Fungal terpene synthase sequences with functional class labels and a homology-grouped cross-validation partition (v2.1)

A curated set of 1089 fungal terpene synthase protein sequences labelled with the class of terpene product (Sesquiterpene 420, Diterpene 373, Triterpene 153, Sesterterpene 143), together with the homology-grouped 5-fold partition used to evaluate the TerpeneBERT classifier. Each record carries its source accession (NCBI GenBank or JGI MycoCosm), organism, GO terms, principal product and SMILES where reported, InterPro and Pfam signatures, a 50%-identity cluster identifier and a cross-validation fold. Why the partition matters. Terpene synthases are a homologous family. A random train/test split places near-identical sequences on both sides and inflates accuracy by roughly 0.09 macro-F1. Folds here are assigned by whole 50%-identity clusters; the maximum cross-fold identity is 0.4987. Relationship to the released model. The TerpeneBERT weights were trained on v2.0, which differed from this file in one record: the sequence of QGA30884.1 carried three spurious N-terminal residues from an unparsed FASTA header. The defect was found after training and is corrected here, so re-running the pipeline on this file will not reproduce the published weights bit for bit. See CHANGELOG.md. Caution. Seven GO terms in the go_terms column are label-pure and leak the class if used as model input. The column is provided for annotation and filtering, not for training. Sequences originating from JGI MycoCosm (36 records) require the acknowledgment given in README.md.