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

Microbiome research has generated an extensive amount of data, resulting in a wealth of publicly accessible samples. Accurate annotation of these samples is crucial for effectively utilizing microbiome data across scientific disciplines. However, a notable challenge arises from the lack of essential annotations, particularly regarding collection location and sample biome information, which significantly hinders environmental microbiome research. In this study, we introduce Meta-Sorter, a novel approach utilizing neural networks and transfer learning, to enhance biome labeling for thousands of microbiome samples in the MGnify database that have incomplete information. Our findings demonstrate that Meta-Sorter achieved a remarkable accuracy rate of 96.7% in classifying samples among the 16,507 lacking detailed biome annotations. Notably, Meta-Sorter provides precise classifications for representative environmental samples that were previously ambiguously labeled as "Marine" in MGnify, thereby elucidating their specific origins in benthic and water column environments. Moreover, Meta-Sorter effectively distinguishes samples derived from human-environment interactions, enabling clear differentiation between environmental and human-related studies. By improving the completeness of biome label information for numerous microbial community samples, our research facilitates more accurate knowledge discovery across diverse disciplines, with particular implications for environmental research.