Sage 2.3.0 Release Candidate 2 now available

[j-wags]

Summary

We’re pleased to announce Sage 2.3.0 Release Candidate 2 (rc2)! Sage 2.3.0 will be the first OpenFF force field to use the AshGC neural network charge model, which was trained to AM1BCC ELF10 charges, and allows for rapid charge assignment for molecules with hundreds or thousands of heavy atoms. Sage 2.3.0 rc2 maintains or improves performance on all of our benchmarks except solvation free energies in nonaqueous solvents.

Please try it out and let us know if you find any substantial improvements, regressions, or other issues in comparison to Sage 2.2.1. If all goes well, we plan to make the full release in a few weeks!

Example

Because AshGC assigns charges so quickly, use cases like quick conformer minimization are now practical. The following python code will minimize 50 different molecules from an input SDF, saving each minimized structure to an SDF file, in about 1 minute:

from openff.toolkit import Molecule, ForceField
from tqdm import tqdm    # for progress bars
mols = Molecule.from_file("inputs.sdf")  # loads 50 different molecules
ff = ForceField("openff_unconstrained-2.3.0-rc2.offxml")
for idx, mol in enumerate(tqdm(mols, desc="Processing molecules")):
    interchange = ff.create_interchange(mol.to_topology())
    interchange.minimize()
    mol.conformers[0] = interchange.positions
    mol.to_file(f"minimized_mol_{idx}.sdf", file_format="sdf")

Two notes on the example:

• The first time NAGL is used in a new environment, there may be a few seconds of extra runtime due to pytorch initialization.
• Warnings are sometimes emitted about protons/stereo/charges, those are from a performance heuristic and don’t change the chemistry of the inputs (we’re working on silencing these)

How to Install

The easiest way to try Sage 2.3.0 rc2 out is by making a new environment:

mamba create -n openff -c conda-forge "openff-toolkit>=0.17"
mamba activate openff

Or you can update an existing environment:

mamba install -c conda-forge "openff-toolkit>=0.17" "openff-interchange>=0.4.6" "openff-nagl>=0.5.3" "openff-nagl-models>=2025.09" "openff-forcefields>=2025.10"

Then, in a workflow, you can load Sage 2.3.0 rc2 with

ff = ForceField("openff-2.3.0-rc2.offxml")

Scientific Details

AshGC v1.0 is our first neural network charge model. It has been trained to produce charges that are comparable to OpenEye AM1-BCC-ELF10. AshGC 1.0 is identical to the previously released AshGC rc3, except it adds support for single-atom charges. More information can be found in the documentation here. AshGC supports most small-molecule chemistries with elements C, O, H, N, S, F, Br, Cl, I, P; a list of excluded SMARTS patterns is included in the documentation.

Sage 2.3.0-rc2 is the second force field candidate in Open Force Field using AshGC v1.0. Both vdW and valence parameters have been re-fit, using AshGC charges, to a dataset of physical properties and QM data.

Sage 2.3.0-rc2 is applicable to drug-like molecules consisting of the elements C, H, O, N, P, S, F, Cl, Br, and I, atomic Xe, and the monoatomic ions Li+, Na+, K+, Rb+, Cs+, F-, Cl-, Br-, and I-. Compared to Sage 2.2.1, Sage 2.3.0-rc2 splits torsions so that a single torsional parameter only covers a central bond with a single multiplicity. It also adds some bond and angle types to improve performance of targeted chemistries, such as small rings.