The Missing Variability Problem in Cancer Prediction: Integrating Multilevel Factors

Abstract Efforts to predict individual cancer risk typically rely on genomic data together with environmental factors external to the organism. Yet even when these domains are considered jointly, a substantial portion of cancer risk variability remains unaccounted. This persistent predictive gap is labeled here as the missing variability problem (MVP). Beyond further refinement of genomic and external environmental factors, existing approaches seek additional sources of variation either by downscaling to molecular processes or by upscaling to tissue-level organization. While these strategies identify relevant factors of variation, they leave open how to jointly use these factors for prediction across levels. This article proposes treating the cellular level as an integrative predictive level at which heterogeneous factors of variation can be coordinated within a single biological bearer. Given the fragmented state of multilevel cancer theory, this article advances a predictive strategy that complements the search for additional factors by asking how risk-relevant factors can be organized once they are identified. The detection of cellular dispositional properties is introduced here as an integrative predictive strategy grounded at this level: cells are characterized by relatively stable differences in how they tend to produce variability across conditions. These differences can be operationalized through measurable features, including levels of molecular noise and patterns of response to tissue-level influences. Framed in this way, cellular dispositions specify how multilevel predictors may become jointly usable. Similar content being viewed by others Introduction Efforts to predict cancer risk have significantly shaped both data-intensive research practices and conceptual frameworks in contemporary biology. Cancer research was among the first domains to move from individual genes to genomic risk scores, and later to the systematic inclusion of environmental factors through the concept of the exposome (Merlin and Giroux 2024). Yet even today, integrating genomic and environmental data accounts for only a limited portion of individual variability in cancer risk (Tomasetti et al. 2017; Plutynski 2021a; Teschendorff 2024). The relevant difficulty is individual-level assessment: genomic and environmental predictors may stratify populations while still imperfectly discriminating the risk of a given individual. In a pan-cancer analysis, Kachuri et al. (2020) added cancer-specific polygenic risk scores to models already including family history and modifiable environmental risk factors. Prediction improved for most cancers, but individual-level discrimination remained moderate, with concordance-index values of 0.635 for breast cancer, 0.664 for melanoma, and 0.716 for colorectal cancer. In simple terms, these models assigned higher predicted risk to the individual who developed cancer earlier in about 64–72% of cases. Thus, even when inherited susceptibility and available environmental and behavioral predictors are used in the same model, individual risk remains imperfectly discriminated. This persistent predictive gap is often treated as a temporary limitation of measurement or data coverage. Here it is approached differently. The gap reflects a more basic difficulty: the problem of integrating predictive factors that are distributed across different levels of biological organization. I refer to this difficulty as the missing variability problem (MVP)—the inability to reliably predict individual cancer risk on the basis of genomic and environmental data. The MVP is a predictive, rather than an explanatory, problem. It bears a limited parallel to the missing heritability problem, usually understood as the gap between high heritability estimates for traits, for example disease risk and behavioral traits, derived from family and twin studies, and the much lower heritability estimated from molecular genome-wide association studies (Matthews and Turkheimer 2022). Matthews and Turkheimer also distinguish several formulations of this problem, including a predictive interpretation. The present parallel concerns only that interpretation: in both cases, known sources of variation remain insufficient for reliable individual-level prediction. The comparison is limited. In the missing heritability problem, molecular genetic predictors do not match the heritability suggested by family and twin studies. In the MVP, substantial cancer-risk variability remains insufficiently predicted even when genomic and environmental predictors are considered together. The MVP appears here as a persistent gap in the prediction of individual cancer risk. One influential way of articulating this gap is provided by the “bad luck” thesis, which situates the remaining unpredictability at the cellular level by treating stochastic events during cell division as a major source of risk. On this account, roughly two-thirds of the variation in cancer risk among tissues is attributed to random mutations arising during DNA replication in normal stem cells, rather than to environmental factors or inherited predispositions (Tomasetti and Vogelstein 2015). Dissatisfaction with this residual indeterminacy motivated efforts to further specify cellular variability by tracing influences below and above the cellular level. These efforts take the form of downscaling and upscaling strategies, which search for additional molecular and tissue-level factors shaping cellular states. These strategies may reduce the predictive gap but remain limited when pursued separately. An integrative task will also arise for any still unidentified factors. This article proposes the cellular level as a site of such integration, because molecular, tissue-level, and environmental influences can become jointly attributable to cells or their lineages. The MVP motivates a shift toward multilevel perspectives on cancer. Predictive factors relevant to cancer risk span molecular, cellular, tissue-level, and organismal processes, as cancer is widely described as a multilevel phenomenon (Green 2021). At the same time, there is no broadly accepted multilevel theory that integrates these processes into a single predictive framework. Philosophical analyses emphasize that existing theoretical approaches rely on distinct ontological and methodological commitments and resist straightforward integration. Cancer-related processes are also processual and context-dependent, which further limits the prospects for a stable hierarchical theory capable of grounding prediction across scales (Bertolaso 2016; Plutynski 2018). Multilevel descriptions therefore tend to coexist, rather than converge. In practice, efforts to close the predictive gap in cancer risk assessment have proceeded by identifying and relocating sources of variability across levels of biological organization. Downscaling approaches seek predictive leverage in molecular sources of variability (Han et al. 2016; Biondo et al. 2023; Teschendorff 2024). Upscaling strategies emphasize tissue-level organization and developmental conditions (Nelson and Bissell 2006; Sonnenschein and Soto 2020; Safaei et al. 2023). Both strategies capture important aspects of carcinogenesis. They also continue to identify further risk-relevant factors as recent work on the microbiota, nervous system, and tumor microenvironment illustrates (Mancusi and Monje 2023; Laplane 2025). Yet factor discovery alone does not specify how such factors should be coordinated with other predictors for risk assessment. Molecular, cellular, tissue-level, and environmental factors may each be relevant, but they do not automatically form a predictive framework when listed side by side. The difficulty addressed in this article is therefore how factors operating at different levels can be organized for risk assessment when they are registered in cellular states, responses, and lineages. A conceptual response to this difficulty is to identify a level of organization at which influences from different levels of organization become jointly tractable as relatively stable features of the same biological unit. Such a level can be described as an integrative predictive level: a level at which multilevel determinants relevant to risk are coordinated into relatively stable organizational states, without presupposing causal completeness or reduction to a single scale. The cell plays this role not because cancer is fundamentally a cellular disease, but because predictive factors shaping variability become accessible at the cellular level. Theoretically, this aligns with dispositional accounts of stemness that treat cells as bearers of lineage-level capacities persisting across divisions (Laplane 2016; Laplane et al. 2019; Suárez 2023). On such accounts, cancer stemness is not a fixed cellular category but a context-sensitive dispositional property, expressed in the capacity of premalignant and malignant cells to generate variable lineages across conditions. Interpreted in this way, cancer-related risk factors—molecular instability and tissue-level factors—can be understood as integrated through cellular dispositions, allowing multilevel predictors to be coordinated at the cellular level for risk assessment. The Missing Variability Problem (MVP) in Cancer Risk Prediction The MVP designates a persistent predictive gap in cancer risk assessment that remains when genomic and environmental factors are jointly considered. In cancer risk prediction, well-known limits persist at the level of genomic prediction and remain after genomic and environmental factors are integrated. The existence of this residual gap becomes especially salient in debates surrounding the “bad luck” thesis, which treats a large fraction of cancer risk as irreducible stochasticity at the cellular level. Attributing this residual variability to the malignant transformation of stem cells identifies a locus of uncertainty, but it does not resolve the predictive problem or indicate how prediction might be improved (Plutynski 2021a). An analogous predictive difficulty appears in microbiology, where genetically identical cells display phenotypic differences under tightly controlled environmental conditions, and this variability is commonly classified as spontaneous rather than rendered predictive (Casali and Merlin 2020; Casali et al. 2025). Here, “spontaneous” does not merely name an explanatory placeholder. It also registers a predictive limitation, namely, the difficulty of identifying stable predictors for when and how such differences emerge within otherwise controlled lineages, without assuming that genetic identity and environmental control exclude molecular noise, epigenetic change, or lineage-dependent cellular states. Work on spontaneous epigenetic variation makes this point concrete by tracking changes that accumulate over divisions while leaving their proximate triggers unresolved. These cases do not provide a direct model for cancer prediction, but they clarify the general form of the problem. In cancer research, this gap becomes visible when the limits of existing predictive approaches are considered in relation to observed variation in cancer risk. One way in which this gap manifests is through the limited predictive power of genomic data. Genomic assays can reveal individual differences in cancer susceptibility based on mutational load and mutation signatures (Alexandrov et al. 2020). Yet only a small number of genetic variants substantially increase cancer risk and can be used for prediction, with mutations in the BRCA genes (BReast CAncer genes 1 and 2) being a canonical example. Whole-genome sequencing further complicated prediction by raising the problem of distinguishing malignant from passenger mutations, many of which are present in cancer cells without contributing to tumor development. As a result, the identification of stable genomic predictors proved far more difficult than initially anticipated (Plutynski 2021b; Pradeu et al. 2023). Across cancer types, polygenic risk scores therefore show modest predictive power, with clinically significant risk increases confined to a small fraction of individuals (Jia et al. 2020). A parallel limitation concerns environmental predictors when considered alongside genomic information. The concept of exposome was introduced to capture environmental contributions to disease risk that escape genomic analysis (Wild 2005). The expectation was not merely to supplement genomic prediction, but to render residual variation predictable through the joint integration of genomic and exposomic data. These expectations were only partially fulfilled (Canali 2019; Merlin and Giroux 2024). Even retrospective analyses fail to account for a substantial proportion of cancer cases through hereditary or environmental factors alone (Wishart 2022). This limitation is not solely a matter of incomplete measurement. It also reflects difficulties in stabilizing environmental predictors across contexts and in specifying how such predictors relate to cellular and tissue-level processes relevant to cancer risk. The predictive gap that persists after the joint consideration of genomic and environmental factors becomes particularly visible in debates surrounding the “bad luck” thesis proposed by Tomasetti and Vogelstein (2015). Drawing on multistep models of carcinogenesis (Fearon and Vogelstein 1990), Tomasetti and Vogelstein argued that roughly two-thirds of the variation in cancer risk can be attributed to stochastic mutational events during stem cell division, with the remaining fraction explained by genetic predisposition and environmental exposure. On this view, the predictive gap is real but largely unbridgeable: beyond counting stem cell divisions, little more can be done to improve prediction. This conclusion has been widely challenged, not only on conceptual grounds but also in light of broader incidence patterns (Plutynski 2021a). If cancer risk were primarily determined by the number of stem cell divisions, it should scale with body size and lifespan. The absence of such scaling across species—captured by Peto’s paradox—shows that cancer risk does not track simple cellular parameters in the expected way (Peto 2016). The paradox highlights a mismatch between population-level regularities and predictions derived from single-level statistical models. Alongside the attribution of residual risk to stochastic events during cell division, researchers increasingly investigate additional sources of cellular variability that are not captured by genomic variation or organism-level environmental exposure. Some of this work focuses on molecular and intracellular processes, including regulatory dynamics and epigenetic change (Han et al. 2016; Biondo et al. 2023; Teschendorff 2024), while other studies emphasize tissue organization and developmental context (Nelson and Bissell 2006; Sonnenschein and Soto 2020; Safaei et al. 2023). These approaches expand the range of factors considered relevant to cancer risk, yet they are typically pursued in parallel and remain only loosely connected. Their mismatch points to a more general difficulty in cancer risk prediction, captured by MVP: the coordination of predictors operating at different levels of biological organization. The problem is not only a shortage of data. Additional data may identify further factors, refine known predictors, or reveal new interactions. Yet such factors still require a framework in which their joint predictive relevance can be articulated across levels of biological organization. Existing approaches tend to address the predictive gap by downscaling toward molecular dynamics or upscaling toward tissue-level organization. While both strategies introduce relevant factors, neither provides an agreed way of integrating them within cancer risk prediction. The MVP concerns limits of predictive coordination across levels of biological organization in cancer risk prediction. Newly identified factors, including microorganisms, retrotransposons, or mitochondrial exchange, may reduce parts of this gap (Laplane 2025). Yet their identification also raises the question of how such factors should be related to other predictors for risk assessment. Integrative Predictive Level and Cancer Risk Prediction Classical accounts of levels of organization emphasize how biological phenomena are distributed across hierarchically related domains, often highlighting the stabilizing role of higher levels and their capacity to constrain or modulate lower-level variability (Wimsatt 1994; Eronen and Brooks 2018). From emergentist perspectives, transitions between levels are also associated with discontinuities and surprises, including cases in which higher-level organization induces novel patterns of variation at lower levels (Needham 1937; Peterson 2016). In predictive contexts, variation is not confined to a single level, but may cascade across levels, with phenotypic properties serving as points of manifestation for influences originating above or below the focal scale (McEntire et al. 2021). These accounts clarify why predicting biological variation often requires attention to multiple levels rather than exclusive focus on any single domain. At the same time, treating levels of organization primarily as sites of causal production or stabilization does not by itself resolve how predictive problems are structured in practice. In research oriented toward risk assessment, the relevant question is not only how processes are distributed across levels, but where heterogeneous influences become jointly tractable for forward-looking inference. Focusing on prediction rather than representation allows this account to set aside debates about whether cancer should be framed primarily as a disease of genes, cells, tissues, or organisms. Instead, levels of organization are approached in terms of their epistemic function in structuring scientific problems by locating predictive tasks at a given level, partitioning them across levels, or coordinating relations between levels (Eronen and Brooks 2018). Different predictive perspectives thus shift attention between levels by foregrounding some descriptions while pushing others into the background (Hochstein 2022). This orientation is also consistent with accounts of data integration that emphasize its situated, purpose-relative character in contemporary biology (Leonelli 2016). Building on this view, the notion of an integrative predictive level can be developed. An integrative predictive level designates a level of biological organization at which heterogeneous processes—spanning molecular, cellular, and tissue domains, among others—become jointly recognizable in ways relevant for prospective risk assessment, without presupposing causal completeness or reduction to a single scale. In contrast to classical discussions of levels that focus on part–whole relations or hierarchical ordering, this notion foregrounds the coordination of predictive factors under conditions of causal entanglement. The integrative predictive level is therefore not an alternative level of organization, but a way of specifying how levels are mobilized to support stable and comparable prognosis in the face of unresolved complexity. Integration here is not achieved by statistically combining predictors across levels. Rather, heterogeneous influences are coordinated in ways that are recognizable in the material organization of a given level, where influences originating above and below that level become jointly expressed as properties of the same organized unit. For instance, at the cellular level, this coordination becomes tractable through phenotypically detectable configurations—such as morphology, chromatin organization, and reproducible response patterns to tissue cues—through which molecular and tissue-level factors become jointly visible for predictive assessment. From a predictive perspective, however, not every level at which variability is detected is equally suited to function as an integrative predictive level for addressing the MVP. To play this role, a level needs to support three interrelated requirements. First, it should allow heterogeneous sources of variability—originating from both lower- and higher-level processes—to be jointly articulated within a single predictive frame (integration). Second, the predictive factors identified at that level should be characterizable in terms that do not collapse into level-specific parameters but retain their significance when relating this level to others (interpretability). Third, the level should enable a stable attribution of predictive properties to a determinate biological unit, such that variability can be referred to a bearer that is itself relevant for prediction (attribution). Taken together, these requirements do not single out a privileged level of organization. Rather, they outline what is practically required for any level to function as an integrative predictive level. The cellular level becomes relevant at this point not as a default explanatory unit, but as a plausible candidate at which these requirements can be jointly met. The search for further cancer-relevant factors can be accommodated within this framework. Cancer neuroscience is a clear case: neural activity and neuron–tumor interactions have become important sources of variation in glioma research, and their effects are often analyzed at the level of cellular populations and tumor niches (Mancusi and Monje 2023). Yet their predictive relevance depends on how these influences are registered in particular tumor cells or cell lineages, for example as altered responsiveness, proliferative behavior, invasive capacity, or persistence over time. Cellular populations and tumor niches organize the distribution and persistence of such effects. The cell or cell lineage remains the bearer to which higher-level influences can be attributed as changes in state, response, and future variability. What distinguishes the cellular level in predictive practice is not explanatory primacy, but the possibility of bringing heterogeneous predictors into the same observable and temporally trackable unit. Molecular processes are often difficult to treat as stable bearers of prediction, while tissue- and population-level descriptions usually capture distributions, constraints, or contexts rather than the unit in which state, response, and future variability are jointly registered. Molecular alterations, epigenetic states, and responses to tissue-level conditions can thus be registered in the same cell or cell lineage and compared across changing conditions in predictive practice. Treating the cell as an integrative predictive level does not commit this account to a view of cancer as a fundamentally cellular process, nor to the claim that malignant progression can be exhaustively explained at the level of individual cells. Cancer unfolds across interacting molecular, cellular, tissue, and organismal processes, none of which can be straightforwardly privileged as the sole locus of causation (Green 2021). This multi-level character is fully acknowledged here. The rationale for focusing on the cellular level is instead practical: it concerns where heterogeneous influences become jointly recognizable in forms suitable for prediction. Consistent with pluralist accounts that emphasize the locality of explanatory and predictive hierarchies in biology (Love 2012), the cellular level functions neither as the exclusive unit of carcinogenesis nor as a substitute for tissue- or organism-level accounts. Cellular populations are a particularly important case, since many predictive questions concern the distribution, expansion, or stabilization of cellular states within tumors. Clonal approaches make this point clear: a clone is a population of cells organized by descent from a single cell and used to capture tumor heterogeneity relevant to development and treatment (Laplane 2025). The integrative task is therefore more precise: population-level patterns become informative insofar as they involve cellular states that can be attributed to individual cells, followed across lineages, and compared across clonal or niche contexts. The cellular level thus serves as a site at which molecular alterations, regulatory disturbances, and contextual influences are rendered visible as comparatively stable organizational configurations. This emphasis on the cellular level should also be distinguished from evolutionary and atavistic narratives that portray cancer as a reversion to unicellularity or as the reactivation of ancestral single-celled programs (Nedelcu 2020). While such perspectives have drawn attention to important features of cancer-related traits, they risk obscuring the fact that multicellularity emerged through the reorganization and redeployment of preexisting cellular capacities rather than their simple suppression, and that selection-relevant dynamics can involve interacting levels rather than a single privileged unit (Laplane et al. 2025). Cancer cells are therefore best understood not as completely autonomous unicellular agents, but as defective and context-dependent entities embedded in higher-level organizational regimes. From this perspective, the cell is a non-accidental object of inquiry: it is the level at which diverse biochemical, regulatory, and structural processes are coordinated into organized states that condition higher-level patterns (Newman 2012). Focusing on the cell as an integrative predictive level accommodates multi-level causation while clarifying how risk-relevant dispositions can be apprehended without reducing prediction to molecular markers or presupposing cellular primacy. This provides a basis for specifying the conditions under which such coordination can support prediction. An integrative predictive level thus is characterized by the integration of heterogeneous sources of variability, the interpretability of predictive factors beyond level-specific parametrization, and the attribution of these factors to a determinate carrier at that level. How these conditions are approached depends on how predictive attention is redistributed relative to the cell, either toward molecular processes or toward higher-level organizational influences. Downscaling: The Molecular Noise Perspective Shifting predictive attention below the cellular scale has made molecular noise an attractive candidate for addressing the MVP. Molecular noise has long been regarded as a significant source of biological variability and therefore warrants careful consideration in this context. It captures forms of variation that are not reducible to genetic differences and that persist even under controlled environmental conditions. This point was emphasized early on by Jacques Monod and, more explicitly, by Richard Lewontin, who used developmental asymmetries in Drosophila melanogaster to argue against genetic determinism (Lewontin and Levins 2000; Lewontin 2000). In Lewontin’s canonical example, genetically identical individuals exhibit asymmetric numbers of sensory bristles beneath their wings: the bilateral symmetry of a single organism breaks down because small stochastic fluctuations in messenger RNA abundance during development determine whether particular cells give rise to bristles. The example illustrates how molecular noise undermines the predictive sufficiency of genome–environment accounts and motivates attention to intrinsic sources of variability. Contemporary biology has extended this insight by treating noise as a pervasive feature of gene expression within genetically identical cell populations. Variability in molecular abundances, conformational states, and reaction kinetics is now widely recognized as contributing to phenotypic heterogeneity, including in cancer-related processes (Jia et al. 2017; Feinberg and Levchenko 2023). Recent work on noise and mechanisms emphasizes that such variability should be understood as arising from the organized functioning of biological systems rather than as a mere by-product of imperfect control or measurement (Tee 2025). Noise, on this view, is not external to mechanisms but emerges from how molecular processes are coordinated and constrained within living systems. In predictive practice, however, molecular noise is typically operationalized in a fragmented manner. Tomasetti and Vogelstein, for example, incorporate noise into their predictive framework primarily as a correction factor accounting for deviations from expected mutation-driven cancer incidence (Tomasetti et al. 2017). Similarly, gene expression noise has long been studied as a prognostic factor, and cancer cells often exhibit elevated noise levels (Zajchowski et al. 2001). Yet many studies linking noise to cancer risk focus narrowly on oscillations in specific genes or proteins—p53 being a prominent case—while treating other sources of stochasticity as background (Han et al. 2016). In such approaches, different stochastic processes are examined in isolation, and “noise” designates a collection of partially independent fluctuations rather than an integrated predictive feature. As a result, noise in cancer prediction is addressed as a set of partially isolated stochastic processes rather than as an integrated predictive feature; although measured at the molecular level, it is not meaningfully attributed to molecules themselves, since individual molecular events do not constitute stable bearers of predictive relevance. These limitations can be addressed at the cellular level. Although noise is detected through molecular measurements, it often becomes predictively relevant when its effects are registered in cellular states associated with lineage-level variability. Molecular noise correlates with other cellular characteristics, including mitochondrial state and chromatin organization, and has been interpreted as an inheritable and evolvable cellular trait (Fraser and Kærn 2009). Treated in this way, noise acquires predictive relevance beyond isolated molecular events. Patterns of noise within individual cells provide insight into epigenetic differentiation pathways and into external influences acting upon them (Biondo et al. 2023). In this sense, heterogeneous stochastic processes are coordinated within a single cellular configuration, allowing predictors of variability to be identified and attributed to a determinate bearer. Upscaling: The Tissue Perspective While the downscaling perspective seeks predictive leverage below the cellular scale, the upscaling perspective shifts attention upward, toward tissue- and organism-level organization. Traditionally, cancer epidemiology treated environmental influences at the level of the organism, while locating carcinogenesis at the cellular level. The expansion of the exposome concept exposed the limits of this division. In its extended form, the exposome includes not only external exposures such as carcinogens and radiation, but also nongenetic processes unfolding at tissue and organismal levels (Merlin and Giroux 2024). This reconceptualization renewed interest in tissue organization and tumor microenvironments as contributors to cancer risk. These developments emerged partly as a response to the geneticization and inframolecularization of cancer research, and to what has been described as the “tissue agnosticism” of early twenty-first-century oncology (Campbell et al. 2024). A purely informational view of cancer, focused on DNA or RNA sequence alterations, proved insufficient. Given the frequency of potentially oncogenic mutations in somatic cells, cancer should occur far more often and at earlier ages than it does. This mismatch motivated renewed attention to the regulatory role of tissue organization, including the capacity of healthy tissue contexts to suppress or even reverse malignant phenotypes (Nelson and Bissell 2006). The most systematic articulation of this perspective is provided by the tissue organization field theory (TOFT). On this view, cancer is primarily a disorder of tissue organization rather than a cellular pathology (Sonnenschein and Soto 2020). Disruptions of tissue structure compromise the constraints that normally regulate proliferation and motility, thereby enabling malignant transformation. TOFT has been influential in highlighting morphogenetic fields, extracellular matrix structure, and long-range cell-cell interactions as central to carcinogenesis. From a predictive perspective, however, tissue-level approaches face a characteristic limitation. Although tissue organization exerts strong regulatory influence, it does not fully integrate or stabilize the cellular properties that are decisive for future variability. Tissue parameters often have differential and sometimes opposing effects depending on cellular state. A well-studied example is extracellular matrix stiffness. Mechanical forces shape stem cell proliferation and differentiation, but increased stiffness inhibits division in normal stem cells while promoting proliferation and heterogeneity in cancer stem cells (Militello and Bertolaso 2022; Safaei et al. 2023). The same tissue-level parameter therefore acquires different predictive meanings depending on cellular phenotype. Related difficulties arise in the coordination of tumor cells. Tissue organization does not fully predefine cellular roles within tumors. Genetically distinct cancer clones can cooperate in ways that are not predictable from tissue structure alone. Exchange relationships, such as trade-offs between growth factor production and invasive capacity, emerge dynamically, even when normal progenitors were not patterned into such roles by tissue organization (Carneiro et al. 2023). These cases indicate that tissue structure shapes, but does not determine, the distribution of cancer-relevant cellular properties. Moreover, the regulatory capacity of tissue is not absolute. In rare but well-documented clinical cases, foreign cancer cells have induced malignancy in human hosts, extending beyond the canonical examples of transmissible cancers in dogs and Tasmanian devils (Rebbeck et al. 2009). Such cases underscore that tissue context alone cannot guarantee suppression of malignancy, and that cancer-relevant properties can remain effective across tissue boundaries. From the perspective of integrative predictive levels, these observations point to a specific limitation of upscaling strategies. Tissue-level descriptions capture relational constraints, but they do not fully satisfy the requirements of integration and attribution. Tissue parameters integrate heterogeneous interactions, but their predictive significance often remains opaque without reference to cellular properties. According to the TOFT, early alterations of tissue structures are associated with later cancer risk; however, these alterations remain difficult to identify and track prospectively. This complicates treating tissue organization itself as a determinate bearer of integrated predictive properties. Tissue influences matter for prediction insofar as they are mediated through cellular states that persist, vary, and propagate across cell lineages. This returns predictive attention to the cellular level. Not by denying the importance of tissue organization, but by identifying the cell as the level at which tissue influences are integrated with lower-level properties in forms suitable for risk assessment. The Cell as a Locus of Predictive Integration The limitations of upscaling and downscaling strategies return the missing variability problem (MVP) to its core difficulty. Tissue-level parameters and molecular stochasticity capture relevant influences, but neither provides a stable bearer of predictive properties. The unresolved question is where variability-generating processes can be coordinated in a way that supports forward-looking risk assessment. This motivates a return to the cellular level as an integrative predictive level. At the cellular level, molecular stochasticity and tissue-level influences can be treated within a single predictive frame. Although noise is measured through molecular assays, its magnitude and structure correlate with stable cellular features, including chromatin organization and metabolic state, allowing variability to be attributed to the cell as a determinate predictive bearer (Fraser and Kærn 2009; Stewart-Ornstein et al. 2012). Similarly, tissue-level signals do not enter prediction as global parameters but as patterned responses that are trackable at the level of individual cells, such as phenotype-dependent reactions to mechanical or biochemical cues (Nelson and Bissell 2006; Safaei et al. 2023). These two dimensions—intrinsic variability and patterned responsiveness—make it possible to compare cells in terms of how they generate and modulate variation under comparable conditions, as reflected in detectable features of cellular organization. Dispositional vocabulary is introduced in this subsection to make explicit why the cell can serve as a stable bearer of coordinated predictors across conditions. The following section shows how this integration can be achieved. In predictive terms, both dimensions support the attribution of dispositional properties to cells, understood as differences in their capacity to vary and to respond in characteristic ways across time and contexts (Austin 2017). While dispositional language is often applied at higher levels of organization, the mechanisms through which variability is realized are described in terms of molecular instabilities and flexibilities (Brigandt et al. 2023). In multicellular systems, this disconnect makes it difficult to locate predictive dispositions at organismal or tissue levels. In cancer, by contrast, the cell provides a site where the generation, persistence, and measurement of variability coincide. This is particularly evident in stem cell populations, which play a central role in predictive accounts associated with the “bad luck” thesis. This connection becomes clearer when stemness is understood dispositionally, as a cell’s capacity to generate lineage-level variation. Contemporary conceptual analyses of stemness distinguish four ontological interpretations of stemness in cancer cells: categorical, systemic, relational, and dispositional (Laplane 2016). Categorical accounts treat stemness as a fixed cellular type. Systemic accounts locate it at the level of tissue organization. Both approaches limit prediction. Categorical views fail to account for the emergence of cancer stem cells from non-stem cells, while systemic views cannot explain the persistence of stem-like behavior across changing tissue contexts. Relational accounts move closer to integration by emphasizing cell–environment interactions. However, relational interpretations remain insufficient for prediction when taken on their own. Cancer stem cells can retain their capacity for asymmetric division even after the conditions that induced it have disappeared. Late cancer recurrence following long periods of quiescence provides a clear illustration of this persistence (Marzagalli et al. 2021). This stability cannot be captured solely in terms of ongoing relations with the tissue environment. The dispositional interpretation identified by Laplane (2016) avoids these limitations. On this view, stemness is an intrinsic cellular property that can be induced by external conditions but is not exhausted by themFootnote 1. It refers to a cell’s capacity to divide asymmetrically, generating a self-renewing and variably differentiated lineage (Lander 2009). Empirically grounded accounts emphasize that this capacity depends on both molecular processes and tissue-level cues (Suárez 2023) yet remains attributable to the cell itself. Cancer stem cells exemplify this structure. They are central drivers of tumor heterogeneity and therapeutic resistance (Rich 2016; Naz et al. 2021), and their behavior reflects a stable disposition to generate variability rather than a transient response to context. The relevant stability is relationally conditioned rather than context-free. The point is temporal: a cell or cell lineage may acquire a state that persists beyond the inducing conditions and later shapes responsiveness, variation, or recurrence in new contexts. Dormant cancer cells illustrate this pattern, since they can remain clinically silent before later reactivation (Marzagalli et al. 2021). Dispositional language is useful here because it refers not to a single observed response, but to a cellular state with context-dependent manifestations. This mode of attribution becomes visible in contemporary work on cell-based biopsy and cytological assessment. Circulating tumor cells are increasingly treated not merely as diagnostic traces but as predictive objects that preserve integrated cellular states shaped by cumulative processes such as genomic instability, dysregulated proliferation, and altered differentiation (Ferreira et al. 2016; Ma et al. 2024). Unlike molecular analyses, which fragment tumor biology into isolated markers, intact cells retain coordinated patterns of variation across molecular, structural, and regulatory dimensions. Morphological features—such as nuclear architecture, chromatin organization, and nuclear–cytoplasmic relations—function as cross-tumor indicators of malignancy precisely because they integrate heterogeneous sources of variability and do not depend on specific molecular pathways or marker panels (Dhar et al. 2016). In this sense, morphology does not operate as a marker in the conventional sense, but as a phenotypic expression of relatively stable cellular dispositions through which influences from multiple levels of organization are coordinated within a determinate cellular bearer. Interpreted in this way, cellular dispositions provide a framework for integrating intrinsic and extrinsic predictors of cancer risk. Higher-level factors, including tissue organization and mechanical constraints, contribute to prediction only insofar as they act through cellular states. Their predictive relevance depends on how cells respond to these influences and on the stability of these responses across divisions. Responses to extracellular biochemical gradients depend on cellular state in this sense. Gradient noise in morphogen and hormone signaling becomes predictively relevant when local signal concentrations meet response thresholds that differ between cellular states (Vetter and Iber 2022). In cancer, intracellular factors can alter these thresholds, so differences among cells in threshold-dependent responses become part of the predictive profile (Romero-Arias et al. 2023). In this sense, two cellular features are central for prediction: levels of molecular noise, which condition the rate of variability within a lineage, and characteristic patterns of response to tissue-level influences. Cellular Dispositions as an Integrative Response to the Missing Variability Problem The missing variability problem (MVP) persists in cancer risk prediction because predictive factors relevant to cancer are identified across multiple levels of biological organization, while prediction becomes feasible when variability can be related to a single, identifiable biological unit. Genomic, epigenetic, and tissue-level influences are well documented, yet their relevance for predicting individual risk remains difficult to articulate in a unified way. The proposal developed in this article is to treat the cellular level as an integrative predictive level. Its relevance to individual risk assessment is practical: cell-based measurements can keep molecular, morphological, and contextual features tied to the same sampled unit, a condition relevant to possible diagnostic or prognostic assays. Integration at this level is material rather than statistical. Statistical integration combines measurements drawn from different levels into composite risk estimates. Material integration consists in coordinating heterogeneous influences as features of the same biological unit. Empirically, this is technically demanding: it requires data linking subcellular features to cellular identity while preserving spatial or lineage context, as in spatial single-cell analysis, single-cell multi-omics, or lineage tracing. At the cellular level, molecular instability and tissue-level influences are not merely correlated with outcomes but become jointly attributable to the same cell or cell lineage. They function predictively not as a single unified variable, but as components of a connected profile of cellular state, responsiveness, and variability. Within this approach, attributing dispositional properties to cells is treated as a predictive strategy. Dispositional properties are understood as capacities of cells to exhibit characteristic patterns of variation across conditions. For predictive purposes, however, such properties are not accessed as primitive or unitary. Rather, they are identified through a structured set of indicators that make cross-level influences tractable at the cellular level: (i) cellular morphology, (ii) epigenetic features that register underlying molecular dynamics, and (iii) stable patterns of responsiveness to tissue-level and other higher-level influences. Together, these components allow influences from below and above the cellular level to be integrated while remaining attributable to the cell itself. Importantly, these dispositional properties are often distinguishable at the cytological level, which makes them suitable for predictive use. A practical difficulty in using the three indicators identified above is that morphology, epigenetic state, and responsiveness to tissue-level conditions may each be informative, but they do not by themselves provide a unified predictor of cellular state. Image-based analyses of breast tumor progression address this difficulty by extracting single-cell features from patient-derived biopsies and using them to classify cells across the transition from normal organization to malignant and metastatic forms (Venkatachalapathy et al. 2021). The relevant features map onto the present framework directly: nuclear shape and related morphometric features capture cellular morphology; chromatin organization and distribution provide image-based access to epigenetic and regulatory state; and spatial relations among nuclei register how cells are embedded in local tissue-mechanical environments (Fig. 1). These features are combined at the level of the same cell, rather than treated as separate variables across molecular, cellular, and tissue levels. In relation to the missing variability problem, this makes a specific form of previously dispersed variability predictively usable: single-cell feature profiles can be compared across normal, malignant, and metastatic tissue states beyond any one indicator taken in isolation. This ordering does not reconstruct genealogical cell lineages; that would require barcoding data, in which cells carry heritable labels that allow their descendants to be identified. The example is therefore best understood as part of a broader shift toward cell-based integration in cancer prediction, where liquid biopsy, cytological assessment, and single-cell imaging treat intact cells as predictive units rather than as carriers of isolated molecular markers (Ferreira et al. 2016; Dhar et al. 2016; Ma et al. 2024). In such practices, morphology, chromatin organization, spatial embedding, and responsiveness to extracellular conditions are coordinated as features of the same cell or cell lineage. These features form a dispositional profile insofar as they indicate not only a present state, but a cell’s capacity to persist, vary, and acquire malignant or metastatic potential. The point of such a dispositional profile is not simply statistical compression. Its predictive value depends on the fact that the measured features are attributed to the same cell or cell lineage, where they describe connected aspects of one biological unit. The relevant predictors are therefore cellular morphology, chromatin and epigenetic state, stable response patterns to extracellular and tissue-level conditions, and lineage-level behavior. Epigenetic features count as cellular properties in this context because they are not treated as isolated molecular events, but as relatively stable regulatory states of the cell that shape its future responses. The three indicators are combined by relating them to a common bearer whose state, responsiveness, and variability can be compared across conditions and over time. The dispositional vocabulary marks this mode of integration. In predictive practice, the difference lies in treating multilevel factors as features of the same cell or cell lineage, combining them as a connected profile rather than by aggregation, and using them to track cellular states along a progression trajectory from normal organization toward malignant and metastatic forms. Interpreted in these terms, molecular noise is not treated as a set of isolated fluctuations. Although it is detected through molecular measurements, it is attributed to the cell as a phenotypic property. Fluctuations at the molecular level become predictively relevant insofar as they are registered in comparable cellular states. Molecular noise plays a key role in developmental processes by disrupting cellular symmetry during division and thereby influencing cell fate decisions (van Heyningen 2024). Normal stem cells exhibit elevated noise levels relative to differentiated cells, while cancer stem cells show even higher noise levels and greater variability in noise across divisions (Capp 2019). Contemporary extensions of the mutator phenotype hypothesis (Loeb et al. 1974; Loeb 2016) treat mutations as only one manifestation of broader intrinsic cellular instability. Phenotypic instabilities can be transmitted across cell lineages via genetic and epigenetic mechanisms, and persistent cancer cell lineages often exhibit elevated noise associated with chromatin instability (Capp and Thomas 2022). Interpreted dispositionally, molecular noise contributes to a cell’s capacity to generate variable descendants and forms one component of a broader dispositional profile, rather than functioning as an isolated molecular parameter. A parallel dispositional interpretation applies to tissue-level influences. Tissue properties affect cellular variability by shaping patterns of response rather than by acting as predictors independently of cellular phenotype. This account of tissue-level influences can incorporate work in cancer biophysics that treats tumor cell collectives as soft-matter systems. This line of research shows that groups of tumor cells can acquire collective physical states that are not derivable from single-cell properties alone and that may contribute to invasion and progression (Massey et al. 2024). Such states help specify how mechanical conditions in tissues and cell ensembles form or stabilize cellular dispositional profiles. At the same time, these profiles remain predictively relevant because they condition how cells and cell lineages later respond to mechanical stimuli. A cell’s response to mechanical stimuli depends on its phenotypic organization: normal and cancerous cells may respond in opposite ways to the same mechanical constraints (Safaei et al. 2023), and both increased and decreased matrix stiffness can promote cancer cell proliferation through different signaling pathways (Wei et al. 2022). These effects are mediated by mechanotransduction pathways involving transmembrane proteins, cytoskeletal elements, and nuclear structures, whose configuration varies across cells and is a central focus of current cancer research (Blanco et al. 2023). Extracellular matrix remodeling adds a feedback dimension to this relation. Cancer and stromal cells can degrade, modify, or produce matrix components, while other cells or cell lineages may differ in their sensitivity to the altered matrix conditions (Winkler et al. 2020). From a dispositional perspective, tissue-level influences matter for prediction insofar as they are reflected in stable cellular response patterns that persist across divisions and condition the likelihood of stemness within a lineage. The practical focus on cells and cell lineages is thus not meant to reconstruct the full mechanical state of the tumor from single cells. It concerns the more limited possibility of tracking or sampling mechanically shaped cellular profiles in cell-based predictive practices, including the mechanical phenotyping of circulating tumor cells (Peralta et al. 2022). Taken together, molecular noise and patterned responsiveness to tissue-level factors are not treated as independent predictors. They are integrated as components of a dispositional profile attributable to individual cells. This integration is material rather than statistical: heterogeneous influences from lower and higher levels are coordinated as properties of a single biological unit, allowing them to enter a connected predictive profile. At the same time, this strategy is subject to clear limitations. Work on tissue fields and cell-state dynamics shows that cancer-relevant cellular states remain open to tissue-level regulation and nongenetic plasticity (Huang et al. 2025). A dispositional profile therefore cannot be inferred from a single cellular output, such as proliferation or marker expression. It requires evidence that patterns of responsiveness remain sufficiently stable across changing conditions, for instance in persistence, lineage behavior, recurrence, or responses to extracellular and tissue-level cues. Not all cellular dispositions relevant to cancer are equally accessible to observation, since biological organization forms dense thickets rather than transparent hierarchies, rendering some properties phenotypically stabilized and detectable while others remain method-dependent or latent (Wimsatt 1994; Eronen and Brooks 2018). This uneven accessibility is well illustrated in cancer cytology, where predictive use is typically restricted to a limited subset of phenotypic features—such as gross cellular morphology or persistent chromatin alterations—while other dispositional traits that shape cancer-related variability remain below the threshold of reliable detection (Zajchowski et al. 2001; Eling et al. 2019). A second, and less obvious, limitation concerns time. Interlevel integration becomes empirically accessible only over extended periods: molecular fluctuations generally need to be followed across time to be recognizable as stable cellular dispositions, and such dispositions often require sustained observation to manifest as patterned behavior in tissue contexts (Emmeche et al. 1997). Experimental cancer organoid systems make this temporal requirement explicit (Tuveson and Clevers 2019), as extended cultivation is typically needed before cellular heterogeneity, lineage structure, and stemness-related properties become phenotypically detectable and predictive. Conclusion The missing variability problem (MVP) characterizes a stable difficulty in cancer risk prediction: a substantial portion of individual risk remains unaccounted for even when genomic and environmental factors are jointly considered. One influential way of articulating this gap is provided by the “bad luck” thesis, which locates the remaining unpredictability at the cellular level by attributing it to stochastic events during stem cell divisions. Efforts to address this gap have moved in two directions—downscaling to molecular noise and upscaling to tissue-level organization. But these strategies typically remain weakly coordinated. The difficulty is therefore not only the lack of candidate factors. It also concerns the absence of a framework for specifying how heterogeneous sources of variability, once identified, can be integrated in a predictive manner. This article has argued that the cellular level provides a point of integration and can be specified as an integrative predictive level. Molecular and tissue-level influences do not enter prediction there as separate parameters, but as coordinated features of a single biological unit. What becomes identifiable at this level are not isolated causes, but dispositions to generate and modulate variability over time. These dispositions are not attributed as elementary properties; rather, their identification is compositional. It proceeds through phenotypically accessible features—such as cell morphology, characteristic patterns of molecular noise, and stable response profiles to tissue-level influences—which together support the attribution of predictive relevance to the cell. In this sense, the cell enables multilevel influences to be jointly apprehended as a connected dispositional profile of a determinate bearer. This does not by itself close the MVP but shows how different empirical strategies for addressing it can be related without reducing cancer prediction to a single scale or mechanism. Notes This also applies to interactions with the immune system. On the discontinuity theory of immunity, immune responses are triggered by abrupt antigenic change, which makes rapidly varying cancer cell lineages more likely to be eliminated, so that a moderate cellular disposition to vary functions as a predictive factor for immune evasion (Pradeu et al. 2013). References Alexandrov LB, Kim J, Haradhvala NJ, Huang MN, Tian Ng AW, Wu Y, Boot A, Covington KR, Gordenin DA, Bergstrom EN, Islam SA et al (2020) The repertoire of mutational signatures in human cancer. Nature 578:94–101. https://doi.org/10.1038/s41586-020-1943-3 Austin CJ (2017) Evo-devo: a science of dispositions. Eur J Philos Sci 7:373–389. https://doi.org/10.1007/s13194-016-0166-9 Bertolaso M (2016) Philosophy of cancer: a dynamic and relational view. Springer, Dordrecht. https://doi.org/10.1007/978-94-024-0865-2 Biondo M, Singh A, Caselle M, Osella M (2023) Out-of-equilibrium gene expression fluctuations in the presence of extrinsic noise. Phys Biol 20:056007. https://doi.org/10.1088/1478-3975/acea4e Blanco B, Gomez H, Melchor J, Palma R, Soler J, Rus G (2023) Mechanotransduction in tumor dynamics modeling. Phys Life Rev 44:279–301. https://doi.org/10.1016/j.plrev.2023.01.017 Brigandt I, Villegas C, Love A, de la Nuno L (2023) Evolvability as a disposition: philosophical distinctions, scientific implications. In: Hansen T, Houle D, Pavličev M, Pélabon C (eds) Evolvability, a unifying concept in evolutionary biology? MIT Press, Cambridge, MA, pp 55–72 Campbell J, Cambrosio A, Basik M (2024) Histology agnosticism: infra-molecularizing disease? Stud Hist Philos Sci 104:14–22. https://doi.org/10.1016/j.shpsa.2024.02.002 Canali S (2019) Evaluating evidential pluralism in epidemiology: mechanistic evidence in exposome research. Hist Philos Life Sci 41:4. https://doi.org/10.1007/s40656-019-0241-6 Capp JP (2019) Cancer stem cells: from historical roots to a new perspective. J Oncol 2019:5189232. https://doi.org/10.1155/2019/5189232 Capp JP, Thomas F (2022) From developmental to atavistic bet-hedging: how cancer cells pervert the exploitation of random single-cell phenotypic fluctuations. BioEssays 44:e2200048. https://doi.org/10.1002/bies.202200048 Carneiro CS, Hapeman JD, Nedelcu AM (2023) Synergistic inter-clonal cooperation involving crosstalk, co-option and co-dependency can enhance the invasiveness of genetically distant cancer clones. BMC Ecol Evol 23:20. https://doi.org/10.1186/s12862-023-02129-7 Casali M, Merlin F (2020) Rethinking the role of chance in the explanation of cell differentiation. In: Levine H, Jolly MK, Kulkarni P, Nanjundiah V (eds) Phenotypic switching: implications in biology and medicine. Cambridge, Academic Press, pp 23–51. https://doi.org/10.1016/B978-0-12-817996-3.00005-0 Casali M, Merlin F, Vianelli A (2025) Dissecting stochasticity in translation: time in start codon selection as a case-study. Philos Theory Pract Biol 17:2. https://doi.org/10.3998/ptpbio.5321 Dhar M, Pao E, Renier C et al (2016) Label-free enumeration, collection and downstream cytological and cytogenetic analysis of circulating tumor cells. Sci Rep 6:35474. https://doi.org/10.1038/srep35474 Eling N, Morgan MD, Marioni JC (2019) Challenges in measuring and understanding biological noise. Nat Rev Genet 20:536–548. https://doi.org/10.1038/s41576-019-0130-6 Emmeche C, Køppe S, Stjernfelt F (1997) Explaining emergence: towards an ontology of levels. J Gen Philos Sci 28:83–117 Eronen MI, Brooks DS (2018) Levels of organization in biology. In: Zalta EN, Nodelman U (eds) The Stanford encyclopedia of philosophy, Winter. https://plato.stanford.edu/entries/levels-org-biology/ Accessed 30 January 2026 Fearon ER, Vogelstein B (1990) A genetic model for colorectal tumorigenesis. Cell 61:759–767. https://doi.org/10.1016/0092-8674(90)90186-I Feinberg AP, Levchenko A (2023) Epigenetics as a mediator of plasticity in cancer. Science 379:eaaw3835. https://doi.org/10.1126/science.aaw3835 Ferreira MM, Ramani VC, Jeffrey SS (2016) Circulating tumor cell technologies. Mol Oncol 10:374–394. https://doi.org/10.1016/j.molonc.2016.01.007 Fraser D, Kærn M (2009) A chance at survival: gene expression noise and phenotypic diversification strategies. Mol Microbiol 71:1333–1340. https://doi.org/10.1111/j.1365-2958.2009.06605.x Green S (2021) Cancer beyond genetics: on the practical implications of downward causation. In: Brooks DS, DiFrisco J, Wimsatt WC (eds) Levels of organization in the biological sciences. MIT Press, Cambridge, MA, pp 195–214. https://doi.org/10.7551/mitpress/12389.003.0014 Han R, Huang G, Wang Y et al (2016) Increased gene expression noise in human cancers is correlated with low p53 and immune activities as well as late stage cancer. Oncotarget 7:72011–72020. https://doi.org/10.18632/oncotarget.12457 Hochstein E (2022) Foregrounding and backgrounding: a new interpretation of levels in science. Eur J Philos Sci 12:23. https://doi.org/10.1007/s13194-022-00457-x Huang S, Soto AM, Sonnenschein C (2025) The end of the genetic paradigm of cancer. PLoS Biol 23:e3003052. https://doi.org/10.1371/journal.pbio.3003052 Jia D, Jolly MK, Kulkarni P, Levine H (2017) Phenotypic plasticity and cell fate decisions in cancer: insights from dynamical systems theory. Cancers 9:70. https://doi.org/10.3390/cancers9070070 Jia G, Lu Y, Wen W et al (2020) Evaluating the utility of polygenic risk scores in identifying high-risk individuals for eight common cancers. JNCI Cancer Spectr 4:pkaa021. https://doi.org/10.1093/jncics/pkaa021 Kachuri L, Graff RE, Smith-Byrne K et al (2020) Pan-cancer analysis demonstrates that integrating polygenic risk scores with modifiable risk factors improves risk prediction. Nat Commun 11:6084. https://doi.org/10.1038/s41467-020-19600-4 Lander AD (2009) The ‘stem cell’ concept: is it holding us back? J Biol 8:70. https://doi.org/10.1186/jbiol177 Laplane L (2016) Cancer stem cells: philosophy and therapies. Harvard University Press, Cambridge Laplane L (2025) Cancer clones revised. Biol Theory 20:253–266. https://doi.org/10.1007/s13752-024-00484-2 Laplane L, Mantovani P, Adolphs R et al (2019) Why science needs philosophy. Proc Natl Acad Sci USA 116(10):3948–3952. https://doi.org/10.1073/pnas.1900357116 Laplane L, Lamoureux A, Richker HI et al (2025) Applying multilevel selection to understand cancer evolution and progression. PLoS Biol 23:e3003290. https://doi.org/10.1371/journal.pbio.3003290 Leonelli S (2016) Data-centric biology: a philosophical study. University of Chicago Press, Chicago Lewontin RC (2000) The triple helix: gene, organism, and environment. Harvard University Press, Cambridge Lewontin R, Levins R (2000) Let the numbers speak. Int J Health Serv 30:873–877 Loeb LA (2016) Human cancers express a mutator phenotype: hypothesis, origin, and consequences. Cancer Res 76:2057–2059 Loeb LA, Springgate CF, Battula N (1974) Errors in DNA replication as a basis of malignant changes. Cancer Res 34:2311–2321 Love AC (2012) Hierarchy, causation and explanation: ubiquity, locality, and pluralism. Interface Focus 2:115–125. https://doi.org/10.1098/rsfs.2011.0064 Ma L, Guo H, Zhao Y et al (2024) Liquid biopsy in cancer: current status, challenges and future prospects. Signal Transduct Target Ther 9:336. https://doi.org/10.1038/s41392-024-02021-w Mancusi R, Monje M (2023) The neuroscience of cancer. Nature 618:467–479. https://doi.org/10.1038/s41586-023-05968-y Marzagalli M, Fontana F, Raimondi M, Limonta P (2021) Cancer stem cells—key players in tumor relapse. Cancers 13:376 Massey A, Stewart J, Smith C, Parvini C, McCormick M, Do K, Cartagena-Rivera AX (2024) Mechanical properties of human tumour tissues and their implications for cancer development. Nat Rev Phys 6:269–282. https://doi.org/10.1038/s42254-024-00707-2 Matthews LJ, Turkheimer E (2022) Three legs of the missing heritability problem. Stud Hist Philos Sci 93:183–191. https://doi.org/10.1016/j.shpsa.2022.02.002 McEntire KD, Gage M, Gawne R et al (2021) Understanding drivers of variation and predicting variability across levels of biological organization. Integr Comp Biol 61(6):2119–2131. https://doi.org/10.1093/icb/icab160 Merlin F, Giroux É (2024) Narratives in exposomics: a reversed heuristic determinism? Hist Philos Life Sci 46:22. https://doi.org/10.1007/s40656-024-00620-y Militello G, Bertolaso M (2022) Stem cells and the microenvironment: reciprocity with asymmetry in regenerative medicine. Acta Biotheor 70:24. https://doi.org/10.1007/s10441-022-09448-0 Naz F, Shi M, Sajid S, Yang Z, Yu C (2021) Cancer stem cells: a major culprit of intra-tumor heterogeneity. Am J Cancer Res 11:55782–55811 Nedelcu AM (2020) The evolution of multicellularity and cancer: views and paradigms. Biochem Soc Trans 48(4):1505–1518. https://doi.org/10.1042/BST20190992 Needham J (1937) Integrative levels: a revaluation of the idea of progress. Clarendon, Oxford Nelson CM, Bissell MJ (2006) Of extracellular matrix, scaffolds, and signaling: tissue architecture regulates development, homeostasis, and cancer. Annu Rev Cell Dev Biol 22:287–309. https://doi.org/10.1146/annurev.cellbio.22.010305.104315 Newman SA (2012) Physico-genetic determinants in the evolution of development. Science 338:217–219. https://doi.org/10.1126/science.1222003 Peralta M, Osmani N, Goetz JG (2022) Circulating tumor cells: towards mechanical phenotyping of metastasis. iScience 25:103969. https://doi.org/10.1016/j.isci.2022.103969 Peterson E (2016) The life organic: the theoretical biology club and the roots of epigenetics. University of Pittsburgh, Pittsburgh Peto R (2016) Epidemiology, multistage models, and short-term mutagenicity tests. Int J Epidemiol 45:621–637. https://doi.org/10.1093/ije/dyv199 Plutynski A (2018) Explaining cancer: finding order in disorder. Oxford University Press, Oxford Plutynski A (2021a) Is cancer a matter of luck? Biol Philos 36:3. https://doi.org/10.1007/s10539-020-09778-8 Plutynski A (2021b) The cancer genome atlas project: data-driven, hypothesis-driven or something in-between? In: Donohue C, Love AC (eds) Perspectives on the Human Genome Project and Genomics. University of Minnesota Press, Minneapolis, pp 77–94 Pradeu T, Jaeger S, Vivier E (2013) The speed of change: towards a discontinuity theory of immunity? Nat Rev Immunol 13(10):764–769 Pradeu T, Daignan-Fornier B, Ewald A et al (2023) Reuniting philosophy and science to advance cancer research. Biol Rev 98(5):1668–1686. https://doi.org/10.1111/brv.12971 Rebbeck CA, Thomas R, Breen M, Leroi AM, Burt A (2009) Origins and evolution of a transmissible cancer. Evolution 63(9):2340–2349 Rich JN (2016) Cancer stem cells: understanding tumor hierarchy and heterogeneity. Medicine 95(1S):S2–S7 Romero-Arias JR, González-Castro CA, Ramírez-Santiago G (2023) A multiscale model of the role of microenvironmental factors in cell segregation and heterogeneity in breast cancer development. PLoS Comput Biol 19:e1011673. https://doi.org/10.1371/journal.pcbi.1011673 Safaei S, Sajed R, Shariftabrizi A, Dorafshan S, Saeednejad Zanjani L, Dehghan Manshadi M, Madjd Z, Ghods R (2023) Tumor matrix stiffness provides fertile soil for cancer stem cells. Cancer Cell Int 23(1):143. https://doi.org/10.1186/s12935-023-02992-w Sonnenschein C, Soto AM (2020) Over a century of cancer research: Inconvenient truths and promising leads. PLoS Biol 18:e3000670. https://doi.org/10.1371/journal.pbio.3000670 Stewart-Ornstein J, Weissman JS, El-Samad H (2012) Cellular noise regulons underlie fluctuations in Saccharomyces cerevisiae. Mol Cell 45:483–493. https://doi.org/10.1016/j.molcel.2011.11.035 Suárez J (2023) What is the nature of stem cells? A unified dispositional framework. Biol Philos 38:43 Tee SH (2025) Noise and mechanisms. Synthese 206:216. https://doi.org/10.1007/s11229-025-05288-w Teschendorff AE (2024) On epigenetic stochasticity, entropy and cancer risk. Philos Trans R Soc Lond B Biol Sci 379:20230054. https://doi.org/10.1098/rstb.2023.0054 Tomasetti C, Vogelstein B (2015) Variation in cancer risk among tissues can be explained by the number of stem cell divisions. Science 347:78–81. https://doi.org/10.1126/science.1260825 Tomasetti C, Li L, Vogelstein B (2017) Stem cell divisions, somatic mutations, cancer etiology, and cancer prevention. Science 355:1330–1334. https://doi.org/10.1126/science.aaf9011 Tuveson D, Clevers H (2019) Cancer modeling meets human organoid technology. Science 364:952–955. https://doi.org/10.1126/science.aaw6985 van Heyningen V (2024) Stochasticity in genetics and gene regulation. Philos Trans R Soc B 379(1900):20230476 Venkatachalapathy S et al (2021) Single cell imaging-based chromatin biomarkers for tumor progression. Sci Rep 11:23041. https://doi.org/10.1038/s41598-021-02441-6 Vetter R, Iber D (2022) Precision of morphogen gradients in neural tube development. Nat Commun 13:1145. https://doi.org/10.1038/s41467-022-28834-3 Wei J, Yao J, Yang C, Mao Y, Zhu D, Xie Y, Liu P, Yan M, Ren L, Lin Y, Zheng Q (2022) Heterogeneous matrix stiffness regulates the cancer stem-like cell phenotype in hepatocellular carcinoma. J Transl Med 20(1):555 Wild CP (2005) Complementing the genome with an exposome: the outstanding challenge of environmental exposure measurement in molecular epidemiology. Cancer Epidemiol Biomarkers Prev 14(8):1847–1850 Wimsatt WC (1994) The ontology of complex systems: levels of organization, perspectives, and causal thickets. Can J Philos Suppl 20:207–274. https://doi.org/10.1080/00455091.1994.10717400 Winkler J, Abisoye-Ogunniyan A, Metcalf KJ, Werb Z (2020) Concepts of extracellular matrix remodelling in tumour progression and metastasis. Nat Commun 11:5120. https://doi.org/10.1038/s41467-020-18794-x Wishart D (2022) Metabolomics and the multi-omics view of cancer. Metabolites 12(2):154 Zajchowski DA, Bartholdi MF, Gong Y, Webster L, Liu HL, Munishkin A, Beauheim C, Harvey S, Ethier SP, Johnson PH (2001) Identification of gene expression profiles that predict the aggressive behavior of breast cancer cells. Cancer Res 61(13):5168–5178 Acknowledgments I am grateful to the anonymous reviewers for their careful and constructive comments, and to Francesca Merlin, Lucie Laplane, Jan Baedke, Alejandro Fábregas-Tejeda, Federico Boem, and Vera Straetmanns for helpful discussions. Funding Open Access funding enabled and organized by Projekt DEAL. No funding was received to assist with the preparation of this manuscript. Author information Authors and Affiliations Corresponding author Ethics declarations Competing Interests The author has no relevant financial or nonfinancial interests to disclose. Additional information Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Rights and permissions Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. About this article Cite this article Shevchenko, S. The Missing Variability Problem in Cancer Prediction: Integrating Multilevel Factors. Biol Theory (2026). https://doi.org/10.1007/s13752-026-00554-7 Received: Accepted: Published: Version of record: DOI: https://doi.org/10.1007/s13752-026-00554-7