Integration of Transferable Prediction of Retention Index and Universal Library Search Enhances Exposome Identification Probability in RPLC/HRMS-Based Non-Targeted Analysis

In non-targeted analysis, spectral matching is one of the most commonly used approaches for annotation or identification of the chemicals of emerging concern in complex samples via high-resolution mass spectrometry (HRMS). Conventional confidence assessment systems struggle to resolve multiple hits during library searches, resulting in ambiguities in identification. In this study, a combination of information extracted from the MS/MS spectra and calibrate-free predicted retention indices (RIs) yielded the probability of true positive (TP) for each hit by machine learning (ML), including a pre-trained molecular fingerprint (MF)-to-retention index (RI) model, a cumulative neutral loss (CNL)-to-RI model trained by 693,681 spectra, and a binary classification model that incorporated 5 significant features from the universal library search algorithm and the difference between the predicted RIs from both RI models. Our results demonstrated a high correlation (R2 = 0.96 and 0.88) between MF-derived and CNL-derived RI values for the model training and testing datasets, respectively, suggesting reduced RI error for TP annotations. When applied to pesticides-spiked samples, the k-nearest neighbors algorithm achieved a weighted F1 score of 0.65 and a Matthews correlation coefficient of 0.30. Integrating ML with reference spectral match enhanced the identification probability of exposome by half in a proof-of-concept study in pesticide contaminants in black tea. This study highlights the potential of integrating ML with reference library matching to enhance the identification probability of contaminants.