Hierarchical Gaussian Process Models of Gene Expression and Transcriptional Regulation
Biological systems are inherently dynamic and time series data provide great insight to understanding them. Such data are most naturally modelled in continuous-time framework that can be directly applied to data with diverse or uneven sampling. Gaussian processes provide a convenient tool for specifying priors over latent continuous-time functions in such models. Such models have previously been proposed for example with a differential equation model of gene regulation. We propose extending these models with a hierarchical Gaussian process that allows modelling diverse experimental setups, such as mixed longitudinal/cross-sectional designs and phylogenetic structure. In the linear differential equation model, this approach can improve performance over our previous work (Honkela et al., PNAS 2010) even in simple transcription factor target ranking, and provides flexibility for modelling more complex data, such as the multi-species Drosophila data of Kalinka et al. (Nature 2010).