AI is rapidly transforming drug discovery, helping scientists identify promising drug candidates faster than ever before. However, new research from the Icahn School of Medicine at Mount Sinai demonstrates that AI still has important blind spots, particularly when it comes to understanding the complex and dynamic nature of proteins.
In a study published in the Journal of the American Chemical Society, researchers identified a previously unknown druggable site in PKMYT1, a protein increasingly recognized as a promising target for cancer treatment. The discovery could pave the way for more selective cancer therapies while highlighting the continued importance of experimental validation alongside AI-driven approaches.
PKMYT1 belongs to a family of proteins known as kinases, which regulate cell growth and division. When these processes become dysregulated, cancer can develop. As a result, kinase inhibitors have become a major class of cancer drugs. However, designing highly selective inhibitors remains challenging because many kinases share similar structural features.
AI missed a crucial binding pocket
Most existing kinase inhibitors target the ATP-binding site, the region responsible for providing energy to the protein. Because this site is highly conserved across many kinases, drugs often affect multiple targets, increasing the risk of unwanted side effects.
The Mount Sinai team set out to identify alternative ways of targeting PKMYT1. Using AlphaFold2, virtual screening techniques and a range of laboratory methods, they discovered a previously hidden binding pocket that could potentially offer a far more selective approach. Remarkably, this binding site was not detected by current state-of-the-art AI models, including AlphaFold3 and other advanced computational tools evaluated during the study.
According to the researchers, the finding demonstrates that proteins are not static structures. Instead, they constantly shift between different conformations, creating temporary opportunities for molecules to bind. These transient states remain difficult for current AI systems to predict reliably.
The study also revealed that very small chemical modifications to a compound could dramatically change how and where it binds to the protein. Such subtle molecular effects further complicate computational predictions.
AI and experimental science
Rather than presenting AI and laboratory research as competing approaches, the study highlights their complementary roles. Researchers first used AI-based protein structure prediction and virtual screening to identify potential compounds. They then validated their findings through X-ray crystallography, biochemical assays and cellular experiments. These laboratory techniques ultimately revealed the hidden binding pocket that AI had overlooked.
The results reinforce a growing consensus within pharmaceutical research: while AI can significantly accelerate early-stage discovery, experimental validation remains essential for confirming biological mechanisms and identifying unexpected findings.
The researchers argue that AI currently performs best when predicting known or common protein structures. Detecting rare, dynamic or unconventional protein states remains a significant challenge.
Future cancer therapies
The newly discovered binding site could provide a foundation for developing a new generation of highly selective cancer drugs targeting PKMYT1. By avoiding conventional ATP-binding sites, future therapies may reduce off-target effects and improve treatment precision.
Beyond its therapeutic potential, the study offers valuable lessons for the future of AI in healthcare and pharmaceutical innovation. The findings may help developers improve AI models so they can better recognize hidden protein conformations and dynamic structural changes.
The research team now plans to optimize compounds that target the newly identified site and investigate whether similar hidden pockets exist in other cancer-related kinases. Their work suggests that the future of drug discovery will depend not only on increasingly sophisticated AI systems, but also on the continued integration of computational predictions with rigorous laboratory science.
Cancer drug development
In 2025, researchers at the Institute of Cancer Research developed an AI-powered technology that uses 3D imaging and deep learning to predict how cancer cells respond to drugs by analysing changes in cell shape. Unlike traditional methods that rely on 2D images, the new approach creates a more realistic representation of cellular behaviour. In tests involving nearly 100,000 3D images of melanoma cells, the system identified the drug being used with 99.3 percent accuracy, even distinguishing between treatments with similar effects.
The technology could significantly accelerate drug development, potentially reducing the process from 10–12 years to around six years less. Researchers estimate the preclinical phase could be shortened from three years to just three months. Beyond melanoma, the tool has shown promise in analysing red blood cells, brain vessels and stem cells. The findings, published in Cell Systems, are being advanced through the spin-off company Sentinal4D.
