Pretrained variational autoencoder for untargeted blood metabolomics, learning metabolite representations that transfer to clinical prediction tasks.
No providers recorded yet. Browse all providers
Single-cell genomics could build foundation models because the community converged on standard preprocessing, which made thousands of independent studies poolable. Untargeted metabolomics never had that. Every lab tunes its own peak-picking parameters, retention times drift across a run, intensities drift across a batch, and the resulting feature matrices from two studies of the same biofluid do not line up. Without a consistent matrix, there is nothing to pretrain on.
mzEmbed is ReviveMed's answer to what becomes possible once that obstacle is removed. It rests on mzLearn, a companion LC/MS signal-detection and alignment algorithm that runs directly from mzML files with no user-set parameters, learning each dataset's characteristics iteratively and correcting retention-time and intensity drift without requiring experimental QC samples. Applied across 22 public serum studies, mzLearn produces a single harmonized feature matrix, and mzEmbed trains an unsupervised variational autoencoder on it — the pretrained generative model that gives metabolomics its transferable representation.
The distinction matters in practice: mzLearn is the feature detector, mzEmbed the framework that pretrains and fine-tunes the VAE on the matrices it emits. The work, by Leila Pirhaji and colleagues at ReviveMed with a collaborator at The Delphina Group, appeared as a January 2025 bioRxiv preprint and was published in Communications Chemistry in December 2025.
Across 15 benchmark datasets, mzLearn detected an average of 11,442 signals against 7,100 for XCMS and 4,655 for ASARI, with a higher true-positive rate (89.0% versus 77.4% and 49.6%) and lower false-positive rate (12.5% versus 17.3% and 18.8%). Its synthetic-QC normalization reached 17% relative standard deviation on withheld QC samples, close to the 16% achieved using real QC samples and well below 23% for total-ion-current normalization.
Pretraining used 22 public serum studies run under HILIC positive-ion chromatography, covering 20,548 blood samples, from which a robustness score isolated 2,736 peak groups measured consistently across studies. Those samples were split 17,465 / 2,055 / 1,028 for training, validation, and test. Hidden-layer widths are set geometrically between input and latent dimensions. Fine-tuning targeted baseline serum metabolomics from 741 clear-cell renal cell carcinoma patients in the CheckMate025 phase III trial (392 on immune checkpoint inhibitor, 349 on mTOR inhibitor), split 443 / 149 / 149. On the held-out test set, transfer learning beat both random initialization and classical baselines: AUC 93.48 versus 82.86 and 91.78 for binary IMDC risk classification, F1 59.55 versus 54.78 and 58.29 for three-class risk, and C-index 67.39 versus 64.45 and 64.61 for overall survival. The adversarially trained predictive model stratified checkpoint-inhibitor-treated patients by survival (log-rank p = 2.1e-3) while showing no stratification in the mTOR arm (p = 0.61), the behavior a treatment-specific signal should show.
The pairing suits translational groups sitting on untargeted LC/MS cohorts too small to train a model from scratch: mzLearn converts raw mzML files into a comparable feature matrix, and a fine-tuned mzEmbed VAE turns that matrix into risk groups, survival estimates, or treatment-response calls. In oncology specifically, the CheckMate025 analysis points at baseline metabolic profiling as a way to refine prognosis beyond established clinical criteria and to flag likely immunotherapy responders before treatment.
This is among the first demonstrations that untargeted metabolomics can support the pretrain-then-transfer pattern that reshaped genomics, and it locates the bottleneck precisely: the barrier was never model architecture but the absence of a consistent feature matrix at scale. Several limits are explicit. mzLearn operates at the MS1 ion level and performs neither de-isotoping nor MS/MS identification, so most top-ranked features remain unannotated; the pretrained model covers HILIC positive mode only, and extending across chromatography types would require multi-modal training; processing is slow, with 2,075 files taking up to 2.5 days on a single machine. Transfer was demonstrated on one external cohort. mzLearn is distributed as a hosted Docker pipeline for non-profit academic use rather than as source, and the mzEmbed repository carries pretraining and fine-tuning scripts under an academic and non-profit use license, without downloadable pretrained weights.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.