Carmna: classification and regression models for nitrogenase activity based on a pretrained large protein language...

Area of Science:

• Biochemistry
• Microbiology
• Bioinformatics

Background:

• Nitrogen-fixing microorganisms are crucial for the global nitrogen cycle.
• Nitrogenase (EC 1.18.6.1) converts atmospheric nitrogen to ammonia.
• Understanding nitrogenase activity regulation is key for agricultural applications.

Purpose of the Study:

• To develop machine learning models for classifying and predicting nitrogenase activity (Carmna).
• To identify key features influencing nitrogenase activity.
• To contribute to the development of efficient bio-fertilizers.

Main Methods:

• Utilized six machine learning algorithms for classification and regression tasks.
• Employed ProtT5 for feature extraction from nitrogenase sequences.
• Incorporated gene expression and codon preference data for model training.

Main Results:

• XGBoost model achieved an AUC of 0.9365 for classification.
• A stacking model with support vector regression yielded an R2 of 0.5572 for regression.
• Identified amino acid proportions, codon preferences, gene expression, and gene distance as factors influencing activity.

Conclusions:

• Machine learning effectively models nitrogenase activity.
• Multiple genomic and expression features are associated with nitrogenase function.
• Findings support the development of enhanced nitrogen-fixing bio-fertilizers.