Geometry-Aware Molecular Representation Learning for Approximating NORMAN-Predicted Retention Time Indices in Nontargeted Screening

Abstract

Retention information provides chromatographic evidence complementary to mass spectrometric data in LC-HRMS nontargeted screening (NTS), yet many existing models rely mainly on 1D SMILES or 2D molecular graphs and make limited use of three-dimensional structure. Here, we developed ViSRTI (Vector-Interaction-Spatial Retention Time Index), a ViSNet-based geometry-aware framework that integrates 3D molecular representations with multimodal descriptors to model the precomputed positive- and negative-mode RTI labels in NORMAN SusDat. On fixed held-out test sets, ViSRTI-G outperformed all three baselines. Relative to the best-performing baseline for each metric, R2 increased by 0.022 (vs GNN-RT) and 0.025 (vs SMILES-BiLSTM), while MAE decreased by 5.39 and 6.56 RTI units (both vs SMILES-BiLSTM) in positive and negative modes, respectively. Incremental feature integration and SHAP-guided retraining identified ViSRTI-GP with KOC and LogP masked for positive mode and ViSRTI-GPT for negative mode, achieving R2 values of 0.9433 and 0.9258 and MAEs of 40.42 and 40.76 RTI units, respectively. Under Bemis–Murcko scaffold-disjoint 5-fold cross-validation, R2 remained at 0.9114 ± 0.0126 and 0.8937 ± 0.0117, indicating substantial agreement for unseen scaffold groups. Laboratory-specific RT-RTI calibration using 21 standards yielded R2 values of 0.8575 and 0.9525 in positive and negative modes. In a benzo[a]pyrene/benzo[e]pyrene case study, calibration-derived expected RTs differed from measured values by only 0.28 and 0.11 min, respectively, and reproduced the observed elution order. Overall, ViSRTI provides a geometry-aware surrogate for NORMAN-predicted RTI and enables laboratory-specific conversion of predicted RTI into complementary retention-time constraints for candidate evaluation in NTS.