A chromatin-structure-guided framework for predictive and interpretable regulatory genomics

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Gene regulation is influenced by 3D chromatin organization, but linear genome models overlook this.
• Existing Hi-C data provides genome-wide conformation but is averaged, biased, and noisy, obscuring specific interactions.

Purpose of the Study:

• To develop a computational framework (CHROME) to identify physically specific chromatin contacts from Hi-C data.
• To integrate these validated contacts into graph representations for improved predictive genomics.
• To enhance the prediction of gene regulation and variant pathogenicity by incorporating 3D genome structure.

Main Methods:

• Developed CHROME, a framework using a self-avoiding polymer ensemble null model to identify nonrandom Hi-C contacts.
• Integrated identified contacts into graph attention networks with sequence and accessibility data.
• Trained models to predict cell-line-specific ChIP-seq profiles, eQTLs, and variant pathogenicity.

Main Results:

• CHROME successfully identifies physically specific chromatin contacts, overcoming Hi-C data limitations.
• Graph representations incorporating these contacts improved prediction of cell-line-specific ChIP-seq profiles over local methods.
• CHROME embeddings enhanced predictions for tissue-specific eQTLs and ClinVar variant pathogenicity, showing cross-cell-type transfer potential.

Conclusions:

• Incorporating physically validated 3D chromatin interactions significantly improves predictive genomics models.
• CHROME offers a method to leverage 3D genome structure for better regulatory prediction and variant interpretation.
• The framework provides interpretability, revealing long-range influences on local gene activity.