Uncovering tissue-specific binding features from differential deep learning.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Transcription factors (TFs) bind DNA cooperatively, influencing tissue-specific gene expression.
• Deep learning models excel at pattern recognition in genomics data.

Purpose of the Study:

• To apply convolutional neural network (CNN) models for discovering sequence features that dictate cooperative and differential TF binding across tissues.
• To predict HOXA2 co-binding sites using models trained on MEIS TF binding data.

Main Methods:

• Analysis of ChIP-seq data for MEIS and HOXA2 TFs in mouse branchial arches.
• Development of deep regularized CNN models for predicting TF binding.
• Evaluation of transfer learning and multitask learning approaches.
• Application of perturbation and gradient-based attribution methods.

Main Results:

• Deep regularized CNN models accurately predict HOXA2 co-binding sites based on MEIS binding data.
• These models significantly outperform shallow CNNs and k-mer methods.
• Identified sequence features driving tissue-specific TF binding.

Conclusions:

• Deep learning, specifically regularized CNNs, offers a powerful approach for identifying tissue-specific TF binding sites.
• This method enhances understanding of gene regulation mechanisms and improves prediction accuracy over existing techniques.