TL;DR
This paper introduces a hybrid machine learning/molecular mechanics (ML/MM) approach to significantly improve the accuracy of alchemical free energy calculations for drug discovery. By combining the speed of MM force fields with the accuracy of ML potentials for intramolecular ligand interactions, they achieve "chemical accuracy" (around 0.5 kcal/mol error) for protein-ligand binding affinities, a crucial step towards more efficient drug design.
What Problem Does This Paper Address?
Predicting how strongly a small molecule (like a drug candidate) will bind to a protein target is a cornerstone of drug discovery. Alchemical free energy calculations using molecular mechanics (MM) force fields are a powerful tool for this, typically achieving 1-2 kcal/mol accuracy. While this is good, reaching "chemical accuracy" (around 0.5 kcal/mol) would drastically reduce the number of compounds needing synthesis and experimental testing. The main bottleneck for MM force fields is their inability to accurately model the intricate intramolecular interactions of ligands, especially torsional energies and their couplings, which can significantly impact binding free energies.
Methods
The authors propose a hybrid ML/MM potential where the ligand's intramolecular interactions are described by a machine learning (ML) model (specifically, the ANI-2x neural network potential), while the interactions between the ligand and the protein/solvent, as well as protein-protein and solvent-solvent interactions, are handled by a standard MM force field (AMBER14SB/TIP3P with OpenFF 1.0.0 "Parsley").
To integrate this hybrid potential into existing alchemical free energy workflows, they developed a post-processing correction method. This method uses a nonequilibrium (NEQ) switching protocol. Standard MM alchemical free energy calculations are first performed. Then, short (10 ps) NEQ simulations are used to calculate a correction term that accounts for the difference in free energy between the MM and the hybrid ML/MM potentials. This correction is applied to the MM-calculated binding free energies to obtain the ML/MM-corrected values. This approach avoids the computational cost of running full alchemical free energy calculations with the ML/MM potential from scratch.
Datasets
The study uses the Schrödinger JACS benchmark set for Tyk2 kinase inhibitors. This dataset consists of 16 congeneric ligands with experimentally determined binding affinities, known to be challenging for standard MM force fields.
Key Insights and Ideas
- Hybrid ML/MM is the sweet spot: Combining the strengths of MM (speed, handling large systems) and ML (accuracy for intramolecular ligand energetics) offers a practical path to higher accuracy without prohibitive computational cost.
- Torsions are key: The significant improvements in binding free energy predictions are likely driven by the ML model's superior ability to capture accurate ligand torsional profiles and couplings, which are often poorly represented in standard MM force fields.
- NEQ corrections are efficient: Using short, bidirectional nonequilibrium simulations as a post-processing step to correct MM free energies is computationally feasible and effective, making the ML/MM approach accessible within existing workflows.
- Ligand conformation matters: The differences in torsional profiles between MM and ML/MM were particularly pronounced for amide functional groups, suggesting that MM force fields may struggle with these common motifs in drug-like molecules.
Key Results
- Standard MM calculations (OpenFF 1.0.0 with AMBER14SB/TIP3P) on the Tyk2 benchmark set yielded an RMSE of 0.97 kcal/mol for absolute binding free energies.
- The hybrid ML/MM approach, using ANI-2x for ligand intramolecular energies and MM for everything else, reduced the RMSE to 0.47 kcal/mol. This represents a significant improvement, bringing the results closer to experimental values.
- The ML/MM corrections were generally positive, indicating that the ML potential slightly disfavors binding compared to the MM potential, with magnitudes up to 4 kcal/mol for some ligands.
- The study demonstrated that 10 ps NEQ switching simulations are sufficient to obtain reliable free energy corrections.
Limitations and Future Work
- Limited element coverage of ML potentials: Current ML potentials like ANI-2x do not cover all elements relevant to drug discovery (e.g., ions, cofactors), limiting their direct application to such systems.
- Computational cost of ML: While faster than QM, ML potentials are still slower than GPU-accelerated MM simulations, though this gap is expected to close.
- Parameterization for condensed-phase properties: ML models may not yet be parameterized to accurately capture condensed-phase properties, meaning MM might still be superior for intermolecular interactions in large systems.
- Naïve ML/MM formulation: The current hybrid formulation is simple and effective but could be improved by more sophisticated definitions of the ML/MM potential, potentially including ML for more terms.
- Interoperability: Developing seamless integration of ML potentials into simulation packages (like OpenMM) is crucial for broader adoption.
- Further investigation of torsion discrepancies: The significant differences in torsion profiles warrant further study, potentially through refitting MM torsion profiles or exploring more advanced ML models.
- Broader applicability: While promising for the Tyk2 system, more extensive studies are needed to confirm the robustness of these accuracy improvements across diverse congeneric series.
Summary updated