PTSP-BERT: Predict the thermal stability of proteins using sequence-based bidirectional representations from...

Area of Science:

• Biotechnology
• Bioinformatics
• Protein Engineering

Background:

• Proteins from thermophilic, mesophilic, and psychrophilic organisms have diverse industrial applications.
• Current methods for determining optimal protein temperatures are laborious, and existing machine learning models are limited to binary classifications.

Purpose of the Study:

• To develop a deep learning model for direct three-class identification of thermophilic, mesophilic, and psychrophilic proteins based on amino acid sequences.
• To improve classification accuracy through dataset balancing and feature weighting techniques.

Main Methods:

• Developed PSTP-BERT, a deep learning model utilizing BERT-bfd features for protein classification.
• Employed SMOTE (synthetic minority oversampling technique) for dataset balancing.
• Utilized a light gradient-boosting machine for feature weighting and selection.

Main Results:

• PSTP-BERT achieved a five-fold cross-validation accuracy of 89.59% and an independent test accuracy of 85.42% for three-class identification.
• The model demonstrated superior performance in binary classification tasks, achieving 93.33% for thermophilic and 88.33% for psychrophilic proteins.
• Independent test accuracy ranged from 89.8% to 92.9%, with new data prediction exceeding 97%.

Conclusions:

• PSTP-BERT offers a significant advancement over existing models for classifying proteins based on their optimal growth temperatures.
• The model provides a convenient and accurate tool for researchers, with source code available on GitHub.