Sparse supervision · spatial fields · local physical reuse
Gal Oren · Boris Fain · Michael Levitt
Accepted at the 2nd SIMBIOCHEM Workshop at NeurIPS 2026
Understand the model · Explore an experiment · Reproduce a result · Model limits
Two water arrangements. Almost the same response-dipole magnitude. Different places for a neighbour to interact.
A molecule changes the electronic environment around it when another molecule approaches. A molecular dipole summarizes part of that change. A response field tells us where it occurs.
GLIDER learns the electrostatic potential associated with this rearrangement. It starts from pretrained polar representations and learns from 48 labelled configurations of 14 solutes: 24 have three water neighbours and 24 have four. The response head is then frozen before the primary transfer tests.
The target is a fixed-geometry difference: evaluate the complex and its separated fragments in the same quantum-chemical basis, then subtract.
This potential describes the mutual electronic response of solute and environment. GLIDER represents it with atom-centred charges and dipoles. The site charges sum to zero, and their combined moment agrees with a separately predicted response dipole.
MACE-POLAR-1-M supplies the frozen geometry features. An equal-weight average of MACE-POLAR-1-M and -1-L complex-minus-fragments predictions supplies the starting response. In comparisons, independently evaluated MACE-POLAR-1-L is the polar baseline. It is one contributor to the prior, rather than an independent physics model. How training and inference work →
| Question | Evidence | Read the data |
|---|---|---|
| Does it transfer to new solutes? | 32–40% lower response-ESP error across 56 solutes and 224 configurations. | Panels I–III |
| Does it extend beyond water? | The original 72 replacements include 62 repulsive QM contacts. In a separate, prediction-frozen follow-up on 36 attractive constructed contacts, GLIDER has 34.9% lower response-ESP error than the unfitted polar baseline. | Original audit · Contact follow-up |
| Can another molecule use the field? | Held-out-water response-coupling MAE: 0.114 → 0.049 kcal mol⁻¹. | Held-out water |
| What does the global branch add? | A matched ablation finds a modest average benefit, with differences across solutes. | Global-branch control |
| Does the response vanish when the inducing fragments separate? | No. From 20 to 100 Å, GLIDER's predicted amplitude grows while its averaged frozen prior declines. | Figure S7 data · Separation protocol |
The non-water follow-up was designed after the original geometry audit, not as a second preregistered test. Its geometry and predictions were fixed before the new QM references, and every selected case is released.
The strongest evidence is for spatial response. Global-dipole gains are less consistent. Panel I missed its broader preregistered joint ESP-and-dipole gate, and all its cases remain available. Selection, label reuse and retained failures →
Compute the response of a solute and its first three waters, then bring in a fourth water. That fourth water enters neither the response-model input nor the base-reference calculation. Its interaction with the frozen field tests whether the spatial prediction is useful beyond its fitting observable.
An exact QM response dipole has an exact global moment, but its single-origin field can be inaccurate nearby. Higher multipoles recover with distance and eventually outperform GLIDER. This experiment measures a frozen response contribution to Coulomb coupling. It does not rank complete force fields by total interaction energy. Explore distance and orientation →
The release is organized around experiments. Each guide points directly to its geometries, references, predictions and results. The paper map connects the current figure and table numbers to these folders.
experiments/
training/ Original 48 configurations, labels and frozen features
panel_1/ 12 solutes · 48 configurations
panel_2/ 24 solutes · 96 configurations
panel_3/ 20 solutes · 80 configurations
nonwater/ 12 solutes · 72 unrelaxed replacement complexes
nonwater_contact/ 12 solutes · 36 separately selected contacts
liquid/ 6 solutes · 24 liquid-derived clusters
shell_size/ 3 solutes · nested 1/3/6/12-water clusters
heldout_water/ 10 solutes · densities, probe states and coupling
heldout_water_pilot/ Separate original six-solute pilot
distance_sweep/ Probe positions, fields, densities and raw evaluations
global_branch/ Matched training ablation and predictions
dissociation/ Fragment-separation test at 3–20 Å
Experiment catalogue · Array formats and units · Current figure data · Frozen checkpoint · Historical provenance
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/score_experiment.py --experiment all
python scripts/reproduce/recompute_statistics.py --output build/statistics
python scripts/verify_companion.pyThis path needs no GPU or new quantum-chemistry calculation. It verifies file hashes and constraints and computes errors from the released reference and prediction arrays. Results go into build/, leaving the archive unchanged.
The paired-interval command uses the manuscript's 100,000-resample seed (20260816), reproducing the primary prospective intervals plotted in the article. Earlier frozen-panel provenance records retain their original seed (20260814); they document the historical analysis rather than the article's harmonized intervals.
For a physical calculation from density matrices, install .[qm] and run:
python scripts/reproduce/recompute_coupling.py --solute methane --water-rank 4The CPU recomputation agrees with the archived QM, GLIDER and baseline coupling values to better than 5 × 10⁻¹¹ kcal mol⁻¹. Reproduction levels and commands →
For prediction on a new geometry, follow the inference guide. It installs the paper-matched MACE implementation and downloads the official M/L checkpoints from their publisher.
The released checkpoint supports analysis and frozen local reuse for the compact, neutral systems tested here. It conserves net response charge and matches its predicted global moment. It does not enforce fragment separability. In a post hoc test, we moved either the sole water in a solute–water pair or an intact water cluster away from the solute. Between 20 and 100 Å, GLIDER's mean predicted response-potential RMS grew from 0.564 to 1.540 mEh/e for one water and from 0.925 to 2.843 mEh/e for the full cluster. The averaged frozen prior declined over both intervals. These are predicted amplitudes on fixed solute-centred probes, not errors against new QM labels. A per-case component diagnostic traces most of the long-range growth to global-moment reconciliation. Inspect Figure S7 or the 3–20 Å and 50–100 Å protocols.
A general molecular-simulation component needs a controlled dissociation limit and self-consistent coupling. Total-energy accuracy, ions and arbitrary condensed phases have not been established by these experiments.
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 citation.
Correspondence: galoren@stanford.edu · levittm@stanford.edu
Original release code is MIT licensed. Third-party software, weights and source structures retain their own licences. Sources and attribution. Generative AI assisted writing, code development, data processing and visual presentation. Scientific conclusions are tied to the calculations and archived records linked above.