Post Doctoral Fellow, Quantitative Systems Pharmacology Pfizer San Diego, California, United States
Disclosure(s):
Mohammadali Alidoost: No financial relationships to disclose
Objectives: Oncology drug development faces significant challenges in linking tumor size dynamics with time-to-event endpoints such as progression-free survival (PFS) [1]. Combining mechanistic modeling with machine learning (ML) methods has proven to be a powerful tool for making clinical endpoint predictions [2]. In this work, we integrated mechanistic modeling with ML methods to predict patient survival probabilities, leveraging both quantitative systems pharmacology (QSP) models and clinical data.
Methods: We generated virtual populations (Vpops) using a modified Metropolis-Hastings algorithm to capture variability between subjects and provide probabilistic estimates of clinical outcomes. Tumor size time series were simulated using a QSP model, and progression/censoring times were assigned based on RECIST criteria. This approach was applied to data from a phase II study of an ALK inhibitor in non-small cell lung cancer (NSCLC) [1]. Using this data, we linked tumor dynamics to PFS by assigning progression/censoring times, crucial for understanding how tumor size impacts clinical endpoints. Moreover, various ML methods were trained to estimate the relationship between tumor size and PFS. Feature importance analysis was employed to assess the contribution of each feature to prediction accuracy. Finally, to evaluate model accuracy, the predictive performance of each method was assessed with metrics such as event-time concordance and Brier score.
Results: Preliminary results from our previous study [1] demonstrated the effectiveness of using QSP models to capture patient heterogeneity and predicted clinical outcomes. We applied ML methods to compare how linking tumor dynamics with PFS changed between the ML and the QSP approaches. By training various ML models on tumor size data, we estimated the relationship between tumor size and PFS and evaluated feature importance. The predictive performance of the ML methods was assessed using Kaplan-Meier curves, event-time concordance, and Brier score. This comparison determined the effectiveness of ML methods in discovering the relationship between tumor dynamics and PFS, providing insights into their potential application in real clinical datasets.
Conclusions: Combining mechanistic modeling with ML offers a powerful approach to evaluate survival prediction for oncology patients. This hybrid model can be used in clinical decision-making by providing predictions of patient outcomes, thereby optimizing treatment strategies. The proposed methodology can be extended to various cancer types and therapies, supporting the development of next-generation treatments and personalized medicine.
Citations: Citations: [1] Braniff, Nathan, et al. "An integrated quantitative systems pharmacology virtual population approach for calibration with oncology efficacy endpoints." CPT: Pharmacometrics & Systems Pharmacology (2025). [2] Butner, Joseph D., et al. "Hybridizing mechanistic modeling and deep learning for personalized survival prediction after immune checkpoint inhibitor immunotherapy." npj Systems Biology and Applications 10.1 (2024): 88.
