Improving biome labeling for tens of thousands of inaccurately annotated microbial community samples based on neural network and transfer learning

Microbiome samples are accumulating at a fast speed, leading to millions of accessible microbiome samples in the public databases. However, due to the lack of strict meta-data standard for data submission and other reasons, there is currently a non-neglectable proportion of microbiome samples in the public database that have no annotations about where these samples were collected, how they were processed and sequenced, etc., among which the missing information about collection niches (biome) is one of the most prominent. The lack of sample biome information has created a bottleneck for mining of the microbiome data, making it difficult in applications such as sample source tracking and biomarker discovery. Here we have designed Meta-Sorter, a neural network and transfer learning enabled AI method for improving the biome labeling of thousands of microbial community samples without detailed biome information. Results have shown that out of 16,507 samples that have no detailed biome annotations, 96.65% could be correctly classified, largely solving the missing biome labeling problem. Interestingly, we succeeded in classify 250 samples, which were sampled from benthic and water column but vaguely labeled as "Marine" in MGnify, in more details and with high fidelity. Whats more, many of successfully predicted sample labels were from studies that involved human-environment interactions, for which we could also clearly differentiated samples from environment or human. Taken together, we have improved the completeness of biome label information for thousands of microbial community samples, facilitating sample classification and knowledge discovery from millions of microbiome samples.