Microstate-Aware Excited-State Descriptors Improve Scaffold-Robust Prediction of Drug Phototoxicity
- 1Stanford Online High School, Stanford University, 415 Broadway, Redwood City, California 94063, United States
Abstract
Drug phototoxicity depends on electronically excited states whose energies and intensities can change with protonation, tautomerism, and conformation, yet most predictive models represent a single parent structure. We evaluated a molecular-state-aware workflow on a curated 450-compound in vitro photoirritation (PIV) benchmark expanded to 1,088 microstates. Across 10 repeated 5-fold scaffold-held-out cross-validations, mean ROC-AUC increased from 0.745 for structure alone to 0.792 after adding parent xTB+sTDA descriptors, 0.807 with microstate summaries, and 0.808 with conformer aggregation; PR-AUC increased from 0.669 to 0.727. Compound-paired bootstrap analysis gave a full-model ROC-AUC gain of +0.0569 (95% CI 0.0234-0.0908) and PR-AUC gain of +0.0533 (0.0133-0.0933). Higher-level Q-Chem calculations preserved screening excited-state ordering (S1 Pearson r = 0.946; T1 r = 0.931), and an additional preselected state panel completed 22/24 TDA jobs with all four parent-alternate microstate pairs available. These pairs showed substantial state-dependent changes, including a 240.4 to 259.0 nm bright-state shift for trinexapac-ethyl. Clinical PIH endpoint transfer was modest (n = 155; ROC-AUC 0.609), consistent with endpoint and chemical-space shift. The results support molecular-state-aware excited-state featurization as a mechanistically grounded strategy for computational photosafety.
Related
- Excited-State Photophysics Improves Scaffold-Robust Phototoxicity Prediction and Reveals Compound-Specific State Effects
- A Machine Learning Approach to logBB Prediction Combining Molecular Docking and Quantum-Chemical Descriptors
- Charge-Transfer-State Descriptors from Donor–Acceptor Interfaces Enable Improved Prediction of Indoor Organic Photovoltaic Efficiency
- Rotamer-Dependent Amino Acid Descriptors as Protein and Peptide Representation for Reactivity and Property Prediction
- Machine learning prediction of DFT-derived electrical conductivity in perovskite and double perovskite materials using physically motivated descriptors
BibTeX
@misc{tsai2026microstateaware,
title = {{Microstate-Aware Excited-State Descriptors Improve Scaffold-Robust Prediction of Drug Phototoxicity}},
author = {Aris Tsai},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009622/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009622/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Microstate-Aware Excited-State Descriptors Improve Scaffold-Robust Prediction of Drug Phototoxicity
AU - Aris Tsai
PY - 2026
DA - 2026/09/29
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009622/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009622/v1
AB - Drug phototoxicity depends on electronically excited states whose energies and intensities can change with protonation, tautomerism, and conformation, yet most predictive models represent a single parent structure. We evaluated a molecular-state-aware workflow on a curated 450-compound in vitro photoirritation (PIV) benchmark expanded to 1,088 microstates. Across 10 repeated 5-fold scaffold-held-out cross-validations, mean ROC-AUC increased from 0.745 for structure alone to 0.792 after adding parent xTB+sTDA descriptors, 0.807 with microstate summaries, and 0.808 with conformer aggregation; PR-AUC increased from 0.669 to 0.727. Compound-paired bootstrap analysis gave a full-model ROC-AUC gain of +0.0569 (95% CI 0.0234-0.0908) and PR-AUC gain of +0.0533 (0.0133-0.0933). Higher-level Q-Chem calculations preserved screening excited-state ordering (S1 Pearson r = 0.946; T1 r = 0.931), and an additional preselected state panel completed 22/24 TDA jobs with all four parent-alternate microstate pairs available. These pairs showed substantial state-dependent changes, including a 240.4 to 259.0 nm bright-state shift for trinexapac-ethyl. Clinical PIH endpoint transfer was modest (n = 155; ROC-AUC 0.609), consistent with endpoint and chemical-space shift. The results support molecular-state-aware excited-state featurization as a mechanistically grounded strategy for computational photosafety.
ER -