EXPERT: Transfer Learning-enabled context-aware microbial source tracking

Abstract

Habitat specific patterns reflected by microbial communities, as well as complex interactions between the community and their environments or hosts’ characteristics, have created obstacles for microbial source tracking: diverse and context-dependent applications are asking for quantification of the contributions of different niches (biomes), which have already overwhelmed existing methods. Moreover, existing source tracking methods could not extend well for source tracking samples from understudied biomes, as well as samples from longitudinal studies.Here, we introduce EXPERT (https://github.com/HUST-NingKang-Lab/EXPERT), an exact and pervasive expert model for source tracking microbial communities based on transfer learning. Built upon the biome ontology information and transfer learning techniques, EXPERT has acquired the context-aware flexibility and could easily expand the supervised model’s search scope to include the context-dependent community samples and understudied biomes. While at the same time, it is superior to current approaches in source tracking accuracy and speed. EXPERT’s superiority has been demonstrated on multiple source tracking tasks, including source tracking samples collected at different disease stages and longitudinal samples. For example, when dealing with 635 samples from a recent study of colorectal cancer, EXPERT could achieve an AUROC of 0.977 when predicting the host’s phenotypical status. In summary, EXPERT has unleashed the potential of model-based source tracking approaches, enabling source tracking in versatile context-dependent settings, accomplishing pervasive and in-depth knowledge discovery from microbiome.

Microbial Dark Matter: From Discovery to Applications

“…However, when the number of samples and biomes increases, running time increases rapidly, preventing large-scale source tracking. This problem could be solved by deep learning solutions by utilizing model-based methods such as neural networks that would enable improvements in both speed and accuracy during source tracking [61] , [62] .…”

“…The first model-based method for source tracking, ONN4MST, already outperforms existing methods [61] for source tracking of known biomes. Further, the EXPERT method employs ONN4MST models to source track in different contexts [62] and has exhibited a high potential to facilitate mining of the microbial dark matter data. The EXPERT models are based on fundamental neural network models and transfer learning approaches, and exhibit high speed and accuracy, even when analyzing very few (a few hundred) samples from understudied biomes.…”

“…However, ONN4MST is limited to searching biomes contained in the pre-built biome ontology, but cannot search understudied biomes. EXPERT [62] is also a deep learning method for microbial source tracking, which employs neural network models and acquired flexibility by applying a transfer learning approach, enabling adaptation to newly introduced biomes. It pre-built three neural network models for source tracking among 1) all possible sources from diverse environments, 2) human-associated sources, and 3) human gut-associated sources.…”

“…Supervised model-based methods [56] accurately quantify the contributions of source biomes for a specific sample but can also adapt to the analysis of samples from less studied biomes [61] . For example, the EXPERT method is able to accurately differentiate CRC stages in patients using a model built from more than 10,000 human microbiome samples from normal individuals [62] .…”

“…Indeed, supervised learning methods are suitable for such analyses. For example, EXPERT has been used to accurately differentiate samples from patients with nearly twenty diseases, in addition to monitoring CRC stages in patients using models constructed from over 10,000 human microbiome samples in normal individuals [62] .…”