Phylogenetic convolutional neural networks in metagenomics

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel with data having inherent spatial relationships, like images.
• Metagenomics data lacks intrinsic neighborhood concepts, limiting traditional CNN application.
• Phylogenetic distance offers a novel proximity measure for metagenomics data.

Purpose of the Study:

• Introduce Ph-CNN, a novel deep learning architecture for metagenomics data classification.
• Adapt CNNs for non-Euclidean biological data using phylogenetic relationships.
• Improve classification accuracy in metagenomics studies.

Main Methods:

• Ph-CNN utilizes patristic distance from phylogenetic trees as a proximity measure.
• Employs a sparsified MultiDimensional Scaling to embed phylogenetic trees in Euclidean space.
• Implemented as a custom Keras layer passing neighborhood information to convolutional layers.

Main Results:

• Ph-CNN demonstrates promising classification performance on synthetic and real-world gut microbiota datasets.
• Outperforms classical algorithms like Support Vector Machines and Random Forest.
• Achieves competitive results compared to a baseline fully connected neural network (Multi-Layer Perceptron).

Conclusions:

• Ph-CNN offers a novel deep learning strategy for metagenomics data classification.
• Effectively adapts CNNs to biological data by leveraging phylogenetic structure.
• Provides a powerful new tool for analyzing complex microbial communities.