Multimodal AI Decodes Extreme Environment Functional Dark Matter Beyond Homology

Abstract

Functional annotation of proteins from extreme environments represents a major bottleneck for bioresource discovery, as a vast reservoir of functional dark matter defies existing homology-based methods. We demonstrate that environmental pressures impart conserved physicochemical energy signatures that co-determine protein function with sequence and structure. Here we developed ACCESS, a multimodal graph neural network employing hierarchical contrastive learning with a tailored label-sample co-embedding to fuse energy, sequence, and structural information and overcome homology scarcity. ACCESS surpasses state-of-the-art methods including BLASTp and CLEAN in annotating low-identity enzymes. Applied to extreme environmental metagenomics, we constructed a function map of extremophile enzymes to expand the biocatalyst library, pinpointed functionally critical residues to guide rational design, and enabled large-scale, function-based macro-evolutionary analyses. This paradigm transcends the limitations of homology, illuminating protein dark matter and accelerating the exploration of the biospheres functional diversity for applications in biotechnology and therapeutic development.