AltereGO: Prediction of Protein Function by Multimodal Fusion of Protein Language Models.

AltereGO improves molecular function prediction for low-homology proteins by fusing ESM2-3B, ProtT5, Ankh, and ProstT5-3Di, achieving the strongest overall classification and ranking performance.

• Why it matters: Protein function annotation lags behind the growth of sequence databases, especially for low-homology proteins where experimental characterization is limited.
• What they did: The framework integrates molecular-function predictors from ESM2-3B, ProtT5, and Ankh plus structure-derived ProstT5-3Di using weighted score-level fusion, evaluated on 495,055 proteins across 8,446 GO terms and a 2,000-protein low-homology test set.
• The result: Four-way fusion outperformed learned embedding-level fusion and, in matched comparisons with eight released methods, significantly improved AUPRC in every comparison and Fmax and Smin in six, with structure-derived gains focused on catalytic, redox, and cofactor-associated functions.