Molecule-aware machine learning for predicting absorption and emission of benzothiadiazole derivatives

Abstract Predicting the absorption and emission wavelengths of 2,1,3-benzothiadiazole (BTD) derivatives across diverse solvents from molecular structure alone remains challenging, and conventional random-split validation can obscure whether models truly extrapolate to new chemical structures or merely interpolate solvent effects for known ones. By enforcing a molecule-aware train–test split, in which all solvent measurements of any given chemical structure are held out together, we show that champion-model $$R^2$$ falls by approximately 27–29 percentage points relative to random splitting, revealing that high random-split accuracy primarily reflects reliable prediction of solvent-induced spectral shifts rather than generalisation to unseen BTD structures. Within this framework, replacing binary with count-based Morgan fingerprints raises single-split $$R^2$$ from 0.747 to 0.941 for absorption and from 0.797 to 0.916 for emission, an improvement rooted in the ability of count fingerprints to encode the recurrence frequency of donor substructures, a key determinant of intramolecular charge-transfer character. A systematically optimised deep neural network reaches $$R^2 = 0.954$$ for absorption on its best of twenty random initialisations (mean $$0.83 \pm 0.12$$ ), the highest value reported for this dataset, yet the molecule-aware benchmark confirms that this performance does not translate to structure-level extrapolation. Comparative evaluation of graph neural networks, multi-task learning, extended descriptor sets, and ensemble methods confirms that none outperforms the count-fingerprint approach at this dataset size. Feature-importance analysis identifies tertiary amine substitution as the dominant structural driver and solvent polarity as the strongest single feature for emission (ranked first or second in all four model/encoding combinations examined, versus 8th–13th for absorption), an asymmetry consistent with the stronger solvatochromism of relaxed charge-transfer states. These results offer practical guidance for computational screening: the models are well suited to solvent selection for known dyes, whereas the design of genuinely new structures demands expanded structural diversity in the training set.