QuantumBind-RBFE: Accurate Relative Binding Free Energy Calculations Using Neural Network Potentials
Francesc Sabanés Zariquiey, Stephen E. Farr, Stefan Doerr, Gianni De Fabritiis · 2025
Read QuantumBind-RBFE: Accurate Relative Binding Free Energy Calculations Using Neural Network Potentials by Francesc Sabanés Zariquiey, Stephen E. Farr, Stefan Doerr, Gianni De Fabritiis (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-01811.txt.
Introduction
Francesc Sabanés Zariquiey and co-authors validated in 2025 a neural approach to drug binding calculations. Predicting how strongly a candidate drug binds its target guides the costly stages of drug design. Standard force fields for the drug molecules limit accuracy. The team instead used a neural network potential, AceFF, built on the TensorNet design, to handle the ligands.
Their method, QuantumBind-RBFE, computes relative binding free energies between similar compounds. On standard benchmarks it beat the GAFF2 force field and an earlier neural potential in accuracy. It nearly matched a top commercial force field as well. It also ran with a time step at least twice as large as earlier neural models, giving real speed gains. Code and model are public for research use.
Catalog record
| CREATOR | Francesc Sabanés Zariquiey, Stephen E. Farr, Stefan Doerr, Gianni De Fabritiis |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 51 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-01811
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
QuantumBind-RBFE: Accurate Relative Binding Free Energy Calculations Using Neural Network Potentials Francesc Sabanés Zariquiey, Stephen E. Farr, Stefan Doerr, Gianni De Fabritiis 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. † † affiliation: Acellera Labs, C Dr Trueta 183, 08005, Barcelona, Spain † † affiliation: Acellera Labs, C Dr Trueta 183, 08005, Barcelona, Spain † † affiliation: Acellera Therapeutics, 38350 Fremont Blvd 203 Fremont CA, 94536 USA † † email: [email protected] † † affiliation: Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), C Dr. Aiguader 88, 08003, Barcelona, Spain † † affiliation: Acellera Therapeutics, 38350 Fremont Blvd 203 Fremont CA, 94536 USA † † affiliation: Institució Catalana de Recerca i Estudis Avançats (ICREA), Passeig Lluis Companys 23, 08010 Barcelona, Spain 1 Introduction Accurate prediction of protein-ligand binding affinities is crucial in drug discovery, particularly in the hit-to-lead and lead optimization stages, where a congeneric series of ligands must be screened efficiently. Among the methods employed, alchemical free energy calculations Jorgensen (2004) ; Abel et al. (2017) ; Armacost et al. (2020) ; York (2023) ; Qian et al. (2024) , including relative binding free energy (RBFE) methods, have gained prominence due to their capacity to deliver reliable estimates of binding affinities across a wide range of compounds. Such techniques are widely used in both academia and the pharmaceutical industry. However, the accuracy of RBFE calculations is hampered by several factors such as convergence of the simulations, omission of relevant chemical effects (e.g. tautomerizations and/or protonation), poor selection of pair edges, poor pose selection and/or the inaccuracy of the ligand force field. In this work, we put a special focus on this last point. Accurately capturing molecular interactions across the diverse landscape of drug-like compounds is a challenge for traditional molecular mechanics (MM) force fields like GAFF Wang et al. (2004) ; Wang et al. (2006) , CGenFF Vanommeslaeghe et al. (2010) ; Vanommeslaeghe et al. (2012) , and OpenFF Qiu et al. (2021) ; Boothroyd et al. (2023) . These force fields can struggle with rare chemical groups and cannot account for key energetic factors, such as polarization and quantum effects. As a result, they can introduce inaccuracies in binding free energy predictions, especially when handling complex ligand chemistry or significant conformational flexibility. Recently, an alternative approach has emerged based on neural network potentials (NNPs) to use fast neural network function approximation of the quantum mechanic’s energy surface Duval et al. (2023) . Based on these architectures, a few methods have been made available which we will consider in this work. MACE-based models Kovács et al. (2023) are very accurate but computationally expensive. This is not a problem in material science where the alternative is very expensive quantum chemistry methods. For biology-oriented applications, ANI-models Devereux et al. (2020) are much faster but at the expenses of accuracy, still providing a very good trade-off. AIMNet2 Anstine et al. (2024) extends ANI with a different architecture to more elements and molecules, still adding on computational costs. The equivariant transform (ET) architecture Thölke and De Fabritiis (2022) and TensorNet Simeon and De Fabritiis (2023) aimed to provide state-of-the-art accuracy without sacrificing computational efficiency, being taylored specifically for drug discovery. These two are integrated in the TorchMD-Net software framework Doerr et al. (2021) ; Thölke and De Fabritiis (2022) ; Pelaez et al. (2024) and a new NNP model has been made available, AceFF Acellera (2024) . Molecular dynamics (MD) simulations based on NNPs promise accurate simulations at reasonable costs compared with density functional theory (DFT) methods, but do represent an increase in computational costs compared to traditional molecular mechanics, yet their future is very promising De Fabritiis (2024) . Machine-learned potentials employ relatively short cutoff distances (typically around 5 Å) for efficiency. However, message-passing architectures—such as those based on TensorNet—effectively extend the receptive field beyond the nominal cutoff. Simeon and De Fabritiis (2023) For example, a single interaction layer with a 5 Å cutoff can yield an effective range of approximately 10 Å. Moreover, in our NNP/MM setup, only the ligand’s intramolecular interactions are modeled by the NNP. Long-range electrostatic is taken care of by the standard classical potential at this time. Recent studies by Karwounopoulos et al. Karwounopoulos et al. (2025) have explored an alternative approach in which torsional profiles in classical force fields are fitted via machine learning as a means to approximate the benefits of using an NNP with mechanical embedding. Their results indicate that this end‐state corrected MM/ML method does not offer significant improvements over standard MM in capturing the energetic contributions of ligand strain. In contrast, NNP/MM approach directly employs a neural network potential to model the complete intramolecular energy surface of the ligand, providing a more comprehensive treatment of ligand strain and other subtle energetic effects. Recently, we explored the integration of NNPs into RBFE calculations with promising results. Sabaneés Zariquiey et al. (2024) Specifically, we used ANI-2x Devereux et al. (2020) with good results. However, ANI-2x cannot handle charged species and has a restricted range of supported atom elements. These limitations restrict the applicability across the diverse chemical space encountered in drug discovery. Recently, alternative approaches have emerged that address these shortcomings by integrating quantum-mechanical insights into the potential. Notably, the QDpi models Zeng et al. (2023) ; Tao et al. (2024) , use a fast QM/MM-MLP framework to model both intra- and intermolecular interactions, including challenging cases such as tautomers and different protonation states. These models have demonstrated improved performance in reproducing properties like proton affinities. Furthermore, Crha et al. extended their QM/MM framework by integrating a machine‐learned potential to predict both the QM region and its fully polarized buffer’s energies, thereby enabling direct alchemical free‐energy perturbations as demonstrated in the methanol-to-methane transformation in water. Crha et al. (2025) In this work, we extend our previous work Sabaneés Zariquiey et al. (2024) by testing QuantumBind-RBFE, our NNP/MM approach Galvelis et al. (2023) for RBFE calculations, using AceFF. The AceFF 1.0 Acellera (2024) model supports a broad range of atom elements, including charged molecules, and is tailored for accurate and efficient RBFE calculations across diverse molecular systems. To validate the performance of RBFE calculations using NNPs, we conducted a comprehensive benchmarking study using established datasets, including