A generalizable framework to comprehensively predict epigenome, chromatin organization, and transcriptome.

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Deep learning models predict genomic features but lack generalizability across tasks and cell types.
• Predicting multiple genomic modalities (epigenome, transcriptome, chromatin organization, enhancer activity) is challenging.

Purpose of the Study:

• To develop a generalizable deep learning framework (EPCOT) for predicting multiple genomic modalities.
• To enable accurate in silico prediction of expensive experimental data like Micro-C and ChIA-PET.
• To uncover biological insights from interpretable genomic representations.

Main Methods:

• EPCOT utilizes a pre-training and fine-tuning framework.
• The model requires only cell-type specific chromatin accessibility profiles as input.
• It learns generic representations applicable across diverse predictive tasks.

Main Results:

• EPCOT accurately predicts multiple genomic modalities for new cell types.
• The framework demonstrates generalizability across different predictive tasks.
• Model interpretation reveals insights into genomic modality mapping, TF binding, and TF impact on enhancer activity.

Conclusions:

• EPCOT provides a powerful and generalizable deep learning approach for predicting complex genomic information.
• The method reduces the need for costly experiments by enabling accurate in silico predictions.
• Interpreting EPCOT facilitates a deeper understanding of gene regulation and cell-type specific genomic functions.