A research team led by Baylor College of Medicine has developed a new strategy combining artificial intelligence and protein analysis to speed up the discovery of small molecules called molecular glues that could treat disease.
Molecular glues act like matchmakers inside cells: they bring a target protein to the cell's natural protein-disposal machinery, which then destroys that protein. This approach offers a way to eliminate harmful proteins entirely, rather than just blocking their function as traditional drugs do.
The researchers focused on a protein called VAV1, which is found mainly in immune cells and helps activate T cells and other immune system components. Abnormal VAV1 activity has been linked to blood cancers and autoimmune diseases. By screening thousands of molecules, the team identified compounds including NGT-201-12 that caused VAV1 levels to drop. Follow-up experiments confirmed these compounds worked through the cell's natural protein-recycling system.
This research is published in Nature Communications.
How AI predicted the molecular interaction
A major challenge in molecular glue research is understanding how these molecules recruit their targets. To address this, the researchers developed GluePlex, a computational workflow that combines artificial intelligence, protein-structure prediction tools, and physics-based modeling. GluePlex predicted how VAV1, CRBN, and the molecular glue come together to form a three-part complex.
The model identified a specific region of VAV1 known as the SH3-2 domain as essential for degradation. Experimental tests confirmed this prediction and pinpointed the exact location the glue uses. The computational prediction was made without any experimental structure of the complex and was subsequently confirmed in the laboratory.
After identifying the mechanism, the team improved the compounds through medicinal chemistry. They introduced chlorine atoms to reduce molecular flexibility and increase degradation efficiency. The researchers also discovered that some compounds degraded another protein, LIMD1, highlighting the importance of evaluating both intended and unintended targets during drug development.
This work demonstrates how artificial intelligence, structural modeling, and proteomics can work together at the earliest stages of drug discovery.