Multimodality, interaction modeling, and multimodule architectures in genomic prediction: A unified conceptual framework

The rapid expansion of genomic, environmental, phenomic, and other high-dimensional data sources has transformed genomic prediction in plant breeding. However, the terms multimodal, interaction modeling, and multimodule architecture are often used inconsistently, generating ambiguity regarding whether they refer to biological assumptions, data integration strategies, or computational design. These dimensions are conceptually independent in the sense that none logically requires or implies the others. Interaction modeling may be implemented within a multimodule architecture, but modular computation does not inherently encode biological interaction. Multimodality refers strictly to the joint use of heterogeneous biological data sources; interaction modeling reflects explicit assumptions about biological dependencies such as genotype-by-environment effects; and multimodularity describes how computation is architecturally organized. We illustrate the proposed framework using conceptual and literature-based examples from wheat breeding, emphasizing interpretation rather than introducing new experimental results. By clarifying terminology and model design principles, this framework aims to improve methodological transparency, facilitate fair comparison among prediction approaches, and strengthen communication between quantitative geneticists, data scientists, and breeding practitioners. While often grouped under the umbrella of artificial intelligence, the approaches used here are more precisely framed as statistical learning methods designed to model and predict measurable genotype-environment-phenotype relationships rather than to generate synthetic or human-like outputs.