When Do Simple Models Win? Machine Learning Architectures for UV Absorption Prediction

Abstract

What determines which machine learning architecture wins for molecular property prediction? We address this question through a controlled comparison of five models spanning four architecture families: fingerprint-based ensembles (Random Forest, XGBoost), a directed message-passing graph neural network (GNN), a bidirectional gated recurrent network (BiGRU), and a pretrained Transformer, using UV absorption wavelength (λmax) prediction as a testbed across 18,415 solute–solvent pairs (after applying the Greenman/Song duplicate-handling protocol; see Methods) with stratified 5-fold cross-validation and two external data sets totaling over 40,000 molecules. The answer depends on the task. For screening within known chemical space, RF with Morgan fingerprints is competitive with the best deep learning model at a fraction of the compute cost: RF and the D-MPNN are statistically indistinguishable (RMSE = 31.50 ± 1.47 vs 33.15 ± 3.27 nm, p = 0.16), and RF trains in 15 min on a CPU with only lightweight grid-search tuning (no GPU) while outperforming all sequence-based models. For exploring novel scaffolds, deep learning is superior: experimental validation on 16 novel UV-absorbing compounds shows that the GNN (MAE = 26.2 nm), Transformer (26.3 nm), and BiGRU (28.6 nm) all outperform RF (38.5 nm), as learned representations generalize where fixed fingerprints cannot. In either case, encoding solvent identity improves all models: a simple SMILES/fingerprint concatenation strategy reduces RMSE by 21–32% on chromophores with multisolvent training coverage (60% of test records) and 6–8% across the full benchmark (p  <  0.02), without domain-specific descriptor engineering. Interpretability analysis confirms that both RF feature importance and BiGRU gradient saliency independently highlight the same chromophore motifs, and that architectures with local inductive bias systematically outperform global-attention Transformers on this locally determined property. Within UV absorption prediction, optimal model choice depends on the relationship between target compounds and available training data: simple fingerprint-based models are an efficient and accurate alternative to deep learning for interpolation within known chemical space, while learned representations from graphs, sequences, or pretraining generalize better to novel scaffolds. We further hypothesize that this pattern (local-bias architectures favored for properties governed by local molecular features) may extend to other locally determined molecular properties; validating this across property types is left as future work.