Evaluating the informativeness of deep learning annotations for human complex diseases.

Area of Science:

• Genomics and Computational Biology
• Human Complex Disease Genetics

Background:

• Deep learning models excel at predicting regulatory effects from DNA sequences.
• The utility of these models for understanding human complex diseases remains unclear.
• Genome-wide SNP annotations are crucial for linking genetic variation to disease.

Purpose of the Study:

• To evaluate the informativeness of deep learning-derived SNP annotations for human complex diseases.
• To assess the performance of DeepSEA and Basenji models in predicting disease heritability.
• To determine if regulatory prediction accuracy correlates with disease relevance.

Main Methods:

• Applied stratified LD score regression to 41 diseases and traits (average N=320K).
• Utilized genome-wide SNP annotations from DeepSEA and Basenji models.
• Aggregated annotations across tissues/cell-types for meta-analyses.

Main Results:

• Deep learning annotations were highly enriched for disease heritability.
• Limited conditionally significant results were observed, with non-tissue-specific and brain-specific Basenji-H3K4me3 showing some significance.
• Conditional informativeness for disease was not directly inferred from regulatory prediction accuracy.

Conclusions:

• Deep learning models have not yet reached their full potential for complex disease insights.
• Further development is needed to enhance the unique information deep learning provides for disease genetics.
• Regulatory prediction accuracy alone is insufficient to gauge a model's value for complex disease.