Inferring genotype-phenotype maps using attention models.

Area of Science:

• Genetics
• Machine Learning
• Quantitative Genetics

Background:

• Predicting phenotype from genotype is a fundamental challenge in genetics.
• Traditional quantitative genetics methods, like linear regression, often assume additive genetic models and struggle with complex epistasis or gene-environment interactions.
• Machine learning, particularly attention-based models, shows promise for capturing complex interactions.

Purpose of the Study:

• To apply attention-based models to predict phenotype from genotype in quantitative genetics.
• To evaluate the performance of attention-based models against traditional methods, especially in scenarios with complex epistasis.
• To explore multi-environment models for genotype-phenotype analysis and transfer learning.

Main Methods:

• Utilized simulated genetic data with varying levels of epistasis.
• Applied attention-based models to both simulated and experimental quantitative trait locus mapping data from budding yeast.
• Developed a multi-environment attention-based model for cross-environment predictions.

Main Results:

• Attention-based models achieved superior out-of-sample predictive performance compared to standard methods, particularly in complex epistatic scenarios.
• The multi-environment model demonstrated successful application of transfer learning for predicting phenotypes in novel environments with limited data.
• The models effectively captured context-dependent genetic interactions.

Conclusions:

• Attention-based models offer a powerful alternative for genotype-phenotype prediction, excelling where traditional methods falter.
• These models can accurately predict complex genetic architectures and facilitate predictions in new environmental contexts.
• The application of attention mechanisms opens new avenues for understanding gene-environment interactions and complex trait genetics.