AI drug discovery faces its defining test: clinical trials

After over a decade of development, AI drug discovery has moved from the lab into the clinic. 

More than 170 AI-designed drugs are now being tested in humans. These trials will prove, for the first time, whether molecules discovered by machines can safely and effectively treat patients.

In Salt Lake City, Recursion has multiple AI-enabled drugs in clinical trials — and a growing pipeline behind them. “Solving disease with AI relies on three main ingredients that have now come together: algorithms, computing, and methodology,” CTO Ben Mabey told The Infinite Loop.

Across the country in Cambridge, Massachusetts, Lila Sciences is building a system in which AI doesn’t just assist in drug discovery, but acts as an autonomous scientist. “The model’s ability to make very good educated guesses at what a good starting molecule might look like has gotten very good,” says CTO Andrew Beam.

In nearby Boston, Insilico Medicine has already reported Phase IIa clinical trial results for its pulmonary fibrosis drug rentosertib — one of the first indications that an AI-discovered small molecule can be used to treat humans. “Many of the common hurdles, like getting the project funded, identifying the target, generating the molecule, and demonstrating results in the clinic, we managed to overcome already,” said CEO Alex Zhavoronkov. “However, the final phases of the clinical trials still remain.”

That remaining stage is where success will be determined — and where most drugs traditionally fail.

For decades, drug discovery was constrained by search. To find a viable molecule, researchers used robotics to physically test millions of existing chemical compounds against a disease target. The process was slow and arduous, as candidates were limited to previously identified chemicals stored in physical libraries. Refining the top contenders into safe, effective medicines would then typically require years of manual trial and error.

AI is changing that by reducing the search space. Models trained on biological and chemical data can prioritize candidates from a mathematical map of countless molecular combinations. They can predict their properties and guide optimization before the physical testing even begins.

Insilico validated the approach with the development of rentosertib. In 2023, the molecule became the first AI-discovered and AI-designed drug to reach Phase II clinical trials. 

It began life when AI identified the protein TNIK as a potential cause of pulmonary fibrosis. Using generative adversarial networks (GANs), the system then assembled a new chemical structure that could lock into the protein’s shape, preventing it from working. Essentially, AI had "imagined" the ideal molecule to treat TNIK. Integrating reinforcement learning, the system then refined this structure to ensure it could survive the human body.

Insilico is now working to scale this approach. Twelve of the company’s drugs have reached the clinical stage, bolstered by two fully automated wet labs, Life Star 1 and 2, which test AI-designed molecules to create a real-time data feedback loop.

According to Zhavoronkov, tasks such as target identification, molecule generation and safety prediction are now performed “almost instantaneously” by AI — albeit with some oversight and modification.

“Even some experiments are run in the end-to-end fashion where AI is controlling the experimental flow,” he said. “However, when it comes to later-stage experimentation and iterative improvement of both the drug and the AI system, we still need a significant number of humans.”

That approach is now translating into commercial traction. In March 2026, Insilico signed a $2.75 billion deal with pharmaceutical giant Eli Lilly to bring its molecules to the global market. The company has now developed at least 28 drugs using AI, nearly half of which are already in clinical development.

Recursion takes a different approach: modeling biology as a data problem. 

The company has built datasets of billions of microscopy images showing how human cells respond to genetic and chemical perturbations. These are combined with transcriptomics and patient data to create machine-learning representations of biological systems.

The goal is to identify relationships between genes, proteins, pathways, and diseases before selecting a drug target. To accelerate this shift, Recursion has developed large foundation models built on transformer architectures.

The models resemble those behind the LLMs produced by frontier AI labs, but with a different data type. “Instead of being applied to text and natural images, we apply them to the language of biology — cellular images, proteins, RNA, DNA,” said Mabey.

Recursion’s infrastructure reflects the scale of that task. The company operates two supercomputers and trains models on more than 40 petabytes of proprietary biological data.

“Deep learning is data hungry, so that has to be paired with data generation,” Mabey said.

These models have been put to work identifying biological relationships, predicting drug effects, and prioritizing compounds for development before any physical synthesis begins.

Recursion also combines its own compute infrastructure with cloud systems to handle spikes in inference and training workloads. This allows the company to scale models without relying entirely on hyperscalers.

Lila Sciences is pushing the model to a new level: integrating AI directly into the experimental process.

Rather than treating the lab as a separate validation stage, the company is building systems in which models generate hypotheses, test them experimentally, and continuously learn from the results. “We’re taking that whole process and putting AI at the centre of it,” said Beam.

In practice, this means turning laboratory instruments into software-controlled systems that AI models can operate directly. Robotics systems then move samples between machines, allowing experiments to run continuously without manual intervention. The company described these systems as “AI science factories.” Beam called them “​​the world’s largest verifier for science.” 

He compared the approach to the shift from trolley cars on rails to self-driving systems. Traditional automated labs, he argued, can repeat predefined workflows, but struggle to adapt dynamically. Lila’s system, by contrast, can assemble new experimental sequences in response to model outputs. 

As the experiments run, they create a cycle of data generation and feedback that allows the models to be continuously enhanced.

Beam described the process as “scientific self-play,” a concept inspired by DeepMind’s AlphaGo mastery of the fiendishly complex board game Go. Initially, the software was trained to predict the moves that humans would make, but that alone didn’t produce superhuman play.

“The model has to be able to play the game itself,” said Beam. “It has to be able to play enough games of this to discover and make moves that people have never made before.”

Lila applies this theory to molecules. The lab provides a substrate, and the AI then makes scientific moves researchers would never have considered. Over time, the system discovers strategies beyond the limits of human design.

The approach is already changing biologics research. Traditionally, designing antibodies for new targets relied on large-scale random mutation and screening. Today, models can generate viable candidates computationally before laboratory optimisation even begins.

But Lila also has a broader goal. Every scientist, the company said, will soon have their own AI collaborator capable of searching literature, generating hypotheses, and designing experiments.

For all the progress in AI-driven discovery, physical human testing remains a tight constraint.

According to research in the US, the clinical phase lasts an average of around 95 months and accounts for 69% of overall R&D costs. Beam put it bluntly: “The biggest bottleneck is obviously still clinical trials.”

After successful testing in labs and animal models, companies must submit an Investigational New Drug (IND) application before beginning human trials. Then the timelines and costs increase sharply. “Once you're on the other side of an IND, you're looking at three to five years and half a billion to a billion dollars worth of money to run that clinical trial,” Beam said.

Approximately 90% of drugs that enter clinical trials fail to reach the market, and those failures constitute the bulk of development costs. But AI can improve the success rates. 

Recursion demonstrated this while developing the cancer drug treatment REC-1245. The process from biological discovery to lead drug candidate took just 18 months — twice as fast as the industry average.

“AI gives us better insights earlier in the process so we have greater certainty that a program will succeed or fail in patients and to identify which patients will be most likely to benefit,” said Mabey.

And it’s not only Recursion that’s reaping the benefits. An analysis published last year found that AI-discovered molecules have an 80–90% success rate in Phase 1 clinical trials. This year, Phase 2 and Phase 3 studies will provide further proof points for dozens of candidates.

Beam is optimistic about the results. He believes AI-designed drugs could reach the market “any year now,” and will eventually become the norm rather than the exception. Completing those clinical trials will be a major step towards this vision. But AI has already moved beyond simply proposing molecules in the lab. It’s now designing drugs that are advancing through human trials — and moving closer to real-world medicine.