A bacterial sensor taxonomy across earth ecosystems for machine learning applications.

Area of Science:

• Microbiology
• Bioinformatics
• Machine Learning

Background:

• Microbial communities inhabit diverse ecosystems, adapting through signal transduction proteins like histidine kinases.
• These sensors detect environmental cues to modulate internal processes, crucial for microbial survival and adaptation.

Purpose of the Study:

• To test the hypothesis that microbial sensor protein profiles (
• fingerprints
• ) can identify ecosystem types and conditions.
• To leverage machine learning to analyze sensor domains for ecological and medical insights.

Main Methods:

• Collected 20,712 metagenomes from host-associated, environmental, and engineered ecosystems.
• Extracted and clustered ~18 million unique sensory domains using MMseqs2.
• Developed gradient-boosted decision tree models to classify ecosystems and predict physical parameters based on sensor abundance.

Main Results:

• Machine learning models accurately classified ecosystem types (87% accuracy) and predicted physical parameters (83% R2 score).
• Identified key sensor domains predictive of specific ecosystems and patient disease states (e.g., oxygen sensing in gut health).
• 98.7% of identified sensor domains were uncharacterized, highlighting potential for new discoveries.

Conclusions:

• Microbial sensor domain profiles serve as reliable indicators of ecosystem type and conditions.
• This approach enables prioritization of uncharacterized sensors for functional discovery and sensor engineering.
• Applications include biotechnology, ecosystem monitoring, and the development of novel medical diagnostics.