Ethics statement

All experimental work was performed according to protocols approved by the Thomas Jefferson University Institutional Animal Care and Use committee.

Animal procedures

All protocols were approved by the Thomas Jefferson University (TJU) Institutional Animal Care and Use Committee. Study subjects included male rats from the Spontaneously Hypertensive Rat (SHR/NHsd) and Wistar Kyoto (WKY/NHsd) strains, corresponding to autonomic dysfunction and control phenotypes, respectively. Rats were purchased from Harlan Laboratories and experimental procedures were carried out one week following animal arrival at our facility. All animals were housed socially in the TJU animal facility. The facilities were maintained at constant temperature and humidity with 12/12 hour light cycles (lights on at Zeitgeber time = 0). We harvested organ tissues at five time points: 4, 6, 8, 12, and 16 weeks of age. Rats were humanely sacrificed via rapid decapitation. CNS tissue was excised and the brainstem was isolated in ice-cold artificial cerebral spinal fluid (10mM HEPES; 140mM NaCl; 5mM KCl; 1mM MgCl₂; 1mM CaCl₂; 24mM D-glucose; pH = 7.4). We simultaneously harvested the adrenal gland, kidney, liver, and left ventricle of the heart. Tissue samples were flash frozen and stored at -80°C. Our original study was designed to include 50 animals (2 genotypes, 5 time points, 5 replicates). One animal deceased prior to the designated time point for organ harvest. Thirty-five animals were included in our study and 2–5 organ samples per strain were obtained at each time point for organs other than the brainstem (S1A Fig); 12 week brainstem tissues (from n = 10 animals) were not included in the present study as these samples were utilized for a parallel study that precluded the gene expression analysis employed here. Five other animals were excluded from our study prior to performing qPCR analysis because either the respective RNA did not pass our quality criteria (see below) or because the tissue was used for other purposes. S1A Fig shows the tissue samples included in our study for each animal.

Molecular biology

Total RNA was extracted from 10–50 mg tissue samples using the Direct-Zol RNA extraction kit, which captures all RNA greater than 18 nucleotides in length (ZYMO Research, Irvine, CA). Samples were DNAse treated and stored at -80°C. Concentration and integrity were assessed with a spectrophotometer (ND-1000 from NanoDrop, Philadelphia, PA). RNA samples with 260/280 (nm/nm) ratio <1.8 and 260/230 ratio 1.8–2.0 were purified with RNA Clean and Concentrator-100 (ZYMO Research, Irvine, CA). High-throughput PCR was implemented as described previously [39, 40]. Intron-spanning PCR primers were designed for 24 assays (see Table 1 in S1 Text). For each sample, 30 ng of total RNA was used. The standard BioMark protocol (Fluidigm, South San Francisco, CA) was employed to reverse transcribe and pre-amplify cDNA samples for 12 cycles using TaqMan PreAmp Master Mix based on the manufacturer’s protocol (Applied Biosystems, Foster City, CA). The qPCR reactions were performed using a 192.24 BioMark Rx Dynamic Array for multiplex gene expression measurement (Fluidigm, South San Francisco, CA). Quantitative PCRs were implemented with 30 cycles (95°C for 15s, 70°C for 5s, 60°C for 60s).

Data processing

We quantified qPCR products by determining threshold cycle (Ct) values. We designed our study with Actb and Gapdh assays as potential reference genes for normalization. To determine whether these assays were appropriate for normalization, we assessed the stability of these genes across samples, as in our previous studies [39], based on a well established method for evaluation of putative reference genes [41]. Our analysis revealed that Actb and Gapdh expression profiles were highly variable across samples, as has been shown previously. No single gene showed consistent stability across all samples. However, the median Ct across all genes was stable across samples, indicating superior utility of the median Ct as a ‘pseudo reference gene’ for data normalization (S1B Fig). Hence, we normalized the raw Ct data based on median expression levels, which were considered to represent reference gene expression levels. For each sample (s) obtained from a specific organ (r) at a specific time point (t), we subtracted the median Ct computed across all genes (g) in that organ: where med(.) is the median of the argument. We next centered the data for comparison across genes based on the median expression level across all samples for each gene: . We used −ΔΔCt values for analyses of gene expression. We omitted Actb and Gapdh from all subsequent analysis due to ambiguity in the functional interpretation of the results. Missing data based on our QC analysis were rare (median = 1.8%, sd = 10% of samples per gene with NA values; median = 4.2%, sd = 9.1% of genes per sample with NA values). Thus, missing data were imputed according to established approaches [6, 42] by replacing missing values with the mean across 10 samples with most similar expression profiles according to Euclidean distance using the impute package in R [43]. Note that we did not impute Brainstem data at age = 12 weeks because we did not obtain the gene expression data from the brainstem samples at this time point. Both raw Ct and normalized data are available (S1 and S2 Files). To examine whether specific samples imparted systematic biases in our results, we implemented Principal Components Analysis in R using the princomp function.

Timeseries analysis

To test whether the temporal dynamics of gene expression differed between autonomic dysfunction and control phenotypes, we applied the Optimal Discovery Procedure (ODP) using the EDGE package in R [44]. Temporal profiles were modeled as natural cubic splines which connect a series of smooth polynomials between knots defined by the degrees of freedom for the spline fit [11, 45] (function ns with df = 3 in the splines package for R [46]). The ODP analysis involved a comparison of a null model to an alternative model. The null model was characterized by a single spline fit to the aggregated autonomic dysfunction and control time-series data for each gene. The alternative model consisted of two splines fit to the respective phenotypes. For each gene, errors between data points and fitted values were summed and squared for the null model (SS⁰) and the alternative model (SS^A). An analogue of the conventional F statistic was computed to evaluate the goodness of fit obtained for the null versus alternative model: F = (SS⁰ − SS^A)/SS^A. The estimated distribution for this statistic was utilized to compute an estimate of the probability of the alternative model under the null hypothesis (p value) with correction for multiple testing (q value). Complete details can be found in [11, 44].

Dynamic network and continuous-time system identification from discrete time-series data

We scaled the data to the range (0, 1) prior to implementing the HMF method. To implement this scaling for expression profile E, we applied the following transformation: E_scaled = (E − min(E))/(max(E) − min(E)). We use E to represent E_scaled in the context of our system identification studies. The following description details the basic theory and procedure underlying the system identification approach using Hartley Modulating Functions (HMF). A signature attribute of this approach is that interaction coefficients can be estimated that jointly describe network dynamics and structure. From a mathematical perspective, a principal advantage of the HMF-based system identification approach is that it obviates the need to compute temporal derivatives of the raw data [47, 48]. Instead, the interaction coefficients k (see Fig 1D) are determined by estimating inner products between both the expression data (and the derivatives thereof) and a set of basis functions—the carefully chosen Hartley modulating functions—and approximating these inner products using the Hartley transform to transform the data into the frequency domain. This procedure facilitates the accurate and robust identification of continuous-time models from discretely sampled data, another principal advantage of the HMF method, as we have demonstrated previously [48].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Conceptual, experimental, and analytic framework for examining multi-organ pathogenesis of autonomic function.

(A, B) The etiology of autonomic dysfunction involves (A) multiple organs in which positive feedback processes involve (B) inflammatory mediators, renin-angiotensin signaling, and sympathetic activity. (C) A number of well studied genes underlie the molecular basis of maladaptive network feedback processes. (D-F) Analysis pipeline for examination of expression patterns of the genes listed in (C) in multiple organs listed in (A). (D) Gene expression measurements were obtained and the data were analyzed using time series statistics and novel network identification approaches. (E) The network identification analysis reconstructed gene expression dynamics and gene regulatory interactions underlying autonomic dysfunction and control phenotypes. (F) Bioinformatic analyses were integrated with analyses of network structure and dynamics to generate novel hypotheses regarding the molecular mechanisms underlying autonomic dysfunction.

https://doi.org/10.1371/journal.pcbi.1005627.g001

Furthermore, our approach entailed the use of powerful regularization techniques that mitigate against overfitting the interaction coefficients [45]. Whereas frequency domain transformations of data have been previously been employed to implement systems identification, these approaches relied on optimization-based estimates of the interaction coefficients [49]. In contrast, using HMF method, we directly estimated the interaction coefficients via regularized regression. Thus, our approach, in principle, can overcome difficulties in parameter estimation that result from non-convex solution spaces characterized by local minima [50].

The remainder of this section starts with a description of the mathematical underpinnings and implementation details underlying our use of the HMF method to identify multi-organ gene regulatory networks. Following this description, we detail further analyses that demonstrate the robustness of our approach.

The expression level E of gene g in organ r at time t was modeled as follows for a data set with samples obtained between time t = 0 and time t = T: where N_r is the number of organs and N_g is the number of genes. The degradation coefficient for E_rg is referred to as γ_rg. For simplicity, we avoided using the degradation term explicitly such that degradation was implicitly incorporated in k_ij for i = r and j = g: (1) The full network can be compactly expressed in matrix form as with the following simplified representation: (2) Note that that parameter matrix K from Eq (2) is equivalent to the Jacobian matrix corresponding to this linear system: J = [J_ij] where J_ij = ∂f_i/∂E_j, f_i = dE_i/dt, and (i, j) each refer to a particular gene-organ combination. Thus, this matrix gives the influence of a gene in column j on a gene in row i.

We estimated the interaction coefficients k by applying the HMF method [47, 48]. This method entails the multiplication of Eq (1) by M different modulation functions ϕ_m (m = 1, 2,…, M). Integrating these products gives the following relation: (3) where ϕ_m = f(t) is chosen such that = 0 for t = 0 and t = T. Note that the times t = 0 and t = T corresponding to sampled ages of 4 and 16 weeks, respectively, such that T = 12 in our computations. The modulating functions ϕ are chosen as follows: where n is the order of the highest derivative of the system described by Eq (1) (i.e., n = 1), cas(x) = sin(x) + cos(x), and . The integrals on the right and left hand sides of Eq (3) can be estimated using the Hartley transform [51], the m-th HMF spectral component of gene expression profile E(t), and the HMF spectra for the i-th derivative of E(t) [47]. These computations are defined respectively as follows: (4) (5) (6) where Importantly, given a solution to Eq (4), numerical solutions to Eqs (5 and 6) can be obtained. The numerical solutions to Eqs (5 and 6) can be used to compute the interaction coefficients based on the following relations: (7) (8) Then Eq (2) can be written as follows (9) where the only unknowns are the interaction coefficients, which can be determined by linear regression. However, Eq (4) must be computed first. Following our previous work [48], we computed Eq (4) by linearly interpolating between average gene expression values at adjacent time points and analytically evaluating the integral: Then we computed the values in Eqs (5 and 6) given n = i = 1 [48]. Note that the analytical solution to the integral in Eq (4) evaluates to zero for n + m − j = 0 (m = −1, j = 0 and m = 0, j = 1). However, allowing Eq (4) to be zero resulted in poor fits of the model to the data. For these cases, we arbitrarily set n + m − j = ϵ with ϵ = 10⁻⁶. This choice of ϵ resulted in robust model fits as detailed below.

The set of interaction coefficients can be obtained for interactions regulating the expression dynamics of each organ/gene combination by solving Eq (9) using a range of m values: In matrix notation, this relation can be expressed as (10) where y is a column vector of terms in which each row entry is computed with a different m value, X is a matrix of terms with a row for each m value and a column for each term corresponding to a particular organ/gene combination, and β is a vector of interaction coefficients with the same length as the number of columns in X. Solving the linear system for β provides interaction coefficients that determine the dynamic regulation of a gene in a given organ, based on the expression profiles of all genes in all organs. However, we wanted to perform system identification for gene regulatory networks spanning multiple organs and genes. To globally estimate the entire set of interaction coefficients using the HMF method, a matrix of the form Eq (11) was utilized. The full matrix was established using instances of Eq (10) for each organ/gene combination considered in the system: (11) where 0 represents a matrix of zeros with the same dimensionality as X. In this formulation, y₁₁ = X₁₁ β⁽¹¹⁾, y₁₂ = X₁₂ β⁽¹²⁾, and y_{Nr Ng} = X_{Nr Ng} β^{(N_r N_g)}. Thus, the solution to Eq (11) provides a global fit to the interaction network including all assayed genes in all sampled organs. The regression problem can be described compactly as y = X β and solved using linear regression. To circumvent overfitting of the model, we applied well established regularization techniques in which the regression coefficients were determined by solving an optimization problem with the following objective function: (12) where ||x||₁ = ∑|x| and . The regularization parameters α ∈ [0, 1] and λ ∈ [λ_min, λ_max] (see below) impose sparsity on the network by forcing the interaction coefficients towards zero [45]. This form of regularization is known as the ‘elastic net’ [52], where the ||β||₁ term represents the ‘lasso’ penalty [53] and ||β||₂ term represents the ‘ridge’ penalty [54]. Thus, α weights the lasso penalty and (1−α) weights the ridge penalty. The elastic net formalism exhibits positive attributes of both regularization techniques with respect to enhancement of network interpretation, based on sparsity of connectivity, and augmentation of prediction accuracy [45].

We performed network identification as described above using a range of m value sets. The regression problem represented by Eq (11) requires that the number of m values exceeds the number of interaction coefficients, N_r ⋅ N_g. Each m value set had the form m = 0, ±1, ±2, …, ±(M − 1), ±M, where M was varied between and [48]. Initial simulations showed that the regression results were not sensitive to the range of the m values, given M ≥ M_min. We selected ten sets of m values that were evenly spaced within the M range. For each of the ten m value sets, we performed the regression analysis for α = 0, 0.2, 0.4, 0.6, 0.8, 1, and we used 10 λ values for each α. The λ values were also varied over a range. The λ range was bounded by λ_max, the minimal λ value associated with a particular α that resulted in all zero coefficients. That is, for λ = λ_max there was no network connectivity. The λ values were incrementally varied from λ_min = λ_max × 10⁻⁴ to λ_max on a logarithmic scale according to the default functionality of the glmnet package used for elastic net regression in R [46, 55].

Optimal solutions to the regularized regression problems (11 and 12) were determined based on simulation results. We simulated the identified network models (2) using MATLAB’s ode45 and ode15s functions, or with R using the lsoda function from the deSolve package [56]. Differences in simulation results, with respect to the choice of numerical integrator, were not visually detectable. To select the optimal fits, we initially considered the sum of squared residuals as an objective function for minimization, , where is the simulation result and is the average for the corresponding gene expression data set. However, according to this objective, the best fits could be those with all zero coefficients. Hence, we revised our objective to penalize fits with low variability: (13) and set ϵ = 10⁻¹⁰ (the choice of ϵ did not make a detectable difference in the selection of the best fit).

Molecular networks underlying the physiology of blood pressure control have been shown to be distinct for autonomic dysfunction versus control [40]. For all network reconstructions, we considered autonomic dysfunction and control phenotypes separately. We initially performed our system identification analysis of a network with all organs (N_r = 5) and genes (N_g = 22) with (N_r ⋅ N_g)² = 12,100 possible connections. We implemented 600 iterations of the system identification algorithm with distinct combinations of the m value range (n = 10), λ value (n = 10), and α value (n = 6). In general, for both phenotypes, small λ values were associated for good fits (i.e., J_sim ∼ min(J_sim)) irrespective of the α value for α > 0 (S2 Fig). For the control, J_sim = min(J_sim) for α = 0.2 at λ = 6.5⋅10⁻⁵ (221 m values). For autonomic dysfunction, J_sim = min(J_sim) for α = 0.4 at λ = 8.3⋅10⁻⁷ (111 m values). Based on these pilot studies, further analyses were implemented with ten λ values, ten sets of m values, and α = 0.2.

System identification robustness analysis

To evaluate the robustness of the HMF Method, we compared the ‘best fit’ network with multiple subnetworks characterized by log(J_sim) < 10 (see S2 Fig). For these comparisons, we considered interaction coefficients that were larger in absolute magnitude than two standard deviations from the median (median ∼ 0 in all cases). Further, we only considered the coefficients for which in each phenotype-specific best fit network and comparison network, and analyses were completed for coefficients that met this criterion for both the best fit network and the comparison network. Spearman rank correlation coefficients were determined and we investigated whether the coefficient sign was sensitive to the regularization parameters using the Fisher's exact test (FET). For the FET analysis, we computed the sign (i.e., +1 or −1) of the coefficients considered in the correlation analysis. The contingency table illustrated in S3 Fig was formulated for the FET. Both Spearman rank and FET p-values were corrected for multiple testing according to the Benjamini-Hochberg method using the qvalue package [57]. We also computed the odds ratio, based on the same data, to quantify the degree of agreement between networks (S3 Fig).

To further evaluate robustness, we examined graph theoretic metrics. A path through a network consists of the sequence of edges between two nodes or vertices. The shortest path length refers to the minimal number of edges connecting two nodes, and the average path length < ℓ > is the mean of all shortest paths between every pair of nodes. This measure is an index of the efficiency with which the network can be navigated. The clustering coefficient C_i generally quantifies the connectivity among all nodes connected to node i. The number of nodes with links to node i is n_i and the number of links amongst the n_i nodes is n_c: C_i = 2n_c/(n_i ⋅ (n_i−1)) [58]. We computed a variant of this measure, the global clustering coefficient, which is the number of closed node triplets (i.e., ‘triangles’) divided by the total number of connected triplets. The global clustering coefficient is also known as the ‘transitivity’ metric [59]. Finally, we assessed the distribution of degrees k_c; that is, the distribution for the number of edges connecting a node to its neighbors, which follows a power law of the following form for many biological networks: . This distribution gives the probability that a given node has k_c connections [58, 60]. To characterize the degree distribution, we fit a power law to the vector of degree frequencies and utilized a Kolmogorov-Smirnov (K-S) test of the null hypothesis that the data were distributed according to the power law. The fit returned an estimate of γ and a p-value indicating the probability of the test statistic given the null hypothesis. Thus, high p-values indicate the absence of evidence in support of the conclusion that the graph under consideration does not exhibit a power law degree distribution. Network analyses were implemented using the igraph package for R [59].

Dynamic pattern analysis

We assessed the patterns of gene expression dynamics across and within organs by temporally ordering the expression profiles identified using the HMF method. We used temporal ordering schemes based on peak timing and valley timing, which were established as follows. The dynamic profiles E_i of each gene i were scaled to E^sc ∈ [0, 1]. We then determined all local extrema for the scaled waveform E^sc. We refer to local maxima as peaks and local minima as valleys, with associated times denoted as t_p and t_v. Putative peaks and valleys ( and ) were considered as veritable extrema if their expression levels relative to the initial time point (i.e., t₀ = 4 weeks) exceeded a given threshold (E^th): For profiles with both peaks and valleys, we determined the first peak or valley that satisfied this condition. We also classified monotonically decaying or increasing profiles as profiles without peak or valleys as defined above. Further, monotonicity required the additional condition that for all x = (p, v). Monotonic increases versus decays were distinguished based on estimates of the profile’s first time derivative; t_p = t₀ and t_v = t₀ for monotonically decaying and increasing profiles, respectively. For visualization, scaled profiles were sorted and plotted according to peak or valley time. For this analysis, we set E^th = 0.5 and E^th0 = 0.1.

Differential network analysis

We evaluated the differences in the multi-organ gene-gene interaction networks between autonomic dysfunction versus control phenotypes. We describe our analyses based on conventional graph theoretic terminology, according to which the network is considered to be a graph G with vertices or nodes V that refer to genes which are connected by edges E: G = (V, E). The edges E are characterized by the interaction coefficients k described above. In particular, we addressed whether edges were added, removed, or switched in sign from autonomic dysfunction to control. We focused on the strongest network connections by considering edges for which the following condition was met: where E_ij is an edge from node V_i to node V_j, med(.) is the median, and sd(.) is the standard deviation. Median values computed across all edges were exactly zero. Edges that did not meet our criteria were set to zero. All edge values were scaled to the interval (−1, 1) by applying E^sc = E/max(|E|). We evaluated differential network properties as follows. For each edge in G, we first computed the difference between unscaled edge values for control (WKY)- versus autonomic dysfunction (SHR)-specific networks as and we defined a threshold edge difference as follows: Then we evaluated each edge for a set of conditions. Edges were considered to be added to the autonomic dysfunction network (absent for control but present for the autonomic dysfunction phenotype) if the following conditions were satisfied: Note that the third condition ensured that an edge could not be considered to be added to the SHR network if that edge was slightly larger than E^th for SHR but slightly smaller than E^th for WKY, such that the actual difference was negligible. Hence, only edge differences that exceeded E^th were considered. Edges were considered to be removed from the SHR network (absent for SHR but present for WKY) if the following conditions were satisfied: Finally, edges were considered to be switched in sign from WKY to SHR (e.g., E_ij = −1 for WKY and E_ij = 1 for SHR) if the following conditions were satisfied: We further filtered the data by analyzing count histograms of ΔE_ij values for added, removed, and switched edges and applying the following respective cutoffs to exclude edges with the smallest differences between SHR and WKY: 0.15, 0.15, and 0.2 (See S4 Fig). Networks characterized by edges that met the aforementioned criteria were visualized using Cytoscape [61].

Bioinformatic analyses

We completed a number of analyses in which we utilized publicly available data and software tools, as described in detail below.

Network motif analysis: We used the mfinder tool to search for specific network motifs associated with cellular adaptation [62, 63]. We visualized the detected subnetworks using Cytoscape [61].

Genetic analysis: The genomes of numerous rat strains have been sequenced, including the strains used in our study: the Wistar Kyoto (WKY/NHsd) and the Spontaneously Hypertensive Rat (SHR/NHsd) [64]. For these strains, single nucleotide variants (SNVs) have been identified relative to the Brown Norway (BN) founder rat strain (genome build RGSC3.4) [64]. We downloaded the SNV data set in the VCF format from the Rat Genome Database during December 2015 (ftp://ftp.rgd.mcw.edu/) [65]. We isolated SNVs that were found in SHR/NHsd but not in WKY/NHsd. To test whether SHR-specific SNVs were present in the coding regions of the genes analyzed in this study, we downloaded the genomic coordinates for exons of interest from the UCSC genome table browser (rn4 rat genome, https://genome.ucsc.edu/) [66] and evaluated the presence of SNVs in these regions using BEDTools [67]. Similarly, to examine gene-proximal regulatory regions, we used the UCSC genome table browser to obtain the genomic coordinates of regions 2kb upstream of the transcription start sites (TSSs) [68, 69]. We identified SNVs in these regions using BEDTools. We further examined whether SNVs in TSS-proximal upstream regions, which are likely to contain regulatory elements [70], could potentially dysregulate gene expression by disrupting transcription factor (TF) binding. We considered regions ±60 bp relative to SNVs identified within 2kb of TSSs, and used the Genomatix software (https://www.genomatix.de/) to evaluate TF binding site (TFBS) enrichment [71, 72]. We identified putative TFBSs by computing Z-scores associated with the difference in counts of TF motif binding site nucleotide matches in the regions of interest as compared to the entire genome (see analysis details in [73]). We considered TFBSs to be statistically significant if Z > 2 was obtained, under the assumption that the Z scores were normally distributed and Z = 2 corresponds to P = 0.045. Visualizations of TFBS motifs were acquired from the Genomatix software interface.

Time-series analysis of multi-organ gene expression reveals differential dynamics underlying disease progression

We characterized the pathogenesis of physiological dysregulation in autonomic dysfunction by examining organs anatomically integrated with the ANS. We temporally profiled the expression of genes implicated in ANS activity, peripheral/central INF, and RAS signaling (Fig 1A–1C). We acquired data over a span of animal ages encompassing the development of cardiovascular and metabolic disease. Arterial hypertension is a key feature of this animal model of autonomic dysfunction. We selected assay time points characterized by pre-hypertension (4, 6, and 8 wk), hypertension-onset (10 and 12 wk), and robust hypertension (16 wk) [40]. To determine whether specific samples or animals included in our study imparted systematic biases in our results, we utilized Principal Components Analysis (PCA). First, we reasoned that if specific samples or groups of samples were systematically biasing our results, such samples would appear to be distinct from the bulk of our samples in the subspace defined by the first two PCs. Such a finding was not obtained. Rather, all samples grouped together in a large aggregate (S1C Fig). To further examine whether specific animals selectively contributed to bias, we implemented PCA separately for each organ. We reasoned that if specific animals were biasing our results, such animals would be associated with pronounced variation across multiple organs. Our analysis showed that animals with pronounced gene expression differences in one organ were often similar to the majority of the other animals when considering the other organs (S1D and S1E Fig). These analyses did not suggest any systematic sample outliers, and hence, all samples analyzed using qPCR were included in our downstream computational analysis. Our analysis framework consisted of statistical approaches, dynamics analysis, network modeling and analysis, and bioinformatics (Fig 1D–1F).

To examine whether gene expression dynamics were distinct in autonomic dysfunction, as compared to the control phenotype, we performed statistical analyses of gene expression dynamics. We implemented the Optimal Discovery Procedure [11] to address whether the temporal profiles for a given gene in a given organ differed significantly between autonomic dysfunction and control phenotypes. In this analysis, temporal profiles of gene expression were characterized by cubic splines fitted to the data (Fig 2). The time series analysis identified numerous statistically significant differences in the temporal dynamics of gene expression (Fig 2, S5 Fig). In particular, all temporal expression profiles from the brainstem showed significant differences in autonomic dysfunction (false discovery rate <0.1, S5 Fig). Overall, these results highlight the extensive differences in dynamics of gene expression between disease and control phenotypes. Note the apparent gene expression differences observed at the 4 wk time point; we address this finding in the discussion. Importantly, at time points in which the disease and non-disease phenotypes exhibited similar gene expression levels, discrepancies in the underlying dynamic trajectories indicated divergent properties of the underlying gene regulatory networks.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Divergent patterns of dynamic gene expression across organs and genes correspond to autonomic dysfunction.

(A) Expression dynamics of pro-inflammatory gene Il1b across organs and phenotypes. (B) Expression dynamics of angiotensin precursor gene Agt. Error bars indicate standard error of the mean. Smooth curves depict natural cubic splines fit to the data.

https://doi.org/10.1371/journal.pcbi.1005627.g002

Robustness analysis of the HMF method

We thoroughly evaluated the robustness of the HMF method for identifying network models of multi-organ gene regulatory networks. We have previously shown that the HMF method achieves robust performance in the identification of gene regulatory network dynamics and structure [48]. We further evaluated the robustness of our network identification results with respect to regression regularization parameters λ and α, structural properties of the identified networks, and sensitivity to exclusion of data subsets.

We performed the following analyses to compare identified networks with the ‘best fit’ network characterized by J_sim = min(J_sim):

1. Spearman rank correlation of interaction coefficient values
2. Fisher’s exact test for counts of coefficient signs (±1)

For each phenotype, we compared the ‘best fit’ network with many subnetworks. For these comparisons, we evaluated Spearman rank correlation coefficients, Fisher’s exact test (FET) p-values, and odds ratios. For the autonomic dysfunction phenotype, a total of 272 comparison networks were identified and we obtained Spearman rank correlation ≥ 0.77 (P < 2.2 × 10⁻¹⁶, S6A Fig left) along with FET P ≤ 2.1 × 10⁻³⁵. All odds ratios approached infinity. For the control phenotype, 289 networks met our criteria for comparison, Spearman rank correlations exceeded 0.79 (P < 2.2 × 10⁻¹⁶, S6A Fig right), and the FET indicated a high degree of interaction coefficient sign similarity (P ≤ 1.0 × 10⁻²²). The corresponding odds ratios approached infinity due to the absence of coefficient sign discrepancies. In general, for both autonomic dysfunction and control, with α > 0 there was a negligible influence of λ on network correlations for λ ∼ λ_min (S6B Fig). While the control network correlations were generally insensitive to the α value, the autonomic dysfunction networks showed a small degree of sensitivity to α. Finally, the range of m values exerted a negligible influence on network correlations. This is illustrated by the aggregates of data points, of each λ/α combination, corresponding to the ten m value ranges assessed (S6B Fig).

We further examined the robustness of out network identification procedure, with respect to regularization parameters λ and α, by determining three key features of network topology [74]:

1. Average path length: < ℓ >
2. Clustering coefficient: C_i
3. Degree distribution: γ

In general, our network analyses revealed that autonomic dysfunction and control networks were characterized by similar topological properties that were largely consistent with those of other biological networks. We observed < ℓ > ∼ 1.5 − 3 (S7A Fig), well within the typical range for biological networks [75, 76]. Similarly, the clustering coefficient values C_i ∼ 0.1 − 0.4 were consistent with previous observations (S7B Fig) [76, 77]. For our analysis of the power law exponent, we considered only fits for which a K-S p-value of P > 0.8 was obtained, as this was consistent with reasonable evidence supporting the power law degree distribution (S7C Fig). The exponent values generally fell in the range of γ ∼ 2−8. Exponent values of 2 < γ < 3 are typically observed for biological networks [60], in partial agreement with our findings. The discrepancy, however, is not surprising given that our analysis included a limited number of functionally related genes, while global networks on the genomic scale would be expected to have distinguishable properties. Further, the elevated values of γ we observed are consistent with the small-world topology associated with the power law distribution. In general, our correlation-based and graph theoretic analyses support the robustness of our system identification approach with respect to reliability of edges identified and global network properties.

While the findings so far support the robustness of our approach to variation in regression parameterization, we wanted to further address whether we could identify similar edges if subsets of the full network were considered. We considered subsets of both organs and genes. Eight organ subsets were considered:

1. Adrenal gland, Brainstem
2. Brainstem, Kidney
3. Brainstem, Left ventricle
4. Kidney, Left ventricle
5. Brainstem, Kidney, Left ventricle
6. Adrenal gland, Kidney, Left ventricle
7. Adrenal gland, Brainstem, Kidney, Left ventricle
8. Adrenal gland, Brainstem, Kidney, Liver, Left ventricle

For each organ subset, we considered four combinations of gene annotations:

1. Inflammation, Autonomic
2. Inflammation, RAS
3. Autonomic, RAS
4. Inflammation, Autonomic, RAS

Thus, we evaluated 32 separate iterations of our network identification approach, with distinct network subsets, for both autonomic dysfunction and control phenotypes. We provide compact examples of identified network structures via parameter heatmap representations (Fig 3A) and corresponding connectivity illustrations (S8 Fig). We considered the coefficient data for the best fit over ten λ values and ten m-value ranges. As described above, we compared each sub-network to the full network based on the Spearman rank correlation and FET, including normalized interaction coefficients k ∈ (−1, 1) that exceeded two standard deviations from the median in both the full network and the comparison network. We also required that at least five coefficient values meet the stated criteria for both networks. For the control network comparisons, our comparison criteria were satisfied in all of the comparisons. Our results showed Spearman rank correlation ≥ 0.77 (P ≤ 3.8 × 10⁻⁴), FET P ≤ 8.1 × 10⁻⁵, and all odds ratios approached infinity. For the autonomic dysfunction phenotype, all sub- networks satisfied our comparison criteria, Spearman rank correlation ≥ 0.71 (P ≤ 4.2 × 10⁻⁵), FET P ≤ 7.4 × 10⁻⁷, and all odds ratios approached infinity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Data-driven model-based network identification shows that connectivity differences are associated with dynamic differences in autonomic dysfunction.

(A) Phenotype-specific network topology representations of gene-gene interaction coefficients. A given coordinate represents the magnitude of influence of the organ-gene in the corresponding column on the organ-gene in the corresponding row (see Fig 1E). (B) Sample dynamic profiles obtained from the simulation of the mathematical models associated with the interactions described in (A). (C) Degree distributions for both in and out degree. These data show that differences in network topology are associated with differential dynamics in autonomic dysfunction versus control.

https://doi.org/10.1371/journal.pcbi.1005627.g003

In considering our correlation-based and graph theoretic analyses across regularization conditions, along with our correlational analyses of subnetwork comparisons, we concluded that our approach is robust to regression parameterization and the inclusion/exclusion of dynamic variables.

System identification yields dynamic model-based phenotype specific regulatory networks

To examine the dysregulation of physiological homeostasis, we applied systems identification to infer network models corresponding to the autonomic dysfunction and control phenotypes. We utilized the HMF method, which entailed the application of the Hartley transform and associated Hartley modulating functions, to infer data-driven dynamic network models. The principal advantages of the HMF method, in comparison to related methods, are that (1) this approach facilitates the identification of continuous-time mathematical models of gene expression from discrete data, and (2) this approach obviates the need to compute temporal derivatives of discrete data by instead utilizing a frequency domain transformation of the data [48].

The network dynamics were represented by ordinary differential equations (ODEs) that described continuous models of gene expression. The network structures were represented by interaction coefficients that describe either gene regulatory interactions in which one gene’s activity promotes the activity of another gene (‘upregulation’), or interactions in which one gene inhibits the activity of another gene (‘downregulation’, Fig 1E). We consider the terms upregulation and downregulation to represent indirect influences in most cases, rather than direct mechanistic influences.

The network models identified for autonomic dysfunction and control phenotypes depicted the topology of gene-gene influences within and across organs (Fig 3A, S8 Fig). It is noteworthy that regulatory interactions within the brainstem were relatively absent in the control phenotype as compared to autonomic dysfunction. Moreover, brainstem genes showed extensive cross-organ influence patterns in autonomic dysfunction, whereas these interactions were comparatively under-represented in the control (S9 and S10 Figs). It is also worth noting that because we did not have 12 wk data for the brainstem samples the identified brainstem networks were more prominently constrained by the earlier time points; we return to this potential limitation in the discussion.

The gene-gene influences underlying the multi-organ network were determined by using our time-series data to computationally infer the interaction coefficients of a mathematical model formulated by ODEs. According to this approach, the HMF method was utilized to infer phenotype-specific regulatory interactions and expression dynamics. Our simulated dynamic models showed considerable agreement with the experimental data (Fig 3B; see S3 File for all time series data along with dynamic model simulations). We refer to each simulated timeseries as an “expression profile” (smooth traces in Fig 3B). Code and annotation information for implementing our models in Matlab, R, and Systems Biology Markup Language are available (S4, S5, S6, S7, S8, S9 and S10 Files).

To further characterize the properties of phenotype-specific networks, we evaluated the respective degree distributions [74]. The in degree of gene x indicates the number of genes that influence gene x (i.e., the number of non-zero values of a given row in Fig 3A). The out degree of gene x denotes the number of genes that gene x influences (the number of non-zero values in the column corresponding to gene x in Fig 3A). Our analysis showed that the autonomic dysfunction network exhibits a similar though distinct degree distribution, as compared to the control network (Fig 3C). Overall, the gene-gene influence matrices (Fig 3A) along with the degree distribution analysis (Fig 3C) showed that differential network topology is associated with dynamic disturbances in autonomic dysfunction that were identified by timeseries analysis (Fig 2, S5 Fig).

Aberrant gene expression cascades underlie the pathogenesis of autonomic dysfunction

Developmental processes are often characterized by coordinated waves of precisely timed gene expression patterns [13, 78]. We hypothesized that an analogous coordinated cascade of dynamic transcriptional patterns regulates disease progression. We investigated whether temporal sequencing of gene expression corresponded to the development of autonomic dysfunction. We analyzed the sequence of peak times observed in the temporal profiles for the autonomic dysfunction and control phenotypes (Fig 4). In this analysis, we considered genes for which prominent expression peaks were not preceded by prominent expression valleys. Profiles with monotonically decreasing expression levels were considered to exhibit peaks at the initial time point of four weeks [13]. Several brainstem profiles showed peak activity at early pre-hypertensive ages including Adrb2, Agt, Il1a, and Il10 (Fig 4A). Similarly, early peaks were observed for kidney Tgfb1, liver Adra1b, and ventricle Th and Ccl5 (Fig 4A). By comparison with the control expression profiles, our analysis revealed a coordinated cascade of gene expression activation patterns that was specific to the autonomic dysfunction phenotype (Fig 4B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Autonomic dysfunction is associated with a disease-specific regulatory cascade of gene expression patterns.

(A) Gene expression profiles from the autonomic dysfunction phenotype were sorted according to the time that the expression profile showed a peak. The analysis shows a cascade or peak times associated with autonomic pathogenesis and corresponding hypertension development (left). Comparison with control profiles (right) show that the autonomic dysfunction cascade is disease-specific. (B) Examples of temporal profiles with relatively early, intermediate, and late peaks corresponding to the autonomic dysfunction (left) and control (right) phenotypes.

https://doi.org/10.1371/journal.pcbi.1005627.g004

We expanded our analysis by examining gene expression profiles with valleys in autonomic dysfunction (S11 Fig). Similar to the results for the peak analysis, we found that disease development was associated with a cascade of valleys that was specific to the autonomic dysfunction phenotype. Moreover, when we completed peak and valley analyses and ordered the expression profiles according to the control phenotype, we found that distinct dynamic patterns were control phenotype-specific (S12 and S13 Figs). Our dynamic pattern analyses showed that distinct sets of temporal profiles exhibit early peaks/valleys, and the respective peak/valley times were associated with various temporal offsets relative to disease onset. To summarize these findings, we analyzed the peak and valley times for autonomic dysfunction versus control (S14 Fig). This analysis highlighted genes that showed similar versus divergent dynamic patterns of peak/valley timing with respect to phenotype (S14 Fig). Our findings indicate that different subsets of the multi-organ network exhibit distinct aberrations in gene expression dynamics, thereby indicating that gene regulatory networks are re-wired under disease conditions.

Differential gene regulatory network wiring is associated with autonomic dysfunction

Given our finding of kinetic and network structure differences corresponding to the development of autonomic dysfunction (Figs 3 and 4), we wanted to further examine the specific molecular underpinnings of these differences. Divergent molecular networks have been associated with cardiovascular disease and aberrant ANS function [40, 79]. Such connectivity differences reflect re-wiring of the molecular networks that determine the gene regulatory interaction dynamics which underly cell, tissue, and organismal physiology [80]. To interrogate the molecular underpinnings of disease-relevant multi-organ network modifications, we examined the network degree distribution, differential network module structure, and specific gene regulatory influences that were re-wired in the disease phenotype [79, 81, 82].

We analyzed the disease-specific out degree distribution and examined the ordering of out degree across genes (Fig 5A). Genes with the highest out degrees included adrenal Ren and brainstem Il10. Interestingly, adrenal Ren showed an pre-hypertensive valley in autonomic dysfunction (Fig 6B) and brainstem Il10 showed a pre-hypertensive peak (Fig 7D), whereas differential dynamics were observed for the control phenotype. These results suggest the possibility that changes in adrenal Ren and brainstem Il10 are key early events that critically influence the pathogenesis of autonomic dysfunction by virtue of their respective kinetics and substantial participation in the regulation of gene expression within and across organs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Differential network structure in autonomic dysfunction.

(A) Out degree is shown according to the degree order of the autonomic dysfunction network. The control network exhibited a divergent out degree pattern. (B) A highly connected module in the autonomic dysfunction network with genes that have prominent influences (columns, regulators) and their targets (rows). Each row and column of the full network matrix (Fig 3A) contains more than 15% of non-zero entries. The corresponding control subnetwork is comparatively sparse. (C) Diagram illustrating gene-gene interactions that are present only in the autonomic dysfunction network. (D) Interactions that are only present in the control network. (E) Autonomic dysfunction-specific interactions that have the opposite sign of the corresponding control interactions (e.g., adrenal Ren upregulates brainstem Tnf in the autonomic dysfunction phenotype but downregulates brainstem Tnf in the control phenotype).

https://doi.org/10.1371/journal.pcbi.1005627.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Divergent dynamics corresponding to differential network structure in autonomic disease.

(A) Dynamic profiles related to gene-gene interactions exclusive to the autonomic dysfunction network. (B) Dynamic profiles related interactions exclusive to the control network. (C) Dynamic profiles related inverted interactions in the autonomic dysfunction phenotype as compared to the control.

https://doi.org/10.1371/journal.pcbi.1005627.g006

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Dynamic profiles support specific network interactions in the brainstem of the autonomic dysfunction phenotype.

(A) Brainstem subnetwork for the autonomic dysfunction phenotype. (B) Interactions involving Agtr1 are highlighted. (C) Example of a three node ‘adaptation motif’ from the autonomic dysfunction phenotype. (D) Dynamic profiles corresponding to the motif from (C) show expected behavior only for the autonomic dysfunction phenotype.

https://doi.org/10.1371/journal.pcbi.1005627.g007

Modules of highly interconnected subnetworks or modules of gene regulatory interactions are critical for the regulation of biological networks [82]. We tested whether we could identify a highly connected module in the autonomic dysfunction network that would be expected to critically regulate autonomic (dys)function. We detected such a module (Fig 5B) which highlighted some of the key influences of genes with high out degrees (Fig 5A). For instance, adrenal Ren upregulated the expression of several genes including liver Adrb1, ventricular Agtr1, and brainstem Il10. In contrast, adrenal Ren downregulated the expression of these genes in the control phenotype (Fig 5B, right). Moreover, adrenal Ren downregulated genes including brainstem Agtr1 in autonomic dysfunction but exerted a corresponding upregulatory influence in the control network (Fig 5B).

To further define the network differences between the autonomic dysfunction and control phenotypes, we queried the networks for gene-gene interactions that were present on only one network as well as interactions that exert opposing influences with respect to phenotype (e.g., interactions that are upregulating in autonomic dysfunction but downregulating in the control network) [83]. We applied a conservative approach that identified only the interactions that showed relatively large differences, yielding a high confidence set of network differences (Methods, “Differential network analysis”). Our analysis revealed that a number of gene regulatory interactions were present exclusively in the autonomic dysfunction network (Fig 5C). For instance, brainstem Adra1b expression dynamics influenced the temporal patterns of Agt and Gja1 expression in the brainstem for the autonomic dysfunction phenotype. To determine if the observed gene regulatory network differences in autonomic dysfunction could be related to differential dynamics of gene expression, we examined the phenotype-specific brainstem expression profiles. This analysis showed that, for the autonomic dysfunction phenotype, Adra1b exhibited a valley following and early peak in Agt and an early valley in Gja1 expression (Fig 6A). Further analysis showed that the decrease in Adra1b preceded decrease in Agt and increase in Gja1. Considered together, these temporal profiles from the autonomic dysfunction phenotype were consistent with the topological interactions amongst these brainstem genes based on the identified network.

We next examined whether there was evidence for control–specific network interactions (Fig 5D). Our analysis revealed that control adrenal Ren upregulated brainstem Il1b and kidney Adrb1 expression. We studied the respective temporal profiles and found that a decrease in adrenal Ren preceded a decrease in brainstem Il1b and kidney Adrb1 in the control phenotype, presumably through reduction of an upregulating influence (Fig 6B). In contrast, for the autonomic dysfunction phenotype, brainstem Il1b and kidney Adrb1 dynamics showed increases during the pre-hypertensive and hypertension-onset time frames, respectively, which could be related to the absence of Ren-mediated influence (Fig 6B).

We concluded our differential network analysis with an evaluation of whether we could identify gene-gene influences that switched sign (e.g., from positive upregulation to negative downregulation) in the autonomic dysfunction phenotype relative to the control network. Our analysis showed evidence for inverted interactions, the majority of which involved adrenal Ren (Fig 5E). For example, adrenal Ren upregulated Tnf but downregulated Agtr1 in the brainstem of the autonomic dysfunction phenotype. The respective expression profiles showed that the attenuation of adrenal Ren expression in autonomic dysfunction preceded a decrease in brainstem Tnf and an increase in brainstem Agtr1. These results could be due to attenuated upregulation and downregulation influences of adrenal Ren on brainstem Tnf and Agtr1, respectively (Fig 6C). Consistent with the inversion of these interactions, the temporal patterns of brainstem Tnf and Agtr1 expression was inverted in the control phenotype relative to that of autonomic dysfunction (Fig 6C). In sum, our studies on differential dynamics and network structure illustrate that our integrated analyses can be applied to identify potential disease mechanisms associated with the organ-specificity and timing of gene expression.

Network motif analysis-based identification of a core module involving the brainstem angiotensin-II receptor transcript

Because our differential network analysis was designed to impose strict criteria for the attribution of phenotypic differences in gene-gene interactions, we elaborated on this analysis. Given the phenotypic disparity in intra-organ gene-gene connectivity in the brainstem (Fig 3A), we specifically focused on the brainstem subnetwork. Further, because connectivity was largely absent in the control brainstem network (S15 Fig), we focused exclusively on the autonomic dysfunction brainstem network (Fig 7A). In particular, we posed the question of whether we could identify network motifs associated with the adaptation to prolonged stimuli [84]. Network motifs that support adaptation have important biological functions and the motif dynamics are readily interpret-able [84, 85]. Hence, we tested whether we could identify three node motifs capable of adaptation in the brainstem network of the autonomic dysfunction phenotype (Fig 7A).

Our analysis of the autonomic dysfunction brainstem subnetwork showed numerous three node feed-forward motifs, many of which could support adaptation according to topological considerations [84] (S16 Fig). Brainstem expression of the angiotensin-II type-1 receptor, encoded by Agtr1, is known to exert prominent influences on ANS function. Hence, we focused on motifs including Agtr1 (Fig 7B). We examined the expression profiles of 3-node motifs including Agtr1. This analysis showed that it was difficult to interpret many of the motifs in terms of the corresponding dynamic expression profiles. For instance, consider a motif in which Dbh upregulates both Agtr1 and Ccl5, while Ccl5 downregulates Agtr1 (S17A Fig). The similarity between the temporal profiles of Ccl5 and Agtr1 is inconsistent with the downregulation of Agtr1 by Ccl5 (S17B Fig, left).

Many other motifs were associated with similar ambiguities regarding interpretation of the dynamics. In contrast, a motif involving Adra1a, Il10, and Agtr1 was associated with temporal dynamics that agreed well with predictions based on motif topology (Fig 7C). Elevated levels of Adra1a were followed by increased levels of Il10 and decreased levels of Agtra1 (Fig 7D, left), consistent with the respective upregulating and downregulating influences of Adra1a on Il10 and Agtr1, as defined by the motif topology (Fig 7C). As Adra1a levels were decreased with time, a subsequent reduction of Il10 and increase of Agtr1 was observed, consistent with the motif structure (Fig 7C and 7D). Despite the existence of numerous feedback loops and complex arrangements of interactions, our analyses revealed that structural features of our multi-organ model could be related to the underlying dynamics in a context-dependent manner. Taken together with our differential network analysis findings, our results suggest that there are core modules that exhibit dynamic behavior consistent with isolated subsets of gene-gene interactions both within and across organs.

Bioinformatic identification of regulatory elements modified by genetic variants associated with autonomic dysfunction

Disease-specific genetic aberrations can contribute to, rather than respond to, the disease state [86]. We tested whether autonomic dysfunction was associated with disease-specific single nucleotide variations (SNVs) that could be related to either protein sequence or regulatory regions of DNA that function as putative transcription factor binding sites (TFBSs). Mutations in TFBSs have been shown to be implicated in complex disease pathology [87], and identification of novel regulatory SNVs could yield important insight as to the mechanistic basis of the alterations of network dynamics and connectivity we observed in autonomic dysfunction.

We initially focused on examining the presence of SNVs in coding regions of genes for which expression was analyzed in our study. We identified several coding SNVs that were all synonymous, such that the corresponding amino acid sequence remains unchanged by the SNV. These SNVs have been previously annotated (rs8154045, rs10546839, rs198233992, rs199262884, rs197275571, rs198523377, rs8174203) [88]. While it is possible that synonymous SNVs could alter gene expression, the functional implications of such variants are currently unclear [89].

Next, we determined whether 2 kb regions upstream of the transcription start sites for our genes of interest contain SNVs in the autonomic dysfunction phenotype. The analysis revealed several SNVs in these gene-proximal upstream regions, which are putative regulatory sites associated with TF binding. Recent analyses have shown that SNVs either within or in proximity to TFBSs can modulate transcriptional regulation [90], with relevance to disease phenotypes [91]. To examine whether the SNVs we identified could potentially disrupt TF binding in gene-proximal regulatory regions, we evaluated the statistical enrichment of TFBSs in the vicinity of the SNVs within these regions (SNV ±60 bp [71]). Our analysis revealed four SNVs with statistically significant enrichment for TFBSs. We found that the upstream region of Adrb1 contains a autonomic dysfunction-specific SNV within a putative binding site for Tfap2a (Fig 8A). Similarly, we found other TFBS enriched sequences near gene-proximal SNVs in the autonomic dysfunction phenotype (Fig 8C, S18 Fig). In addition to Tfap2a, the analysis identified TFBSs for Ebf1 and Ybx1, with sites proximal to SNVs associated with Adrb1, Agt, Il1b, and Tnf in the autonomic dysfunction phenotype (Fig 8C, S18 Fig). Our analyses suggest that altered connectivity and dynamics in autonomic dysfunction could be related to genetic variations that influence the expression of functionally important genes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Autonomic dysfunction-specific single nucleotide variants in regulatory regions upstream of prominent genes.

(A) Example spatial relations amongst an SNV, TFBS, and transcription start site. (B) Transcription factor binding motif associated with Tfap2a along with the respective sequences in the autonomic dysfunction and control phenotypes. A single nucleotide variation (SNV) occurs in a key element of the putative binding site for the autonomic dysfunction phenotype. (C) Summary of TFs with bioinformatically identified TFBSs in or near autonomic dysfunction-specific SNVs upstream and proximal to genes implicated in physiological homeostasis.

https://doi.org/10.1371/journal.pcbi.1005627.g008

Conclusion

Our process dynamics-based approach presents a novel experimental and analytic paradigm for the dissection of mechanisms underlying disease pathogenesis. Our analyses yielded several hypotheses for novel mechanisms of autonomic dysfunction development. Our dynamic and network analyses are based on a mathematical model that can be further utilized for simulations and computational analyses directed to unravel new disease mechanisms and optimize treatment strategies [94, 119]. Our data-driven model can also serve as a basis for identifying putative biomarkers for prediction of disease onset [20]. Future directions include refining the temporal resolution of the time series analysis, examining a broader range of ages, and utilizing an expanded repertoire of molecular and functional assays. Investigation into whether network interactions undergo re-wiring throughout the progression of disease [76] will be important for understanding the mechanisms underlying how therapeutic interventions can restore pathological molecular networks to a healthy state [120]. Targeted perturbations will be critical to confirming mechanistic influences of gene regulatory interactions. In sum, our novel integrated approach can be more broadly applied to examine the developmental dynamics of numerous chronic disease states. Such analyses are expected to yield novel mechanistic knowledge that can facilitate the identification of novel biomarkers and therapeutic treatment strategies based on the structure and dynamics of multi-organ interactions underlying organismal homeostasis.