The Quasi-Bound State as a Predictor of Relative Binding Free Energy

Relative binding free energy (ΔΔ<i>G</i>_bind) predictions have become the main approach to evaluate the potency of a congeneric series of compounds. They are enabled by alchemical transformations coupled to free energy methods, tools that have become essential in drug design. Yet, they are computationally expensive and are limited to small compound sets and relatively simple transformations. The ever-increasing size of virtual screening databases demands faster methods to assess virtual hits. Here, we show that the structural robustness of protein-ligand complexes, measured as the free energy necessary to reach a quasi-bound state (Δ<i>G</i>_QB) by Dynamic Undocking (DUck), is well suited to detect outliers in the structure-activity continuum (i.e., activity cliffs), which are particularly challenging for knowledge-based approaches. On different congeneric series of HSP90α, CDK2, and BACE1 inhibitors, we demonstrate that Δ<i>G</i>_QB can deliver excellent predictions. Despite the local nature of the measurement, these are in some cases comparable to the much more computationally demanding alchemical transformation methods. We find that for systems following a one-step dissociation model, Δ<i>G</i>_QB informs about the free energy of the transition state, allowing us to predict relative binding kinetics and, when the series present relatively constant on-rates, also ΔΔ<i>G</i>_bind. This work has important implications for drug discovery, as it shows that within a well-defined applicability domain, high-throughput computational dissociation studies can deliver ΔΔ<i>G</i>_bind predictions that compare well with rigorous alchemical transformation methods at a fraction of the cost.