Predicting microbiomes through a deep latent space.

Area of Science:

• Microbiome research
• Computational biology
• Machine learning

Background:

• Microbial communities significantly impact their environment by altering nutrient and chemical availability.
• Understanding microbial composition is crucial for enhancing productivity and health.
• Sequencing facilities can be inaccessible or costly, necessitating alternative prediction methods.

Purpose of the Study:

• To computationally predict microbial composition using accessible environmental features.
• To develop a deep learning model for microbiome prediction.
• To enable microbiome engineering strategies with limited resources.

Main Methods:

• Utilized heterogeneous autoencoders to create a deep latent space representation of microbial abundance.
• Designed a predictive model using environmental features (e.g., plant age, temperature, precipitation) as input.
• Applied transfer learning for prediction in distinct scenarios with limited data.

Main Results:

• Reconstructed Maize rhizosphere microbial composition (717 taxa) from a 10-value latent space with high fidelity (>0.9 Pearson correlation).
• Successfully predicted microbial composition from environmental variables with 0.73 Pearson correlation and 0.42 Bray-Curtis dissimilarity.
• Demonstrated the ability to predict microbiome composition under hypothetical future climate change conditions.

Conclusions:

• The proposed deep latent space model accurately predicts microbial composition from environmental data.
• This approach offers a cost-effective and accessible alternative to sequencing for microbiome analysis.
• The model supports microbiome engineering by predicting current and future microbial community structures.