AceFF: A State-of-the-Art Machine Learning Potential for Small Molecules

TL;DR

AceFF addresses the challenge of generalizing interatomic potentials to drug-like molecules by introducing TensorNet2 with neutral charge equilibration, enabling learned partial charges and long-range Coulomb terms alongside a fast, equivariant short-range energy. Trained on a large DFT dataset of drug-like compounds, AceFF-2 achieves near-DFT accuracy with competitive MD-scale speed, and demonstrates robust performance across torsion scans, strained conformers, large ligands, and MLIP/MM simulations. The work provides extensive benchmarks against state-of-the-art MLIPs and traditional force fields, highlighting strengths in charge handling and smooth potential energy surfaces while documenting limitations for highly charged, out-of-domain species. The authors release AceFF-2 and supporting data/tools to the community, establishing a rigorous, transferable protocol for evaluating MLIPs in drug discovery and enabling practical MLIP/MM workflows.

Abstract

We introduce AceFF, a pre-trained machine learning interatomic potential (MLIP) optimized for small molecule drug discovery. While MLIPs have emerged as efficient alternatives to Density Functional Theory (DFT), generalizability across diverse chemical spaces remains difficult. AceFF addresses this via a refined TensorNet2 architecture trained on a comprehensive dataset of drug-like compounds. This approach yields a force field that balances high-throughput inference speed with DFT-level accuracy. AceFF fully supports the essential medicinal chemistry elements (H, B, C, N, O, F, Si, P, S, Cl, Br, I) and is explicitly trained to handle charged states. Validation against rigorous benchmarks, including complex torsional energy scans, molecular dynamics trajectories, batched minimizations, and forces and anergy accuracy demonstrates that AceFF establishes a new state-of-the-art for organic molecules. The AceFF-2 model weights and inference code are available at https://huggingface.co/Acellera/AceFF-2.0.

Figures (10)

• Figure 1: The AceFF-2 architecture TensorNet2 builds upon TensorNetTensorNet with the new components colored in blue. The main change is that after the tensor embedding, and each tensor interaction block, a set of partial charges $q$ and weights $w$ are computed and undergo a neutral charge equilibration (NQE) procedure as done in AIMNet2anstine2024aimnet2. The partial charges are then combined with edge features during the tensor interaction. The predicted charges are used to compute a Coulomb energy $E_\text{Coulomb}$ alongside the standard short range node-wise MLIP energy term $E_i$. The dashed arrow represents where iteration occurs for more than one interaction layer. The pictured diagram corresponds to TensorNet2 1-layer. The grey dashed circles around the example molecule illustrate the short cutoffs used in the tensor embedding and interaction. The blue dashed circle represents the long range of the total charge (a global feature feature), NQE block, and the Coulomb interaction.
• Figure 2: Torsion scans. a. Sellers et al. torsion scan benchmarksellers2017comparison. The orange lines are the median value. The methods are ordered from left to right by median MAE compared to the coupled cluster reference data. The y-axis has been intentionally capped at 2.0 kcal/mol to focus on the differences between the MLIP models. An image of one of the molecules with the torsion angle indicated is pictured. Values marked by * were taken from a previous publication sabanes_zariquiey_quantumbind-rbfe_2025. All other data points were recalculated. b. Behara et al torsion scan benchmark behara2024. The orange lines are the median value. The methods are ordered from left to right by median MAE compared to the coupled cluster reference data. The scatter points are color-coded by the molecular charge. The y-axis has been intentionally capped at 4.0kcal/mol to focus on the differences between the MLIPs. Data points of methods marked with * were taken from the original databehara2024, all other data points were recalculated. On both plots, a guide line has been drawn at 1 kcal/mol.
• Figure 3: Schrödinger ligand test set evaluation. a. Energy error vs force error for AceFF-1.0, color-coded by ligand total charge. b. Energy error vs force error for AceFF-2, color-coded by ligand total charge. The axis range is the same in sub-plots a and b to aid comparison. The inset graph in sub-plot b shows a zoomed-in view of the data points. c. Force errors for all tested models. The orange line is the median value. The data points are color-coded by charge.
• Figure 4: Potential energy of ethane C-C bond length scan for different MLIPs compared to a reference of DFT with $\omega$B97m-V/tzvppd.
• Figure 5: Inference speed of different models vs number of atoms. a. Zoomed in on the x-axis to highlight speed for a small number of atoms. b. Full range of tested atom sizes. Guidelines have been drawn which indicate where 1, 10, and 100 ns/day speeds are if a simulation was run with a 1fs timestep. Methods marked with an * have been run with CUDA graphs enabled. Aimnet2-CUDA graphs marked with † runs out of memory with more than 1500 atoms. This is due to the $N^2$ implementation of the CUDA graph-compatible code version.