Data-Driven Prediction of Ultraviolet-Visible Spectra From Molecular Structure Using Attention-Based Graph Neural Networks.

Predicting UV-vis spectra from molecular structure is an important but relatively underexplored task. In this study, we developed a deep learning approach based on graph neural networks (GNNs) to directly predict UV-vis spectra from SMILES representations. Several GNN architectures were evaluated, among which Attentive Fingerprint achieved the best performance with an R2 of 0.8630, mean absolute error of 0.0389, root mean squared error of 0.0740, cosine similarity of 0.9589, Pearson correlation coefficient of 0.9521, and distance metric of 1.5758 on the test set. Further analysis showed that the model might be able to capture meaningful structure-property relationship across diverse chemical classes. The model was also evaluated on external datasets, where it maintained good predictive performance, indicating its generalizability. These results demonstrate that attention-based GNNs provide an effective approach for UV-vis spectrum prediction and may offer useful insights into the molecular regions contributing to model predictions, with potential applications in molecular analysis, high-throughput screening, and compound design. To facilitate practical use, we deployed the model as an interactive web application (https://spectra-prediction.streamlit.app/), which enables direct prediction of UV-vis spectra from SMILES inputs.