A Y Combinator session on AI for biology asks whether AI can cure cancer, yet its cancer evidence is one personalised case, a DNA model trained on a single human genome and no cleared AI-designed drug. What the record supports is a narrower claim about analysing one patient’s tumour data after standard options have run out.
Table of Contents
No drug designed by AI has won FDA clearance, according to a Wall Street Journal report that Benzinga summarised on 5 November 2026. Drugmaker executives told the paper that AI helps pick targets and suitable patients but cannot resolve biology’s unknowns alone. That is the record behind every promise of AI cancer treatment, and the promises keep coming: Anthropic’s chief executive posted in September that curing most human disease within five to 10 years may be possible.
A video published on 8 November 2026 asks the question in its title. Y Combinator’s Paper Club session, “Is AI Now Good Enough To Cure Cancer?”, gives its first talk to Alfredo Gonzalez, who argues that cancer has become a per-patient data problem. The session is a Y Combinator production, and Gonzalez’s company, Omanta, went through the same organisation’s Summer 2026 batch, so the host also has a stake in the startup on stage.
Only one of the four talks is about cancer
The chapter list puts the cancer talk at 8:26 and ends it at 29:50, about 21 minutes. The three talks that follow cover virtual cells and autonomous research (Yusuf Roohani of Intology), what a model learns from DNA (Garyk Brixi, first author of the Evo 2 paper) and few-shot prompting for biological design (Divya Nori). None of them reports a result in a tumour. Roohani described his results as computer simulations that still need lab validation, and the wet-lab results in the other talks came from bacteriophage genomes, a reverse transcriptase and an enhancer tested in a mouse brain.
The session description sells a broader thesis: sequencing costs have collapsed, biology produces more data than anyone can mine, and the models are catching up. That is a claim about AI for biology. The title narrows it to cancer without a cancer result to match.
The company’s name is unstable too. Auto-generated captions and at least one news write-up call it AMA, while Y Combinator’s own pages list Omanta, founded in 2026, with six employees in San Francisco. The pages also describe it as a personalised research lab for one patient at a time.
Per-patient analysis extends a public programme
Gonzalez says the Cancer Genome Atlas sequenced a couple of hundred patients per tumour type a decade ago, so researchers could only look for averages, while each patient today yields about 20,000 gene-level data points plus copy number, RNA expression and splicing that hospitals are not set up to mine. The National Cancer Institute puts the atlas at 11,000 patients across 33 cancer types, ten of them rare, which averages about 330 patients per type by simple division. His order of magnitude holds.
His central example, a rare brain tumour that clustered with mesenchymal tumours, is also older than the AI claim. NCI lists among the atlas findings that cancers of different tissues can share the same alterations and be more alike than tumours from the same tissue. What is new is running that comparison for one named patient instead of a cohort.
The cost story needs a second look as well. A whole genome now costs under $1,000, according to one detailed account of a single patient’s care, but the National Human Genome Research Institute’s cost series leaves out downstream analysis such as identifying variants and interpreting results. Interpretation is the expensive step, and it is the work a per-patient lab sells.
The remission case rests on a radioligand therapy
The flagship case is Sid Sijbrandij, the GitLab co-founder with a rare bone cancer of the spine. Omanta’s launch post says Gonzalez was genomics lead for the personalised campaign behind Sijbrandij’s complete remission and led a year-long review of 25 terabytes of sequencing and molecular data. The most detailed independent account, an 18 January 2026 essay by Elliot Hershberg, describes what that data changed. After the cancer returned in 2024, single-cell analysis by computational biologists showed tumour cells expressing fibroblast markers, including FAP. A service in his care team then surfaced an experimental German radioligand therapy that targets FAP. The tumour shrank enough for surgery, and afterwards T cells made up 89% of infiltrating immune cells, up from 19%, following a checkpoint inhibitor, a peptide vaccine, an oncolytic virus and the radiotherapy together.
Two details limit the AI claim. Hershberg does not describe the radioligand therapy as AI-designed, and he calls the case anomalous: Sijbrandij is a billionaire who needed five single-patient FDA applications, each accepted within 48 hours, plus a team of specialists. By Sijbrandij’s own estimate it costs $1 billion to get a drug approved and $1 million to dose one person with a personalised therapy. The talk adds two unnamed patients, a rare brain tumour and a patient whose last three therapies were chemotherapies that the analysis judged a poor fit, and it gives no outcome for either.
Drugmakers count zero AI-designed approvals
The Wall Street Journal interviews that Benzinga summarised carry the opposing evidence. AstraZeneca’s chief executive, Pascal Soriot, called the idea that AI will cure cancer “a little bit of a jump.” Vertex’s executive chairman, Jeff Leiden, said the company ran its data set of 10,000 compounds past roughly 50 AI companies and none produced a working drug. McKinsey’s Alex Devereson said about a quarter of mid-stage to late-stage drug failures trace to off-target safety problems, which current AI models struggle to predict.
These are interested parties. The executives sell drugs, and AstraZeneca is paying Insitro $40 million upfront for an AI collaboration worth up to about $1.5 billion with milestones, so the scepticism is about attribution, not abandonment. The co-founder of Generate Biomedicines said compute alone is no differentiator and that winners will pair it with proprietary biological data and the ability to close the loop between generation and measurement. That is close to Roohani’s own argument for experiment-guided loops over bigger datasets, and it places the dispute in what AI is credited with, not in whether it is used.
The best AI drug result so far treats lung scarring
The nearest thing to a clinical result for an AI-designed drug is not a cancer trial. Insilico Medicine’s rentosertib, whose target and molecule both came from generative AI, raised forced vital capacity by a mean 98.4 mL at the 60 mg dose against a 20.3 mL decline on placebo in a 71-patient trial for idiopathic pulmonary fibrosis, published in Nature Medicine on 3 June 2025. Seven patients stopped because of liver injury or dysfunction, and only 12 of 18 patients on the high dose, 67%, completed it, against 88% on placebo.
The authors name small arms, participants who were all residents of China and a short follow-up as limits, and Insilico, the drug’s developer, has its own stake in the result. Trade analyses cited in Drug Discovery and Development put the rate at which AI-originated compounds clear phase 2 at roughly the same 40% as conventional drugs. A phase 2a result is a step toward clearance, which the Wall Street Journal report says no AI-designed drug has reached.
Evo 2 reads cancer genes but has met one human genome
Evo 2, the DNA model behind the Brixi talk, is the best-documented technology in the session. Nature published it in March 2026: a model with 40 billion parameters trained on 9.3 trillion DNA tokens, a context window of one million letters and open weights. A classifier built on its embeddings separates loss-of-function from harmless BRCA1 variants with an AUROC of 0.95 on the test set, where 0.5 is chance and 1.0 is perfect, a task with direct breast-cancer relevance.
The limits sit in the paper. The only human genome in the training data is the reference genome, and the zero-shot predictions rest on variation across species. Brixi told the audience that training on everyone in the room would add redundancy, not cancer signal. Press coverage of the Nature paper notes that generated sequences do not necessarily represent replication-competent genomes, and the authors write that supervised fine-tuning or reinforcement learning with experimental feedback may be needed to improve the functions of generated DNA. That makes Evo 2 a tool for interpreting variants today, in this reading, and not a source of cancer therapies.
A narrower claim the evidence supports
The evidence supports a claim that fits one patient and not a population. When standard options are exhausted, a team that sequences deeply, reanalyses the data and matches it to experimental therapies can find a target that standard reports missed, as in Sijbrandij’s case. It does not support the word cure, AI-designed cancer drugs reaching patients, or outcomes for anyone beyond that documented case. StartupHub’s summary lists what the talk left unanswered: who pays for a full-time machine-run lab per patient, what regulators will accept from an agent’s repurposing call, and when a designed protein treats a person instead of a benchmark.
For a patient with exhausted options and the oncologist who treats them, the practical point sits in Gonzalez’s answer on tissue. Most hospitals preserve samples in formalin, which rules out organoid work, and Hershberg says that fixative destroys most of the nucleic acid information. Whether fresh or frozen tissue can be kept is a question for the care team, not medical advice, and the clinician keeps the final say on treatment, as Gonzalez agreed. Two developments would change this reading: a first FDA clearance for an AI-designed drug, and published outcomes for more than one Omanta patient.
Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency
This article is an original analysis supported by the sources cited below
Is AI Now Good Enough To Cure Cancer? | YC Paper Club
The session itself and its transcript: the four talks, the per-patient cancer claims, virtual cells, Evo 2 and in-context prompting, plus the speakers’ own answers to audience questions.
Omanta: A personal research lab dedicated to each patient.
The company’s own launch post: its Summer 2026 batch, its description of the GitLab co-founder case and the 25 terabytes of data reviewed, all as claims by the company.
Omanta: A research lab for one | Y Combinator
Confirmed the company name, its 2026 founding, six employees and San Francisco base, which differ from the AMA spelling in captions and one news write-up.
The Cancer Genome Atlas (TCGA)
NCI’s account of the atlas: 33 cancer types, 11,000 patients, ten rare cancers and the finding that tumours of different tissues can resemble each other.
DNA Sequencing Costs: Data
NHGRI’s method for its cost-per-genome series, including the downstream analysis steps, such as variant identification and interpretation, that the figures leave out.
Genome modeling and design across all domains of life with Evo 2
The Evo 2 paper text: 9.3 trillion training tokens, one-million-letter context, the BRCA1 classifier result, the single human reference genome and the authors’ stated limits.
AI trained on 9 trillion DNA letters predicts harmful mutations and designs new genomes
Independent coverage of the March 2026 Nature publication, including the caveat that generated sequences are not necessarily replication-competent genomes.
Going Founder Mode On Cancer
The most detailed account of the GitLab co-founder’s treatment: the FAP finding, the German radioligand therapy, the 19% to 89% T-cell shift, five FDA applications and his cost estimate.
Drugmaker CEOs Say AI Alone Can’t Cure Diseases, Challenge Silicon Valley’s Bold Predictions: ‘To Say AI Is Going to Cure Cancer…’
Benzinga’s summary of a Wall Street Journal report: no AI-designed drug cleared, Vertex’s test of about 50 AI firms, the McKinsey safety estimate and AstraZeneca’s Insitro deal.
Insilico’s rentosertib clears a phase 2a hurdle
Trade report on the 71-patient phase 2a trial of an AI-designed lung drug, with the 98.4 mL gain, the liver-related dropouts and the authors’ stated limits.
YC Founder Says Cancer Is Now a Data Problem
A same-week news summary of the cancer talk, including the AMA spelling and the list of questions the talk left unanswered.
Is AI Now Good Enough To Cure Cancer? | YC Paper Club – Whatfinger Startup And Small Business
A repost of the video description with its 8 November 2026 date and the chapter timings that show the cancer talk running from 8:26 to 29:50.
| Citing this article? Brief excerpts are welcome. Please credit Webiano.digital, name the author where stated, and include a link to https://webiano.digital and to this original article. Full or substantial republication requires prior written permission. Read our Copyright and Content Use Policy. |
This article was prepared with the assistance of artificial intelligence tools. The content underwent expert human review, and Webiano Digital & Marketing Agency assumes editorial responsibility for its final version and publication.
