State-space modeling of dynamic genetic networks

TL;DR

This paper introduces a dynamic state space modeling approach using an EM algorithm with Kalman smoothing to infer gene regulatory networks from high-throughput time course microarray data, revealing key regulatory genes and interactions.

Contribution

It proposes a novel state space model with an EM algorithm for reverse engineering transcriptional networks from microarray data, including hidden states and model selection.

Findings

Identified 4 hidden states in T-cell data.

Detected key regulatory genes like FYB, CCNA2, AKT1, and JUNB.

Discovered specific gene interactions such as Jun-B with SMN1.

Abstract

The genomic reality is a highly complex and dynamic system. The recent development of high-throughput technologies has enabled researchers to measure the abundance of many genes (in the order of thousands) simultaneously. The challenge is to unravel from such measurements, gene/protein or gene/gene or protein/ protein interactions and key biological features of cellular systems. Our goal is to devise a method for inferring transcriptional or gene regulatory networks from high-throughput data sources such as gene expression microarrays with potentially hidden states, such as unmeasured transcription factors (TFs), which satisfies certain Markov properties. We propose a dynamic state space representation. Our method is based on an EM algorithm with an incorporated Kalman smoothing algorithm in the E-step, a bootstrap for confidence intervals to infer the networks and the AIC for model selection. The state space model is an approach with proven effectiveness to reverse engineer transcriptional networks. The proposed method is applied to time course microarray data obtained from well established T-cell. When we applied the method to the T-cell data, we obtained 4, as the optimum number of hidden states. Our results support interesting biological properties in the family of Jun genes. The following genes were mostly seen as regulatory genes. These genes includes FYB, CCNA2, AKT1, TRAF5, CASP4, and CTNNB1. We found interaction between Jun-B and SMN1, and CDC2 activates Jun-D. We found few significant interactions or one-to-one correspondence among the 4 putative transcription factors. Among the important key genes in terms of outward-directed edges, we found genes such as CCNA2, JUNB, CDC2, CASP4, JUND to have a high degree of connectivity. R Computer source code is made available at our website at http://www.math.rug.nl/stat/Main/Software.