Current genomic deep learning models display decreased performance in cell type specific accessible regions.

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Deep learning models predict epigenetic features like chromatin accessibility from DNA sequence.
• Regulatory elements (CREs) are crucial for gene regulation but form a small genomic fraction.
• Cell type-specific CREs are vital for complex disease heritability.

Purpose of the Study:

• Evaluate genomic deep learning models in cell type-specific chromatin accessibility regions.
• Compare general-purpose models with tissue/task-specific models.
• Identify strategies to improve model performance in functionally important genomic regions.

Main Methods:

• Assessed performance of Enformer and Sei models across varying cell type specificities.
• Compared general-purpose models with single-task and high-capacity multi-task models.
• Analyzed model performance in general genome-wide and specific accessible regions.

Main Results:

• Genomic deep learning model accuracy varies across the genome, decreasing in cell type-specific accessible regions.
• Single-task learning and high-capacity multi-task models improve performance in cell type-specific regions.
• Improved reference sequence predictions did not consistently enhance variant effect predictions.

Conclusions:

• Genomic deep learning model performance is reduced in cell type-specific accessible regions.
• Strategies like increasing model capacity can enhance performance in these key regulatory regions.
• Novel approaches are needed to improve variant effect prediction accuracy.