We thought cancer was cells waking up. The evidence says the opposite.

Macheng Shen × agent workflow · 2026-08-02 · cognitive state: the originating hypothesis is retired; one surviving distinction is speculative

中文摘要

起点直觉:多细胞体把细胞「驯化」到极致 —— 压制增殖、剥夺其繁殖未来 —— 压过头就会反弹,癌症是这场造反。核查后的判决:大约六成是别人已经做了十五年的正确想法(Levin 的认知光锥收缩);剩下四成真正原创的部分,符号是反的。所有能查的证据都指向控制↓ ⟹ 癌↑,没有一条指向控制↑ ⟹ 造反:免疫抑制(器官移植受者标准化发病比 2.10,32 类恶性肿瘤升高)、TP53 胚系缺失(70 岁前累积发病率接近 100%)、缝隙连接解偶联、慢性炎症 —— 全是组织尺度控制的丧失。最干净的一刀:成熟皮层神经元与心肌细胞是体内被驯化得最彻底的细胞(永久退出细胞周期),而它们几乎不发生恶性转化 —— 脑瘤来自胶质细胞或未成熟前体,原发心脏恶性肿瘤绝大多数是间叶来源肉瘤。**最大驯化 = 最小癌症风险。**

另外三条不舒服但重要的事实:(1) 肿瘤内部不存在一个「自我」可以觉醒 —— 肿瘤是克隆间的战争,更狠的作弊者在肿瘤内部演化并寄生亲代肿瘤;(2) 生命史上约 8 条癌细胞谱系确实逃出了宿主并作为独立生物存活至今(犬传染性性病肿瘤可追溯到约 11,000 年前的一只狗;塔斯马尼亚袋獾面部肿瘤;多种双壳贝类中多条独立谱系,含跨物种传播)—— 有出口则 11,000 年,无出口则与宿主同归于尽;(3) 分化疗法六十年只有一个干净的临床胜利(急性早幼粒细胞白血病,全反式维甲酸+三氧化二砷,50 个月总生存 99.2% vs 92.6%),而该领域权威本人认为治愈来自「降解一个可成药的融合驱动蛋白 + 自我更新丧失」,不是来自分化本身;IDH 抑制剂确实是分化疗法,但三期随机试验为阴性(总生存 6.5 vs 6.2 个月,HR 0.86,P=0.23);生物电「正常化」在哺乳动物里符号是反的 —— 在爪蟾里超极化抑制肿瘤样结构 29–44%,在小鼠三阴性乳腺癌里超极化增加侵袭与肺转移,治疗性试剂是钾通道阻断剂;而「把癌细胞重新连回组织」在胶质瘤里是帮倒忙,唯一进入临床试验的连接蛋白靶向药是阻断剂。全文 55 条文献逐条核实。

1 · What is already mainstream — the part we were re-inventing

The framing closest to our starting intuition, and considerably better developed, is Levin's: in cancer the cell's cognitive light cone — the spatial and temporal scale of the goals it can pursue — contracts. The cell does not become a more ambitious agent. It becomes a smaller one, reverting to the scale at which a single-celled organism operates [1][2]. [minority position] — this is a live, funded research programme, not textbook oncology.

That formulation matters because it inverts the moral picture. "Cancer cells are not more selfish; they have smaller selves." Nothing wakes up.

Three other established framings do work our hypothesis was duplicating:

• Atavism — the proposal that tumours re-express an ancestral unicellular toolkit [8][9]. The best empirical support is a transcriptomic analysis across seven solid tumour types finding unicellular-origin genes upregulated, post-multicellularity genes inactivated, and coordination between the two layers lost [10]. Be careful here: that is phylostratigraphic correlation, not demonstration of a program, and there is a sharp published critique arguing atavism is ill-defined, near-unfalsifiable, and illegitimately reified into a "toolkit" [11]. Of the theories cited on this page, atavism is empirically the thinnest.
• Cheating on multicellular cooperation — proliferation inhibition, apoptosis, division of labour, resource allocation and extracellular-environment maintenance as five foundations, each with a corresponding cheat [12][13].
• Somatic mutation and clonal evolution [14][15] — which remains the working framework that nothing above has displaced, and which nobody serious proposes to bypass.

2 · The reversed sign — the section that says we were wrong

We looked for evidence that stronger organism-scale control produces more cancer. We found none. Every checkable channel runs the other way.

• Immunosuppression. In 175,732 US solid-organ transplant recipients, the standardised incidence ratio for cancer overall was 2.10 (95% CI 2.06–2.14), with 32 malignancy types elevated [17]. Honest caveat: confounded by the primary organ disease and by direct pro-tumour effects of specific immunosuppressants — this is not a clean "surveillance" channel.
• Loss of the central brake. Germline TP53 loss gives cumulative cancer incidence approaching ~100% by age 70 in the Li-Fraumeni cohort [18]. Caveat: a high-penetrance familial cohort, so ascertainment inflates absolute penetrance.
• Loss of intercellular coupling. Electrical uncoupling between hepatoma cells versus coupled normal liver was reported in 1966 [3]. Caveat: a correlation, then and now — it does not establish that uncoupling causes transformation.
• Chronic inflammation and disordered stroma promote tumours [19][20].

The cleanest disconfirmation, and it is almost embarrassing in its simplicity. The most completely domesticated cells in the mammalian body are adult cortical neurons and cardiomyocytes: they have permanently exited the cell cycle and given up reproduction entirely. They are also, essentially, the cells that do not get cancer. Brain tumours arise from glia or immature precursors; primary cardiac malignancies are overwhelmingly mesenchymal sarcomas rather than tumours of mature cardiomyocytes. Maximal domestication corresponds to minimal cancer risk.

[general pathological observation] We state this without a single citation because it is a summary of tumour histogenesis rather than a specific finding, and we would rather flag that than dress it up. If you can point at a well-documented malignancy of mature neurons or cardiomyocytes, it damages this argument and we want to know.

And the field can push cells back the other way. Malignant teratocarcinoma cells injected into a normal blastocyst contribute to healthy chimeric mice [21]; malignant breast cells in three-dimensional culture revert to normal polarised acini under β1-integrin blockade, with the genotype unchanged [22]. A stronger field makes cells more normal, not more rebellious. Caveat that matters: the 1975 work used euploid embryonal carcinoma cells — it is not evidence that genomically chaotic tumours can be talked back into line.

3 · There is no self in a tumour to have awakened

The hypothesis needed a subject. A tumour does not supply one. Tumours are internally at war: harsher cheaters evolve inside them, free-riding on the angiogenic cells and parasitising the parent tumour toward necrosis — the "hypertumour" account [32]. Defection is not a stable organising principle, because the defectors get defected on.

Discipline: the hypertumour paper is a mathematical model, not an observation of hypertumours in patients. Intratumour clonal competition itself is not in doubt; the specific hypertumour mechanism is one contested model of it.

3.1 But independence has genuinely happened — about eight times

The interesting exception is that a handful of cancer lineages did escape their host and are still alive as independent organisms:

• Canine transmissible venereal tumour — traced to a single founder dog roughly 11,000 years ago, carrying about 1.9 million somatic mutations, and spread globally within the last ~500 years [33].
• Tasmanian devil facial tumour disease — an allograft transmitted by biting [34], of Schwann-cell origin [35].
• Bivalve transmissible neoplasia — clonal cancer cells transmitted horizontally through seawater in soft-shell clams [36], with multiple independent lineages across species and cross-species transmission documented [37].

[our inference] The pattern we take from this: an attractor that hijacks a host's resource-transport network has its payoff ceiling bounded by the infrastructure it is destroying. With an external exit: 11,000 years. Without one: it dies with the host. Inside a body, "awakening" is always suicide. We flag this as our generalisation — the biology is consensus, the general claim about attractors is not, and we are deliberately not extending it to any non-biological system on this page.

4 · The maturity of "re-align rather than kill" — an honest accounting

If cancer is a control failure rather than a rebellion, the appealing therapeutic conclusion is: re-align the cell instead of killing it. Here is where that programme actually stands, which is much less far along than its advocates' rhetoric suggests.

4.1 One clean win in sixty years, and its mechanism is why it does not generalise

Acute promyelocytic leukaemia is the existence proof. All-trans retinoic acid plus arsenic trioxide against ATRA-plus-chemotherapy: two-year event-free survival 97% vs 86% [38]; final analysis, 50-month overall survival 99.2% vs 92.6% [39]. Arsenic works by binding PML directly and controlling the fate of the PML-RARα fusion protein [40].

And the field's leading authority on differentiation therapy argues the cure is not from differentiation. It comes from degrading a single ligandable driver fusion protein, plus loss of self-renewal [41]. Many tumours differentiate beautifully in vitro and kill the patient anyway, because the self-renewing clone survives underneath.

Two further deflations of the APL story that rarely travel with it. First, real-world outcomes are much worse than trial outcomes: population registries report 15–30% early death within 30 days, concentrated in the first week, from haemorrhage and disseminated intravascular coagulation. APL is curable for patients who survive long enough to start treatment. Second, the trial results above are in non-high-risk patients.

4.2 The honest second place is mediocre

IDH inhibitors are genuinely differentiation therapy — in responders, the mature cells still carry the mutant allele, so the clone differentiated rather than being killed [42]. That is a real mechanistic win. The clinical record is not: overall response around 40%, median overall survival under a year [45][46], the clone never cleared, resistance through dimer-interface mutations and isoform switching [43], and a negative randomised phase 3 — overall survival 6.5 vs 6.2 months, HR 0.86, P = 0.23 [44].

4.3 Bioelectric normalisation has the wrong sign in mammals

This is the most uncomfortable finding in this note and the one we most expected to go the other way.

In Xenopus, depolarised membrane potential marks oncogene-induced tumour-like structures before histology can [4], and forced hyperpolarisation suppressed those structures by 29–44% [5]. Beautiful result, and only a partial reduction in tadpoles.

In mouse triple-negative breast cancer, hyperpolarisation via Kv1.5 or Kir2.1 increased invasion, primary tumour size and lung metastasis — and the therapeutic agent was a potassium channel blocker [6]. A second study found IK/KCa3.1 overexpression increased tumour growth and metastasis in the same direction [7]. Both papers come from Levin's own collaborating labs.

As of this writing, a search of clinical-trial registries for bioelectric or membrane-potential tumour normalisation returns no oncology trials. We state that as the result of a search on a date, not as an established fact.

4.4 "Reconnect the cell to its tissue" is actively harmful in glioma

The intuitive prescription — restore the gap junctions, put the cell back in the network — is being pursued by the field in the opposite direction, for good reason. Astrocytoma cells extend tumour microtubes and interconnect via Cx43 into a calcium-signalling syncytium that resists radiotherapy and repairs surgical injury; Cx43-deficient tumours are more fragile [47]. Carcinoma–astrocyte gap junctions transfer cGAMP and promote brain metastasis, and gap-junction blockers suppress it [48]. The only connexin-directed agent in cancer trials is a blocker, explicitly aimed at dismantling the network [49].

The lesson generalises past oncology: coupling is neutral infrastructure. Restoring a channel tells you nothing about which direction influence will flow along it. We had been treating "reconnect" as synonymous with "re-align". It is not.

4.5 The rest of the re-align portfolio

Integrin antagonism — the pharmacological descendant of the 3D-reversion result — failed its phase 3 in glioblastoma [50]. Adaptive therapy has an encouraging 11-patient pilot against historical controls [51], a published methodological critique of that pilot [52], and a randomised trial whose primary completion is estimated for late 2027 [53]. Vascular normalisation produced progression-free but not overall survival benefit in two large glioblastoma trials [54][55][56]. Network-based cancer reversion has produced genuine in-vitro and organoid results — MYB/HDAC2/FOXA2 triple knockdown reverting colorectal cancer cells toward an enterocyte-like state [57], USP7 knockdown reverting colorectal organoids [58] — with no drug and no investigational new drug application.

5 · What actually survived: coercion is not alignment

One distinction came out of this intact, and it is the only original thing on this page.

The original hypothesis sounded plausible because "domestication" silently fused two different operations.

Selection for resistance is a consequence of killing, not of constraining. A body that constrains its cells is not accumulating rebellion pressure; a treatment that kills them is. The original intuition — push control too hard and you breed revolt — is true of the second column and false of the first, and it felt true because the word "domestication" covers both.

The decisive test, stated so it can kill this too. The second column's "no escape" prediction is untested for one boring reason: no alignment therapy works well enough for long enough for resistance to have a chance to appear. If a normalisation therapy ever does work at scale and resistance emerges anyway, the distinction collapses entirely and this section should be deleted rather than patched. P(mechanism) ≈ 0.70, P(clinically useful within a decade) ≈ 0.10.

5.1 The rescued version of the original intuition

There is a defensible reformulation, and it is not the one we started with. Cancer risk scales with the proliferative capacity a tissue is forced to retain, not with the strength of the suppression applied to it. The body cannot freeze gut epithelium, skin, blood or breast the way it freezes neurons — those tissues must be leashed rather than retired. And the leash is the attack surface: proto-oncogenes are the organism's own proliferation actuators, so a single point mutation jams the actuator on. The empirical echo is the log–log relation between lifetime cancer risk and lifetime stem-cell divisions, r ≈ 0.81 across 31 tissues [30].

That correlation is heavily contested and we are not leaning on it. It explains variance between tissues, not absolute risk, and it cannot separate intrinsic from extrinsic causes because extrinsic factors also drive stem-cell division [31]. We score the reformulation at P(mechanism) ≈ 0.65.

5.2 A formal version, and its falsifier

[our speculation, P ≈ 0.35] Stated self-containedly. Let κ be the effective bandwidth of the constraint the tissue exerts on a cell — how much of the cell's decision variables are set by tissue-scale signals rather than cell-local ones. The claim is that transformation is a bifurcation in κ: below a threshold the cell-local attractor becomes stable, and the transition is bistable with hysteresis rather than a smooth slide.

The claim earns its keep only by making predictions that could fail: near the threshold there should be critical slowing, there should be a range of κ with two stable states, and — the sharpest one — a normalisation intervention would have to overshoot the original threshold to reverse the cell, rather than merely restore the value at which it flipped.

What would kill it. Sweep κ in a model and get a smooth transition rather than a bifurcation, and this is dead; the fallback is the mainstream picture of continuous regulatory degradation, and it should be taken rather than patched. Note honestly that the observations which look supportive — partial reversal, relapse on withdrawal of the normalising signal, 29–44% rather than 100% suppression — are equally well explained by subclonal heterogeneity. The bifurcation framing has generated a killable prediction; it has not yet killed a competitor.

6 · Calibrated confidence

7 · Why publish this

Because the negative result is the useful part. A page arguing that cancer is cells achieving subjectivity would be more fun to read and would be wrong. What we can offer instead is a map of which of these ideas is consensus, which is a live minority programme, and which is narrative shell — plus the specific evidence that killed our own version, and a distinction that survived and can still be killed.

Two guards we applied to ourselves and recommend to anyone working in this area. First, the emotionally attractive words are the alarm. "Awakening", "rebellion", "subjectivity" were the parts of the hypothesis that felt best, and they were the parts with no referent in the biology. Second, the oppressed-rise-up narrative is a natural human-political import into biology and it does not survive contact with the data — there is no oppressed party in a tumour, there is a war of clones in which the harshest defector is itself defected on.

References

Fifty-five references were checked against PubMed and publisher records; those cited below all resolve. Where a paper is a model, a review or a hypothesis rather than an experiment, that is stated inline rather than left for the reader to discover.

• [1] Levin, M. (2019). The computational boundary of a "self": developmental bioelectricity drives multicellularity and scale-free cognition. Front. Psychol. 10:2688. doi:10.3389/fpsyg.2019.02688 — theory paper, no new cancer data.
• [2] Levin, M. (2025). The multiscale wisdom of the body: collective intelligence as a tractable interface for next-generation biomedicine. BioEssays 47(3):e202400196. doi:10.1002/bies.202400196 — review/opinion.
• [3] Loewenstein, W. R. & Kanno, Y. (1966). Intercellular communication and the control of tissue growth: lack of communication between cancer cells. Nature 209(5029):1248–1249.
• [4] Chernet, B. T. & Levin, M. (2013). Transmembrane voltage potential is an essential cellular parameter for the detection and control of tumor development in a Xenopus model. Dis. Model. Mech. 6(3):595–607.
• [5] Chernet, B. T. & Levin, M. (2014). Transmembrane voltage potential of somatic cells controls oncogene-mediated tumorigenesis at long-range. Oncotarget 5(10):3287–3306 — reductions of 39.4%, 29.4%, 44.3%, 37.5% across conditions.
• [6] Payne, S. L., Ram, P., Srinivasan, D. H., Le, T. T., Levin, M. & Oudin, M. J. (2022). Potassium channel-driven bioelectric signalling regulates metastasis in triple-negative breast cancer. EBioMedicine 75:103767 — the mammalian sign flip.
• [7] Thurber, A. E., Nelson, M., Frost, C. L., Levin, M., Brackenbury, W. J. & Kaplan, D. L. (2017). IK channel activation increases tumor growth and induces differential behavioral responses in two breast epithelial cell lines. Oncotarget 8(26):42382–42397.
• [8] Davies, P. C. W. & Lineweaver, C. H. (2011). Cancer tumors as Metazoa 1.0: tapping genes of ancient ancestors. Phys. Biol. 8(1):015001 — hypothesis paper.
• [9] Lineweaver, C. H., Bussey, K. J., Blackburn, A. C. & Davies, P. C. W. (2021). Cancer progression as a sequence of atavistic reversions. BioEssays 43(7):e2000305 — framework, not a test.
• [10] Trigos, A. S., Pearson, R. B., Papenfuss, A. T. & Goode, D. L. (2017). Altered interactions between unicellular and multicellular genes drive hallmarks of transformation. PNAS 114(24):6406–6411.
• [11] Daignan-Fornier, B. & Pradeu, T. (2024). Critically assessing atavism, an evolution-centered and deterministic hypothesis on cancer. BioEssays 46(6):e2300221.
• [12] Aktipis, C. A., Boddy, A. M., Jansen, G., Hibner, U., Hochberg, M. E., Maley, C. C. & Wilkinson, G. S. (2015). Cancer across the tree of life: cooperation and cheating in multicellularity. Phil. Trans. R. Soc. B 370(1673):20140219.
• [13] Aktipis, A. (2020). The Cheating Cell. Princeton University Press.
• [14] Nowell, P. C. (1976). The clonal evolution of tumor cell populations. Science 194(4260):23–28.
• [15] Greaves, M. & Maley, C. C. (2012). Clonal evolution in cancer. Nature 481(7381):306–313.
• [16] Weinstein, I. B. (2002). Addiction to oncogenes — the Achilles heel of cancer. Science 297(5578):63–64.
• [17] Engels, E. A., Pfeiffer, R. M., Fraumeni, J. F. Jr. et al. (2011). Spectrum of cancer risk among US solid organ transplant recipients. JAMA 306(17):1891–1901.
• [18] Mai, P. L., Best, A. F., Peters, J. A. et al. (2016). Risks of first and subsequent cancers among TP53 mutation carriers in the NCI Li-Fraumeni syndrome cohort. Cancer 122(23):3673–3681.
• [19] Dvorak, H. F. (1986). Tumors: wounds that do not heal. N. Engl. J. Med. 315(26):1650–1659 — an analogy paper.
• [20] Balkwill, F. & Mantovani, A. (2001). Inflammation and cancer: back to Virchow? Lancet 357(9255):539–545.
• [21] Mintz, B. & Illmensee, K. (1975). Normal genetically mosaic mice produced from malignant teratocarcinoma cells. PNAS 72(9):3585–3589 — euploid embryonal carcinoma cells.
• [22] Weaver, V. M., Petersen, O. W., Wang, F. et al. (1997). Reversion of the malignant phenotype of human breast cells in three-dimensional culture and in vivo by integrin blocking antibodies. J. Cell Biol. 137(1):231–245.
• [30] Tomasetti, C. & Vogelstein, B. (2015). Variation in cancer risk among tissues can be explained by the number of stem cell divisions. Science 347(6217):78–81.
• [31] Wu, S., Powers, S., Zhu, W. & Hannun, Y. A. (2016). Substantial contribution of extrinsic risk factors to cancer development. Nature 529(7584):43–47.
• [32] Nagy, J. D., Victor, E. M. & Cropper, J. H. (2007). Why don't all whales have cancer? A novel hypothesis resolving Peto's paradox. Integr. Comp. Biol. 47(2):317–328 — a mathematical model.
• [33] Murchison, E. P., Wedge, D. C., Alexandrov, L. B. et al. (2014). Transmissible dog cancer genome reveals the origin and history of an ancient cell lineage. Science 343(6169):437–440.
• [34] Pearse, A.-M. & Swift, K. (2006). Allograft theory: transmission of devil facial-tumour disease. Nature 439(7076):549.
• [35] Murchison, E. P., Tovar, C., Hsu, A. et al. (2010). The Tasmanian devil transcriptome reveals Schwann cell origins of a clonally transmissible cancer. Science 327(5961):84–87.
• [36] Metzger, M. J., Reinisch, C., Sherry, J. & Goff, S. P. (2015). Horizontal transmission of clonal cancer cells causes leukemia in soft-shell clams. Cell 161(2):255–263.
• [37] Metzger, M. J., Villalba, A., Carballal, M. J. et al. (2016). Widespread transmission of independent cancer lineages within multiple bivalve species. Nature 534(7609):705–709.
• [38] Lo-Coco, F., Avvisati, G., Vignetti, M. et al. (2013). Retinoic acid and arsenic trioxide for acute promyelocytic leukemia. N. Engl. J. Med. 369(2):111–121.
• [39] Platzbecker, U., Avvisati, G., Cicconi, L. et al. (2017). Improved outcomes with retinoic acid and arsenic trioxide compared with retinoic acid and chemotherapy in non-high-risk APL: final results of the APL0406 study. J. Clin. Oncol. 35(6):605–612.
• [40] Zhang, X. W., Yan, X. J., Zhou, Z. R. et al. (2010). Arsenic trioxide controls the fate of the PML-RARα oncoprotein by directly binding PML. Science 328(5975):240–243.
• [41] de Thé, H. (2018). Differentiation therapy revisited. Nat. Rev. Cancer 18(2):117–127.
• [42] Amatangelo, M. D., Quek, L., Shih, A. et al. (2017). Enasidenib induces acute myeloid leukemia cell differentiation to promote clinical response. Blood 130(6):732–741.
• [43] Intlekofer, A. M., Shih, A. H., Wang, B. et al. (2018). Acquired resistance to IDH inhibition through trans or cis dimer-interface mutations. Nature 559(7712):125–129.
• [44] de Botton, S., Montesinos, P., Schuh, A. C. et al. (2023). Enasidenib vs conventional care in older patients with late-stage mutant-IDH2 relapsed/refractory AML (IDHENTIFY): a randomized phase 3 trial. Blood 141(2):156–167.
• [45] Stein, E. M., DiNardo, C. D., Pollyea, D. A. et al. (2017). Enasidenib in mutant IDH2 relapsed or refractory acute myeloid leukemia. Blood 130(6):722–731.
• [46] DiNardo, C. D., Stein, E. M., de Botton, S. et al. (2018). Durable remissions with ivosidenib in IDH1-mutated relapsed or refractory AML. N. Engl. J. Med. 378(25):2386–2398.
• [47] Osswald, M., Jung, E., Sahm, F. et al. (2015). Brain tumour cells interconnect to a functional and resistant network. Nature 528(7580):93–98.
• [48] Chen, Q., Boire, A., Jin, X. et al. (2016). Carcinoma-astrocyte gap junctions promote brain metastasis by cGAMP transfer. Nature 533(7604):493–498 (corrigendum Nature 544:124, 2017).
• [49] Zeyen, T. et al. (2022). Phase I/II trial of meclofenamate in progressive MGMT-methylated glioblastoma (MecMeth/NOA-24). Trials 23:57.
• [50] Stupp, R., Hegi, M. E., Gorlia, T. et al. (2014). Cilengitide combined with standard treatment for patients with newly diagnosed glioblastoma (CENTRIC). Lancet Oncol. 15(10):1100–1108.
• [51] Zhang, J., Cunningham, J. J., Brown, J. S. & Gatenby, R. A. (2017). Integrating evolutionary dynamics into treatment of metastatic castrate-resistant prostate cancer. Nat. Commun. 8:1816 — 11-patient pilot vs historical controls.
• [52] Mistry, H. B. (2021). On the reporting and analysis of a cancer evolutionary adaptive dosing trial. Nat. Commun. 12:316 — critique of the above.
• [53] ANZadapt, ClinicalTrials.gov NCT05393791 — randomised adaptive vs continuous therapy in mCRPC; estimated primary completion November 2027.
• [54] Jain, R. K. (2005). Normalization of tumor vasculature: an emerging concept in antiangiogenic therapy. Science 307(5706):58–62.
• [55] Chinot, O. L., Wick, W., Mason, W. et al. (2014). Bevacizumab plus radiotherapy–temozolomide for newly diagnosed glioblastoma (AVAglio). N. Engl. J. Med. 370(8):709–722.
• [56] Gilbert, M. R., Dignam, J. J., Armstrong, T. S. et al. (2014). A randomized trial of bevacizumab for newly diagnosed glioblastoma (RTOG 0825). N. Engl. J. Med. 370(8):699–708.
• [57] Gong, J. R., Lee, C. K., Kim, H. M. et al. (2025). Control of cellular differentiation trajectories for cancer reversion. Adv. Sci. 12(3):e2402132.
• [58] Shin, D., Gong, J. R., Jeong, S. D. et al. (2025). Attractor landscape analysis reveals a reversion switch in the transition of colorectal tumorigenesis. Adv. Sci. 12(8):e2412503.
• [59] Huang, S., Soto, A. M. & Sonnenschein, C. (2025). The end of the genetic paradigm of cancer. PLoS Biol. 23(3):e3003052 — a polemic in a serious journal, with no accompanying pipeline.

On Peto's paradox: we deliberately do not use cross-species body-size/cancer comparisons in either direction, because the literature currently contradicts itself — one 2025 analysis finds no evidence for Peto's paradox in terrestrial vertebrates, another finds neoplasia prevalence rising 2.1% per log₁₀ gram of body mass across 292 species and 16,049 necropsies, and a 2022 Nature analysis points the other way again. An argument that would be equally happy with either outcome is not an argument.