PeNGaRoo, a combined gradient boosting and ensemble learning framework for predicting non-classical secreted proteins

Area of Science:

• Microbiology
• Computational Biology
• Bioinformatics

Background:

• Gram-positive bacteria utilize protein secretion systems for host infection and biotechnological applications.
• Identifying 'non-classical' secreted proteins is challenging due to the absence of signal peptides and diverse secretion pathways.
• Existing computational methods for predicting non-classical secreted proteins are limited by dataset size and feature complexity.

Purpose of the Study:

• To develop an improved computational method for predicting non-classical secreted proteins from sequence data.
• To leverage experimentally validated datasets and advanced machine learning techniques for enhanced prediction accuracy.
• To create a user-friendly online tool for the discovery of these important bacterial proteins.

Main Methods:

• Construction of a high-quality dataset of experimentally verified non-classical secreted proteins.
• Comprehensive analysis and individual performance assessment of various sequence-based features.
• Development of a two-layer Light Gradient Boosting Machine (LightGBM) ensemble model with particle swarm optimization for parameter tuning.

Main Results:

• The developed ensemble model achieved high performance metrics: 0.900 accuracy, 0.903 F-value, 0.803 Matthew's correlation coefficient, and 0.963 AUC.
• The model outperformed existing state-of-the-art predictors on an independent test dataset.
• An accessible online predictor, PeNGaRoo, was developed based on the optimal ensemble model.

Conclusions:

• The proposed LightGBM ensemble model significantly enhances the prediction of non-classical secreted proteins.
• The PeNGaRoo web server provides a valuable resource for researchers studying Gram-positive bacteria.
• This work facilitates the discovery of effector proteins and inspires future advancements in prediction methodologies.