Transfer Protocol Matters More Than Source-Task Selection in Low-Data Molecular Property Prediction: An Audit of QM9 Pretraining on Computed Oxidation Potentials and Redox Reaction Energies
Abstract
Pretraining a network on QM9 quantum-chemical labels and transferring it to a data-scarce molecular endpoint is a common strategy in chemical machine learning, and a growing literature proposes tasksimilarity descriptors for deciding which source task to transfer from. We audit both decisions on two non-synthetic computational chemistry datasets — DFT-derived oxidation potentials (OxPot, 15,238 molecules) and redox reaction energies (RedDB, 15,551 molecules) — using 24 independently initialised QM9 source models spanning eight quantum-chemical tasks. We first establish, in a controlled synthetic benchmark with known informative coordinates, that magnitude-pruning masks can be aggregated into interpretable task-level feature-usage signatures that recover ground-truth feature overlap (pooled mean Spearman 0.2749, versus 0.1755 for unpruned first-layer weights and 0.0614 for a random control). Eleven signature series were compared; their relative ordering proved sensitive to pooling, with gradient saliency ahead in the unmatched pooled summary (0.2855) and behind by 0.211 in a paired same-setting audit. This mechanism did not translate into deployable source selection: on the two molecular datasets every selector produced negative mean transfer gain under full-weight transfer (feature-similarity top-1: −0.124, negative-transfer rate 95.8%; pseudo-label top-1: −0.177, 100%), and an agreement gate limited harm mainly by abstaining in 87.5% of cases. We then evaluated 144 matched case–source pairs under four transfer protocols (576 transfer runs). At the target-case level, resetting the output head improved relative gain by 0.087 (case-bootstrap 95% CI 0.070–0.106; 18/18 cases positive), and restricting transfer to the first layer added 0.044 (95% CI 0.034–0.055; 18/18 positive). Protocol changes produced the largest and most consistent shifts in realised gain. Only under the most conservative protocol did any dataset reach non-negative mean transfer (OxPot, +0.003; RedDB remained negative). For this setting, transfer protocol and an explicit abstention option deserve priority over increasingly elaborate sourcesimilarity descriptors.
Related
- Geometry-Aware Line Graph Transformer Pretraining for Molecular Property Prediction
- Accurate and Rapid Vibrational Spectra of Molecular Crystals from Machine-Learned Interatomic Potentials
- Detection of Molecular Oxidation States in Dye Sensitized Solar Cells-Relevant Organic Dyes Eosin Y Through Optical UV/Vis Absorption Spectrum Changes
- Galvanic Coupling Links Heterogeneous and Molecular Palladium Catalysis in Aerobic Alcohol Oxidation
- Prediction of aluminium distributions in zeolites using basin-hopping and machine learning potentials
BibTeX
@misc{zhang2026transfer,
title = {{Transfer Protocol Matters More Than Source-Task Selection in Low-Data Molecular Property Prediction: An Audit of QM9 Pretraining on Computed Oxidation Potentials and Redox Reaction Energies}},
author = {Yueru Zhang},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008164/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008164/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Transfer Protocol Matters More Than Source-Task Selection in Low-Data Molecular Property Prediction: An Audit of QM9 Pretraining on Computed Oxidation Potentials and Redox Reaction Energies
AU - Yueru Zhang
PY - 2026
DA - 2026/09/01
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008164/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008164/v1
AB - Pretraining a network on QM9 quantum-chemical labels and transferring it to a data-scarce molecular endpoint is a common strategy in chemical machine learning, and a growing literature proposes tasksimilarity descriptors for deciding which source task to transfer from. We audit both decisions on two non-synthetic computational chemistry datasets — DFT-derived oxidation potentials (OxPot, 15,238 molecules) and redox reaction energies (RedDB, 15,551 molecules) — using 24 independently initialised QM9 source models spanning eight quantum-chemical tasks. We first establish, in a controlled synthetic benchmark with known informative coordinates, that magnitude-pruning masks can be aggregated into interpretable task-level feature-usage signatures that recover ground-truth feature overlap (pooled mean Spearman 0.2749, versus 0.1755 for unpruned first-layer weights and 0.0614 for a random control). Eleven signature series were compared; their relative ordering proved sensitive to pooling, with gradient saliency ahead in the unmatched pooled summary (0.2855) and behind by 0.211 in a paired same-setting audit. This mechanism did not translate into deployable source selection: on the two molecular datasets every selector produced negative mean transfer gain under full-weight transfer (feature-similarity top-1: −0.124, negative-transfer rate 95.8%; pseudo-label top-1: −0.177, 100%), and an agreement gate limited harm mainly by abstaining in 87.5% of cases. We then evaluated 144 matched case–source pairs under four transfer protocols (576 transfer runs). At the target-case level, resetting the output head improved relative gain by 0.087 (case-bootstrap 95% CI 0.070–0.106; 18/18 cases positive), and restricting transfer to the first layer added 0.044 (95% CI 0.034–0.055; 18/18 positive). Protocol changes produced the largest and most consistent shifts in realised gain. Only under the most conservative protocol did any dataset reach non-negative mean transfer (OxPot, +0.003; RedDB remained negative). For this setting, transfer protocol and an explicit abstention option deserve priority over increasingly elaborate sourcesimilarity descriptors.
ER -