The conduct of most first-in-human oncology drug trials is conceptually incoherent, and unethical.


April 14, 2019

This is a lay explainer for my bioRxiv preprint Precautionary Coherence Unravels Dose Escalation Designs.

What is it about?

When a new cancer drug is tried for the first time in cancer patients, we lack answers even to very basic questions about the drug: What range of doses can patients tolerate? What bad effects does the drug cause? These are the types of questions that we try to answer in ‘Phase 1’ dose-finding trials.

These dose-finding trials are most commonly run using a group-wise ‘dose-escalation’ design. Such trials enroll patients sequentially, and dose them in small groups called ‘cohorts’—usually of 3 patients each. The first cohort gets a low dose of the drug that is expected to be quite safe, and are monitored for some period of time to see if they experience toxicities. If not, then the study dose is ‘escalated’, which means that a new cohort is enrolled to receive a higher dose. This process is repeated usually until moderate (but not severe) toxicities are seen. The dose where this happens is then thought to strike a reasonable balance between being too low (which might not help anyone) and too high (which causes too much harm).

This way of running a dose-finding trial treats the enrolled patients like manufactured products in destructive testing: Take 3 units of the product, and drop them from 3 ft onto a concrete floor. If they suffer minimal damage, take 3 fresh units and drop them from 4 ft. And so on, until you find the greatest height where the damage stays strictly cosmetic, and the product still works. Obviously, this manner of product testing depends on all the manufactured units being identical. (Imagine how senseless such testing would be, if it were done on a grab-bag of mobile phones of many different makes and models built to different standards of ruggedness.)

But the patients enrolled in a dose-finding study are of course not identical ‘units’! In light of the differences between patients, think about that moment in a dose-escalation trial when the second cohort are enrolled at an escalated dose. At this point in time, you have a cohort of (say) 3 patients already in the study, who have been tolerating a low dose of the drug. Wouldn’t they be the best candidates to try a new, higher dose of the drug—a dose that has never been tried in humans before? I think so. Nevertheless, the common practice is to enroll 3 new patients into the escalation cohort. This practice violates a principle that I introduce in this paper under the name ‘precautionary coherence’. The reason for this name is simple: when trying a higher dose for the first time, caution demands this be done preferentially in willing patients who have already tolerated a lower dose, rather than drug-naive patients.

The core point of this paper is that designing a dose-finding trial in accordance with ‘precautionary coherence’ (PC) requires abandoning the practice of dose-escalation described above. Rather, a ‘PC’ dose-finding trial must use a dose-titration design, where each patient individually starts at a low dose and gradually increases to whatever dose is best for him or her. (To help me communicate this core point succinctly, I have chosen to build upon a distinction between ‘escalation’ and ‘titration’ that has sometimes appeared in the literature. But I should point out that if you read other articles on this topic, you will mainly find “intra-patient dose escalation” used to designate what I term ‘titration’. For example, many Phase 1 study reports in the literature will say something like, “intra-patient dose escalation was not allowed.”)

In order to lay out the paper’s main argument, I take the most widely used and familiar dose-finding design in oncology—the so-called ‘3+3’ design—and develop a ‘PC’ version of it. This requires me (1) to set forth rules for running a ‘3+3/PC’ trial and then (2) to develop a way to visualize the events that happen during such a trial, and what conclusions we can draw from those events. I also (3) run computer simulations that show how a 3+3/PC trial can individualize dosing to give better outcomes for patients.

Here is an animation of the paper’s Figure 1, showing a 3+3/PC design in action:

A 3+3/PC trial in action. Note how the ‘dose-survival curve’ on the right accumulates the experience of trial participants titrating through the ‘OX’ plot on the left.

Why is it important?

This work caps a yearlong effort to criticize the practice of ‘1-size-fits-all’ dosing in oncology, and to promote the opposite principle: dose individualization. Whereas some of my earlier work in 2017 was theoretical, speculative and even philosophical, this paper is intensely pragmatic. Its main argument is free from highly technical or abstract ideas. I expect that most oncology trialists can read this paper, understand it, and implement a 3+3/PC study of their own.


I would like to thank Ying Yuan of MD Anderson, for first bringing to my attention the applicability of coherence principles in dose-finding methodology. His September 25 Short Course “Bayesian Adaptive Designs for Immunotherapy and Drug Combination Trials” at the 2017 ASA Biopharmaceutical Section Regulatory-Industry Statistics Workshop greatly stimulated my thinking on this topic. (This acknowledgement is not meant to suggest that Dr. Yuan endorses the argument I make in this paper.)