AI in Drug Discovery and Development

Developing a new drug generally takes 10 to 15 years and carries a 90 percent failure rate with a hefty price tag of $1 billion to $2 billion.¹ Artificial intelligence (AI) is changing that. By shortening these timelines, reducing costs and improving the odds of success, AI is beginning to reshape every stage of drug discovery and development. It is not replacing scientists, but it is helping them to make better decisions faster. 

Many aspects of drug development have been painstakingly slow because they depend on unraveling complex biological systems. Disorders such as cancer, autoimmune disorders and neurodegenerative conditions involve thousands of interacting genes, proteins and biochemical pathways that are difficult for even the most experienced researchers to fully understand in a timely manner. 

But, AI has the ability to recognize patterns within enormous datasets, an ability nearly impossible for humans. It can also identify previously overlooked relationships between genes and diseases, offer a prediction of how drug candidates may interact with target proteins and estimate whether a compound is likely to succeed or fail.

That’s why, while still in its early stages, AI may become one of the most influential technologies in modern drug development. Let’s examine each stage of development, explore some recent advances and assess the limitations. 

The Catalyst

One of the fundamental problems scientists have faced in biomedical research is determining the three-dimensional structure of proteins. A protein’s shape largely determines its function. When researchers know the shape of a protein involved in a disease, they are often better able to design drugs that bind to it more effectively. Historically, solving these structures has been one of the slowest and most technically challenging tasks in biomedical research. 

For decades, researchers have relied on experimental techniques, such as X-ray crystallography, cryo-electron microscopy and nuclear magnetic resonance spectroscopy, to determine protein structures. While these methods produce accurate results, they are costly and time-consuming. And some proteins are so difficult to analyze that researchers cannot determine their structures at all. 

All this started to change in 2018 when Google DeepMind, Google’s AI research laboratory, first introduced the promise of an AI-driven protein structure prediction with a model called AlphaFold. It demonstrated that deep learning could predict protein structures with unprecedented accuracy, setting the stage for its successor. 

Two years later, Google DeepMind introduced AlphaFold2 — an AI model that could predict protein structures from amino acid sequences with near-experimental accuracy, and it could do so in minutes or hours. The work earned Demis Hassabis, PhD, a British artificial intelligence (AI) researcher and entrepreneur and John Jumper, PsyD, an American chemist and computer scientist, a share of the 2024 Nobel Prize in chemistry for developing AlphaFold, while David Baker, a biochemicist, shared the prize for computational protein design.² 

Their work was a game-changer. Instead of solving one protein structure at a time, researchers could use the model to predict the structures of virtually all known proteins — currently more than 200 million — and it made those predictions available through a database to scientists around the world. 

Andrei Lupas, PhD, a molecular biologist at what was then known as the Max Planck Institute for Developmental Biology (now the Max Planck Institute for Biology Tübingen), had spent more than 10 years trying to determine the structure of a protein his laboratory was studying before AlphaFold2 successfully predicted it.³ “This will change medicine,” Lupas said at the time. “It will change research. It will change bioengineering. It will change everything.”⁴ 

Since then, AlphaFold3 was released in 2024, taking the process a step further. AlphaFold2 focused on predicting the three-dimensional structure of proteins. Building on AlphaFold2, AlphaFold3 expanded beyond predicting what a protein looks like to predicting how it interacts with other molecules. That distinction is important because drug discovery depends not only on understanding a protein’s structure but also on understanding how it interacts with other biologically important molecules.

For instance, could a potential drug bind to its intended target? How tightly? Might it also interact with DNA, RNA or other proteins? Instead of simply revealing a protein’s structure, AlphaFold3 helps researchers understand the molecular interactions that influence how future medicines may work.

From Prediction to Creation

AlphaFold’s success sparked an explosion of AI innovation. Competitors such as Boltz-1, Chai-1 and Meta’s ESMFold have expanded the field, offering researchers additional options for predicting protein structures, although they differ in design, speed and capabilities. Yet predicting protein structures was only the beginning. 

Researchers soon began asking whether AI could help design the next generation of medicines. Rather than simply analyzing existing data, these new AI systems can propose novel molecules, predict their properties and help researchers identify the most promising candidates for further development. Some of these AI-designed therapies are already reaching patients through clinical trials.

Patients with idiopathic pulmonary fibrosis (IPF) — a progressive and often fatal lung disease with limited treatment options — have reason for cautious optimism, thanks to Insilico Medicine’s AI platform. It identified a novel biological target, designed a new molecule (INS018_055) and advanced it into clinical trials in approximately 30 months,⁵ which is much faster than the traditional timeline. Phase IIa clinical trials were encouraging, and they progressed into larger clinical studies. The results suggest the drug may slow disease progression. 

Seventy-one IPF patients in China, between July 2023 and June 2024, were randomly assigned to receive the drug Rentosertib (formerly INS018_055/ISM001-055) 30 mg once daily, 30 mg twice daily or a placebo for 12 weeks. Investigators reported Rentosertib was safe and generally well-tolerated across all dosage groups. The 60 mg once-daily group showed the strongest improvement in lung function with the mean forced vital capacity increasing by approximately 98 mL over the 12-week period. However, the placebo group experienced a mean decline of approximately 20 mL.⁶ 

That same study reported favorable changes in several serum biomarkers associated with pulmonary fibrosis, which suggests the drug affected the biological pathways it was designed to target. Larger trials are needed to confirm these findings.

In July 2026, Insilico announced the drug had entered Phase III clinical testing.⁷ The Phase IIa results, coupled with the advancement to Phase III, represent one of the first demonstrations that an AI-generated drug candidate can progress from AI-assisted discovery to encouraging clinical outcomes. That doesn’t prove AI has led to a cure for IPF, and Rentosertib may not ultimately receive regulatory approval. But, the fact that AI played a central role in generating a drug candidate capable of advancing this far is encouraging.

The Next Generation and Beyond

Besides IPF, AI-driven platforms are being used to accelerate multiple stages of early drug discovery across a growing range of therapeutic areas.

Isomorphic Labs is using AI to design therapies for cancer and immunological and inflammatory disorders in collaboration with Novartis, Eli Lilly and Johnson & Johnson.⁸

Recursion Pharmaceuticals is using its AI-powered Recursion OS to identify drug targets and design new therapies for oncology and rare diseases. The company’s Phase II investigational therapy REC-4881 for familial adenomatous polyposis emerged from therapeutic insights generated by Recursion OS.⁹ 

Xaira Therapeutics, which launched in 2024, is being more secretive with its AI-first platform, only revealing they are aiming at “really hard targets and large molecule therapeutics,” with immunology being one area of interest.¹⁰ 

While all of this is going on, researchers are exploring ways to push the technology even further. One of the most promising areas of research is the use of AI to design complex therapeutic modalities, including antibodies, engineered proteins and RNA-based therapies.¹¹ These are often more difficult to design than traditional small-molecule drugs because of their size and structure. Researchers hope AI will help identify and optimize these therapies more quickly, which would potentially open new treatments for diseases that have proven resistant to conventional drugs.

Scientists are also investigating how AI can make drug discovery more personalized. For example, rather than developing a single therapy for all patients with a disease, future AI platforms may be able to analyze genomic, proteomic and clinical data to identify treatments best suited for specific populations or individuals with unique genetic mutations.¹² 

Other researchers are focusing their efforts on AI-assisted drug repurposing that evaluates existing medications for new therapeutic uses.¹³ This sort of use of AI technology is at the early stages of development, but it does illustrate how AI’s role may expand well beyond today’s drug discovery process.

AI Limitations

Even as AI continues to advance, it remains a tool with significant limitations. AI is only as good as the data on which it’s trained. Incomplete or biased datasets, poor-quality experimental data and underrepresentation of certain diseases or populations can all reduce the accuracy and reliability of AI models.

AI models often perform well when tested using the datasets on which they were trained, but their accuracy may decline when applied to new patient populations, different patient populations, laboratory settings or diseases with limited available data. When accuracy drops for any of these reasons, the model is said to have poor generalizability. Ensuring AI models perform reliably across diverse populations remains an active area of research.

As such, AI predictions still require laboratory validation. While AI models can predict protein structures, drug-target interactions, toxicity and binding affinity, every prediction still needs to be confirmed experimentally. In short, AI generates hypotheses, rather than proof. 

AI can model many interactions, but biology is incredibly complex. Living systems involve immune responses, metabolism, off-target effects and disease heterogeneity. Promising compounds can still fail because living systems are more complex than computer models can fully capture.

AI also has a “black box” problem, meaning there is often little transparency in how complex models arrive at their outputs. Even when deep-learning systems produce highly accurate results, they may not explain how or why they reached those conclusions. That lack of interpretability makes researchers and regulators cautious, particularly because FDA reviewers generally expect evidence that can be independently evaluated.

Finally, regulatory and ethical challenges remain. Standards for transparency, reproducibility and validation continue to evolve, and U.S. Food and Drug Administration guidance for AI-assisted drug development is still being refined. 

A Scientific Enhancement

Despite these challenges, AI is transforming drug discovery in ways that would have seemed impossible a decade ago. It’s helping researchers solve complex problems more quickly and efficiently than ever before. The early successes of AlphaFold and AI-generated drug candidates entering clinical trials demonstrate these are not just theoretical concepts but rather practical tools that are influencing modern medicine.

At the same time, AI is not a substitute for careful scientific investigation. Computer models cannot replace laboratory experiments, clinical trials or the expertise of biologists, chemists, physicians and regulatory scientists. Every prediction generated by an algorithm must still be validated through rigorous testing to ensure new therapies are both safe and effective. As AI systems continue to evolve, improvements in data quality, model transparency and regulatory oversight will be essential in realizing their full potential.

Perhaps AI’s greatest contribution is not replacing the scientific method but enhancing it. By uncovering patterns that might otherwise remain hidden, speeding up the process and helping researchers prioritize the most promising drug candidates, AI allows scientists to focus their expertise where it matters most. 

While many challenges remain, AI is poised to become an indispensable partner in the search for tomorrow’s medicines.