Multi-scale variational autoencoder for imputation of missing values in untargeted metabolomics using whole-genome...

Area of Science:

• Computational Biology
• Genomics
• Metabolomics

Background:

• Missing data is a significant challenge in mass spectrometry-based metabolomics, potentially causing biased and incomplete analyses.
• Integrating whole-genome sequencing (WGS) with metabolomics data offers a promising strategy to improve data imputation accuracy.

Purpose of the Study:

• To develop a novel method for imputing unknown metabolites in metabolomics data by leveraging whole-genome sequencing (WGS) information.
• To enhance the accuracy and completeness of metabolomics datasets through integration with genomic data.

Main Methods:

• A multi-scale variational autoencoder was employed to jointly model burden scores, polygenic risk scores (PGS), and linkage disequilibrium (LD) pruned single nucleotide polymorphisms (SNPs).
• This approach facilitates feature extraction and imputation of missing metabolomics data by learning latent representations from both omics datasets.
• The method imputes missing metabolomics values by utilizing genomic information.

Main Results:

• The proposed method demonstrated superior performance over conventional imputation techniques on empirical metabolomics datasets with missing values.
• Using 35 template metabolites, burden scores, PGS, and LD-pruned SNPs, the method achieved R²-scores > 0.01 for 71.55% of metabolites.

Conclusions:

• Integrating WGS data into metabolomics imputation enhances data completeness and downstream analyses, leading to more accurate investigations of metabolic pathways and disease associations.
• The findings highlight the benefits of using WGS data for metabolomics imputation and the importance of multi-modal data integration in precision medicine.