Multitask knowledge-primed neural network for predicting missing metadata and host phenotype based on human...

Area of Science:

• Microbiome research
• Computational biology
• Machine learning in health

Background:

• Microbiome signatures are linked to human diseases, prompting machine learning for prediction.
• Challenges exist in accuracy, generalizability, and interpretability of microbiome-based predictions.
• Host factors like age and gender confound microbiome analysis and predictions.

Purpose of the Study:

• To develop a unified model, MicroKPNN-MT, for predicting human phenotype from microbiome data and metadata.
• To enhance prediction accuracy and generalizability by incorporating host metadata.
• To predict missing metadata from microbiome data, improving model robustness.

Main Methods:

• Developed MicroKPNN-MT, an extension of the MicroKPNN framework.
• Integrated host metadata (age, gender) as input features.
• Utilized additional decoders to predict metadata from microbiome data when unavailable.
• Applied the model to the mBodyMap dataset covering healthy individuals and 25 diseases.

Main Results:

• MicroKPNN-MT demonstrated potential as a predictive tool for multiple diseases.
• The model successfully predicted missing metadata from microbiome data.
• Incorporating real or predicted metadata significantly improved disease prediction accuracy.
• Enhanced generalizability of predictive models by leveraging metadata.

Conclusions:

• MicroKPNN-MT offers a unified approach for microbiome-based disease and metadata prediction.
• The integration or prediction of host metadata is crucial for robust microbiome-based health predictions.
• This model advances the application of machine learning in understanding the human microbiome's role in health and disease.