EpiGePT: a pretrained transformer-based language model for context-specific human epigenomics.

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Large language models (LLMs) show promise in genomics due to similarities between natural language and biological sequences.
• Existing LLMs face limitations in predicting epigenomic signals, particularly concerning chromatin interactions and unseen cellular contexts.
• Understanding cell-type-specific gene regulation is crucial for deciphering complex biological processes.

Purpose of the Study:

• To develop a novel transformer-based model, EpiGePT, for accurate prediction of context-specific human epigenomic signals.
• To enhance the predictive power of LLMs in genomics by incorporating 3D genome architecture and transcription factor dynamics.
• To improve the prediction of genetic variant impacts on gene regulation in diverse cellular environments.

Main Methods:

• Developed EpiGePT, a transformer architecture integrating transcription factor activities and 3D genome interaction data.
• Trained and evaluated EpiGePT on human epigenomic datasets, focusing on cell-type-specific signal prediction.
• Compared EpiGePT's performance against existing state-of-the-art methods for epigenomic prediction tasks.

Main Results:

• EpiGePT demonstrates superior performance in predicting context-specific human epigenomic signals compared to current methods.
• The model excels in predicting cell-type-specific long-range chromatin interactions.
• EpiGePT accurately predicts the impact of genetic variants on epigenomic regulation.

Conclusions:

• EpiGePT advances the prediction of epigenomic signals by effectively integrating diverse genomic data, including chromatin interactions.
• The model offers a powerful tool for understanding cell-type-specific gene regulation and the functional consequences of genetic variants.
• The developed prediction service provides a valuable resource for the genomics and epigenetics research community.