QuantumBind-RBFE: Accurate Relative Binding Free Energy Calculations Using Neural Network Potentials

TL;DR

This work demonstrates that a neural-network potential, AceFF-1.0, integrated via an NNP/MM framework (QuantumBind-RBFE), can improve relative binding free energy predictions over traditional MM force fields for diverse, drug-like ligands, including charged species. Using a dual-topology RBFE workflow with alchemical perturbations and ATM, AceFF-1.0 shows lower RMSE and MAE and better ranking (Kendall τ) than GAFF2, and competitive performance relative to OPLS4 across multiple targets. The study also reports practical gains from increasing the NNP/MM timestep to 2 fs, achieving substantial speedups with comparable accuracy, while identifying charge-space limitations that motivate future AceFF improvements. Overall, the results place AceFF-1.0 as a robust, publicly accessible option for RBFE calculations, enabling more accurate and scalable drug-discovery workflows, albeit with caveats on charge coverage and training data diversity.

Abstract

Accurate prediction of protein-ligand binding affinities is crucial in drug discovery, particularly during hit-to-lead and lead optimization phases, however, limitations in ligand force fields continue to impact prediction accuracy. In this work, we validate relative binding free energy (RBFE) accuracy using neural network potentials (NNPs) for the ligands. We utilize a novel NNP model, AceFF 1.0, based on the TensorNet architecture for small molecules that broadens the applicability to diverse drug-like compounds, including all important chemical elements and supporting charged molecules. Using established benchmarks, we show overall improved accuracy and correlation in binding affinity predictions compared with GAFF2 for molecular mechanics and ANI2-x for NNPs. Slightly less accuracy but comparable correlations with OPLS4. We also show that we can run the NNP simulations at 2 fs timestep, at least two times larger than previous NNP models, providing significant speed gains. The results show promise for further evolutions of free energy calculations using NNPs while demonstrating its practical use already with the current generation. The code and NNP model are publicly available for research use.

Figures (10)

• Figure 1: Description of the NNP/MM scheme. While the ligand is simulated with a neural network potential (NNP), the rest of the system is treated with classical molecular mechanics (MM)
• Figure 2: Seller's torsion scan benchmark, comparing torsion accuracy for a series of potentials against a CCSD(T)/CBS baseline.
• Figure 3: (Left) Root Mean Squared Error (RMSE) and (right) Kendall tau correlation for the $\Delta G$s of each protein-ligand system calculated in combination with different approaches: GAFF2 (teal), reported estimates using FEP+ with the OPLS4 forcefield(yellow) and AceFF 1.0 (blue).
• Figure 4: Scatterplots of predicted $\Delta G$ values for each evaluated system using AceFF 1.0. The green and yellow shaded areas represent absolute error thresholds of 1 kcal/mol and 2 kcal/mol, respectively. Additional metrics, including mean absolute error (MAE), root mean square error (RMSE), Kendall tau correlation ($\tau$), and top 30% and top 5 compound identification accuracy, are also displayed. 95% confidence interval values for the relevant metrics are shown in brackets.
• Figure 5: Identification of top compounds across datasets. The left plot shows the accuracy in identifying compounds within the top 30%, while the right plot focuses on the top 5 compounds in each series. Results are compared across the evaluated methods: AceFF 1.0, GAFF2, and OPLS4 with FEP+.