Adam Nasim1,2, Dr Imran Nasim2,4, Mihály Leiwolf3
Surrey University, University of Poitiers, IBM
Introduction/Objectives: Tumour xenograft models are widely regarded as the gold standard for evaluating the preclinical efficacy of oncology drug candidates, enabling direct comparisons of therapeutic efficacy and facilitating the prediction of effective human dose ranges through modeling and simulation. However, these experiments are time-consuming, costly, and lie on the critical path of drug development. In this study, ML statistical forecasting methods are applied to pharmacometrics for the first time to extrapolate data from patient-derived xenograft experiments, with the goal of significantly reducing the time required to obtain meaningful insights. By comparing the performance of AI-driven approaches against classical pharmacokinetic-pharmacodynamic (PKPD) models, we aim to assess which method provides more accurate predictions and whether AI can offer additional value in accelerating drug discovery processes. Methods: 8 preclinical tumor growth models coupled with a cell loss function and incorporating drug effect were tested to describe PDX data [1,2] for three molecules (ribociclib, binimetinib, and buparlisib) used to treat six cancer types. The best suited model was selected based BIC, goodness of fit plots, the precision of parameter estimates, and the accuracy of predictions. For each molecule, the parameters were estimated on the complete dataset (~ 300 days) and different early time windows ranging from 20 to 50 days. The parameters re-estimated on early data were used to simulate for the complete time series and compared with the model fit for the complete dataset. Models such as exponential smoothing (ETS), complex exponential smoothing (CES) [3], and Theta [4] try to capture trend and seasonality to make long-term time series predictions and adjust these predictions to the short term by modeling noise in the data. These statistical ML forecast models have been trained on early time windows (20-50 days) to make forecasts until for the remaining part of the complete time series (300 days). The results were compared to the PKPD extrapolations and PKPD parameters were re-estimated on the ML forecasts and compared with the estimates on the original data. Results: PKPD models adequately describe the data for the first 100 days of rich data, with a decreasing performance for later sparse data. The quality of predictions stays similar for different estimation time windows, but for ribociclib and binimetinib, parameter estimates vary based on the training time length. Re-estimated parameters on the ML forecasts for binimetinib and buparlisib, show important differences between the original estimates as well as between the estimates for the forecast data coming from different ML models with a decreased inter-individual variability is compensated by a high residual unexplained variability. ML forecast methods are more sensitive to training window size. For binimetinib, they provide more accurate predictions than PKPD, for ribociclib, they have similar performance, and for buparlisib, they underperform the classical pharmacometric analysis. ML models provide an adequate overall description of tumour dynamics, however, fail to provide precise individual forecasts. Due to their ability to describe overall tendencies, they do not capture well smaller variations in tumour volume. The automatic forecast models available in StatsForecast offer little control over hyperparameters. More complex statistical forecast models with an in-depth evaluation of the hyperparameter space may yield a better prediction of tumour dynamics. Both PKPD and ML models could be improved by the incorporation of covariates and static features. Conclusions: If suitable AI models are identified they have the potential to reduce the timelines of PDX experiments by extrapolation, however further evaluation is needed and more sophisticated models tested prior to their application in model-informed drug development.
(1) Gao, H., Korn, J.M., Ferretti, S., Monahan, J.E., Wang, Y., Singh, M., Zhang, C., et al.: High-throughput screening using patient-derived tumor xenografts to predict clinical trial drug response. Nat. Med. (2015) (2) Niessner, H., H¨usch, A., Kosnopfel, C., Meinhardt, M., Westphal, D., Meier, F., Schilling, B., Sinnberg, T.: Exploring the in vitro and in vivo therapeutic potential of BRAF and MEK inhibitor combination in NRAS-mutated melanoma. Cancers (Basel) (2023) (3) Svetunkov, I., Kourentzes, N., Ord, J.K.: Complex exponential smoothing. Nav. Res. Logist. (2022) (4) Petropoulos, F., Nikolopoulos, K.: Optimizing Theta Model for Monthly Data. In: Proceedings of the 5th International Conference on Agents and Artificial Intelligence – Volume 1: ICAART. (2013)
Reference: PAGE 33 (2025) Abstr 11586 [www.page-meeting.org/?abstract=11586]
Poster: Methodology – AI/Machine Learning