State-Constrained SHAP Distills Polymer Property Models into Transferable Structural Rules for Fixed-Interval Screening
- 1Department of Electrical Engineering & Information Systems, Graduate School of Engineering, University of Tokyo, Tokyo 113-8656, Japan
- 2Department of Chemistry, University College London, London WC1H 0AJ, United Kingdom
Abstract
Polymer informatics has accelerated data-driven prediction of polymer properties, yet accurate prediction does not directly enable discovery. Here, we introduce state-constrained SHAP (SC-SHAP), an attribution-to-rule distillation framework for fixed-interval polymer screening. Using XGBoost as the teacher, SC-SHAP reorganizes SHAP and SHAP-interaction attributions on chemically interpretable MACCS fingerprints into support-qualified structural states defined by required-present and required-absent substructures. These states provide auditable property ordering, executable structural conditions, and candidate-level rule traces, while the learned library is reused across target windows with interpolation-only boundary calibration. Evaluation on experimental PoLyInfo datasets for glass-transition temperature (T g), tensile stress at break (TSS), and electrical conductivity (EC) yields Spearman correlations between state scores and mean measured properties of 0.908 ± 0.008, 0.817 ± 0.069, and 0.835 ± 0.025, respectively. Benchmarks include direct XGBoost, MACCS-GCN, MACCS-Transformer, and rule-based screening. In the fixed-interval screening benchmark, SC-SHAP achieves higher mean scores than direct XGBoost, MACCS-GCN, MACCS-Transformer, and WRAcc on the smaller TSS and EC datasets. Multiple seeds and alternative structural splits further assess transfer to structurally distinct polymers. By unifying nonlinear knowledge distillation, chemically explicit rules, and cross-window reuse, SC-SHAP provides an interpretable screening framework spanning thermal, mechanical, and electrical polymer properties.
Related
- Development of a Deep Reinforcement Learning Approach for Transition State Search in Molecular Systems: A Transferable Model Across Structural and Dimensional Variations
- Efficient and Transferable Machine Learning Potentials for Zn-Coordinated Zeolitic Imidazolate Frameworks Through Structural Diversity Sampling and Transfer Learning.
- Insights into the Structural-Transport Property Relationship of the Size-Dependent Organic Molecules on Metal Surfaces
- Physical property-based screening of polyurethane foam for gold enrichment from rock gold ores
- Physics-Constrained Conformal Screening: A Certified Machine Learning Framework for the Rational Design of Dual Glycoepitope-Imprinted Polymers Recognizing CA 19-9 and MUC1-Tn
BibTeX
@article{wang2026stateconstrained,
title = {{State-Constrained SHAP Distills Polymer Property Models into Transferable Structural Rules for Fixed-Interval Screening}},
author = {Weihao Wang and Haoou Ruan and Keith T. Butler and Akiko Kumada and Masahiro Sato},
journal = {Journal of Chemical Information and Modeling},
year = {2026},
month = oct,
doi = {10.1021/acs.jcim.6c02120},
url = {https://doi.org/10.1021/acs.jcim.6c02120},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - State-Constrained SHAP Distills Polymer Property Models into Transferable Structural Rules for Fixed-Interval Screening
AU - Weihao Wang
AU - Haoou Ruan
AU - Keith T. Butler
AU - Akiko Kumada
AU - Masahiro Sato
PY - 2026
DA - 2026/10/08
JO - Journal of Chemical Information and Modeling
DO - 10.1021/acs.jcim.6c02120
UR - https://doi.org/10.1021/acs.jcim.6c02120
AB - Polymer informatics has accelerated data-driven prediction of polymer properties, yet accurate prediction does not directly enable discovery. Here, we introduce state-constrained SHAP (SC-SHAP), an attribution-to-rule distillation framework for fixed-interval polymer screening. Using XGBoost as the teacher, SC-SHAP reorganizes SHAP and SHAP-interaction attributions on chemically interpretable MACCS fingerprints into support-qualified structural states defined by required-present and required-absent substructures. These states provide auditable property ordering, executable structural conditions, and candidate-level rule traces, while the learned library is reused across target windows with interpolation-only boundary calibration. Evaluation on experimental PoLyInfo datasets for glass-transition temperature (T g), tensile stress at break (TSS), and electrical conductivity (EC) yields Spearman correlations between state scores and mean measured properties of 0.908 ± 0.008, 0.817 ± 0.069, and 0.835 ± 0.025, respectively. Benchmarks include direct XGBoost, MACCS-GCN, MACCS-Transformer, and rule-based screening. In the fixed-interval screening benchmark, SC-SHAP achieves higher mean scores than direct XGBoost, MACCS-GCN, MACCS-Transformer, and WRAcc on the smaller TSS and EC datasets. Multiple seeds and alternative structural splits further assess transfer to structurally distinct polymers. By unifying nonlinear knowledge distillation, chemically explicit rules, and cross-window reuse, SC-SHAP provides an interpretable screening framework spanning thermal, mechanical, and electrical polymer properties.
SN - 1549-9596
SN - 1549-960X
ER -