Abstract: TH-OR20
Network Analysis of Paired Plasma-Urine Metabolomes Reveals Genetic Determinants of Metabolite Clusters
Session Information
- Augmented Intelligence and Digital Health Advances
October 24, 2024 | Location: Room 2, Convention Center
Abstract Time: 04:40 PM - 04:50 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Schlosser, Pascal, Universitatsklinikum Freiburg, Freiburg, Baden-Württemberg, Germany
- Hackenberg, Maren, Universitatsklinikum Freiburg, Freiburg, Baden-Württemberg, Germany
- Monteiro-Martins, Sara, Universitatsklinikum Freiburg, Freiburg, Baden-Württemberg, Germany
- Haug, Stefan, Universitatsklinikum Freiburg, Freiburg, Baden-Württemberg, Germany
- Kottgen, Anna, Universitatsklinikum Freiburg, Freiburg, Baden-Württemberg, Germany
Background
Metabolites are often part of biochemical pathways, regularly in complex relations that may be influenced by genetic variants in the involved enzymes. Many of the metabolic reactions driving the biological network are non-linear (e.g. cyclic). Motivated by the success of deep learning-based approaches to detect complex patterns, we investigated whether such models can provide representations of metabolomics that capture their complex, non-linear relations and provide a link to the underlying genetics.
Methods
We present an integrated analysis of data from both plasma (1,096 metabolites) and urine (1,139 metabolites) via an unsupervised method for clustering and local dimension reduction in 4,850 individuals from the GCKD study, a prospective cohort of adults diagnosed with CKD. First, we group the metabolites into modules using a sparse hierarchical clustering approach. On each module, we fit a principal component (PC) analysis and train a variational autoencoder (VAE), a type of neural network, to infer low-dimensional representations of the central factors of variation underlying the data. We then search for genetic associations of these module representations by performing genome-wide association study (GWAS).
Results
We identified 165 modules of correlated metabolites. GWAS identified significant associations (p-values<5e-8/165) with81 VAE- and 150 PC- informed representations of these clusters. An association at NAT8 showed 575 magnitudes lower p-values when investigating metabolite modules (p-value=2.1e-2217) compared to analyses of its individual metabolites. The strongest association was observed for PYROXD2 (module p-value=1.4e-3275). This module integrated the plasma and urine levels of six methyllysine-related metabolites, and displayed independent associations with genetic variants in PYROXD2, NAT8, their interaction, eGFR and UACR in multivariate models. This suggests interactions between these two enzymes and a core role of plasma N6-methyllysine in these associations.
Conclusion
Our results show that VAE- and PC-based representations of metabolic pathways can facilitate the detection of genetic associations beyond those identified by single-metabolite GWAS, indicating existing linear and non-linear genetic regulation on a pathway level.
Funding
- Commercial Support – Bayer Pharma AG