Gene regulatory network inference using sparse probabilistic models

The main task of systems biology is to uncover mechanisms that regulate complex processes that take place in biological cells, especially the mechanisms of gene regulation. This project aims to identify gene regulatory interactions taking place in the early development of neural tube. Solutions proposed in this work for identification of transcription factors and their target genes are mostly based on Bayesian sparse linear models. Sparse linear regression is used to predict gene expression given a set of transcription factor expressions. Even though transcriptional regulation mechanisms are frequently non-linear, a sparse linear regression model succeeds in identifying some of the transcription factors that are already known to play role in the studied biological processes. The probabilistic sparse model is compared to its frequentist equivalent, the lasso. Both algorithms are evaluated on neural tube development data and Bayesian model produces superior results. The probabilistic sparse linear model is also extended to take into account non-linear interactions between transcription factors. A simple interaction model is presented and it is integrated into the probabilistic sparse linear model. The resulting model is a novel method to incorporate non-linearity into linear regression. Variational approximate inference algorithm is derived for the extended model with transcription factor interactions. This model is then applied to a subset of potentially important transcription factors and it succeeds to find a very sparse solution. Again, the results from the extended model succeed to identify transcription factors that seem to have important role. Both probabilistic models for gene regulation provide a set of straightforwardly testable predictions of potentially interesting transcription factors.