Decoding longitudinal microbiome trajectories: an interpretable machine learning approach for biomarker discovery and...

Area of Science:

• Microbiome research
• Computational biology
• Machine learning applications in health

Background:

• Longitudinal microbiome data offers insights into disease development and progression.
• Predictive modeling of dynamic microbial data presents analytical challenges.
• Identifying time-varying microbial biomarkers is crucial for early diagnosis and intervention.

Purpose of the Study:

• To introduce LP-Micro, a robust and interpretable machine learning framework for longitudinal microbiome analysis.
• To enhance understanding of pathogenetic mechanisms through time-varying microbial effects.
• To improve prediction accuracy for disease outcomes using microbiome data.

Main Methods:

• Longitudinal microbial feature screening using polynomial group lasso.
• Disease outcome prediction with machine learning models (XGBoost, deep neural networks).
• Interpretable association testing via permutation feature importance.

Main Results:

• LP-Micro accurately identifies disease-related microbiome taxa in simulations.
• The framework demonstrates improved prediction accuracy over existing methods.
• Applications in childhood dental disease and post-bariatric surgery weight loss show high prediction accuracy.

Conclusions:

• LP-Micro effectively analyzes longitudinal microbiome data for predictive modeling.
• The framework highlights critical time points and microbial changes associated with disease outcomes.
• Findings align with clinical expectations and advance microbiome research applications.