phyLoSTM: a novel deep learning model on disease prediction from longitudinal microbiome data.

Area of Science:

• Microbiology
• Computational Biology
• Genomics

Background:

• The human microbiome is dynamic and influenced by diet and medical interventions.
• Longitudinal microbiome sequencing data presents challenges due to variable timepoints and imbalanced datasets.

Purpose of the Study:

• To introduce phyLoSTM, a novel deep learning framework for analyzing longitudinal microbiome data.
• To predict disease outcomes by integrating microbiome dynamics with host environmental factors.

Main Methods:

• phyLoSTM combines Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs).
• The framework handles variable timepoints and employs weight balancing for imbalanced data.
• Model performance was tested on simulated and two real-world longitudinal human microbiome studies.

Main Results:

• phyLoSTM achieved high AUCs: 0.897 on simulated data, 0.762 on the DIABIMMUNE study (food allergy), and 0.713 on the DiGiulio study (preterm delivery).
• Performance improvements over Random Forest were 5%, 19%, and 8% respectively.
• The method effectively evaluates microbiome composition changes contributing to outcome prediction.

Conclusions:

• phyLoSTM demonstrates superior predictive accuracy for longitudinal human microbiome studies.
• The framework offers a novel approach for disease prediction by analyzing temporal microbiome data and host factors.