Despite advances in Bayesian adaptive designs ^{1, 2} and model-based dose-finding ³ , oncology dose-finding studies remain conceptually in the thrall of ‘the’ maximum tolerated dose (MTD). This fallacious concept stands opposed to the long-recognized heterogeneity of cancer patients’ pharmacokinetics and pharmacodynamics (PK/PD), and to the diversity of their individual values and goals of care. Under this conceptual yoke, these dose-finding studies constitute a significant choke-point in drug development, where a severe discount may be applied to the potential value in new molecules through the hobbling of subsequent ‘efficacy’ trials by inadequate individual-level dosing ⁴ .

Strangely, Bayesian innovation in dose-finding studies has proceeded apace without issuing a meaningful challenge to the inherently frequentist conception of an MTD as determined by whole-cohort frequencies of dose-limiting toxicities (DLTs). Thus, even as Bayesianism has made progress toward the ethical imperative of efficient use of data ⁵ in such studies, it has neglected to confront the distinct ethical dimension of individualism ⁶ . This seems a great irony, as the dynamic learning model of Bayesianism is equally suited, and indeed equally essential, to solving the latter problem.

This paper demonstrates individualized dose-finding in a simulated Phase I study of a cytotoxic chemotherapy drug for which neutropenia constitutes the critical dose-limiting toxicity. Importantly, myelosuppression is interpreted also as a monotone index of therapeutic efficacy, without the added complication of a dose-response ‘plateau’ ⁷ such as postulated for molecularly targeted agents (MTAs). This creates a problem setting where simple heuristics apply, simplifying the demonstration undertaken here. The aim of this exercise is to elaborate a concrete setting in which ‘dose-finding study’ may readily be seen as a misnomer. Under the view advanced here, early-phase studies of this kind should be conceived as dose titration algorithm tuning (DTAT) studies.

The idea that ‘dose finding studies’ should yield dose titration algorithms (DTAs) is not new. More than a quarter-century ago, Sheiner and colleagues ⁸ advocated a learn-as-you-go concept for “dose-ranging studies”, addressing concerns about “parallel-dose designs” that are not far removed from the motivations for the present paper. As in the advocacy of Sheiner et al., parametric models play an important role in this paper, although in keeping with a spirit of pragmatism I relax this dependence to some extent by means of a semiparametric dynamic on-line learning heuristic.

A hypothetical cytotoxic drug is considered, modeled notionally after docetaxel, to be infused in multiple 3-week cycles. The pharmacokinetics are taken to follow a 2-compartment model with parameters as estimated for docetaxel in a recent population pharmacokinetic study ⁹ . Chemotherapy-induced neutropenia (CIN) is taken to follow a myelosuppression model due to Friberg et al. ¹⁰ . Together, these models form a population pharmacokinetic/pharmacodynamic (PK/PD) model within which DTAs may be simulated and tuned for optimality. For simulation purposes, and anticipating the future value of ready access to a variety of inference procedures in follow-on work, this PK/PD model is implemented in R package pomp ¹¹ . R version 3.3.2 was used ¹² .

Basic behaviors of the models are illustrated by simulation graphics generated for 25 individuals randomly generated from the population PK/PD model. Properties with specific relevance to absolute neutrophil count nadir ( ANC _nadir )-targeted dose titration are then investigated, with an eye to demonstrating the predictability of nadir timing. In particular, an approximate linearization of neutrophil nadir level and timing is demonstrated, achieved through suitable power-law transformation of infusion doses and logarithmic transformation of neutrophil concentration. Within this transformed parameter space, a simple recursive DTA is defined on the basic heuristic of the Newton-Raphson method for root-finding. For simplicity, monitoring of CIN is not modeled endogenously to this algorithm, but is treated as exogenous such that nadir timing and level are known precisely. A ‘DTAT’ study is simulated and visualized for 25 patients, with the tuning parameters of the recursive titration algorithm held fixed. The visualization supports a discussion of how these parameters might be tuned over the course of a Phase I study. All simulations and figures in this paper were generated by a single R script, archived on OSF ¹³ .

Recursive nonlinear filtering, as implemented in the extended Kalman filter (EKF) or its more modern adaptations ¹⁴ , constitutes a powerful conceptual framework for approaching model-based dose titration ¹⁵ . Indeed, the ‘tuning’ in ‘DTAT’ was itself suggested by the practice of tuning a Kalman filter ¹⁶ for optimal performance.

For present purposes, however, it suffices to implement a model-free recursive titration algorithm built on the Newton-Raphson method, with a numerically-estimated derivative based on most recent infusion doses and their corresponding ANC nadirs. In this algorithm, a relaxation factor ω = 0.75 is applied to any proposed dose increase, with safety in mind. Whereas the slope of log( ANC _nadir ) with respect to d o s e 4 is expected to be strictly negative at steady state, hysteresis effects arising during initial steps of dose titration do sometimes yield positive numerical estimates for this slope; so the slope estimates are constrained to be ≤ 0. The infusion dose for cycle 1 is 50 mg, and the cycle-2 dose is calculated conservatively using a slope −2.0, which is larger (in absolute terms) than for any of our simulated patients except id1 and id13; see Figure 4. For reference, these starting values for the tuning parameters of the titration algorithm are collected in Table 3.

With the illustrative purpose of this article again in mind, we treat neutropenia monitoring as an exogenous process yielding precise nadir timing and levels. This enables a demonstration of the main point without the encumbrance of additional modeling infrastructure peripheral to the main point.

If one considers Figure 5 as a sequence of titration outcomes emerging in serially enrolled study subjects, it becomes clear that even quite early in the study it will seem desirable to ‘retune’ the titration algorithm. For example, provided that course-1 CIN monitoring is implemented with sufficient intensity to deliver advance warning of an impending severely neutropenic nadir, so that timely colony-stimulating factor may be administered prophylactically ¹⁷ , then upon review of the titration courses in the first 10 subjects it may well appear desirable to increase dose ₁ from 50mg to 100mg. Likewise, given the third-dose overshooting that occurs in 4 of the first 10 subjects, it may seem desirable to adjust the relaxation factor ω downward. Of note, at any given time any such proposed retuning may readily be subjected to a ‘dry run’ using retrospective data from all convergent titration courses theretofore collected. (Hysteresis effects would however be inaccessible to a strictly data-driven dry run absent formal modeling that captures such effects.) Furthermore, the ‘tuning’ idea readily generalizes to the fundamental modification or even wholesale replacement of a DTA. For example, the overshooting seen for subjects id8, id10, id12 and id23, inspires further thought about refining (or replacing) the admittedly very naive Newton-Raphson method employed herein. (At the very least, DTAs deployed in actual DTAT studies must incorporate fail-safe upper bounds both on absolute dose and on proportional dose increases.)

A further dimension of ‘tuning’ that must be discussed is the potential for driving the tuning parameters using statistical models built on baseline covariates. Surely, to the extent that the great heterogeneity in final dosing evident in Figure 5 could be predicted based on age, sex, weight or indeed on pharmacogenomic testing, then dose ₁ should be made a function of these covariates. The recalibration of such models as data accumulate from successive study subjects is very much a part of the full concept of ‘tuning’ I wish to advance.

Finally, whereas I have discussed ‘tuning’ here largely in terms of reflective, organic decision-making such as occurs in the creative refinement of algorithms or in data-driven statistical model development, I do not mean to exclude more formal approaches to algorithm tuning. A decision-theoretic framing of the tuning problem should enable formal algorithm tuning to be specified and carried out meaningfully. Such framing would also have the salutary effect of bringing into view objectively the important matter of patients’ heterogeneity with respect to values and goals of care. It seems quite likely that the balance of benefits from aggressive titration versus harms of toxicities will generally differ from one patient to another. Dose titration algorithms should most emphatically be tuned to these factors as well. For example, if a patient with more advanced disease and short expected survival nevertheless decided to enroll in a Phase I DTAT study to pursue the possibility of therapeutic benefit, then this patient’s decision would indicate a subjective weighting of benefits vs. harms favoring a higher starting dose and more aggressive titration.

It is where pharmacometrics meets the field of optimal control that the current literature seems to make its closest point of contact with the DTAT concept I am advancing here. In optimal-control investigations of chemotherapy ^{18– 23} , as in DTAT, relatively large decision spaces are explored. Indeed, the infinite-dimensional spaces of control functions posited for exploration in optimal control applications dwarf the finite-dimensional spaces of tuning parameters in DTAT as dramatically as the latter dwarf the finite sets of discrete doses trialed in now-standard Phase I studies. This intermediate ‘cardinality’ of DTAT reflects an important advantage in an era when, to almost universal chagrin, the detested 3+3 dose-finding design retains its hegemony due partly to widespread resistance to modeling ²⁴ . In such an era, optimal control applications that involve detailed mathematical modeling of tumor biology and dynamics sadly seem consigned to the fringes of practice. Acceptance of such ambitious problem formulations, expressing as they do the spirit of a future age, must await deep cultural changes in the medical sciences and clinical practice.

As easy as it is, however, to disparage ‘resistance to modeling’ as some kind of antediluvian attitude, this resistance does rightfully assert the importance of unmodeled complexities that necessitate application of organic forms of clinical judgment ²⁵ . It should be clear from the above discussion of ‘tuning’ that DTAT readily accommodates and veritably invites scrutiny, supervision and modification by clinical judgment. For example, if during the course of our DTAT study adverse effects other than neutropenia were to emerge as occasional dose-limiting toxicities, then the full concept of ‘tuning’ advanced above would invite dynamic, ‘learn-as-you-go’ modifications of the titration algorithm. Such modifications could begin with decreasing the relaxation factor ω, but might also involve efforts to classify and predict these new DLTs, and to incorporate such new understanding explicitly into the DTA yielded by the study. Indeed, whatever philosophical challenge DTAT embodies is likely to take the form of requiring an intensified commitment to clinical judgment, in a learn-as-you-go world where the always-provisional nature of medical knowledge must frankly be acknowledged ^{6, 26} .

I have presented the DTAT principle here embedded in the context of a specific simulation study. This creates the need explicitly to demarcate what I wish to advance as essential in DTAT, from what is merely incidental to the illustration offered here. DTAT makes its fundamental contribution in putting forward a new abstraction (the DTA with its tuning parameters) capable of embodying knowledge objectively ²⁷ , to supersede a fallacious abstraction (‘the’ MTD) that almost completely lacks this capability. I use the term fallacious advisedly, meaning to identify ‘the’ MTD specifically with what Whitehead called the “ fallacy of misplaced concreteness [which] consists in neglecting the degree of abstraction involved when an actual entity is considered merely so far as it exemplifies certain categories of thought ²⁸ .” Indeed, the 1.5 orders of magnitude spanned by the MTD _is of Figure 5 show the degree of abstraction involved in ‘the’ MTD to be so egregious as to render this concept plainly useless for embodying what we need to learn from Phase I oncology studies.

Several aspects of the illustration offered here should be clearly understood as not essential to the DTAT principle. Firstly, notwithstanding the important heuristic role it has played in the development (and even the naming) of DTAT, recursive filtering in no way delimits DTAT. In fact, I now rather suspect that full-information methods will push recursive filtering to the sidelines in practical DTAT applications, and that whatever utility recursive filtering retains will derive from its use as a vehicle for illustrating DTAT to clinicians, perhaps in nomogram forms ¹⁵ . Secondly, although neutropenia-targeted dosing of a cytotoxic chemotherapy drug has provided a most propitious context for the present simulation, DTAT need not be thought limited to such drugs. In the important area of immuno-oncology, common dose-limiting toxicities (DLTs) admit monitoring on time-scales comparable to the chemotherapy induced neutropenia (CIN) simulated here. For example, the cytokine release syndrome (CRS) that accompanies chimeric antigen receptor (CAR-)T cell therapies typically arises within 1 week of administration (even earlier with concomitant high-dose IL-2) and constitutes a clinical syndrome that admits multivariate monitoring on numerous quantitative clinical and laboratory measures ²⁹ . Even molecularly targeted agents (MTAs), for which late toxicities have attracted the lion’s share of attention ³⁰ , remain accessible to the DTAT principle—with DTA learning occurring on a longer time scale. Of course, a DTA that reacts to diverse, lower-grade MTA toxicities ³¹ that patients experience and evaluate subjectively may resemble a process of ongoing shared decision making (with the oncology care team) more closely than it resembles the impersonal calculations we typically think of as ‘algorithmic’. But with a suitably broadened understanding of ‘algorithm‘—one that accommodates what might typically be termed protocols—the DTAT (or perhaps, DT PT) principle continues to apply. In such applications, supervision and modification by clinical judgment as mentioned above clearly comes to the fore. But even then, the development and application of scoring systems for patient-reported clinical symptoms and quality of life would enable dose titration protocols (DT Ps) to be described objectively in quite ‘algorithmic’ terms that would preserve the applicability of a ‘tuning’ concept.

I have advanced a concept of dose titration algorithm tuning (DTAT), drawing illustrative and orienting connections with recursive filtering and optimal control. I have concretely illustrated key elements of DTAT by simulating neutrophil-nadir-targeted titration of a hypothetical cytotoxic chemotherapy drug with pharmacokinetics and myelosuppressive dynamics patterned on previously estimated population models for docetaxel. I have also discussed the applicability of DTAT to other types of anti-cancer therapy. I believe DTAT presents a prima facie case for discarding the outmoded concept of ‘the’ maximum tolerated dose (MTD) of cancer therapeutics. This argument should be of interest to a wide range of stakeholders, from cancer patients with a stake in receiving optimal individualized ‘MTD _i ’ dosing, to shareholders in pharmaceutical innovation with a stake in efficient dose-finding before Phase III trials.

The data referenced by this article are under copyright with the following copyright statement: Copyright: � 2017 Norris DC

Data associated with the article are available under the terms of the Creative Commons Zero "No rights reserved" data waiver (CC0 1.0 Public domain dedication).

Open Science Framework: Code and Figures for v1 of F1000Research submission: Dose Titration Algorithm Tuning (DTAT) should supersede the Maximum Tolerated Dose (MTD) concept in oncology dose-finding trials, doi 10.17605/osf.io/vwnqz ¹³

Daniela Conrado (Associate Director, Quantitative Medicine at Critical Path Institute) confirms that the author has an appropriate level of expertise to conduct this research, and confirms that the submission is of an acceptable scientific standard. Daniela Conrado declares she has no competing interests.