Molecular absorption and fluorescence property prediction via a multimodal consensus model: Winning solution for the transmittance category of the 2nd EUOS/SLAS Joint Challenge.

The 2nd EUOS/SLAS Joint Challenge was a competition aimed at developing reliable computational models for predicting transmittance and fluorescence properties from the chemical structures of approximately 100,000 compounds. This paper describes the method that achieved 1st place in the blind test of the Transmittance category. Our approach is based on a consensus strategy that integrates diverse predictions from multiple model types, including gradient boosting decision trees, message passing neural networks (MPNNs), and Uni-Mol2, via a weighted ensemble, covering multimodal molecular representations in 1D, 2D, and 3D. 5-fold cross-validation based on Bemis-Murcko scaffolds was employed to improve generalizability, and focal loss was adopted to handle extreme class imbalance. Post hoc comparison showed that the individual model leading under cross-validation differed from the Public-test leader in all four subtasks, whereas the final ensemble ranked between first and fifth across tasks, supporting the robustness of the consensus strategy to model selection.