A Partial Granger Causality Approach to Explore Causal Networks Derived From Multi-parameter Data

Abstract

Background: Inference and understanding of gene networks from experimental data is an important but complex problem in molecular biology. Mapping of gene pathways typically involves inferences arising from various studies performed on individual pathway components. Although pathways are often conceptualized as distinct entities, it is often understood that inter-pathway cross-talk and other properties of networks reflect underlying complexities that cannot by explained by consideration of individual pathways in isolation. In order to consider interaction between individual paths, a global multivariate approach is required. In this paper, we propose an extended form of Granger causality can be used to infer interactions between sets of time series data.

Results: We successfully tested our method on several artificial datasets, each one depicting various possibilities of connections among the participating entities. We also demonstrate the ability of our method to deal with latent and exogenous variables present in the system. We then applied this method to a highly replicated gene expression microarray time series data to infer causal influences between gene expression events involved in activation of human T-cells. The application of our method to the T-cell dataset revealed a set of strong causal links between the participating genes, with many links already experimentally verified and reported in the biological literature.

Conclusions: We have proposed a novel form of Granger causality to reverse-engineer a causal network structure from a time series dataset involving multiple entities. We have extensively and successfully tested our method on synthesized as well as real time series microarray data.