FluorCast: Solvent-aware 2D-3D feature fusion for predicting fluorophore photophysics
4 affiliations
- 1Chandra Family Department of Electrical and Computer Engineering Cockrell School of Engineering
- 2University of Texas at Austin
- 3Department of Chemistry
- 4Memorial University of Newfoundland
Abstract
Machine-learning (ML) models can accelerate the prediction of fluorescence properties, but their apparent performance depends strongly on molecular representation and on the evaluation strategy. Here we present FluorCast, a solvent-aware ML framework assembled from three public datasets—ChemFluor, Deep4Chem, and FluoDB-Lite—comprising 66,820 standardized records, 33,965 unique chromophores, and 1,368 nonmissing canonical environment identifiers. Conventional Morganfingerprint models were benchmarked for absorption and emission maxima, photoluminescence quantum yield (PLQY), and Stokes shift under random, molecule-grouped, and scaffold-grouped splits, revealing substantial performance degradation with increasing structural novelty. To incorporate three-dimensional (3D) molecular information, we evaluated frozen 512-dimensional ConforFormer embeddings alone and in feature-level fusion with Morgan fingerprints. On matched test sets, fusion reduced the mean absolute error (MAE) relative to Morgan fingerprints alone in 11 of 12 target-split combinations, although molecule-grouped PLQY prediction worsened and several scaffold-grouped gains were small. An external set of five recently reported fluorophores with no exact training-set match provided a further test; fusion reduced absorption MAE from 65.67 to 24.49 nm, whereas Morgan-only models performed better for emission and Stokes shift, and PLQY performance was comparable. These results indicate that learned 3D conformational embeddings provide complementary, target-dependent information rather than a universal replacement for two-dimensional (2D) molecular fingerprints.
Related
- Predicting Adsorbate-Solvent Interaction Energies in Zeolite Pores Using Convolutional Neural Networks with Attention Mechanisms.
- Dual Fluorophore System for Direct Measurement of Acidity of Aerosol Particles
- One-Electron Modulation of Photophysics in a Through-Space Donor–Acceptor System
- Fluorophore-Conjugated Diphenylalanine Peptides Reveal Self-Sorting in a Multicomponent Supramolecular System
- Solution structure of a peptide-conjugated aggregation-induced-emission (AIE) fluorophore bound to a protein
BibTeX
@misc{leung2026fluorcast,
title = {{FluorCast: Solvent-aware 2D-3D feature fusion for predicting fluorophore photophysics}},
author = {Chris Leung and Yuming Zhao},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010685/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010685/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - FluorCast: Solvent-aware 2D-3D feature fusion for predicting fluorophore photophysics
AU - Chris Leung
AU - Yuming Zhao
PY - 2026
DA - 2026/10/19
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010685/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010685/v1
AB - Machine-learning (ML) models can accelerate the prediction of fluorescence properties, but their apparent performance depends strongly on molecular representation and on the evaluation strategy. Here we present FluorCast, a solvent-aware ML framework assembled from three public datasets—ChemFluor, Deep4Chem, and FluoDB-Lite—comprising 66,820 standardized records, 33,965 unique chromophores, and 1,368 nonmissing canonical environment identifiers. Conventional Morganfingerprint models were benchmarked for absorption and emission maxima, photoluminescence quantum yield (PLQY), and Stokes shift under random, molecule-grouped, and scaffold-grouped splits, revealing substantial performance degradation with increasing structural novelty. To incorporate three-dimensional (3D) molecular information, we evaluated frozen 512-dimensional ConforFormer embeddings alone and in feature-level fusion with Morgan fingerprints. On matched test sets, fusion reduced the mean absolute error (MAE) relative to Morgan fingerprints alone in 11 of 12 target-split combinations, although molecule-grouped PLQY prediction worsened and several scaffold-grouped gains were small. An external set of five recently reported fluorophores with no exact training-set match provided a further test; fusion reduced absorption MAE from 65.67 to 24.49 nm, whereas Morgan-only models performed better for emission and Stokes shift, and PLQY performance was comparable. These results indicate that learned 3D conformational embeddings provide complementary, target-dependent information rather than a universal replacement for two-dimensional (2D) molecular fingerprints.
ER -