Lola FÉRÉ GELY 1,2, Karl BRENDEL 1, Déborah HIRT 2
1 Ipsen Innovation, Pharmacometrics (Paris, FRANCE), 2 Université Paris Cité (Paris, FRANCE)
Introduction: In oncology, high rates of therapeutic failure (particularly in late stages of clinical development) highlight the need to improve early understanding of tumor response mechanisms. To better identify patients likely to benefit from treatment, Mouse Clinical Trials (MCTs), including Patient-Derived Xenografts (PDXs), offer a relevant preclinical framework to evaluate anticancer efficacy while capturing inter- and intra-tumor variability. Usually, a large number of PDXs from diverse tumor types are included and treated at a single dose level. MCTs generate complex and high variability data, combining longitudinal tumor volume measures with genetics molecular data. The analysis of such data poses major methodological challenges, including response modeling, discrimination of different types of responder profiles, and identification of predictive genetic characteristics. Although several methods are routinely used to classify responders and non-responders, no clear consensus has emerged1,2. Several existing methods demonstrated correct identification of different types of responses but relies on (1) statistical clustering (K-means)3 or (2) empirical slope-based approach (piecewise model)4 combined with a Latent Class Mixed Models (LCMM) algorithm (gathering similar trajectories based on longitudinal data).
Objective: In this work, we attempt to develop a methodological framework combining a more physiological model (non-linear model) with the LCMM algorithm. We aim to create a rapid identification of tumor responder profiles in a MCT, characterize tumor response heterogeneity and support therapeutic decision-making with post-classification genetic analysis.
Methods: A combined strategy was implemented in which tumor growth dynamics were first modeled using a mechanistic nonlinear mixed-effects model5. Individual parameter estimates were then used as inputs for LCMM classification algorithm, allowing a more physiologically informed identification of response profiles, without prior assumptions on class membership. Models with multiple latent classes were tested and selected based on classification criteria and parsimony. Forty-seven PDX tumor growth longitudinal trajectories were analyzed, representing two indications (melanoma and colorectal cancer), with corresponding mutation status (BRAFV600E or BRAFnonV600E) collected for each PDX1. The performance and interpretability of the proposed method was assessed by comparison with established tumor growth inhibition (TGI) metrics, computed from model-based simulations. TGI is calculated at the final time of the 28-days follow-up (lastTGI) representing the relative difference between treated and control tumor volumes from baseline. Genetic characteristics represented in each class were analyzed and compared with literature.
Results: This LCMM-physiological approach led to a 4-classes model with high (7/47), moderate response (5/47), low (10/47) and no (25/47) response. It allows a correct discrimination of response profiles, characterized by 4 different distributions of their parameters (kge, growth rate/constant or kkill, shrinkage rate/constant). The genetic features associated with these classes were consistent with existing literature knowledge, with a higher proportion of BRAFV600E mutations and melanoma pharmacological models in the high-response classes than in the low-response classes (71% VS 16% for BRAFV600E mutation and 100% VS 52% for the melanoma indication, in the high- and low-response classes, respectively). In addition, the discriminatory power of this approach was assessed by comparison with established TGI criteria, calculated on observed data at the last time of follow-up (lastTGI). TGI-based response distribution across classes confirmed a clear gradient of treatment effect : majority of regression profiles in first class, second class mostly composed of high inhibition cases, third class characterized by mixed high and moderate inhibition, and class D mainly associated with moderate and low inhibition. This strategy provides a flexible framework for potential future applications using more complex tumor models (incorporating delays or resistance mechanisms), pending confirmation across additional case studies. In contrast, alternative methods such as K-means clustering, latent covariate identification, and within-/between-subject mixture modeling were tested and compared but did not yield more conclusive or biologically informative classifications.
Conclusions: This framework combining nonlinear tumor growth modeling and latent class algorithm based on physiologically meaningful individual parameters led to a consistent stratification of PDXs, while providing enhanced biological interpretability. This approach allows the integration of more complex tumor dynamics such as delayed treatment effects or resistance mechanisms. Overall, the proposed methodology represents a promising tool to identify target populations at later stages of drug development, link genetic characteristics to treatment response, and increase mechanistic understanding of anticancer efficacy.
References:
1. Gao, H. et al. High-throughput screening using patient-derived tumor xenografts to predict clinical trial drug response. Nat Med 21, 1318–1325 (2015).
2. Guo, S., Jiang, X., Mao, B. & Li, Q.-X. The design, analysis and application of mouse clinical trials in oncology drug development. BMC Cancer 19, 718 (2019).
3. Hartigan, J. A. & Wong, M. A. Algorithm AS 136: A K-Means Clustering Algorithm. Journal of the Royal Statistical Society. Series C (Applied Statistics) 28, 100–108 (1979).
4. Savel, H., Meyer-Losic, F., Proust-Lima, C. & Richert, L. Statistical classification of treatment responses in mouse clinical trials for stratified medicine in oncology drug discovery. Sci Rep 14, 934 (2024).
5. Simeoni, M. et al. Predictive Pharmacokinetic-Pharmacodynamic Modeling of Tumor Growth Kinetics in Xenograft
Reference: PAGE 34 (2026) Abstr 12194 [www.page-meeting.org/?abstract=12194]
Poster: Drug/Disease Modelling - Oncology