Abstract
Due to the large numbers of transcription factors (TFs) and cell types, querying binding profiles of all TF/cell type pairs is not experimentally feasible, owing to constraints in time and resources. To address this issue, we developed a convolutional-recurrent neural network model, called FactorNet, to computationally impute the missing binding data. FactorNet trains on binding data from reference cell types to make accurate predictions on testing cell types by leveraging a variety of features, including genomic sequences, genome annotations, gene expression, and single-nucleotide resolution sequential signals, such as DNase I cleavage. To the best of our knowledge, this is the first deep learning method to study the rules governing TF binding at such a fine resolution. With FactorNet, a researcher can perform a single sequencing assay, such as DNase-seq, on a cell type and computationally impute dozens of TF binding profiles. This is an integral step for reconstructing the complex networks underlying gene regulation. While neural networks can be computationally expensive to train, we introduce several novel strategies to significantly reduce the overhead. By visualizing the neural network models, we can interpret how the model predicts binding which in turn reveals additional insights into regulatory grammar. We also investigate the variables that affect cross-cell type predictive performance to explain why the model performs better on some TF/cell types than others, and offer insights to improve upon this field. Our method ranked among the top four teams in the ENCODE-DREAM in vivo Transcription Factor Binding Site Prediction Challenge.
