Prediction of Singlet-Triplet Gaps Using Self-Attention–Based Regression with Model Integration into a Molecular Editor
Abstract
The singlet–triplet gap (STG) is an important property of photofunctional materials, but obtaining reliable values can be challenging. Experimentally, its determination often requires a combination of advanced spectroscopic techniques, while computational approaches can be time-consuming and demand specialized expertise. Therefore, a low-cost and user-friendly approach for obtaining reasonable STG estimates could make this property more accessible to non-experts. In this work, we evaluate regression models based on different molecular representations using the INVEST dataset from the Aspuru-Guzik group that contains ADC(2) energy gaps for ~70K molecules. We compare transformer models operating on SMILES representations, graph-based models (AttentiveFP), and fingerprint-based models. While AttentiveFP performs best for molecules similar to those in the training set, pre-trained transformers generalize better to unseen chemical space after fine-tuning. Motivated by these complementary strengths, we integrate our best-performing models into an online 2D chemical structure editor, enabling users to obtain STG estimates directly from drawn molecular structures. The editor can be accessed at http://ketcher-predict.ttk.hu/.
Related
- Weighted Views Representation with Multi-Head Coulomb Attention for Molecular Property Prediction
- Molecular Property Prediction with Well-Calibrated Uncertainties via Graph Neural Networks and Conformal Quantile Regression
- Measuring molecular singlet oxygen (1O2*) from atmospheric photosensitizers: Intercomparison of techniques, irradiation setups, data analysis and protocol recommendations
- Graph attention framework for site-selectivity prediction in undirected C(sp3)–H borylation
- Integration of Network Toxicology and Machine Learning to Prioritize Candidate Molecular Features Associated with NNK-Related Lung Cancer
BibTeX
@misc{feher2026prediction,
title = {{Prediction of Singlet-Triplet Gaps Using Self-Attention–Based Regression with Model Integration into a Molecular Editor}},
author = {Péter Pál Fehér and András Stirling and Andras Stirling},
howpublished = {ChemRxiv},
year = {2026},
month = aug,
doi = {10.26434/chemrxiv.15008084/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008084/v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Prediction of Singlet-Triplet Gaps Using Self-Attention–Based Regression with Model Integration into a Molecular Editor
AU - Péter Pál Fehér
AU - András Stirling
AU - Andras Stirling
PY - 2026
DA - 2026/08/31
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008084/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008084/v1
AB - The singlet–triplet gap (STG) is an important property of photofunctional materials, but obtaining reliable values can be challenging. Experimentally, its determination often requires a combination of advanced spectroscopic techniques, while computational approaches can be time-consuming and demand specialized expertise. Therefore, a low-cost and user-friendly approach for obtaining reasonable STG estimates could make this property more accessible to non-experts. In this work, we evaluate regression models based on different molecular representations using the INVEST dataset from the Aspuru-Guzik group that contains ADC(2) energy gaps for ~70K molecules. We compare transformer models operating on SMILES representations, graph-based models (AttentiveFP), and fingerprint-based models. While AttentiveFP performs best for molecules similar to those in the training set, pre-trained transformers generalize better to unseen chemical space after fine-tuning. Motivated by these complementary strengths, we integrate our best-performing models into an online 2D chemical structure editor, enabling users to obtain STG estimates directly from drawn molecular structures. The editor can be accessed at http://ketcher-predict.ttk.hu/.
ER -