Model-informed approach for risk management of bleeding toxicities for bintrafusp alfa, a bifunctional fusion protein targeting TGF-β and PD-L1

Purpose Bintrafusp alfa (BA) is a bifunctional fusion protein composed of the extracellular domain of the transforming growth factor-β (TGF-β) receptor II fused to a human immunoglobulin G1 antibody blocking programmed death ligand 1 (PD-L1). The recommended phase 2 dose (RP2D) was selected based on phase 1 efficacy, safety, and pharmacokinetic (PK)–pharmacodynamic data, assuming continuous inhibition of PD-L1 and TGF-β is required. Here, we describe a model-informed dose modification approach for risk management of BA-associated bleeding adverse events (AEs). Methods The PK and AE data from studies NCT02517398, NCT02699515, NCT03840915, and NCT04246489 (n = 936) were used. Logistic regression analyses were conducted to evaluate potential relationships between bleeding AEs and BA time-averaged concentration (Cavg), derived using a population PK model. The percentage of patients with trough concentrations associated with PD-L1 or TGF-β inhibition across various dosing regimens was derived. Results The probability of bleeding AEs increased with increasing Cavg; 50% dose reduction was chosen based on the integration of modeling and clinical considerations. The resulting AE management guidance to investigators regarding temporary or permanent treatment discontinuation was further refined with recommendations on restarting at RP2D or at 50% dose, depending on the grade and type of bleeding (tumoral versus nontumoral) and investigator assessment of risk of additional bleeding. Conclusion A pragmatic model-informed approach for management of bleeding AEs was implemented in ongoing clinical trials of BA. This approach is expected to improve benefit-risk profile; however, its effectiveness will need to be evaluated based on safety data generated after implementation. Supplementary Information The online version contains supplementary material available at 10.1007/s00280-022-04468-6.

relationships between observed drug concentration and time or covariates of interest). Precision of the parameter estimates was assessed via the covariance matrix, derived from the Fisher information matrix obtained in the covariance step. Stability of the models was assessed on the basis of convergence criteria, goodness-of-fit plots, successful covariance steps, condition number (<1000), and correlation between parameters. Model selection was based on the difference in the objective function value (OFV) between 2 nested models for hierarchical models (p<0.05), decrease in unexplained variability, improvement in goodness-of-fit plots, and scientific and (patho)physiological plausibility of the model. To evaluate the predictive performance of the models, prediction-corrected visual predictive checks (pcVPCs) were performed. Between-subject variability in CL and V 1 was assumed to be log-normally distributed (exponential), and for maximal change in CL over time (Imax) it was assumed to be normally distributed (additive). Residual unexplained variability was described using a combined additive and proportional error model. Covariate analysis included the following steps: (1) refinement of pre-defined list of covariates of interest, including assessment of correlation among covariates and proportion of missing covariate values; (2) inclusion of the covariates from the refined list on CL and V1 simultaneously, resulting in the full model; (3) full model reduction, including backwards elimination via SCM functionality in PsN.

Supplemental Tables
Supplementary Table 1  Numbers are odds ratios (95% CIs). bleed_GI, gastrointestinal bleeding event of any grade; bleed_TEAE, treatment-related bleeding event of any grade; BTC, biliary track cancer; Cavg,SD, average concentration over the dosing interval Note: data were analyzed by logistic regression, using the following equation: where is the probability of adverse event , 0 is the probability of absent any predictors, 1... is a vector of predictors, and 1... is a vector of coefficients describing the effects of 1... on .

Supplemental Figures
Supplemental Figure 1. Goodness-of-fit plots for the final population pharmacokinetic model.
Purple points are individual data points; yellow lines are locally weighted smoothing curves.

Supplemental Figure 2. Prediction-corrected visual predictive checks for the full pharmacokinetic model for bintrafusp alfa.
Points are individual observations; lines represent medians, 5th and 95th percentiles of the binned observed data; and shaded areas represent 95% prediction intervals around the median and the 5th and 95th percentiles based on the simulations. The X-axis shows time since previous dose for every observation.

NCT04246489.
The blue line and shaded area represent model-predicted AE probability (median and 95% CI); pink circles represent observed AE incidence by quartiles of exposure and are placed at the 12.5th, 37.5th, 62.5th, and 87.5th percentiles of the exposure distribution (i.e. the median for each exposure quartile); the error bars represent 95% CIs; pink dotted lines represent boundaries of exposure quartiles; and purple dots indicate individual patient data. The odds ratio was 1.59 (95% CI 1.22-2.11) per 100 μg/mL for Bleed_TEAE and 1.69 (95% CI 1.2-2.42) per 100 μg/mL for Bleed_GI. AE, adverse event; Cavg,SD, average bintrafusp alfa concentration over the dosing interval; bleed_TEAE, treatment-related bleeding event of any grade; bleed_GI, gastrointestinal bleeding events of any grade.