Uncovering tissue-specific binding features from differential deep learning

Abstract

Motivation: Transcription factors (TFs) can bind DNA in a cooperative manner, enabling a mutual increase in occupancy. Through this type of interaction, alternative binding sites can be preferentially bound in different tissues to regulate tissue-specific expression programmes. Recently, deep learning models have become state-of-the-art in various pattern analysis tasks, including applications in the field of genomics. We therefore investigate the application of convolutional neural network (CNN) models to the discovery of sequence features determining cooperative and differential TF binding across tissues. Results: We analyse ChIP-seq data from MEIS, TFs which are broadly expressed across mouse branchial arches, and HOXA2, which is expressed in the second and more posterior branchial arches. By developing models predictive of MEIS differential binding in all three tissues we are able to accurately predict HOXA2 co-binding sites. We evaluate transfer-like and multitask approaches to regularising the high-dimensional classification task with a larger regression dataset, allowing for creation of deeper and more accurate models. We test the performance of perturbation and gradient-based attribution methods in identifying the HOXA2 sites from differential MEIS data. Our results show that deep regularised models significantly outperform shallow CNNs as well as k-mer methods in the discovery of tissue-specific sites bound in vivo. Availability: For implementation and models please visit https://doi.org/10.5281/zenodo.2635463.

HOX paralogs selectively convert binding of ubiquitous transcription factors into tissue-specific patterns of enhancer activation

“…We systematically extracted differential MEIS binding across the BAs (Fig. 6-Supplemental Fig.1) and found, using convolutional neural network (CNN) models, that differential classification of MEIS binding is sufficient to uncover HOX motif features (Phuycharoen et al, 2019); specifically, the fraction of MEIS peaks higher in BA2 and in PBA (= lower BA1) is highly enriched in sequence features matching HOX-PBX motif (Fig. 6B).…”

“…Out of 215830 MEIS peaks, 101055 are in common between the three tissues; MEIS peaks were combined and re-centered using DiffBind. B. CNN models of MEIS differential peaks uncover enrichment of tissue-specific sequence motifs as described in (Phuycharoen et al, 2019). MEIS binding was classified in six categories (i.e.…”

“…MEIS differential sequence features are detected by recently published differential convolutional neural network (CNN) structure (Phuycharoen et al, 2019). For in silico binding site knockout we trained a convolutional neural network (CNN) model for multitask regression of MEIS and HOX RPKM binding level.…”

“…For in silico binding site knockout we trained a convolutional neural network (CNN) model for multitask regression of MEIS and HOX RPKM binding level. The CNN was trained by transfer learning, using convolution parameters from a previously published 1-convolutional layer MEIS RPKM model (Phuycharoen et al, 2019). Convolutional filters were transferred to a new model, which was then trained on a subset of MEIS regions also bound by HOX, to simultaneously predict log2RPKM values in 2 replicates of Hoxa2 in BA2, 2 from Hoxa3 in PBA, and one replicate of MEIS in BA1, BA2 and PBA.…”