Model-based method for transcription factor target identification with limited data

Abstract

We present a computational method for identifying potential targets of a transcription factor (TF) using wild-type gene expression time series data. For each putative target gene we fit a simple differential equation model of transcriptional regulation, and the model likelihood serves as a score to rank targets. The expression profile of the TF is modeled as a sample from a Gaussian process prior distribution that is integrated out using a nonparametric Bayesian procedure. This results in a parsimonious model with relatively few parameters that can be applied to short time series datasets without noticeable overfitting. We assess our method using genome-wide chromatin immunoprecipitation (ChIP-chip) and lossof-function mutant expression data for two TFs, Twist, and Mef2, controlling mesoderm development in Drosophila. Lists of topranked genes identified by our method are significantly enriched for genes close to bound regions identified in the ChIP-chip data and for genes that are differentially expressed in loss-of-function mutants. Targets of Twist display diverse expression profiles, and in this case a model-based approach performs significantly better than scoring based on correlation with TF expression. Our approach is found to be comparable or superior to ranking based on mutant differential expression scores. Also, we show how integrating complementary wild-type spatial expression data can further improve target ranking performance.Bayesian inference | Gaussian process inference | gene regulation | gene regulatory network | systems biology T ranscription factors are key nodes in the gene regulatory networks that determine the function and fate of cells. An important first step in uncovering a gene regulatory network is the identification of target genes regulated by a specific transcription factor (TF). A common approach to this problem is to experimentally locate physical binding of TF proteins to the DNA sequence in vivo using a genome-wide chromatin immunoprecipitation (ChIP) experiment (1, 2). However, recent studies suggest that many observed binding events are neutral and do not regulate transcription, while regulatory binding events often occur at enhancers that are not proximal to the target gene that they control (3, 4). Therefore, the task of identifying transcriptional targets requires the integration of ChIP binding predictions with evidence from expression data to help associate binding events with target gene regulation. If there is access to expression data from a mutant in which the TF has been knocked out or overexpressed, then differential expression of genes between wild type and mutant is indicative of a potential regulatory interaction (5, 6). Available spatial expression data for the TF and the putative target can also provide support for a hypothesized regulatory link.A problem with the above approach is that the creation of mutant strains is challenging or impossible for many TFs of interest. Even when available, mutants may provide very limited information because of redundancy or...

Ranking of gene regulators through differential equations and Gaussian processes

“…In other words we are interested in ranked regulator prediction instead of ranked target prediction. The proposed approach is otherwise fairly similar with respect to the methods to that in [3].…”

“…For a given TF this likelihood can be measured for all potential target genes and they can then be ranked as putative targets. This idea was exploited by [3] who validated their results using ChIP data and were able to show that model-based approaches can do considerably better than simple correlationbased approaches.…”

Identifying latent dynamic components in biological systems

“…Normally distributed error terms, as in (10), imply that the noise level is independent of the magnitude of the measurements. Although this assumption is commonly made, it is not appropriate for all biological applications, for example, when concentrations are measured over time.…”

Biophysically Motivated Regulatory Network Inference: Progress and Prospects

“…The most commonly used differential equation models are ordinary differential equation and stochastic differential equation models where the TF effect on the transcription rate is learned using a greedy search (one target gene at a time). Several challenges remain for learning dynamic models including (1) treating the parameterization of these large networks as a proper global system by simultaneously fitting all parameters [50] , (2) modeling latent states such as TF activity [51,52] , (3) explicitly modeling activator, repressor [53] , degradation, and target expression with distinct biophysically correct distributions, and (4) determining correct methods for using these models to design optimal experiments.…”