NFEmbed: modeling nitrogenase activity via classification and regression with pretrained protein embeddings

Area of Science:

• Agricultural Science
• Bioinformatics
• Microbiology

Background:

• Synthetic nitrogen fertilizers cause environmental issues like eutrophication and reduced crop yields.
• Nitrogen-fixing microorganisms offer a sustainable alternative by utilizing the nitrogenase enzyme.
• Expressing functional nitrogenase in cereal crops is a potential strategy to enhance nitrogen fixation.

Purpose of the Study:

• To predict microbial strains with high nitrogenase activity using machine learning.
• To enable screening and ranking of potential nitrogen-fixing strains based on genomic data.
• To develop advanced computational tools for biofertilizer discovery.

Main Methods:

• Exploration of protein language model embeddings for prediction.
• Development of two stacking ensemble models: NFEmbed-C and NFEmbed-R.
• Utilized machine learning algorithms including k-Nearest Neighbors, Random Forest, Decision Tree Regressor, eXtreme Gradient Boosting Regressor, and Support Vector Regressor.

Main Results:

• Both NFEmbed-C and NFEmbed-R models outperformed state-of-the-art methods.
• NFEmbed-C achieved 0.949 sensitivity, 0.892 F1 score, and 0.784 Matthews Correlation Coefficient.
• NFEmbed-R demonstrated strong performance with an R² score of 0.783 and low MSE (0.158) and RMSE (0.398).

Conclusions:

• Machine learning effectively predicts nitrogenase activity in microbial strains.
• The developed models provide a powerful tool for identifying superior biofertilizer candidates.
• This approach supports sustainable agriculture by reducing reliance on synthetic nitrogen fertilizers.