Hierarchical Microbial Functions Prediction by Graph Aggregated Embedding.

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Predicting microbial metabolic function from 16S rRNA sequencing data is crucial for understanding microbial communities.
• Current methods struggle with non-human samples and uncultured microbes (operational taxonomy units - OTUs).
• Functional profiling is often hindered by imbalanced datasets and long-tailed distributions of microbial functions.

Purpose of the Study:

• To develop a novel computational framework, Hierarchical micrObial functions Prediction by graph aggregated Embedding (HOPE), for accurate microbial function prediction.
• To address the limitations of existing methods in profiling challenging microbial communities.
• To improve the prediction of KEGG Orthology (KO) functions, particularly in the context of imbalanced data.

Main Methods:

• HOPE integrates microbial co-occurrence network topology with k-mer sequence composition.
• It embeds these features into a latent space, preserving OTU relationships.
• A hierarchical multitask learning module is employed to handle the imbalanced distribution of microbial functions.

Main Results:

• HOPE significantly outperforms baseline methods (HOPE-one, HOPE-seq, GraphSAGE) across multiple 16S rRNA sequencing datasets (abalone, human, and shrimp gut).
• The model demonstrates strong generalization capabilities on diverse microbial communities.
• HOPE effectively addresses the challenge of predicting functions for OTUs from non-human samples.

Conclusions:

• HOPE provides a robust and accurate method for predicting microbial metabolic functions.
• The framework's ability to handle data imbalance and generalize makes it valuable for microbiome research.
• The underlying principles of HOPE are adaptable to other biological prediction tasks, such as gene function prediction.