Refining biome labeling for large-scale microbial community samples: Leveraging neural networks and transfer learning

Area of Science:

• Microbiology
• Bioinformatics
• Environmental Science

Background:

• Microbiome research generates vast datasets, necessitating accurate sample annotations for effective data utilization.
• Incomplete metadata, particularly collection location and biome information, impedes environmental microbiome research.
• Publicly available microbiome samples often lack crucial details for comprehensive analysis.

Purpose of the Study:

• To introduce Meta-Sorter, a novel approach for enhancing biome labeling of microbiome samples.
• To improve the accuracy and completeness of metadata for environmental microbiome samples in the MGnify database.
• To facilitate more precise knowledge discovery in microbiome research, particularly in environmental studies.

Main Methods:

• Utilized neural networks and transfer learning for biome classification.
• Applied Meta-Sorter to thousands of MGnify database samples with incomplete biome annotations.
• Developed a method to distinguish between environmental and human-associated microbiome samples.

Main Results:

• Meta-Sorter achieved 96.7% accuracy in classifying 16,507 microbiome samples.
• Precisely classified previously ambiguous "Marine" samples into benthic and water column origins.
• Effectively differentiated samples from human-environment interactions.

Conclusions:

• Meta-Sorter significantly enhances biome label completeness for microbial community samples.
• Improved annotations facilitate more accurate knowledge discovery, especially in environmental research.
• The tool aids in distinguishing human-related studies from purely environmental ones, refining research focus.