3D-IsofunGO: input data and cached features for structure-aware isoform function prediction

Input data and cached features accompanying the 3D-IsofunGO paper — a structure-aware method for protein isoform function prediction that fuses ESM-2 protein language model embeddings with AlphaFold2 contact-map representations (encoded by a pre-trained StructGCN) via late fusion, evaluated on 971 Molecular Function Gene Ontology terms with isoform-level labels. Two archives are provided: 3d-isofungo-data-core.zip (~1.2 GB): GO labels, GO graph and term features, the fixed train/validation/test split, information-accretion weights, normalized per-branch features and k-NN graphs, and per-model predictions — sufficient to reproduce all tables and figures. 3d-isofungo-data-raw.zip (~11.4 GB): k-mer baseline features, the isoform interaction network, AlphaFold2 Cα contact maps, scalar pLDDT scores, the merged input protein FASTA (32,769 isoforms), and the GO ontology — for full reproduction from raw inputs. Code and usage instructions: https://github.com/atakmty/3D-IsofunGO