Acellera AceFF 1.1 is a next-generation Neural Network Potential (NNP) designed for Relative Binding Free Energy (RBFE) calculations in drug discovery.
It addresses key limitations of traditional molecular mechanics (MM) force fields and earlier NNP models, including restricted atom types, limited charge support, and computational inefficiencies.
The model leverages the TensorNet architecture[1] and the NNP software library TorchMD-Net [2] to provide accurate predictions for diverse drug-like compounds, supporting all key chemical elements and charged molecules.
AceFF 1.1 improves the stability of molecular dynamics simulations, supports 2 fs timesteps, and achieves state-of-the-art accuracy with fewer outliers in RBFE predictions.
Description
AceFF 1.1 is the second version of a new family of potentials released by Acellera.
It uses TensorNet 1-layer trained
on Acellera's internal proprietary dataset of molecular forces and energies using the wB97M-V/def2-tzvppd level of theory and VV10 dispersion corrections.
The training set was built on PubChem.
We extracted the SMILES and generated molecules, filtering out molecules larger than 30 atoms.
We kept only molecules with the elements H, B, C, N, O, F, Si, P, S, Cl, Br, and I.
The AceFF1.1 dataset includes an additional 1 million conformations of molecules with up to 30 atoms and more diverse charges, building upon the AceFF1.0 dataset.
Bechmarks
Wiggle150
The table shows the results on the Wiggle150 benchmark. We include AIMNet2 and ANI-2x for comparison.
Method
MAE (kcal/mol)
RMSE (kcal/mol)
AceFF1.1
2.51
3.18
AceFF1.0
2.73
3.32
AIMNet2
2.39
3.13
ANI-2X
4.41
5.41
Performance of NNPs on Wiggle150 benchmark
Schrodinger ligands test set
We create our own hold-out test set by labelling 650 ligands from the Schrodinger public binding free energy benchmark (Jacs, Merk, and charge_annhil sets) with AceFF DFT level of theory. We evaluate the Force MAE of the AceFF predictions.
ligand_testset
Key Features
Broad Applicability: Supports diverse drug-like molecules, including charged species and rare chemical groups.
High Accuracy: Benchmark-tested on the JACS dataset, demonstrating performance comparable to or better than MM-based methods (e.g., GAFF2, FEP+).
Improved Stability: Enables a 2 fs timestep for NNP/MM simulations, significantly reducing computational costs.
Integration-Friendly: Available for RBFE calculations via HTMD.
Open Science: The model and all benchmarking data are accessible on GitHub for not-for-profit usage.
Usage
AceFF 1.1 is designed for use alone or in an NNP/MM approach, where the ligand is treated with the neural network potential and the environment with molecular mechanics.
Example notebooks are available in Google Colab, demonstrating the use of AceFF with OpenMM and ASE.
Run ML potential molecular simulations of a small molecule using ACEMD with this tutorial , e.g. to minimize.
For a tutorial on running mixed protein-ligand simulations, refer to NNP/MM in ACEMD.
Applications
Drug Discovery: Optimizing lead compounds in hit-to-lead and lead optimization stages using free energy methods.
Binding Free Energy Calculations: Accurate and efficient RBFE predictions for diverse molecular systems.
Molecular dynamics: Capturing higher-body terms than traditional MM force fields, AceFF can be used for structure minimization and dynamics of small molecules.
Limitations
Small molecules only: AceFF 1.1 is trained on specifically curated and extended PubChem data. However, proteins, water, etc are not part of the dataset. AceFF 2.0 will be capable of simulations of proteins.
Time step: Use time steps of 2fs to run dynamics with hydrogen mass repartitioning.
Only -1,0,1 charges: For simplicity, we have trained only on these types of charged molecules; do not use it on +2,-2, etc. AceFF 1.2 will fix this.
References
[1] Simeon, Guillem, and Gianni De Fabritiis, Tensornet: Cartesian tensor representations for efficient learning of molecular potentials, Advances in Neural Information Processing Systems 36 (2024), https://arxiv.org/abs/2306.06482
[2] Raul P. Pelaez, Guillem Simeon, Raimondas Galvelis, Antonio Mirarchi, Peter Eastman, Stefan Doerr, Philipp Thölke, Thomas E. Markland, Gianni De Fabritiis, TorchMD-Net 2.0: Fast Neural Network Potentials for Molecular Simulations, J. Chem. Theory Comput. 2024, 20, 10, 4076–4087, https://arxiv.org/abs/2402.17660
[3] Francesc Sabanés Zariquiey, Stephen E. Farr, Stefan Doerr, Gianni De Fabritiis, QuantumBind-RBFE: Accurate Relative Binding Free Energy Calculations Using Neural Network Potentials, https://arxiv.org/abs/2501.01811 (2025).
