Theories of Carcinogenesis

Speculative If the rate-limiting step in cancer is the tissue context in which a mutant cell sits rather than the mutation itself — which the measurements in section 3 make possible and no experiment has established — then a class of interventions that does not currently exist becomes worth building: treatments and preventives aimed at the field rather than at the cell, which do not have to kill anything to work. The condition comes first because it has not been met, and section 7 says what would meet it.

Established The magnitude is set by the size of the disease and by the fact that the dominant theory has produced almost nothing preventive. The IARC estimates 20.6 million new cancer cases and 9.8 million cancer deaths in 2024, rising to a projected 34.4 million cases in 2050, a 67% increase, with one person in five developing cancer in a lifetime. Frontier Fifty years of somatic-mutation-theory research has produced a remarkable therapeutic catalogue and a thin preventive one: targeted drugs are priced per patient per year and arrive after diagnosis, while the interventions that have actually moved population incidence — tobacco control, hepatitis B vaccination, human papillomavirus vaccination — are all interruptions of promotion or of infection, not of mutation.

Speculative The number that matters is a fraction nobody agrees on, and that disagreement is the whole prize. One camp puts the intrinsic, replication-error share of lifetime cancer risk high; the other puts extrinsic factors at 70 to 90% of risk and intrinsic at under 10 to 30%. Frontier Those are not different datasets, they are different estimators applied to the same data, and settling which is right would reallocate a global research and prevention budget rather than add to it.

Frontier Who gets it is the part that distinguishes this from most of the corpus. Mutation-centric progress accrues to health systems that can pay for targeted agents and companion diagnostics; promotion-centric progress accrues through regulation, air quality and vaccination, which are cheap per head and work best in exactly the low- and middle-income settings where the 2050 burden growth is concentrated. Handwave The condition, stated plainly: none of this follows unless the two theories can be told apart by an experiment, and section 5 argues that the experiment which could do it was last run properly in 2004.

This brief is not about what causes cancer. It is about whether the field has the right kind of causal theory — whether cancer is a disease of cells that acquire mutations, or a disease of tissues whose organisation fails, with the mutations a consequence, a marker, or one input among several. The two candidate theories make different predictions about what an experiment should show, and the field has spent forty years running experiments that do not discriminate between them.

Established The somatic mutation theory holds that cancer is a clonal disease originating in a single cell that accumulates mutations in genes controlling proliferation, and that the default state of a metazoan cell is quiescence. It is the working theory of essentially all cancer genomics, it underwrites the entire targeted-therapy programme, and it has evidence nobody disputes: leukaemias defined by specific translocations, and drugs that work by inhibiting the product of the mutated gene. Frontier Its weak point is stated inside its own literature — the mutations it names as causal are present, under positive selection, in tissue that is not cancer and never becomes cancer.

Speculative The tissue organization field theory, argued since the 1990s by Carlos Sonnenschein and Ana Soto at Tufts, inverts two premises at once. Carcinogenesis is a tissue-level disorder of the reciprocal interactions between epithelium and stroma — development gone wrong rather than a cell gone rogue — and the default state of cells is proliferation, so the thing needing explanation is the constraint, not the growth. Frontier Its experimental case rests on a small number of tissue-recombination results, its critics say it is untestable at scale, and its proponents say the same of the theory it would replace. David Vaux has argued for the somatic mutation theory against them in print; the exchange is one of the few head-to-head comparisons the field has staged.

Frontier The position most working cancer biologists actually hold is a third one that is rarely written down as a theory: mutation is necessary but not sufficient, and the missing ingredient is promotion. This is the initiation-promotion model of the 1940s, given molecular content by results like the 2023 demonstration that particulate air pollution causes lung adenocarcinoma by inflaming tissue that already carries oncogenic mutations rather than by creating them. Speculative It has the awkward property of being compatible with almost any result, which is a strength as a description and a weakness as a theory.

Speculative The framing this page tests is that the somatic mutation theory is failing and its rival is winning, and it does not survive. What survives is narrower and more useful: the sufficiency claim inside the somatic mutation theory is dead on the evidence, the tissue organization field theory has never been tested at the statistical power its own central experiment requires, and the third position has become dominant without anyone specifying what would refute it. Handwave Three theories, one of which is falsified in its strong form, one of which is untested, and one of which is unfalsifiable, is not a resolved question.

Established Start with the anchor this brief was commissioned on, because checking it produced a correction. The commission cites “Martincorena: thousands of driver mutations per square centimetre of normal skin and oesophagus, NOTCH1 clones at higher frequency in normal tissue than in the cancers”. Established The second half is right and the first half is off by more than an order of magnitude. The 2015 Science paper (348, 880–886, 22 May 2015) sequenced 74 cancer genes across 234 biopsies of sun-exposed eyelid epidermis from four individuals, each biopsy 0.8 to 4.7 mm², and reports two to six mutations per megabase per cell and about 140 driver mutations per square centimetre — not thousands. Established What it does report in the thousands is a different quantity: 3,760 somatic mutations detected across the biopsy set, and, in the 2018 oesophagus work, several hundred mutations per cell in donors in their twenties rising to over 2,000 per cell in older adults. Frontier Total mutations per cell and driver mutations per square centimetre are different measurements and the sweep merged them. The correction does not weaken the point — 140 driver mutations per square centimetre of skin that is not cancer and will not become cancer is already the most damaging number in this subject — but a brief that quotes the wrong one can be dismissed on the number rather than on the argument.

Established The oesophagus half of the anchor verified exactly, and it is the sharper result. Ultra-deep targeted sequencing of 74 cancer genes across 844 normal oesophageal samples from nine transplant donors aged 20 to 75 found NOTCH1 mutations in 30% to 80% of normal oesophagus in middle-aged and elderly people, against about 10% of oesophageal squamous cell carcinomas, with TP53 mutant cells at 5 to 10% of normal epithelium. Frontier A gene called a cancer driver is commoner in the tissue that stays healthy than in the tumour it supposedly drives. Speculative That single comparison is the hardest fact the somatic mutation theory has to accommodate, and accommodating it requires the theory to say that NOTCH1 mutation is selected in normal epithelium and selected against in carcinoma — which is a coherent claim and is not the claim the theory is usually asserted to make.

Established The population-scale version of the measurement arrived in 2025 and it is much larger. Lawson and colleagues applied targeted duplex sequencing at an error rate below five errors per billion base pairs to 1,042 oral epithelium swabs and 371 blood samples from the TwinsUK cohort, median age 68. They find 46 to 49 genes under positive selection and estimate more than 62,000 driver mutations across the sampled clones — NOTCH1 alone accounting for 20,185, TP53 for 7,942, FAT1 for 7,826. Established Mutations accumulate linearly with age at about 18 single-nucleotide variants and 2 indels per cell per year; by ages 65 to 85 roughly 10% of oral cells carry a NOTCH1 driver and 3% carry a TP53 driver. Established Selection is real and measured: dN/dS around 2 for missense and 8 for truncating mutations in driver genes, with negative selection at dN/dS 0.69 against truncating mutations in essential genes. Frontier The authors’ own conclusion is the load-bearing one: clone sizes plateau rather than expanding without limit, which they describe as evidence of major barriers to carcinogenesis downstream of the driver mutation.

Established The cleanest natural experiment on sufficiency is iatrogenic, and it was published in December 2025. Whole-genome analysis of 168 cancer-free tissue samples from 16 organs in 22 patients treated for metastatic cancer measured the mutational cost of treatment directly: platinum chemotherapy adds about 89 mutations per cell per cycle in blood and 25 per cycle in lung; more than 25% of driver mutations in treated normal tissue, TP53 included, were attributable to the treatment. Frontier The corresponding cancer excess is small. Platinum treatment in young adults carries a hazard ratio of about 2.4 for a later solid malignancy, against roughly a 25-fold increase in lung cancer risk from smoking. Speculative A deliberate, measured, massive increase in driver-mutation burden produces a modest increase in cancer, which is as close to a controlled test of the sufficiency claim as human data will ever get.

Established The comparative-biology result points the same way and is independent of all of the above. Somatic mutation rates measured in 208 intestinal crypts from 56 individuals across 16 mammal species range from 47 substitutions per genome per year in humans to 796 in mice, with naked mole-rats at 93 — and the burden at the end of a species’ lifespan varies only by about threefold, mean 3,206 substitutions, across animals whose lifespans differ by orders of magnitude. Established Mutation rate correlates inversely with lifespan at r = 0.85, and body mass barely matters once lifespan is controlled. Frontier Peto’s paradox therefore survives: a whale is not protected from cancer by a lower mutation rate, so whatever protects it is not about mutation supply.

Established Now the other side of the ledger, and it is smaller than the somatic-mutation side by orders of magnitude in both sample size and funding. The central tissue-recombination result is Maffini, Soto, Calabro, Ucci and Sonnenschein in the Journal of Cell Science (117, 1495–1502, 2004). Rat mammary epithelial cells and mammary stroma were exposed to the carcinogen N-nitrosomethylurea separately and recombined in all four combinations. Established Neoplastic transformation occurred only when the stroma had been exposed, whether or not the epithelial cells had been; carcinogen-treated epithelium implanted into untreated stroma formed normal ducts. Frontier The paper does not report per-arm animal numbers or tumour incidences in its abstract, and this session could not retrieve them — which is precisely the problem section 7 is about.

Frontier And the result that cuts against the strong version of the same theory, which neither camp quotes often. Hayward and colleagues (Cancer Research 59, 5002, 1999) built prostate tissue recombinants under the renal capsule of immunodeficient mice: carcinoma-associated fibroblasts combined with initiated human prostatic epithelium produced grafts approaching 500 times the wet weight of any other combination, reaching median masses of 97 to 198 mg against 10 mg or less elsewhere. Established The same carcinoma-associated fibroblasts combined with normal, uninitiated epithelium produced nothing. Speculative In rat mammary gland the stroma was sufficient; in human prostate it was not. Both results are real, both are twenty years old, and no experiment has been run to determine whether the difference is the tissue, the carcinogen, the host, or the sample size.

Speculative A third recombination line supports the stroma-causal claim by genetics rather than by chemistry. Raising expression of the mesenchymal growth factor FGF10 in the stromal compartment alone produces multifocal prostatic intraepithelial neoplasia or adenocarcinoma in the adjacent, genetically unaltered epithelium. Frontier An engineered change confined to the stroma is sufficient to make the epithelium neoplastic, which is a somatic-genetic cause located in the wrong tissue for the standard theory and a stromal cause with an explicit mutation for the rival one.

Established The experiment most often cited as decisive for the tissue theory is the nuclear-transfer result, and its own authors read it the other way. Hochedlinger and colleagues (Genes & Development 18, 1875–1885, 2004) transferred a melanoma nucleus into an enucleated oocyte, derived embryonic stem cells, and made chimeric mice; the melanoma genome contributed to melanocytes, lymphocytes and fibroblasts, so the malignant phenotype was not fixed by the genome alone. Established But 11 of 11 chimeras given the inducing signal developed melanoma at an average latency of 19 days, against 25% penetrance and 60-day latency in the donor model, and a third also developed rhabdomyosarcoma. Frontier The authors concluded that irreversible chromosomal alterations carried by the donor nucleus contributed to tumorigenic potential. Speculative The experiment shows reprogramming suppresses the phenotype and the genome still predisposes, which is the honest reading and is not what either camp quotes.

Established The strongest recent result for a non-mutational origin is in flies, and it is clean. Parreno and colleagues in Giacomo Cavalli’s group (Nature 629, 688–696, 2024) depleted the Polycomb protein PH from Drosophila larval eye discs for 24 hours and restored it within 48; 100% of eye discs became tumours, with loss of polarity and of differentiation markers. Established Whole-genome sequencing of 12 tumours found 92.8% of variants below allele frequency 0.2 and none shared between tumours — no recurrent driver. The JAK–STAT ligands upd1, upd2 and upd3, plus zfh1 and chinmo, stayed derepressed after the protein returned. Speculative A transient, fully reversible perturbation with no mutation produced a self-sustaining cancer state at complete penetrance. Handwave In an insect eye disc, with a genome and a tissue architecture unlike a mammal’s, and nobody has repeated it in a vertebrate.

Established On the promotion side the 2023 air-pollution result is the most quantitative thing the third position has. Hill, Lim, Weeden and colleagues (Nature 616, 159–167, 5 April 2023) related particulate matter under 2.5 µm to the incidence of EGFR-driven lung cancer across 32,957 cases in four within-country cohorts, and found oncogenic EGFR mutations in 18% of histologically normal lung samples and KRAS mutations in 53%, across 295 individuals in three cohorts. Established In mice the pollutant drew macrophages into the lung, released interleukin-1β, and pushed EGFR-mutant alveolar type II cells into a progenitor-like state; an exploratory arm of a cardiovascular trial of interleukin-1β blockade had already reported reduced incident lung cancer. Frontier The authors argue explicitly for promotion rather than initiation: the mutations were already there and the exposure acted on the tissue.

Established Two further measured results undercut a premise both theories share. Single-cell and barcoding analysis of 260,922 cells from normal, inflamed and neoplastic intestine, with sequencing of human sporadic colorectal polyps, found those polyps to be polyclonal in origin, with a later polyclonal-to-monoclonal transition as lesions advance. Established And the elimination machinery is measurable: normal epithelial cells clear 70 to 90% of RasV12-transformed neighbours at a one-in-ten ratio, with disruption of the LILRB3 recognition pathway cutting clearance by about 70% and EphA2 silencing by about 50%. Speculative If precancers start from many lineages and normal tissue routinely deletes transformed cells, then “a single cell acquires mutations and becomes a tumour” is not a description of the usual case.

Frontier The epidemiology has a fact neither theory predicts and both must explain. Cancer incidence in young adults is rising on a birth-cohort pattern: early-onset incidence up more than 15% since 2000, colorectal incidence in Americans born in the 1990s roughly twice that of those born in the 1940s at the same age, with clonal expansion to carcinoma apparently faster in early-onset colorectal cancer than in later-onset disease. Speculative Somatic mutation burden accumulates with age and cannot by itself produce a birth-cohort effect; something that changed in the mid-twentieth century changed the risk, and the candidates on offer — diet, obesity, microbiome, specific exposures — are promotion-side variables.

Speculative And the record of what is not established should be as flat as the record of what is. No experiment has shown a human cancer arising with no contributing somatic mutation. No experiment has shown stroma-only carcinogen exposure producing tumours in a modern, powered, pre-registered design with the resulting tumours sequenced. Speculative The intrinsic-versus-extrinsic split remains contested between under 10 to 30% intrinsic and 70 to 90% extrinsic, and the dispute is about estimators rather than data. Handwave Anyone who tells you this question has been settled in either direction is describing their field’s consensus rather than the evidence.

Frontier The live positions are unusually well separated, and several are held by people who agree about every measurement and nothing else. They are set out as positions with holders.

Frontier Position one: the somatic mutation theory, in its necessary-and-sufficient form. Cancer is a clonal genetic disease; find the drivers, target them. Established Its strongest evidence is therapeutic and is not in dispute: translocation-defined leukaemias, and drugs that work by inhibiting the product of a named mutant gene. Speculative The sufficiency half is dead on the numbers in section 3 — 140 drivers per square centimetre of normal skin, NOTCH1 commoner in healthy oesophagus than in oesophageal cancer, a quarter of drivers in treated normal tissue attributable to chemotherapy at a hazard ratio of 2.4. Frontier Almost nobody defends sufficiency in print, and the theory is nonetheless taught, funded and reported as though sufficiency held.

Speculative Position two: the tissue organization field theory. Carcinogenesis is development gone wrong at the level of tissue organisation; the default state of cells is proliferation; carcinogens act on the stroma; mutations are consequence or epiphenomenon. Frontier Its experimental case is three or four recombination results, of which the strongest is twenty years old and the best-known counterexample is in the same technique. Handwave Its weakest move is rhetorical rather than empirical: treating every somatic-mutation anomaly as support, when an anomaly for one theory is not evidence for another unless that other predicted it. Speculative Its strongest move is to point out that its central experiment has never been repeated at modern power, which is true and is somebody’s fault.

Frontier Position three: mutation is necessary but not sufficient, and promotion supplies the rest. This is where most working cancer biologists sit, and the 2023 air-pollution result is its best evidence: particulate matter acting on lung that already carries oncogenic EGFR mutations in 18% of normal samples, through macrophage recruitment and interleukin-1β. Speculative It absorbs the normal-tissue-mutation data without strain and it absorbs the stromal data without strain, which is the problem. Handwave No version of this position currently states a result that would refute it, and a theory that accommodates every observation is doing description rather than explanation.

Speculative Position four: origins can be purely epigenetic. A 24-hour depletion of one Polycomb protein produced tumours in every Drosophila eye disc tested, with no recurrent mutation across twelve sequenced tumours and permanent derepression of JAK–STAT ligands after the protein returned. Frontier Douglas Hanahan added “nonmutational epigenetic reprogramming” to the hallmarks framework in 2022 as an enabling characteristic distinct from genome instability, which is the mainstream conceding the category. Handwave The distance from an insect imaginal disc to a mammalian epithelium is the whole question, and nobody has crossed it.

Frontier Position five, the quietest and possibly the most important: the binding variable is selection, not mutation supply. Mutation burden at end of lifespan varies only about threefold across sixteen mammal species; oral clone sizes plateau rather than expanding; normal cells delete 70 to 90% of transformed neighbours at low mutant density. Speculative On this reading cancer risk is set by how strongly a tissue polices its clones, and mutation is the abundant input rather than the scarce one. Speculative It is compatible with both major theories and reframes both, which is why it is rarely presented as a position at all.

Speculative Position six is about the logic rather than the evidence: are the two theories even incompatible? Philosophers of science have argued that the somatic mutation theory and the tissue organization field theory address different levels of description and can be nested; Sonnenschein and Soto reject that as a way of absorbing their theory without testing it. Handwave This page notes that the reconciliation move is only available if somebody states which observations each theory forbids, and neither camp has done it.

Frontier Position seven is a live quantitative dispute with policy consequences. Tomasetti and Vogelstein attributed much of the variation in cancer risk between tissues to the number of stem-cell divisions — replication error, or “bad luck”. Established Wu, Powers, Zhu and Hannun replied in Nature 529 that intrinsic factors account for under 10 to 30% of lifetime risk, putting extrinsic factors at 70 to 90%, using mathematical modelling, mutation-accumulation rates and immigrant-cohort epidemiology. Frontier Tomasetti and colleagues replied that one of those methods returns over 85% extrinsic even in constructed cases where extrinsic factors have no role at all. Speculative The exchange is unresolved, it is about estimators rather than measurements, and both sides’ press coverage was worse than either paper.

Speculative Position eight: the single-cell-of-origin premise, which both major theories share, may be wrong. Human sporadic colorectal polyps are polyclonal in origin, becoming monoclonal as they advance. Handwave If early lesions routinely begin as cooperating multi-lineage populations, then the mutational-clonal-expansion narrative and the field-disruption narrative are both approximations of something with a different shape, and neither camp has a version of its theory written for that case.

Frontier The binding bottleneck is that the experiment which discriminates between the two theories was last run at 2004 scale, and nobody has rerun it. Tissue recombination is, in the words of the standard protocol, labour intensive and technically demanding, with biases from physical dissociation, viral transduction and transplantation; there is no standard design, no power analysis in the literature, and no multi-laboratory replication. Speculative The two landmark results point in opposite directions — stroma sufficient in rat mammary gland, stroma insufficient in human prostate — and the difference has never been investigated. Handwave A field that has spent two decades arguing about an experiment it has not repeated is arguing about a memory.

Frontier The second is that initiation is unobservable in humans and the field says so in its own reviews. The transition from normal through precancer to cancer within intact tissue remains poorly understood largely because early-stage clinical samples and suitable models are scarce; everything in section 3 is either a cross-sectional snapshot of normal tissue or a retrospective reconstruction from a tumour that already exists. Speculative No one has ever watched a human cancer begin.

Established The third is methodological and specific: the measures the field has detect selection, not causation. dN/dS ratios of 2 for missense and 8 for truncating variants in driver genes establish that these mutations confer a proliferative advantage in normal tissue. Frontier They do not establish that the advantage leads anywhere, and the same analysis finds clone sizes plateauing. Speculative A mutation can be under strong positive selection in a tissue and be irrelevant to that tissue’s cancer risk, and NOTCH1 in oesophagus appears to be exactly that.

Frontier The fourth is that there is no assay for the non-mutational half of carcinogenesis. The Ames test and its descendants detect genotoxicity cheaply and at scale; nothing comparable exists for tumour promotion, epigenetic reprogramming, chronic inflammation or receptor-mediated effects, which is why regulatory classification of non-genotoxic carcinogens rests on two-year rodent bioassays and expert weight-of-evidence judgement. Speculative A theory of carcinogenesis that cannot be operationalised as a screening assay does not reach the point of regulatory use, whatever its scientific merits.

Speculative The fifth is the model-organism gap, and the Polycomb result shows its shape exactly. A 24-hour perturbation producing 100% tumour penetrance with no recurrent mutation is the cleanest demonstration of non-mutational origin anywhere — in a fly eye disc, a tissue with no stroma in the mammalian sense, no immune surveillance of the mammalian kind and a genome under different constraint. Handwave Nobody has proposed the mammalian version, and until somebody does the result is an existence proof about insects.

Handwave The sixth is not technical: neither theory has stated what would refute it, so the argument cannot end. The somatic mutation theory retreats from sufficiency to necessity when challenged; the tissue theory absorbs every somatic-mutation anomaly as support; the promotion synthesis accommodates everything by construction. Speculative This is a solvable institutional problem, it costs nothing, and section 10 records writing the convention as a formal requirement.

Established This subject waits on measurement capability rather than on results other briefs produce, and the capability has been arriving fast. Duplex sequencing at error rates below five per billion base pairs made it possible to count driver mutations in a buccal swab rather than in a tumour; single-cell and spatial multi-omics, lineage tracing and barcoding made polyclonality detectable at all. Frontier Every destabilising result in section 3 is younger than the instrument that produced it, which is the usual pattern when a paradigm starts to leak.

Frontier One dependency is a model system nobody on this map produces: a mammalian tissue in which epithelium and stroma can be manipulated independently, longitudinally, and at a sample size that supports a pre-registered comparison. The microphysiological platforms assessed in Microphysiological Systems are the nearest thing and were built for toxicology and drug response rather than for carcinogenesis; Lab-Grown Organs supplies the tissue-construction technique. Speculative Neither community is trying to settle a theory dispute, and their constructs would be worth more to this question for it.

Speculative A second dependency is exposure measurement, and it is the binding one for the promotion position. A birth-cohort effect in early-onset cancer is a statement about what changed in the environment, and testing any specific candidate requires the individual-level, longitudinal exposure record that Exposomics is built to produce and has not yet produced at cohort scale. Frontier Without it, “promotion by modern exposures” is a hypothesis with a shape and no subject.

Speculative Traffic with the rest of this map runs mostly outward. Precision Medicine inherits the consequences directly, because a companion diagnostic is a bet that a mutation is causal in the tissue where it is found; Cellular Rejuvenation and Longevity Therapies inherit the clonal-expansion measurements as an ageing phenotype; Immune Engineering inherits the surveillance half of the selection story. Handwave Nothing on this map blocks this brief. What blocks it is one experiment nobody has funded.

Frontier Every position in section 4 has a stated test, and they rank by what a result would settle rather than by how hard they are.

Speculative Tier one, and the decisive experiment for this brief: stroma-only carcinogen exposure in a tissue-recombination design at modern statistical power, pre-registered, in more than one tissue and more than one laboratory, with the resulting tumours whole-genome sequenced. The commission named this test and it is the right one. Established Delivered: the 2004 rat mammary result in which neoplastic transformation occurred only when the stroma had been exposed to N-nitrosomethylurea, regardless of whether the epithelium had; the 1999 prostate result in which carcinoma-associated fibroblasts grew initiated epithelium by up to 500-fold and did nothing at all to normal epithelium; and the FGF10 result in which a stroma-confined genetic change makes adjacent unaltered epithelium neoplastic. Speculative Not delivered: any of them at modern power, any of them pre-registered, any of them replicated across laboratories, and — the omission that matters most — sequencing of the tumours that arose from unexposed epithelium. Handwave The experiment is therefore not a new idea but a repetition with instruments that did not exist in 2004, which is exactly the kind of work no funding mechanism is designed to buy.

Frontier Tier two is one addition to tier one and it may be worth more than the rest of the design: sequence the tumours from the stroma-only arm. If tumours arising in carcinogen-naive epithelium implanted into carcinogen-exposed stroma carry canonical driver mutations, the tissue theory has been shown to operate through mutation and the argument narrows enormously. Speculative If they carry none, the somatic mutation theory has a counterexample in a mammal that it currently does not have. Handwave Neither answer has been obtained because the 2004 experiment predates routine tumour sequencing and nobody has gone back.

Established Tier three is a natural experiment already running, with its first readout published in December 2025. Cancer treatment deposits a measured, enormous mutational load in normal tissue — about 89 mutations per cell per platinum cycle in blood, over a quarter of driver mutations in exposed normal tissue attributable to treatment — in a population that is then followed for decades. Frontier The hazard ratio for a later solid malignancy after platinum in young adults is about 2.4, against roughly 25-fold for smoking and lung cancer. Speculative Prospective follow-up of treated survivors, stratified by measured induced mutation burden, is the largest and cheapest test of the sufficiency claim available to anyone, and it requires no new laboratory at all.

Speculative Tier four tests the promotion position where it makes its sharpest prediction: block the promoter and watch incidence. Interleukin-1β blockade reduced incident lung cancer in an exploratory arm of a cardiovascular outcomes trial, and the mechanism proposed in 2023 — macrophage recruitment and a progenitor-like state in EGFR-mutant alveolar cells — predicts that the effect should be concentrated in people whose normal lung already carries those mutations, which is 18% of sampled normal lung. Frontier A prevention trial stratified on normal-tissue mutation status has never been run, and it is the experiment that would make promotion a theory with a prediction rather than a description.

Speculative Tier five is the mammalian version of the Polycomb result. A transient, fully reversible epigenetic perturbation in a mouse epithelium, followed by whole-genome sequencing of any tumours that arise, to establish whether purely epigenetic initiation happens in an animal with a stroma and an immune system. Handwave The fly result gives 100% penetrance from a 24-hour depletion with no recurrent mutation; nobody has stated what the mammalian experiment would even look like, and a failure to reproduce would be as informative as a success.

Speculative Tier six manipulates the elimination machinery instead of the mutation supply. Normal cells clear 70 to 90% of transformed neighbours at low mutant density, and disrupting LILRB3 or EphA2 cuts that clearance by 70% and 50% respectively. Frontier The discriminating experiment is to disable cell competition in vivo at fixed mutation burden and measure cancer incidence. Speculative If incidence rises sharply with no change in mutation, the binding variable is policing rather than damage, which is position five in section 4 and currently has no direct in vivo test.

Frontier Tier seven is a counting exercise with a large payoff: how many human cancers are polyclonal in origin? Colorectal polyps are; the question is whether that generalises. Speculative Both major theories are written for a single cell of origin, so a high polyclonal fraction would require both of them to be rewritten rather than one of them to be chosen.

Handwave Tier eight costs nothing and is the precondition for the rest: a pre-registered statement, from each camp, of the observations its theory forbids. Speculative Until that exists, every result in section 3 will be claimed by everyone, which is what has happened to all of them so far.

Established And the experiments that are not tests of this question and are regularly reported as though they were. The clinical success of a drug targeting a mutant protein shows the mutation is a good therapeutic target in an established tumour; it does not show the mutation caused the tumour. Speculative Mutational signatures show which exposure damaged the DNA, not whether that damage was what made the cell cancerous. Handwave Both are routinely offered as settling the causal question, and neither addresses it.

Frontier There is no engineering payoff here and this page will not manufacture one. Every requirement below is a requirement for an instrument or a study design that would answer a question, which is why section 10 records a funding condition as a formal requirement.

Speculative The specification for the decisive test, assembled from what the component experiments actually used. Tissue recombinants of dissociated epithelium and stroma in collagen, implanted under the renal capsule of immunodeficient hosts, read out at 20 to 85 days; all four exposure combinations run concurrently; at minimum two tissues, since rat mammary gland and human prostate already disagree; group sizes set by a published power calculation rather than by convenience, against the historical arms of five to nine grafts; pre-registered primary endpoint; and whole-genome sequencing of every lesion, against matched normal from the same donor. Handwave Nothing in that list is technically hard. All of it is procedurally unusual, and that is why it has not happened.

Established The measurement half of the specification is already routine and genuinely impressive. Duplex sequencing at under five errors per billion base pairs, with cumulative duplex coverage in the hundreds of thousands, applied to a non-invasive buccal swab; whole-genome analysis of 168 normal samples across 16 organs resolving 166,732 substitutions; barcoding and single-cell analysis of 260,922 intestinal cells. Frontier The field can now count driver mutations in healthy tissue far more precisely than it can say what they mean.

Speculative The requirement nobody has specified is the assay. Regulatory carcinogen identification needs a cheap, scalable, validated in vitro test for non-genotoxic carcinogenesis to sit alongside genotoxicity testing, and no such test has been validated. Handwave Until one exists, the seven of IARC’s ten key characteristics that are not about DNA reactivity are assessed from bioassays and expert judgement rather than from a bench method, and the practical theory of carcinogenesis in use by regulators is whatever the available assays can see.

Established The nearest brief on this map is Precision Medicine, and the two are consecutive rather than overlapping. That page assesses what happens once you accept that a mutation is the treatable thing: companion diagnostics as a regulated device class, the oncology funnel from 38% sequenced-and-matchable down to about 1.5% end-to-end benefit, and the asymmetry between predicting harm and predicting benefit. Frontier This page asks the question one step upstream — whether the mutation is the causal thing at all — and the seam is sharp: every companion diagnostic is a bet that a mutation found in a tissue is driving what happens in that tissue, and section 3 shows that bet failing in normal oesophagus.

Speculative Three neighbours supply instruments rather than arguments. Microphysiological Systems and Lab-Grown Organs build the epithelium-plus-stroma constructs the decisive experiment needs, for toxicology and transplantation reasons; Exposomics would supply the individual-level exposure record that any promotion hypothesis requires. Frontier None of the three is trying to settle a theory dispute, which is why their measurements would count for more here than a partisan’s.

Frontier Two neighbours inherit the result directly. Cellular Rejuvenation and Longevity Therapies take on the awkward corollary of the clonal-expansion data: a tissue full of positively selected driver clones is the normal state of an old body, so any intervention that improves the fitness of old cells is also an intervention on those clones. Speculative Immune Engineering owns the surveillance half of the selection story, and Microbiome Engineering owns one of the promotion-side variables that the hallmarks framework formally adopted in 2022.

Frontier Two adjacencies are argumentative rather than technical, and they are the ones this page leans on. Philosophy of Science owns the machinery this dispute needs and does not use: whether two theories at different levels of description are rivals or nested, and what it means that neither states what would refute it. Speculative And the sibling briefs in this category share the same structural defect in unrelated subjects — Objective Collapse Models is a theory family whose parameters can be extended faster than experiment removes them, and Prebiotic Chemical Networks is a field whose central predicate cannot be failed. Handwave Three pre-paradigmatic subjects, three missing refutation conditions, and in each case writing one would cost nothing.

Established The institutional fact that most changes the shape of this argument is that the regulators already left strict somatic mutation theory behind, and the research funders did not. The IARC Monographs programme evaluates carcinogenic hazard against ten key characteristics of carcinogens, consolidated from two 2012 workshops and twenty-four mechanistic endpoints. Established Only three of the ten concern DNA reactivity or genomic instability. The other seven are epigenetic alteration, oxidative stress, chronic inflammation, immunosuppression, receptor-mediated effects, immortalisation, and altered cell proliferation, death or nutrient supply — and the framework is now used by the US Environmental Protection Agency and the National Toxicology Program as well. Frontier The operational theory of carcinogenesis in regulatory use is already a promotion-inclusive one; the operational theory in grant review is not.

Established The scale that makes the allocation question serious is in the IARC’s own estimates: 20.6 million new cases and 9.8 million deaths in 2024, projected to 34.4 million cases by 2050, a 67% increase, with lung cancer leading at 12.8% of cases and 19.1% of deaths. Speculative Almost all of that growth is in countries that will not be able to buy targeted therapy at current prices, which makes the theory question a distributional question rather than an academic one.

Frontier The funding structure has a specific blind spot and the decisive experiment falls straight into it. A pre-registered, multi-laboratory, multi-tissue repetition of a 2004 recombination result is not target discovery, not a trial, not a technology, and produces no therapeutic asset; it is theory testing, which no major cancer funder has a mechanism for. Speculative The result is that the single experiment both camps agree would be informative has gone unrun for two decades while both camps continued to publish about it.

Speculative There are three institutional requirements this field lacks, and none of them is money. The first is a refutation condition from each camp, written before the next result. Speculative The second is a venue and a norm for theory-testing replication in cancer biology, since the recombination literature has no ring trial, no standard protocol and no published power analysis. Handwave The third is a validated in vitro assay for non-genotoxic carcinogenesis, without which regulatory practice will keep resting on two-year rodent bioassays and expert judgement while the science argues about mechanisms no screen can see.

Frontier The largest hazard here is not a laboratory risk. It is that the causal theory determines who benefits, and the two candidate theories favour different populations. A mutation-centric programme produces targeted agents priced per patient and companion diagnostics that require sequencing infrastructure; a promotion-centric programme produces air-quality regulation, vaccination and food policy, which cost little per head and work without a health system. Speculative With 34.4 million annual cases projected by 2050 and the growth concentrated where oncology is least available, the choice of theory is a choice about which half of the world a research budget is for.

Established Second, the public-messaging failure around the intrinsic-versus-extrinsic dispute, which is the clearest case in modern oncology of a technical disagreement being read as a moral one. The stem-cell-division analysis was widely reported as showing that most cancer is bad luck, and the rebuttal putting extrinsic factors at 70 to 90% was read as restoring blame to behaviour. Frontier Neither paper said either thing, both were arguing about how to partition variance between tissues, and the public conclusion — that prevention is or is not worth doing — does not follow from either. Handwave A field whose central quantitative dispute reliably converts into a debate about whose fault cancer is has a communication problem that is also an ethical one.

Speculative Third, overdiagnosis, which the normal-tissue data turns from a screening problem into a conceptual one. If 140 driver mutations sit in every square centimetre of middle-aged eyelid skin, if NOTCH1 clones occupy 30 to 80% of normal oesophagus, and if clone sizes plateau rather than progressing, then a molecular early-detection test sensitive enough to find driver mutations will find enormous numbers of lesions that would never have harmed anyone. Frontier The measurements that undermine the sufficiency claim are the same measurements that set the false-positive floor for every liquid-biopsy programme, and the detection field and the theory field mostly do not talk.

Speculative Fourth, a point about how a minority theory is treated, stated carefully because it cuts both ways. The tissue organization field theory has been marginal for thirty years, its proponents publish in review venues more than in experimental ones, and their central experiment has not been repeated — which can be read as the field correctly declining to chase a weak hypothesis, or as a hypothesis never given the resources to be tested. Handwave The distinguishing evidence would be a funded, powered attempt that failed, and it does not exist. Speculative Refusing to run a cheap discriminating experiment and then citing the absence of evidence from it is the specific failure mode this page is most concerned about.

Frontier Fifth, and to the field’s credit: the results most damaging to the dominant theory were produced by its own strongest practitioners. The normal-skin and normal-oesophagus measurements came out of the same cancer-genomics programme whose premise they complicate, and the 2025 population-scale work concluding that clone growth plateaus is from the same group. Established Interest running that hard against a finding is the strongest evidential signal available in any field, and it should be said out loud.

Speculative The concrete consequence of settling this is a reallocation, not a discovery. If mutation is the binding constraint, the future of cancer control is molecular, expensive, individual and post-diagnostic. If promotion is, it is environmental, cheap, collective and pre-diagnostic. Frontier The interventions that have actually reduced population cancer incidence so far — tobacco control, hepatitis B and human papillomavirus vaccination — are all of the second kind, which is a fact the first framing has never had to explain.

Speculative The second consequence is about ageing societies and it is uncomfortable. A body in its seventies is a mosaic of positively selected driver clones: about 10% of oral cells carrying a NOTCH1 driver, 3% a TP53 driver, most of the oesophagus colonised. Speculative If clone sizes plateau because tissue policing holds, then anything that weakens that policing — immunosuppression, chronic inflammation, or possibly some rejuvenation interventions — releases a population of clones that is already there. Handwave That is a prediction nobody has tested and a risk nobody has priced, and it belongs to Longevity Therapies as much as to this page.

Speculative A third consequence is epistemic and generalises beyond cancer. The somatic mutation theory is the best-funded, best-instrumented causal theory in biomedicine, and its sufficiency claim was falsified by the very sequencing programme built to confirm it. Frontier The lesson available to any field with a dominant theory and a new instrument is that the instrument usually breaks the theory before it confirms it, and the interesting question is how long the field takes to say so.

Speculative And the outcome more likely than either resolution: no decisive experiment is run, and the theories converge by attrition. The somatic mutation theory quietly abandons sufficiency, the tissue theory is absorbed as “the microenvironment matters”, and cancer biology proceeds with a description that fits everything and forbids nothing. Frontier That is the modal outcome, it costs nothing, and it leaves the 2050 burden projection exactly where it was.

These horizons track measurement capability and theory closure rather than any engineering programme, because there is no engineering programme here to track:

• 10 yr: Frontier Expect normal-tissue mutation atlases to extend from skin, oesophagus, blood and oral epithelium to most accessible organs, at population rather than donor scale, and expect them to keep finding driver clones where there is no disease. Speculative Expect the prospective follow-up of chemotherapy-induced mutation burden to deliver its first hard sufficiency readout, since the exposure is measured and the cohort already exists. Speculative Do not expect the tissue-recombination repetition, because nothing in the current funding structure buys it.
• 25 yr: Speculative This is the window in which a mammalian test of purely epigenetic initiation plausibly runs, and in which cell-competition manipulation at fixed mutation burden becomes routine enough to answer whether policing or damage is the binding variable. Speculative A positive result on either would move the field’s centre of gravity more than any single genomic dataset has. Handwave A null on both would leave the somatic mutation theory where it is now: necessary, insufficient, and unreplaced.
• 50 yr: Handwave On this horizon the plausible development is consolidation rather than victory: a quantitative model in which mutation supply, tissue policing and promotion each carry a measured share of risk, with the shares differing by organ. Speculative That would end the theory dispute by dissolving it, which is how most disputes of this shape have ended, and it would require the experiments in section 7 to have been run rather than argued about.
• 100 / 250+ yr: Handwave Beyond useful forecasting. The defensible statement is about shape: a field whose dominant theory was falsified in its strong form by its own instruments either states a successor or stops making causal claims. Handwave Any specific claim about what a settled theory of carcinogenesis would deliver clinically is an assertion, and this page does not make one.

• Depends on No typed dependency edge to another brief on this map, and the absence is informative rather than an omission: what this subject waits on is one repeated experiment and one assay, neither of which any brief here produces. Microphysiological Systems and Lab-Grown Organs build the epithelium-and-stroma constructs the decisive test needs but are not blocking, because those constructs are being built for toxicology and transplantation regardless. Precision Medicine is the nearest brief by subject and runs downstream of this one: it assesses what follows from treating a mutation as the actionable thing, where this page asks whether it is the causal thing.
• Requires (not on this map) Six conditions, in the order of the tokens above. The repetition. The 2004 rat mammary result and the 1999 prostate result disagree about whether stroma alone suffices, and neither has been repeated at modern power, pre-registered, or run across laboratories. The sequencing. Nobody has sequenced tumours arising from carcinogen-naive epithelium in carcinogen-exposed stroma, and that single addition would tell you whether the tissue route runs through mutation or around it. The refutation conditions. Neither camp has written down what its theory forbids, which is why every result in section 3 has been claimed by everyone. The assay. Genotoxicity has the Ames test; the seven of IARC’s ten key characteristics that are not about DNA reactivity have two-year rodent bioassays and expert judgement. The mammalian epigenetic test. A 24-hour Polycomb depletion gives 100% tumour penetrance with no recurrent mutation in a fly eye disc, and no one has proposed the vertebrate version. The money. A funding mechanism for testing a theory rather than finding a target, which no major cancer funder currently operates.
• Enables A quantitative partition of cancer risk between mutation supply, tissue policing and promotion, which is what prevention policy needs and does not have, and a false-positive floor for molecular early detection, which Precision Medicine needs and does not have. Downstream, Exposomics and Immune Engineering inherit a sharper target list. No typed enabling edge is claimed, because a theory dispute that has not been adjudicated enables nothing on a schedule.
• Adjacent Within this map: Precision Medicine for what follows downstream, Microphysiological Systems and Lab-Grown Organs for the constructs, Exposomics for exposure measurement, Cellular Rejuvenation and Longevity Therapies for the ageing-mosaic corollary, Immune Engineering and Microbiome Engineering for two promotion-side variables, Philosophy of Science for the theory-choice machinery, and the sibling Category XI briefs Objective Collapse Models and Prebiotic Chemical Networks for the same missing-refutation-condition pathology in other subjects. Outside the map: cancer epidemiology, regulatory toxicology, duplex sequencing, developmental biology and the philosophy of levels of explanation supply every number on this page.

Established “Cancer is a genetic disease.” Cancer cells carry mutations and those mutations matter therapeutically; that is not the same claim. Established Normal sun-exposed eyelid skin carries about 140 driver mutations per square centimetre and 18 to 32% of its cells carry positively selected cancer-gene mutations; normal oesophagus carries NOTCH1 mutant clones in 30 to 80% of tissue by middle age. Speculative The defensible statement is that cancer is a genetic disease in the same sense that a fire is a chemical reaction: true, and not an account of why this building burned and the one next door did not.

Established “Driver mutations drive cancer.” NOTCH1 mutations are found in 30 to 80% of normal middle-aged oesophagus and in roughly 10% of oesophageal squamous cell carcinomas. Frontier A gene under strong positive selection in normal tissue can be selected against during carcinoma development. Speculative “Driver” is a label for a mutation with a proliferative advantage in some context, and the field kept the word after the evidence stopped supporting the metaphor.

Established “More mutations means more cancer.” Platinum chemotherapy adds about 89 mutations per cell per cycle in blood and accounts for over a quarter of driver mutations in exposed normal tissue, and the subsequent hazard ratio for a solid malignancy in young adults is about 2.4 — against roughly 25-fold for smoking. Established Across sixteen mammal species the mutation burden at end of lifespan varies only about threefold while lifespan varies by an order of magnitude. Speculative Mutation supply is abundant; whatever is scarce is something else.

Frontier “The somatic mutation theory has been refuted.” Its sufficiency claim has. Its necessity claim has not: no human cancer has been shown to arise with no contributing somatic mutation, and the therapeutic record of targeting mutant proteins is real. Speculative What is true is that the version taught, funded and reported is the sufficiency version, and it is the one the evidence does not support.

Speculative “The tissue organization field theory has been refuted.” It has not been tested at the power its own central experiment requires. The 2004 rat mammary result stands unrepeated; the 1999 prostate result, in which carcinoma-associated fibroblasts did nothing to normal epithelium, is the strongest counterexample and is also unrepeated. Handwave Two unreplicated experiments pointing opposite ways is not a refutation, it is an unfinished argument, and both camps have treated the silence as agreement with themselves.

Frontier “Nuclear transfer proved cancer is epigenetic.” A melanoma nucleus transferred into an oocyte produced embryonic stem cells that made melanocytes, lymphocytes and fibroblasts, so the malignant phenotype was suppressed by reprogramming. Established In the same paper, 11 of 11 chimeras developed melanoma at an average 19-day latency against 25% penetrance and 60 days in the donor model, and the authors concluded that irreversible chromosomal alterations contributed. Speculative The experiment is quoted by the tissue camp for its first half and by nobody for its second.

Speculative “Cancer begins when one cell goes wrong.” Human sporadic colorectal polyps are polyclonal in origin, with a polyclonal-to-monoclonal transition as lesions advance. Handwave Both major theories assume a single cell of origin, so this is a premise they share and one that the data is starting to take away from both of them.

Established “Most cancer is bad luck” and “most cancer is preventable” are both misreadings of the same argument. The dispute is over how to partition between-tissue variation in risk, with one side putting intrinsic replication error high and the other putting extrinsic factors at 70 to 90%, and a subsequent reply arguing that one of the estimators returns over 85% extrinsic even in constructed cases where extrinsic factors contribute nothing. Frontier It is a methodological dispute about an estimator, it is unresolved, and neither answer tells an individual whose fault anything is.

Frontier “Regulators classify carcinogens by whether they damage DNA.” The IARC framework uses ten key characteristics, of which three concern electrophilicity, genotoxicity and genomic instability and seven do not — epigenetic alteration, oxidative stress, chronic inflammation, immunosuppression, receptor-mediated effects, immortalisation, and altered proliferation, death or nutrient supply. Speculative The regulatory apparatus adopted a promotion-inclusive theory of carcinogenesis over a decade ago, and the research funding structure has not followed.

Established And one correction this page makes to the summary of this subject in circulation, including in its own commission. The received statement is that Martincorena found thousands of driver mutations per square centimetre of normal skin and oesophagus, with NOTCH1 clones commoner in normal tissue than in the cancers, and that the decisive next step is tissue recombination and stroma-only carcinogen exposure at modern statistical power. Established The NOTCH1 half is exactly right. The density is not: the measured figure is about 140 driver mutations per square centimetre of eyelid epidermis, and the thousands belong to a different quantity — total somatic mutations per cell, which exceeds 2,000 in older adults’ oesophagus. Speculative The decisive experiment is correctly named and is not half delivered so much as half abandoned: it was run twice, in 1999 and 2004, in two tissues, with opposite results, at sample sizes in the single digits, before tumour sequencing was routine. Handwave The reason it has not been redone is not that it is hard. It is that repeating a twenty-year-old experiment to adjudicate a theory is not a fundable activity, which makes this a failure of institutions rather than of instruments.