Hybrid Graph Neural Network Ensemble with Stacked Meta-Learning for Laboratory-Conditioned Chromatographic Retention-Time Prediction

Abstract

Abstract Accurate prediction of chromatographic retention time (RT) is valuable for compound identification in untargeted metabolomics and forensic toxicology, yet validation must account for repeated molecular structures, chemical-space coverage, and heterogeneous laboratory sources. Here we present a hybrid ensemble that combines Graph Attention Networks (GAT), Graph Convolutional Networks (GCN), and a fingerprint neural network (FPNN) with an ExtraTrees meta-learner and learned laboratory embeddings. The HighResNPS-derived modeling table contains 3776 structure–laboratory observations from 23 forensic toxicology laboratories, corresponding to 1357 InChIKey connectivity groups. Under molecular-identity-grouped outer validation repeated three times, the final ensemble achieved R2 = 0.977 ± 0.002 and MAE = 0.360 ± 0.016 min; under scaffold-aware validation, it achieved R2 = 0.928 ± 0.004 and MAE = 0.670 ± 0.013 min. Paired cluster-bootstrap comparisons show that combining the three networks lowers MAE versus each alone, whereas the meta-learner improves on simple averaging only for identity-grouped prediction. Prospective maximum-Tanimoto diagnostics define structural similarity to the development set, while molecule-aware perturbation analysis is reported as the sensitivity of the GAT component rather than a mechanism of the complete ensemble. The resulting scope is laboratory-conditioned RT prediction for new chemical groups within the 23 represented laboratory labels. A browser-based demonstration is available at https://huggingface.co/spaces/AI4deeperScience/chromatography-rt-predictor.