DeepIMB: Imputation of non-biological zero counts in microbiome data

Area of Science:

• Microbiome research
• Bioinformatics
• Computational biology

Background:

• Non-biological zero counts are prevalent in microbiome data due to low sequencing depth and sampling variation.
• These zeros distort taxon abundance, obscure true biological signals, and complicate downstream analyses.

Purpose of the Study:

• To introduce DeepIMB, a deep learning imputation method for microbiome data.
• To accurately identify and impute non-biological zeros while preserving biological integrity.

Main Methods:

• DeepIMB uses a two-phase approach: identifying non-biological zeros with a gamma-normal mixture model and imputing them using a deep neural network.
• The imputation model integrates taxon abundances, sample covariates, and phylogenetic distances to capture complex data relationships.

Main Results:

• DeepIMB accurately imputes non-biological zeros and preserves genuine biological signals.
• Simulation studies demonstrated DeepIMB's superior performance over existing methods in terms of mean squared error, Pearson correlation, and Wasserstein distance.

Conclusions:

• DeepIMB effectively resolves the issue of non-biological zeros in microbiome datasets.
• This method advances microbiome research by improving data quality and the reliability of subsequent analyses.