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 from DNA sequence.
• Cis-regulatory elements (CREs) are crucial for gene regulation but form a small genomic fraction.
• Cell type-specific CREs harbor significant complex disease heritability.

Purpose of the Study:

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

Main Methods:

• Assessed deep learning model performance across varying cell type specificity in accessible regions.
• Compared general-purpose models (Enformer, Sei) with tailored models.
• Investigated impact of model capacity and training strategies on performance.

Main Results:

• Model accuracy varies genome-wide, decreasing in cell type-specific accessible regions.
• Increased model capacity (single-task or high-capacity multi-task) improves performance in these regions.
• Reference sequence prediction improvements do not consistently enhance variant effect predictions.

Conclusions:

• Genomic deep learning model performance is not uniform across the genome.
• Performance is notably reduced in cell type-specific accessible regions.
• Strategies exist to enhance model performance for cell type-specific regulatory elements.