Transferable Collective Variable to accelerate Protein-Ligand (Un)Binding Transitions via Explainable Machine Learning and Intriguing Role of Ligand Solvation

In the realm of biophysics and therapeutics, understanding the unbinding process of drugs is of paramount importance. The behavior of such systems is heavily influenced by their thermodynamic and kinetic properties, making it essential to estimate the ligand binding free energies and the rate of ligand dissociation. However, these processes are often characterized by rare event transitions that exceed the capabilities of standard brute-force molecular dynamics simulations.

Enhanced sampling simulations offer a potential solution, but their effectiveness hinges on the choice of suitable collective variables (CVs), a task that proves challenging for intricate systems like protein-ligand complexes.

In this research, a novel approach is introduced to derive optimized CVs from the transition state region (TS) using an interpretable machine learning (ML) model called Elastic Net. By incorporating some physically intuitive order parameters from the TS region, the derived CVs significantly accelerate ligand binding and unbinding transitions, enabling rapid free energy surface (FES) convergence across a wide range of systems.

These systems include buried and solvent-exposed active sites such as the Trpsin-benzamidine complex, host-guest systems, and sodium epoxidase.

A noteworthy finding of this study is the significant contribution of ligand hydration in the optimized CV, highlighting the crucial role of the solvent in driving ligand binding and unbinding transitions. The estimated binding free energies for various protein-ligand complexes closely align with experimental results while maintaining a low computational cost.

Furthermore, the derived optimized CV proves transferable across diverse systems, providing a robust and scalable framework for drug discovery and the investigation of complex biomolecular recognition.