Inferring genotype-phenotype maps using attention models.

Area of Science:

• Genetics
• Machine Learning
• Quantitative Genetics

Background:

• Predicting phenotype from genotype is a core challenge in genetics.
• Traditional quantitative genetics relies on linear regression models, often assuming additive effects and pairwise epistasis.
• These models struggle with complex epistasis and gene-environment interactions.

Purpose of the Study:

• To apply attention-based machine learning models to genotype-phenotype prediction.
• To evaluate the performance of attention models against traditional methods, especially in complex genetic scenarios.
• To explore multienvironment models for joint analysis and transfer learning.

Main Methods:

• Utilized simulated data with varying epistatic complexity.
• Applied attention-based models to genotype-phenotype mapping.
• Tested models on experimental data from a budding yeast quantitative trait locus mapping study.
• Developed and analyzed multienvironment attention-based models.

Main Results:

• Attention-based models showed superior out-of-sample predictive performance in epistatic regimes compared to standard methods.
• The models effectively captured complex gene interactions.
• Multienvironment models demonstrated potential for transfer learning in novel environments.

Conclusions:

• Attention-based models offer a powerful alternative for genotype-phenotype prediction, particularly for complex genetic architectures.
• These models can handle intricate epistatic interactions and gene-environment effects.
• The developed architectures facilitate efficient analysis across multiple environments and enable predictive transfer learning.