Hugging Face's logo

Acellera
/
AceForce-1.0

chemistry
molecular simulations
machine learning potentials
neural network potentials
drug discovery
Model card Files
AceForce-1.0 / README.md
giadefa's picture
giadefa
Update README.md
265e2d1 verified
preview code
|
raw
history blame
4.69 kB
metadata
license: other
license_name: license.txt
license_link: LICENSE
tags:
  - chemistry
  - molecular simulations
  - machine learning potentials
  - neural network potentials
  - drug discovery

AceForce 1.0

Organization(s): Acellera Therapeutics, inc
Contact: [email protected]
License: Refer to the linked publication or accompanying license file for usage rights.

Overview

AceForce 1.0 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 to provide accurate predictions for diverse drug-like compounds, supporting all key chemical elements and charged molecules. AceForce 1.0 improves the stability of molecular dynamics simulations, supports 2 fs timesteps, and achieves state-of-the-art accuracy with fewer outliers in RBFE predictions.

NOTE: Soon, the model will be publicly available here. For pre-releases, write to [email protected] for consideration.

Description

AceForce 1.0 is the first 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 20 atoms. We kept only molecules with the elements H, B, C, N, O, F, Si, P, S, Cl, Br, and I, and a formal charge of -1,0,1.

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.

Performance Highlights for Relative Binding Free Energies (RBFEs)

  • Error Metrics: AceForce 1.0 exhibits lower RMSE and higher Kendall tau correlations compared to traditional MM force fields across various protein targets.
  • Outlier Reduction: Achieves competitive accuracy at both 1 kcal/mol and 2 kcal/mol error thresholds.
  • Ranking Ability: Effectively identifies top-performing compounds, achieving >75% accuracy in several benchmark datasets.

Usage

AceForce 1.0 is designed for use in NNP/MM workflows, where the ligand is treated with the neural network potential and the environment with molecular mechanics. It can also be used to simulate pure NNP of small molecules.

  1. Run ML potential molecular simulations of a small molecule using ACEMD with this tutorial , e.g. to minimize.
  2. 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, ACEFORCE can be used for structure minimization and dynamics of small molecules.

Limitations

  • Small molecules only: ACEFORCE 1.0 is trained on specifically curated and extended PubChem data. However, proteins, water, etc are not part of the dataset. ACEFORCE 2.0 will be capable of simulations of proteins.
  • Time step: Use time steps of 2fs or 3fs to run dynamics with hydrogen mass repartitioning.
  • Only -1,0,1 charges: For simplicity we have trained only on these type of charged molecules; do not use it on +2,-2, etc. Aceforce 1.1 will fix this.

References

  • QuantumBind: Next-Generation Accurate Relative Binding Free Energy Calculations Using the AceForce Neural Network Potentials, in preparation.