Plant breeding has entered a new era, transitioning from a predominantly unimodal, genomics-focused paradigm to one defined by multimodal data integration. Dense genomic markers, long-term environmental datasets, physiological and omics measurements, and high-throughput phenotyping platforms are now routine components of modern breeding programs. Paradoxically, while plant breeding now operates in a multimodal data ecosystem, most prediction models remain effectively unimodal, relying on a single type of information at a time-genomics alone, enviromics alone, physiology alone, or phenomics alone. This overlooks a fundamental biological reality: complex traits emerge from the interaction of genetic potential, environmental conditions, and dynamic physiological responses. Across the life sciences, the integration of heterogeneous biological data has become a central strategy for improving prediction and discovery. We argue that genomic prediction must become truly multimodal to reflect the biological reality that complex traits emerge from interacting genetic, environmental, and physiological processes. We clarify multimodality as distinct from commonly conflated concepts such as multi-environment or multi-trait prediction and outline strategies for integrating heterogeneous data sources using multi-kernel and multi-stream modeling frameworks.
Crossa et al. (Mon,) studied this question.