A temporal-aware machine learning framework enables microbial community dynamics prediction with personalized...

Area of Science:

• Microbial Ecology
• Computational Biology
• Precision Medicine

Background:

• Forecasting microbial community dynamics from sparse longitudinal data is challenging for precision medicine and ecological monitoring.
• Existing models often rely on data imputation and assume population-level dynamics, limiting personalized predictions.

Purpose of the Study:

• To develop a personalized, temporal-aware framework for accurate microbial abundance forecasting from incomplete longitudinal data.
• To enable accurate predictions without the need for data imputation.

Main Methods:

• Proposed MicroProphet, a framework utilizing a time-aware Transformer architecture.
• Reconstructed subject-specific microbial trajectories using only the initial 30% of observed time points.
• Employed an attention mechanism to capture critical transitional states.

Main Results:

• Demonstrated robust cross-ecosystem generalizability across synthetic communities, human gut microbiomes, infant gut development, and corpse decomposition.
• Achieved high predictive accuracy and biological interpretability.
• Enabled early detection of disease-associated microbial shifts and optimized timing for microbiome interventions.
• Accurately inferred decomposition timelines in forensic settings.

Conclusions:

• MicroProphet transforms incomplete microbiome data into actionable, individualized forecasts.
• Lays the foundation for temporal-aware systems in microbial ecology and precision health.