IDMIL: an alignment-free Interpretable Deep Multiple Instance Learning (MIL) for predicting disease from...

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• The human microbiome plays a crucial role in health and disease.
• Metagenomics enables the study of microbial communities but current predictive models are computationally intensive and limited by reference databases.

Purpose of the Study:

• To develop a novel, accurate, and efficient method for predicting human diseases from whole-metagenomic data.
• To overcome the limitations of existing alignment-based and reference-dependent approaches.

Main Methods:

• Formulation of disease prediction as a Multiple Instance Learning (MIL) problem.
• Application of deep convolutional neural networks (CNNs) with a neural attention mechanism for feature extraction and interpretability.
• An alignment-free approach that does not require assembly or reference sequence databases.

Main Results:

• The proposed MIL and deep-CNN framework significantly outperforms existing methods in disease prediction accuracy.
• The attention mechanism provides interpretability by identifying disease-correlated microbial sequence groups.
• The alignment-free method is fast and scalable for large-scale metagenomic datasets.

Conclusions:

• The developed approach offers a powerful and efficient tool for leveraging metagenomic data in precision medicine.
• This method enhances our understanding of the microbiome's role in disease and facilitates the discovery of novel microbial biomarkers.