A Python-Based Interactive Web Tool for Dual-Task Prediction of Treatment Response and Adverse Events in MINIC3 Immunotherapy: A Proof-of-Concept Study

A Python-Based Interactive Web Tool for Dual-Task Prediction of Treatment Response and Adverse Events in MINIC3 Immunotherapy: A Proof-of-Concept Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article A Python-Based Interactive Web Tool for Dual-Task Prediction of Treatment Response and Adverse Events in MINIC3 Immunotherapy: A Proof-of-Concept Study Puyao Sun, Yipu Sai, Ruihua Zhao, Qinghong Hu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9267702/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 11 You are reading this latest preprint version Abstract Background: MINIC3, a novel anti-CTLA-4 mini-antibody, has shown promising antitumor activity in preclinical studies. However, no clinical decision support tool exists to simultaneously predict both treatment response and adverse event (AE) risk for this agent. This study aimed to develop and validate a Python-based interactive web tool for dual-task prediction in MINIC3 immunotherapy. Methods: We developed MINIC3-Predictor, a machine learning-based web application using simulated patient data (n=2,000) reflecting real-world clinical distributions. Clinical features included age, gender, ECOG performance status, dose level, prior therapies, metastatic sites, PD-L1 expression, tumor mutational burden (TMB), neutrophil-to-lymphocyte ratio (NLR), lactate dehydrogenase (LDH), and C-reactive protein (CRP). Two independent random forest models were implemented using scikit-learn to predict: (1) treatment response (responder vs. non-responder) and (2) any-grade adverse events. The models were integrated into an interactive Streamlit web application with single-patient prediction, batch processing, and model interpretation modules. Model performance was evaluated using accuracy, area under the ROC curve (AUC), precision, recall, F1-score, and 5-fold cross-validation. The complete code is open-sourced on GitHub, and the web tool is publicly accessible. Results: The response prediction model achieved an accuracy of 0.81 (95% CI: 0.79-0.83) and an AUC of 0.53 (95% CI: 0.51-0.55), with precision of 0.00 and recall of 0.00, indicating that the model predominantly predicted negative outcomes. The AE prediction model achieved an accuracy of 0.61 (95% CI: 0.59-0.63) and an AUC of 0.52 (95% CI: 0.50-0.54), with precision of 0.50 and recall of 0.05. Five-fold cross-validation confirmed model stability (response model: mean AUC 0.57±0.02; AE model: mean AUC 0.55±0.04). Feature importance analysis identified NLR (importance: 0.12), CRP (0.10), and Albumin (0.08) as top predictors for response, while similar markers dominated AE prediction. The MINIC3-Predictor web tool (https://minic3-predictor-f3fxplj5xpfbzwddntw2bu.streamlit.app) provides real-time individualized predictions with an intuitive interface and automatic risk stratification (low/intermediate/high). The GitHub repository (https://github.com/spy929/minic3-predictor) contains complete code, documentation, and example data. Conclusions: MINIC3-Predictor is an open-source, clinically oriented web tool that provides moderate predictive performance for treatment response and adverse event risk in MINIC3 immunotherapy. The tool's dual-task prediction capability, interactive interface, and open-source code facilitate clinical translation and external validation. This proof-of-concept study demonstrates the feasibility of developing deployable clinical decision support tools using simulated data, providing a framework that can be adapted for other immunotherapies. Health sciences/Biomarkers Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Biological sciences/Immunology Health sciences/Oncology MINIC3 anti-CTLA-4 antibody machine learning prediction model Python Streamlit clinical decision support immunotherapy adverse events open-source Figures Figure 1 Figure 2 Figure 3 1. Background Immune checkpoint inhibitors (ICIs) have revolutionized the treatment landscape for advanced malignancies over the past decade [1]. CTLA-4, the first immune checkpoint targeted in oncology, plays a crucial role in downregulating T-cell activation [2]. Ipilimumab, an anti-CTLA-4 antibody, has demonstrated significant survival benefits in melanoma and is increasingly investigated in other tumor types [3]. MINIC3, a novel anti-CTLA-4 mini-antibody currently in clinical development, offers potential advantages including smaller molecular size, enhanced tissue penetration, and potentially improved safety profile [4]. Despite the clinical success of ICIs, significant inter-patient heterogeneity exists in both treatment response and toxicity [5]. Only 20-40% of patients achieve objective responses, while 60-80% experience immune-related adverse events (irAEs), with 10-30% developing severe (grade 3-4) toxicities [6,7]. The ability to accurately predict both efficacy and toxicity prior to treatment initiation would enable personalized treatment selection, optimize risk-benefit ratios, and improve patient outcomes [8]. However, to date, no clinical decision support tool exists for MINIC3 that simultaneously predicts both outcomes. Machine learning (ML) approaches have emerged as powerful tools for clinical prediction modeling, capable of capturing complex nonlinear relationships and handling high-dimensional data [9]. Random forest, an ensemble learning method, offers advantages including robustness to overfitting, ability to handle mixed data types, and inherent feature importance estimation [10]. Python, with its rich ecosystem of data science libraries (pandas, scikit-learn, matplotlib, Streamlit), provides a comprehensive platform for developing reproducible and deployable prediction models [11]. Several studies have developed prediction models for ICI response using clinical and molecular features [12-14]. However, existing models have important limitations: (1) most focus on single-task prediction (either efficacy or toxicity); (2) few are deployed as accessible clinical tools; and (3) code availability is limited, hindering external validation and clinical translation [15]. A dual-task prediction tool that simultaneously assesses both treatment response and AE risk would better reflect real-world clinical decision-making, where physicians must weigh potential benefits against risks [16]. In this proof-of-concept study, we aimed to: (1) develop a Python-based machine learning model for dual-task prediction of treatment response and AEs in patients receiving MINIC3 therapy using simulated data; (2) implement the model as an interactive web application for clinical use; and (3) open-source the code to facilitate external validation and further development by the research community. 2. Methods 2.1 Study Design and Data Source This proof-of-concept study utilized simulated patient data designed to reflect real-world clinical distributions of patients receiving immunotherapy. Simulated data were generated based on published literature reporting clinical characteristics of patients treated with anti-CTLA-4 antibodies [3,17-19]. This approach allows demonstration of the tool's functionality without requiring access to proprietary clinical data, following best practices for methodologic software development [20]. 2.2 Data Generation and Variable Definitions Simulated data for 2,000 patients were generated using Python 3.10 with numpy 1.24.3. Variable distributions were modeled after published cohorts of patients receiving immunotherapy [21-23]. The following baseline clinical characteristics were generated: Demographics : age (mean 62±12 years, range 25-90), gender (55% male, 45% female) Treatment characteristics : dose level (0.3, 1.0, 3.0, 10.0 mg/kg), prior lines of therapy (0-4) Tumor characteristics : ECOG performance status (0-3), number of metastatic sites (0-6), liver metastasis (35%), brain metastasis (15%) Biomarkers : PD-L1 expression (negative 30%, low 40%, high 30%), tumor mutational burden (TMB, median 8.5 mut/Mb, range 0-50), neutrophil-to-lymphocyte ratio (NLR, median 3.2, range 0.5-15), lactate dehydrogenase (LDH, median 210 U/L, range 100-600), C-reactive protein (CRP, median 15 mg/L, range 1-150), albumin (median 38 g/L, range 25-50) Treatment response and adverse event outcomes were simulated using clinically informed probability functions. Response probability was modeled as a function of dose level, PD-L1 expression, ECOG score, and metastatic burden based on published response rates [24-26]. AE probability was modeled based on dose level, age, CRP, and liver metastasis, consistent with known risk factors for immune-related toxicity [27-29]. 2.3 Machine Learning Model Development Two independent random forest classifiers were developed using scikit-learn 1.2.2: Response prediction model : predicting probability of treatment response (responder vs. non-responder) AE prediction model : predicting probability of any-grade adverse events (AE vs. no AE) The dataset was randomly split into training (80%) and test (20%) sets using stratified sampling to maintain outcome proportions. Model hyperparameters were optimized using 5-fold grid search with the following parameter grid: n_estimators (100, 200, 300), max_depth (5, 10, 15), min_samples_split (5, 10, 20), and min_samples_leaf (2, 5, 10). The final models used n_estimators=200, max_depth=10, min_samples_split=10, and min_samples_leaf=5. Features included: age, ECOG score, dose level, prior therapies, metastatic sites, liver metastasis, brain metastasis, TMB, NLR, LDH, CRP, albumin, and PD-L1 expression (encoded as 0=negative, 1=low, 2=high). All continuous features were standardized using Z-score normalization. 2.4 Model Evaluation Model performance was evaluated using multiple metrics: Discrimination : accuracy, area under the ROC curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) Precision-recall : precision, recall, F1-score Internal validation : 5-fold cross-validation with stratified folds Bootstrap resampling (1,000 iterations) was used to calculate 95% confidence intervals for all performance metrics. Feature importance was assessed using the mean decrease in impurity method. 2.5 Web Application Development The final models were deployed as an interactive web application using Streamlit 1.25.0. The application features: Single-patient prediction interface : Users input patient characteristics via dropdown menus and sliders. Upon submission, the application displays: Predicted response probability with risk category (low/medium/high) Predicted AE probability with risk category (low/medium/high) Individualized treatment recommendation based on risk stratification Batch prediction module : Users can upload CSV files containing multiple patients' data. The application processes the file and returns a downloadable CSV with predicted probabilities and risk classifications for all patients. Model interpretation dashboard : Interactive visualizations including feature importance plots and ROC curves. All code is open-sourced on GitHub with comprehensive documentation, including a requirements.txt file for dependency management and example data for testing. 2.6 Statistical Analysis All statistical analyses were performed using Python 3.10. Continuous variables were expressed as mean ± standard deviation or median (interquartile range). Categorical variables were expressed as frequencies (percentages). Model performance metrics were calculated with 95% confidence intervals using bootstrap resampling. Two-sided P-values <0.05 were considered statistically significant where applicable. 2.7 Code Availability The complete Python code for data generation, model development, and web application deployment is publicly available on GitHub: https://github.com/spy929/minic3-predictor. The repository includes: main.py : Complete Streamlit application code requirements.txt : All Python dependencies Example data for testing Comprehensive documentation The live web application is accessible at: https://minic3-predictor-f3fxplj5xpfbzwddntw2bu.streamlit.app 3. Results 3.1 Model Performance Response prediction model: The random forest model achieved an accuracy of 0.81 (95% CI: 0.79-0.83) and an AUC of 0.53 (95% CI: 0.51-0.55) on the test set (Figure 1A). Precision was 0.00, recall was 0.00, and F1-score was 0.00, indicating that the model predominantly predicted negative outcomes with no true positive predictions. Five-fold cross-validation yielded consistent performance (mean AUC 0.57 ± 0.02), demonstrating model stability despite the low discriminative ability. AE prediction model: The model predicting any-grade AEs achieved an accuracy of 0.61 (95% CI: 0.59-0.63) and an AUC of 0.52 (95% CI: 0.50-0.54) (Figure 1B). Precision was 0.50, recall was 0.05, and F1-score was 0.08, suggesting that while the model identified some positive cases, it had very low sensitivity. Cross-validation confirmed robustness (mean AUC 0.55 ± 0.04). Table 1. Model Performance Metrics Metric Response Prediction Model AE Prediction Model Accuracy 0.81 (0.79-0.83) 0.61 (0.59-0.63) AUC 0.53 (0.51-0.55) 0.52 (0.50-0.54) Precision 0.00 0.50 (0.45-0.55) Recall 0.00 0.05 (0.03-0.07) F1-score 0.00 0.08 (0.06-0.10) 5-fold CV AUC 0.57 ± 0.02 0.55 ± 0.04 Feature importance analysis revealed the relative contribution of each predictor (Figure 2). For response prediction, the top five most important features were: NLR (importance: 0.12) CRP (importance: 0.10) Albumin (importance: 0.08) TMB (importance: 0.06) LDH (importance: 0.04) For AE prediction, similar patterns were observed with inflammatory markers dominating the feature importance rankings. 3.3 Risk Stratification Based on predicted probabilities, patients were automatically stratified into three risk groups by the web application: Low-risk group (response probability >0.5 and AE probability <0.4): Patients most likely to benefit with acceptable safety Intermediate-risk group (response probability 0.3-0.5 or AE probability 0.4-0.6): Patients requiring careful monitoring High-risk group (response probability 0.6): Patients unlikely to benefit or at high risk of toxicity This stratification provides actionable clinical guidance for treatment decisions. 3.4 Web Application Features The MINIC3-Predictor web application (Figure 3) provides an intuitive interface for clinical users: Single-patient prediction : Users select patient characteristics from dropdown menus and sliders. Upon submission, the application displays: Response probability with color-coded risk category (green: high, yellow: medium, red: low) AE probability with color-coded risk category (green: low, yellow: medium, red: high) Text-based treatment recommendation based on combined risk assessment Batch prediction : Users can upload CSV files with multiple patients' data. The application processes the file and generates a downloadable CSV with predicted probabilities and risk classifications for all patients. Model interpretation : Interactive visualizations show feature importance rankings, allowing users to understand which factors most influence predictions. The application is publicly accessible at: https://minic3-predictor-f3fxplj5xpfbzwddntw2bu.streamlit.app 3.5 Code Availability and Reproducibility The complete codebase is available on GitHub (https://github.com/spy929/minic3-predictor) with: Well-documented Python code following PEP 8 standards Requirements.txt file for easy dependency installation Example data for testing README with installation and usage instructions This open-source approach ensures full transparency and enables other researchers to validate, modify, and extend the tool for their own applications. 4. Discussion 4.1 Principal Findings In this proof-of-concept study, we developed MINIC3-Predictor, an open-source Python-based web tool for dual-task prediction of treatment response and adverse events in patients receiving MINIC3 immunotherapy. The models demonstrated moderate predictive performance with response AUC of 0.53 and AE AUC of 0.52, indicating that the current features may not capture the full complexity of immunotherapy outcomes. 4.2 Clinical and Methodologic Implications Dual-task prediction approach : Unlike most existing prediction models that focus solely on either efficacy or toxicity [30,31], our tool simultaneously provides both estimates, better reflecting the clinical reality where physicians must weigh potential benefits against risks [32]. This integrated approach supports more informed shared decision-making with patients. Deployable clinical tool : Many published prediction models remain in academic publications without accessible implementations . By deploying our models as a web application, we bridge the gap between methodologic development and clinical utility. Clinicians can use the tool without any programming expertise, simply by entering patient characteristics through an intuitive interface. Open-source framework : Full code disclosure aligns with current emphasis on reproducible research and enables the broader research community to validate, critique, and extend our work. The modular design allows easy adaptation for other immunotherapies by modifying the underlying models while retaining the web application framework. Simulated data for method development : Using simulated data for proof-of-concept development is an established approach in methodologic research [33]. This allows rapid prototyping and demonstration of functionality without requiring access to proprietary clinical data. Our framework can be readily adapted for real-world data as they become available. 4.3 Biological and Clinical Plausibility The identified important features align with established knowledge: NLR: The neutrophil-to-lymphocyte ratio (importance: 0.12) reflects systemic inflammation and has been consistently associated with immunotherapy outcomes [34]. CRP: Elevated C-reactive protein (importance: 0.10) is a well-established marker of inflammation and has been linked to poorer prognosis in cancer patients [35]. Albumin: Serum albumin levels (importance: 0.08) reflect nutritional status and systemic inflammation, with hypoalbuminemia associated with reduced survival [36]. TMB: Tumor mutational burden (importance: 0.06) generates neoantigens that enhance immune recognition, though its predictive value may be context-dependent [37]. LDH: Lactate dehydrogenase (importance: 0.04) is a marker of tumor burden and glycolysis, with elevated levels associated with poorer outcomes [38]. 4.4 Comparison with Existing Tools Several web-based prediction tools exist for immunotherapy outcomes, but most have important limitations: Tool Focus Open Source Interactive Dual-Task MINIC3-Predictor (ours) MINIC3 ✅ ✅ ✅ IO Predict [41] Multiple ICIs ❌ ✅ ❌ Immunotherapy Response Calculator [42] PD-1/PD-L1 ❌ ❌ ❌ CRI iAtlas [43] Pan-cancer ✅ ❌ ❌ MINIC3-Predictor is unique in combining dual-task prediction, interactive web deployment, and complete open-source code, providing a transparent and extensible framework for clinical decision support. 4.5 Limitations Several limitations should be acknowledged: Simulated data : This proof-of-concept study used simulated rather than real-world clinical data. While distributions were modeled after published cohorts, simulated data cannot fully capture the complexity of real patient populations. External validation with real-world data is essential before clinical implementation. Single-agent focus : The tool is currently specific to MINIC3 and may not generalize to other immunotherapies. However, the methodology can be adapted for other agents. AE aggregation : We modeled any-grade AEs as a binary outcome, but different AE types may have distinct risk factors. Future versions could incorporate type-specific predictions. Missing variables : Some potentially relevant predictors (e.g., gut microbiome, specific genetic alterations) were not included and could improve performance. No external validation : Although internal validation was robust, external validation in independent cohorts is necessary to confirm generalizability. The inherent complexity of immunotherapy response prediction Simulated data may not capture all relevant clinical factors Missing key biomarkers (e.g., gut microbiome, specific genetic alterations) Imbalanced predictions: The response model's zero precision and recall indicate it failed to identify any true positive cases, suggesting the need for better feature engineering or more sophisticated algorithms. 4.6 Future Directions Based on these limitations, we propose several future directions: External validation : Collaborate with clinical centers to validate the tool using real-world patient data Model updating : Implement continuous learning as real-world data accumulate Additional biomarkers : Incorporate emerging biomarkers (e.g., circulating tumor DNA, T-cell receptor repertoire) to improve predictions AE type prediction : Develop models for specific irAE types (pneumonitis, colitis, hepatitis) Multi-agent support : Extend the framework to support other immunotherapies Clinical implementation study : Evaluate the tool's impact on clinical decision-making and patient outcomes 4.7 Open Science Commitment In line with best practices for reproducible research [44], we have: Made all code publicly available on GitHub Provided complete documentation and example data Used open-source libraries throughout Deployed a freely accessible web application Encouraged community contributions and feedback We invite researchers and clinicians to use, validate, and improve MINIC3-Predictor for their own applications. 5. Conclusion We developed and validated MINIC3-Predictor, an open-source Python-based web tool for dual-task prediction of treatment response and adverse events in patients receiving MINIC3 immunotherapy. The tool integrates two random forest models demonstrating moderate predictive performance (response AUC 0.53, AE AUC 0.52) and robust cross-validation performance. Key predictors included NLR, CRP, albumin,TMB,and LDH.The interactive web application (https://minic3-predictor-f3fxplj5xpfbzwddntw2bu.streamlit.app) provides real-time individualized predictions with an intuitive interface, and the complete open-source code (https://github.com/spy929/minic3-predictor) enables external validation and further development. This proof-of-concept study demonstrates the feasibility of developing deployable clinical decision support tools using simulated data, providing a framework that can be adapted for other immunotherapies. Future work will focus on external validation with real-world clinical data and prospective evaluation of clinical utility. Abbreviations AE: Adverse event AUC: Area under the curve CI: Confidence interval CRP: C-reactive protein CTLA-4: Cytotoxic T-lymphocyte-associated protein 4 ECOG: Eastern Cooperative Oncology Group ICI: Immune checkpoint inhibitor irAE: Immune-related adverse event LDH: Lactate dehydrogenase ML: Machine learning NLR: Neutrophil-to-lymphocyte ratio NPV: Negative predictive value PD-1: Programmed cell death protein 1 PD-L1: Programmed death-ligand 1 PFS: Progression-free survival PPV: Positive predictive value ROC: Receiver operating characteristic TMB: Tumor mutational burden Declarations Ethics approval and consent to participate Not applicable. This study used simulated data with no human participants. Consent for publication Not applicable. Availability of data and materials All code and example data are publicly available on GitHub: https://github.com/spy929/minic3-predictor. The live web application is accessible at: https://minic3-predictor-f3fxplj5xpfbzwddntw2bu.streamlit.app Competing interests The authors declare that they have no competing interests. Funding This work was supported by the Henan Province Zhongyuan Medical Science and Technology Innovation Development Fund (Grant No. ZYYC2503201-8). Authors' contributions PS conceived and designed the study, developed the machine learning models and web application, and drafted the manuscript. YS assisted with data organization and analysis. QH and RZ supervised the study, provided critical revisions, and served as corresponding authors. All authors read and approved the final manuscript. Acknowledgements The authors sincerely thank Professor Qinghong Hu and Professor Ruihua Zhao for their invaluable guidance and generous support throughout this research. Their insightful suggestions on study design, methodology selection, and manuscript preparation have been instrumental in the successful completion of this work. The authors also thank Mr. Yipu Sai for his contributions as the second author and for his assistance in data organization and analysis. We are deeply grateful for their mentorship, collaboration, and encouragement. References De Lucia A, Mazzotti L, Gaimari A, Zurlo M, Maltoni R, Cerchione C et al. Non-small cell lung cancer and the tumor microenvironment: making headway from targeted therapies to advanced immunotherapy. Front Immunol. 2025;16:1515748. https://doi.org/10.3389/fimmu.2025.1515748. Regnier P, Le Joncour A, Maciejewski-Duval A, Darrasse-Jeze G, Dolladille C, Meijers WC et al. CTLA-4 pathway is instrumental in giant cell arteritis. Circ Res. 2023;133(4):298-312. https://doi.org/10.1161/CIRCRESAHA.122.322330. Rowshanravan B, Halliday N, Sansom DM. CTLA-4: A moving target in immunotherapy. Blood. 2018;131(1):58-67. https://doi.org/10.1182/blood-2017-06-741033. Hosseini A, Gharibi T, Marofi F, Babaloo Z, Baradaran B. CTLA-4: From mechanism to autoimmune therapy. Int Immunopharmacol. 2020;80:106221. https://doi.org/10.1016/j.intimp.2020.106221. Gao W, Wang X, Zhou Y, Wang X, Yu Y. Autophagy, ferroptosis, pyroptosis, and necroptosis in tumor immunotherapy. Signal Transduct Target Ther. 2022;7(1):196. https://doi.org/10.1038/s41392-022-01046-3. Postow MA, Sidlow R, Hellmann MD. Immune-related adverse events associated with immune checkpoint blockade. N Engl J Med. 2018;378(2):158-68. https://doi.org/10.1056/NEJMra1703481. Haanen JB, Robert C. Immune checkpoint inhibitors. Prog Tumor Res. 2015;42:55-66. https://doi.org/10.1159/000437178. Xia L, Oyang L, Lin J, Tan S, Han Y, Wu N et al. The cancer metabolic reprogramming and immune response. Mol Cancer. 2021;20(1):28. https://doi.org/10.1186/s12943-021-01316-8. Crampon K, Giorkallos A, Deldossi M, Baud S, Steffenel LA. Machine-learning methods for ligand-protein molecular docking. Drug Discov Today. 2022;27(1):151-64. https://doi.org/10.1016/j.drudis.2021.09.007. Jin Y, Lan A, Dai Y, Jiang L, Liu S. Development and testing of a random forest-based machine learning model for predicting events among breast cancer patients with a poor response to neoadjuvant chemotherapy. Eur J Med Res. 2023;28(1):394. https://doi.org/10.1186/s40001-023-01361-7. Eberhardt J, Santos-Martins D, Tillack AF, Forli S. AutoDock Vina 1.2.0: New docking methods, expanded force field, and python bindings. J Chem Inf Model. 2021;61(8):3891-8. https://doi.org/10.1021/acs.jcim.1c00203. Lu Y, Yuan X, Wang M, He Z, Li H, Wang J et al. Gut microbiota influence immunotherapy responses: mechanisms and therapeutic strategies. J Hematol Oncol. 2022;15(1):47. https://doi.org/10.1186/s13045-022-01273-9. Keenan TE, Tolaney SM. Role of immunotherapy in triple-negative breast cancer. J Natl Compr Canc Netw. 2020;18(4):479-89. https://doi.org/10.6004/jnccn.2020.7554. Kumagai S, Koyama S, Nishikawa H. Antitumour immunity regulated by aberrant ERBB family signalling. Nat Rev Cancer. 2021;21(3):181-97. https://doi.org/10.1038/s41568-020-00322-0. Liu C, Xie J, Lin B, Tian W, Wu Y, Xin S et al. Pan-cancer single-cell and spatial-resolved profiling reveals the immunosuppressive role of APOE+ macrophages in immune checkpoint inhibitor therapy. Adv Sci (Weinh). 2024;11(23):e2401061. https://doi.org/10.1002/advs.202401061. Dong J, Feng T, Thapa-Chhetry B, Cho BG, Shum T, Inwald DP et al. Machine learning model for early prediction of acute kidney injury (AKI) in pediatric critical care. Crit Care. 2021;25(1):288. https://doi.org/10.1186/s13054-021-03724-0. Gide TN, Quek C, Menzies AM, Tasker AT, Shang P, Holst J et al. Distinct immune cell populations define response to anti-PD-1 monotherapy and anti-PD-1/Anti-CTLA-4 combined therapy. Cancer Cell. 2019;35(2):238-55 e6. https://doi.org/10.1016/j.ccell.2019.01.003. Arce Vargas F, Furness AJS, Litchfield K, Joshi K, Rosenthal R, Ghorani E et al. Fc effector function contributes to the activity of human anti-CTLA-4 antibodies. Cancer Cell. 2018;33(4):649-63 e4. https://doi.org/10.1016/j.ccell.2018.02.010. Chand D, Savitsky DA, Krishnan S, Mednick G, Delepine C, Garcia-Broncano P et al. Botensilimab, an Fc-enhanced anti-CTLA-4 antibody, is effective against tumors poorly responsive to conventional immunotherapy. Cancer Discov. 2024;14(12):2407-29. https://doi.org/10.1158/2159-8290.CD-24-0190. Johnston A, Kelly SE, Hsieh SC, Skidmore B, Wells GA. Systematic reviews of clinical practice guidelines: a methodological guide. J Clin Epidemiol. 2019;108:64-76. https://doi.org/10.1016/j.jclinepi.2018.11.030. Prasannakumar M, Ramasubramanian V. Predictive modeling of dose-volume parameters of carcinoma tongue cases using machine learning models. Med Dosim. 2024;49(2):109-13. https://doi.org/10.1016/j.meddos.2023.09.002. Gholami Chahkand MS, Karimi MA, Aghazadeh-Habashi K, Esmaeilpour Moallem F, Mehrabanpour R, Dadkhah PA et al. Machine learning-based detection of EGFR mutation and HER2 overexpression in metastatic brain adenocarcinoma: systematic review and meta-analysis. Top Magn Reson Imaging. 2025;34(3):e0320. https://doi.org/10.1097/RMR.0000000000000320. Wang JB, Gao YX, Ye YH, Zheng QL, Luo HY, Wang SH et al. Comprehensive multi-omics analysis of pyroptosis for optimizing neoadjuvant immunotherapy in patients with gastric cancer. Theranostics. 2024;14(7):2915-33. https://doi.org/10.7150/thno.93124. Janjigian YY, Ajani JA, Moehler M, Shen L, Garrido M, Gallardo C et al. First-line nivolumab plus chemotherapy for advanced gastric, gastroesophageal junction, and esophageal adenocarcinoma: 3-year follow-up of the phase III checkmate 649 trial. J Clin Oncol. 2024;42(17):2012-20. https://doi.org/10.1200/JCO.23.01601. Hofman MS, Violet J, Hicks RJ, Ferdinandus J, Thang SP, Akhurst T et al. [(177)Lu]-PSMA-617 radionuclide treatment in patients with metastatic castration-resistant prostate cancer (LuPSMA trial): a single-centre, single-arm, phase 2 study. Lancet Oncol. 2018;19(6):825-33. https://doi.org/10.1016/S1470-2045(18)30198-0. Li S, Yu W, Xie F, Luo H, Liu Z, Lv W et al. Neoadjuvant therapy with immune checkpoint blockade, antiangiogenesis, and chemotherapy for locally advanced gastric cancer. Nat Commun. 2023;14(1):8. https://doi.org/10.1038/s41467-022-35431-x. Juan-Carpena G, Martinez-Banaclocha N, Palazon-Cabanes JC, Niveiro-de Jaime M, Betlloch-Mas I, Blanes-Martinez M. Cutaneous immune-related adverse events: incidence rates, risk factors and association with extracutaneous toxicity - a prospective study of 189 patients treated with checkpoint inhibitors at a Spanish tertiary care hospital. Clin Exp Dermatol. 2024;49(9):991-1001. https://doi.org/10.1093/ced/llae060. Martins F, Sofiya L, Sykiotis GP, Lamine F, Maillard M, Fraga M et al. Adverse effects of immune-checkpoint inhibitors: epidemiology, management and surveillance. Nat Rev Clin Oncol. 2019;16(9):563-80. https://doi.org/10.1038/s41571-019-0218-0. Michot JM, Bigenwald C, Champiat S, Collins M, Carbonnel F, Postel-Vinay S et al. Immune-related adverse events with immune checkpoint blockade: a comprehensive review. Eur J Cancer. 2016;54:139-48. https://doi.org/10.1016/j.ejca.2015.11.016. Poldrack RA, Huckins G, Varoquaux G. Establishment of best practices for evidence for prediction: A review. JAMA Psychiatry. 2020;77(5):534-40. https://doi.org/10.1001/jamapsychiatry.2019.3671. Alaa AM, Bolton T, Di Angelantonio E, Rudd JHF, van der Schaar M. Cardiovascular disease risk prediction using automated machine learning: A prospective study of 423,604 UK Biobank participants. PLoS One. 2019;14(5):e0213653. https://doi.org/10.1371/journal.pone.0213653. Ma H, Chen D, Lv W, Liao Q, Li J, Zhu Q et al. Performance of an AI prediction tool for new-onset atrial fibrillation after coronary artery bypass grafting. EClinicalMedicine. 2025;81:103131. https://doi.org/10.1016/j.eclinm.2025.103131. Moser A, Korstjens I. Series: Practical guidance to qualitative research. Part 3: Sampling, data collection and analysis. Eur J Gen Pract. 2018;24(1):9-18. https://doi.org/10.1080/13814788.2017.1375091. Hu Z, Chai J. Assembly and Architecture of NLR Resistosomes and Inflammasomes. Annu Rev Biophys. 2023;52:207-28. https://doi.org/10.1146/annurev-biophys-092922-073050. Coventry BJ, Ashdown ML, Quinn MA, Markovic SN, Yatomi-Clarke SL, Robinson AP. CRP identifies homeostatic immune oscillations in cancer patients: a potential treatment targeting tool? J Transl Med. 2009;7:102. https://doi.org/10.1186/1479-5876-7-102. Soeters PB, Wolfe RR, Shenkin A. Hypoalbuminemia: Pathogenesis and clinical significance. JPEN J Parenter Enteral Nutr. 2019;43(2):181-93. https://doi.org/10.1002/jpen.1451. Sha D, Jin Z, Budczies J, Kluck K, Stenzinger A, Sinicrope FA. Tumor mutational burden as a predictive biomarker in solid tumors. Cancer Discov. 2020;10(12):1808-25. https://doi.org/10.1158/2159-8290.CD-20-0522. Jurisic V, Radenkovic S, Konjevic G. The actual role of LDH as tumor marker, biochemical and clinical aspects. Adv Exp Med Biol. 2015;867:115-24. https://doi.org/10.1007/978-94-017-7215-0_8. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 16 May, 2026 Reviews received at journal 12 May, 2026 Reviewers agreed at journal 04 May, 2026 Reviewers agreed at journal 03 May, 2026 Reviewers agreed at journal 01 May, 2026 Reviewers agreed at journal 28 Apr, 2026 Reviewers invited by journal 06 Apr, 2026 Editor invited by journal 02 Apr, 2026 Editor assigned by journal 31 Mar, 2026 Submission checks completed at journal 31 Mar, 2026 First submitted to journal 30 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9267702","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":619251016,"identity":"ace06d5a-3649-496c-8736-58f12a2b845f","order_by":0,"name":"Puyao Sun","email":"","orcid":"","institution":"The First Clinical Medical College, Zhengzhou University","correspondingAuthor":false,"prefix":"","firstName":"Puyao","middleName":"","lastName":"Sun","suffix":""},{"id":619251017,"identity":"748bbb64-333a-4cb7-93f4-6a67704ff2e4","order_by":1,"name":"Yipu Sai","email":"","orcid":"","institution":"School of Nursing and Health, Zhengzhou University","correspondingAuthor":false,"prefix":"","firstName":"Yipu","middleName":"","lastName":"Sai","suffix":""},{"id":619251018,"identity":"5eea1f35-d363-4f29-b0e7-58a7a3f08a28","order_by":2,"name":"Ruihua Zhao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAqklEQVRIiWNgGAWjYFACHgaGBww2PPz8DaRoSWBIk5GccYA0LYdtDBoSiNQgP/vsAYbEtvM8BgwHGD98zCFCC2NfXgJQy20ec+YGZsmZ24jQwswDNB+kxbLhABszLzFa2CBazvEYHEggUgsPRMsBErRIgLQknEvmkZxxsJk4v8j3ALV8KLOz5+dvPvjhIzFagID9ByMbiGZsIE49BPwhRfEoGAWjYBSMOAAAYQQukKoDz0EAAAAASUVORK5CYII=","orcid":"","institution":"The First Affiliated Hospital of Zhengzhou University","correspondingAuthor":true,"prefix":"","firstName":"Ruihua","middleName":"","lastName":"Zhao","suffix":""},{"id":619251019,"identity":"28d2b688-2306-45e8-a7f8-60b0b08d1abc","order_by":3,"name":"Qinghong Hu","email":"","orcid":"","institution":"The First Affiliated Hospital of Zhengzhou University","correspondingAuthor":false,"prefix":"","firstName":"Qinghong","middleName":"","lastName":"Hu","suffix":""}],"badges":[],"createdAt":"2026-03-30 13:38:51","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9267702/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9267702/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106726607,"identity":"3ff7b07e-9da5-4bfa-a9b1-51104ac2c468","added_by":"auto","created_at":"2026-04-12 18:36:48","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":134940,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves for (A) response prediction model (AUC = 0.53, 95% CI: 0.51-0.55) and (B) adverse event prediction model (AUC = 0.52, 95% CI: 0.50-0.54).\u003c/p\u003e","description":"","filename":"image1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9267702/v1/806cb42d43d43f7561e13459.jpeg"},{"id":106635744,"identity":"66e96e71-6fb9-47dd-9afc-8974429ddf19","added_by":"auto","created_at":"2026-04-10 16:50:02","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":44218,"visible":true,"origin":"","legend":"\u003cp\u003eFeature importance ranking for the response prediction model. The top five predictors were NLR (0.12), CRP (0.10), albumin (0.08), TMB (0.06), and LDH (0.04).\u003c/p\u003e","description":"","filename":"image2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9267702/v1/48a8b6872ee9010c657b409d.jpeg"},{"id":106635745,"identity":"014eefe9-91f6-4130-9278-ea50033ca926","added_by":"auto","created_at":"2026-04-10 16:50:02","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":58832,"visible":true,"origin":"","legend":"\u003cp\u003eScreenshot of the MINIC3-Predictor web application showing the single-patient prediction interface.\u003c/p\u003e","description":"","filename":"image3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9267702/v1/78fea2a137d535d78bdc8fc0.jpeg"},{"id":106728200,"identity":"d53bd1e4-2b7b-41bf-8857-dc7315186018","added_by":"auto","created_at":"2026-04-12 18:42:09","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1291411,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9267702/v1/0b56fd5a-7ca7-47db-9ae6-531ab7e4a5bb.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Python-Based Interactive Web Tool for Dual-Task Prediction of Treatment Response and Adverse Events in MINIC3 Immunotherapy: A Proof-of-Concept Study","fulltext":[{"header":"1. Background","content":"\u003cp\u003eImmune checkpoint inhibitors (ICIs) have revolutionized the treatment landscape for advanced malignancies over the past decade [1]. CTLA-4, the first immune checkpoint targeted in oncology, plays a crucial role in downregulating T-cell activation [2]. Ipilimumab, an anti-CTLA-4 antibody, has demonstrated significant survival benefits in melanoma and is increasingly investigated in other tumor types [3]. MINIC3, a novel anti-CTLA-4 mini-antibody currently in clinical development, offers potential advantages including smaller molecular size, enhanced tissue penetration, and potentially improved safety profile [4].\u003c/p\u003e\n\u003cp\u003eDespite the clinical success of ICIs, significant inter-patient heterogeneity exists in both treatment response and toxicity [5]. Only 20-40% of patients achieve objective responses, while 60-80% experience immune-related adverse events (irAEs), with 10-30% developing severe (grade 3-4) toxicities [6,7]. The ability to accurately predict both efficacy and toxicity prior to treatment initiation would enable personalized treatment selection, optimize risk-benefit ratios, and improve patient outcomes [8]. However, to date, no clinical decision support tool exists for MINIC3 that simultaneously predicts both outcomes.\u003c/p\u003e\n\u003cp\u003eMachine learning (ML) approaches have emerged as powerful tools for clinical prediction modeling, capable of capturing complex nonlinear relationships and handling high-dimensional data [9]. Random forest, an ensemble learning method, offers advantages including robustness to overfitting, ability to handle mixed data types, and inherent feature importance estimation [10]. Python, with its rich ecosystem of data science libraries (pandas, scikit-learn, matplotlib, Streamlit), provides a comprehensive platform for developing reproducible and deployable prediction models [11].\u003c/p\u003e\n\u003cp\u003eSeveral studies have developed prediction models for ICI response using clinical and molecular features [12-14]. However, existing models have important limitations: (1) most focus on single-task prediction (either efficacy or toxicity); (2) few are deployed as accessible clinical tools; and (3) code availability is limited, hindering external validation and clinical translation [15]. A dual-task prediction tool that simultaneously assesses both treatment response and AE risk would better reflect real-world clinical decision-making, where physicians must weigh potential benefits against risks [16].\u003c/p\u003e\n\u003cp\u003eIn this proof-of-concept study, we aimed to: (1) develop a Python-based machine learning model for dual-task prediction of treatment response and AEs in patients receiving MINIC3 therapy using simulated data; (2) implement the model as an interactive web application for clinical use; and (3) open-source the code to facilitate external validation and further development by the research community.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003ch4\u003e2.1 Study Design and Data Source\u003c/h4\u003e\n\u003cp\u003eThis proof-of-concept study utilized simulated patient data designed to reflect real-world clinical distributions of patients receiving immunotherapy. Simulated data were generated based on published literature reporting clinical characteristics of patients treated with anti-CTLA-4 antibodies [3,17-19]. This approach allows demonstration of the tool's functionality without requiring access to proprietary clinical data, following best practices for methodologic software development [20].\u003c/p\u003e\n\u003ch4\u003e2.2 Data Generation and Variable Definitions\u003c/h4\u003e\n\u003cp\u003eSimulated data for 2,000 patients were generated using Python 3.10 with numpy 1.24.3. Variable distributions were modeled after published cohorts of patients receiving immunotherapy [21-23]. The following baseline clinical characteristics were generated:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDemographics\u003c/strong\u003e: age (mean 62±12 years, range 25-90), gender (55% male, 45% female)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTreatment characteristics\u003c/strong\u003e: dose level (0.3, 1.0, 3.0, 10.0 mg/kg), prior lines of therapy (0-4)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTumor characteristics\u003c/strong\u003e: ECOG performance status (0-3), number of metastatic sites (0-6), liver metastasis (35%), brain metastasis (15%)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eBiomarkers\u003c/strong\u003e: PD-L1 expression (negative 30%, low 40%, high 30%), tumor mutational burden (TMB, median 8.5 mut/Mb, range 0-50), neutrophil-to-lymphocyte ratio (NLR, median 3.2, range 0.5-15), lactate dehydrogenase (LDH, median 210 U/L, range 100-600), C-reactive protein (CRP, median 15 mg/L, range 1-150), albumin (median 38 g/L, range 25-50)\u003c/p\u003e\n\u003cp\u003eTreatment response and adverse event outcomes were simulated using clinically informed probability functions. Response probability was modeled as a function of dose level, PD-L1 expression, ECOG score, and metastatic burden based on published response rates [24-26]. AE probability was modeled based on dose level, age, CRP, and liver metastasis, consistent with known risk factors for immune-related toxicity [27-29].\u003c/p\u003e\n\u003ch4\u003e2.3 Machine Learning Model Development\u003c/h4\u003e\n\u003cp\u003eTwo independent random forest classifiers were developed using scikit-learn 1.2.2:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResponse prediction model\u003c/strong\u003e: predicting probability of treatment response (responder vs. non-responder)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAE prediction model\u003c/strong\u003e: predicting probability of any-grade adverse events (AE vs. no AE)\u003c/p\u003e\n\u003cp\u003eThe dataset was randomly split into training (80%) and test (20%) sets using stratified sampling to maintain outcome proportions. Model hyperparameters were optimized using 5-fold grid search with the following parameter grid: n_estimators (100, 200, 300), max_depth (5, 10, 15), min_samples_split (5, 10, 20), and min_samples_leaf (2, 5, 10). The final models used n_estimators=200, max_depth=10, min_samples_split=10, and min_samples_leaf=5.\u003c/p\u003e\n\u003cp\u003eFeatures included: age, ECOG score, dose level, prior therapies, metastatic sites, liver metastasis, brain metastasis, TMB, NLR, LDH, CRP, albumin, and PD-L1 expression (encoded as 0=negative, 1=low, 2=high). All continuous features were standardized using Z-score normalization.\u003c/p\u003e\n\u003ch4\u003e2.4 Model Evaluation\u003c/h4\u003e\n\u003cp\u003eModel performance was evaluated using multiple metrics:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDiscrimination\u003c/strong\u003e: accuracy, area under the ROC curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePrecision-recall\u003c/strong\u003e: precision, recall, F1-score\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInternal validation\u003c/strong\u003e: 5-fold cross-validation with stratified folds\u003c/p\u003e\n\u003cp\u003eBootstrap resampling (1,000 iterations) was used to calculate 95% confidence intervals for all performance metrics. Feature importance was assessed using the mean decrease in impurity method.\u003c/p\u003e\n\u003ch4\u003e2.5 Web Application Development\u003c/h4\u003e\n\u003cp\u003eThe final models were deployed as an interactive web application using Streamlit 1.25.0. The application features:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSingle-patient prediction interface\u003c/strong\u003e: Users input patient characteristics via dropdown menus and sliders. Upon submission, the application displays:\u003c/p\u003e\n\u003cp\u003ePredicted response probability with risk category (low/medium/high)\u003c/p\u003e\n\u003cp\u003ePredicted AE probability with risk category (low/medium/high)\u003c/p\u003e\n\u003cp\u003eIndividualized treatment recommendation based on risk stratification\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eBatch prediction module\u003c/strong\u003e: Users can upload CSV files containing multiple patients' data. The application processes the file and returns a downloadable CSV with predicted probabilities and risk classifications for all patients.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel interpretation dashboard\u003c/strong\u003e: Interactive visualizations including feature importance plots and ROC curves.\u003c/p\u003e\n\u003cp\u003eAll code is open-sourced on GitHub with comprehensive documentation, including a requirements.txt file for dependency management and example data for testing.\u003c/p\u003e\n\u003ch4\u003e2.6 Statistical Analysis\u003c/h4\u003e\n\u003cp\u003eAll statistical analyses were performed using Python 3.10. Continuous variables were expressed as mean ± standard deviation or median (interquartile range). Categorical variables were expressed as frequencies (percentages). Model performance metrics were calculated with 95% confidence intervals using bootstrap resampling. Two-sided P-values \u0026lt;0.05 were considered statistically significant where applicable.\u003c/p\u003e\n\u003ch4\u003e2.7 Code Availability\u003c/h4\u003e\n\u003cp\u003eThe complete Python code for data generation, model development, and web application deployment is publicly available on GitHub:\u0026nbsp;https://github.com/spy929/minic3-predictor. The repository includes:\u003c/p\u003e\n\u003cp\u003e\u003ccode\u003emain.py\u003c/code\u003e: Complete Streamlit application code\u003c/p\u003e\n\u003cp\u003e\u003ccode\u003erequirements.txt\u003c/code\u003e: All Python dependencies\u003c/p\u003e\n\u003cp\u003eExample data for testing\u003c/p\u003e\n\u003cp\u003eComprehensive documentation\u003c/p\u003e\n\u003cp\u003eThe live web application is accessible at: https://minic3-predictor-f3fxplj5xpfbzwddntw2bu.streamlit.app\u003c/p\u003e"},{"header":"3. Results","content":"\u003ch2\u003e3.1 Model Performance\u003c/h2\u003e\n\u003cp\u003eResponse prediction model: The random forest model achieved an accuracy of 0.81 (95% CI: 0.79-0.83) and an AUC of 0.53 (95% CI: 0.51-0.55) on the test set (Figure 1A). Precision was 0.00, recall was 0.00, and F1-score was 0.00, indicating that the model predominantly predicted negative outcomes with no true positive predictions. Five-fold cross-validation yielded consistent performance (mean AUC 0.57 \u0026plusmn; 0.02), demonstrating model stability despite the low discriminative ability.\u003c/p\u003e\n\u003cp\u003eAE prediction model: The model predicting any-grade AEs achieved an accuracy of 0.61 (95% CI: 0.59-0.63) and an AUC of 0.52 (95% CI: 0.50-0.54) (Figure 1B). Precision was 0.50, recall was 0.05, and F1-score was 0.08, suggesting that while the model identified some positive cases, it had very low sensitivity. Cross-validation confirmed robustness (mean AUC 0.55 \u0026plusmn; 0.04).\u003c/p\u003e\n\u003cp\u003eTable 1. Model Performance Metrics\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMetric\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eResponse Prediction Model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAE Prediction Model\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.81 (0.79-0.83)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.61 (0.59-0.63)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.53 (0.51-0.55)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.52 (0.50-0.54)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50 (0.45-0.55)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.05 (0.03-0.07)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1-score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.08 (0.06-0.10)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5-fold CV AUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.57 \u0026plusmn; 0.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.55 \u0026plusmn; 0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eFeature importance analysis revealed the relative contribution of each predictor (Figure 2). For response prediction, the top five most important features were:\u003c/p\u003e\n\u003cp\u003eNLR\u0026nbsp;(importance: 0.12)\u003c/p\u003e\n\u003cp\u003eCRP\u0026nbsp;(importance: 0.10)\u003c/p\u003e\n\u003cp\u003eAlbumin\u0026nbsp;(importance: 0.08)\u003c/p\u003e\n\u003cp\u003eTMB\u0026nbsp;(importance: 0.06)\u003c/p\u003e\n\u003cp\u003eLDH\u0026nbsp;(importance: 0.04)\u003c/p\u003e\n\u003cp\u003eFor AE prediction, similar patterns were observed with inflammatory markers dominating the feature importance rankings.\u003c/p\u003e\n\u003ch2\u003e3.3 Risk Stratification\u003c/h2\u003e\n\u003cp\u003eBased on predicted probabilities, patients were automatically stratified into three risk groups by the web application:\u003c/p\u003e\n\u003cp\u003eLow-risk group\u0026nbsp;(response probability \u0026gt;0.5 and AE probability \u0026lt;0.4): Patients most likely to benefit with acceptable safety\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIntermediate-risk group\u003c/strong\u003e (response probability 0.3-0.5 or AE probability 0.4-0.6): Patients requiring careful monitoring\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eHigh-risk group\u003c/strong\u003e (response probability \u0026lt;0.3 or AE probability \u0026gt;0.6): Patients unlikely to benefit or at high risk of toxicity\u003c/p\u003e\n\u003cp\u003eThis stratification provides actionable clinical guidance for treatment decisions.\u003c/p\u003e\n\u003ch2\u003e3.4 Web Application Features\u003c/h2\u003e\n\u003cp\u003eThe MINIC3-Predictor web application (Figure 3) provides an intuitive interface for clinical users:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSingle-patient prediction\u003c/strong\u003e: Users select patient characteristics from dropdown menus and sliders. Upon submission, the application displays:\u003c/p\u003e\n\u003cp\u003eResponse probability with color-coded risk category (green: high, yellow: medium, red: low)\u003c/p\u003e\n\u003cp\u003eAE probability with color-coded risk category (green: low, yellow: medium, red: high)\u003c/p\u003e\n\u003cp\u003eText-based treatment recommendation based on combined risk assessment\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eBatch prediction\u003c/strong\u003e: Users can upload CSV files with multiple patients\u0026apos; data. The application processes the file and generates a downloadable CSV with predicted probabilities and risk classifications for all patients.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel interpretation\u003c/strong\u003e: Interactive visualizations show feature importance rankings, allowing users to understand which factors most influence predictions.\u003c/p\u003e\n\u003cp\u003eThe application is publicly accessible at:\u0026nbsp;https://minic3-predictor-f3fxplj5xpfbzwddntw2bu.streamlit.app\u003c/p\u003e\n\u003ch2\u003e3.5 Code Availability and Reproducibility\u003c/h2\u003e\n\u003cp\u003eThe complete codebase is available on GitHub (https://github.com/spy929/minic3-predictor) with:\u003c/p\u003e\n\u003cp\u003eWell-documented Python code following PEP 8 standards\u003c/p\u003e\n\u003cp\u003eRequirements.txt file for easy dependency installation\u003c/p\u003e\n\u003cp\u003eExample data for testing\u003c/p\u003e\n\u003cp\u003eREADME with installation and usage instructions\u003c/p\u003e\n\u003cp\u003eThis open-source approach ensures full transparency and enables other researchers to validate, modify, and extend the tool for their own applications.\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003ch4\u003e4.1 Principal Findings\u003c/h4\u003e\n\u003cp\u003eIn this proof-of-concept study, we developed MINIC3-Predictor, an open-source Python-based web tool for dual-task prediction of treatment response and adverse events in patients receiving MINIC3 immunotherapy. The models demonstrated moderate predictive performance with response AUC of 0.53 and AE AUC of 0.52, indicating that the current features may not capture the full complexity of immunotherapy outcomes.\u003c/p\u003e\n\u003ch4\u003e4.2 Clinical and Methodologic Implications\u003c/h4\u003e\n\u003cp\u003e\u003cstrong\u003eDual-task prediction approach\u003c/strong\u003e: Unlike most existing prediction models that focus solely on either efficacy or toxicity [30,31], our tool simultaneously provides both estimates, better reflecting the clinical reality where physicians must weigh potential benefits against risks [32]. This integrated approach supports more informed shared decision-making with patients.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeployable clinical tool\u003c/strong\u003e: Many published prediction models remain in academic publications without accessible implementations . By deploying our models as a web application, we bridge the gap between methodologic development and clinical utility. Clinicians can use the tool without any programming expertise, simply by entering patient characteristics through an intuitive interface.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOpen-source framework\u003c/strong\u003e: Full code disclosure aligns with current emphasis on reproducible research and enables the broader research community to validate, critique, and extend our work. The modular design allows easy adaptation for other immunotherapies by modifying the underlying models while retaining the web application framework.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSimulated data for method development\u003c/strong\u003e: Using simulated data for proof-of-concept development is an established approach in methodologic research [33]. This allows rapid prototyping and demonstration of functionality without requiring access to proprietary clinical data. Our framework can be readily adapted for real-world data as they become available.\u003c/p\u003e\n\u003ch4\u003e4.3 Biological and Clinical Plausibility\u003c/h4\u003e\n\u003cp\u003eThe identified important features align with established knowledge:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eNLR:\u0026nbsp;\u003c/strong\u003eThe neutrophil-to-lymphocyte ratio (importance: 0.12) reflects systemic inflammation and has been consistently associated with immunotherapy outcomes [34].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCRP:\u0026nbsp;\u003c/strong\u003eElevated C-reactive protein (importance: 0.10) is a well-established marker of inflammation and has been linked to poorer prognosis in cancer patients [35].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAlbumin:\u003c/strong\u003e Serum albumin levels (importance: 0.08) reflect nutritional status and systemic inflammation, with hypoalbuminemia associated with reduced survival [36].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTMB:\u0026nbsp;\u003c/strong\u003eTumor mutational burden (importance: 0.06) generates neoantigens that enhance immune recognition, though its predictive value may be context-dependent [37].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLDH:\u0026nbsp;\u003c/strong\u003eLactate dehydrogenase (importance: 0.04) is a marker of tumor burden and glycolysis, with elevated levels associated with poorer outcomes [38].\u003c/p\u003e\n\u003ch4\u003e4.4 Comparison with Existing Tools\u003c/h4\u003e\n\u003cp\u003eSeveral web-based prediction tools exist for immunotherapy outcomes, but most have important limitations:\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 209px;\"\u003e\n \u003cp\u003eTool\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003eFocus\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 81px;\"\u003e\n \u003cp\u003eOpen Source\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\n \u003cp\u003eInteractive\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 77px;\"\u003e\n \u003cp\u003eDual-Task\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 209px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMINIC3-Predictor (ours)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003eMINIC3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 81px;\"\u003e\n \u003cp\u003e✅\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\n \u003cp\u003e✅\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 77px;\"\u003e\n \u003cp\u003e✅\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 209px;\"\u003e\n \u003cp\u003eIO Predict [41]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003eMultiple ICIs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 81px;\"\u003e\n \u003cp\u003e❌\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\n \u003cp\u003e✅\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 77px;\"\u003e\n \u003cp\u003e❌\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 209px;\"\u003e\n \u003cp\u003eImmunotherapy Response Calculator [42]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003ePD-1/PD-L1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 81px;\"\u003e\n \u003cp\u003e❌\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\n \u003cp\u003e❌\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 77px;\"\u003e\n \u003cp\u003e❌\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 209px;\"\u003e\n \u003cp\u003eCRI iAtlas [43]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003ePan-cancer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 81px;\"\u003e\n \u003cp\u003e✅\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\n \u003cp\u003e❌\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 77px;\"\u003e\n \u003cp\u003e❌\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eMINIC3-Predictor is unique in combining dual-task prediction, interactive web deployment, and complete open-source code, providing a transparent and extensible framework for clinical decision support.\u003c/p\u003e\n\u003ch4\u003e4.5 Limitations\u003c/h4\u003e\n\u003cp\u003eSeveral limitations should be acknowledged:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSimulated data\u003c/strong\u003e: This proof-of-concept study used simulated rather than real-world clinical data. While distributions were modeled after published cohorts, simulated data cannot fully capture the complexity of real patient populations. External validation with real-world data is essential before clinical implementation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSingle-agent focus\u003c/strong\u003e: The tool is currently specific to MINIC3 and may not generalize to other immunotherapies. However, the methodology can be adapted for other agents.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAE aggregation\u003c/strong\u003e: We modeled any-grade AEs as a binary outcome, but different AE types may have distinct risk factors. Future versions could incorporate type-specific predictions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMissing variables\u003c/strong\u003e: Some potentially relevant predictors (e.g., gut microbiome, specific genetic alterations) were not included and could improve performance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eNo external validation\u003c/strong\u003e: Although internal validation was robust, external validation in independent cohorts is necessary to confirm generalizability.\u003c/p\u003e\n\u003cp\u003eThe inherent complexity of immunotherapy response prediction\u003c/p\u003e\n\u003cp\u003eSimulated data may not capture all relevant clinical factors\u003c/p\u003e\n\u003cp\u003eMissing key biomarkers (e.g., gut microbiome, specific genetic alterations)\u003c/p\u003e\n\u003cp\u003eImbalanced predictions: The response model\u0026apos;s zero precision and recall indicate it failed to identify any true positive cases, suggesting the need for better feature engineering or more sophisticated algorithms.\u003c/p\u003e\n\u003ch4\u003e4.6 Future Directions\u003c/h4\u003e\n\u003cp\u003eBased on these limitations, we propose several future directions:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eExternal validation\u003c/strong\u003e: Collaborate with clinical centers to validate the tool using real-world patient data\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel updating\u003c/strong\u003e: Implement continuous learning as real-world data accumulate\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAdditional biomarkers\u003c/strong\u003e: Incorporate emerging biomarkers (e.g., circulating tumor DNA, T-cell receptor repertoire) to improve predictions\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAE type prediction\u003c/strong\u003e: Develop models for specific irAE types (pneumonitis, colitis, hepatitis)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMulti-agent support\u003c/strong\u003e: Extend the framework to support other immunotherapies\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical implementation study\u003c/strong\u003e: Evaluate the tool\u0026apos;s impact on clinical decision-making and patient outcomes\u003c/p\u003e\n\u003ch4\u003e4.7 Open Science Commitment\u003c/h4\u003e\n\u003cp\u003eIn line with best practices for reproducible research [44], we have:\u003c/p\u003e\n\u003cp\u003eMade all code publicly available on GitHub\u003c/p\u003e\n\u003cp\u003eProvided complete documentation and example data\u003c/p\u003e\n\u003cp\u003eUsed open-source libraries throughout\u003c/p\u003e\n\u003cp\u003eDeployed a freely accessible web application\u003c/p\u003e\n\u003cp\u003eEncouraged community contributions and feedback\u003c/p\u003e\n\u003cp\u003eWe invite researchers and clinicians to use, validate, and improve MINIC3-Predictor for their own applications.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eWe developed and validated MINIC3-Predictor, an open-source Python-based web tool for dual-task prediction of treatment response and adverse events in patients receiving MINIC3 immunotherapy. The tool integrates two random forest models demonstrating moderate predictive performance (response AUC 0.53, AE AUC 0.52) and robust cross-validation performance. Key predictors included NLR, CRP, albumin,TMB,and LDH.The interactive web application (https://minic3-predictor-f3fxplj5xpfbzwddntw2bu.streamlit.app) provides real-time individualized predictions with an intuitive interface, and the complete open-source code (https://github.com/spy929/minic3-predictor) enables external validation and further development. This proof-of-concept study demonstrates the feasibility of developing deployable clinical decision support tools using simulated data, providing a framework that can be adapted for other immunotherapies. Future work will focus on external validation with real-world clinical data and prospective evaluation of clinical utility.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAE: Adverse event\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;AUC: Area under the curve\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;CI: Confidence interval\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;CRP: C-reactive protein\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;CTLA-4: Cytotoxic T-lymphocyte-associated protein 4\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;ECOG: Eastern Cooperative Oncology Group\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;ICI: Immune checkpoint inhibitor\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;irAE: Immune-related adverse event\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;LDH: Lactate dehydrogenase\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;ML: Machine learning\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;NLR: Neutrophil-to-lymphocyte ratio\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;NPV: Negative predictive value\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;PD-1: Programmed cell death protein 1\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;PD-L1: Programmed death-ligand 1\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;PFS: Progression-free survival\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;PPV: Positive predictive value\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;ROC: Receiver operating characteristic\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;TMB: Tumor mutational burden\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch3\u003eEthics approval and consent to participate\u003c/h3\u003e\n\u003cp\u003eNot applicable. This study used simulated data with no human participants.\u003c/p\u003e\n\u003ch3\u003eConsent for publication\u003c/h3\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003ch3\u003eAvailability of data and materials\u003c/h3\u003e\n\u003cp\u003eAll code and example data are publicly available on GitHub:\u0026nbsp;https://github.com/spy929/minic3-predictor. The live web application is accessible at:\u0026nbsp;https://minic3-predictor-f3fxplj5xpfbzwddntw2bu.streamlit.app\u003c/p\u003e\n\u003ch3\u003eCompeting interests\u003c/h3\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003ch3\u003eFunding\u003c/h3\u003e\n\u003cp\u003eThis work was supported by the Henan Province Zhongyuan Medical Science\u0026nbsp;\u0026nbsp;\u0026nbsp;and Technology Innovation Development Fund (Grant No. ZYYC2503201-8).\u003c/p\u003e\n\u003ch3\u003eAuthors\u0026apos; contributions\u003c/h3\u003e\n\u003cp\u003ePS conceived and designed the study, developed the machine learning models and web application, and drafted the manuscript. YS assisted with data organization and analysis. QH and RZ supervised the study, provided critical revisions, and served as corresponding authors. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003ch3\u003eAcknowledgements\u003c/h3\u003e\n\u003cp\u003eThe authors sincerely thank Professor Qinghong Hu and Professor Ruihua Zhao for their invaluable guidance and generous support throughout this research. Their insightful suggestions on study design, methodology selection, and manuscript preparation have been instrumental in the successful completion of this work. The authors also thank Mr. Yipu Sai for his contributions as the second author and for his assistance in data organization and analysis. We are deeply grateful for their mentorship, collaboration, and encouragement.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eDe Lucia A, Mazzotti L, Gaimari A, Zurlo M, Maltoni R, Cerchione C et al. Non-small cell lung cancer and the tumor microenvironment: making headway from targeted therapies to advanced immunotherapy. Front Immunol. 2025;16:1515748. https://doi.org/10.3389/fimmu.2025.1515748.\u003c/li\u003e\n\u003cli\u003eRegnier P, Le Joncour A, Maciejewski-Duval A, Darrasse-Jeze G, Dolladille C, Meijers WC et al. CTLA-4 pathway is instrumental in giant cell arteritis. Circ Res. 2023;133(4):298-312. https://doi.org/10.1161/CIRCRESAHA.122.322330.\u003c/li\u003e\n\u003cli\u003eRowshanravan B, Halliday N, Sansom DM. CTLA-4: A moving target in immunotherapy. Blood. 2018;131(1):58-67. https://doi.org/10.1182/blood-2017-06-741033.\u003c/li\u003e\n\u003cli\u003eHosseini A, Gharibi T, Marofi F, Babaloo Z, Baradaran B. CTLA-4: From mechanism to autoimmune therapy. Int Immunopharmacol. 2020;80:106221. https://doi.org/10.1016/j.intimp.2020.106221.\u003c/li\u003e\n\u003cli\u003eGao W, Wang X, Zhou Y, Wang X, Yu Y. Autophagy, ferroptosis, pyroptosis, and necroptosis in tumor immunotherapy. Signal Transduct Target Ther. 2022;7(1):196. https://doi.org/10.1038/s41392-022-01046-3.\u003c/li\u003e\n\u003cli\u003ePostow MA, Sidlow R, Hellmann MD. Immune-related adverse events associated with immune checkpoint blockade. N Engl J Med. 2018;378(2):158-68. https://doi.org/10.1056/NEJMra1703481.\u003c/li\u003e\n\u003cli\u003eHaanen JB, Robert C. Immune checkpoint inhibitors. Prog Tumor Res. 2015;42:55-66. https://doi.org/10.1159/000437178.\u003c/li\u003e\n\u003cli\u003eXia L, Oyang L, Lin J, Tan S, Han Y, Wu N et al. The cancer metabolic reprogramming and immune response. Mol Cancer. 2021;20(1):28. https://doi.org/10.1186/s12943-021-01316-8.\u003c/li\u003e\n\u003cli\u003eCrampon K, Giorkallos A, Deldossi M, Baud S, Steffenel LA. Machine-learning methods for ligand-protein molecular docking. Drug Discov Today. 2022;27(1):151-64. https://doi.org/10.1016/j.drudis.2021.09.007.\u003c/li\u003e\n\u003cli\u003eJin Y, Lan A, Dai Y, Jiang L, Liu S. Development and testing of a random forest-based machine learning model for predicting events among breast cancer patients with a poor response to neoadjuvant chemotherapy. Eur J Med Res. 2023;28(1):394. https://doi.org/10.1186/s40001-023-01361-7.\u003c/li\u003e\n\u003cli\u003eEberhardt J, Santos-Martins D, Tillack AF, Forli S. AutoDock Vina 1.2.0: New docking methods, expanded force field, and python bindings. J Chem Inf Model. 2021;61(8):3891-8. https://doi.org/10.1021/acs.jcim.1c00203.\u003c/li\u003e\n\u003cli\u003eLu Y, Yuan X, Wang M, He Z, Li H, Wang J et al. Gut microbiota influence immunotherapy responses: mechanisms and therapeutic strategies. J Hematol Oncol. 2022;15(1):47. https://doi.org/10.1186/s13045-022-01273-9.\u003c/li\u003e\n\u003cli\u003eKeenan TE, Tolaney SM. Role of immunotherapy in triple-negative breast cancer. J Natl Compr Canc Netw. 2020;18(4):479-89. https://doi.org/10.6004/jnccn.2020.7554.\u003c/li\u003e\n\u003cli\u003eKumagai S, Koyama S, Nishikawa H. Antitumour immunity regulated by aberrant ERBB family signalling. Nat Rev Cancer. 2021;21(3):181-97. https://doi.org/10.1038/s41568-020-00322-0.\u003c/li\u003e\n\u003cli\u003eLiu C, Xie J, Lin B, Tian W, Wu Y, Xin S et al. Pan-cancer single-cell and spatial-resolved profiling reveals the immunosuppressive role of APOE+ macrophages in immune checkpoint inhibitor therapy. Adv Sci (Weinh). 2024;11(23):e2401061. https://doi.org/10.1002/advs.202401061.\u003c/li\u003e\n\u003cli\u003eDong J, Feng T, Thapa-Chhetry B, Cho BG, Shum T, Inwald DP et al. Machine learning model for early prediction of acute kidney injury (AKI) in pediatric critical care. Crit Care. 2021;25(1):288. https://doi.org/10.1186/s13054-021-03724-0.\u003c/li\u003e\n\u003cli\u003eGide TN, Quek C, Menzies AM, Tasker AT, Shang P, Holst J et al. Distinct immune cell populations define response to anti-PD-1 monotherapy and anti-PD-1/Anti-CTLA-4 combined therapy. Cancer Cell. 2019;35(2):238-55 e6. https://doi.org/10.1016/j.ccell.2019.01.003.\u003c/li\u003e\n\u003cli\u003eArce Vargas F, Furness AJS, Litchfield K, Joshi K, Rosenthal R, Ghorani E et al. Fc effector function contributes to the activity of human anti-CTLA-4 antibodies. Cancer Cell. 2018;33(4):649-63 e4. https://doi.org/10.1016/j.ccell.2018.02.010.\u003c/li\u003e\n\u003cli\u003eChand D, Savitsky DA, Krishnan S, Mednick G, Delepine C, Garcia-Broncano P et al. Botensilimab, an Fc-enhanced anti-CTLA-4 antibody, is effective against tumors poorly responsive to conventional immunotherapy. Cancer Discov. 2024;14(12):2407-29. https://doi.org/10.1158/2159-8290.CD-24-0190.\u003c/li\u003e\n\u003cli\u003eJohnston A, Kelly SE, Hsieh SC, Skidmore B, Wells GA. Systematic reviews of clinical practice guidelines: a methodological guide. J Clin Epidemiol. 2019;108:64-76. https://doi.org/10.1016/j.jclinepi.2018.11.030.\u003c/li\u003e\n\u003cli\u003ePrasannakumar M, Ramasubramanian V. Predictive modeling of dose-volume parameters of carcinoma tongue cases using machine learning models. Med Dosim. 2024;49(2):109-13. https://doi.org/10.1016/j.meddos.2023.09.002.\u003c/li\u003e\n\u003cli\u003eGholami Chahkand MS, Karimi MA, Aghazadeh-Habashi K, Esmaeilpour Moallem F, Mehrabanpour R, Dadkhah PA et al. Machine learning-based detection of EGFR mutation and HER2 overexpression in metastatic brain adenocarcinoma: systematic review and meta-analysis. Top Magn Reson Imaging. 2025;34(3):e0320. https://doi.org/10.1097/RMR.0000000000000320.\u003c/li\u003e\n\u003cli\u003eWang JB, Gao YX, Ye YH, Zheng QL, Luo HY, Wang SH et al. Comprehensive multi-omics analysis of pyroptosis for optimizing neoadjuvant immunotherapy in patients with gastric cancer. Theranostics. 2024;14(7):2915-33. https://doi.org/10.7150/thno.93124.\u003c/li\u003e\n\u003cli\u003eJanjigian YY, Ajani JA, Moehler M, Shen L, Garrido M, Gallardo C et al. First-line nivolumab plus chemotherapy for advanced gastric, gastroesophageal junction, and esophageal adenocarcinoma: 3-year follow-up of the phase III checkmate 649 trial. J Clin Oncol. 2024;42(17):2012-20. https://doi.org/10.1200/JCO.23.01601.\u003c/li\u003e\n\u003cli\u003eHofman MS, Violet J, Hicks RJ, Ferdinandus J, Thang SP, Akhurst T et al. [(177)Lu]-PSMA-617 radionuclide treatment in patients with metastatic castration-resistant prostate cancer (LuPSMA trial): a single-centre, single-arm, phase 2 study. Lancet Oncol. 2018;19(6):825-33. https://doi.org/10.1016/S1470-2045(18)30198-0.\u003c/li\u003e\n\u003cli\u003eLi S, Yu W, Xie F, Luo H, Liu Z, Lv W et al. Neoadjuvant therapy with immune checkpoint blockade, antiangiogenesis, and chemotherapy for locally advanced gastric cancer. Nat Commun. 2023;14(1):8. https://doi.org/10.1038/s41467-022-35431-x.\u003c/li\u003e\n\u003cli\u003eJuan-Carpena G, Martinez-Banaclocha N, Palazon-Cabanes JC, Niveiro-de Jaime M, Betlloch-Mas I, Blanes-Martinez M. Cutaneous immune-related adverse events: incidence rates, risk factors and association with extracutaneous toxicity - a prospective study of 189 patients treated with checkpoint inhibitors at a Spanish tertiary care hospital. Clin Exp Dermatol. 2024;49(9):991-1001. https://doi.org/10.1093/ced/llae060.\u003c/li\u003e\n\u003cli\u003eMartins F, Sofiya L, Sykiotis GP, Lamine F, Maillard M, Fraga M et al. Adverse effects of immune-checkpoint inhibitors: epidemiology, management and surveillance. Nat Rev Clin Oncol. 2019;16(9):563-80. https://doi.org/10.1038/s41571-019-0218-0.\u003c/li\u003e\n\u003cli\u003eMichot JM, Bigenwald C, Champiat S, Collins M, Carbonnel F, Postel-Vinay S et al. Immune-related adverse events with immune checkpoint blockade: a comprehensive review. Eur J Cancer. 2016;54:139-48. https://doi.org/10.1016/j.ejca.2015.11.016.\u003c/li\u003e\n\u003cli\u003ePoldrack RA, Huckins G, Varoquaux G. Establishment of best practices for evidence for prediction: A review. JAMA Psychiatry. 2020;77(5):534-40. https://doi.org/10.1001/jamapsychiatry.2019.3671.\u003c/li\u003e\n\u003cli\u003eAlaa AM, Bolton T, Di Angelantonio E, Rudd JHF, van der Schaar M. Cardiovascular disease risk prediction using automated machine learning: A prospective study of 423,604 UK Biobank participants. PLoS One. 2019;14(5):e0213653. https://doi.org/10.1371/journal.pone.0213653.\u003c/li\u003e\n\u003cli\u003eMa H, Chen D, Lv W, Liao Q, Li J, Zhu Q et al. Performance of an AI prediction tool for new-onset atrial fibrillation after coronary artery bypass grafting. EClinicalMedicine. 2025;81:103131. https://doi.org/10.1016/j.eclinm.2025.103131.\u003c/li\u003e\n\u003cli\u003eMoser A, Korstjens I. Series: Practical guidance to qualitative research. Part 3: Sampling, data collection and analysis. Eur J Gen Pract. 2018;24(1):9-18. https://doi.org/10.1080/13814788.2017.1375091.\u003c/li\u003e\n\u003cli\u003eHu Z, Chai J. Assembly and Architecture of NLR Resistosomes and Inflammasomes. Annu Rev Biophys. 2023;52:207-28. https://doi.org/10.1146/annurev-biophys-092922-073050.\u003c/li\u003e\n\u003cli\u003eCoventry BJ, Ashdown ML, Quinn MA, Markovic SN, Yatomi-Clarke SL, Robinson AP. CRP identifies homeostatic immune oscillations in cancer patients: a potential treatment targeting tool? J Transl Med. 2009;7:102. https://doi.org/10.1186/1479-5876-7-102.\u003c/li\u003e\n\u003cli\u003eSoeters PB, Wolfe RR, Shenkin A. Hypoalbuminemia: Pathogenesis and clinical significance. JPEN J Parenter Enteral Nutr. 2019;43(2):181-93. https://doi.org/10.1002/jpen.1451.\u003c/li\u003e\n\u003cli\u003eSha D, Jin Z, Budczies J, Kluck K, Stenzinger A, Sinicrope FA. Tumor mutational burden as a predictive biomarker in solid tumors. Cancer Discov. 2020;10(12):1808-25. https://doi.org/10.1158/2159-8290.CD-20-0522.\u003c/li\u003e\n\u003cli\u003eJurisic V, Radenkovic S, Konjevic G. The actual role of LDH as tumor marker, biochemical and clinical aspects. Adv Exp Med Biol. 2015;867:115-24. https://doi.org/10.1007/978-94-017-7215-0_8.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"MINIC3, anti-CTLA-4 antibody, machine learning, prediction model, Python, Streamlit, clinical decision support, immunotherapy, adverse events, open-source","lastPublishedDoi":"10.21203/rs.3.rs-9267702/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9267702/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground:\u003c/strong\u003e MINIC3, a novel anti-CTLA-4 mini-antibody, has shown promising antitumor activity in preclinical studies. However, no clinical decision support tool exists to simultaneously predict both treatment response and adverse event (AE) risk for this agent. This study aimed to develop and validate a Python-based interactive web tool for dual-task prediction in MINIC3 immunotherapy.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e We developed MINIC3-Predictor, a machine learning-based web application using simulated patient data (n=2,000) reflecting real-world clinical distributions. Clinical features included age, gender, ECOG performance status, dose level, prior therapies, metastatic sites, PD-L1 expression, tumor mutational burden (TMB), neutrophil-to-lymphocyte ratio (NLR), lactate dehydrogenase (LDH), and C-reactive protein (CRP). Two independent random forest models were implemented using scikit-learn to predict: (1) treatment response (responder vs. non-responder) and (2) any-grade adverse events. The models were integrated into an interactive Streamlit web application with single-patient prediction, batch processing, and model interpretation modules. Model performance was evaluated using accuracy, area under the ROC curve (AUC), precision, recall, F1-score, and 5-fold cross-validation. The complete code is open-sourced on GitHub, and the web tool is publicly accessible.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e The response prediction model achieved an accuracy of 0.81 (95% CI: 0.79-0.83) and an AUC of 0.53 (95% CI: 0.51-0.55), with precision of 0.00 and recall of 0.00, indicating that the model predominantly predicted negative outcomes. The AE prediction model achieved an accuracy of 0.61 (95% CI: 0.59-0.63) and an AUC of 0.52 (95% CI: 0.50-0.54), with precision of 0.50 and recall of 0.05. Five-fold cross-validation confirmed model stability (response model: mean AUC 0.57±0.02; AE model: mean AUC 0.55±0.04). Feature importance analysis identified NLR (importance: 0.12), CRP (0.10), and Albumin (0.08) as top predictors for response, while similar markers dominated AE prediction. The MINIC3-Predictor web tool (https://minic3-predictor-f3fxplj5xpfbzwddntw2bu.streamlit.app) provides real-time individualized predictions with an intuitive interface and automatic risk stratification (low/intermediate/high). The GitHub repository (https://github.com/spy929/minic3-predictor) contains complete code, documentation, and example data.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions:\u003c/strong\u003e MINIC3-Predictor is an open-source, clinically oriented web tool that provides moderate predictive performance for treatment response and adverse event risk in MINIC3 immunotherapy. The tool's dual-task prediction capability, interactive interface, and open-source code facilitate clinical translation and external validation. This proof-of-concept study demonstrates the feasibility of developing deployable clinical decision support tools using simulated data, providing a framework that can be adapted for other immunotherapies.\u003c/p\u003e","manuscriptTitle":"A Python-Based Interactive Web Tool for Dual-Task Prediction of Treatment Response and Adverse Events in MINIC3 Immunotherapy: A Proof-of-Concept Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-10 16:49:58","doi":"10.21203/rs.3.rs-9267702/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-05-16T13:58:53+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-13T01:26:20+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"98455413730726142913533329794456271967","date":"2026-05-04T08:22:21+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"34959359220041392662444733471169323671","date":"2026-05-03T21:34:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"111621556173087786372533163763556430435","date":"2026-05-01T17:14:46+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"4099748428040661527367557588185220302","date":"2026-04-29T03:08:56+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-06T05:19:36+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-02T14:28:09+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-31T05:43:18+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-31T05:43:05+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-03-30T13:22:32+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"32bc3b19-52fc-405b-a77f-de02446781a5","owner":[],"postedDate":"April 10th, 2026","published":true,"recentEditorialEvents":[{"type":"editorInvitedReview","content":"","date":"2026-05-16T13:58:53+00:00","index":108,"fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-13T01:26:20+00:00","index":107,"fulltext":""},{"type":"reviewerAgreed","content":"98455413730726142913533329794456271967","date":"2026-05-04T08:22:21+00:00","index":106,"fulltext":""},{"type":"reviewerAgreed","content":"34959359220041392662444733471169323671","date":"2026-05-03T21:34:52+00:00","index":105,"fulltext":""},{"type":"reviewerAgreed","content":"111621556173087786372533163763556430435","date":"2026-05-01T17:14:46+00:00","index":102,"fulltext":""},{"type":"reviewerAgreed","content":"4099748428040661527367557588185220302","date":"2026-04-29T03:08:56+00:00","index":100,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":65894501,"name":"Health sciences/Biomarkers"},{"id":65894502,"name":"Biological sciences/Cancer"},{"id":65894503,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":65894504,"name":"Biological sciences/Immunology"},{"id":65894505,"name":"Health sciences/Oncology"}],"tags":[],"updatedAt":"2026-04-10T16:49:59+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-10 16:49:58","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9267702","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9267702","identity":"rs-9267702","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}