mbSparse: an autoencoder-based imputation method to address sparsity in microbiome data

Area of Science:

• Microbiome research
• Bioinformatics
• Computational biology

Background:

• Gut microbiota plays a vital role in host physiology.
• High sparsity (numerous zeros) in microbiome data poses significant analytical challenges.
• Existing methods struggle with accurate imputation of sparse microbiome data.

Purpose of the Study:

• To develop a novel deep learning-based algorithm, mbSparse, for accurate imputation of sparse microbiome data.
• To evaluate the performance of mbSparse compared to existing methods.
• To assess the utility of mbSparse in a colorectal cancer analysis.

Main Methods:

• Developed mbSparse, an imputation algorithm using a feature autoencoder and a conditional variational autoencoder (CVAE).
• Leveraged deep learning for learning sample representations and data reconstruction.
• Applied mbSparse to simulated and real microbiome datasets, including colorectal cancer data.

Main Results:

• mbSparse achieved superior imputation accuracy, reducing mean squared error by up to 4.1 compared to existing methods.
• In colorectal cancer analysis, mbSparse increased the detection of disease-associated taxa from 7 to 27 and improved predictive accuracy (AUC from 0.85 to 0.93).
• mbSparse effectively restored over 88% of removed counts, preserving taxonomic relationships with a Pearson correlation of 0.9354.

Conclusions:

• mbSparse offers a powerful deep learning solution for accurate microbiome data imputation, overcoming challenges posed by data sparsity.
• The CVAE component is crucial for mbSparse's enhanced accuracy.
• mbSparse improves biological insights and predictive power in microbiome-associated disease studies.