DeepToA

DeepToA predicts the theater of activity (ToA) of microbiomes from metagenomic sequencing data by integrating taxonomic and functional profiles to link microbial composition to environmental context.

Key Features:

• Deep-Learning Framework: Employs an ensemble deep-learning approach trained on taxonomic and functional metagenomic profiles.
• Taxonomic Profiles (node2vec): Transforms hierarchical taxonomic profiles into numerical vectors using node2vec to capture relationships within microbial communities.
• Functional Profiles with Textual Enrichment: Integrates functional data with textual descriptions of protein families or domains to enrich input features.
• Dimension Reduction and Clustering: Applies clustering-based dimension reduction to mitigate sparsity in taxonomic data prior to model training.
• Ensemble Learning Approach: Combines multiple models within an ensemble framework to enhance prediction robustness and reduce overfitting.
• Feature Importance Analysis (SHAP): Uses SHAP (SHapley Additive exPlanations) values to identify taxonomic and functional features that contribute to ToA predictions.

Scientific Applications:

• Microbial Ecology: Characterizes how microbial community taxonomic and functional composition corresponds to environmental contexts such as soil, gut, or skin.
• Human Health: Relates microbiome composition at specific body sites to theaters of activity relevant to health and disease studies.
• Environmental Science: Assesses microbial contributions to ecosystem functions and biogeochemical cycles by assigning environmental ToAs.

Methodology:

Metagenomic profiles from MGnify are preprocessed with node2vec embeddings for taxonomic data and clustering for dimension reduction, functional data are enriched with textual information for protein families/domains, an ensemble of deep-learning models is trained on the processed data, and the models are evaluated on a dataset classified into 10 theaters of activity achieving 98.30% accuracy.