Computational predictions for protein sequences of COVID-19 virus via machine learning algorithms

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in virology

Background:

• The COVID-19 pandemic necessitates understanding viral protein variations across human populations.
• Global spread of SARS-CoV-2 impacts different countries uniquely.

Purpose of the Study:

• To classify human protein sequences of COVID-19 based on country of origin.
• To develop a machine learning model for predicting geographic protein sequence patterns.

Main Methods:

• Utilized conjoint triad (CT) method for data preprocessing, converting amino acids to numerical groups.
• Implemented two data labeling strategies: country code assignment and binary element representation.
• Employed machine learning algorithms, including linear Support Vector Machine (SVM), for sequence classification.

Main Results:

• Achieved 100% accuracy and sensitivity, with 90% specificity using binary labeling and linear SVM.
• Demonstrated higher classification accuracy for countries with more available protein sequence data, such as the USA.
• Identified data imbalance as a significant challenge, with US data comprising 76% of the total sequences.

Conclusions:

• The developed machine learning model effectively classifies COVID-19 protein sequences by country.
• The model shows potential as a predictive tool for understanding geographic variations in viral proteins.
• Addressing data imbalance is crucial for improving the model's generalizability across all countries.