Scientists at Mount Sinai have uncovered a previously hidden pocket in a cancer-related protein that state-of-the-art AI systems failed to detect, highlighting both the power and significant limitations of artificial intelligence in drug discovery. The finding suggests that proteins are far more dynamic and flexible than current computational models assume, and that experimental validation remains essential even in an era of advanced machine learning.

What Makes This Discovery Important for Cancer Drug Development?

The research team focused on PKMYT1, a protein known as a kinase that controls how cells grow and divide. Because this process can malfunction in cancer, PKMYT1 has become a promising target for new cancer drugs. Most experimental drugs designed to block kinases work by targeting the ATP-binding site, the region where the protein uses the cell's energy supply to function. However, many kinases share nearly identical ATP-binding sites, making it difficult for drugs to distinguish between the desired target and other kinases, which can lead to unwanted side effects.

The newly discovered hidden pocket offers a potential solution to this problem. By targeting this alternative binding site instead of the crowded ATP-binding region, researchers could design more selective cancer drugs that avoid some of the toxicity and specificity challenges associated with traditional kinase inhibitors.

How Did AI Miss This Critical Binding Site?

The research team used AlphaFold2, a leading AI system for predicting protein structures, along with virtual screening to identify molecules that might interact with PKMYT1. They then performed X-ray crystallography, biochemical testing, and cellular studies to confirm how the molecules behaved. When they tested whether current computational approaches could predict the newly discovered binding mode using AlphaFold3 and Boltz-2, along with molecular dynamics simulations, the AI systems failed to identify it.

"Our study shows both the power and the limitations of AI in drug discovery. AI was very accurate when predicting known protein shapes, but it missed a completely unexpected binding pocket that we could only uncover experimentally," said Avner Schlessinger, Professor of Pharmacological Sciences and Director of the AI Small Molecule Drug Discovery Center at the Icahn School of Medicine at Mount Sinai.

Avner Schlessinger, Professor of Pharmacological Sciences at Icahn School of Medicine at Mount Sinai

One of the most striking findings was that a very small chemical modification caused the molecule to switch from binding in the hidden pocket to binding in a conventional way. This discovery reveals that proteins are incredibly dynamic and sensitive to subtle molecular changes, constantly shifting between different shapes rather than existing in a single fixed form.

Steps to Improve AI Systems for Future Drug Discovery

• Incorporate Dynamic Protein Modeling: Train AI systems to recognize that proteins are flexible structures that shift between multiple conformations, not static shapes frozen in a single state.
• Integrate Experimental Validation: Combine computational predictions with laboratory experiments like X-ray crystallography and biochemical testing to catch binding sites that AI models overlook.
• Expand Training Data: Use results from experimental discoveries, including unexpected binding pockets, to retrain and refine AI models so they learn to predict previously hidden protein states.

The Mount Sinai team plans to develop more potent compounds that target the newly discovered site and investigate whether similar hidden pockets exist in other cancer-related kinases. They also hope to refine computational methods so AI systems can better predict these hard-to-detect protein shapes in the future.

What Does This Mean for the Future of AI in Materials and Drug Science?

While AI has transformed drug discovery by accelerating the identification of promising molecular candidates, this research underscores that artificial intelligence is not a replacement for human expertise and experimental work. The findings suggest that future AI systems must be designed with a deeper understanding of protein dynamics and the limitations of current computational approaches.

Interestingly, the broader field of AI-driven materials science is exploring similar collaborative approaches. Researchers at Argonne National Laboratory are developing a roadmap for using large language models, or LLMs, to accelerate battery research and materials discovery. These AI systems could text-mine hundreds of millions of research papers, analyze performance datasets, and even help coordinate multiple AI agents working on different aspects of materials science simultaneously.

"LLMs can be integrated with existing battery research tools, such as simulation software and material property databases. This can help scientists create AI-powered, self-driving laboratories that accelerate the research process through automation," explained Guiliang Xu, an Argonne chemist and corresponding author of the battery research review.

Guiliang Xu, Chemist at Argonne National Laboratory

However, the Argonne researchers also identified significant technical challenges that must be addressed before AI can fully realize its potential in materials science. These include the need for high-quality training data, computational efficiency, and protocols for effective collaboration among multiple AI agents. Importantly, they noted that researchers traditionally publish only successful results, but for AI systems to work optimally, they also need to be trained on failure data, such as materials with poor experimental performance.

The Mount Sinai discovery reinforces a critical lesson for the entire field: as AI becomes more sophisticated, the partnership between computational tools and experimental science becomes more important, not less. The hidden pocket in PKMYT1 represents not a failure of AI, but rather a reminder that breakthrough discoveries often require human intuition, rigorous experimentation, and the humility to recognize what machines cannot yet see.