Artificial intelligence is beginning to produce something more important than medical chatbots and automated diagnoses: experimental treatments for diseases that have resisted conventional drug discovery for decades.
AI systems can search genetic data, medical literature, protein structures and enormous chemical libraries far faster than a human research team could examine them manually. They can identify biological targets, propose new molecules and predict which candidates are most likely to interact with a disease-causing pathway.
The progress is real, but the word “cure” needs caution. Most AI-generated treatments remain in laboratories or early clinical trials. Artificial intelligence can help researchers find a promising direction, but it cannot prove that a treatment is safe, effective or durable without years of laboratory testing and carefully controlled human studies.
“Incurable” Does Not Always Mean Impossible to Treat
Diseases are often described as incurable when medicine cannot reverse their underlying cause. Patients may still receive treatments that reduce symptoms, slow progression or prevent complications.
The distinction is especially important for rare diseases. The World Health Organization says more than 300 million people live with one of over 7,000 known rare conditions, and more than 95% still lack an effective treatment. Many are progressive genetic disorders that begin during childhood. (WHO’s rare-disease resolution recognises rare diseases as a global health priority.)
Traditional pharmaceutical development often struggles with these conditions because patient populations are small, biological data are limited and the expected commercial return may not justify a large research programme.
AI changes the economics of the earliest stages. It can compare information across thousands of conditions and millions of possible compounds, helping researchers find candidates that might otherwise remain unnoticed. That does not remove the expense of manufacturing and clinical trials, but it may make starting the search more practical.
AI Can Identify Both the Target and the Drug
Conventional drug discovery frequently begins with a biological target, such as a protein involved in inflammation, tumour growth or tissue damage. Researchers then test compounds to find one capable of changing that target’s behaviour without causing unacceptable harm elsewhere in the body.
AI can assist at both stages. It can analyse gene expression, proteins, disease pathways and clinical information to suggest a target. Generative models can then design previously unknown chemical structures predicted to interact with it.
The US Food and Drug Administration says AI is now appearing throughout the drug-development lifecycle, including nonclinical research, clinical development, manufacturing and post-market monitoring. The agency had already reviewed more than 500 drug and biological-product submissions containing AI components between 2016 and 2023, and says the number has continued to increase. (The FDA’s AI drug-development overview explains how regulators are approaching the technology.)
This means AI is no longer confined to speculative laboratory demonstrations. It is contributing to real development programmes being reviewed by regulators, although only a small number of genuinely AI-originated drugs have reached later-stage human testing.
An AI-Discovered Lung Drug Has Reached Phase 2
The clearest example is rentosertib, formerly known as ISM001-055. The experimental medicine is being developed for idiopathic pulmonary fibrosis, or IPF, a progressive condition in which scar tissue gradually damages the lungs.
Existing medicines can slow decline in some patients, but they do not reverse the disease. Researchers used an AI platform to identify TNIK as a potential target involved in fibrosis and then used generative chemistry to design a molecule capable of inhibiting it.
A peer-reviewed Phase 2a study published in Nature Medicine tested rentosertib in 71 patients for 12 weeks. Participants received one of three doses or a placebo. Rates of treatment-emergent adverse events were broadly similar between the groups, although liver toxicity and diarrhoea contributed to some discontinuations.
Patients receiving the highest dose showed an average increase of 98.4 millilitres in forced vital capacity, a measure of lung function, compared with an average decline of 20.3 millilitres in the placebo group. The finding is encouraging, but the trial was small, short and designed primarily to examine safety. Larger and longer studies are needed before researchers can determine whether the drug meaningfully slows IPF or improves survival.
Rentosertib is therefore not proof that AI has cured pulmonary fibrosis. It is proof that an AI-discovered target and AI-generated molecule can progress far enough to produce a measurable signal in a controlled human trial.
AlphaFold Is Giving Researchers New Biological Maps
A treatment cannot be designed effectively when researchers do not understand the shape and behaviour of the molecule causing the problem.
Proteins perform much of the work inside cells, and their three-dimensional structures influence how they function. Experimentally determining those structures can be expensive and slow, particularly for unusual proteins associated with neglected diseases.
Google DeepMind’s AlphaFold changed that process by predicting protein structures from amino-acid sequences. The freely available AlphaFold database contains predictions for more than 200 million structures and has been used by over three million researchers in more than 190 countries. More than 30% of AlphaFold-related research is focused on understanding disease. (Google DeepMind’s five-year AlphaFold review describes the platform’s research impact.)
AlphaFold 3 extends the concept by predicting interactions involving proteins, DNA, RNA and small drug-like molecules. Such models could help researchers understand where a medicine might bind and whether a previously neglected protein offers a practical treatment target.
These predictions do not replace laboratory experiments. A convincing digital structure can still be incomplete or incorrect in a living cell. Its value is that researchers can begin with a stronger hypothesis rather than spending months or years exploring every possibility equally.
AI Could Accelerate Personalised Gene Therapies
Some rare diseases are caused by a mutation unique to one patient or a very small number of families. Traditional drug development is particularly difficult because there may never be enough patients for a conventional large trial.
In 2025, researchers created a personalised base-editing treatment for an infant with severe carbamoyl-phosphate synthetase 1 deficiency. The therapy was designed and manufactured in approximately six months and produced early clinical benefits, although the child requires long-term monitoring. The treatment itself was not simply generated by a chatbot, but it demonstrated how rapidly a patient-specific genetic medicine can now be developed. (The personalised therapy was reported in The New England Journal of Medicine and reviewed by Nature.)
AI could make similar projects faster by helping scientists select editing systems, design guide sequences, anticipate unwanted changes and plan experiments.
Stanford Medicine’s CRISPR-GPT research, for example, describes an AI “copilot” that can generate experimental plans, analyse gene sequences and predict possible off-target edits. The system has guided laboratory experiments, but it is a research tool rather than an autonomous treatment designer approved for patient care.
Regulators Are Adapting to Treatments for Tiny Patient Groups
AI can generate more rare-disease candidates, but conventional clinical-trial requirements may be impossible when only a few people have a particular mutation.
In February 2026, the FDA proposed a framework that could allow certain personalised genetic treatments to rely on small, well-controlled studies when traditional trials are not feasible. Developers would still need a credible biological mechanism, early evidence of benefit, informed consent, manufacturing controls and continued collection of real-world safety and effectiveness data.
The framework could complement AI-assisted discovery by creating a regulatory route for treatments that would never have enough eligible patients for a study involving hundreds or thousands of participants.
Greater flexibility does not mean weaker responsibility. An experimental therapy designed for one patient can still cause serious or irreversible harm. Small studies make transparent evidence, independent oversight and long-term follow-up more important, not less.
AI Still Cannot Solve the Hardest Part of Drug Development
An algorithm can predict that a molecule should work and still be wrong.
Human biology contains interactions that may be missing from training data or poorly represented in laboratory models. A compound can bind perfectly to its intended target but fail to reach the correct tissue, break down too quickly, interact with another protein or produce toxicity that was not predicted.
This is why successful computer modelling is only the beginning. A treatment must survive laboratory validation, animal or alternative preclinical testing, manufacturing assessment and several phases of human trials.
The FDA and European Medicines Agency have published joint principles calling for AI used in drug development to remain human-centred, risk-based, clearly documented and continuously assessed. The principles also emphasise data governance, defined contexts of use and multidisciplinary oversight. (The FDA’s good-AI-practice principles reflect the concern that impressive models can still produce unreliable medical evidence.)
AI Is Creating More Chances, Not Instant Cures
The most realistic promise of AI is not that one supercomputer will suddenly eliminate cancer, dementia or every rare genetic disorder.
Its strength is creating more credible opportunities for scientists to test. It can reveal a protein that had been overlooked, propose a molecule that would have been difficult to design manually or make a one-patient genetic therapy faster to plan.
For diseases that have attracted little investment or resisted decades of research, creating more viable candidates is itself a major advance. The first AI-discovered treatments entering human trials show that the technology can move beyond prediction and into medicine.
The decisive evidence will come later. A true breakthrough requires treatments that survive large trials, improve patients’ lives, remain safe over time and become affordable enough to reach the people who need them.
AI is not curing incurable diseases yet. It is beginning to make some of them look less impossible.
