Kinetics Beyond Pose: A Structure-independent Framework for koff Prediction Across Diverse Protein Families

Predicting the ligand–receptor dissociation rate constant (koff) remains to this day a key challenge in drug discovery, despite its well-established relevance to residence time and in vivo drug efficacy, since it depends on the dynamic interaction between the ligand and the receptor site that cannot be immediately reconstructed from equilibrium conditions alone. In this work, we developed a transformer-based machine-learning framework able to predict the koff value relying solely on the protein amino acid sequence and ligand SMILES string, without requiring explicit 3D protein–ligand complex structures. Evaluated on a chemically and biologically diverse dataset, our model achieved competitive performance in cross-validation and on a held-out test set, while also showing robust behavior on the most represented protein families. Our results indicate that promising koff prediction performance can be obtained with a structure-independent representation, by exploiting pretrained protein and molecular language representations as a scalable strategy for kinetics-aware drug discovery.