Multi-scale inference of genetic trait architecture using biologically annotated neural networks

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Genome-wide association (GWA) studies are crucial for identifying genetic variants associated with complex traits.
• Current methods for association mapping and enrichment analysis often operate independently.
• Integrating biological knowledge into analytical frameworks can improve the power and interpretability of GWA studies.

Purpose of the Study:

• To introduce Biologically Annotated Neural Networks (BANNs), a novel nonlinear probabilistic framework for association mapping in GWA studies.
• To develop a method that allows simultaneous SNP-level mapping and SNP-set enrichment analyses.
• To create a fully interpretable neural network model that leverages biological annotations.

Main Methods:

• Developed BANNs, a feedforward neural network with partially connected architectures based on biological annotations.
• Modeled SNP-level effects and aggregated SNP-set effects using interpretable layers.
• Treated network weights and connections as random variables with prior distributions reflecting genetic effects at different genomic scales.
• Employed variational inference for posterior summaries to enable simultaneous mapping and enrichment analyses.

Main Results:

• Simulations demonstrated that BANNs outperform state-of-the-art association mapping and enrichment approaches across diverse genetic architectures.
• BANNs were successfully applied to real GWA data from mouse models and human cohorts (Framingham Heart Study).
• The framework replicated known associations for high and low-density lipoprotein cholesterol content in UK Biobank data.

Conclusions:

• BANNs provide a powerful and interpretable framework for dissecting the genetic architecture of complex traits.
• The simultaneous mapping and enrichment capabilities of BANNs offer a significant advancement in GWA study analysis.
• BANNs demonstrate broad applicability and effectiveness in both simulated and real-world genetic datasets.