Shiny binary modeling platform: a reproducibility-oriented software platform for structured binary modeling with integrated interpretation, validation support, model comparison, and archive export.

Background: Structured binary modeling is often implemented through scattered scripts and temporary files, leaving preprocessing recipes, benchmarks, deployed models, interpretation outputs, and archive materials difficult to trace. This limits practical reproducibility and later reuse. Results: We developed Shiny Binary Modeling Platform, an R/Shiny application that links data setup, preprocessing retention, candidate-model benchmarking, final model deployment, SHAP-based interpretation, external-validation support, model comparison, and archive export in one project workflow. In a demonstration using 6383 traditional Chinese medicine-derived compound records, a source-derived pure-toxic-source label, and 12 ADME/physicochemical descriptors, the platform evaluated 113 candidate model specifications under stratified 10-fold cross-validation. Majority-class downsampling was applied only within training folds; assessment folds retained the original imbalanced distribution. The descriptor set was used to exercise the software workflow and should not be interpreted as a comprehensive endpoint-specific toxicity-QSAR feature set. No independent external validation dataset was used, so reported performance represents internal cross-validation only. Stepglm [both]+RF-a two-stage specification using both-direction stepwise logistic-regression screening followed by random forest-served as the primary performance-summary model for ROC/PR reporting and structured export. With pure_toxic_label = 1 as the toxic/positive class, its merged out-of-fold PR-AUC was 0.2469; at threshold 0.5, toxic-class precision/PPV, recall/sensitivity, F1, and NPV were 0.2288, 0.4091, 0.2935, and 0.9509. Archived global and local SHAP outputs were re-audited separately: the global outputs shown in were traced to the retained Lasso + RF historical reference project, whereas the local waterfall output shown in was traced to the Stepglm [both]+RF primary performance-summary model. Conclusion: This is a software article rather than a final endpoint-specific prediction study. The current use case demonstrates integrated benchmarking, deployment, internal OOF reporting, audited SHAP-output handling, model comparison, and reproducible export. External-validation support and standalone prediction-app export are implemented capabilities, but the single-dataset demonstration does not support independent external-performance, deployment-readiness, or competitive toxicity-QSAR claims.