Interpreting artificial neural networks to detect genome-wide association signals for complex traits

Area of Science:

• Genetics
• Computational Biology
• Bioinformatics

Background:

• Investigating complex diseases is difficult due to combined genetic and environmental factors.
• Genome-Wide Association Studies (GWAS) have identified many variants but are limited by assumptions like linearity and epistasis.
• New computational methods are needed to overcome limitations of traditional genetic association studies.

Purpose of the Study:

• To apply artificial neural networks (ANNs) to genome-wide genotype data for complex trait prediction.
• To identify potentially associated loci (PAL) for complex traits using feature importance scores from ANNs.
• To develop a method for estimating P-values for ANN-identified PAL and validate the approach on real-world data.

Main Methods:

• Trained ANNs on genome-wide genotype data to predict simulated and real complex traits.
• Utilized post hoc interpretability methods to extract feature importance scores for PAL identification.
• Developed a P-value estimation approach for detected PAL and applied it to a schizophrenia cohort.

Main Results:

• Simulations confirmed precise detection of associated loci using strict criteria.
• ANN analysis identified multiple loci in the Estonian Biobank schizophrenia cohort missed by linear methods.
• Identified PAL showed concordance with known schizophrenia and bipolar disorder loci, enriched for brain morphology and function genes.

Conclusions:

• Artificial neural networks offer a powerful, comprehensive approach to enhance complex disease gene discovery.
• ANNs can identify novel genomic loci beyond the scope of conventional GWAS.
• Further advancements in ANNs hold significant potential for improving the identification of genetic associations in complex diseases.