Acellera/AceFF-2-RESP-1

Acellera AceFF 2-RESP-1

Acellera Therapeutics, inc · info@acellera.com · Apache 2.0

A RESP-charge variant of AceFF-2: predicts energy, forces, and per-atom RESP partial charges for electrostatic embedding in NNP/MM (e.g. RBFE). A modified TensorNet-2 model [Farr2026] with an added charge head, trained on a subset of the AceFF-2 dataset (the Schrödinger dataset, as it is referred to in the AceFF-2 paper). Paper: AceFF.

Benchmarks

RESP charges

Predicted vs reference RESP charges (650 conformers, 30,278 atoms, Q −2…+2).

Overall MAE 0.033 e, RMSE 0.049 e.

Q	Conformers	Atoms	Charge MAE (e)	Charge RMSE (e)	Force MAE (eV/Å)	Energy MAE (eV)
−2	23	1048	0.0381	0.0490	0.0762	0.0605
−1	80	3559	0.0298	0.0406	0.0458	0.0858
+0	422	19425	0.0341	0.0509	0.0462	0.0790
+1	122	6078	0.0302	0.0450	0.0533	0.0876
+2	3	168	0.0355	0.0571	0.0543	0.2897
All	650	30278	0.0330	0.0486	0.0486	0.0818

Usage

Currently requires two branches (not yet merged): the torchmd-net tensornet2_resp branch torchmd/torchmd-net@resp_model for loading/inference, and the atm electrostatic-embedding branch Acellera/atm@electrostatic_embedding for RBFE. Both will be upstreamed.

import torch
from torchmdnet.models.model import load_model

model = load_model("aceff-2-resp-1.ckpt", model="tensornet2_resp", derivative=True).eval()
z = torch.tensor([8, 1, 1])                                   # water
pos = torch.tensor([[0.,0.,0.], [0.757,0.586,0.], [-0.757,0.586,0.]])
batch = torch.zeros(3, dtype=torch.long); q = torch.zeros(1)  # total charge
energy, forces, charges = model(z=z, pos=pos, batch=batch, q=q)
print(charges.reshape(-1))                                    # per-atom RESP charges

Limitations

• Small molecules only (elements H, B, C, N, O, F, Si, P, S, Cl, Br, I).
• 2 fs timestep with hydrogen mass repartitioning.
• Molecular charge states −2…+2 only.

Citation

If you use AceFF-2-RESP-1, please cite our preprint: Evaluating Electrostatic Embedding MLIP/MM for Relative Binding Free Energy Calculations, arXiv:2608.13355, DOI: 10.48550/arXiv.2608.13355.

References

[Simeon2024] Simeon, De Fabritiis. TensorNet. NeurIPS 36 (2024). https://arxiv.org/abs/2306.06482

[Pelaez2024] Pelaez et al. TorchMD-Net 2.0. J. Chem. Theory Comput. 2024, 20, 4076. https://arxiv.org/abs/2402.17660

[Zariquiey2025] Sabanés Zariquiey et al. QuantumBind-RBFE. https://arxiv.org/abs/2501.01811

[Farr2026] Farr et al. AceFF. https://arxiv.org/abs/2601.00581