Will AI Cure Cancer? What the Evidence Shows
AI could help make more cancers curable by improving how treatments are discovered, designed, and matched to patients. It is already contributing to cancer research. But there is no evidence-based date when AI will cure every cancer, and faster computation alone cannot guarantee a successful treatment.
The recent melanoma vaccine results show why the question deserves a serious answer. They also show why it matters to separate the contribution of an algorithm from the outcome of a clinical trial.
Cancer is a group of diseases, and some cancers can already be cured with existing care. Progress will likely arrive through better answers to particular problems: preventing a recurrence, overcoming drug resistance, or finding a treatment for someone whose cancer has few options.
Did AI help produce the melanoma vaccine results?
AI was involved in designing the individualized vaccine. That is supported by the developer's account of its technology. The clinical benefit, however, belongs to the treatment combination tested in patients.
On August 19, 2026, Moderna and Merck announced positive phase 3 results for intismeran autogene, also called V940 or mRNA-4157, combined with Keytruda (pembrolizumab). The study enrolled 1,137 patients with high-risk melanoma after complete surgical removal of their tumors.
Compared with Keytruda alone, the combination improved recurrence-free survival and distant metastasis-free survival. These outcomes measure time until recurrence or death, and distant spread or death, respectively. The initial announcement did not provide detailed efficacy figures. Follow-up for overall survival was continuing, and the vaccine remained investigational.
Where AI fits
In a 2023 explanation of its vaccine platform, Moderna described algorithms that analyze sequencing data from tumor and blood samples. They select up to 34 candidate neoantigens: unusual protein targets arising from mutations in that patient's tumor.
The selection problem is complicated. A tumor can contain many mutations, but only some are likely to produce targets that the immune system can recognize effectively. AI helps rank the possibilities so the vaccine can carry instructions for a chosen set.
The trial did not compare an AI-designed vaccine with an otherwise equivalent vaccine designed without AI. It therefore cannot tell us what percentage of the benefit came specifically from AI. The result reflects the vaccine, its delivery, the immune response, Keytruda, and the treatment setting together.
AI's role in the development process is real; claiming that AI itself cured those patients goes beyond the evidence. Our explanation of what mRNA vaccines are and how they work covers the underlying biology.
How can AI speed up cancer drug discovery?
Drug development involves a series of decisions under uncertainty. Researchers need to choose a biological target, find something that affects it, test whether that effect helps, and establish whether the treatment can be used safely.
AI can make some of those decisions better informed and help researchers prioritize which experiments to run. In this setting, AI includes specialized models trained on molecular, laboratory, or clinical data. It is a broader field than a chatbot answering medical questions.
Finding useful targets
A model can examine gene activity, protein interactions, and experimental results for clues about what a cancer depends on. It may suggest a protein or pathway worth investigating.
The crucial next question is whether changing that target actually harms the cancer without causing unacceptable damage elsewhere. An association in a dataset is a starting point for experiments.
Designing and prioritizing molecules
Models can propose chemical structures and predict properties that make some candidates more promising than others. This can narrow a large search before researchers spend resources making and testing compounds.
Tools such as AlphaFold 3 also predict the structures of molecular complexes, including interactions involving proteins and small molecules. Such predictions can help researchers investigate how a candidate might fit its target.
A convincing predicted fit does not establish that a molecule will become a useful medicine. It still needs to reach the tumor, remain active long enough, and work at a tolerable dose.
Learning from experiments
The strongest opportunity is a repeated cycle: models suggest experiments, laboratories test the suggestions, and new results improve the next round. Automation can make that cycle faster.
This could reduce time spent pursuing weak ideas. Whether it shortens the entire path to an effective cancer treatment must be measured separately from how quickly a model generates a candidate.
What evidence has reached human studies?
There are examples beyond computer predictions, but they answer different questions.
An AI-assisted drug reached a randomized trial
A 2025 phase 2a study in Nature Medicine tested rentosertib, a drug developed using AI for both target identification and molecule design. It involved 71 people with idiopathic pulmonary fibrosis, a lung-scarring disease.
The 12-week trial primarily assessed safety and also found an encouraging lung-function signal at the highest dose. It was small and short, and treatment discontinuations included liver toxicity and diarrhea. Larger, longer studies were needed.
This is evidence that an AI-assisted discovery process can produce a candidate worth testing in people. It is not a cancer trial, and it does not prove that AI-designed drugs generally outperform other drugs.
An AI model predicted responses to cancer drugs
NIH researchers reported a system called PERCEPTION in 2024. It used information about gene activity in individual tumor cells to predict drug response and resistance.
Researchers evaluated it using existing patient datasets, including multiple myeloma, breast cancer, and lung cancer. This was a proof-of-concept study. Patients were not randomly assigned to receive AI-selected treatment versus standard care to determine whether the model improved survival.
That distinction matters: predicting an outcome from recorded data is an earlier achievement than showing that using a prediction changes outcomes for new patients.
Will everyone get a medicine made for their own cancer?
Personalized cancer care already exists, but the phrase can describe several different approaches.
| Approach | What is personalized? |
|---|---|
| Selecting an existing treatment | Tumor testing helps identify which available drug or clinical trial might suit the patient. A new drug does not have to be manufactured. |
| Making an individualized vaccine | The vaccine's RNA instructions are tailored to selected targets from the patient's tumor, using a shared delivery and production approach. |
| Inventing a new drug for one person | A new molecule is designed and would need evidence about its production, dosing, safety, and effectiveness. This is a much broader undertaking. |
For the first approach, biomarker testing may identify a genetic change linked to a treatment that already exists. AI could help interpret complex results, but precision medicine does not depend on AI for every decision.
For individualized vaccines, the promise is closer to the scenario of analyzing a tumor and creating a matching treatment. Changing the RNA instructions within a platform is different from inventing a completely new chemical drug for each patient.
Sequencing also has limits. It reads biological information; it does not automatically tell us which finding matters most or what treatment will work. A mutation may have no available drug, may be irrelevant to tumor growth, or may be present in only part of the cancer.
Tumor testing is also different from testing inherited cancer risk. Many changes found in a tumor developed during the person's life and were not inherited from a parent.
Could AI make a personalized cancer vaccine in 48 hours?
One prominent version of this claim came from Oracle's Larry Ellison. At the January 21, 2025, White House AI infrastructure announcement, he described a future linking cancer detection, tumor sequencing, and robotic production of an individualized mRNA vaccine in about 48 hours. The remarks appear in the official event transcript.
That statement described an ambition. It did not demonstrate that a patient could receive a clinically proven, individually manufactured cancer treatment within two days of diagnosis.
There are several separate clocks:
- Getting the information: collecting a suitable sample, sequencing it, and checking data quality.
- Designing the treatment: interpreting the results and choosing targets or candidates.
- Producing a dose: manufacturing, testing quality, and arranging delivery.
- Proving the benefit: establishing safety, effectiveness, and how long the benefit lasts.
AI and automation may shorten parts of this chain. Making one step very fast does not mean the entire chain takes the same amount of time.
There is another distinction: producing a patient's batch within a treatment platform that has already been studied is different from developing that platform in the first place. A manufacturing claim should say where its clock starts and which checks are included.
Why is cancer still difficult, even with much better AI?
A tumor can contain several populations of cells
Cells within the same tumor may differ genetically and respond differently to treatment. A drug can eliminate many cells while leaving a resistant population behind.
Cancers can also change during treatment. A target may change, another growth pathway may take over, or the surrounding tissue may help cancer cells survive. Better analysis can help researchers study these problems, but the biology still has to be overcome.
Drug effects extend beyond the target
A drug that works in a laboratory may behave differently in a person. Distribution through the body, metabolism, interactions with other medicines, and effects on healthy tissue all matter.
This is why increasingly capable models still need experimental evidence. The FDA's drug-development outline distinguishes discovery from preclinical research, human studies, regulatory review, and ongoing safety monitoring.
The data may leave important people out
A model can perform well in one dataset and less well at another hospital or in a population poorly represented during training. Differences in sample preparation, equipment, previous treatment, and patient characteristics can affect what it learns.
Validation should therefore ask whether the tool works on independent data and whether using it improves care. A high prediction score alone does not answer both questions.
Access remains part of the problem
An effective treatment has limited impact if patients cannot obtain testing, reach a treatment center, or receive the medicine in time. Personalized manufacturing and specialist care also require resources.
AI may help with some of these tasks, but lower computing costs do not automatically translate into affordable treatment. Access needs to be measured alongside scientific performance.
What would convincing progress look like?
The strongest evidence would show improvements that patients experience: fewer recurrences, longer survival, better quality of life, or less treatment toxicity. Reliable gains in manufacturing speed and access would also matter.
A shrinking tumor can be encouraging without establishing a cure. Likewise, a trial showing fewer recurrences during follow-up does not guarantee that cancer will never return. The outcome measured and the length of follow-up need to stay attached to the headline.
When assessing a new AI cancer claim, ask three questions:
- What did AI actually do? Select targets, design a molecule, predict treatment response, or perform another task?
- Where was it tested? In a simulation, in a laboratory, using past patient records, or in a prospective clinical trial?
- What improved? Prediction accuracy, research speed, treatment outcomes, or access to care?
There is a credible path for AI to help more people survive cancer. The melanoma vaccine program illustrates one route: combining computational selection with sequencing, mRNA technology, immunotherapy, and clinical testing.
How far and how quickly that progress will extend remains uncertain. Better cures for particular cancers, more durable control of others, and fewer people receiving ineffective treatment would all be meaningful advances—even without a single moment when someone announces that AI has cured cancer.
This article provides general information and does not replace individual medical advice.