|
Abstract INTRODUCTION:
Targeted LC–MS/MS metabolomics is increasingly deployed at clinical scale, but reproducibly processing large datasets at scale remains a bottleneck. Single-sample peak-picking workflows with manual curation introduce inter-analyst variability, miss low-signal peaks, and do not scale to panels of hundreds of metabolites across thousands of samples. To enable standardized, automated processing of a large clinical metabolomics resource, we developed MRMpipe specifically for the Biobanque Québécoise de la COVID 19 (BQC19). MRMpipe is a general-purpose clinical LC–MS/MS analysis software that integrates cross-sample consensus peak detection, adaptive retention-time (RT) alignment, and assay-specific machine-learning quality scoring. It accepts vendor raw files as input and produces quantification-ready results, requiring minimal manual data curation. Here, we report a multi-site analytical evaluation conducted using MRMpipe, developed to support standardized processing of targeted metabolomics data from the BQC19 (1,595 samples, 8 hospitals, six clinical cohorts). We performed analytical validation against stable-isotope-labelled internal standards, quality control (QC)-based normalization, and expert curation.
METHODS:
A total of 1,595 plasma samples were collected from eight hospitals across six clinically defined cohorts within the BQC19, including hybrid-immunity case/control arms, hospitalized and out-patient, as well as long-COVID positive and negative phenotypes. The study was approved by the CHUM Research Ethics Board under the BQC19 multicenter ethics framework. Samples were analyzed by targeted LC–MS/MS using a curated clinical metabolomics panel with stable-isotope-labelled internal standards, run across 40 analytical batches. To enable standardized processing of this large, multi-site dataset and prepare it for release to the scientific community, we developed MRMpipe, which processed all samples end-to-end, from vendor files to quantification ready output. The software performs signal conditioning, adaptive retention-time (RT) alignment, and cross-sample consensus peak detection. A two-tier machine-learning quality scorer flags each metabolite-sample pair and assigns one of three outcomes (accept, review, reject). The quality scorer ranks competing peak candidates using peak shape descriptors, signal-to-noise metrics, RT and ion-ratio consistency across the batch. All processing steps and decisions are recorded in a persistent audit trail. Analytical performance was benchmarked against expert manual curation on a representative subset of samples following established targeted LC-MS/MS analysis protocols.
RESULTS:
Across the full 1,595-sample cohort, MRMpipe produced consistent, high-quality quantitative outputs at scale. Adaptive RT alignment reduced cross-sample variability across all eight hospitals. Cross-sample consensus peak detection improved recovery of low-abundance signals that are typically missed by single-sample peak-picking approaches, while the machine-learning quality scorer consistently distinguished confident from ambiguous integrations. Precision after internal-standard-correction met predefined clinical acceptance criteria across all batches, and pooled QC samples exhibited stable quantification across sites, indicating effective absorption of batch and site effects. On the curated subset, MRMpipe outputs closely matched expert manual integrations, while sharply reducing hands-on review time. Together, these results demonstrate that MRMpipe provides a reproducible and scalable analytical workflow for large, multi-site targeted LC-MS/MS datasets.
CONCLUSION:
This work demonstrates that transparent machine learning methods can be integrated into clinical targeted LC–MS/MS workflows without compromising reproducibility or interpretability. By using assay-specific models trained on curated reference data, MRMpipe avoids the “pretrained model does not fit my LC” failure mode that has limited adoption of AI-based tools in the context of clinically sourced data. MRMpipe is a general-purpose clinical LC–MS/MS analysis software, developed specifically to enable standardized processing and validated release of large, multi-site metabolomics datasets such as those generated by the BQC19. The analytical evaluation reported here, supported by orthogonal internal standard validation, demonstrates a practical and deployable path toward standardized targeted LC–MS/MS analysis at scale.
|
= Discovery stage. (57.21%, 2026)
= Translation stage. (23.38%, 2026)
= Clinically available. (19.40%, 2026)
