KinForm: Trained models, protein embeddings, binding-site scores, and sequence ID mapping

This dataset provides all runtime assets needed to reproduce KinForm results and generate predictions offline using the repository’s code. It includes pre-trained models, precomputed protein embeddings, optional binding-site probability scores, and the sequence ID mapping used across assets. Contents results/sequence_id_to_sequence.pkl A pickled dict mapping stable sequence IDs (e.g., “Sequence 11894”) to amino-acid sequences. Used to align embeddings and binding-site scores to IDs. results/trained_models/ Serialized scikit-learn regressors trained for kcat and KM across KinForm configurations (e.g., UniKP, KinForm-L, KinForm-H). Ready for use with main.py to train/evaluate/predict. results/binding_sites/ Pseq2Sites per-residue binding-site probability outputs in TSV format. Columns: PDB: sequence ID (matches the keys in sequence_id_to_sequence.pkl) Pred_BS_Scores: stringified list of length L (sequence length), where the i-th value is the probability that residue i is in the binding site. results/protein_embeddings/ Precomputed mean and weighted protein embeddings used as model inputs, including ESM2 (layers 26 and 29), ESMC (layers 24 32), and ProtT5 (layer 19 and last layer). For details on methods and benchmarks, see the preprint: https://arxiv.org/abs/2507.14639. Use with https://github.com/Digital-Metabolic-Twin-Centre/KinForm.