Decoding functional genes across species with annotation-independent machine learning

Dissecting the genetic basis of complex traits across species is a challenge for traditional low-throughput, species-specific methods. To overcome these limitations, we introduce a computational framework that integrates cross-species comparative genomics and machine learning to identify functional genes for shared phenotypes, independent of prior annotation. Our method links orthologous gene group (OG) profiles to specific traits, enabling the powerful discovery of functional genes across diverse evolutionary clades. For plant-arbuscular mycorrhizal symbiosis, a notoriously difficult system, our model pinpointed a regulatory network by identifying 27 core plant-AM symbiosis genes among its top 50 candidate OGs, including the critical receptor SYMRK. The approach also proved highly effective for identifying functional genes related to the tested vertebrate skeletal development and multiple bacterial traits. Most notably, for bacterial motility, our model not only identified 63 known motility genes from the top 100 candidate OGs (of which 78 are present in Escherichia coli) but also guided the experimental validation of three novel essential genes. This annotation-independent strategy represents a paradigm shift in functional genomics, offering a scalable and universal engine to decode the genetic architecture of complex traits and illuminate the vast functional dark matter across the tree of life.