CS700:Graduate Seminar in Computer Science & Informatics
Olgert Denas, Department of Mathematics and Computer Science
Recent advances in sequencing technologies are producing complementary measurements at large numbers of genome locations, cell types, and time points. These data have made possible the global study of complex cell processes such as gene regulation. However, they also demand computational models that make a small number of assumptions on the generating mechanisms, scale well with data dimensionality, and provide a reasonably interpretable output.
We propose and motivate here deep artificial neural networks (ANNs) as models for gene regulation. We then provide a systematic way of extracting knowledge from the trained model in an interpretable way. We validate our model on gene regulatory processes during Erythropoiesis and uncover a number of global regulatory mechanisms and interplays between the participating TFs supported by recent empirical studies.
While powerful, ANNs are usually considered non-trivial to train and hard to interpret. Hopefully, our analysis and implementation has shown the viability of such models and will be an incentive for their wider adoption.
Code available on http://bitbucket.org/gertidenas/dimer