A patient with idiopathic multicentric Castleman disease, a rare immune disorder with no standard cure, was told he was going to die. An artificial intelligence model then identified a repurposed drug as a top candidate for his condition, and he recovered. That single case captures a broader shift now playing out across medicine: AI systems are scanning billions of molecular combinations, predicting protein structures, and designing drugs from scratch for diseases that have resisted treatment for decades.
An AI-Designed Drug Shows Real Results in Lung Fibrosis
The strongest clinical evidence so far comes from a drug that AI both conceived and refined. Insilico Medicine used generative models to identify TNIK, a protein linked to tissue scarring, as a target for idiopathic pulmonary fibrosis, a progressive lung disease with limited treatment options. The company then used AI to design a small-molecule inhibitor called INS018_055, also known as rentosertib, which showed robust activity across preclinical fibrosis models and advanced through Phase 1 safety testing.
In a randomized Phase 2a trial published in Nature Medicine, the drug moved into human efficacy testing for idiopathic pulmonary fibrosis. Biomarker analysis from a 12-week study found a dose-dependent improvement in forced vital capacity, a key measure of lung function. At the highest dose, patients showed a mean FVC improvement of approximately 98.4 mL compared with a decline in the placebo group, according to data presented in the American Journal of Respiratory and Critical Care Medicine. The biomarker signatures also revealed antifibrotic and anti-inflammatory activity, suggesting the drug was hitting its intended biological pathways.
What makes this case unusual is that every step, from target identification to molecule design, relied on AI rather than conventional screening. As highlighted in peer-reviewed commentary on the trial, the AI-driven workflow for TNIK differed sharply from standard drug development approaches, which typically start with a short list of well-characterized targets and then sift through huge libraries of compounds. Here, algorithms proposed the target itself and then iteratively generated and scored candidate molecules, compressing years of wet-lab work into a much shorter cycle.
The evidence for TNIK as a fibrosis target was described as strong in some areas and weaker in others, a reminder that even promising AI-generated candidates face the same biological uncertainty as traditionally discovered drugs. Still, the fact that a generative pipeline produced a molecule now showing measurable benefit in patients represents a concrete advance, not a theoretical one. It also offers a template for tackling other fibrotic disorders, from liver cirrhosis to kidney scarring, where conventional discovery efforts have repeatedly stalled.
Protein Prediction Changed the Starting Line
None of this would move as quickly without the structural biology revolution that began with AlphaFold. The system, developed by Google DeepMind, demonstrated highly accurate protein prediction in work published in Nature, giving researchers three-dimensional maps of proteins that had previously taken months or years to determine experimentally. That leap turned protein structure from a rare, rate-limiting resource into something closer to a searchable commodity.
The open-access AlphaFold Protein Structure Database, maintained by EMBL-EBI and Google DeepMind, now provides AI-predicted structures used widely across academic and industry research, accelerating the earliest phase of drug design: understanding what a target looks like and how a molecule might bind to it. Instead of guessing at the contours of a binding pocket, chemists can now start with a detailed model and use computational docking or generative models to explore chemical space more efficiently.
Newer systems extend this idea beyond single proteins. AlphaFold 3, for example, predicts structures of biomolecular interactions, including complexes with ligands and nucleic acids, according to work reported in Nature. That matters because drugs rarely interact with a single isolated protein; they bind to proteins in complex with other molecules, and predicting those interactions accurately is a key step for structure-based discovery. In parallel, generative approaches such as RFdiffusion have demonstrated the ability to design entirely novel proteins with specified structures and functions, including therapeutic binders, as described in recent structural biology research. Together, these tools mean researchers can now both predict what exists and engineer what does not, closing a loop between understanding and invention.
Drug Repurposing Saves Lives AI Screening Missed
Designing new molecules is only one path. AI is also proving effective at finding new uses for drugs that already exist, often at a fraction of the cost and time required for de novo discovery. A model developed by Dr. David Kelley’s group, described in a BBC Future feature, mines large biomedical datasets to uncover unexpected connections between compounds and diseases. In one example, amphetamines typically used to treat ADHD were flagged as candidates for other conditions, with what was once considered a side effect becoming the main therapeutic action.
The most dramatic illustration involved a terminally ill patient with idiopathic multicentric Castleman disease. AI identified a top drug candidate for his condition through repurposing analysis, and the patient was effectively cured, according to reporting from the University of Pennsylvania. For rare diseases with small, geographically dispersed patient populations, traditional clinical trials are often economically unviable. AI-driven analysis of biomedical knowledge graphs can predict treatment candidates for these neglected conditions, allowing researchers to prioritize a handful of promising therapies for focused, small-scale studies instead of launching broad, expensive screening programs.
Researchers are also using AI to rapidly scan existing compounds against infectious threats. As described by MIT scientist James Collins in the same BBC coverage, machine-learning models can evaluate millions of molecules in silico to find potential antibiotics. Collins notes that his team can now, in a matter of days, screen vast chemical libraries for activity against pathogens, a process that would have taken months or years using traditional lab assays alone. This approach has already surfaced candidates against drug-resistant infections such as gonorrhea and MRSA, illustrating how repurposing and de novo discovery can blur together when algorithms roam across chemical and biological space without regard to old category boundaries.
From Lab Breakthroughs to Real-World Safety
The speed and breadth of AI-driven discovery raise a practical question: how do regulators, clinicians, and patients keep up with potential risks? Even when a compound has been on the market for years, a new use can reveal unexpected side effects in different populations or at different doses. That makes post-marketing surveillance especially important for repurposed drugs and algorithm-designed molecules alike.
In the United States, one key tool is the federal adverse event reporting infrastructure, which allows clinicians, patients, and manufacturers to log suspected drug-related problems. The government’s online safety reporting portal is designed to capture these real-world signals, feeding them into pharmacovigilance systems that look for patterns across thousands of reports. As AI accelerates the front end of discovery, systems like this become the backstop, helping ensure that rare but serious harms are detected and acted on quickly.
Some researchers argue that AI itself should be applied to pharmacovigilance, scanning free-text clinical notes, social media, and electronic health records for early signs of trouble. Others caution that black-box models could obscure rather than clarify causal links if their reasoning is not transparent. The likely outcome is a hybrid: human regulators supported by algorithmic tools, with clear audit trails and a premium on interpretability when safety decisions are at stake.
Promise, Limits, and the Road Ahead
Across these examples, a pattern emerges. AI excels at navigating complexity: sifting through vast chemical libraries, mapping intricate protein structures, and traversing tangled biomedical knowledge graphs. It can propose hypotheses at a scale and speed that would be impossible for human researchers alone. Yet every success story still depends on traditional science at critical points: validating a target in animal models, running carefully controlled clinical trials, and monitoring safety once a drug reaches the market.
Experts quoted in recent reporting emphasize that AI is not a magic wand but a powerful amplifier. It can compress the search for new therapies from years to months, and in some cases from months to days, but it cannot change the underlying biology or the ethical obligations of drug development. Negative trials will still happen; some AI-designed candidates will fail for the same mundane reasons as their conventional predecessors.
Even so, the landscape is clearly shifting. A few years ago, AI-designed drugs and algorithm-guided repurposing were largely speculative. Now, an AI-generated molecule for lung fibrosis has produced encouraging clinical signals, knowledge-graph analysis has rescued a patient with a lethal immune disorder, and structure-prediction tools are woven into the fabric of modern biomedical research. The next wave of breakthroughs will likely come not from AI alone, but from teams that treat it as a collaborator, one that can explore more possibilities than any human group, but still needs human judgment to decide which paths are worth following.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
