Interpretable machine learning platform connecting multi-omics data to crop phenotype

Overview

This solution is an interpretable machine learning platform designed to unravel the causal relationships between plant genotype, intermediate molecular traits, and observable phenotype. By integrating host genome, root-associated rhizosphere microbiomes, seedling and root metabolomes, and leaf transcriptomes, the platform delivers a holistic view of how genetic and microbial variation drives plant morphological traits. The approach is particularly relevant for crop improvement under nutrient stress, where traditional genomic prediction models often fall short.

The platform addresses a critical challenge in agricultural biotechnology: understanding how below-ground root traits and rhizosphere microbiomes interact with host genetics to influence plant performance. Validated on a maize association panel, a breeding-related population, and a sorghum diversity panel under two nitrogen conditions, the model demonstrates strong cross-species transferability using homologous genes between maize and sorghum.

Technical specifications

Core methodology:

• Multi-layer neural network architecture extending linear mediation models to non-linear frameworks, supporting three- to five-layer causal chains
• Integration of genotype, microbiome, root metabolites, and transcriptome profiles as sequential middle layers
• Bayesian approaches for statistical significance and genomic interpretability
• Markov Chain Monte Carlo (MCMC) methods for inferring distributions of genetic variant and omics effects
• Weighted-sum prediction of complex traits from inferred effect distributions
• Cross-species model transferability between maize and sorghum using homologous gene mapping

Key findings and performance:

• Genome-wide association studies identified 622 plant loci linked to 104 microbial groups in the maize rhizosphere
• Key microbe mediators identified, including Massilia putida, known to boost plant growth and nitrogen uptake under low nitrogen conditions
• Genomic prediction accuracy improved by 5% to 40% when microbiomes were incorporated, with the largest gains under low nitrogen conditions

Technology readiness level

The platform has been validated through three completed studies demonstrating host genetic control over rhizosphere microbiomes, identification of gene and microbe mediators in causal chains, and improved genomic prediction accuracy with microbiome integration. Preliminary analyses have confirmed the three-layer mediation model, and ongoing work is expanding the framework to four- and five-layer architectures. The model is currently being trained on maize and tested on sorghum, positioning the technology at an advanced research and validation stage ready for collaborative refinement and field deployment.