Characterizing collaborative transcription regulation with a graph-based deep learning approach.

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Current deep learning models for epigenome analysis rely solely on DNA sequences.
• Transcription factor (TF) collaboration and regulatory activities extend beyond linear DNA, involving 3D chromatin organization.
• Integrating 3D chromatin structure is crucial for understanding complex TF interactions.

Purpose of the Study:

• To develop a novel graph-based neural network, ECHO, for predicting chromatin features.
• To leverage 3D chromatin organization data from Micro-C contact maps to characterize TF collaborations.
• To improve the prediction accuracy of chromatin features compared to existing sequence-based methods.

Main Methods:

• Developed ECHO, a graph neural network incorporating 200-bp high-resolution Micro-C contact maps.
• Utilized gradient-based attribution methods for model interpretability.
• Analyzed the impact of chromatin contacts at different distances on feature prediction.

Main Results:

• ECHO predicted 2,583 chromatin features with significantly higher AUROC and AUPR than sequence-based models.
• Demonstrated that chromatin contacts at varying distances differentially impact chromatin feature prediction.
• Identified key TF binding motifs, collaborative binding events, and relevant neighboring sequence patterns through attribution analysis.

Conclusions:

• ECHO effectively integrates 3D chromatin organization for superior epigenome prediction.
• The model provides insights into complex, divergent TF collaborative regulatory mechanisms.
• ECHO's interpretability facilitates the identification of critical genomic elements and interactions driving chromatin features.