Machine Learning Approaches for Predicting Virus-Human Protein-Protein Interactions: An Evaluation of Retroviral Interaction Networks

Abstract

Virus-human protein-protein interactions (VHPPI) are key to understanding how viruses manipulate host cellular functions. This study constructed a retroviral-human PPI network by integrating multiple public databases, resulting in 1,387 interactions between 29 retroviral and 1,026 human genes. Using minimal sequence similarity, we generated a pseudo-negative dataset for model reliability. Five machine learning models--Logistic Regression (LR), Support Vector Machine (SVM), Naive Bayes (NB), Decision Tree (DT), and Random Forest (RF)--were evaluated using accuracy, sensitivity, specificity, PPV, and NPV. LR and KNN models demonstrated the strongest predictive performance, with sensitivities up to 77% and specificities of 52%. Feature importance analysis identified GC content and semantic similarity as influential predictors. Models trained on selected features showed enhanced accuracy with reduced complexity. Our approach highlights the potential of computational models for VHPPI predictions, offering valuable insights into viral-host interaction networks and guiding therapeutic target identification. SignificanceThis study addresses a crucial gap in antiviral research by focusing on the prediction of virus-host protein-protein interactions (VHPPI) for retroviruses, which are linked to serious diseases, including certain cancers and autoimmune disorders. By leveraging machine learning models, we identified essential host-pathogen interactions that underlie retroviral survival and pathogenesis. These models were optimized to predict interactions accurately, offering valuable insights into the complex mechanisms that retroviruses use to manipulate host cellular processes. Our approach highlights key host and viral proteins, such as ENV_HV1H2 and CD4, that play pivotal roles in retroviral infection and persistence. Targeting these specific interactions can potentially disrupt the viral lifecycle while minimizing toxicity to human cells. This study thus opens avenues for the development of selective therapeutic strategies, contributing to more effective and targeted antiviral interventions with fewer side effects, marking a significant step forward in computational virology and drug discovery.