Imputing metabolomics with graph denoising autoencoders

Sarker, Kowshika

Permalink

https://hdl.handle.net/2142/127281 Copy

Description

Title

Imputing metabolomics with graph denoising autoencoders

Author(s)

Sarker, Kowshika

Issue Date

2024-12-09

Director of Research (if dissertation) or Advisor (if thesis)

Zhai, ChengXiang

Department of Study

Siebel School Comp & Data Sci

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• metabolomics
• imputation
• graph denoising autoencoder
• graph neural network

Abstract

Metabolomics is an efficacious modality to extract impactful insights in numerous biomedical applications. Metabolomic datasets often exhibit significant sparsity containing many missing entries. Over the years, numerous metabolomic imputation approaches have been studied. Denoising autoencoders have proven powerful for analyzing noisy data in various domains. On the other hand, graph representations are very popular in biochemical research. In this work, we study the efficacy of graph denoising autoencoders (GDAEs) - the mechanism of which is an integration of denoising autoencoders with graph representations - for the imputation of metabolomic data, as this potential avenue is yet unexplored. We propose a GDAE-based metabolomics imputation approach and benchmark it on three metabolomic datasets comparing with the imputation quality of eight existing methods. We also benchmark the imputation methods based on their effect on downstream classification and clustering tasks. We simulate different patterns and proportions of missing entries in metabolomes and compare the level of difficulty of imputation for different missingness. We also inspect the effect of imputation quality on downstream performance. Based on the empirical evidence, we conclude that GDAE-based imputation is an impactful data preprocessing paradigm for metabolomic analyses.

Graduation Semester

2024-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/127281

Copyright and License Information

Copyright 2024 Kowshika Sarker

Owning Collections