A Letter to the Editor of Clinical Cancer Research, desk-rejected Oct 19, 2018:
To the Editor: Pursuing a “realistic” depiction of the impact of dose-finding methods upon oncology drug development, Conaway and Petroni¹ conduct an extraordinarily intricate simulation study comprising the entire sequence of regulatory trials. This end-to-end simulation lets the authors forego dose-finding performance measures that “[do] not resonate with clinicians,”¹ and instead link dose-finding design choice to a concrete event: regulatory approval. Yet this outcome seems to appeal more directly to investors concerned with marketing pharmaceutical innovation, than to clinicians concerned with caring for individual patients. Indeed, the realism clinicians desire must be sought in the one detail conspicuously absent from this otherwise comprehensive simulation study.
In a simulation so intricate, it seems remarkable that dose-limiting toxicity (DLT) probabilities are specified separately at distinct dose levels (see Table 1)¹, rather than obtained as quantiles of an individualized maximum tolerated dose (MTDᵢ) distribution². Even as a purely technical consideration, this would economize on arbitrarily chosen parameters in an already overwrought simulation. More importantly, it would require consideration of dose individualization designs. Model-based one-size-fits-all dose-finding methods have failed to “resonate with clinicians” precisely because they have failed to distinguish themselves clinically from their still-dominant up-and-down algorithmic competitors like 3+3. Both types of design depend equally on a simplistic DLT conception² arising from WWII-era explosives research³. Three decades of frustration should teach model-based advocates that model-basedness alone is not enough.
To “resonate with clinicians,” dose-finding designs must acknowledge clinically meaningful problems such as optimal-dose heterogeneity, and guide clinicians toward solutions. Concomitantly, the clinically offensive analytical convenience of ‘target toxicity rate’ must be abandoned. Biostatistics has at its disposal all necessary machinery for pursuing a dose individualization agenda, if only it would embrace one neglected topic: _recursive filtering_⁴.
But progress toward dose individualization may now circumvent Biostatistics altogether. A simple, clinically meaningful ‘precautionary coherence’ (PC) principle yields a ‘3+3/PC’ design that outperforms even theoretically ideal one-size-fits-all dose finding⁵. While biostatisticians perseverate on promoting model-based one-size-fits-all dose-finding designs, clinician-investigators (and pharmaceutical investors) will have recourse to this simple alternative, having superior ethical and performance characteristics.
Oncology journals could encourage development of dose individualization methods, simply by requiring such simulation studies to obtain DLT probabilities from explicitly modeled optimal-dose distributions. This is an extremely modest proposal. In the study by Conaway and Petroni¹, this would have required changing just one line of simulation code, discussing optimal-dose heterogeneity in the text, and explaining its neglect by the dose-finding methods considered.
References
1. Conaway, M. R. & Petroni, G. R. The impact of early phase trial design in the drug development process. Clinical Cancer Research (2018). doi:10.1158/1078-0432.CCR-18-0203
2. Norris, D. C. One-size-fits-all dosing in oncology wastes money, innovation and lives. Drug Discovery Today 23, 4–6 (2018). doi:10.1016/j.drudis.2017.11.008
3. Dixon, W. J. & Mood, A. M. A Method for Obtaining and Analyzing Sensitivity Data. Journal of the American Statistical Association 43, 109–126 (1948). doi:10.1080/01621459.1948.10483254
4. Norris, D. C. Dose Titration Algorithm Tuning (DTAT) should supersede ‘the’ Maximum Tolerated Dose (MTD) in oncology dose-finding trials. F1000Research 6, 112 (2017). doi:10.12688/f1000research.10624.3
5. Norris, D. C. Precautionary Coherence Unravels Dose Escalation Designs. bioRxiv 240846 (2017). doi:10.1101/240846