IWAE Cross-Platform Metabolomics Imputation
Summary¶
Smith, Pinto, Zagkos, et al. (2026) treat cross-platform metabolomics harmonisation as a missing-data problem: metabolites measured on one platform (Metabolon) but not another (National Phenome Centre LC-MS, NPC) are imputed as though "missing," using an ensemble of importance-weighted autoencoders (IWAEs) trained on 979 Airwave Health Monitoring Study samples with both platforms measured. This is presented as the first method to entirely recreate one metabolomics dataset from another, rather than merely aligning overlapping metabolites, and is validated by checking whether imputed-metabolite associations with clinical outcomes (BMI, CRP) match associations from the real, directly-measured metabolites.
Method¶
- Metabolite clustering: UMAP embedding of the combined Metabolon-NPC feature space identified five clusters — Fatty acids, Sphingolipids, Phospholipids, Acyl-glycerides, and an Others catch-all — used to train cluster-specific imputation models rather than a single global model, since a single model trained on all features was found to prioritize lipid-imputing signal at the expense of other metabolite classes given the modest sample size (N=979).
- IWAE ensemble with greedy selection: an IWAE is trained per cluster (test-set MAE 0.125–0.190 across clusters) plus a full-feature model (test-set RMSE 0.251); a greedy selection procedure across the ensemble chooses the best-performing imputation for each target metabolite.
- Validation via downstream association concordance: rather than relying on imputation accuracy alone, the imputed Metabolon dataset (2,971 independent samples) was combined with real Airwave Metabolon, Airwave NPC, and UK Biobank NMR datasets to compare observational associations with BMI and C-reactive protein (CRP) between real and imputed metabolite values.
Performance¶
- Across all imputed Metabolon metabolites, held-out test-set correlation with real values averaged ρ = 0.61 (SD 0.09, IQR 0.55–0.67).
- A "well-imputed" subset of 199 metabolites (22% of the Metabolon panel) achieved R² ≥ 0.55 with low uncertainty (R² variance ≤ 0.025); of these, 43 were metabolites unique to Metabolon (not directly present in NPC), demonstrating genuine cross-platform inference rather than trivial pass-through of overlapping features.
- Imputation performance was highly class-dependent: the Lipid superclass imputed best (mean R² = 0.49, SD 0.25 — also the highest variance), while Xenobiotics, Amino acids, Peptides, Cofactors/Vitamins, Carbohydrates, Energy, and Nucleotide metabolites all imputed substantially worse (mean R² 0.11–0.24) — attributed to the LC-MS input platform predominantly capturing lipid species. Swapping the input platform to NMR data materially improved imputation of non-lipid classes (peptides, carbohydrates) in a secondary analysis (Airwave 2), supporting the interpretation that imputation quality tracks the presence of closely related input features rather than lipid biology specifically.
- Association concordance: standardized beta-coefficients for the 199 well-imputed metabolites correlated strongly between real and imputed values (ρ = 0.93 for BMI, ρ = 0.89 for CRP), with minimal mean difference (BMI Δ = 0.005 ± 0.04; CRP Δ = 0.005 ± 0.04) and only 5% (BMI) / 3% (CRP) of metabolites showing a coefficient difference beyond 1.96 SD. Comparable concordance held against independent real NPC LC-MS (BMI Δ = -0.013 ± 0.04; CRP Δ = -0.019 ± 0.04) and UK Biobank NMR (BMI Δ = -0.007 ± 0.05; CRP Δ = 0.01 ± 0.04) datasets.
Interpretation and Limitations¶
The authors position the method for cross-cohort harmonisation and replication of already-established metabolite-phenotype associations, not as a substitute for direct metabolite quantification. Imputation quality is bottlenecked by how closely related the input platform's features are to the target metabolite class; broadening input platforms (e.g., adding NMR or reverse-phase LC-MS) is proposed to extend accurate imputation beyond lipids. The authors also suggest (but did not test) that SHAP-based feature importance on the trained IWAEs could help annotate unknown LC-MS features by identifying which input features drive a given metabolite's imputation.
Availability¶
No public code repository; code is available from the corresponding author on reasonable request. Implemented in TensorFlow 2.4.1 / Python 3.7.9, with platform feature-matching via the M2S MATLAB package.
See Also¶
- LIPID Imputation — an elastic-net-based approach to the analogous cross-platform harmonisation problem for plasma lipidomics rather than general metabolomics.
- Nightingale Health NMR Platform — an NMR metabolomics platform of the kind this imputation approach could extend to as an input source.