Loading…
Contrastive encoder aligning NMR metabolomics to the plasma proteome, adding proteome-level disease risk signal to cohorts with no proteomics.
No providers recorded yet. Browse all providers
Deep plasma proteomics is expensive enough that profiling roughly 54,000 UK Biobank participants required a precompetitive consortium of 13 pharmaceutical companies. NMR metabolomics is cheap enough that it was run on nearly half a million people in the same cohort. The two assays are biologically coupled — apolipoproteins and enzymes such as LPL, CETP and LCAT govern the lipoprotein particles the NMR panel quantifies — so the intuitive fix is to predict the missing proteins from the metabolites and feed that to a risk model.
AugMent takes the opposite position: reconstruction optimizes the wrong quantity. Its encoder is trained with a CLIP-style contrastive objective whose only requirement is that each participant's metabolomic embedding sit closer to their own proteomic profile than to anybody else's in the batch. Nothing in the loss asks for a named protein's concentration, and the trained encoder recovers only 135 of 2,923 proteins at Pearson r > 0.3, against 1,923 for a reconstruction transformer on the same pairs. It nonetheless predicts disease better than any of them, and a temperature ablation shows the trade is causal: weakening the contrastive term raises median per-protein recovery from 0.07 to 0.25 while lowering prediction.
The model comes from Simon Rasmussen's group at the University of Copenhagen, with collaborators at ETH Zurich and ZS Associates. It sits beside MetFoundation, which pretrains on the same Nightingale panel but stays inside one modality, and shares its frozen-encoder-plus-shallow-probe recipe with COMICAL, which aligns genotypes to brain-MRI phenotypes in the same biobank.
The encoder is a four-layer multilayer perceptron with ReLU activations, mapping 170 Nightingale NMR metabolite concentrations through hidden layers of 256, 1,024 and 2,048 units into a 2,923-dimensional space matched to the Olink Explore 3072 plasma proteome — about 8.4 million parameters, plus a learnable inverse-temperature scalar clamped at 100. Metabolomic embeddings and raw proteomic profiles are L2-normalized, and a symmetric cross-entropy loss over the pairwise similarity matrix drives matched participants onto the diagonal. Missing values are filled by a denoising autoencoder and both layers residualized against age, sex and BMI beforehand. Training used 34,821 paired UK Biobank participants at a learning rate of 1 × 10⁻⁴ for 90 epochs.
Applied frozen to 442,559 UK Biobank participants with metabolomics but no proteomics, the augmented representations improved C-index over raw metabolomics in 68 of the 88 diseases where measured proteomics had held an advantage — 14 significant after FDR correction, none significantly worse — led by chronic nephritic syndrome (C-index 0.89, Δ 0.067), Crohn's disease (0.68, Δ 0.052) and hypertensive renal disease (0.91, Δ 0.044). Across the remaining 509 endpoints 295 improved, 20 significantly, at a median Δ of 0.010. The counts are conservative by construction: with 1,000 bootstrap resamples the smallest attainable two-sided p-value is 0.002, so significance is collective rather than individual at this resolution.
The practical target is population-scale epidemiology, where a deep proteomic assay covers a nested subset of a biobank and NMR metabolomics covers all of it. Investigators encode their metabolomics-only cohort through the frozen model and fit their usual survival or classification models on the result, gaining discrimination in renal, inflammatory and respiratory conditions the NMR panel does not measure directly, at no extra assay cost. The recipe generalizes wherever one molecular layer is deeply characterized in a minority of participants and a cheaper layer covers everyone.
AugMent addresses a formulation multi-omics integration has largely left alone: joint models require both layers at inference, while cross-modal imputation targets the missing layer without regard to downstream utility. The finding that participant-level discrimination beats feature-level fidelity is the more portable result, arguing against per-feature reconstruction as the default objective for cross-omics work. The limits are stated plainly by the authors: median gains near 0.01 C-index suit population stratification rather than individual clinical decisions, phenotypes defined by a directly measured metabolite such as hypercholesterolemia are diluted by projection into proteomic space, and both cohorts are predominantly of European ancestry. The MIT-licensed code ships a synthetic-data generator, but neither cohort is redistributable and no trained checkpoint is released.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.