Current Trend and Performance Evaluation of Machine Learning Methods for Predicting Host-Pathogen Protein-Protein Interactions

Abstract

Host-pathogen interactions (HPI) play a vital role in the study of infectious disease mechanisms. Experimental identification of the interaction between a host and pathogen has shown over the years to be expensive based on time, labour, and cost which has led to the discovery of not so many interactions when compared to the potential interactions that could be identified. Hence the reason for the adoption of machine learning-based methods for host-pathogen interaction prediction which has proven to be not just cost and time-effective but also outcome-reliable. Therefore, this work identified and evaluated the top five most frequently used machine learning-based HPI prediction algorithms viz Random Forest, Support Vector Machine, Deep Neural Network, Logistic Regression, and Naïve Bayes. In addition to the top five identified models, Deep Forest, an ensemble deep learning model was evaluated. This is because Deep Forest is one of the recently adopted models used for HPI prediction. The models identified for HPI prediction were benchmarked against two of the most frequently used host-pathogen combinations that were also identified in this work which include human-parasite and human-bacteria. Amongst the models evaluated, Deep Forest, an ensemble of cascade forest models recorded the highest predictive accuracy closely followed by random forest while Naïve Bayes presented the lowest performance. Other metrics measured in this work are sensitivity, specificity, precision, F1 score, Matthews Correlation Coefficient (MCC), and Area Under ROC (AUC).