AI Curing Cancer
How AI is compressing the search for cancer cures
“AI will cure cancer” has been one of those sentences that sounds profound until you examine it.
In fact “AI will cure cancer” is making the rounds again in 2026 with Elon joining the chorus.
Which cancer, as there are over 100?
What does cure mean?
How does a language model cross the distance between predicting a molecule and keeping a human being alive?
Then Anthropic published an experiment that made the sentence more concrete.
Claude did not cure cancer. It did not design a cancer drug. It did not treat a mouse, much less a person.
What it did was run a protein-design campaign.
That distinction matters. The near-term breakthrough is not a miracle molecule emerging from a chatbot. It is the conversion of scientific work into a machine-scalable loop: research, hypothesize, design, test, learn, and repeat.
The experiment is becoming software.
And that may be the development that eventually changes cancer.
What Anthropic Actually Did
Many drugs work by binding to a protein and changing what it does. The drug may block a growth signal, reveal a cancer cell to the immune system, or carry a toxic payload to a tumor. Before any of that can happen, the molecule must recognize the right target and bind to it with useful strength and specificity.
Anthropic asked whether Claude could design small proteins that perform this recognition task.
A human expert wrote a detailed protocol. Anthropic’s team selected targets, supplied literature, compute, and access to specialist open-source tools. Claude then acted less like an oracle and more like a campaign manager. It researched targets, selected possible binding sites, installed and operated protein-design and structure-prediction software, generated candidates, evaluated them, and decided which ones to advance.
Outside laboratories at Adaptyv Bio and Twist Bioscience synthesized and tested the proposed proteins. According to Anthropic’s technical report, the campaign produced 1,320 interpretable designs across 15 targets. Of those, 354 bound in the experimental assay, for an overall hit rate of 26.8 percent. At least one binder succeeded for 14 of the 15 targets. Depending on the setup, hit rates ranged from roughly 22 to 35 percent.
Some of the targets (EGFR, PD-L1, and VEGF-A) are already important in cancer biology. Ninety of the successful designs reportedly bound more tightly than 10 nanomolar, and 42 more tightly than 1 nanomolar. Lower is better.
Those are real results.
They are also bounded results.
The humans defined the playing field. Claude operated inside a substantial scaffold created by experts. The targets were mostly known, several had been used as protein-design benchmarks, and the work did not include cellular activity, animal efficacy, toxicity, pharmacokinetics, manufacturing, or clinical benefit. There was no matched campaign proving that Claude beat an equally resourced team of human protein designers.
The honest description is still remarkable: an AI system coordinated a complicated computational workflow and produced hundreds of physical molecules that independent laboratories could test.
This is major progress since 2024 when I entered into a competition to use AI to develop novel medicine. Read the four part series here:
The Important Product Is the Loop
Protein design itself is not new. De novo miniproteins with very strong affinities existed before this experiment. The specialist tools Claude used were built by human scientists. Anthropic did not invent the field.
What Anthropic is trying to invent is a more general scientific operator.
The model reads papers, uses narrow scientific software, manages compute, compares outputs, recovers from errors, and maintains a long objective across many steps. The intelligence is not just in Claude.
It is in the whole system: model, protocol, tools, data, compute, evaluation, wet lab, and human judgment.
That system matters because modern science is fragmented. One expert understands the disease. Another understands protein structure. Another runs an assay. Another analyzes the result. Information waits in queues between them.
An agent can work across those boundaries continuously. It can generate far more candidates than a person would inspect manually, keep an audit trail of why each one was selected, and begin the next round as soon as experimental results return.
Anthropic’s experiment was not yet a complete self-improving laboratory. But the architecture is visible.
The goal is a closed loop in which machines do not merely answer scientific questions. They help run the process that produces reliable answers.
Cancer Is Not One Problem
There will probably never be one cure for cancer because cancer is not one disease.
The National Cancer Institute recognizes more than 100 types. Each patient’s cancer can contain a different combination of genetic changes. Even within a single tumor, different cells can carry different mutations. Treatment kills sensitive cells while resistant cells survive, reproduce, and change the problem.
Cancer is evolution taking place inside the body.
So “AI will cure cancer” should not mean that one machine discovers one universal pill. It should mean that AI helps turn more cancers into preventable, detectable, curable, or manageable conditions.
That requires progress across the entire stack.
AI can help detect tumors in medical images and pathology slides before a human can see an obvious pattern. It can combine genomic, proteomic, imaging, and clinical data to classify a tumor more precisely. It can search the scientific literature for hidden relationships, identify biological dependencies, and propose targets that cancer cells rely on more than healthy cells do.
It can design small molecules, antibodies, proteins, RNA medicines, and cell therapies. It can predict which candidates are likely to fold, bind, enter tissue, or fail because of toxicity. It can help robotic laboratories decide which experiment would be most informative next.
And on the clinical side, it can match patients to trials, predict treatment response, find useful drug combinations, monitor recurrence, and adapt therapy as the tumor changes.
The NCI already describes AI as operating across cancer mechanisms, screening, diagnosis, drug discovery, surveillance, and care. That breadth is the point. Cancer is a systems problem. AI is useful because it can operate across systems.
But an AI prediction is not biological truth.
The wet lab remains the court of appeal. Animal studies remain imperfect but necessary. Human trials remain necessary.
Machines can compress the search but they can’t repeal reality.
From Molecular Recognition to Cancer Killing
A protein binder is best understood as a molecular key.
Sometimes the key can block the lock directly. A binder to EGFR might interfere with a signal that tells a tumor cell to grow. A binder to PD-L1 might make it harder for the tumor to suppress an immune attack. A binder to VEGF-A might interrupt the creation of blood vessels that feed a tumor.
But binding can also be used as an address.
Attach the binder to a toxin, a radioactive atom, or another drug, and it may carry the payload toward cells displaying the target. Put it into a diagnostic tracer, and it may help reveal the location of a tumor. Use it as the recognition domain of a CAR T cell, and it may tell an engineered immune cell what to attack. Connect one binder to a tumor and another to a T cell, and it may bring killer and target together.
This is where de novo protein design becomes more than a laboratory curiosity.
A 2024 study in Nature Biomedical Engineering used computationally designed binders against EGFR and CD276 as the recognition domains in CAR T cells for glioblastoma. In preclinical experiments, those designs improved surface expression, proliferation, resistance to exhaustion, and antitumor performance compared with CARs using conventional antibody-derived single-chain fragments.
A 2025 Science study designed binders to cancer-associated peptide–HLA complexes. Installed in CAR T cells, one helped induce killing of NY-ESO-1-positive melanoma cells in vitro. That approach is especially interesting because it can make fragments of proteins from inside a cancer cell visible at the cell surface.
Neither result is a human cure.
Both show how a designed binder can become one replaceable component inside a larger therapeutic machine.
The Hard Parts Stay Hard
A high-affinity binder aimed at the wrong target is a precisely engineered mistake.
Cancer therapy still requires selecting targets that tumors need and healthy tissues can spare. It requires exquisite specificity, sufficient exposure, acceptable immune reactions, stable manufacturing, and a way to stop resistant cells from escaping. A medicine must work in the turbulent environment of a human body, not merely in a clean assay.
History supplies a useful warning. A de novo-designed IL-2/IL-15 mimic called Neo-2/15 produced impressive antitumor activity in mouse models in a 2019 Nature paper. A descendant, NL-201, entered a Phase 1 clinical trial and showed target engagement. The company later discontinued development after reviewing preliminary data, expected benefit and risk, and the resources required to continue.
That is not proof that de novo proteins fail. It is proof that the distance from elegant design to useful medicine is enormous.
AI does not remove that distance.
It can help us search and travel it faster.
The Cost of Being Wrong
The most important economic variable in science may be the cost of a falsified hypothesis.
If a bad idea consumes months of expert labor and scarce laboratory capacity, researchers can test only a few ideas. They become conservative. Institutions fund what can be defended, not always what is most informative.
If machines can research a target overnight, generate thousands of candidates, choose a diverse set, document their reasoning, and send the best designs into automated assays, the cost of being wrong falls.
That changes behavior.
Scientists can explore stranger hypotheses. Small teams can run campaigns that once required specialized departments. Negative results become structured data for the next round instead of dead ends buried in a notebook.
The advantage is not that the machine is always right.
It is that civilization can afford to learn more often.
Buying More Time
This is where the cancer story becomes a longevity story.
Cancer risk rises with age because damage accumulates and the systems that suppress abnormal cells become less reliable. The same aging body is also contending with cardiovascular disease, neurodegeneration, immune decline, fibrosis, and the loss of regenerative capacity.
Extending healthy life will not come from one longevity pill. It will come from a portfolio of machines that help us detect damage earlier, understand it more deeply, repair it more precisely, and repeat that process across every major cause of death.
Longevity escape velocity is the point at which medical progress adds more than a year to our remaining healthy life for every year that passes.
It is an evocative idea that plays out as science and technology accelerate the progress of each other.
A person survives one cancer because a model detects it early.
That buys enough time for a better immunotherapy.
The immunotherapy buys enough time for improved cardiovascular repair.
Better cardiovascular medicine buys enough time for treatments that slow neurodegeneration.
Each advance becomes a bridge to the next.
The path may look less like immortality and more like a staircase.
AI matters because it can steepen the staircase and accelerate how rapidly we climb it. It can make every researcher more capable, every laboratory more productive, every patient record more informative, and every experimental cycle shorter. The same general machinery that coordinates a protein-design campaign can eventually coordinate work on cell replacement, gene regulation, immune rejuvenation, organ repair, and the molecular damage of aging.
There will be failures. There will be impressive molecules that die in trials. There will be models that sound certain and are wrong. Biology will retain veto power.
Techno-optimism is not the belief that those constraints disappear.
It is the belief that intelligence is a tool for overcoming constraints, and that we are learning to manufacture more of it.
Anthropic did not show us a cure for cancer. It showed us an early version of a machine that can participate in the search. Give that machine better models, better tools, robotic laboratories, clinical data, rigorous evaluation, and scientists who know where its confidence ends, and the rate of biological learning begins to compound.
The first job of the machine is not to make us immortal.
It is to buy us time.
Then to use that time to buy more time.
That is how escape velocity begins.
Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!
Thank you for helping us accelerate Life in the Singularity by sharing.
I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.
Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.
To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏
