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

# VeloBind — trained models and reproducibility archive Companion data deposit for the paper *"VeloBind: A Compact, Structure-Free Modelfor Calibrated Protein–Ligand Binding Affinity Prediction"* (Mathur & Hassan, 2026). VeloBind predicts protein–ligand binding affinity (pKd) from protein sequence andligand SMILES alone, with calibrated, protein-family-conditional conformalprediction intervals. Source code: https://github.com/HassanLABAI/VeloBind ## Contentsmodels/ Full 45-model gradient-boosted ensemble (3 algorithms × 3 seeds× 5 folds), RidgeCV meta-learner, isotonic calibrator, and theconformal calibration objects. This is the model used for allheadline results (CASF-2016 R = 0.847).single/ Compact 12 MB single-LightGBM variant + conformal calibrator(CASF-2016 R = 0.834) for constrained deployment.preprocessors/ Ligand feature scaler required for inference.predictions_*.csv Per-complex predictions on CASF-2013 and CASF-2016.metrics_full.csv, extra_metrics.json, RESULTS.mdAll reported metrics, in one place.pdb_class_cache.csv PDB → functional-class map used for the family analysis.