EpiGePT: a Pretrained Transformer model for epigenomics.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Large language models (LLMs) show promise for genomics due to similarities between natural language and biological sequences.
• Current LLMs for genomics have limitations, including the exclusion of chromatin interactions and inability to predict in novel cellular contexts.

Purpose of the Study:

• To develop EpiGePT, a transformer-based pretrained language model for predicting context-specific epigenomic signals and chromatin contacts.
• To overcome limitations of existing models by incorporating transcription factor (TF) activities and 3D genome interactions.

Main Methods:

• EpiGePT utilizes a transformer architecture, a type of LLM, pretrained on genomic data.
• The model integrates context-specific TF activities and 3D genome interaction data into its training.
• Performance was evaluated on diverse epigenomic signal prediction tasks against existing methods.

Main Results:

• EpiGePT demonstrated superior performance in predicting various epigenomic signals compared to current methods.
• The model successfully enabled cross-cell-type prediction of long-range chromatin interactions.
• EpiGePT provided insights into the functional impact of genetic variants across different cellular contexts.

Conclusions:

• EpiGePT enhances the utility of LLMs in genomics by enabling context-specific predictions and incorporating crucial biological information.
• The model's ability to predict chromatin interactions and variant effects in new contexts advances the study of gene regulatory mechanisms.