GenNet

GenNet applies biologically informed deep learning to predict phenotypes from genetic variants in population genomics.

Key Features:

• Interpretable Neural Network Architectures: Constructs neural networks by embedding biological knowledge from public databases so that only biologically plausible connections are included, enabling visualization of how genetic variants contribute to phenotypic traits.
• Memory Efficiency: Restricts network connectivity to biologically relevant links to reduce memory usage for large-scale genomic datasets.
• End-to-End Pipeline: Implements processing from input genetic data to phenotype prediction within a unified framework for interpretable models.

Scientific Applications:

• Phenotype Prediction: Applied to predict seventeen phenotypes, identifying replicated genes such as HERC2 and OCA2 associated with hair and eye color.
• Genetic Association Discovery: Identified novel associations including ZNF773 and PCNT for schizophrenia.
• Pathway-Level Interpretation: Highlighted pathways predictive of schizophrenia, including ubiquitin-mediated proteolysis, the endocrine system, and viral infectious diseases.

Methodology:

Leverages deep learning to integrate genetic data with biological knowledge by constructing neural networks that reflect meaningful genetic connections; connections can be defined based on gene annotations or other biological information, and during training the network assigns higher weights to connections important for phenotype prediction.