Predicting bacterial phenotypic traits through improved machine learning using high-quality, curated datasets

Area of Science:

• Microbiology and Bioinformatics
• Computational Biology and Genomics

Background:

• Predicting prokaryotic phenotypes (observable traits) is crucial for biotechnology, environmental science, and evolutionary biology.
• Understanding genotype-phenotype relationships can unlock new applications for microbial communities.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting prokaryotic physiological properties from protein family inventories.
• To generate novel phenotypic data for bacterial strains and make it publicly available.

Main Methods:

• Utilized standardized datasets from the BacDive database.
• Modeled eight physiological properties using protein family inventories.
• Evaluated model performance with multiple metrics and examined biological implications.

Main Results:

• Achieved high confidence values in predictions, highlighting the importance of data quality and quantity.
• Generated 50,396 new data points for 15,938 strains, now available in BacDive.
• Developed open-source software applicable to other datasets, including metagenomic data.

Conclusions:

• Machine learning provides a reliable method for inferring bacterial phenotypes from genotypes.
• The generated data and open-source tools will significantly advance microbial research and applications.
• This approach can aid in assessing microbial potential for environmental applications like bioremediation.