to NH groups of individual amino acid residues (contact distance <5 Å), and provide essentially the same type of information as traditional RNA-protein cross-linking assays, but in a site-specific manner. A sample containing 15N-labelled ORF578–120 : RNA 7merS in a ratio of 1∶0.5 was prepared (higher RNA concentrations prevented measurements due to excessive signal broadening). Selective saturation of RNA signals with a series of radiofrequency pulses resulted in a significant decrease in signal intensity of the backbone amides in 1H-15N correlation spectra (relative to the reference spectrum with off-resonance saturation) of mainly aa107–120 and aa81–92, and to a lesser extent, aa94–105, and even less, aa64–79 (Fig. 1A and Fig. S1E). (The typical effects of RNA→ORF57 saturation transfer on selected example signals from amides non-adjacent and adjacent to RNA are shown on the bottom right traces of the Figure “Typical effects of complex formation and RNA→protein ST” introduced later in the Results section.) The RNA→ORF57 ST experiment was repeated as a control with non-specific RNA 7merN, but no site-specific saturation transfer, and hence no direct interaction, was detected even when using a 2-fold excess of RNA (Fig. 1A and Fig. S1E). Based on the results of saturation transfer mapping, which are also in line with signal perturbation mapping, we conclude that ORF578–120 contacts the specific RNA motifs directly using primarily its regions aa107–120 and aa81–92, with additional contribution from residues within aa94–105 and aa64–79. No significant binding was detected with non-specific RNA of similar length.
Previously we mapped aa103–120 as the ALYREF interaction site in HVS ORF57 [30]. Here, residues within the same region were also implicated in binding with a specific RNA motif. Given the multi-functional importance of this region, we endeavored to characterize it structurally. The secondary structure prediction algorithms Psipred [40] and Agadir [41] suggest ORF57 aa108–118 should be α-helical. Our experimental NMR data, namely dihedral angles derived from TALOS+ [42], 15N[1H] NOE experiments, and presence of characteristic i to i+3 NOEs for a shorter peptide ORF57103–120 (see Fig. S2), also all demonstrate that the ORF578–120 site aa107–118 exists in α-helical conformation; therefore this region was named “R-b helix”.
Previously the structure of the complex of ALYREF fragment aa54–155 (ALYREF54–155) with ORF578–120 could not be determined due to an unfavorable chemical exchange regime, causing signal broadening for the interacting residues [30]. In view of the importance of the aa103–120 region for both ALYREF and RNA binding, and differences in local structure of ALYREF-binding regions of ICP27 [30] and ORF57, we pursued the structure of the ORF57-ALYREF complex interface. We employed a short ORF57103–120 construct, which displayed much improved spectra and less exchange behavior (Fig. S3). The atomic resolution structure of the ALYREF54–155 - ORF57103–120 complex was determined using a total of 2427 non-redundant NOEs, 122 of which were intermolecular (Table 1 and Fig. S4). Previous signal perturbation mapping indicated that ORF57 aa103–120 comprise the ALYREF-binding site [30]; the new data defined the site more precisely as aa106–120 (Fig. S3C). Within the complex, the ORF57 peptide is α-helical for aa108–119, contacting the loops L1 and L5 on the α-helical face of ALYREF (Fig. 2). The binding site on ALYREF is composed of a hydrophobic patch formed by the sidechains of L82, V86, L94, Y135, V138, L140 and M145, with E93 and E97 contributing to ionic interactions (Fig. 2D). The aromatic sidechain of W108ORF57 is positioned at one end of this hydrophobic patch of ALYREF in close proximity to the sidechain of V86 (Fig. 2D). The majority of the remaining hydrophobic contacts of ORF57 are formed by V112 and the aliphatic part of the R113 sidechain, with A109, A115 and A116 also contributing. The positive charged R113ORF57 sidechain is positioned between ALYREF residues E93 and E97 which are therefore likely to form salt bridges. The structure reveals unexpected differences in binding conformations and molecular recognition of two functionally-similar viral adaptor proteins, HVS ORF57 and HSV-1 ICP27, on essentially the same site on ALYREF (Fig. 3 and Fig. S3D), despite the presence of deceptively similar recognition triads identified earlier [30].
ALYREF is known to bind RNA weakly and non-specifically, primarily using its flexible N- and C-terminal domains [7]. Using NMR signal perturbations, non-specific 15-mer RNA (15merN) oligonucleotide binding to the ALYREF fragment aa1–155 (ALYREF1–155) was previously mapped to RGG motifs situated within its unstructured N-terminus and also to loops L1 and L5 of the RRM domain [10]. To address how well ALYREF binds to the viral mRNA specifically recognized by ORF57, we firstly explored how ALYREF binding to RNA depends on the length of the viral oligo sequence. Chemical shift mapping was carried out with [15N]-REF1–155 using equimolar 7merS or 14merS (Fig. 1B). The data indicated that the short RNA 7merS causes small perturbations almost exclusively within the RRM domain in loops L1, L3 and L5, whereas the longer 14merS caused signal broadening within the N-terminal aa12–48 along with minor shift changes within the RRM. The extent and location of signal changes is similar to that observed previously using non-specific 15-mer RNA (15merN) [10], suggesting that ALYREF itself cannot discriminate between viral and non-viral RNA, and binds it only weakly. Saturation transfer experiments confirmed that 14merS RNA contacts ALYREF in the N-terminal region containing RGG motifs (Fig. 1B and Fig. S1F), at the same site where non-specific RNA binding occurs. The measurement of the Kd for ALYREF1–155 interaction with RNA oligonucleotides using fluorescence unfortunately could not be completed due to increase in sample turbidity upon RNA addition, likely caused by non-specific protein aggregation. From NMR titration data the lower-limit Kd estimates for 7merS and 14merS binding were >100 µM and >50 µM, respectively. These values are significantly higher than the values characterizing the specific binding of the 14merS to viral ORF578–120 (7.57 µM, Fig. S7A), and closer to the Kd for non-specific binding of 7merN to ORF578–120 (38.8 µM, Fig. S7B). Overall, these estimates show that the viral RNA motif is specifically recognized and binds with viral ORF578–120 but not ALYREF1–155, suggesting that in the cell the viral mRNA would be initially preferentially recognized and bound by viral ORF57.
As shown above, the ORF57 region aa106–120 is involved in specific viral RNA binding, but it can also be utilized for ALYREF binding. This raises the question: which binding partner does this particular region select when all three components are present? Here we used solution NMR experiments to investigate directly if these local interactions are indeed competitive, and which of these is stronger and hence is preferentially selected when the ternary complex is formed. Residue-specific signal changes in the ORF57-ALYREF complex upon addition of unlabelled RNA were monitored using an IDIS-TROSY experiment [38], which allows the observation of separate 1H-15N-correlation spectra of two differentially-labeled proteins in the same sample. At a stoichiometric 1∶1 ratio of [15N,13C]- ORF578–120 and [15N]-REF1–155, the backbone amides at the protein-protein interface are exchange broadened (including aa106–120 of ORF57), however signals from other regions of both proteins are clearly observable, and their pattern is characteristic of ORF57-ALYREF complex formation. Having been assigned to particular amino acid residues, they were able to report site-specific changes in the protein-protein interactions in response to ligand binding (Fig. 1, Fig. S5 and Fig. S6).
In an initial experiment, a stoichiometric equivalent of a shorter specific RNA 7merS was added to the differentially-labeled ORF57-ALYREF complex. In contrast to the substantial broadening of aa64–120 observed on addition of 7merS to free ORF578–120, no broadening and only small shifts in ORF57 signals were observed when ALYREF was present (Fig. 1). This indicates that ALYREF reduces ORF57 binding to specific viral RNA, protecting its binding site, aa106–120. A control experiment using ALYREF54–155, which lacks the N-terminal RNA binding site, produced similar results, also suggesting that the region aa106–120 of ORF57 has higher affinity for ALYREF than for a specific RNA oligo 7merS (Fig. S6). The ALYREF signal changes induced by the addition of 7merS were marginal. In a related experiment, to directly follow the displacement of RNA from ORF57, one equivalent of the 7merS was added to [15N]-ORF578–120 causing substantial signal broadening in aa64–120. Then one equivalent of [15N,13C]-REF1–155 was added to the same sample, resulting in recovery of all ORF57 signals except for those which became instead involved in the ALYREF interaction (aa106–120) and remained broad. IDIS-TROSY spectra of both ALYREF and ORF57 confirmed formation of the protein-protein complex as the fingerprint pattern of observable ORF57 signals was consistent with ORF57 bound to ALYREF, but not RNA. These direct experiments demonstrate that virus-specific RNA is displaced from ORF57 aa106–120 by the competitive binding of ALYREF to this region, and not by the preferential binding of RNA to ALYREF.
As the short RNA 7merS cannot be bound efficiently by the ORF57-ALYREF complex, subsequent experiments were performed using a longer specific RNA 14merS. Importantly, as evidenced by the presence of a large number of relatively sharp signals from amide groups of both proteins (Fig. 4, Fig. 5, Fig. S5, Fig. S6), the complexes formed retained a high degree of flexibility, even for residues directly involved in interactions. There were no NOE signals observed between RNA and proteins, making it impossible to apply standard techniques for full 3D structure determination of the ternary complex. Therefore, to obtain information regarding the spatial organization of this largely fluid assembly, a saturation-transfer version of isotopically-discriminated TROSY [38] experiment (ST-IDIS-TROSY) was created and used to detect directly, in residue-specific manner, where exactly RNA contacts the ALYREF1–155 and ORF578–120 in the ternary complex (Fig. 4). A sample was prepared containing a 1∶1∶1 mixture of [15N,13C,2H]-ORF578–120, [15N,2H]-REF1–155 and non-labeled 14merS (∼40 kDa complex in total). Protein deuteration was used to improve the quality of spectra and reduce possible artifacts due to spin diffusion effects [37]. RNA proton signals were selectively saturated by radiofrequency pulses [37], [39], and changes in IDIS-TROSY [38] peak intensities were monitored to reveal which amide groups are situated in close proximity (<5 Å) to RNA moieties, observing fingerprint spectra from both proteins at once (Fig. 4B). The examples of typical changes in individual signals (from interacting and non-interacting sites) in response to complex formation and saturation transfer are shown for illustration on Fig. 5, with residue-specific results presented in Fig. S1G,H, and an overview is included in Fig. 1. When the saturation transfer effect was initially calculated from the ratio I5.85/I21.0 obtained with on-resonance ribose proton (5.85 ppm) and off-resonance (21.0 ppm) saturation, we noticed a significant amount of non-specific saturation transfer to virtually all serine residues in ORF57. We explained that by the inadvertent saturation of serine hydroxyl groups. To compensate for this effect, we have used two different saturation schemes. In the first scheme, we selectively saturated two RNA resonances (moieties) with similar chemical shifts (5.75 and 5.85 ppm), and calculated the ratio of signal intensities I5.75/I5.85 (Fig. S1G). If one of the saturated RNA moieties is positioned closer to a protein amide group (and within 5 Å) than the other, then the amount of cross-saturation transfer from them to this amide will not be equal. Hence, where the ratio I5.75/I5.85 deviates from unity, it highlights residues adjacent to RNA. The close positioning of saturating frequencies on the other hand should cross-saturate broad hydroxyl signals to a similar extent, compensating for this artifact. In the second scheme, the on-resonance saturation was centered at 12.0 ppm and off-resonance at 21.0 ppm, and I12.0/I21.0 ratio calculated (Fig. S1H). The RNA signals at 12.0 ppm were broad and not observable, but this frequency was chosen as it is characteristic for RNA imino protons. Both saturation schemes led to similar mapping results: whereas many protein amide resonances remained unaffected by the RNA signal saturation, several regions in ALYREF and ORF57 in the ternary complex were clearly highlighted (Fig. 1 and Fig. S1G,H). The most pronounced ST effect was observed for the arginine-rich N-terminal region of ALYREF aa24–48, with parts of the RRM domain also affected (Fig. 1B). The region aa79–100 within ORF57 was also highlighted by saturation transfer, as seen by the deviations of the I5.75/I5.85 and I12.0/I21.0 ratios from unity. The increase in estimated error margins within the regions affected by ST (Fig. S1G,H) is explained by signal broadening, leading to a reduction of signal intensities for amides in contact with RNA. The presence of ALYREF in the sample clearly reduces the size of the ORF57 site available for RNA binding (Fig. 1A). In the presence of ORF57, the saturation transfer from RNA 14merS to ALYREF becomes more pronounced (i.e., larger deviation of Ifreq1/Ifreq2 from unity), suggesting that RNA is retained by ALYREF within the ternary complex more efficiently than by ALYREF alone.
The NMR experiments therefore all indicate that ALYREF partially displaces the viral RNA initially bound specifically to ORF57, but retains it within the complex. In the ternary complex ORF57 aa106–120 directly interacts with the ALYREF RRM, whereas flexible flanking regions of ALYREF (aa24–48) and ORF57 (aa81–92), and to lesser extent, parts of helix 2 of the ALYREF RRM, jointly keep hold of the viral RNA molecule. Interestingly, amide signals from flexible protein regions which become involved in direct contacts with RNA (as evidenced by RNA-protein saturation transfer), are only partially broadened in the complex. They had intensities higher than signals from the folded regions, but lower than signals from the unfolded non-interacting regions (examples of this behavior can be seen in Fig. 5 and Fig. S5). This suggests that the interaction with RNA in these conditions was somewhat transient and did not lead to the formation of a rigid 3D structure.
Fluorescence measurements were used to quantify the overall strength of the ALYREF1–155 and ORF578–120 interaction in the absence and presence of a specific fragment of viral RNA. Both protein constructs possess tryptophan residues, one of these (W108 of ORF57) forms part of their binding interface, and is buried upon protein complex formation. Control experiments showed that ALYREF1–155 and ORF578–120 both have a fluorescence intensity maximum at 355 nm which is not shifted by 14merS RNA addition (however, ALYREF samples become turbid due to non-specific aggregation, complicating measurements). The formation of equimolar ALYREF1–155 - ORF578–120 complex leads to a blue shift of the emission maxima of the sample (Fig. S7C), in agreement with burial of the tryptophan sidechain in a hydrophobic environment. The blue shift, quantified by λbcm, becomes more pronounced with increasing concentrations of equimolar ALYREF1–155 - ORF578–120 in the sample, allowing an estimation of the apparent Kd for this interaction as 2.56±0.20 µM (Fig. S7C). Addition of 14merS to pre-mixed 10 µM equimolar ALYREF1–155 - ORF578–120 complex caused both a decrease in fluorescence intensity (, reflecting the change from binary protein-RNA to ternary complex formation), and a blue signal shift (, reflecting the change from binary protein-protein to ternary complex formation). Non-linear fit of the two dependencies together to the three-equation equilibrium model using DynaFit software [43] (see Fig. 6A) allowed an estimation of the apparent macroscopic Kd for the RNA binding to ORF57-ALYREF as 1.55±0.24 µM. The overview of Kd's determined for this simplest equilibrium model for ternary complex formation [44] is presented on a thermodynamic cycle in the inset of Fig. 6B. The estimated value of KdOR+A = 0.52 µM for binding of ALYREF to ORF57-RNA complex can be inferred from the thermodynamic equilibrium considerations [44]. As the measured Kd values for the formation of binary complexes are significantly higher than for the ternary complexes (e.g., KdO+R = 7.57 µM>KdOA+R = 1.55 µM), the assembly shows clear-cut cooperative behavior [44]. These results therefore reveal that the ternary complex formation, leading to introducing RNA to ALYREF, is thermodynamically driven by the overall cooperativity. Further analysis of this simplest equilibrium model using COPASI simulations illustrates a dramatic increase in the population of ALYREF molecules bound to RNA when the ORF57 is present (Fig. 6B). We have also run COPASI simulations for an extended binding model, where a very weak non-specific binding of RNA to ALYREF (with estimated Kd>50 µM, see above) is taken into account (Fig. 6C), and two different Kd values for non-specific binding are assumed for calculations as examples. The COPASI simulations for each model demonstrate that the presence of ORF57 in stoichiometric amounts significantly increases the concentration of ALYREF in complex with RNA (i.e., [OAR]), compared to the background level of non-specific ALYREF-RNA complex (i.e., [AR], Fig. 6B,C). Even with the most conservative estimates (assuming the lowest value of Kd = 50 µM for non-specific ALYREF-RNA binding), for the concentrations used in this example the amount of virus-specific RNA in complex with ALYREF increases more than 3.6 times.
Interestingly, adding a large excess of RNA to the 2.5 µM ALYREF-ORF57 mixture displayed a more complex behavior of signal shift (Fig. S7D): the blue shift observed with only a small excess of RNA was partially reversed, consistent with protein-protein complex dissociating at higher RNA excess. Intuitively this result is expected if RNA over-saturates the binding site on ORF57, competitively displacing ALYREF from R-b helix. This competitive behavior at very high [RNA] cannot be adequately described by currently parameterized simple equilibrium models, such as shown on Fig. 6B,C, which only account for cooperativity.
The experimental fluorescence equilibrium binding data thus reveal the overall cooperativity in the ternary complex formation when the components are present at or near stoichiometric amounts, and support a role of ORF57 as an adaptor introducing RNA to ALYREF. The fluorescence data are also consistent with a local competitiveness of ORF57-ALYREF and ORF57-RNA interactions: this competitiveness becomes apparent at macroscopic (i.e., molecular) level only if RNA is in significant excess. These observations fit well with the NMR experiments which show the ability of ALYREF to partially displace RNA from R-b helix of ORF57, while forming ternary complex.
The results of NMR mapping of RNA binding regions of ORF578–120 were confirmed by UV cross-linking using purified protein and radio-labeled RNA oligonucleotide, performed as previously described [10]. The ORF57 mutants Y81A+R82A, R88A+F89A and W108A+R111A+V112A all significantly reduced the efficiency of cross-linking with RNA 14merS (Fig. 7A). Substitution of residues W108,R111,V112, which are the most important for ALYREF binding [30], also has the strongest reductive effect on RNA binding, confirming independently that the RNA- and ALYREF-binding sites overlap. The control mutation D110A+E114A marginally increases the efficiency of RNA cross-linking, this is likely to be due to reduction in electrostatic repulsion between this protein mutant and RNA.
To independently confirm the NMR observations in regard to RNA oligonucleotide binding with ORF578–120, ALYREF1–155 and their complex, we performed in vitro reconstitution assays followed by UV cross-linking experiments. ORF578–120 showed a strong RNA-binding activity for 7merS and 14merS in sharp contrast to GST-ALYREF which bound weakly with both RNAs (Fig. 7B,C). When ORF578–120 was incubated with RNA prior to mixing with GST-ALYREF, followed by GST affinity purification of the resulting complexes and UV cross-linking, there was a drastic reduction of the RNA cross-linked to ORF578–120 and a concomitant increase in the RNA cross-linked onto GST-ALYREF. Therefore the RNA-binding activity of ORF578–120 is severely reduced upon interaction with ALYREF, whereas the amount of RNA in contact with ALYREF increases in the ternary complex. This independent data obtained at a molecular (i.e, macroscopic) level concurs fully with the NMR data obtained at a residue-specific level of detail. Interestingly, the increase in ALYREF-RNA cross-linking efficiency observed experimentally in the presence of ORF57 fits well with the numerical estimates using COPASI simulations shown on Fig. 6C.
The combination of structural and interaction data allows us to suggest a model that explains how the adaptor protein ORF57 from HVS introduces viral mRNA to cellular mRNA export factor ALYREF, functioning as a molecular “hijacker”. As previously noted, ALYREF itself binds mRNA weakly and non-specifically, and cannot discriminate between cellular and viral transcripts, needing other proteins to recruit mRNA and strengthen this binding to a functionally significant level. This non-specific binding can be observed and mapped here by weak saturation transfer from RNA to protein (Fig. 1). In its free form the N-terminal region aa8–120 of ORF57 is flexible and mainly unstructured, apart from the short α-helix aa108–118 which we named R-b helix. It is anticipated that the positively charged region aa61–120 interacts transiently with the negatively charged part of the ORF57 polypeptide chain aa12–28, keeping the molecule in a loosely “closed” conformation. When the specific viral mRNA motif binds, it is recognized by the extensive ORF57 region aa64–120 which comprises R-b helix (see Fig. 8). On its own, R-b helix is unable to bind RNA, but its presence, as well as the presence of the flanking region, are essential. It can be envisaged that the RNA-binding region of ORF57 may form a hairpin or other compact structure to offer an extensive network of contacts recognizing and holding the viral RNA molecule (Fig. 8).The fragment aa64–120, which is rich with arginines, serines and aromatic residues, forms direct contacts with RNA, but without forming a stable 3D structure, and this RNA binding is expected to release the negatively-charged N-terminal part of ORF57. ORF57 is not able to recognize and bind mRNA which lacks specific viral motifs, ensuring that this viral adaptor selects only viral transcripts for further export. Although the R-b helix of ORF57 participates in recognition and binding of viral mRNA, it has much higher affinity for ALYREF binding. Therefore in the presence of ALYREF (see Fig. 8), the R-b helix is released from RNA and binds the RRM domain of ALYREF instead. However at that point the adjacent flexible regions of both ORF57 and ALYREF, which are also involved in RNA binding, are brought together, and the viral RNA molecule is not released but is held in place by the synergetic action of these flanking regions (Fig. 8). The overall cooperativity of ternary complex assembly is demonstrated by the fluorescence measurements and supported by the remodeling assay, while the local competitiveness of RNA and ALYREF binding to R-b helix is directly demonstrated by the NMR data, and additionally supported by the fluorescent measurements showing ternary complex dissociating when over-titrated with RNA. The RNA binding within the stoichiometric ternary complex is mapped using direct NMR measurement of spatial proximity of individual amino acid residues of both ORF57 and ALYREF to RNA (Fig. 8). When the ternary RNA-ORF57-ALYREF complex is formed, the N-terminal regions of both ORF57 and ALYREF in contact with RNA retain significant flexibility (as evidenced by NMR signal shapes and lack of RNA-protein NOEs). The main NXF1-binding region of ALYREF aa15–36 [5], [10] partially overlaps with region involved in viral mRNA binding, and remains sufficiently exposed. This allows us to speculate that at the next stage of the pathway the NXF1 would bind to this region, partially displacing the viral mRNA, which however will be held in the vicinity by the rest of the binding site, formed by synergetic interactions of ALYREF and ORF57. ALYREF binding would then help switch NXF1 into the high-affinity RNA binding mode, forcing it to accept viral mRNA for export via the nucleopore, using the host pathway. We suggest that the presence of partially-overlapping binding sites, and a combination of competitive interactions at the level of specific sites, and cooperative binding at the macroscopic level, may provide a general molecular mechanism for targeted successive RNA transfer between protein molecules along the export pathway.
The functional role of the viral protein ORF57, to introduce specific viral RNA to the cellular protein ALYREF, thereby facilitating the export from the nucleus of unspliced viral mRNA via the NXF1-dependent cellular pathway, is fulfilled via several critical properties. First, ORF57 must recognize a specific viral mRNA, which contains characteristic sequence motifs, and ignore cellular mRNA. Second, it should interact with cellular ALYREF, and induce binding of a specific viral mRNA to ALYREF, allowing ALYREF then to pass it over to NXF1 at the next stage of the pathway. Therefore, the ability to recognize and reversibly bind RNA is crucial for correct and efficient functioning of viral ORF57 as a “molecular highjacker” of the host cell pathway. Here, we have reconstituted in vitro and studied in detail the function of the core responsible for the recruitment of viral RNA to ALYREF, using the essential protein fragments of ORF57(aa8–120) and ALYREF(aa1–155) which interact with each other and with viral RNA. A range of experimental approaches was used to reveal the molecular mechanism of viral RNA recognition and following transfer from ORF57 to ALYREF with atomic-to-residue-level resolution.
The position of viral mRNA binding sites on ORF57 was previously broadly localized to aa8–120 [20], [21]. This region is largely unstructured, and contains multiple arginines between residues 62 and 120 which may potentially mediate RNA binding, although this was not confirmed previously. Here, we used NMR to characterize the binding sites more precisely. To directly identify the residues in proximity to RNA we employed RNA→protein cross-saturation transfer experiments [37], [39] which revealed two main ORF578–120 regions in direct contact with RNA, aa107–120 and aa81–92, with residues from aa94–105 and aa64–79 also contributing. Although the general “polyelectrostatic effect” [45] is expected to attract mostly negatively-charged RNA to positively-charged regions and thus explain the weak RNA-binding affinity of ALYREF, the amino acid sequence determinants of a sequence-specific RNA recognition by viral ORF57 are still to be explored. It should be noted that the specific RNA recognition by flexible charged polypeptide regions is not unprecedented, and similar examples have been described in the literature [46]–[49]. The mapping of RNA binding regions using NMR is also fully supported by the biochemical data. Mutations of selected residues (Y81+R82, R88+F89 and W108+R111+V112) to alanine within these regions reduced the efficiency of UV RNA-ORF578–120 cross-linking, confirming their involvement in mediating RNA binding (Fig. 7A). Moreover, earlier we probed the physiological effect of a number of mutations in this region via an ex vivo assay for cytoplasmic accumulation of an HVS ORF47 reporter mRNA, which reflects the ability of full-length ORF57 (bearing site-specific mutations) to form an export competent ribonucleoprotein particle [30]. Mutations within R-b helix of W108A, R111A, V112A, R119A and R120A and their combinations all substantially reduced the efficiency of the mRNA cytoplasmic accumulation, which was previously interpreted as a confirmation of the functional significance of ORF57 – ALYREF interaction for ORF57-mediated nuclear export of viral mRNA [30]. In view of the new experimental data, the same R-b-helix region is also directly involved in viral mRNA binding, meaning that the physiological effect of these mutations cannot be solely attributed to blocking ORF57 – ALYREF interactions, but these mutations will likely affect the whole process of how viral mRNA is recognized and introduced to ALYREF. Additionally, mutations R79A+V80A and R94A+I95A in the flanking region also reduced noticeably the cytoplasmic accumulation [30], which now can be explained by the involvement of this region in mRNA binding. These previously obtained physiological assay data obtained with full-length proteins thus corroborate the functional importance of the molecular regions characterized here in detail using shorter molecular fragments. Future in vivo studies using all these ORF57 mutations introduced in live HVS, and monitoring their effect on the process of cell infection and time-dependent localization of interacting components, are expected to clarify further the order of binding events mediated by the R-b-helix and flanking RNA-binding regions in the context of live cell. The mutations most detrimental to HVS are expected to involve residues situated on the R-b helix which participate in both ALYREF and viral RNA binding. The results presented here provide a map for further extensive mutational analysis, and a framework for its functional interpretation in vivo.
The ALYREF-ORF57 structure determined here shares some similarities with the ALYREF-ICP27 structure [30] (Fig. 3) as both viral peptides make similar contacts with a patch on ALYREF's surface. There are however clear differences between the ALYREF-ICP27 and ALYREF-ORF57 complexes, not discovered in the earlier signal perturbation mapping analysis [30]. The ORF57 fragment is helical and mainly contacts the looped side of ALYREF, whereas ICP27 has an extended conformation and stretches along the groove formed by α-helices of ALYREF RRM (Fig. 3). As a result, the ALYREF-ORF57 complex is superficially more similar to U2AF homology motif (UHM) recognition [50] of Trp containing peptides (Fig. 3D), although ALYREF lacks the signature Arg-X-Phe UHM interacting motif [51]. It is therefore apparent that variations in sequences and local structures of viral adaptor proteins can achieve similar binding with the promiscuous RRM domain of ALYREF [52]. This may explain the lack of an obvious conserved “ALYREF-binding motif” in other herpes viral adaptor proteins, such as KSHV ORF57 or EBV EB2 [17], [18], [53].
We previously suggested a recognition triad for ALYREF, namely W105, R107, L108 in ICP27 and W108, R111, V112 in ORF57 [30]. The functional role of this triad in ICP27-ALYREF binding in vivo, and for efficiency of viral mRNA export and HSV-1 production was also studied in detail recently [54]. In light of the new structural data presented here, the importance of R113ORF57 should also be emphasized as it plays a similar role to that of R107ICP27 [30]. The quantitative differences in mutational effects of the triad residues W105AICP27/W108AORF57 and L108AICP27/V112AORF57 on binding with ALYREF [30] can now be explained by the subtle differences in their structural context. The structure of ORF57-ALYREF complex also provides an explanation for the weak affinity of a short ORF57 aa105–115 peptide and ORF578–120 double mutant R119A+R120A [30], as both constructs are likely to disrupt helix formation and increase the entropic cost of binding.
Using the combination of traditional atomic-resolution structure determination and novel saturation-transfer experiments we were able to follow the process of RNA transfer from ORF57 to ALYREF in a site-specific manner, and suggest a model of how the ternary complex is assembled (Fig. 8). It is interesting that the signal intensity of the N-terminal region aa22–48 of ALYREF does not reduce significantly upon RNA binding in the presence of ORF57, suggesting that this interaction is relatively transient. Overall, we conclude that ORF57 simply bridges the interaction between viral RNA and cellular ALYREF, cooperatively enhancing the formation of the ternary complex, without allosterically remodeling ALYREF. The overall cooperativity of the ternary complex formation is demonstrated by the quantitative Kd measurements for the different complexes within the thermodynamic cycle (Fig. 6B). The cooperativity of interactions explains the partial RNase sensitivity of the ALYREF-ORF57 complex reported previously [21]. The same cooperativity may also explain why the presence of ORF57 in the nucleus of an infected cell does not cause indiscriminate export of non-viral mRNAs which lack the specific viral sequence motif. The estimates using NMR and fluorescence experiments showed that the apparent Kd of binding of specific RNA oligonucleotide to equimolar ALYREF1–155 – ORF578–120 complex (1.55 µM) is lower than to ALYREF1–155 (>50 µM) or ORF578–120 (7.57 µM) individually. The inferred Kd of ALYREF binding to ORF57-RNA complex (Fig. 6B) is 0.52 µM, two orders of magnitude tighter than for ALYREF-RNA binding. These estimates support that the ternary complex overall is stabilized cooperatively, despite the presence of local competition between RNA and ALYREF for ORF57 R-b helix region aa106–120.
It is notable that the values of NMR signal shifts within the flexible regions were not a good indicator of the formation of transient complexes with RNA. It is likely to be due to a substantial fluidity of the complex leading to extensive chemical shift averaging over the conformational ensemble; however the saturation transfer from RNA protons to protein amides served as a more reliable indicator of local binding. The ternary complex formed here may present a good example of fuzzy complexes, the existence of which was postulated recently [55], [56]; more specifically, it would fit the “flanking” model, where the short R-b helix acts as a clamp forming more rigid part of the complex interface, whereas transient interactions within flexible flanking regions contribute to the overall stability of the complex. Such a mode of recognition, using fairly short linear motifs located within flexible regions of viral proteins, is expected to provide evolutionary advantages for quick adaptability of viruses [56]–[59], and may fit well with a necessity to bind and remodel NXF1, and dismantle the RNA-ALYREF-ORF57 complex at the next step of the viral mRNA export pathway.
The protein constructs used here comprise the main binding elements for the assembly of the specific ORF57-RNA-ALYREF ternary complex, which is responsible for introducing the herpesviral mRNA to the cellular export factor ALYREF. However, both native ALYREF and ORF57 contain additional regions which may contribute to the binding of longer viral mRNA molecules, adding to the overall cooperativity of this assembly, and strengthening it further. The viral mRNA transcripts, which are much longer than oligos used in the current study, would provide additional contact points for RNA-binding regions within the C-terminal regions of both full-length ALYREF [7], [8] and ORF57 [23]. Therefore, the quantitative measurements of binding reported here for the essential core of the ternary complex provide only a lower affinity estimate for the full-length complex. We speculate that once the ternary ORF57-mRNA-ALYREF complex encounters NXF1-p15, viral mRNA will be displaced from the N-terminus of ALYREF where NXF1 binds in its place [5], [10]. Viral mRNA at that moment will still be retained by ALYREF's RRM domain together with ORF57, presenting it to NXF1. Binding of ALYREF switches NXF1 into a high-affinity RNA-binding mode [5], [6], forcing it to accept the foreign viral mRNA and commit it to export to the cytoplasm. As indicated earlier, the adaptor proteins homologous to HVS ORF57 (ICP27 in HSV-1, ORF57 in KSHV) are expressed by all herpesviridiae. The location of RNA binding sites within these proteins is poorly conserved, and the exact location of binding sites with cellular adaptors such as ALYREF [17] and UIF [19] often is unknown or not evident due to the lack of recognizable sequence motifs responsible for such binding. Even when superficial similarity exists, as in the case of ALYREF recognition triad residues suggested earlier for HSV-1 ICP27 and HVS ORF57 [30], here we showed that in fact the structural details of the molecular recognition are significantly different, despite binding occurring in the same cleft on the surface of the RRM of ALYREF. This finding means that it is probably too early for modeling and predictions to be used to discover the binding interfaces between RNA and viral and cellular proteins, and detailed experimental structural studies need to be continued for this molecular pathogen-host system, with different herpesviruses using diverse strategies for molecular recognition to achieve a similar functional outcome, such as viral mRNA export. Continuation of similar studies for signature viral adaptors from more medically relevant herpesviruses, such as HSV-1 or KSHV involved in cancer, may possibly identify new drug targets for novel treatments.
This work for the first time suggests a detailed mechanism for the assembly of the key ternary RNA-ORF57-ALYREF complex leading to herpesvirus highjacking the host nuclear export pathway. We show the importance of partially-overlapping multifunctional binding sites and combination of competitive and cooperative binding events as a likely mechanism for the orderly assembly and disassembly of mRNA nuclear export complexes and molecular transfer, adding to the emerging knowledge in this area [35], [36], [60], [61].
All proteins were expressed in E. coli BL21-RP cells (Novagen). For NMR studies, murine ALYREF (also called REF2-I) isoform constructs aa1–155 (ALYREF1–155), aa54–155 (ALYREF54–155), as well as ORF5756–140 and ORF578–120, were produced as described previously [30]. ALYREF protein constructs are identical to protein REF used in [30]; only name was changed due recent gene naming conventions [31]. ORF57103–120 peptide was produced as a GST-fusion construct as described for ICP27103–138 [30]. Post gel filtration, all samples were buffer exchanged using an Amicon ultrafiltration cell to the NMR buffer (20 mM phosphate pH 6.2, 50 mM NaCl, 50 mM each of L-Arg, L-Glu and β-mercaptoethanol, 10 mM EDTA); L-Arg and L-Glu were added to reduce protein aggregation and improve sample stability [62]. Unlabelled synthetic peptide ORF57103–120 was obtained from Peptide Protein Research Ltd (UK). For UV cross-linking, hexa-histidine ORF578–120, hexa-histidine GB1 and GST-ALYREF fusions were purified by affinity chromatography and dialyzed against RB100 buffer (25 mM HEPES pH 7.5, 100 mM KOAc, 10 mM MgCl2, 1 mM DTT, 0.05% Triton, 10% Glycerol).
All experiments were carried out at 30°C on Bruker DRX600, DRX700 and Varian Inova 800 MHz spectrometers equipped with cryoprobes, and a Bruker DRX800 with a room temperature probe. Standard triple-resonance experiments were used to assign spectra: ORF57103–120 in free form and with a 3-fold excess of unlabelled ALYREF54–155 added, ORF5756–140 in the free form, and ALYREF54–155 with a 3-fold excess of ORF57103–120 synthetic peptide added. ALYREF54–155, ALYREF1–155 and ORF578–120 were assigned previously [10], [30]. Spectra were processed using NMRpipe [63] and Topspin 2.1 (Bruker) and analyzed using Sparky (University of California). Distance restraints obtained from 3D 15N- and 13C- edited NOESY-HSQC experiments (τm 130 ms) and dihedral restraints from TALOS+ [42] were used in structure calculations by CYANA [64]. Additionally, intermolecular contacts were unambiguously identified using 13C-edited, 12C-filtered NOESY-HSQC (τm 150 ms) spectra acquired on a Varian Inova 800 MHz spectrometer. In this experiment only NOE crosspeaks between 1H-13C moieties of 13C,15N-labelled ALYREF and 1H(12C) of unlabelled ORF57 were observed [65], [66]. A final ensemble contained 20 structures with lowest target function values. Images were generated using Pymol (DeLano Scientific). Chemical shift assignments were submitted to the BioMagResBank for free ORF57103–120 (bmr17664), free ORF5756–140 (bmr17663) and the ALYREF-ORF57 complex (bmr17693). Structure coordinates and experimental constraints for ALYREF-ORF57 complex were deposited into the Protein Data Bank (2yka). Ramachandran plot statistics for residues in most favored regions, additional allowed regions, generously allowed regions, disallowed regions calculated for structured ALYREF74–152+ORF57106–120 are: 79.8%, 20.2%, 0%, 0%.
IDIS-TROSY spectra [38] were acquired using 1∶1 mixtures of 0.4 mM 13C,15N-labelled ORF578–120 and 15N-labelled ALYREF1–155 or ALYREF54–155, followed by additions of RNA. RNA oligonucleotides were obtained from Sigma. Two oligos contained the ORF57-specific motif [32] GAAGAGG (7merS) and CAGUCGCGAAGAGG (14merS), and two were non-specific CAGUCGC (7merN) and CAGUCGCAUAGUGCA (15merN; this oligo is identical to that used previously [10]). Irradiation of resolved RNA signals [39] in cross saturation transfer (ST) [37] versions of standard Bruker-library HSQC, TROSY and IDIS-TROSY [38] was achieved by using a selective Gaussian pulse train (lasting 0.7 s in total) using a series of 8.5 ms 180 degree pulses (B1 = 60 Hz). The saturation pulse train was tagged at the end of relaxation delay of 2.3 s, immediately prior to the first hard proton pulse. RNA peaks were selectively irradiated by centering the pulse train at 5.85, 5.75 or 12.00 ppm frequencies, with off-resonance irradiation at 21 ppm. Ratios of amide signal intensities in equivalent spectra, obtained with saturation at two different frequencies freq1 and freq2 (as indicated) were obtained. Residues were highlighted as close to RNA in space if the ratio of signal intensities Ifreq1/Ifreq2 differed significantly from unity (by more than three standard deviations (SD), calculated from the Ifreq1/Ifreq2 variability observed within non-interacting regions).
RNA oligonucleotides were end-labelled with [γ32P]-ATP using Polynucleotide Kinase (Fermentas). UV cross-linking with proteins was performed as previously described [10]. For the in vitro reconstitution assay, 10 or 100 µg ORF578–120 was incubated with 5 µg radiolabelled and cold RNA (7merS or 14merS) at room temperature for 10 minutes. The mixture was added to 20 µg of GST-tagged full length ALYREF (aa1–218) immobilized onto Glutathione-coated beads (GE Healthcare) in RB100 buffer. Beads were washed and complexes were eluted in native conditions (50 mM TRIS pH 7.5, 100 mM NaCl, 40 mM reduced glutathione) before being subjected to UV-irradiation or not. Proteins were resolved on 15% SDS-PAGE stained with Coomassie blue and analyzed by PhosphoImaging.
Purified proteins were transferred into buffer F (20 mM phosphate pH 6.2, 50 mM NaCl, 50 mM L-Arg+L-Glu, 5 mM EDTA, 1 mM TCEP) by 3 overnight dialysis steps and then concentrations determined by UV absorption (280 nm). Measurements were carried out on a Varian Cary Eclipse fluorimeter, with excitation at 280 nm and emission monitored over 290–600 nm at a scan rate of 120 nm/min. Titrations were carried out with at least 1 min of equilibration time after each addition. ORF578–120-RNA (OR) titrations were performed using 13 µM ORF57, titrations of 1∶1 ORF57-ALYREF (OA) with RNA used an initial protein concentrations of 2.5 and 10 µM. Blue shift in emission maximum was caused by protein-protein (OA and OAR) complex formation and was quantified as a change in barycentric mean values (“centre of mass” of the peak), where λ is emission wavelength (320–390 nm) and Iλ is fluorescence intensity at this wavelength. Binding of RNA to ORF57 (in OR and OAR complexes) was quantified by measuring a decrease in integral fluorescent intensity . Apparent macroscopic dissociation constants Kd for binary complexes were obtained by non-linear regression fit of and dependences on the total concentration of added component, using either a standard quadratic equation, or DynaFit software (BioKin Ltd) [43] which produced the same results. Value of apparent macroscopic Kd for ternary complex formation was obtained by titrating RNA (14merS) to 10 µM 1∶1 ALYREF-ORF57 mixture, followed by simultaneous fitting of the associated changes in and to the non-redundant three-equation equilibrium model (Fig. 6A) using DynaFit [43]. Changes in normalized fluorescence parameters, caused by the increase in [OA] or/and [R], were related with concentrations as , and , with the values of the response coefficients a, b, c and d and normalization factors n and m obtained during the nonlinear fit (so that a = 1≈b and c = 1≈d, and all concentration expressed in µM units). Further simulations of equilibrium reactions within different binding models were conducted using COPASI software [67].
|
10.1371/journal.pntd.0002757 | Gentamicin-Attenuated Leishmania infantum Vaccine: Protection of Dogs against Canine Visceral Leishmaniosis in Endemic Area of Southeast of Iran | An attenuated line of Leishmania infantum (L. infantum H-line) has been established by culturing promastigotes in vitro under gentamicin pressure. A vaccine trial was conducted using 103 naive dogs from a leishmaniosis non-endemic area (55 vaccinated and 48 unvaccinated) brought into an endemic area of southeast Iran. No local and/or general indications of disease were observed in the vaccinated dogs immediately after vaccination. The efficacy of the vaccine was evaluated after 24 months (4 sandfly transmission seasons) by serological, parasitological analyses and clinical examination. In western blot analysis of antibodies to L. infantum antigens, sera from 10 out of 31 (32.2%) unvaccinated dogs, but none of the sera from vaccinated dogs which were seropositive at >100, recognized the 21 kDa antigen of L. infantum wild-type (WT). Nine out of 31 (29%) unvaccinated dogs, but none of vaccinated dogs, were positive for the presence of Leishmania DNA. One out of 46 (2.2%) vaccinated dogs and 9 out of 31 (29%) unvaccinated dogs developed clinical signs of disease. These results suggest that gentamicin-attenuated L. infantum induced a significant and strong protective effect against canine visceral leishmaniosis in the endemic area.
| A 24 month vaccine trial was conducted using 103 leishmania free dogs in an area of southeast Iran endemic for visceral leishmaniosis. The dogs were vaccinated with gentamicin-attenuated line of Leishmania infantum. No local and/or general indications of disease were observed in the vaccinated dogs immediately after vaccination with an attenuated line of Leishmania infantum. Nine out of 31 (29%) unvaccinated dogs, but none of those vaccinated, were positive for presence of Leishmania DNA by PCR. In western blots, sera from 10 out of 31 (32.2%) unvaccinated dogs, but none of the sera from vaccinated dogs, recognized the 21 kDa antigen of Leishmania infantum wild-type. One out of 46 (2.2%) vaccinated dogs and 9 out of 31 (29%) unvaccinated dogs developed clinical signs of disease. The attenuated Leishmania infantum induced a significant and strong protective effect against Leishmania infantum infection in the field.
| Leishmania infantum (L. infantum) is a causative agent of visceral leishmaniasis (VL), which is a severe and frequently lethal protozoan disease of humans and dogs. Canine visceral leishmaniosis (CVL) is widely distributed in large areas of Europe, South America, the Middle-East, Central Asia, China, and Africa, particularly in the countries of the Mediterranean Basin [1], [2]. In Iran, at least seven endemic foci in dogs have been identified including the Baft district in the southeast of the country where there is a high seroprevalence in domestic dogs [3]. Dogs are the principal reservoir of L. infantum and can be an important threat to public health. Control of the disease in dogs has been shown to reduce the human incidence [4], [5]. Although there have been a number of vaccine trials, there is currently no effective and completely safe vaccine against any form of leishmaniasis. A successful vaccine against Leishmania is most likely to be either an attenuated line or a subunit vaccine based on antigens with demonstrable protective function [6], [7]. Subunit and attenuated vaccines can be highly effective and induces protection against pathogen [8], [9], [10]. We previously reported that a cultured attenuated line of L. infantum, identified as L. infantum H-line, was selected by culturing promastigotes in vitro under pressure of gentamicin [11]. Gentamicin, which has frequently been added to cultures of Leishmania to prevent bacterial contamination [12], [13], is an aminoglycoside that interacts with RNA in prokaryotic cells [14]. The precise mechanism of bactericidal activity of aminoglycosides is not fully understood, but some hypotheses include disruption of ribosomal activity by breaking up polysomes, misreading of mRNA during protein synthesis and production of abnormal or nonfunctional proteins. Comparative proteomics profiling of the attenuated line identified key changes in parasite thiol-redox metabolism [15]. Thiol-redox metabolism is crucial for Leishmania which is exposed to an oxidative burst when they encounter their mammalian macrophage host cell [16]. L. infantum H-line is more susceptible to oxidative stress, and thus a change in thiol-redox metabolism in this line may explain its loss of virulence [15]. L. infantum H-line invaded but was unable to survive within bone marrow derived macrophages of BALB/c mice in vitro [11]. Moreover, the attenuated line failed to spread to, and within, the visceral organs of BALB/c mice and dogs over a 12 week observation period [17]. Immunohistochemical investigation showed no parasites in the popliteal lymph node (PLN) of immunized dogs whereas there were parasites in the PLN of 60% of dogs infected with L. infantum WT [18]. No clinical signs and histopathological abnormalities were found in the dogs immunized with the attenuated line of parasite over 2 years post-immunization [17], [19]. Dogs immunized with the attenuated line parasites elicited a Th1 response and were protected against experimental CVL [19]. We previously reported that Western blot analysis of antibodies to the 21 kDa antigens of L. infantum H-line and WT might be a useful technique for distinguishing between dogs vaccinated with L. infantum H-line and dogs naturally infected with L. infantum WT in epidemiologic studies [20].
In the present study, for the first time, we show the impact of L. infantum H-line vaccine against natural infection in dogs in a highly endemic area of Iran over a 24 month follow-up.
Promastigotes of L. infantum JPCM5 (MCAN/ES/98/LIM-877), were cultivated in complete haemoflagellate minimal essential medium (HOMEM) (GIBCO) supplemented with 10% (vol/vol) heat-inactivated fetal calf serum (HI-FCS) (Labtech International). L. infantum H-line was generated in the same medium supplemented with 10% (vol/vol) HI-FCS and gentamicin at 20 µg/ml (Sigma) [11]. Stationary phase promastigotes of the attenuated line were harvested after 48 subpassages and a suspension at a concentration of 5×108 cells/ml in PBS was prepared.
The field study was conducted in 3 villages, Dehsard, Khosrowabad and Dehsarar of Baft County (56.2147°E, 28.2727°N), Kerman Province, in the southeast of Iran (Fig. 1a). The area has a desert climate and the total annual rainfall is 309 mm with a minimum of 3 mm in July and maximum of 120.9 mm in April. The minimum mean monthly relative humidity is 26% (June) and the maximum is 56% (January). Initially, 77 household dogs were examined for clinical signs of the disease and tested for the presence of specific anti-Leishmania antibody by an immunofluorescence assay (IFA).
A vaccine trial was conducted on 103 dogs (55 vaccinated and 48 unvaccinated). The protocol for vaccination of the dogs had been reviewed and approved by the Medical Ethics and Animal and Use Care Committee of the Kerman Medical University (study protocol number KA/89/15), in accordance with the Guide for the Care and Use of Laboratory Animals Eighth Edition. The animals were kept under typical local conditions of food and housing and sampled with the owners' consent. All dogs with clinical signs of disease were sacrificed to avoid unnecessary suffering.
On the basis of the rate of seropositivity detected in the 77 dogs tested (see above), and to anticipate a number of dogs being lost during the follow-up period of 24 months, a total 103 dogs were used in this study. At follow-up prior of starting, we expected 60% seropositive in the unvaccinated group and 20% seropositive in vaccinated group, with a confidence level of 0.95 and power of 0.9, and ratio of 1.29 sample size (i.e. n2/n1 equivalent to 45/35). This estimated a sample of 31 in unvaccinated group and 40 in vaccinated group. Ten unvaccinated dogs and 13 vaccinated dogs were considered to lose at follow-up prior in this study. One hundred and three healthy male German shepherd cross dogs from non-endemic areas (Kerman city 225 Km northwest of Dehsard) (Fig. 1b) between 6–18 months old were used. All of the animals were negative for presence of leishmanial DNA and serum specific anti-Leishmania IgG antibody. The dogs had previously been vaccinated against canine parvovirus and rabies and were also treated with the anthelmintic drugs praziquantel and pyrantel. The weight and age were recorded for each dog and they were randomly divided into 2 groups (55 dogs in vaccinated group and 48 dogs in unvaccinated group). The dogs in the vaccinated group were injected subcutaneously (s.c) with 100 µl of the suspension of stationary phase promastigotes in the foreleg of the animals. The unvaccinated dogs were injected subcutaneously with 100 µl of PBS also in the foreleg. The dogs were transferred into the endemic area over a period of 1.5 months before June 2010 and re-homed at households within the endemic area. Information about the risk of the procedures was given to persons who became owners of dogs. We included vaccinated and control dogs in each house whenever possible, in order to match their degree of exposure to natural infection.
The dogs were followed up over 24 months, from June 2010 to cover four sandfly seasons, which occur in June and September in the endemic areas of the southeast of Iran [21]. The efficacy of the vaccine was evaluated by clinical examination and serological and parasitological analyses. Active disease surveillance measures were implemented in each of the study villages. A trained worker was located in the Health House in the village, and together with our team had responsibility for disease monitoring. The follow-up was performed at 3, 6, 9, 12, 18, 20 and 24 months after starting trial. The peripheral blood samples were taken for complete blood cells (CBC) count and biochemical parameters including serum total protein, serum albumin and serum globulin. The clinical signs of disease were classified according to a modified version of leishvet guidelines as described previously [22]. Briefly, Stage I, Mild disease, animals exhibiting peripheral lymphadenomegaly or papular dermatitis, creatinine <1.4 mg/dl, non-proteinuric, negative to low levels of antibody. Stage II, Moderate disease, animals, which apart from the signs listed in stage I, may exhibit diffuse or symmetrical cutaneous alterations (exfoliative dermatitis/onychogryphosis, ulcerations (planum nasale, footpads, bony prominences, mucocutaneous junctions), anorexia, weight loss, fever, and epistaxis, mild non-regenerative anemia, hyperglobulinemia, hypoalbuminemia, normal renal profile, creatinine <1.4 mg/dl, low to high levels of antibody. Stage III, Severe disease, animals which apart from the signs listed in stages I and II, may exhibit signs originating from immune-complex lesions such as vasculitis arthritis and glomerulonephritis, chronic kidney disease, creatinine 1.4–2 mg/dl, medium to high levels of antibody. Stage IV, Very severe disease, animals which, apart from the signs listed in stages I, II and III, may exhibit signs pulmonary thromboembolism, or nephrotic syndrome and end stage renal disease (creatinine >5 mg/dl) with medium to high levels of antibody. Twenty six out of 103 (25.2%) dogs left or died from a disease unrelated to leishmaniosis over the 24 month period follow-up. All unvaccinated dogs were sacrificed by intravenous injection of thiopental sodium 33% (5 ml/kg) [23] at the end of the study. Five ml of peripheral blood were taken from the foreleg vein of each dog in EDTA for isolating parasite DNA for PCR test, and 2 ml for separation of sera for IFA and Western blotting test. The samples were stored at −20°C.
Specific anti-Leishmania total IgG antibody was measured by IFA as described previously [24]. Briefly, slides were coated with washed promastigotes of L. infantum WT and air dried. Slides were incubated with twofold dilutions of the test samples in humid moist conditions for 60 min at 37°C. Excess antibody was washed off the slides and bound antibody was detected using fluorescein isothiocyanate (FIT) conjugated sheep anti-dog IgG (Sigma), diluted at 1∶400 in PBS-0.01% Evans blue. Specific anti-Leishmania IgG1 and IgG2a antibodies were measured by ELISA as described previously [19]. Briefly, serum was prepared from the clotted blood of dogs at 20-month post follow-up and stored at −20°C. Each well of flat-bottom microtitre plates was coated with 1 µg of soluble Leishmania antigen in 0.1 M carbonate buffer pH 9.6 and incubated at 4°C overnight. Following 3 washes in PBS (pH 7.4) containing 0.05% Tween-20, the plates were blocked with 200 ml of blocking buffer (1% BSA in PBS), and incubated at 37°C for 1 h. After 3 washes, 100 µl of serially diluted serum sample (1∶100 starting dilution in PBS/BSA 1%) was added to the wells and incubated for 2 h at 37°C. Bound antibodies were detected by 50 µl/well goat anti-dog IgG1 conjugated to HRP at 1∶500 dilution and for detection of IgG2, 50 µl/well sheep anti-dog IgG2 conjugated to HRP at 1∶5000 dilution (Bethyl Laboratories, Montgomery, TX, USA). The plates were incubated at 37°C for 1 h and subsequently washed 6 times. One-hundred µl of TMB substrate were added to each well. The reaction was stopped after 15 min incubation at room temperature using H2SO4 (0.5 M) (50 µl). Absorbances were measured at 405 nm on an ELISA reader.
The Western blot technique was applied as described previously [25]. Briefly, stationary phase promastigotes of L infantum H-line or wild-type parasite (1×107 cells per lane) were washed with ice-cold PBS three times, and disrupted by sonication. An equal volume of sample buffer [0.1 M Tris (Merck), 12% sodium dodecyl sulfate (Merck), 10% glycerol (Merck), 5% β-mercaptoethanol (Merck), 0.1% bromophenol blue, pH 8.0] was mixed and the solution denatured at 95°C for 5 min. Promastigote lysates were fractioned individually on a 12% SDS-PAGE gel and subsequently transferred onto a nitrocellulose membrane (Sigma-Aldrich). The blots were individually incubated with 1∶50 diluted sera in PBS containing 3% skimmed milk at room temperature for 18 h. The blots were incubated with 1∶10000 diluted goat anti-dog IgG-heavy and light chains antibody horseradish peroxidase (HRP) conjugated (Bethyl Lab. Inc) in PBS containing 5% skimmed milk at room for 2 h. The blots were washed as above, incubated with ECL Plus chemiluminescent substrate (GE Healthcare), and exposed to X-ray film.
The possible presence of leishmanial DNA was assayed in peripheral blood and PLN necropsy. DNA was extracted (Promega, Columbus, OH, USA), according to the manufacturer's instructions and stored at −20°C until use. PCR amplification was carried out in 50 µl reaction volumes using 0.5 pmol of the kinetoplastid-specific primers K13A (5′-GTGGGGGAGGGGCGTTCT-3′) and K13B (5′-ATTTTACACCAACCCCCAGTT-3′) [26]. The amplification products were analysed by 1.5% agarose gel and visualized under UV light. A positive control containing genomic DNA of Leishmania-infected dog and negative control without template DNA were included.
VL cases were monitored from the records of each of the pediatric wards of Afzelipour Medical Centre at Kerman University of Medical Sciences. As VL cases might occasionally be referred to other hospitals, further information on VL cases were obtained from Centre for Disease Control and Prevention in Kerman Medical University.
Statistical analyses were performed with statistical package EpiTool (available at http://epitools.ausvet.com.au) to determine the number of animals in the vaccine trial. For quantitative data, Student's t-test was used to calculate differences the levels of IgG between two groups. Comparison between the levels of IgG1 and IgG2 antibodies was performed using Paired sample t-test. Chi-square test was used to examine the relationship between the number of dogs which IFA titers of IgG antibody were at >1∶100 between 2 groups. Fisher's exact test was used to calculate difference leishmanial DNA between 2 groups. Data are expressed as the mean ± standard deviation mean (SDM) for each group. Differences were considered significant when P<0.05.
Seventy seven household dogs, living in the endemic areas, were examined for clinical signs of the disease and tested for presence of specific anti-Leishmania IgG antibody. Specific anti-Leishmania antibody by IFA was found in 31 out of 77 (40.2%) of the householder dogs (≥1∶100). Sixty two out of 77 (80.5%) animals were asymptomatic.
The efficacy of the vaccine was evaluated after 4 sandfly transmission cycles by clinical examination, and serological and parasitological analyses. A vaccine trial was conducted on 103 dogs (55 vaccinated and 48 unvaccinated). No local indications including swelling, and pain at the injection site and no general indications of disease including anorexia, apathy, vomiting and diarrhoea were observed after the vaccine administration. Twenty three dogs (9 vaccinated and 14 unvaccinated) (22.3%) left the study after a change in residence or disappearance. Three unvaccinated dogs (2.9), died. Two of these dogs had to be put down because of accidental injury and one died from a disease unrelated to leishmaniosis.
All vaccinated dogs gave positive titers of specific anti-Leishmania antibodies whereas, all but 2 unvaccinated dogs were seronegative over the 3 month follow-up (Fig. 2). Fluctuations of the mean levels of antibody were observed in the sera of vaccinated dogs over the 20 month follow-up (Fig. 2) but did not rise. In contrast, the mean levels of antibody increased in the sera of unvaccinated dogs over the same period (Fig. 2). There was a significant difference between mean levels of antibody in the sera of vaccinated and unvaccinated dogs (P<0.001). The cut-off for which animals were considered seropositive was established to be a positive IFA results at serum dilutions of >1∶100. As shown in Table 1, twelve out of 31 (38.7%) unvaccinated dogs and 2 out of 46 (4.3%) vaccinated dogs were seropositive at >1∶100. The rest of dogs, 19 out of 31 (61.2%) unvaccinated dogs and 44 out of 46 (95.7%) vaccinated dogs were seropositive ≤1.100 over the 24 month follow-up. The number of unvaccinated dogs which were seropositive at >1∶100 was significantly higher than vaccinated dogs (P<0.0005). Specific anti-Leishmania IgG1 and IgG2 antibodies were present in the sera of dogs vaccinated with L. infantum H-line, predominantly of the IgG2 subclass. In the sera of vaccinated dogs, the level of IgG1 was significant lower than the level of IgG2 (P<0.001) whereas, the level of IgG1 was significantly higher than the level of IgG2 in the sera of unvaccinated dogs (P<0.05).
Two sera from vaccinated dogs which were seropositive at >1∶100, dogs V20 and V33, recognized the 21 kDa antigen of L. infantum H-line but not of L. infantum WT (Fig. 3). Ten out of 31 (32.2%) unvaccinated dogs which were seropositive at >1∶100 recognized the 21 kDa antigen of L. infantum WT, but not of L. infantum H-line (Table 1). As shown in Table 1, sera from 2 unvaccinated dogs, C2 and C61 were seropositive at >1∶100 but did not recognize the 21 kDa antigen of both L. infantum H-line and L. infantum WT. Sera from all dogs in both groups which were seropositive at ≤1∶100 did not recognize any antigens of L. infantum H-line or wild-type parasites.
The presence of leishmanial DNA and clinical signs of disease in vaccinated and unvaccinated dogs after 4 sandfly seasons are summarized in Table 1. No leishmanial DNA was found in the vaccinated dogs. In contrast, 9 out of 31 (29%) unvaccinated dogs were positive for the presence of leishmanial DNA over the 24 month period follow-up. As shown in Table 1, eight of 12 (66.7%) unvaccinated dogs with high levels of antibody (>1∶100) became PCR positive. The number of unvaccinated dogs that were PCR positive was significantly higher than that in the vaccinated dogs (P<0.002). All but 1 vaccinated dogs, [(2.2%), remained free of clinical abnormalities over the 24 months period of observation. Among the unvaccinated dogs, 9 out of 31 [(29%), two dogs, C8 and C1 in stages II and III of clinical signs of disease, respectively], presented one or more clinical signs of disease (Table 1).
No VL cases were referred to the Afzalipour Medical Centre or recorded in the Centre for Disease Control and Prevention from the area of study more than 3 years since June 2010.
This is the first study to demonstrate efficacy of an attenuated L. infantum vaccine against natural CVL in dogs in a highly endemic area. The progression of leishmaniosis in dogs is associated with humoral response and depression of cellular immunity [27]. We previously reported that L. infantum H-line induced a CD4+Th1 response which was characterized by the production of relatively higher levels of IFN-γ and lower levels of IL-10 compared with those in the dogs infected with wild-type parasite [18], [19]. In contrast to L. infantum WT, the attenuated parasite was unable to multiply and survive in the visceral organs of immunized dogs [17] and remained localized in the skin at the site where the promastigotes were injected. It has been reported that promastigotes of L. infantum WT develop to amastigote forms in infected macrophages at the site of inoculation and the infection may spread, resulting in a systemic form [1]. Dissemination of the parasite in the visceral organs of symptomatic dogs is the result of the development of a non-protective Th2 response [28]. Subcutaneous vaccination with the attenuated line in the foreleg of the dogs, an area covered with hair, will prevent or significantly reduce the likelihood of uptake of attenuated line parasites by sandflies, which tend to feed only on areas of exposed skin. This observation alleviates concerns about the possibility of reversion to virulence by the attenuated line during passage through sandflies.
We reported subclasses of IgG in vaccinated dogs and correlated higher IgG2 with protection provided by L. infantum H-line [19]. In the present study, we found the level of IgG1 was significant lower than the level of IgG2 in the sera of vaccinated dogs. It has been reported that specific anti Lieshmania IgG1 in dogs is associated with the development of disease, whereas IgG2 antibody is associated with asymptomatic infection [29].
The present study was carried out in 3 villages, in the district of Dehsard, Baft County in the southeast of Iran highly endemic for CVL [3], [30], [31]. In a preliminary study, we found that 40.2% of the household dogs were seropositive for L. infantum (>1∶100). However, the prevalence of CVL in domestic dogs in this area might be higher than 40.2%. It has been reported that 37% of seronegative asymptomatic dogs from an endemic area were positive by the PCR with skin tissue [32]. It is recognized that introducing 103 dogs into an area could disturb the ecological dynamics between dogs, parasite and vectors. Householders whose dogs were seronegative did not allow their dogs to be used in the study. Thus dogs from a non-endemic area were brought in for the trial and whenever possible each household included unvaccinated and vaccinate dogs in order to equalize their degree of exposure to the risk of natural infection. The seroposivity of unvaccinated dogs was higher than that of household dogs, living in the area. It has been reported that some dog breeds such as the German shepherd are more susceptible to development of CVL [33], [34].
In the unvaccinated group 12 out of 31 dogs (38.7%) dogs were highly seropositive (>1∶100) which 35.5% of them developed signs of CVL over 24 months of period of monitoring. The specific anti-Leishmania IgG antibody was raised in the sera of the dogs vaccinated with L. infantum H-line [17], [19]. We found that except 2 unvaccinated dogs, all sera from the vaccinated dogs and unvaccinated dogs which were seropositive at >1∶100 recognized the 21 kDa antigens of L. infantum H-line or WT. It is in agreement with another study that band of 21 kDa has the highest immuno-reactivity and the most often recognized in the case of CVL [35]. In the present study we found sera from vaccinated dogs recognized the 21 kDa antigen of L. infantum H-line whereas, sera from unvaccinated dogs, which were natural infected with L. infantum WT recognized the 21 kDa antigens of L. infantum WT (Table 1). This observation is in agreement with our previous study that Western blot analysis of antibodies to the 21 kDa antigens of L. infantum H-line and WT is very useful method for distinguishing between dogs vaccinated with L. infantum H-line and dogs experimentally infected with L. infantum WT [20]. We found that sera from 2 unvaccinated dogs, C2 and C61, which were seropositive at >1∶100 did not recognize any antigens of L. infantum H-line or L. infantum WT (Table 1). It has been reported that the sensitivity and specificity of Western blotting is greater than IFA for diagnosis of CVL in dogs [25]. IFA cross-reaction antibody between L. infantum and other diseases such as Ehrlichias canis (E. canis) and Babesia canis and also some kind of clinical signs of disease might be possible [36], [37], [38]. Moreover, E. canis infection might induce immunosuppression [39] and therefore the immune system is not able to develop the protective immunity induced by the attenuated line and the vaccinated animal is not immune to natural challenge. This observation may be useful to explain the sign of disease in 2 vaccinated dogs,V20 and V33, over 24 months period monitoring and suggests in the future, we need to check for the presence of E. canis during vaccination with the attenuated L. infantum.
A number of vaccines such as FLM, LiESAp-MDP have shown a degree of effectiveness against experimental CVL in dogs [40], [41]. Our study is the first vaccine trial in dogs that might show an impact of a vaccine and in reducing the occurrence of VL in the local human population. It has been reported that human seropositive (>1∶800) in this area was 1.55% and approximately half of 108 registered cases were from Baft [3]. No VL cases were recorded from the area of study 3 years since these data were collected. Impact of this vaccine on human population should be confirmed in the further studies. The results presented clearly demonstrated that a gentamicin-attenuated line L. infantum vaccine induced a significant and strong protective effect against CVL in dog and holds considerable promise for vaccination of dogs against CVL in the field.
|
10.1371/journal.ppat.1003071 | MiniCD4 Microbicide Prevents HIV Infection of Human Mucosal Explants and Vaginal Transmission of SHIV162P3 in Cynomolgus Macaques | In complement to an effective vaccine, development of potent anti-HIV microbicides remains an important priority. We have previously shown that the miniCD4 M48U1, a functional mimetic of sCD4 presented on a 27 amino-acid stable scaffold, inhibits a broad range of HIV-1 isolates at sub-nanomolar concentrations in cellular models. Here, we report that M48U1 inhibits efficiently HIV-1Ba-L in human mucosal explants of cervical and colorectal tissues. In vivo efficacy of M48U1 was evaluated in nonhuman primate (NHP) model of mucosal challenge with SHIV162P3 after assessing pharmacokinetics and pharmacodynamics of a miniCD4 gel formulation in sexually matured female cynomolgus macaques. Among 12 females, half were treated with hydroxyethylcellulose-based gel (control), the other half received the same gel containing 3 mg/g of M48U1, one hour before vaginal route challenge with 10 AID50 of SHIV162P3. All control animals were infected with a peak plasma viral load of 105–106 viral RNA (vRNA) copies per mL. In animals treated with miniCD4, 5 out of 6 were fully protected from acquisition of infection, as assessed by qRT-PCR for vRNA detection in plasma, qPCR for viral DNA detection in PBMC and lymph node cells. The only infected animal in this group had a delayed peak of viremia of one week. These results demonstrate that M48U1 miniCD4 acts in vivo as a potent entry inhibitor, which may be considered in microbicide developments.
| This report describes the protective effect of a CD4 peptide mimetic against HIV infection on human mucosal explants and further on, when used in a microbicide gel, against a SHIV challenge in cynomolgus macaques. Evidence is given that our “miniprotein engineering” strategy, which corresponds to miniaturizing a binding protein by transferring its binding site onto a small scaffold protein, followed by an extensive optimization of this miniprotein led to an active peptide with promising pre-clinical antiretroviral activity in the case of the gp120/CD4 interaction. Five out of six animals, pre-treated with a 0.3% miniCD4 gel were fully protected from SHIV162P3 vaginal challenge. On the whole, we demonstrated that such a small CD4 mimetic peptide could represent a powerful preventive agent against sexual HIV transmission, validating miniaturized protein interface design up to the discovery of potential new drugs.
| In complement to an effective vaccine against HIV transmission, development of potent anti-HIV microbicides remains important strategies to consider for HIV prevention [1]–[2]. The CAPRISA-004 clinical trial of 1% tenofovir gel provides the first encouraging results from a randomized phase IIb efficacy trial with a 39% reduced risk of HIV acquisition among HIV-uninfected woman [3]. In this study tenofovir gel was applied in a pericoital fashion both before and after intercourse. While encouraging, the observed levels of protection were less than optimal reaching only 50% in those reporting >80% adherence to the dosing strategy. More recently trial of a once-daily dosing regimen with tenofovir gel (VOICE) failed to demonstrate any detectable efficacy in at risk women. These studies underline the need to develop additional microbicide candidates with complementary or synergistic activity. Indeed, the combination of different antiretroviral candidates acting on different steps of the viral cycle and possessing non-overlapping resistance profiles might be a key to increasing the levels of protection observed in the CAPRISA-004 trial. However, only few microbicide candidates, in early clinical development, have been tested for efficacy against vaginal or rectal SIV/SHIV challenge in macaques, providing proof of concept for accelerated clinical development.
The use of effective therapeutic drugs, such as tenofovir and more recently dapivirine tested in phase III trial, targeting post-entry events in the viral replication cycle (reverse transcription, integration and viral maturation) remains attractive [3], however the impact of their dual use in prevention and therapy on the potential induction of resistance evolution remains controversial. In contrast drugs that target the initial steps of viral attachment to CD4 and one of two co-receptors (CCR5 or CXCR4) triggering viral fusion with susceptible targets cells (entry inhibitors) are not widely used in therapy. Some CCR5-inhibitors have demonstrated efficacy against vaginal challenge with simian-human immunodeficiency virus (SHIV) in macaque models [4]–[6]. While active against monotropic R5 virus exclusively utilizing the CCR5 co-receptor, these compounds would provide no protection against viruses either able to utilize both the CXCR4 and CCR5 co-receptors (dual tropic, R5X4 virus) or against monotropic X4 viruses exclusively utilizing CXCR4 for viral entry. However, compounds active against CD4 binding would likely be active against all three phenotypes of virus: R5, X4 and R5/X4. Few drugs target CD4 expression, one notable exception being CADA, yet to be evaluated in non human primate (NHP) studies [7]. Furthermore the consequence of interference with this key immunoregulatory protein on the immune system has not been fully characterized. In contrast the use of compounds directly targeting the CD4 binding site in the viral envelope protein (gp120) is less likely to have any immunomodulatory effects, providing a promising alternative strategy. Previous studies demonstrated that a small molecule inhibitor of gp120-CD4 interaction (BMS-378806) could protect NHP from vaginal challenge with SHIV162P3 [8]. However, the breadth of activity for this compound against the diversity of natural viral isolates was too narrow to warrant further development as a microbicide candidate.
Here we describe the evaluation of a miniCD4 mimetic, M48U1 presented on a stable 27 amino acid scaffold. Its close structural mimicry of CD4 endows it with broad activity across a wider range of pseudotyped virus expressing envelope of HIV-1 isolates [9]. We report the activity of M48U1 against HIV infection of human cervical and colorectal human tissue explants. Furthermore we present NHP studies to access both the pharmacokinetics and efficacy of M48U1 against vaginal challenge with SHIV162P3.
Small peptide mimetics of CD4, called miniCD4s, have been developed by our group for a number of years through a process of binding site transfer and further extensive optimization of a scorpion toxin scaffold [10]–[11]. Structural studies confirmed the close binding similarity between CD4 and these compounds [12]–[13]. Among the most recent generation of miniCD4s, M48U1, which no longer contains any residue common with macaque or human CD4, was identified as possessing sub-nanomolar affinity for gp120 with potent antiviral activity against HIV-1 subtype B and C pseudoviruses with ED50 values ranging from 0.07 to 2 nM and 0.4 to 37 nM, respectively [9].
These positive results prompted further assessment of the activity of M48U1 in genital and colorectal tissue explants models. Nonpolarized cervical and colorectal explant tissues were pre-treated with M48U1 for 1 h at 37°C followed by exposure to HIV-1Ba-L for 2 h in the continued presence of compound. M48U1 inhibited mucosal explant infection in the low nanomolar range with IC50 value of 42 nM (9–193 nM) and 4 nM (1–28 nM) for cervical tissue (Figure 1A) and in colorectal tissue (Figure 1B), respectively. Furthermore, M48U1 effectively inhibited the dissemination of HIV-1Ba-L to co-cultured T cells (PM-1) by migratory cells (dendritic cells and CD4 T cells) that emigrate out of cervical explants during the first 24 h of culture [14] (Figure 1C) with an IC50 of about 42 nM (11–157 nM). A small diminution in response was seen at the highest concentration of drug tested. Presumably this effect is due to the variability of the model or it could also reflect some loss in solubility of M48U1 at neutral pH. In contrast, the miniCD4 is very soluble at pH 4.5 (up to 30 mg/mL), which is the pH of the HEC- gel formulation (see Materials and Methods) used for in vivo challenge. M48U1 formulated in the HEC-gel was thus tested in colorectal tissue and the HIV-1Ba-L infection inhibition was as efficient as unformulated peptide (Figure 1D), displaying very tight error bars. Interestingly, M48U1 retained its anti-HIV activity in this gel even when kept for several months at 4°C. These data demonstrated that M48U1 miniCD4 was active against both direct mucosal tissue infection and dissemination of infection by migratory dendritic cells in ex vivo models. As some microbicides were shown to have poor biocompatibility and to induce inflammatory responses, we first evaluated the cytotoxicity of M48U1 using an enhanced colorimetric MTT assay. In co-cultures of monocyte-derived dendritic cells (MO-DC) and allogeneic CD4+ T cells, cultures of ME-180 endocervical cells [9], here, in penile tissue and in a T-cell line (PM-1), M48U1 showed no apparent cytotoxicity at the highest concentrations tested (Figure 2), suggesting good biocompatibility for the miniCD4. This was in contrast to the significant toxicity when the same tissue or cells were treated with nonoxynol-9 (Figure 2). Prior to macaque challenge studies, the activity of M48U1 was determined in TZM-bl cells against the simian-human chimeric virus (SHIV162P3) or pseudotyped SHIV162P3, with IC50 values of 25 nM and 9 nM, respectively (Table 1). These results were consistent with the nanomolar activity of M48U1 previously reported against various HIV-1 isolates [9].
Based on these encouraging data, macaque studies were initiated to assess the effectiveness of M48U1 against vaginal challenge with SHIV162P3. To determine optimal dose and appropriate time of viral challenge, formulation and pharmacokinetic studies were performed in uninfected animals. Previous reports indicated that inhibitor concentrations >1000 fold above the in vitro IC50/90 are required to protect macaques from in vivo mucosal challenge [4], [8]. Therefore, M48U1 was formulated at 0.3% (i.e. 3 mg/g of gel – corresponding to about 1 mM) in a gel comprising 1.5% hydroxyethylcellulose (HEC), 0.1% sorbic acid and 2.5% glycerol, at pH 4.6. One hour after M48U1-HEC gel application, peptide concentration in vaginal fluid was still about 75% of its concentration in the gel and this value dropped to 10% after 4 hours (Figure 3).
For SHIV challenge studies two groups of 6 naïve and sexually mature female cynomolgus macaques (Macaca fascicularis) were treated with medroxyprogesterone acetate (Depo-Provera, Pfizer, 30 mg) 40 days in advance. Animals received 2 g of HEC gel with (M48U1 group) or without (placebo group) M48U1 miniCD4. The gel was atraumatically instilled into the vaginal vault one hour before challenge. Intravaginal challenge was carried out with a 10-fold animal infectious dose 50% (AID50), corresponding to 2,500 tissue culture infectious dose 50% (TCID50) of SHIV162P3 (obtained from the NIH AIDS Research and Reference Reagent Program [15]), inoculated in 50% human seminal plasma [16].
As expected, after this stringent ‘high dose’ challenge of progesterone-treated females, all the control animals became infected as evidenced by detection of virus in plasma, followed by seroconversion. In contrast, five out of six animals, treated with 0.3% M48U1 were fully protected resulting in a significant difference compared to control animals (p = 0.0152; Fisher exact test). For these five animals (Figure 4), plasma viral loads remained below detection limit (60 copies of vRNA/mL) throughout the duration of the study, they did not seroconvert and the SHIV DNA copy numbers in lymph nodes also remained below the limit of detection (10 copies/million cells). The plasma of the single animal (# 19302) infected in the M48U1 gel group was fully evaluated for possible miniCD4-resistant viruses. Plasma was collected 21 days after challenge, and the extracted virus genes were sequenced and pseudotyped. All the breakthrough viruses tested were fully susceptible to M48U1 in TZM-bl assay. Some mutations were observed (notably the Q507R mutation) but none of them were shown to be responsible for any form of resistance. We don't find any reported resistance mutations [17] in such breakthrough viruses. In addition, protected animals did not differ from control animals in terms of MHC haplotypes associated with control of viremia [18]–[20] (Supplemental Figure S1). Mucosal treatments could potentially induce local inflammatory response which may interfere with infection and microbicide effect. Quantities of cytokines and chemokines in vaginal fluids of macaques 25 h after treatment with placebo- or M48U1-gel were not significantly different (Supplemental Figure S2), except lower G-CSF (60%) in M48U1- vs placebo-treated animals. However, it is possible that imperfect gel delivery in the infected animal or the 0.3% dosing of M48U1 were not sufficient to achieve 100% protection. Further dose/response studies are required to eliminate this last possibility.
In conclusion, we have demonstrated that the CD4 mimetic M48U1 miniprotein can efficiently protect macaques from SHIV challenge, indicating that this small peptide, acting as a fusion entry inhibitor, could represent a new preventive agent against sexual transmission of HIV-1 when formulated as a microbicide. Thanks to its stably folded scaffold, this peptide possesses high stability and resistance to temperature and protease degradation. Contrary to other entry inhibitors, miniCD4 targets the virus and not a human receptor. Moreover, it is not used in current HIV therapy. This might provide a significant advantage as it may prevent the spread of viruses that become resistant to current treatments. Our study provided evidence that the molecules, which interact with the CD4-binding site of HIV envelope, are highly potent at blocking HIV entry and represent valuable microbicide candidates. Finally, these results show that “our miniprotein engineering strategy”, which corresponds to miniaturizing a protein by transferring its binding site onto a small scaffold followed by optimization, could represent an effective method to develop new binding-inhibitory drugs with modified physicochemical and pharmacokinetic properties.
Adult cynomolgus macaques (Macaca fascicularis) were imported from Mauritius and housed in the facilities of the “Commissariat à l'Energie Atomique et aux Energies Alternatives” (CEA, Fontenay-aux-Roses, France). Non-human primates (NHP, which includes M. fascicularis) are used at the CEA in accordance with French national regulation and under national veterinary inspectors (CEA Permit Number A 92-032-02). The CEA is in compliance with Standards for Human Care and Use of Laboratory of the Office for Laboratory Animal Welfare (OLAW, USA) under OLAW Assurance number #A5826-01. All experimental procedures were also conducted accordingly to European guidelines for animal care (European directive 86/609, “Journal Officiel des Communautés Européennes”, L358, December 18, 1986). The use of NHP at CEA is also in accordance with recommendation with the newly published European Directive (2010/63, recommendation N°9). No suffering was specifically associated with vaginal treatment of macaques. The animals were used under the supervision of the veterinarians in charge of the animal facility. This study was part of the European Microbicides Project (EMPRO, EU contract number LSHCT-2004-503558) which NHP studies were accredited by ethical committee “Comité Régional d'Ethique pour l'Expérimentation Animale Ile-De-France Sud” under statement number 04-001 of April 7th, 2004.
The M48U1 peptide containing a p-(cyclohexylmethyloxy)phenylalanine residue at position 23 was synthesized at Pepscan Presto Inc. (Lelystad, The Netherlands) by solid phase peptide synthesis and purified after refolding by reverse-phase high performance liquid chromatography as described elsewhere [21].
SHIV162P3 virus was obtained through AIDS Research and Reference Reagent Program, NIAID, NIH from Drs Janet Harouse, Cecilia Cheng-Mayer, Ranajit Pal and the DAIDS, NIAID [15]. HIV-1Ba-L, provided by this Program was passaged through activated PBMC for 11 days.
Cervical tissue was obtained from patients undergoing planned therapeutic hysterectomy at St. George's Hospital in London, United Kingdom. Surgically resected specimens of colorectal tissue were collected at St. George's Hospital, London, United Kingdom. Penile tissue (glans) was obtained from individuals undergoing gender reassignment surgery at Charing Cross Hospital. All tissues were collected after signed informed consent was received from all patients and under protocols approved by the Local Research Ethics Committee (National Research Ethic Service NHS). On arrival in the laboratory, resected tissue was cut into 2–3 mm3 explants comprising both epithelial and stromal tissue or muscularis mucosae, depending on the tissue, as described previously [14], [22].
Cervical explants were cultured in RPMI 1640 medium supplemented with 2 mM L-glutamine, 10% fetal calf serum (FCS), and antibiotics (100 U/mL of penicillin, 100 µg/mL of streptomycin) Cervical explants were pre-incubated with ten-fold dilutions of M48U1 for 1 h. Tissue was then exposed to HIV-1Ba-L (at 104 TCID50/mL equivalent to an MOI of 1) for 2 h and then washed 4 times with PBS to remove compound and/or virus. Explants were then transferred to a fresh tissue culture plate [14]. Following overnight incubation, tissue explants were moved to a fresh tissue culture plate and migratory cells left in the original plate were washed twice with PBS and co-cultured with 4×104 PM-1 cells/well without compound in 96-well plates for trans-infection assays. Tissue explants and cellular co-cultures were cultured for 15 days in the absence of compound.
Colorectal explants were maintained with DMEM containing 10% FCS, 2 mM L-glutamine and antibiotics (100 U/mL of penicillin, 100 µg/mL of streptomycin, 80 µg/mL of gentamicin). Colorectal explants were incubated with ten-fold dilutions of M48U1 for 1 h before exposure to HIV-1Ba-L (at 103 TCID50/mL equivalent to an MOI of 1). Explants were then washed 4 times with PBS to remove unbound compound and/or virus. Tissue explants were then transferred onto gelfoam rafts (Welbeck Pharmaceuticals, UK) and cultured for 15 days as previously described [22] in the absence of compound.
For all tissue explant models and migratory/PM-1 cell co-cultures approximately 50% of the supernatants of explants and cellular co-cultures were harvested every 2 to 3 days following re-fed with fresh media in the presence or absence of compound. The extent of virus replication was determined by measuring the p24 antigen concentration in supernatants (HIV-1 p24 ELISA, AIDS Vaccine Program, National Cancer Institute, Frederick, MA) [22]. All culture and incubations were carried out at 37°C in an atmosphere containing 5% CO2.
Penile tissues and PM-1 T cells were incubated in complete RPMI medium supplemented with ten-fold dilutions of M48U1 for 24 h at 37°C. Nonoxynol-9 microbicide (tested at 500 ng/mL) was employed as cytotoxic reference. The compounds were removed and the tissue explants and cells were incubated for 2 h in medium containing 500 µg/mL of 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT). For tissue, after releasing the formazan in ethanol overnight, cytotoxicity was determined by measuring the optical density of the purple formazan product at 570–690 nm using Synergy HT multi-detection microplate reader. For PM-1 cells, the absorption of the formazan product was determined after lysis of the cells.
Primary HIV-1SF162 or SHIV162P3 were briefly amplified in PHA-IL2 activated human PBMC. Corresponding Env expressing constructs were prepared by DNA amplification of the complete Env gene from PBMC co-cultures and subsequent cloning into pSV7d or pcDNA4/TO expression vector (Invitrogen BV, Groningen, The Netherlands). Sequencing of the pseudovirus construct and phylogenetic analysis of the complete gp160 confirmed the relation between the pseudovirus and the associated original virus.
Pseudoviruses were generated in a 24-well plate by co-transfection of 2×105 human embryonic kidney (HEK) 293T cells (obtained from ATCC) with pNL4-3. LucR-E- (400 ng, NIH AIDS Research and Reference Reagent Program) and HIV-1SF162 or SHIV162P3 Env expressing plasmid (1 µg) using the Calcium phosphate method (ProFection Mammalian Transfection Systems, Promega Benelux BV, Leiden, The Netherlands). After 24 h, the culture medium (Dulbecco's Minimum Essential Medium - DMEM) was replaced with medium containing 1 mM sodium butyrate and incubated for another day. Two days post-transfection, pseudoviruses were harvested and passed through Millex 0.45 µm filters. 10% fetal bovine serum (FBS) was added to filtered pseudoviruses, then the suspensions were aliquoted in 1 mL tubes and stored at −80°C until use.
Adherent CD4 and CCR5 expressing TZM-bl cell line with a luciferase reporter gene under HIV LTR control (NIH AIDS Research and Reference Reagent Program) [23] was cultured in DMEM containing 1% L-glutamine, 10% heat-inactivated FBS, and 50 µg/mL gentamicin.
Primary viruses and pseudoviruses were preliminarily titrated in TZM-bl cells and used at a concentration that resulted in 105 relative light units (RLU). The inhibitory activity of M48U1 against either HIV-1SF162, SHIV162P3 or corresponding pseudoviruses was measured as follows: 50 µL of (pseudo)virus suspension and 50 µL of a serial dilution of M48U1 or 50 µL medium (negative control) were pre-incubated for 30 min. Next, 100 µL of TZM-bl cells (at 105/mL) supplemented with 30 µg/mL DEAE dextran were added to each well and the 96-well plates were incubated for 48 h. Subsequently, 120 µL of supernatants were removed and 75 µL of Steadylite HTS (Perkin Elmer, Life Sciences, Zaventem, Belgium) were added. Next the luciferase activity (proportional to the amount of infectious virus particles) was measured using a TriStar LB941 luminometer (Berthold Technologies GmbH & Co.KG., Bad Wildbad, Germany) and expressed as relative light units (RLU). Finally, the inhibitory activity was calculated in GraphPad Prism 5.03 using non-linear regression (GraphPad Software, San Diego, CA, USA). All cultures and incubations were carried out at 37°C in 5% CO2 atmosphere.
M48U1 was formulated at 3 mg/g in an aqueous vehicle containing sorbic acid (0.1%), glycerol (2.5%) and gelling polymer hydroxyethylcellulose (HEC, final concentration 1.5%). The pH of the gels was 4.6±0.1, within the range of the normal, premenopausal vaginal pH. The osmolality (292±3 mOsm/kg) fell within the range of physiological fluids, thereby avoiding safety issues with hyperosmolar microbicide gels. Placebo gels were prepared in the same way but did not contain any M48U1.
Two grams of the previously described gel containing M48U1 at 3 mg/g were injected, using a French catheter connected to ready-to-use syringe, into the vaginal vault of two naïve female cynomolgus macaques (Macaca fascicularis). Vaginal fluids were sampled using Weck-Cel surgical sponges before and 1 h, 4 h, 6 h, 24 h, 48 h and 72 h after gel application. Only one sampling was performed for each animal. Multiple samplings would increase the statistical analysis but would result in a decrease of the miniCD4 concentration in the vagina for the next samplings. The miniCD4 concentration of each sample was independently assessed 3 times and the resulting data was represented by mean and standard error of the mean (SEM). To determine these concentrations, MS spectra were registered with a 4800 MALDI-TOF/TOF mass spectrometer (Applied Biosystems, Foster City, USA) using an internal standard consisting of M48U3 (similar peptide with a slight difference of residue at position 23) [9].
Two groups of 6 naïve female cynomolgus macaques (Macaca fascicularis), treated with 30 mg medroxyprogesterone acetate (Depo-Provera, Pfizer) 40 days before the challenge, were atraumatically injected into the vaginal vault 1 h before challenge with 2 g of HEC gel with (M48U1 group) or without (placebo group) M48U1. Intravaginal challenge was carried out with 10 AID50 of SHIV162P3, corresponding to 2,500 TCID50, inoculated in 50% human seminal plasma [16].
Blood was collected in EDTA tubes at different time points, from 20 days before to 78 days after infection. Plasma viremia for each sample was evaluated using quantitative RT-PCR for measurement of viral RNA copy numbers. Detection limit of this method is 60 RNA copies/mL and quantification limit is 300 RNA copies/mL [24]. The SHIV DNA copy numbers in lymph nodes were also measured in animals with undetectable plasma viremia by quantitative PCR, using primers amplifying the gag regions of SHIV. Detection limit is 10 copies per million of cells [25].
Vaginal fluids were collected 24 h after viral challenge, i.e. 25 h after intravaginal treatment. Cytokines and chemokines (G-CSF, GM-CSF, IFN-γ, IL-1β, IL-1Ra, IL-2, IL-5, IL-6, IL-8, IL-10, IL-12/23(p40), IL-17, IL-18, MCP-1, MIP-1α, MIP-1β, TNF-α) were quantified using non-human primate Milliplex kit (Millipore) and a Bioplex2000 instrument (Biorad).
|
10.1371/journal.ppat.1002155 | Cross-Species Transmission of a Novel Adenovirus Associated with a Fulminant Pneumonia Outbreak in a New World Monkey Colony | Adenoviruses are DNA viruses that naturally infect many vertebrates, including humans and monkeys, and cause a wide range of clinical illnesses in humans. Infection from individual strains has conventionally been thought to be species-specific. Here we applied the Virochip, a pan-viral microarray, to identify a novel adenovirus (TMAdV, titi monkey adenovirus) as the cause of a deadly outbreak in a closed colony of New World monkeys (titi monkeys; Callicebus cupreus) at the California National Primate Research Center (CNPRC). Among 65 titi monkeys housed in a building, 23 (34%) developed upper respiratory symptoms that progressed to fulminant pneumonia and hepatitis, and 19 of 23 monkeys, or 83% of those infected, died or were humanely euthanized. Whole-genome sequencing of TMAdV revealed that this adenovirus is a new species and highly divergent, sharing <57% pairwise nucleotide identity with other adenoviruses. Cultivation of TMAdV was successful in a human A549 lung adenocarcinoma cell line, but not in primary or established monkey kidney cells. At the onset of the outbreak, the researcher in closest contact with the monkeys developed an acute respiratory illness, with symptoms persisting for 4 weeks, and had a convalescent serum sample seropositive for TMAdV. A clinically ill family member, despite having no contact with the CNPRC, also tested positive, and screening of a set of 81 random adult blood donors from the Western United States detected TMAdV-specific neutralizing antibodies in 2 individuals (2/81, or 2.5%). These findings raise the possibility of zoonotic infection by TMAdV and human-to-human transmission of the virus in the population. Given the unusually high case fatality rate from the outbreak (83%), it is unlikely that titi monkeys are the native host species for TMAdV, and the natural reservoir of the virus is still unknown. The discovery of TMAdV, a novel adenovirus with the capacity to infect both monkeys and humans, suggests that adenoviruses should be monitored closely as potential causes of cross-species outbreaks.
| Infection from adenoviruses, viruses that cause a variety of illnesses in humans, monkeys, and other animals, has conventionally been thought to be species-specific. We used the Virochip, a microarray designed to detect all viruses, to identify a new species of adenovirus (TMAdV, or titi monkey adenovirus) that caused a deadly outbreak in a colony of New World titi monkeys at the California National Primate Research Center (CNPRC), and also infected a human researcher. One-third of the monkeys developed pneumonia and liver inflammation, and 19 of 23 monkeys died or were humanely euthanized. The unusually high death rate (83%) makes titi monkeys unlikely to be natural hosts for TMAdV, and the genomic sequence of TMAdV revealed that it is very different from any other known adenovirus. The researcher developed an acute respiratory illness at the onset of the outbreak, and was found to be infected by TMAdV by subsequent antibody testing. A clinically ill family member with no prior contact with the CNPRC also tested positive. Further investigation is needed to identify whether TMAdV originated from humans, monkeys, or another animal. The discovery of TMAdV suggests that adenoviruses should be monitored closely as potential causes of cross-species outbreaks.
| Adenoviruses, first isolated in the 1950s from explanted adenoid tissue, are double-stranded nonenveloped DNA viruses that naturally infect many vertebrates, including humans and nonhuman primates. The human adenoviruses in the Mastadenovirus genus, comprised of all mammalian adenoviruses, are classified into 7 species A-G, and at least 51 different serotypes (and 5 proposed types, HAdV-52 to HAdV-56) have been described to date [1], [2]. Adenoviruses are the cause of an estimated 5–10% of febrile illnesses in children worldwide [3]. Some serotypes, such as human adenovirus type 14 (HAdV-14), have been associated with severe and potentially fatal outbreaks of pneumonia in residential facilities and military bases [4]. Adenoviruses have also been associated with other clinical syndromes including conjunctivitis, hepatitis, and diarrhea [5]. In nonhuman primates, most epidemiologic studies of adenoviruses have focused on their identification in fecal samples from asymptomatic animals [6], [7], [8]. Overt respiratory disease associated with simian adenoviruses has also been observed [9]. Although adenoviruses are significant pathogens, genetically modified strains are being actively explored as potential vectors for vaccines and gene therapy [10].
Infection by adenoviruses has generally been thought to be species-specific. Human adenoviruses do not usually replicate in monkey cells in the absence of helper viruses [11], and do not productively infect rodents (and vice versa) [12]. Studies of sera from animal handlers and zoo workers exposed to chimpanzees in captivity fail to detect antibodies to chimpanzee adenoviruses [13], [14]. However, recent serological surveys have found antibodies to New World and Old World monkey adenoviruses in donor human sera from regions where the monkeys are endemic [14], [15]. In addition, phylogenetic analyses of adenoviruses from greater apes reveal that they fall precisely into “human” adenoviral species B, C, and E [7]. The high degree of sequence relatedness within members of each species suggests that at least some adenoviral strains may be capable of infecting both nonhuman primates and humans.
Beginning in May of 2009, a deadly outbreak of fulminant pneumonia and hepatitis occurred in a closed colony of New World titi monkeys of the Callicebus genus at the California National Primate Research Center (CNPRC). Routine microbiological testing for an infectious etiology was negative. We previously developed the Virochip (University of California, San Francisco) as a broad-spectrum surveillance assay for identifying viral causes of unknown acute and chronic illnesses [16], [17], [18], [19], [20], [21], [22]. The Virochip, a pan-viral microarray containing ∼19,000 probes derived from all viral species in GenBank (n∼2500) [21], [23], has been previously successful in detection of novel outbreak viruses such as the SARS coronavirus [22], [24] and the 2009 pandemic H1N1 influenza virus [23]. Here we apply the Virochip to identify a novel and highly divergent adenovirus as the cause of the titi monkey outbreak. In addition, we present clinical and serological evidence that this virus may have infected a researcher at the CNPRC and a family member, thus demonstrating for the first time the potential for cross-species infection by adenoviruses.
In early 2009, the CNPRC housed 65 titi monkeys in one quadrant of an animal building. The index case, a healthy adult titi monkey, presented on May 14, 2009 with cough, lethargy, and decreased appetite (Fig. 1A, T1). Despite aggressive treatment with intravenous fluids and antibiotics, the animal developed severe respiratory distress and was humanely euthanized 5 days later. A second case presented 4 weeks later near the entrance to the building (Fig. 1A, T54). In the interim period, 3 healthy titi monkeys had been relocated from a separate building (Fig. 1A, T2, T3, and T19), with 2 of the 3 monkeys placed into the cage formerly occupied by the index case, reflecting a total at-risk population of 68. Over the ensuing 2 months, 21 additional monkeys, including one of the relocated monkeys, presented with clinical signs similar to those shown by the index case (attack rate = 23/68, or 34%) (Figs. 1A and 1B). Clinical signs in affected animals included cough, lethargy, poor appetite, tachypnea, and abdominal breathing. These symptoms progressed to overt respiratory distress and death or humane euthanasia within an average of 8 days. Chest radiographs typically revealed diffuse interstitial pulmonary changes and bronchoalveolar consolidation indicative of pneumonia, with right middle lobe predominance (Fig. 1C). Animals displaying clinical signs were quarantined and aggressively treated by veterinarians with supplemental oxygen, anti-inflammatory medications, bronchodilators (nebulized albuterol), broad-spectrum antibiotics, and antivirals (oseltamivir and/or ribavirin). In total, 19 animals died or were euthanized due to the illness during the outbreak (case fatality rate = 19/23, or 83%). Only 4 monkeys survived, even though the majority of sick animals (17/23, or 74%) consisted of apparently healthy adults and juveniles. Interestingly, none of the 133 rhesus macaques (Macaca mulatta) housed in the same building became sick during the outbreak, and neither did any of the Old World monkeys from surrounding outdoor colonies of rhesus and cynomolgus macaques (Macaca fascicularis).
Gross necropsy findings were similar in all titi monkeys and were characterized primarily by diffuse, consolidated pneumonias, with occasional evidence of fibrinous pleuritis, pericardial/pleural edema, and hemorrhage (Fig. 1D-1). Some livers, spleens, and lymph nodes were found to be abnormally enlarged. Hepatic necrosis and hemorrhage, along with ascites, were occasionally appreciated. On histologic examination, the normal cellular architecture of the lung and trachea was destroyed, and prominent intranuclear inclusion bodies were observed in the liver, lung, and trachea (Figs. 1D-2 and 1D-3).
A routine microbiological workup for infectious causes of the outbreak, including bacterial, mycoplasma, and fungal cultures, was negative. Respiratory viral testing failed to detect evidence of respiratory syncytial virus, adenovirus, influenza virus A and B, human metapneumovirus, and parainfluenza virus types 1, 2, and 3.
Given the clinical presentation of a severe acute viral respiratory illness and the appearance of intranuclear inclusion bodies on histological examination, we strongly suspected that a virus that had eluded detection by conventional assays was the cause of the titi monkey outbreak. Nasal, lung, and liver swab samples collected during necropsy were analyzed using the Virochip [21], [23]. Microarrays were analyzed using ranked Z-scores to assess the highest-intensity viral probes [18]. From a lung swab sample from an affected monkey, 4 of the top 80 probes on the Virochip corresponded to adenoviruses. Other viruses or viral families with ≥4 probes among the top 80, including chimpanzee herpesvirus (Herpesviridae), bovine viral diarrhea virus (Flaviviridae), and endogenous retroviruses (Retroviridae), were regarded as less likely to cause fulminant pneumonia and hepatitis, so were not pursued any further. The 4 adenovirus probes mapped to 2 different gene regions corresponding to the DNA polymerase and penton base (Fig. 2A). Interestingly, the 4 viral probes were derived from 2 different Adenoviridae genera (SAdV-23, simian adenovirus 23, PAdV-A, porcine adenovirus A, and HAdV-5, human adenovirus 5, in the Mastadenovirus genus; FAdV-D, fowl adenovirus D, in the Aviadenovirus genus), suggesting the presence of a divergent adenovirus that was not a member of any previously known species.
To confirm the Virochip finding of an adenovirus, we used consensus primers to amplify a 301 bp fragment from the hexon gene by PCR [25]. The fragment shared ∼86% nucleotide identity with its closest phylogenetic relatives in GenBank, SAdV-18, an Old World vervet monkey adenovirus, and the human species D adenoviruses. The newly identified adenovirus was designated TMAdV, or titi monkey adenovirus. Specific PCR for TMAdV was then used to screen body fluids and tissues from affected monkeys (Table 1). PCR results were positive from post-necropsy liver and lung tissues as well as from sera, conjunctival swabs, oral swabs, and nasal swabs collected at time of quarantine in 8 different affected monkeys, but were negative from a throat swab from an asymptomatic animal whose other 5 cage mates had become sick. In addition, nasal swabs were negative in 3 asymptomatic, minimal-risk titi monkeys housed in a separate building. To confirm the presence of virus in diseased tissues, we examined lung tissue from affected monkeys by transmission electron microscopy, revealing abundant icosahedral particles characteristic of adenovirus filling the alveoli (Fig. 1D-4).
Next, to assess persistent subclinical infection from TMAdV, we analyzed serum samples from at-risk asymptomatic or affected surviving monkeys 2 months after the outbreak (n = 41). All post-outbreak serum samples were negative for TMAdV by PCR (Table 1). To assess potential TMAdV shedding, stool samples collected from all cages housing titi monkeys 2 months post-outbreak were analyzed by PCR (n = 27), and were found to be negative. In addition, we checked for TMAdV in rectal swab samples from rhesus macaques housed in the same building as the titi monkeys (n = 26) and in pooled droppings from wild rodents (n = 2) living near the titi monkey cages. All macaque and rodent fecal samples were negative for TMAdV by PCR.
We also sought to determine whether PCR assays commonly used to detect human adenoviruses in clinical or public health settings could detect TMAdV. Adenovirus PCR was performed on a TMAdV-positive clinical sample, a TMAdV culture, and a human adenovirus B culture (as a positive control) using an additional 5 pairs of primers, according to previously published protocols [26], [27], [28] Three of the 5 primer pairs, designed to detect human respiratory adenoviruses B, C, and E, failed to amplify TMAdV [27]. The remaining 2 pairs of primers, both derived from highly conserved sequences in the hexon gene [26], [28], were able to detect TMAdV in culture as well as directly from clinical material.
To facilitate whole-genome sequencing of TMAdV, deep sequencing of a lung swab from one affected titi monkey and lung tissue from another affected monkey was performed. Out of ∼11.9 million high-quality reads, 2,782 reads and 3,767 reads aligned to the SAdV-18 genome by BLASTN (Fig. 2B, blue) and TBLASTX (Fig. 2B, transparent blue), respectively, with reads mapping to sites across the genome. De novo assembly of the complete TMAdV genome from reads that aligned to SAdV-18 was not possible due to insufficient sequence coverage (<46%). The poor apparent coverage was the result of high sequence divergence of the TMAdV genome from SAdV-18, which hindered the identification of most of the 16,524 actual deep sequencing reads derived from TMAdV (Fig. 2B, red). Thus, after partial assembly of TMAdV using overlapping reads aligning to the SAdV-18 genome, remaining gaps were closed by specific PCR. The complete genome of TMAdV was found to be 36,842 base pairs in length, with a base composition of 20.8% A, 29.8% C, 29.8% G, and 19.6% T, and a GC content of 59.6%, comparable to that of adenoviral species Groups C, D, and E in the Mastadenovirus genus. The deduced genomic structure of TMAdV was also similar to that of other mastadenoviruses and consists of 34 open reading frames (Fig. 2C).
Whole-genome phylogenetic analysis placed TMAdV in an independent species group separate from the known human adenoviral species A–G (Fig. 3). Among all 95 fully-sequenced adenovirus genomes in GenBank, the closest simian adenoviral relatives to TMAdV were SAdV-3, SAdV-18, and SAdV-21, with pairwise nucleotide identities ranging from 54.0% to 56.3% (Fig. 4). The closest human adenoviral relatives were the species D adenoviruses, which share 54.3% to 55.1% identity to TMAdV, with human adenoviruses of other species slightly less similar (51.1%–54.6%). The placement of TMAdV into a separate group by phylogenetic analysis was also observed when looking individually at the hexon, polymerase, penton base, and fiber genes (Fig. S1). Scanning nucleotide pairwise identity plots revealed that, among the major adenovirus genes, the DNA polymerase and hexon are more conserved, whereas the E1A and fiber are more divergent (Fig. 4). The significant overall sequence divergence of TMAdV from known human and simian adenoviruses is highlighted by the finding that PAdV-A (porcine adenovirus A), a non- primate mammalian adenovirus, shared only a slightly less similar whole-genome pairwise identity to TMAdV of 47.0% (Fig. 4). In fact, in the DNA polymerase gene, TMAdV shared a pairwise identity with PAdV-A of 67.2%, comparable to its pairwise identities with the other human adenoviruses, 59%–71.7% (Figs. 4 and S1). Although TMAdV was found to be highly divergent from other adenoviruses, different isolates of TMAdV from 3 affected titi monkeys were remarkably conserved, sharing 100% identity across the full-length hexon gene (data not shown).
The high level of sequence divergence in TMAdV held true at the amino acid level as well, with amino acid identities relative to other mastadenoviruses ranging from 20.8% to 27.5% for the fiber, the most divergent protein, to 68.7%–78.2% for the hexon (Table 2). Although bearing low sequence similarity to other adenoviruses, the penton base of TMAdV contained an RGD motif that presumably binds αv integrins. By both nucleotide and amino acid comparisons, the closest phylogenetic relative to TMAdV in GenBank overall was SAdV-3 (Fig. 4; Table 2). Bootscanning analysis revealed no evidence for recombination of TMAdV with other adenoviruses at either the whole-genome or individual gene level (Fig. S2).
The main neutralization determinant for adenoviruses, the epsilon determinant (ε), is formed by loops 1 and 2 in the hexon protein [29]. The epsilon determinant of TMAdV was significantly divergent from that of other mastadenoviruses, with amino acid identities in loop 1 varying from 30.6% to 44.8% and in loop 2 varying from 54.4% to 67.0% (Table 2). This observation suggested that cross-neutralization of TMAdV with sera reactive against other human/simian adenoviruses is unlikely.
We next attempted to culture TMAdV in an A549 (human lung adenocarcinoma) cell line, a BSC-1 (African green monkey kidney epithelial) cell line, and PMK (primary rhesus monkey kidney) cells (Fig. 5). Direct inoculation of cell cultures with a lung swab sample from an affected titi monkey produced a weak initial cytopathic effect in macaque BSC-1 and human A549 cells at day 7. However, despite multiple serial passages, we were unable to propagate the infected cell culture supernatant in either BSC-1 or PMK cells. In contrast, propagation in human A549 cells resulted in viral adaptation by passage 6 and generation of a fully adapted strain of TMAdV by passage 10 that was able to productively infect all 3 cell lines. Thus, culturing and propagation of TMAdV were successful in a human A549 cell line, but not in established or primary monkey kidney cell lines.
In hindsight, only one individual at the CNPRC reported becoming ill during the titi monkey outbreak, the researcher in closest, daily contact with the animals. Symptoms began near the onset of the outbreak, although whether they began prior to or after identification of the index case is unclear. The researcher, with a past medical history of multiple sclerosis, initially developed symptoms of a viral upper respiratory infection (URI), including fever, chills, headache, and sore throat, followed by a dry cough and “burning sensation in the lungs” that was exacerbated by a deep breath or coughing. The researcher endorsed a history of recurrent upper respiratory infections, and did not regard the illness as related to the titi monkey outbreak. Although symptoms persisted for 4 weeks, at no time did the researcher seek medical care, and no antibiotics were taken during the illness.
We carried out contact tracing to identify family members and other individuals in close contact with the researcher. Interestingly, two family members in the household also developed flu-like symptoms about 1–2 weeks after the researcher initially became sick. Their symptoms – fever, cough and muscle aches – appeared milder than those of the researcher and completely resolved within 2 weeks. Neither individual sought medical care for these symptoms, and notably, neither had ever visited the CNPRC.
To explore a potential link between the outbreak and associated illness in humans, we blindly tested available sera from titi monkeys (n = 59), rhesus macaques housed in the same building (n = 36), CNPRC personnel and close contacts (n = 20), and random human blood donors (n = 81) for evidence of recent or prior infection by TMAdV by virus neutralization (Fig. 6). Nineteen serum samples from 15 at-risk affected (symptomatic) titi monkeys were tested. Among 3 affected titi monkeys surviving the outbreak, 2 monkeys mounted a vigorous neutralizing Ab response to TMAdV, with negative pre-outbreak Ab titers (<1∶8) but antibody titers 2 months after the outbreak of >1∶512, while 1 monkey exhibited a positive but much weaker response. Affected titi monkeys who died during the outbreak exhibited a wide range of neutralizing Ab titers, from <1∶8 to >1∶512 (those without Ab likely died before mounting a response).
To investigate the possibility of subclinical infection by TMAdV, we also examined serum samples from asymptomatic titi monkeys (n = 40) and nearby rhesus macaques (n = 36), collected 2 months after the outbreak. Fourteen of 40 asymptomatic titi monkeys tested (35%) had antibody to TMAdV, indicating that the incidence of subclinical infection was significant (Fig. 1A; Fig 6). In fact, one of the 14 asymptomatic titi monkeys with positive Ab titers was located in the minimal-risk building. In contrast, only 1 of 36 rhesus macaque samples was positive, with an Ab titer of 1∶16. The 1 antibody-positive rhesus serum sample was negative by specific PCR for TMAdV (data not shown), as was stool from the cage in which the rhesus monkey was housed (Table 1).
Approximately 4 months after the outbreak, serum samples were collected from CNPRC personnel in direct contact with the titi monkeys. Serum samples were also collected from the two family members of the clinically ill CNPRC researcher 1 year after the outbreak. Only two samples were found positive for neutralizing Abs to TMAdV: (1) Ab titers for the clinically ill researcher were 1∶32, and (2) Ab titers for one of the family members of the clinically ill researcher were 1∶8.
Among 81 random blood donors from the Western United States, 2 individuals (2/81, 2.5%) had positive Ab titers of 1∶16 and 1∶8. Pooled rabbit sera containing antibodies to human adenovirus serotypes 1 through 35, representing species A–E, were unable to neutralize TMAdV (data not shown). Thus, the results of our serological survey appear unlikely to be due to nonspecific cross-reactivity from prior exposure to known human adenoviruses.
In this study, we employed a pan-viral microarray assay, the Virochip, to identify a novel adenovirus associated with a fulminant pneumonia outbreak in a colony of New World titi monkeys. Despite the absence of an animal model, which precludes a strict fulfillment of Koch's postulates, there are several lines of evidence implicating this novel adenovirus, TMAdV, as the cause of the outbreak. First, conventional testing for other pathogens, including other viruses by Virochip, was negative, and affected monkeys did not respond to empiric therapy with antibiotics or antivirals (ribavirin and oseltamivir in anecdotal use are not effective against adenoviral infections) [30]. Second, the clinical presentation of pneumonia and hepatitis is consistent with the known spectrum of disease associated with adenoviral infections. Third, TMAdV sequence was recovered by PCR in various body fluids and tissues from affected monkeys, including blood, respiratory secretions, and lung/liver tissue (Table 1). Fourth, the finding of intranuclear inclusions in diseased tissues, as well as direct visualization of adenoviral-like particles (TMAdV) in lung alveoli by electron microscopy (Figs. 1D-2 to 1D-4), support a primary role for TMAdV in the pathogenesis of tissue injury in affected monkeys. Finally, there was a significant neutralizing Ab response in surviving animals, with 2 monkeys having titers undetectable prior to the outbreak but rising to >1∶512 at convalescence (Fig. 6).
Although TMAdV retains the core genomic features common to all adenoviruses (Fig. 2C), phylogenetic analysis clearly places TMAdV within a separate branch, with no closely related neighbors (Figs. 3 and S1). A phylogenetic distance of >10% combined with the lack of cross-neutralization defines TMAdV as a new species [31]. Since emerging adenovirus strains such as HAdV-14 and HAdV-D22/H8 (otherwise known as HAdV-D53) are known to arise from recombination events among related ancestral strains [32], [33], we performed bootscanning analysis to look for such events in TMAdV. The bootscanning analysis, however, failed to show evidence of recombination, likely because closely related and/or ancestral strains to TMAdV have not yet been identified.
Entry of adenoviruses into cells involves an initial attachment of the fiber knob to the cell receptor, followed by internalization via a secondary interaction of the penton base with αv integrins [34], [35]. The presence of an RGD motif in the TMAdV penton base implies that the virus uses αv integrins for internalization [35]. However, the high sequence divergence in the fiber protein (Table 2), as well as the absence of fiber motifs conserved among adenoviruses that bind CAR [36], [37] (coxsackievirus-adenovirus receptor) or CD46 [38], [39], [40] (data not shown), suggest that neither of these two human adenoviral receptors may be the attachment receptor for TMAdV. Further studies will be necessary to identify the preferred cellular attachment and internalization receptors for TMAdV.
Despite its isolation from affected titi monkeys, we were unable to propagate TMAdV in both established (BSC-1) and primary (PMK) monkey kidney cells (Fig. 4). The virus, however, grew efficiently in a human A549 lung adenocarcinoma cell line. One explanation for this finding is that TMAdV may be unable to productively infect cells derived from Old World monkeys (e.g. rhesus and African green monkeys). An alternative possibility is that successful propagation of TMAdV may depend on infection of a specific host cell type, such as A549 lung, and not BSC-1 or PMK kidney cells. Nevertheless, after 10 passages in human A549 cells, the fully adapted strain of TMAdV exhibits an extended host range with the ability to productively infect both monkey and human cells. This observation implies that TMAdV possesses an inherent capacity to cross the species barrier and infect both humans and nonhuman primates. Efforts to identify host range and cell tropism of TMAdV, as well as the specific sequence changes responsible for adaptation to growth in cell culture, are currently underway.
The virulence of TMAdV in healthy and apparently immunocompetent titi monkeys (83% case fatality rate) is highly unusual for infections by adenovirus. In humans, deaths due to adenovirus infections or outbreaks are generally low (up to 18% for pneumonia associated with HAdV-14 [4]). Furthermore, severe infections from human adenoviruses are more commonly associated with older age, immunosuppression, and chronic underlying conditions such as kidney failure [4], [41]. Young, healthy individuals are in general much less likely to succumb to adenoviral-related illness. The severity of TMAdV-related illness in affected titi monkeys suggests that this species of monkey may not be the natural host for the virus. The failure to detect fecal shedding of TMAdV in convalescent or asymptomatic animals also suggests that the virus does not normally infect titi monkeys (Table 1).
Although the exact origin of TMAdV remains unclear, we can speculate on several possibilities. One possibility is that a cross-species “jump” from captive macaques to a susceptible colony of titi monkeys precipitated the outbreak. As there have been no new introductions of monkeys into the closed colony for the past 2 years, this conjecture relies on asymptomatic infection and transmission of TMAdV in the captive rhesus/cynomolgus macaque population at the CNPRC. CNPRC personnel who visited macaque areas would occasionally enter titi rooms with no change in personal protective equipment, thus providing a potential route of transmission for the virus. In addition, specific antibodies were detected in 1 of 36 (2.8%) asymptomatic rhesus macaques housed in the same building (Fig. 6), indicating that TMAdV has the capacity to infect this species of Old World monkey. Notably, the closest identified phylogenetic relative to TMAdV among the complete genomic sequences available in GenBank is a rhesus monkey adenovirus, SAdV-3 (Fig. 4; Table 2). Furthermore, serological evidence for cross-species adenoviral transmission events between different nonhuman primate species has been previously reported in the literature [42].
Although we failed to detect TMAdV in rodent droppings found near titi monkey cages (Table 2), it is still possible that the virus arose from an unknown animal reservoir. In this regard, the high sequence divergence of TMAdV relative to the known human/simian adenoviruses (Fig. 3), and comparable sequence similarity in the polymerase gene to a porcine adenovirus (Figs. 3 and S1) are striking. The four-week interval between the index case and the second case appears overly long given a typical incubation period for adenovirus infections of no more than 1 week [43]. This may be explained by our finding of a high rate of subclinical infection by TMAdV in asymptomatic titi monkeys (35%), but may also be due to separate introductions of TMAdV into the colony from an as-yet unidentified reservoir.
Our study data also support the potential for cross-species transmission of TMAdV between monkeys and humans. The researcher's fever, cough, and pleuritic symptoms (“burning sensation in the lungs”) are consistent with the development of a prolonged viral respiratory illness. Interestingly, pleurisy has been specifically reported in association with certain human adenovirus infections [44]. The clinical presentation, time of illness concurrent with the onset of the outbreak, and presence of neutralizing Abs in convalescent serum all strongly point to primary infection of the researcher by TMAdV. The detection of weakly neutralizing Abs (1∶8) in a serum sample from a sick family member of the researcher also suggests that TMAdV may be capable of human-to-human transmission. The decreased levels of neutralizing Abs to TMAdV in the researcher (1∶32) and a family member (1∶8) relative to those in infected titi monkeys (up to >1∶512) are consistent with a recent study showing much higher levels of neutralizing antibodies in chimpanzees than in humans with adenovirus infections, possibly due to more robust adenovirus-specific T-cell responses in humans than in monkeys [45].
Several lines of evidence support the contention that the direction of TMAdV transmission was zoonotic (monkeys to humans) rather than anthroponotic (humans to monkeys). First, the closest known relative to TMAdV in GenBank is SAdV-3, an Old World monkey adenovirus (Fig. 3; Table 2). Second, our results show that PCR assays for human adenoviruses in common use are capable of detecting TMAdV. Although sequencing of PCR amplicons for human adenoviruses is not performed routinely in diagnostic virology, TMAdV would presumably have been detected previously in large-scale studies of hexon sequencing of Ad field isolates if it were circulating in the community [46], [47]. Finally, the available sequence data in GenBank is heavily biased towards human adenoviruses, and much less is known about the potential diversity of the simian adenoviruses. We also cannot formally exclude the possibility that the outbreak arose from anthroponotic transmission. In our study, 2 of 81, or 2.5% of random adult blood donors exhibited borderline titers of neutralizing antibody to TMAdV, indicating either a low prevalence of TMAdV in the human population or cross-reactivity to a related virus (although no evidence of cross-reactivity was found with HAdV serotypes 1 through 35). Future large-scale studies of TMAdV seroepidemiology will be needed to better understand transmission of TMAdV between monkeys and humans. Nevertheless, our discovery of TMAdV, a novel adenovirus with the capacity to cross species barriers, highlights the need to monitor adenoviruses closely for outbreak or even pandemic potential.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The use and care of all animals followed policies and guidelines established by the University of California, Davis Institutional Animal Care and Use Committee (IACUC) and CNPRC (Animal Welfare Assurance #A3433-01). The protocol for the maintenance and breeding of the titi monkey colony was approved by the University of California, Davis IACUC (Protocol #15730). No specific animal research protocol was drafted for this study, as only excess clinical samples were analyzed for diagnostic purposes. Animals in extreme respiratory distress were humanely euthanized by veterinarians. Extensive veterinary care was provided to all animals affected by the outbreak in order to minimize pain and distress.
Serum samples from staff at the CNPRC, close contacts, and random adult blood donors were collected under protocols approved by institutional review boards of the University of California, Davis (Protocol #200917650-1) and University of California, San Francisco (Protocol #H49187-35245-01). Specifically, written informed consent was obtained from staff at the CNPRC and close contacts for analysis of their samples. Any potentially identifying information has been provided with the explicit permission of the individuals involved.
Sera from random blood donors were obtained from the Blood Systems Research Institute (San Francisco, CA); sera were derived from affiliated donor banks in California (Blood Centers of the Pacific, San Francisco, CA), Nevada (United Blood Service, Reno, NV), and Wyoming (United Blood Services, Cheyenne, Wyoming) and de-identified prior to analysis.
The California National Primate Research Center (CNPRC), which houses over 5,000 nonhuman primates, is a part of the National Primate Research Centers Program and is accredited by the Association for the Assessment and Accreditation of Laboratory Animal Care, International (AAALAC). At the beginning of 2009, the CNPRC maintained a colony of 74 titi monkeys (Callicebus cupreus) and a colony of over 4,500 rhesus macaques (Macaca mulatta). No new animals have been introduced into either colony for over 2 years. All titi monkeys are maintained in small social groups, while rhesus macaques are maintained in small or large social groups. All animal facilities are maintained in compliance with United States Department of Agriculture specifications.
Eighty-eight percent of the titi monkey population (n = 65) were housed in 1 quadrant of an indoor animal building, and all titi monkeys demonstrating clinical signs originated from this building (i.e. “at-risk” room) (Fig. 1A). Rhesus macaques (n = 133) were housed in the other 3 quadrants of this same building, and surrounding the building were outdoor housing units with rhesus macaques and cynomolgus macaques (Macaca fasicularis). Three additional titi monkeys were moved into the at-risk room less than 2 weeks after presentation of the index case, reflecting a total at-risk population of 68 animals. The remaining 6 titi monkeys were housed in an indoor animal building greater than 500 yards from the at-risk population (i.e. “minimal-risk” room).
The outbreak lasted approximately 3 months from May to August of 2009. Affected titi monkeys died from 3–24 days after appearance of clinical signs, with an average time to death or humane euthanasia of 8 days. Clinical and epidemiological data, including daily census reports, were tracked and recorded by veterinary and management staff. All personnel entering the titi monkey rooms (both at-risk rooms and minimal-risk rooms) needed to pass within approximately 20 feet of macaque enclosures prior to entry. CNPRC personal protective equipment (PPE) policy requires a change of PPE between entrance/exit of animal rooms housing different species. Staff compliance of this policy may have been compromised. Measures have since been taken by CNPRC management to ensure compliance with existing policies.
Bacterial, mycoplasma, and fungal cultures were performed at the CNPRC. Clinical samples were also sent to an outside laboratory (Focus Diagnostics, Cypress, CA) for respiratory viral testing by centrifugation-enhanced shell vial culture followed by direct fluorescent antibody staining for 8 viruses (respiratory syncytial virus, adenovirus, influenza virus A and B, parainfluenza virus types 1, 2, and 3, and human metapneumovirus).
Gross and histopathological analyses of post-mortem tissues were performed by a board-certified veterinary pathologist specializing in nonhuman primate/laboratory animal medicine, a branch of Primate Services at the CNPRC. At necropsy, tissue samples from the trachea, lung, and liver were collected and fixed in 10% formalin. Tissues were routinely processed and embedded in paraffin. 3-µm sections were stained with hematoxylin and eosin (HE) and examined by light microscopy. For transmission electron microscopy, tissue fragments (2×2 mm) were excised from paraffin blocks of lung, deparaffinized, and processed as previously described [48].
Total nucleic acid was extracted from body fluid or swab samples using commercially available kits (Qiagen, Valencia, CA). 200 µL of sample were passed through a 0.22 µm filter (Millipore, Temecula, CA) to remove bacteria and cellular debris and then treated with Turbo DNase (Ambion, Culver City, CA) to degrade host genomic DNA prior to extraction. For tissue samples, lung or liver tissue was homogenized in a 15 mL Eppendorf tube using a disposable microtube pestle (Eppendorf, San Diego, CA) and scalpel, and RNA extraction was then performed using TRIzol LS (Invitrogen, Carlsbad, CA), followed by isopropanol precipitation and two washes in 70% ethanol. Extracted nucleic acid was amplified using a random PCR method to generate cDNA libraries for Virochip and deep sequencing analyses as previously described [18], [21].
The current 8×60 k Virochip microarrays used in this study contain 19,058 70mer probes representing all viral species in GenBank, and combine probes from all previous Virochip designs [17], [18], [21], [23]. Four probes derived from 2 different Adenoviridae genera (SAdV-23, PAdV-A, HAdV-5, and FAdV-D) yielded an adenovirus signature on the Virochip that was found to be TMAdV. With the exception of SAdV-23, these highly conserved probes are part of the core Virochip design and were derived from all available adenoviral sequences in GenBank as of 2002 [21]. One explanation why more high-intensity probes to simian adenoviruses were not seen by Virochip analysis is that the genomes of many simian Ads, including SAdV-3 and SAdV-18 (the two closest phylogenetic relatives to TMAdV in GenBank), were not sequenced until after 2004 [7], [49], and thus their genomes are not as well-represented on the Virochip microarray.
Virochip analysis was performed as previously described [21], [23]. Briefly, samples were labeled with Cy3 or Cy5 fluorescent dye, normalized to 10 pmol of incorporated dye, and hybridized overnight using the Agilent Gene Expression Hybridization kit (Agilent Technologies, Santa Clara, California). Slides were scanned at 3 µm resolution using an Agilent DNA Microarray Scanner. Virochip microarrays were analyzed with Z-score analysis [18], hierarchical cluster analysis [50], and E-Predict, an automated computational algorithm for viral species prediction from microarrays [51]. Only Z-score analysis, a method for assessing the statistical significance of individual Virochip probes, yielded a credible viral signature on the microarray.
We initially used consensus primers derived from a highly conserved region of the hexon gene to confirm the Virochip finding of an adenovirus by PCR [25]. After recovering the full genome sequence, we then designed a set of specific PCR primers for detection of TMAdV, TMAdV-F (5′-GTGACGTCATAGTTGTGGTC) and TMAdV-R (5′-CTTCGAAGGCAACTACGC). The TMAdV-specific quantitative real-time PCR was performed on a Stratagene MX3005P real-time PCR system using a 25 µL master mix consisting of 18 µL of water, 2.5 µL of 10X Taq buffer, 1 µL of MgCl2 (50 mM), 0.5 µL of deoxynucleoside triophosphates (dNTPs; 12.5 mM), 0.5 µL of each primer (10 µM), 0.5 µL SYBR green, 0.5 µL of Taq polymerase (Invitrogen, Carlsbad, CA), and 1 µL of extracted nucleic acid. Conditions for the PCR reaction were 40 cycles of 94°C for 30 s, 55°C for 30 s, and 72°C for 30 s. Amplicons were purified on a 2% agarose gel, cloned into plasmid vectors using TOPO TA (Invitrogen, Carlsbad, CA), and sent to an outside company (Elim Biopharmaceuticals, Hayward, CA) for Sanger sequencing in both directions using vector primers M13F and M13R.
To assess linearity and limits of sensitivity for the TMAdV PCR assay, 12 serial log dilutions were made of a standard plasmid constructed by cloning the 157-bp TMAdV amplicon into a TOPO plasmid vector. Purified plasmid clones at each serial dilution were quantified using a Nanodrop spectrophotometer and then spiked into negative serum, stool, or oral swab sample matrix, each matrix consisting of a pool of 10 sera, 10 stool samples, or 3 oral swabs, respectively. For each sample type, a standard curve for the TMAdV PCR assay was calculated from 3 PCR replicates at each dilution of nucleic acid extracted from the spiked matrix (data not shown). To determine limits of sensitivity for the assay, probit analysis of results from 6 PCR replicates of 7 serial log dilutions (from a starting concentration of ∼1.2×105 copies/mL) was performed using SPSS 16.0 (SPSS Inc., Chicago, IL). By probit analysis, the 95% limit of detection for TMAdV was 781, 377, or 35 viral genome equivalents/mL for serum, stool, or oral swab samples, respectively (data not shown).
To facilitate whole-genome sequencing of TMAdV, we prepared amplified cDNA/DNA libraries for deep sequencing from lung tissue and a lung swab sample from 2 different monkeys using previously published protocols [23], [52]. Briefly, randomly amplified libraries were cleaved with a Type IIs restriction endonuclease (GsuI) and truncated adapters were ligated on the resulting strand ends. Full-length adapters containing strict 6-nt barcodes were added via an additional 15 cycles of PCR. Amplified libraries were size-selected on a 2% agarose gel at approximately 350 bp average length and then sent to an outside company (Elim Biopharmaceuticals, Hayward, CA) for deep sequencing on an Illumina Genome Analyzer IIx (Illumina, San Diego, CA). Paired-end reads were sequenced for 73 cycles in each direction. Paired-end reads were subsequently filtered to eliminate low-complexity sequences with a Lempel-Ziv-Welch (LZW) compression ratio below 0.4 [53], split into individual reads, classified by barcode, and stripped of any remaining primer sequences using BLASTN alignments (word size = 11, E-value = 1×10−5). After low-complexity filtering and barcode/primer trimming, 11,950,557 sequence reads remained, with each read consisting of 67 nucleotides, for a total of ∼800 megabases of sequence. Reads were then aligned using BLASTN (word size = 11, E-value = 1×10−5) and TBLASTX (word size = 11, E-value = 1×10−5) to the genome sequence of SAdV-18 (Fig. 2B). Overlapping reads aligning to SAdV-18 were used to assemble portions of the TMAdV genome with Geneious software (version 3.6.5) [54], employing the SAdV-18 genome as a reference sequence and requiring a 20-bp minimum overlap and 95% overlap identity. Aligning reads were also used to design PCR primers to close remaining gaps in the TMAdV genome. Amplicons derived from specific TMAdV PCR primers were gel-purified, cloned, and sequenced as described above. The 5′ end corresponding to the inverted terminal repeat (ITR) of TMAdV was obtained by PCR using a forward degenerate consensus primer and a reverse TMAdV-specific primer. The 3′ end was recovered using a forward primer near the 3′ end of the genome and a reverse primer derived from 5′-ITR sequence.
To identify predicted coding regions in the TMAdV genome, we used the fully annotated genome sequence of SAdV-21 in GenBank as a reference. First, we aligned the two genomes and identified all ORFs that were present with Geneious [54]. We then selected the candidate ORF that best matched the corresponding ORF in the annotated reference genome. For adenovirus genes that are spliced (e.g. E1A), the identification of a GT-AG intron start-stop signal was used to pinpoint the correct ORF. To confirm the accuracy of the coding sequence, the sequence of each identified ORF was aligned to a database containing all adenoviral proteins in GenBank by BLASTX.
To generate whole-genome and individual gene nucleotide phylogeny trees, all 95 fully sequenced unique adenovirus genomes were first downloaded from GenBank. Multiple sequence alignments were then performed on a 48-core Linux system using ClustalW-MPI [55]. Trees were constructed after bootstrapping to 1000 replicates by the neighbor-joining method (based on Jukes-Cantor distances) in Geneious [54], [56]. Pairwise alignments were calculated using Shuffle-LAGAN (window size, 400 bp; step size 40 bp; translated anchoring), a glocal alignment algorithm that is able to calculate optimal alignments by using both local alignments and global maps of sequence rearrangements (e.g. duplications of the fiber gene in adenovirus genomes with 2 fibers) [57]. Sliding window analysis of the Shuffle-LAGAN pairwise alignments was performed using the online mVISTA platform [58]. More accurate alignments were obtained with Shuffle-LAGAN than with either ClustalW-MPI or Geneious (data not shown). Bootscanning analysis was performed according to the Kimura 2-parameter method using 1000 replicates with Simplot (version 3.5.1) [59]. Pairwise amino acid amino acid alignments between predicted TMAdV proteins and corresponding proteins in other adenoviruses (Table 2) were performed using Geneious [54].
A549 (human lung adenocarcinoma) and BSC-1 (African green monkey kidney epithelial) cell lines as well as PMK (primary rhesus monkey kidney) cells are routinely maintained at the Viral and Rickettsial Disease Laboratory (VRDL) branch of the California Department of Public Health. Media consisting of Hank's medium (for A549 cells) or Dulbecco's modified Eagle's medium (DMEM) (for BSC-1 cells) were supplemented with 1×nonessential amino acids (Invitrogen, Carlsbad, CA), 10% fetal bovine serum, 100 U of penicillin/mL and 100 µg of streptomycin/mL. PMK cells were maintained in tubes containing growth media and antibodies to SV-40 and SV-5 polyomaviruses (Viromed, Pasadena, CA). Clinical samples were clarified by centrifugation for 10 min×4000 g and passaged through a 0.2-µm filter. Cell culture passages were subjected to 3 freeze-thaw cycles and clarified as above. After achieving 80–90% confluency, cell culture media were changed to maintenance media with 2% FBS and were inoculated with 200 µL of clinical sample or 100 µL of passaged viral supernatant. Viral replication was monitored over 14 days by visual inspection under light microscopy for cytopathic effect (CPE). To confirm the generation of infectious virus, viral supernatants were quantitated by an end-point dilution assay.
A virus stock of TMAdV (passage 7) was produced on human A549 cells, aliquoted, and quantitated by end-point dilution. To perform the virus neutralization assay, 55 µL of viral supernatant at a concentration of 100 TCID50 and 55 µL of serum (starting at a 1∶8 dilution) were mixed and incubated for 1 hour at 37°C. As a control for each serum sample, 55 µL of culture media and 55 µL of serum were mixed and treated in an identical fashion. While mixtures were incubating, A549 cells grown in T-25 plates were trypsinized and 4,000 cells in 100 µL of media were added to each well of a 96-well plate. After incubation, 100 µL of mixture were inoculated into appropriate wells containing 4,000 cells per well and the entire plate was placed in a 37°C 5% CO2 incubator. Cells in the plate wells were observed for evidence of CPE every other day for 1 week. For wells that showed inhibition of viral CPE, the corresponding serum samples were diluted in six 2-fold steps and then retested. The reciprocal of the highest dilution that completely inhibited viral CPE was taken as the neutralizing antibody titer.
To assess cross-neutralization of TMAdV by known human adenoviruses, 7 pools of in-house reference sera at the VRDL (rabbit hyperimmune sera) and collectively containing antibodies to human adenovirus serotypes 1 through 35 were available for testing. For each pool, 55 µL of rabbit sera and 55 µL of viral supernatant at a concentration of 100 TCID50 were mixed, incubated for 1 hour at 37°C, and inoculated onto A549 cells in wells of a 96-well plate as described above. Cells in the plate wells were observed for evidence of CPE every other day for 1 week.
All Virochip microarrays used in this study have been submitted to the NCBI GEO database (study accession number GSE26898; microarray accession numbers GSM662370-GSM662391; microarray design accession number GPL11662). The annotated, whole-genome sequence of TMAdV has been submitted to GenBank (accession number HQ913600). Deep sequencing reads have been submitted to the NCBI Sequence Read Archive (accession number SRA031285).
|
10.1371/journal.pcbi.1000483 | Evolutionary Triplet Models of Structured RNA | The reconstruction and synthesis of ancestral RNAs is a feasible goal for paleogenetics. This will require new bioinformatics methods, including a robust statistical framework for reconstructing histories of substitutions, indels and structural changes. We describe a “transducer composition” algorithm for extending pairwise probabilistic models of RNA structural evolution to models of multiple sequences related by a phylogenetic tree. This algorithm draws on formal models of computational linguistics as well as the 1985 protosequence algorithm of David Sankoff. The output of the composition algorithm is a multiple-sequence stochastic context-free grammar. We describe dynamic programming algorithms, which are robust to null cycles and empty bifurcations, for parsing this grammar. Example applications include structural alignment of non-coding RNAs, propagation of structural information from an experimentally-characterized sequence to its homologs, and inference of the ancestral structure of a set of diverged RNAs. We implemented the above algorithms for a simple model of pairwise RNA structural evolution; in particular, the algorithms for maximum likelihood (ML) alignment of three known RNA structures and a known phylogeny and inference of the common ancestral structure. We compared this ML algorithm to a variety of related, but simpler, techniques, including ML alignment algorithms for simpler models that omitted various aspects of the full model and also a posterior-decoding alignment algorithm for one of the simpler models. In our tests, incorporation of basepair structure was the most important factor for accurate alignment inference; appropriate use of posterior-decoding was next; and fine details of the model were least important. Posterior-decoding heuristics can be substantially faster than exact phylogenetic inference, so this motivates the use of sum-over-pairs heuristics where possible (and approximate sum-over-pairs). For more exact probabilistic inference, we discuss the use of transducer composition for ML (or MCMC) inference on phylogenies, including possible ways to make the core operations tractable.
| A number of leading methods for bioinformatics analysis of structural RNAs use probabilistic grammars as models for pairs of homologous RNAs. We show that any such pairwise grammar can be extended to an entire phylogeny by treating the pairwise grammar as a machine (a “transducer”) that models a single ancestor-descendant relationship in the tree, transforming one RNA structure into another. In addition to phylogenetic enhancement of current applications, such as RNA genefinding, homology detection, alignment and secondary structure prediction, this should enable probabilistic phylogenetic reconstruction of RNA sequences that are ancestral to present-day genes. We describe statistical inference algorithms, software implementations, and a simulation-based comparison of three-taxon maximum likelihood alignment to several other methods for aligning three sibling RNAs. In the Discussion we consider how the three-taxon RNA alignment-reconstruction-folding algorithm, which is currently very computationally-expensive, might be made more efficient so that larger phylogenies could be considered.
| In 1968, Francis Crick hypothesized that the first ribosome consisted entirely of RNA, without any protein cofactors [1]. A domain structure for this primeval ribosome was recently proposed [2]. To synthesize such a reconstructed ribosome or reconstructions of other evolutionarily significant RNAs such as group II introns [3] or telomerase [4], it will be necessary to develop methods that can predict the sequences and structures of ancient RNAs based on the divergent sequences of their many descendants.
An inspection of RNA alignments, such as those in the RFAM database [5], suggests that an evolutionary model for RNA structure must eventually include multiple layers of detail: point substitutions, covariant substitutions of base-pairs [6],[7], indels [8], local changes in secondary structure such as helix slippage [9], and changes in domain structure [2]. Stochastic context-free grammars (SCFGs), which can efficiently detect the long-range correlations of RNA base-pairing structures, are natural probabilistic models of such phenomena and have been used for ncRNA homology detection [10]–[13], gene prediction [14],[15], folding [16],[17] and alignment [18]–[20].
By analogy with models of substitution processes, which are well-understood [21], we may take the problem of building phylogenetic models of RNA evolution and split it into two halves. The first half is the development of a pairwise model, describing the probability distribution of a descendant () conditional on its immediate ancestor (). In substitution processes, the pairwise model is a conditional substitution matrix. Often (but not always) the pairwise model, representing a finite evolutionary time , is derived from an instantaneous model of change over an infinitesimal time interval, i.e., a continuous-time Markov chain (parametrized by a rate matrix). Obtaining the transition probabilities of this chain (via exponentiation of the rate matrix) yields a pairwise model whose parameters are smoothly-varying functions of . A pairwise model represents an individual branch of a phylogenetic tree, with representing the length of that branch.
The second half of the phylogenetic modeling problem involves extending the model (and related inference algorithms) from a single branch to a complete phylogeny, i.e., from a pairwise model of two sequences to a multiple-sequence model of many sequences. In a typical situation, the sequences at the leaves of the tree are observed but those at internal nodes are not. Questions of interest then include:
For substitution models, there has been extensive work focused on answering each of these questions. Given a pairwise substitution model, questions A and B can be answered exactly by Felsenstein's pruning algorithm [22] and question C can be answered by the peeling algorithm (first presented for pedigree analysis by Elston and Stewart [23]). The estimation of evolutionary histories (question D) has been addressed by exact summarization [24] and sampling [25] approaches. Another representation of answers A–C is that the pruning and peeling algorithms (combined) are just the sum-product algorithm on a directed graphical model [26], yielding exact marginal distributions for unobserved variables. Graphical models also suggest general-purpose sampling approaches in addition to the exact sum-product algorithm.
The two halves of the reconstruction problem — developing a pairwise model and then extending it to multiple sequences — are largely independent. Felsenstein's pruning algorithm, for example, is essentially blind to the parametric form of the pairwise substitution model; it just assumes that a substitution matrix is provided for every branch. Subsequent models developed by other researchers can be plugged into the pruning algorithm without modification [27],[28].
We therefore addressed the problem of modeling the indel-evolution of multiple structured RNAs in a similarly-modular fashion by separating the creation of pairwise and multiple-sequence models. In previous work, we addressed the first (pairwise) part of the RNA reconstruction problem by describing a simple continuous-time model of RNA structural evolution [29]. This model corresponded to a Pair SCFG with a time-dependent parametrization which we used to simultaneously align and predict the structure of pairs of related RNAs. The focus of the present work is to solve the second (multiple-sequence) part of the RNA reconstruction problem by giving a general procedure for extending a pairwise model to multiple sequences related by a phylogenetic tree. This process yields a multiple-sequence SCFG, a natural model of the evolutionary relationships between multiple structured RNAs.
The main contributions of this paper are (1) an algorithm that transforms a phylogenetic ensemble of pair grammars, representing models on branches of a phylogenetic tree, into a coherent, multiple-sequence SCFG, (2) dynamic programming (DP) algorithms for performing inference under this multiple-sequence SCFG, and (3) freely-available software implementing algorithms (1) and (2) for the simplified case of a three-taxon star-topology tree. While the idea of composing conditionally-normalized models on trees is intuitive, the resulting models can be very complex, even for simple models of RNA evolution, making (1) necessary. Studies of related indel models have suggested that an implementation of dynamic programming (DP) algorithms on a three-taxon tree is sufficient to draw samples from the posterior distribution of ancestral sequences on more complex tree topologies, using Markov Chain Monte Carlo or MCMC [30]–[32], suggesting that (2) and (3) are, in principle, sufficient for analyzing trees relating many sequences.
We show that our algorithm produces a multiple-sequence grammar which is much more compact than suggested by naive approaches to model construction. We provide analyses of the asymptotic complexities of models constructed using our procedure and provide estimates of the time and memory required to reconstruct the structures of several RNA families for the case of a three-taxon phylogeny, which we have implemented in the program Indiegram. While by these estimates only the smallest sequences currently fit into affordable memory, thereby preventing us from applying our method to many problems of interest, a simulation study suggests that we can hope to accurately reconstruct ancestral structures over long evolutionary time, even in the presence of structural divergence.
In the Discussion, we speculate on algorithmic extensions that may reduce memory requirements, inspired by related work in reconstructing DNA and protein sequences.
We describe below a general method for constructing a multiple-sequence stochastic grammar for alignment, folding and ancestral reconstruction of RNA, given a phylogenetic tree and a description of the evolutionary process acting along each branch.
Our problem statement is this: Given a phylogenetic tree relating several structured RNAs and a description of the evolution of a structured RNA along a single branch of the tree (in the form of a Pair SCFG), (1) find the corresponding phylogenetic multiple-sequence grammar and (2) use that grammar to reconstruct, a posteriori, the evolutionary histories of the RNAs. We assume here that the phylogeny, including both the tree topology and branch lengths, is given.
This paper focuses on model construction and inference algorithms rather than the heuristics which will be necessary to make these algorithms fast enough for analysis of many biological datasets. As discussed below, the complexity of general inference algorithms is prohibitively high for many problems of interest. However, this complexity can be significantly reduced by incorporating outside knowledge. For example, if we know the consensus structure of several sequences or their individual structures, then we can constrain our algorithms accordingly. Similarly, we might consider only ancestral structures which are compatible with a given multiple sequence alignment, or a relatively small set of candidate alignments (as in the ORTHEUS program [33]). Such constraints are commonly used by programs for SCFG-based RNA sequence analysis such as QRNA [34], Stemloc [18] and CONSAN [19]. Alignment and structural constraints can be combined [18].
In the following sections we introduce more precise definitions for two-sequence models of RNA structure and outline our algorithms for (1) combining these two-sequence models on a phylogenetic tree and (2) using the composite phylogenetic grammars for inference.
We discuss the general problem of creating state-space models of the evolution of related sequences, beginning with models of substitution processes acting at independent sites (as studied in likelihood phylogenetics) and generalizing to models of indels, first in primary sequences and then in sequences with conserved secondary structure.
A stochastic model for the evolution of one sequence (the ancestor, ) into another (the descendant, ) over an interval of time () can be described by a joint distribution, . This joint distribution can be factored, , where is the marginal distribution over ancestral sequences and is the conditional distribution over descendant sequences given an ancestral sequence. In terms of phylogenetics, the conditional distribution describes the evolution along a branch of length .
It is possible to “multiply” two such models together. More precisely, one multiplies two conditional distributions and sums out the intermediate sequence. Thus, successive evolution along two branches is modeled by the distributionand we can sum sequence out of this, obtaining the distributionfor the composite branch .
This formalism underlies likelihood phylogenetics. Working under the independent-sites assumption, is the 'th element of the joint substitution matrix for a single site and is the corresponding element of the conditional matrix. The conditional matrix is in fact the matrix exponential , where is the substitution rate matrix [24]. Composition of two branches just amounts to a matrix multiplication.
A similar formalism can be used to describe the evolution of whole sequences with indels. Suppose that the joint distribution is the distribution modeled by a pair hidden Markov model (Pair HMM) [11], a probabilistic model of the evolution of two sequences under the approximation that only adjacent characters are directly correlated, and the marginal is the distribution of a single-sequence HMM, a probabilistic model of single sequences under the same approximation. The conditional distribution then corresponds to a conditional Pair HMM, a discrete-state machine which transforms one sequence (the input, ) into another (the output, ). Following computational linguists, we call this conditionally-normalized state machine a string transducer or simply a transducer [35]. Because of its conditional normalization, this state machine is distinct from a standard Pair HMM. A Pair HMM has two outputs and and emits symbols to both of those outputs, while a transducer absorbs symbols from the input and emits symbols to the output . Despite this distinction, Pair HMMs and transducers share very similar inference algorithms; for example, is computed using a direct analogue of the Forward algorithm [11].
We extend this formalism to the case of structured RNA as follows. Let and now represent structured RNA sequences or, more precisely, parse trees. A single-sequence SCFG models the marginal ; a jointly-normalized Pair SCFG [11] models the the joint distribution . The conditional distribution is modeled by a conditionally-normalized Pair SCFG. Following terminology from computational linguistics [36], we call this conditionally-normalized grammar a parse-tree transducer.
String transducers are special cases of parse-tree transducers, just as HMMs are special cases of SCFGs. Henceforth, we will drop the distinction between strings and parse trees. We will also refer interchangeably to “states” (in the state-machine representation) and “nonterminals” (in the grammar representation). Likewise, we will refer interchangeably to “state paths” (machines) and “parse trees” (grammars).
We can use the concepts of factoring probability distributions introduced in the two-sequence framework to model the common descent of many homologous sequences. Given a phylogenetic tree and a two-sequence model, we wish to obtain a multiple-sequence SCFG describing the common descent of the observed sequences.
A singlet transducer (which emits, but does not absorb, symbols) lies at the root of the phylogeny and serves as a generative model of the ancestral sequence. To represent the evolution of an ancestral sequence into many descendant sequences, we place a branch transducer on each branch of the phylogeny.
Throughout this paper we frequently refer to two and three-taxon (star) phylogenies. In all cases, the sequence is assumed to be the (unobserved) ancestral sequence and the sequences , , and the (observed) extant sequences.
In this section, we describe dynamic programming (DP) algorithms for inferring the alignment, structure and evolutionary history of multiple related RNAs, using the multiple-sequence SCFG we have derived.
The transducer composition algorithm described above constructs a phylogenetic SCFG for both ancestral and extant sequences. A parse tree for this SCFG represents a structural and evolutionary explanation of the extant sequences, including a complete ancestral reconstruction. Consequently, given a set of extant sequences, many of the questions of interest to us can be reduced to searches over, or summarizations of, the set of possible parse trees.
Well-known algorithms already exist for maxing or summing over SCFG parse tree likelihoods. The Cocke-Younger-Kasami (CYK) algorithm performs maximum-likelihood (ML) inference; the Inside algorithm can be used to sum over parse trees or sample them a posteriori; and the Inside-Outside algorithm yields posterior probabilities for individual parse tree nodes [11].
All of these algorithms are, however, complicated (at least in our models) by the existence of “null cycles” in the grammar. A null cycle is a parse tree fragment that is redundant and could be removed, such as a detour through states () that could be replaced by a direct transition (). Biologically, null cycles correspond to fragments of ancestral sequence that were universally deleted and therefore are unobserved in any of the extant sequences. These unobserved fragments can be unbounded in length (and so, therefore, can the parse tree). Within the CYK, Inside and Outside recursions, this causes cyclic dependencies which cannot be resolved.
Below we describe a method to eliminate null cycles from the ensemble model by transforming any SCFG to an equivalent acyclic SCFG. We then present multiple-sequence versions of the CYK, Inside and Outside algorithms.
While some sort of null-cycle elimination is often required in order to deal with cyclic dependencies, there are several ways to accomplish this other than the algorithm presented below. A simpler approach (that only works for the CYK algorithm) appears in the computational linguistics literature [37]. We have also developed a heuristic for CYK that simply ignores null cycles as well as an iterative approximation that loops several times over cyclically-dependent cells of the DP matrix until the estimate starts to converge. For conciseness, we have omitted descriptions of these methods, presenting only the exact elimination algorithm.
We implemented our model construction algorithm on the three-taxon star phylogeny. Given a singlet transducer modeling ancestral structures and a branch transducer modeling structural evolution, our Perl modules generate C++ code for the corresponding jointly-normalized three-sequence (Triplet) SCFG. Any model of structural evolution which can be represented as a Pair SCFG and factored into singlet and branch transducers is permitted as input to the packages, allowing for flexible, automated model design. The available software is described in Text S3.
We illustrated our method for building models of structured sequences using a model which was introduced in previous work, the TKF Structure Tree [29], a simplified probabilistic model of the evolution of RNA structure.
The TKF Structure Tree (TKFST) model is based on the Thorne-Kishino-Felsenstein (TKF) model of the stochastic evolution of primary sequences via indel events [44]. In the original TKF model, sequence evolves under a time-homogeneous linear birth-death-immigration process [45]. Single characters (“links”) are inserted with rate and deleted with rate . At equilibrium, sequences obey a geometric length distribution with parameter . Although this model has flaws (e.g., it lacks affine gap penalties, rate heterogeneity and context-dependent mutation rates), it illustrates many of the key ideas used by more sophisticated indel models, notably the possibility for systematic derivation of pairwise alignment automata from first principles via analysis of birth-death processes [44],[46].
The TKF Structure Tree model is an extension of the TKF model to RNA structure. In this model, loop and stem regions are mutually nested (Figure 8): the parameter determines the proportion of links within loop sequences that are nested stems, and every stem sequence has a nested loop at the end. Single bases are inserted and deleted in loops with rates and ; similarly, base-pairs are inserted and deleted in stems with rates and . Both loops and stems have geometric length distributions with parameters and . Insertions of a new stem into an existing loop sequence (or deletions of an existing stem) occur at the same rate as single-base insertions (or deletions) and can model large-scale structural changes (Figure 9).
We parametrized the singlet and branch transducers of the TKFST model using estimates reported by a phylo-grammar for RNA secondary structure prediction, PFOLD [16], and an implementation of pairwise alignment for the TKF Structure Tree model, Evoldoer [29]. The equilibrium distributions of unpaired and paired nucleotides of the singlet and branch transducers, as well as the substitution models of unpaired and paired nucleotides of the branch transducers, were derived from the substitution rate matrices of the PFOLD program. These rate matrices, which have proven useful for RNA structure prediction [16],[17],[47], were derived from the Bayreuth tRNA database [48] and the European large subunit rRNA database [49].
This continuous-time model corresponds to a Pair SCFG and as such fits neatly into our modeling framework once the probability distribution is appropriately factored into marginal and conditional distributions (generated by singlet and branch transducers). Tables 1 and 2 show the states and transitions of the singlet transducer (single-sequence SCFG) which generates ancestral sequence under the Structure Tree model. Tables 3 and 4 show the states and transitions of the branch transducer (conditionally-normalized Pair SCFG) which evolves a sequence and structure along a branch of the phylogenetic tree.
The equilibrium distribution and transition probabilities between states of the TKFST model can be expressed in terms of functions of the evolutionary time along a branch and the insertion and deletion rates and of the model. The length of ancestral sequences is geometric in (Table 2), defined as . The three functions , and which govern the transition probabilities in Table 4 are defined for loop sequences asand similarly for stem sequences [29].
The above-described TKFST SCFGs must be transformed slightly before they can be loaded into Indiegram. The grammars are presented in Indiegram format in Text S4.
A few other useful statistics for the TKFST model: the expected number of links in a loop sequence is and in a stem sequence . Since of the links in a loop sequence are nested stems, and since each stem has twice as many nucleotides as it has links (since each link is a base pair), the expected number of bases in a loop sequence isThe expected number of bases in a stem sequence isThe expected number of bases that are created/removed when a loop-sequence link is inserted/deleted isThe expected number of stems directly rooted in a given loop sequence is and the expected number of stems directly rooted in, or indirectly descended from, a given loop sequence is (note that this is also the expected total number of loop sequences indirectly descended from a given loop sequence). Therefore, in the equilibrium structure, the expected number of stems is ; of loops, ; of unpaired bases, ; and of base-pairs, . In a tree with total branch length , the expected number of single-base deletions is ; of base-pair deletions, ; and of substructure deletions, .
The TKFST model, like the original TKF model, probably needs refinements in order to accurately model many structural RNAs. For example, it fails to model certain phenomena observed in natural RNA structures (such as base-stacking or tetraloops) and in alignments of those structures (such as helix slippage). We assessed its appropriateness as a model of RNA structural evolution by conducting benchmarks of its capabilities for (1) multiple sequence alignment of structured RNAs, summing over all possible structures, and (2) structure prediction of homologous structured RNAs and comparing its performance to Stemloc (one of the better-performing pairwise SCFGs used for RNA multiple alignment [20]). The results of these benchmarks, reported in Table 5 and Table 6, suggest that TKFST is a useful guide for deriving more complicated models of RNA evolution: while it has relatively poor sensitivity (but high positive predictive value) as a base-pairing predictor, it is competitive with one of the most accurate RNA multiple sequence alignment programs [20].
TKFST's poorer performance at base-pairing prediction is likely due to its much-simpler model of RNA structure. The richer grammar, as described in [18], is much more complex than TKFST: excluding the substitution model, it has 14 free parameters (compared to TKFST's 4), uses an affine gap penalty (compared to TKFST's linear gap penalty), and explicitly models structural features such as multiple-branched loops, symmetric/asymmetric bulges, and minimum loop lengths. Unlike TKFST, the richer grammar is structurally unambiguous: a one-to-one mapping exists from structures to parse trees. Although we use the TKFST model as an illustrative example of a Pair SCFG that can be extended with our method, the model is not fundamental to our approach and can be replaced by a different and more realistic pairwise model, such as the Stemloc pairwise SCFG used in these comparisons [20]. We anticipate that further improvements should be possible by reviewing other comparisons of SCFGs at structure prediction, such as the study of [17].
We used our model-construction algorithm to build the grammar corresponding to the TKFST model acting on a star phylogeny with three (extant) leaf sequences and a single (unobserved) ancestral sequence. We chose this phylogeny for two reasons: (1) it is the simplest extension of the well-studied, standard two-sequence (Pair SCFG) model and (2) algorithms on a phylogeny with three leaves should be sufficient for ergodic sampling of reconstructions on any larger phylogeny, using, e.g., a Gibbs-sampling MCMC kernel [31] or a progressive suboptimal-alignment sampling heuristic [33].
The statistics of the TKFST model on the three-taxon phylogeny illustrate the advantages of our procedure for model construction. While the singlet and branch transducers are relatively simple—the singlet transducer, shown in Table 2, has 7 total states and 2 bifurcation states and the branch transducer, shown in Table 4, has 21 total states and 6 bifurcation states—the ensemble model of three extant sequences is very complex. The naive exponential upper-bound gives a maximal state space of size states. Using our uninformed search algorithm, we determined that there are 287 accessible states and 686 possible transitions between these states (compare with the transitions estimated with the exponential calculation). After performing the transformations described in “Exact elimination of null cycles in SCFGs” to eliminate useless windback states, the ensemble model has a reduced state space with 230 states, albeit at the cost of extra transitions, bring the total to 1,789 transitions (here we are trading reduced memory complexity, which is linear in the number of states, for increased time complexity, which is linear in the number of transitions). Note that both before and after the reduction in complexity, the total number of states and transitions are less than the approximate bounds of states and transitions suggested in “The composition algorithm”. Nonetheless, the extreme complexity of the ensemble model, despite the simplicity of the underlying model of RNA structure, makes clear the necessity for automated procedures for model construction. Dataset S1 gives the state space of the ensemble model constructed by the search algorithm and Dataset S2 the reduced model after eliminating windback states; both are in Graphviz format for visualization and show the state of the singlet transducer generating ancestral sequence as well as the states of the branch transducers generating observed sequences.
We implemented constrained maximum-likelihood inference of the structural alignment and ancestral structure of three extant sequences in a C++ program (Indiegram). For tractability, Indiegram uses the concept of fold envelopes described earlier to limit the fold space considered by the CYK algorithm, permitting structural information for the three extant sequences to be (optionally) supplied as input. If no structural information is supplied, then Indiegram uses a single-sequence SCFG to estimate a set of plausible folds [18], which are used to constrain the CYK algorithm.
The inference algorithms in Indiegram could be further constrained to enforce, for example, a fixed multiple alignment or a consensus structure for extant sequences. While experimentally-determined structures of individual RNAs are relatively rare, curated deep sequence alignments, such as those constructed for ribosomal RNAs [50], are frequently available for characterized RNA families. By constraining the inference algorithms with such sequence alignments, the memory and time complexity of the algorithms could be dramatically reduced. Such constraints can be naturally expressed with “alignment envelopes,” the alignment-space analogue of fold envelopes [18]. However, in this paper we focus on model construction and inference algorithms and postpone exploration of heuristics and constraints of these algorithms for future work.
While reconstructing large RNAs such as ribosomal subunits is currently computationally-inaccessible without further heuristics to constrain our algorithms, reconstructing small RNAs of biological interest will soon be feasible. Table 7 shows estimates of the memory and time required to reconstruct biologically-interesting subunits of the nanos 3′ translational control element and tRNAs, as well as two small RNAs which show significant structural divergence, the Y RNAs and Group II introns, and therefore promise to be interesting candidates for ancestral reconstruction. The reconstructed structures for three nanos 3′ translational control elements (TCEs) and three tRNAs, which could be analyzed given current computational limitations, can be found at http://biowiki.org/IndieGram; however, the phylogenetic trees relating the tested sequences have short branch lengths, making the reconstruction problem easy by forcing the reconstructed structures to be essentially-identical to those of one of the extant RNAs.
Guided by our experience with the nanos 3′ TCE and tRNA, where the reconstruction problem was made easy by the presence of a close outgroup, we conducted a simulation study of the dependence of reconstruction accuracy on outgroup branch length, with the further goal of comparing the performance of our reconstruction method (when simulating directly from the model) to simpler reconstruction methods that ignore either structure or phylogeny. (We here use the term “outgroup” loosely to denote the variable-length branch in our three-taxon study, where the other two branches are held at unit length.)
We simulated the evolution of RNAs under the TKFST model along three-taxon phylogenies (with one internal node), where we kept the branch lengths of two sibling species constant and varied the branch length of the outgroup between at steps of size . Parameters used in the simulation were and for loop sequence and and for stem sequence; the probability of a stem insertion was . These yielded a mean loop length of 5 bp and a mean stem length of 2.33 bp, with substructure indels per alignment. We selected alignments to reconstruct by requiring that there be at least two ancestral stems, loops of and stems of ; to reduce the complexity of our algorithms we additionally required that the sequences have .
We then attempted three-way multiple alignment (and, in some cases, reconstruction of the ancestor) using a variety of statistical inference algorithms. We sought insight as to the relative importance of the following factors in reconstructing ancestral RNA: (i) modeling the secondary structure; (ii) modeling the phylogenetic topology & branch lengths; (iii) using posterior-decoding algorithms to maximize the expected alignment accuracy, rather than picking the single most likely alignment [20],[51],[52].
The alignment programs we used in this benchmark were Indiegram (exact ML inference of alignment and ancestral structure, given phylogeny, descendant structures and correct model); Stemloc (a greedy ML heuristic, ignoring phylogeny in favor of a single-linkage clustering of the descendant structures); Stemloc-AMA (a posterior-decoding heuristic, maximizing the alignment's expected accuracy rather than its likelihood); and Handel (ML alignment under various indel models that ignore secondary structure completely). In detail, the reconstruction methods were
Stemloc : the Stemloc program was used to align the three sequences via single-linkage clustering with a Pair SCFG [18]. The structures of the leaf sequences were provided, but not the phylogenetic branch lengths. Instead of modeling a true phylogeny by introducing unobserved ancestral sequences, it just does single-linkage clustering of the observed sequences.
Stemloc-AMA : the Stemloc program was used to align the three sequences in “sequence annealing” mode, a posterior decoding method that attempts to optimize AMA, a sum-over-pairs alignment accuracy metric [20]. The structures of the leaf sequences were provided, but not the phylogenetic branch lengths. This program uses the same underlying pair SCFG as Stemloc, but instead of maximizing likelihood, it attempts to maximize an alignment accuracy metric.
TKF91: with the TKF91 model [44], the Handel package [30],[39],[40] was used to align the three extant sequences and reconstruct the ancestor. The correct phylogenetic tree and branch lengths were supplied (as they were for the Indiegram benchmark). The insertion, deletion and substitution rates for the TKF91 model were set equal to those of the loop submodel of TKFST. This may be understood as a naive sequence-only reconstruction that completely ignores basepair structure (i.e. the stem sub-model of TKFST).
Long Indel: with a single-event trajectory approximation to the long indel model [53], the Handel package was used to align the three extant sequences and reconstruct the ancestor. The correct phylogenetic tree and branch lengths were supplied. The deletion and substitution rates were set equal to those of the loop submodel of TKFST. The mean indel length was set equal to , the mean number of bases that are created/removed by an insertion/deletion in the loop submodel of TKFST; the mean equilibrium sequence length (and thereby the insertion rate) was equal to , the mean number of bases in TKFST at equilibrium (“A simple model of RNA structural evolution” has formulae for these quantities in terms of the TKFST rate parameters). This model improves on the previous model (TKF91) by introducing affine gap penalties.
We measured alignment accuracy, under the simplifying assumption that this correlates well with ancestral reconstruction accuracy.
We first consider the perfect alignment rate; that is, the number of times each method gets the alignment exactly correct. Theory predicts that Maximum Likelihood inference, using the correct model and parameters, should be asymptotically optimal (if one only counts perfect guesses). Inspecting Figure 10, we find this to be almost the case; the exception is when the outgroup is very distant and the bins may be undersampled (the departure from prediction that we observe for low-identity alignments is not statistically significant: when the optimal success rate drops below , then 125 trials are probably insufficient to compare two near-optimal methods). We also note that the ML version of Stemloc is near-optimal, despite the Stemloc pair-SCFG being slightly different from the TKFST pair-SCFG in parameterization and structure (e.g. Stemloc 's grammar does not allow insertion/deletion of entire substructures). The ML version of Stemloc is also observed to have a slightly higher perfect alignment rate than the posterior-decoding version (Stemloc -AMA). Finally, we note that the structure-blind models (TKF91 and Long Indel) perform consistently worse than the structure-aware methods; furthermore, both linear (TKF91) and affine (Long Indel) gap-penalties perform equally bad in this test (note that the TKFST model, from which the true alignments were simulated, does not allow long-indel events, which may partly explain why affine gap-penalties do not help in this benchmark).
A subtly different ranking emerges from consideration of the alignment accuracy. In Figure 11, we abandon the all-or-nothing metric of counting only perfect alignments, instead using a metric that shows what proportion of the alignment is correct. Specifically, we plot the Alignment Metric Accuracy (AMA) as a function of outgroup branch length. AMA measures the proportion of residues which are correctly aligned, averaged over all pairs of sequences [54]. Figure 4 reveals that Stemloc-AMA (which attempts to find the alignment with the maximum expected AMA) edges out both Indiegram and the ML version of Stemloc (both of which attempt to find the alignment with the maximum likelihood). These results, compared to the subtly different story told by the perfect alignment rate, underscore the point that benchmark results for alignment methods can depend exquisitely on the choice of accuracy metric. The superiority of ML methods is only assured in terms of perfect alignment rate, and not necessarily other accuracy metrics.
Taken together, these results suggest that the most important factor distinguishing the various models we have examined is the incorporation of some form of basepair structure: structure-blind Handel (regardless of linear vs affine gap penalty) performs much worse than the structure-aware SCFG methods. Intuitively, this is to be expected: whenever a basepair-aware method aligns one half of a basepair, it gets the other nucleotide correctly aligned for free. In benchmarks of RNA multiple alignment programs, structure-aware scoring schemes routinely outperform structure-blind scoring schemes [55],[56]. Since we know that modeling structure is very important, it's not too surprising that it turns out to be the most important of the factors we considered.
The second most important, amongst the factors we have considered in this experiment, is selection of the most appropriate objective function for the task at hand (c.f. perfect alignment rate vs AMA), followed by use of the correct posterior-decoding algorithm for the chosen objective function (c.f. Stemloc vs Stemloc-AMA). This is a subtle but important point: before deciding exactly what inference algorithm we're going to use to reconstruct ancestral sequences, we need to decide whether we want to maximize (a) the probability that our reconstructed sequence is 100% correct, (b) the expected number of nucleotides that are correctly reconstructed, (c) the expected number of base-pairs that are correctly reconstructed, (d) the expected number of stems that are correctly reconstructed, (e) some other metric. Each of these metrics would require a slightly different inference algorithm.
Lastly, the fine details of the scoring scheme—including branch lengths, substitution scores, gap penalties and so forth—appear to be the least important of the factors we considered, yielding observable differences only when all other aspects of the inference procedure were more-or-less equal. While such details of the model may affect reconstruction quality, they appear to have very minor influence on alignment quality.
Following the conception of paleogenetics [57], a large number of synthetic reconstructions of ancestral protein sequences have been reported in the literature [58]–[65]. There is also scientific interest in reconstructing DNA sequences [33], [66]–[71]. Given the importance of the RNA world hypothesis to current discussions of the origin of life [72]–[78], the many modern-day relics of this world [79]–[82] and the recent proposal of a structural model for the primordial ribosome [2], we believe that phylogenetic reconstruction of ancient RNA is a significant problem, deserving of strong bioinformatics support.
The work reported in this paper builds on extensive prior art in the areas of evolutionary modeling and ancestral reconstruction. Reviewing all of this would take several books, but we can note some key references. The reconstruction of ancient sequences was first proposed in 1963 by Pauling and Zuckerkandl [57]; current applications of this idea, mostly using substitution models, are surveyed in the book edited by Liberles [83]. Many algorithms in phylogenetics implicitly reconstruct substitution histories, whether by parsimony [84],[85] or likelihood [22]. There is a substantial body of work to model indels on phylogenies [30], [35], [39], [44], [53], [86]–[92]. Recent work has extended these ideas to the reconstruction of indel histories [93],[94], particularly at the genomic scale [33],[95]. There is also prior work in computational linguistics on the theory of transducers for sequences [96] and parse trees [36],[37],[97],[98] (from which we take the terms “string transducer” and “parse-tree transducer”). We draw on the bioinformatics literature for SCFGs [10],[11],[99], especially Pair SCFGs [14],[18],[19] and phylogenetic SCFGs [16]. In particular, an early example of a pairwise conditional model for structure-dependent RNA evolution was given by Eddy et al [12]. A conditional framework similar to ours in some respects is described by Sakakibara et al [100]. The dynamic programming inference algorithms for multiple-sequence SCFGs are closely related to the protosequence algorithm of Sankoff [42].
While we have focused on the TKF Structure Tree model in our Results, our model-construction algorithm is applicable to any model of the evolution of secondary structure which can be expressed as a Pair SCFG. Realistic structural and thermodynamic effects—such as base-stacking or loop length distributions—can, in principle, be incorporated. Other phenomena of RNA evolution may prove more difficult: modeling helix slippage with a branch transducer is awkward, let alone more radical changes in structure; pseudoknots, too, are impossible with the models we have described here. Even so, variants of our models could be used for proposing candidate alignments for more accurate scoring by such models.
An implementation of inference algorithms for models on the three-taxon phylogeny is sufficient to construct a MCMC sampling algorithm over many sequences on an arbitrary phylogeny. A sketch of such a sampling algorithm is as follows: at each step of the sampling algorithm, we re-sample the sequence and structure of the ancestral node , conditioned on the sequences and structures of , and . The structural alignment of all four sequences can change at each step, providing for fast mixing and guaranteeing ergodicity. This move is similar to the sampler proposed by [31] for models with a HMM structure. Note that this, in principle, permits construction of a crude sampler to simultaneously infer phylogeny as well, by proposing and accepting or rejecting changes to the underlying tree as well as the implied structural alignment.
Reconstructing structural changes of large RNAs using the three-way sampling kernel which we have described would require resources far in excess of those currently available; barring the availability of supercomputers with terabytes of memory, such algorithms will only be feasible for short RNAs (Table 7). A promising direction is to consider variations on the three-way sampling kernel, such as the importance-sampling approach described for the TKF model by [32]. This approach first proposes an ancestor by aligning extant sequence to (ignoring ); then, in a second step, the proposed is independently aligned to . The proposed three-way alignment and reconstruction is then randomly accepted (or rejected) using a Hastings ratio based on the three-way transducer composition. The complexity of this kernel is the same as the pairwise case; with suitable constraints, this is feasible for RNA grammars on present hardware, at least for ribosomal domains (if not yet whole subunits—although pairwise alignment of those should also be possible soon). The approach of Redelings and Suchard therefore merits future consideration in the context of modeling the evolution of RNAs on a tree.
An alternative MCMC scheme for sampling RNA phylogeny, structure and alignment was developed for the SimulFold program [101]. SimulFold does not use a strictly normalized probabilistic model, resulting in some oddities in the ways that structure and indels interact (for example, it does not penalize deletion of one half of a basepair). Currently, it is not clear how appropriate SimulFold would be for ancestral reconstruction, although it has several advantages (e.g., explicit treatment of pseudoknots). Of course, MCMC kernels are inherently adaptable to other purposes: the MCMC moves developed for SimulFold may be useful for inference under different models.
This paper focuses on the case where the tree topology is known, but many of the methods which we have described can be extended to the more general case where none of the possible constraints (phylogeny, structure or alignment) are final. For example, the probabilistic framework readily allows us to compare likelihoods of two different phylogenetic trees by constructing a composite transducer for each tree. Thus, the MCMC samplers described above for alignments could, in principal, be extended to phylogenies (albeit at a computational cost).
While MCMC provides the most information about the posterior distribution of evolutionary histories, in practice a maximum likelihood inference may be adequate (and typically much faster). The progressive profiling used by the Ortheus program for reconstructing ancestral genomes is promising [33]. This approach is similar to a progressive multiple alignment algorithm, in that it proceeds via a single postorder (leaf-to-root) traversal of the phylogeny. As each node is visited, a profile is generated for that node, by aligning the profiles of its children to a composite transducer using DP, then sampling a finite number of traceback paths through the DP matrix. The profile is not linear: the sampled paths instead form a reticulate network, a.k.a. a partial order graph [102]. An equivalent of Ortheus for RNA reconstruction should be possible, representing the intermediate profiles using transducers.
Given the excellent performance of Stemloc-AMA 's sequence annealing, particularly when measured using its own scoring metric (AMA), such posterior-decoding methods should also be considered for reconstruction.
In summary, the evolutionary models and algorithms we have described form a systematic theoretical platform on which we can test different optimization and sampling strategies for studying the structural evolution of RNA gene families in detail. Stochastic grammars are powerful tools for this task, although they will not be the only tools we need, particularly as we move towards modeling RNA evolution in greater detail. Our hope is that these algorithms will allow us to test and refine our understanding of RNA evolution by computational reconstruction and (eventually) direct experimental investigation of early ribonucleic machines.
|
10.1371/journal.pmed.1002656 | Delivery outcomes in term births after bariatric surgery: Population-based matched cohort study | Obesity increases the risk of adverse delivery outcomes. Whether weight loss induced by bariatric surgery influences these risks remains to be determined. The objective was to investigate the risk of adverse delivery outcomes among post-surgery women compared with women without bariatric surgery history but with similar characteristics.
We identified 801,443 singleton live-born term births (≥37 gestational weeks) in the Swedish Medical Birth Register between 1 January 2006 and 31 December 2013, of which 1,929 were in women with a history of bariatric surgery and a pre-surgery weight available from the Scandinavian Obesity Surgery Registry. For each post-surgery delivery, up to 5 control deliveries were matched by maternal pre-surgery BMI (early-pregnancy BMI used for controls), age, parity, smoking, education, height, country of birth, and delivery year (N post-surgery deliveries:matched controls = 1,431:4,476). The main outcome measures were mode of delivery, induction of labor, post-term pregnancy (≥42 + 0 gestational weeks), epidural analgesia, fetal distress, labor dystocia, peripartum infection, obstetric anal sphincter injury (perineal tear grade III–IV), and postpartum hemorrhage. Among the women with a history of bariatric surgery, the mean pre-surgery BMI was 42.6 kg/m2, the median surgery-to-conception interval was 1.4 years, and the mean BMI loss between surgery and early pregnancy was 13.5 kg/m2 (38 kg). Compared to matched control women, post-surgery women were less likely to have cesarean delivery (18.2% versus 25.0%, risk ratio [RR] 0.70, 95% CI 0.60–0.80), especially emergency cesarean (6.8% versus 15.1%, RR 0.40, 95% CI 0.31–0.51). Post-surgery women also had lower risks of instrumental delivery (5.0% versus 6.5%, RR 0.73, 95% CI 0.53–0.98), induction of labor (23.4% versus 34.0%, RR 0.68, 95% CI 0.59–0.78), post-term pregnancy (4.2% versus 10.3%, RR 0.40, 95% CI 0.30–0.53), obstetric anal sphincter injury (1.5% versus 2.9%, RR 0.46, 95% CI 0.25–0.81), and postpartum hemorrhage (4.6% versus 8.0%, RR 0.58, 95% CI 0.44–0.76). Since this study was not randomized, a limitation is the possibility of selection bias, despite our efforts using careful matching.
Bariatric-surgery-induced weight loss was associated with lower risks for adverse delivery outcomes in term births.
| Obesity is prevalent and is a major health problem in pregnancy and childbirth.
Bariatric surgery induces large and sustained weight loss and is becoming more common in Sweden and other developed countries.
Whether bariatric-surgery-induced weight loss influences delivery outcomes is not well described.
We compared 1,431 term births in women with bariatric surgery history with 4,476 population control births matched on pre-surgery BMI in cases and early-pregnancy BMI in controls, as well as several other maternal characteristics.
Women with bariatric surgery history had lower risk of instrumental delivery and cesarean delivery during labor.
Women with bariatric surgery history, compared with controls, had substantially lower risks for post-term pregnancy, induction of labor, epidural analgesia, and delivery complications such as labor dystocia, fetal distress, peripartum infection, obstetric anal sphincter injury, and postpartum hemorrhage.
Bariatric-surgery-induced weight loss was associated with lower risks for adverse delivery outcomes in term births.
Although bariatric surgery appears beneficial for maternal delivery outcomes, other adverse pregnancy and infant outcomes have to be considered when counseling women on the safety of giving birth after bariatric surgery.
| Obesity in women of childbearing age has increased in the US and many other developed countries during the last decades [1]. The proportion of US women 20–39 years of age with obesity (body mass index [BMI] ≥ 30 kg/m2) was 35.7% in 2015–2016, and the corresponding proportion for class III obesity (BMI ≥ 40 kg/m2) was 7.8% [2]. Maternal obesity is associated with adverse delivery outcomes including induction of labor, cesarean delivery, labor dystocia, fetal distress, and postpartum hemorrhage [3].
Unfortunately, effective treatment options for obesity are limited. Bariatric surgery is the only treatment to date that induces large and sustained weight loss [4,5]. Although bariatric surgery is accepted as reasonably safe, uncertainty remains regarding the risks for a subsequent pregnancy and delivery [6–8]. We have previously reported lower risks of gestational diabetes and large-for-gestational-age births as well as higher risks of small-for-gestational-age birth and preterm birth in post-surgery women compared to controls matched for pre-surgery BMI [9,10]. Hence one would expect a lower risk for delivery complications.
Three meta-analyses on bariatric surgery and delivery outcomes reported non-significant differences for cesarean delivery and postpartum hemorrhage between women with and without a history of bariatric surgery [11–13]. Studies on bariatric surgery and delivery outcomes have generally been small, used heterogeneous control groups [14–19], or lacked a comparison group accounting for pre-surgery BMI [6,15,16,18–24]. We have identified only 3 previous studies on delivery outcomes that included a control group matched on pre-surgery BMI [14,17,25], of which all included fewer than 140 post-surgery deliveries. No previous study to our knowledge has investigated obstetric anal sphincter injury (OASIS; i.e., perineal tear) in women with bariatric surgery history compared with controls.
We conducted a population-based study using data from nationwide Swedish registers on singleton live term and post-term births (i.e., gestational age ≥37 completed weeks). We compared the risk of adverse delivery outcomes including cesarean delivery, instrumental delivery, post-term pregnancy (gestational age ≥42 completed weeks), induction of labor, labor dystocia, fetal distress, peripartum infection, OASIS, and postpartum hemorrhage among deliveries to women with versus without a bariatric surgery history but with otherwise similar characteristics.
In Sweden, prenatal and delivery care are tax funded, and the participation rate in the prenatal care program is almost 100%. The Swedish Medical Birth Register includes information on more than 98% of all births in Sweden since 1973 [26]. Information is prospectively collected during pregnancy, delivery, and the neonatal period using standardized prenatal, obstetric, and neonatal records. This matched cohort study was approved by the regional ethics committee in Stockholm, Sweden (No. 2013/730-31), and was conducted on de-identified data. By using the unique personal identification number assigned to each Swedish resident [27], data from the Medical Birth Register was linked to the Scandinavian Obesity Surgery Registry (SOReg), and the National Patient Register.
SOReg was established nationwide in 2007 and covers approximately 98.5% of all bariatric procedures in Sweden [28]. Local data from a few hospitals were available beginning in 2004. The register includes pre-surgery and follow-up information.
The National Patient Register includes diagnostic and surgical information on hospital admissions and hospital-based outpatient care visits, coded according to the Swedish versions of the International Classification of Diseases–10th revision (ICD-10) and the Nordic Medico-Statistical Committee Classification of Surgical Procedures, both used in Sweden from 1997 and onwards [29].
We did not work from a pre-specified analysis plan, but instead used the methods described in 2 previous studies [9,10]. We had access to 876,068 deliveries recorded in the Medical Birth Register between 1 January 2006 and 31 December 2013. We excluded births to mothers without a valid personal identification number at the time of delivery, who could not be linked to other registers for assessment of bariatric surgery status. We also excluded mothers with multiple births (since they differ regarding delivery outcomes). From the remaining 844,956 singleton deliveries with linkable mothers (S1 Fig), we excluded stillbirths (0.32%) and preterm births (<37 completed gestational weeks; 4.8%), as well as deliveries where gestational age was missing (0.04%). After these exclusions, 801,443 singleton live births at or after 37 gestational weeks remained, of which 3,105 occurred after bariatric surgery (performed between 1 January 1983 and 31 December 2013). Of these, 1,956 were performed between 1 January 2006 and 31 December 2013 and were included in SOReg. From SOReg, we recorded the bariatric surgery date, the type of procedure, and pre-surgery BMI (N = 1,929 had data on this variable).
We created a matched control cohort using births to women without bariatric surgery history (according to the National Patient Register and SOReg). Up to 5 control births were matched without replacement to each post-surgery birth; once a birth to a woman without bariatric surgery history was selected as a control, the same birth could not be selected again. The matching factors were maternal age (± 2 years), parity (primiparous or multiparous), pre-surgery BMI category (using early-pregnancy BMI in controls; 30 to <35, 35 to <40, 40 to <45, 45 to <50 or ≥50 kg/m2), early-pregnancy smoking status (non-smoker, 1–9 daily cigarettes, ≥10 daily cigarettes, or missing), educational level (<10, 10–12, or >12 years, or missing), height (<155, 155 to <165, 165 to <175, or ≥175 cm), country of birth (Nordic or non-Nordic), and delivery year (2006–2013; ±1 year).
Pre-surgery BMI was calculated from measured weight and height before surgery. Early-pregnancy BMI was calculated from measured weight and self-reported height at the first prenatal visit (median gestational week 10); self-reported smoking status was also registered at the first prenatal visit. To reduce measurement error and missingness, we used the median height from all registered pregnancies. Mother’s country of birth was retrieved from the Medical Birth Register and categorized into Nordic (Sweden, Denmark, Norway, Finland, Iceland) or non-Nordic.
The main outcomes of the study were cesarean delivery, instrumental delivery (forceps or vacuum extraction), induction of labor, post-term pregnancy (gestational age ≥ 42 completed weeks), epidural analgesia, and delivery complications including labor dystocia, fetal distress, peripartum infection, OASIS, and postpartum hemorrhage (defined as blood loss ≥ 1,000 ml within 24 hours following delivery).
Information on mode of delivery, induction of labor, epidural analgesia, and OASIS was obtained from standardized delivery record information. Information on labor dystocia (ICD-10: O62.0, O62.1, O62.2, O62.8, O62.9, O66.9, O63.0, O63.1, O63.9), fetal distress (ICD-10: O68.9), postpartum hemorrhage (ICD-10: O72, O67.8), and peripartum infection (ICD-10: O41.1, O75.2, O75.3, O85.9, O86.0) were obtained from diagnoses at hospital discharge. Cesarean delivery was further divided into elective (before labor start) and emergency. For emergency cesarean delivery, we also investigated indication, classified into dystocia, fetal distress, or other.
In order to characterise the relationship between BMI and the selected delivery outcomes, we used cubic splines applied to the whole study population with early-pregnancy BMI available (excluding births to women with previous bariatric surgery; remaining N = 729,867); we characterized the relationship between BMI as a continuous variable and the outcomes with knots placed at BMI 18.5, 22, 25, 30, 35, and 40 kg/m2.
The association between history of bariatric surgery (versus matched controls) and delivery outcomes was estimated by risk ratios (RRs) with 95% confidence intervals (CIs) by using modified Poisson regression models [30], conditioning on the matching set, with each set consisting of 1 post-surgery birth and up to 5 matched control births.
Ninety-eight percent (N = 1,896) of bariatric surgery procedures were gastric bypass, and 1.7% (N = 33) were sleeve gastrectomy, gastric banding, or biliopancreatic diversion with duodenal switch. The median surgery-to-conception interval was 1.4 years (interquartile range 0.8–2.3), and the median surgery-to-delivery interval was 2.1 years (interquartile range 1.5–3.0). Compared to pregnant women in the general population without bariatric surgery history, women with such history were older and were more likely to be obese, to smoke, to be of Nordic origin, and to be multiparous (all P < 0.001; Table 1). These distributional differences were eliminated by the matching procedure (Table 1).
Women with bariatric surgery history did not differ from matched controls regarding pre-surgery obesity status or age distribution, but still had a higher mean pre-surgery BMI (mean difference 0.64 kg/m2; 95% CI 0.55–0.73) and were on average 54 days older (95% CI 26–81; Table 1). In the surgery group, the mean BMI loss between surgery and early pregnancy was 13.5 kg/m2 (SD 4.2), corresponding to a weight loss of 38 kg (SD 12), or 31.7% (SD 8.5%) (Table 1; S2 Fig).
In Fig 1 we present the relation between maternal early-pregnancy BMI and adverse delivery outcomes based on 729,867 deliveries in women without bariatric surgery history. There was an association between increased BMI and the risk of cesarean delivery, induction of labor, post-term pregnancy, labor dystocia, fetal distress, postpartum hemorrhage, and peripartum infection. There was no association between BMI and epidural analgesia, and a negative association between BMI and OASIS.
Absolute risks for delivery outcomes in women with bariatric surgery history and matched controls are presented in Fig 2 (S3 Fig also includes unmatched population comparators), and adjusted relative risks in Fig 3. Compared to matched population controls, post-surgery women were less likely to have cesarean delivery (18.2% versus 25.0%; RR 0.70, 95% CI 0.60–0.80), with the decreased risk observed for emergency cesarean delivery (6.8% versus 15.1%; RR 0.40, 95% CI 0.31–0.51), but not for elective cesarean delivery (12.3% versus 11.7%; RR 1.02, 95% CI 0.85–1.22). When analyzing by indication for emergency cesarean delivery, there was a lower risk of cesarean delivery with labor dystocia and fetal distress indication. The risk of instrumental delivery was 5.0% versus 6.5% in post-surgery women compared to matched controls (RR 0.73, 95% CI 0.53–0.98). Post-surgery women had lower risks of induction of labor, post-term pregnancy, and epidural analgesia. Risks were also reduced for labor dystocia, fetal distress, peripartum infection, OASIS, and postpartum hemorrhage.
Effect modification by parity was found for cesarean delivery (primarily driven by differences in elective cesarean sections; P = 0.02), with no difference versus controls in parous women (Fig 4). Statistically significant effect modification was also found for induction of labor (greater risk reduction in primiparous than parous women; P = 0.002) and post-term pregnancy (greater risk reduction in parous than primiparous women; P = 0.04; Fig 4).
Regarding variations across surgery-to-conception interval categories, we found statistically significant effect modification for instrumental delivery, with greater protective effect for women who became pregnant during the first year after surgery (Fig 5).
In the main analysis, we did not account for clustering due to women giving birth more than once after bariatric surgery (N = 142; 9.9%). Therefore, we also performed a sensitivity analysis including only the first birth after surgery, resulting in similar estimates as in the main analysis (S1 Table).
In a mediation analysis where we adjusted for all matching variables and birth weight (mean birth weight in women with history of bariatric surgery was 3,458 g, and in the matched controls 3,813 g; P < 0.001), only minimal attenuation in results was observed as compared to the main model (S2 Table).
In this nationwide matched prospective cohort study, women with bariatric surgery history had lower risk of instrumental delivery and cesarean delivery during labor than controls matched on pre-surgery BMI. Furthermore, the risks of post-term pregnancy, induction of labor, epidural analgesia, labor dystocia, fetal distress, peripartum infection, OASIS, and postpartum hemorrhage were substantially lower in women with bariatric surgery history compared with the matched controls. In analyses stratified by parity, the reduced risk for cesarean delivery and OASIS was restricted to primiparous women. Furthermore, the reduced risk for instrumental delivery was only observed in women with <1 year between surgery and conception.
Reduction in cesarean deliveries was most pronounced for emergency cesareans, whereas the proportion of elective cesareans did not differ between women with bariatric surgery history and BMI-matched controls. Lesko and Peaceman [17] studied the risk of cesarean section in 70 women after bariatric surgery as compared to 140 controls matched on pre-surgery BMI, but did not report a lower risk. A US study by Abenhaim et al. using controls with class III obesity (BMI ≥ 40 kg/m2) found a lower risk of cesarean delivery overall [23], whereas other studies have not reported lower risk [6,24,25]. When we investigated indication for emergency cesarean, we found strong associations of emergency cesarean with labor dystocia and fetal distress, which no previous study to our knowledge has investigated. The lower risk estimates for labor induction in post-surgery deliveries observed in our study are in contrast with the findings of Abenhaim et al. [23] and an Israeli study of 326 cases compared with obese controls, which reported increased risk of labor induction in post-surgery deliveries [22]. A lower risk of post-term pregnancy but not labor induction was reported in a Danish cohort study with controls matched by early-pregnancy (rather than pre-surgery) BMI [6]. Rates of labor induction in the present study were lower in post-surgery deliveries. This could partially be attributed to a lower proportion of post-term pregnancy, which is a major indication for labor induction.
The almost halved risk of postpartum hemorrhage in post-surgery women is important, given that postpartum hemorrhage is a major cause of maternal morbidity and also mortality, and is supported by previous findings in studies that included obese controls [23,24] and controls matched on pre-surgery BMI [17].
Although BMI was negatively associated with OASIS risk, we observed a decreased risk for OASIS in women with bariatric surgery history compared with matched controls. OASIS is associated with incontinence and reduction in quality of life and also has major implications for reproductive health [32]. This is a novel finding, and could be seen despite the negative correlation of OASIS risk with maternal BMI [33].
Previous studies have reported a higher proportion of small-for-gestational-age infants and lower proportion of large-for-gestational-age infants born to mothers with bariatric surgery history [9,17,23], suggesting a shift towards lower risk of excessive fetal growth. The lower birth weight is also influenced by a lower proportion of post-term pregnancy, which is a major cause for macrosomia.
We hypothesized that the probable mechanism behind a lower risk of adverse delivery outcomes was the reduction in fetal size at birth. However, in the mediation analysis, introduction of birth weight did not have a major impact on the reduced risk of delivery outcomes observed. Hence, bariatric-surgery-induced weight loss appears to have beneficial effects on delivery outcomes independent of reduction in fetal growth. The reduction in delivery complications in mothers with bariatric surgery such as postpartum hemorrhage, OASIS, and peripartum infection may be caused by shorter duration of first and second stage labor, with fewer interventions and examinations and other factors associated with severe obesity in the mother. Furthermore, reduction in birth weight also decreases risk of uterine atony as well as birth canal laceration. We have previously reported a lower risk of gestational diabetes after bariatric surgery, which could lower the risk of labor induction, and also lower birth weight compared to matched control women [9]. Although there was no association between increased BMI and instrumental delivery, we found a reduced risk of instrumental delivery in women with previous bariatric surgery compared to matched population controls, which was unexpected.
Our study was of sufficient size to detect clinically meaningful differences in delivery outcomes between women with bariatric surgery history and matched controls. Further, for the women with bariatric surgery history, we had access to and matched on pre-surgery BMI instead of early-pregnancy BMI, which is often used but addresses a different research question. Pre-surgery matching answers the question of the effect of bariatric surgery itself, including weight loss and other metabolic/anatomical effects, on outcomes during pregnancy and delivery, while early-pregnancy BMI matching corresponds to a more clinical question, i.e., whether 2 women with similar early-pregnancy BMI, but one with a history of bariatric surgery and the other without such history, are expected to have similar outcomes during pregnancy [6,21,34].
Some studies have compared women with previous bariatric surgery to obese women or morbidly (class III) obese women in early pregnancy [20,22–24]. Such an approach is less precise than comparing based on pre-surgery BMI, which in our study ranged between 30 and 70 kg/m2. In contrast to the studies by Abenhaim et al. [23] and Parker et al. [24], which were based on delivery hospital discharge codes for pregnancies complicated by previous bariatric surgery, we used prospective exposure information from the high-quality SOReg, which covers 98% of all bariatric surgery procedures in Sweden (data are entered into the register by the surgeon who performed the bariatric surgery procedure). This approach is a major strength of our study and ensures that misclassification of the exposure is very limited in our study as compared to studies relying on discharge codes.
The study is based on data from the Swedish national health registers linked with SOReg. These health registers have a high quality, although there may be a proportion of up to 5% missing data for some variables [26,28,29].
This study was not randomized: Randomisation requiring pregnancy after bariatric surgery and a control intervention would be impossible to implement. The observational design may be affected by selection bias, despite our efforts using careful matching. Another potential limitation is that we used pre-surgery data regarding maternal BMI for post-surgery births, but data from early pregnancy for matched controls. Given the median surgery-to-conception interval of 1.4 years, we believe this was a reasonable approach.
Despite matching by BMI category, women with bariatric surgery history had significantly higher BMI at matching than controls. This is likely to result in a conservative bias, as greater BMI is associated with increased risks for most of the investigated outcomes. In the matching procedure, 26% of post-surgery pregnancies were excluded, foremost in the highest BMI category (BMI ≥ 50 kg/m2), because no matched controls could be identified. We were able to take into account the influence of country of birth but did not have access to information on ethnicity, which is a limitation affecting generalisability.
The vast majority of bariatric surgery procedures were gastric bypass, a procedure with greater malabsorptive effects than, for example, sleeve gastrectomy or gastric banding. Hence, our results may not be generalisable to women with other procedures.
A limitation was that we did not have data on indication for induction of labor. Furthermore, dystocia was based on diagnoses at discharge from the delivery hospital and was therefore more prevalent in women with instrumental or cesarean delivery as compared to non-instrumental vaginal delivery. However, we do not think that dystocia diagnosis was influenced by bariatric surgery history, and any misclassification would therefore be non-differential.
Bariatric-surgery-induced weight loss was associated with lower risk of cesarean and instrumental delivery and substantially lower risks of maternal complications during delivery and the early postpartum period. Given the magnitude of the observed effects, bariatric surgery may be an important procedure for improving delivery outcomes in obese and morbidly obese women. However, there is also an increased risk of complications for the infant, including small-for-gestational-age and preterm birth, that has to be taken into consideration.
|
10.1371/journal.pcbi.1003742 | Membrane Interaction of Bound Ligands Contributes to the Negative Binding Cooperativity of the EGF Receptor | The epidermal growth factor receptor (EGFR) plays a key role in regulating cell proliferation, migration, and differentiation, and aberrant EGFR signaling is implicated in a variety of cancers. EGFR signaling is triggered by extracellular ligand binding, which promotes EGFR dimerization and activation. Ligand-binding measurements are consistent with a negatively cooperative model in which the ligand-binding affinity at either binding site in an EGFR dimer is weaker when the other site is occupied by a ligand. This cooperativity is widely believed to be central to the effects of ligand concentration on EGFR-mediated intracellular signaling. Although the extracellular portion of the human EGFR dimer has been resolved crystallographically, the crystal structures do not reveal the structural origin of this negative cooperativity, which has remained unclear. Here we report the results of molecular dynamics simulations suggesting that asymmetrical interactions of the two binding sites with the membrane may be responsible (perhaps along with other factors) for this negative cooperativity. In particular, in our simulations the extracellular domains of an EGFR dimer spontaneously lay down on the membrane in an orientation in which favorable membrane contacts were made with one of the bound ligands, but could not be made with the other. Similar interactions were observed when EGFR was glycosylated, as it is in vivo.
| Epidermal growth factor receptor (EGFR) molecules are of central importance in cellular communication. Embedded in the cell membrane, these receptors bind epidermal growth factor (EGF) molecules outside the cell and translate this binding into specific biochemical signals inside the cell, which in turn trigger cell proliferation, migration, or differentiation. EGFR dysfunction has been implicated in a variety of cancers, and EGFR-targeting drugs are commonly used in cancer treatments. It has been widely assumed that the extracellular portion of an EGFR molecule protrudes perpendicularly from the cell membrane. In detailed, atomic-level computer simulations, however, we find that it lies down on the membrane, placing its EGF-binding site adjacent to the membrane surface. We further show that EGF may interact with EGFR in two distinct ways (with or without the involvement of the membrane). This may explain the experimental finding that an EGF molecule binds to EGFR more weakly at higher EGF concentration. This phenomenon, which is a manifestation of an underlying negative cooperativity, is an important but poorly understood characteristic of EGFR activity. In this study, we also model and analyze the glycan chains attached to EGFR, which are integral to its behavior in living cells.
| The epidermal growth factor receptor (EGFR), a member of the Her (ErbB) family of cell-surface receptors, is critical to a variety of cellular processes and is implicated in the development of several forms of cancer and other diseases [1]–[3]. In normal cells, EGFR activation is initiated by the binding of extracellular ligands from the epidermal growth factor (EGF) family [4]–[7], giving rise to the formation of active EGFR dimers, which transmit intracellular signals. It was first shown 30 years ago [8]–[12] that the Scatchard plots of EGF binding to EGFR are nonlinear (concave up), which is indicative of heterogeneous binding affinity. It has been further suggested that the heterogeneity in EGFR ligand binding may play an important role in determining the signaling response to different ligand concentrations [12]–[14].
More recently, in a study conducted by Pike and colleagues [15], a characterization of EGFR ligand binding based on a simultaneous fitting of binding isotherms from cells with different levels of EGFR expression showed that EGFR ligand binding may be described by a simple model (shown in Fig. 1A). In this model, which is consistent with earlier results [16], negative cooperativity underlies the heterogeneity of EGFR ligand binding [15], [16]: The binding affinity of a ligand at one EGFR binding site is smaller when the other site is occupied).
The structural origin of this negative cooperativity has been unclear. The existence of two structurally distinct binding sites in the doubly liganded dimer of Drosophila EGFR (dEGFR) [17] is consistent with its binding cooperativity. In crystal structures of the doubly liganded human EGFR dimer, however, the two binding sites are structurally virtually identical [18], [19]. A recent investigation, based on structural and biochemical analyses [20], suggested that the ligand-binding cooperativity may be explained by a conformational change in the ectodomain dimer. Although it is very plausible that such a scenario explains some of the negative cooperativity in EGFR, it is not clear that it represents the only, or even the main, contribution.
Here we use molecular dynamics (MD) simulations to investigate the structural basis of the negative cooperativity in the ligand binding of human EGFR. We simulated the human EGFR ectodomain dimer anchored to a lipid membrane by EGFR transmembrane (TM) helices. In our simulations of both singly and doubly liganded ectodomains, the dimer began in an upright orientation, with the dimer's long axis perpendicular to the membrane, then spontaneously rotated and lay down on the membrane in such a way that one of the binding sites faced the membrane, while the other faced the bulk solvent. The ligand in the membrane-facing site developed extensive favorable interactions with the membrane, and our approximate free energy calculations suggest that these interactions contribute a significant fraction of the ligand binding free energy. These findings are consistent with Förster resonance energy transfer (FRET) experiments [21], [22], which showed that some EGFR-bound ligands are positioned within 40 Å of the membrane, while others are positioned beyond 70 Å. In further simulations of glycosylated EGFR ectodomains, we found that the ectodomain orientation and the membrane interaction of the bound ligand appear compatible with the in vivo glycosylation of EGFR.
Based on our simulation findings, we suggest that the negative cooperativity in human EGFR ligand binding may arise in part from broken symmetry between the two bound ligands in an orientation in which the ectodomain rests on the membrane. Specifically, the simulations showed that the membrane favorably interacts with ligands bound to EGFR ectodomains resting on the membrane, and the results suggest that the high-affinity binding to an unliganded EGFR dimer may be attributed to this previously largely overlooked ligand-membrane interaction (Fig. 1A). Such a structural explanation is supported by the experimental finding that when high-affinity ligand binding is abolished, the distance between bound ligands and the membrane increases [22]. The mechanism we propose here also offers a straightforward explanation of the observation that negative ligand-binding cooperativity in human EGFR is observed only when the receptor is embedded in the membrane.
We first simulated a dimer of the ectodomains (domains I, II, III, and IV) and the single-helix TM segment of EGFR. Based on available crystal structures [18], [19], [23], the ectodomains were prepared in the form of an EGF-bound, back-to-back symmetric dimer, and the TM helices were prepared in the form of a TM dimer, with the N-terminal GxxxG-like motifs (where G represents glycine, or another amino acid with a small side chain) as the dimer interface [24]–[26]. The ectodomains were initially positioned upright, approximately perpendicular to the membrane (Fig. 1B). In the simulation, the ectodomains lay down toward the membrane surface, and approximately 1.7 µs into the simulation, one of the bound ligands developed extensive interactions with the membrane (Fig. 1C) in a partial-resting orientation of the ectodomain dimer. Later in the course of the simulation, the ectodomains lay flat on the membrane, producing a full-resting orientation (Fig. 1D). The simulation observations are supported by FRET measurements [21], [22], which have indicated that the EGFR ectodomain dimer may rest on the membrane. The simulation indicates that such orientations of the ectodomains are made possible because the linker segments between the ectodomains and the TM helices are not fully rigid, as has been previously suggested [23]. We previously simulated EGFR ectodomain monomers tethered to membrane-embedded TM helices. Although the ectodomain is ligand-free in the simulations, it came to rest on the membrane from an upright orientation in such a way that, if it were ligand-bound, the ligand would be in contact with the membrane (Fig. 6A in ref. [26]).
Notably, the orientation of the ectodomain dimer with respect to the membrane broke the symmetry between its two bound ligands: One of the ligands (the membrane-facing ligand) but not the other (the solvent-facing ligand) was in contact with the membrane (Fig. 1C). Once the ectodomain dimer rested on the membrane, the membrane-facing ligand developed extensive favorable interactions with the lipids. In particular, the hydrophobic residues of the EGF ligand, such as Pro7 and Leu8, were found to enter the hydrophobic interior of the membrane's extracellular leaflet (Fig. 1C, inset). As discussed in detail in later sections, similar membrane interactions were also observed for one of the bound ligands in two other EGFR dimer simulations we performed.
Although it is challenging to accurately calculate biomolecular binding free energies in simulation, generalized Born models can often provide a rough estimate. We estimated the binding free energy of each bound ligand using the molecular mechanics/generalized Born volume integration (MM/GBVI) model (see Methods) [27]. We initially calculated what we refer to as the ligand-protein interaction energy, which is an estimate for the free energy if there is little change in the protein structure on binding. As shown in Fig. 1E, the simplest application of the MM/GBVI method yields an interaction energy of an EGF ligand with EGFR of ∼80 kcal mol−1. The ligand-membrane interaction energy contributes an additional ∼25 kcal mol−1 to the membrane-facing ligand but almost nothing to the solvent-facing one.
The EGF-EGFR interaction energy of over 80 kcal mol−1 is much higher than the experimental value of the EGF binding free energy (10–15 kcal mol−1 [15]). As noted above, however, the computational quantity does not include the conformational free energy cost incurred when EGFR adopts its ligand-bound conformation during EGF binding. Using the structures of the active [18], [19] and inactive [26] dimers, the MM/GBVI model estimates the cost of EGFR's transition to the ligand-bound conformation to be 89 kcal mol−1 per monomer in the EGFR dimer. This is qualitatively consistent with our previously reported simulations [26], in which the ligand-bound conformation was not stable without the EGF ligands, and it suggests that the favorable EGF-EGFR interaction is approximately canceled out by EGFR adopting the unfavorable ligand-bound conformation. Despite this apparently satisfactory cancellation, we suspect that the individual canceling terms may still be overestimates, and we regard the use of the generalized Born model in the present context as qualitative.
To assess whether interactions with the membrane could contribute significantly to the experimental ligand-binding free energy, we do not only look at the large (∼25 kcal mol−1) estimated binding energy, but we also look at how this energy compares to the estimated ligand-protein interaction energy. Since the protein-membrane energy is a substantial fraction even of the large ligand-protein interaction energy, the generalized Born model supports the conclusion that the membrane-facing ligand binds to the EGFR dimer with higher affinity than does the solvent-facing one.
Intriguingly, we found that the hydrophobic patch formed by Pro7 and Leu8 is largely conserved in vertebrate EGF molecules (Fig. 1F). Pro7 is especially well conserved in vertebrate EGF. Leu8 is less conserved, but this position is hydrophobic in the majority of vertebrate EGF members despite being solvent accessible.
In vivo, human EGFR is glycosylated on the ectodomains [28], and 10 of the receptor's 12 potential glycosylation sites are found to be fully or partially occupied by a variety of large, branching glycans [29], [30]. It is conceivable that the relatively bulky glycans may preclude an EGFR ectodomain dimer from resting on the membrane. To test whether this is the case, we modeled and simulated an EGFR ectodomain–TM dimer system with full glycosylation (Fig. 2A).
We decorated the EGFR ectodomains at the 10 identified glycosylation sites [29], [30] with three types of glycans (BiS1F1, Man6, and Man8) that are common in EGFR glycosylation (see Methods, Fig. S1, and Table S1). The simulation shows that the glycosylation does not disrupt the ligand-membrane contacts (Fig. 2B) we observed in nonglycosylated EGFR. In the course of the simulation starting from the partial-resting orientation (Fig. 1C), the orientation of the ectodomains with respect to the membrane remained unchanged, with the membrane-facing ligand embedded in the membrane and the other ligand facing the solvent. The simulation showed that, in addition to interglycan interactions, the polar glycans interact extensively with the protein and the lipid head groups. The flexibility of glycans allowed the ectodomains to rest on the membrane. The glycans were found to be distributed adjacent to the protein surfaces, rather than protruding into the solvent (Fig. S2). The membrane-facing ligand of the glycosylated EGFR exhibited the same degree of membrane interactions in the simulation (Fig. 2C) as that of the nonglycosylated EGFR (as indicated by the values of t = 0 in Fig. 2C). In fact, the membrane embedding of the glycosylated-EGFR membrane-facing ligand was slightly deeper than that of the membrane-facing ligand in nonglycosylated EGFR.
We calculated that the membrane interaction contributes an additional estimated ∼30% to the membrane-facing ligand's MM/GBVI binding energy but virtually nothing to that of the solvent-facing ligand (Fig. 2C). We thus conclude that robust interaction between EGFR-bound ligands and the membrane, which may contribute to the heterogeneous ligand binding in an EGFR dimer, is accessible to glycosylated as well as nonglycosylated EGFR.
Having demonstrated that the ectodomains in a two-ligand EGFR dimer may rest on the membrane and that the two ligands may differ in their interactions with the membrane due to the ectodomain's orientation, we further simulated the one-ligand EGFR dimer. These simulations suggest that the ectodomains may rest on the membrane and that the ligand in a one-ligand EGFR dimer may also develop favorable interactions with the membrane, thus providing a structural model for high-affinity binding in the one-ligand EGFR dimer (Fig. 1A).
Because a crystal structure of a singly liganded ectodomain of an EGFR dimer is not yet available, we made a model based on the crystal structure of the two-ligand ectodomain dimer by removing one bound ligand from the crystal structure [26]. We here simulated the one-ligand ectodomain dimer in this conformation, connected with the TM segments, three times. In all three simulations, the ectodomain dimer, which was initially in an upright orientation, spontaneously lay down on the membrane (Fig. 3B), allowing the ligand to come into contact and develop extensive interactions with the membrane. We again calculated the MM/GBVI energy of the ligand's interaction with the membrane (Fig. 3C). The results suggest that the membrane interaction is energetically favorable, and that the free energy increase associated with the ligand's membrane interaction is a significant fraction of the free energy arising from its interaction with EGFR.
The favorable nature of the ligand-membrane interaction strongly suggests that the membrane-facing binding sites are associated with high-affinity binding. This notion is notably supported by the observation from FRET experiments that abolishing the high-affinity binding leads to a significant increase in the average distance between the ligands and the membrane [22]. Assuming a thermodynamically equilibrated system, in one-ligand EGFR dimers the ligands predominantly occupy the high-affinity sites facing the membrane (Fig. 1A).
We also performed a similar simulation of a fully glycosylated, one-ligand ectodomain dimer attached to the TM segments. Starting form an upright conformation, the ectodomain dimer again spontaneously lay down, with its bound ligand coming into contact with the membrane (Fig. 3B and 3C). This simulation suggests that a glycosylated one-ligand dimer may also prefer to rest on the membrane in such a way that its bound ligand faces the membrane, and that the ligand-membrane interactions are energetically favorable
A simulation study of EGFR [22] has previously suggested that ectodomain interactions with the membrane may be at the root of the observed negative cooperativity of ligand binding. It was further suggested that the negative cooperativity may arise from the ectodomain's transition to a dEGFR-like asymmetric conformation, induced by interactions with the membrane. Although our simulations also suggest the important role of EGFR ectodomain interactions with the membrane, our simulations did not show a robust transition from a symmetric to an asymmetric conformation in the ectodomain dimer. Fig. 4A shows that, other than minor deviations due to the inherent flexibility of the loop regions, the dimer's two ectodomain subunits were nearly conformationally identical in our simulations. In particular, the conformations of domain II in the two subunits are highly similar, whereas in the dEGFR dimer the domain II is straight in one subunit and bent in the other, which ultimately leads to the different conformations of the two binding sites. This is illustrated in Fig. 4B, where the angle characterizing the bending of domain II is plotted. The angles of the two EGFR subunits were approximately the same in our simulations of the two-ligand dimer, much as they are in crystal structures [18], [19]. Our MM/GBVI calculation supports the notion that the two receptors of the two-ligand EGFR dimer maintain similar binding-site conformations while resting on the membrane: The two ligands have comparable MM/GBVI interaction energies with the receptors (Fig. 5B), including cases in which the receptors are glycosylated (Fig. 2C).
On the other hand, for the one-ligand dimer, which assumes asymmetric conformations, the angles differ significantly between the two subunits. Similarly, in our simulations the average root-mean-square deviation (RMSD) of the Cα atoms of domains I, II, and III between the two subunits was significantly lower for the two-ligand ectodomain dimer (2.7±0.3 Å when glycosylated and 2.6±0.5 Å when nonglycosylated) than for the one-ligand dimer (4.4±0.2 Å when glycosylated and 4.7±0.3 Å when nonglycosylated).
We performed three independent simulations of the two-ligand, nonglycosylated EGFR dimer. As discussed above, in one of these simulations, the ectodomain dimer first assumed a partial-resting orientation and eventually lay flat on the membrane (Fig. 1; also shown as Simulation 1 in Fig. 5). In another simulation (Fig. 5A, Simulation 2), the system arrived at a similar partial-resting orientation and remained there to the end of the simulation. In the third simulation (Fig. 5A, Simulation 3), the ectodomain dimer was found to rest sideways on the membrane surface. What is common to all three simulations, however, is that only one bound ligand made extensive contact with the membrane (Fig. 5B, upper panels) despite the variation in the orientation of the ectodomain dimer. The buried surface area of the membrane-facing ligand was consistently greater than that of the solvent-facing ligand (Fig. 5B): 2,300±100 Å2 versus 1,600±100 Å2. A substantial portion (up to 200 Å2) of the approximately 700-Å2 difference is due to the embedding of Pro7 and Leu8 in the membrane. Our MM/GBVI calculations also consistently suggest that ligand-membrane interactions contribute a significant fraction to the free energy of ligand-receptor binding (Fig. 5B).
Earlier FRET measurements indicated that EGF bound to EGFR dimers falls into two groups: one in which the bound ligand is close to the membrane, and another in which it is farther away. Specifically, the FRET results showed that the N termini of the EGF molecules in the “close” group are no more than 35–40 Å from the membrane, and the N termini of the EGF molecules in the “far” group are no closer than 69–71 Å from the membrane [21], [22]. Our simulation results (Figs. 2C and 5B) agree with these data (see the description of the distance measurements in the Methods section) and the simulation conformations are similar to those in the structural model proposed by Kästner et al. based on their FRET results [31]. In all of our simulations, the membrane-facing ligands were close to the membrane surface (∼10 Å), and thus belong to the former population. This population may also include the solvent-facing ligands in cases in which the ectodomain dimer rests flat on the membrane, such as at the end of Simulation 1, where the membrane distance is ∼40 Å for the solvent-facing ligand. The latter population, on the other hand, may consist of the solvent-facing ligands in dimers such as those observed in Simulations 2 and 3, as well as those in the upright dimers (with distances of ∼80–120 Å).
Combining the observations from the simulations with those from the FRET measurements, we suggest that it is unlikely that the negative cooperativity of ligand binding can be attributed to a single specific orientation of the ectodomain dimer. We instead suggest that the cooperativity is associated with an ensemble of ectodomain-dimer orientations, with the shared feature that the high-affinity ligand binding occurs at the membrane-facing binding site. This provides a straightforward explanation for the experimental observation that abolishing high-affinity ligand binding increases the average ligand-membrane distance [22]. Additionally, our simulations showed that free EGF molecules may interact favorably with and be attached to the membrane (Fig. S3). This simulation finding, combined with the observation that Spitz ligands (which bind to dEGFR) need to be palmitoylated (and thus attached to membrane) to activate dEGFR in vivo [32], raises the possibility that the ligand-binding process of EGFR may occur at the membrane surface.
Our simulations suggest that an EGFR ectodomain dimer may rest on the membrane, and that the interaction between a bound ligand and the membrane may lead to a breaking of the symmetry between the two ligands, thus contributing to the negative cooperativity of EGFR ligand binding (Fig. 1A). Our investigation is in part inspired by the FRET measurements of ligand distance from the membrane; based on these results, the orientation of EGFR ectodomains relative to the membrane was suggested to affect ectodomain conformations and give rise to the negative cooperativity [21], [22], [31]. The mechanism we propose here is particularly supported by the FRET finding that abolishing high-affinity ligand binding leads to a significant increase in the average distance between EGFR-bound ligands and the membrane [22]. Our simulations of glycosylated EGFR (to our knowledge the first simulations of a fully glycosylated receptor) showed that the mechanism we propose is compatible with EGFR glycosylation: A glycosylated ectodomain dimer may also rest on the membrane, and the attached glycans do not preclude interactions between the EGFR-bound ligand and the membrane.
In this investigation, we have largely focused on EGFR dimers because they are central to the negative cooperativity of EGFR ligand binding [14]. EGFR monomers may also bind ligands at a high affinity comparable to that of EGFR dimers [15], but the ectodomain structure of the ligand-bound EGFR monomer has not yet been resolved. In previous MD simulations, we showed that an EGFR monomer is similar to an EGFR dimer in that its ectodomains also rest on the membrane in a way that would allow membrane contact with the bound ligand [26]. From this observation, which is independent of any specific conformation of the ectodomains, it may be inferred that the high affinity of ligand binding in EGFR monomers could also be explained by favorable interactions between the membrane and the bound ligands.
Our simulations suggest that the ectodomains of an EGFR dimer may rest on the membrane and that a bound EGF ligand may be in direct and energetically favorable contact with the membrane. Our earlier simulations also suggest that EGFR monomer ectodomains may also rest on the membrane [26]. This does not imply, however, that the ectodomains are fixed on the membrane in well-defined orientations. It is likely that, on a timescale much longer than our simulations, the ectodomains convert from one orientation to another in a dynamic equilibrium. While the orientations in which the ectodomains rest on the membrane may predominate, the ectodomains likely access the other orientations that could be crucial to the process of ligand binding or EGFR dimerization.
A recent study [20] proposed that a conformational change from the so-called “flush” to the “staggered” arrangement between the two extracellular subunits in an EGFR dimer (Fig. 6A) may be at the root of the binding cooperativity of EGFR. While such a binding-cooperativity mechanism differs from the mechanism we propose here, these two mechanisms are not mutually exclusive. In agreement with the finding of Liu et al. [20] based on crystal structures, our simulations show that the two-ligand EGFR dimer prefers the staggered conformation and that the one-ligand and ligand-free EGFR dimers prefer the flush conformation [26]. Intriguingly, the ectodomain interaction with the membrane and the glycosylation of EGFR appear to strengthen this trend (Fig. 6B). From this observation, we suggest that the membrane may be of critical importance to the negative cooperativity of EGFR ligand binding, not only for its asymmetric interactions with the bound ligands, but also for its effect on the accessible conformational space of the ectodomain dimers. Further investigation is certainly needed to quantify the contribution of the conformational dynamics of the ectodomains and the contributions of ligand-membrane interactions to the ligand-binding cooperativity of EGFR. Further investigation would also be needed to clarify whether the membrane interactions of the ectodomains have any role in autoinhibition. We have not addressed this question, but we have previously shown that the membrane interactions of the EGFR kinase domain do play an autoinhibitory role [25], [26].
Experiments have shown that the ligand-binding cooperativity of EGFR is apparently missing for isolated EGFR dimer ectodomains in solution [12]. It was shown that the negative cooperativity may be partially recovered when the membrane is included in experiments of EGFR ectodomains attached to the TM helices [33]. Our suggested mechanism for the negative binding cooperativity, in which the membrane plays a central role, offers a straightforward explanation for these findings. If the asymmetry between the bound ligands in an EGFR dimer, and thus the binding cooperativity, is indeed associated with the difference in the interactions of bound ligands with the cell membrane, the absence of the membrane would naturally eliminate the binding cooperativity. Likewise, the lack of cooperativity for detergent-solubilized EGFR [34] may be explained by the absence of an extended membrane capable of interacting with EGFR-bound ligands. It has been shown that mutations at the intracellular domains of EGFR yield nearly linear Scatchard plots [33]. Although these Scatchard plots could reflect a weakened negative cooperativity due to these mutations, and thus suggest that the root of the negative cooperativity may lie beyond the ectodomains and the membrane, there is an alternative explanation: that the dimerization prior to ligand binding, which is a prerequisite of the binding cooperativity [15], was weakened, leading to both a near-linear Scatchard plot and a difficulty in using the plot to reliably quantify binding cooperativity [20].
Our investigation of the relationship between the EGFR ectodomains and the cell membrane using atomistic, long-timescale MD simulations suggests a structural mechanism for the negative cooperativity of ligand binding of EGFR dimers; in this mechanism, the ectodomains may rest on the membrane, and the presence of the membrane may break the symmetry between the two binding sites. These results add further support to the emerging view that interactions between EGFR and the membrane play a central role in many aspects of the regulation of EGFR signaling [25], [26], [34]–[36].
The simulations were performed on a special-purpose supercomputer, Anton [37], using the Amber ff99SB-ILDN [38]–[40] force field, combined with the ff99SB* backbone correction [41] for proteins, the CHARMM C36 force field [42] for lipids, and TIP3P [43] as the water model. The simulated systems were solvated in water with 0.15 M NaCl, with residue protonation states corresponding to pH 7. Additional Na+ ions were included to neutralize the net charges of the proteins (−3 for the extracellular domains of each EGFR, −4 for each EGF ligand) and the POPS lipids. As an equilibration stage, the protein backbone atoms were first restrained to their initial positions using a harmonic potential with a force constant of 1 kcal mol−1 Å−2. The force constant was linearly scaled down to zero over at least 50 ns. Simulations were performed in the NPT ensemble with T = 310 K and P = 1 bar using the MTK algorithm [44] with 20-ps relaxation time. Water molecules and all bond lengths to hydrogen atoms were constrained using M-SHAKE [45]. The simulation time step was 1 fs for the equilibration stage and 2 fs for production simulations; the r-RESPA integration method was used, with long-range electrostatics evaluated every 6 fs [46].
The glycosylation of EGFR was modeled based on the mass-spectrometry analysis of the CL1-0 cell line [30], which is broadly consistent with similar analysis on CL1-5 and A431 cell lines [29], [30]. Since EGFR glycan attachments in the cell are very diverse—for every glycosylation site there is a large number of different glycan types that can be attached to it—we chose glycans among the most commonly found at the specific sites. These three common types are BiS1F1, Man6, and Man8 (Fig. S1 and Table S1). The glycan structures for the initial models were obtained using the Glycam web service [47] and then adjusted in VMD [48] to avoid clashes with protein and membrane. The simulations were performed with the GLYCAM06 force field [49] applied to the glycans.
The simulated systems included the ectodomain–TM dimers with two EGF molecules bound (three simulations of 2.6, 1.2, and 2.1 µs; ∼315,000 atoms) and with one EGF molecule bound (three simulations of 2.5, 2.3, and 0.9 µs; ∼300,000 atoms), a two-ligand glycosylated ectodomain–TM dimer (3.0 µs; ∼310,000 atoms), a one-ligand glycosylated ectodomain–TM dimer (8.3 µs; ∼300,000 atoms), and a single EGF molecule (see SI; two simulations of 8.9 and 8.3 µs; ∼62,000 atoms); a membrane was included in every case. Each system is set up such that each dimer is at least 25 Å from its periodic image.
The model membrane consisted of POPC lipids, with 30% (molar) POPC randomly replaced by POPS in the intracellular leaflet of the bilayer (only for the ectodomain–TM simulations) to approximately mimic the charge distribution in the cellular membrane [26], [50]. Modeling, analysis, and visualization were performed using VMD [48].
The distance between the EGF N terminus and the membrane, namely the distance from the N terminus to the plane through the phosphates of the extracellular lipid layer, was computed in a manner consistent with the FRET measurements [22].
The EGF-EGFR interaction energy estimation was based on the molecular mechanics/generalized Born volume integration (MM/GBVI) model [27] and performed using MOE software (Chemical Computing Group) [51]. The EGF-receptor binding energy was calculated for each snapshot from the difference of the energy of the EGF-receptor complex and the sum of isolated EGF and receptor energies. The EGF-membrane energy was calculated analogously. The conformational free energy of EGFR extracellular dimers was estimated based on the published coordinates of the full-length ligand-bound and ligand-free EGFR dimers [26] after energy minimization. Our calculations included domains I, II, III, and IV. The MM/GBVI energy is −34287.4 kcal mol−1 for the ligand-free dimer and −34110.2 kcal mol−1 for the ligand-bound dimer (the EGF ligands were not included in the calculation), and thus the conformational free energy cost for each monomer is 88.6 kcal mol−1.
|
10.1371/journal.pgen.1000143 | FHY1 Mediates Nuclear Import of the Light-Activated Phytochrome A Photoreceptor | The phytochrome (phy) family of photoreceptors is of crucial importance throughout the life cycle of higher plants. Light-induced nuclear import is required for most phytochrome responses. Nuclear accumulation of phyA is dependent on two related proteins called FHY1 (Far-red elongated HYpocotyl 1) and FHL (FHY1 Like), with FHY1 playing the predominant function. The transcription of FHY1 and FHL are controlled by FHY3 (Far-red elongated HYpocotyl 3) and FAR1 (FAr-red impaired Response 1), a related pair of transcription factors, which thus indirectly control phyA nuclear accumulation. FHY1 and FHL preferentially interact with the light-activated form of phyA, but the mechanism by which they enable photoreceptor accumulation in the nucleus remains unsolved. Sequence comparison of numerous FHY1-related proteins indicates that only the NLS located at the N-terminus and the phyA-interaction domain located at the C-terminus are conserved. We demonstrate that these two parts of FHY1 are sufficient for FHY1 function. phyA nuclear accumulation is inhibited in the presence of high levels of FHY1 variants unable to enter the nucleus. Furthermore, nuclear accumulation of phyA becomes light- and FHY1-independent when an NLS sequence is fused to phyA, strongly suggesting that FHY1 mediates nuclear import of light-activated phyA. In accordance with this idea, FHY1 and FHY3 become functionally dispensable in seedlings expressing a constitutively nuclear version of phyA. Our data suggest that the mechanism uncovered in Arabidopsis is conserved in higher plants. Moreover, this mechanism allows us to propose a model explaining why phyA needs a specific nuclear import pathway.
| In response to changes in the environment, animals can take shelter while the sessile plants must adapt to the prevalent conditions. Great plasticity in growth and development are striking examples of how plants cope with a changing environment. In plants, light is both a source of energy and an essential informational cue perceived by several classes of photoreceptors. Phytochrome-mediated light signaling is particularly well studied, because these photoreceptors control all aspects of the plant life cycle. The phytochromes are cytoplasmic in the dark and must enter the nucleus upon light activation to initiate signal transduction. How this important light-regulated event is achieved is poorly understood. Here we describe the function of an evolutionary conserved protein called FHY1 for Far-red elongated HYpocotyl 1. We demonstrate that FHY1 interacts with a light-activated phytochrome in the cytoplasm, allowing the complex to be transported into the nucleus. Interestingly, if this phytochrome can enter the nucleus by another mechanism, FHY1 is no longer required for seedling development, indicating that a major function of FHY1 is to chaperone an activated phytochrome into the nucleus. Our experiments suggest that this mechanism uncovered in Arabidopsis is widely conserved among flowering plants.
| Plants are sessile organisms and therefore have to adapt growth and development to the environmental conditions at their site of germination. Light is one of the most important factors directing such adaptive responses and it is involved in many developmental steps throughout the life of plants [1],[2]. To detect intensity, quality (wavelength) and direction of incident light plants have evolved a set of photoreceptors monitoring red/far-red (R/FR), blue/UV-A and UV-B [3]–[7]. The phytochrome family of red/far-red photoreceptors plays a key role in seed germination, leaf and stem development, circadian rhythms, shade avoidance and induction of flowering [8]. Although in higher plants phytochromes are not the primary photoreceptors controlling phototropism and chloroplast movements, the phytochromes modulate these responses [9]–[11].
Phytochromes are homodimeric chromoproteins containing the linear tetrapyrole phytochromobilin as chromophore. They photoconvert between two spectrally distinct forms: the red-light-absorbing Pr and the biologically active far-red light-absorbing Pfr form [3],[12]. As the absorption spectra of the two forms overlap the photoconversion is not complete in either direction. Irradiation with light therefore results in a wavelength-specific equilibrium between the Pr and Pfr forms, with only ∼2% Pfr in far-red light and ∼85% Pfr in red light [13]. Under natural conditions the Pfr/Pr ratio differs dramatically depending on the position of the plant within the community (canopy shade versus open environment) [14],[15].
In Arabidopsis the phytochrome gene family consists of five members (PHYA–E), among which PHYA and PHYB play the most prominent functions [16]. phyB is the major red light receptor and mediates the red/far-red reversible low fluence response (LFR). Other members of the phytochrome family contribute to responses primarily controlled by phyB. In contrast, responses to continuous far-red light (high irradiance response, HIR) and to single light pulse of very low fluence light (VLFR) depend exclusively on phyA [1],[3],[12]. Photoreceptor mutants have reduced fitness but only the phyA mutant is conditionally lethal, highlighting the importance of this photoreceptor [17],[18]. Its functional importance is further revealed by the high degree of sequence conservation among all angiosperms [19]. phyA is also crucial for the modulation of phototropin responses such as the enhancement of phototropism [10],[11].
The subcellular localization of phytochromes is tightly regulated by light. They localize to the cytosol in the dark but translocate into the nucleus upon light activation, where they interact with several transcription factors (e.g. PIFs, phytochrome interacting factors) [20]–[24]. Given that light-activated phytochromes localize to the nucleus and interact with transcription factors, it is not surprising that 10–20% of the genes in Arabidopsis are subject to regulation by red and/or far-red light [25]. Consequently, nuclear accumulation of the photoreceptor is a key step in both phyA and phyB signaling [26]–[29]. The C-terminal half of phyB presumably contains an Nuclear Localization Signal (NLS), which is masked in the dark by the N-terminal half of the photoreceptor. Light triggers a conformational change, potentially unmasking the NLS and allowing nuclear transport of phyB [30]. This model predicts that the general nuclear import machinery is sufficient for phyB nuclear transport. In contrast, it has recently been shown that nuclear accumulation of phyA depends on two plant specific proteins called FHY1 and FHL [11],[26],[27]. Importantly, these proteins are not required for nuclear accumulation of phyB and for phyB signaling [26],[27]. FHY3 and FAR1, two transposase-related transcription factors, directly control FHY1 and FHL transcription and thus indirectly affect phyA nuclear accumulation [31].
FHY1 and FHL are small proteins (202 and 181 aa, respectively) containing an NLS and a Nuclear Export Sequence (NES) [32],[33]. High similarity between FHY1 and FHL is confined to the 36 most C-terminal amino acids. This small domain is necessary and sufficient for the light-regulated interaction with phyA in vitro and it is essential for function in vivo [26],[32]. Our previous work has shown that FHY1 and FHL are essential for phyA nuclear accumulation but the molecular mechanism involved remains elusive [26],[27]. Three models can explain the requirement of FHY1/FHL for light-regulated nuclear accumulation of phyA. i) FHY1/FHL may be essential for nuclear import of phyA and work as adapter proteins using their NLS and phyA binding-site to link phyA to the general nuclear import machinery. Alternatively, phyA would enter the nucleus independently of FHY1/FHL but ii) FHY1/FHL action may be required to stabilize phyA and protect it from degradation or iii) to trap it in the nucleus and prevent it from being exported back into the cytosol. In this report we provide strong evidence for a model, in which FHY1 and FHL work as adaptor proteins facilitating nuclear transport of phyA. Our data reveal an intriguing system for regulated nuclear transport of a cargo protein that does not contain an NLS of its own.
The high degree of sequence conservation among phyA in angiosperms suggests that the same might be true for phyA signaling components, such as FHY1 and FHL [19]. Yet, the amino acid identity between them is below 30% although they are functional homologs [33]. The only motifs conserved in FHY1 and FHL are the NLS (and to a minor degree the NES) in their N-terminal region and the phyA binding-site at the C-terminus. A database search for FHY1/FHL homologs revealed the presence of FHY1-like proteins in numerous plant species. This is interesting given the key function of FHY1/FHL in phyA signaling in Arabidopsis. The only motifs conserved between all the FHY1-like proteins found in the database and Arabidopsis FHY1/FHL are the NLS and the C-terminal phyA binding-site (Figure 1A). In contrast, the ∼150 aa linking the NLS and the motif essential for interaction with phyA are too diverse to be aligned. Together with the finding that the FHY1/FHL homologs from both rice and dandelion complement the fhy1 mutant phenotype (data not shown) this suggests that FHY1-like proteins may be defined as proteins containing an NLS and an “FHY1 type” phyA binding-site separated by a ∼150 aa spacer. To test whether this definition holds true we generated an artificial FHY1 consisting of an SV40 NLS and the C-terminal 36 aa of Arabidopsis FHY1 (FHY1 167–202 = FHY1 CT) with Yellow Fluorescent Protein (YFP) as a spacer in between. fhy1 mutant seedlings expressing this artificial FHY1 under the control of the CaMV 35S promoter were hypersensitive to FR, similar to fhy1 seedlings complemented with P35S∶YFP-FHY1 (Figure 1B). Furthermore, the artificial FHY1 accumulated in the nucleus and colocalized with phyA in light-induced nuclear speckles (Figure 1C, D) thus behaving like Arabidopsis FHY1/FHL [26],[27]. We therefore conclude that the NLS and the phyA binding-site of FHY1/FHL are necessary and sufficient for phyA nuclear accumulation.
Given that both the NLS and the phyA-interaction domain of FHY1 are sufficient for FHY1 activity we tested whether adding the NLS to phyA directly would be enough to promote nuclear localization of phyA fused to the Green Fluorescent Protein (GFP). phyA null mutants transformed with either PHYA-GFP (Figure 2A, B) or PHYA-NLS-GFP (Figure 2C–2F) driven by the PHYA promoter were analyzed microscopically. As previously described [23] nuclear accumulation of phyA-GFP was light-dependent (Figure 2A, B). In contrast, in lines expressing phyA-NLS-GFP nuclear localization was constitutive (Figure 2C, D). Nuclear bodies appeared extremely rapidly upon light excitation in phyA-NLS-GFP plants. When nuclei of etiolated phyA-NLS-GFP seedlings were imaged without a light treatment or immediately after a 5 sec red light pulse a smooth nucleoplasmic staining was observed (Figure 2E, data not shown). However as little as 1 minute after a 5 sec red light pulse nuclear bodies appeared in those nuclei (Figure 2F).
The phenotypic consequences of expressing a constitutively nuclear version of phyA was evaluated by comparing wild type, phyA and phyA transformed either with a construct encoding PHYA-GFP, PHYA-NLS-GFP or PHYA-NLS. Western blot analysis of dozens of independent transgenics showed that while we obtained lines expressing wild-type levels of phyA-GFP at a reasonable frequency (10–20%) we never found lines expressing high levels of either phyA-NLS or phyA-NLS-GFP (data not shown). For our phenotypic analysis we used two homozygous single insertion lines for each construct. phyA-GFP line 1 expressed wild-type levels of phyA while phyA-GFP line 2 expressed phyA levels comparable to the highest expressing phyA-NLS-GFP lines we obtained (Figure S1). Despite the relatively low levels of phyA, the phyA-NLS-GFP lines rescued the FR-HIR phenotype of phyA mutants very efficiently for hypocotyl elongation and anthocyanin accumulation (Figure 3A, B). Moreover, the phyA-NLS and phyA-NLS-GFP lines also showed a normal phyA-mediated VLFR response for inhibition of hypocotyl elongation in response to pulses of FR light (Figure 3C). It should also be noted that, despite having constitutively nuclear phyA, phyA-NLS (-GFP) lines did not show a cop (constitutively photomorphogenic) phenotype, indicating that nuclear import of phyA is not sufficient to trigger a light response (Figure 3 and data not shown).
The nuclear localization of phyA-NLS-GFP in darkness (Figure 2C), a condition, where there is much reduced phyA-FHY1 interaction, suggested that phyA-NLS-GFP nuclear accumulation did not require FHY1. In order to test this hypothesis genetically we crossed phyA phyA-NLS-GFP with fhy1 mutants and selected siblings in the F2 that were homozygous for phyA, fhy1 and the transgene. Microscopic analysis of such seedlings demonstrated that neither nuclear accumulation nor light induced formation of nuclear bodies of phyA-NLS-GFP required FHY1 (Figure 4C, D, G, H). In control experiments we confirmed that for phyA-GFP plants light-dependent nuclear import was strongly dependent on FHY1 (Figure 4A, B, E, F and data not shown) [26],[27]. We concentrated our analysis on fhy1 mutants because fhy1 has a much stronger phenotype than fhl [33].
Given that nuclear accumulation of phyA-NLS-GFP did no longer require FHY1, we tested whether fhy1 mutants expressing phyA-NLS-GFP had a normal light response to continuous FR light. Interestingly, both the hypocotyl elongation and anthocyanin accumulation phenotypes of fhy1 mutants were efficiently rescued by phyA-NLS-GFP but not by phyA-GFP (Figure 5). Our data thus indicate that FHY1 becomes dispensable in seedlings expressing phyA-NLS-GFP, suggesting that during the FR-HIR FHY1 is only necessary to control nuclear accumulation of phyA.
It was recently shown that FHY3 and FAR1, two closely related transcription factors, directly regulate the expression of FHY1 and FHL [31]. Given that phyA-NLS-GFP could rescue the fhy1 phenotype, we hypothesized that this construct may also be capable of rescuing fhy3 mutants, in which the major defect appears to be reduced FHY1 and FHL levels. We restricted our analysis to fhy3 mutants because FHY3 plays a significantly more important role for this response than FAR1 [31],[34]. We thus crossed fhy3 with phyA-NLS-GFP plants and analyzed homozygous wild type and mutant fhy3 siblings. Our phenotypic characterization of the response to far-red light showed that while phyA-NLS-GFP rescued the fhy3 mutant phenotype phyA-GFP could not (Figure 6). Our results are thus consistent with the notion that the major function of FHY1 and FHY3 is to respectively operate a directly and indirectly control of phyA nuclear accumulation.
The only functionally important and widely conserved parts of FHY1 are the NLS and the phyA-interaction domain (Figure 1) [26],[32],[33]. Moreover, nuclear accumulation of phyA-NLS-GFP occurred independently of light and FHY1 (Figures 2 and 4). Taken together these data support the notion that FHY1 mediates light-dependent nuclear import of phyA upon interaction in the cytoplasm. A prediction of this model is that over-expression of either native or artificial FHY1 lacking the NLS should sequester phyA in the cytoplasm and thus result in a dominant negative phenotype.
To test this hypothesis we omitted the SV40 NLS in the artificial FHY1 or replaced it by an NES and transformed the constructs (i.e. (NES-) YFP-FHY1 CT) into wild-type plants. As the fusion proteins encoded by the constructs are below the size exclusion limit of the nuclear pore [35] they can enter the nucleus by diffusion but do not accumulate there due to the absence of an NLS. The NES containing version, which is predicted to be actively exported from the nucleus, localized mainly to the cytosol (Figure 7B). As predicted by the nuclear import model, seedlings expressing these constructs were strongly hyposensitive to FR (Figure 7A). This phenotype is consistent with the previous finding that FHY1 containing a disrupted NLS does not complement the fhy1 phenotype but rather results in an almost complete loss of FR sensitivity [32]. Western blot analysis confirmed that the phyA levels were normal in seedlings expressing (NES-) YFP-FHY1 CT thus excluding the possibility that reduced amounts of phyA were responsible for the dominant negative phenotype (Figure S2). However, NES-YFP-FHY1 CT strongly inhibited phyA nuclear accumulation when crossed into plants expressing Cyan Fluorescent Protein (CFP) tagged phyA (Figure 7B). This suggests that in the cytosol NES-YFP-FHY1 CT competes with endogenous FHY1/FHL for binding to phyA (-CFP) and thereby interferes with phyA (-CFP) nuclear transport.
It was previously shown that FHY1 and its paralogue FHL are required for nuclear accumulation of phyA [26],[27]. The analysis of mutants clearly demonstrates that FHY1 plays the predominant function for both phyA nuclear accumulation and phyA-mediated light responses [27],[33]. This is presumably due to the roughly 15-fold higher level of FHY1 mRNA compared to FHL [33]. We therefore restricted our analysis to the fhy1 single mutant background, i.e. in the presence of functional FHL. Both FHY1 and FHL interact with light-activated phyA through a conserved C-terminal domain [26]. However, the mechanism, by which these proteins enable nuclear localization of phyA, remains to be established. Our phylogenetic analysis shows that, similarly to phyA, FHY1-related proteins are widely distributed among angiosperms (Figure 1A), suggesting conservation of this aspect of phyA signaling. Moreover, this analysis shows that among FHY1-like proteins only the amino-terminal NLS, which is essential for the interaction with importin alpha (Figure S3), and the carboxy-terminal phyA-interaction domain are conserved. It has previously been shown that both these domains of FHY1 are necessary for function [32]. Our analyses now show that they are also sufficient for FHY1 activity and that the ∼150 aa in between do not perform an essential function. The simplest model (hereafter termed “import” model) accounting for those results is that upon light excitation phyA interacts with FHY1 in the cytoplasm and that this complex enters the nucleus using the general nuclear import machinery (Figure S6).
According to this model adding a strong (and exposed) NLS to phyA should render phyA nuclear accumulation both light- and FHY1-independent. Our experiments show that these predictions are fulfilled in plants expressing phyA-NLS-GFP (Figures 2 and 4). In addition, when an FHY1 variant lacking the NLS sequence is over-expressed in wild-type plants this construct sequesters phyA in the cytoplasm and results in a dominant-negative de-etiolation phenotype (Figure 7). These observations are fully consistent with the notion that FHY1 mediates light-regulated phyA nuclear import by binding selectively to the active Pfr form of phyA in the cytosol and, thereby, linking phyA in a regulated manner to the nuclear import machinery (Figure S6). Our findings indicate that during de-etiolation in far-red light the system essential for nuclear localization of phyA, i.e. FHY3 and FHY1, can be replaced by simply attaching an NLS to phyA. It is however highly unlikely that such plants do not show a decrease in fitness under more natural conditions. The complex system relying on FHY3/FAR1 and FHY1/FHL is highly conserved in evolution (Figure 1A) [31] and FHY1-like proteins from dandelion and rice can compensate for the absence of FHY1 in Arabidopsis (data not shown). The strict conservation of FHY1-like proteins in angiosperms (in the sense of proteins containing a phyA binding-site linked to an NLS) points to a common molecular mechanism of phyA nuclear import and underlines the importance for regulated subcellular localization of phyA. An obvious advantage of the FHY1/phyA system over targeting phyA to the nucleus using an NLS is that it allows for co-existence of nuclear and cytosolic phyA pools and that the pool sizes can be regulated. This may be especially important with regard to possible cytosolic functions of phytochromes as recently described [11]. Nuclear import of phyB does not rely on the FHY1/FHY3 pathway but is light regulated nevertheless [23],[26],[27],[36],[37]. The FHY1-mediated nuclear import described here may explain how phyA can be imported so rapidly in response to light and how this import is possible under light conditions where the pool of Pfr is extremely small [13]. Such conditions are typically encountered for phyA-controlled light responses, such as the VLFR and the FR-HIR [1].
Two alternative scenarios for FHY1 function have been proposed, in which nuclear transport of phyA would not depend on FHY1-like proteins and may even be light-independent (i.e. both Pr and Pfr are transported) [26]. In these models (hereafter referred to as the “FHY1 nuclear anchor” and “protection” models) phyA could either be trapped in the nucleus or protected from degradation by binding to FHY1. As the phyA/FHY1 interaction is light dependent, these models would explain the light regulated nuclear accumulation of phyA as well. Yet, these hypotheses are inconsistent with our data for several reasons. In etiolated seedlings phyA protein levels are much higher than FHY1 (data not shown). This renders both the “FHY1 nuclear anchor” and the “protection” models difficult to envisage unless one FHY1 molecule would bind to multiple phyA proteins. In the “import” model one FHY1 molecule would transport one phyA dimer per cycle resulting in nuclear accumulation of large numbers of phyA molecules after multiple transport cycles. In addition, the subcellular localization of phyA-NLS-GFP was not affected in the fhy1 mutant background (Figure 4), which is only compatible with the “nuclear import” model. The normal localization of phyA-NLS-GFP in fhy1 mutants is also supported functionally, given that this construct complements fhy1 (Figure 5). Moreover, western blot analyses show that FHY1 does not affect phyA protein levels in far-red light irrespective of whether phyA enters the nucleus using FHY1 [11]. Moreover the abundance of constitutively nuclear phyA-GFP was also unaffected in the fhy1 background (Figure S4). These data indicate that FHY1 does not act by protecting phyA from degradation once the photoreceptor entered the nucleus.
Although phyA strongly accumulates in the nucleus in response to irradiation with FR in vivo spectroscopic measurements indicate that not significantly more than ∼2% of the total phyA is in the Pfr form under such conditions [38]. This strongly suggests that in FRc the major fraction of nuclear phyA is in the Pr and not the Pfr form [12]. Furthermore, yeast two hybrid experiments show that the light-induced interaction of FHY1 and phyA is R/FR reversible, suggesting that the phyA/FHY1 complex rapidly dissociates upon conversion of Pfr to Pr (Figure S5). It is, however, inherent to the “FHY1 nuclear anchor” and “protection” models that FHY1 has to be bound to phyA to inhibit its export into the cytosol or protect it from degradation. Again, the only model compatible with our findings is the “import” model, where an interaction for a limited time period would be sufficient to allow accumulation of phyA in the nucleus. A constitutive interaction of phyA and FHY1 may even interfere with phyA nuclear accumulation as it may block recycling of FHY1. Once in the nucleus phyA would be trapped in the “import” model – irrespective of whether it is in the Pr or Pfr form – because it is too big to exit the nucleus by diffusion. Taken together our findings strongly support the import model (Figure S6).
After accumulation in the nucleus phyA interacts with various transcription factors (e.g. PIFs) [20],[21],[24]. It is noteworthy that nuclear body formation is still light dependent for phyA-NLS-GFP (Figure 2). Moreover, formation of these subnuclear structures does not require FHY1 (Figure 4) although FHY1 and phyA have been found in light-induced nuclear bodies (Figure 1) [11],[26],[27]. The light-induced nuclear bodies may thus represent sites of phyA-PIF interaction as has previously been reported [39],[40]. Complementation of the fhy1 mutant by phyA-NLS (-GFP) shows that the interaction of phyA and downstream signaling components does not require FHY1. Rather, binding of FHY1 may prevent the interaction of phyA and effectors. If dissociation of the phyA/FHY1 complex were a prerequisite to initiate downstream signaling this would be an additional argument against the “FHY1 nuclear anchor” and “protection” models. Answering these questions will provide a “molecular” link between phyA nuclear accumulation and initiation of the signaling cascade(s) leading to transcriptional regulation of 10–20% of the genes in the Arabidopsis genome [25],[41],[42].
Adding a strong NLS to phyA results in light- and FHY1-independent nuclear accumulation of the protein. Nevertheless, dark-grown seedlings expressing such constitutively nuclear localized phyA display a normal morphology in darkness and still show normal light responses (phyA-mediated VLFR and HIR) (Figure 3). The fluence-rate dependency and the need for sustained excitation are hallmarks of the HIR [1] and it is well established that nuclear accumulation per se is an HIR [23],[43]. Yet, maximal hypocotyl growth inhibition and anthocyanin accumulation in seedlings expressing constitutively nuclear localized phyA are still fluence-rate dependent and require continuous irradiation (Figure 3). Thus, the “physiological HIR” does not derive exclusively from the HIR characteristics of phyA nuclear accumulation, indicating that in wild-type plants more than only one step in phyA signaling is an HIR. The phenotype of plants expressing constitutively nuclear phyA is thus clearly distinct to the partial det/cop phenotype of a mutant expressing a constitutively Pfr-like phyA [44]. Thus, control of phyA nuclear accumulation does not seem to play an essential role to prevent initiation of downstream signaling in the absence of light, which is crucial for the highly sensitive VLFR. The different affinities of phyA in the Pr and Pfr forms for downstream signaling components such as PIF1 and PIF3 may be sufficient to inhibit the induction of a VLFR in the dark.
Despite having a low total level of phyA (only around 25% of wild-type levels) inhibition of hypocotyl elongation and promotion of anthocyanin accumulation is very efficiently complemented in the phyA-NLS and phyA-NLS-GFP lines (Figure 3 and S1). These results suggest that nuclear phyA abundance (rather than total phyA levels) primarily controls these light responses. The strong phenotype of the fhy1 fhl and fhy3 far1 double mutants, which do not contain detectable levels of phyA in the nucleus, further supports this view [26],[31],[33]. Thus, nuclear accumulation of both phyA and phyB has been shown to be functionally important in Arabidopsis [26],[27],[29]. While these studies show that this is an important step of the signal transduction cascade for numerous phytochrome responses, they by no means exclude the possibility for cytoplasmic activities of the phytochromes. Cytoplasmic phytochrome responses are widely described in cryptogam species [45]–[47] and a recent paper indicates that cytoplasmic phyA may be required for the modulation of the phototropic response in Arabidopsis [11].
The vast majority of proteins enters the nucleus either passively or by active, importin-mediated transport [35]. However, there are nuclear localized proteins, which are too big to pass through the nuclear pore by diffusion but still do not contain an NLS. Similar to phyA many of these proteins use a piggyback mechanism and rely on the NLS of an interacting protein for nuclear transport [48]–[56]. Yet, in contrast to phyA, most of these proteins seem to interact with the NLS containing protein constitutively [48],[49],[51],[52],[56] or they are even part of a stable oligomeric complex with one of its components providing an NLS [53],[55]. Often the NLS containing protein also performs an essential function besides nuclear transport [49]–[52],[54]. Compared to the piggyback systems described above, the FHY1/phyA system is unique inasmuch as i) nuclear transport of the cargo protein is regulated by a conformational change of phyA [27] and ii) the NLS containing protein is dedicated exclusively to nuclear transport of the cargo protein given that FHY1 becomes dispensable in a strain where phyA possesses it own NLS (Figure 5).
To obtain the PPHYA∶PHYA-NLS-GFP5 construct (CF461), we inserted the following sequence AALQKKKRKVGGAAA between phyA and GFP5 of CF161 [27] using standard molecular biology techniques (NLS is underlined). PPHYA∶PHYA-NLS (CF460) is the same construct except that there is a stop codon directly after the last codon of the NLS sequence (i.e. does not contain GFP5).
Transgenic plants expressing phyA-NLS (CF460) and phyA-NLS-GFP (CF461) under the control of the PHYA promoter were obtained by transforming the constructs (CF460, CF461) into phyA-211 mutants by Agrobacterium-mediated transformation [57]. Transgenic plants were selected on 0.5× Murashige & Skoog (MS) medium (Duchefa), 0.7% agar (Sigma) with 30 µg/ml kanamycin. Single insertion lines were selected by determining the kanr/kans ratio in T2. Homozygous progeny of two representative single insertion lines for each construct were used for further studies.
pphyA40-phyA (contains PPHYA∶PHYA-CFP∶TerRbcS) is a T-DNA vector derived from pCHF40-phyA (contains P35S∶PHYA-CFP∶TerRbcS) and was used to generate plants expressing PHYA promoter driven phyA-CFP. pphyA40-phyA and pCHF40-phyA were obtained as described for pphyA30-phyA and pCHF30-phyA but contain ECFP (Clontech) instead of EYFP [26].
pCHF70-, pCHF72- and pCHF73-FHY1 167–202 are T-DNA vectors used to generate plants expressing CaMV 35S promoter driven YFP-FHY1 CT, NLS-YFP-FHY1 CT (artificial FHY1) and NES-YFP-FHY1 CT. Details regarding cloning of these constructs can be found in Text S1.
pCHF70-, pCHF72- and pCHF73-FHY1 167–202 were used for Agrobacterium-mediated transformation of Ler and fhy1-1 (only pCHF72-FHY1 167–202), pphyA40-phyA for transformation of phyA-201 [57]. Transgenic plants were selected on soil using BASTA (AgrEvo). Unless indicated otherwise, homozygous progeny of single insertion lines (1∶3 segregation of the selection marker) were used for the experiments.
Lines co-expressing either NLS- or NES-YFP-FHY1 CT and phyA-CFP were obtained by genetic crossing of Ler P35S∶NLS/NES-YFP-FHY1 CT and phyA-201 PPHYA∶PHYA-CFP (Ler ecotype). The F1 generation was used for microscopic analysis.
The phyA-211 fhy1-1 plants expressing phyA-NLS-GFP were obtained by crossing phyA-211 PPHYA∶PHYA-NLS-GFP5 (Col ecotype) into fhy1-1 (Ler ecotype) background. In F2 siblings were selected that were homozygous for the transgene and phyA-211 and either wild-type (i. e. phyA-211 FHY1 PPHYA∶PHYA-NLS-GFP5, in Col×Ler background) or mutant (i. e. phyA-211 fhy1-1 PPHYA∶PHYA-NLS-GFP5, in Col×Ler background) for FHY1. In all experiments with phyA-211 fhy1-1 PPHYA∶PHYA-GFP5 the phyA-211 FHY1 PPHYA∶PHYA-NLS-GFP5 in Col×Ler background was used as wild-type control.
phyA-211 fhy3-1 PPHYA∶PHYA-NLS-GFP plants were obtained by crossing phyA-211 PPHYA∶PHYA-NLS-GFP (Col ecotype) into fhy3-1 (Col ecotype) background. In F2 seedlings homozygous for phyA-211, fhy3-1 and the transgene were selected.
The Columbia (Col-0) and Landsberg erecta (Ler) ecotype of A. thaliana were used as wild type. phyA-211 [58], cop1-4 [59] and fhy3-1 [60],[61] are in Col while fhy1-1 [61],[62] and phyA-201 [58] are in Ler. phyA-211 PPHYA∶PHYA-GFP5 (A-GFP1), phyA-211 fhy1-1 PPHYA∶PHYA-GFP5, phyA-211 fhy3-1 PPHYA∶PHYA-GFP5 and fhy1-1 P35S∶YFP-FHY1 were previously described [27]. A second phyA-211 PPHYA∶PHYA-GFP5 line (A-GFP2) which was obtained during the screen described previously [27] was used because its phyA-GFP protein level is close to the phyA-NLS-GFP protein level in the lines we obtained.
Measurements of hypocotyl length in continuous FR light and anthocyanin accumulation were performed as described [63]. For hypocotyl length seedlings were grown on half-strength MS, 0.7% agar while for anthocyanin accumulation seedlings were grown on half-strength MS, 0.7% agar supplemented with 1.5% sucrose. The VLFR of hypocotyl elongation and its transition to the HIR was investigated essentially as described [64]. Briefly, chilled seeds were exposed to red light for 6 h followed by 18 h of incubation in darkness before transfer to pulses of FR (3 min) given at different dark intervals (117 min, 57 min, 27 min or 0 min = continuous FR). Hypocotyl length was measured to the nearest 0.5 mm after 3 d of treatment and is expressed relative to dark controls. Data are means and SE of at least 11 replicate boxes (10 seedlings per box).
Microscopic analyses in Figures 2A–D and 4A–D were performed with a Leica DM 600B equipped with Leica LTR6000 laser (software LAS, Leica Application Suite) using GFP and DAPI filter sets and a 20× air objective. 3-day-old dark-grown seedlings were directly observed under the microscope (dark condition). For light conditions, 3 day-old-dark-grown seedlings were pretreated for 10 min with white light before they were observed under the microscope.
For microscopic analyses in Figures 1C and 1D, 2E and 2F, 4E–H and 7B a Zeiss Axioscope 2 equipped with a 63× oil-immersion objective and GFP, YFP and CFP specific filter sets was used. The seedlings used for microscopy were grown as described in the figure legends.
Materials and Methods for Figures S1–S6 can be found in Text S1.
|
10.1371/journal.pgen.1004301 | Evolutionarily Diverged Regulation of X-chromosomal Genes as a Primal Event in Mouse Reproductive Isolation | Improper gene regulation is implicated in reproductive isolation, but its genetic and molecular bases are unknown. We previously reported that a mouse inter-subspecific X chromosome substitution strain shows reproductive isolation characterized by male-specific sterility due to disruption of meiotic entry in spermatogenesis. Here, we conducted comprehensive transcriptional profiling of the testicular cells of this strain by microarray. The results clearly revealed gross misregulation of gene expression in the substituted donor X chromosome. Such misregulation occurred prior to detectable spermatogenetic impairment, suggesting that it is a primal event in reproductive isolation. The misregulation of X-linked genes showed asymmetry; more genes were disproportionally downregulated rather than upregulated. Furthermore, this misregulation subsequently resulted in perturbation of global transcriptional regulation of autosomal genes, probably by cascading deleterious effects. Remarkably, this transcriptional misregulation was substantially restored by introduction of chromosome 1 from the same donor strain as the X chromosome. This finding implies that one of regulatory genes acting in trans for X-linked target genes is located on chromosome 1. This study collectively suggests that regulatory incompatibility is a major cause of reproductive isolation in the X chromosome substitution strain.
| Reproductive isolation characterized by male sterility and decreased viability is important for speciation, because it suppresses free genetic exchange between two diverged populations and accelerates the genetic divergence. One of the reproductive isolation phenomena, hybrid sterility (sterility in hybrid animals), is possibly caused by deleterious interactions between diverged genetic factors brought by two distinct populations. The polymorphism not only in protein-coding sequences but also in transcriptional regulatory sequences can cause the genetic incompatibility in hybrid animals. However, the precise genetic mechanisms of hybrid sterility are mostly unknown. Here, we report that the expression of X-linked genes derived from one mouse subspecies was largely misregulated in the genetic background of another subspecies. The misregulated expression of the X-linked genes subsequently affected the global expression of autosomal genes. The results collectively indicate that hybrid sterility between the two mouse subspecies is caused by misregulation of gene expression due to genetic incompatibility in the transcriptional regulatory circuitry. Such genetic incompatibility in transcriptional regulation likely underlies reproductive isolation in general.
| Reproductive isolation is a typical consequence of deleterious epistatic interactions between genes that have evolutionarily diverged in species or subspecies [1]–[3]. One of the most common types of postzygotic reproductive isolation is sterility of interspecific (or intersubspecific) hybrid progeny in F1 or later intercross or backcross generations. Although numerous sterility factors are mapped genetically, only a limited number of responsible genes have been cloned in mammals and non-mammalian vertebrates [4]. A confounding factor that makes it difficult to identify sterility-causing genes is that these genes function properly in their parental pure species (or subspecies), and deleterious interactions (i.e., genetic incompatibility) between them only occur in the hybrid genetic background [4]. Genetic incompatibility occurs in various levels, not only in physical interactions between responsible gene products (e.g., proteins), but also in the balance between expression levels of the responsible genes. Using hybrid animals between two mouse subspecies, Goncalves et al. [5] reported that the expression of approximately 30% of non-imprinted autosomal genes expressed in the liver is diverged by variants in cis-regulatory elements and variants in regulatory factors acting in trans. Furthermore, the rates of variants in cis-regulatory elements is higher than those in regulatory factors acting in trans. Such evolutionary divergence in transcriptional regulation may contribute to phenotypic differences and genetic incompatibilities underlying reproductive isolation.
Experimental genetic studies of animals including mice (Mus musculus) and fruit fly (Drosophila) suggest that the X chromosome often plays a central role in the sterility of hybrid males between species (or subspecies). The genetic loci that cause hybrid disruptions appear to be preferentially located on the X chromosome, which is referred to as the “large X-effect” [6]. The most convincing hypothesis to explain the large X-effect is that male reproductive genes preferentially accumulate on the X chromosome because of the immediate exposure of beneficial recessive mutations due to hemizygosity in males, and such X-linked reproductive genes evolve rapidly by positive selection [7], [8]. On the other hand, transcriptional inactivation during spermatogenic meiosis (i.e., meiotic sex chromosome inactivation [MSCI]) and postmeiotic stages (i.e., postmeiotic sex chromosome repression [PMSR]) influence the genetic content of the X chromosome. Thus, spermatogenic genes expressed during the premeiotic stage are preferentially enriched on the mouse X chromosome [9], [10] and the human X chromosome [11]. Despite the accumulation of premeiotically expressing genes on the X chromosome, many hybrid male mice show spermatogenic defects during meiotic and postmeiotic stages such as spermiogenesis [12]–[15]. Thus far, such discrepancy has not been fully understood.
Two mouse subspecies, M. m domesticus and M. m. musculus, diverged about 0.5 million years ago [16], and they manifest the early-stage of hybrid incompatibility phenotypes. Laboratory crosses between M. m. domesticus and M. m. musculus often yield fertile females, but sterile males [17]. The first mammalian hybrid sterility gene, PR domain containing 9 (Prdm9), was identified as the gene responsible for F1 hybrid sterility between M. m. musculus and M. m. domesticus [18]. Incompatibility between two Prdm9 alleles alone is not sufficient to drive reproductive isolation. Instead, the gene dosage of Prdm9 and combinations of particular Prdm9 alleles in conjunction with functional incompatibility with other X-linked gene(s) are important factors [19].
Natural habitats of M. m. musculus and M. m. domesticus overlap in Europe forming a hybrid zone, where hybrid populations exhibit reduced fertility and barriers to gene flow. It is known that X chromosomal genes have more limited flow beyond the hybrid zone than autosomal ones, suggesting a major role for the X chromosome in the reproductive isolation between the two subspecies [20]–[24]. The prominent role of the X chromosome was also supported by laboratory studies using F2 male progeny between the strains derived from the two subspecies [25] and the chromosome substitution strains, in which the X chromosome of the host strains (C57BL/6J [B6], predominantly derived from M. m. domesticus or wild M. m. domesticus-derived LEWES/EiJ) is substituted by the counterparts of the wild M. m. musculus-derived strains [15], [26]. We also reported a similar male-specific reproductive phenotype characterized by complete sterility in another chromosome substitution strain, B6-ChrXMSM, in which the entire length of the X chromosome is substituted by the counterpart of the M. m. molossinus-derived MSM strain in the genetic background of the B6 strain [13], [27], [28]. M. m. molossinus is an evolutional hybrid between M. m. musculus and M. m. castaneus, and its nuclear genome is predominantly derived from M. m. musculus [29]–[32]. Therefore, these observations imply that the same or a similar genetic basis underlies the male sterility of the X chromosome substitution strains that carry M. m. musculus-derived or M. m. molossinus-derived X chromosome in the genetic background of M. m. domesticus.
The most prominent and first detectable phenotype in B6-ChrXMSM males is a small number of primary spermatocytes during the first cycle of spermatogenesis, suggesting a defect in meiotic entry [13]. This phenotype ultimately results in a decreased number of meiotic and post-meiotic germ cells, and in turn contributes to a reduction of testis weight. Our previous quantitative trait locus (QTL) analysis mapped the responsible gene(s) for the reduced testis weight of adult males to the distal region of the X chromosome [27].
In this study, we report a sub-fertile phenotype in a newly established partial chromosome substitution strain, B6-ChrXTMSM, which has the MSM-derived distal half of the X chromosome in the genetic background of the B6 strain. To detect the primary event at the premeiotic stage, we conducted genome-wide expression profiling by microarray analyses of the testes of prepubertal B6-ChrXMSM and B6-ChrXTMSM males. Compared with the B6 strain, we found differential expression for 20% of MSM-derived X-linked genes that mostly show downregulation. Furthermore, the altered expression of X-linked genes subsequently evokes perturbation of genome-wide transcriptional regulation of autosomal genes. Notably, the differential expression in B6-ChrXTMSM is substantially restored in F1 progeny generated by crossing B6-ChrXTMSM females and B6-Chr1MSM males, in which chromosome 1 is substituted by the counterpart of the MSM strain in the genetic background of the B6 strain. This observation suggests that chromosome 1 contains upstream regulatory genes for X-linked target genes and the genetic incompatibility between trans-acting regulatory genes on chromosome 1 and cis-regulatory elements of the X-linked target genes is involved in the differential gene expression in B6-ChrXMSM and B6-ChrXTMSM testes. Intriguingly, we found that the differential expression also occurs in genes on chromosome 1 of F1 male progeny, implying that a similar phenomenon may generally occur in any donor chromosome of the inter-subspecific chromosome substitution strains. Our data suggest that the transcriptional regulatory system has diverged at the whole genome level during mouse evolution. Furthermore, the incompatibility of the diverged gene regulation between the two mouse subspecies results in reproductive isolation of X chromosome substitution strains.
In mice, spermatogenesis begins a few days after birth. Spermatogonia, which represent the mitotic stages of spermatogenic cells, occupy the basal compartment of seminiferous tubules, where they proliferate and differentiate to give rise to spermatocytes. Primary spermatocytes subsequently go through two rounds of meiotic division to form four haploid round spermatids. Finally, round spermatids transform into sperm. Primary spermatocytes are first detected at around 10 days postpartum (dpp), and the first meiotic prophase continues for 10–12 days [33].
Histological observation of the testes of B6-ChrXMSM and B6-ChrXTMSM strains showed no detectable impairment at 8 dpp, and a perceptible change was initially apparent at 10 dpp ([13] and Figure S1A). At 14 dpp, we noted a clear defect. Primary spermatocytes were abundant in the seminiferous tubules of control animals, but rarely observed in B6-ChrXMSM and B6-ChrXTMSM tubules ([13] and Figure S1A). To estimate the frequency of primary spermatocytes at prophase of meiosis I, testicular cells from 18-day-old males were immunostained with an antibody against SYCP3, a component of the synaptonemal complex and marker of early spermatocytes. The frequency of SYCP3-positive spermatocytes among all testicular cells was significantly low in B6-ChrXMSM and B6-ChrXTMSM males compared with that in B6 males, suggesting an impairment of meiotic entry in the two strains (B6, N = 4, 46.9±7.8%; B6-ChrXMSM, N = 4, 6.4±2.7%; B6-ChrXTMSM, N = 4, 18.0±4.0%; Bonferroni-corrected P<0.01 by two-tailed Student's t-test; [13] and Figure S1B). A small number of meiotic spermatocytes likely cause low production of sperm and a subsequent reduction of the testis weight. Consistent with this notion, adult B6-ChrXMSM and B6-ChrXTMSM males showed a significant reduction of testis weight compared with that of B6 males (weight of paired testes: B6, N = 8, 224.3±6.9 mg; B6-ChrXMSM, N = 3, 144.0±8.2 mg; B6-ChrXTMSM, N = 8, 138.1±11.1 mg; Bonferroni-corrected P<0.01 by two-tailed Student's t-test; [27] and Figure S1C).
Our previous composite interval mapping and newly performed interval mapping showed that the QTL responsible for the reduced testis weight was located at the distal region of the X chromosome with a significantly high LOD score ([27] and Figure S1D). The highest peak was detected at 147 Mb by the interval mapping (LOD score = 27.5). The degree of reduction in testis weight of B6-ChrXMSM males was similar to that of B6-ChrXTMSM males. Moreover, these two strains carry MSM alleles at the QTL, which affect the testis weight. Thus, the same biological basis may underlie the reduced testis weight and small number of meiotic spermatocytes in the two strains.
To detect the primary event that occurs in testicular cells prior to meiotic entry, we conducted whole-genome transcriptional profiling by microarray analysis of total RNA from whole testes of 5- and 7-day-old males. At these days of age, testicular histology of the two strains was indistinguishable from that of B6 ([13] and Figure S1A). The cellular composition of the testes at 5 dpp is relatively simple. Spermatogonia and Sertoli cells comprise 8.0% and 88.5% of the seminiferous epithelium, respectively [33]. At 7 dpp, a proportion of undifferentiated spermatogonia undergoes proliferation and gives rise to differentiated spermatogonia, resulting in a gradual increase of the proportion of spermatogonia [34], [35]. For transcriptional profiling, we used an Affymetrix Mouse Genome 430 2.0 array. The donor (MSM) chromosomes in the chromosome substitution strains were derived from M. m. molossinus in which the genome is significantly diverged from the laboratory mouse genome [36]. For this reason, it is not appropriate to use assignment of the presence or absence of the hybridization signal by comparison of the perfect match (PM) and mismatch (MM) probes. Therefore, we calculated the gene expression level with each probe set using the robust multichip average (RMA), as implemented in GeneSpring GX software, which considers only PM probes in its estimation of the expression level with each probe set [37]. Furthermore, single nucleotide polymorphisms in PM probes possibly cause mishybridization, which may lead to undercounting the expression signals of MSM-derived alleles. To avoid this occurrence, significantly polymorphic probe sets were excluded from the analysis. Similarities of all PM probe sequences of the Mouse Genome 430 2.0 array were searched against the MSM sequence reads (DRA000194) by Megablast [36]. To evaluate polymorphisms in the probe sets, we performed scoring using a classifier based on the number of identified probes in the MSM genome and the number of perfectly matched probes with the MSM sequence. The polymorphism score and number of probe sets for each polymorphism score are shown in Figure S2. Probe sets with polymorphism scores of ≤10 were judged as “conserved” probe sets, and “polymorphic” probe sets (a polymorphism score of >10) were excluded from the analysis. Polymorphism score for all probe sets is available on the FTP site of the NIG Mouse Genome Database (ftp://[email protected]/pub/msmdb/Affy_Probe_Info_MSM.zip). The polymorphic probe sets excluded from the present analysis were listed in Table S1.
We performed the transcriptional profiling using microarray expression data of B6-ChrXMSM testes at 5 dpp. Of the 1,376 probe sets for genes located in the genomic region from 5,341,800 to 163,344,914 bp on the X chromosome, 373 polymorphic probe sets were excluded from the analysis. Of the 43,678 probe sets for autosomal and non-polymorphic X chromosomal genes, 16,546 remained after filtering by the expression signal intensity (see Methods). To investigate the effect of the MSM-derived X chromosome on genome-wide gene expression, we plotted base-2 logarithms of the fold changes of expression levels for 16,546 transcripts in B6-ChrXMSM testes relative to those in the B6 strain. Results of the transcriptional profiling of the X chromosome and control chromosome 3, which has a physical size similar to that of the X chromosome, are shown in Figure 1A. Results of other chromosomes are shown in Figure S3. We next applied filtering to volcano plots that are constructed using fold change values and statistical significance by the Benjamini-Hochberg FDR corrected moderated t-test (fold change ≥1.50; P<0.05). Transcripts with significantly differential expression detected by filtered probe sets are indicated by red dots in Figure 1A and S3. The frequency of such transcripts in each chromosome is shown in Figure 1B. Both upregulated and downregulated transcripts relative to the B6 strain were notably enriched on the X chromosome. In B6-ChrXMSM testes, upregulated and downregulated transcripts from the X chromosome were detected by 3.4% (15/437) and 9.84% (43/437) of the probe sets, respectively. In contrast, upregulated and downregulated transcripts from autosomes were detected by 0.95% and 0.84% of the probe sets, respectively (Table 1 and Figure 1B). Interestingly, the direction of differential expression in B6-ChrXMSM testes was asymmetric as indicated by a disproportionately higher number of downregulated transcripts. This asymmetry was also detected for transcripts from the autosomes, although it was less obvious than that for transcripts from the X chromosome.
We examined whether fold change variance was different between X chromosomal and autosomal transcripts by the Ansari-Bradley test for centralized data, which is a commonly used rank-based test [38]. The results showed that the fold change variance of the X chromosomal transcripts was significantly larger than that of the autosomes (Ansari-Bradley test, P≤1.00×10−16). To test whether the different variances of the fold changes were independent of the expression level of the transcripts, we classified the probe sets into three groups by the raw signal intensity. The Ansari-Bradley test showed that the variance of the fold change for the X chromosome and autosomes was significantly different for both the low expression group (raw signal intensity between 10 and 100) and the intermediate expression group (raw signal intensity between 100 and 1000) (Table S2). We could not test for the high expression group (raw signal intensity of >1000), because such signal intensities were not observed for the X chromosome. For autosomes, we found 208 transcripts in the high expression group, among which more than half (125/208 transcripts) encode nuclear or mitochondrial ribosomal proteins. Because the haploinsufficiency of ribosomal protein genes immediately affects translational activity in many mutant mice, only a small number of ribosomal genes are located on the X chromosome, which is protected from X-inactivation in female somatic cells and male germ cells, such as MSCI and PMSR [39], [40].
We next performed transcriptional profiling of B6-ChrXMSM testes at 7 dpp. Of the 43,678 autosomal and non-polymorphic X chromosomal probe sets, 17,890 remained after filtering by the expression signal intensity. Similar to 5 dpp, differential expression of X chromosomal genes relative to the B6 strain was observed in B6-ChrXMSM testes at 7 dpp. Upregulated and downregulated transcripts were detected by 5.93% (28/472) and 14.83% (70/472) of the probe sets, respectively (Figure 1A, 1B and S4 and Table 1). In contrast, upregulated and downregulated transcripts from autosomes were detected by 1.19% and 3.83% of the probe sets, respectively. Differential expression detected by the X chromosomal probe sets was more significant than that detected by the autosomal probe sets, which was irrespective of the expression levels (Ansari-Bradley test, P≤1.00×10−16; Table S2). The differential expression of both X chromosomal and autosomal transcripts became more obvious at 7 dpp compared with that at 5 dpp (Ansari-Bradley test, autosomes, P≤1.00×10−16; X chromosome, P = 2.98×10−5). Differentially expressed transcripts in 5- and 7-day-old B6-ChrXMSM testes are listed in Tables S3 and S4. These data indicate that transcripts with highly differential expression are preferentially located on the X chromosome. Gene ontology (GO) analysis revealed that differentially expressed transcripts on autosomes and the X chromosome were drastically biased toward those involved in meiotic processes such as synapsis, synaptonemal complex, M-phase of the meiotic cell cycle, and meiotic chromosome organization as well as general male gamete production (Figure 1C).
Next, to examine whether MSM-derived cis-regulatory elements are responsible for the differential expression in B6-ChrXMSM, we compared the expression signal intensities of all X-linked genes expressed in B6-ChrXMSM testes with that in B6 or MSM testes. The Pearson correlation coefficient of B6-ChrXMSM to B6 and B6-ChrXMSM to MSM was 0.9031 and 0.9495, respectively (Figure 2A). We tested whether the correlation coefficient of B6-ChrXMSM to MSM is significantly larger than that of B6-ChrXMSM to B6 or not. We applied the logarithm transformation of the data to approximate the distribution of the data by the normal distribution and then obtained the approximate P-value by the test using the Fisher's z-transformation. The results showed that the correlation coefficient of B6-ChrXMSM to MSM was significantly larger than that of B6-ChrXMSM to B6 (P = 6.97×10−8). When the signal intensity of B6-ChrXMSM was compared with that of MSM, 93.5% (29/31) of upregulated transcripts (indicated in purple dots in Figure 2A) and 45.7% (32/70) of downregulated transcripts (indicated in blue dots in Figure 2A) were converged in a less than 1.5-fold change. This result suggests that these transcripts might reflect transcriptional regulation by MSM-derived cis-regulatory elements on the X chromosome.
The B6-ChrXTMSM strain has a MSM-derived genome between 86,497,454 and 165,344,914 bp of the X chromosome. Consequently, we could investigate how the subspecies origin of the genomic region is attributable to the differential expression of the X chromosome. We performed transcriptional profiling of B6-ChrXTMSM testes at 5 dpp. Of the 697 probe sets on the distal half of the X chromosome, 176 polymorphic probe sets were excluded from the analysis. Of the remaining 43,888 probe sets for autosomal and X chromosomal transcripts, 17,109 were used after filtering by the expression signal intensity. We plotted base-2 logarithms of the fold changes of transcript expression levels in B6-ChrXTMSM testes relative to those in the B6 strain. The variance of fold changes of the transcripts from genes located on the distal half of the X chromosome was significantly larger compared with that of other chromosomal regions including the proximal half of the X chromosome (Ansari-Bradley test, P≤1.00×10−16; Figures 3A and 5). The upregulated and downregulated transcripts were detected by 2.51% (6/239) and 9.62% (23/239) of the probe sets, respectively, for the distal half of the X chromosome, whereas differentially expressed transcripts for autosomes and the proximal half of the X chromosome were detected at a apparently lower frequency (Figure 3B and Table 2).
We next performed microarray analysis of testes from a congenic strain in which only a 37.8 Mb interval between 125,512,711 and 163,344,520 bp on a more distal region of the X chromosome was substituted by the MSM-derived genome in the B6 background. We observed similar differential expression only in the MSM-derived genome, as was observed in B6-ChrXMSM and B6-ChrXTMSM strains (Figure S6). This result indicated that the differential expression is strictly restricted to the genomic region derived from the MSM strain.
When we performed transcriptional profiling of B6-ChrXTMSM testes at 7 dpp, 18,449 probe sets remained after filtering by the expression signal intensity. Consistent with B6-ChrXMSM testes, we found genome-wide differential expression in B6-ChrXTMSM testes at 7 dpp was more obvious compared with that at 5 dpp (Ansari-Bradley test, autosomes and the proximal half of the X chromosome, P≤1.00×10−16, the distal half of the X chromosome, P = 2.61×10−4; Figures 3A and S7). The frequency of the differentially expressed transcripts detected by the probe sets for each chromosome is shown in Figure 3B and Table 2. Such transcripts in 5- and 7-day-old B6-ChrXTMSM testes are listed in Tables S5 and S6, respectively. At 7 dpp, differentially expressed transcripts in B6-ChrXTMSM testes largely overlapped with those in B6-ChrXMSM testes (70.6%, 686/971) (Figure 3C). They consisted of not only transcripts from the distal half of the X chromosome (7.0%), but also those from autosomal regions (91.0%) (Figure 3D and Table S7). This finding implies that differential expression in the distal half of the X chromosome subsequently affects genome-wide transcription at 7 dpp.
We also compared the signal intensity of all transcripts from the distal-half of the X chromosome in B6-ChrXTMSM testes with that in B6 or MSM strains. The Pearson correlation coefficient of B6-ChrXTMSM to B6 was 0.9124, whereas that of B6-ChrXTMSM to MSM was 0.9589 (Figure 2B). The correlation coefficient of B6-ChrXTMSM to MSM was significantly larger than that of B6-ChrXTMSM to B6 by the test using the Fisher's z-transformation (P = 4.02×10−6). When the expression in B6-ChrXTMSM was compared with that in MSM, 77.8% (14/18) and 40.5% (17/42) of upregulated (purple dots in Figure 2B) and downregulated transcripts (blue dots in Figure 2B), respectively, were converged in a less than 1.5-fold change.
To assign a chromosome that interacts with X chromosomal genes and is responsible for the genetic incompatibility in B6-ChrXTMSM males, we produced F1 male progeny generated from crosses of B6-ChrXTMSM females with males of other chromosomal substitution strains, and investigated their reproductive phenotypes. We generated crosses with a total of 22 chromosome substitution strains of which 13 crosses showed significant restoration of testis weight (Bonferroni-corrected P<0.05 by the two-tailed Student's t-test; Figure S8). Among F1 male progeny from crosses with chromosomal substitution strains, (B6-ChrXTMSM×B6-Chr1MSM)F1 males (hereafter abbreviated as B6-Chr1MSM/B6XTMSM) exhibited the most effective restoration of testis weight (Figure S8). To determine whether the restoration of testis weight is caused by proper spermatogenesis or other reasons including hyperplasia of Leydig cells, we investigated the phenotype of B6-Chr1MSM/B6XTMSM testes. The testicular histology showed that meiotic spermatocytes observed in B6-Chr1MSM/B6XTMSM testes were not abundant as observed in B6-ChrXTMSM testes at 14 dpp, but spermatocytes were increased in B6-Chr1MSM/B6XTMSM testes at 18 dpp (Figure 4A). This result indicated that the progression of spermatogenesis in B6-Chr1MSM/B6XTMSM males is delayed at the early stage, but becomes similar to that of the B6 strain at a later stage. Immunostaining analysis showed that the frequency of spermatocytes was restored significantly in B6-Chr1MSM/B6XTMSM testes at 18 dpp (Figure 4B).
To examine the timing of spermatogenesis restoration in B6-Chr1MSM/B6XTMSM testes, we performed transcriptional profiling by microarray analysis of testes from B6-Chr1MSM/B6XTMSM males at 5 and 7 dpp. Because the B6-Chr1MSM/B6XTMSM strain was heterozygous for B6 and MSM alleles on chromosome 1, polymorphic probe sets on chromosome 1 were excluded from this microarray analysis. Of the remaining 42,781 probe sets, 15,377 probe sets were used after filtering by the expression signal intensity using 5 dpp samples. Fold changes of expression signals in B6-Chr1MSM/B6XTMSM testes relative to that in the B6 strain revealed that the genes on the distal half of the X chromosome were expressed differentially as was observed in B6-ChrXTMSM testes (Figures 5A and 5B, Table 3, and Figure S9). The differentially expressed transcripts in 5 dpp B6-Chr1MSM/B6XTMSM testes are listed in Table S8.
When we performed the transcriptional profiling with 7 dpp B6-Chr1MSM/B6XTMSM testes, we found that the expression patterns were noticeably variable among the individuals. We conducted principal component analysis (PCA), as implemented in GeneSpring GX software, using expression data for all genes of three individuals each of B6 and B6-ChrXTMSM, and eight individuals of B6-Chr1MSM/B6XTMSM. The results showed that the coordinate values were clustered for B6 and B6-ChrXTMSM individuals (Figure 6A). In contrast, the coordinate values of the eight B6-Chr1MSM/B6XTMSM individuals were largely scattered between the clusters of B6 and B6-ChrXTMSM. Two B6-Chr1MSM/B6XTMSM individuals (a and b in Figure 6A) were classified into the same cluster of B6, suggesting that genome-wide transcriptional regulation was restored in these individuals. The other two B6-Chr1MSM/B6XTMSM individuals (c and d in Figure 6A) were rather close to the cluster of B6-ChrXTMSM, indicating that they remained in the original state. Another four B6-Chr1MSM/B6XTMSM individuals belonged to neither B6 nor B6-ChrXTMSM clusters.
To characterize the restoration of B6-Chr1MSM/B6XTMSM testes in more detail, the B6-Chr1MSM/B6XTMSM individuals were classified as “restored” or “non-restored” types. Individuals a, b, and e belonged to the restored type and were relatively close to the B6 cluster on the x-axis of principal component 1. The remaining individuals belonged to the non-restored type. Of the 42,781 probe sets, 18,669 remained for analysis after filtering by the expression signal intensity. Notably, the high degree of genome-wide differential expression that occurred in B6-ChrXMSM and B6-ChrXTMSM testes at 7 dpp was not observed in the restored type of B6-Chr1MSM/B6XTMSM individuals (Figure 5A and 5B, Table 3, and Figure S10). The Ansari-Bradley test showed that the variance of the fold changes in restored B6-Chr1MSM/B6XTMSM individuals was significantly smaller than that in B6-ChrXTMSM individuals, which was irrespective of the chromosomal regions (autosomes and the proximal half of the X chromosome, P≤1.00×10−16; the distal half of the X chromosome, P = 5.25×10−5). This result suggested that the gene expression pattern of the restored B6-Chr1MSM/B6XTMSM individuals was recovered to some extent from the differential expression toward the B6 type. When we examined the expression pattern of the top 50 differentially expressed genes in B6-ChrXTMSM testes at 7 dpp, we found that the expression levels of these genes were considerably restored in B6-Chr1MSM/B6XTMSM testes (Figure 6B). In contrast, when the fold changes of expression for transcripts in non-restored B6-Chr1MSM/B6XTMSM males relative to B6 males were plotted, the variance of the transcripts remained to be similar to B6-ChrXTMSM males (Figure S11). Although B6-Chr1MSM/B6XTMSM individuals showed variable expression profiles, all tested adult B6-Chr1MSM/B6XTMSM individuals (N = 14) indicated the restoration of testis weight (Figure S8). Thus, theses observations suggested that the developmental stage at 7 dpp might be a transient period of change from the differential expression profile of B6-Chr1MSM/B6XTMSM testes to the B6 type.
Although the general expression profile was re-established in the restored type of B6-Chr1MSM/B6XTMSM individuals, transcripts from the distal half of the X chromosome appeared to remain in a differentially expressed state (Figure 5A and 5B, and Table S9). Next, we compared the signal intensity of transcripts of B6-Chr1MSM/B6XTMSM on the distal half of the X chromosome with that of B6 or MSM. The Pearson correlation coefficient of B6-Chr1MSM/B6XTMSM to B6 was 0.9400, and that of B6-Chr1MSM/B6XTMSM to MSM was 0.9838 (Figure 6C). The correlation coefficient of B6-Chr1MSM/B6XTMSM to MSM was significantly larger than that of B6-Chr1MSM/B6XTMSM to B6 by the test using the Fisher's z-transformation (P = 1.27×10−14). This result suggests that genetic incompatibilities in the transcriptional regulation of differentially expressed X-linked genes are resolved to some extent by MSM alleles of upstream regulatory genes on chromosome 1.
Interestingly, we noticed a high frequency of differentially expressed genes on chromosome 1 as compared to those on other autosomes in B6-Chr1MSM/B6XTMSM individuals at both 5 and 7 dpp, although they carried MSM alleles heterozygously throughout chromosome 1 (Figures 5A and 5B). The Ansari-Bradley test showed that the variance of fold changes of transcripts from chromosome 1 was significantly larger than that of other chromosomal regions except the distal X chromosome (5 dpp, P≤1.00×10−16; 7 dpp, P≤1.00×10−16). This finding suggests that genes in the MSM-derived genome on chromosome 1 are also differentially expressed in B6-Chr1MSM/B6XTMSM testes.
To re-examine the expression pattern of X-linked genes in B6-ChrXTMSM and B6-Chr1MSM/B6XTMSM testes, the expression levels of selected genes were measured by real-time quantitative RT-PCR. The expression levels of eight genes that locate near by the QTL responsible for reduced testis weight were significantly low in the B6-ChrXTMSM testes than in the B6 testes, which was consistent with the microarray data (Figure S12). The expression levels of these genes were restored in B6-Chr1MSM/B6XTMSM testes, and their expression levels tended to shift toward those of MSM testes rather than B6 testes. The eight genes included three candidate genes responsible for genetic incompatibility: TAF7-like RNA polymerase II, TATA box binding protein (TBP)-associated factor (Taf7l), and two nuclear mRNA export factors (Nxf2 and Nxf3).
To identify the genes responsible for the restoration of differential expression in B6-Chr1MSM/B6XTMSM testes, we performed QTL analysis. We produced F1 male progeny from crosses between B6-Chr1MSM females and wild-type B6 males. Then, the (B6-Chr1MSM×B6)F1 males were crossed with B6-ChrXTMSM females to obtain 314 male progeny for the QTL analysis (Figure S13A). Because the male progeny had recombination at various sites in chromosome 1 between B6 and MSM, we could map the QTLs affecting testis weight. As a result, we found continuously high LOD scores over the large region of chromosome 1 by single marker analysis and interval mapping (Figure S13B and S13C). The highest peak was observed at 64.5 Mb by the interval mapping (LOD score = 21.8). This indicates that at least one QTL responsible for the genetic incompatibility are located possibly at the region of 40–80 Mb on chromosome 1.
We clearly show that approximately 20% of genes located in MSM-derived X chromosomal regions are differentially expressed in the genetic background of the B6 strain. The most important finding in this study is that such differential expression in prepubertal testes occurs prior to the reproductive phenotype of the X chromosome substitution strains. Furthermore, perturbation of genome-wide gene expression possibly caused by the differential X-linked gene expression might enhance the reproductive phenotype. Thus, the ultimate phenotype of the reproductive isolation is the sum of deleterious effects by the differential expression of many genes.
Previous studies showed extensive overexpression of X-chromosomal genes in testes of sterile F1 males generated from cross of M. m. musculus females and M. m. domesticus males due to a disruption of MSCI [41], [42]. By contrast, our study showed that the differential expression of MSM-derived genes in the X chromosome substitution strains occurs at premeiotic stage of spermatogenesis, and the differential expression is bidirectional, upregulation and downregulation. Moreover, it was observed not only for the X chromosomal genes but also for genes on chromosome 1. Thus, mechanisms underlying the overexpression of X-chromosomal genes in the F1 males might be different from that in the X chromosome substitution strains.
The cell population in seminiferous tubules of the prepubertal testis consists of undifferentiated spermatogonia called spermatogonial stem cells (SSCs), differentiated spermatogonia, and their supporting Sertoli cells. To identify which cell type showed differential expression in the B6-ChrXMSM (or B6-ChrXTMSM) testis, we referred to a previous study by Yang et al. [35]. In their study, microarray analyses were carried out using OCT4-positive SSCs and OCT4-negative cells, including differentiated spermatogonia and somatic cells, which were isolated from 7 dpp testes. OCT4 is an important transcription factor involved in maintenance of stem cell pluripotency. We found that the differentially expressed transcripts in the X chromosome substitution strains were present in the gene lists enriched in both OCT4-positive and OCT4-negative cells (Figure S14), suggesting that the differential gene expression occurs in both SSCs and non-SSC cells. Interestingly, all differentially expressed transcripts that were enriched in SSCs were downregulated, whereas those enriched in non-SSC cells showed preferential upregulation (Figure S14). The implication of the observed skew in differential expression is unknown at present. We could not exclude the possibility of delayed spermatogonial development, though the testes appeared to be similar histology in B6 and the X chromosome substitution strains, which might influence the transcription profiles. To refer to the generality of misregulation in each type of testicular cells, further analysis using isolated cell population will be required.
For transcripts from X-linked genes in B6-ChrXMSM and B6-ChrXTMSM, approximately 77–79% transcripts were not differentially expressed relative to the B6 strain, implying that the transcriptional regulation of these genes is conserved between MSM and B6 strains (Figure 7A and 7B). Approximately 15–16% of the transcripts were downregulated, while the remaining 6–7% showed upregulation. By comparing the expression signal intensities of differentially expressed genes in X chromosome substitution strains with those of the same genes in the MSM strain, we found that these genes could be classified into two types. The first type showed expression levels comparable to those in the MSM strain, whereas the second type showed expression levels that were different from those in the MSM strain (Figure 7A and 7B and Table S10 and S11). The first type likely reflects transcriptional regulation by MSM-derived cis-regulatory elements on the X chromosome. The second type might reflect the transcriptional misregulation possibly caused by the incompatibility between B6-derived trans-acting regulators and MSM-derived cis-regulatory elements for X-linked target genes. Notably, the proportions of these two types were significantly different between upregulated and downregulated transcripts. In B6-ChrXMSM (or B6-ChrXTMSM), 80–90% of upregulated transcripts belonged to the first type (Figure 7A and 7B). In contrast, more than 50% of downregulated transcripts belonged to the second type. It is plausible that the genetic incompatibility of the transcriptional regulation readily causes a decrease of gene expression rather than an increase.
In this study, we found that a number of genes on the MSM-derived X chromosome and chromosome 1 were differentially expressed in the genetic background of the B6 strain. This observation suggests that genome-wide transcriptional (or post-transcriptional) regulation has evolutionarily diverged during mouse subspeciation, and that the diverged expression is not specific to the X chromosome, but rather a general phenomenon observed for genes in substituted chromosomes. Despite the diverged expression of genes on chromosome 1, the parental B6-Chr1MSM strain does not exhibit any reproductive phenotypes [43], and GO analysis showed that differentially expressed genes were not preferentially categorized into reproductive GO classes (data not shown). Male reproductive genes preferentially accumulate on the X chromosome, because beneficial X-linked genes are readily selected because of the hemizygosity in males [8], [44], [45]. Moreover, genes expressed during premeiotic stages are not subjected to sex chromosome inactivation. All these conditions accelerate accumulation of male reproduction-related genes on the X chromosome, which are expressed at the early stage of spermatogenesis. Consequently, substitution of an X chromosome from different subspecies might have disproportionally more significant effects on premeiotic spermatogenesis than that of autosomes.
Although a set of differentially expressed genes on the X chromosome and autosomes is responsible for the reproductive phenotype of X chromosome substitution strains, some genes may exert predominant effects on spermatogenic and cellular processes. Our QTL analysis of the reduced testis weight detected high LOD score on the distal one-third of the X chromosome. This region contains multiple differentially expressed genes that are known to be involved in spermatogenesis and housekeeping functions. We then explored X-linked genes that are differentially expressed in the testes of X chromosome substitution strains relative to that in B6 testes, and are thought to predominantly affect the testis weight. Based on the functional gene annotations, we focused on three genes in the relevant region. First, Taf7l is reduced in abundance of the mRNA in the B6-ChrXMSM and B6-ChrXTMSM testes to approximately one-third of that in B6 testes. Taf7l encodes a male germ cell-specific paralogue of the transcription factor IID (TFIID) subunit TAF7. TFIID is a highly conserved general transcription factor that is required for transcription of protein-coding genes by RNA polymerase II. Taf7l knockout male mice exhibit age-dependent spermatogenic defects, a decreased testis weight, and significantly low production of sperm [46]. Biochemical studies have demonstrated that TAF7L is closely associated with the TFIID subunit TBP in meiotic and postmeiotic male germ cells [47]. Thus, a reduction of Taf7l mRNA may affect the transcriptional activity of genes related to meiotic and postmeiotic processes in germ cells. Two other genes were nuclear mRNA export factors, Nxf2 and Nxf3, which are specifically expressed in germ cells and Sertoli cells in the testis, respectively [48], [49]. Our data showed that the gene expression level of Nxf2 and Nxf3 in B6-ChrXMSM and B6-ChrXTMSM testes was decreased to one-third and one-tenth of that in B6 testes, respectively. Nxf2-deficient male mice exhibit meiotic arrest and are sterile [48]. NXF family members are known to play roles in not only nuclear mRNA export but also various aspects of post-transcriptional mRNA metabolism [50]. It is notable that the above three genes, which were significantly downregulated in B6-ChrXMSM and B6-ChrXTMSM testes and restored the expression in B6-Chr1MSM/B6XTMSM testes, were categorized as the second type in terms of expression levels relative to the MSM strain. Coevolution between cis-regulatory elements and trans-acting factors is more frequently observed in sex-specific genes than in other genes [51], which might accelerate the transcriptional divergence of these genes.
The differential gene expression was observed genome-widely at 7 dpp in the testes of B6-ChrXMSM and B6-ChrXTMSM strains. At this stage, 70% of the differentially expressed genes were common to B6-ChrXMSM and B6-ChrXTMSM testes, and 90% of the common genes were located on autosomes. These data suggest that the significantly reduced expression of these predominant genes located on the distal half of the X chromosome affect global gene expression at the later stages (7 dpp onward) in B6-ChrXMSM and B6-ChrXTMSM testes. GO analysis revealed that many of the genome-wide differentially expressed genes are involved in meiosis, including stimulated by retinoic acid gene 8 (Stra8) on chromosome 6 and cellular retinoic acid binding protein I (Crabp1) on chromosome 9. The abundance of these transcripts was significantly reduced in B6-ChrXMSM and B6-ChrXTMSM testes. Retinoic acid (RA) is known to be essential for germ cells to enter meiosis in both the ovary and testis. RA stimulates Stra8 expression to induce expression of downstream genes required for transition into meiosis [52], [53]. CRABP1 is a cellular RA-binding protein that is expressed in spermatogonia [54]. The reduced expression of these master regulatory genes required for meiotic entry is a likely cause of the significant reduction of meiotic spermatocytes in prepubertal B6-ChrXMSM and B6-ChrXTMSM testes. We also found other genes downregulated in the X chromosomal substitution strains, which are known to be involved in meiosis. These genes included three synaptonemal complex proteins genes (Sycp1, Sycp2, and Sycp3), DMC1 dosage suppressor of mck1 homolog (Dmc1) that functions in meiosis-specific homologous recombination, testis-expressed genes (Tex11 and Tex15), maelstrom homolog (Mael), and structural maintenance of chromosomes 1B (Smc1b). The reduced expression of these genes might explain the defect of synapsis observed in early spermatocytes of B6-ChrXMSM and B6-ChrXTMSM testes [13]. Thus, the misregulated gene expression in premeiotic spermatogonia may influence the cellular processes in germ cells at meiotic and possibly postmeiotic stages.
In B6-Chr1MSM/B6XTMSM testes, the introduction of MSM-derived chromosome 1 restored the reproductive phenotypes and transcriptional misregulation. At 5 dpp, the misregulation was not restored in B6-Chr1MSM/B6XTMSM testes, and the delayed restoration began at 7 dpp, which is reminiscent of the fact that emergence of meiotic spermatocytes was delayed in B6-Chr1MSM/B6XTMSM testes. However not all X-linked transcripts were restored in B6-Chr1MSM/B6XTMSM testes. These findings imply that genetic factor(s) on chromosome 1 are not sufficient for complete restoration of B6-ChrXTMSM phenotypes. Consistent with this notion, the reduced testis weight in B6-ChrXTMSM strain was restored to some extent by the introduction of other MSM-derived autosomes. Thus, an intricate genetic mechanism is involved in the reproductive isolation between mouse subspecies.
In summary, our study provides comprehensive characterization of the transcriptional profile in X chromosome substitution strains, demonstrating that transcriptional regulation divergence between the two mouse subspecies contributes to the improper gene expression in the genetic background of different subspecies. Evolutionary divergence in transcriptional regulation explains the phenotypic differences between subspecies and occasionally causes incompatibilities attributed to diseases and reproductive disorders that contributes to reproductive isolation. Our study has revealed an insight into gene expression divergence in mammals and the occurrence of speciation.
The MSM strain was established and maintained at the National Institute of Genetics (NIG), Mishima, Japan. B6 mice were purchased from CLEA Japan (Tokyo, Japan), and maintained at NIG. The full set of chromosome substitution strains was established using MSM as the chromosome donor and B6 as the host (background) strain [43], and were maintained at NIG. Almost the entire length (5.3–166.3 Mb) and distal region (86.5–163.3 Mb) of the X chromosome were substituted by the counterparts of the MSM strain in B6-ChrXMSM and B6-XTMSM strains, respectively (Figure S15) [27], [43]. In the B6-Chr1MSM strain, almost the entire length (3.2–193.4 Mb) of chromosome 1 was substituted by the counterpart of the MSM strain (Figure S15) [43]. All animal experiments were approved by the Animal Care and Use Committee of NIG.
After euthanasia of prepubertal males, their testes were dissected, immersed in RNAlater (Ambion, Austin, TX), and stored at −20°C. Total RNA from the whole testis was extracted using QIAzol Lysis Reagent and an RNeasy Mini kit (Qiagen. Valencia, CA). DNase digestion of the purified RNA with RNase-free DNase (Qiagen) was performed according to the manufacturer's protocol. RNA quality was checked with an Agilent Bioanalyzer 2100 (Agilent Technologies, Santa Clara, CA), and only RNA samples with >9 RNA integrity number were used in experiments. Biotinylated amplified RNA was generated from 100 ng total RNA using an Affymetrix GeneChip 3′IVT Express Kit, and then hybridized to an Affymetrix Mouse Genome 430 2.0 array (Affymetrix, Santa Clara, CA) following the manufacturer's instructions. For each strain or genotype, three distinct mice were tested separately. For the B6-Chr1MSM/B6XTMSM strain, a total of eight mice were tested, because they showed variable expression profiles.
Microarray data were analyzed with GeneSpring GX software (Agilent Technologies). The expression data were normalized using the RMA method. Because genotypes at boundary regions of each chromosome substitution strain are not defined precisely, probe sets in these regions were excluded from the expression analysis (Figure S15). We also excluded probe sets with an unknown chromosome location. Chromosomal locations for the mouse reference genome (mm9) were obtained from the Ensembl genome browser (http://asia.ensembl.org/index.html). Y-linked probe sets were omitted because their number was extremely small. As mentioned in the Results, polymorphic probe sets were filtered out. Only probe sets with a raw signal intensity of >10 in at least one of the paired strains were used for the analysis. GO analysis and calculation of P-value were performed by GeneSpring GX software. The Ansari-Bradley test and the Fisher's z-transformation were conducted by the R package. The expression data was deposited in the NCBI Gene Expression Omnibus (GSE50687).
The testes of 3- to 4-month-old euthanized males were dissected and weighed. For histological analysis, the testes were placed in fresh Bouin's fixative at room temperature. Excess fixative was removed with 70% ethanol. The tissues were then dehydrated and embedded in paraffin for microtome sectioning. The sections (6 µm) were stained with hematoxylin and eosin. To count meiotic spermatocytes, we used 18-day-old males. Immunocytochemistry was performed as described previously [13] using an antibody against SYCP3 (Novus Biologicals, Littleton, CO). After immunocytochemistry, the slides were counterstained with Hoechst 33258. The proportion of meiotic spermatocytes was measured by the frequency of SYCP3-positive spermatocytes among all testicular cells.
A total of 314 male progeny from crosses between B6-ChrXTMSM females and (B6-Chr1MSM×B6) F1 males were used to map quantitative trait loci for testis weight and relative testis weight per body weight (RTW). All males were genotyped at 80 single nucleotide polymorphism (SNP) markers on chromosome 1 by using the Sequenom MassARRAY iPLEX Gold Assay (Sequenom) as previously described [36]. Genotype and trait data from all male progeny were indicated in Table S12. Single marker analysis and interval mapping (walk speed: 1 Mb) were performed by using QTL Cartographer (http://statgen.ncsu.edu/qtlcart/).
|
10.1371/journal.pcbi.1000750 | Structural Characteristics of Novel Protein Folds | Folds are the basic building blocks of protein structures. Understanding the emergence of novel protein folds is an important step towards understanding the rules governing the evolution of protein structure and function and for developing tools for protein structure modeling and design. We explored the frequency of occurrences of an exhaustively classified library of supersecondary structural elements (Smotifs), in protein structures, in order to identify features that would define a fold as novel compared to previously known structures. We found that a surprisingly small set of Smotifs is sufficient to describe all known folds. Furthermore, novel folds do not require novel Smotifs, but rather are a new combination of existing ones. Novel folds can be typified by the inclusion of a relatively higher number of rarely occurring Smotifs in their structures and, to a lesser extent, by a novel topological combination of commonly occurring Smotifs. When investigating the structural features of Smotifs, we found that the top 10% of most frequent ones have a higher fraction of internal contacts, while some of the most rare motifs are larger, and contain a longer loop region.
| Structural genomics efforts aim at exploring the repertoire of three-dimensional structures of protein molecules. While genome scale sequencing projects have already provided us with all the genes of many organisms, it is the three dimensional shape of gene encoded proteins that defines all the interactions among these components. Understanding the versatility and, ultimately, the role of all possible molecular shapes in the cell is a necessary step toward understanding how organisms function. In this work we explored the rules that identify certain shapes as novel compared to all already known structures. The findings of this work provide possible insights into the rules that can be used in future works to identify or design new molecular shapes or to relate folds with each other in a quantitative manner.
| Under physiological conditions most proteins self-assemble into unique structures that dictate their interactions with other molecules and determine their function. Protein structures can be decomposed into individually folding units, so called folds [1]. A fold is determined from the number, arrangement, and connectivity (topology) of secondary structure elements [2]. Manually curated [3], semi–automated [4] and automated approaches [5], [6] classify protein folds by organizing them into hierarchical systems. Due to the lack of a clear understanding of how to define and classify folds, these various subjective approaches carry substantial inconsistencies [2], [7]. Meanwhile, recent studies paint a more nuanced picture of the fold universe of proteins, one that is more continuous in nature, where some higher density hubs formed by related structures correspond to and connect known folds [8], [9], [10], [11]. Part of the motivation to rethink the nature of the protein fold universe is provided by the apparent success of molecular modeling efforts that use short amino acid segments from known protein structures to build up novel folds [12]. Additional motivation comes from anecdotal examples that identify structures representing transitions between previously described folds, which either results in a unification of different fold families or suggests removing fold definitions altogether [13], [14]. One such example is described for the RIFT domain, where it is suggested that starting from an ancestral RIFT domain a strand invasion and a strand–swap event (with subsequent duplication and fusion events) resulted in the emergence of the swapped hairpin and double-psi beta barrel folds, respectively [15]. These folds cannot be interconverted with simple topological modifications, such as circular permutation, although their common evolutionary origin has been established.
Since the definition of complete folds is ambiguous, one has to consider structural definitions of smaller (local) entities, such as supersecondary structure elements, that could describe protein folds and the structure universe in a more quantitative and systematic nature. Supersecondary structure elements are defined as a number of regular secondary structure elements that are linked by loops (e.g. Rossmann, helix-turn-helix, four strand Greek key, β-meander motifs etc.). Folds are formed by the overlapping combination of various supersecondary elements, which are shared among different proteins and sometimes highly repeated within the same one. This observation prompted the theory of a relic peptide world [16], which proposes that modern, stable proteins are the results of duplication, mutation, shuffling and fusion of a limited set of relic peptides. Various efforts have tried to explore possible tool sets of supersecondary elements, such as antiparallel ββ-sheets [17], αββ and ββα motifs [18], αα-turn motifs [19], four helix bundles [20] and so on. Building on these earlier efforts, we introduced a new, general, supersecondary structure classification that fully describes all known protein structures [21]. In this schema a basic supersecondary motif, which we will refer to as Smotif, is composed of two regular secondary structure elements linked by a loop. Smotifs are characterized in protein structures by the types of sequential secondary structures and the geometry of the orientation of the secondary structures with respect to each other, as described by four internal coordinates [21], [22]. The definition for supersecondary structure elements for Smotifs is different from other studies or from the above mentioned textbook examples and it is rooted in practical reasoning. In this study we explored Smotifs of only two connected secondary structures because for this subset we had indication from prior work that the number of possible combinations are limited. Also, if we used a definition that has higher number of connected secondary structures e.g. 3 or more, the number of combinations would be very large and would prevent us from a systematic classification. Recently, we demonstrated that Smotifs with loop fragments having lengths up to 12 residues, together with their bracing secondary structure elements are exhaustively sampled in the Protein Data Bank (PDB). We also demonstrated that the available set of Smotifs has been essentially unchanged at least for the last 5 years, despite that during this time the sequence databases have doubled and a significant number of new folds have emerged [23].
These previous observations motivated us to analyze the occurrence of Smotifs among protein folds and explore the question of what is really unique about a structure that is identified as “novel”. Does the emergence of a novel fold coincide with the emergence of novel Smotifs that are integrated into a structure with known ones? Is it possible to generate novel folds solely from existing Smotifs? What are the rules that guide combinations of Smotifs to an apparently novel fold? Is the novelty of a certain Smotif or the novelty of combining well-known Smotifs the driving force behind the appearance of novel folds? These questions might be relevant to shed light on the rules governing protein structure evolution. There are practical considerations to understanding the actual limits of the definition and novelty of a fold. Exploring these issues can aid in developing more accurate structure modeling tools and support the design and realization of new and experimentally accessible molecular shapes.
We explored the frequency of occurrences of all Smotifs in all protein folds. We established an exhaustive library of 324 types of Smotifs, as classified by their geometry, for each of the four combinations of possible bracing secondary structure elements. We have shown that this geometrical classification of Smotifs correctly captures local structural similarity (see Definition of optimal classification of Smotif geometry in Material and Methods). Previously we have shown that Smotifs are useful for loop prediction because loop conformations (as defined by the orientation of the embracing secondary structures) up to 10–12 residues are exhaustively sampled in PDB [21], [23]. We further refined this observation by exploring the increase of coverage of Smotifs in PDB over time (Fig. 1). Approximately 10 years ago all categories of Smotifs were already represented by at least one example.
The occurrence of Smotif geometries in different types of protein folds is uneven (Fig. 2). There are some Smotifs whose geometries are ubiquitous, and occur in many different folds, while others are specific to a few. Fig. 2 displays a ββ class Smotif (a) that is highly represented across different folds, corresponding to a geometry that tightly aligns two ββ-strands and, thus, allows many non-bonded contacts to be formed. Meanwhile another Smotif within the ββ class (b), which is structurally similar but where one of the β-strands is tilted, has a very low occurrence within known folds. Similar trends can be observed for αα, αβ, and βα Smotifs: Smotifs forming extensive non-bonded interactions occur more frequently in known folds. We explored the normalized number of intra-motif non-bonded contacts as a function of Smotif frequency and found an exponential correlation between the number of contacts and frequency of motif usage (correlation of r = 0.83 as fitted on a logarithmic scale), indicating that the most frequent motifs (top 10%) are forming more contacts. However, there is not a statistically significant correlation for the rest of the Smotif frequencies (Fig. 3).
Another suspected factor for Smotif preferences is their size, as large Smotifs simply cannot fit into smaller folds. Here we found no clear tendency except once again the top 10% most frequent Smotifs, which indeed tend to be smaller (on average 12 (σ = 6) residues total within the bracing secondary structures, without counting the variable number of loop residues, while motifs at all other frequencies are generally formed by 16 residues (σ = 8)). The longer the loop connecting the bracing secondary structures, the more likely that contacts will be formed between non-proximal secondary structures: e.g. a ββ-type Smotif that connects together strands of two β-sheets. A correlation was found between the length of the loop within Smotifs and the frequency of Smotif usage in folds among the 50% least frequent Smotifs. However, Smotifs extracted from new folds do not show correlation between Smotifs size or loops length and the frequency of Smotifs: i.e. new folds are not necessarily formed by large Smotifs and do not necessarily have particularly long loops (data not shown).
We also explored whether solvent accessibility is correlated with the frequency of Smotifs, as one could suspect that buried, conserved cores would be formed by frequently occurring Smotifs and structural regions outside the common core would have a trend to comprise a higher proportion of rare Smotifs, due to a less restrictive structural environment. However, we could not find any statistically significant correlation between the frequency of Smotifs and their exposure (Fig. S1).
Since the repertoire of Smotifs seems to have come close to saturation (Fig. 1) [23], this prompts the question of what is really unique about a fold structure when it is identified as “novel”. Detecting novel folds is a non-trivial question. Automated structural comparisons are often followed by manual inspection to characterize new protein structures. We have explored proteins that were classified as novel at the time of their discovery in two expert validated sources, in the archives of SCOP [3] and in the series of CASP experiments [24]. We found that proteins that were considered novel folds at CASP 3–6 meetings (years 1998–2004) and in SCOP 1.73, 1.75 (years 2007–2009) do not have any novel Smotif geometries that were not present in previously solved structures. In other words, none of the Smotifs of novel folds have a unique geometry (Table 1). For instance, as early as the third round of CASP Meetings in 1998 [25], all of the targets identified as novel folds by the experts could have been reconstructed using Smotifs from known protein structures. If, in our Smotif comparison, we required not only a match in the geometry between the Smotifs in the novel structures and those in the solved structures, but also required identical lengths of the flanking secondary structures, still less than 6% of the Smotifs in novel folds at CASP meetings would not have a match in already known structures. Similarly, we have checked the motif composition of new folds from the archives of SCOP in the 1.73 (2007 November) and 1.75 (2009 June) releases. These contain a total of 233 new folds from 1140 proteins. Similar to the CASP targets, none of these novel folds had a Smotif that was not already observed in a previously known fold. With the stricter definition, that requires a fit of the length of the bracing secondary structures, still less than 1% proved to be novel Smotifs. Initially, we found 47 Smotifs (out of the 8056 analyzed) that appeared to be new. However, after manual inspection, it turned out that these are all explained by an artifact of replacing obsolete PDB entries with newer ones, with a corresponding newer date. The above observations suggest that recently solved novel folds do not imply the emergence of new Smotifs, and that a protein with a novel fold can be constructed using Smotifs from already existing protein folds. As an illustration, T0181 (PDB code: 1nyn), a new fold submitted to CASP5, can be constructed from 7 overlapping Smotifs, all of which can be located in previously solved structures of other proteins representing a variety of different folds (Fig. 4).
When we explored the frequency of occurrence of Smotifs in the non-redundant set of known folds, we observed that novel folds have a larger fraction of Smotifs that have a low frequency of occurrence in the PDB (Fig. 5 CASP dataset; see Fig. S3 and S4 for distribution of Smotif frequency calculated for SCOP 1.75 and SCOP 1.73 respectively). On the other hand, superfolds [26], those that are adopted by many different sequences often with different functions, are built by Smotifs that occur with medium or high frequencies in existing folds. This implies that novel folds are composed of a new permutation of existing Smotifs and, specifically, a structure will have a greater likelihood of being “novel” if the structure is enriched with rarely occurring Smotifs. This phenomenon becomes especially apparent when the relative frequency of occurrences of Smotifs drops below 0.09 (Fig. 5, Fig. S2, Fig. S3).
Two examples of the above observations are illustrated in Fig. 6. The first example is the new fold target T0181, discussed above (PDB code: 1nyn; Fig. 6A). The second example is a member of the immunoglobulin fold (PDB code: 1gyv; Fig. 6B), which is one of the most populated folds. Target 181, a new fold structure, can be decomposed into 7 Smotifs, where five are considered low frequency (i.e. frequency smaller than 0.01, or less than 1%). On the other hand, for a representative structure of the immunoglobulin fold (SCOP fold descriptor 48725, Immunoglobulin-like beta sandwich), the opposite situation occurs. Five out of the 7 Smotifs that comprise the structure are very well represented (high frequency) in the pool of Smotifs (Fig. 6B).
One could speculate that some novel folds were recently discovered simply because of difficulty in experimental determination, i.e. these structures are harder to solve. We used the XtalPred program [27] to predict the crystallizability of 347 new folds and 2802 known folds, all solved approximately in the same time period (since SCOP 1.73 released in 2007). We found that new folds from the most recent SCOP release 1.75 indeed have a small tendency to be less feasible for experiments. However, XtalPred and other prediction methods for protein crystallizability heavily rely on known homologs of a query sequence. The rationale is that if a protein with a similar sequence has been solved before it usually indicates that this particular protein family is more experimentally tractable. This artifact is illustrated in our analysis by the fact that while new folds from SCOP 1.75 do show less favorable XtalPred scores as compared to known folds, this difference disappears in case of new folds of SCOP 1.73 (Fig. 7).
Another plausible way to generate new folds is to combine, otherwise common Smotifs in an unusual sequence, to result in a new topology. To explore this, we calculated a Novelty Z-score for each protein, which was obtained from the product of individual Smotif frequencies. The hypothesis is that if the Novelty Z-score of some novel folds is similar to that of known folds, then the novelty for these cases must be a consequence of a never before seen combination of otherwise common Smotifs rather than a result of being constructed from rare Smotifs. And while new folds from the CASP dataset do show a distribution of Novelty Z-scores biased towards low values (Fig. S4), in the case of SCOP 1.75 (Fig. S5) and SCOP 1.73 (Fig. S6), most novel folds are indistinguishable from already known structures in terms of their overall Novelty Z-scores, which indicates that these structures may indeed be a new topological arrangement of common Smotifs. However, one may note the more frequent extreme negative outliers in the distributions for the novel folds in these datasets (averages and standard deviations are −1.03±1.1, 0.25±1.35 and 0.0±1.0 for CASP dataset, SCOP 1.75, and SCOP 1.73, respectively). This means that although novel folds are often built using a higher proportion of rare Smotifs, in many cases these folds are novel because their Smotifs are assembled in an unusual sequence. This is illustrated with Target T0201 (CASP 6) and the S50S ribosomal protein L6P (PDB code 1s72 chain E) that share 3 out of 6 of their Smotifs (Fig. 8). However the sequential arrangement of these shared Smotifs is different, yielding different topologies.
Since the early nineteen-nineties, it has been clear that the universe of protein folds is much more limited and redundant than the sequences adopting them [28]. Structural biology and the recently launched Structural Genomics efforts have discovered a large subset of possible fold shapes. Many predictions suggest that most of the folds are already known [28], [29], [30]. Meanwhile, by solving many of the possible folds, the characteristic differences earlier described among fold definitions has become more blurred [8], [10], [31]. In practice, discovering all possible folds may be an impossible task, partly because it is clear now that the definition of folds is highly subjective [2], and partly because the distribution of folds is extremely uneven: while only a dozen superfolds seem to populate half of a typical genome, and only about 200 folds populate 2/3 of it, it is possible that many thousands of more rarely occurring shapes need to be discovered to reach 80–90% coverage of all possible shapes that were established during evolution [32][33].
In this work we explored the entirety of protein shapes from the perspective of their Smotif building blocks, which can be defined more objectively than the folds themselves, and which are observed to be nearly completely sampled in the currently known structures. Using this repertoire of Smotifs, we observed that novel folds can be distinguished from already discovered ones by the presence of rare Smotifs and, less often as an unusual combination of otherwise common Smotifs. The most frequently used motifs have a higher average number of internal contacts, while some of the rarest motifs are larger, and contain longer linker regions. These observations may be useful starting points for future works to identifying or designing sequences that are likely to constitute “novel” folds.
While in this work we defined Smotifs according to practical considerations and did not investigate if these Smotifs or subset of them could also serve as possible units for structural evolution, it is noteworthy to mention other studies that identified similar structural elements as possible building blocks of structural hierarchy using different approaches. The so called Closed Loops were identified by their close Cα-Cα contacts from solution structures and found to have a nearly standard size (27 residues +/−5). This typical size distribution of Closed Loops was supported by polymer statistics, as it is the theoretical optimal size for loop closure and subsequently suggested to be a universal building block of protein folds [34], [35]. In another approach, dynamic Monte Carlo simulation of alpha carbon chain of the nearest 24 neighbor in a lattice model identified clusters of “most interacting residues”, which serve as anchors for protein folding [36]. These anchors were found to be conserved hydrophobic clusters of residues that keep together the so called Tightened End Fragments, which essentially correspond to the Closed Loop definition. Finally a most recent paper updates on the idea of ancient relic peptides of length 20–40 residues that co-occur in different structural contexts, and suggested to be an ancestral pool of peptide modules [37].
All structures from CASP 3,4,5,6 meetings [38] that were manually identified as “novel folds” at the time of the experiment: CASP3 (protein identification (PDB code): T0052 (2ezm), T0059 (1d3b), T0063 (1bkb), T0067 (1bd9), T0071 (1b9k), T0080 (1bnk), and T0083 (1dw9)), CASP 4 (T0086 (1fw9), T0116_3a (1ewq), T0116_3b (1ewq), T0120_1 (1fu1), and T0124 (1jad)), CASP5 (T0129 (1izm), T0149_2 (1nij), T0161 (1mw5), and T0162_2 (1izn)) and CASP6 (T0201 (1s12), T0209_2 (1xqb), T0216_1 (1vl4), T0216_2 (1vl4), T0238 (1w33), T0242 (2blk) and T0248_2 (1td6)) were collected. Four tailored datasets of previously solved protein structures were generated for comparisons with the “novel” folds of each CASP experiment (see below). The tailored datasets did not contain any structure that was deposited after June 1998 (6,366 entries), June 2000 (10,199 entries), June 2002 (15,234 entries) and June 2004 (22,076 entries) to compare with targets from CASP3, CASP4, CASP5, and CASP6 respectively. Similarly, four SCOP [3] database releases were used for calculating motif frequencies (see below): SCOP 1.39 (CASP3 new fold set), SCOP 1.53 (CASP4 new fold set), SCOP 1.61 (CASP5 new fold set), and SCOP 1.69 (for CASP6 new fold set). Since CASP meetings start in June and SCOP databases were released after June during the same year, all structures that were present in the SCOP database with a deposition date after June were removed.
Similarly, we have downloaded all “new folds” from the SCOP 1.73 and 1.75 releases, 123 and 110 folds, respectively, that are part of a total of 1140 proteins. The list of new folds for earlier releases can be found at SCOP via History link (http://scop.mrc-lmb.cam.ac.uk/scop/index_prevrel.html).
A Smotif is defined by two consecutive regular secondary elements (i.e. α-helix or β-strand), connected by a loop. The N and C-terminal regular secondary structures of a Smotif are referred as SS1 and SS2, respectively. Motif geometry refers to the local spatial arrangement of SS1 with respect to SS2 as introduced in [22] using four internal coordinates. Briefly, SS1 and SS2 were represented by their principal moments of inertia (M1 and M2). If P1 and P2 are the end point of SS1 and start point of SS2, and L is the vector between P1 and P2, then plane Π is defined by M1 and L and plane Γ is defned by M1 and the normal to plane Π. Geometry of a Smotif is expressed by four measures: the distance (D) between the C-terminal of SS1 and the N-terminal of SS2 (distance between P1 and P2) and three angles: a hoist (δ): angle between L and M1, a packing (θ): angle between M1 and M2, and a meridian (ρ): angle between M2 and plane Γ (Fig. 2 in [21]).
A library has been established that classifies each Smotif in all PDB structures. This library is organized in a two-level hierarchy: in the first level of classification, (i) Smotifs are identified according to the type of bracing secondary structures: αα, αβ, βα and ββ according to the definition of secondary structure by the DSSP program [39]. At the second level, (ii) Smotifs are grouped according to their geometry, as described above [21], [22]. A protein structure can, therefore, be expressed as a string of overlapping Smotifs where the SS2 from one Smotif constitutes the SS1 in the following Smotif.
The geometrical values used in the second level of classification are distributed in a continuous space. Distance is distributed between 0 and 40 Å. (values larger than 40 Å are assigned to 40), δ and θ angles span from 0 to 180 degrees, and the ρ angle spans from 0 to 360 degrees. In order to compare Smotif geometries, the parameter spaces of geometrical values were binned, where each bin is defined by the 4 parameters described above. A range of binning sizes and parameter intervals were explored for the four variables in order to get the sharpest partitioning power of the geometrical space with the smallest number of possible bins (Fig. S7). The quality of the binning was assessed by calculating the RMSD (Root Mean Square Deviation) and the LGA scores [40] upon structural superposition for all Smotifs that were classified in the same or different geometrical bin. The optimal bin partitioning for each parameter was obtained by studying the distribution of distance and angle values of Smotifs in SCOP 1.71 proteins and resulted in only 324 types of Smotif definitions using the following binning values: 4 Å bins for distance, 60 degree bins for δ and θ starting at 0 degree, and 60 degree bins for ρ, starting at 30 degree. At this level of bin resolution the RMSD upon structural superposition of more than 75% of Smotifs that belong to the same geometrical bin falls below 1 Å (Fig. S7).
A program that defines Smotifs is available upon request from the authors.
All protein structures that were identified as “new folds” from SCOP releases 1.73 and 1.75 and CASP 3–6 meetings were decomposed into Smotifs. In case of SCOP, each release identifies the new folds in comparison to the rest of the folds while in case of the CASP sets a Smotif library extracted from a backdated PDB was prepared for each CASP meeting. Within the pairs of datasets, Smotifs in SCOP new and existing folds and Smotifs from CASP new folds and the corresponding Smotif library from previously solved structures, were compared to evaluate the existence of identical Smotifs in the novel folds and the previously defined folds. The first comparison was based on the type of secondary structures and the geometry (D, hoist, packing, and meridian) of Smotifs. In a second, stricter comparison, the lengths of the flanking secondary elements (SS1 and SS2) were also compared. If these lengths differed by more than 2 or 4 residues in the case of strands or helices, respectively, the Smotifs were considered different.
To avoid redundancy when calculating the frequencies of Smotif occurrences for each four-dimensional geometric bin, only a single protein was selected from each protein fold (as defined by SCOP database). Since fold families contain more than one protein structure and structures that belong to the same fold may have a variable number of Smotifs this selection process was repeated 100 times, randomly selecting a different protein in each analysis. Therefore, the frequency of occurrence of a given geometrical bin is the average of counts computed from 100 rounds of analysis for each family.
Each of the proteins in the database was converted into a string of Smotifs. Thus, a protein having 5 regular secondary structures would be expressed as a string of 4 overlapping Smotifs. For each protein, a normalized probability score of observing such a string of Smotifs was calculated:(1)where N is the number of Smotifs and fr is the frequency of the Smotif i as calculated previously. Individual scores were converted into statistical Z-scores using the mean (μ) and standard deviation (σ) of the population of scores, as (2)(2)
Internal contact ratio was calculated as the number of non-bonded atomic contacts (i.e. H-bonds, polar contacts, hydrophobic contacts) between SS1 and SS2 divided by Smotif size. Contacts were defined by the Contact of Structural Units (CSU) program [41]. CSU is based on the detailed analysis of interatomic contacts and interface complementarity. For every structural unit CSU calculates the solvent accessible surface of every atom and determines the contacting residues and type of interactions they undergo including all putative hydrogen bond contacts.
Protein crystallizability was predicted with the XtalPred server [27]. XtalPred predicts protein crystallizibility by combining nine features: length, length of predicted disorder, Gravy index, insertion score, instability index, percent of coil structure, isoelectric point. Based on these features the protein is assigned to one of five crystallization classes: optimal, suboptimal, average, difficult and very difficult. Each class represents different crystallization success rate observed in TargetDB [42]. Three SCOP domain datasets were compiled for submission to XtalPred; domains from “new folds” as defined in (1) SCOP 1.75 and (2) in SCOP 1.73, respectively, and (3) domains in SCOP 1.75 that were added since the release of SCOP 1.73 and that were not new folds. This ensures that we are focusing on proteins that were solved approximately in the same time but were classified differently in terms of novelty. The amino acid sequences of the domains were obtained from the ASTRAL website (astral-scopdom-seqres-gd-all-1.75.fa, astral-scopdom-seqres-gd-all-1.73.fa). Sequence redundancy was removed among the domains using CDHIT clustering [43] at 95% sequence identity threshold. The SCOP 1.75 and 1.73 “new fold” domains dataset contained 170 and 177 representative sequences (517 and 558 redundant sequences), respectively, and the SCOP 1.75 “known fold” dataset contained 2802 representative sequences (out of 13,043 redundant ones). Each amino acid sequence was submitted to XtalPred to calculate the crystallizability class.
The corresponding PDB structure, chain identification and residue range was located for each Smotif (369,859 Smotifs in total). We calculated ACC values (water exposed surface area or number of water molecules in contact with the residue) using the DSSP program [44]. The average solvent accessibility of Smotifs was calculated by averaging the ACC values over all residues of the Smotif. We also calculated average ACC values by excluding loop residues, which are usually exposed, for each Smotif, but the conclusions were not affected.
|
10.1371/journal.pcbi.1002218 | Robust Signal Processing in Living Cells | Cellular signaling networks have evolved an astonishing ability to function reliably and with high fidelity in uncertain environments. A crucial prerequisite for the high precision exhibited by many signaling circuits is their ability to keep the concentrations of active signaling compounds within tightly defined bounds, despite strong stochastic fluctuations in copy numbers and other detrimental influences. Based on a simple mathematical formalism, we identify topological organizing principles that facilitate such robust control of intracellular concentrations in the face of multifarious perturbations. Our framework allows us to judge whether a multiple-input-multiple-output reaction network is robust against large perturbations of network parameters and enables the predictive design of perfectly robust synthetic network architectures. Utilizing the Escherichia coli chemotaxis pathway as a hallmark example, we provide experimental evidence that our framework indeed allows us to unravel the topological organization of robust signaling. We demonstrate that the specific organization of the pathway allows the system to maintain global concentration robustness of the diffusible response regulator CheY with respect to several dominant perturbations. Our framework provides a counterpoint to the hypothesis that cellular function relies on an extensive machinery to fine-tune or control intracellular parameters. Rather, we suggest that for a large class of perturbations, there exists an appropriate topology that renders the network output invariant to the respective perturbations.
| Cellular signaling networks have to function reliably and with high fidelity in an uncertain environment. In this paper, we investigate the topological principles to achieve such robust signal processing in living cells. Specifically, we identify the topological organizing principles that enable a signaling network to keep the stationary intracellular concentrations of certain molecules, such as active signaling compounds, within tightly defined bounds – despite conditions of uncertainty and in the face of multiple perturbations. We demonstrate that an appropriate topological organization renders the output of the pathway invariant against a large class of possible detrimental fluctuations, such as changes in energy states or total protein concentrations. Furthermore, we show that the topological requirements for robust signal processing can be formalized in terms of a linear vector space, denoted as invariant perturbation space, that predicts the robustness properties of the network. Constructing this invariant perturbation space for the Escherichia coli chemotaxis pathway reveals that the pathway is indeed invariant with respect to most dominant perturbations that would otherwise significantly hamper information transmission. Our framework provides a counterpoint to the hypothesis that cellular function relies on an extensive machinery to fine-tune or control intracellular parameters.
| All living cells rely on the capacity to respond to intra- or extracellular signals and have evolved a dedicated biochemical machinery to continuously sense, transmit, and process a variety of internal and environmental cues. A key requisite for reliable signal processing is the capability of living cells to keep the stationary intracellular concentrations of certain molecules, such as active signaling compounds, within tightly defined bounds – despite conditions of uncertainty and in the face of multiple perturbations. While the apparent insensitivity of key intracellular concentrations, and hence of cellular function, to detrimental influences is widely recognized as a salient property of cellular signaling, knowledge of the precise mechanisms underlying these instances of pathway robustness is still fragmentary [1]–[6].
Here, we report a simple, yet highly efficient, novel formalism that pinpoints the necessary architecture for concentration robustness in living cells. We assert and substantiate by mathematical proof and experimental evidence that certain classes of network architectures render the functional output of the network, as represented by a set of steady state protein concentrations, invariant to a large class of perturbations. Our approach emphasizes robustness as a structural property of a network as a whole, rather than as a consequence of parameter-tuning or individual positive or negative interaction loops [3], [7], and offers a novel paradigm to understand the topological organization of cellular signaling networks. Differing from earlier approaches, our framework accounts for perturbations of large magnitude and is not restricted to a particular class of network kinetics, such as mass-action systems [5]. Applications include the robustness of input-output relationships with respect to variations in total component concentrations, reaction parameters, abundances of common resources like ATP, RNA polymerases, and ribosomes, as well as detrimental effects of pathway crosstalk, and variations in temperature. Our focus is on perturbations whose time scales are slow compared to the intrinsic dynamics of the pathway.
To establish the mechanisms of robust signaling, we consider a multi input-multi output signaling network, whose temporal behavior is described by a set of ordinary differential equations for the state variables, , e.g., , where the indices indicate different variables or reaction fluxes . The equations can be organized into the more compact form,(1)where denotes the stoichiometric matrix. The reaction fluxes are specified by functions that depend on the variables and a set of parameters . We require the existence of a – not necessarily unique – stationary state that obeys the steady state condition with . In the following, we assume that the functionality of the network is encoded in the steady state of a subset of output variables, defined as , whose concentration values depend on a set of intra- or extracellular signals. The remaining intermediate variables are defined by . The system is said to exhibit local concentration robustness with respect to a particular parameter if a sufficiently small perturbation in this parameter does not affect the stationary concentrations of the output variables, . Mathematically, the perturbation is characterized by the vector of logarithmic partial derivatives with elements , evaluated at the stationary state.
As the main result of the work, we now seek to identify stringent conditions on the network architecture – rather than on kinetic parameters – such that the robustness property holds for perturbations of large magnitude. To this end, we first recall the conditions for local concentration robustness. Utilizing results from linear control theory, local robustness can be ascribed to two scenarios: Either the perturbation has no effect on any stationary concentration within the network. In this case, the vector is an element of a vector space spanned by the columns of a matrix – with being a basis of the right nullspace of the scaled stoichiometric matrix, defined such that . Or, more generally, the perturbation propagates through the network and affects the stationary concentration of some or all of the non-robust intermediate variables , albeit without affecting the set of output variables . In this case, it can be shown that the perturbation vector is an element of the joint vector space spanned by the columns of and the columns of a matrix . The latter matrix is given by the logarithmic partial derivatives of reaction rates with respect to the intermediate variables , with elements . We note that the elements of correspond to the kinetic orders or scaled elasticities of the reaction fluxes and attain integer values for the case of reaction networks that follow mass-action kinetics [8]. Taken together, a necessary and sufficient condition for local concentration robustness is therefore that the vector is an element of the vector space spanned by the columns of and , or equivalently, that the rank condition,(2)is fulfilled. Here, the notation denotes a concatenation of the columns of both matrices. To ascertain local concentration robustness the rank condition is evaluated at the particular stationary state. See Materials and Methods and Text S1 for details and proof.
In general, local concentration robustness is not a sufficient condition to allow for robust signal processing in living cells. The fluctuations encountered by biological systems, such as variations in component concentrations arising from stochasticity in gene expression, are typically of large magnitude and cannot be described by local perturbations at a particular stationary state. Our aim is therefore to establish precise conditions for global concentration robustness. Specifically, a system is said to exhibit global concentration robustness with respect to a particular parameter if the stationary concentrations of the set of output variables is invariant with respect to perturbations in . Thereby, may take any value within a biophysically feasible perturbation set and is not restricted to small variations.
To obtain a viable criterion to judge global concentration robustness, we therefore extract from the local vector space, spanned by the columns of , the largest subspace that does not depend on the choice of kinetic parameters, and hence, the specific stationary state. This subspace, denoted as the invariant perturbation space , defines the largest vector space that guarantees local robustness at any stationary state of the system. Consequently, a perturbation of increasing magnitude that is confined to the invariant perturbation space may gradually affect the intermediate variables, but does not affect the designated output variables. The condition for global concentration robustness is then given by , or, equivalently, as , where denotes a matrix whose columns span the vector space .
We emphasize that the matrix and its associated vector space are independent of kinetic parameters and therefore represent a genuine structural property of any signaling network. Proof and an algorithm is relegated to Materials and Methods and the SI, here we only outline its construction using a simple example.
To illustrate the construction of the invariant perturbation space, we consider the simple pathway shown in Figure 1. Here, the output variable of the pathway is subject to strong fluctuations in its synthesis rate . Rather than aiming to suppress the detrimental perturbations, the pathway employs an intermediate variable that compensates perturbations and ensures global concentration robustness of . The pathway is described by two differential equations for the time-dependent behavior of the concentrations of and , respectively,(3)For brevity, and as the only assumption on the rate equations and kinetic parameters, we require that the pathway gives rise to a unique stationary state for each value of . To obtain insight about the concentration robustness of the variable with respect to , we construct the invariant perturbation space, derived from the concatenated matrix . The matrix is given by the logarithmic partial derivatives of reaction rates with respect to the intermediate non-robust variable . We obtain(4)where denotes the unknown state-dependent logarithmic partial derivative with respect to the variable . In general, the precise value of depends on the functional form of the rate equations, the value of the perturbation , and the kinetic parameters.
The matrix can be constructed algorithmically from the stoichiometric matrix. We obtain,(5)where and denote the stationary flux values.
To obtain a matrix representation of the invariant perturbation space, we now need to identify the largest parameter-independent subspace spanned by the columns of . To this end, we note that the vector space spanned by the columns of a matrix remains invariant under elementary matrix operations (EMO), such as multiplication of a column by the same non-zero factor or the addition of an arbitrary multiple of one column to another. Applying a set of suitable EMOs, we obtain(6)We note that in this particular case, the invariant perturbation space is of the same dimension as the local vector space. In general, however, not all dimensions of the local space are retained, see Section III of Text S1 for an example.
To test for global concentration robustness of the variable with respect to , we now have to evaluate the rank condition . The perturbation is characterized by the vector(7)where denotes the unknown state-dependent value of the logarithmic partial derivative. It can be straightforwardly ascertained that the rank condition for global concentration robustness is fulfilled, irrespective of the value of . Hence, the variable exhibits global concentration robustness with respect to perturbations in its synthesis rate.
We note that our simple example is a well-known instance of robust perfect adaptation [9], [10]. Biologically, the variable acts as an integrator, under the condition that the degradation rate of is independent of the concentration of itself. Utilizing our approach, the invariant perturbation space can be constructed algorithmically for any given reaction network. The condition for global concentration robustness can then be ascertained by a simple numerical test and does not require extensive computations or additional expert knowledge.
To further illustrate the construction of the invariant perturbation space, we briefly consider the robustness of a canonical two-component system – one of the simplest and best-studied examples of robust signaling. Bacterial two-component systems typically consist of a membrane-bound sensor kinase that senses a specific stimulus and a cognate response regulator that modulates the signal response. Reliable functioning of two-component systems often requires that the output of the pathway, the concentration of phosphorylated response regulator as a function of an external stimulus, is not compromised by fluctuations in total protein concentrations of both components. The robustness of bacterial two-component systems with respect to such concentration fluctuations was investigated previously [11], [12]. In particular, Batchelor and Goulian [11] identified that the principal mechanism for concentration robustness is due to a bifunctional histidine kinase that phosphorylates and dephosphorylates its cognate response regulator.
Figure 2 depicts a simplified model of the respective system. The histidine kinase () is phosphorylated by an external ligand. The phosphorylated kinase () transfers the phospho-group to the unphosphorylated response regulator (). The pathway output is the concentration of the phosphorylated diffusible response regulator (). Importantly, dephosphorylation of the response regulator () requires the participation of the bifunctional histidine kinase (). Utilizing our approach, we seek to confirm that, in this case, the stationary concentration of is invariant to variations in the expression levels of both proteins. For brevity, we again consider a highly simplified system and focus on the construction of the invariant perturbation space. In particular, the formation of protein complexes is neglected and all phosphorylation reactions are assumed to follow mass-action kinetics. A solution of the full system, including an explicit account of conserved moieties, is provided in Text S1 (Section VII).
To obtain the invariant perturbation space, we first derive the matrix of logarithmic partial derivatives of reaction rates with respect to the non-robust variables , , and . We assume that both proteins are synthesized and degraded with unknown rates and – using the simplifying assumption that degradation (or dilution) acts only on the unphosphorylated forms and . The unknown partial derivatives of the degradation reactions are denoted as and , respectively. The remaining reactions are assumed to follow mass-action kinetics, resulting in partial logarithmic derivatives of unit value. Specifically, the phosphorylation rate is dependent on the concentration of the unphosphorylated form , the phosphotransfer rate depends upon the concentration of and , and the dephosphorylation rate finally depends on the concentration of the phosphorylated response regulator , as well as the unphosphorylated form of the bifunctional kinase. The matrix is given in Figure 2B.
As the next step, we need to identify the nullspace of the scaled stoichiometric matrix . The nullspace of the unscaled stoichiometric matrix is readily available using standard tools of linear algebra. The representation of the unscaled nullspace is subsequently scaled with the unknown steady state reaction rates, such that , , and . A representation of the scaled nullspace is provided in Figure 2B. Taken together, we again obtain the invariant perturbation space as the maximal subspace spanned by the columns of independent of kinetic parameters or steady state reaction rates. A matrix representation of the invariant perturbation space is given in Figure 2C.
We assume that the system is perturbed by unknown variations in the synthesis rates of both proteins, and , respectively. The corresponding partial derivatives with respect to unknown perturbations are denoted as and and shown in Figure 2C. To ascertain global concentration robustness of , we confirm that the rank condition is indeed fulfilled. Hence, the output of the pathway, the steady state concentration of , is invariant to perturbations in the synthesis rates of both components.
We note that, in general, our approach does presuppose that the system gives rise to a biologically feasible steady state solution for . This requirement usually entails additional constraints on the possible reaction rates and kinetic parameters. For example, robustness of is only feasible under the condition that the total expression of the response regulator exceeds the steady state solution for . Below we present a generalization of the rank condition to account for additional constraints on molecule concentrations (see also Text S1, Section VIII).
Our approach is applicable to a variety of different scenarios, including several special cases which are discussed in the following. In particular, our approach relies on an interpretation of the elements of the matrix – the logarithmic partial derivatives of reaction rates with respect to the intermediate variables. For typical biochemical rate equations, these partial derivatives are nonlinear functions of kinetic parameters and therefore usually represent unknown and state-dependent quantities. However, as demonstrated above, our approach is still applicable in such a situation and does not require extensive knowledge of the functional form of the rate equations. In the most general case, each logarithmic partial derivative is represented by an unknown non-zero value within the matrix . The resulting invariant perturbation space is required to be independent of these unknown derivatives. Hence, the invariant perturbation space is predominantly a structural property of the network and is identical for structurally equivalent networks. See Text S1 for details.
However, in some cases the elements of the matrix can be constraint further, owing either to particular functional forms of the rate equations or to simplifying assumptions that allow to approximate more complicated rate equations. An example of the former are generalized mass-action (GMA) kinetics of a reaction rate ,(8)For GMA kinetics, the partial logarithmic derivatives correspond to the exponents and are often considered to be constant quantities. Consequently, the partial logarithmic derivatives may be represented as constant entries within the matrix . In this case, the invariant perturbation space is particularly straightforward to obtain.
As an example of simplifying assumptions, we note that complex rate equations are often approximated by more simple equations corresponding to specific kinetic regimes. In particular, a Michaelis-Menten equation can be approximated by a mass-action term or a constant for substrate concentrations far below or far above the Michaelis constant, respectively. In this case, the logarithmic partial derivative is approximately constant or zero, respectively. However, any result from applying the criterion for global concentration robustness is only valid as long as the assumptions underlying the approximation are fulfilled.
As yet, we have only considered reaction networks in the absence of mass-conservation relationships or conserved moieties. However, often the total concentration of some compounds can be considered as approximately constant over the relevant time-scales, giving rise to additional dependencies between variables. In this case, the system of differential equations for the independent state variables, is augmented by a set of dependent state variables , whose values are determined by a set of mass conservation equations. The full system of equations governing the time evolution of the system is(9)(10)with the vector denoting the total concentration of each molecular component. The matrix denotes a link matrix and usually consists of integer elements. To incorporate these dependencies within our approach, we must modify the definition of the matrix to account for the logarithmic partial derivatives with respect to the dependent variables. See Text S1 for details. Using the augmented matrix , our approach proceeds as described above. As a corollary, we then obtain a simple criterion to judge global concentration robustness with respect to perturbations in conserved total concentrations [5], [6], see Text S1 (Section VII.B).
Our approach differs from a number of previous approaches to investigate robustness of biochemical reaction networks [1], [5], [6], [13]. The formalism is not restricted to systems described by mass-action kinetics, but is applicable a wide range of ODE-based descriptions of biochemical networks. Likewise, we do not focus on specific types of perturbations, such as variations in conserved moieties [5] or temperature [13]. Rather, our approach is applicable to any perturbation that can be described by a vector of partial derivatives of reaction rates – of which variations in conserved moieties, as well as of temperature are particular examples. We also mainly envision a scenario, where the perturbations are slow compared to the intrinsic fluctuation-compensation dynamics of the pathway. In particular, we consider the steady state of a selected subset of variables to represent the robust output of the system. Transient fluctuations in the vicinity of this state are not considered. However, the scenario described in this work indeed holds for many instances of cellular robustness. For example, in the case of gene expression noise, the observed fluctuations in expression levels are usually at least an order of magnitude slower than the phosphorylation dynamics in subsequent signaling pathways. Hence such fluctuations can be compensated by post-translational mechanisms – as described within this work. Similar arguments apply for several dominant fluctuations typically encountered by cellular signaling pathways, such as variations in temperature or abundance of common resources like ATP.
To substantiate the explanatory power achieved by an interpretation of a complex cellular signaling network in terms of its associated invariant perturbation space, we now consider the robustness of the E. coli chemotaxis pathway. The topology of the pathway is depicted in Figure 3. The pathway responds to changes in concentrations of chemoeffectors such as certain amino acids or sugars by altering the phosphorylation state of the diffusible response regulator CheY. The concentration of free phosphorylated CheY () – the central output quantity of the pathway – then determines swimming behavior of the cell. Robust and precise regulation of is a prerequisite for high chemotaxis efficiency and is maintained in the face of multifarious perturbations, most notably ATP availability, stochasticity in component abundance [14], and receptor cluster assembly [15], [16]. However, seemingly contradicting its functional objective, the pathway is rather sensitive to variations in the expression of some of its constituent proteins. For example, it was shown that a two-fold overexpression of CheZ or CheY levels already result in an 50% decrease of experimentally observed chemotactic performance, as determined by the size of swarm rings on soft agar plates [17].
To reveal the mechanisms underlying the remarkable robustness that nonetheless allows reliable functioning of the pathway, we construct the invariant perturbation space as described above. The concatenated matrix is obtained by considering the stoichiometric matrix and the kinetic dependencies shown in Figure 3. See SI (Section V) for details of the derivation. A parameter independent representation of the invariant perturbation space is shown in Figure 4A. To investigate the robustness of the pathway, we first consider changes in chemoeffector concentration (L), perturbations in the expression of CheA (A) and CheW (W), as well as variations in receptors (T) and ATP availability (ATP). The corresponding perturbation vectors are shown in Figure 4B. In each case, the corresponding perturbation vector is an element of the invariant perturbation space and the rank condition for global concentration robustness of is fulfilled. Hence, the diffusible response regulator indeed exhibits global robustness of its stationary concentration with respect to these five highly detrimental influences.
Next, we consider changes in the expression of the individual proteins CheR (), CheB (), CheY (), and CheZ (). The corresponding perturbation vectors are given in Figure 4C. As can be ascertained by inspection of the rank condition, the respective perturbation vectors are not elements of the invariant space – in good agreement with the rather high sensitivity exhibited by the pathway in response to variations in the expression of these proteins [17]. Nonetheless, the observed total concentrations of CheR, CheB, CheY, and CheZ are not “fine-tuned” and are known to exhibit considerable variability under various conditions. To explain this alleged paradox, we have to take the sequential arrangement of genes into operons, as shown in Figure 3B, into account. A closer inspection of Figure 4 then reveals that perturbations that arise from concerted fluctuations in protein concentrations, induced by stochastic synthesis of meche operon transcripts, are within the invariant perturbation space. And, indeed, coupling of expression levels of chemotaxis proteins adjacent on an operon has been experimentally shown to positively correlate with chemotactic efficiency and to underlie active selection during chemotactic spreading on soft agar plates [18]. Generalizing from this example, we expect that gene organization into operons and expression from polycistronic mRNA is a generic, evolutionary driven, mechanism to alleviate detrimental effects of stochasticity in gene expression. In the context of our framework, coupling of expression on the transcriptional [14] and translational level [18], reduces the effective dimensionality of a perturbation, thereby enabling an invariant perturbation space of lower dimension to compensate and counteract the detrimental effects of fluctuations. In this sense, strong transcriptional and translational coupling is closely related to the robustness conveyed by bifunctional enzymes [5]. For the E. coli chemotaxis pathway strong coupling of genes expressed from one operon is evident in cells expressing yellow and cyan fluorescent protein fusions to CheY and CheZ, respectively, from one bicistronic plasmid construct, as shown in Figure 5A [14], [19]. The striking invariance of the pathway output upon a seven fold concerted increase in the transcriptional activity of the chemotaxis operons following the deletion of the anti sigma factor FlgM is shown in Figure 5B [14], [19].
As argued previously [20], the benefits of co-variation to reduce the effective dimensionality of perturbations are likely to confer a selective advantage strong enough to drive the assembly of genes into operons. Our results also highlight the functional importance of seemingly redundant or insignificant interaction characteristics, whose functional relevance is difficult to ascertain without an appropriate theoretical framework. A striking example is the catalyzed dephosphorylation of CheY by CheZ, as opposed to the uncatalysed dephosphorylation of CheB. While such a difference often seems extraneous to reliable signal transduction, such differences also shape the invariant perturbation space and are therefore crucial to achieve robust signal processing. A further example of a relevant interaction characteristic is the competitive binding of CheY and CheB to CheA, which results in a phosphotransfer rate to CheB that scales as . While not fine-tuned on the parameter level, this qualitative dependence is a prerequisite for robustness of the pathway output and in excellent agreement with experimental findings [21]. In this sense, our approach also offers a theoretical framework to investigate the functional relevance of given reaction characteristics – beyond their role in straightforward signal transmission.
The interpretation of a complex cellular signaling network in terms of its associated invariant perturbation space has profound implications for our ability to understand and eventually rationally engineer robust biological circuits. There is increasing evidence that the utilization of post-transcriptional noise compensatory networks is a widespread mechanism in prokaryotic signaling. Experimentally ascertained examples include instances of two-component systems [1], [11], [12], the regulation of the glyoxylate bypass [22], and the sporulation network of B. subtilis [20]. In each case, an evolved network topology relegates potentially detrimental fluctuations in compound concentrations to its associated invariant perturbation space – rather than utilizing an expensive machinery to fine-tune native expression levels. We expect that similar mechanisms will provide an indispensable backbone for synthetic biology. Guided by the algorithmic construction of the invariant perturbation space, a key strategy for synthetic biology is to either maximize the invariant perturbation space by rationally rewiring the specificity of protein interactions [23], [24], or correlating perturbations among components, by placing genes on polycistronic mRNA or by building fusion constructs – in each case circumventing the need to fine-tune parameters that are experimentally hard to control. Our algorithm is applicable to large systems and requires only qualitative information on kinetic interactions. Our results allow us to clarify several long-standing issues relating to the emergence of cellular robustness. In particular, we hypothesize that the ubiquitous existence of puzzling, seemingly redundant, interaction loops that characterize our current understanding of cellular pathways is deeply rooted in as yet unrecognized mechanisms to counteract functional fragilities [10], [25]. In this sense, an interpretation of signalling architecture in terms of its invariant perturbation space offers a novel paradigm to understand cellular robustness, with the prospect to rationally engineer robust signaling circuits or target cellular defects.
In the following, we outline the conditions for local concentration robustness, as stated in Eq. (2). We employ a logarithmic expansion of the stationary form of Eq. (1), , with , to linear order in a perturbation and the resulting changes in the state variables ,(11)with denoting a square matrix with entries on the diagonal. The expansion coefficients are(12)The relative perturbation and its response are defined as , , and .
In the absence of the condition for robustness of the pathway output, , the expansion Eq. (11) has a unique solution for that quantifies the local linear response to a sufficiently small perturbation in parameters. The existence of the solution is guaranteed by the requirement that the Jacobian of the system is of full rank and hence invertible, implied by the dynamic stability of the considered steady state. Similar consideration are extensively utilized within, for example, Metabolic Control Analysis [8], [13], [26], [27].
However, the requirement of concentration robustness, , removes the degrees of freedom that correspond to (changes in) the output variables . In this case, Eq. (11) translates into the condition(13)In general, Eq. (13) is overdetermined, that is, no solution exists and the condition cannot be fulfilled. Eq. (13) has a unique solution if and only if at least one of the following two conditions holds: Either the columns of the matrix are elements of the right nullspace of the matrix , spanned by the columns of the matrix . In this case, we obtain and, necessarily, . Or, the columns of the matrix are linearly dependent on the columns of the matrix . In mathematical terms, these two conditions can be summarized in the equation(14)Here, the columns of span the right nullspace of , such that . The notation denotes a concatenation of the columns of the matrices and , as described in the main text. See also SI (Sections II and IV) for a rigorous derivation.
In the following, we outline the formal definitions and proof for global concentration robustness. For conciseness, we consider only generalized mass action (GMA) networks without conserved moieties. The general case, including a formal derivation of the conditions for global concentration robustness, is described in SI, Section IV. The biochemical network is defined as in Eq. (1). We consider a perturbation that takes values in a physically reasonable, connected set . For a GMA network, the reaction rates are given by for reaction rates affected by the perturbation and for reaction rates not affected by the perturbation. The concentration vector is split into as described in the main text. The network is assumed to have a perturbation-dependent steady state which is asymptotically stable for all in a physically reasonable, connected perturbation set .
The property of global concentration robustness is then formally defined as follows: For any values of the reaction rate parameters and any choice of the functions , the steady state output concentration vector is constant over .
The global invariant perturbation space as discussed in the main text for a GMA network is given by , where denotes the image or range of the matrix. Thereby, are the columns of the matrix with elements , i.e. the logarithmic derivatives of the reaction rate vector with respect to , and is a matrix whose columns span the space of the vectors which are in the kernel of for all in the kernel of .
To obtain a condition for global concentration robustness, we consider the vectors whose elements are zero whenever the reaction rate is not affected by the perturbation . If all such vectors are element of the space , then the network has global concentration robustness. Conversely, if there exists such a which is not in the space , then there exists rate parameters and functions for which the steady state output concentration is not constant over , and thus the network does not have global concentration robustness. Computationally, the condition can be tested by the rank condition , where is any matrix whose columns span the space .
The signal transduction of the E. coli chemotaxis pathway can be described to good accuracy by the interplay of the core components, the methyl accepting chemoreceptors (Tar, Tap, Tsr, Trg), the methyltransferase CheR, the methylesterase CheB, the response regulator CheY and its designated phosphatase CheZ (see Box 1). The total concentrations of these proteins are approximately , , , , , , and M. The concentration includes all receptors where CheR and phosphorylated CheB can bind to with high affinity, via a pentapeptide sequence at the carboxyl termini of the Tar and Tsr receptors. The set of mass action equations that determine the phosphorylation level of free diffusible response regulator proteins, , are listed below.
In the following, we consider the stationary case of the chemotaxis equations. We thereby employ the approximations as , , as , and . The simplified set of stationary equations read(22)(23)(24)(25)where we have resolved the complexes and and introduced the stationary functions and as defined above for time independent mean methylation level and fixed ligand concentration . A derivation of the entries in Figure 4 is provided in Text S1.
|
10.1371/journal.pgen.1003983 | Basolateral Mg2+ Extrusion via CNNM4 Mediates Transcellular Mg2+ Transport across Epithelia: A Mouse Model | Transcellular Mg2+ transport across epithelia, involving both apical entry and basolateral extrusion, is essential for magnesium homeostasis, but molecules involved in basolateral extrusion have not yet been identified. Here, we show that CNNM4 is the basolaterally located Mg2+ extrusion molecule. CNNM4 is strongly expressed in intestinal epithelia and localizes to their basolateral membrane. CNNM4-knockout mice showed hypomagnesemia due to the intestinal malabsorption of magnesium, suggesting its role in Mg2+ extrusion to the inner parts of body. Imaging analyses revealed that CNNM4 can extrude Mg2+ by exchanging intracellular Mg2+ with extracellular Na+. Furthermore, CNNM4 mutations cause Jalili syndrome, characterized by recessive amelogenesis imperfecta with cone-rod dystrophy. CNNM4-knockout mice showed defective amelogenesis, and CNNM4 again localizes to the basolateral membrane of ameloblasts, the enamel-forming epithelial cells. Missense point mutations associated with the disease abolish the Mg2+ extrusion activity. These results demonstrate the crucial importance of Mg2+ extrusion by CNNM4 in organismal and topical regulation of magnesium.
| Magnesium is an essential element for living organisms. Its absorption occurs at the intestine through the barrier comprised of epithelial cells. In this process, transcellular Mg2+ transport across epithelia, involving both entry from one side and extrusion from the other side, is important. Previous studies have revealed the role of Mg2+-permeable channel protein in Mg2+ entry into the epithelial cells. However, the identity of proteins involved in Mg2+ extrusion to the inner parts of body has remained unknown. Mice genetically engineered not to express CNNM4, which localizes to the epithelial membrane facing to the inner parts of body, show hypomagnesemia due to the defect in magnesium absorption. Functional analyses using culture cells directly reveal that CNNM4 can extrude intracellular Mg2+ to the outside of cells. These results indicate that CNNM4 mediates transcellular Mg2+ transport across the intestinal epithelia. Furthermore, we also show that these CNNM4-lacking mice also have a defect in amelogenesis, which is consistent with the disease symptoms of Jalili syndrome that is known to be caused by mutations in the CNNM4 gene.
| Magnesium is an essential element involved in a wide variety of biological activities. Homeostasis of the magnesium level is strictly regulated by intestinal absorption and renal reabsorption, in which epithelia function as a barrier that permits selective and regulated transport of Mg2+ from apical to basolateral surfaces. Genomic analyses of familial cases of hypomagnesemia have identified key molecules directly involved in these processes. CLDN16, encoding claudin-16/paracellin-1, and CLDN19, encoding claudin-19, are mutated in recessive familial hypomagnesemia with hypercalciuria and nephrocalcinosis [1], [2]. These genes are highly expressed in the thick ascending limb of Henle's loop in the kidney and encode tight junction proteins, which form a cation-selective paracellular channel and drive the flux of Mg2+ between adjacent epithelial cells [3]. Another key molecule is TRPM6; mutations of TRPM6 cause recessive hypomagnesemia with secondary hypocalcemia [4], [5].
TRPM6 is a member of the transient receptor potential melastatin-related (TRPM) protein family and constitutes a Mg2+-permeable ion channel that localizes to the apical membrane of epithelial cells in the intestine and kidney [6]. In addition, it has also been shown that TRPM7, a close relative of TRPM6, plays an essential role in magnesium homeostasis in mice [7]. Therefore, TRPM6/TRPM7 plays a primary role in the apical entry of Mg2+ into cells, which is the first step in transcellular Mg2+ absorption across the epithelial barrier, another major Mg2+ transport pathway. To accomplish Mg2+ absorption, epithelial cells need to extrude Mg2+ via their basolateral membrane by opposing the inward-oriented driving force on Mg2+ imposed by the electrical membrane potential. Such a transcellular Mg2+ transport mechanism, involving both apical entry and basolateral extrusion, is evolutionarily conserved from Caenorhabditis elegans [8], [9], but molecules involved in basolateral Mg2+ extrusion have not been identified.
Ancient conserved domain protein/cyclin M (CNNM) constitutes a family of 4 integral membrane proteins that possess an evolutionarily conserved but uncharacterized domain from bacteria [10]. Recent genomic analyses have revealed a link between CNNM genes and magnesium homeostasis. Several single nucleotide polymorphisms in CNNM genes are associated with the serum magnesium level [11] and mutations in CNNM2 cause familial dominant hypomagnesemia [12]. The bacterial ortholog of these proteins in Salmonella, CorC, has been suggested to participate in Mg2+ efflux [13], while ectopically expressed CNNM2 in Xenopus oocytes showed voltage-dependent transport of several divalent cations, including Mg2+ [14]. Moreover, expression of a splice-variant of CNNM2 could restore the growth of a Mg2+-deficient Salmonella strain [15]. However, a study on CNNM2 expressed in HEK293 cells showed that it mediates a Na+ current [12]. Therefore, the importance of CNNMs in Mg2+ transport still remains unknown. Moreover, it has been reported that mutations in CNNM4 cause Jalili syndrome, which is characterized by recessive amelogenesis imperfecta (AI) and cone-rod dystrophy (CRD) [16], [17]. However, the molecular mechanism that links CNNM4 dysfunction to these pathological conditions and its relationship with magnesium homeostasis remain to be determined.
In this study, we generated CNNM4-knockout mice; these mice showed defects in amelogenesis and intestinal Mg2+ absorption. Endogenous CNNM4 is highly expressed in the mature ameloblasts and intestinal epithelia, and localizes at their basolateral membrane. Functional analyses at the molecular and organismal levels revealed a common role for CNNM4 in mediating transcellular Mg2+ transport by basolateral Mg2+ extrusion.
To reveal the physiological function of CNNM4, we generated CNNM4-knockout mice. For this purpose, we used a commercially available embryonic stem (ES) cell clone, which possesses the neomycin-resistance gene cassette inserted in the genomic region between the first and second exons of CNNM4 by homologous recombination (Figure 1A). Chimeric heterozygous mice were obtained by blastocyst injection of the ES cells, and CNNM4-knockout mice were obtained by breeding. Successful recombination in the genomic DNA obtained from CNNM4+/− and CNNM4−/− mice was confirmed by Southern blotting (Figure 1B) and routine genotyping was done by PCR (Figure 1C). The gene cassette contains the splice acceptor sequence that forces mRNA splicing to occur artificially at the acceptor sequence, and the resulting mRNA is truncated after the second exon. Indeed, immunoblotting analyses with the anti-CNNM4 antibody (Figure S1) confirmed that CNNM4−/− mice lack expression of endogenous CNNM4 protein (Figure 1D). Both CNNM4+/− and CNNM4−/− mice were viable, with no gross abnormalities.
Immunoblotting analyses of lysates obtained from various organs showed that CNNM4 is highly expressed in the small intestine and colon (Figure 2A), consistent with the previously reported analyses at mRNA level [18]. We next performed immunohistochemical staining to examine the expression pattern in the colon. As shown in Figure 2B, positive CNNM4 signals were specifically observed at the mucosal epithelial layer, with no significant signals at the muscular layer. Counterstaining of the tissue samples obtained from CNNM4−/− mice showed no positive signals, thus confirming that the signal at the mucosal epithelia properly reflects the localization of endogenous CNNM4.
To precisely determine the subcellular localization of CNNM4, we also performed immunofluorescence microscopy. Low-magnification images confirmed the specific expression of CNNM4 in the mucosal epithelia (Figure 2C). In the high-magnification images, positive signals for CNNM4 were mostly observed at the plasma membrane, but were clearly separated from those for F-actin, immediately beneath those for ZO-1 (Figure 2D). F-actin staining strongly labels the apical membrane of the intestinal epithelia [19], and ZO-1 is a marker for tight junctions in the colonic mucosa [20], which form a physical border between the apical and the basolateral membranes. Thus, these results imply a basolateral localization of CNNM4 in the colon epithelia.
To further confirm the basolateral localization of CNNM4, we ectopically expressed CNNM4-FLAG in MDCK cells, which maintain a highly polarized epithelial character in culture. As shown in Figure S2, the expressed CNNM4-FLAG proteins co-localized with Na+/K+ ATPase (basolateral marker), immediately beneath ZO-1.
The fact that CNNM4, a putative Mg2+ transporter, localizes to the basolateral membrane of the intestinal epithelia suggests the involvement of CNNM4 in the regulation of magnesium homeostasis. To explore this possibility, we analyzed the magnesium levels in CNNM4−/− mice maintained on a normal diet (CLEA Rodent Diet CE-2 containing 0.34% magnesium). Magnesium quantitation, using the colorimetric reagent Xylidyl Blue-I, showed that CNNM4−/− mice had a significantly lower serum magnesium concentration: an approximately 18% decrease was observed in comparison to CNNM4+/+ mice (Figure 3A). Moreover, the magnesium level in urine was drastically reduced, by approximately 71% (Figure 3A). These results demonstrate that CNNM4−/− mice have altered magnesium regulation. To examine whether this alteration was specific to magnesium, we used inductively coupled plasma-emission spectroscopy (ICP-ES) to examine the levels of several major metal elements in serum. As shown in Figure 3B, the levels of sodium, potassium, and calcium were not affected in CNNM4−/− mice, whereas the magnesium level was significantly reduced. We then observed mice fed a magnesium-deficient diet (containing 0.0027% magnesium) and found a significant increase in mortality in CNNM4−/− mice (Figure 3C), indicating that CNNM4−/− mice have abnormal magnesium homeostasis.
Magnesium homeostasis is regulated by the balance between intestinal absorption and renal excretion. The decrease in renal excretion can be considered to reflect a compensatory response to maintain magnesium levels during hypomagnesemia caused by intestinal malabsorption. To directly measure the effect on intestinal absorption, we analyzed the magnesium content in feces. As shown in Figure 3D, there was significantly higher excretion of magnesium in feces in CNNM4−/− mice (22% increase compared to CNNM4+/+ mice), without a significant difference in the quantity of food ingested. These symptoms are very similar to those of the TRPM7-mutant mice, which have defects in intestinal magnesium absorption [7]. Collectively, these results indicate that CNNM4-deficiency results in malabsorption of magnesium at the intestine.
To clarify the molecular function of CNNM4, we first examined the effect of CNNM4-overexpresion on the intracellular levels of major metal elements by using ICP-ES. As shown in Figure 4A, HEK293 cells transfected with CNNM4-FLAG contained more sodium and less magnesium in comparison to control vector-transfected cells, consistent with the occurrence of Mg2+ extrusion. Other analyzed elements (potassium, calcium, and zinc) showed no significant differences. We next performed imaging analyses with Magnesium Green, a fluorescent indicator for Mg2+. HEK293 cells transfected with CNNM4-FLAG were first loaded with Mg2+ by bathing them in a solution containing 40 mM Mg2+, which was then exchanged with a Mg2+-free solution to artificially promote Mg2+ extrusion. As shown in Figure 4B, the intensity of fluorescent signals in cells expressing CNNM4-FLAG (confirmed by immunofluorescence microscopy, performed after the imaging analyses) rapidly decreased immediately after Mg2+ depletion, whereas only a very subtle decrease was observed in empty vector-transfected cells. Thus, CNNM4 is able to stimulate Mg2+ extrusion.
The electrical potential across the plasma membrane forces Mg2+ to move inward into cells, and thus, energy supply is needed to actively extrude Mg2+ to the outside. Many proteins involved in active transport across the plasma membrane utilize the large electrochemical potential of Na+. To determine the importance of extracellular Na+ in Mg2+ extrusion, we first performed Mg2+ extrusion assays by replacing Na+ in the medium with another cation, N-methyl-D-glucamine (NMDG). In this case, Mg2+ extrusion was completely abolished (“−Na+” in Figure 4B). We also performed time-lapse imaging analyses for 10 min (Figure 4C and Video S1). Mg2+ depletion in the medium caused Mg2+ extrusion in CNNM4-expressing cells (Phase 1) and addition of 40 mM Mg2+ restored intracellular Mg2+ (Phase 2). In the absence of extracellular Na+, Mg2+ depletion did not induce Mg2+ extrusion (Phase 3), but restoration of Na+ instantaneously caused Mg2+ extrusion (Phase 4). Such tight coupling between the presence of extracellular Na+ and the occurrence of Mg2+ extrusion further supports the notion that CNNM4 stimulates Na+/Mg2+ exchange; this is also consistent with the sodium increase observed in CNNM4-expressing cells (Figure 4A).
To determine whether the rapid restoration of intracellular Mg2+ in 40 mM Mg2+ media is caused by the reverse action of CNNM4, we performed similar time-lapse imaging analyses using cells treated with Cobalt (III) hexammine (CoHex), which broadly inhibits channel-mediated Mg2+ influx [21], [22]. CoHex treatment significantly inhibited the Mg2+ recovery (Figure S3A), suggesting that some Mg2+ channels are involved in the Mg2+ recovery process. For more detailed characterization of the Mg2+ uptake in CNNM4-expressing cells, we performed a quantitative imaging analyses by using a less-sensitive, but ratiometric fluorescent probe Mag-fura2. Cells were bathed in extracellular solutions containing various concentrations of Mg2+ and Na+. Unlike 40 mM extracellular Mg2+, 10 mM Mg2+ was not sufficient to load CNNM4-expressing cells when extracellular Na+ was set to 78.1 mM (Figure S3B). However, when extracellular Na+ was depleted (0 mM), CNNM4-expressing cells incorporated significant amount of Mg2+ even at 10 mM. Furthermore, we observed that even though the Mg2+ level in CNNM4-expressing cells was lower than that in the control cells before loading, it became much higher after the loading procedure with 10 mM Mg2+, 0 mM Na+ solution, and then returned to the basal level when extracellular Mg2+ was removed. These data strongly suggest the occurrence of the reverse action of CNNM4 and corroborate our notion that CNNM4 stimulates Na+/Mg2+ exchange.
To characterize the molecular function of CNNM4 in more detail, we next performed electrophysiological analyses on CNNM4 expressed in HEK293 cells. As shown in Figure S4A–C, CNNM4 expression induced no significant electronic currents, while CNNM2 expression generated an inward current of Na+, as reported previously [12]. To directly measure Mg2+ extrusion, we next performed simultaneous Mg2+ imaging and electrophysiological recording experiments. The exchange of the extracellular solution with an Mg2+-free solution stimulated rapid Mg2+ decrease without inducing significant electronic currents in CNNM4-expressing cells (Figure S4D–E). These results suggest the possibility that CNNM4 might exchange 2 Na+ and 1 Mg2+, and thus, it is electroneutral. Therefore, we performed quantitative imaging analyses of intracellular Na+ and Mg2+ by using ratiometric fluorescent probes, sodium-binding benzofuran isophthalate (SBFI) and Mag-fura2, respectively. As shown in Figure 4D, Mg2+ depletion from the extracellular medium induced not only the decrease of intracellular Mg2+ but also the increase of intracellular Na+. In addition, the molar ratio of increased Na+ and decreased Mg2+ was calculated to be 2.14∶1, which is roughly consistent with the electroneutral exchange of Na+ and Mg2+ (2∶1). To quantitatively assess the dependency of Mg2+ extrusion on the presence of extracellular Na+, we performed Mg2+ extrusion assays by changing the concentration of extracellular Na+. Extracellular Na+ accelerated Mg2+ extrusion in a dose-dependent manner, and the Hill coefficient was calculated to be 1.90, a value close to 2 (Figure 4E). This result suggests that there are 2 or more Na+-binding sites in CNNM4, which also agrees with the characteristic of 2 Na+/1 Mg2+ exchanger.
One of the common features of Jalili syndrome, which is caused by mutations in CNNM4, is AI, the malformation of tooth enamel [16], [17]. We noticed that CNNM4−/− mice displayed abnormal teeth with chalky-white discoloration (Figure 5A), which is typically observed in mice with defective amelogenesis. This phenotype was apparent as early as 3 weeks of age and was observed in all CNNM4−/− mice examined. To characterize the abnormality in amelogenesis, we subjected maxillary incisors to analyses with scanning electron microscopy (SEM). The low-magnification images showed that the thickness of the enamel layer in CNNM4−/− mice was not so different from that in CNNM4+/+ mice (Figure 5B). However, the high-magnification images showed that the enamel rods were immature and the inter-rod area was increased in CNNM4−/− mice (Figure 5C). We then subjected the samples to composition analyses using energy dispersive X-ray spectrometry (EDX). As shown in Figure 5D, the levels of both calcium and phosphorus were significantly decreased in CNNM4−/− mice, confirming the occurrence of hypomineralization.
To explore the role of CNNM4 in amelogenesis, we performed immunohistochemical staining to examine the localization of CNNM4 in the enamel-forming tissue. Enamel formation occurs in the area covered by ectodermally-derived epithelial cells, so-called ameloblasts [23]. The ameloblasts first deposit a complex extracellular matrix composed of enamel proteins (secretory stage), and then come to maturity, with a shortened morphology, and promote mineralization of the enamel (maturation stage). During the secretory stage, positive signals of CNNM4 were observed specifically at the stratum intermedium (SI) layer, but not in the ameloblasts (Figure 6A–B). However, the expression pattern significantly changes at the maturation stage, with strong positive signals in the ameloblasts themselves.
Mature ameloblasts are known to undergo repetitive cycles of transdifferentiation between ruffle-ended (RA) and smooth-ended (SA) ameloblasts, which can be discerned by ZO-1-staining [24]. Immunofluorescence staining showed that CNNM4 exists throughout the basolateral membrane immediately beneath the ZO-1 signals in the RA-type ameloblasts, which possess dot-like accumulations of ZO-1 at the cell-cell contact sites facing the enamel-forming area (Figure 6C). It should be noted that this basolateral localization pattern of CNNM4 in RA-type ameloblasts is quite similar to that observed in intestinal epithelia (Figure 2D), suggesting that CNNM4 promotes Mg2+ removal from the maturing enamel. Indeed, the elemental analyses of the mature enamel with EDX indicated that the magnesium levels were significantly increased in CNNM4−/− mice (Figure 5D).
To ascertain the functional importance of Mg2+ extrusion by CNNM4, we examined whether missense point mutations in CNNM4, which have been reported to occur in the patients of Jalili syndrome [16], [17], have any effects on Mg2+ extrusion activity. We tested the effect of two different point mutations, viz., S200Y and L324P, both of which occur in the evolutionarily conserved DUF21 domain (Figure 7A). When these mutants were expressed in HEK293 cells, they localized to the plasma membrane, similarly to wild-type (WT) CNNM4 (Figure 7B). However, both mutants showed very weak, if any, Mg2+ extrusion activity in comparison to WT CNNM4 (Figure 7C). Therefore, a dysfunction in Mg2+ extrusion, caused by mutations in this gene, probably underlies this little understood human disease.
Another feature of Jalili syndrome is CRD, which is characterized with the degeneration of rod and cone photoreceptors in the retina [16], [17]. To investigate the integrity of retinal function of CNNM4−/− mice, we performed histological and electroretinogram (ERG) analyses. To observe the retinal histology, we stained retinal sections from 2-month-old (young adult) CNNM4−/− mice with toluidine blue. We found that the retinal layers were normal and no symptom of retinal degeneration was observed in the retina of CNNM4−/− mice (Figure S5A). We also performed immunofluorescent analysis in the CNNM4−/− retina, using markers of photoreceptor, bipolar, and horizontal cells. Outer segments of rod and cone photoreceptors stained with anti-rhodopsin and cone opsins (M-opsin and S-opsin) are normal in the CNNM4−/− retina (Figure S5B). Cone photoreceptor synaptic terminals stained with Peanut Agglutinin (PNA) are also localized normally in the outer plexiform layer (OPL). Photoreceptor synaptic ribbons stained with the anti-Ctbp2 antibody showed horseshoe-like structure in the vicinity of dendritic tips of bipolar cells stained with the anti-mGluR6 antibody both in CNNM4+/+ and CNNM4−/− mice, and dendrites of rod ON-bipolar cells stained with anti-PKC-α antibody and processes of horizontal cells stained with the anti-Calbindin antibody were properly extended into the OPL in the CNNM4−/−retina. To evaluate the retinal function, we recorded ERGs from CNNM4−/− mice. As shown in Figure S5C, no obvious difference was observed between 2-month-old CNNM4+/+ and CNNM4−/− mice in their ERGs under both scotopic and photopic conditions, which reflects the functions of rods and cones, respectively (a-wave in scotopic condition 1.0 log stimuli: +/+, 280±57 µV; −/−, 251±23; unpaired t-test: p = 0.6204; a-wave in photopic condition 1.0 log stimuli: +/+, 11.3±1.9 µV; −/−, 8.1±0.9; p = 0.1966; b-wave in scotopic condition 1.0 log stimuli: +/+, 619±115 µV; −/−, 563±44; p = 0.6366; b-wave in photopic condition 1.0 log stimuli: +/+, 163±24 µV; −/−, 126±23; p = 0.2937; +/+, n = 5; −/−, n = 6).
Retinal dysfunction occasionally becomes evident with age. Indeed, knockout mice for RP3, one of causative genes of human hereditary retinal diseases [25], do not show an apparent loss of the retinal cells at 1 month of age, but degeneration of photoreceptor cells has occurred at 6 months [26]. Therefore, histological analyses of the retina of 6-month-old CNNM4−/− mice were performed. However, we did not observe any signs of histological abnormalities (Figure S5A–B). We also recorded ERGs from 6-month-old CNNM4−/− mice and again observed normal ERGs under both scotopic and photopic conditions (Figure S5C).
In this study, we have shown that CNNM4 localizes to the basolateral membrane of epithelial cells and extrudes Mg2+. Theoretically, Mg2+ extrusion requires an energy supply to overcome the inward-oriented force on Mg2+ diffusion imposed by the membrane potential. A Na+-coupling Mg2+ extrusion mechanism has long been suggested, and indeed, various types of mammalian cells possess Na+/Mg2+ exchange activity [27], [28]. It was recently reported that SLC41A1 can biochemically function as a Na+/Mg2+ exchanger when expressed in HEK293 cells [29]. It is expressed ubiquitously [30], and the ectopically expressed SLC41A1 in MDCK cells localizes at the basolateral membrane [31]. Therefore, SLC41A1 may also be involved in the regulation of directional Mg2+ transport across the intestinal epithelia. However, it should be noted that the speed of Mg2+ extrusion by CNNM4 (reaching plateau after 1∼2 min) is much faster than that by SLC41A1 (after ∼10 min) [29]. Such a rapid Mg2+ extrusion has not been reported in the previous studies characterizing the endogenous Mg2+ extrusion systems in non-intestinal cells [27], [28]. Thus, CNNM4 appears to be a qualitatively different, high capacity type of Mg2+ extrusion molecule, which may have a specialized role in the intestinal epithelia. Magnesium absorption from the intestine is essential for magnesium homeostasis, and 100–150 mg magnesium is daily absorbed from the intestine in humans [32]. To absorb such a large amount of magnesium through the intestinal epithelia, the magnesium transport system in the intestine should be highly active. It is known that both paracellular and transcellular pathways are functional and play important roles in the intestinal magnesium absorption [33]. In the transcellular pathway, Mg2+ entry into the intestinal epithelial cells is mediated by apically localized Mg2+-permeable channels TRPM6/7 that can rapidly incorporate Mg2+ [6], [7]. Therefore, it is very reasonable that Mg2+ extrusion from the basolateral membrane is mediated by high capacity transporters, such as CNNM4, to achieve efficient transcellular Mg2+ transport through intestinal epithelia.
CNNM4−/− mice showed a defect in magnesium absorption, but were viable, without any significant observable phenotype when fed a normal diet. CNNM proteins comprise a family of 4 related proteins, CNNM1–4 [10], and thus, the mild phenotype of CNNM4−/− mice can be ascribed to the functional complementation by other CNNM family proteins. CNNM4 is expressed in the intestine, but not in the kidney, and thus, it will not affect renal reabsorption, the other key process in the regulation of magnesium homeostasis. The amount of magnesium reabsorbed from the glomerular filtrate is estimated to be about 10 times that absorbed from digested food. Therefore, the absence of CNNM4 in the kidney raises the next important question of what molecule is responsible for Mg2+ extrusion from distal convoluted tubule (DCT) cells in the kidney, where TRPM6 is expressed at the apical membrane and where transcellular Mg2+ transport occurs [34]. Two previous papers have reported strong expression and localization of CNNM2 at the basolateral membrane of the DCT cells [12], [18]. Therefore, it can be assumed that CNNM2 plays an important role in renal reabsorption of magnesium at the DCT by mediating transcellular Mg2+ transport cooperatively with TRPM6. It should be noted here that SLC41A1 is also expressed in the DCT cells and its gene mutation causes nephronophthisis-related disorder [31]. Because the affected patients did not exhibit any abnormalities in serum or urine magnesium level, the authors speculated that the disease phenotype might result from perturbed intracellular magnesium homeostasis [31]. Future studies using gene knockout mice and detailed analyses of the biochemical properties of these molecules, CNNM2 and SLC41A1, will grant more insight into the individual roles in renal magnesium control.
CNNM4 is mutated in Jalili syndrome, which is characterized by recessive AI and CRD [16], [17]. Our CNNM4−/− mice showed no signs of abnormalities in the retinal tissue architecture and function (Figure S5). In contrast, we observed a clear amelogenesis-defective phenotype. In the enamel-forming tissue, CNNM4 is strongly expressed at the basolateral membrane in RA-type ameloblasts. Such a basolateral localization is similar to that observed in the intestinal epithelia and suggests that CNNM4 is involved in the vectorial transport of Mg2+ from the enamel-forming areas through the ameloblasts. Indeed, RA-type ameloblasts have tight junctions in the region adjacent to the enamel-forming areas, and form a niche in which active ion transport occurs [24]. The precise role of Mg2+ in the enamel-forming process remains unknown, but the striking expression of CNNM4 in RA-type mature ameloblasts suggests that Mg2+ needs to be removed from the enamel tissue to promote mineralization of enamel. Indeed, it has been reported that the magnesium content of the enamel is inversely correlated with the extent of mineralization [35]. Further characterization of CNNM4-knockout mice will contribute to a better understanding of this intriguing process in which the most solid tissue in the body is generated.
We appropriately treated mice to ameliorate suffering, according to the guidelines for proper conduct of animal experiments (issued by the Science Council of Japan), and received approval for this study from the institutional review board of Osaka University.
We purchased an ES clone (ID: EPD0426_1_C08) from EUCOMM, in which the neomycin-resistant gene cassette had been inserted in the genomic region between the first and second exons of CNNM4 by homologous recombination. The ES cells were used to generate germline chimeras that were bred with C57BL/6J females to generate CNNM4-knockout mice. Southern blot analyses were performed to confirm appropriate recombination. Genomic DNA of mice was digested with EcoRV and hybridized with the external or neo probes. Genotyping PCR was performed using the following primer set: 5′-TAACTGTTGGAAGGCTGAGG-3′ and 5′-AGGCAGGGGCTCCCTTTCAT-3′. Mice were maintained under standard specific pathogen-free conditions.
Human CNNM4 cDNA was purchased from Invitrogen (IMAGE: 30340626). Amino acid substituted mutants S200Y and L324P were generated with the QuickChange Site-Directed Mutagenesis Kit (Agilent). An anti-CNNM4 rabbit polyclonal antibody was raised in rabbits immunized with bacterially expressed His-CNNM4 proteins (amino acids 546–775) and purified with corresponding GST-tagged recombinant proteins. Anti-ZO-1 mouse monoclonal antibody was generated in the previous study [36] and provided by Dr. Masahiko Itoh (Dokkyo Medical University) and Dr. Mikio Furuse (Kobe University). Anti-mGluR6 guinea pig polyclonal antibody was described previously [37]. Anti-Na+/K+ ATPase mouse monoclonal antibody (#05-369) and anti-M-opsin rabbit polyclonal antibody (AB5405) were purchased from Merck Millipore. Anti-FLAG rabbit polyclonal antibody (F7425) and anti-PKCα rabbit polyclonal antibody (P4334) were purchased from Sigma-Aldrich. Anti-Ctbp2 mouse monoclonal antibody (612044) was purchased from BD Biosciences. Anti-Rhodopsin (LB-5597) and anti-Calbindin (PC253L) rabbit polyclonal antibodies were purchased from LSL and Calbiochem, respectively. Anti-S-opsin goat polyclonal antibody (sc-14363) was purchased from Santa Cruz Biotechnology. Alexa Fluor 488-conjugated anti-rabbit IgG was purchased from Invitrogen and Sigma-Aldrich. Alexa Fluor 488-conjugated anti-mouse IgG was purchased from Sigma-Aldrich. Alexa Fluor 568-conjugated anti-mouse IgG, and rhodamine-labelled phalloidin were purchased from Invitrogen. Cy3-conjugated anti-rabbit, -goat and -guinea pig IgGs were purchased from Jackson ImmunoResearch Laboratories. Rhodamine-labeled PNA (RL1072) was purchased from Vector Laboratories.
HEK293 cells and MDCK cells were cultured in Dulbecco's modified Eagle's medium supplemented with 10% fetal bovine serum and antibiotics. Transient expression and knockdown were achieved using LipofectAmine2000 (Invitrogen) to transfect cells with plasmids or siRNAs according to the manufacturer's instruction. Plasmid constructs in the pCMV-Tag 4 vector (Agilent Technologies) were used for expression of CNNM4. For knockdown experiments, duplex siRNAs against human CNNM4 (Invitrogen), which target the following sequence: CNNM4-siRNA, 5′-GCGAGAGCAUGAAGCUGUAUGCACU-3′, were used. As control, we used siRNA representing a scrambled sequence of CNNM4-siRNA, 5′-GCGACGAAAGUGUCGGUAUCGAACU-3′.
For intestine preparation, intestines were dissected from 2-month-old mice, embedded in OCT compound (Sakura Finetechnical), frozen in liquid nitrogen, and then sectioned into at 10-µm sections using a cryostat (Leica). The sections were mounted on glass slides, air-dried, and fixed with phosphate-buffered saline (PBS) containing 4% paraformaldehyde (PFA) for 10 min at 4°C. For mandible preparation, 6-week-old mice were anesthetized and fixed by perfusion with PBS containing 4% PFA. Mandibles were dissected out, fixed with PBS containing 4% PFA for 12 h at 4°C, decalcified with 10% EDTA for 2 weeks, dehydrated with xylene through a graded ethanol series, and embedded in paraffin. Sections (4-µm thick) were cut using a microtome (Leica), and then mounted on glass slides. Slides were heat-treated in Pascal, a pressure chamber (Dako) and cooled at room temperature after deparaffinization and rehydration. Both frozen and paraffin-embedded sections were then incubated with PBS containing 0.3% H2O2. After blocking with PBS containing 3% fetal bovine serum and 10% bovine serum albumin for 1 h at room temperature, specimens were incubated with the primary antibodies overnight at 4°C, followed by incubation with the peroxidase-conjugated secondary antibodies for 1 h at room temperature. Immunostaining was developed with diaminobenzidine and counterstained with Mayer's haematoxylin. The specimens were observed under a microscope (BX41 equipped with a DP20 camera; Olympus). Differential Interference Contrast (DIC) images were collected using an inverted microscope (IX71 equipped with a DP20 camera; Olympus).
Cells cultured on coverglasses were washed with PBS and fixed with 1% formaldehyde for 15 min at room temperature. When stained for ZO-1, cells were permeabilized with 0.5% TritonX-100 in PBS for 10 min at room temperature. When stained for Na+/K+-ATPase, cells were permeabilized with 0.1% TritonX-100 for 5 min at room temperature. After blocking with PBS containing 3% fetal bovine serum and 10% bovine serum albumin (blocking buffer) for 1 h, cells were incubated for 12 h with the primary antibody diluted in blocking buffer. After 3 washes with PBS, cells were incubated for 30 min with the appropriate secondary antibodies diluted in blocking buffer. Cryosections of intestines and paraffin-embedded sections were prepared as described above. When stained for ZO-1, sections were permeabilized with ice-cold acetone for 3 min after fixation. Fixed sections were blocked and incubated with the primary and secondary antibodies as for cultured cells. After washing with PBS, coverglasses were mounted with Aqueous Mounting Medium PermaFluor (Thermo SCIENTIFIC) and observed with a confocal scanning laser microscope (FLUOVIEW FV1000; Olympus). The procedure of immunofluorescent analysis of retinas was described previously [38], [39]. Mouse eyes were fixed with PBS containing 4% PFA for 30 min or 5 min, embedded in OCT compound, frozen, and sectioned. Frozen 20 µm sections were blocked with PBS containing 5% normal goat serum and 0.5% Triton X-100 for 30 min, and then incubated with primary antibodies for 4 h at room temperature. Slides were washed with PBS three times for 5 min each time and incubated with secondary antibodies for 2 h at room temperature. The specimens were observed with a confocal scanning laser microscope (LSM510; Carl Zeiss).
ERG responses were measured after overnight dark adaptation using PuREC system with LED electrodes (Mayo Corporation) [40]. 2- and 6-month-old mice were anesthetized with an intraperitoneal injection of ketamine and xylazine. The mice were stimulated with stroboscopic stimuli of 1.0 log cd-s/m2 (photopic units) maximum intensity. 4 levels of stimulus intensities ranging from −4.0 to 1.0 log cd-s/m2 were used for the scotopic ERG recordings, and 4 levels of stimuli ranging from −0.5 to 1.0 log cd-s/m2 were used for the photopic ERGs. Animals were light adapted for 10 min before the photopic ERG recordings. 8 and 16 responses were averaged for photopic (−4.0 and −3.0 log) and all scotopic recordings, respectively.
Mice were fed either a normal diet containing 0.34% magnesium (CLEA Rodent Diet CE-2, CLEA Japan) or a magnesium-deficient diet containing 0.0027% magnesium (CLEA Japan). Blood samples were obtained from 8-week-old mice. These were incubated at 4°C overnight, and serum was then collected by centrifugation at 1,000× g for 20 min at 4°C. Urine and feces samples were collected from 2-month-old mice by using metabolic cages (CLEA Japan). Feces were air-dried, incubated with 1 N nitric acid (1∶10; wt∶volume) overnight, and then centrifuged. The magnesium concentration of the supernatant was determined using Xylidyl Blue-I (Wako) according to the manufacturer's instructions.
Serum samples were mixed with HCl at a final concentration of 1% and incubated at 95°C for 2 h. Samples were then subjected to elementary analysis with ICPS-8100 (Shimadzu), according to the manufacturer's instructions. The mean of triplicate measurements was used to represent the result of a single sample. The results were normalized to total protein levels, which were determined by the Bradford method.
Mg2+-imaging analyses with Magnesium Green were performed as follows. HEK293 cells were incubated with Mg2+-loading buffer (78.1 mM NaCl, 5.4 mM KCl, 1.8 mM CaCl2, 40 mM MgCl2, 5.5 mM glucose, 5.5 mM HEPES-KOH, pH 7.4), including 2 µM Magnesium Green-AM (Invitrogen), for 45 min at 37°C. The cells were rinsed once with loading buffer and viewed using a microscope (IX81 equipped with a DP30BW camera and a USH-1030L mercury lamp; Olympus). Fluorescence was measured every 20 sec (excitation at 470–490 nm and emission at 505–545 nm) under the control of the Metamorph software (Molecular Devices). Then, the buffer was changed to −Mg2+ buffer (MgCl2 in the loading buffer was replaced with 60 mM NaCl), or to −Mg2+−Na+ buffer (NaCl in −Mg2+ buffer was replaced with NMDG-Cl). The data are presented as line plots (mean of 10 cells). After imaging analyses, cells were fixed with PBS containing 3.7% formaldehyde and subjected to immunofluorescence microscopy to confirm protein expression. Cobalt (III) hexammine was purchased from SIGMA.
pIRES-HcRed plasmids [41] for expressing CNNM2 or CNNM4 were transfected into HEK293 cells with FuGENE6 (Roche). After 24 h, cells were plated on glass coverslips coated with poly-L-lysine (SIGMA) and maintained in normal culture media plus 40 mM MgCl2 until use. Patch-clamp experiments under the whole-cell configuration were performed according to Stuiver et al., [12] with minor modifications. The experiments were performed with Axopatch 200B amplifier and Clampex 9.2 data acquisition system (Molecular Devices), and borosilicate patch pipettes had resistances of 5–10 MΩ after filled with the intracellular solution. Voltage steps (1 sec in duration) from the holding potential of 0 mV to potentials between −120 to +70 mV with 10 mV increment were delivered every 4 sec. The density current was obtained from the peak current at −110 mV and was normalized with the membrane capacitance of the cell. The extracellular solutions was 80 mM Na-gluconate, 0 or 20 mM MgSO4, 10 mM HEPES (pH 7.35 adjusted with Tris). The intracellular solution was 120 mM NMDG, 120 mM 2-(N-morpholino)-ethanesulfonic acid hydrate, 2 mM MgSO4, 10 mM HEPES (pH 7.2 adjusted with H2SO4). All solutions were adjusted to 295–305 mOsm with sucrose.
Simultaneous Mg2+-imaging and electrophysiological recording experiments were performed with IX71 microscope (Olympus) equipped with iXon EM-CCD camera (Andor Technology) and a xenon lamp in Lambda DG-4 illumination system (Sutter Instrument). Borosilicate patch pipettes had resistances of 3–5 MΩ after filled with the intracellular solution containing 2 µM Magnesium Green (non-AM form, Invitrogen). Cells were voltage clamped to −10 mV, and the imaging was started after the fluorescent intensities from the cell became stabilized (20–35 min after the establishment of the whole-cell configuration). The fluorescence was measured every 20 sec (excitation at 470–490 nm and emission at 505–545 nm). Mg2+-loading buffer and −Mg2+ buffer were used as extracellular solutions. The intracellular solution was 2 mM MgCl2, 2 mM NaCl, 5 mM EGTA, 140 mM KCl, 5 mM HEPES (pH 7.25 adjusted with KOH). All solutions were adjusted to 295–305 mOsm with sucrose.
HEK293 cells were transfected with expression plasmids for CNNM4, and maintained in normal culture media plus 40 mM MgCl2 until use. Mg2+ extrusion assays were performed with the abovementioned protocol, with following modifications. Cells were loaded with 2 µM Mag-fura2-AM or 3 µM SBFI-AM (Invitrogen) and viewed using the IX81 microscope (Olympus) equipped with ORCA-Flash 4.0 CMOS camera (Hamamatsu Photonics) and USH-1030L mercury lamp (Olympus). The fluorescence was measured every 20 sec (excitation at 330–350 nm and 370–390 nm, and emission at 505–545 nm), and −Mg2+ buffer with various Na+ concentrations (prepared by replacing NaCl with NMDG-Cl) was used to stimulate Mg2+ efflux.
Intracellular concentrations of free Mg2+ and Na+ ([Mg2+]i and [Na+]i, respectively) were determined from the following equation:R: the ratio of the signal intensity with 330–350 nm excitation (F1) to that with 370–390 nm excitation (F2) (R = F1/F2). Rmax: the maximum value of R. Rmin: the minimum value of R. Q: the ratio of the signal intensity with 370–390 nm excitation under minimum Mg2+ or Na+ concentration to the signal intensity with 370–390 nm excitation under maximum Mg2+ or Na+ concentration (F2min/F2max). Kd: 1.5 mM for Mag-fura2 [42] and 11.3 mM for SBFI [43], respectively. Rmin, Rmax, Fmin, Fmax were obtained after each experiment. For Mag-fura2, Rmin, Fmin were recorded by addition of 6 µM 4-Bromo-A23187 (Wako) and 10 mM EDTA, and Rmax, Fmax were recorded by incubating the cells under −Mg2+ buffer plus 6 µM 4-Bromo-A23187 and 50 mM MgCl2. For SBFI, Rmax, Fmax were recorded by incubating the cells under −Mg2+ buffer supplemented with 5 µM Gramicidin (Wako), and Rmin, Fmin were recorded by incubating the cells under the Na+-depleted buffer (a −Mg2+ buffer which NaCl is replaced with KCl) with 5 µM Gramicidin. The cells were fixed with PBS containing 3.7% formaldehyde after fluorescence measurement and subjected to immunofluorescence microscopy to confirm protein expression. Difference of [Na+]i and [Mg2+]i just after Mg2+ depletion (between time = 0 and 20 sec) was used to determine the initial velocity of Na+ influx (V0 (Na)) and Mg2+ efflux (V0 (Mg)), respectively. The ratio of CNNM4-dependent Na+ influx versus Mg2+ efflux was calculated as follows:Vmax, KA, and Hill coefficient were determined by SigrafW software [44].
HEK293 cells were transfected with expression plasmids for CNNM4, and maintained in normal culture media until use. Mg2+ loading assays were performed with the abovementioned protocol for ratiometric imaging, with following modifications. Cells were incubated in −Mg2+ buffer with 2 µM Mag-fura2-AM for 10 min, 37°C. The cells were once rinsed with −Mg2+ buffer and viewed using the same apparatuses. Then, the extracellular solution was changed to buffers with various Mg2+ and Na+ concentrations (buffer with low Na+ concentrations were prepared by replacing NaCl with NMDG-Cl) and incubated for 4 min to load Mg2+. Finally, the extracellular solution was changed to −Mg2+ buffer to stimulate Mg2+ efflux.
Maxillae dissected from 2-month-old mice were fixed with 70% ethanol for 5 days, dehydrated in ascending alcohol series, and embedded in methyl methacrylate. After embedding, cutting specimen, and surface polishing, the samples were then coated with room-temperature ionic liquid (1-butyl-3-methylimidazolium tetrafluoroborate), which work as an electric conductor and enables the observation of biological specimen by an SEM [45]. Samples were mounted with carbon adhesion tape on a specimen holder for SEM. Backscattered and secondary electron images were obtained with SEM (VE-9800: Keyence). The composition changes were analyzed with EDX (VE9800: EDAX) attached to the SEM at an accelerating voltage of 8 keV.
|
10.1371/journal.pgen.1002145 | Foxp2 Regulates Gene Networks Implicated in Neurite Outgrowth in the Developing Brain | Forkhead-box protein P2 is a transcription factor that has been associated with intriguing aspects of cognitive function in humans, non-human mammals, and song-learning birds. Heterozygous mutations of the human FOXP2 gene cause a monogenic speech and language disorder. Reduced functional dosage of the mouse version (Foxp2) causes deficient cortico-striatal synaptic plasticity and impairs motor-skill learning. Moreover, the songbird orthologue appears critically important for vocal learning. Across diverse vertebrate species, this well-conserved transcription factor is highly expressed in the developing and adult central nervous system. Very little is known about the mechanisms regulated by Foxp2 during brain development. We used an integrated functional genomics strategy to robustly define Foxp2-dependent pathways, both direct and indirect targets, in the embryonic brain. Specifically, we performed genome-wide in vivo ChIP–chip screens for Foxp2-binding and thereby identified a set of 264 high-confidence neural targets under strict, empirically derived significance thresholds. The findings, coupled to expression profiling and in situ hybridization of brain tissue from wild-type and mutant mouse embryos, strongly highlighted gene networks linked to neurite development. We followed up our genomics data with functional experiments, showing that Foxp2 impacts on neurite outgrowth in primary neurons and in neuronal cell models. Our data indicate that Foxp2 modulates neuronal network formation, by directly and indirectly regulating mRNAs involved in the development and plasticity of neuronal connections.
| Foxp2 codes for an intriguing regulatory protein that provides a window into unusual aspects of brain function in multiple species. For example, the gene is implicated in speech and language disorders in humans, song learning in songbirds, and learning of rapid movement sequences in mice. Foxp2 acts by tuning the expression levels of other genes (its downstream targets). In this study we used genome-wide techniques to comprehensively identify the major targets of Foxp2 in the embryonic brain, in order to understand its roles in fundamental biological pathways during neurodevelopment, which we followed up through functional analyses of neurons. Most notably, we found that Foxp2 directly and indirectly regulates networks of genes that alter the length and branching of neuronal projections, an important route for modulating the wiring of neural connections in the developing brain. Overall, our findings shed light on how Foxp2 directs particular features of nervous system development, helping us to build bridges between genes and complex aspects of brain function.
| Forkhead-box protein P2 is a highly conserved vertebrate protein, belonging to an important group of transcription factors [1]. By modulating the expression of downstream target genes, forkhead-box proteins influence a diverse array of processes, including cell-cycle regulation, signal transduction, differentiation, patterning and metabolism [2]. They thereby play crucial roles during embryogenesis, in postnatal development and in the mature organism, and many have been linked to disease states [3]. The P subgroup is a divergent branch of forkhead-box proteins that share a distinctive DNA-binding domain located near the C-terminal end of the protein, as well as zinc-finger/leucine-zipper motifs that mediate dimerization, and a glutamine-rich region towards the N-terminus [4], [5].
Functional evidence from multiple species implicates Forkhead-box protein P2 in particularly intriguing aspects of brain development and function [1]. (Here we adopt the standard accepted nomenclature to refer to the protein in different species: FOXP2 in humans, Foxp2 in mice, FoxP2 in other chordates, with the corresponding gene names in italics [6].) In humans, damage to one copy of the FOXP2 gene causes a rare neurodevelopmental disorder, characterised by difficulties mastering sequences of mouth movements during speech, as well as impaired language processing [4], [7], [8]. Heterozygous disruptions of the mouse orthologue (Foxp2) yield dramatic reductions in synaptic plasticity of cortico-striatal brain circuits, associated with deficits in learning of rapid motor skills [9]. Mouse pups with homozygous Foxp2 mutations show more severe neural effects – gross motor impairments, delayed postnatal maturation of the cerebellum and dramatic reductions in emission of ultrasonic vocalisations – against a background of reduced weight-gain and postnatal lethality [9]-[11]. In addition, the avian orthologue (FoxP2) is required for normal vocal learning in songbirds [12], [13]. Selective knockdown of the gene in a key striatal nucleus in juvenile zebrafinches leads to incomplete and inaccurate imitation of tutor songs [14].
Studies of both human FOXP2 and mouse Foxp2 identified similarly strong CNS (central nervous system) expression during embryogenesis, which is confined to neurons (absent from glial cells) and enriched in various brain structures, including deep layers of the developing cortical plate, and parts of the striatum, thalamus and cerebellum [15],[16]. These embryonic expression patterns appear highly concordant in the different species, and show remarkable overlaps with sites of pathology identified by neuroimaging of human children and adults carrying FOXP2 mutations [16], [17]. Neural expression of the gene continues postnatally and into adulthood [4], [15], and is also observed in certain non-neural tissues, most notably the distal alveolar lung epithelium, and the outflow tract and atrium of the cardiovascular system [18].
The above observations of well-conserved and specific CNS expression patterns [15], [16] suggest that Foxp2 is likely to have important functions in neurodevelopment. Nevertheless, as data continue to accumulate regarding its impacts on the postnatal brain [9], [11], [14], the specific roles of Foxp2 in the developing CNS remain largely elusive. One route for gaining insights into the biological processes controlled by a transcription factor is to define the regulatory networks that are directly downstream of it [1]. An efficient strategy for identifying direct targets exploits chromatin immunoprecipitation (ChIP) methods to screen the tissue of interest [19]. Two previous investigations have coupled ChIP with hybridisation to promoter microarrays (ChIP-chip) in order to uncover binding sites of FOXP2 in human foetal brain tissue [20] and in human neurons grown in culture [21]. Both screens were of limited scope – the microarrays in these studies comprised fragments from the 5′ ends of ∼5,000 loci [20], [21], representing a small percentage of the known gene promoters in the genome. Neither study combined ChIP data with large-scale expression analyses. A more recent report used mRNA expression profiling in human neuronal models transfected with different versions of FOXP2 to explore regulatory differences between the human and chimpanzee orthologues, but did not include any ChIP screening for direct targets [22].
In the present study, we performed a systematic large-scale in vivo ChIP-chip screen of the embryonic mouse brain, coupling Foxp2-ChIP to high-density arrays with oligonucleotides tiled across >17,000 promoters. We robustly established the empirical significance of our ChIP results in wild-type brains by determining the null distribution of signals generated by matched control tissue from littermates that expressed no Foxp2 protein. Under strict empirical thresholds that minimised false positive signals, we isolated a set of 264 high-confidence in vivo targets. Gene ontology (GO) analyses of the ChIP-chip data, as well as genome-wide expression profiling and in situ hybridisations of wild-type and mutant mice, converged on neurite outgrowth as one of the most prominent biological themes associated with Foxp2 function in the embryonic CNS. We went on to directly demonstrate, using neuronal cell models and primary neurons from the embryonic mouse brain, that Foxp2 alters expression of neurite-outgrowth targets and thereby influences neurite process length and branch number.
In vivo Foxp2-ChIP screening was carried out using brains harvested from embryonic mice. Experiments were performed with mice that were wild-type for Foxp2, as well as homozygous littermates that do not express any Foxp2 protein (Foxp2-S321X mutants; see Materials and Methods) [9]. The different types of sample were screened in parallel, undergoing identical experimental manipulations and data processing. In this context, the homozygous mutant mouse tissue acts as an ideal negative control [21]. Since such samples completely lack Foxp2 protein (see Figure S1 and [9]), fragments that are pulled down by Foxp2-ChIP in these cases give an unbiased empirical indication of background noise and false positive rates yielded by the procedure. Whole mouse brains from wild-type or mutant mice were harvested at embryonic day 16 (E16), corresponding to a timepoint at which particularly high levels of Foxp2 expression are observed in the developing CNS [16]. Chromatin isolated by Foxp2-ChIP was labelled and hybridised to DNA microarrays covering the promoter regions of ∼17,000 mouse transcripts (Agilent Technologies), using total input DNA as a reference sample. Each promoter on these arrays is represented by an average of twenty-five 60-mer probes spanning ∼5.5 kb upstream and ∼2.5 kb downstream of the transcription start site, allowing peak regions of binding to be precisely defined (Figure 1). Moreover, the presence of multiple probes for each promoter scattered throughout the array gives independent enrichment values within the same promoter, which aids discrimination of real biological targets from false positive events. Specifically, since the shearing process during ChIP produces overlapping fragments of chromatin, true targets should show evidence of enrichment for multiple probes across the promoter region, while promoters with only a single enriched probe are most likely to be false positive results.
In order to identify enriched promoters, Foxp2-ChIP data were analysed as per Materials and Methods. Briefly, array data from independent biological replicates (three independent ChIP experiments hybridised to one each of three array sets) were normalised for each genotype (wild-type or mutant control) separately. Normalised array data (excluding probes with a negative average enrichment across replicate experiments) were subjected to a sliding window analysis, using a similar method to that employed in genome-wide ChIP-chip studies of other forkhead transcription factors [23]. Each probe was assigned a value (window-adjusted score) based on the median fold enrichment of itself and its neighbouring probe on either side (within 500 bp upstream and 500 bp downstream), and then probes were ranked based on this window score.
By analysing the distribution of window scores observed in the mutant null control experiments we were able to derive an empirical threshold for significance, which could then be applied to the wild-type data. We found that window scores greater than or equal to 0.974 (corresponding to ∼2-fold enrichment) excluded 99% of the data-points in the mutant null control experiments. When we applied this threshold to data from wild-type experiments, we identified a set of 1,217 promoter regions that were consistently enriched by Foxp2-ChIP over 3 replicates in wild-type mouse brains (Table S1). On inspection of the locations of the enriched probes throughout the mouse genome, no positional bias was observed (Figure S2). Since some of the enriched regions lay close to the transcriptional start site (TSS) of more than one gene, the 1,217 promoter regions corresponded to 1,253 genes. Of note, using the same analysis parameters, only 147 genes were enriched in the mutant null controls, suggesting a low false discovery rate. Nevertheless, in order to minimize false-positive findings, we excluded any enriched genes from the wild-type dataset that also had window scores exceeding the 99% threshold in the mutant control dataset. This filtering process yielded a slightly smaller set of 1,164 putative targets (Table S2).
When we applied stricter thresholds to the wild-type data, selecting only those promoters in which at least one probe gave a window score of ≥ 1.5, we identified a shortlist of 259 promoter regions. Since a small number of peak regions lay directly between the TSSs of two different genes, these 259 promoters corresponded to a slightly higher total of 266 genes. Crucially, the same analyses of the entire mutant null control dataset identified only a single gene in the genome with a window score of ≥ 1.5 (the Pigt gene), indicating an extremely low rate of false positives under these stricter selection criteria. We excluded two genes from the strict wild-type shortlist (Pigt and Zfp496) since they contained probes that exceeded the 99% threshold (i.e. window score of >0.974) in mutant null controls (Figure S3). The outcome of these analyses was a final curated shortlist of 264 high-confidence in vivo targets (Table S3).
Given that DNA is sheared randomly during the ChIP process, we would expect a true Foxp2 binding event to be represented by a peak of enrichment at a target promoter. This peak would result from the sheared DNA forming a series of overlapping fragments, with the region closest to the binding site showing the highest degree of enrichment (i.e. highest number of fragments pulled down during immunoprecipitation) and with progressively less enrichment observed as the distance to the binding site increases on either side. Figure 1A gives typical examples of the enrichment peaks observed for putative targets from our Foxp2-ChIP dataset. Examination of corresponding data from mutant control experiments emphasises the relative lack of enrichment in nulls that lack Foxp2 protein, indicating that the enrichment in wild-type samples results from highly specific Foxp2-mediated interactions. Furthermore, we followed up a subset of candidates with qPCR, consistently confirming their enrichment (Figure 1B).
Enriched regions represented in the shortlist of high-confidence targets were assessed in silico for any over-represented sequence motifs (see Text S1). This analysis did not enforce a priori conditions of motif sequence, other than a length restriction of 8 bases. This meant that rather than limiting our search to occurrences of known patterns in the promoters, we obtained an unbiased list of motifs that were characteristic of the Foxp2-ChIP target promoters. Eight sequences (motifs A-H) were found to be significantly over-represented (p<0.05) in the shortlist of high confidence target promoter sequences (Table 1). Importantly, the three most commonly identified over-represented motifs from this unbiased search (A–C) were partial or complete matches to well established FOX/FOXP/FOXP2 binding motifs (RYMAAYA/TATTTRT/AATTTGT), providing additional strong support for the biological relevance of our findings. A further over-represented motif (motif D) did not match the known consensus motifs and was detected in 182 promoters out of the 247 promoter regions that could be surveyed from the Foxp2-ChIP shortlist (See Text S1; Figure S4A). Thus, we reasoned that it may represent a novel putative Foxp2 binding sequence. EMSA (Electrophoretic Mobility Shift Assay) experiments demonstrated strong specific binding of FOXP2 to this motif (Figure S4B), when located in putative Foxp2 target promoter sequences, such as those for Nrn1, Nfat5 and Sema6d. However, not all occurrences of this motif were strongly bound by Foxp2, suggesting that while the site is capable of being bound by Foxp2 protein, the binding is context specific – as is regularly seen for other FOX family binding sites [24].
In addition to the use of in vivo ChIP to uncover target genes that are directly bound by Foxp2 (direct targets), we assessed regulatory cascades further downstream (indirect targets) via an expression profiling approach. Again we focused on E16 mouse brain tissue, analysing the same genotypes (wild-type mice and their homozygous Foxp2-S321X littermates, 5 and 6 biological replicates, respectively) on the same genomic background as used for the ChIP experiments. While ChIP identifies DNA-binding events of Foxp2-positive cells, expression profiling is expected to be more sensitive to tissue heterogeneity. Therefore we selected a key site of high Foxp2 expression with considerable prior evidence of functional relevance [9], [14]–[17], the ganglionic eminences (developing striatum and pallidum). Analysis of genome-wide expression data (see Materials and Methods for details) identified 340 genes that were differentially expressed (p<0.01) between wild-type and Foxp2-S321X homozygous mutant samples (Table S4). 180 of these genes showed reduced expression in absence of Foxp2 protein, while the remaining 160 genes showed increases (Table S4). Of these 340 genes, 19 genes (5.6%) were found in common with the ChIP-chip target gene list (Table S5), including those with known CNS functions, such as Nell2 (neural epidermal growth factor-like like 2), Nrn1 (neuritin), Cck (cholecystokinin), and Alcam (activated leukocyte cell adhesion molecule). Notably, the human orthologues of Nrn1 and Cck have been independently proposed as top direct targets in small-scale ChIP screens of human foetal tissue [20].
We went on to determine whether any biological themes were over-represented within the direct targets (promoter bound by Foxp2) and indirect targets (not bound by, but regulated downstream of Foxp2), using unbiased GO analyses. The Foxp2-ChIP and expression profiling datasets were each assessed independently using the WebGestalt program [25], identifying functional categories that were significantly enriched (Figure 2 and Figure S5). In the Foxp2-ChIP dataset we observed significant over-representation of genes involved in processes including cell motility and migration, chromatin architecture and assembly, synaptic transmission, and a number of categories associated with RNA metabolism such as regulation of RNA stability and mRNA processing. In the expression profiling dataset significant categories included regulation of transcription, actin cytoskeleton organisation and biogenesis, and cellular protein catabolism. Consistent with previous studies [20], [21], nervous system development, neurogenesis and multiple G-protein signalling categories — including G-protein coupled receptor signalling (ChIP), and G-protein signalling and Wnt receptor signalling (expression) — were significant in both datasets.
We next performed in situ hybridisation on brains from wild-type and Foxp2-S321X E16 embryos, to further assess major targets suggested by the ChIP and expression profiling screens. Consistent with previously published data [16], in addition to the developing striatum, Foxp2 expression at this developmental stage is highest in the diencephalon (developing thalamus), midbrain and cerebellar primordium (Figure 3). The in situ hybridisation data confirmed regulation of Shhrs (also known as Dlx6as1 or Evf1/2), a transcript showing greater than 200-fold increased expression levels in S321X mice. This noncoding RNA is highly specific to the ganglionic eminences in the embryo and is known to play a vital role in the control of the homeodomain transcription factors Dlx5 and Dlx6 [26], [27]. These data illustrate that loss of Foxp2 can influence transcripts central to key neurodevelopmental processes in vivo. We then focused on target genes common to both ChIP and expression profiling datasets, to determine whether expression changes could be observed, not only in the ganglionic eminences, but also elsewhere in the developing brain (Figure 3). Indeed, Nell2, Nrn1 and Cck all demonstrated clear increases in expression in the developing basal ganglia at E16 in the Foxp2 mutant compared to wild-type, in agreement with the array data (Figure 3 and Figure S5), providing further evidence that they are indeed direct targets, repressed by Foxp2. Significantly, Nrn1, a gene important for neuronal outgrowth [28], showed strongly increased expression in mutants in additional regions where Foxp2 is typically expressed, including the developing thalamus and cerebellum (Figure 3). Similarly, Cck shows additional increases in expression in the cerebellum (Figure 3).
Certain putative direct Foxp2 targets with known roles in the CNS, such as Ywhah (tyrosine 3-monooxygenase/tryptophan 5-monooxygenase activation protein, eta polypeptide) and Wasf1 (WASP family 1), are ubiquitously expressed in the developing mouse brain [29]. However, other genes are thought to have more localised and/or temporally defined patterns of expression. To ascertain whether these targets are co-expressed with Foxp2, further in situ hybridisation was carried out at E16. The results demonstrate that genes from the ChIP dataset with established CNS functions are found in a range of Foxp2-expressing brain regions (Figure S6).
The GO analyses of independent Foxp2 target data from genome-wide ChIP and expression profiling uncovered a consistent and statistically significant over-representation of genes involved in neurite development & morphogenesis, axon extension and axon guidance pathways (Table 2). Multiple GO categories associated with such processes were significantly over-represented in at least one of the datasets, and several of these functional classes were significant in both datasets including neurogenesis, neuron projection development and axonogenesis (Figure 2 and Table 2). Furthermore, when we investigated KEGG pathways associated with these datasets, we observed enrichment of genes in a number of pathways (Figure 4A and Table S6), one of the most significant of which was the axon guidance pathway (p = 4.73×10−8 and p = 3.00×10−4 in Foxp2-ChIP and expression profiling datasets, respectively). Interestingly, a number of different but interacting genes within this pathway were identified in the two datasets (Figure 4B), suggesting that direct and indirect targets may represent different aspects of the same functional downstream effects of Foxp2-mediated regulation. In sum, our unbiased genomic screens for Foxp2-dependent gene networks predicted that neurite outgrowth and axon guidance should be key biological themes associated with Foxp2 function in the developing brain.
We went on to test this prediction through genetic and functional analyses of neuronal cell models and primary neurons. First, we independently assessed whether differences in Foxp2 expression affect expression of putative direct target genes involved in neurite outgrowth, using a well-validated murine cell model. Neuro2a cells are neuron-derived cells that can be differentiated to take on a more neuron-like identity via exposure to retinoic acid. These cells were stably transfected with Foxp2 or with an empty vector control, and then grown in media either with or without the addition of retinoic acid. Cells that were over-expressing Foxp2 demonstrated consistent expression changes for multiple direct target genes that were identified in our E16 ChIP screen and implicated in neurite outgrowth pathways (Figure 5A). Significant repression of target gene expression was observed both pre- and post-differentiation; however most of the neurite-outgrowth genes showed greater fold changes following differentiation. Next, we formally assessed the hypothesis that changes in Foxp2 levels, and the concomitant alterations in expression of neurite-outgrowth genes, yield detectable differences in the growth of neurites in these cells. After 24 and 48 hours of retinoic acid treatment (matching the timepoints for analyses target gene expression) we performed blind scoring of cells and observed striking qualitative differences in neurite outgrowth when cells over-expressed Foxp2, as compared to sham-transfected controls. Cells that had been transfected with Foxp2 prior to differentiation consistently displayed increased neurite length, in a manner that was easy to distinguish from controls (Figure 5B).
To further assess the in vivo relevance of these findings, we examined whether there were corresponding phenotypic effects mediated by functional Foxp2 in neurons of the developing brain. We isolated primary neurons from the ganglionic eminences of E16 mouse brains, matching the region and timepoint used for our original target screening. Here, we aimed to directly test whether the Foxp2-positive neurons derived from the developing basal ganglia show altered neurite outgrowth when the gene is mutated. The assay was facilitated by availability of a mouse model (Foxp2-R552H) in which the protein is expressed at normal levels, but is nevertheless dysfunctional [9]. R552H mice recapitulate an aetiological mutation that causes speech and language deficits in a large human family. This change yields a substitution in the DNA-binding domain which severely impairs the transcription factor function of the mutant protein [30], such that the overall phenotype of homozygous R552H animals is very similar to that observed for mice which completely lack Foxp2 [9]–[11]. However, unlike the Foxp2-null mice, R552H homozygotes still express detectable levels of the protein, allowing us to clearly identify Foxp2-positive cells in our primary cultures via antibody staining. This represents an important measure, given the heterogeneous nature of the dissected material used to generate the primary culture.
We again observed obvious differences in neurite outgrowth associated with presence of functional Foxp2 (Figure 6A). A blinded analysis revealed statistically significant changes in quantitative measures of neurite outgrowth for Foxp2-expressing neurons from wild-type embryos as compared to those from homozygous Foxp2-R552H littermates (Figure 6B). In particular, the latter showed significant reductions in total outgrowth (p = 0.001); mean (p<0.001), median (p = 0.008) and maximum process length (p<0.001); and average number of branches (p = 0.003). Thus, loss of Foxp2 function in striatal neurons that normally express this transcription factor yields significant reductions in multiple indices of neurite outgrowth. When Foxp2-negative cells from the wild-type cultures were compared to equivalent cells from mutants (Figure 6C), it was only the total outgrowth that met significance (p = 0.013).
These findings are strongly in agreement with differences in levels of Foxp2 expression, neurite outgrowth and correlated physiological properties between the two major subpopulations of striatal medium spiny neurons (MSNs) in vivo. While both striatonigral (Drd1a) and striatopallidal (Drd2) MSNs continue to increase their dendritic area well into adulthood, Drd1a MSNs develop significantly more dendrites [31]. This dichotomy in dendritic growth contributes to key physiological differences between both MSN populations, although the underlying mechanisms remain unknown [31]. Furthermore, studies of cultured striatal neurons demonstrate that Drd1a MSNs have larger dendritic trees than Drd2 MSNs, invoking intrinsic mechanisms [31]. To study whether these intrinsic differences in dendritic growth correlate with Foxp2 expression levels, we investigated mice expressing enhanced green fluorescent protein (EGFP) either mainly in Drd1a or Drd2 MSNs [32]. We found that Foxp2 shows consistently high expression in striatonigral Drd1a MSNs and very low expression in Drd2 MSNs throughout the striatum (Figure S7), further supporting roles for Foxp2 in neurite outgrowth.
Although early studies of Foxp2 orthologues in multiple species suggested that it may play crucial roles in neurodevelopment [15], [16], the exact nature of such roles has not been established. Indeed, much of the existing knowledge regarding neuronal functions of this transcription factor instead concerns its impacts on the postnatal CNS [9], [14]. In the present study we employed a high-throughput functional genomic strategy to shed new light on the in vivo activities of Foxp2-dependent pathways in the developing CNS.
Of note, among the biological themes that we identified, our comprehensive ChIP-chip and expression profiling in midgestation brain tissue independently and consistently highlighted gene networks underlying neurite development and morphogenesis, axon extension and axon guidance. These findings drove us to specifically assess the impact of the Foxp2 gene on neurite outgrowth phenotypes in genetically manipulated neuronal cell models and primary neurons from embryos of mutant mice. Our functional experiments confirmed regulation of the highlighted gene networks and indicated that wild-type Foxp2 thus enhances multiple facets of neurite development in vivo, including outgrowth process length and branch number. The data suggest that the mode of action may be predominantly cell autonomous, since the functional effects were mainly restricted to the subset of Foxp2-expressing cells within a mixed population of neurons. This possibility of cell-autonomous effects is an interesting hypothesis that could be clarified in further studies.
Our neurite outgrowth findings are in line with new evidence regarding the functional impact of evolutionary differences between FOXP2 orthologues [33]. For example, it is known that this transcription factor underwent two amino-acid substitutions on the human lineage after splitting from the chimpanzee lineage, leading to speculation that such changes may have been important for evolution of spoken language. In a recent study, researchers inserted the relevant substitutions into the endogenous Foxp2 gene of mice, and observed that striatal neurons had significantly longer dendrites and increased synaptic plasticity [33]. By contrast, we have shown that mice with loss of function of Foxp2 have statistically significant reductions in neurite outgrowth (Figure 6 in the present paper) and decreased synaptic plasticity [9].
Furthermore, the identification of potential regulatory links between Foxp2 and neural connectivity may be informative for wider discussions regarding the evolution of vocal learning [34]. Auditory-guided vocal learning is a rare trait that is only found in a small number of animal groups; the best understood examples include speech acquisition in humans and learning of song by certain bird species. As noted above, while human FOXP2 has been implicated in speech abilities [4], [7], [8], avian FoxP2 is required for normal song-learning in songbirds [12], [13], supporting the view that this is a molecule with broader relevance for vocal-learning in multiple species. Intriguingly, it has been independently proposed that specific changes in patterns of neural connectivity in the brains of vocal learners account for the differences in their speech/song behaviours relative to other closely-related species that lack such abilities [34]–[36]. Perhaps evolutionary differences in FoxP2 orthologues may contribute to altered patterns of connectivity in the different species, and thereby help to explain differential capacities for vocal learning. Since we did not assess the impact of evolutionary changes in the present study, this remains an open question for future investigation using comparative functional genomics.
To our knowledge, the current report represents the first large-scale in vivo characterisation of direct and indirect Foxp2 targets in the embryonic brain. It is of interest to consider how the present findings relate to published screens that used more limited ChIP surveys [19]–[21], or that employed expression profiling [22], [33]. The extent of direct overlap with previous datasets is complicated by three confounding factors. First, there are differences in scope of screening; the prior ChIP-chip investigations only queried a small subset of known promoters [20], [21]. Second, there are differences in species under investigation. Previous target screens largely focused on human and/or chimpanzee FOXP2, and the differences between the two [19]–[22], [33], while here we chose to comprehensively define the pathways regulated by murine Foxp2. Mouse models offer considerable advantages for functional genomics, and careful integration of murine data with those from other species will enhance our understanding of evolutionary roles of this gene. Finally, the majority of earlier studies screened neuron-like cells grown in culture [19]–[21], and no investigation of this transcription factor has reported integrated use of genome-wide ChIP and expression profiling to screen the same tissue.
Nevertheless, many important consistencies are observed between the different datasets, particularly in the biological themes and processes that they implicate. For example neurite outgrowth pathways and synaptic plasticity are over-represented in all FoxP2 ChIP-chip datasets across different species and neuronal cell-type, in vitro and in vivo [20], [21]. These processes are closely related during the development of neuronal networks. Genes controlling neurite outgrowth or axon guidance during early development have crucial roles in maturation and stabilisation of synaptic connectivity at later stages and eventually in activity-dependent synaptic plasticity in the mature brain throughout life (such as neurotrophins, semaphorins and ephrins) [37], [38]. Hence, the strong impact of Foxp2 on neurite outgrowth during one particular stage at E16 might even reflect major Foxp2 functions that are relevant throughout the development and maintenance of neuronal networks. A case in point is provided by our data demonstrating that Nrn1 is a highly robust downstream target. The Nrn1 gene encodes neuritin, which is already expressed at embryonic stages of development and was initially identified as a downstream effector of neuronal activity and neurotrophin-induced neurite outgrowth [28]. Nrn1 not only showed one of the strongest enrichment signals in our in vivo ChIP experiments, but was independently detected as a target in our systematic expression profiling experiments of equivalent tissue and by in situ hybridisation – the corresponding human homologue was also one of the top direct targets reported in a small-scale ChIP screen of human foetal brain tissue [20].
A number of additional genes, which overlap with earlier studies, merit further comment. The Cck gene, which showed convergent evidence in our embryonic ChIP experiments, expression profiling screens and in situ hybridisation analyses, was reported as a direct target in both prior published human ChIP-chip studies [20], [21]. Lmo4 (Lim domain only 4) was found to be indirectly downregulated by Foxp2 in our analyses of embryonic brain tissue (Table S4) and the human orthologue LMO4 was similarly repressed by FOXP2 in previous expression profiling studies of human neuron-like cells by Konopka and colleagues [22]. Interestingly, in that earlier study using cellular models, this indirect target was repressed both by human and chimpanzee versions of FOXP2 [22]. LMO4 encodes a transcription factor that plays important roles in cortical patterning, and is one of the few genes known to show asymmetric expression in the embryonic human brain [39]. Efnb2 (Ephrin-b2), a well-validated direct target (Figure 1, Figure 4, Figure 5) was identified in the Konopka et al. study as one of a small number of genes that may be differentially regulated by human and chimpanzee FOXP2 orthologues [22]. This gene is of particular interest since it is implicated in neurite outgrowth and axon guidance (and also synaptic plasticity) in the basal ganglia and related brain structures [40]. In addition, Nell2, a validated ChIP and expression array target (Figure 3), has also been linked to neurite outgrowth [41], and has recently been shown to promote neuronal survival by trans-activation by estrogen [42].
Given the substantially enhanced scope of ChIP screening in the present study, we were able to identify many interesting novel targets that could not be isolated in the earlier work. For example, our high-confidence shortlist of direct targets includes Pak3 – a downstream effector of the Rho family of GTPases which plays critical roles in pathways restraining neurite growth [43]; Nptn (neuroplastin) – encoding a synaptic glycoprotein involved both in development/maintenance of synaptic connections [44] and in long-term plasticity [45]; Wasf1 – a gene that regulates activity-induced changes in dendritic spine morphogenesis [46] and is involved in actin remodelling during axon growth [47]; the neuronal semaphorins Sema4f [48] and Sema6d [49]; as well as Ywhah (also known as 14-3-3), which encodes an adapter protein implicated in presynaptic plasticity [50] (Figure 1, Figure 4, Figure 5; Table S3). Although the screening tissue was embryonic brain, many of the relevant genes have functions that go beyond this to also influence neural plasticity at later stages. Overall, this dataset will be important for directing follow-up studies of Foxp2-dependent pathways and assessing their involvement in traits such as acquisition of motor-skills [9], vocal learning [14], and spoken language [1]. While it is likely to be an indirect target of Foxp2 regulation, it is noteworthy that Evf1/2 (Shhrs) showed such highly increased expression in Foxp2-S321X mice. It has been shown that the Evf2 RNA molecule co-operates with the Dlx2 protein to activate the Dlx5/6 enhancer element [27]. Thus it is interesting that both the DLX1/2 and DLX5/6 loci have been implicated in autism via independent studies, including a common polymorphism in the DLX5/6 enhancer itself [51]–[53].
Of 340 genes showing differential expression (p<0.01) between mutant and wild-type ganglionic eminences, only 19 (∼5%) corresponded to putative direct targets of Foxp2 from the ChIP-chip screens. Thus, most of the expression differences observed in the transcriptional profiling experiments are unlikely to represent direct modulation due to Foxp2 binding, but could instead represent cascade effects further downstream (i.e. loss of Foxp2 directly alters expression of a relatively small subset of genes, which in turn indirectly affect many others). Discrepancies between the ChIP-chip and expression profiling datasets may also result from our experimental design: the former could potentially detect binding events of Foxp2-expressing neurons anywhere in the brain, while the latter was targeted specifically at the ganglionic eminences, a region showing particularly high Foxp2 levels. Foxp2 target genes that are not expressed in this structure could therefore be observed in the ChIP study, but would not be detected in the expression analysis. An example of such a target is Sema3a. The promoter of this gene was bound by Foxp2 in our ChIP study (Figure 1), but its expression only overlaps with Foxp2 expression in the cerebellum (Figure S6). Nevertheless, it is not unusual in studies of transcription factor function to observe substantial differences between promoter occupancy maps and transcriptional profiling data. It is well established that transcription factors can be poised ready at particular genomic sites, awaiting important co-factors, before modulating expression of the relevant targets [2], [54], [55].
The present investigation queried the vast majority of known promoters in the genome, but we acknowledge that the screening strategy is unable to uncover potential regulatory sequences that lie outside classical promoter regions. In earlier work, based on low-throughput shotgun sequencing of human FOXP2-ChIP fragments, we identified a FOXP2-bound element in the first intron of CNTNAP2 (contactin-associated-protein-like-2) a gene implicated in language impairments and autism [19]. Although the mouse genome contains an orthologous region to the human FOXP2-bound regulatory element of CNTNAP2, this was not represented on the arrays used in this study, and hence it escaped detection. When we carried out ChIP-PCR experiments using the same mouse embryonic brain tissue as used for ChIP-chip we demonstrated clear Foxp2 occupancy of the orthologous region in mouse Cntnap2. Specific enrichment was observed in the wild-type brains; while no enrichment was found in equivalent tissue from the mutant null controls (see Figure S8 and Table S7). Studies are now underway using ‘ChIP-seq’ techniques (coupling ChIP to next-generation-sequencing) to allow a fully unbiased view of FOXP2/Foxp2 binding throughout the genome.
Among the validated direct targets of Foxp2 identified in our study there were a number of microRNA (miRNA) molecules, including mir-124a and mir-137. miRNAs are an extensive class of short (∼18–23 nucleotide) noncoding molecules which provide extra layers of dynamic control in networks of gene expression [56]. miRNAs are abundant in the brain and implicated in critical aspects of nervous system development and function, ranging from early neurogenesis and proliferation [57], through neural differentiation and dendrite morphogenesis [58], to adaptive mechanisms in mature neurons, including learning and memory [59]. They play pivotal roles in processes such as neurite outgrowth, axonal pathfinding and synaptic plasticity, mechanisms for which localised rapid control of protein synthesis is paramount [58], [59].
In conclusion, the use of in vivo genomic screening strategies in the developing embryonic brain has proved to be a powerful approach for understanding the biology of Foxp2, one of the most intriguing transcription factors of the CNS. This starting point led us to functional characterisation of new mechanisms of Foxp2 action, in particular the modulation of networks involved in neurite outgrowth, axonogenesis and other core aspects of neural development. Future studies will define how these regulatory networks differ between distinct species, what role miRNAs play in Foxp2-related pathways and phenotypes and will investigate whether it is possible to rescue the established neurobiological effects associated with loss of Foxp2 function, through manipulation of key targets. Ultimately, such work promises to fully uncover the functional pathways that connect Foxp2 with plasticity of the developing CNS.
In vivo Foxp2-ChIP in embryonic mouse brain tissue was performed according to the protocol previously described by Vernes and colleagues [21]. Each of the three replicates included whole brain tissue (from the telencephalon to the brain stem at the level of the foramen magnum) at E16 (embryonic day 16), a developmental timepoint of high Foxp2 expression [16], pooled from 5–6 mice of matching genotype. Experiments were carried out either with wild-type embryos, or homozygous Foxp2-S321X mutants as negative controls. S321X mutants carry an early nonsense mutation that disrupts Foxp2; the resulting combination of nonsense-mediated RNA decay and protein instability leads to a complete lack of detectable Foxp2 protein in the brain [9]. The wild-type embryos and mutant controls used in these experiments were all matched littermates, backcrossed for at least ten generations into a C57BL/6J strain, maximizing the homogeneity of the genomic background. Although homozygous mutants display developmental delays and reduced cerebellar growth after birth, they show no gross anomalies in brain anatomy or development during embryogenesis [9]. All animal work was carried out conforming to the regulatory standards of the UK Home Office, under Project Licence 30/2016.
E16 mouse brains were extracted, snap frozen in liquid nitrogen and stored at −80°C until use. Each whole brain was weighed, then chopped finely with a razor on ice. Brains were pooled to achieve a total weight of between 0.3 and 0.5 g of tissue (between 5–6 brains per replicate) and resuspended in 5 ml PBS. A 1/10 volume (500 µl) of cross-linking buffer was added prior to 15 minutes incubation with agitation at room temperature. Formaldehyde was quenched via the addition of 125 mM glycine. Cross-linked tissue was washed in PBS before brief mechanical homogenisation. Pellets were then incubated in two in vivo ChIP lysis buffers at room temperature for ten minutes each: Buffer 1 (50 mM HEPES-KOH pH = 7.5, 140 mM NaCl, 1 mM EDTA, 10% glycerol, 0.5% NP-40, 0.25% Triton X-100, protease inhibitors); Buffer 2 (200 mM NaCl, 1 mM EDTA, 0.5 mM EGTA, 10 mM Tris pH = 8, protease inhibitors). After collection via centrifugation, nuclei were resuspended in 5 ml sonication buffer (10 mM Tris-HCl pH = 8, 1 mM EDTA, 0.5 mM EGTA, protease inhibitors). Samples underwent 15 rounds of 20-second sonication pulses at 30% power, with 60 seconds on ice between each round (Branson Digital Sonifier - S450D). Agarose gel electrophoresis was used to confirm that fragment size was 300–1000 bp. Cells were centrifuged at 10,000 g and 4°C for 10 minutes to remove cell debris. 10 µg of polyclonal rabbit anti-Foxp2 antibody (C-terminal antibody, Geschwind Laboratory, UCLA) [20] pre-coupled to 100 µl Dynal M-280 rat anti-rabbit IgG magnetic protein-A beads was added and incubated at 4°C, rotating overnight. Beads were washed five times in RIPA buffer and once in TE buffer. Chromatin was eluted from beads in TE buffer with 1% SDS at 65°C for 10 minutes with agitation. The chromatin was then incubated at 65°C overnight to reverse cross-links. Purified chromatin was amplified via Ligation Mediated PCR (LMPCR) according to published protocols [60]. Size and purity of DNA was assessed via spectrophotometry and gel electrophoresis.
2 µg of amplified immunoprecipitated chromatin, or total input DNA was fluorescently labelled with Cy5 and Cy3 respectively using random primers provided in the BioPrime DNA labelling system (Invitrogen). The labelling reaction was allowed to proceed for 16 hours at 37°C, before purification by sodium acetate precipitation. Hybridisation to mouse promoter arrays (Agilent Technologies, #G4490A) was carried out by the UCLA microarray core facility, according to the manufacturer's instructions. Arrays consisted of 60-mer oligonucleotides spanning ∼8 kb (5.5 kb upstream and 2.5 kb downstream of TSS) at each of ∼17,000 mouse promoter regions. Probes were spaced on average, between 100–300 bp apart, with approximately 25 probes for each promoter region. Three littermate matched sets of pooled wild-type or mutant control chromatin samples were applied to microarrays, each using its respective input DNA sample as the internal reference on the array. Thus, the three wild-type and three mutant control datasets represent signals obtained from a total of 34 individual mouse embryos.
Array images were scanned using the Axon GenePix 4000B. Data were retrieved and initial quality control carried out using the Axon GenePix 4000B software package. All promoter coordinates and probes were mapped with reference to the NCBI m36 mouse assembly. Microarray data analysis was carried out using the mArray package for R [61]. LOESS normalisation and background correction was performed within each array. Data were normalised between arrays using quantile normalisation, and mean values were calculated from three biological replicates (wild-type or mutant control experiments) for each probe - called ‘probe scores’, such that a score of 1 corresponds to 2-fold enrichment in ChIP versus total input DNA. All negative probe scores were assigned a value of zero. A ‘window-adjusted score’ for each probe was then calculated as the median value of each probe score and its nearest neighbour on either side. Neighbouring probes were only considered if they fell within 500 bp upstream or 500 bp downstream of the central probe. This window size was based on the average size of the labelled DNA fragments, estimated to be approx 1000 bp. Thus, a true binding event would likely be indicated by positive scores of multiple neighbouring probes within a 1000 bp window. In cases where there were less than three probes located within this 1000 bp window showing a signal greater than background then the window-adjusted score was set to zero. This process helps to guard against artificial skewing of enrichment values at edges of promoter regions.
The use of mutant null controls enabled us to robustly assess the empirical significance of wild-type ChIP results. The data from the mutant control experiments were used to estimate a null distribution of window scores; that is, the non-specific signals produced by the Foxp2-ChIP protocol even when there is no Foxp2 protein available for pulldown. (Note that a subset of binding events in mutant null controls could potentially be due to crossreactivity of Foxp2 antibodies with closely related proteins, such as Foxp1 or Foxp4, that may bind to the same promoter. However, prior work with the antibody used here suggests that levels of crossreactivity are extremely low [20].) From this null distribution of window scores we calculated the threshold which excluded 99% of all datapoints in controls. This threshold could then be applied to the wild-type array data.
Chromatin isolated during Foxp2-ChIP in mutant and wild-type mouse brains was amplified using a semi-quantitative PCR technique, as described previously [21], using primers directed towards the peak regions of enrichment (Table 3). The β-actin housekeeping gene promoter was used as a negative control.
The ganglionic eminences, sites of particularly high embryonic Foxp2 expression [16], were dissected from E16 brains of six wild-type mice and six homozygous Foxp2-S321X mutant littermates. For each embryo, the left- and right-hemisphere ganglionic eminences were pooled in TRIzol reagent and RNA was extracted using the QIAGEN RNeasy kit, according to the manufacturer's instructions. RNA yield was measured using a NanoDrop ND-1000 spectrophotometer (NanoDrop Technologies, Wilmington, DE), and its quality was assessed using RNA6000 Nano Assays on an Agilent Bioanalyzer 2100 (Agilent Technologies, Santa Clara, CA).
Gene expression profiling was performed using whole-genome mouse BeadChip arrays from Illumina (San Diego, California, USA), which include 45,281 probes representing 31,492 mouse transcripts. In brief, 500 ng of total RNA was reverse transcribed to synthesize first- and second-strand complementary DNA (cDNA). Following purification on spin columns, in vitro transcription was used to synthesize biotin-labelled complementary RNA (cRNA). 1500 ng of biotin-labelled cRNA was hybridized to Mouse WG-6 V2 Expression BeadChip arrays (Illumina Inc., San Diego, CA) at 55°C for 18 h. The hybridized arrays were washed and labelled with streptavidin-Cy3 according to the manufacturer's protocols before being scanned with the Illumina Bead Array Scanner. Raw data were exported from the Illumina BeadStudio software (v3.4.0) for further processing and analysis using the R statistical software [62] and BioConductor packages [63]. Signal data and detection scores were extracted for each of the 12 samples. Signal data were background corrected by subtracting the average signal from the negative control probes on each array, prior to being transformed and normalised using the ‘VSN’ package [64]. Quality control analyses, including hierarchical clustering and principal component analysis (PCA), identified one outlier sample (from the wild-type group). This sample had very low signal compared to other samples while hybridisation and labelling metrics were normal, suggesting a sample problem rather than a technical issue. It was sufficiently outlying to remove from further analysis and the remaining 11 samples were re-normalised. The dataset was then filtered to remove probes that were not detected (detection score <0.95 in all samples), resulting in a final dataset of 24,479 probes.
Statistical analysis was performed using the ‘Linear Models for Microarray Analysis’ (Limma) package [65]. Differential expression between mutant and wild-type animals was assessed using a linear model that included effects for genotype and litter. Raw p-values were corrected for multiple testing using the false discovery rate (FDR) controlling procedure of Benjamini and Hochberg [66]. Fourteen probes were significant at a FDR of 5%. The larger set of 340 probes significant at p<0.01 was used for further biological investigation. We performed permutation tests on the genotype labels (11 choose 5), taking litter effects into account, and found that ≥340 genes were differentially expressed at p = 0.01 in only 9 out of the possible 108 permutations (∼8%). Gene annotation was added to the final probe list using the relevant annotation file (MouseWG-6_V2_0_R0_11278593_A.txt) from the Illumina website (http://www.illumina.com).
Neuro2a (murine neuroblastoma) cells were cultured in ‘Growth media’: Modified Eagles Medium (MEM) (Sigma) supplemented with 10% Foetal Calf Serum (Sigma), 2 mM L-glutamine (Sigma) and 2 mM Penicillin/Streptomycin (Sigma). Cells were grown at 37°C in the presence of 5% CO2. Stable Neuro2a cell-lines overexpressing Foxp2 protein or non-expressing controls, were generated via transfection with pcDNA3.1/Foxp2 (mouse isoform I - untagged) or the empty vector, using Genejuice (Novagen) according to the manufacturers' instructions. Cells were cultured in complete medium supplemented with 500 µg/ml G418 (Calbiochem) as a selective agent. Resistant single colonies were isolated 20 days after transfection, then cultured and expanded independently in the presence of G418 (500 µg/ml). Expression of recombinant Foxp2 was confirmed using qRT-PCR and Western blotting with two polyclonal antibodies recognizing different epitopes of the protein (goat N-terminal antibody, Santa Cruz Biotechnology [30]; rabbit C-terminal antibody, Geschwind Laboratory, UCLA [20]). Three Foxp2-transfected clones with a high and consistent level of expression and three empty vector clones were chosen for use in further experiments. Neuro2a cells were differentiated via the addition of Modified Eagles Medium supplemented with 2% Foetal Calf Serum (Sigma), 2 mM L-glutamine (Sigma), 2 mM Penicillin/Streptomycin (Sigma) and 20 µM all-trans retinoic acid (‘Differentiation media’).
RNA was extracted from three independent clones of Neuro2a cells stably transfected either with murine Foxp2 or the empty control vector following culture in growth media or differentiation media (for 24 or 48 hours). Total RNA was extracted from cells harvested in TRIzol reagent using the RNeasy kit (QIAGEN) according to manufacturers' instructions. Reverse transcription was performed as described previously [21]. Small molecular weight RNA was harvested using the Purelink miRNA isolation kit (Invitrogen) according to manufacturers' instructions. In order to assess miRNA expression levels, the small molecular weight RNA was polyadenylated prior to reverse transcription using the NCode miRNA First-strand cDNA synthesis kit (Invitrogen), as per the manufacturers' protocol.
PCR reactions utilised SYBR Green supermix (BioRad) as described [21]. Primers specific for candidate genes and the control housekeeping genes GAPDH/Gapdh (glyceraldehyde 3-phosphate dehydrogenase) and U6 (small nuclear RNA) were designed using PrimerBank [67] (Table 4). Quantitative PCR reactions were performed on the iQ5 thermal cycler real-time PCR detection system (BioRad) according to manufacturers' instructions. Melting curve analysis was performed to assess the specificity of the amplification. Data analysis was performed using iCycler software (BioRad), and quantification was via the comparative CT method [68]. Fold changes are reported in response to Foxp2 expression compared to cells transfected with an empty vector, following normalisation to an internal control, the GAPDH housekeeping gene (for gene expression) or U6 small nuclear RNA (for miRNA expression). Data are expressed as mean of values from three independent clones ± standard error of the mean (SEM). Statistical significance was assessed using Students t-tests (two-tailed).
Ganglionic eminences from both hemispheres were dissected from wild-type and homozygous Foxp2-R552H E16 littermates [9]. R552H mice carry a missense mutation affecting a conserved arginine residue located in the Foxp2 DNA binding domain, matching an aetiological mutation found in a well-characterised multigenerational family with speech and language disorder (the KE family) [4]. R552H homozygous mice demonstrate comparable phenotypes to homozygous Foxp2 knockouts [9]. This suggests that the introduction of this mutation yields a stable, but non-functional protein product, a conclusion that is supported by in vitro functional studies [30]. Dissections were performed in dissection buffer (15 mM HEPES, 0.01% NaHCO3, 25 mM glucose in HBSS-CMF) and dissected regions were immediately chopped on ice and pelleted at 800 RPM and 4°C for 5 minutes. The pellet was incubated in papain solution (20 units/ml papain, 1 mM L-cysteine, 0.5 mM EDTA, 100 units/ml DNaseI, in dissection buffer) on ice for 5 minutes then at 37°C for 10 minutes, agitating regularly. The enzymatic reaction was halted by addition of Ovo-BSA solution (10 mg/ml ovomucoid, 10 mg/ml BSA in dissection buffer). Cells were pelleted at 1000 RPM for 5 minutes at 4°C and the pellet was washed then re-suspended in complete medium (neurobasal media (Sigma) supplemented by 2 mM Glutamax (Sigma). 2 mM Penicillin/Streptomycin (Sigma) and 1X B27 supplement). Suspension was triturated using plastic and glass pipettes to dissociate any remaining cell clumps before passing the cell suspension through a 70 µm cell strainer. Single cell suspensions were seeded onto laminin and poly-D-lysine coated coverslips (BD Biosciences) at a density of 6.3×104 cells per well into 24 well plates and grown at 37°C in the presence of 5% CO2 in complete medium.
After 4 days in culture, cells were fixed using 4% Paraformaldehyde solution for 15 minutes at room temperature and permeablised in wash solution (0.1% Triton X-100 in TBS). Antibodies were diluted in Blocking Solution (1% Fish Gelatine, 0.1% Triton X-100, 5% BSA in PBS). Cells were co-stained at 4°C overnight, using two primary antibodies; an anti-MAP2 rabbit polyclonal antibody (Chemicon) and an anti-Foxp2 mouse monoclonal antibody recognising an epitope near the C-terminal end of the protein (Gift from Prof. A. Banham). Detection was then facilitated via four rounds of antibody incubation, which allowed magnification of the Foxp2 signal. Cells were incubated with anti-rabbit TRITC (Alexa Fluor 568, Molecular Probes) plus anti-mouse biotinylated (BA9200, Vector Labs) secondary antibodies, followed by incubation with anti-rabbit TRITC plus anti-biotin FITC (Alexa Fluor 488, Molecular Probes) antibodies, each for 1 hour, shaking under limited light exposure. This secondary/tertiary antibody incubation was then repeated under the same conditions. Nuclei were visualised using mounting media containing a DAPI counterstain (VectaShield). Cells were viewed on a Nikon Eclipse TE2000U fluorescence inverted microscope. Images were captured using a Hamamatsu black and white C4742-95 Orca hi-sensitivity CCD camera with IPLab imaging software (Scanalytics Inc), and analysed using the neurite outgrowth function of Metamorph Version 7.6 (Molecular Devices). Statistical analyses were carried out using ANCOVA (analysis of covariance) for genotype and embryo, followed by post-hoc Sidak correction. Data are expressed as the mean ± standard error of the mean (SEM).
In situ hybridization was carried out on 10 µM frozen sections of E16.5 embryo heads as previously described [69]. An approximately 500 bp fragment of each target transcript was subcloned into pCR4-TOPO (Invitrogen) for dioxygenin-labelled riboprobe synthesis. Primer sequences for the riboprobes are available on request. Equivalent parasagittal sections were hybridized in parallel from three wild-type and three homozygous Foxp2-S321X mutant embryos and all slides were developed for 16 hours, or 6 hours in the case of Shhrs. In all cases a sense-strand negative control riboprobe gave no specific signal (data not shown).
|
10.1371/journal.ppat.1003758 | Conservative Sex and the Benefits of Transformation in Streptococcus pneumoniae | Natural transformation has significant effects on bacterial genome evolution, but the evolutionary factors maintaining this mode of bacterial sex remain uncertain. Transformation is hypothesized to have both positive and negative evolutionary effects on bacteria. It can facilitate adaptation by combining beneficial mutations into a single individual, or reduce the mutational load by exposing deleterious alleles to natural selection. Alternatively, it may expose transformed cells to damaged or otherwise mutated environmental DNA and is energetically expensive. Here, we examine the long-term effects of transformation in the naturally competent species Streptococcus pneumoniae by evolving populations of wild-type and competence-deficient strains in chemostats for 1000 generations. Half of these populations were exposed to periodic mild stress to examine context-dependent benefits of transformation. We find that competence reduces fitness gain under benign conditions; however, these costs are reduced in the presence of periodic stress. Using whole genome re-sequencing, we show that competent populations fix fewer new mutations and that competence prevents the emergence of mutators. Our results show that during evolution in benign conditions competence helps maintain genome stability but is evolutionary costly; however, during periods of stress this same conservativism enables cells to retain fitness in the face of new mutations, showing for the first time that the benefits of transformation are context dependent.
| Transformation of environmental DNA can provide bacteria with a means to adapt quickly to a changing environment. While this can benefit microbes by facilitating the spread of antibiotic resistance, it can also be harmful if it causes the loss of beneficial alleles from a population. Therefore, it is unclear what evolutionary factors enable transformation to persist in bacterial populations. We used the naturally transformable opportunistic pathogen Streptococcus pneumoniae to investigate the long-term benefits of transformation. We compared the fitness of laboratory populations of S. pneumoniae after 1000 generations of evolution. Half of these populations were naturally transformable (competent) while the other half was deficient for this function. At the same time, half of the evolving populations were periodically exposed to short periods of mild stress. We find that competence reduces the average fitness gain of evolving populations, but this cost is mitigated in populations facing mild stress. Using whole genome sequencing, we discovered that functional competence reduces the total number of fixed mutations and prevents hyper-mutable cells from increasing in frequency. Our results suggest that competence in S. pneumoniae is a conservative process acting to preserve alleles, rather than an innovative one that persists because it recombines beneficial mutations.
| Natural transformation is an important cause of genome evolution in bacteria, but the evolutionary factors maintaining natural transformation, or competence, in bacteria remain uncertain [1], [2], [3], [4]. Transformation is widely believed to have evolved to facilitate adaptation, especially in a clinical context where transformation occurs at a high rate and may allow pathogens to evade antibiotics or immune surveillance [2], [5], [6], [7]. However, transformation can both be beneficial and costly to bacterial cells. It can speed up adaptation by combining non-antagonistically epistatic beneficial mutations into a single individual [8], [9], [10], [11], similar to the Fisher-Muller effect in Eukaryotes. It can also reduce the mutation load by combining deleterious alleles into a common background, which more efficiently exposes these mutations to natural selection [12], [13], [14], [15]. Similarly, transformation can eliminate deleterious alleles if these are replaced by transformed DNA with the wild-type sequence [12]. Such a function has been inferred in naturally transforming Neisseria, where high rates of transformation with ‘self-DNA’ leads to conservation of core regions of the genome [16]. Alternatively, transformation can impose significant fitness costs because it is energetically expensive and bacterial cells may incorporate damaged or mutated environmental DNA that reduces bacterial fitness [12]. Thus, although the mechanisms that regulate bacterial transformation are well understood, the evolutionary factors that maintain this process are not.
Experiments designed to quantify the evolutionary effects of bacterial transformation have been equivocal. While Baltrus et al. [17] showed that transformable strains of Helicobactor pylori evolved more rapidly than non-transformable strains, Bacher et al. [18] found the reverse in Acinetobacter baylyi. Indeed, wild-type strains in this experiment evolved lower rates of transformation during laboratory culture. Although these experimental differences may have been caused by details particular to the species investigated, they may have also been caused by experimental approaches that inadvertently exposed different benefits or costs of competence. For example, studying yeast, Saccharomyces cerivisiae, Gray and Goddard [19] showed that sex only increased adaptation in a stressful environment, and that this was especially pronounced in populations of cells with a high mutation rate. Similar studies designed to partition the effects of these experimental factors have thus far not been attempted using bacteria.
Here, we use an experimental evolution approach to examine the long-term effects of transformation on the naturally competent opportunistic pathogen Streptococcus pneumoniae. Replicate cultures of a competent strain and an isogenic mutant unable to become competent were evolved for 1000 generations in two different chemostat environments. Half of the populations were evolved in a constant benign environment, while the other half was exposed to twice-weekly pulses of a sub-minimum inhibitory concentration (MIC) of kanamycin. This treatment was used to examine context-dependent effects of transformation. Although the concentration of kanamycin we used did not influence the growth rate of cells in our experimental populations (Figure S1), it is sufficient to induce cellular stress and also induce natural competence [20]. In brief, we found that competent cells evolved in a benign environment increased in fitness less than non-competent cells; however, this cost of competence was alleviated when cells evolved in an environment where they were challenged with periodic stress. Using whole genome sequencing of evolved isolates, we show that non-competent populations fixed more mutations than competent ones and are furthermore were more likely to evolve to become mutators. Our results provide direct experimental evidence that the effects of transformation are context dependent. In benign conditions the effects of transformation are conservative and evolutionarily costly; however, this same conservativism benefits cells living in stressful environments.
To quantify the effects of competence on the evolution of S. pneumoniae we evolved replicate populations of a competent wild-type and an isogenic non-competent mutant in chemostats for 1000 generations. We introduced periodic stress in half of the evolving populations by applying sub-MIC concentrations of kanamycin to chemostat vessels twice each week. Kanamycin was added as a single injection to the chemostat sampling port in order to achieve a final concentration of 5 ug/mL, which is approximately 20-fold below the MIC for this strain. At the chemostat flow rates used, the kanamycin was eliminated within 7 hours. For the majority of time, these populations therefore experienced the same environment as the unstressed populations. At the start of this experiment, all ancestral populations exhibited pronounced oscillations of several orders of magnitude in cell density (between 109/ml and 105/ml), as previously described with this species [21]. However, within 500 generations these fluctuations were uniformly lost [22], and all populations retained a stable density of 109 cells/ml. As a consequence of this change, which made it impossible to directly compete evolved and ancestral cells, fitness differences between treatments after experimental evolution were estimated using pair-wise competition assays between differentially marked terminal populations [23]. This allowed us to directly estimate fitness differences between evolved strains, and to quantify the relative fitness of competent and non-competent populations under benign and periodically stressed conditions.
In contrast to the expectation that competence accelerates adaptation, we found instead that evolved competent populations of S. pneumoniae were significantly less fit than non-competent populations (Fig. 1, restricted maximum likelihood (REML) mixed model compared to 0: t = −2.22; p<0.001). This evolutionary cost of competence corresponds to a fitness difference between evolved populations of ∼0.098/hr, which implies that under benign conditions the effects of competence on adaptation are significantly negative. By contrast, the relative fitness of competent and non-competent populations evolved in the presence of periodic stress was indistinguishable (Fig. 1, REML mixed model compared to 0: t = −0.18; p = 0.741) indicating that these conditions significantly offset the evolutionary costs of competence (REML mixed model: F = 2.077; p = 0.017).
Exposure to periodic stress could offset the costs of competence by increasing adaptation of competent cells and/or by reducing the rate of adaptation of non-competent cells. The results from competition assays support both explanations. When we estimated the relative fitness of stressed versus unstressed competent cells we found that stressed populations were significantly more fit (∼0.052/hr) than populations evolved in an unstressed environment (Fig. 1, REML mixed model compared to 0: t = 1.27; one-tailed p = 0.045). By contrast, we found the reverse relationship for non-competent cells, and of roughly equal magnitude; stressed non-competent populations were significantly less fit (∼0.067/hr) than populations evolved in an unstressed environment (Fig. 1, REML mixed model compared to 0: t = −1.67; one-tailed p = 0.012). The sum of these two effects on the rate of adaptation, i.e. the benefit for stressed competent populations and the cost for stressed non-competent populations, is similar to the ∼0.098/hr cost for competent cells noted above (Fig. 1). The benefits to competence of stress therefore arise in two ways: by increasing adaptation of competent cells and by enabling competent cells to avoid fitness reductions attributable to exposure to stress.
To analyze the influence of competence and stress on genome evolution during experimental evolution we obtained the complete genome sequences of evolved clones from each of the 16 populations together with their respective ancestors. Competence could potentially alter mutation fixation in two ways, both caused by the fact that competence unites mutations from separate cells into a common genetic background. First, if recombined mutations are beneficial either alone or in combination, competence could increase the fixation rate because recombinant cells would be predicted to increase in frequency. Second, if recombined mutations are deleterious, alone or in combination, competence could reduce the fixation rate because recombinant cells would be exposed to natural selection and eliminated from the experimental population. A similar reduction in fixation rate would be anticipated if competence replaces new mutations in the host genome with donor DNA containing the wild-type allele, thereby “correcting” mutations. We furthermore predict that sub-MIC antibiotic stress will have a general increase on mutation fixation, owing to potentially mutagenic effects of kanamycin [24]. Sequencing of evolved genomes identified a total of 421 synonymous mutations and 1282 non-synonymous mutations across all evolved lines. Substitutions were not evenly distributed across treatments and populations, however. Consistent with the second possibility outlined above, we found that the total number of mutations in competent populations was significantly lower than in non-competent populations (Fig. 2A; GLM: z = −9.344, df = 1 p<0.0001). Moreover, the total number of mutations was significantly higher in populations experiencing periodic stress for both competent and non-competent populations (Fig. 2A; GLM: z = 7.379, df = 1 p<0.0001).
We next determined the mutation rate of each population based on the total number of mutations. This method assumes that the mutation rate was constant over the period of evolution and that there are no back mutations. We found that three of eight non-competent populations substituted between two and eight-fold more mutations than the other populations (Fig. 2B), suggesting that these lineages had evolved to become genetic mutators. However, because this estimate assumes a constant mutation rate during the 1000 generations of experimental evolution, rather than a changing rate due to the fixation of mutator alleles, mutators emerging at the end of experimental evolution may not be identified. To address this limitation, we first searched for mutations in genes diagnostic for bacterial mutators (e.g. DNA-repair genes polC, dnaQ, rpoABC and mutL [25], see Table S1 for a comprehensive list). We focused here on non-synonymous mutations because these are more likely to cause functional defects in the relevant genes. Next we determined the mutation rate of all evolved lineages relative to their respective ancestor using a phenotypic assay designed to detect the frequency of spontaneous mutants to resistance to either rifampicin or streptomycin [26]. Using the first approach, we detected significantly more non-synonymous mutations in DNA repair genes in non-competent than in competent populations (Fig. 2C: GLM one-tailed: χ2 = 162.00, df = 1, p = 0.0478), an effect that is most pronounced in genes for mismatch repair (Fig. 2B; GLM: Mismatch repair: χ2 = 9.875, df = 1, p = 0.001). Next, using a phenotypic assay, we found that the mutation frequency of non-competent lineages was significantly higher than competent populations (Fig. 2D; Two-way ANOVA: F1,11 = 10.189; p = 0.009), although the mutation frequency of ancestral strains were indistinguishable (t-test with unequal variances: t = 1.1336, df = 7.112, p = 0.2937). In contrast to re-sequencing results, these assays found no overall effect of stress on the mutation frequency (Fig. 2D; Two-way ANOVA: F1,11 = 2.699; p = 0.129), nor an interaction between stress and competence (Fig. 2C; Two-way ANOVA: F1,11 = 0.0003; p = 0.9876). Thus both at the genetic and phenotypic levels, our data support a model where competence reduces mutation fixation and limits the emergence of mutator phenotypes, but this conservatism comes possibly at the expense of reduced adaptation under benign growth conditions.
Transformation can dramatically benefit S. pneumoniae by facilitating the evolution of drug resistance and the emergence of novel modes of virulence [27], [28], [29]. However, these benefits in pathogenic bacterial lineages under strong antibiotic selection tell only part of the story, and may not reflect the effects of transformation more broadly. Using an experimental evolution approach, we found that competence benefited cells by reducing the mutation load and limiting the emergence of mutators (Fig. 2). Additionally, competent populations reached higher fitness when evolving in the presence of periodic stress; equally, exposure to periodic stress decreased the rate of evolution of non-competent populations (Fig. 1). Although we applied an extremely mild stress in our experiment (Figure S1), it is notable that the kanamycin concentration we used is sufficient to induce competence in wild-type strains [20]. It is therefore possible that benefits to competence in populations that experienced drug-stress was the result of increased recombination, which could have off-set the cost of transformation in a benign environment by slightly increasing their rate of adaptation. By contrast, non-competent cells exposed to kanamycin may face greater costs because kanamycin causes an inability to repair ribosomal decoding errors, which can subsequently lead to DNA damage and increase the mutation rate [24]. These stress-dependent benefits of competence may be particularly important in the human nasopharynx, where S. pneumoniae is exposed to unpredictable and severe stress from drug exposure, immune surveillance and from coexisting bacterial competitors.
Transformation is predicted to benefit bacterial species with high mutation rates by reducing their mutation load [12]. Using complete genome sequences, we estimate that the average mutation rate in S. pneumoniae is ∼3.8×10−8 per bp per generation. This corresponds to U = 0.08 mutations per genome per generation, or about 200-fold higher than Escherichia coli [30], yet similar to other naturally transformable opportunistic pathogens, such as H. pylori and H. influenzae [30], [31], [32], [33]. Despite these high rates of mutation we were surprised to find that some of the non-competent strains evolved even higher rates of mutation than their ancestor during this long-term experiment (Figs. 2A and 2B). These genotypic results were confirmed phenotypically (Fig. 2D), and suggest that our genomic data underestimates the difference in fixation rates in competent and non-competent populations. Although we are uncertain what caused the difference in mutation rates between competent and non-competent lineages to arise, one strong possibility is that transformation separates mutator alleles from the mutations they cause. Thus while mutations in DNA repair genes leading to mutators may arise equally in both competent and non-competent cells, they are lost before they become common in competent lineages [34]. Accordingly, competent lineages fix fewer mutations overall. Under benign conditions this may limit adaptation while causing minimal harm to non-competent populations. However, non-competent cells suffer to a greater degree when faced with stress, because they cannot revert to a less loaded state, and because stress may exacerbate the negative fitness effects of new mutations [35], [36]. In a similar recent study with the yeast Saccharomyces cerevisiae sex neutralised the deleterious effects of hyper mutation on the rate of adaptation [19]. The neutralisation of potentially deleterious mutations, e.g. those that lead to hyper mutation, is an example of how transformation can function to conserve genome integrity. Similar effects are inferred in the naturally transformable bacterial genus Neisseria where the number of species-specific DNA uptake sequences (i.e. small sequence tags that identify that a DNA fragment is derived from a particular species) are more frequent in the core genome. This indicates that these core genes are often the target of ‘selfing’ events, by which. transformation stabilizes the integrity of key genes [16]. Although S. pneumoniae is much more promiscuous than Neisseria when it comes to environmental DNA choice, the mechanism that induces competence is assumed to lead to the uptake of DNA released from lysed cells of the same and/or very closely related species [37], which suggests that the conservative benefits we observe from competence in our study may extend more broadly to pneumococci in the natural environment.
Bacteria in nature face unpredictable patterns of stress and mutation. Our results suggest that these conditions, together with an intrinsically high mutation rate, favour the maintenance of transformation while infrequent stress may facilitate its loss. Notably, surveys of naturally competent species such as H. influenzae and B. subtilis have revealed that transformation rates among clones within species can vary by over 6 orders of magnitude [38], [39], [40]. Similar variation exists in S. pneumoniae [41], [42], [43], indicating that competence is readily gained and lost in this species. In summary, we conclude that competence in S. pneumoniae is a conservative process acting to preserve alleles, rather than an innovative one that persists because of benefits it provides by recombining beneficial mutations.
Strains used in this study were derived from Rx1 and its isogenic non-competent derivative FP5, which is unable to secrete the competence stimulating peptide, CSP [42]. Spontaneous rifampicin or streptomycin resistant mutants were isolated from each strain, and then four independent colonies of each type were further sub-cloned and stored at −80°C. These four independent ancestors were selected based on similar growth rates (Figure S2) and mutation rates (mean rates ±95% CI Non-competent: 1.84*10−7±1.56*10−7, competent: 8.41*10−8±7.51*10−8; t-test with unequal variances: t = 1.1336, df = 7.112, p = 0.2937). These 16 total clones (2 strains×2 drug resistance types×4 replicates) represented the ancestral populations for experimental evolution. Cultures were grown in ¼ CTM pH 7.8 (Complete Transformation Medium), which per litre consists of: 7.5 g Tryptic Soy Broth; 0.25 g yeast extract; 6 g NaCl at pH 7.8. This environment supported high levels of transformation (Figure S3). Blood agar plates (Tryptic Soy Agar (TSA)+3% horse blood), supplemented, where necessary, with either 4 µg/mL rifampicin or 100 µg/mL streptomycin, were used to enumerate cell density within chemostats and for colony counting during competition assays.
Experimental evolution and competition assays were performed in custom-made chemostats with a 25 mL working volume and a flow rate of 4 mL/hr, whilst maintained at 37°C [29]. Chemostat cultures were inoculated and maintained as described previously [21], and sampled every 50 generations of growth. Samples were stored at −80°C as freezer stocks in ¼ CTM pH 7.8+25% v/v glycerol at an OD600 of 0.20, corresponding to a density of 2×108 cells mL−1.
Sixteen chemostat populations were inoculated with independently picked clones from the original antibiotic resistant strain to generate 4 replicates each of a 2*2 treatment design pairing competence and stress. The replicates in each treatment were equally split between the two differently marked versions of Rx1 (competent strain) and Fp5 (non-competent strain). Half of the populations were exposed twice a week to low doses of kanamycin introduced directly into the chemostat to simulate short periods of stress. Kanamycin concentrations were 5 µg/mL upon introduction, but declined with the normal outflow rate of the chemostat. This concentration of kanamycin had no effect on the growth rate of cells (Figure S1), but is sufficient to cause ribosomal decoding errors during protein production, which promotes the induction of competence [20], [44]. For simplicity, this treatment is referred to as “periodic stress” and the basal treatment as “benign”. Each strain was evolved independently, thereby avoiding potential effects of cross-induction of competence or competence-induced cell-lysis [45], [46]. Every week, after approximately 50 generations, a 1 mL sample was taken from each population and tested for the presence of the correct marker and absence of the opposite marker. Contaminated populations were restarted with 50 µL of the previous sampled uncontaminated time point. Populations were maintained for 20 weeks, which corresponds to about 1,000 generations.
Fitness was determined by comparing the change in relative densities of two reciprocally marked evolved populations in a chemostat in mixed culture over a 32-hour span. This time period was chosen because it is within the period that the periodically stressed populations spend in the benign environment between doses of kanamycin. Competition assays were initiated by inoculating chemostats with equal densities of each competitor. Chemostats were sampled immediately and then again after 32 hours to determine the relative densities of each competitor. The Malthusian parameters per hour were then calculated for each strain based on the density of each strain at the start and end of the competition, as described previously [47]. The selection rate constant was then calculated as the difference between Malthusian parameters as described previously [23]. First, we tested for a significant fitness difference between competitors for each treatment by comparing a restricted maximum likelihood mixed model against an intercept of zero, corresponding to equal fitness. In the mixed model, replicate fitness assays of competitor combinations were nested as a random effect within the fixed effect of treatments (absence/presence periodic stress and absence/presence competence). Second, we used the restricted maximum likelihood (REML) mixed model, again with replicate fitness assays as a random factor within treatments, to test for fitness differences between treatments (periodic stress or competence as a fixed factor). All analyses were done in R with package LME4. P-values were estimated by MCMC simulation with 10,000-fold replication using the p.vals command from the languageR package.
Clonal isolates from each of the 16 evolved populations as well as all four ancestral strains were sequenced using the SOLiD4 platform at the University of Manchester genomics facility. Genomic DNA was obtained using phenol-chloroform isolation and ethanol precipitation [48]. SOLiD data were normalised to an equal number of reads (8,315,863 per strain) for each sample using a custom perl script, that randomly sampled the reads from the original dataset (getRandomTags_Index_fastq.pl) developed by I. Donaldson. The normalisation equalized the size of the datasets to the strain with the lowest number of reads thereby normalising the quality of the consensus sequences. The normalised reads were then mapped against the fully sequenced reference strain S. pneumoniae R6 (genbank accession: NC_003098 = AE007317, an easily accessible version of the genome database can be found at http://www.streppneumoniae.com) using BFAST (0.6.4e) using default colour space methodology giving an average coverage depth of 151-fold. Mapped reads were then locally realigned around INDELs using SRMA (0.1.15). SNPs and small INDELs were then determined from the resulting BAM-files using the Geneious package (Geneious 5.4, Auckland, New Zealand; [49]. A SNP or INDEL was called when the change to the reference was supported in 60% of the reads with at least a coverage depth of 20 reads using the variant calling tool in Geneious 5.4 to minimise false positives and negative SNP calls. The 60% support is slightly less stringent than previous studies, but this is compensated by the high average coverage depth of 151 (±SE 2.53) reads/base [50], [51]. Subsequently, variant tables extracted from Geneious were used in the Galaxy online tool set [52], [53] to identify mutations for each evolved clone compared to its ancestor. Parallel changes were then double checked by hand in the UCSC microbial genome browser [54] to eliminate false positives. The resulting mutation tables were used for further analysis.
To determine the effect of periodic stress and competence the total numbers of mutations were compared in a GLM model with a Poisson distribution using R. Subsequently, two different methods were used to determine the mutation rate of evolved populations. First, the per base per generation mutation rate was calculated from genome data from the total number of mutations (See Table S2 for mutation rates based on coding mutations and synonymous mutations only); 95% confidence intervals for these mutation rates were determined in R using a Poisson distribution. This analysis assumes that the number of mutations is relatively small, that there are no back mutations, and that the mutation rate was constant over the period of evolution.
Second, the mutation frequency of terminal lineages (i.e. the number of spontaneous mutants with either rifampicin or streptomycin resistance/total population density) was determined following the methods in [26] to estimate the mutation rate of terminal populations isolated after their final generation of experimental evolution. Each strain and its corresponding ancestor was grown overnight at 37°C+5% CO2. Cells were then washed and concentrated by centrifugation and re-suspended in 100 µL 0.8% NaCl. 10 µL spots at several different dilutions (five spots per dilution) were plated onto blood agar plates supplemented with either 4 µg/mL rifampicin or 100 µg/mL streptomycin, whilst total cell densities were determined on unsupplemented plates. Mutation frequency was estimated as the ratio of the number of mutants to the total population size. Assays were performed in triplicate for each genotype, and the relative mutation frequency was determined as the ratio of the mutation frequency of each evolved lineage to its corresponding ancestor. Mean relative mutation frequencies were log-transformed before analysis using a two-way ANOVA.
Finally, parallel changes in DNA repair genes were examined at the level of each gene and functional group, as determined from the KEGG classifications for S. pneumoniae R6 (http://www.genome.jp/dbget-bin/www_bget?gn:T00060). The table of non-synonymous SNPs was used to create a SNP-by-gene table by scoring presence/absence of at least one SNP in a given gene for each strain (see Table S1 for a detailed summary of the genes involved and the amino acid changes found per evolved line). Functional group associations were created from the total SNP-by-gene table by summarising presence and absence of SNPs for genes associated with a functional group according to the KEGG-database. Generalised linear models were used to test for differences between treatments for parallel non-synonymous mutations in functional groups.
|
10.1371/journal.ppat.1002040 | Human Neutrophil Clearance of Bacterial Pathogens Triggers Anti-Microbial γδ T Cell Responses in Early Infection | Human blood Vγ9/Vδ2 T cells, monocytes and neutrophils share a responsiveness toward inflammatory chemokines and are rapidly recruited to sites of infection. Studying their interaction in vitro and relating these findings to in vivo observations in patients may therefore provide crucial insight into inflammatory events. Our present data demonstrate that Vγ9/Vδ2 T cells provide potent survival signals resulting in neutrophil activation and the release of the neutrophil chemoattractant CXCL8 (IL-8). In turn, Vγ9/Vδ2 T cells readily respond to neutrophils harboring phagocytosed bacteria, as evidenced by expression of CD69, interferon (IFN)-γ and tumor necrosis factor (TNF)-α. This response is dependent on the ability of these bacteria to produce the microbial metabolite (E)-4-hydroxy-3-methyl-but-2-enyl pyrophosphate (HMB-PP), requires cell-cell contact of Vγ9/Vδ2 T cells with accessory monocytes through lymphocyte function-associated antigen-1 (LFA-1), and results in a TNF-α dependent proliferation of Vγ9/Vδ2 T cells. The antibiotic fosmidomycin, which targets the HMB-PP biosynthesis pathway, not only has a direct antibacterial effect on most HMB-PP producing bacteria but also possesses rapid anti-inflammatory properties by inhibiting γδ T cell responses in vitro. Patients with acute peritoneal-dialysis (PD)-associated bacterial peritonitis – characterized by an excessive influx of neutrophils and monocytes into the peritoneal cavity – show a selective activation of local Vγ9/Vδ2 T cells by HMB-PP producing but not by HMB-PP deficient bacterial pathogens. The γδ T cell-driven perpetuation of inflammatory responses during acute peritonitis is associated with elevated peritoneal levels of γδ T cells and TNF-α and detrimental clinical outcomes in infections caused by HMB-PP positive microorganisms. Taken together, our findings indicate a direct link between invading pathogens, neutrophils, monocytes and microbe-responsive γδ T cells in early infection and suggest novel diagnostic and therapeutic approaches.
| The immune system of all jawed vertebrates harbors three distinct lymphocyte populations – αβ T cells, γδ T cells and B cells – yet only higher primates including humans possess so-called Vγ9/Vδ2 T cells, an enigmatic γδ T cell subset that uniformly responds to the majority of bacterial pathogens. For reasons that are not understood, this responsiveness is absent in all other animals although they too are constantly exposed to a plethora of potentially harmful micro-organisms. We here investigated how Vγ9/Vδ2 T cells respond to live microbes by mimicking physiological conditions in acute disease. Our experiments demonstrate that Vγ9/Vδ2 T cells recognize a small common molecule released when invading bacteria become ingested and killed by other white blood cells. The stimulation of Vγ9/Vδ2 T cells at the site of infection amplifies the inflammatory response and has important consequences for pathogen clearance and the development of microbe-specific immunity. However, if triggered at the wrong time or the wrong place, this rapid reaction toward bacteria may also lead to inflammation-related damage. These findings improve our insight into the complex cellular interactions in early infection, identify novel biomarkers of diagnostic and predictive value and highlight new avenues for therapeutic intervention.
| The cellular immune system consists of an ‘innate’ arm of phagocytes and antigen-presenting cells, and an ‘adaptive’ arm of antigen-specific lymphocytes capable of developing immunological memory. Yet, there is increasing evidence of considerable crosstalk between the two [1]. Innate responses directly influence the shape and outcome of adaptive T cell responses, and vice versa specialized T cell subsets feedback to innate cells [2]. Among these interactions, the regulation of neutrophil-mediated inflammatory responses by Th17 cells has received enormous attention over the past few years [3], and with the emergence of novel T cell subsets additional networks are being proposed so that each polarized T cell eventually pairs with an innate counter player [4]–[7].
The necessity to integrate complex signals in order to mount the most effective defense is best illustrated by the existence of ‘unconventional’ T cells bridging the classical divide between innate and adaptive immunity, such as natural killer T cells, mucosal-associated invariant T cells, intestinal intraepithelial CD8αα+ T cells and dendritic epidermal γδ T cells [8]–[14]. These often tissue-associated lymphocytes are characterised by restricted T cell receptor (TCR) repertoires that allow them to respond rapidly to a limited range of conserved structures. Unconventional T cells readily assume a plethora of effector functions, ranging from sentinel tasks and targeted killing to engaging with keratinocytes, fibroblasts, phagocytes and antigen-presenting cells as well as other lymphocyte.
γδ T cells expressing a Vγ9/Vδ2 TCR – Vγ2/Vδ2 according to an alternative nomenclature – are only found in humans and higher primates and differ fundamentally from all other conventional and unconventional T cells [15]. Activated Vγ9/Vδ2 T cells produce a range of cytokines, kill infected and transformed target cells, regulate survival and differentiation of monocytes and maturation of dendritic cells, provide B cell help and present antigens to CD4+ and CD8+ T cells [11], [12], [16], [17]. They expand considerably in many infections, at times to >50% of all circulating T cells within a few days [18], and respond selectively in a non-MHC restricted manner to the microbial metabolite (E)-4-hydroxy-3-methyl-but-2-enyl pyrophosphate (HMB-PP) [19]. HMB-PP is an intermediate of the non-mevalonate pathway of isoprenoid biosynthesis that is present in many bacteria and in malaria parasites but not in humans [17]–[19]. The rapid and sensitive response of Vγ9/Vδ2 T cells to a broad range of pathogens evokes cardinal features of innate immunity. Indeed, HMB-PP fulfills Janeway's criteria for a ‘pathogen-associated molecular pattern’ in that it is (i) invariant among different species; (ii) a product of a pathway unique to micro-organisms; and (iii) essential in microbial physiology [17]. Yet, HMB-PP recognition is not mediated via germline-encoded pattern recognition receptors but involves the re-arranged Vγ9/Vδ2 TCR [20]–[22].
Bacteria that possess the non-mevalonate pathway and hence produce HMB-PP comprise some of the most detrimental human pathogens such as the causative agents of cholera, diphtheria, plague, tuberculosis and typhoid, but also numerous commensal and opportunistic species in the mucosal flora, skin and feces [19], [23]. In all these micro-organisms, HMB-PP is an essential intracellular metabolite, and it is not clear whether and how it is released by invading bacteria and becomes visible to the immune system as soluble molecule. Indeed, earlier studies with mycobacteria suggested that uptake of whole bacteria by monocytes, macrophages, or DCs may be required for the recognition by Vγ9/Vδ2 T cells [24]–[27]. Neutrophils are the first immune cells infiltrating the site of infection and the main phagocytes responsible for early pathogen clearance, and growing evidence points toward a crucial role of γδ T cells in regulating neutrophil responses in mouse models of infection, hypersensitivity and autoimmunity [8], [12]. Yet, the interplay between γδ T cells and neutrophils has not been addressed in detail [28], [29]. Our present data demonstrate that Vγ9/Vδ2 T cells readily respond to neutrophils harboring phagocytosed bacteria, and that this response is strictly dependent on the ability of these bacteria to produce HMB-PP and cell-cell contact of Vγ9/Vδ2 T cells with accessory monocytes. The majority of circulating Vγ9/Vδ2 T cells shows migration properties similar to monocytes [30], suggesting that these two cell types are co-recruited to the site of inflammation and interact with each other at early stages of infection [17], [31]. Our present findings thus indicate a direct link between invading pathogens, neutrophils, monocytes and microbe-responsive γδ T cells, and suggest novel diagnostic and therapeutic approaches in acute infection.
Neutrophils are short-lived phagocytes that undergo spontaneous apoptosis in vitro unless rescued by survival signals. We previously demonstrated that activated human Vγ9/Vδ2 T cells induce monocytes to survive and differentiate into inflammatory dendritic cells [31]. Here, HMB-PP stimulated Vγ9/Vδ2 T cells had a similar survival effect on autologous neutrophils and readily rescued them from undergoing apoptosis (Figure 1). This effect was selective and dependent on the number of Vγ9/Vδ2 T cells and the concentration of HMB-PP. An increase in neutrophil survival could already be observed at ratios of only 1 γδ T cell per 100 neutrophils and at HMB-PP concentrations as low as 0.1–1 nM. Activation of Vγ9/Vδ2 T cells in these cultures was confirmed by upregulation of CD69 and secretion of interferon (IFN)-γ (Figure S1 in Text S1). The low γδ T cell numbers and HMB-PP concentrations needed to promote neutrophil survival in vitro are likely to have physiologic relevance.
Activated neutrophils mobilize intracellular stores of CD11b to the cell surface and shed CD62L, thus enhancing their potential to undergo firm adhesions with endothelial cells and extravasate at the site of inflammation. In line with their anti-apoptotic effect on neutrophils, Vγ9/Vδ2 T cells induced upregulation of CD11b and loss of CD62L in surviving neutrophils in an HMB-PP dependent manner (Figure 2A). Importantly, synthetic HMB-PP alone did not have any activity on neutrophils in the absence of γδ T cells (Figure 1, Figure 2 and data not shown).
Rapid recruitment of neutrophils involves the chemotactic action of CXCL8 (IL-8) produced at the site of inflammation, and increased endothelial permeability mediated by tumor necrosis factor (TNF)-α. Analysis of the supernatants from the above experiments revealed that co-cultures of neutrophils and HMB-PP stimulated Vγ9/Vδ2 T cells produced considerable amounts of both CXCL8 and TNF-α, in a dose-dependent manner and at levels comparable to lipopolysaccharide (LPS) stimulated neutrophils (Figure 2B). Another cytokine implicated in neutrophil recruitment is IL-17, which in a number of infection models is readily produced by murine γδ T cells [8]. While activated Vγ9/Vδ2 T cells readily produce TNF-α, IFN-γ and granulocyte/macrophage colony-stimulating factor (GM-CSF) [31], [32], we were unable to detect IL-17 in our co-cultures indicating that under the conditions tested human γδ T cells failed to secrete relevant levels of IL-17 (data not shown). This is reminiscent of recent findings that human αβ T cells including human Th17 cells modulate neutrophils (which lack the IL-17 receptor C chain) in an IL-17–independent manner through a combination of TNF-α, IFN-γ and GM-CSF [33]. In our cultures, blocking experiments demonstrated that TNF-α played a key role in the γδ T cell-mediated effect on neutrophils, as judged by a partial inhibition of neutrophil survival and a reduction of CD11b expression in the presence of soluble TNF-α receptor (sTNFR), while neutralizing antibodies against GM-CSF and IFN-γ had no significant effect (Figure 3).
Taken together, these data show that Vγ9/Vδ2 T cells become activated by soluble HMB-PP in the presence of autologous neutrophils and that they provide potent stimulatory signals inducing neutrophil survival and activation. The interaction of the two cell types leads to the rapid release of the pro-inflammatory mediators CXCL8 and TNF-α into the microenvironment, thereby potentially maintaining neutrophil influx at the site of infection.
Under physiological conditions, invading pathogens are rapidly taken up by newly recruited neutrophils. We therefore tested whether Vγ9/Vδ2 T cells respond to neutrophils harboring phagocytosed bacteria in a similar manner as they respond to soluble HMB-PP. In order to do this, we set up triple cultures consisting of neutrophils, monocytes and Vγ9/Vδ2 T cells, mimicking physiological conditions at the site of infection.
Human neutrophils readily took up green fluorescent protein (GFP) expressing Escherichia coli, Listeria innocua and Mycobacterium smegmatis, with >95% of the neutrophils being GFP+ within 30 min (Figure 4A and data not shown). Triple co-cultures of neutrophils harboring different strains of M. smegmatis with autologous Vγ9/Vδ2 T cells and monocytes led to rapid γδ T cell activation, as evidenced by upregulation of CD69 and expression of TNF-α and IFN-γ within 20 hours (Figure 4B and data not shown). Activation profiles were similar to those seen in control cultures with non-infected neutrophils in the presence of synthetic HMB-PP, demonstrating that Vγ9/Vδ2 T cells respond to bacterial degradation products released or presented by neutrophils. For the sake of clarity and simplicity all activation data in the following sections are shown as proportion of CD69+ TNF-α+ γδ T cells in the cultures although the cells were always co-stained for IFN-γ as well. The proportion of CD69+ IFN-γ+ and TNF-α+ IFN-γ+ γδ T cells followed essentially the same pattern throughout this study and led to the same conclusions.
In order to investigate the correlation between the ability of bacteria to produce HMB-PP and their capacity to stimulate Vγ9/Vδ2 T cells, we designed experiments to distinguish a specific γδ T cell response to HMB-PP from a possible background stimulation by the plethora of other microbial compounds acting via pattern recognition receptors. Thus, we generated a M. smegmatis transfectant stably expressing a second copy of the gene encoding HMB-PP synthase (gcpE) and hence overproducing HMB-PP compared to the parental wildtype (wt) strain [34] (Figure S2 in Text S1). As a second bacterial model we utilized HMB-PP producing and HMB-PP deficient strains of the non-pathogenic Gram-positive bacterium Listeria innocua [35], [36] (Figure S3 in Text S1).
Compared with M. smegmatis wt bacteria, higher levels of Vγ9/Vδ2 T cell activation were observed when using M. smegmatis-gcpE+ (Figure 5). Furthermore, considerable Vγ9/Vδ2 T cell activation was seen with phagocytosed L. innocua-gcpE+, a strain in which HMB-PP artificially accumulates, but not with the naturally HMB-PP deficient L. innocua wt strain that was >100× less potent (Figure 5). The double transfectant L. innocua-gcpE+lytB+, in which HMB-PP becomes converted into the downstream reaction products isopentenyl pyrophosphate (IPP) and dimethylallyl pyrophosphate (DMAPP), resulted in no detectable Vγ9/Vδ2 T cell activation (data not shown). These data demonstrate that the response of Vγ9/Vδ2 T cells to neutrophils harboring phagocytosed bacteria depends on the ability of these bacteria to produce HMB-PP and suggest that phagocytosis and subsequent degradation of bacteria in neutrophils leads to either presentation of HMB-PP on the cell surface or the release of soluble HMB-PP into the microenvironment.
Vγ9/Vδ2 T cells, monocytes and neutrophils share a responsiveness toward inflammatory chemokines and are the earliest leukocytes recruited to sites of infection. Vγ9/Vδ2 T cell responses in vitro are greatly facilitated by contact with monocytes as ‘feeder cells’, which most likely act by ‘presenting’ HMB-PP to Vγ9/Vδ2 T cells and by providing contact-dependent signals [17]. In support of our previous observation of a substantial HMB-PP dependent crosstalk between Vγ9/Vδ2 T cells and monocytes leading to optimum γδ T cell activation [31], the response of Vγ9/Vδ2 T cells to neutrophils harboring phagocytosed L. innocua-gcpE+ was largely dependent on the presence of monocytes. Omission of monocytes from the co-cultures resulted in greatly reduced expression levels of CD69, TNF-α and IFN-γ, compared to triple co-cultures (Figure 6A and data not shown), suggesting that monocytes provide essential help for the recognition of bacteria by Vγ9/Vδ2 T cells and increase the sensitivity of the response especially at low HMB-PP concentrations.
We speculated that this accessory effect might have stemmed from contact-dependent interactions of monocytes with either neutrophils or γδ T cells and tested this hypothesis in transwell cultures where neutrophils harboring phagocytosed L. innocua-gcpE+ in the lower chamber were separated from Vγ9/Vδ2 T cells in the upper chamber. As shown in Figure 6B, cell-cell contact between monocytes and Vγ9/Vδ2 T cells was crucial for the response to phagocytosed bacteria, while no contact was needed between Vγ9/Vδ2 T cells and neutrophils, and neither between monocytes and neutrophils. These data indicate that upon phagocytosis of HMB-PP+ bacteria, neutrophils release soluble factors that efficiently stimulate Vγ9/Vδ2 T cells, while monocytes provide important contact-dependent accessory signals.
Since neutrophils harboring bacteria were able to stimulate Vγ9/Vδ2 T cells in a transwell system, we next examined whether cell-free culture supernatants derived from infected neutrophils stimulated Vγ9/Vδ2 T cells in a similar manner. Indeed, Vγ9/Vδ2 T cells readily responded to supernatants from neutrophils harboring L. innocua-gcpE+ but not from neutrophils harboring L. innocua wt bacteria, as evidenced by expression of CD69, TNF-α and IFN-γ (Figure 6C and data not shown). Importantly, short-term pre-treatment of L. innocua-gcpE+ supernatants with shrimp alkaline phosphatase abrogated the bioactivity on Vγ9/Vδ2 T cells completely (Figure 6C), evoking the known sensitivity of mycobacterial HMB-PP to dephosphorylation and the relative inactivity of the dephosphorylated products [37]–[39]. Control experiments confirmed that alkaline phosphatase affected the response of Vγ9/Vδ2 T cells to synthetic HMB-PP but not to the phosphatase-resistant diphosphonate analogue, HMB-PCP [40], demonstrating that the presence of alkaline phosphatase in the cultures had no inhibitory effect on the cells' ability to express CD69, TNF-α and IFN-γ (Figure 6D and data not shown). We conclude that upon phagocytosis of HMB-PP+ bacteria neutrophils release soluble HMB-PP into the microenvironment where it becomes accessible to monocytes and Vγ9/Vδ2 T cells.
In order to assess the clinical relevance of our findings, we expanded our panel of bacteria by including clinical isolates of pathogens that are frequently associated with community- and hospital-acquired infections and pose serious threats to public health (Table S1 in Text S1). Of note, neutrophils harboring HMB-PP+ pathogens but not neutrophils harboring HMB-PP− pathogens induced in Vγ9/Vδ2 T cells the co-expression of CD69, TNF-α and IFN-γ (Figure 7A and data not shown). This response was largely independent of the presence of other pathogen-associated molecular patterns such as LPS as both Gram-negative (Acinetobacter baumannii, Enterobacter cloacae, Klebsiella pneumoniae, Pseudomonas aeruginosa) and Gram-positive bacteria (M. smegmatis) capable of producing HMB-PP stimulated Vγ9/Vδ2 T cells equally. Direct addition of alkaline phosphatase to these co-cultures abrogated the HMB-PP dependent cytokine responses, confirming soluble HMB-PP as common Vγ9/Vδ2 T cell stimulator in these species (Figure 7A, Figure S4 in Text S1). The bioactivity of culture supernatants harvested after 5 hours from neutrophils harboring the above bacteria corresponded to levels of 0.1–10 nM HMB-PP, as titrated against a HMB-PP standard (data not shown). Residual levels of CD69 expression after phosphatase treatment may have been due to incomplete degradation of HMB-PP and to indirect stimulation of Vγ9/Vδ2 T cells by other microbial compounds such as LPS acting on neutrophils or monocytes [41], [42]. In contrast to HMB-PP producing species, HMB-PP deficient Gram-negative (Chryseobacterium indologenes) and Gram-positive bacteria (Enterococcus faecalis, L. innocua, Staphylococcus aureus) did not elicit Vγ9/Vδ2 T cells responses above background as demonstrated by the complete lack of TNF-α (Figure 7A, Figure S4 in Text S1) and IFN-γ (data not shown). These findings illustrate the extraordinary specificity Vγ9/Vδ2 T cells for HMB-PP, even in the abundant presence of other microbial products and despite high levels of monocyte and/or neutrophil-derived mediators such as IL-1β, IL-6 and CXCL8 that were present in our triple co-cultures regardless of the HMB-PP status of the phagocytosed bacteria (data not shown).
γδ T cells expand rapidly in acute bacterial infections [18]. We therefore tested whether phagocytosed pathogens could induce expansion of 5-(and 6-)carboxyfluorescein diacetate succinimidyl ester (CFSE)-labeled Vγ9/Vδ2 T cells. As shown in Figure 7B, Vγ9/Vδ2 T cells proliferated considerably in the presence of supernatants derived from neutrophils harboring HMB-PP+ Enterobacter cloacae but not from neutrophils harboring HMB-PP− Chryseobacterium indologenes or Staphylococcus aureus. Similarly to the immediate up-regulation of CD69, TNF-α and IFN-γ, the proliferation of Vγ9/Vδ2 T cells in response to Enterobacter cloacae was HMB-PP dependent and could be abrogated by alkaline phosphatase. Expanding Vγ9/Vδ2 T cells also up-regulated the high affinity IL-2 receptor, CD25 (Figure 7B) and became responsive to exogenously added IL-2, which enhanced the proliferative response even further (data not shown). Blocking experiments demonstrated a crucial requirement of soluble and contact-dependent signals for optimum stimulation of Vγ9/Vδ2 T cells. TNF-α was recently implicated in Vγ9/Vδ2 T cell proliferation in response to IPP and IL-2 [43], and blocking of lymphocyte function-associated antigen-1 (LFA-1, CD11a/CD18) efficiently disrupted cluster formation with monocytes [31]. Here, both Vγ9/Vδ2 T cell proliferation and CD25 up-regulation in response to supernatants derived from neutrophils harboring HMB-PP+ Klebsiella pneumoniae (Figure 7B) or Enterobacter cloacae (data not shown) were readily inhibited by sTNFR and anti-CD11a antibodies but not by anti-IFN-γ antibodies. Finally, addition of anti-Vγ9 antibodies completely abrogated the Vγ9/Vδ2 T cell proliferation in response to HMB-PP (data not shown) and Enterobacter cloacae supernatants (Figure S7 in Text S1), confirming a requirement for the TCR [44]. Taken together, our findings demonstrate that Vγ9/Vδ2 T cells are rapidly activated by a broad range of HMB-PP producing pathogens, leading to TCR, LFA-1 and TNF-α dependent γδ T cell expansion.
We next addressed whether the dichotomy between HMB-PP+ and HMB-PP− pathogens in their potential to trigger γδ T cells in vitro is replicated under physiological conditions in vivo. As clinical correlate for HMB-PP+ and HMB-PP− infections, we analyzed episodes of acute bacterial infections in peritoneal dialysis (PD) patients, in whom the peritoneal catheter affords continuous and non-invasive access to the inflammatory infiltrate (Table S2 in Text S1). PD-associated peritonitis is characterized by a considerable influx of neutrophils and monocytes into the peritoneal cavity [45], [46], where the two cell types may become targets for local or infiltrating γδ T cells [31], [47]. Here, in a total of 24 newly recruited patients examined on the first day of acute peritonitis (i.e. before administration of antibiotics), both the total number and the frequency of peritoneal Vγ9/Vδ2 T cells were elevated in HMB-PP+ infections compared to HMB-PP− infections, suggesting increased recruitment and/or proliferation in response to HMB-PP released by bacteria (Figure 8). Moreover, local activation was evidenced by higher percentages of Vγ9/Vδ2 T cells expressing CD69 in the HMB-PP+ patient group. In contrast, we did not see any significant differences in the numbers and frequencies of peritoneal neutrophils, monocytes/macrophages and total CD3+ T cells, regardless of the HMB-PP status of the causative pathogen (Figure S5 in Text S1). Similarly, while the proportion of Vγ9/Vδ2 T cells within peritoneal CD3+ T cells was clearly elevated in HMB-PP+ infections, CD4+ and CD8+ T cells showed no such bias (Figure S6 in Text S1).
As Medzhitov stated recently, “inflammation is beneficial in appropriate amounts but can easily become detrimental when excessive because of its tissue-damaging potential” [48]. PD patients constitute a particularly vulnerable group where inflammatory events can have profound consequences [49]–[51]. We speculated that local activation of γδ T cells may contribute to inflammation-related damage and tested whether the occurrence of clinical complications in PD patients depends on the capacity of the causative pathogen to produce HMB-PP. Our analysis of 26 patients treated at the University Hospital of Wales, Cardiff, UK, demonstrated that infections with HMB-PP+ bacteria were associated with worse outcomes, evidenced as higher mortality rates and higher incidences of technique failure (i.e., cessation of therapy due to catheter removal, transfer to hemodialysis or patient death), while HMB-PP− bacteria caused milder disease (Figure 9). Of note, we were able to validate this pattern in two larger and entirely independent cohorts treated in Australia (ANZDATA Registry) and at the University Hospital of North Staffordshire, Stoke-on-Trent, UK (Figure 9). In order to rule out that this pattern was not due to differences in Gram staining (and hence endotoxin-related), we divided the group of HMB-PP+ bacteria further into Gram+ and Gram− species. Our outcome analysis demonstrates that even within the Gram+ group, bacteria capable of producing HMB-PP were associated with worse outcomes compared to HMB-PP− pathogens (Figure 9), suggesting that the HMB-PP producing capacity of the causative pathogen might be of predictive value for the clinical outcome from bacterial peritonitis (Table S3 in Text S1).
In order to identify potentially useful diagnostic and prognostic biomarkers of inflammation severity and outcomes from bacterial infection, we measured a large number of immunological parameters. These analyses identified elevated frequencies of peritoneal Vγ9/Vδ2 T cells on day 1 as possible predictor of subsequent technique failure within three months after infection (Table 1). Similarly, expression of the activation marker HLA-DR by peritoneal Vγ9/Vδ2 T cells on the day of admission was associated with increased mortality. No other parameters tested including the numbers and frequencies of neutrophils, monocytes or CD4+ and CD8+ T cells reached statistical significance (data not shown). Among soluble factors in peritoneal effluent, only elevated levels of TNF-α on day 1 indicated higher rates of technique failure and mortality (Table 1), while no such correlation was seen for other cytokines and chemokines, including GM-CSF, IFN-γ, IL-1β, IL-2, IL-6, IL-10, IL-12p70, IL-17, IL-22, CXCL8, CXCL10 and sIL-6R (data not shown).
Our findings of a rapid γδ T cell response to neutrophil-engulfed HMB-PP producing pathogens and its potential detrimental consequence in episodes of acute peritonitis may not only be of diagnostic and predictive value for affected patients, they also highlight possible new avenues of therapeutic intervention in bacterial infections. HMB-PP is an intermediate of the non-mevalonate pathway of isoprenoid biosynthesis, in which the first enzymatic step catalyzed by 1-deoxy-d-xylulose-5-phosphate reductoisomerase (Dxr) can be inhibited by fosmidomycin (Figure S8A in Text S1), a natural antibiotic produced by Streptomyces lavendulae [52], [53]. We therefore speculated that the effect of fosmidomycin pre-treatment of bacteria may serve a dual purpose in treating acute infections: by directly inhibiting an essential pathway in a broad range of pathogens and by abrogating HMB-PP driven inflammatory responses.
Tests with selected clinical isolates of common pathogens demonstrated that the majority of HMB-PP+ bacteria (Enterobacter cloacae, Klebsiella pneumoniae, Pseudomonas aeruginosa) was susceptible to overnight treatment with fosmidomycin (with the exception of Acinetobacter baumannii as expected [54]), with a mean inhibitory concentration (MIC) of 1–32 µg/ml depending on the strain (Table S1 in Text S1). Of note, fosmidomycin also acted on multidrug-resistant strains including bacteria harboring the recently discovered ‘New Delhi’ metallo-β-lactamase 1 (NDM-1) [55], [56] (Davey MS, Tyrrell JM et al., submitted for publication). In contrast to the efficient killing of most HMB-PP+ bacteria, the HMB-PP− bacteria Chryseobacterium indologenes, Enterococcus faecalis and Staphylococcus aureus were not affected by fosmidomycin (Table S1 in Text S1).
We next investigated the potential of short-term fosmidomycin treatment to affect γδ T cell activation by inhibiting the bacterial HMB-PP biosynthesis. Prior exposure of bacteria to fosmidomycin for 1 hour did not affect uptake by neutrophils as demonstrated using Escherichia coli-gfp+ (Figure S8B in Text S1), and neither did it affect gross bacterial viability as confirmed by re-plating treated Enterobacter cloacae on antibiotic-free plates in order to overcome the competitive inhibition by fosmidomycin (Figure 10A). Yet, pre-incubation of Escherichia coli, Enterobacter cloacae and Klebsiella pneumoniae with fosmidomycin for only 1 hour prior to phagocytosis by neutrophils clearly abrogated their capacity to stimulate Vγ9/Vδ2 T cells. This inhibitory effect on γδ T cell responses was evident both for activation of Vγ9/Vδ2 T cells in triple co-cultures with neutrophils harboring fosmidomycin-treated bacteria (Figure 10A and data not shown) as well as for activation (Figure S8B in Text S1) and proliferation (Figure 10B) of Vγ9/Vδ2 T cells in response to cell-free supernatants from neutrophils harboring fosmidomycin-treated bacteria. Together these results indicate that fosmidomycin not only has a direct antibacterial effect but also possesses immediate anti-inflammatory properties by inhibiting γδ T cell-driven responses (Figure 11), thus making the non-mevalonate pathway an attractive novel drug target for the treatment of acute infection.
Despite its relevance in early infection, the immediate crosstalk of γδ T cells, monocytes and neutrophils in the presence of bacterial pathogens has not been addressed in detail. This is particularly the case in humans who possess a unique γδ T cell population uniformly targeting an invariant non-self-metabolite, HMB-PP. Previous reports already associated the activation of Vγ9/Vδ2 T cells with the production of HMB-PP by microbes. This link was mainly based on the observation that Vγ9/Vδ2 T cell levels are often elevated in the blood of patients infected with HMB-PP producing pathogens [18] and that bacterial extracts prepared from those species activate Vγ9/Vδ2 T cells in vitro much better than extracts prepared from HMB-PP deficient micro-organisms [19], [35], [57]. Other investigators have speculated that Vγ9/Vδ2 T cells respond in vivo toward infected host cells with dysregulated isoprenoid metabolism leading to accumulation of isopentenyl pyrophosphate (IPP) regardless of the presence or absence of HMB-PP [58]. Here we unequivocally demonstrate that Vγ9/Vδ2 T cells respond to live bacteria upon phagocytosis by neutrophils, that this response is strictly HMB-PP dependent, and that it is amplified by the presence of monocytes providing crucial accessory signals. While it has remained puzzling how the immune system actually ‘sees’ an intracellular metabolite that is unlikely to be secreted or released by live micro-organisms, our findings show that biologically relevant traces of HMB-PP escape into the microenvironment after phagocytosis of extracellular bacteria by neutrophils. These conditions are likely to occur during the acute stage of the infection when Vγ9/Vδ2 T cells and monocytes are co-recruited to the site of inflammation [17] where they encounter neutrophils engaged in clearing invading pathogens (Figure 11).
The present findings explain how HMB-PP may become released at the site of infection. However, the molecular mechanism of HMB-PP recognition by Vγ9/Vδ2 T cells remains poorly understood. Our observation that monocytes were required for Vγ9/Vδ2 T cell responses to phagocytosed bacteria offers important clues. Monocytes and monocyte-derived macrophages or DCs were shown before to provide accessory help and may constitute a pivotal trigger for Vγ9/Vδ2 T cell responses to different bacterial pathogens. In the case of direct infection of monocytic cells, HMB-PP derived from intracellular bacteria may reach the cell surface bound to a presenting molecule [24], [25], [27]. In the case of extracellular bacteria, monocytes may take up or bind soluble HMB-PP released by professional phagocytes and present it to Vγ9/Vδ2 T cells (Figure 11). The HMB-PP presenting pathway remains elusive but may involve cell surface F1-ATPase [59], together with tight cell-cell interactions via LFA-1/ICAM-1 [31], [60], while it is independent of MHC class I, MHC class II, β2-microglobulin or CD1 [61]. Of note, any chemical modification of the molecular structure of HMB-PP abrogates its bioactivity by several magnitudes, such that the closely related natural metabolites IPP and DMAPP are >10,000 times less active in vitro [39], [40], [62], [63]. This is supported by our previous [35], [64], [65] and present demonstration that HMB-PP deficient bacteria (but which produce IPP and DMAPP) fail to stimulate cytokine production by Vγ9/Vδ2 T cells. Treatment with fosmidomycin or alkaline phosphatase abrogated the Vγ9/Vδ2 T cell responses to HMB-PP producing bacteria and emphasized the importance of HMB-PP for the induction of IFN-γ and TNF-α. However, fosmidomycin or alkaline phosphatase treated cultures as well as cultures involving HMB-PP deficient bacteria did show residual levels of CD69 expression, in line with a role for direct or indirect sensing of microbial TLR ligands [41], [42], [66] that is likely to amplify the overall response. In this respect it is intriguing that our present study identified a crucial role for TNF-α in supporting Vγ9/Vδ2 T cell proliferation, a cytokine which is readily produced not only by activated Vγ9/Vδ2 T cells themselves but also by neutrophils and monocytes exposed to microbial compounds such as LPS. This is in stark contrast to other cytokines produced by innate immune cells such as IFN-α and IFN-β which may induce upregulation of CD69 on Vγ9/Vδ2 T cells but fail to co-stimulate Vγ9/Vδ2 T cell proliferation [32]. Taken together, we identified an inflammatory crosstalk of Vγ9/Vδ2 T cells, neutrophils and monocytes in the presence of HMB-PP producing bacteria that can be manipulated at various check-points: (i) the antibiotic fosmidomycin abrogates the microbial HMB-PP production and thus renders bacterial pathogens invisible for Vγ9/Vδ2 T cells; (ii) alkaline phosphatase degrades free HMB-PP released by neutrophils into the microenvironment; (iii) blocking antibodies against the TCR prevent the recognition of HMB-PP by Vγ9/Vδ2 T cells; (iv) blocking antibodies against CD11a disrupt the LFA-1/ICAM-1 dependent contact between γδ T cells and monocytes needed for γδ T cell stimulation; (v) and sTNFR neutralizes soluble TNF-α which is released by all three cell types in response to microbial ligands and acts as growth factor for Vγ9/Vδ2 T cells and survival factor for neutrophils (Figure 11).
How does the HMB-PP dependent crosstalk of Vγ9/Vδ2 T cells, monocytes and neutrophils in vitro translate into the situation in vivo in acutely infected patients? Studies in patients with systemic inflammatory response syndrome suggested a significant role for Vγ9/Vδ2 T cells as early responders after severe insult and identified a correlation between Vγ9/Vδ2 T cell activation and clinical scores [67]. Our own findings in patients with PD-related peritonitis support this notion and demonstrate that the capacity of the causative pathogen to produce HMB-PP and local infiltrates of activated Vγ9/Vδ2 T cells on day 1 are indicative of acute inflammatory responses and may predict the subsequent clinical outcome from infection. It is becoming increasingly clear that the nature of the infection is a major determinant of outcome, and future interventions may well have to focus on subgroups of patients with different forms of infection [68]. A careful re-analysis of peritonitis outcomes from validated registry data [69]–[76] confirms that HMB-PP+ bacteria cause clinically more severe infection and emphasizes the need to pay more attention to detailed host-pathogen interactions. Bacterial infection remains a leading cause of morbidity and mortality worldwide, not the least due to the alarming spread of antibiotic-resistant pathogens that is posing an enormous challenge on clinical practice, public healthcare and biomedical research [55], [56]. In most cases, antimicrobial treatment is largely empirical as microbiological culture results are typically not available for 2–3 days. Moreover, many times no organism can be identified, with rates of culture-negative infections occasionally reaching 50% [77]. In contrast, laboratory analyses of immune cells and/or soluble mediators may provide valuable information within a few hours and both aid diagnosis and refine treatment. In this respect, fosmidomycin or related HMB-PP inhibitors might constitute useful combination partners for antibiotic therapy especially in critical cases such as acute peritonitis or sepsis where the intervention has to commence before the nature of the causative pathogen is known, and where activated Vγ9/Vδ2 T cells may contribute to poor clinical outcome. Fosmidomycin targets an essential pathway in a broad range of pathogens [54], [78], [79] and simultaneously abrogates Vγ9/Vδ2 T cell responses [80], [81]. Of note, the inhibitory effect of fosmidomycin on the Vγ9/Vδ2 T cell bioactivity was detectable after only 1 hour of treatment and well below its MIC, i.e. under conditions that would easily be achievable in patients. Other ways of specifically manipulating γδ T cell mediated responses may include the use of anti-γδ TCR antibodies or γδ TCR antagonists such as BrH-PCP [82]. Given the importance of TNF-α for γδ T cell proliferation and the association of peritoneal TNF-α levels with morbidity and mortality, one may also advocate the use of TNF-α blocking reagents for the treatment of acute PD-related peritonitis [83].
Taken together our experiments demonstrate that Vγ9/Vδ2 T cells recognize with HMB-PP a small common molecule released by the majority of invading bacteria when they become phagocytosed by neutrophils. Stimulation of Vγ9/Vδ2 T cells at the site of infection is likely to amplify the local inflammatory response with important consequences for pathogen clearance and the development of microbe-specific immunity. However, if triggered at the wrong time or the wrong place, this rapid reaction toward most bacteria may also lead to inflammation-related damage and detrimental clinical outcome. These findings improve our insight into the complex cellular interactions in early infection, identify novel biomarkers of possible diagnostic and predictive value and highlight new avenues for therapeutic intervention.
This study was conducted according to the principles expressed in the Declaration of Helsinki and under local ethical guidelines (Bro Taf Health Authority, Wales). The study was approved by the South East Wales Local Ethics Committee under reference number 04WSE04/27. All patients provided written informed consent for the collection of samples and subsequent analysis.
The Cardiff study population included 39 adult patients who were receiving PD at the University Hospital of Wales, Cardiff, UK, and were admitted with acute peritonitis between September 2008 and October 2010 (Table S2 in Text S1). Eight stable patients with no infection in the previous 3 months were included in this study as non-infected controls. In addition, microbiological and survival data were obtained from all 739 adult patients who were receiving PD between 1987 and 2008 at the University Hospital of North Staffordshire, Stoke-on-Trent, UK; and from all 2,542 Australian adult patients from the Australia and New Zealand Dialysis Transplant (ANZDATA) Registry who were receiving PD between 2003 and 2008 (Table S2 in Text S1). Diagnosis of acute peritonitis was based on the presence of abdominal pain and cloudy peritoneal effluent with >100 WBC/mm3. Infections were grouped into culture-positive and culture-negative episodes, according to the result of the microbiological analysis of the effluent. Bacteria species identified in culture-positive infections were grouped into HMB-PP+ and HMB-PP−, based on the presence or absence of HMB-PP in the microbial metabolism [17], [19]. Endpoints of outcome analyses were 14th and 90th day mortality and technique failure (catheter removal, transfer to hemodialysis, and/or patient death). In order to rule out a history of previous antibiotic treatment as a potentially confounding factor, outcome studies were restricted to patients with first-episode peritonitis, excluding cases of fungal infection or unrecorded culture results.
Bacteria strains used in this study are listed in Table S1 in Text S1. Mycobacterium smegmatis strains were grown aerobically at 37°C in liquid Lemco medium (Oxoid) supplemented with 10 mg/ml peptone, 5 mg/ml NaCl and 0.25% Tween 80 (Sigma), or on solid Lemco plates with 15 mg/ml agar (Fisher). Listeria innocua strains were grown aerobically at 37°C in liquid brain heart infusion medium (Oxoid) or on agar plates. Escherichia coli laboratory strains and multi-drug resistant clinical isolates of Acinetobacter baumannii, Chryseobacterium indologenes, Enterobacter cloacae, Enterococcus faecalis, Klebsiella pneumoniae, Pseudomonas aeruginosa and Staphylococcus aureus were grown in liquid LB broth and on solid Columbia blood agar (Oxoid). Where appropriate, antibiotics were added to the medium: M. smegmatis-gcpE+, 100 µg/ml hygromycin B; L. innocua-gcpE+, 7.5 µg/ml chloramphenicol; E. coli-gfp+, 100 µg/ml ampicillin (all from Sigma). Bacterial susceptibilities to fosmidomycin were determined by microbroth dilution method, according to the Clinical and Laboratory Standards Institute guidelines [84]. A log 2 dilution series of 0.06 to 128 µg/ml allowed the identification of the minimal inhibitory concentration (MIC) where bacterial growth was absent. No defined break points have been acknowledged for fosmidomycin [85], therefore resistance was defined as concentrations >128 µg/ml.
Peritoneal cells were harvested from chilled overnight dwell effluents [31], [47]; cell-free supernatants were stored at −70°C. PBMC were isolated from peripheral blood of healthy volunteers using Lymphoprep (Axis-Shield). Vγ9+ T cells (>98%) were purified from PBMC using monoclonal antibodies (mAbs) against Vγ9-PE-Cy5 (Immu360; Beckman-Coulter) and anti-PE microbeads (Miltenyi). Monocytes (>98%) were purified using anti-CD14 microbeads (Miltenyi). Neutrophils (>95%) used for bacterial phagocytosis were isolated from peripheral blood using a Lymphoprep gradient followed by dextran sedimentation [86]. Remaining erythrocytes were lysed with ammonium chloride solution (150 mM NH4Cl, 10 mM KHCO3, 0.1 mM EDTA), and neutrophils were washed in HBSS without Mg2+ and Ca2+ (Sigma) and resuspended to a final cell concentration of 2×106/ml in HBSS with Mg2+ and Ca2+ (Sigma) supplemented with 10% human serum. Neutrophils (>98%) used in Vγ9/Vδ2 T cell co-culture experiments were isolated from peripheral blood by initial dextran sedimentation followed by centrifugation through discontinuous Percoll gradients [87]. The cell culture medium used throughout this study was RPMI-1640 with 2 mM L-glutamine, 1% sodium pyruvate, 50 µg/ml penicillin/streptomycin and 10% fetal calf serum (Invitrogen).
Single colonies were grown in culture broth for 18 hours, and bacteria were washed in PBS and resuspended in HBSS with Mg2+ and Ca2+ supplemented with 10% human serum. Freshly isolated neutrophils were incubated with bacteria at a multiplicity of infection (MOI) of 0.1–100 bacteria per neutrophil for 30–60 min at 37°C, with gentle shaking. In some experiments, bacteria were pre-treated with 0.5–25 µg/ml fosmidomycin for 1 hour prior to phagocytosis. Actual MOIs of all bacterial inocula used were determined by plating out serial dilutions on agar plates and expressed as colony forming units (CFU) per neutrophil. Non-phagocytosed bacteria were washed off three times. For microscopic analyses, neutrophils harboring GFP-expressing bacteria were washed, counterstained with DAPI (Sigma) and fixed in 2% paraformaldehyde. Images were acquired on a Nikon Eclipse 80i fluorescence microscope equipped with a Nikon DXM 1200F camera and processed with Adobe Photoshop. For the generation of cell-free supernatants, neutrophils pre-incubated with bacteria as described above were cultured for 5 hours in complete RPMI-1640. Supernatants were then harvested and cells removed by centrifugation at 12,000 g for 10 min. Samples were stored at −20°C and thawed a maximum of 5 times. For some experiments, neutrophil supernatants were treated with 0.015 U/µl shrimp alkaline phosphatase for 30 min at 37°C.
Unless indicated otherwise, neutrophils were co-cultured for 20 hours in complete RPMI-1640 medium (further supplemented with 8 µg/ml colistin (Sigma) for assays involving multi-drug resistant clinical isolates) with autologous monocytes and γδ T cells at a ratio of 10 neutrophils and 1 monocyte per 1 γδ T cell (10∶1∶1), in the absence or presence of 0.015 U/µl shrimp alkaline phosphatase (Roche). Proliferation assays using γδ T cells that had been pre-labeled with CFSE (Molecular Probes) were incubated for 4–6 days. Controls included co-cultures in the absence or presence of 1–100 nM synthetic HMB-PP [40] or 1–100 ng/ml LPS from Salmonella abortus equi (Sigma). In transwell experiments, neutrophils were separated from γδ T cells by 0.4 µm pore polycarbonate membranes (Fisher Scientific). Cell-free supernatants derived from neutrophils after phagocytosis of bacteria were tested in monocyte-γδ T cell co-cultures (1∶1) at a dilution of 1 in 3. Blocking reagents used were anti-IFN-γ (25718; R&D Systems); anti-CD11a (TS1/22) from Dr Ruggero Pardi (DIBIT-Scientific Institute San Raffaele, Milano, Italy); and sTNFR p75-IgG1 fusion protein (Enbrel; Amgen); alone or in combination at 10 µg/ml each. Anti-TCR-Vγ9 (Immu360; Beckman Coulter) was used at 1.25 µg/ml.
Cells were acquired on an eight-color FACSCanto II (BD Biosciences) and analyzed with FloJo 7.6 (TreeStar), using monoclonal antibodies against CD3 (UCHT1), CD15 (HI98), CD25 (M-A251), CD62L (DREG-56), CD69 (FN50), CD86 (2331) and HLA-DR (L243) from BD Biosciences; TCR-Vγ9 (Immu360) and CD40 (mAB89) from Beckman Coulter; and CD11b (ICRF44) and CD14 (61D3) from eBioscience; together with appropriate isotype controls. Cells of interest were gated based on their appearance in side scatter and forward scatter area/height, exclusion of live/dead staining (fixable Aqua; Invitrogen) and surface staining: CD3− CD14− CD15+ neutrophils, CD3− CD14+ CD15− monocytes, and CD3+ Vγ9+ γδ T cells. Apoptotic cells were identified using Annexin V (BD Biosciences). For detection of intracellular cytokines, 10 µg/ml brefeldin A (Sigma) was added to cultures 5 hours prior to harvesting. Surface-stained cells were labeled using the Fix & Perm kit (eBioscience) and monoclonal antibodies against IFN-γ (45.15; BD Biosciences) and TNF-α (188; Beckman Coulter).
Soluble cytokines in cell culture supernatants were detected using ELISA kits for IL-1β, IL-6 and IL-17 (R&D Systems); and IFN-γ, TNF-α and CXCL8 (BD Biosciences). All samples were measured in duplicate on a Dynex MRX II reader. Cell-free peritoneal effluents were analyzed on a SECTOR Imager 600 (Meso Scale Discovery) for TNF-α, GM-CSF, IFN-γ, IL-1β, IL-2, IL-6, IL-10, IL-12p70, CXCL8 (IL-8) and soluble IL-6 receptor (sIL-6R). In addition, IL-17, IL-22 and CXCL10 in peritoneal effluents were measured in duplicate on a Dynex MRX II reader, using conventional kits (R&D Systems).
Data were analyzed using two-tailed Student's t-tests (GraphPad Prism 4.0), and considered significant as indicated in the figures and tables: *, p<0.05; **, p<0.01; ***, p<0.001. Cumulative survival curves as a function of time were generated using the Kaplan-Meier approach and compared using the log rank test (SPSS 16.0).
|
10.1371/journal.pgen.1001393 | Comparative Analysis of Proteome and Transcriptome Variation in Mouse | The relationships between the levels of transcripts and the levels of the proteins they encode have not been examined comprehensively in mammals, although previous work in plants and yeast suggest a surprisingly modest correlation. We have examined this issue using a genetic approach in which natural variations were used to perturb both transcript levels and protein levels among inbred strains of mice. We quantified over 5,000 peptides and over 22,000 transcripts in livers of 97 inbred and recombinant inbred strains and focused on the 7,185 most heritable transcripts and 486 most reliable proteins. The transcript levels were quantified by microarray analysis in three replicates and the proteins were quantified by Liquid Chromatography–Mass Spectrometry using O(18)-reference-based isotope labeling approach. We show that the levels of transcripts and proteins correlate significantly for only about half of the genes tested, with an average correlation of 0.27, and the correlations of transcripts and proteins varied depending on the cellular location and biological function of the gene. We examined technical and biological factors that could contribute to the modest correlation. For example, differential splicing clearly affects the analyses for certain genes; but, based on deep sequencing, this does not substantially contribute to the overall estimate of the correlation. We also employed genome-wide association analyses to map loci controlling both transcript and protein levels. Surprisingly, little overlap was observed between the protein- and transcript-mapped loci. We have typed numerous clinically relevant traits among the strains, including adiposity, lipoprotein levels, and tissue parameters. Using correlation analysis, we found that a low number of clinical trait relationships are preserved between the protein and mRNA gene products and that the majority of such relationships are specific to either the protein levels or transcript levels. Surprisingly, transcript levels were more strongly correlated with clinical traits than protein levels. In light of the widespread use of high-throughput technologies in both clinical and basic research, the results presented have practical as well as basic implications.
| An old dogma in biology states that, in every cell, the flow of biological information is from DNA to RNA to proteins and that the latter act as a working force to determine the organism's phenotype. This model predicts that changes in DNA that affect the clinical phenotype should also similarly change the cellular levels of RNA and protein levels. In this report, we test this prediction by looking at the concordance between DNA variation in population of mouse inbred strains, the RNA and protein variation in the liver tissue of these mice, and variation in metabolic phenotypes. We show that the relationship between various biological traits is not simple and that there is relatively little concordance of RNA levels and the corresponding protein levels in response to DNA perturbations. In addition, we also find that, surprisingly, metabolic traits correlate better to RNA levels than to protein levels. In light of current efforts in searching for the molecular bases of disease susceptibility in humans, our findings highlight the complexity of information flow that underlies clinical outcomes.
| An underlying assumption in many biological studies is the concordance of transcript and protein levels during the flow of information from DNA to phenotype. Clearly, protein levels are greatly influenced by post-translational processing and inherent variations in stability but, in general, it is assumed that perturbations of transcript levels are substantially correlated with protein levels. The extent to which this occurs, however, remains poorly understood and understanding the relationships across scales, from DNA to phenotype, has both practical and basic implications. For example, “genetical genomics” studies examine transcript levels as a function of genetic variation and use this information to construct models, such as interaction networks, to explain complex phenotypes [1]–[8]. Systems based approaches, in particular, have relied heavily on transcriptome data [9].
Concordance of protein and transcript levels has been studied in yeast and plants. A recent comparative study in a yeast segregating population showed that there is a significant but modest correlation between transcript and protein levels [10]. Moreover, this report also found that, in general, loci that influence protein abundance are different from those affecting transcript abundance. A similar comparative analysis of molecular phenotype mapping in Arabidopsis [11] was reported subsequently. In this report the authors investigated the commonality of hotspot loci (defined as loci affecting a large number of traits within each biological class) across various biological scales and observed a general theme consistent with the phenotypic buffering of perturbations affecting molecular phenotypes as one looks to scales further away from the DNA variation (e.g. proteome vs transcriptome). Both of these reports emphasize the value gained from bringing together information from various biological scales, as each dataset will add new information to the phenotypic effect of DNA variation.
We now report global analysis of transcript-protein relationships in mice using a genetic approach involving thousands of naturally occurring perturbations. For this, we have utilized a recently developed panel of permanent inbred strains of mice, termed the Hybrid Mouse Diversity Panel (HMDP), that allows high resolution mapping of complex traits [12]. We chose to examine protein and transcript levels in liver given the importance of the organ in metabolic traits relevant to disease.
The experimental design of our study is depicted in Figure 1. To study the relationship between transcript and protein levels globally, we examined 97 inbred strains of mice of the HMDP representing a wide range of genetic diversity, including ∼11,000,000 single nucleotide polymorphisms as well as copy number variations [13], [14]. As we have shown previously, this population includes thousands of expression quantitative trait loci (eQTL) that can be mapped in the population using association analysis with correction for population structure using a mixed model algorithm [12]. The resolution achieved in this way is, on average, one to two orders of magnitude narrower than that using linkage analysis [12]. Livers from the 97 strains were quantitatively analyzed for global transcript levels using the Affymetrix HT-MG-430A platform and for protein levels using LC-MS employing AMT tag approach for identification and 16O/18O labeling for quantification [12], [15]. In the latter, each individually processed and unlabeled sample is spiked with the 18O labeled “universal” reference pool (i.e. the pool made from mixing together the same amount of isolated proteins from all samples) providing an internal standard for accurate measurement of protein abundance across biological samples. This dual-quantification, which combines the label-free and isotope labeling techniques, has been shown to be significantly superior over label-free methods in terms of quantification precision [15] and offers a simple, robust, and a more precise alternative to other proteomic techniques for studying variations in protein levels across large biological samples. In the LC-MS dataset, we also included 10 technical replicates from the C57BL/6J strain to measure the reproducibility of the sample preparation and technology which we describe in detail below.
One technical issue for proteomic analysis relates to how peptides with non-synonymous coding SNPs would present themselves in the correlation and association analyses. In order to annotate peptides detected by the tandem LC- MS/MS fragmentation patterns, the mass spectra are matched against a pre-established known reference database. In our case this database, which was built from the pool of all the inbred strains in the HMDP panel, was created by annotating the peptides against the reference sequence (C57BL/6J strain) followed by filtering out those peptides which have non-synonymous coding SNPs documented in public database for any of the HMDP strains. As a result of this preprocessing step, the peptides identified by LC-MS were limited to those that did not contain known non-synonymous SNPs in their amino acid sequences.
From the original 5363 peptides measured, we selected peptides that a) had less than 50% missing measurements in the whole population, b) had no internal lysine or arginine, and c) aligned uniquely to one Ensembl gene. Fifty four percent of peptides (2893 peptides) passed these initial selection criteria. To assess the quality of the measurements, we investigated the amount of technical noise in the peptides selected. Having the control technical replicates allowed us to measure the reproducibility of the LC-MS measurement and assess whether the variation in the levels of the selected peptides in the HMDP population was due to technical or genetic variation. The distribution of the variance in the control mice and in the HMDP panel are shown in Figure 2A (the blue histogram). The mean and median across the ten replicates were 0.19 and 0.08 (the grey histogram), respectively, suggesting that, for most peptides, the measurements were robust. In contrast, the distribution of the variance was much broader in the genetic population where the mean and median of variances across all the peptides were 0.2 and 0.3 respectively (Figure 2A, the blue histogram).
The relationship between RNA levels and peptide levels across the HMDP genetic perturbations would be a function of the genetic variation in the peptide levels as well as the degree of nongenetic/technical variations in peptide quantification. Thus, we defined a “signal to noise” measure for each peptide as the ratio of the total variance in the HMDP over the variance in the ten replicates. The variance in the ten replicates would be due to nongenetic biologic variance as well as technical variance (herein termed “noise”) while the total HMDP variance would include genetic variance, nongenetic biologic variance, and technical variance. Accordingly, a large value of signal to noise could either mean large genetic variation, small nongenetic variation, or both. Conversely, a smaller value for signal to noise would either mean small genetic variation, large nongenetic variation, or both. As can be seen in Figure 2B, signal to noise ratios varied significantly across different peptides.
We complemented the LC/MS studies for a small set of proteins (11) by performing immunoblot quantitation in 9 of the HMDP strains. Over half of the peptides exhibited significant discrepancies in relative levels using the two methods and those with small “signal to noise” ratios (small genetic variation and/or large noise component) exhibited reduced correlations with the immunoblotting results (p-value = 3.3×10−5). Although immunoblotting is semi-quantitative, these results suggest that peptides with a large signal to noise ratio will provide the best estimates of the true relationship between RNA levels and protein levels. (Immunoblot results are presented in Table S1, Table S2, and Figure S1).
Transcript levels in inbred stains were measured by profiling three mice for each strain, using Affymetrix MOE430A platform, and taking the average of expression over the three biological replicates. This design provided us with an opportunity to a) better estimate the “true” values for the mRNA levels in each strain and b) estimate the heritability of each probeset across the HMDP population. The distribution of heritability estimates is shown in Figure 2C. Consistent with previous reports [16], we detected a broad spectrum of heritability estimates for the transcript levels ranging from 0.07 to 0.95. Using ANOVA, we assigned significance to the heritability values obtained for each probeset and found that for as many as 50% of the probesets (11248 probesets), there was a significant (p-value<0.05 for strain term) genetic component affecting the transcript levels.
Previous reports have documented conflicting results about the reproducibility measurements generated by microarray platforms [17]–[21]. To investigate this in our dataset, we compared the expression levels measured by the Affymetrix microarray to the expression levels measured by next generation sequencing (NGS) in small subset of inbred strains. Using Illumina's Genome Analyzer we profiled the liver transcriptome of one C57BL/6J and one DBA/2J mouse and generated ∼17,000,000 sequences for the fragmented mRNAs for each strain. Using Tophat [22] and Cufflink [23] algorithms, we were able to uniquely align and count 4,800,000 and 7,000,000 sequences for C57BL/6J and DBA/2J respectively (See Materials and Methods). After sequence alignment and quantification of the read counts, we compared the transcriptome of the each strain against the microarray data generated by the Affymetrix MOE430a platform (Figure 2D). This comparison revealed a high concordance between the data obtained from two technologies (Spearman correlation coefficient of 0.69 for both C57BL/6J and DBA/2 samples). These results were similar to a previous cross-platform comparative study [24] and indicated that, for most transcripts, microarray data produce a highly reliable estimate of transcript levels.
Based on the results reported above, in order to enrich for high quality data and provide a better estimate of true relationships between transcript and protein data, we focused on transcript and peptide data with significant genetic and biological variation. For peptide data, we used the signal to noise ratio parameter to further filter the noisy peptides from our dataset. The cutoff we chose for filtering peptides was a signal to noise ratio of 2. After removing the “noisy” measurements, we were left with 1543 peptides (see Dataset S1 for the expression and Dataset S2 for the annotation of the peptides). These 1543 peptides represented 486 Ensembl Genes from which 39% were represented by only one peptide and the remaining 60% were represented by two or more peptides (Figure 2E). The most abundant number of peptides (>20 peptides per gene) was found for 2 genes, aldehyde dehydrogenase 1 family, member L1 (Aldh1l1), and carbamoyl-phosphate synthetase 1 (Cps1). For the transcript data, we focused only on those transcripts that had a) a significant genetic component underlying their variation (heritability p-value<0.05), and b) unambiguous annotation in the Ensembl database. To comprehensively compare the transcripts and peptides, we also included those probesets that were annotated as the same Ensembl gene/transcript as one of the peptides in the protein data. This resulted in the total of 9896 probesets (representing 7185 Ensembl genes) from the initial 22670 probesets.
To investigate the range of gene products present in the filtered datasets, we generated a separate list of “GO Slim” terms for each of the three major GO categories (Cellular Compartment or “CC”, Molecular Function or “MF”, and Biological Process or “BP”) and used the “GO Term Mapper” website (http://go.princeton.edu/cgi-bin/GOTermMapper) to classify and count the number of proteins and transcripts in each of the 3 major GO categories (Table S3). We also compared our results to the background set (all genes annotated by Mouse Genome Informatics) and using Fisher's Exact test calculated the degree of enrichment or underrepresentation for each GO class. Figure S2 depicts the results of this analysis for the Cellular Compartment GO terms. While the proteome and transcriptome datasets represent a wide range of gene products present in various cellular compartments, the compartments are not equally represented. In the protein data, mitochondrial genes were overwhelmingly the largest set and in the transcript data the nuclear compartment was at the top of the list for enriched CC GO terms. The enrichment analysis showed that for both protein and transcript datasets the majority of the GO terms tested were significantly over and under-represented and these differences were more pronounced for the protein data (Table S3). A likely explanation for this observation is that the LC-MS analysis of the liver provides a biased sampling of the proteome data (due to the abundance and/or cellular location of the protein). The significant differences between the transcript data and the background set may be partially explained by the bias introduced in the design of the Affymetrix microarray. Alternatively, since the transcript data by the virtue of filtering represent the significantly heritable subset of all transcripts present on the Affymetrix chip, one could postulate that in some cellular compartments or cellular processes transcripts are more or less likely to exhibit common genetic variation.
We next examined the degree of concordance between the transcript and protein levels. For this, we compared the transcript and peptide measurements for every peptide-probeset pair that mapped to the same Ensembl gene. This “gene-level” analysis included 2010 peptide-probeset pairs (1342 peptides and 607 probesets) representing 396 Ensembl genes. Figure 3A shows the correlation coefficient distribution for these 2010 peptide-probeset pairs. Highly significant positive correlation (p-value<1e-06, r>0.46) between RNA and protein was found for 21% of the genes (85 out of 396) and ∼15% of the peptide-probeset pairs (291 out of 2010). The most significant correlation (r = 0.87) was found for the glyoxalase 1 gene (Glo1) where the peptide and transcript of this gene correlated (Figure S3). Overall, we found that the relationship between mRNA and protein levels was modest (mean r = 0.27) and for 39% of the pairs (761 of 2010) the mRNA and protein levels did not correlate significantly at the nominal 0.05 p-value threshold. Our estimate of average correlation between mRNA and protein was slightly higher than those reported in other organisms, perhaps due to recent improvements in the LC-MS technology and/or statistical power.
Next, we examined if the amount of technical noise and/or lack of genetic variation could explain the modest correlation between mRNA and peptide data. For this we classified each peptide based on the signal to noise ratio (defined earlier) and looked at the median correlation between mRNA and peptides within each group. As shown in Figure 3B, we found that as the ratio of signal to noise increases so does the correlation between the mRNA and peptide levels of the gene. In fact the median correlation for the least noisy group, comprising peptides with signal to noise ratio >90%, was twice as large as the noisiest group of peptides (peptides with signal to noise ratio <60%). These results suggests that the modest correlation between peptides and mRNA observed in our study is partially due to either the presence of significant nongenetic variation or small genetic variation in some proteins.
Aside from lack of genetic variation in peptides, another plausible explanation for the lack of high correlation between peptides and probesets could be the analytic approach chosen to calculate correlations. In our study, we estimated the relationship between mRNA and proteins by examining the correlations between pairs of peptides and probesets that were annotated to the same gene without considering the isoform information for that gene. The choice of analytic approach presented here was mainly due to the limitation of the technology we used to measure the transcript levels. The probesets on the Affymetrix microarrays are designed to hybridize mainly to the transcripts 3′ end. Such design will fail to accurately measure the levels of isoforms which are identical at the 3′ end but are differentially regulated at the transcript level. The inability to measure isoform specific expression can clearly impact the mRNA-protein correlation results for certain peptides which represent specific isoforms as LC-MS data may include peptides unique to a gene's isoform. Figure 4A and 4B illustrate an example of differential isoform regulation identified in the LC-MS data. Acox1 (acyl-Coenzyme A oxidase 1, palmitoyl) is a peroxisomal gene involved in fatty acid beta-oxidation pathway and metabolism of very long chain fatty acids, and its deficiency causes pseudoneonatal adrenoleukodystrophy [25] in humans. This gene produces four protein-coding products (Acox1-001, Acox1-002, Acox1-003, and Acox1-201 as denoted in Ensembl genome browser) shown in Figure 4A (bottom panel). All isoforms except for “Acox1-002” include exon 4 of this gene. In LC-MS data, 20 peptides were measured for this protein. One of these 20 peptides (“GHPEPLDLHLGMFLPTLLHQATEEQQER”) maps to the exon 4 sequence of this gene, thus, does not represent the “Acox1-002” isoform which skips this exon. Examining the expression profile and correlation of these 20 peptides revealed that all peptides representing “Acox1-002” isoform are highly intercorrelated (mean r = 0.86, Figure 4B) and exhibit a similar expression profile (Figure 4A, top panel), but none have either similarity in expression profile or significant correlation with the peptide mapping to the exon 4 which is skipped in Acox1-002 isoform (mean r = 0.23, Figure 4A top panel, and 4B). This suggests that Acox1-002 isoform (with the skipped exon 4) is the main isoform underlying the significant correlation among 19 of the 20 peptides identified by LC-MS in our genetic population. This example illustrates that the LC-MS data contain information on differential regulation of isoforms, in contrast to the microarray data.
To investigate if our inability to measure isoform specific expression by microarrays could explain the lack of concordance between mRNA and peptides, we utilized the next generation sequencing data generated for the two inbred strains described earlier. This dataset provided us with an opportunity to examine the transcript level expression of the exons measured by NGS with the protein level expression of exons measured by the LC-MS. To investigate this, for each peptide we calculated the count of exons in RNA-Seq data for two strains. We then compared the DBA to B6 ratio of each exon in the peptide data to the DBA to B6 ratio of normalized sequence counts (reported in FPKM units) in the RNA-Seq data. The results of these comparisons are shown in Figure 4C. Similar to the gene-level analysis, the exon level analysis of all the filtered data also suggested a modest relationship between the exon counts in the mRNA data vs exon levels in the LC-MS data (r2 = 0.02). This global analysis provided no support for the presence of differential splicing/isoform regulation as being a significant factor in the mRNA and protein overall relationships observed between LC-MS and microarray data. The relationship between RNA-Seq data and LC-MS peptide data is particularly strong (r = 0.42) for those peptides exhibiting a strong correlation (r>0.5) with microarray data (Figure 4C).
In an alternate approach to study the effect of differential splicing on the correlation pattern, we examined our LC-MS data at the isoform level and compared the results to the gene level analysis. For this, we grouped various peptides of each protein into unique and mutually exclusive clusters of known isoforms as defined by the Ensembl database. In this classification we allowed peptides to only represent one Ensembl protein ID and excluded any peptide which matched with two or more Ensembl proteins. Focusing on clusters with at least two peptides assigned to a cluster, we calculated the within cluster correlation of of peptides and compared the average within-cluster correlation to the average correlation of peptides at the gene level analysis. The average correlation of peptides at the gene level analysis was estimated at 0.47. In comparison, the average within cluster correlation of peptides representing the same isoforms was estimated to be 0.52. Combined with the NGS analysis described earlier, the small and nonsignificant increase in the peptide concordance after taking into account the isoform membership provides little support for differential regulation of splicing/isoform expression as a significant factor underlying the observed modest correlation between transcripts and proteins.
In light of the modest correlation observed between the transcript and protein pairs, we examined the relationship of each of these two datasets with clinical traits. In our HMDP panel, we have previously measured a set of 42, some interrelated, metabolic traits (see Materials and Methods). In this analysis, in order to make a direct comparison across the two datasets, we once again focused on the 396 genes for which we had at least one peptide and one transcript measurement. At the 5% false discovery rate, we observed that three quarters of probesets (457 from the total 607) significantly correlated with at least one of the clinical traits. In contrast, at the same false discovery rate, only 28% of the total peptides (380 out of 1342) showed significant correlation with at least one of the 42 phenotypes. Despite the fact that the starting number of peptides was twice the number of probesets (1342 and 607), the total number of significant correlations for the peptides was only about half the number found for the probesets (2206 vs 1107). The same biased pattern was also observed at other statistical thresholds as shown in Table 1. In addition to probeset-pair analysis, we also carried a similar analysis at the gene level to estimate what fraction of starting genes (396 total genes) a) exhibit consistent relationship with clinical traits both at the transcript level and the protein level b) exhibit trait relationships unique to either of the two molecular phenotypes. From the 396 genes, 325 genes had at least one significant correlation at the 5%FDR with clinical phenotypes and 162 had at least one significant correlation with phenotypes at the protein level (Table 1). At the transcript level, the total number of significant correlations amounted to 1781 vs 556 found at the protein level. From these, 234 relations were found to be common for transcript and protein of the same genes and 1547 were unique to transcripts only (Figure 5A). Despite this overwhelming bias toward better correlation of transcripts, we also found 322 unique relations at the protein level (Table 1 and Figure 5A). Altogether, about half the significant protein-trait correlations also exhibited transcript-trait correlations, but only 15% of the significant transcript-trait correlations exhibited corresponding protein-trait correlations.
We sought to examine whether the concordance between protein and transcript data was dependent on the biological function and/or cellular location of the gene product. For this we restricted the list of genes within each of the 3 major GO_slim terms described earlier to the 396 genes for which we had at least one probeset and one peptide measured. We then defined the average relationship between protein and transcript products of the genes within each GO category by computing the correlation between the gene products and taking the average of these correlations. The three panels in Figure S4 show the average correlations of the transcript and protein product of the genes grouped by their assigned GO categories. Striking differences in the concordance between proteins and transcripts across some of the GO categories were observed. For example, for Cellular Compartment GO terms (CC), we found that peroxisomal and ER genes have on average a better correlation between protein and transcript products than other cellular compartments. We also found that for some of the GO categories the similarity between protein and transcript levels was almost non-existent (for example in BP the “cell growth” class, Figure S4C). To assess the significance of these observations, for each GO class we created 100,000 bootstrap datasets (each the size of the number of genes assigned to the respective GO category) containing correlation coefficient p-values randomly selected from the pool of peptide-probeset correlation p-values. We then assessed the significance of observed averaged correlation p-value for each class by comparing it to the distribution of the averaged p-values in the bootstrapped dataset (Table S4). In some GO groups, we found a class of genes for which the relationship between the transcript level and protein level is significantly better than for other GO groups. We also found GO classes in which the transcript levels and protein levels of the genes were significantly discordant (i.e. the relationship between protein and transcript was significantly less than what would be expected by chance). For example, in MF we found that genes classified as having a role in “electron carrier activity” had a strong relationship among the protein and transcript levels (p-value = 8.9e-03) and this relationship is significantly compromised for genes with “transporter activity” (p-value = 4e-05). Another example of discordant group was genes involved in the translation process (p-value<1e-05). Interestingly, the “translation” category has been proposed recently to be involved in phenotypic buffering in a yeast genetic interaction network. Overall, these results indicate that cellular compartments and biological processes vary in the degree to which the linear relationships between transcript levels and their protein products are conserved.
We also examined the level of concordance among transcripts and proteins of genes that are members of the same biological pathway. For this, we focused on 212 biological pathways on the KEGG website (http://www.genome.jp/kegg/). We annotated the peptides and probesets according to their pathway membership as determined by their Ensembl gene IDs. Ninety nine out of 212 pathways contained genes for which we had both more than one transcript and more than one protein measured. Focusing on these 99 pathways, we then performed the following three correlation analyses: 1) correlation between peptides belonging to the same pathway, 2) correlation between probesets belonging to the same pathway, 3) correlation between probesets and peptides belonging to the same pathway. Comparing the results of these three analyses suggests that overall within-transcript correlation of biological pathway genes is higher than within-protein correlations (0.20 vs 0.14 mean Spearman correlation coefficients) (Figure 5B), and transcript-protein correlations are the weakest of all (mean Spearman correlation coefficient of 0.11). From the 99 pathways, 79 pathways had better between-transcript correlations than between-protein correlations and 20 had better between-protein correlations. We also observed that for most pathways when there was good concordance between the transcripts there was also good concordance between the peptides of that pathway (r = 0.39, Figure 5B).
Next, we examined the genetic loci regulating protein and transcript levels. It is known that the presence of SNP within probe sequence can affect hybridization of the mRNA [26], leading to both type-I and type-II errors in the genomewide association analysis. In our Affymetrix dataset, as expected, we also observed a significant effect of SNPs on genomewide association results for fraction of the probes, as judged by comparing the significance level for local eQTLs between probesets before and after masking of probes containing publicly available SNPs (see Text S1 and Table S5 for details, and Dataset S3 for the list of probes which were masked from each probeset due to the presence of SNP). To minimize this technical artifact, we removed all the SNP-containing probes from their corresponding probeset before normalization of the data and eliminated all the probesets which contained 8 or more probes with SNPs (∼300 probesets fell in this category). Therefore, all the data reported below were generated from masked probesets.
We performed genomewide association on both transcriptome and proteome datasets using 95,854 SNPs with minor allele frequencies greater than 10% obtained from the Broad Institute (http://www.broadinstitute.org/mouse/hapmap) and Wellcome Trust Center (WTCHG) (see Materials and Methods for details). To account for the population structure and genetic relatedness among strains in the genome-wide association mapping, we utilized the Efficient Mixed Modeling Algorithm (EMMA) [27]. Furthermore, haplotype analysis of the inbred strains has shown the presence of over 60,000 haplotype blocks of varying size throughout the genome of inbred strains [28]. Since the presence of these blocks could be a source for overestimation of extent of genetic regulation and false positive associations, for each transcript and protein we removed significant associations due to high linkage disequilibrium (defined as R-squared of 0.5 or larger between genotypes). Since the transcript and protein data have different variance properties, which may subsequently affect our statistical power to detect associations in the two different datasets, we avoided the use of the same statistical cutoff for each dataset. Instead, in order to achieve a comparable genome-wide cutoff across the two datasets, we made use of false discovery rate and compared the two association results by restricting the genome-wide mapping results of each dataset to a list of associations with a similar false discovery rates. The results are summarized in Table 2 and the eQTL profile for the combined set is depicted in Figure 6A. At the 5% genome-wide FDR cutoff (p-value<1.7e-05) we identified 14463 associations for the transcript data (referred to as “eQTL” for expression QTL). At this cutoff stringency, 63% of the transcripts (6299 out of 9896) mapped to at least one locus and roughly one third of the transcripts (3651 out of 9896) mapped to two or more loci (Table 2). In contrast, at the same 5% FDR (p-value<9.6e-06), we only found 1368 significant associations for the proteins (referred to as “pQTL” for protein QTL). The fraction of total proteins with significant association was 672 genes (43%) from which 339 mapped to more than one locus (Table 2). In general, the mapping data for molecular phenotypes can be subdivided into “local eQTLs/pQTLs” to highlight the presence of genetic variation near/within the gene controlling the transcript or protein levels and “distant eQTLs/pQTLs” to discover trans-acting gene-locus interactions at the genetic level [29]. An empirical calculation of haplotype blocks in the HMDP panel (based on continuous stretch of SNPs with the R-squared value above 0.5) showed an average size of 0.73 Mb and a range from less than a kb to 11 Mb (median = 0.25 Mb). Given this fine mosaic structure in the HMDP genotypes, we defined a local eQTL/pQTL as an eQTL/pQTL with the peak SNP located in the 4 Mb window flanking 2 MB on either side of the transcription start site and transcription termination site of the gene. Based on this, from the total of 14463 significant associations in the transcript data, 2066 were local and 12397 were distant eQTLs. In the protein data the numbers of local and distant eQTLs were 144 and 1224, respectively. The proportion of variance explained by the peak SNP in local pQTLs was 44%, local eQTL was 42%, distant pQTL was 27%, and distant eQTL was 23%. The difference in proportion of variance explained between the distant pQTLs and distant eQTLs was highly significant (Student t-test p-value<1e-16), however, a similar comparison showed no significant difference between the mean effect sizes of local pQTL and local eQTLs (Figure S5). For each dataset, the proportion of variance explained by the local SNPs was significantly larger, as expected, as compared to the distant SNPs.
A previous study in plants found that a small number of loci regulated the levels of many proteins [11]. Accordingly, we examined our data for the existence of similar “hotspots”. In the transcript data, we found that the 14463 eQTLs mapped to 9108 distinct peak SNPs. Over 85% of these SNPs (8034 out of 9108) were associated with either one or two transcripts and only a small fraction (334 SNPs) were associated with five or more transcripts. In the protein data, 1368 significant pQTLs mapped to 1088 distinct SNPs across the genome. From these 1088 SNPs, 930 were associated with a single protein, 100 were associated with 2 proteins, and 14 SNPs were associated with 5 or more proteins. From the 1368 peak markers associated with protein levels 438 (32%) were also a peak SNP for one or more transcripts. To investigate if the distinct peak SNPs found in the transcript and protein data map near each other, we divided the genome into 2 Mb bins and using a 50 kb sliding window counted the number of associations in each bin. In the transcript data, the median eQTL number/window was 8 and the highest number of associations was found for bins on Chr 4 (from 98.7 Mb to 100.8 Mb) with 71 eQTLs, Chr 5 (from 80 Mb to 83.1 Mb, from 112.4 Mb to 114.4 Mb) with 79 and 75 eQTLs in each respectively, Chr 7 (from 143.2 Mb to 146.2 Mb) with 78 eQTLs, Chr 8 (from 93.0 Mb to 95.1 Mb) with 71 eQTLs, Chr 17 (from 43.8 Mb to 46.4 Mb) with 80 eQTLs, and Chr 18 (from 55.0 Mb to 57.5 Mb) with 76 eQTLs. In the protein data, however, most associations were randomly distributed except for a clustering of associations on Chr 3 (from 36.5 Mb to 38.6 Mb) with 20 pQTLs and Chr 11 (from 94.3 Mb to 96.7 Mb, and from 114.1 Mb to 118.1 Mb) with 19 and 21 pQTLs respectively (Figure 6B). These results contrast to the previously published reports where hotspots containing hundreds or thousands of eQTLs were observed [30], [31]. This could be partially explained by both our ability to map molecular phenotypes with higher precision in the HMDP panel and the relatively stringent genome-wide threshold chosen to carry out the analysis. The eQTL hotspots on Chr 4 (from 98.7 Mb to 100.8 Mb) resides 6 Mb proximal and the Chr 5 hotspot (from 80 Mb to 83.1 Mb) resides 25 Mb distal to the Chr 5 hotspot reported recently in mouse-hamster radiation hybrid cell line [32]. Despite the relative close distance in mapping, however, we did not find a significant overlap between the genes mapping to these two loci in the two studies.
The global look at the eQTL profiles of the transcriptome and proteome described above suggested that transcripts are more extensively regulated at the genetic level than proteins. However, since the transcriptome data is more comprehensive than the protein data, the differences observed between two datasets might be due to sampling bias. In order provide a measure of similarity for genetic regulation of proteins and transcripts we restricted the data to the set of 396 genes for which we had both protein and transcript measurements available. As mentioned earlier, in this restricted dataset the 396 genes are represented by twice as many peptides as probesets (1343 peptides and 607 probesets). Similar to the genome-wide global analysis, we avoided the use of single statistical cutoff to compare association results across the transcript and peptide datasets, as each dataset has its own variance properties. Instead, we compiled separate lists of significant associations for each dataset using the same FDR cutoff. Since the FDR threshold is driven by the distribution of p-values in each dataset, this allowed us to compare the two lists directly without setting a single statistical cutoff for both datasets. Limiting the mapping data to those associations that met the 5%FDR cutoff in each dataset (p-value<1.7e-05 for transcripts and p-value<9.6e-06 for proteins) we found that despite mapping twice as many peptides as probesets the number of significant associations were roughly equal (939 and 1083 significant associations for probesets and peptides, respectively). This suggests that transcripts are twice as likely to be genetically regulated as are peptides. Next, we performed a gene level analysis where we assigned the associations obtained in each data set to their respective genes and for each gene investigated the degree of similarity in genetic regulation across the protein and transcript dataset. As summarized in Table 2 and consistent with the probeset/peptide analysis described earlier, we found that the number of genes under genetic regulation, as judged by fraction of total genes with at least one significant genome-wide association, favors the transcript dataset. Overall, from the initial 396 genes, 75% (297/396) of the genes with transcript products had at least one significant result vs 61% (242/396) of the genes in the LC-MS dataset. The number of genes with multiple eQTL and pQTL was 205 and 171, respectively. We also looked at the comparison across datasets after classifying the mapping results into local and distant eQTL and pQTL. For distant associations, 281 genes mapped to 799 distinct loci in the transcript dataset and 236 genes mapped to 874 unique genomic locations in protein dataset. Overlapping the association results from the two datasets for distant eQTL/pQTLs, we found that only 25 loci overlap with each other. From these 25 loci, 7 loci had the same peak SNP between the pQTL and the eQTL and in the remaining 18 the distance of peak SNP between the eQTL and pQTL ranged from 2.6 kb to 1.6 Mb. For local eQTLs, we found approximately twice as many local eQTLs as local pQTLs for the 396 genes (79 vs 46). To examine the extent of overlap between local QTLs, we considered a pQTL and an eQTL shared if they mapped within 2 Mb of each other. Using this definition, there were 26 local QTLs shared between the protein and transcript products of the gene. From these common QTLs, 8 mapped to the same peak SNP in the genome-wide association and 18 others mapped in various proximities of each other ranging from 23 bp to 1.8 Mb. The number of shared local QTLs suggests that majority of local pQTLs (26/46 = 56%) are likely to be conserved at the transcript level, and only 1/3 (26/79 = 32%) of eQTLs are conserved at the protein level.
Since local eQTLs are less likely to contain false positives [33], we utilized them to assess if our definition of significance based on FDR had any effect of the results of the comparative analysis we described above. For this, since the transcript local eQTL counts outnumbered the peptide counts, we set a fixed threshold for significance in the transcript dataset (5% FDR, p-value<1.69e-05), counted the number of pQTL overlaps with the significant eQTLs at varying statistical cutoffs, and asked if the increase in the overlap was more than what would be expected by chance. For this we examined results at 5% FDR cutoff, 10% FDR cutoff, and 25% FDR cutoff. At the 5%FDR (p-value<9.58e-06) there were 46 local pQTLs from which 26 overlapped with the 79 eQTLs. Decreasing the p-value stringency to detect association in the protein data, however, did not significantly increase the overlap between eQTL and pQTL. At the 10% FDR (p-value<2.95e-05) we detected 56 pQTL from which 28 overlapped with eQTLs (2 more than 5%FDR), and at the 25% FDR (p-value<0.0002) we detected 68 pQTLs for which the overlap with eQTL was only 29, one more than 10% FDR cutoff and 3 more than the 5%FDR cutoff. These non-significant changes in overlap between eQTL and pQTL suggest that the lack of overlap between eQTL and pQTL as presented earlier was not due to the genomewide significance thresholds set for each dataset.
We should emphasize that one limitation of our study originated from our study design where we utilized different number of mice per strain to estimate the transcript and the peptide levels. For transcript levels we profiled the RNA from 3 mice per strain and estimated transcript levels for genes by averaging over the data obtained for three mice, but for the LC-MS data we only sampled one mouse per strain. This design, by its nature, results in a higher power to detect genome-wide associations and significant correlations with clinical traits for transcript data in comparison to the peptide data. In fact, mapping transcript levels by taking only the data from one of the three microarray data for each strain gave us on average 36% fewer local eQTLs in comparison to what we had obtained by averaging the expression phenotypes over the three microarrays (Text S1 and Table S6). This difference would not change the overall conclusions regarding the commonality and the differences observed between the peptide and transcript genome-wide mapping results.
We report a comparative analysis of the genetic regulation of the transcriptome and proteome in a mammalian system. By examining the effects of thousands of genetic perturbations simultaneously on transcript and protein levels in the HMDP, we were able to investigate the global nature of relationships between the two. Since the HMDP was typed for numerous clinical/physiologic traits, we were also able to study the relationships of these to transcript levels as compared to protein levels. Finally, we examined the commonality of genetic drivers affecting transcript and protein levels. We discuss these points in turn below.
We performed the comparison of protein and transcript levels using two separate approaches. In one approach we comprehensively compared the LC-MS peptide measurements to the microarray expression estimates. In the second approach we examined the relationship between the expression of exons representing the peptides identified by the LC-MS to the expression of exons counted in the next generation sequence data. In addition, to address sources of technical and biologic variation in our measurements, we filtered peptides with significant nongenetic variation. In all these analyses we found that the relationship between the protein expression and transcript expression was modest at best and in only 50% of the cases did this relationship reach nominal statistical significance. We also found that the amount of genetic variation is a predictor of concordance between peptides and transcripts.
Our data complements the data previously published for yeast and plant indicating similar modest protein-transcript relationship. As compared to yeast, we found a slightly higher estimate of protein-transcript concordance (0.27 vs 0.18 correlation) when considering all the peptide measurements and significantly higher estimates (0.42 vs 0.18 correlation) when considering peptides with large genetic variability. The higher estimates reported here are likely to be more reflective of the true relationship between protein and transcript levels as compared to the previous reports mainly due to the choice of technology used to measure protein levels in our study. We utilized the differential labeling technique as put forth by Qian and colleagues where the label free samples are combined with an internal control labeled with heavy isotope [15]. This mixture is then quantified and the results are reported as the ratio of sample to the pool for each identified peptide during mass spectrometry. This strategy, which offers the advantage of overcoming peptide level variation due to platform robustness, has been shown to more precisely quantify peptides as compared to label free methods [15]. This was evident in our study as well where we showed that in general the variance in technical replicates was low, with an overall narrow distribution across the peptides quantified.
Biologically, the modest relationship between the proteins and transcripts is likely to be explained in part by molecular events such as translational efficiency, alternative splicing, folding, assembly into complexes, transport and localization, covalent modification, secretion, and degradation, all of which affect protein levels independently of transcripts. The importance of these post-transcriptional processes is highlighted by a recent report showing that the presence of genetic variation in some of these post-transcriptional processes is associated with certain human diseases [34]. We acknowledge that the design of our study and our most comprehensive dataset, which was generated by Affymetric microarrays capturing the 3′ end of transcripts, prevented us from comprehensively addressing the issue of differential splicing. However, using two complementary approaches, NGS and concordance level of peptides, we examined the possibility of differentially regulated isoforms as a predictor for the lack of concordance between microarray data and LC-MS data. In neither case did differential splicing appear to contribute importantly to the lack of transcript-protein correlation.
An unexpected finding was the stronger association of transcript levels with clinical traits as compared to protein levels with clinical traits. This is likely due in part to the greater technical difficulties for the quantification of proteins as compared to transcripts, but the differences were quite striking and there may be additional explanations. One possible explanation is that the molecular phenotypes are reactive to the clinical phenotypes (rather than being causal) and that there is increased buffering at the protein level. Apart from the strength of the trait associations, the protein and transcript associations in many cases did not overlap. For example, less than 15% of clinical trait-transcript correlations were replicated when traits were correlated with the corresponding proteins.
At the genetic level we also found marked differences in the number and locations of loci controlling protein and transcript levels. When we directly examined protein-transcript pairs corresponding to the same gene, we found that the transcript data had twice as many associations as the protein data. One plausible explanation for the existence of the differential genetic regulation between proteins and transcripts is that of “phenotypic buffering” as put forth previously [11]. An alternative explanation, however, would be that in general the more removed a phenotype is from the DNA variation, the more complex the phenotype becomes. Thus, protein levels would be affected by all the factors influencing transcript levels as well as numerous additional factors. The consequence of increasing complexity in the phenotype is that less of the variation in phenotype would become linked to a single DNA variation.
In summary, we highlighted the similarities and differences in genetic regulation of protein and transcript levels. Although a component of the observed differences in regulation is likely to be technical, particularly with respect to the protein levels, it is clear that the proteomics and transcriptomics provide nonoverlapping information. Thus, these data have important implications for systems biology approaches that utilize such high throughput data. They also raise fundamental questions about the complexity of the relationships between various biological scales involved in complex genetic traits.
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. All experiments in this paper were carried out with UCLA IACUC approval.
Male mice from the HMDP panel, approximately 6–10 weeks of age, were purchased from Jackson Labs and were fed Purina Chow (Ralston-Purina Co., St. Louise, MO) at 16 weeks of age. All mice were maintained on a 12 h light/dark cycle. At 16 weeks of age, whole body fat, fluids and lean tissue mass of mice were determined using a Bruker Optics Minispec nuclear magnetic resonance (NMR) analyzer (The Woodlands, TX, USA) according to the manufacturer's recommendations. We also calculated the total mass of the mice, sum of lean mass, free fluid, and fat mass, and body fat percentage, fat mass/total mass. Following a 16-hour fast, mice were weighed and then bled retro-orbitally under isoflurane anaesthesia. Complete blood counts were performed using a Heska CBC-Diff analyzer (Heska Corp, Loveland, CO, USA). An external control sample with known analyte concentration was run in each plate to ensure accuracy. Glucose levels were determined using commercially available kits from Sigma (St Louis, MO, USA). Insulin levels were measured using commercial ELISA kits (ALPCO Diagnostics). All measurements were performed in triplicate according to the manufacturer's instructions. Plasma lipids were determined as previously described [35]. Mice were euthanized by cervical dislocation and the mass of individual tissues and fat depots (heart, kidney, retroperitoneal fat pad, epididymal fat pad, subcutaneous fat pad, and omental fat pad) were determined by dissecting and weighing each tissue/pad separately after the mice were euthanized. Following this, liver tissues were dissected out, flash frozen in liquid nitrogen, and kept at −70 degrees until further processing.
At 16 weeks of age, the liver tissues of the mice were dissected out, flash frozen in liquid nitrogen, and kept at −70 degrees until further processing. For RNA profiling the RNA from 3 mice per strain were hybridized to Affymetrix Mouse Genome HT_MG-430A arrays. Frozen liver samples were weighed and homogenized in Qiazol according to the manufacturer's protocol. Following homogenization, RNA extraction was performed using Qiagen's RNeasy kit (cat# 74104). Ninety two strains of mice had three biological replicates, five strains had two biological replicates and two strains had one biological replicate each. All RNA samples were cleaned using a Biosprint96 (Qiagen, Valencia, CA) with RNA cleanup beads (Agencourt Bioscience, Beverly, MA) following manufacturer's protocol with adaptations for use with the Biosprint. The quality of the total RNA from the samples was monitored by the Agilent 2100 Bioanalyzer (Agilent Technologies, Palo Alto, CA) and RNA quantity was measured with a NanoDrop (NanoDrop Technologies, Inc. Wilmington, DE) following the manufacturer's instructions. All samples were arrayed into three 96 well microtiter plates following a randomized design format that places samples from the same strain on different plates to better estimate variance across testing strains. All target labeling reagents were purchased from Affymetrix (Santa Clara, CA). Double-stranded cDNAs were synthesized from 1 ug total RNA through reverse transcription with an oligo-dT primer containing the T7 RNA polymerase promoter and double strand conversion using the cDNA Synthesis System. Biotin-labeled cRNA was generated from the cDNA and used to probe Affymetrix Mouse Genome HT_MG-430A arrays. The HT_MG-430A Array plate consists of 96 single MG-430A arrays arranged into standard SBS 96 well plate format. All cDNA and cRNA target preparation steps were processed on a Caliper GeneChip Array Station from Affymetrix. Array hybridization, washing and scanning were performed according to the manufacturer's recommendations. Scanned images were subjected to visual inspection and a chip quality report was generated by the Affymetrix's GeneChipOperating System (GCOS) and Expression console Affymetrix). Two of 288 chips were excluded due to low QC scores. The image data was processed using the Affymetrix GCOS algorithm utilizing quantile normalization or the Robust Multiarray method (RMA) to determine the specific hybridizing signal for each gene. Expression data can be obtained from Geo database (GSE16780). To avoid the effect of SNP on hybridization, we matched the location of ∼14 million SNPs from dbSNP database (NCBI) to the location of the individual probes on the genome. If the location of the probe had a matching SNP within it we flagged the probe and exclude it from the cdf file prior to RMA normalization. If a probeset contained SNP in 8 or more 25-mer probes, we excluded the probeset from the analysis. The cleaned datasets were then background corrected and normalized using the affy package (from bioconductor) using rma, pmonly, and median-polish normalization methods.
RNA isolation for Next Generation Sequencing followed the same protocol as the one described above for the microarray data. For the RNA-Seq experiment, two inbred mice (C57BL/6J and DBA/2J) were chosen. Library preparation was carried using illumina's mRNA-Seq 8-Sample Prep Kit protocol (Illumina, cat# RS-100-0801). In brief, 1 to 10 ug of RNA was used for library construction. In the first step the poly-A containing mRNA molecules were purified using poly-T oligo-attached magnetic beads. Next, the purified mRNA was fragmented into small pieces using divalent cations, followed by double stranded cDNA synthesis using random primers, adenylated at the 3′ end and ligated to the sequencing adapters. The ligated products were then separated on 2% agarose gel, 200 bp fragments were selected and PCR amplified using PE 1.0 and PE 2.0, and purified using QIAquick PCR Purification Kit (QIAGEN, part # 28104). The final library concentration was verified by Bioanalyzer. Sequencing reaction was performed by Illumina Genome Analyzer 2.0 at UCLA Human Genetics microarray core. Raw sequences were uploaded onto Galaxy website (at http://main.g2.bx.psu.edu/) and using the Tophat software [22] was aligned against the reference genome (M. musculus, mm9) downloaded from UCSC. Alignment was performed by setting the parameter for misalignment to one. Relative abundance of transcripts (in Fragments Per Kilobase of Exon per Million read sequence units) was estimated using the Cufflink software [23] and the Ensembl's Mus_musculus NCBIM37 as the reference annotation file.
Male mice were euthanized using isoflurane followed by cervical dislocation at 6–10 weeks of age. The liver tissue was immediately frozen in dry ice until further processing. The 97 samples corresponding to different mouse strains plus some extra samples from C57BL/6J mouse were randomized into 10 batches of 10 samples. Each batch was processed separately prior to quantitative LC-MS analysis. Before the LC-MS analysis the batches were put together and the sample list was randomized one more time. The extraction and digestion of the proteins was performed using a commonly used protocol based on denaturation of protein in 8 M urea followed by digestion with trypsin. Briefly, approximately 5 mg of liver tissue was resuspended in 100 ul denaturing solution (8 M urea, 50 mM Tris-HCl pH 8.0 and 1 mM EDTA) and homogenized with a motorized pestle. Upon homogenization, the total protein content was measured by Bicinchoninic Protein Assay (BCA, Pierce, Rockford, IL) and the 500 ug aliquots were taken from each sample for further processing. DTT was added to a concentration of 10 mM in sample, then to solubilize and unfold the proteins the samples were incubated for 30 min at 37oC with shaking. Cysteine residues were alkylated by adding iodoacetamide up to 40 mM concentration and incubating for 1 hour at 37oC, with shaking, in the dark. For protein digestion the samples were diluted 10-fold with 50 mM ammonium bicarbonate (pH 7.8), supplemented with 1 mM CaCl2, 10 ug of trypsin and incubated for 3 hours at 37oC with shaking. The sample digests were purified with solid phase extraction using C18 columns (Discovery DSC-18, SUPELCO, 52601-U), lyophilized and resuspended in 25 mM ammonium bicarbonate pH 7.8. The peptide amounts were estimated with BCA assay. On the average the amount of purified tryptic peptides was 200 ug. To generate the 18O reference sample, 20 ug of each sample was pooled, then boiled 10 minutes, followed by immediate cooling for 10 minutes. The boiling/cooling steps were performed to inactivate trypsin (this step helps to avoid back-exchange of 18O-labeled peptides). The pooled reference was then subjected to solution-phase tryptic 18O exchange, followed by quenching of tryptic activity with formic acid. The pooled sample was then added in equal amounts with each individual sample for quantitation purposes.
Construction of a library of proteins and tryptic peptides present in the liver is an important step for follow-up quantitation. 10 ug aliquots from all 97 strains were pooled together and subjected to LC fractionation by strong cation exchange (SCX) chromatography on a 200 mm×2.1 mm Polysulfoethyl A column (PolyLC, Columbia, MD) preceded by a 10 mm×2.1 mm guard column, using a flow rate of 0.2 mL/min. LC separations were performed using an Agilent 1100 series HPLC system (Agilent, Palo Alto, CA). Mobile phase solvents consisted of (A) 10 mM ammonium formate, 25% acetonitrile, pH 3.0 and (B) 500 mM ammonium formate, 25% acetonitrile, pH 6.8. Once loaded, isocratic conditions at 100% A were maintained for 10 min. Peptides were separated by using a gradient from 0–50% B over 40 min, followed by a gradient of 50–100% B over 10 min. The gradient was then held at 100% solvent B for another 10 min. Following lyophilization, all thirty fractions collected during this gradient were dissolved in 25 mM ammonium bicarbonate and stored at −80OC.
Each SCX fraction was analyzed with an automated custom-built capillary HPLC system coupled online to an LTQ ion trap mass spectrometer (Thermo Fisher, San Jose, CA) by using an electrospray ionization interface. The reversed phase capillary column was prepared by slurry packing 3-µm Jupiter C18 particles (Phenomenex, Torrance, CA) into a 75 µm i.d.×65 cm fused silica capillary (Polymicro Technologies, Phoenix, AZ). The mobile phase solvents consisted of (A) 0.2% acetic acid and 0.05% TFA in water and (B) 0.1% TFA in 90% acetonitrile. An exponential gradient was used for the separation, which started with 100% A, and gradually increased to 60% B over 100 min. The instrument was operated in a data-dependent mode with an m/z range of 400–2000. Ten most abundant ions from each MS scan were selected for further MS/MS analysis by using a normalized collision energy setting of 35%. Dynamic exclusion was applied to avoid repeat analyses of the same abundant precursor ion.
The SEQUEST software (Thermo Fisher) was used to search the MS/MS data against the mouse International Protein Index (IPI) database (version 3.52 http://www.ebi.ac.uk/IPI). Human keratins and porcine trypsin were added into the database as expected contaminants. Trypsin cleavage specificity was required for all of the considered peptides. The following criteria were used to filter raw SEQUEST results: 1) Xcorr≥1.9 and DeltaCn2≥0.21 for charge state +1; 2) Xcorr≥2.5 and DeltaCn2≥0.26 for charge state +2; 3) Xcorr≥2.8 and DeltaCn2≥0.32 for charge state +3. These criteria provide the maximum number of peptide identifications not exceeding 1% false discovery rate (FDR). To estimate the FDR of peptide identifications we searched against a reversed database as previously described [36].
Relative peptide and protein quantitation was based on ratios between intensities of natural 16O isotope containing peptides and reference peptides labeled with stable 18O isotope at the carbonyl group at the C-terminus of the peptide. To create a reference sample we pooled together 20 ug aliquots from all strains and labeled the C-termini with 18O isotopes using trypsin catalyzed exchange in the presence of heavy H218O water as described above and elsewhere [37]. Prior to the LC-MS analysis 3.75 ug aliquots from each individual sample were mixed with the same amount of 18O-labeled reference sample.
The 7 ug aliquots were analyzed on a LTQ-Orbitrap mass spectrometer that was interfaced with a 75 um i.d.×65 cm long LC column packed with 3 um Jupiter C18 particles (Phenomenex). The mobile phase solvents consisted of (A) 0.2% acetic acid and 0.05% TFA in water and (B) 0.1% TFA in 90% acetonitrile. An exponential gradient was used for the separation, which started with 100% A and gradually increased to 60% B over 100 min. LC-MS datasets were analyzed by in-house software VIPER [38] that detected features in mass – elution time space and assigned them to peptides in AMT tag database as described elsewhere [39], [40]. Typically an LC-MS run identifies ∼3,500 16O/18O peptide pairs that co-elute with a 4.0085 Da mass difference.
As we mentioned before, the relative abundances of tryptic peptides were calculated as the ratio between light and heavy isotopes. The relative abundances then were normalized with EigenMS procedure [41] to correct systematic biases that may arise for example from unequal sample loading, batch-to-batch differences in sample processing and LC column variability. Briefly, the EigenMS procedure discovers the systematic trends (so-called eigenpeptides) in the data using singular value decomposition and then removes contributions of those eigenpepides from each peptide. For all data analysis purposes the peptide and protein intensities were log2 transformed and zero-centered by subtracting the peptide or protein specific means taken across all the samples.
To determine the protein levels by immunobloting, liver samples were homogenized in RIPA including phosphatase and protease inhibitors (Santa Cruz Biotech sc-24948), and protein determination were done using the Biorad Dc Assay. Protein samples were boiled following addition of Laemmli loading dye, separated on Invitrogen precast gels, and transferred to PVDF membranes. Membranes were rinsed in 1× TBST (Cell signaling #9997) blocked in 5% skim milk-TBST, rinsed in TBST, and incubated with primary antibodies diluted in 3% BSA-TBST for 1 hr at 23degC or overnight at 4degC. Membranes were washed in TBST and incubated with an HRP-conjugated anti rabbit IgG KPL (#474-1516) 1/5000 in 5% skim milk-TBST. Membranes were washed again, incubated in ECL-plus, and signal detected using a Biorad Chemidoc or film. Densitometry was done using the Biorad Quantity One software. The following list of antibodies and working dilutions were used for each protein: Fasn (Cell Signaling cat #3180, 1/2000), Acyl (Cell Signaling cat #4332, 1/2000), Ywhae (Cell Signaling cat #9635, 1/2000), Vim (Cell Signaling cat#3932, 1/1000), Rkip (Cell Signaling cat#5291, 1/2000), Gapdh (Cell Signaling cat#3683, 1/5,000), Glo1 (Sigma Chemical SAB1100242, 1/20,000), GstA4 (Sigma Chemical SAB1100244, 1/20,000), AnxA5 (Sigma Chemical AV36687, 1/2000), Hao1 (Sigma Chemical AV42480, 1/2000), Aldh3A2 (Sigma Chemical HPA014769, 1/20,000), Actin (Sigma Chemical A2066, 1/5,000), Acox1 (Abnova PAB4367, 1/2000).
For transcript data we applied three filtering steps based on 1) genetic heritability, 2) probeset annotation. We have profiled 3 mice per strain which allowed us to estimate the broad sense heritability for each transcript. Broad sense heritability for each transcript was measured using ANOVA where strain information was used as a grouping factor. The broad sense heritability which is defined as the ratio of genetic variance over total variance for the phenotype was estimated by dividing the sum of squares of the strain information factor over total sum of squares in the ANOVA. The significance of heritability was established if the p-value for the strain information term in ANOVA was below the nominal 0.05 threshold. The selection cutoff for including gene in the analysis based on heritability was set to heritability p-value of 0.05. From the 22700 probesets 10186 probesets did not meet this cutoff. For annotation filtering, we acquired the Ensembl Gene ID for each Affymetrix probeset and selected those probesets that were only annotated to only one Ensembl gene. From the initial 22700 probesets, 4401 probesets had ambiguous annotation (either did not map to a gene or mapped to more than 2 Ensembl genes). Overall, 9896 probesets met both filtering criteria (significant heritability and unique Ensembl annotation). For the protein data, the initial filtering steps were based on 1) eliminating peptides with excessive missing values which would otherwise have unreliable mapping information, 2) eliminating peptides with missed internal cleavage sites which cause unreliable measurement. To annotate peptides we utilized the SpliceCenter web-based tool [42] to obtain the location of the exon each peptide represents. Peptides which mapped to multiple exons of more than one gene (as determined by SpliceCenter) were excluded from the analysis because of ambiguous annotation. The genomic exon coordinates for each peptide was then used to query the Ensembl database to acquire the Ensembl Gene IDs. In the second step of filtering peptides that had more than the value of 2 for the ratio of total variance over technical variance were chosen. Overall, 1543 peptides met the two-stage filtering described above.
Inbred strains were previously genotyped by the Broad Institute (http://www.broadinstitute.org/mouse/hapmap), and they were combined with the genotypes from Wellcome Trust Center for Human Genetics (WTCHG). Genotypes of RI strains at the Broad SNPs were inferred from WTCHG genotypes by interpolating alleles at polymorphic SNPs among parental strains, calling ambiguous genotypes missing. Of the 140,000 SNPs available, 95,854 were informative with an allele frequency greater than 10% and missing values in less than 10% of the strains. These SNPs were used for both protein and transcript genome-wide association analysis.
We applied the following linear mixed model to account for the population structure and genetic relatedness among strains in the genome-wide association mapping [27]: y = μ+xβ+u+e.
In the formula, μ represents mean, x represents SNP effect, u represents random effects due to genetic relatedness with Var(u) = σg2K and Var(e) = σe2, where K represents IBS (identity-by-state) matrix across all genotypes in the HMDP panel. A restricted maximum likelihood (REML) estimate of σg2 and σe2 are computed using EMMA, and the association mapping is performed based on the estimated variance component with a standard F-test to test β≠0. We applied EMMA (Efficient Mixed Model Association) as an R implementation of a linear mixed model. The percent of variance explained for each molecular phenotype was calculated using the SNP effect calculated from EMMA by defining it as 1-(variance of residuals/variance of original phenotypes). It should be noted that since EMMA is orders of magnitude faster than other implementations commonly used, we were able to perform statistical analyses for all pairs of transcripts and genome wide markers in a few hours using a cluster of 50 processors. Both pQTL and eQTL were defined as “local” if the peak association SNP position was within a 4 Mb interval, flanking 2 Mb on either side of the transcription start and end of the gene under regulation. Genome-wide cutoff: Genome-wide cutoffs were calculated as the false discovery rates using the “qvalue” package for FDR calculation in the R statistical software [43]. Due to the computational complexity associated with evaluating q-values for over 400 million p-values, we computed the FDRs by taking the average FDR for 100 samples each containing 5 million randomly selected p-values from the original calculated p-values. FDR calculation was carried out separately for the protein and transcript dataset.
All statistical analyses and data visualizations were carried out using the R statistical software (available at http://cran.r-project.org/). Classification of proteins and transcripts to various GO categories was accomplished using Mouse Genome Informatics website at Jackson Laboratories and the GO ontology tool at Princeton. For each probe and each peptide, we first obtained the MGI IDs using the MGI batch query tool at http://www.informatics.jax.org/ [44]. Using MGI IDs we utilized the GO Term Mapper at http://go.princeton.edu, which is based on map2slim algorithm [45] to obtain the GO annotations and summary statistics. The background geneset used in this analysis was the list of all genes annotated by MGI. To assess the significance of the correlation coefficients observed in the PT-pair GO analysis, for each GO category we created 100,000 bootstrap datasets each equal in size to the number of genes assigned to the GO term. Bootstrapping was carried out randomly and without replacement from the pool of 584 original correlation p-values among the PT-pairs. The significance of the observed average p-values for each GO term is reported as the two-tailed test against the empirical distribution created by the corresponding 100,000 permutation set. All the correlation coefficients and corresponding p-values reported in the paper are calculated using the bicor function in the WGCNA R package [46]. The main advantage of using bicor, which performs biweight midcorrelation calculation, over Pearson's correlation is based the robustness of the correlation coefficient measurement to the presence of outliers in the data.
|
10.1371/journal.ppat.1006130 | Leishmania HASP and SHERP Genes Are Required for In Vivo Differentiation, Parasite Transmission and Virulence Attenuation in the Host | Differentiation of extracellular Leishmania promastigotes within their sand fly vector, termed metacyclogenesis, is considered to be essential for parasites to regain mammalian host infectivity. Metacyclogenesis is accompanied by changes in the local parasite environment, including secretion of complex glycoconjugates within the promastigote secretory gel and colonization and degradation of the sand fly stomodeal valve. Deletion of the stage-regulated HASP and SHERP genes on chromosome 23 of Leishmania major is known to stall metacyclogenesis in the sand fly but not in in vitro culture. Here, parasite mutants deficient in specific genes within the HASP/SHERP chromosomal region have been used to investigate their role in metacyclogenesis, parasite transmission and establishment of infection. Metacyclogenesis was stalled in HASP/SHERP mutants in vivo and, although still capable of osmotaxis, these mutants failed to secrete promastigote secretory gel, correlating with a lack of parasite accumulation in the thoracic midgut and failure to colonise the stomodeal valve. These defects prevented parasite transmission to a new mammalian host. Sand fly midgut homogenates modulated parasite behaviour in vitro, suggesting a role for molecular interactions between parasite and vector in Leishmania development within the sand fly. For the first time, stage-regulated expression of the small HASPA proteins in Leishmania (Leishmania) has been demonstrated: HASPA2 is expressed only in extracellular promastigotes and HASPA1 only in intracellular amastigotes. Despite its lack of expression in amastigotes, replacement of HASPA2 into the null locus background delays onset of pathology in BALB/c mice. This HASPA2-dependent effect is reversed by HASPA1 gene addition, suggesting that the HASPAs may have a role in host immunomodulation.
| Millions of people around the world are at risk of infection with single-celled Leishmania parasites that cause a wide range of infectious diseases of the immune system, the leishmaniases. There is no effective vaccine for these infections while available drugs have toxic side-effects and resistance is an increasing problem. Human infection occurs through the bite of infected blood-feeding sand flies. Leishmania parasites live in the sand fly gut, where they complete a complex series of developmental changes to become infective to mammals. The parasites also modify the insect gut to promote their own transmission. Little is known about the molecular regulation of these processes. Recently, we showed that deletion of a small group of related genes from the parasite prevented completion of its development in the sand fly, suggesting a role for these sequences in transmission to the host. This study clarifies the expression pattern of these genes during parasite development and shows that the observed stalling of development is accompanied by changes in parasite-sand fly gut interactions and a loss of parasite transmission. These target genes also influence disease development in the mammalian host, confirming critical roles for their encoded proteins throughout the parasite life cycle.
| Kinetoplastid parasites of the genus Leishmania cause a diverse spectrum of mammalian infectious diseases, the leishmaniases, ranging from cutaneous and mucosal pathologies to potentially fatal visceral infections [1]. Endemic human cases have been reported on all continents except Australia and Antarctica [2]. Leishmania parasites are transmitted by female phlebotomine sand flies (Diptera: Psychodidae: Phlebotominae) of the genera Phlebotomus (Old World) and Lutzomyia (New World) [3]. All mammalian-infective Leishmania species (spp.) belong to two characterized subgenera, L. (Leishmania) [4] and L. (Viannia) [5]. Within the sand fly, parasites of both subgenera undergo metacyclogenesis, a series of morphological and functional changes that produce mammalian-infective metacyclic promastigotes. This process can be triggered, at least in vitro, by various factors including nutrient depletion, reduction of pH and tetrahydrobiopterin levels [6,7]. As recently described [8], mammalian-resident intracellular amastigotes transform into midgut adapted, proliferative promastigotes post blood meal (PBM) and these in turn produce nectomonads (synonymous with elongated nectomonads [9,10]) that mediate midgut attachment [11]. Nectomonads then transform into proliferative leptomonads (synonymous with short nectomonads [9,10]), which produce promastigote secretory gel (PSG) containing filamentous proteophosphoglycan (fPPG) [12,13]. Finally, leptomonads transform into the mammalian-infective, midgut-detached metacyclics [14] and haptomonads which colonise and degrade the stomodeal valve (SV) [15,16].
While little is known about the molecular regulation of metacyclogenesis, several genes have been identified that are specifically expressed in late stages of the process. The best-characterised of these are found at the L. (L.) major cDNA16 locus on chromosome 23 and are termed the L. (Leishmania) species-specific HASP (hydrophilic acylated surface protein) and SHERP (small hydrophilic endoplasmic reticulum-associated protein) genes (HASPA1, SHERP1, SHERP2, HASPB and HASPA2; [17,18]). The locus is conserved in other L. (Leishmania) species but divergent in L. (Viannia) species [19]. The three HASPs are highly related proteins with identical N- and C-terminal regions. HASPB, the best characterized of these proteins, is trafficked to and tethered at the parasite cell surface by co-translational N-myristoylation and post-translational palmitoylation at its N-terminal SH4 domain [20,21]. HASPB contains extensive amino acid repeat domains in its central region and these show inter- and intraspecific variations while bearing some resemblance to peptidoglycan and immunoglobulin-binding domains of several bacterial surface proteins [22–24]. HASPB is specifically expressed in L. (L.) major metacyclics and amastigotes, but is only detectable in amastigotes in L. (L.) mexicana [19,23,25]. The HASPA genes, which do not encode amino acid repeats, have identical 5’ untranslated regions (UTRs) and open reading frames (ORFs) but different 3’UTR sequences. These contribute to distinct mRNA expression patterns: HASPA2 mRNA is expressed early in procyclics and peaks in metacyclics, while HASPA1 mRNA is upregulated in metacyclics and amastigotes [17,24,26]. SHERP, a small membrane associated protein, is expressed predominantly in metacyclic parasites, where it localizes to the cytosolic faces of the endoplasmic reticulum and mitochondrion [27] and can fold in the presence of membrane phospholipids, supportive of a role in protein-protein interactions [28]. The in vitro binding of SHERP to vacuolar H+-ATP synthase components involved in subcellular compartment acidification has led to the hypothesis that SHERP may impact on parasite autophagocytosis [28], a process shown to be essential for metacyclogenesis [29].
Genetic deletion of the whole cDNA16 locus in L. (L.) major, by homologous recombination, generated mutants that were stalled in metacyclogenesis within the sand fly, predominately in the nectomonad stage [30]. By contrast, the same mutants showed no significant phenotype when maintained in in vitro culture with the culture-generated metacyclics proving more virulent than the parental parasite line (Friedlin V1; FVI) in BALB/c mice [31]. In the same study, episomal-replacement of the full cDNA16 locus into the null background led to unregulated HASP and SHERP overexpression and avirulence [31]. Conversely, more recent reintegration of the whole cDNA16 locus into its former location on chromosome 23 re-established parental line gene regulation and rescued metacyclogenesis in vivo [30]. The replacement of an episomal HASPB copy alone into the null background suggested that HASPB was key for the completion of metacyclogenesis in vivo [30]. However, since unregulated episomal expression can cause misleading phenotypes in Leishmania, verification of this observation by HASPB reintegration into the original cDNA16 locus became essential.
The main focus here was to investigate the contribution of the individual HASPs and SHERP to metacyclogenesis within the sand fly midgut and to host transmission, utilising a broad panel of newly-validated genetic mutants, generated and rigorously tested in vitro for this study. While these aims were not fully met, passaging all mutant lines through sand flies to investigate the respective contribution of the HASP and SHERP proteins to metacyclogenesis in vivo revealed clear differences in gene expression and parasite behaviour when comparing in vitro and in vivo conditions. This stimulated a first investigation into the potential impact of midgut factors on gene regulation in Leishmania. Further experiments with a sub-group of mutants confirmed that completion of metacyclogenesis, PSG formation, SV colonisation and parasite-to-host transmission are dependent on HASP and SHERP genes. Use of these mutants also allowed us, for the first time, to address the previously established differences in HASPA1 and HASPA2 mRNA expression at the protein level and to investigate the respective contribution of HASPA1 and HASPA2 to amastigote virulence in vivo.
To address our understanding of metacyclogenesis and host transmission in vivo, we focused firstly on the in vitro generation and characterization of new HASP/SHERP replacement mutant lines; secondly, on the in vivo infectivity of cultured mutant lines; and finally, on the impact of genes in the cDNA16 locus on parasite development in the sand fly and transmission by sand fly bite.
For this study, a total of seventeen new L. (L.) major HASPA1, SHERP, HASPB and/or HASPA2 replacement lines were generated by homologous recombination of newly synthesized gene replacement constructs (S1 Fig) into the original chromosomal location of the cDNA16 locus within the null background of the previously characterized cDNA16 double deletion mutant (cDNA16 dKO [31]; Tables 1 and S1). The alternative approach of targeted gene disruption/deletion of individual genes was not possible technically due to high levels of sequence identity and repetition within the cDNA16 locus. For clarity, and due to strong similarities in mutant phenotypes, the data for only 6 representative gene-replacement lines are shown and discussed here (HASPB sKI, SHERP sKI, HASPA1 sKI, HASPA2 sKI, HASPA1/2 sKI & HA1/2+S2/HB sKI; Table 1). Additional data from other lines and clones can be viewed in the supplementary files. L. (L.) major FVI (the wild type parental line) and the previously characterized cDNA16 dKO and full cDNA16 locus single replacement mutant lines (cDNA16 sKI [30]) served as controls in all experiments.
Briefly, newly generated HASP and/or SHERP replacement lines were rigorously tested in vitro to ensure correct construct integration and regulated expression in the former cDNA16 locus, using a series of standard protocols, and selected clones (at least two per genotype) were then used for further analyses. Clones were initially screened by PCR (S2 Fig), followed by Southern blot and qPCR analysis to ensure correct integration of HASP and SHERP gene replacement constructs (Figs 1A, 1B, 1C and S2–S5). For Southern blot analysis, genomic DNA (gDNA) from selected clones was SacI digested, size separated and probed with suitable DIG-labelled DNA fragments (Figs 1B and S4). The HASP probe hybridized to HASPA2, HASPA1 and HASPB in FVI, detecting the previously observed 7.6 Kb, 4.3 Kb and 2.2 Kb fragments, respectively [30]. The HASPA1 sKI, HASPA1/2 sKI, HASPA2 sKI and HASPB sKI mutants also showed single fragments of the expected sizes (6.1 Kb, 9.3 Kb, 7.5 Kb and 6.7 Kb, respectively), while no fragments were observed in the cDNA16 dKO and SHERP sKI mutant lines. The HA1/2+S2/HB sKI mutant line, containing all the cDNA16 locus gene types, showed two expected fragments: one of 9.3 Kb equivalent to that detected in the HASPA1/2 sKI mutant line (due to integration of the same HASPA1/2 construct); the second of 2.2 Kb, matching the HASPB fragment detected in FVI, as expected. The SHERP probe confirmed the previously observed 1.8 Kb and 1.6 Kb fragments in FVI [30] and hybridized to an expected 4.2 Kb fragment in SHERP sKI and to an expected 1.6 Kb fragment in HA1/2+S2/HB sKI, matching one of the SHERP fragments in FVI. The BSD (blasticidin resistance) gene [32] was only detected in the HASPA1 and SHERP/HASPB (S2/HB) constructs; a single fragment each in the HASPA1 sKI (6.1Kb) and in the HA1/2+S2/HB sKI (2.6 Kb) mutant lines as expected, with no BSD hybridising fragments in FVI and cDNA16 dKO. The NEO (neomycin resistance) gene [33], present in the HASPA1/HASPA2 (HASPA1/2 or HA1/2), HASPA2, HASPB and SHERP constructs, generated fragments of the same size as those detected by the HASP and SHERP probes since there was no SacI restriction site between the gene of interest and the resistance marker cassette (Figs 1A and S3). The Southern analysis also excluded the presence of multi-copy episomal constructs in the parasite lines.
Gene copy number in selected clones was verified by gene-specific qPCR (Figs 1C and S5). The Na/H antiporter-like protein gene, present as a single copy on chromosome 23, was used as control for data normalization. A value of 1(±0.2) was predicted for all targets (HASPA1, SHERP, HASPB and HASPA2 gene constructs) present as single gene copies. Only HASPA1/2 sKI and HA1/2+S2/HB sKI, which contained both HASPA1 and HASPA2 in the HASPA1/2 (HA1/2) construct were expected to generate a value of 2(±0.5) for HASPA. As shown in Fig 1C (see also S5 Fig), the expected values were obtained for each target gene within the mutant lines.
Two to three selected clones from each line were then passaged through BALB/c mice to restore parasite infectivity and stable, regulated gene construct expression (which is known to diminish for the HASP and SHERP genes following prolonged parasite passage in culture [21]). Immunoblotting was used to verify stage-specific protein expression from the integrated HASP and SHERP genes, using whole lysates generated from low-passage parasites cultured over a 6 day time course from day 2 to day 7 post inoculum (p.i.), in comparison with HASPB, HASPA and SHERP expression patterns in FVI (Fig 1D). While only one clone per line is shown here, additional clones showed comparable results for each parasite line, respectively (S6 Fig), as demonstrated in S7 Fig for HASPA2 sKI and HASPA1/2 sKI. Due to its small molecular mass and biophysical properties, SHERP is inherently difficult to transfer onto nitrocellulose, reducing blot quality. However, the results for HASPB and SHERP showed that integration of the HASPB and SHERP constructs into the cDNA16 locus was sufficient to reproduce the parental line (FVI) protein expression patterns. The unexpected differences in HASPA expression between HASPA1 sKI, HASPA1/2 sKI and HASPA2 sKI mutant lines are discussed further below. The constitutively expressed L. (L.) major N-myristoyltransferase protein (NMT; [36]) served as a loading control in all protein analyses.
Selected mutants were further tested for any growth defects due to genetic manipulation. Low-passage parasites were inoculated at 105 parasites/ml into M199 medium and cell numbers monitored every 24 hr for 7 days. This growth assay was performed in 3–4 consecutive repeats, inoculating fresh M199 medium to a final concentration of 105 parasites/ml using the previous cultures at late log-phase (day 3 p.i.). No significant fitness defect was observed in any of the mutant lines growing as in vitro promastigotes in comparison to the parental line (FVI; Figs 1E and S8). All selected clones reached stationary phase by day 3–4 p.i., correlating with peak parasite density, followed by a similar rate of decline in live parasite density to day 7 p.i. with no statistically significant differences.
Since HASPA1 and HASPA2 have identical ORFs, it is impossible to distinguish these two proteins by antibody probing in wild type parasites [26]. Generation of individual HASPA1 and HASPA2 replacement mutant lines in this study allowed analysis of protein expression from these two genes, previously only possible at the mRNA level [17,18]. The immunoblot data from promastigotes (Figs 1D, S6 and S7) showed a similar HASPA expression in HASPA2 sKI as in FVI and cDNA16 sKI, peaking at day 7 p.i. when cultures were enriched in metacyclics. This correlated with the previously established HASPA2 mRNA expression pattern in FVI [17,18], although the correlation was less clear for HASPA2-only combination mutants with HASPB and/or SHERP (S6 Fig). In contrast, HASPA protein was not detectable in HASPA1 sKI and in HASPA1-only combination mutants cultured up to 7 days (Figs 1D and S6) regardless of the clone tested, while low level mRNA expression had previously been detected in metacyclic cells [17,18].
Conversely, immunoblots of whole lysates of intracellular amastigotes purified from skin lesions of BALB/c mice showed expression of HASPA in the HASPA1 sKI amastigotes, but no detectable HASPA expression in the HASPA2 sKI amastigotes (Fig 2E). Overall, this analysis has for the first time demonstrated that protein expression from the individual HASPA genes is stage-specific in L. (L.) major, with HASPA1 expressed in amastigotes and HASPA2 in promastigotes.
Interestingly, the presence of both HASPA1 and HASPA2 in the same construct appeared to enhance and deregulate HASPA expression in HASPA1/2 sKI and HA1/2+S2/HB sKI promastigotes (Fig 1D; and in other HASPA1-HASPA2 combination mutants with HASPB or SHERP; S6 Fig) regardless of the clone tested (S7 Fig), suggesting that either HASPA1 contributes to HASPA expression in early culture stages in the presence of HASPA2 or that the HASPA2 gene has lost parental line regulation in the HASPA1/2 construct (Figs 1D and S6). The latter explanation implicates cDNA16 locus structural constraints as important for correct HASPA regulation. This is a valid hypothesis since the HASPA1/2 construct was assembled with the same DNA fragments as the HASPA1 and HASPA2 constructs (S1 Fig), the sequence was verified post construct assembly and the different HASPA expression patterns in HASPA1 sKI and HASPA2 sKI correlated with the previously established mRNA expression patterns [17,18]. However, this unexpected HASPA1/2 mutant phenotype did not result in an unexpected phenotype in vivo.
Our previous work had shown that deletion of the full cDNA16 locus enhanced footpad lesion development in BALB/c mice following inoculation with cDNA16 dKO as compared to FVI promastigotes [31]. Due to our new observations on stage specific expression of HASPA1 and 2 in amastigotes and promastigotes, respectively, we wanted to revisit the original in vivo phenotype and address whether deletion and replacement of these genes affected footpad pathology. For that purpose, two clones from each of seven parasite lines (FVI, cDNA16 dKO cDNA16 sKI, HASPA1 sKI, HASPA2 sKI, HASPA1/2 sKI and HA2+S2/HB sKI; refer to S9 Fig for clone ID) were grown in vitro to late-stationary phase (day 7 p.i.) and assessed for the presence of metacyclic parasites. Given the lack of metacyclic-specific markers in L. (L.) major other than SHERP and HASPB, the focus of this study, the efficiency of metacyclogenesis in vitro was monitored by two complementary approaches: by morphometric analysis of fixed parasites (Figs 2A and S10), in comparison with existing sand fly data, and by metacyclic purification following peanut lectin agglutination (PNA, Fig 2B).
Morphometry showed that all mutant lines had generated metacyclics (Figs 2A and S10) with FVI and HASPA1 sKI being the most efficient (>70%) and cDNA16 dKO and HASPA2 sKI the poorest (<55%). Statistical significant differences occurred only between: FVI and cDNA16 dKO, P = 0.003; FVI and HASPA2 sKI, P = 0.003; HASPA1 sKI and cDNA16 dKO, P = 0.011; HASPA1 sKI and HASPA2 sKI, P = 0.014. However, metacyclics from all cultured lines and clones were similar in size and morphology (S10 Fig), suggesting no difference in metacyclic quality. Leptomonads were present at a similar percentage (~20%) in all cultured parasites tested.
Metacyclic enrichment by PNA is subject to some losses due to passive entrapment of the highly motile parasites in the lectin aggregates. Despite the observed reduction in metacyclic numbers for some lines as compared to morphometry (Fig 2A and 2B), overall the agglutination test did not reveal any significant differences in the capacity of the different parasite lines to undergo metacyclogenesis (Fig 2B), as verified by morphometric analysis (S10 Fig). Comparing three replicates, FVI generated a mean of 55.8%, cDNA16 dKO 55.2%, cDNA16 sKI 57.7%, HASPA1 sKI 47.9%, HASPA1/2 sKI 51.5%, HASPA2 sKI 58.2% and HA2+S2/HB sKI 47.2% metacyclics in day 7 cultures. Based on these findings, the numbers of metacyclic parasites inoculated into the footpads of BALB/c mice were very similar for all clones tested.
Given the demonstrated consistency of metacyclogenesis in these clones, and to allow direct comparison with previous observations, cultured promastigotes were harvested and injected at 3x107 parasites per BALB/c footpad without prior metacyclic purification. Disease progression for two independent clones per mutant line (Figs 2C, S9a and S9c) was tested by weekly footpad measurements followed by limiting dilution assay (LDA) of footpad homogenates once ~2 mm lesion diameter was reached (Figs 2D, S9b and S9d). This analysis showed that both cDNA16 dKO clones were fast to develop severe lesions (within 5 weeks p.i.), as previously observed [31], although this was not due to elevated numbers of parasites within the lesions (Figs 2C, 2D and S9). The parental line (FVI) clones required ~8 weeks p.i. to reach the same level of lesion development, while cDNA16 sKI required ~6–7 weeks. Interestingly, both HASPA2 sKI clones showed a delayed onset of footpad lesion development compared to FVI, which was also repeatedly observed during routine passaging of parasites through BALB/c mice, using these two distinct clones. HASPA2 sKI-infected BALB/c mice took significantly longer (>10 weeks p.i.; P<0.001) to produce comparable footpad lesions to those observed in the other mutants. Conversely, lesion development in HASPA1 sKI and HASPA1/2 sKI infected footpads had a similar time course to cDNA16 sKI (~6–7 weeks). Lesion development in HA2+S2/HB sKI was comparable to FVI, suggesting that addition of a HASPB and/or SHERP copy restored FVI virulence levels. Further investigation will be required to resolve the mechanisms whereby the HASPs and SHERP contribute to disease outcome.
Sádlová et al. [30] showed that stalling of metacyclogenesis due to cDNA16 locus deletion is only observed in the sand fly vector and not in culture. In this new study, each of the L. (L.) major mutant lines (Tables 1 and S1) was fed independently at 106 early log-phase promastigotes/ml blood to P. (P.) papatasi and/or P. (P.) duboscqi, both L. (L.) major-specific vector species [8]. A total of 2,736 sand flies (S2 Table) were sampled at different time points (day 2, 5, 9 and 12 PBM or at day 6 and 12 PBM only) and infection loads, parasite localization and parasite morphology assessed (Figs 3–5 and S11–S14).
Infection load data by microscopy (Figs 3A and S11) revealed that P. (P.) duboscqi supports significantly higher (P<0.001) parasite numbers in the midgut than P. (P.) papatasi. In P. (P.) papatasi, significant differences in parasite load were observed between FVI and the three mutant lines tested, cDNA16 dKO, HASPB sKI and SHERP sKI, after blood meal defecation (day 5, 9 and 12 PBM; P< 0.001; Figs 3A and S11). While FVI, cDNA16 dKO and HASPB sKI showed significantly increased parasite loads from day 2 PBM to day 12 PBM (P<0,001; P = 0.009; P = 0.032, respectively), SHERP sKI survived comparatively poorly (~40% infected at day 12 PBM) with significantly decreased persistence of infection from day 2 PBM to day 12 PBM (P = 0.027). Interestingly, SHERP sKI survival in P. (P.) duboscqi was not affected (~85% at day 12 PBM) and developed as well as other mutant lines tested (Figs 3A and S11). In general, parasite lines survived well in P. (P.) duboscqi, showing significant increases in parasite numbers (P≤0.002) from day 6 PBM to day 12 PBM (S11 Fig). The microscopically evaluated parasite loads were verified for day 12 PBM samples by qPCR for 30 infected female sand flies per parasite line (Figs 3B and S12). Although microscopy tended to underestimate infection loads compared to qPCR, there was generally good correlation between light microscopic and qPCR data with the exception of cDNA16 dKO and SHERP sKI in P. (P.) papatasi, which showed significantly higher parasite loads by qPCR than by microscopy (Figs 3B and S12). However, while microscopic analysis only evaluates live parasites, qPCR does not discriminate between live and dead parasites containing genomic DNA. No significant differences were established between parasite lines tested in P. (P.) duboscqi at day 12 PBM by qPCR with the exception of FVI and HASPA1 sKI compared to HASPA2 sKI (P = 0.01 and 0.004, respectively) and HA1/2+S2/HB sKI (P = 0.03 and 0.01, respectively).
In parallel, parasite localization was assessed during the course of sand fly infection. Parasites were generally observed in the endoperitrophic space by day 2 PBM and within the midgut lumen by day 5/6 PBM after blood meal defecation (Figs 4, S13 & S3–S5 Tables). On rare occasions, blood meal remnants were present in the AMG and hindgut by day 5/6 PBM. Invasion of the TMG (thoracic midgut) was observed to varying degrees by day 5/6 PBM and this increased in frequency and intensity by day 9 and 12 PBM (Figs 4, S13 & S3–S5 Tables). Infections with FVI and cDNA16 sKI concentrated strongly in the TMG by day 9 and 12 PBM and these were accompanied by TMG distention. All other mutant lines tested did not show this distension effect, with parasites being either evenly spread from the cardia to the posterior of the AMG or largely found in the AMG by day 12 PBM. An in vitro assay was used to demonstrate that parasite osmotaxis was not significantly compromised in any of the mutant lines compared to FVI; this could, therefore, be excluded as a factor affecting parasite accumulation in the TMG (Fig 5A). The observed TMG distention in late stage infection in FVI and cDNA16 sKI parasites only was indicative of the presence of PSG, an attachment matrix for leptomonads and nectomonads but proposed to be traversable by metacyclics (communication by Matthew E. Rogers). Mutant line infections lacked this gel, as observed by light microscopy (Fig 5B), suggesting compromised PSG generation in the mutant lines in vivo. To further investigate this observation, PSG was extracted from infected sand flies by pooling 10 infected midguts per infecting line. Samples were dot-blotted to activated nitrocellulose membranes and probed with the LT15 antibody that recognises L. major fPPGs, major components of the PSG [38]. PSG was only detected in FVI and cDNA16 sKI, but was undetectable in the other mutants tested. While parasites from all tested lines reached the cardia, significant differences were observed in the efficiency of SV colonization. FVI and cDNA16 sKI were the only lines tested to efficiently colonize the SV (Figs 4, 5B and 5C). While SHERP sKI and HASPB sKI were observed to attach at low numbers to the SV in P. (P.) papatasi (29.4% and 38.4% of analysed infected sand flies, respectively), mutant lines infecting P. (P.) duboscqi colonized the SV only weakly (<5% of cases; Fig 4), probably due to reduced haptomonad generation, the only parasite forms that attach to the SV. This hypothesis could not be verified due to the lack of any haptomonad specific markers.
Parasite morphology was analysed on Giemsa-stained gut smears from day 5/6, 9 and 12 PBM, using measurements of flagellum length, cell body length and width with separate analysis in AMG and TMG. This analysis showed significant differences in midgut metacyclogenesis between L. (L.) major lines at different time points (Figs 6 and S14). Only FVI and cDNA16 sKI produced metacyclics efficiently by day 12 PBM (P<0.001), compared to all other lines tested in the same vector species, although FVI metacyclic generation was significantly more efficient in P. (P.) duboscqi by day 12 PBM than in P. (P.) papatasi (P<0.001). These observations confirm that all other mutant lines tested could not complete metacyclogenesis in vivo, although they did so in vitro (Fig 2A and 2B). In addition, metacyclics were preferentially present in the TMG, while leptomonads were equally represented in the AMG and TMG (Figs 6 and S14) and nectomonads were preferentially found in the AMG. Thus all mutant lines, except cDNA16 sKI, resembled the cDNA16 null background with very few metacyclic-like parasites present at day 12 PBM and no gradient of differentiated parasites towards the TMG, correlating with the lack of parasite accumulation in the TMG and PSG-deficiency in these lines. Leptomonad generation in P. (P.) duboscqi showed no significant differences at day 12 PBM between all tested lines, except for HASPA1 sKI, which generated leptomonads very inefficiently (P<0.001, compared to all other lines tested in the same vector species). In P. (P.) papatasi, differences in leptomonad generation were only observed in SHERP sKI (P<0.001, compared to all other lines tested in the same vector species), which produced leptomonads as inefficiently as HASPA1 sKI in P. (P.) duboscqi. Differences in leptomonad generation between FVI and all other mutant lines by day 12 PBM were more pronounced in P. (P.) papatasi. Overall, no mutant line tested rescued the full parental line (FVI) phenotype, with the exception of cDNA16 sKI.
To further characterize HASPB and SHERP expression in the newly generated mutant lines, parasites derived either from culture or from sand fly midguts were fixed, antibody labelled for HASPB or SHERP and analysed by confocal microscopy. Both proteins were clearly detected in metacyclics derived from cultured FVI, cDNA16 sKI, HASPB sKI and SHERP sKI, respectively, while cDNA16 dKO promastigotes did not show fluorescence, as expected (Fig 7). FVI and cDNA16 sKI recovered from sand fly midguts were positive for HASPB and SHERP expression, too. However, unlike in vitro, HASPB sKI and SHERP sKI from sand fly midguts produced no detectable HASPB or SHERP signal, respectively. These observations suggest that either the gene regulation observed in these mutants in vitro is not replicated in vivo or that parasite differentiation is compromised in vivo, leading to a loss of HASPB and SHERP expression.
To investigate whether the lack of HASPB and SHERP expression was due to altered regulation at the protein or mRNA level, qRT-PCR was performed on midgut and culture-derived parasite mRNA. The HASPB mRNA levels of FVI and cDNA16 sKI, both having completed metacyclogenesis in vivo, showed elevated HASPB mRNA levels at day 6 PBM in vivo, subsequently declining towards day 12 PBM (Fig 8). Interestingly, HASPB sKI from midguts, which had not produced a detectable fluorescent HASPB signal (Fig 7), showed the same expression pattern, although total mRNA levels were lower than in FVI at all time-points, despite statistical compensation for the two HASPB copies in FVI compared to the one copy in HASPB sKI. In culture derived parasites, FVI and HASPB sKI showed similar expression patterns with peak expression of HASPB mRNA at day 7 p.i., although initial mRNA levels were higher (~1.39-fold) at day 3 p.i. in HASPB sKI compared to FVI. HASPB mRNA expression in cDNA16 sKI had already peaked at day 5 p.i. and declined by day 7 p.i. The downregulation of HASPB mRNA in vivo towards day 12 PBM was unexpected, since HASPB is expressed in metacyclics and amastigotes, and contradicts the in vitro observed mRNA expression pattern, which peaked at the final time-point (day 7 p.i.), when metacyclics were at their densest (Fig 8). For HASPA mRNA analysis, the HASPA2 sKI mutant was used, given that HASPA2 is known to be upregulated in early log-phase growth [18]. In the case of HASPA, the expression patterns of cDNA16 sKI and HASPA2 sKI were comparable to FVI in vitro, but distinct from FVI in vivo, while similar between cDNA16 sKI and HASPA2 sKI (Fig 8). In the case of SHERP, cDNA16 sKI and SHERP sKI had similar mRNA expression patterns both in vitro and in vivo, but these were distinct from FVI in both conditions (Fig 8). Overall, these results suggest that dysregulation of gene expression is not responsible for the lack of HASPB and SHERP detection in sand fly derived parasites, since the single-replacement lines either resembled FVI and/or cDNA16 sKI, which both develop normally in vivo.
Since it has previously been reported that culture conditions can influence promastigote development [39], we investigated whether the 20% FCS supplemented M199 medium used for in vitro culture influenced HASPB and SHERP expression, as compared to growth in 5% sucrose, mimicking the sugar-rich plant sap that provides nutrients in the sand fly midgut after blood meal defecation in vivo. Parasites grown for 2 days in M199 were washed and suspended in 5% sucrose/PBS solution, prior to collection of protein samples every 24 hours of the growth cycle. Comparative immunoblots of the whole parasite lysates (S15 Fig) provided no evidence for downregulation of HASPB and SHERP in HASPB sKI and SHERP sKI, respectively, when grown in 5% sucrose conditions, thereby excluding culture conditions as the source of differential HASPB and SHERP expression.
An alternative explanation for the observed differences in HASP and SHERP expression from replacement constructs in the sand fly could include a role for vector-derived regulatory factors. To further investigate this hypothesis, selected L. (L.) major lines were incubated either with (+) or without (-) homogenized midguts of uninfected blood-fed sand flies harvested at day 6 and 12 PBM. Parasite growth was monitored every 24 h and lysates harvested at day 6 p.i. were immunoblotted and analysed using ImageJ. This analysis showed that the addition of midgut homogenate affected parasite growth in culture (Fig 9A). All tested lines grew more slowly with day 6 PBM midgut homogenate as compared to negative controls, failing to reach stationary phase by day 6 p.i. Conversely, parasite growth rates with day 12 PBM midgut homogenates were comparable to negative controls until day 3 p.i. when the negative controls reached stationary phase with subsequent decline in cell numbers. By contrast, the day 12 homogenate-supplemented parasite population continued to expand slowly until day 6 p.i. All tested lines showed a significant increase in growth when day 12 homogenates were compared to day 6 homogenates (FVI: P = 0.027; cDNA16 dKO: P = 0.016; cDNA16 sKI: P = 0.013; HASPA2 sKI: P = 0.015). FVI and cDNA16 sKI grown with day 6 PBM homogenate also showed a reduction of detectable HASPA and HASPB compared to the negative control (FVI: 2.7-fold and 1.1-fold; cDNA16 sKI: 2.2-fold and 3.8-fold, respectively; Fig 9B and 9C). Cultures grown with day 12 midgut homogenate showed more limited reduction in HASPA and HASPB levels compared to the negative control (FVI: 1.5-fold and 1.1-fold; cDNA16 sKI 1.4-fold and 1.2-fold, respectively), although these differences could be a consequence of the observed differences in parasite growth and potential slowing of metacyclogenesis. HASPA2 sKI and HASPB sKI were also tested in this way (Fig 9D). Compared to FVI and cDNA16 sKI, HASPA2 sKI showed a similar response in cultures spiked with day 6 and 12 PBM midgut homogenates, while HASPB sKI showed increased HASPB expression in response to both midgut homogenates (Day 6 PBM: 1.45-fold; Day 12 PBM 1.2-fold).
Since completion of metacyclogenesis, PSG plug formation and SV degradation have been hypothesised to be essential for successful parasite transmission, we wanted to experimentally confirm the failure of our metacyclogenesis-impaired mutant lines to be transmitted in vivo. Due to the complexity of these experiments, we were limited to testing only a small subset of mutants: cDNA16 dKO and HASPB sKI (both unable to complete metacyclogenesis, produce PSG or colonize the SV in P. (P.) duboscqi) as representative lines hypothesised to be non-transmissible; and cDNA16 sKI and FVI (competent for metacyclogenesis, PSG production and SV colonisation) as parasite lines predicted to be successfully transmitted to a suitable host.
These experiments were conducted at the National Institute of Health (NIH), USA, using P. (P.) duboscqi, the same vector species used at the Charles University in Prague, CZ. Due to technical issues, the original FVI line was replaced with the FVI line available at the NIH (FVI (NIH)); both FVI lines are derived from the same original parent. Infection quality and metacyclogenesis progression in sand flies were monitored by parasite counting and morphometry, using light microscopy over a 14 day course PBM. FVI (NIH) presented similar infection loads (Fig 10A) as previously observed by qPCR for P. (P.) duboscqi in Prague (Fig 3B). While showing on average lower infection loads compared to FVI, HASPB sKI also showed comparable results between previous qPCR data and dissection (Figs 3B and 10A). cDNA16 dKO and cDNA16 sKI showed weaker infections compared to previous qPCR results (Figs 3B and 10A). The frequency (%) of metacyclics per midgut parasite load was also assessed, showing that only FVI (NIH) and cDNA16 sKI produced metacyclics efficiently in the vector (Fig 10B), as previously observed in the P. (P.) duboscqi colony from Prague (Fig 6), although FVI (NIH) was more efficient than cDNA16 sKI.
For the transmission experiments, sand flies infected with FVI (NIH), cDNA16 dKO, cDNA16 sKI or HASPB sKI were exposed on day 14 PBM to both ears of anaesthetised naïve mice. Genomic DNA was collected from the exposed ears and subjected to qPCR for parasite detection. As expected, the results showed successful parasite transmission for cDNA16 sKI and FVI (NIH) lines (37.5% and 27.3%, respectively, of analysed ear biopsies positive for parasite DNA), while cDNA16 dKO and HASPB sKI showed no evidence of transmission (Fig 10C). Interestingly, cDNA16 sKI was more efficiently transmitted than the FVI (NIH) control, despite the higher parasite loads and metacyclic content detected on average in FVI (NIH). This may relate to the higher virulence of cDNA16 sKI infections in mice (Fig 2A and 2B). However, for successful transmission, the arbitrary nature of sand fly feeding under these experimental conditions must also be considered. Since feeding success does not correlate with transmission success, this variable is difficult to quantify [40]. Either way, previous use of the FVI (NIH) line in this setting suggested that transmission success is commonly higher and comparable to cDNA16 sKI.
This study builds on our previous findings that the cDNA16 locus is essential for completion of metacyclogenesis in the sand fly midgut and that episomal expression of a HASPB gene copy can partially rescue the parental phenotype [30]. Here, we aimed to determine whether metacyclogenesis requires one gene from the cDNA16 locus, a subset of the HASP and/or SHERP genes, or the whole cDNA16 locus under parental gene regulation signals. Our in depth analysis has demonstrated a more complex picture.
The mRNA data reported here suggest that correctly regulated HASPB expression is the important event in metacyclogenesis in L. (L.) major, with cDNA16 sKI showing a similar expression pattern as FVI in vivo (Fig 8). This interpretation is further supported by the observed differences in HASPA and SHERP expression between the FVI and mutant lines, including cDNA16 sKI, suggesting that these genes are less likely to be required for completion of metacyclogenesis. The high expression of HASPB protein in L. (L.) major metacyclics [23] and the observation by Sádlová et al. (2010) that unregulated episomal expression of HASPB promoted metacyclogenesis completion further support this hypothesis [30]. Our data cannot directly prove that HASPB is the key player in metacyclogenesis in vivo, however, as the HASP and SHERP genes re-integrated alone into the cDNA16 locus are not expressed in the sand fly midgut. This is in contrast to the in vitro data collected during characterization of the same mutant lines, which show convincingly that the HASP and SHERP constructs are expressed and stage-regulated as in wild type parasites. Interestingly, replacement of the whole cDNA16 locus into its former location does recover the parental line phenotype in vivo, including detectable HASPB and SHERP product expression (Fig 7), while the step-by-step replacement of HASP and SHERP genes does not (see HA1/2+S2/HB sKI), suggesting that either the order and genomic context of the HASP and SHERP genes are vital to their correct regulation in vivo but not in vitro, or that trans-factors present in vivo play a role in parasite gene regulation in the sand fly.
A confounding factor here is the observed weak correlation between mRNA and protein abundance in Leishmania species [41]. In trypanosomes, mRNAs can be stored in cytoplasmic RNA granules until translation, potentially accounting for this discrepancy [42]. Recently, formation of mRNA granules has also been described for Leishmania under stress conditions, such as starvation [43], previously demonstrated to be a trigger for metacyclogenesis in vitro [6,7]. In the case of stage-regulated genes in Leishmania, as in other protozoan species [44], mRNA up-regulation can occur at a developmental stage prior to protein expression [41]. Although not definitively shown in Leishmania, it is feasible that these mRNAs are stored in RNA granules prior to entry into the translational machinery. Such sequestration could accelerate adaptational responses in dynamically changing environments, such as in the sand fly midgut, thereby promoting parasite survival. It is not known whether HASP and SHERP mRNAs are stored in RNA granules nor what signals might operate in vivo to initiate their subsequent translation. However, one testable hypothesis might propose that HASP and SHERP mRNAs are indeed sequestered cyoplasmically and “bypassed” in vitro due to a lack of midgut-specific signals in culture, thereby explaining the observed differences in HASP and SHERP mRNA and product expression in vitro and in vivo.
The above observations suggest that HASP and SHERP function could be related to cellular mechanisms required for survival and development under sand fly midgut conditions. In this dynamic environment, alterations in pH, temperature, amino acid and digestive enzyme content occur during parasite differentiation, changes that are only partially reproduced under culture conditions. The importance of in vivo passaging to restore virulence in L. (L.) major and other Leishmania spp. has been previously demonstrated [40,45,46]. Prolonged in vitro culture frequently causes parasite avirulence [47], which can be partially restored by repeated passaging through a susceptible mammalian host or sand fly vector [45]. Similar effects have been reported in other parasitic systems: culture-adapted Trypanosoma brucei bloodstream parasites, for example, have a 1,000x lower antigen switching rate for their variable surface glycoproteins (VSG) than in natural isolates [48,49]. Interestingly, the addition of crude uninfected blood-fed midgut homogenates to cultured Leishmania promastigotes in this current study indicated potential regulatory and temporal influences on parasite growth and HASPA and HASPB expression in the period following the blood meal time (Fig 9). Considering that the midgut is subject to dynamic changes over the course of Leishmania infection, reflecting the transition from blood meal to sugar digestion, our data suggest that Leishmania could utilise changes in the midgut content to drive its gene regulation and subsequent proliferation and differentiation. Well-established examples of this type of parasite/host interaction can be found in other kinetoplastid species such as T. brucei, during differentiation into stumpy forms [50,51] and trypomastigotes [52–54] and during gametocytogenesis in Plasmodium spp. [55].
It is also possible that differences in mutant gene expression are chromosomal context dependent. The observation that the LmjcDNA16 sKI line is the only one of the range of mutants generated to express HASPB and SHERP in vivo (Fig 7) and rescue metacyclogenesis (Figs 4 and 6), as described above, suggests that the order and context in which the HASP and SHERP genes occur in the cDNA16 locus are relevant for correct in vivo regulation. Accurate polycistronic RNA processing for the production of mature mRNAs in Leishmania requires the correct positioning of downstream gene 5’-splice acceptor sites relative to the upstream gene polyadenylation sites [56]. Clearly our constructs, with their flanking sequences and DHFR-flanked antibiotic-resistance genes, are expressed appropriately in vitro but not in vivo, suggesting that other factors are important in regulating expression from our constructs in the sand fly midgut. While further investigation into HASP and SHERP function is highly desirable, the lack of HASP and SHERP mutant phenotypes in vitro hampers further rapid investigation due to the complexity of parasite manipulation in the sand fly midgut and the limiting amounts of biological material recoverable from infected sand flies.
Sádlová et al. [30] showed that cDNA16 dKO parasites grown in vitro secrete fPPG, constituent of the PSG, into the culture medium. By contrast here, only FVI and cDNA16 sKI secrete detectable amounts of fPPG in the sand fly midgut (Fig 5C) leading to formation of PSG that is detectable by light microscopy (Fig 5B). While further analysis using a panel of relevant antibody probes would be desirable, our current data suggest that PSG is not produced in the cDNA16 mutant lines in vivo. It is unclear why fPPG secretion and PSG formation should be impaired in vivo in the sand fly midgut but not in vitro, and how the HASPs and SHERP contribute to this process. The PPG synthetic pathway, while not completely described in Leishmania, involves secretion via the ER and Golgi in other eukaryotic cells [57], two cellular compartments that the HASPs and SHERP do not appear to enter. Rather, N-myristoylated HASPs are palmitoylated on the cytosolic face of the Golgi, prior to transport to the plasma membrane [20], while SHERP associates intracellularly as a peripheral membrane protein [27], interacting in vitro with a sub-unit of a vacuolar type H(+)-ATPase that functions in acidification [28]. There is no evidence to suggest that SHERP affects secretion via the ER/Golgi route, but it cannot be excluded that this small protein influences fPPG synthesis and/or secretion indirectly by its interactions with intracellular membrane constituents [27]. Alternatively, the lack of fPPG secretion could be a consequence of incomplete metacyclogenesis in the HASP and SHERP mutants preventing maturation of fPPG-secreting leptomonads. This seems unlikely, given that leptomonad forms were observed by morphometry in comparable numbers in HA1/2+S2/HB sKI, FVI and cDNA16 sKI, while PSG formation in the TMG was only observed in FVI and cDNA16 sKI infections. Since there are no molecular markers for mature leptomonads, it cannot be excluded that the parasite forms observed by morphometry were functionally immature, however. Further investigation is required to determine the relationship between HASP and SHERP deletion and the lack of PSG formation.
PSG has been shown to play a role in TMG colonization in the sand fly [13,58], which could explain why several mutant lines lacking fPPG failed to establish mature infections and to accumulate in the TMG, in particular, in P. (P.) papatasi. The PSG plug forces sand flies to regurgitate prior to blood meal intake, supporting parasite transmission into the host skin [38]. Once in the skin, glycans donated by the fPPGs promote the recruitment of neutrophils and macrophages [38], prior to macrophage invasion [59,60]. Further, PSG induces the alternative activation of macrophages, promoting arginase-1 activation and antagonising nitric oxide synthase 2 (NOS2), thereby facilitating parasite survival [60]. Therefore, a lack of PSG may impair sand fly transmission and reduce the likelihood of parasite survival in the mammalian host. While we were able to show impaired transmission in our PSG-deficient mutant lines in vivo, the same lines cultured in vitro were all able to establish persistent infection in BALB/c mice following sub-cutaneous high dose needle infection of 107 late-stage parasites. Interestingly, several PSG-deficient lines, in particular cDNA16 dKO, HASPA1 sKI and HASPA1/2 sKI, were more virulent in BALB/c mice than FVI, while other lines were significantly less virulent (e.g. HASPA2 sKI). However, it must be noted that in vitro, these parasite lines were able to complete metacyclogenesis and to secrete fPPG [30]. In vivo-derived parasites from these PSG-deficient lines show a lack of metacyclic generation, suggesting that parasites would not be infective, even if they were transmitted from the sand fly to a mammalian host. Since our data showed that our mutant lines are not transmissible, with the exception of cDNA16 sKI, this question could not be addressed in a natural transmission scenario. Even so, our results confirmed experimentally that completion of metacyclogenesis, PSG secretion and SV colonization, which are all hallmarks of mature parasite infections in sand flies, are essential for successful parasite transmission, as is the presence of the cDNA 16 gene locus.
On artificial challenge in BALB/c mice with our mutant lines, significant differences were observed in pathology between mutant lines. As previously observed by McKean et al. (2001), the full cDNA16 locus deletion mutant (cDNA16 dKO) caused faster footpad swelling than the parental line (FVI; Fig 2C). Interestingly, the phenotype of the full cDNA16 locus replacement line (cDNA16 sKI) was intermediate, suggesting gene dose dependency of the observed phenotype. Since HASPA1 sKI has a similar phenotype to cDNA16 sKI, while HA2+S2/HB sKI has a similar phenotype to FVI, this suggests that the cDNA16 sKI intermediate phenotype is HASPA1 gene dependent. Although not directly demonstrated, these data suggest that the deletion of the cDNA16 locus products increases inflammatory responses (either directly or by pleiotropic effects), a phenotype moderated by re-introduction of one copy of the locus containing half the wild type copy number of HASP and SHERP genes. This hypothesis is supported by the observation in our previous study [31] that an episomal cDNA16 replacement mutant, overexpressing HASP and SHERP, was avirulent. While HASPA1 sKI showed a similar phenotype as cDNA16 sKI, the HASPA2 sKI mutant line showed a significantly delayed onset of lesion formation, despite equivalent inoculum and in vitro capacity for metacyclic formation (Fig 2A and 2B). This observation was of interest because HASPA2 protein expression is promastigote-specific, while HASPA1 protein expression is amastigote-specific (Figs 1D and 2E), correlating with the mRNA expression data for these genes [17]. Introducing a HASPA1 copy with a HASPA2 copy (HASPA1/2 sKI) caused a similar phenotype as HASPA1 sKI (Fig 2C). Interestingly, the parasite burden per footpad at 2 mm swelling was similar for all tested lines; only the time point when these were reached varied between the lines. These observations could indicate functional differences in the HASPA proteins during the parasite life cycle, potentially affecting initial metacyclic parasite survival post inoculation or amastigote proliferation rather than disease progression per se; a surprising conclusion given their identical protein sequences.
Based on the results presented here, we confirm a role for both the HASP and SHERP proteins in Leishmania metacyclogenesis in the sand fly, leading to parasite transmission, and propose a subsequent role for the HASP proteins in establishing infection in the mammalian host. We also show for the first time that sand fly midgut extracts collected post-blood meal affect parasite behaviour in vitro over time as the midgut luminal content changes. While biochemical fractionation and in depth analysis of the midgut lysates are now required to advance these studies, they may offer a novel approach to simulate in vivo phenotypes in vitro while also affirming the importance of vector-parasite interactions in vivo.
L. (L.) major Friedlin V1 (MHOM/IL/81/Friedlin/VI; FVI) was used as a parental line/wild type in this study, as for all previous work on this locus. The previously described L. (L.) major 4.8 cDNA16 double deletion (cDNA16 dKO; ΔcDNA16::HYG/ΔcDNA16::PAC; [31]) and cDNA16 single replacement lines (cDNA16 sKI; ΔcDNA16::HYG/ΔcDNA16::PAC/ΔPAC::cDNA16+ NEO [30]) were used as controls. New L. (L.) major HASP and SHERP replacement lines were generated by homologous recombination with newly synthesized linear DNA constructs into the former cDNA16 locus within the null or other mutant backgrounds, using the nucleofection kit (Amaxa) according to the suppliers’ guidelines. The DNA constructs contained one or two HASP and/or SHERP gene(s) with their native 5’ and 3’ UTRs and a selectable antibiotic resistance marker gene (either NEO or BSD). Correct genomic integration of DNA constructs was verified by Southern blot and quantitative PCR (qPCR) followed by parasite passage through BALB/c mice (described below). Promastigotes of all L. (L.) major lines were routinely cultured in Medium 199 (M199) (supplemented with 20% v/v heat-inactivated foetal calf serum, 1% v/v penicillin-streptomycin) as described [31]. For parasite in vitro differentiation, parasites grown until late log-phase in M199 were harvested, washed and suspended in 5% sucrose/PBS (described below).
Metacyclic purification by peanut agglutination (PNA; Sigma) was originally described elsewhere [61]. We used 50 μg/ml PNA for 15 min. at room temperature with regular agitation of parasite suspensions to separate agglutinated promastigote forms from metacyclics by slow centrifugation (175xg). Metacyclic ratios in day 7 cultures were established by pre- and post-agglutination counts on a haemocytometer.
The protocol for Southern analysis has been described elsewhere [30]. Genomic DNA (gDNA) was extracted using DNeasy Blood & Tissue columns (Qiagen), SacI-digested, separated by 0.8% agarose gel electrophoresis and blotted onto positively charged nitrocellulose membranes. Digoxigenin (DIG) labelled probes were used for detection with the DIG-development system (Roche).
Whole parasite lysates in Laemmli buffer (20 ml 0.5 M Tris-HCl [pH 6.8], 3.08 mg DTT, 40 ml SDS [10%], 50 mg Bromophenol Blue, 20 ml Glycerol [100%] and sterile Milli-Q water (MQH2O) to 100 ml) were separated by SDS-PAGE as described [31] and blots analysed using affinity-purified polyclonal rabbit antibodies against SHERP [27], HASPB (336; [31]) or non-affinity-purified polyclonal HASP antibodies for HASPA-detection. A secondary anti-rabbit HRP antibody (Sigma) was used with ECL Prime (Amersham) for detection. A polyclonal antibody against L. (L.) major N-myristoyltransferase (NMT; [36]) was used as a loading control. In case of the growth condition assay, band intensity on immunoblots were analysed using ImageJ. Values were normalized first for each gene individually (HASPA2, HASPB and NMT, respectively) against negative control to compensate variations within the image. Then normalized HASPA and HASPB values were normalized against the corresponding NMT loading control to adjust for loading variations prior to calculating the fold-difference between (+) and (-) conditions.
PSG was extracted from 10 infected midguts pooled into 50 μl PBS, as described [60]. Whole lysates of debris pellets and 6x spun PSG containing supernatants were blotted on to activated nitrocellulose membranes using an adapted protocol [60]. Dot blots were probed for PSG with the LT15 antibody against phosphoglycan disaccharide repeats [PO4-6Galb1-4Mana1-]x [38] and treated as for the immunoblots using an anti-mouse HRP secondary antibody (Sigma).
The osmotaxis assay was adapted from Leslie et al. [62] and Oliveira et al. [63]. Briefly, plain glass capillary tubes (75 mm length, 0.8 inner/1 mm outer diameter) were filled with wash and incubation (WIS) buffer (30 mM β-glycerophosphate disodium salt, 87 mM NaCl, 27 mM KCl, 2 mM CaCl2, 2 mM MgCl2, 0.004% enriched Bovine Serum Albumin [pH 7.1]) containing 1% agarose ± 100 mM of sucrose, leaving exactly 1 cm void (~5 μl). Once set, the void was filled with WIS buffer. The glass capillaries were equilibrated in WIS buffer for ~30 min. at room temperature on a rocking table. Parasites were grown to late log-phase/early stationary-phase in M199, harvested, washed twice in WIS buffer and suspended to a final concentration of ~2.5x107 cells/ml in WIS buffer. The equilibrated glass capillary tubes were dipped into the parasite suspension at a slight angle (6 capillary tubes with 100 mM sucrose and 6 without sucrose were used per strain) and incubated at 26°C for 1 h. The WIS buffer in the capillary void was removed, mixed with 1% formaldehyde in saline solution and applied to a haemocytometer for parasite counting.
At the Charles University, Prague, sand flies were maintained at 26°C and high humidity (75%) on 50% sucrose solution and a 14 h light/10 h dark photoperiod as described [64]. Sand fly infections and analysis were carried out as described by Sádlová et al. [30]. For infections, colony-bred Phlebotomus (Phlebotomus) papatasi and P. (P.) duboscqi (Turkey and Senegal strains, respectively; reared at the Department of Parasitology, Charles University, Prague) were fed on infected (106 parasites/ml) heat-inactivated rabbit blood through a chick-skin membrane for up to 2 h in the dark. Unfed sand flies were separated from engorged females, which were sampled at day 2, 5, 9 and 12 post blood meal (PBM) or at day 6 and 12 PBM only. Light microscopic analysis of dissected midguts was used to establish parasite localization (endoperitrophic space, AMG to cardia, attached to the stomodeal valve either weakly or strongly) and parasite load per midgut for all dissection points; scoring adapted from Myšková et al. [65] as light (<100 parasites/gut), moderate (100–1000 parasites/gut), heavy (>1000 parasites/gut) and very heavy (>>1000 parasites/gut). Dissected midguts were scored by two independent researchers. In addition, 30 infected female sand flies from day 12 PBM were individually frozen at -20°C in buffer for DNA extraction and parasite loads were scored by qPCR as described [30]. Sand fly infection experiments were repeated at least three times per tested line.
At the National Institute of Health (NIH), Rockville, USA, sand flies were maintained in similar conditions. For infection, colony-bred 2- to 4-day-old P. (P.) duboscqi females (Mali strain; reared at the Laboratory of Malaria and Vector Research, NIAID) were infected by artificial feeding through a chick skin membrane on defibrinated rabbit blood (Spring Valley Laboratories, Sykesville, MD) containing 350 units/ml penicillin, 350 μg/ml streptomycin, and 3–4 x 106 procyclics or amastigotes/ml (P1-P5) from four L. major lines (FVI, cDNA16 dKO, cDNA16 sKI and HASPB sKI) initially isolated from BALB/c footpad lesions. After 3 h of feeding in the dark, fully blood-fed sand flies were separated. For the initial experiment, the total number of parasites and percent metacyclics per midgut were established at different days after infection (D2, D7, D9 and D14; data on for D14 shown in Fig 10) for each group. Thereafter, in two independent repeats, the total number of parasites and the percentage of metacyclics were determined on the day of transmission. After dissection, each midgut was placed in 50 μl of PBS in a microcentrifuge tube, macerated with a plastic pestle (Kimble Chase) and parasites counted using a haemocytometer; metacyclic forms were distinguished by morphology and movement [66].
Quantitative PCR was performed on genomic DNA samples of selected new mutant clones to establish integrated copy number using Fast SYBR Master Mix (Applied Biolscience) in the OneStep qPCR system (Life technologies) with the OneStep software v.2.2.2 according to the supplier’s guidelines. Target genes were detected with gene-specific primers and results were normalized against the Na/H antiporter-like protein gene (see S6 Table). The protocol for detection of parasite load is described elsewhere [30,65]. Primers specific for kinetoplast minicircle DNA were used for parasite detection as described by Mary et al. [67] (S6 Table).
Parasites were grown for 3 days in 10 ml complete M199 at 26°C, pelleted and washed twice in sterile PBS at room temperature. Parasites resuspended in 10 ml 5% sucrose/PBS were incubated for an additional 4 days at 26°C. 107 parasites were collected every 24 h and lysed in Laemmli buffer for Western analysis.
Sand flies were fed on heat inactivated uninfected rabbit blood through a chick-skin membrane for up to 2 h in the dark. 50 blood fed midguts were dissected at day 6 and 12 PBM into 200 μL M199 + Amikin (250 μg/ml) + penicillin (60 μg/ml) + fluorocytosin (1.5 mg/ml). The midguts were homogenised and filter through a 0.22 μm filter spinning column (Ultrafree—MC, GV dutapore). The midgut extract was added to 4 ml of M199 and 1 ml was aliquoted into culture tubes. M199 was chosen as the medium of choice given its high serum content, which reduced the impact of additional protein introduced via the midgut homogenate. FVI, LmjcDNA16 dKO, LmjcDNA16 sKI and LmjHASPB sKI were inoculated into the 1 ml medium, respectively, to a final concentration of 105 parasites/ml and left to grow for 6 days. Parasite density was established by parasites counting on a haemocytometer. At day 6 p.i., parasites were pelleted and washed twice in PBS before lysis in 50 μl 1x Laemmli buffer at 95°C for 10 min.
As described previously [30], midguts were divided into AMG and TMG, smeared on to glass slides, fixed with 100% methanol and stained with Giemsa. 160 randomly chosen parasites were imaged (using a 100x oil-immersion objective, Olympus BX51 fluorescent microscope, Olympus DP70 camera) per midgut section per day per strain on at least three randomly chosen gut smear slides. For cultured parasites, cell pellets were washed twice in saline solution (0.9% NaCl), applied to polylysine slides for 15 min. and the excess suspension tipped off. Glass slides were air dried, fixed with 100% methanol, rinsed with water, stained with Giemsa and analysed by microscopy. 100 randomly chosen parasites per slide were imaged (using a 63x oil-immersion objective, Zeiss Axioplan microscope, Optronics 60800 camera system).
For immunofluorescence analysis, cultured parasites were harvested, washed, fixed in 4% formaldehyde and applied to polylysine glass slides as described [21,68]. Parasites derived from sand flies were analysed in 100% methanol fixed gut smears as above. Cells were permeabilized with 0.2% Triton-X 100/PBS and blocked with Image-iT FX signal enhancer (Invitrogen) prior to analysis of HASPB or SHERP expression using polyclonal anti-HASPB (336; [32]) and anti-SHERP [27] antibodies, respectively, followed by Alexa Fluor 488 Dye (Invitrogen) secondary antibody. Samples were mounted either in Vectashield or Mowviol with 4', 6-diamidino-2-phenylindole (DAPI; Vector) and imaged with Zeiss LSM 510 or 710 META confocal microscopes.
Flagellum and cell body length and width were measured with Image J [69]. Parasites were classified into four groups adapted from Walters (1993) and Ciháková and Volf (1997): (i) procyclics < 14 μm body length ≥ 4 μm, flagellum shorter than cell body; (ii) nectomonads: body length ≥ 14 μm; (iii) leptomonads: body length < 14 μm and flagella length < 2 times body length and (iv) metacyclics: body length < 14 μm and flagella length ≥ 2 times body length. Paramastigotes and haptomonads were not considered. In vivo, haptomonad presence or absence was determined by the state of SV colonization.
Experiments routinely used 6–8 weeks old BALB/c mice (Harlan Laboratories, UK). For footpad infections, 3x 107 late-stage stationary parasites were injected in 30 μl PBS (adapted from Depledge et al. [70]). For routine parasite passage to re-establish virulence, amastigotes were harvested from draining lymph nodes 8–10 weeks post footpad infection and inoculated into M199 for incubation at 26°C.
To establish parasite virulence, footpads of groups of 5 female BALB/c mice were infected, ears were marked and mice placed randomly in cages. Mice were chosen blindly for footpad measurement once a week until footpad lesion reached 2 mm, at which point they were sacrificed. Parasite burdens per footpad were established after mouse-scarification by a limiting dilution assay (LDA) of infected footpads, adapted from Titus et al. [71] and Lima et al. [72] and calculated by the online Leishmania LDA analysis tool available on the Imperial College London homepage: (http://wwwf.imperial.ac.uk/theoreticalimmunology/llda/) [73].
Amastigotes were isolated using a protocol adapted from Paape and Aebischer (2011)[74]. BALB/c mice were infected at the base of the tail on both sides by needle inoculation of 3x107 late stationary parasites. After lesions developed (8–10 weeks p.i.), mice were sacrificed and lesion material excised with a scalpel, weighed, forced through a 70 μm cell strainer into homogenization buffer (20 mM HEPES-KOH, pH 7.3, 0.25 M sucrose supplemented with cOmplete Mini proteinase inhibitor cocktail [Roche]) and washed once in homogenization buffer. Amastigotes were released by forcing the cell suspension through a 25-gauge needle. Nuclei were removed by centrifugation at 100xg for 2 min. The supernatants were loaded onto a discontinuous sucrose gradient: 20, 40, and 60% (w/w) sucrose in HEPES saline (30 mM HEPES-KOH, pH 7.3, 0.1 M NaCl, 0.5 mM CaCl2, 0.5 mM MgCl2) [75], centrifuged for 25 min. at 700xg. Amastigotes were isolated from the 40/60% sucrose interface, diluted in PBS and washed once in PBS.
Experiments at the NIH were performed using 6- to 8-week-old BALB/c mice (Charles River Laboratories Inc.), maintained under pathogen-free conditions. To assess the capacity of mutants to transmit to BALB/c mice, 10–20 (as available) infected P. (P.) duboscqi females from each group (FVI, cDNA16 dKO, cDNA16 sKI and HASPB sKI) on day 14 PDM were exposed to each mouse ear in the first experiment. Thereafter, validation of transmission was carried out for the groups where mature infections containing metacyclics were observed (FVI and cDNA16 sKI). Mice were anesthetized intraperitoneally with a mixture of ketamine (100 mg/kg) and xylazine (10 mg/kg). Infected sand flies were placed in vials with a meshed surface and applied to the ears using custom-made clamps. The flies were allowed to feed for two hours in the dark. Mouse ears were removed two hours after exposure to infected flies and frozen at -70°C until processed. Total genomic DNA was extracted using the DNeasy tissue kit following the manufacturer’s protocol (Qiagen). A total of 50 ng of sample DNA was amplified in triplicate by real time PCR (Biorad c1000 thermal cycler and cfx96 real time system) using primers JW11 and JW12 [76] together with a Ld3C6 fluorescent probe [77] targeting kinetoplast minicircle DNA. Parasite numbers were determined by the real time system software based on a standard curve of serially diluted L. major (WT Friedlin V1) DNA. The cut-off was based on values obtained for naïve DNA controls.
Animal experiments in York were approved by the University of York Animal Procedures and Ethics Committee and performed under UK Home Office license (‘Immunity and Immunopathology of Leishmaniasis’ Ref # PPL 60/4377). All animal experimental procedures at the NIH were reviewed and approved by the National Institute of Allergy and Infectious Diseases (NIAID) Animal Care and Use Committee under animal protocol LMVR4E. The NIAID DIR Animal Care and Use Program complies with the Guide for the Care and Use of Laboratory Animals and with the NIH Office of Animal Care and Use and Animal Research Advisory Committee guidelines.
All statistical analysis was done with the SPSS software v.20-22 (IBM) or GraphPad Prism 5. P<0.05 was considered to be significant. Categorical data were analysed by χ2 test. Normally distributed continuous data were analysed by one-way ANOVA and post-hoc Tukey multiple comparison test. Non-normally distributed data were normalized by log10 or square root transformation, as appropriate, and submitted to parametric analysis where possible. Otherwise, non-normally distributed data were analysed by Kruskal-Wallis and post-hoc Dunn’s multiple comparison test. Growth assay and footpad lesion data were analysed by repeat measure ANOVA and post-hoc Tukey multiple comparison test or by Freidman test and post-hoc Dunn’s multiple comparison test.
|
10.1371/journal.pntd.0001015 | Oviposition Site Selection by the Dengue Vector Aedes aegypti and Its Implications for Dengue Control | Because no dengue vaccine or antiviral therapy is commercially available, controlling the primary mosquito vector, Aedes aegypti, is currently the only means to prevent dengue outbreaks. Traditional models of Ae. aegypti assume that population dynamics are regulated by density-dependent larval competition for food and little affected by oviposition behavior. Due to direct impacts on offspring survival and development, however, mosquito choice in oviposition site can have important consequences for population regulation that should be taken into account when designing vector control programs.
We examined oviposition patterns by Ae. aegypti among 591 naturally occurring containers and a set of experimental containers in Iquitos, Peru. Using larval starvation bioassays as an indirect measure of container food content, we assessed whether females select containers with the most food for their offspring. Our data indicate that choice of egg-laying site is influenced by conspecific larvae and pupae, container fill method, container size, lid, and sun exposure. Although larval food positively influenced oviposition, our results did not support the hypothesis that females act primarily to maximize food for larvae. Females were most strongly attracted to sites containing immature conspecifics, even when potential competitors for their progeny were present in abundance.
Due to strong conspecific attraction, egg-laying behavior may contribute more to regulating Ae. aegypti populations than previously thought. If highly infested containers are targeted for removal or larvicide application, females that would have preferentially oviposited in those sites may instead distribute their eggs among other suitable, previously unoccupied containers. Strategies that kill mosquitoes late in their development (i.e., insect growth regulators that kill pupae rather than larvae) will enhance vector control by creating “egg sinks,” treated sites that exploit conspecific attraction of ovipositing females, but reduce emergence of adult mosquitoes via density-dependent larval competition and late acting insecticide.
| Controlling the mosquito Aedes aegypti is of public health importance because, at present, it is the only means to stop dengue virus transmission. Implementing successful mosquito control programs requires understanding what factors regulate population abundance, as well as anticipating how mosquitoes may adapt to control measures. In some species of mosquitoes, females choose egg-laying sites to improve the survival and growth of their offspring, a behavior that ultimately influences population distribution and abundance. In the current study, we tested whether Ae. aegypti actively choose the containers in which they lay their eggs and determined what cues are most relevant to that process. We also explored whether females select containers that provide the most food for their larval progeny. Surprisingly, egg-laying females were most attracted to sites containing other immature Ae. aegypti, rather than to sites containing the most food. We propose that this behavior may contribute to density-dependent competition for food among larvae and play a larger role than previously thought in regulating Ae. aegypti populations. We recommend that accounting for, and even taking advantage of, this natural behavior will lead to more effective strategies for dengue prevention.
| Dengue viruses are transmitted to humans primarily by the mosquito Aedes aegypti and represent an increasing public health concern in tropical and subtropical regions worldwide. Because no vaccine or antiviral therapy is commercially available, controlling the mosquito vector is the only current means to prevent dengue outbreaks [1]. Contemporary control campaigns, rather than attempting to eradicate Ae. aegypti, aim to suppress mosquito populations below a threshold density at which they no longer support viral amplification [2]. Controlling adult mosquitoes is made challenging by the behavior of domestic Ae. aegypti. Adult Ae. aegypti rest inside homes, typically on clothing, curtains, bedspreads, and furniture, items that cannot be sprayed with residual insecticides [3]. Aerosol space sprays consist of small airborne droplets of insecticide designed to kill adult mosquitoes on contact, but difficulty in reaching indoor adult resting sites can limit their efficacy [4]. Even when space sprays are effective in reducing adult populations, effects are transient due to the continuing emergence of new adults or immigration from untreated areas [3], [5]. Insecticide-treated materials (curtains, water container covers, and bednets) have shown promise in reducing Ae. aegypti populations [6], [7], but the impact of these reductions on dengue transmission has not been determined.
Currently, the World Health Organization recommends directing routine Ae. aegypti control toward the immature stages [2]. Ae. aegypti females lay eggs singly just above the water line, often in man-made containers located in the home or yard (buckets, drums, tires, and vases, etc.) [8]–[10]. Eggs hatch when inundated, and larvae develop by filter feeding and browsing for microorganisms and organic detritus [11], [12]. Control approaches such as container removal (source reduction) and larvicide application aim to reduce the number of new emerging adults in the population [2]. Traditional models of Ae. aegypti assume that population dynamics are regulated predominantly by density-dependent competition for food during early larval stages and little affected by oviposition rates [13], [14]. Based on these models, some researchers have assumed that all containers suitable for larval development receive an excess of eggs, thereby leading all larvae to experience density-dependent competition [15]. This is the rationale behind targeted source reduction (a WHO-recommended control strategy) and the expectation that eliminating containers that, for example, produce 75% of adults will lead to a proportionate decrease in the overall adult population [15]–[17]. Much remains unclear, however, about the factors regulating Ae. aegypti adult production and how reducing, but not eliminating, containers will ultimately affect adult abundance.
In some mosquito species, female choice in oviposition site is adaptive and can influence population distribution and dynamics [18], [19]. Females can enhance survival and development of their offspring by selecting egg-laying sites that reduce exposure to predators and competitors [19], [20], or increase access to food [21], [22]. In general, understanding insect egg-laying decisions may provide additional insight into the factors affecting population regulation and aid in predicting how populations will respond to control measures [23]. Oviposition preferences by Ae. aegypti have been studied in the laboratory [24]–[31], but to a lesser extent in the field [32]–[36]. Research has typically involved varying one or two oviposition site factors at a time and observing the number of eggs laid in response (reviewed in [24]). Such studies reveal the types of abiotic and biotic stimuli potentially affecting oviposition, but yield limited information on the relative importance of these stimuli in nature [24], [37].
The goals of our study were to test whether free-ranging Ae. aegypti females make active choices regarding where they oviposit and to identify factors influencing oviposition. Although selective oviposition has been demonstrated using small oviposition traps in the field [32]–[35] or water-storage containers in an enclosure [36], we examined for the first time females' oviposition choices among naturally-occurring containers in homes throughout a large, dengue-endemic city. We also investigated the consequences of oviposition site selection for offspring fitness by testing whether females choose sites to maximize the amount of food available for their progeny. Food availability is known to affect components of mosquito fitness such as offspring survival, development time, and adult size [38]. Lastly, we considered the implications of selective oviposition behavior for Ae. aegypti population regulation and the success of targeted larval control strategies.
Our study was conducted in Iquitos (73.2°W, 3.7°S, 120 m above sea level), a city of approximately 380,000 people located in the Amazon Basin, Department of Loreto, Northeastern Peru [10], [39]–[41]. Rain falls during all months of the year and average temperature and relative humidity are fairly consistent [42]. During our study period from July 2007 to August 2009, mean monthly temperature ranged from 24.8°C (±1.1 SD) in June 2008 to 26.5°C (±1.1 SD) in December 2008. Average relative humidity ranged from 80.2% (±4.1 SD) in August 2007 to 86.2% (±4.4 SD) in April 2009. More detailed climate data for the years 2007 to 2009 are given in the Supporting Information (Table S1).
In response to the unreliable municipal water supply, Iquitos residents store water in containers [40]. Household containers are filled in three primary ways: 1) from spigots in the home or neighborhood (manually filled), 2) intentionally placed outside to collect rain water (rain-filled), and 3) filled with rain water as a result of being untended outside (unmanaged). Method of filling is correlated with the frequency of water turnover and amount of organic detritus present in containers, with manually filled containers kept the cleanest and unmanaged containers collecting the most organic material. Containers in Iquitos generally lack predators of larval Ae. aegypti, such as copepods or fish (ACM and JW, unpublished data), but do occasionally contain immature Culex which may act as competitors [10]. Ae. aegypti are reproductively active all year in Iquitos. Of the roughly 290,000 containers examined by Morrison et al. [10], 7.3% contained Ae. aegypti larvae and/or pupae.
We measured larval resistance to starvation (RS, number of days larvae survive without food) as an indirect measure of per capita food availability in containers [38]. In general, mosquito larvae that consume more food are able to store more energy reserves and resist starvation longer [13], [44]. During the above-described survey of Iquitos containers, 5 to 25 third instar larvae were removed from containers in the field and transferred to individual plastic cups (5 cm diameter×6 cm height) filled to 2/3 capacity with bottled drinking water. Third instars were used for bioassays because fourth instar Ae. aegypti frequently pupate when starved [44]. Cups were placed indoors in our field laboratory, where larvae were exposed to natural light and temperature. Water was changed every two days to prevent accumulation of waste and microbial growth [51]. Time to death (in days) was recorded for each larva. Because starvation times were not distributed symmetrically for larvae from each container, the median larval RS was used as the measure of central tendency for the data for each container.
Spearman rank correlation was used to identify any association between larval RS and egg density (mean eggs laid per day/circumference). Data were stratified according to whether or not all larvae had been removed from containers on the first survey day. To account for potential effects of larval abundance and container size, data also were stratified by larval abundance (≤50 larvae vs. >50 larvae) and container capacity (≤20 L vs. >20 L).
For 12 weeks during June to August 2009, we carried out an experimental study manipulating both the presence of conspecific larvae and accumulation of organic material in containers and recorded oviposition by wild females. This experiment was replicated in three central Iquitos residences, the courtyard of our field laboratory and in the yards of two other houses selected based on the consistent presence of Ae. aegypti and homeowner willingness to participate. At each residence, three identical 6-liter blue plastic buckets (20 cm diameter×23 cm height) were placed close to one another (0.5 m apart) to minimize differences in container position. Hourly at each house, ambient temperature and relative humidity were recorded using a Hobo® ProV2 data logger (U23-001; Onset Computer Corporation, Pocasset, MA) and water temperature was recorded in one container per house using a Hobo® Pendant logger (UA-002-64).
We created three container treatments: A (unmanaged, with larvae), B (unmanaged, no larvae), and C (manually filled, no larvae). Unmanaged containers (A and B) were filled with four liters of tap water and allowed to accumulate organic debris for 12 weeks, whereas manually filled containers (C) were cleaned and refilled with new tap water every other day. Fifty first instar Ae. aegypti larvae were introduced into treatment A containers every two weeks starting on the first day. Oviposition was monitored by lining the inside of buckets with strips of brown paper towel to collect eggs. Every second day, papers were exchanged and the number of eggs counted as described above. On egg collection days, we also temporarily removed larvae from treatment A containers to determine their developmental stage and count them. Larvae were then returned to the container from which they originated.
To estimate the accumulation of organic detritus and bacterial growth in unmanaged containers, a thoroughly mixed water sample was measured for cloudiness using a turbidity tube [52] and dissolved oxygen content using an Ecological Test Kit (Rickly Hydrological Company, Columbus, OH). Water samples were returned to containers after testing. In all containers, tap water was added every few days to replace water lost to evaporation. Any pupae were removed to prevent emergence of adult Ae. aegypti.
We monitored oviposition in 591 containers in 448 households across Iquitos. Ae. aegypti eggs were deposited in 51.8% of surveyed containers (306 of 591). Egg counts per container per day were strongly skewed, with the majority of containers receiving 0 to 50 eggs (median = 2, mean = 41), and a few containers receiving hundreds of eggs (Figure 1). All mosquitoes reared from collected eggs were Ae. aegypti, which we found to be the only Aedes species present in domestic containers throughout Iquitos. The presence of Ae. aegypti larvae in households was independent from whether or not adult females were caught during entomological surveys (χ2 = 1.897, df = 1, p = 0.169). Culex mosquitoes were occasionally present in the same containers (5.2% of all containers surveyed, 11.3% of Ae. aegypti-positive containers), but were easily distinguished by morphology. We did not find any containers colonized only by Culex.
After controlling for collection period, three variables were significant predictors of whether females laid eggs in containers: Ae. aegypti larvae, exposure to sunlight (≥20% of day), and absence of a container lid (Table 2). The probability of oviposition increased when sites held conspecific larvae (β = 1.658; 95% CI = [1.286, 2.030]; p<0.001), an effect which remained consistent regardless of larval abundance or whether larvae had been removed from containers during the day(s) prior to egg collection. Containers located outside and exposed to sunlight (≥20% of the day) were more likely to receive eggs compared to indoor containers (β = 0.601; 95% CI = [0.114, 1.089]; p = 0.016) and shaded outdoor containers (sunlight<20% of the day) (β = 0.538; 95% CI = [0.124, 0.952]; p = 0.011). No difference was detected between shaded outdoor containers and indoor containers (β = 0.063; 95% CI = [−0.413, 0.540]; p = 0.795). Oviposition decreased when containers were covered with lids (β = −0.706; 95% CI = [−1.430, 0.017]; p = 0.056).
Among containers receiving eggs, the number of eggs laid was affected by larval abundance, whether larvae were removed prior to oviposition, pupae, fill method, circumference, and (circumference)2 (Table 3). Females laid more eggs when over 50 conspecific larvae were present in containers (β = 0.759; 95% CI = [0.483, 1.035]; p<0.001). Among sites from which larvae were removed prior to egg collection, however, a significant increase in egg abundance was observed only when more than 100 conspecific larvae had been present (β = 0.838; 95% CI = [0.126, 1.549]; p = 0.021). More eggs were laid in containers that held Ae. aegypti pupae, regardless of whether they had been removed (β = 0.448; 95% CI = [0.141, 0.754]; p = 0.004). Among the three fill methods, unmanaged containers received a larger number of eggs than rain and manually filled containers (β = 0.387; 95% CI = [0.092, 0.681]; p = 0.010); there was no difference between rain and manual filling (β = 0.073; 95% CI = [−0.241, 0.387]; p = 0.647). Container circumference had a positive effect on egg abundance (β = 0.011; 95% CI = [0.005, 0.017]; p<0.001), whereas the impact of (circumference)2 was negative (β = −0.00002; 95% CI = [−0.00003, −0.000004]; p = 0.013). When the regression equation was plotted, egg abundance increased with container size initially, but eventually leveled off as containers approached 270 cm in circumference (Figure 2). No significant interactions were identified between predictor variables in either regression model.
Third instar larvae were collected for starvation bioassays from 113 containers. For the majority of containers, median larval RS was between 5 to 15 days (range 0 to 28 days) (Figure 3). There were no significant correlations between median RS and the mean density of eggs laid per day (all other larvae retained, n = 59 containers, Spearman's ρ = 0.15; all other larvae removed, n = 54 containers, Spearman's ρ = 0.0008). No correlations were evident when the data were also stratified by larval abundance or container capacity (data not shown).
Ambient temperature and relative humidity were measured for the first four weeks and were consistent among the three study locations (field laboratory, house 1, and house 2) (Table 4). Water temperature (Table 4) was recorded for eight weeks and found to be similar for the field laboratory and house 1. Due to logger malfunction, water temperature was not recorded at house 2. Because Iquitos climate was relatively consistent during June to August 2009 (Table S1), we expect the data recorded at each location to be indicative of the entire study period.
Conspecific larvae were present in treatment A containers and absent from treatment B and C containers throughout the experiment. The number of Ae. aegypti eggs laid in each container per week was influenced by container treatment (ANOVA F = 77.70; df = 2, 4; p<0.001) and week (ANOVA F = 6.47; df = 11, 88; p<0.001), but not by house (ANOVA F = 4.45; df = 2, 4; p = 0.096). Females laid the most eggs in unmanaged containers with larvae (A) and the fewest in containers with clean water and no larvae (C) (Figure 4). The number of eggs laid fluctuated over time in all container treatments, peaking in weeks 4 and 5, and again in week 11.
Water in unmanaged containers (A and B) increased in turbidity (ANOVA F = 41.55; df = 6, 30; p<0.001) (Figure 5a) and decreased in dissolved oxygen content over time (ANOVA F = 10.19; df = 6, 30; p<0.001) (Figure 5b), signs of rising levels of organic detritus and microbial growth. Water turbidity and dissolved oxygen content were not influenced, however, by the presence of larvae (treatment A vs. B) (turbidity: ANOVA F = 3.16; df = 1, 2; p = 0.217; oxygen: ANOVA F = 0.19; df = 1, 2; p = 0.704) or location (turbidity: ANOVA F = 5.12; df = 2, 2; p = 0.163; oxygen: ANOVA F = 4.65; df = 2, 2; p = 0.177). Although water assays did not quantify large solid detritus such as leaves, unmanaged containers in each location received similar amounts of detritus due to their proximity to one another. Taken together, our experimental results indicate that food levels were similar among treatment A and B containers, and that differences in oviposition among the two were attributable to the presence of larvae.
In nature, Ae. aegypti egg distribution among containers was influenced by a combination of factors, including the presence of conspecific larvae and pupae, container fill method, sun exposure, container size, and the presence of a lid. Although the negative effect of container lid was likely due to presence of a physical barrier [54], consistent patterns with respect to the remaining variables suggest that gravid Ae. aegypti females actively choose among potential oviposition sites. Female Ae. aegypti responded most strongly to the presence of conspecific immatures, both in terms of the probability of oviposition and the number of eggs laid. This correlation was not due to more frequent presence of adult females in houses with colonized containers. In our study, the presence of colonized containers was not associated with the capture of adult females during entomological surveys. Furthermore, Getis et al. observed a cohort effect among Ae. aegypti in Iquitos; infested containers typically held a single cohort of Ae. aegypti developing in synchrony, rather than multiple overlapping cohorts [39]. Thus, successive life stages were spatially correlated, but there was no correlation between larval and adult abundance at the household level. After adjusting for conspecific immatures, we did not observe an effect of Culex larvae or pupae on Ae. aegypti oviposition in our multivariate models.
For Ae. aegypti, attraction of gravid females to containers with immature conspecifics may seem at first counter-productive. Field populations are thought to be limited foremost by density-dependent competition for food during the early larval stages [9], [14], [55]. In addition, studies have documented that high larval densities negatively impact several components of mosquito fitness, including larval survivorship [56]–[58], development rate [55], [59], adult lifespan [60], adult size [59], [61], and female fecundity [62], [63]. From this standpoint, it would seem advantageous for ovipositing females to avoid conspecifics as competitors to their own progeny. Interestingly, conspecific attraction has been observed across numerous animal taxa (e.g., reviewed in [64]–[66]), such as birds, mammals, reptiles, fish, and insects, including other mosquitoes [22], [67], [68]. The drawbacks of increased competition may be counter-balanced by the benefit of using conspecifics as a reliable cue of habitat quality [24], [69]. Conspecific attraction has been described as a means for females to exploit information collected by others. Rather than gathering information on a multitude of environmental factors potentially affecting offspring growth, a process constrained by energy, time, and/or sensory capabilities, females may be able to quickly assess habitat suitability by observing the reproductive success of previous females [70]. In the case of Ae. aegypti, we speculate that conspecific larvae and pupae may serve as signals that a site experiences infrequent water turn-over and desiccation, and contains adequate food, two conditions necessary for successful larval development.
Due to an inherent trade-off between gaining information on habitat suitability and increasing competition for offspring, we expected conspecific attraction to be tempered by aversion to containers with high larval densities. Laboratory assays have demonstrated a dose-specific oviposition response that increased with conspecific densities up to ∼1 larva/mL and decreased thereafter [28], [71]. In our study, conspecific larvae were always attractive, perhaps because larval densities in Iquitos were far lower (average = 0.03 larvae/mL, SE = 0.006) than the densities found to repel females in laboratory experiments. Only 1.2% of Ae. aegypti-colonized containers had densities greater than 1 larva/mL. We suspect that few containers in Iquitos ever reach repellent densities.
We observed that free-ranging Ae. aegypti laid more eggs in sites that had recently held conspecifics compared to those that had not, suggesting that conspecific attraction is mediated by chemical cues. The preference for conspecific-conditioned water has been noted in the laboratory [72] and attributed to semiochemicals produced by larval-associated bacteria [71]. Semiochemicals may act as attractants to help females locate cryptic sites, and/or as stimulants to promote egg-laying [30]. Some laboratory studies have revealed preference of ovipositing Ae. aegypti for sites containing conspecific eggs [27], [31], leading to the discovery of oviposition-inducing egg semiochemicals [29]. Because our survey required daily collection of eggs, we were unable to investigate in the field the effect of conspecific eggs on Ae. aegypti oviposition site selection. Interestingly, when investigators separated the components of these semiochemicals, some components elicited attractive/stimulating responses, whereas others produced repellent/deterrent responses. Depending on their concentration, attractive chemicals can also become repellent [29], [30]. If applied properly, chemical mediators of oviposition behavior have potential to be useful for Ae. aegypti control.
Container characteristics such as fill method, sun exposure, and size played a secondary role in oviposition choice. During our observational and experimental studies, more eggs were laid in unmanaged containers and few eggs were laid in manually filled containers. Because unmanaged containers collect the most organic detritus and manually filled containers are kept cleanest, this pattern is consistent with females selecting oviposition sites based on the availability of larval food. If females act primarily to maximize food for their offspring, we would expect the number of eggs laid per container to increase proportionate to food availability. From our starvation assays, however, we were unable to demonstrate any correlation between the median larval survival time, an indirect measure of food availability, and the number of eggs laid per container. Although this result implies that female Ae. aegypti did not oviposit to maximize food for their progeny, several limitations of our study could have affected our ability to test this relationship. First, starvation bioassays were conducted on larvae already present in containers at the start of surveys, and thus provided information on container food content over the past few days or weeks, rather than at the moment of oviposition. Because our study design necessitated collecting eggs to quantify oviposition, measuring starvation times of pre-existing third instars was the best alternative. Second, the third instar larvae we collected likely hatched at different time points and results from their starvation bioassays could be confounded by differences in age and time they had to feed.
We also observed more Ae. aegypti eggs deposited in containers exposed to sunlight (≥20% of the day). Larval development is highly temperature-dependent [73], [74]. A recent biophysical model of Ae. aegypti development in Australia predicted that, when containers are not prone to desiccation, sun-exposed containers reach warmer temperatures and support more generations of Ae. aegypti than shaded containers [75]. Females may have a selective advantage if they are able to detect containers with warmer water where their offspring develop faster. This, however, appears to contradict data from Puerto Rico by Barrera et al. [76], who found that immature Ae. aegypti were more abundant in shaded containers with low water temperature (≤29°C), indicating that females oviposited more frequently in containers shielded from full sunlight. Due to environmental differences between Iquitos and Puerto Rico, our criteria for shaded vs. exposed may have varied from those used by Barrera et al. [76]. Outdoor containers in Puerto Rico commonly receive sun exposure >50% of the day (ACM, unpublished data), in contrast to Iquitos, where abundant tree coverage limits sun exposure to only 10–40% of the day for most outdoor containers. We cannot directly compare our data to that of Barrera et al. [76] because metrics were not provided for container categories of “full sun,” “partial sun,” or “shaded.” We were not able to measure water temperature in each surveyed container. Maximum daily water temperatures from our experimental containers were typically 27–28°C, suggesting that water temperatures are lower in Iquitos compared to Puerto Rico.
Attraction to large oviposition sites has been demonstrated in Ae. aegypti [36] as well as other mosquito species [67], possibly because large sites collect more food or are resistant to desiccation. We found that the number of Ae. aegypti eggs laid increased with container circumference up to a threshold around 270 cm, after which oviposition leveled off, indicating that perhaps the relative advantage of large container size diminishes as containers become bigger. Due to the low occurrence in Iquitos of containers exceeding 270 cm in circumference (n = 26 of 591 containers, 4% of surveyed containers), we could not assess the relationship further between increasing container size and oviposition.
A major limitation of our study design was the inability to examine effects of container material and/or texture on oviposition. Container texture affects Ae. aegypti oviposition, with females preferring to lay their eggs on rough surfaces [34], [37]. Because we lined containers with strips of paper towel to transport eggs back to the field laboratory, we artificially made container surfaces homogeneous. In a previous Iquitos field study, we showed that females laid more eggs in cement containers compared to plastic or metal containers when all were unlined and similar in size [45]. Additional experimental studies should be conducted to investigate the importance of container material to oviposition site choice when conspecific presence and abundance, fill method, sun exposure, and container size are varied.
Ae. aegypti oviposition site choice appears to be flexible, potentially reflecting a mix of site selection strategies across the population. A small portion of females may act as “founders” (e.g., [77]), choosing non-colonized sites based on environmental indicators of quality, whereas the majority of females respond predominantly to conspecific cues. Alternatively, each female may partition her egg batch so that most eggs are laid in colonized containers, when colonized containers are available, and a smaller fraction elsewhere. It should be noted that these scenarios are not mutually exclusive; for any female, the decision to reject or accept a particular site may change with time. For example, results from studies on herbivorous insects demonstrated that ovipositing females typically become more accepting of low-ranking sites as search time progresses (reviewed in [78]). Recent theoretical work on animal decision rules suggests that when individuals are limited by time, number of options, and accuracy with which they can assess site quality, decisions should be based on the best-of-n rule [79]. If female Ae. aegypti use this rule, they are likely to assess a fixed number of sites (n) and choose the perceived best among them, rather than searching longer for a site that meets specific criteria. Such a rule could explain the oviposition patterns we observed in Iquitos; colonized containers tend to be utilized when found, but other site characteristics (size, sunlight, and organic detritus) are used to judge site quality if the n sites do not include a colonized container. This remains to be confirmed in the field. Decision rules used by Ae. aegypti to select oviposition sites merit further investigation.
Female choice of oviposition site may have greater impact on Ae. aegypti population dynamics than previously thought. We propose that, due to strong conspecific attraction, oviposition site selection could lead to dense aggregations of larvae and actually contribute to density-dependent regulation. This phenomenon may explain why larvae in the field frequently develop under food-limiting conditions [9], [38], [80]. It is likely that while some colonized sites become crowded, other suitable larval development sites remain empty. A companion study in Iquitos indicated comparable survival and development rates when larvae were reared in water collected from colonized vs. non-colonized containers in the field, suggesting no difference in food content (STS, unpublished data). These results imply that availability of larval food is not the primary determinant of oviposition choices and agree well with our larval starvation data presented herein. A similar study in Trinidad, West Indies, revealed no difference in nutrient levels between water-storage drums colonized or not by Ae. aegypti [81].
Our results have direct implications for strategies to control Ae. aegypti. Targeting containers that produce the most Ae. aegypti adults for removal or larvicide treatment will reduce mosquito populations in the short term. Sustained population suppression, however, will be difficult to achieve by these means. Elimination of highly productive containers (or the immature Ae. aegypti within) will likely shift new eggs to alternative suitable containers. If immature conspecifics are no longer available as a strong oviposition cue, females that would have concentrated their eggs in those highly productive sites may instead oviposit among suitable, previously unoccupied containers based on food availability and/or sun exposure. Strategies that kill mosquitoes late in their development (i.e., insect growth regulators (IGRs) that kill pupae [82], [83] rather than larvae) will enhance vector control by creating “egg sinks,” treated containers that exploit conspecific attraction of ovipositing females, but reduce emergence of adult mosquitoes via density-dependent larval competition and late acting insecticide. For an egg sink strategy, it would be best to employ IGRs that have no repellent effects on ovipositing females, such as pyriproxyfen [84] or methoprene [85]. Pyriproxyfen is of particular interest because adult females are able to transfer the IGR to other oviposition sites [84], [86]. Thus, pyriproxyfen-treated containers could potentially serve as both egg sinks and sources for insecticide dissemination. The success of this approach would depend on oviposition patterns of individual females.
Alternatively, rather than relying on conspecific larvae, control tools could be designed to capitalize on the attractant or stimulant properties of semiochemicals influencing Ae. aegypti oviposition responses in the field. Bacteria-derived oviposition attractants could be used to lure females to lethal ovitraps or stimulants could be used to increase their exposure to insecticide-impregnated substrates [30]. The fact that wild Ae. aegypti are quite selective when choosing oviposition sites may be the basis for development of new strategies and products for control of dengue virus vectors.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.