Predicting complex phenotypes using multi-omics data in maize.

Understanding and predicting complex traits in plants remains a fundamental challenge due to the emergent nature of most phenotypes and their dependence on genetic, regulatory, and environmental interactions. Accurate prediction of traits and identification of underlying genetic elements have broad applications for plant breeding, systems biology, and biotechnology. Here, we tested if multi-omic datasets could improve predictive accuracy of 129 diverse maize phenotypes across 9 environments using genomic markers, field-based transcriptomic data from 2 locations, and drone-derived phenomic data of vegetative indices. We trained and compared linear (rrBLUP) and nonlinear (support vector regression) models using single- and multi-omics inputs. Multi-omics models consistently outperformed single-omics models for most traits, with genomic and transcriptomic inputs contributing distinct biological features. Phenomic features alone yielded the lowest predictive power but improved predictions for specific trait categories like root architecture. Transcriptomic datasets enabled cross-environment prediction, demonstrating that gene expression patterns from one field site could accurately predict traits measured in another. Environment-specific expression of benchmark flowering time genes highlighted the value of transcriptomics in capturing genotype-by-environment (G × E) interactions not detectable through genomic data alone. Analysis of model feature weights further indicated that predictive signal is distributed across many genes, consistent with complex traits such as yield arising from coordinated, network-level processes rather than a small number of dominant loci. These findings demonstrate that integrating transcriptomic and phenomic data with genotypes enhances trait prediction, improves model generalizability across environments, and provides deeper insight into the genetic and regulatory architecture of agriculturally important traits in maize.