Mitigating dataset imbalance in molecular maximum absorption wavelength prediction via self-supervised learning and scaffold diversification

Near-infrared (NIR) absorbing dyes, especially those with maximum absorption wavelengths (λ_max) in the NIR region, show promising potential for phototherapy and bioimaging. However, their limited representation in open-source datasets hinders the development of data-driven predictive models. To address this issue, the NIRExDs dataset was constructed by manually supplementing existing open-source datasets with 2805 entries extracted from the literature, resulting in a 2.3-fold and 19.2-fold increase in the number of dyes in the NIR-I and NIR-II spectral regions, respectively. The self-supervised learning model Uni-Mol demonstrated superior performance in predicting λ_max compared to traditional supervised models, particularly in the long-tailed and sparsely distributed NIR-II region, with a more significant MAE reduction compared to other spectral regions. Analysis based on the prediction results revealed that models trained on scaffold-split datasets can expose limitations related to insufficient molecular scaffold diversity. Inspired by this finding, an optimized dataset, NIRExDs-1, was developed by adding a small number of structurally similar compounds to NIRExDs dataset. Models trained on NIRExDs-1 exhibited improved predictive performance on both internal and external test sets compared to those trained on the original NIRExDs dataset. Furthermore, scaffold split showed significantly weaker capability in capturing solvent effects than random split, primarily due to its inability to fully utilize measurement data of similar molecules across different solvent environments. Finally, Uni-Mol reproduced experimentally observed red-shift trends under classical molecular design strategies, confirming its ability to capture generalizable chemical rules.