Generative biology, the use of AI to design molecules and program biological functions, has moved from laboratory promise into clinical testing. At Ai4 2026, biotechnology leaders described systems that can create proteins, generate small molecule drugs and write functional DNA, with several AI developed medicines now in human trials and one program in advanced phases of development
This shift could change how quickly new medicines reach patients, how researchers tackle diseases that have resisted conventional drug discovery and how much biological risk society is willing to accept. The same models that can create a therapeutic protein or a new drug candidate may also lower the barrier to engineering dangerous biological systems.
The larger question is whether generative biology can become a reliable medical discipline rather than an impressive design tool. Can AI produce treatments that survive the complexity of the human body, meet regulatory standards and improve patient outcomes, or will the technology prove better at generating candidates than delivering cures?
“What if drug development could become more programmatic, repeatable, and scalable?” said Gevorg Grigoryan, cofounder and chief technology officer of Generate Biomedicines. “Today, every drug is still something of a heroic, one-off effort.”
Generative Biology Wants To Turn Drug Hunting Into Engineering
Traditional drug discovery resembles a long search through an almost endless list of chemical options. Researchers identify a biological target, test huge numbers of compounds and eliminate nearly all of them. A promising molecule must bind to the right target, avoid the wrong ones, reach the correct tissue, remain stable, pass toxicology studies and prove useful in patients.
Generative biology starts from a different approach. Rather than screen whatever nature or a chemical library happens to offer, researchers ask a model to create a molecule with a chosen set of traits.
Grigoryan argued that future systems could retain what scientists learn about molecular binding, manufacturing, toxicity and clinical response, then carry those lessons into the next drug program. A failed trial would still hurt. It might no longer become a dead end.
“That is how we think about generative biology,” he said. “It is not only about generating new molecules. It is about guiding biological processes in the human body toward a desired outcome, while doing so with modularity, composability, and scalability.”
Generate Biomedicines is putting that theory into practice with therapeutic proteins. Its lead program, GB-0895, is a long acting antibody aimed at TSLP, a protein involved in airway inflammation. The company began two global Phase III studies for severe asthma, with plans to enroll about 1,600 patients in more than 40 countries.
AI Designed Drugs Have Reached Their Clinical Reckoning
Insilico founder and CEO Alex Zhavoronkov told the Ai4 audience that Insilico had roughly 31 development candidates at the preclinical candidate stage or beyond. The company, which develops small molecule drugs, has advanced one of the most mature pipelines built around generative AI. He said one program had entered Phase III, three were in Phase II and eight were in Phase I.
“Every time you nominate and advance a development candidate, there are roughly 1,200 individual steps involved,” Zhavoronkov said. “AI can accelerate many of those steps. But you have to be good at all of them to deliver a successful drug.”
He shared that AI can help researchers choose targets, generate molecules, predict chemical properties and plan experiments. While it can advance research, it cannot solve problems around poor biology, weak clinical execution, unreliable manufacturing or a trial aimed at the wrong patients.
Insilico’s leading program is rentosertib, an oral small molecule that targets TNIK and is being studied for idiopathic pulmonary fibrosis, a severe disease that scars the lungs. In July 2026, Insilico initiated a randomized Phase III trial that is expected to enroll 320 participants at 47 centers in China. While a Phase III trial does not prove that the AI worked, it does show that a candidate survived long enough to face the kind of study that can expose small efficacy gains, hidden safety problems and early assumptions that did not hold.
No generative AI designed medicine has yet secured full market approval, but Insilico’s progress brings that possibility closer, but the decisive evidence will come from clinical results, not the method used to draw the molecule.
“These drugs have reached human patients,” Zhavoronkov said. “None is approved yet, but we are getting closer.”
The Human Body Remains The Hardest Dataset
Radical Numerics CEO Eric Nguyen described current AI drug design as “the tip of the iceberg.”
The body is a shifting system. Genes, age, immunity, metabolism, disease history and previous treatments alter how a drug behaves. Two people can receive the same dose and experience very different outcomes.
“The next frontier is understanding how a drug will behave in a complex system like the human body,” Nguyen said. “That is inherently a multimodal problem.”
A molecule may bind perfectly to its intended target and still fail to reach the correct tissue. It may trigger an immune reaction or it may break down too quickly. It may also potentially affect another pathway that researchers did not measure. It may work in laboratory animals and do little for people.
Grigoryan said AI models must remain tied to extensive measurement.
“Machine-learning insights do not emerge from thin air,” he said. “They require measurable data to improve and remain grounded.”
The strongest biotechnology platforms may be those that close the loop between computation and experiment. In this scenario, a model proposes a molecule and a laboratory tests it. The result returns to the model. This process is slower than the story surrounding generative AI, but it is far more useful.
Aging Gives AI Drugmakers Their Largest Possible Market
Zhavoronkov’s vision extends beyond faster drug discovery. He wants Insilico to attack aging itself.
The company studies biological changes that occur from birth through later life, then looks for targets linked to both aging and recognized diseases. A drug might first be developed for fibrosis, cancer or another diagnosable condition. Researchers could later study whether the same mechanism affects the broader decline associated with age.
“We build models that seek to understand aging from birth to death, identify therapeutic targets that are implicated both in aging and in disease, and then design small molecules for specific diseases with the hope that some may eventually be repurposed or extended to address aging itself,” Zhavoronkov said.
He called such candidates dual purpose therapeutics.
“Aging is a problem everyone has,” he said. “Regardless of how hard we work, we eventually decline and become more susceptible to disease.”
Zhavoronkov suggested that GLP-1 medicines, now used widely for diabetes and weight management, might eventually show a measurable longevity benefit.
“GLP-1 drugs are a new class of medicines that help regulate weight and metabolism,” he said. “They may be the first therapeutics to provide a meaningful longevity benefit, perhaps even in otherwise healthy people. Remember my words. I think we will see papers on that.”
Longevity research attracts serious scientists, aggressive investors and dubious claims in nearly equal measure. AI may help identify links between aging and disease that human researchers missed. It cannot shorten the years needed to show that healthy people live longer. Long term studies would need to separate any drug effect from benefits tied to weight loss, cardiovascular health and better metabolic control.
Still, the impacts of AI in longevity research can be substantial. “If you give one additional year of life to everyone on the planet, you generate roughly 8.3 billion life-years,” Zhavoronkov said. “That is the equivalent of well over 100 million human lifetimes at current life expectancy. The impact would be enormous.”.
The Emerging Role of Regulation in AI Drug Development
Regulators are preparing for a growing volume of AI supported evidence. In January 2026, the U.S. Food and Drug Administration and the European Medicines Agency published 10 principles for good AI practice in drug development. The principles cover data governance, model performance, human oversight, risk assessment and a clearly defined context for each system’s use. The message to drugmakers is that regulators will not accept “the model said so” as evidence.
“Think about modern language models,” Nguyen said. “They are trained on human language. We are now training models on the direct substrate of life.”
The danger further compounds when considering potentially malicious use of AI in drug development.
“When you tell people that AI generated a virus, it scares them,” Nguyen said. “In this case, the virus infects bacteria, not humans. But the experiment illustrates the fine line between using science to improve human health and using, or misusing, the same science to create something dangerous.”
Nguyen argued that biological design and biological defense must advance together. That could mean stronger sequence screening, tighter laboratory controls, improved pathogen surveillance and defensive models trained to detect dangerous constructs before they are manufactured.
“In our view, these are two sides of the same coin,” he said. “The technology that allows you to program biological function with DNA and create an organism also lowers the barrier for someone who wants to create something dangerous.”
