Toward a General Cure for Cancer: Why We Might Need an AGI Researcher
One-sentence summary:
Rather than relying solely on incremental, subtype-by-subtype progress, the world should fund and govern a globally accessible AGI cancer researcher — a language-based, tool-using system coupled to autonomous laboratories — explicitly tasked with discovering general interventions that restore growth control across tumors.
What unites cancers — and why a general strategy is thinkable
Cancer’s common denominator is loss of growth control: cells sustain proliferative signaling, evade suppressors, resist death, and bypass intrinsic and extrinsic checkpoints. These capabilities recur across diverse tissues even as the initiating lesions vary; the hallmarks of cancer framework distills this convergence and has been repeatedly refined with new dimensions, including phenotypic plasticity, non-mutational epigenetics, and the tumor microenvironment [1–3]. Large-scale pan-cancer efforts show conserved disruptions to pathways across histologies, supporting the premise that cross-tumor interventions can be reasoned about without denying heterogeneity [4, 5].
Clinical practice already acknowledges shared vulnerabilities. Tumor-agnostic approvals — pembrolizumab for MSI-H/dMMR tumors and for TMB-high disease; larotrectinib for NTRK fusions — demonstrate that targeting a common failure mode (mismatch-repair deficiency, hypermutation, or an oncogenic fusion) can cut across tissue of origin. These are not “general cures,” but they prove the concept that lineage-agnostic levers exist [6–9].
Two routes to a general cure
The direct route is today’s default: mobilize human expertise to map mechanisms, integrate multi-omic and clinical data, run trials, and — through cumulative advances — approach broadly applicable interventions.
The indirect route is technically plausible now: build an artificial general intelligence (AGI) capable of cross-domain scientific reasoning at the effective scale of the global research enterprise and assign it the problem “discover a general cure for cancer.” Such a system would read and synthesize literature; generate and prune hypotheses; design, schedule, and interpret experiments; integrate multi-modal data; and orchestrate long-horizon programs with automated labs and human clinicians.
Importantly, this augments rather than displaces scientists: by coupling models to self-driving laboratories, hypotheses must survive experiment, and error-prone reasoning is bounded by closed-loop validation and independent safety gates. The constituent technologies exist and are improving: “robot scientist” systems in biology and modern autonomous experimentation in chemistry, materials, and protein engineering have demonstrated end-to-end loops from hypothesis to result [10, 11].
The practice is not without precedent. When a concise, hand-checkable route was out of reach in mathematics, the Four-Colour Theorem was settled via a computer-assisted proof, later formalized in a proof assistant. A comparable modus operandi — computational procedures that scale beyond unaided human capacity, paired with verification — could apply to discovering pan-cancer interventions [12, 13].
Why a language-centric AGI is a plausible substrate for science
Skeptics doubt language-model-based systems can reliably “do science.” Yet a substantial literature treats language as a scaffold for abstract thought and self-regulation, with “inner speech” coordinating cognition. If much high-level reasoning is conducted through language, scaling language-native models — augmented with tools, external memory, and verification — becomes a plausible route to general scientific competence (not the only route, but a powerful one) [14–16].
Just as relevant, contemporary reasoning models improve when afforded more test-time compute: structured deliberation, multiple solution paths, and self-consistency checks raise accuracy on hard problems. For frontier queries — e.g., “design a multi-component regimen that durably restores growth control across contexts” — reliability will hinge as much on inference-time thinking and verification as on pretraining scale [17–19].
A pragmatic program sketch
A credible AGI-for-cancer effort should be practical and staged:
Knowledge integration. Unified corpora (literature, protocols, negative results) with governed access to clinical and genomic datasets; privacy-preserving analytics (e.g., federated learning) to respect data sovereignty and patient consent while enabling cross-site modeling. Frameworks from the Global Alliance for Genomics and Health (GA4GH) and large biobanks illustrate workable governance [20, 21].
Hypothesis generation and curation. Multi-omic integration to surface lineage-agnostic vulnerabilities (DNA-damage response, cell-cycle checkpoints, apoptotic priming, telomere maintenance), systematically cross-validated against pan-cancer resources [3, 5].
Automated experimentation. High-throughput perturbation (CRISPR/RNAi, small-molecule/protein modalities) run by self-driving labs to test and refine mechanistic claims, with pre-registered plans and independent replication nodes [10, 11].
Design against evolution. From the outset, incorporate adaptive-therapy principles — explicitly optimizing combinations and schedules to delay or prevent resistance — so candidates are evaluated not just for response but for evolutionary stability [22, 23].
Translational pathways. Use tumor-agnostic regulatory precedents as a template for trials centered on shared biomarkers or pathway states, then iterate toward progressively broader indications [6–9].
The economics: train once, think many times
Public estimates indicate that compute for recent frontier models already costs in the tens to low hundreds of millions of dollars, with several analyses projecting over $1 billion training runs within the next few years if current trends persist. The Stanford AI Index and Epoch AI both document rapid growth in training cost and compute; importantly, the total cost to reason on hard scientific questions will also include substantial inference-time computation (deliberation, retrieval, simulation, verification). Planning should assume wide error bars but a clear directional trend: training may be expensive, and sustained “thinking” on a problem this hard will be, too [24, 25].
Who pays — and how to keep it fair, safe, and sustainable
At this scale, durable funding and legitimate governance exceed the remit of any single philanthropist, firm, or nation. Analogues exist: CERN’s council and shared capital facilities; ITER’s cost-sharing treaty; pooled health-financing and access mechanisms in the Global Fund and Gavi; and UNESCO’s Open Science Recommendation [26–30].
A treaty-backed, UN-anchored vehicle could (i) pool contributions, (ii) procure compute, power, and data at utility scale, (iii) embed auditability, safety testing, and biosecurity review, (iv) guarantee open-science outputs proportionate to public funding, and (v) allocate inference budgets to a short list of humanity-scale questions — with a general cure for cancer as a flagship.
Because public investors will expect safeguards, such a body should align with the WHO’s guidance on artificial intelligence in health and with emerging international safety norms (for example, the Bletchley Declaration). Concretely: staged capability evaluations, independent red-teaming, strict separation between in silico design and wet-lab actuation, tiered access to high-risk tools, and human-in-the-loop oversight for any experimental transition — especially in biology [31, 32].
Data stewardship should be designed in, not bolted on. Federated learning and federated analysis let models learn from distributed datasets without moving patient data; GA4GH’s frameworks provide principled governance; and biobank access regimes supply working templates for consent, oversight, and transparency [20, 21].
Biosecurity must be non-negotiable. Wet-lab integration should comply with existing nucleic-acid synthesis-screening frameworks (for example, the U.S. HHS guidance for providers of synthetic double-stranded DNA) and the International Gene Synthesis Consortium’s Harmonized Screening Protocol, with procurement requirements that restrict access to screened providers and benchtop devices. Independent review boards should have veto authority over any agent-design capabilities, and audit logs should be mandatory for all tool calls that could touch sequence design or synthesis [33, 34].
Environmental externalities must be managed up front. Training and large-scale inference consume energy and, depending on siting and cooling, water. Best-practice engineering — efficient hardware and algorithms, carbon-free energy procurement, low-water cooling, circularity targets — materially reduces impact. Credible social license will require binding carbon and water budgets, transparent accounting, and independent auditing [35, 36].
What success would mean
If such an undertaking succeeded, it would culminate nearly two centuries of pathology and oncology — from Virchow’s cellular pathology, through the molecular and genomic eras, to a general restoration of growth control delivered with machine reasoning. The field’s telos — durably ending malignant proliferation — would be realized not by displacing human scientists, but by completing their project with a new kind of colleague.
References
[1] Hanahan D. & Weinberg R.A. The Hallmarks of Cancer. Cell 100, 57–70 (2000).
[2] Hanahan D. & Weinberg R.A. Hallmarks of Cancer: The Next Generation. Cell 144, 646–674 (2011).
[3] Hanahan D. Hallmarks of Cancer: New Dimensions. Cancer Discovery 12, 31–46 (2022).
[4] Vogelstein B. et al. Cancer Genome Landscapes. Science 339, 1546–1558 (2013).
[5] ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium. Pan-Cancer Analysis of Whole Genomes. Nature 578, 82–93 (2020).
[6] Le D.T. et al. Mismatch-repair deficiency predicts response of solid tumors to PD-1 blockade. N. Engl. J. Med. 372, 2509–2520 (2015).
[7] Marabelle A. et al. Tumour mutational burden and outcomes with pembrolizumab (KEYNOTE-158). Ann. Oncol. 31, 131–141 (2020).
[8] Drilon A. et al. Larotrectinib in TRK fusion-positive cancers. N. Engl. J. Med. 378, 731–739 (2018).
[9] U.S. FDA. Pembrolizumab for TMB-H solid tumors. News Release (2020).
[10] King R.D. et al. The Automation of Science. Science 324, 85–89 (2009).
[11] MacLeod B.P. et al. Self-driving laboratory for accelerated materials discovery. Matter 3, 613–628 (2020).
[12] Appel K. & Haken W. Every planar map is four colorable. Illinois J. Math. 21, 429–567 (1977).
[13] Gonthier G. Formal Proof — The Four-Color Theorem. Notices Amer. Math. Soc. 55, 1382–1393 (2008).
[14] Alderson-Day B. & Fernyhough C. Inner Speech: Development, Cognitive Functions, Phenomenology, and Neurobiology. Psychol. Bull. 141, 931–965 (2015).
[15] Carruthers P. The Cognitive Functions of Language. Behav. Brain Sci. 25, 657–726 (2002).
[16] Lupyan G. & Clark A. Words and the World. Curr. Dir. Psychol. Sci. 24, 279–284 (2015).
[17] Wang X. et al. Self-Consistency Improves Chain-of-Thought Reasoning. arXiv:2203.11171 (2022).
[18] Yao S. et al. Tree of Thoughts: Deliberate Problem Solving with LLMs. arXiv:2305.10601 (2023).
[19] OpenAI. Learning to Reason with LLMs (o1). Technical Report (2024).
[20] Global Alliance for Genomics and Health. Framework for Responsible Sharing of Genomic and Health-Related Data. (2018).
[21] Bycroft C. et al. The UK Biobank Resource with Deep Phenotyping and Genomic Data. Nature 562, 203–209 (2018).
[22] Gatenby R.A. et al. Adaptive Therapy. Cancer Res. 69, 4894–4903 (2009).
[23] Gatenby R.A. & Brown J.S. Evolution and Ecology of Resistance in Cancer Therapy. Cold Spring Harb. Perspect. Med. 8, a033415 (2018).
[24] Stanford HAI. AI Index Report 2024. (2024).
[25] Epoch AI. Trends in AI Training Compute and Costs. (2024).
[26] CERN. Convention and Governance. (1953 – present).
[27] ITER Organization. Agreement on the Establishment of the ITER Project. (2006).
[28] The Global Fund. Results Report. (2025).
[29] Gavi, the Vaccine Alliance. Governance and Board Resources. (2025).
[30] UNESCO. Recommendation on Open Science. (2021).
[31] World Health Organization. Ethics and Governance of Artificial Intelligence for Health. (2021).
[32] UK Government. The Bletchley Declaration. (2023).
[33] U.S. HHS. Screening Framework Guidance for Providers of Synthetic Double-Stranded DNA. (2010).
[34] International Gene Synthesis Consortium. Harmonized Screening Protocol v2.0. (2017).
[35] Patterson D. et al. The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink. IEEE Computer 55, 18–28 (2022).
[36] Masanet E. et al. Recalibrating Global Data Center Energy-Use Estimates. Science 367, 984–986 (2020).
Note: This essay is adapted from a scholarly draft intended for submission as a Nature Comment; it has not been peer-reviewed.