Leveraging viral genome sequences and machine learning models for identification of potentially selective antiviral agents

Figures

• 

  Fig. 1. Comprehensive analysis and predictive modeling of virus-selective drugs. A Heatmap visualization of pairwise genomic similarity among common pathogenic viruses. Green indicates sequence dissimilarity, while orange denotes sequence similarity. B Distribution of approved and investigational antiviral drugs. The left panel displays the overall distribution of these drugs, while the right panel shows the number of drugs corresponding to each type of virus. C Flowchart outlining the feature components used for selective antiviral drug modeling. D Performance evaluation of the optimal predictive models and their corresponding parameter combinations. Results are presented as mean ± standard deviation (SD), with error bars representing the SD from 20 independent iterations. AUC-ROC area under the receiver operating characteristic curve, BA balanced accuracy, ECFP4 extended connectivity fingerprint 4, FASTA fast alignment search tool for DNA/RNA sequences; F&T Fisher’s exact test and t-test, FS feature selection, HBV hepatitis B virus, HCV hepatitis C virus, HHV human herpesvirus, HIV human immunodeficiency virus, HPV human papillomavirus, HSV herpes simplex virus, MCC Matthews…

• 

  Fig. 2. Comprehensive analysis and predictive modeling of pan-antiviral drugs. A Distribution of non-cytotoxic approved and investigational drugs in the Tox21 cell viability assays. Compounds were selected based on a stringent criterion of being inactive (non-cytotoxic) in at least 30 assays. B Structural similarity analysis of approved and investigational antiviral drugs compared to non-cytotoxic pharmaceutical compounds (NCPC) using the Tanimoto coefficient. C Performance evaluation of the optimal predictive models and their respective parameter combinations. Results are presented as mean ± standard deviation (SD), with error bars representing the SD from 20 independent iterations. AUC-ROC area under the receiver operating characteristic curve, BA balanced accuracy, FS feature selection, MCC Matthews correlation coefficient, NB naïve Bayes, NCPC non-cytotoxic pharmaceutical compounds, NNET neural network, RF random forest, SVM support vector machine, XGB eXtreme gradient boosting.

• 

  Fig. 3. Virtual screening for potential anti-SARS-CoV-2 drugs. A Workflow outlining the identification process for potential anti-SARS-CoV-2 drugs. B Comparative distribution of non-cytotoxic anti-SARS-CoV-2 compounds (NCACs) and cytotoxic anti-SARS-CoV-2 compounds (CACs). C Structural similarity analysis of the NCACs and CACs using the Tanimoto coefficient. D Performance evaluation of the optimal predictive models and their respective parameter combinations. Results are presented as mean ± standard deviation (SD), with error bars representing the SD from 20 independent iterations. AUC-ROC area under the receiver operating characteristic curve, BA balanced accuracy, PAM pan-antiviral drug model, FET Fisher’s exact test, FS feature selection, MCC Matthews correlation coefficient, NB naïve Bayes, NNET neural network, SAM selective antiviral drug model, RF random forest, ROSE random over-sampling examples, NCSM non-cytotoxic anti-SARS-CoV-2 model, SVM support vector machine, XGB eXtreme gradient boosting.

• 

  Fig. 4. Concentration-response curves of representative SARS-CoV-2 PP entry inhibitors. Results are presented as mean ± standard deviation (SD), with error bars representing the SD of two independent experiments. PP pseudotyped particle, TOX cytotoxicity assay, IC50 half-maximal inhibitory concentration.

• 

  Fig. 5. Concentration-response curves of representative SARS-CoV-2 RdRp inhibitors. Results are presented as mean ± standard deviation (SD), with error bars representing the SD of two independent experiments. RdRp, RND-dependent RNA polymerase, IC50 half-maximal inhibitory concentration.