DeepCOMBI: explainable artificial intelligence for the analysis and discovery in genome-wide association studies

Area of Science:

• Genetics
• Artificial Intelligence
• Bioinformatics

Background:

• Deep learning significantly enhances prediction performance in data science.
• Explainable AI (XAI) extracts knowledge from deep learning models by interpreting results.
• Genome-wide association studies (GWAS) aim to identify genetic variants associated with phenotypes.

Purpose of the Study:

• To explore genetic architectures of phenotypes using XAI in GWAS.
• To introduce DeepCOMBI, a novel three-step algorithm for genetic architecture exploration.
• To improve the power and precision of identifying genetic associations.

Main Methods:

• DeepCOMBI trains a neural network for phenotype classification.
• Layer-wise relevance propagation is used to explain classifier decisions.
• Importance scores identify relevant genomic locations for hypothesis testing.

Main Results:

• DeepCOMBI outperforms traditional P-value thresholding and baseline methods in power and precision.
• Two novel disease associations were identified: rs10889923 for hypertension and rs4769283 for type 1 diabetes.
• Findings were validated using independent studies published up to 2020.

Conclusions:

• DeepCOMBI offers a powerful approach for exploring genetic architectures in GWAS.
• The integration of XAI enhances the interpretability and utility of deep learning in genetics.
• This method successfully identified novel genetic associations, advancing disease research.