The single most transformative technical development of the period is widely agreed to be the sequencing revolution. The landmark 2005 demonstration of genome sequencing in microfabricated high-density picolitre reactors opened the door to reading tumour genomes at scale, an advance that would have been unimaginable when the journal launched. From that technological foundation flowed the systematic discovery of cancer genes, the classification of tumours by their mutational signatures and the rise of precision oncology as a clinical discipline. The identification of recurrent mutations of the BRAF gene in human cancer in 2002, before the sequencing revolution fully matured, had already signalled what was to come: single genetic lesions, once found, could define entire treatment strategies.

Clinical actionability has expanded at a remarkable pace. Analyses quantifying the expanding landscape of clinical actionability for patients with cancer show that a steadily growing fraction of tumours now harbour alterations for which targeted drugs or guided treatment decisions exist. The poster child of this era remains BRAF V600E inhibition in melanoma, but the paradigm has matured in sophistication. When colon cancers proved unexpectedly unresponsive to BRAF inhibition, researchers discovered that feedback activation of EGFR was protecting the cells—a finding that led directly to the FDA-approved combination of BRAF and EGFR inhibitors for BRAF-mutant colorectal cancer, a therapy developed in René Bernards’ laboratory at the Netherlands Cancer Institute. The lesson embedded in that story reshaped the field: cancers are wired for resilience, and single-agent thinking is rarely sufficient.

That lesson now drives the logic of synthetic lethality, one of the most productive conceptual frameworks of the past decade. Rather than attacking oncogenes directly, synthetic lethal strategies exploit vulnerabilities created by a tumour’s alterations. A striking recent example is the demonstration that amplification of the cyclin E1 gene CCNE1, common in aggressive ovarian and other cancers, creates a dependence on the PKMYT1 kinase that can be pharmacologically exploited. This approach extends druggability into territory long considered untouchable, and drug discovery scientist Paul Workman of the Institute of Cancer Research has championed the broader project of drugging the cancer genome—developing chemical inhibitors and small-molecule research tools against targets once dismissed as intractable.

Perhaps no target symbolises both the promise and the frustration of the era better than RAS. Mutant RAS genes were identified in the early 1980s and were long considered undruggable, a verdict repeated in reviews for decades. Yet the past few years have delivered direct RAS inhibitors, and a 2026 phase three trial reported that the RAS inhibitor daraxonrasib outperformed chemotherapy in previously treated metastatic pancreatic cancer—one of the most difficult malignancies to treat. Allan Balmain, whose laboratory uses mouse models to trace the cells that acquire initiating oncogenic mutations, notes that interpreting such advances requires a deeper understanding of tumour evolution itself: when a mutation arises, in which cell it arises, and what environmental forces promote or suppress its expansion.

That evolutionary perspective has been genuinely subversive. Sequencing of normal human skin revealed a high burden of somatic mutations and pervasive positive selection operating in histologically normal tissue—mutant clones competing for space in organs that look entirely healthy. Even more provocative, work in mosaic mouse skin has shown that injury can prevent the expansion of Ras-mutant cells, overturning the intuitive assumption that wounding uniformly promotes carcinogenesis. Complementing this, chemically induced skin tumours have been traced to long-lived stem cells of the upper hair follicle, underscoring that the identity of the cell of origin is a decisive variable in tumour outcome. Cancer, in this view, is not simply a cell-autonomous genetic disease but an ecological process embedded in tissue dynamics, environmental exposure and time.

The cancer stem cell concept illustrates how an important idea can be discovered, oversimplified and then properly understood only over decades. The founding observation—that human acute myeloid leukemia is organized as a hierarchy originating from a primitive hematopoietic cell—was published in 1997, before the review period began. The intervening years saw the concept generalized, contested and eventually refined. As Hans Clevers and Eduard Batlle argued in revisiting the field, functional stemness depends on cellular context and state rather than immutable identity, and epithelial cell plasticity allows cells to cross lineage boundaries that once seemed fixed. The modern synthesis holds that many tumour cells can transiently adopt stem-like behaviour, which has important consequences for therapy resistance and relapse, even if the simplest hierarchical models proved too rigid.

Technology has also reshaped how tumours are studied. Clevers’ pioneering work on organoids—self-organizing, three-dimensional cultures derived from adult stem cells—created patient-derived models that preserve the genetic and functional features of the original tissue, bridging the long-standing gap between cell lines and patients. Organoid biobanks now support drug screening, gene-function studies and personalised treatment predictions, and the technology has been commercialised widely, with Clevers holding patents and co-founding companies in the field. Alongside organoids, powerful tools such as CRISPR-based functional genomics, exemplified by the synthetic lethality screens of the Bernards laboratory, allow systematic interrogation of cancer vulnerabilities directly in human cells.

Beyond genes, the anniversary reflections give substantial weight to metabolism and the tumour microenvironment, areas that were marginal when the journal began but are now central. Tumour-derived lactic acid was shown to functionally polarise tumour-associated macrophages, demonstrating that a metabolic by-product can reprogram immune cells within the microenvironment. Cancer-associated cachexia, the devastating wasting syndrome that contributes to a large fraction of cancer deaths, has recently been connected to mechanistic insight in human studies, replacing decades of descriptive work. Karen H. Vousden, whose research spans p53 biology and metabolism, has explored how dietary composition interacts with tumour metabolism, and precision nutrition companies are now testing whether manipulating non-essential amino acids can enhance therapy in patients. The message is that cancer cannot be understood as a tumour in isolation; it is a systemic disease reshaping the entire body.

The newest force transforming the field is artificial intelligence. The transformer architecture described in ‘Attention is all you need’ underpins models that now predict biomolecular interactions with remarkable accuracy, as demonstrated by AlphaFold 3, and AI agents are being designed to empower biomedical discovery end to end—from hypothesis generation to experimental execution. Marinka Zitnik of Harvard Medical School leads work on multimodal and agentic AI systems paired directly with laboratory experiments, and is team lead of AURORA within Cancer Grand Challenges. The authors’ cautious optimism is tempered by experience: past waves of overhyped technologies promised more than they delivered, and AI will need to demonstrate genuine predictive power in clinics, not only impressive benchmarks, to transform outcomes.

Woven through the reflections is an honest accounting of unmet expectations. Not every large-scale endeavour—from indiscriminate high-throughput screening to simplistic gene-expression signatures—delivered what early enthusiasts forecast. Combination therapy remains difficult to develop because toxicity, not efficacy, is usually the limiting factor. Drug resistance evolves faster than new drugs arrive, and late-stage solid tumours continue to claim most of the lives that cancer kills. Yet the trajectory is unmistakable: five-year survival has risen for many malignancies, targeted combinations have turned once-fatal diseases into chronic or curable conditions, and the mechanistic depth of the field is unrecognisable compared with 2001. The authors’ shared conclusion is that the next 25 years will be judged by how well the field integrates these hard-won insights—evolutionary, ecological, metabolic and computational—into treatments that reach every patient, not just those with actionable mutations in privileged tumour types.

Subject of Research: A 25-year retrospective on transformative advances and unmet expectations in cancer research

Article Title: Reflecting on 25 years of cancer research: transformative advances and unmet expectations

Article References: Balmain, A., Bernards, R., Clevers, H., Vousden, K. H., Workman, P., & Zitnik, M. (2026). Reflecting on 25 years of cancer research: transformative advances and unmet expectations. Nature Reviews Cancer. https://doi.org/10.1038/s41568-026-00974-4

Image Credits: AI Generated

DOI: 10.1038/s41568-026-00974-4

Keywords: cancer research, Nature Reviews Cancer, tumour evolution, precision oncology, synthetic lethality, RAS inhibitors, cancer stem cells, organoids, tumour metabolism, cachexia, artificial intelligence, drug discovery

The review, authored by Giulio Caravagna of the University of Trieste and Area Science Park, Trevor A. Graham of the Centre for Evolution and Cancer at The Institute of Cancer Research in London, and Andrea Sottoriva of Human Technopole in Milan, makes a deceptively simple but profound point. Sequencing is a snapshot, whereas evolution is a dynamic process. A single tumour biopsy tells us which mutations are present and at what frequencies, but it does not, by itself, reveal the underlying dynamics that produced them. Bridging this gap requires models. Population genetics, the discipline developed over the past century to understand how allele frequencies change in natural populations under the influence of mutation, selection and drift, provides exactly the framework needed to transform static measurements of variant allele frequencies into quantitative estimates of clonal fitness and evolutionary timings.

At the heart of the framework lies the concept of the site frequency spectrum, the distribution of mutations across different variant allele frequencies within a tumour sample. In a neutrally evolving tumour, one in which no clone enjoys a fitness advantage over its neighbours, mathematical theory predicts a characteristic power-law tail: a predictable excess of mutations at progressively lower frequencies, each arising in expanding lineages as passengers hitchhiking along with the growing clone. This ‘neutral tail’ has become a signature of neutral tumour evolution, first identified across cancer types in work led by the same research groups. Deviations from this expected distribution are the fingerprints of selection. When subclones carrying driver mutations expand faster than neutral expectations, the frequency spectrum distorts in characteristic and quantifiable ways, allowing researchers to estimate the strength of selection rather than merely guess at it.

The authors show how modern computational tools exploit these principles. Subclonal deconvolution, the process of resolving a bulk sequencing sample into its constituent clonal and subclonal populations, has traditionally been treated as a clustering problem. But clustering alone cannot distinguish between a tumour shaped by strong selection and one dominated by spatially constrained growth, where genetic diversity accumulates by neutral drift in separate geographic compartments. By embedding population genetic models directly into inference algorithms, for example approaches combining machine learning with branching process theory, researchers can distinguish genuine selection from the illusions created by tumour architecture and sampling bias. Simulations of mutation, drift and selection can be tuned until the synthetic genomic data they produce match patient samples, converting descriptive allele frequencies into estimates of evolutionary parameters such as selection coefficients and the timing of clonal expansions.

Timing is one of the most clinically valuable outputs of this framework. Clock-like mutational processes, such as the spontaneous deamination of methylated cytosines, accumulate at approximately constant rates, providing a molecular clock against which key events can be dated. Studies of clear cell renal cell cancer have used these principles to time landmark events in tumour evolution, revealing that many chromosomal catastrophes occur astonishingly early, sometimes decades before diagnosis. More recent theoretical work has shown that patient age itself can help distinguish selection from causation in cancer genomes, since a mutation that arises early and confers a growth advantage leaves a different statistical imprint than one that simply accumulates with time. Dating driver events, genome doublings and the origins of metastatic seeds transforms the tumour genome from a parts list into a chronicle.

The review also confronts the limitations and confounders that complicate evolutionary inference from real data. Bulk sequencing averages across millions of cells, obscuring rare subclones and entangling spatial structure with temporal dynamics. Multi-region sequencing and single-cell approaches help, but each introduces its own biases: sampling depth, copy number alterations that distort allele frequencies, and the fundamental fact that a biopsy represents only a fragment of a spatially extended population. Copy number changes in particular must be carefully modelled, since amplifications and deletions shift variant allele frequencies in ways that can mimic or mask selection. The authors emphasise that assumptions embedded in population genetic models, such as well-mixed populations or constant growth rates, must be tested rather than taken for granted, because spatially constrained tumour growth can generate patterns that superficially resemble selection in the absence of any fitness advantage.

Beyond genetics, the framework extends to epigenetic inheritance and phenotypic plasticity, two dimensions of cancer evolution that the standard genetic models handle poorly. Epigenetic states such as DNA methylation patterns are heritable across cell divisions and can be under selection, yet they are reversible and can switch stochastically, creating a one-to-many relationship between genotype and phenotype. Quantitative models adapted from evolutionary theory, including those describing phenotypic plasticity and stochastic switching in fluctuating environments, offer a way to measure the heritability, transition rates and fitness consequences of non-genetic states. This matters enormously for therapy, because drug-tolerant persister cells frequently arise through epigenetic reprogramming rather than genetic mutation, and their dynamics determine whether resistance emerges in weeks or years.

The ecological dimension of tumour evolution receives similar treatment. Cancers are not just populations of competing clones; they are ecosystems in which cells cooperate, cheat and interact with stromal and immune cells. Game theory and eco-evolutionary models capture frequency-dependent selection, in which the fitness of a clone depends on the composition of its neighbourhood, something classical population genetics assumes away. These models have practical consequences. Adaptive therapy strategies, which aim to maintain sensitive clones that suppress resistant ones rather than eradicate the tumour wholesale, draw directly on ecological and game-theoretic reasoning. Similarly, immune selection on neoantigens shapes both tumour antigenicity and response to checkpoint inhibitors, and can be quantified using selection metrics adapted from population genetics.

Ultimately, the review’s central message is one of reframing. Cancer genomes should be read not as catalogues of mutations but as quantitative records of evolutionary processes, in which every allele frequency, every frequency-spectrum distortion and every signature of mutational timing encodes information about the dynamics that produced them. The authors argue that population genetics provides the foundation for understanding and, ultimately, predicting the trajectories of cancer evolution. If the framework fulfils its promise, the implications for precision oncology are substantial: forecasts of relapse timing grounded in measured evolutionary parameters, treatment strategies designed around the predictable dynamics of resistance, and clinical decisions informed not merely by which mutations a tumour carries today, but by the evolutionary forces that will shape what it becomes tomorrow.

The intellectual roots of this framework stretch well beyond oncology. Population genetics matured through the study of natural populations, and its migration into cancer research has been gradual, beginning with early attempts to reconstruct individual tumour histories from genetic data and gaining momentum as sequencing costs fell. A parallel body of work on clonal haematopoiesis and on pre-malignant lesions in normal tissues has reinforced the point that the evolutionary processes described in the review operate long before a tumour is diagnosed. Studies of normal breast tissue, for example, have revealed rare aneuploid epithelial populations and copy number alterations shared with frank cancers, suggesting that the same population genetic tools can illuminate the earliest steps of carcinogenesis across a continuous spectrum from healthy tissue to invasive disease.

Selection in tumours is not exclusively positive. Adapted metrics such as the ratio of synonymous to non-synonymous mutations, borrowed directly from classical genetics, have revealed universal patterns of selection in cancer and somatic tissues, while complementary analyses indicate that negative selection acts on essential cellular functions and on the immunopeptidome, pruning mutations that would compromise basic biology or expose cells to immune attack. Copy number amplifications of wild-type regions may further allow tumours to tolerate otherwise deleterious coding mutations, a reminder that the interplay between different classes of genomic alteration can obscure simple fitness calculations. At population scale, large studies of somatic mutation across many individuals now provide the statistical power to measure these forces with unprecedented precision.

Practical questions of study design also fall within the framework’s remit. How many biopsies are needed to confidently identify truly clonal mutations in a heterogeneous tumour, and how sampling schemes affect the inferred frequency spectrum, are problems that can themselves be solved with evolutionary models rather than ad hoc rules. The review’s figures trace this progression, from conceptual overviews of genomic readouts through the dynamics of variant allele frequencies, the confounding effects of spatial structure, and the distinct evolutionary routes by which tumours respond to therapy and relapse. Together they map a research programme in which the reliability of every evolutionary claim is tied explicitly to the sampling strategy and the model assumptions behind it.

Subject of Research: Application of population genetics models to infer tumour evolutionary dynamics from cancer genome sequencing data.

Article Title: A guide to understanding tumour evolution through the lens of population genetics

Article References: Caravagna, G., Graham, T. A., & Sottoriva, A. (2026). A guide to understanding tumour evolution through the lens of population genetics. Nature Reviews Cancer. https://doi.org/10.1038/s41568-026-00973-5

Image Credits: AI Generated

DOI: 10.1038/s41568-026-00973-5

Keywords: tumour evolution, population genetics, cancer genomics, clonal selection, variant allele frequency, subclonal deconvolution, genetic drift, mutational signatures, phenotypic plasticity, adaptive therapy, tumour heterogeneity, molecular clock