(M-036) Mechanistic Insights by Combining Quantitative Systems Pharmacology (QSP) and Machine Learning: An Inflammatory Bowel Disease Case Study

Yamato Sano – Clinical Pharmacology & Bioanalytics, Development Japan – Pfizer R&D Japan, Japan; Liliana Jelu – Translational Clinical Sciences, Pfizer Research & Development – Pfizer, United States; Bryce Johnson – Inflammation and Immunology, Pfizer Research & Development – Pfizer, United States; Valerie Clerin – Inflammation and Immunology, Pfizer Research & Development – Pfizer, United States; Richard Allen – Translational Clinical Sciences, Pfizer Research & Development – Pfizer, United States; Nessy Tania – Translational Clinical Sciences, Pfizer Research & Development – Pfizer, United States

Objectives: Quantitative Systems Pharmacology (QSP) modeling has been valuable in understanding disease pathophysiology, exploring therapeutic targets, and optimizing treatment strategies. However, the complexity of a model often makes complete interpretation of simulation results difficult. Here we provide a framework for gaining deeper mechanistic understanding from a QSP model output by applying machine learning (ML) to the model outputs. As a case study, we demonstrate the results by considering virtual population simulations of a QSP model of Inflammatory Bowel Disease (IBD) which can predict fecal calprotectin (FCP) response and endoscopic improvement.

Methods: A QSP model for IBD and a virtual population for ulcerative colitis (UC) were previously developed [1,2,3] and calibrated to published clinical data of FCP response and endoscopy scores [4,5]. The model was used to predict the effect of anti-p40, anti-TL1A, and combination therapy [3]. From the virtual population simulations, 38 model variables (immune cells and cytokines at baseline and end of induction) were used as input features for the subsequent ML models. Next, Random Forest, Support Vector Machine, Neural Network, and XGBoost were selected as candidate ML-based responder classifier models. Responders were defined as virtual patients who achieved endoscopic scores of 0-1 based on the predicted FCP. Predictive performance for each model was evaluated by AUC for the Receiver Operating Characteristic (ROC) curve, as well as sensitivity and specificity (80% for the training set and 20% for the test set). Finally, the importance of each feature was calculated based on the selected model.

Results: The random forest model showed the highest AUC for the ROC curve among the four candidate classifiers (AUC: 0.87-0.95 across all therapeutic scenarios) with reasonable sensitivity and specificity. Using the random forest model, we found that tissue neutrophils, CRP, epithelial damage, IL-6, IL-8, and TNF-α at the end of induction were identified as strong predictors for both therapies. For anti-p40, TL1A and Th1 were identified as key features differentiating between responders and non-responders: specifically, responders tended to have higher Th1 at baseline while non-responders have elevated TL1A at baseline and end of induction. Key features identified for anti-TL1A were IL-12, IL-17, and GM-CSF, all of which tended to be elevated in non-responders at baseline and end of induction. Our analysis revealed that these therapies have both some overlapping and distinct immune pathway features that correlate with treatment response.

Conclusions: We presented a framework of combining QSP and ML approaches to gain additional mechanistic insights that might not be immediately apparent due to the complexity of a QSP model. Application of this approach on model parameters could also aid in determining sensitive parameters and enrich virtual population generation for different clinical responses. In the future, this methodology can be applied to additional QSP models to investigate key features that differentiate between responders and non-responders or other clinical outcomes with complex underlying mechanisms.

Citations: 1. Rogers KV, et al. (2021) Clinical and Translational Science, 14(1):239-48.
2. Rogers KV, et al. (2021) Clinical and Translational Science, 14(1):249-59.
3. Fang Y, et al. (2024) 15th American Conference of Pharmacometrics.
4. Sands BE, et al. (2019) New England Journal of Medicine, 381(13):1201-14.
5. Danese S, et al. (2022) Clinical Gastroenterology and Hepatology, 1;20(12):2858-67.

Keywords: QSP, Machine Learning, IBD