Work

Sparse Supervision Turns Polar Pretraining into Transferable Molecular Response Fields

Year: 2026
Type: preprint
Abstract: Molecular neighbours redistribute electrons, changing the potential experienced by nearby matter. A molecular dipole summarizes this response but does not locate it. We ask whether polar pretraining can be redirected towards the spatial interaction-induced response with only tens of labelled configurations. Geometry-Learned Induced Dipole and Electrostatic Response (GLIDER) learns the fixed-geometry, complex-minus-fragments potential through constrained atom-centred charges and dipoles. Its 48 training configurations span 14 solutes: 24 have three water neighbours and 24 have four. One fixed model was then tested prospectively on 56 response-unseen solutes in 224 water environments, reducing response-potential error by 32–40% relative to the strongest eligible public polar baseline used without response-specific fitting. A separate post-audit follow-up on 36 constructed attractive contacts with three neutral non-water neighbours retained a 34.9% field-error reduction with the same fixed checkpoint. In a post hoc ten-solute test, the predicted field reduced held-out-water frozen response-coupling energy mean absolute error from 0.114 kcal mol−1 for that baseline to 0.049 kcal mol−1, with corresponding near-field torque and orientation gains. Together, these tests show that sparse response supervision can turn polar pretraining into a transferable spatial field with measurable local electrostatic utility. Code and data: https://github.com/Scientific-Computing-Lab/GLIDER_AI
Language: English
Open Access status: green
OpenAlex ID: W7235665095
Cites: 33
Cited by:
Related to: 10