A directed learning strategy integrating multiple omic data improves genomic prediction.

Area of Science:

• Plant Breeding and Genetics
• Bioinformatics and Computational Biology
• Genomics and Omics

Background:

• Genomic prediction (GP) accelerates molecular plant breeding by using genome-wide markers to predict phenotypes.
• Current GP methods using only genomic data have plateaued, and prior omics predictions neglected genomic information.
• Integrating diverse omics data offers a potential solution to overcome current prediction limitations.

Purpose of the Study:

• To develop a novel GP strategy, MLLASSO, integrating multiple omics data (genomics, transcriptomics, metabolomics).
• To improve the predictability of complex traits like rice yield by capturing higher-order gene interactions.
• To identify genetically predictable genes (GPGs) and explore their biological significance in trait regulation.

Main Methods:

• Developed MLLASSO, a multilayered least absolute shrinkage and selection operator model.
• Integrated genomic, transcriptomic, and metabolomic data into a single predictive framework.
• Iteratively learned three layers of genetic features (GFs) supervised by observed transcriptome and metabolome.

Main Results:

• MLLASSO significantly improved rice yield predictability from 0.1588 (GP alone) to 0.2451.
• Identified GPGs whose expression is accurately predicted by genetic markers, serving as strong predictors for complex traits.
• Discovered that GPGs are enriched in eQTL genes and trait-related transcriptional factor families, indicating biological relevance.

Conclusions:

• MLLASSO represents a directed learning strategy supervised by intermediate omics data, enhancing prediction reliability and robustness.
• Learned genetic features possess biological implications for gene expression regulation, linking molecular data to complex traits.
• The integration of multiple omics data via MLLASSO offers a powerful approach for advancing plant breeding and trait prediction.