VeloBind: Pre-trained models, features, and predictions for structure-free protein-ligand binding affinity prediction

Pre-trained ensemble models, extracted feature matrices, and prediction outputs for VeloBind, a structure-free protein-ligand binding affinity predictor using frozen ESM-2 embeddings and gradient-boosted ensembles. Achieves Pearson R = 0.8469 on CASF-2016 using sequence + SMILES input only. Contents:- velobind_models.zip: 45 fold models (LGBM + CatBoost + XGBoost, 3 seeds × 5 folds), meta-learner, scalers, applicability domain files- velobind_features.zip: NPZ feature matrices for training set, CASF-2016, and CASF-2013- velobind_predictions.zip: Prediction CSVs and full metrics with bootstrap confidence intervals