A dynamical system model for predicting gene expression from the epigenome
James Brunner, Jacob Kim, Timothy Downing, Eric Mjolsness, Kord M. Kober
11 Epigenome Prediction of Gene Expression using a Dynamical System Approach
James Brunner † Center for Individualized Medicine Microbiome Program, Mayo ClinicRochester, MN 55901, USA † Jacob Kim
Timothy Downing
Eric Mjolsness
Kord M. Kober ‡ Department of Physiological Nursing andBakar Computational Health Sciences Institute, University of California San FranciscoSan Francisco, CA 94143, USA ‡ Gene regulation is an important fundamental biological process. The regulation of gene ex-pression is managed through a variety of methods including epigentic processes (e.g., DNAmethylation). Understanding the role of epigenetic changes in gene expression is a funda-mental question of molecular biology. Predictions of gene expression values from epigeneticdata have tremendous research and clinical potential. Dynamical systems can be used togenerate a model to predict gene expression using epigenetic data and a gene regulatorynetwork (GRN). Here we present a novel stochastic dynamical systems model that predictsgene expression levels from methylation data of genes in a given GRN.
Keywords : Epigenetic Modification, Gene Regulatory Networks, Piecewise DeterministicMarkov Processc (cid:13) a r X i v : . [ q - b i o . M N ] A ug
1. Introduction
Gene regulation is an important fundamental biological process. It involves a number ofcomplex sub-processes that are essential for development and adaptation to the environment(e.g., cell differentiation and response to trauma ). Understanding gene expression patternshas broad scientific and clinical potential, including providing insight into patterns of funda-mental molecular biology (e.g., gene regulatory networks) and a patient’s response to disease(e.g., HIV infection ) or treatment (e.g., chemotherapy-induced peripheral neuropathy ).The regulation of gene expression is managed through a variety of methods, includingtranscription, post-transcriptional modifications, and epigenetic processes. One epigeneticprocess, DNA methylation, occurs primarily at the cytosine base of the molecule that isadjacent to guanine (i.e., CpG site). DNA methylation of promoter and gene body regionscan act to regulate gene expression by repressing (e.g., ) or activating (e.g., ) transcription.Understanding the role of epigenetic changes in gene expression is a fundamental questionof molecular biology and predicting gene expression from epigenetic data is an active area ofresearch. Predictions of gene expression values from epigenetic data have tremendous researchand clinical potential. For example, DNA is inexpensive to collect and is easy to store. It offersboth genetic (i.e., genotype) and epigenetic (i.e., methylation status) information in a stableformat. This information is obtainable from a large number of ongoing and previously existingbiobanks. Gene expression (i.e., RNA), however, is more difficult and more expensive to obtain.Given the unique type of information that gene expression can provide (i.e., the presence andquantity of the functional product of a gene), it will be very useful and economical if geneexpression values could be reliably predicted from methylation information.A dynamical systems model can predict gene expression using epigenetic data and a generegulatory network (GRN) by simulating hypothesized mechanisms of transcriptional regu-lation. Such models provide predictions based directly on these biological hypotheses. Wedevelop an interaction network model that depends on epigenetic changes in a GRN.One major advantage to the dynamical systems approach is obtaining a distribution ofgene expression beyond just expression status. Furthermore, a stochastic dynamical systemprovides us with a distribution of gene expression estimates, representing the possibilities thatmay occur within the cell.Recent studies have developed models to predict gene expression levels with deep con-volutional neural networks from genome sequence data and with regression models frommethylation data. Previous studies have developed models to predict expression status (e.g.,on/off or high.low) with gradient boosting classifiers from histone modification data, andwith machine learning classification methods from methylation data and from methylationand histone data combined. To our knowledge, there are no studies that have taken a dy-namical systems approach to predicting gene expression from methylation data and a GRN.Given the opportunity presented by dynamical systems approaches and the potential prac-tical utility, we present a novel stochastic dynamical systems model for predicting gene ex-pression levels from epigenetic data for a given GRN.
2. Methods2.1.
Assumptions
Our model assumes fundamentally that transcription of DNA is (relatively) fast and doneat a linear rate determined by the bound or unbound state of transcription factor bindingsites. We assume that binding and unbinding of transcription factors is a (relatively) slowand stochastic process, with propensity proportional to availability of transcription factor.Our model is built on the hypothesis that the propensity of transcription factor bindingis influenced non-linearly by epigenetic modification of the binding site. We assume that“background” transcription (i.e. transcription occurring when no transcription factor is bound)is only enough to allow down regulation to exist. In model training and testing, genes withmissing expression data are assigned ’zero’ expression values.
Model Equations
We model gene regulation using a piecewise-deterministic Markov process (PDMP) asintroduced in Davis 1984 (see also ) given by the equations: B i ( t ) = B i (0) + Y i (cid:18)(cid:90) t (1 − B i ( τ )) λ i µ i µ i + ( α i ) ν i ( κ i · g ) dτ (cid:19) − Y i (cid:18)(cid:90) t ˆ λ i B i ( τ ) dτ (cid:19) (1)and ddt g j = γ j + ( φ j · B ) − d j g j (2)where B i ( t ) ∈ { , } , is a boolean random variable representing the bound/unbound state ofa binding site region of DNA and g i is the transcript amount the genes modeled. Equation (1)is given as the sum of two Poisson jump processes Y i ( h i ( t )) and Y ( h i ( t )) which take values in Z ≥ , and are piecewise constant between randomly spaced discrete time points (the bindingand unbinding events). The propensities h i ( t ) and h i ( t ) are taken to be linear functions of the available tran-scription factors, which is assumed to be the same as the transcript variables g j . We take thevalues κ ij ∈ { , } ; these parameters along with the set of φ ji represent the structure of theunderlying gene regulatory network.We include the term µ i µ i + ( α i ) ν i (3)to represent the impact of epigenetic modification on the binding propensity of transcriptionfactors. In this term, α i is the measured epigenetic modification to a transcription factorbinding site (e.g. percentage of methylated bases). Equation (3) is a sigmoidal function whichis either strictly increasing or strictly decreasing depending on the sign of ν i . If ν i > , then thisterm decreases, implying that epigenetic modification decreases transcription factor binding.Conversely, if ν i < , the model implies that epigenetic modification increases transcriptionfactor binding.Finally, we use a linear ODE for the value of the transcripts g j . We take φ ji ∈ {− , , } based on the structure of the underlying gene regulatory network. We include baseline tran-scription γ j and decay d j . Because we use a linear ODE in Eq. (2), we can solve exactlybetween jumps of B . Master Equation and Equilibrium Distribution
It is common practice in the study of reaction networks modeled as stochastic jump pro-cesseDetail of figure 9 in manuscript.s to represent the process using so called “chemicalmaster equation”, which is the Kolmogorov forward equation for the jump process. Thegenerator for a PDMP can be defined (see Aza¨ıs 2014 for details). We define a density P ( B i , g ( t ) , t ) = P i ( g ) , i = 1 , ..., | B | for each possible state B i of B such that (cid:80) | B | i =1 P i ( g ) = P ( g ) is the probability distribution for the vector g , and each P i satisfies dP i ( g , t ) dt = | g | (cid:88) j =1 (cid:0) γ j + ( φ j · B i ) − d j g j (cid:1) ∂P i ( g , t ) ∂g j + (cid:88) ( j : (cid:107) B j − B i (cid:107) =1) | B | (cid:88) k =1 (cid:20) B ik (1 − B jk ) λ k µ k µ k + ( α k ) ν k ( κ k · g ) + ˆ λ k B jk (1 − B ik ) (cid:21) P j ( g , t ) − | B | (cid:88) k =1 (cid:20) (1 − B ik ) λ k µ k µ k + ( α k ) ν k ( κ k · g ) + ˆ λ k B ik (cid:21) P i ( g , t ) . (4) Model Sampling & Equilibrium Distribution Estimation
We can estimate the equilibrium distribution for the PDMP using kernel density estimation(KDE) with a Gaussian kernel. To compute the marginal distributions on the various genevariables, which we can estimate with a 1 dimensional kernel. In a realization of the process,binding or unbinding events occur at times t i such that [ t , t ) (cid:116) [ t , t ) (cid:116) · · · (cid:116) [ t n − , T ) = [ t , T ) and we may compute g ( t ) exactly in each interval. The estimation is then f i ( x ) = n − (cid:88) k =0 (cid:90) t k +1 t k √ πh exp (cid:32) − (cid:0) x − (cid:2) e − d i ( t − t k ) ( g i ( t k ) − S ki ) + S ki (cid:3)(cid:1) h (cid:33) dt (5)where S ki = d i ( γ i + (cid:80) j φ ij B j ) . In Fig. 1, we show schematically how f i ( x ) is estimated from arealization of the process B ( t ) , g ( t ) . For each binding site i α li κ i λ i ˆ λ i µ i ν i B i ( t ) For each gene j φ j d j γ j g j ( t ) f lj ( x ) = R T l K j ( g j ( t ) , x ) dt For each data point l Fig. 1. Plate diagram of the process to estimate the marginal PDF f j ( x ) of each gene’s transcript level according to our model.Parameters in diamonds are read from data, parameters in hexagons are determined by the structure of the network, parametersin circles must be fit to the model by maximizing likelihood over a training data set, and parameters in stars are the state variablesof the dynamical model. Notice that the dynamical model implies that the state variables depend on each other, meaning thisnetwork of dependence is not acyclic . The kernel K j ( x, y ) used to estimate likelihood is Gaussian kernel in only component j (i.e., it is the Guassian kernel after orthogonal projection onto dimension j ). Model Parameter Estimation
While the parameters κ ij , φ ji and γ j are determined by the structure of the underlying generegulatory network, assumed to be known, and the epigenetic parameter α i is assumed mea-surable, we must still estimate the remaining parameters λ i , ˆ λ i , µ i , ν i and d j . We estimate theseparameters using a negative log-likelihood minimization procedure using stochastic gradientdescent. Sample paths used to estimate the gradient of the likelihood are generated using amodified form of Gillespies algorithm which handles time-dependent jump propensities. Thisprocedure involves approximating the gradient of the map from parameter set to log-likelihoodso that we can use a gradient descent method. In Fig. 2, we give a schematic representationof how ˆ L D is estimated from a set of realizations of the model, each realization correspondingto a single data sample. For each binding site i α li κ i λ i ˆ λ i µ i ν i B i ( t ) For each gene j φ j d j γ j g lj g j ( t ) L l = T l R T l K ( g ( t ) , g l ) dt For each data point l ˆ L D = − P | DataSet | l =1 ln( L l ) Fig. 2. Plate diagram of the process to estimate total likelihood of a data set according to our model. Parameters in diamondsare read from data, parameters in hexagons are determined by the structure of the network, parameters in circles must be fit tothe model by maximizing likelihood over a training data set, and parameters in stars are the state variables of the dynamicalmodel. Notice that the dynamical model implies that the stat variables depend on each other, meaning this network of dependenceis not acyclic . The kernel K ( x, y ) used to estimate likelihood is Gaussian. Gradient Estimation
In order to estimate ∇ ˆ L D, ω for use in optimization, we can use the generator of the system.To do this, we must rewrite the likelihood estimate L g , α ( θ ) as a sum of partial likelihoods L i g , α defined for each possible state of the vector B . Then, because we are estimating the likelihoodat an equilibrium distribution, we may use Eq. (4) and assume that dP i dt = 0 to compute each ∇ L i g , α . See supplemental file for a derivation of ∇ L i g , α using this method. Creation of model from known GRN
We use a gene regulatory network assumed to be known to create a model ofgene regulation that includes transcription factor binding dynamics. To do this, we as-sociate binding sites with the genes that they regulate, and use these associations tocreate a bipartite graph. Due to the available level of detail in the data set, werely only on associations between binding sites and targets. We therefore assume thatany regulator of a given target binds with every site associated with that target. g g g g g g B x ) B x ) B x ) B x ) B y ) B y ) Fig. 3. (left) Underlying gene regulatory network. (right) Bipartite networkof the underlying GRN.
In order to capture the effectsof different regulating transcrip-tion factors that may bind tothe same site, we create “du-plicate” variables that representthe same binding site but boundwith different transcription fac-tors. The result is that we haveonly one transcription factor foreach binding site variable. Asshown in Fig. 3, each edge ofa graph representing the originalgene regulatory network is replaced by a path from transcription factor to binding site to tar-get gene. Note that the model as described in Section 2.2 allows a more general topology, andour procedures for parameter fitting and model analysis do not depend on this construction.
3. Model Evaluation3.1.
Gene Regulatory Network
Gene to gene interactions ( κ ) were defined using the Discriminant Regulon ExpressionAnalysis (DoRothEA) framework. Transcription factor (TF) to target interactions were de-fined using the DoRothEA highest confidence interactions and scored are just 1 or -1 for up-regulating and downregulating, respectively. Binding site to target edges ( φ ) were defined byCpG methylation sites which were associated with changes in transcript expression (eCpG). eCpG probes were identified for genes with which a change in expression was associated witha change in methylation state in participants from the Multi-Ethnic Study of Atherosclerosis(MESA). These CpG probes were then classified into a status which described the geographicrelationship between CpG probe and the associated gene (e.g., “TRANS”, “Promoter”, “TSS”,“Distal”). The binding site regions used in this study for a gene were defined by the proximalstatus classifications of the MESA data (i.e., ‘Promoter’ or ‘TSS’) and were scored as the meanof all of CpG probes in the status class. For evaluation, we identified a set of genes previouslyidentified as deferentially expressed in individuals with PTSD as compared to controls. Ofthese, we identified 278 regulator to target maps (kappa) which we then used to identify othertargets to include. The final set included 512 gene to gene relationships comprised of 303unique target genes. A GRN was built using these 303 genes as input producing a final net-work with 74 genes with 65 sites (network shown in supplemental file). Of these 74 genes, 29had sufficient regulatory information (i.e., an associated binding site and transcription factor)for which parameters could be estimated and expression distributions generated.
Model Training and Testing Data
Matched epigenetic and gene expression data were obtained from the Grady TraumaProject (GTP) study. Methylation and gene expression was measured from whole blood fromAfrican American participants. Methylation data were obtained from the NCBI Gene Expres-sion Omnibus (GEO) (GSE72680) and measured using the Illumina HumanMethylation450
BeadChip. Methylation status was quantified as a beta score. A total of 19,258 eCpG probeswere used as input. Beta scores for CpG sites within the same region for a gene (i.e., ‘Promoter’or ‘TSS’) were aggregated together as the average. Probes were merged in 1,885 regions. Re-gions that were not detected in any sample (i.e., “NaN”) were removed. Gene expression datawere obtained from GEO (GSE58137) measured with the Illumina HumanHT-12 expressionbeadchip. Two batches of non-overlaping samples were quantified using two versions of thebeadchip (V3.0: n=243 participants, mean expression intensity = 189.96, IQR = 49.88 to106.60; V4: n=106 participants, mean expression intensity = 321.58, IQR = 88.85 to 139.78).Intensity scores were log2-transformed. Gene expression probes were first annotated to EN-TREZ ID and then annotated to the symbol using the HUGO annotation. Genes withmultiple gene expression measurements (i.e., multiple probes) were represented by the lastone in the list when processed. The final merged datasets contained n=243 samples for theparticipants measured for expression on the V3.0 beadchip and n=97 for the participantsmeasured for expression on the v4.0 beadchip.
Fig. 4. The mean and variance over time for simulation of gene AK3 for sample 5881.Fig. 5. Negative log-likelihood estimate of training data set forsplit 99x at each iteration of maximum-likelihood gradient de-scent procedure.
Matched data from participants mea-sured for expression on the v3.0 beadchip(n=243) were used for evaluation. Thisprimary dataset was split into trainingand testing datasets, containing 80%/20%(n=195, n=48 samples, respectively). Toavoid the impact of a particular split, werepeated the shuffle process 100 times. Foreach split of the data, parameter estima-tion was performed on the training set andequilibrium distributions of the predictedexpression levels were generated using thetesting set. We evaluated the performanceof the method as the difference between themeasured and predicted value for each gene.
Estimation of equilibrium distribution
Fig. 6. Equilibrium distribution plots for CCM2 for individual (a) 6110 in shuffle 43, and (b) 6742 in shuffle 77.Table 1. Average root mean square errors for eachgene across the 100 shuffles with fitted parametersand random parameters.
GeneSymbol AverageRMSEfittedparameters AverageRMSErandomparameters Difference Relativeperformance
AHR 2.196 6.113 3.918 2.785AK3 0.757 10.459 9.702 13.810ALOX5 2.172 6.797 4.625 3.129BAG3 1.276 7.251 5.976 5.685BAK1 1.637 8.114 6.478 4.958CCM2 0.876 10.842 9.966 12.377CD19 0.925 9.389 8.464 10.148CD4 1.232 8.982 7.750 7.290CTSH 1.262 8.278 7.016 6.561CXCR5 4.425 7.444 3.020 1.683CYP27A1 0.886 8.535 7.650 9.636FBXO32 1.163 8.939 7.777 7.689FCER1A 1.536 9.926 8.391 6.463GSTM1 0.792 11.160 10.369 14.101ICAM4 2.101 7.583 5.482 3.609IRF1 3.357 10.636 7.280 3.168LDHA 2.693 10.870 8.178 4.037LTA4H 3.193 11.846 8.653 3.710MT1X 1.222 9.468 8.247 7.749NR1D2 0.932 7.585 6.654 8.142OAS1 1.162 9.794 8.633 8.431RPL39L 1.211 7.995 6.785 6.602RRM2B 1.085 7.658 6.573 7.057SCP2 4.598 7.354 2.757 1.600SLC20A1 1.877 10.684 8.807 5.691SREBF1 1.784 6.582 4.799 3.691SURF6 1.107 9.236 8.130 8.347VWA5A 2.744 7.310 4.567 2.664ZNF654 3.254 8.015 4.762 2.464
In order to conserve computational resources,we used as a stopping condition for sample pathsimulation the number of random numbers thatwere needed for simulation to some time point.This was number was set to random draws.To evaluate if the gradient descent process was be-ing ended too early, one additional shuffle was per-formed (”99x”) which used an existing shuffle (”99”)but with parameters allowing for less constrictivestopping conditions maxsteps=10000 StoppingCon-dition=100. While simulating, we saved mean andvariance estimates for the distribution for interme-diate time-points in order to approximate the rateof convergence to an equilibrium distribution. Weuse a finite difference estimate for the rate of changeof estimated mean and variance to determine if thestopping condition that we used is appropriate. Wecan see from Fig. 4 that longer simulation time maybe necessary for accurate estimates of equilibrium.
Performance evaluation
To evaluate the performance of our fitting procedure on gene expression predictions wegenerated predictions using a randomly generated parameter set. To avoid the impact ofa particular split, we randomly generated parameters for each of the previously generatedsplits (i.e., the 100 splits previously used for training and testing). Ten random estimateswere generated for each shuffle giving 1000 predictions for each gene generated using randomparameters. We evaluated the performance of the parameter fitting method as the ratio ofroot-mean-square error of the predicted value given by randomly chosen and fitted parameters:
Relative performance = Average RMSE(random parameters)Average RMSE(fitted parameters) .
4. Results
Fig. 7. Equilibrium distribution plots generated from random parameters for AK3 for individual (a) 6522 for random parameterset 0 in shuffle 0, and (b) 8331 random parameter set 7 in shuffle 13.
Estimated parameters
Fig. 8. (Top) Predicted versus ob-served expression values and (Bottom)residuals for the test samples for all 100shuffles for AK3. Each shuffle is col-ored.
Parameters were estimated for all genes using the proce-dure detailed in Section 2.5. Figure 5 shows the likelihood ofthe training dataset during the course of the gradient descentprocedure.4.0.2.
Equilibrium distributions
Equilibrium distributions were generated for 29 genes forwhich we had sufficient connectivity information available.Example equilibrium distributions for the CCM2 gene for twosamples from different shuffles are shown in Fig. 6. Averageroot mean square errors for each gene across the 100 shuffles isreported in Table 1. We observed the highest performance forAK3 (average RSME = 0.757, Fig. 8) and lowest for SCP2(average RSME = 4.598). In this evaluation, our model bi-ases towards underestimating the expression levels (see sup-plemental file). Relative to a random set of parameters, ourfitted parameters improved the predictions by a factor of 1.6to over 14 (Table 1). For example, the predicted expressiondistribution for AK3 generated from random parameters areshow in Fig. 9 and the predicted versus observed values areshown in Fig. 7.
5. Discussion
In this study, we demonstrate that gene expression levelscan be accurately predicted from methylation state of a pro-moter region and a GRN. Our model successfully uses quantitative data describing epigeneticmodification of transcription factor binding sites to generate a probability distribution whichdescribes the possible level of transcript. To our knowledge, this is the first study to developand evaluate a stochastic dynamical systems model predicting gene expression levels from epigenetic data for a given GRN.By using a dynamical systems approach, our model generates an estimation of gene ex-pression given DNA methylation based on the mechanistic hypothesis of differential bindingaffinity of a transcription factor caused by epigenetic modification. Furthermore, the dynam-ical systems approach allows study of complex regulatory networks, including those whichcontains cycles. This method has the potential for broad practical usage. DNA is broadlyavailable in banked tissue in the absence of RNA. By measuring methylation from these spec-imens, RNA expression value can be estimated. In addition to predicting RNA expressionfrom specimens, our model can also be used to evaluate for functional effects of changes inmethylation states at particular sites on gene expression levels. Fig. 9. (Top) Predicted versus ob-served expression values and (Bottom)residuals for the test set generated from10 sets of random parameters for all100 shuffles for AK3. Each shuffle andparameter set is colored.
Overall we see good performance in the predictions of themodel with fitted parameters (e.g., Fig. 8) and dramatic im-provements to prediction relative to a randomly generatedset of parameters (e.g., Fig. 9). Although the predictions aresomewhat dependent on the selection of training data, and insome cases, the observed value and predicted value are notwell represented by the equilibrium distribution the overallprediction of the occurrence of gene expression remains accu-rate (i.e., on/off or low/high). Poor predictions could be theresult of slow convergence in the maximum-likelihood param-eter estimation. In order to estimate parameters, we must it-eratively generate model predictions for a range of parametervalues. This is done by generating equilibrium distributionsin a method similar to Markov-Chain Monte Carlo (MCMC)sampling.The estimated fit of the model to training data improvedover iterations of the procedure (e.g., Fig. 5). However, themean and standard deviations from the equilibrium distribu-tions do not converge as quickly as we would like (e.g., Fig. 4).This slow convergence, and the necessity for repeated esti-mations, mean that computational time is a limiting factor.Future analyses should simulate longer to identify the appro-priate cut offs given the data used, and thus improve the fitof the model parameters.We were able to accurately predict gene expression despitethe limited size of our GRN. We expect that model predic-tions will improve with more regulatory information. Although we queried for 302 genes, ourtranscription factor to target and binding site reference data produced a gene regulatory net-work with 74 genes, of which 29 had sufficient regulatory information to be predicted. Inparticular, dynamical systems should perform well in gene regulatory networks with complextopologies, including those with loops. We were unable to evaluate a more complicated GRNfrom all reference regulatory data due to computational constraints. As such, we evaluated with the toy model of PTSD data which was acyclic.While the use of a stochastic dynamical system has distinct advantages over morestatistically-driven methods, there are of course limitations to the approach as well. In partic-ular, our model is based on the assumption that epigenetic modification effects the propen-sity of the random process of transcription factor binding and unbinding. Furthermore, ourmodel assumes that DNA transcription is a comparatively fast (and so approximated as deter-ministic) process that depends on transcription factor binding. Finally, our model implicitlyassumes that RNA translation to protein product is immediate. In addition to the funda-mental limitations of the model, we also limit the scope of our testing to linear productionof DNA transcript, depending on transcription factor binding status. Future efforts will befocused on improving the prediction accuracy, improving prediction robustness across folds,and evaluating across other gene regulatory networks and gene sets.In conclusion, we have developed a dynamical systems approach which accurately predictsgene expression from methylation state and a GRN. Our results support the idea that methy-lation patterns of cis-promoter regions are associated with gene expression levels. Advances ingene regulatory information will continue to improve the predictions of our model by providingmore structure to the GRNs. In addition, we have a route forward to develop optimizationswhich can step up the ease of use and scaling of our approach. Finally, our approach is broadlyaccessible and can be used for a diverse array of research projects which have DNA samplesavailable but in which gene expression data are missing or limited or in studies evaluating thefunctional effects of changes in methylation state on gene expression.
6. Code availability
Supplemental files, including further mathematical details, are available at https://doi.org/10.5281/zenodo.3970970 . All code are available at https://github.com/kordk/stoch_epi_lib .
7. Acknowledgments
This project was initially conceived as an interdisciplinary project as part of the “ShortCourse in Systems Biology - a foundation for interdisciplinary careers” at the Center forComplex Biological Systems at the University of California Irvine held Jan. 21 - Feb. 8, 2019in Irvine, CA (NIH GM126365). This study was funded by the National Cancer Institute(NCI, CA233774). Preprint of an article submitted for consideration in Pacific Symposium onBiocomputing 2020 World Scientific Publishing Company . References
1. Reik W.
Stability and flexibility of epigenetic gene regulation in mammalian development . Nature,vol. 447(7143):pp. 425–32, 2007.2. Cobb JP, et al. Application of genome-wide expression analysis to human health and disease .Proc Natl Acad Sci U S A, vol. 102(13):pp. 4801–6, 2005.3. King MC et al. Evolution at two levels in humans and chimpanzees . Science, vol. 188(4184):pp.107–116, 1975.24. Singh KP, et al. Mechanisms and Measurement of Changes in Gene Expression . Biol Res Nurs,vol. 20(4):pp. 369–382, 2018.5. Bosinger SE, et al. Gene expression profiling of host response in models of acute HIV infection .J. Immunol., vol. 173(11):pp. 6858–6863, 2004.6. Kober K, et al. Differential Methylation and Expression of Genes in the Hypoxia Inducible Factor1 (HIF-1) Signaling Pathway Are Associated With Paclitaxel-Induced Peripheral Neuropathy inBreast Cancer Survivors and with Preclinical Models of Chemotherapy-Induced Neuropathic Pain .Mol Pain, vol. 16:p. 1744806920936502, 2020.7. Stephens KE, et al. Epigenetic regulation and measurement of epigenetic changes . Biol Res Nurs,vol. 15(4):pp. 373–381, 2013.8. Razin A et al. DNA methylation and gene function . Science, vol. 210(4470):pp. 604–610, 1980.9. Eden S et al. Role of DNA methylation in the regulation of transcription . Curr Opin Genet Dev,vol. 4(2):pp. 255–9, 1994.10. Spruijt CG et al. DNA methylation: old dog, new tricks?
Nat Struct Mol Biol, vol. 21(11):pp.949–54, 2014.11. Anderson DF, et al. On classes of reaction networks and their associated polynomial dynamicalsystems . Journal of Mathematical Chemistry, 2020.12. Agarwal V et al. Predicting mRNA Abundance Directly from Genomic Sequence Using DeepConvolutional Neural Networks . Cell Rep, vol. 31(7):p. 107663, 2020.13. Zhong H, et al. Predicting gene expression using DNA methylation in three human populations .PeerJ, vol. 7:p. e6757, 2019.14. Ebert P, et al. Epigenome-based prediction of gene expression across species . bioRxiv, 2018.15. Klett H, et al. Robust prediction of gene regulation in colorectal cancer tissues from DNA methy-lation profiles . Epigenetics, vol. 13(4):pp. 386–397, 2018.16. Li J, et al. Using epigenomics data to predict gene expression in lung cancer . BMC Bioinformatics,vol. 16 Suppl 5:p. S10, 2015.17. Davis MH.
Piecewise-deterministic Markov processes: a general class of non-diffusion stochasticmodels . Journal of the Royal Statistical Society: Series B (Methodological), vol. 46(3):pp. 353–376, 1984.18. Zeiser S, et al. Simulation of genetic networks modelled by piecewise deterministic Markov pro-cesses . IET systems biology, vol. 2(3):pp. 113–135, 2008.19. Crudu A, et al. Hybrid stochastic simplifications for multiscale gene networks . BMC systemsbiology, vol. 3(1):p. 89, 2009.20. Anderson DF et al. Stochastic analysis of biochemical systems , vol. 1. Springer, 2015.21. Anderson DF, et al. Product-form stationary distributions for deficiency zero chemical reactionnetworks . Bulletin of mathematical biology, vol. 72(8):pp. 1947–1970, 2010.22. Wang Y, et al. Parameter inference for discretely observed stochastic kinetic models using stochas-tic gradient descent . BMC systems biology, vol. 4(1):p. 99, 2010.23. Aza¨ıs R, et al. Piecewise deterministic Markov processrecent results . In
ESAIM: Proceedings ,vol. 44, pp. 276–290. EDP Sciences, 2014.24. Anderson DF.
A modified next reaction method for simulating chemical systems with time de-pendent propensities and delays . The Journal of chemical physics, vol. 127(21):p. 214107, 2007.25. Garcia-Alonso L, et al. Transcription Factor Activities Enhance Markers of Drug Sensitivity inCancer . Cancer Research, vol. 78(3):pp. 769–780, 2018.26. Varley KE, et al. Dynamic DNA methylation across diverse human cell lines and tissues . GenomeRes., vol. 23(3):pp. 555–567, 2013.27. Breen MS, et al. PTSD Blood Transcriptome Mega-Analysis: Shared Inflammatory Pathwaysacross Biological Sex and Modes of Trauma . Neuropsychopharmacology, vol. 43(3):pp. 469–481,2018.328. Yates B, et al. Genenames.org: the HGNC and VGNC resources in 2017et al. Genenames.org: the HGNC and VGNC resources in 2017