Submitted:
13 November 2026
Posted:
16 November 2026
You are already at the latest version
Abstract
Predicted retention times can narrow candidate lists in liquid chromatography–high-resolution mass spectrometry (LC–HRMS). Deep-learning models report errors near the variability between repeated measurements, but their credibility depends on the evaluation design. We reviewed deep-learning retention-time prediction for non-polymeric small molecules in reversed-phase liquid chromatography (RP-LC). Searches of PubMed, Scopus, and Web of Science through August 3, 2026, identified 594 records; 33 studies were eligible. Fifteen studies reported mean absolute errors (MAEs) on the METLIN Small Molecule Retention Time (SMRT) dataset across at least seven differently filtered subsets. Among 13 models evaluated under a broad retained-only convention, the spread from best to worst was 18 s, compared to the 36 s mean repeat-measurement variability observed on SMRT reference standards. Unseen chromatographic methods were rarely tested; one model’s error rose from 29.3 s internally to 596.8 s on a held-out gradient. Random molecular partitions remained common, including in studies using cross-validation or transfer-set deduplication. Risk of bias was low in 18 studies for model development and 19 for performance evaluation, but in only seven for leakage control and one for generalizability. Reported accuracy depends as much on benchmark construction as on architecture. To support reliable compound identification, these models need structure-aware and method-aware validation with uncertainty estimates.
Keywords:
retention-time prediction
; quantitative structure–retention relationship (QSRR)
; reversed-phase liquid chromatography (RPLC)
; deep learning
; chemometrics
; metabolomics
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.