Environmental Predictability in Antarctic Metagenomes: DNA, Protein, and Structure-Aware Representations

Antarctic marine metagenomes contain extensive genomic diversity, yet a large fraction of their genes remains functionally uncharacterized. We present a large-scale benchmark based on the ACE Antarctic catalog, comprising 218 metagenomes, 89.7 million non-redundant genes, 49.0 million valid protein translations, and 44.6 million genes without functional annotations. Genes were grouped into expression clusters, whose abundances were predicted from environmental variables using random forests. Repeated $R^2\$-based evaluations assigned 19.15 million prioritized gene–protein records to HIGH or LOW ecological association across particle-attached and free-living fractions. We compare interpretable DNA and protein baselines with frozen representations from GenomeOcean, a genomic language model, ESM2, a protein language model, ESM3 structure-aware representations, and concatenated gene–protein embeddings. Homology-aware MMseqs2 splits, paired gene–protein assignments, duplicate removal, and training-only preprocessing were used to prevent information leakage. Across six ecological prediction tasks, GenomeOcean achieved AUROC values of 0.746–0.825, while ESM2 reached 0.723–0.788, consistently outperforming simple composition baselines. Protein, structural, and fused representations provided complementary information under fraction and transfer shifts. In functional open-set evaluation, ESM2 achieved a Pfam OOD AUROC of 0.860 with an energy detector, compared with 0.513 for GenomeOcean and 0.735 for early fusion. These results show that environmental predictability and functional novelty are representation-dependent: genomic, protein, and structure-aware models capture complementary signals rather than defining a universally optimal embedding.