Is cancer a natural kind? On the one hand, the question is whether we are right to split cancer into the categories we use. According to Plutynski (([2018]), cancer nosology yields “a multimodal and cross-cutting family of classificatory schemes” which seems to warrant “pluralist realism” about cancer. On the other hand, the question is whether our concept of ‘cancer’ simply lumps together facts according to human interests, which does not allow for useful generalizations. Cancer anthropocentrism is challenged by a rising approach in oncology, namely comparative oncology, which investigates cancer in all species (e.g., (Aktipis et al. [2015]; Schiffman and Breen [2015]; Albuquerque et al. [2018]). According to its proponents, a major advantage of this approach is that it frees us from human practical biases and yields better generalizations about cancer (Aktipis [2020]), indeed even leading to a “universal theory of cancer biology” (Dujon et al. [2021]). What can we hope to learn from the nonanthropocentric approach of comparative oncology? Although comparative oncology can challenge our category of cancer, it is itself fraught with the problem of which phenomena should count as cancer. Some researchers are very inclusive and take any tumor to be cancer, including in invertebrates and plants; others prefer to limit the category to what can invade other tissues and metastasize. Criteria and extension of cancer are clearly interdependent, which often gives the impression of a certain arbitrariness in comparative oncology (what people find is a direct reflection of the definition of cancer they started with). To break this circle, we argue that it is better to embrace a provisional and heuristic anthropocentrism, which begins with hypothetical generalizations about human cancers. Human cancers may be special, but they are the best-known cancers. Precise hypotheses relative to these generalizations can then be tested in other species. We propose in particular to identify “comparative paradoxes”, i.e., claims about cancer that should hold in various species, but actually do not hold. For instance, Peto’s paradox is the puzzle of why large organisms don’t get cancer more often than small ones. The implicit generalization is that disordered cells should be proportional to the number of cells, given that cancer is caused by the accumulation of random mutations in individual cells, but this is precisely what observation contradicts (Abegglen et al., 2015). Whenever this generalization does not hold and the explanation does not reside in the fact that some animals have evolved highly specific anticancer mechanisms, it has the potential to challenge our most general conceptions of cancer. In addition to Peto’s paradox, we will discuss two other paradoxes: the connection between cancer and longevity and the immunological control over cancer. Prominent authors in comparative oncology have argued that the main advantage of this approach is to reconceptualize cancer as a much broader phenomenon: cancer is best understood as a form of “cheating” (Aktipis et al 2015). Instead, we claim that its main advantage is to assess and revise our most entrenched convictions about how cancer works.

Cancer is often conceptualized in terms of selective conflict between cell and organism (Greaves 2015, Aktipis 2020). On this view, cancer involves a form of multi-level selection in which the cancerous cell phenotype is favored by selection at the cell level but opposed by selection at the organism level. Recently, Gardner (2015) and Shpak and Lu (2016) have argued that cancer is not a true case of multilevel selection, because cancer is an evolutionary dead-end. I argue that this “evolutionary dead-end” argument is powerful but not decisive.

Cancer cells keep accumulating alterations leading to a diversification through space and time. This diversity in the composition of cancer cells represents a major challenge for cancer treatment as it is difficult (if possible, at all) to find a treatment that works on all the cancer cells. The clonal evolution model introduces evolutionary tools in oncology to make sense of this evolution of the cancer cells. In this model cancer cells are regrouped in clones—populations of cells that share a common identity (traditionally a common set of driver mutations) inherited from a common ancestor cell. This allows us to reconstruct the evolutionary history of the tumour (and metastases), and to track its evolution through time. This has, for example, allowed researchers to identify mutations involved in resistance to targeted therapies (e.g., EGFRT790M induces resistance to first generation EGFR inhibitors). But reconstruction of the clonal evolution is far from an easy task, both pragmatically and conceptually. The conceptual issue can be easily grasped by just indicating that the two main characteristics of clones —genealogy and identity— pull in opposite directions. Genealogically speaking, all cancer cells have a common ancestor, the first transformed cell, so each cancer is one big clone. But cancer cells of a given cancer are all unique. Thus, regrouping cancer cells into clones requires making a choice with regards to which criterion to use. Traditionally, the choice is to regroup cells according to their driver mutations, as these are conceived as the only mutations that impact cells’ properties, and they are easily tractable including in clinics. We take issue with this choice. In this talk, we will first deconstruct the notion of clone in oncology, highlighting that it relies on the following dubious assumptions: (1) driver mutations can be distinguished from passengers; (2) driver mutations provide a good proxy of cancer cell phenotype; (3) intraclonal heterogeneity can be ignored. This will lead us to argue that the notion of clone must be revised. Second, we will argue in favour of a change in the understanding of clonal identity. Our suggestion is to regroup cells according to their similarities, distinguishing clonal (lineage-dependent) similarities, from non-clonal (lineage-independent) similarities (e.g., similarities that are stochastic or induced by phenotypic plasticity). Both types of similarities can contribute to explaining cancer cells properties, such as response to treatment. But only the former can contribute to clonal evolution as whatever causes the similarity is inherited by descendant cells of that lineage. Third, we will explore the benefit of this conceptual turn. It opens new research programs on how to analyse the evolutionary dynamics in cancer cells, experimentally and computationally. We will show the first results of an original experimental set-up we have developed to focus on the inheritance of functional properties, with no prior assumption regarding what exactly causes the observed lineage-dependent similarities.

Cancer researchers and clinicians speak of both cancer “drivers” and “actionable” mutations. In this paper, we explore how these two concepts are overlapping, and how they are different. Cases like the BCR-ABL1 gene fusion found in people with chronic myelogenous leukemia (CML) have served as exemplars in clinical teaching and research about the value of cancer genomics cancer diagnosis and treatment, but we argue that there are good reasons to think that the CML case is exceptional. With the completion of the cancer genome atlas project (TCGA) there is a growing realization that there are many more “drivers” than anticipated, placing an ever-larger wedge between the notion of “drivers” and “actionable” genes, in ways that have shifted the conversation about the relevance of cancer genomic data to diagnosis, prognosis and treatment. Clinicians now require a more fine-grained, contextual, and hierarchical ranking of significant variants for cancer diagnosis and treatment. We document here the shifts in the presuppositions driving the use of AI and genomic data in cancer diagnosis. We delineate the different ways that variants can be used in clinical activities and explain how this maps on to the distinction between "actionable" vs. "driver" mutations. For instance, the “driver” concept initially emerged in cases where molecular features of particular cancers were well-characterized, such as CML. In this case, a specific mutation provided important clinical information. However, the concept has since expanded to cover a broader set of genes found to be recurrently mutated in specific cancers using “Big Data” and AI approaches. Identification of “driver” mutations in this manner led to the splitting off of “driver” from the concept of “actionable” mutations. The latter refers to a subset of mutations which serve as biomarkers for particular treatments. While these concepts overlap in certain cancers, in others, it is crucial to keep them distinct. For instance, in molecularly heterogenous diseases, such as myelodysplastic syndromes (MDS) and acute myeloid leukemia (AML), it is very important to not conflate them. Although genetic risk stratification often guides treatment decisions, variants in such models are not “actionable” in the sense of being specific treatment targets. This debate over how to demarcate “drivers,” versus “actionable” mutations is tied to a larger debate about the proper role of AI in biomedicine. The use of AI to identify “driver” genes does a dual service: on the one hand, it provides at best correlative, predictive information; on the other, it also indicates a potential causal role. The concept of “actionable” mutations attempts to move beyond the correlative. In this way, the trajectory of cancer research aims to move from identifying “drivers” to distinguishing “actionable” mutations. AI approaches may tell us little as yet about the specific causal role they play, or whether we might expect to successfully intervene their downstream products or associated pathways, raising questions regarding the scope and limits of these methods in translational cancer research.