Predicting myelosuppression from phase I data: Which model should we use?
Núria Buil-Bruna, Alienor Berges, S. Y. Amy Cheung, Martin Johnson, Helen Tomkinson
Quantitative Clinical Pharmacology, AstraZeneca, Cambridge, UK
Objectives: Ever since the semi-mechanistic model to describe drug-induced myelosuppression was published by Friberg et al. [1] (M1), several generalisations have been developed. For example, Mangas-Sanjuan et al. [2], incorporated cell cycle dynamics to improve the model’s ability to predict toxicity under alternative schedules (M2), Bender et al. [3], incorporated an additional cell kill effect to describe the downward drift in cell-count profiles seen in some patients (M3), and Quartino et al. [4], improved the model’s ability to accurately describe nadir cell counts (M4). Here we aim to assess these models for their ability to predict tolerable dose and schedules based on early phase I data.
Methods: A simulation and estimation procedure was performed. PK data was generated for a hypothetical small molecule. For each extended model scenario, potency parameters were scaled to ensure the 4th dose escalation predicted 33% grade 3/4 (G3/4) events. Each model was used to simulate 300 small datasets mimicking cycle 1 (21 days) of a phase I dose escalation study (weekly schedule: 3 days on/4 days off). Datasets were reestimated with all four models which were then used to simulate long-term toxicity (up to 8 cycles, 189 days) of a large population receiving different drug schedules (continuous and a variety of intermittent schedules). Bias (mean error, ME) was obtained by comparing incidence of G3/4 events to the true simulation model.
Results: All models successfully converged when reestimating their training datasets. M3 had more difficulties in numerical estimation across all scenarios (average successful rate 78.5% compared to 100% of other models). This is probably due to cycle 1 data being insufficient to estimate cumulative toxicity.
Assuming M4’s scenario represents realistic nadirs, all models performed broadly similarly on estimating cycle 1 nadirs. The greatest bias was seen in M1 (ME M1= -8.7%, M2=-5.4%, M3=+1.8% points), which underpredicted %G3/4.
Assuming M2’s scenario represent a realistic depiction of the impact of different schedules on toxicity, all models performed similarly with an average ME between 5% and 7% points.
Assuming M3’s scenario represents a realistic long-term cumulative toxicity scenario, all models underpredicted %G3/4 after 8 cycles (M1 ME=-18.7%, M2 ME = -12.7%, M3 ME = -8.9% points).
Conclusions: All three extended models have advantages over the original Friberg model. Choice of extended model in a phase I setting depends primarily on whether cumulative toxicity is observed/expected.
References:
[1] Friberg et al, Journal of Clinical Oncology (2002)
[2] Mangas-Sanjuan, et al. J Pharmacol Exp Ther (2015)
[3] Bender, et al. Cancer chemotherapy and pharmacology (2012)
[4] Quartino, et al. Investigational new drugs (2012)