Excited-State Photophysics Improves Scaffold-Robust Phototoxicity Prediction and Reveals Compound-Specific State Effects
- 1Stanford University, 415 Broadway, Redwood City, California, 94063
Abstract
Phototoxicity is controlled by light absorption and access to reactive excited states, but predictive models usually encode only a single ground-state structure. Here, a physically staged representation was evaluated on 450 compounds (192 positive, 258 negative) with 10 repeated five-fold scaffold-grouped cross-validations. Structure alone gave ROC-AUC 0.745 ± 0.018 and PR-AUC 0.668 ± 0.027. Adding parent-state xTB/sTDA photophysics increased these metrics to 0.792 ± 0.012 and 0.716 ± 0.020; microstate aggregation reached 0.807 ± 0.012 and 0.726 ± 0.017, while conformer expansion plateaued at 0.808 ± 0.010 and 0.729 ± 0.017. In 10,000 paired compound bootstraps, the largest robust ranking increment was the structure-to-parent-photophysics step (ΔROC-AUC +0.0457, 95% CI +0.0274 to +0.0643). Microstates produced a smaller rank-based increment but reduced Brier error by 0.0080 (95% CI 0.0024–0.0134). Higher-level Q-Chem calculations preserved S1/T1 ordering (r = 0.946/0.931). Matched-state calculations showed large photophysical changes for avobenzone and 2-ethylanthraquinone but weak changes for pradofloxacin and ketoprofen. The results identify excited-state information as the dominant reproducible addition beyond structure and show that molecular-state effects are chemically meaningful but heterogeneous rather than universally accuracy-enhancing.
Related
- Microstate-Aware Excited-State Descriptors Improve Scaffold-Robust Prediction of Drug Phototoxicity
- Plasma Protein Binding Prediction under Scaffold Split: Tree Models, ChemBERTa, Graph Networks, and Late Fusion
- Blood–Brain Barrier Permeability Prediction by Integrating Deep Learning Representations with Chemical Compound Knowledge
- Scaffold-Controlled Evaluation of Molecular Representations for Antimalarial Activity Prediction
- β-Sitosterol–Naringenin Co-Loaded Nanophytosomes: Encapsulation, Compound-Specific Release Kinetics, Machine Learning-Assisted Release Prediction, and Cell-Dependent Anticancer Activity
BibTeX
@misc{tsai2026excitedstate,
title = {{Excited-State Photophysics Improves Scaffold-Robust Phototoxicity Prediction and Reveals Compound-Specific State Effects}},
author = {Aris Tsai},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009615/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009615/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Excited-State Photophysics Improves Scaffold-Robust Phototoxicity Prediction and Reveals Compound-Specific State Effects
AU - Aris Tsai
PY - 2026
DA - 2026/09/29
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009615/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009615/v1
AB - Phototoxicity is controlled by light absorption and access to reactive excited states, but predictive models usually encode only a single ground-state structure. Here, a physically staged representation was evaluated on 450 compounds (192 positive, 258 negative) with 10 repeated five-fold scaffold-grouped cross-validations. Structure alone gave ROC-AUC 0.745 ± 0.018 and PR-AUC 0.668 ± 0.027. Adding parent-state xTB/sTDA photophysics increased these metrics to 0.792 ± 0.012 and 0.716 ± 0.020; microstate aggregation reached 0.807 ± 0.012 and 0.726 ± 0.017, while conformer expansion plateaued at 0.808 ± 0.010 and 0.729 ± 0.017. In 10,000 paired compound bootstraps, the largest robust ranking increment was the structure-to-parent-photophysics step (ΔROC-AUC +0.0457, 95% CI +0.0274 to +0.0643). Microstates produced a smaller rank-based increment but reduced Brier error by 0.0080 (95% CI 0.0024–0.0134). Higher-level Q-Chem calculations preserved S1/T1 ordering (r = 0.946/0.931). Matched-state calculations showed large photophysical changes for avobenzone and 2-ethylanthraquinone but weak changes for pradofloxacin and ketoprofen. The results identify excited-state information as the dominant reproducible addition beyond structure and show that molecular-state effects are chemically meaningful but heterogeneous rather than universally accuracy-enhancing.
ER -