Associate Director Bristol Myers Squibb, United States
Disclosure(s):
Mayu Osawa, PhD: No financial relationships to disclose
Background/
Objectives: Tumor growth dynamic modeling has been widely used to predict late outcomes and support decision-making in solid tumors [1]. In hematology area, resent research shows that M-protein may serve as a surrogate biomarker linked to progression-free survival (PFS) for patients with RRMM [2]. This work aims to establish a model platform linking M-protein to PFS for patients with RRMM across multiple drugs, treatment regiments and studies, and then assess the predictive value of the model platform for an independent clinical trial.
Methods: A total 2434 subjects with RRMM from 6 Phase II/III studies (12 treatment combinations of 5 individual drugs) were included to develop the model. Sequential modeling and joint modeling approaches [3] were applied to characterize the link between M-protein to PFS. Stein model, Claret model, Wang model, Bonate model, and two-population model [4] were explored in the development of structure model for M-protein. PFS was modeled by a parametric hazard model, testing if best fit is achieved by exponential, Weibull, Gompertz and log-logistic distributions. Current M-protein (normal or log-transformed), first derivative (slope) of M-protein, AUC of M-protein from baseline to current, current M-protein normalized to baseline (log-transformed) and current M-protein percent change from baseline were tested as a link function. The effect of baseline covariates on hazard function were assessed by univariate screening and stepwise forward selection (p < 0.01). The developed model was then used for an external validation to evaluate predictive performance for a study (n=682) not included in the model development, using all data or truncated data (at 1, 3, 6 months) with baseline covariates as input.
Results: Claret M-protein model provided a good fit to describe the M-protein data across the different studies. The final PFS model was a log-logistic hazard model, with baseline albumin, baseline M-protein, baseline beta-2 microglobulin, time since disease diagnosis, myeloma type (IGG or non-IGG), number of prior myeloma treatments, MM stage (I, II, III or missing), ECOG performance status, treatment group (1, 2, or 3 drugs), on the hazard function. The best link function was log-transformed current M-protein normalized to baseline. The model showed a good performance across the drugs and treatments in both sequential model and joint model. The predicted PFS for the study in external validation was in agreement with the observed PFS, and both full and truncated M protein data as input predicted PFS beyond 5 years.
Conclusions: The model platform described in this work demonstrated promise in its predictive performance across drugs and treatments. The final model included covariates that appear to make sense in the context of RRMM. As such, this modeling platform can be used as predictive tool to simulate long term disease progression and guide decisions for future trials.
Citations: [1] Bruno R et al. Clin Cancer Res. 2020;26(8);1787-1795 [2] Cheng Y et al. eJHaem. 2022; 3(3);815-827, [3] Desmée S et al. AAPS J. 2015;17:691-699 [4] LIXOFT TGI library. https://mlxtran.lixoft.com/model-libraries/tgi-library/
