| # External data disclosure | |
| Extra training rows were used from **CatPred-DB** (https://huggingface.co/datasets/RosettaCommons/CatPred-DB; Boorla and Maranas, Nature Communications 2025, doi:10.1038/s41467-025-57215-9; license CC BY 4.0). Source files used (train files for training; val/test files only to exclude their enzymes): kcat/kcat_train.csv, kcat/kcat_test.csv (only to EXCLUDE its enzymes), kcat/kcat_val.csv (only to EXCLUDE its enzymes), km/km_train.csv, km/km_test.csv (only to EXCLUDE its enzymes), km/km_val.csv (only to EXCLUDE its enzymes). | |
| Filtering applied before training (to avoid leakage): | |
| 1. SMILES canonicalised with RDKit and mapped to the competition's SMILES strings where identical. | |
| 2. Removed (SMILES, sequence) pairs already in the competition training set. | |
| 3. Removed every enzyme sequence that appears in the public validation inputs, and every near-copy (3-mer cosine >= 0.8). | |
| 4. Only the CatPred-DB *_train.csv files are used; every enzyme that appears in its *_val.csv or *_test.csv files is removed. | |
| The competition test set is private, so we cannot check it directly; if any test items are present in this external data, the organisers may remove them from the evaluation, as announced. The exact rows used are in `external_data/catpred_filtered_rows.parquet`. Protein embeddings use the pretrained ESM2 model `facebook/esm2_t12_35M_UR50D` (MIT license). | |