GEMimp: An Accurate and Robust Imputation Method for Microbiome Data Using Graph Embedding Neural Network.

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Microbiome research highlights the link between microbial composition and human health/disease.
• Microbiome data analysis is challenged by sparsity (prevalence of zero counts).
• Data sparsity can skew abundance distributions and reduce the reliability of findings.

Purpose of the Study:

• To introduce GEMimp, an imputation method to enhance microbiome data robustness.
• To evaluate GEMimp's performance against existing state-of-the-art imputation techniques.

Main Methods:

• GEMimp utilizes the node2vec algorithm with Breadth-First Search (BFS) and Depth-First Search (DFS) for random walks.
• This method learns low-dimensional representations of taxonomic units to reconstruct similarity networks.
• Comparative analysis included SAVER, MAGIC, and mbImpute on simulated and real-world datasets.

Main Results:

• GEMimp achieved the highest Pearson correlation coefficient compared to raw data, outperforming other methods.
• GEMimp demonstrated proficiency in identifying significant and disease-related taxa.
• The method effectively mitigated sparsity issues in both simulated and real-world microbiome datasets (e.g., T2D, CRC).

Conclusions:

• GEMimp significantly improves the analysis of sparse microbiome data.
• By alleviating sparsity, GEMimp facilitates downstream analyses and advances microbiological research.
• The method enhances the detection of disease-associated microbial taxa.