Experimental assessment of static and dynamic algorithms for gene regulation inference from time series expression data

Miguel Lopes,Gianluca Bontempi

doi:10.3389/fgene.2013.00303

Abstract

Accurate inference of causal gene regulatory networks from gene expression data is an open bioinformatics challenge. Gene interactions are dynamical processes and consequently we can expect that the effect of any regulation action occurs after a certain temporal lag. However such lag is unknown a priori and temporal aspects require specific inference algorithms. In this paper we aim to assess the impact of taking into consideration temporal aspects on the final accuracy of the inference procedure. In particular we will compare the accuracy of static algorithms, where no dynamic aspect is considered, to that of fixed lag and adaptive lag algorithms in three inference tasks from microarray expression data. Experimental results show that network inference algorithms that take dynamics into account perform consistently better than static ones, once the considered lags are properly chosen. However, no individual algorithm stands out in all three inference tasks, and the challenging nature of network inference tasks is evidenced, as a large number of the assessed algorithms does not perform better than random.

Highlights

The measurement of gene expression levels, by using microarrays or high throughput technologies, makes it possible to infer statistical dependencies between the expression of two genes
Results obtained using three different datasets show that dynamic models perform better on the inference of gene regulatory interactions from time series, than static models such as Bayesian networks
The advantage of the adaptive lag models is that they can potentially infer interactions characterized by higher and variable lags. Their performance depends on the maximum allowed lag, lmax, and care should be taken when defining this parameter: if it is set to an unrealistic high value, in the range of many hours, eventually interactions will be estimated at that range, hurting the network inference performance

Summary

Introduction

The measurement of gene expression levels, by using microarrays or high throughput technologies, makes it possible to infer statistical dependencies (e.g., correlations) between the expression of two genes. In other terms if two genes are part of a regulatory interaction, their expression levels over time are expected to be correlated with a certain lag and the time order is expected to elucidate the respective promoter/target roles. Such lag is unknown a priori and should be properly learned from data. The adoption of temporal dynamic models makes the learner less biased but necessarily more exposed to high variance In spite of this intuition, and there are some comparisons between dynamic and static methods in the literature on gene regulatory networks, these are not systematic or extensive

Objectives

Methods

Results

Discussion

Conclusion