GenNet framework: interpretable deep learning for predicting phenotypes from genetic data.

Area of Science:

• Population genomics
• Computational biology
• Bioinformatics

Background:

• Deep learning application in population genomics is hindered by computational demands and a lack of model interpretability.
• Existing methods often lack biological context, limiting the discovery of meaningful genetic associations.
• Interpretable models are crucial for understanding the genetic architecture of complex traits and diseases.

Purpose of the Study:

• To introduce GenNet, an open-source deep learning framework designed for phenotype prediction from genetic variants.
• To develop interpretable and memory-efficient neural network architectures by integrating biological knowledge.
• To enable researchers to gain novel insights into the genetic basis of complex traits and diseases.

Main Methods:

• GenNet constructs neural network architectures by embedding biological knowledge from public databases, ensuring biologically plausible connections.
• The framework utilizes interpretable and memory-efficient neural network designs.
• The framework was applied to seventeen diverse phenotypes for validation.

Main Results:

• GenNet successfully identified known genes (e.g., HERC2, OCA2) for hair and eye color.
• Novel genes (e.g., ZNF773, PCNT) associated with schizophrenia were discovered.
• Predictive biological pathways for schizophrenia, including ubiquitin-mediated proteolysis, endocrine system, and viral infectious diseases, were identified.

Conclusions:

• GenNet provides a freely available, end-to-end deep learning solution for phenotype prediction.
• The framework facilitates the development and use of interpretable neural networks in genomics research.
• GenNet offers a powerful tool for uncovering the genetic architecture of complex traits and diseases with enhanced biological relevance.