Adequate trials: How the search for a cure shaped leukemia diagnosis

How many leukemias?

Today, two main criteria are used to classify leukemia: cell type and cell maturity. The intersection of these criteria defines its four most common categories: Acute Lymphoblastic Leukemia (ALL), Acute Myeloblastic Leukemia (AML), Chronic Lymphocytic Leukemia (CLL), and Chronic Myeloid Leukemia (CML).

To contemporary observers, this classification is unproblematic. It guides the treatment of hundreds of thousands of patients worldwide and underlies billions of dollars’ worth of research (Leukemia and Lymphoma Society Report 2016). Yet, until the mid-1960s the classification and diagnosis of leukemia were messy affairs. A leukemic patient was unlikely to receive the same diagnosis from different doctors; not only did doctors disagree on what type of leukemia a specific patient had, they disagreed on what types of leukemia existed (Johnsson, 1949).

How were the controversies around leukemia classification finally resolved? The most straightforward explanation would be that technoscientific advances allowed hematologists to solidify diagnoses based on cell type. Tracing developments in histology and pathology, however, easily rules out such an explanation: the two major breakthroughs in cell classification took place in the early 1900s and the 1980s. Yet the diagnostic grid emerged in the 1960s, that is either six decades too late, or two decades too early.

The first evidence to support the AML/ALL/CML/CLL division appeared in the early 20th century with progress in embryonic cell research (Naegeli, 1912). Rather than resulting in a solid diagnostic grid, these discoveries fueled diagnostic controversy. For decades, scientists were looking under the same microscope, but seeing very different cell entities.

The mid-1970s was marked by the arrival of a new ‘miraculous’ device: the flow-cytometer. Utilizing laser technology and fluorescent dyes, flow-cytometers expanded the ‘biological platform’ of cell classification (Keating & Cambrosio, 2003). Aiding the ‘subjective’ and ‘qualitative’ microscopic gaze with ‘objective’ and ‘quantitative’ precision, flow-cytometers rendered classificatory controversies obsolete.

Looking at the historical timeline, we find that the new classification followed neither the advances in cell morphology, nor the diffusion of flow-cytometers. ¹ If the advent of technological ‘arbitrators’ was not the key, then how can we explain the emergence of the diagnostic grid? I argue that the major development that enabled hematologists resolve the controversy in the 1960s was the emergence of clinical research into cancer chemotherapy. In an attempt to prove that chemical cancer therapy was possible, oncologists restructured the institutions and practices of cancer medicine, including the classifications of leukemias and leukemic patients.

To develop this argument, I propose to study leukemia diagnosis as a social category and process (Blaxter, 1978). Social science has given considerable attention to medical classifications, with a specific call for a ‘sociology of diagnosis’ (e.g. Jutel & Nettleton, 2011). I further this conversation by arguing that the analytical key to explaining the formation and transformation of diagnosis is the relationship between diagnosis and therapy.

My account builds on the existing sociological literature but diverges significantly from dominant approaches. Most research on diagnoses and on the relation between diagnosis and therapy falls under one of two categories. The sociology of the professions assumes that doctors gain and secure the power to classify when they demonstrate their ability to solve existing problems successfully. Even if theorists see the key to such success as a question of attribution (Eyal, 2019), according to this approach drugs become transformative when they work. Other approaches, such as the medicalization literature, analyze drug effects as social constructions. Drugs shape society, culture, and medical practice itself, when they channel powerful actors.

Building on work in Science and Technology Studies (STS) I take the middle path. To explain how therapies become impactful, it is not enough to consider them as mere ‘social constructions’. Yet, it is similarly problematic to argue that therapies become transformative when they produce physiological effects independent of any social context.

I argue that therapies can become powerful when they work just well enough to mobilize support and resources. Analyzing the history of cancer chemotherapy, I show that when drugs delivered problematic and limited results, experts pursued therapeutic promise by redefining what counts as a ‘problem’ and what qualifies as a ‘solution’. Oncologists created concrete organizational, epistemic, and material infrastructures that enabled anticancer drugs to produce effects and that enabled oncologists to interpret these effects as therapeutic.

To trace the process by which the invention of new drugs brought about the invention of new disease categories, I develop the concept ‘adequate trials’. With this concept, I capture a specific mode of innovation that combines both the testing and exploration of new technologies; instead of merely testing drugs to see if they work, oncologists created the conditions that enabled them to work.

I show that the commitment to ‘adequate trials’ and to the restructuring of cancer medicine was the result of two displacements. First, oncologists translated their deep-seated normative commitment to patients into a commitment to drugs themselves. Second, oncologists replaced the question Does this drug work? with the question What does it mean for this drug to work? I then turn to identify the organizational, scientific, and political contingencies that shaped these displacements and show how they led to the rewriting of diagnoses. By identifying the conditions that gave rise to oncological practice, I theorize the processes by which rigid testing becomes a medium for normative and epistemic change. In this sense, I contribute to our understanding of how technology and society are co-produced (Jasanoff, 2004).

Social studies of medical classifications

It is tempting to think that disease categories evolve and become increasingly specific as we learn to carve nature at its joints. Work in STS, however, reminds us that diagnoses are not neutral outcomes of scientific progress but are equally shaped by politics and culture (Bowker & Star, 2000; Brown, 1995; Epstein, 2009). To explain the formation and transformation of diagnoses, we must turn to sociological theory.

Three sociological approaches are obvious candidates for the study of the leukemia controversy: the sociology of the professions, the medicalization literature, and studies of standardization. I will briefly review each of these literatures to articulate their implications for the case of leukemia and to show why they fail to produce convincing accounts.

From the perspective of the sociology of the professions, diagnoses are tools used by competing groups of experts in struggles over professional jurisdiction (Abbott, 1988). Diagnoses frame various aspects of social life as targets for specific interventions, thus enabling professions to claim them under their jurisdictions.

Indeed, the solidification of the various categories of blood cancer went hand-in-hand with the emergence of oncology as a new professional group. In the mid-1940s and through most of the 1960s, oncology was a marginalized medical specialty, overseen by the internal medicine board and practiced in a handful of institutions. By the mid-1970s, however, oncology had become one of the major biomedical fields in the United States (Bud, 1978; Keating & Cambrosio, 2011). As oncology evolved, the classification of cancer became sophisticated and elaborate. Resources were dedicated to the study of increasingly specific tumorous subtypes, and oncology itself split into subspecialties.

Although the solidification of leukemia diagnoses is clearly related to the emergence of oncology, theories of jurisdictional claims have serious drawbacks when applied to the case of leukemia. According to the sociology of the professions, doctors expand their jurisdictions by competing over new domains of practice. Following this logic, we should expect to find competition over the territory of leukemia. However, no one was competing over leukemia. Borrowing Abbott’s terminology, we could say that leukemia was a territory that no one wanted ‘to get stuck with’. As Krueger shows in her work on childhood cancers, doctors distanced themselves from leukemic patients and especially from leukemic children. Because it was seen as a death sentence, leukemia revealed the incompetence of the medical profession and threatened its status. Krueger (2008, p. 54) quotes a participant in a 1966 conference to describe the general sentiment of pediatricians toward leukemic children: ‘How can I get rid of this patient? Nobody likes doctors who take care of dying kids; how can I unload this patient?’

Another possible explanation is that oncology emerged with the discovery of potential anticancer drugs. Leukemia, one could argue, was about to become an object of competition. This objection, too, is unconvincing. The early experimentations with cytotoxic drugs were gory endeavors. While experimentation gave hope to patients and their families, its everyday realities were gruesome and, in many ways, harsher than the disease itself.

The cytotoxic compounds destroyed patients’ immune systems and the blood’s ability to clot, and patients would succumb to fatal infections or uncontrollable bleedings. Of all leukemic children treated at the National Cancer Institute (NCI) between 1965 and 1971, 60%–80% died not from cancer but from viral, fungal, or bacterial infections related to the treatment (Levine et al., 1972). As medical experience accumulated, doctors documented harsh side-effects, including permanent damage to vital organs, loss of cognitive abilities, damage to mental health, and a ‘second cancer’ induced by therapy itself.

Children cycled from illness, through short periods of remissions, and then inevitably relapsed. Clinical photographs taken of patients with acute leukemia graphically demonstrated the dreadful manifestation of the disease and its attendant toxic therapies. A long list of complications included: oral ulcers, leukemic infiltration of the eyes, chronic nosebleeds, uncontrollable hemorrhaging and life-threatening infections. … The condition debilitated the sufferers and necessitated frequent trips to and from the clinic as well as extended hospital stays. (Krueger, 2008, p. 86)

While treatment of leukemia did expand medical territory, ² it also jeopardized the medical profession’s claim over its jurisdiction. Oncologists risked the lives and wellbeing of American children. Not only parents, but pediatricians and family doctors refused to refer children to oncologists. For decades, chemotherapy was hardly a basis for gaining status and legitimacy, either within the medical profession or with regulators, patients, or the American public (DeVita & DeVita-Raeburn, 2015).

The medicalization literature teaches us that medical diagnoses are used to reframe social problems as medical ones in order to gain control over deviant populations (Zola, 1972). The most common form of medicalization is perhaps the turning of historically non-medical problems, such as homosexuality and addiction, into medical ones (Conrad & Schneider, 1992). Another medicalization strategy is the decreasing of thresholds for inclusion within diagnostic categories, similar to that observed with the category for ‘depression’ (Healy, 1997).

Although cancer patients were not seen as a deviant population, medicalization still had important implications for the case of leukemia. The development of cancer medicine was inherently linked to state power and the increasing capacity to monitor and manage populations. Cancer medicine depended on the ability of the state to enter the domestic space and construct the ‘health’ of the American population as an object of governance (via educational programs and public screening campaigns for example). Government investment in cancer research sustained the autonomy of the pioneering oncologists who worked at the NCI, among them Emil Frei and Emil Freireich. This autonomy enabled oncologists to sidestep community-based practitioners who would not subject their patients to toxic experimentation.

While the role of state power must be factored into the account, it fails to explain why the controversy around leukemia classification was resolved in oncology clinics. Oncologists needed state support, but how did they manage to enlist the government as an ally? How did oncologists link their ‘scandalous’ practices to conduits of state power? I will return to this question in the discussion on therapy.

Finally, the literature on standardization is also key to the case of leukemia. Standardization is needed for: (1) creating the objects of scientific investigation (Kohler, 1994; Landecker, 2007), (2) creating the controlled conditions under which regularities can be observed (Knorr-Cetina, 1999; Latour, 1999), (3) establishing ‘technologies of trust’ that sustain scientific collaboration (Porter, 1996) and (4) to mitigate attempts of regulatory intervention (Timmermans & Berg, 2003). Leukemia classification is a classic standardization story. This paper shows that ‘hard’ evidence did not precede the formation of leukemia diagnosis. To the contrary, standardization of leukemia diagnosis enabled researchers to collect the evidence necessary to retrospectively justify cell-based classification.

To articulate the key role of standardization in the case of leukemia, we must examine how obstacles to standardization are overcome. The standardization literature demonstrates that when diagnostic categories become standardized, more practices, objects and infrastructures will soon follow (Star & Lampland, 2008). The creation of the DSM-II for example, led to standardization not only of diagnoses, but also of the psychopharmaceutic industry, third-party reimbursement, and patient identity (Timmermans & Epstein, 2010, p. 78, citing Lakoff, 2005).

Multiple mechanisms have been suggested for explaining the ‘imperialism’ of standards. Standards tend to become tangible and enmeshed in infrastructures and practice. They tend to become transparent and to make residuals and deviation impossible. Standard also perform regulatory functions and are used to endow/deny legitimacy (Bowker & Star, 2000).

Following this literature, we should expect standardization to expand outwards and inwards. In the case of leukemia, we find that attempts at standardization quickly reached an impasse. The standardization of the diagnostic category (ALL/AML/CLL/CML) and of the diagnostic tool (differentiation between lymphoblasts and myeloblasts) were mutually dependent. To standardize the disease category, hematologists needed to standardize cell classification, and to standardize cell classification, they had to standardize the disease categories. Hematologists spent six decades stuck in a ‘double-bind’ and could only get out with the exogenous pressure of therapeutics. I will return to this point in the following section.

Diagnosis and therapy

Theories of jurisdictional struggle and medicalization fail to explain the case of leukemia because both assume that the medical profession’s ability to gain power depends on their ability to successfully solve existing problems (whether they assume that drugs ‘work’, or that they function as mere social constructions). From this perspective, it is unclear how oncologists managed to expand their jurisdiction or align with state power, when they so clearly failed to deliver politically useful results.

A sociology of therapy is useful here. I argue that experts gain power not only by solving existing problems but also by reinventing what counts as a ‘problem’ and what qualifies as a ‘solution’. ³ Drawing on Eyal (2013) I move from a sociology of the professions to a ‘sociology of tasks and problems’. I propose that experimentation with new therapies drove oncologists to restructure the task and problems of cancer medicine. More specifically, I show that experimentation with antineoplastic drugs in the 1960s, provided hematologists with the incentives and tools to solve the longstanding leukemia controversy.

In order to explain the transformative power of chemotherapy, I first dispel some sociological notions about therapy. A common approach for theorizing the role of treatment in diagnosis is the ‘diagnosis by treatment’ approach (Aronowitz, 2001). It refers to a trivial medical practice: Doctors often use their patients’ response to treatment to assign them a diagnostic category. In this case, the classification scheme remains constant; the only problem is to confirm that the patient falls under the suspected category.

Another popular approach’ is the ‘wild empiricism’ narrative. We can find its traces in popular science narratives (Mukherjee, 2011), in sociological work that tangentially engaged with early chemotherapy (Fujimura, 1996), and it is often adopted by oncologists themselves. The story goes as follows: In the 1940s, basic sciences were making very little progress in cancer research. There was practically nothing doctors could do for patients with inoperable cancers. Desperate to help patients, oncologists refused to be held back. Instead of waiting for scientists to figure out cancer etiology, they decided to treat patients. Oncologists took the handful of cytotoxic compounds in their arsenal and began shooting in all directions – testing compounds on whatever cancer they could put their hands on and tweaking dosages and schedules until they turned up a response. At the time, oncologists were highly criticized for their ‘shotgun empiricism’. Their methods were considered ‘irrational’, ‘unscientific’, and even ‘scandalous’. In later accounts the same practices were praised for their visionary courage: Oncologists paved the way for future medicine by demonstrating that chemical cancer treatment was even possible.

There are two main problems with these common notions about therapy and chemotherapy. First, they assume the stability of diagnostic categories. The ‘wild empiricism’ story, for example, celebrates Frei and Freireich’s discovery of a cure for ALL. Such a statement is highly anachronistic. When Frei and Freireich began their experimentations, this subtype of leukemia did not exist. It would be more accurate to say that when Frei and Freireich induced remissions in a subset of leukemic children, they discovered ALL. Oncologists did not use drugs to determine if patients had ALL or CLL, as suggested in the ‘diagnosis by treatment’ approach, but rather to create these categories.

Another problem is that both accounts assume that therapy has social and political impact when it is successful. Yet, as we have seen, this is not quite the case with cancer chemotherapy. It would be difficult, if not impossible, to calculate if early oncologists were helping more than harming. To explain how chemotherapy gained power I propose a rather different model. I argue that therapies gain transformative potential when they are not quite successful. Therapies transform institutions and practices when they work just well enough to produce hope (Delvecchio Good et al., 1990).

To explain these transformations, the following section will introduce the concept of ‘adequate trials’ and show how specific organizational, political, scientific, and jurisdictional pressures led oncologists (1) to channel the normative commitment to save patients into a commitment to the drugs themselves, and (2) to shift from a focus on drug testing to an inquiry into what it means for a drug to work.

Adequate trials

‘Adequate trials’ was a key term for oncologists in the 1960s. It articulated oncologists’ concerns that candidate drugs would not receive a ‘fair chance’ to demonstrate their potential due to normative, bureaucratic, scientific, and clinical obstacles. I use this term to characterize a specific mode of technological innovation that is relevant for 1960s oncology as well as other techno-scientific sites. This mode of innovation, I argue, is distinct because it combines both the exploration and testing of new technologies.

Technological innovation has multiple aspects. Actors often explore the potential of new technologies by actively transforming social environments to create new possible usages and to encourage technological diffusion. Actors can push for the creation of new infrastructures and grids, as well as shape consumer needs and preferences, to promote technologies (e.g. Cowan, 1985). Just as often, actors must engage in the testing of new technologies to establish if a given technology works under a fixed set of parameters.

Scholars in STS have studied the social processes that shape the setting and interpretation of such parameters and demonstrated their contextual flexibility (Bijker et al., 2012). Yet even when considering the flexibility of test parameters, there are significant differences between exploration and testing. Studies in STS and of innovation have differentiated open-ended searches from testing (e.g. Opazo, 2018). The concept of ‘adequate trials’ highlights the boundaries and transitions between these modes of innovation. Under what conditions do rigid tests and their goals become more malleable?

Oncologists working in the early 1960s rigorously tested drugs and did not engage in open-ended explorations. They also did not simply ask if drugs worked but created the conditions under which drugs could work. What incentives, pressures and constraints generated the ‘adequate trials’ approach? I provide a summary of the conditions that were underlying oncologists’ practical orientations. This summary also serves as a roadmap for the following sections.

Three major factors solidified oncologists’ commitment to cancer chemotherapy. (i) Early therapeutic enthusiasm about cancer chemotherapy was enough for raising (ii) initial government funding, which in turn secured (iii) institutional autonomy for oncologists working at the NCI. We should note that the early therapeutic enthusiasm was grounded in two realities.

First, leukemic children showed remarkable responses to cytotoxic therapy. Second, leukemic adults were far less likely to respond when treated with the same drugs. The fact that oncologists made progress treating dying children gave them considerable political and rhetorical leverage. It was much easier to justify radical procedures and all-out attacks when the life of a young child was at stake. The persistent difference in response rates between children and adults was also important, because it suggested disease specificity and reinforced the question of causality. These factors allowed oncologists to leverage limited clinical achievements into therapeutic hope.

In the absence of any of these factors, it is highly likely that the immediate failure to produce satisfactory results would have been enough to discourage oncologists from undertaking further experimentation. Oncologists would have simply given up. Working under these specific conditions, however, oncologists developed an unwavering commitment to the vision of chemical cancer therapy.

While these three factors consolidated the undertaking, four major contingencies shaped oncologists’ disposition toward the restructuring of goals and tasks: (i) scarcity of candidate drugs – only a handful of the hundreds of thousands of compounds screened by the NCI made it past the laboratory; (ii) scientific and clinical uncertainty – there were no baselines for determining if a drug ‘works’ or what were ‘acceptable’ risks and side-effects, and there were countless ways to administer drugs, each producing very different therapeutic and toxic effects ⁴ ; (iii) underdeveloped infrastructure – trials were conducted by a network of research groups spread across the country, so materials and practices were unstandardized and results variously biased and tainted by noise; (iv) jurisdictional pressures – the harsh realities of chemotherapy made it difficult to enlist patients, their families, and other medical professionals. Defending the legitimacy of chemotherapy and patient recruitment proved to be challenging.

Working under these conditions, oncologists became highly invested in each of the candidate drugs in their arsenal. To give drugs a realistic opportunity to succeed, oncologists became willing to stretch out the experimental, organizational, and rhetorical resources at their disposal. The very notion of chemical cancer therapy was being worked out, so oncologists had to let go of the grand goals and rationales that originally guided the enterprise. The only way to move forward was by keeping their eyes on the ‘ends-in-view’: They had to continuously invent and revise short-term, realistic, and provisional endpoints and markers for the evaluation of both drugs and the experimental design itself. These included outcome variables, inclusion and exclusion criteria, as well as diagnoses themselves.

More specifically, to create the conditions under which drugs could work, oncologists regrouped leukemic patients according to treatment response, creating the incentive and infrastructure that allowed the standardization of cell classification. Similarly, they had to reinvent metrics and methods for drug evaluation. While oncologists’ achievements were revolutionary, they looked nothing like the original clinical goals from when they had begun their search. To justify experimentation with chemotherapy, one had to come up with alternative yardsticks beyond ‘cured’. Cancer chemotherapy trials depended on the development of complex new tools for monitoring patients, measuring whether patients were getting better, and calculating how much better, for example, scales for defining degrees of remission, as well as calculation of progression-free survival and time-until-relapse (CALG-B, 1960).

To link the discussion on ‘adequate trials’ to the theorizing of diagnosis, I draw on Zerubavel (1996) and Mol (2003) to propose that diagnoses do not mirror ontological types, but are practical classification resulting from technological interventions. Therefore, we should expect to find multiple, orthogonal and overlapping ways to divide and classify phenomena. Moreover, we should expect diagnostic classifications to be drawn and redrawn as technologies of medical intervention change. As I argued above, however, therapy is not an isolated technology plugged into an existing matrix to solve existing problems. A sociological account must approach therapy as a complex set of institutions and practices, in the context of which the very goals of therapy, as well as the social role of the profession providing it, are determined. In this paper, I analyze diagnoses as part of the institutional context that shapes the practical orientations of doctors and researchers (Clarke & Fujimura, 1992; Rheinberger, 1997).

Data and methods

For this research, I have collected and analyzed previously unstudied archival materials from the leading institutions in hematology and oncology in the 1940s–1970s: most importantly, the trial protocols of the Cancer and Leukemia Group B, the Nathan Mental Papers Collection at the National Library of Medicine, and the William Dameshek Papers Collections at the Historical Archive of the Harvard Medical School. ⁵ These diverse materials enabled me to conduct a comprehensive review of documents produced by scientists, oncologists, and government regulators and administrators (within and outside NCI and NIH).

In addition, I reviewed titles and abstracts of all papers published in The Cancer Chemotherapy Report (n = 1584), the central oncology journal in the first three decades of the NCI, and analyzed all papers concerning leukemia (n = 159). I then traced relevant citation networks to study key papers published between 1900 and 1970 in additional medical journals (e.g. The Lancet, the Canadian Medical Association Journal).

I conducted seven semi-structured interviews with key figures in oncology who were involved in early cancer clinical trials in order to test my interpretations of the archival materials and inquire about informal practices that could be missing from written documentation. Interviews lasted an average of one hour and were recorded and transcribed.

Since I focus on the work of oncologists in the NCI and academic centers, analyses of patients’ perspective, the pharmaceutical industry and community hospitals serving the majority of Americans are missing from this account. There were several reasons for these choices; the pharmaceutical industry was simply not a major actor in the early years of cancer chemotherapy. ⁶ The perspective of patients and their families, as well as the everyday practice of cancer medicine in community settings are essential pieces of the puzzle. Yet because so little scholarly attention has been paid to early chemotherapy trials, despite their enormous historical significance, I found it critical to focus on their technical and bureaucratic details. I hope that research into the early experimentation with chemotherapy will inform future research into the experiences of patients and the practice of cancer medicine between 1960 and 1980.

Leukemia classifications before 1960

Until the early 20th century, leukemia was divided into two major categories: acute and chronic. The first had a sudden onset and progressed quickly. The latter developed slowly until reaching an ‘acute phase’ (Norman & Gwyn, 1930). This distinction was based primarily on clinical observation. Acute and chronic leukemias were assumed to involve different types of cancerous white blood cells – immature lymphocytes and mature myelocytes, respectively.

This classification was grounded in contemporary histology. In the late 19th century, hematologists could differentiate between healthy white blood cells, but immature cells did not develop the functions by which healthy cells were differentiated. It was therefore difficult to classify the primitive cells characteristic of acute leukemia. At the time, researchers commonly identified myelocytes by their granularity, but since primitive cells do not present granula, acute leukemia was considered a ‘lymphatic’ disorder (Johnsson, 1949). Advances in embryonic cells research soon complicated this distinction. In 1905, Naegeli first described ‘myeloblasts’ – immature leukemic cells that were myelocytic precursors. As reports on acute myelogenous leukemia accumulated, a heated debate ensued: Some hematologists adhered to the distinction between lymphatic and myelogenous leukemias, while others suspected that ‘lymphatic leukemia’ was a spurious category resulting from unrefined histology. In the early 20th century, wrote Asselstine (1932), all acute cases were classified as lymphocytic. If anything other than ‘a lymphatic blood picture’ was found, the diagnosis was considered to be ‘prejudiced’. According to Asselstine, however, since 1905 the tendency had been to classify most cases as myelogenous.

Contemporary histological tests were insufficient to solve the controversy. Dameshek (1962), a leading hematologist who became President of the American Society of Hematology in 1965, asserted that cell morphology is not enough to trace a cell line to the mature form and to differentiate granulocytic from lymphocytic leukemia. Naegeli’s discovery spurred the development of new chemical tests, but these were practically self-defeating. The oxidase reaction introduced by Dunn (1913) is a striking example. A positive oxidase reaction could confirm that the blasts were myelocytic precursors, but a negative reaction could only confirm that the blasts are immature. It could not eliminate the possibility that the blasts were myelogenous or affirm that they were lymphocytes precursors. The classification of cells was highly unstandardized and depended on the specific practices/beliefs of individual researchers:

The cell types observed in individual patients could appear to some to be characteristic of acute leukemia, and to others of an ‘acute phase’ or chronic leukemia, according to their predilections. (Dameshek, 1962, p. 19)

As cell morphology could not determine diagnoses, doctors aimed to link histology to clinical patterns. Dameshek (1962) described the limitation of cell morphology and the need to use clinical data to interpret the microscopic blood picture:

How many cases of acute leukemia can be classified according to … cellular type is a matter of debate and often personal conviction. We believe that the study of the morphology of the blasts alone, by ordinary staining methods alone, cannot lead to their inclusion in one or the other of the cell series … Perhaps correlation of the physical examination with the blood picture may be as accurate. (p. 4)

The identification of lymphoblasts depended on the collection of standardized blood samples from patients with acute leukemia. To identify these patients, however, one had to have already determined that the number of lymphoblasts in the blood picture were consistent with a diagnosis of acute leukemia. Daston and Galison (2007) have argued that what scientists see under a microscope depends on sociohistorical configurations of objectivity. In the case of leukemia, we find that to standardize the interpretation of the blood picture, hematologists needed to standardize prior clinical classification of patients from which the blood came. As both classification of cells and of patients were co-dependent, it was impossible to move forward and establish new classificatory schemes of cells and leukemias.

One the one hand, hematologists hoped that better diagnostics would enable them to standardize cell classification:

There is still much argument about the influence of the cell type on the chances of remission, and this cannot be expected to become resolved until there is a uniformly acceptable classification of the acute leukemias. (Dameshek, 1962, p. 19)

One the other hand, it was impossible to improve the diagnoses of patients when cell classification was lagging behind. As diagnostic chaos ruled, doctors provided patients with only a general diagnosis of leukemia and avoided the question of type:

In this stage of uncertainty about … the differences between individual blasts we prefer to … regard many cases as ‘leukemia acute’, rather than to make definitive assumption about their derivation from any one cell line. (Dameshek, 1962, p. 5)

In 1962, a group of researchers associated with the Leukemia Group B published a study motivated by ‘the suspicion’ that the imprecision of clinical knowledge is ‘due to the use, in different institutions (or in the same institution in different times), of varying criteria for cytologic classification of acute leukemia’ (Lee et al., 1962, p. 151). In a blind study comparing 381 classifications of 42 leukemic cases, the researchers found that 30.4% of the reports failed to receive a diagnosis confirmed by a clear majority, and 9 out of 42 cases failed to receive any agreed diagnosis (p. 152).

How did researchers get out of this bind? How did cell type gain enough diagnostic value to validate the distinction between four different types of leukemias?

What is an anticancer drug?

In the mid-1940s, researchers first discovered that cytotoxic compounds showed antitumor effects in leukemic patients. Wartime experimentation with warfare gases led Goodman et al. (1946) to discover the antitumor effects of nitrogen mustard, while Farber (1948) reached similar results with folic-acid antagonists. These discoveries sparked an interest in cancer chemotherapy. In 1955, the US government established the CCNSC, an agency mandated to coordinate and carry out large-scale screening and testing of compounds for antineoplastic activity. Assembling the infrastructure and orchestrating the trials, however, proved a challenging task.

As we shall see in the following sections, the scarcity of candidate-drugs, uncertainty around the very notion of ‘anticancer drugs’, obstacles for standardization of practices and materials across trials as well as challenges to recruitment and allocation of patients were key factors that blended the boundaries between testing and inventing.

In clinical testing, it is standard to evaluate drugs by comparing new compounds with existing ones. This practice disguises some of the complexities of evaluation: What criteria should guide the measurement of therapeutic gains? And when is therapeutic gain substantial enough for us to say that a drug is ‘working’? Usually, we do not need to confront these questions, because we can trust existing conventions. We simply need to decide if a new drug is performing better or worse than existing ones.

In the case of cytotoxic compounds in the 1950s and 1960s, no existing therapies could provide baselines for evaluation. Antimetabolites and alkylating agents produced clinical improvements, but these were unsystematic and difficult to measure. Moreover, experimentation in animals and humans showed that cytotoxic drugs produced highly variable effects depending on how they were put to use. Response depended not only on the drug itself, but on type and stage of cancer and on treatment regimen. Variation in dosage and treatment intervals (e.g. intermittent, daily), as well as in preparation (e.g. oral, intravenous) and use of supportive therapies (e.g. antibiotics, steroids) made a huge difference in response. ‘Our greatest unsolved problem is the most effective rhythm of administration for each type of drug … the choice between intermittent or daily therapy … has not been determined for each of the antitumor drugs’ (Wright, 1962, p. 72). Negative results could equally mean that the drug is ineffective, or that oncologists do not know how and on whom to use it.

Drug toxicity further complicated the experimentation. Oncologists struggled to strike the right balance between efficacy and toxicity, working with narrow ‘therapeutic margins’. Increasing the dose could easily kill patients. Administering too low a dose could easily lead researchers to discard a miracle drug.

Toxicity studies in animal models were of limited use. In a comprehensive review, Owens (1962) concluded that toxicity studies in rodents, dogs and monkeys were inadequate for predicting toxicity in humans on all key outcomes (nervous system toxicity, gastrointestinal toxicity, hepatic toxicity, bone marrow toxicity, and renal toxicity). Owens criticized the ‘arbitrary’ convention of using one-tenth of the animal model dosage to determine the Maximum Tolerated Dose (MTD). Similarly, fellow oncologist Louis (1962) disapproved of the irrationality of contemporary procedures for exploring the relations between toxicity and efficacy:

In general, four major procedures for drug administration have evolved. One method relies on the physician’s intuition to determine intervals for and increments of change in dosage. On the basis of such an approach involving 3 or 4 patients, a regimen might be established which ever after is held sacred. A second method assumes that therapeutic effect should become evident shortly after initiation of a specific dose, and dose levels are rapidly increased so that unusually severe toxicity results. A third approach is similar to an all-or-none phenomenon in that huge doses are given on the assumption that toxic or near toxic levels are required to produce maximum oncolytic effect. With this method, agents are considered therapeutic failures when toxicity occurs without oncolytic effect. The fourth method, which appears to be a more scientific and rational approach, engenders the concept that dose levels should be determined by a balance between therapeutic effectiveness and drug toxicity. This method, however, allows for doses to be tailored to the individual patient’s requirement so that uniform data is not obtained since no two patients receive the same dose. (p. 99)

Louis’ review demonstrates the extent of controversy and uncertainty surrounding early cytotoxic therapy. Each approach involved untested assumptions about the progression of cancer and about the therapeutic mechanisms of anticancer drugs. Even in relatively straightforward testing in mice, miscalculation of toxicity/cumulative-toxicity could have easily led researchers to mistake drug toxicity with inefficacy.

Another factor that increased the stakes was the scarcity of candidate drugs. The CCNSC screened hundreds of thousands of compounds in animals and tissue cultures, but over 90% were rejected at as they proved to be either too toxic or ineffective (CCR 1, 1959; CCNSC Advisory Council Meeting Minutes, 1955, 1956; Zubrod, 1960, 1984; Zubrod et al., 1977). Only a handful of compounds showed acceptable ‘therapeutic ratios’ (i.e. neoplastic activity in dosages that were considerably below lethal) and left the laboratory. Oncologists were therefore heavily invested in candidate drugs. Discarding compounds was not to be done lightly, without making sure that drugs were tested under optimal conditions for demonstrating efficacy.

These conditions fostered a very specific orientation. Oncologists were highly committed to giving candidate drugs ‘adequate trials’, but it was unclear what such trials were. There was no straightforward way to decide if drugs worked or not, and no standard protocol for discarding drugs. Further, it was simply impossible to test every potential method and schedule for administering new compounds. As oncologists couldn’t prove that a drug is not working, the task of drug testing had to be redefined.

The minutes from a CCNSC meeting that took place in August 1967 demonstrate that the main question for oncologists at the time was not Is this drug working? but What does it mean for a drug to work?

(1) How does one measure ‘effect’ of a new drug under consideration? Should the median, the mean or some other population statistic of individual survival be used? … (2) how can the effect of different dose regimen be compared? Is it possible to interpolate between regimens to get the ‘best’ treatment schedule? … (3) what are proper experimental designs to deduce the ‘best’ regimen and the influence of drugs in general? What is the meaning of ‘best’?

The quotation marks in this text are very telling. The meeting participants were not asking if a specific drug has an effect but what an ‘effect’ is and how they could measure it. Oncologists were inventing the very notion of anticancer drugs, blending clinical, scientific, and normative concerns: What should be counted as ‘therapeutic benefit’? How much risk or harmful side-effects should be tolerated? Trying to make progress with these questions, oncologists openly adopted a few arbitrary numerical values for designing drug trials. Protocol 22 (1962, p. 12) of the Cancer and Leukemia Group B provides an example of one of many definitions of an ‘adequate trial’:

Adequate trial is defined as a dose capable of inducing either remission or toxic manifestation requiring interruption or diminution of drug dose administered for an accumulated period of six weeks. Drug dose will be considered adequate irrespective of length of administration beyond three weeks, if drug toxicity and progressive disease coexist, or if response occurs. … It is arbitrarily assumed that only drugs capable of inducing response in not less than 30% of patients will be of interest for the group. Nine consecutive negative patients will be sufficient to exclude that such degree of activity exist for any given drug.

This excerpt provides an excellent example of the logic of ‘adequate trials’. The group decided on a threshold for acceptable response rate. But the response rate of what group? Oncologists were not only trying to find drugs that would work for 30% of the patients, but also to figure out which diagnoses should define the relevant 100% of the patient population.

Same drugs, different results

Another challenge concerned the standardization of trials. Most clinical testing was not conducted in-house but by a network of cooperative groups spread across the country. To facilitate this collaborative work, the CCNSC had to invest extensively in eliminating variation in practices and materials. As a series of reports published in 1962 indicates, the main problem of drug testing was that oncologists were testing the same compounds but getting vastly different results (CCR, 1962). ‘In an evaluation of studies of chemotherapeutic drugs in man, one cannot help but be struck by the divergent results reported with the same agent, despite careful clinical evaluation by experienced observers’ (Wright, 1962, p. 69, emphasis mine).

Marvin Schneiderman, a cancer researcher at the NCI, wrote disapprovingly that no conclusive results regarding the efficacy of 5-FU had been obtained, despite it having been tested in over 1412 patients by 1961. ‘The reports available are characterized by a wide range in percent of patients showing objective response’ (Schneiderman, 1962, p. 107). Schneiderman (p. 108) compiled a list of all the factors that might have contributed to inconsistency in 5-FU trials. The list included: (1) the lack of standardization in the quality of 5-FU, (2) the ‘patient-site mix’ (i.e. the organs or locations of cancer included in the trial), (3) the criteria for inclusion and exclusion of patients from the final statistical analysis, (4) fuzzy definitions of ‘response’, (5) inconsistencies in drug administration (schedule and dosage), and (6) the ‘clinical material’, or the kind of patient treated (e.g. degree of illness, sex, age). To examine variability on each variable, Schneiderman pooled data from all published papers on 5-FU. He concluded that the supply of 5-FU had been adequately standardized but that all other factors were highly likely to introduce inconsistencies.

Keating and Cambrosio (2011) offer a detailed account of the initiatives taken by the NCI and the cooperative groups to standardize cancer trials. These include standardization of chemical substances and plant extracts, securing of pure strains of inbred animals, as well as devising protocols for clinical and laboratory research and providing services of data management and statistical consultation. In this article, I focus on yet another key aspect of trial standardization – the effort to homogenize the samples of cancer patients.

Concluding his analysis, Schneiderman expressed particular concern over variation in ‘clinical material’. Oncologists recruited subjects who differed on key variables that affect clinical outcomes, and they also failed to report on sample makeup in publications. As long as the ‘clinical material’ was not properly standardized, argued Schneiderman, no systematic progress could be made.

Optimizing ‘clinical material’

Because of the limited understanding of cancer, diagnostic categories were unstable and uninformative in the design of early clinical trials (CCNSC Advisory Council Meeting Minutes, 1956). No scientific principles or clinical experience could help predict which drugs were likely to work in which types of cancer. Nor could they help in tailoring treatment protocols to specific cancers. Researchers were conducting ‘exploratory trials’; they tested drugs as broadly as possible, taking any type of cancer ‘they could get their hands on’ (Gehan, 1962). It was because of their improvised and intuitive nature, however, that exploratory studies did not produce conclusive results.

As new compounds and funding became available through the CCNSC, the problem of matching patients and trials became a serious clinical/bureaucratic challenge. The CCNSC and the network of cooperative groups worked to set up ‘rational’ procedures for patient recruitment. In a paper titled ‘Selection of patients for evaluation of chemotherapeutics procedures in advanced cancer’, Karnofsky (1962), a leading oncologist from Sloan Kettering, wrote:

It is estimated that there are more than 500,000 new cases of cancer in the United States each year. … Of this number about 350,000 will have unresectable, recurrent, or advanced disease. If all patients with unresectable neoplastic disease could be assigned to planned chemotherapeutic studies, and if adequate funds and medical and hospital facilities were available, therapeutic trials could be designed and completed with optimal efficiency. … In our experience the investigator often does not have the opportunity to select specific patients for his investigations. The factors responsible for patient availability usually result in clinical material considerably different from what the investigator might select under ideal conditions. (p. 73)

Oncologists faced an operational challenge: How could the NCI and cooperative groups locate and allocate the ideal patients for each clinical trial? As patients and trials were spread across the US, this was a serious obstacle.

Karnofsky’s paper raised yet another problem. Regardless of the type of cancer they had, patients were terrible ‘clinical material’. Patients ended up in clinical trials only after they failed in standard therapy. The ethical reasoning was clear: Only those who ‘had nothing to lose’ should be subjected to potentially ineffective or harmful procedures. Such patients were often more vulnerable to drug toxicity and were less likely to survive long enough for the drug to ‘kick in’. Oncologists worried that drugs that could potentially produce remissions in less advanced patients would be prematurely discarded.

Even when suitable patients were found, it was often difficult to enroll them. David Nathan, who became chief of hematology at Boston’s Children Hospital after completing an internship at the NCI in 1967, tells that he had to take children one by one from the pediatrics department to the oncology ward, because pediatricians refused to refer them (interview, 2017). Takao Ohnuma tells that in Rosewell Park the patients who ended up in cancer trials were often leukemic patients who had no access to care for alleviating symptoms. Patients would come to the hospital to receive antibiotics and blood transfusions, and were convinced to try curative chemotherapy (interview, 2018). George Canellos, reflecting on his days at the NCI in the late 1960s, tells a similar story. Oncologists had more success recruiting patients from distant areas (e.g. rural Virginia) than from the areas closely neighboring the NCI. This was because patients from the District of Columbia had good medical care and they avoided oncologists, while patients with limited access to medical care were more likely to come to the NCI (interview, 2017).

Another problem was that tumors quickly developed drug resistance. It also became apparent that patients who were treated with one drug might become resistant to another – a phenomenon oncologists termed ‘cross resistance’. Patients in experimental trials were more likely to develop ‘cross resistances’, because they had already been treated with one of the first chemotherapies. This created selection bias: Patients who ended up in trials were not representative of the broader population of cancer patients because their tumors had been modified by previous treatments.

In face of these obstacles, it was crucial to understand how to wisely ‘use’ patients – how to best recruit, sort, and match patients and trials. These considerations turned researchers’ attention to the question of patient classification. In order to rationalize the enrollment of patients into clinical trials and to optimize drug testing, researchers suggested classifying patients according to their response to drugs. They hoped to generate new data that would enable them to predict which patients would respond to which treatments. They were about transform the field of diagnostic classifications in oncology.

New patient classifications

To standardize and optimize patient selection, oncologists began to insist on formal criteria of patient recruitment in protocols (Hall, 1962, p. 391). More importantly, they began to see drug testing as a tool for rewriting cancer classification:

In the early trials of a new agent the type and range of response to treatment are not predictable and all patients meeting the criteria for patient selection are treated. As data accumulate, it is necessary to sort out and classify those patients responsive and those unresponsive to treatment, those who tolerate an adequate course, and those who develop early or severe toxicity. … From the accumulation of data of this type on drugs that exhibit some degree of therapeutic activity, the clinical situations responsive to the treatment can be identified. … Patients whose clinical picture falls into the responsive pattern can then be selected to expand the study. (Karnofsky, 1962, p. 75)

Many reported differences in response rates between centers would be eliminated if investigators classified their therapeutic results … in relation to patterns and stages, rather than lumping clinical results together for a particular type of cancer. This does not seem to be too onerous a recommendation, despite the complaints of many clinical investigators. We readily classify apples to identify patterns such as Delicious, Winesap, and McIntosh, and stage apples as to whether they are green, ripe, or overripe. We even note if they have been artificially colored. Through the offices of the CCNSC, equally useful pattern and stage classifications for the major types of advanced cancer can be established by study and agreement, so that clinical trials can delineate these situations in which a patient responds in a predictable manner to a chemotherapy. (Karnofsky, 1962, p. 77).

As available diagnoses could not guide clinical decision making, researchers devised a new, reversed logic. Instead of using diagnostic data to guide their experimentation with new therapies, they used patients’ responses to therapy in order to classify them. Oncologists did not use drugs to see if a specific patient fell under an existing category, but rather to create new classificatory knowledge. The plan was to divide patients into responders and non-responders and to collect information that could potentially explain the difference between the two groups. Such information included morphology of cancer cells, age, sex, and numerous other variables. First, these variables were used for analyzing variation in response across these subgroups. In the next step, meaningful differences informed the assignment of different treatments for different groups. Karnofsky formulated a drug-focused, rather than a patient-focused goal: To make sure that no drug got patients who were unresponsive to it.

Oncologists hoped to accumulate data on ‘clinical patterns’ (i.e. histories of ‘treated diseases’) in order to create a new typology of untreated disease. By creating new ways to describe and examine treated cancer, oncologists ‘multiplied its ontological domains’ (Mol, 2003). They had transformed the meaning of existing categories (by linking them to new data), created new categories and eliminated others.

New cell morphology?

A canonical dataset used by hematologists prior to the 1960s was a collection of cases (572 children and 179 adults) compiled by Tivey in 1954 and titled ‘The natural history of untreated acute leukemia’. Although distinctions by cell type were present in the report, they only minimally informed the analysis. The reader was asked to bear in mind the poor state of cell morphology and was cautioned to treat cell-type data as mere approximations. Data on adult and children were pooled together, and attempts to breakdown the series did not establish conclusive differences between them.

Diagnosis is frequently incomplete. Criteria exist for the morphological differentiation of the several types of leukemia, but the application of these criteria is extremely difficult in the more acute forms, hence many authors report data simply as ‘acute leukemia’. (Tivey, 1954, p. 322)

While the diagnosis of leukemia was in disarray, advances in cytotoxic compounds stirred up the field:

The recorded history of effective palliative treatment of acute leukemia dates from the introduction of Aminopterin by Farber [in] 1948. Since then, clinical trials of a number of chemotherapeutic agents have been reported. … Some of these compounds have shown an apparent differential success in the treatment of the various morphologic types of leukemia, and even differences in the effectiveness of treatment of the same disease in the adult and the child, it will be necessary to consider the characteristic pattern of each disease in appropriate detail. (Tivey, 1954, p. 322 emphasis mine)

As chemotherapy was making great strides, it became clear ‘children are much more likely to respond to chemotherapeutic agents or to steroids than adults’ (Dameshek, 1962, p. 13). The question was why. Do children and adults suffer from ‘the same disease’? Or from different leukemias? To answer this question, researchers returned to the problem of cell morphology:

There is a fair degree of unanimity among authors … that the great majority of childhood acute leukemia belong to the lymphocytic type. … This blast cell differs morphologically from those found more commonly in adult acute leukemia, some of which are undoubtedly from the granulocytic variety. (Dameshek, 1962, p. 5)

Dameshek presented an early version of the hypothesis that children and adults suffer from different leukemias involving different cell types (ALL/AML). The difference in response to treatment reinforced the question of cell morphology, but in a renewed form: Do children and adults have different types of blasts? Dameshek had not yet made a distinction between ‘disease types’. Soon enough, however, the work of oncologists gave these differences ontological weight and operational meaning. The differences between leukemias began to matter.

Consider Figure 1, the appendix from of the second trial conducted by the Cancer and Leukemia Group B in 1957. Although all leukemic patients were pooled together in a single trial, it was suspected that age and cell type were both clinically significant variables. The statistical design, therefore, matched pairs of patients across three age categories and three categories of cell type. The residual categories of ‘unclassified cell type’ clearly indicates that cell morphology was being ‘worked out’ and that standardization was still underway.

OPEN IN VIEWER

Figure 1.

Leukemia classification, age, and cell classification.

While studies that grouped children and adults were still common (e.g. CALGB protocol 21 1962), oncologists increasingly targeted either children or adults and differentiated patients by cell type. It became less and less meaningful to pool together leukemic patients.

Consider the table from a study conducted by Carter et al. (1962) (See Table 1). The study focused exclusively on children and differentiated between no less than four morphological types (lymphocytic, monocytic, granulocytic, erythroleukemia). It compared three different regiments (6-MP at different doses or a combination of 6-MP and hormonal therapy) to test links between morphology and drug efficacy. As researchers began comparing the cell morphology of adults and children, they also compared cell morphology within samples of leukemic children. Such refined data on cell type would have been meaningless just a few years earlier. In the 1960s a new question became possible: Do patients with different types of leukemic cells respond differently to treatment? Cell morphology became a diagnostic tool for designing subsequent trials.

OPEN IN VIEWER

Table 1.

Classification of patients by morphological type of leukemia and drug response (Carter et al., 1962, p. 156).

Patient response according to specific drug regimen and type of leukemia
Type of leukemia	6-MP (2.5 mg/kg)		6-MP (6.6 mg/kg)		6-MP (2.5 mg/kg) + DON (0.25 mg/kg)
	Patient reviewed	Patients attaining complete or partial remission	Patient reviewed	Patients attaining complete or partial remission	Patient reviewed	Patients attaining complete or partial remission
Undifferentiated	31	18	40	20	35	29
Lymphocytes	15	11	14	11	18	17
Granulocytes	5	1	5	1	5	2

Soon after, age became a proxy for cell type, and childhood leukemia was identified with ALL. ⁷ In this context, the standardization of morphological classification became of urgent concern.

The difference between the various types of blasts may … no longer simply be of academic importance, for in these days of more or less specific chemicals for specific types of leukemias, it is often highly desirable to know the precise character of the type of leukemia which is to be treated. A number of more searching diagnostic procedures have therefore been introduced, many in recent years, which may be thought of as neo-morphology. (Dameshek, 1962, p. 6)

While Dameshek acknowledged the gap between what hematologists should and could see under a microscope, he announced that ‘a new diagnostic criteria has been made available’. Referring to ‘suspect’ cell organs (e.g. nuclei, mitochondria, and organelle) that might hold the key for cell differentiation, he targeted the entities that needed to be made visible in standard stains. Hematologists, he argued, were at the brink of developing ‘newer optical methods’ that would enable them to see the difference between primitive cancerous cells. Dameshek even listed the technologies that promised to help microscopes catch up with the clinic, among which were ‘supravital staining’, and time-lapse phase motion pictures (Dameshek, 1962, p. 7).

Ironically, Dameshek was both right and wrong in his prediction. None of the technologies he mentioned were the key to the standardization of hematological classifications in leukemia. Ambiguity was only eliminated with the advent of flow-cytometry and immunophenotyping of surface proteins. Dameshek was right, however, to celebrate the arrival of a new diagnostic criteria. Hematologists did learn to differentiate between blasts, but they did so using old optic methods. Responses-to-treatment were used to subdivide patients with different leukemias and thus to validated morphological distinction. Oncologists were able created a rich corpus of cell-staining knowledge that was solid enough to become part of cancer clinical trials (interviews with Joseph Bertino, 2017, Franco Muggia, 2017). Morphology still underlies laboratory work to today, alongside immunophenotyping (Keating & Cambrosio, 2003). Not all distinctions and cell types studied by researchers ended up in diagnostic tables. Monocytes, for example, were eventually subsumed under the category of myelogenous leukemia. But ALL, AML, CML, and CLL became the fundamental classificatory categories in onco-hematology.

Conclusion

In this article, I have examined the relations between diagnosis and therapy and demonstrated that the development of new drugs can transform the very way we classify and understand diseases. I have argued that therapy has transformative potential not only when it works, but when it works just well enough to mobilize support and motivate the restructuring of the tasks and problems of medicine. Drug development under conditions of uncertainty, scarcity of candidate drugs, under-organization, and jurisdictional pressures led oncologists to develop a very specific practical orientation. Oncologists moved away from rigid drug testing and instead reinvented what it meant for drugs to ‘work’, carefully manufacturing the conditions under which drugs could ‘work’.

The existing literature has mostly discussed clinical trials as mechanisms for political legitimation used to adjudicate competing claims about efficacy and harm (e.g. Marks, 1997). Building on the work of Greene (2007) and Keating and Cambrosio (2011), this article demonstrates that clinical trials can also function as experimental spaces that are conducive for developing radically innovative practices.

To investigate the experimental function of clinical trials, I do not focus exclusively on technological and scientific innovation but shift the analytic focus onto the normative entrepreneurship of oncologists. Therapy, by definition, links technologies with normative and ethical concerns (Jasanoff, 2004). The case of leukemia chemotherapy trials shows that oncologists’ epistemic and scientific innovation cannot be understood independently from their bureaucratic, organizational, and normative ingenuity.

The concept ‘adequate trials’ is relevant not only for biomedicine, but connects to other technological fields. I’ll briefly discuss its relevance to AI and consider the case of autonomous vehicles. We can note one meaningful similarity between the history of AI and the development of chemotherapy; a deep commitment to unlock the potential of machine intelligence existed side by side with fundamental obstacles for making machines produce ‘human-like’ performance (Collins, 1990; Suchman, 1987). The persistent difficulties in making machines ‘speak’, ‘see’, or ‘drive’ meant that these tasks had to be redefined and the original goals reformulated, in order to discover in what ways machines could be made to ‘work’. Ensmenger (2012), for example, shows how these modifications took place even in a field that was specifically chosen because it was easy to formalize, namely chess.

Marres’s research on autonomous vehicles provides another productive example. In their work on tests and testing, Marres and Stark (2020) argue that contemporary technologies force us to rethink our notions of scientific experimentation; more and more technological innovations are not developed and tested in the isolated and controlled site of the laboratory. Innovative processes increasingly necessitate tests that take place within social settings. Consequently, they argue, society itself is being increasingly manipulated in tests that are both scientific and social experiments. In her work on street trials of autonomous vehicles, Marres (2020) provides a convincing case in point. In order to train the algorithms and computational systems of driverless cars, such vehicles must be put on the streets and made to interact with ‘real traffic’. The placing of autonomous vehicles on the streets transforms the routines orchestrating the shared use of public roads (some road users are denied access to certain areas, some lanes are blocked, traffic controllers in yellow vests appear on the scene, etc.). Just as machines are socialized to interact with humans, the trial socializes various road users to interact with autonomous vehicles. Cancer chemotherapy had to create a new notion of a drug and new disease categories. Street trials, if successful, will transform our notions concerning the use of public space and the respective rights and responsibilities of its users.

The case of leukemia is a story of diagnostic specification – oncologists split up a two-cell grid into four. This story fits a bit too well with the popular exception that with scientific progress, diagnoses will become increasingly narrow and true-to-nature. The analysis of diagnoses and therapy, however, can capture quite the opposite trend. The search for therapy leads experts to lump diagnostic categories together just as it leads them to split them up. The case of leukemia itself displays such reversal. As soon as they established the distinction between ALL/CLL/AML/CML, oncologists created the category of ‘blood cancers’ by lumping together leukemias and lymphomas. Such a category was necessary in order to transpose practices and models from treatment of leukemia to treatment of lymphoma and solid tumors (interview with George Canellos, 2017). Leukemia became a model for other cancers and was ‘put together’ with lymphomas under the category of ‘liquid cancers’, both in theory and in practice.

In an era of ‘precision medicine’, when cancer classification is undergoing intense fragmentation with critical costs for distributive justice, it is especially important to push back against deterministic approaches to diagnoses and to reground medical classification in the institutions and practices of therapy.