Predicting cell type-specific epigenomic profiles accounting for distal genetic effects

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Interpreting genome-wide association studies (GWAS) requires understanding how genetic variants influence the epigenome.
• Profiling these effects across the non-coding genome is experimentally challenging due to scalability issues.
• Current machine learning models are limited to the cell types they were trained on.

Purpose of the Study:

• To develop a computational model that can predict epigenetic signals in previously unseen cell types.
• To incorporate distal DNA interaction effects for more accurate epigenetic imputation.
• To improve the generalizability of deep learning models for genomic predictions.

Main Methods:

• Introduced Enformer Celltyping, a deep learning model integrating distal DNA interaction effects (up to 100,000 base pairs).
• Utilized DNA and chromatin accessibility data for epigenetic imputation.
• Developed a framework for evaluating genetic variant effect prediction models using regulatory quantitative trait loci (eQTL) mapping studies.

Main Results:

• Enformer Celltyping outperforms existing best-in-class approaches in predicting epigenetic signals.
• The model demonstrates generalizability across diverse cell types and genomic regions.
• The evaluation framework highlighted limitations in current genomic deep learning models.

Conclusions:

• Enformer Celltyping offers a scalable and accurate method for epigenetic imputation in novel cell types.
• The model advances the interpretation of genetic variants in the non-coding genome and GWAS.
• It provides a tool for studying cell type-specific genetic enrichment in complex traits.