The code, frozen model and evidence behind Sparse Supervision Turns Polar Pretraining into Transferable Molecular Response Fields
Gal Oren · Boris Fain · Michael Levitt
Accepted at the 2nd SIMBIOCHEM Workshop at NeurIPS 2026
The idea · The evidence · Explore the data · Reproduce · Citation
Exact QM fields from paper Fig. 1c. Blue marks negative response potential and red positive response potential; the molecular drawings and field layers are preserved from the paper source.
A neighbouring molecule does not experience a dipole magnitude. Its atoms sample an electrostatic potential at particular locations. The two water arrangements above have nearly the same induced-dipole magnitude (0.464 and 0.468 D), yet place their negative-potential regions differently. Their full dipole vectors need not be equal.
GLIDER (Geometry-Learned Induced Dipole and Electrostatic Response) learns the interaction-induced change in that potential. At fixed geometry, the QM target is the potential of the complex minus the potentials of its separately evaluated fragments, all in the same counterpoise-consistent basis:
The model represents this response with atom-centred charges and dipoles. Its output conserves total response charge and reconciles the site moment with a separately predicted molecular response dipole. The negative gradient of the potential gives the electric field. This is the mutual response of the fragments, not a unique decomposition into polarization and charge transfer.
The recipe is deliberately small: 48 response-labelled solute-plus-three-water environments, spanning 14 chemistries, train only the GLIDER response head. Geometry-dependent features from MACE-POLAR-1-M and a starting response built from frozen MACE-POLAR-1-M/-1-L predictions supply the pretrained inputs. Neither foundation checkpoint is fitted to these response labels. We use independently evaluated, response-unfitted MACE-POLAR-1-L as the polar baseline.
The same frozen checkpoint was evaluated prospectively on 56 solutes absent from GLIDER response supervision, in 224 three-water environments. Exposure to these identities during broad foundation-model pretraining is unknown. Response-ESP error fell by 32–40% relative to the polar baseline across three separately acquired panels; 54 of 56 solutes improved.
The strongest result is spatial response transfer. Global response-dipole gains are less decisive, and Panel I did not pass a broader preregistered gate requiring both ESP and dipole improvements. Those cases and the frozen checkpoint remain in the release. See the three panels and paired intervals →
In a separate 12-solute test (72 cases), the same model was evaluated with NH₃, CH₃OH and CH₃CN environments. Response-ESP NRMSE was 0.450 for GLIDER versus 0.617 for the polar baseline, a 27.0% reduction. Both neighbour identity and environment size changed, so this is a transfer test rather than an isolated substitution experiment. Inspect the per-species data →
A fourth water probes a response computed for a solute plus W1–W3. W4 is absent from both the model input and the base-response QM calculation. At W4, the frozen, one-way response-coupling energy MAE is 0.049 kcal mol⁻¹ for GLIDER versus 0.114 kcal mol⁻¹ for the polar baseline.
This is a Coulomb response component, not total interaction or solvation energy and not self-consistent polarization. An exact QM response dipole at one predefined origin is a useful global-moment control in the near field. Higher exact multipoles recover with distance and eventually outperform GLIDER in the fixed-charge far-field test. See the distance control →
Every figure and table in the paper has a paper-to-data map. The underlying CSVs retain their source filenames, which sometimes differ from the final paper's figure numbers; the map resolves them.
| Start here | What you can inspect |
|---|---|
| Prospective results | Figure and panel mapping, 56-solute aggregate and paired comparisons, plus frozen output tables. |
| Geometries, predictions and QM references | The three prospective panels and the camera-ready numeric source tables. |
| Frozen checkpoint | The five-member response head and its training manifest. |
| Chronology and provenance | Prediction-before-reference manifests, byte-preserved source code, hashes and the retained Panel-I gate result. |
| Use on a geometry | Inference wrapper, example input and third-party checkpoint setup. |
The Figure 1c image layers are exact source assets. The other gallery graphics are reading aids: quantitative charts are rendered from released CSVs, while the architecture diagram is conceptual. No manuscript PDF or LaTeX package is included here.
The lightweight path verifies the released checkpoint and archives and recomputes prospective aggregate statistics; it does not rerun quantum chemistry or change the frozen model. Python 3.11+ is recommended.
git clone https://github.com/Scientific-Computing-Lab/GLIDER_AI.git
cd GLIDER_AI
python -m venv .venv
source .venv/bin/activate
python -m pip install -e .
python scripts/reproduce/verify_release.py
python scripts/reproduce/recompute_statistics.py --root . --output /tmp/glider-statistics
python scripts/verify_companion.pyFor a new geometry, install the model extra, the paper-matched MACE source, and its official MACE-POLAR-1 M/L checkpoints. The inference guide gives the complete command; third-party weights are fetched from their publisher and are not redistributed here. To regenerate the visual gallery, install .[figures] and run python scripts/render_figures.py and python scripts/render_readme_field.py.
Please cite Gal Oren, Boris Fain and Michael Levitt, “Sparse Supervision Turns Polar Pretraining into Transferable Molecular Response Fields,” 2nd SIMBIOCHEM Workshop at NeurIPS 2026. Machine-readable details are in CITATION.cff. Software is MIT licensed; third-party models retain their own licences. Correspondence: galoren@stanford.edu and levittm@stanford.edu.
Scope: the tested systems are neutral, closed-shell organic solutes with the specified molecular neighbours. Ions, arbitrary condensed phases, self-consistent embedding, complete interaction energies and molecular-dynamics performance remain outside the evidence in this release.