Sofia Stathopoulos1, Giuseppe Pasculli1, Pauline Bambury1, Fianne Sips2, Jane Knöchel1, Maud Beneton1, Paolo Messina1, Daniel Röshammar1, Marco Virgolin2
1InSilicoTrials Technologies S.p.A., 2InSilicoTrials Technologies B.V.
Objectives Machine learning (ML)-based disease progression modeling offers an advanced framework for optimizing drug development and clinical trial design in oncology (1). This work presents a multi-model integration approach leveraging pharmacokinetic (PK) and surrogate biomarker pharmacodynamic (PD) modeling based on early phase study results, and ML-driven survival predictions in external patient populations to enable in-silico clinical trial predictions of late-stage studies. By linking drug-specific models with generalizable disease progression models, this framework supports optimal dose selection, improves patient inclusion/exclusion criteria, and may increase the overall probability of success. Methods A modular modeling approach was implemented to integrate distinct but interdependent models describing drug exposure (based on Phase 1-2 data), biomarker response (based on Phase 2 data), and survival probability (based on literature data of a larger prostate cancer population). The orchestration between these models was designed on the InSilicoTrials’ platform, whose workflow builder enables to combine data and models built with different environments (NONMEM, R, Matlab, Python, etc.) via a drag-and-drop interface. The resulting workflow consisted of: 1. A dataset of prostate cancer patient survival, based on the work of Peng et al. 2023 (2). 2. A PK developed in NONMEM, describing drug disposition under various dosing regimens using compartmental differential equations. Drug exposure was modeled as a function of clearance, volume of distribution, and absorption rates, with inter-individual variability (IIV) incorporated via a log-normal distribution. 3. A PKPD model implemented in R, describing prostate-specific antigen (PSA) dynamics as a function of drug concentration, using an indirect response model. 4. A ML-based model, developed in Python, to predict survival probability (survival model) given the patient’s covariate and the patient’s PSA response calculated with the PK and PKPD models. The model was implemented as a gradient boosting survival model using scikit-survival (3) and its hyperparameters were optimized using Optuna (4). Furthermore, SHAP was used to explain how the inputs of the ML model contribute to its predictions (5). The in-silico trial simulation was configured to explore multiple scenarios, varying patient demographics (age, weight, PSA levels), trial designs, sample sizes, and dosing regimens. Illustrative dashboards were set up to be used by the wider project team, showing key figures and tables of drug exposure, PSA response and survival outcome, offering practical guidance for informed decision-making. Results The PK, PKPD and ML survival model were successfully connected using the workflow builder of the platform. Simulations demonstrated a clear dose-response relationship, where increased drug exposure resulted in greater PSA suppression, leading to higher predicted survival probabilities when plugged into the ML model. Explaining the ML model’s predictions using SHAP identified key predictive features of survival, highlighting not only PSA levels but also age and other patient characteristics (e.g., marital status and race) as influential covariates. Sensitivity analyses quantified the relative contribution of each variable to survival, while interactive visualization tools allowed exploration of treatment effects. Clinical trial simulations showed the impact of baseline PSA values (as inclusion criteria) and dose on the expected Phase 3 probability of success. Conclusions This study established a generalizable framework for integrating drug specific PKPD models for early phase biomarkers with general AI-driven disease-progression modeling for predicting clinical outcomes, enhancing predictive power in oncology clinical trials. The ability to interchange drug-specific models while maintaining a consistent disease progression model provides a scalable solution for clinical trial design and dose optimization across drug programs. The InSilicoTrials’ platform facilitates such integrations of models and data (across internal and external information sources), enabling explorations of alternative trial designs to optimize the probability of late phase study success. Future applications will extend this methodology to other indications, integrating multi-omics datasets and real-world evidence as well as refine precision medicine strategies. Expanding this hybrid mechanistic-ML framework across multiple indications will support adaptive trial designs, biomarker-driven patient selection, and internal decision-making in oncology drug development.
1. Liu Q, Huang R, Hsieh J, Zhu H, Tiwari M, Liu G, et al. Landscape Analysis of the Application of Artificial Intelligence and Machine Learning in Regulatory Submissions for Drug Development From 2016 to 2021. Clin Pharmacol Ther. 2023 Apr;113(4):771–4. 2. Peng ZH, Tian JH, Chen BH, Zhou HB, Bi H, He MX, et al. Development of machine learning prognostic models for overall survival of prostate cancer patients with lymph node-positive. Sci Rep. 2023 Oct 27;13(1):18424. 3. S. Pölsterl, “scikit-survival: A Library for Time-to-Event Analysis Built on Top of scikit-learn,” Journal of Machine Learning Research, vol. 21, no. 212, pp. 1–6, 2020. 4. Akiba T, Sano S, Yanase T, Ohta T, Koyama M. Optuna: A next-generation hyperparameter optimization framework. InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining 2019 Jul 25 (pp. 2623-2631). 5. Lundberg SM, Lee SI. A unified approach to interpreting model predictions. Advances in neural information processing systems. 2017;30.
Reference: PAGE 33 (2025) Abstr 11685 [www.page-meeting.org/?abstract=11685]
Poster: Methodology – AI/Machine Learning