A stacking-based algorithm for antifreeze protein identification using combined physicochemical, pseudo amino acid...

Area of Science:

• Biochemistry and Bioinformatics
• Computational Biology
• Protein Science

Background:

• Antifreeze proteins (AFPs) are crucial in medical and food industries due to their ice-structuring capabilities.
• Accurate identification of AFPs is essential for their effective application and further research.

Purpose of the Study:

• To develop and validate a novel stacking-based classifier for identifying antifreeze proteins.
• To analyze the key features contributing to the recognition of antifreeze proteins.

Main Methods:

• Feature extraction using reduction properties, scalable pseudo amino acid composition, and physicochemical properties.
• Training a stacking model with LightGBM, XGBoost, and RandomForest algorithms, with Logistic regression as the final matching layer.
• Validation using test sets and an independent validation set.

Main Results:

• The proposed stacking classifier achieved a recognition accuracy of 98.3% on the test set and 98.5% on the validation set.
• Analysis identified H^r*H^r, H^rH^r, and Sc-PseAAC_1 as crucial numerical features.
• Hydrophobicity, secondary structure, charge, van der Waals forces, and solvent accessibility were confirmed as important factors for antifreeze capability.

Conclusions:

• The developed stacking classifier is highly effective for identifying antifreeze proteins.
• The study highlights the importance of specific protein features and structural properties in antifreeze activity.
• The findings provide valuable insights for the design and application of antifreeze proteins.