Cross-platform metabolomics imputation using importance-weighted autoencoders

TL;DR

This paper introduces a new method using autoencoders to accurately impute metabolomics data across different platforms, enabling better data synthesis and replication.

Contribution

The study introduces a novel ensemble of importance-weighted autoencoders for cross-platform metabolomics imputation.

Findings

The imputed metabolites showed high correlation with real values (µρ = 0.61).

43 metabolites unique to Metabolon were well-imputed with ≥ 55% variance captured.

Imputed metabolites showed strong concordance with clinical outcomes like BMI and CRP.

Abstract

Metabolomics data are often generated through different platforms and quantification methods which makes their synthesis and large-scale replication challenging. This study developed an ensemble of importance-weighted autoencoders to perform cross-platform metabolomics imputation between two metabolomics platforms, Metabolon and National Phenome Centre (NPC) at Imperial College, using 979 samples from the Airwave Health Monitoring Study. The generated samples were highly correlated with real values across all metabolites (µρ = 0.61 (0.55–0.67)). The well-imputed subset contained 199 metabolites (22%), capturing ≥ 55% variance (R² ≥ 0.55) with minimal uncertainty (R² variance ≤ 0.025), including 43 metabolites unique to Metabolon. The concordance of associations in 2,971 validation samples between real and imputed metabolites with two clinical outcomes, body mass index (BMI) and C-reactive protein (CRP), were highly correlated (ρBMI = 0.93; ρCRP = 0.89) with minimal mean difference (BMI µΔ = 0.005 (0.04); CRP µΔ = 0.005 (0.04)). Similar concordance occurred with equivalent UK Biobank (BMI µΔ = −0.007 (0.05); CRP µΔ = 0.01 (0.04)) and NPC (BMI µΔ = −0.013 (0.04); CRP µΔ = −0.019 (0.04)) metabolites. This methodological innovation offers a scalable and accurate method for cross-platform imputation, enabling the aggregation of metabolomics data from different epidemiological studies for replication and meta-analyses.

Figures