DeepMicro

DeepMicro applies deep representation learning with autoencoder architectures to learn low-dimensional representations of high-dimensional microbiome and human microbiota profiles for disease prediction.

Key Features:

• Autoencoder architectures: Utilizes various autoencoder architectures to transform high-dimensional microbiome profiles into low-dimensional representations.
• Deep representation learning: Implements deep representation learning to capture structure in sparse, high-dimensional microbiome data.
• Dimensionality reduction: Reduces input dimensionality from hundreds of thousands of features to robust low-dimensional embeddings.
• Sparse data handling: Addresses sparsity and low sample sizes typical of microbiome datasets.
• Classification integration: Employs learned representations as inputs to machine learning classification models trained to predict disease outcomes.
• Comparative performance: Demonstrates superior disease-prediction performance compared to approaches based on strain-level marker profiles across multiple datasets.
• Speed and optimization: Accelerates model training and hyperparameter optimization by 8X to 30X relative to traditional methods.
• Evaluation scheme: Applies a rigorous evaluation scheme across multiple datasets.

Scientific Applications:

• Disease outcome prediction: Predicts disease outcomes from microbiome and human microbiota profiles using learned low-dimensional representations.
• Method benchmarking: Benchmarks representation-learning approaches against strain-level marker profile methods across multiple datasets.
• Microbiota–health analysis: Enables exploration of predictive signals in human microbiota related to health and disease.

Methodology:

Utilizes various autoencoder architectures for deep representation learning to convert high-dimensional, sparse microbiome profiles into low-dimensional embeddings, uses those embeddings as inputs to machine learning classification models to predict disease outcomes, and applies a rigorous evaluation scheme across multiple datasets while reporting 8X to 30X faster training and hyperparameter optimization compared to traditional strain-level marker profile methods.