Switched Dynamical Latent Force Models for Modelling Transcriptional Regulation

In order to develop statistical approaches for transcription networks, statistical community has proposed several methods to infer activity levels of proteins, from time-series measurements of targets' expression levels. A few number of approaches have been proposed in order to outperform the representation of fast switching time instants, but computational overheads are significant due to complex inference algorithms. Using the theory related to latent force models (LFM), the development of this project provide a switched dynamical hybrid model based on Gaussian processes (GPs). To deal with discontinuities in dynamical systems (or latent driving force), an extension of the single input motif approach is introduced, that switches between different protein concentrations, and different dynamical systems. This creates a versatile representation for transcription networks that can capture discrete changes and non-linearities in the dynamics. The proposed method is evaluated on both simulated data and real data, concluding that our framework provides a computationally efficient statistical inference module of continuous-time concentration profiles, and allows an easy estimation of the associated model parameters.