Retention-Time Bridging across Gradients for LC-HRMS Feature Correspondence

Abstract

Untargeted liquid chromatography–high-resolution mass spectrometry (LC–HRMS) supports high-throughput metabolomics, but rapid gradients can compress chromatographic separation and limit downstream annotation. A practical annotation strategy is to analyze all samples using a rapid untargeted method and then reinject selected biological and pooled quality-control (QC) samples under a longer-gradient method to improve separation, MS/MS acquisition, and standard-based confirmation. Because short- and long-gradient feature tables cannot be matched reliably by m/z alone, we developed RTBridge, a same-sample, feature-table-based retention-time (RT) bridging workflow that establishes feature correspondence without requiring structural descriptors or quantitative structure–retention relationship models. Unique anchors were selected using within-method mass uniqueness, cross-method mass agreement, and an abundance-concordance score (ACS). Six empirical RT-transfer models were evaluated: linear regression, piecewise linear interpolation, locally estimated scatterplot smoothing (LOESS), piecewise cubic Hermite interpolation (PCHIP), spline/generalized additive modeling (GAM), and monotone GAM. Using all selected mode-specific anchors, PCHIP produced 5,595 primary feature pairs across four matrix/mode data sets, with a mean median RT error of 0.0232 min. GAM-based residual screening reduced the number of selected training anchors from 1,518 to 1,185, primary-pair recovery from 5,595 to 4,867, and verification support from 161 to 151, while decreasing the mean median RT error to 0.0212 min. We therefore recommend PCHIP RT transfer using all selected mode-specific anchors without residual-based hard deletion, followed by abundance-concordance ranking. RTBridge establishes feature correspondence rather than metabolite identification. Potential extensions include platform or column transfer, batch harmonization, cross-matrix correspondence, and interlaboratory integration but will require direct experimental validation.