Disentangling genotype and environment specific latent features for improved trait prediction using a compositional...

Area of Science:

• Genetics
• Plant Breeding
• Bioinformatics

Background:

• Traditional predictive models in plant breeding use Principal Component Analysis (PCA) and Autoencoders (AE) for high-dimensional data.
• These methods struggle to separate genotype-specific and environment-specific features, limiting trait prediction accuracy for complex traits.

Purpose of the Study:

• To develop a novel method that disentangles genotype and environment features for enhanced predictive modeling.
• To improve the accuracy of predicting plant traits influenced by both genetic and environmental factors.

Main Methods:

• Developed a compositional autoencoder (CAE) with a hierarchical architecture to decompose high-dimensional data.
• The CAE framework effectively separates entangled latent features into distinct genotype-specific and environment-specific components.
• Applied the CAE to a maize diversity panel dataset for trait prediction.

Main Results:

• The CAE demonstrated superior modeling of environmental influences compared to existing methods.
• Achieved a 7-fold improvement in predictive performance for the 'Days to Pollen' trait.
• Showcased a 10-fold improvement in predictive performance for the 'Yield' trait.
• Outperformed Principal Component Analysis (PCA), Partial Least Squares Regression (PLSR), and vanilla autoencoders.

Conclusions:

• Disentangling latent features using CAE provides a powerful tool for precision breeding and genetic research.
• The CAE significantly enhances trait prediction models, advancing agricultural and biological sciences.
• This approach offers a more accurate understanding of genotype-environment interactions in plant traits.