Baylor’s GluePlex AI helps map a molecular glue that trash-tags immune protein VAV1

Some disease proteins are hard to switch off with ordinary drugs, so researchers have started asking the cell to throw them away instead. Scientists at Baylor College of Medicine have used an AI pipeline called GluePlex to work out how a small molecule marks VAV1, an immune-cell signaling protein tied to blood cancers and autoimmune disease, for destruction. The work appeared in Nature Communications on September 12, 2026, and Baylor highlighted it on September 29.

The molecules involved are called molecular glues. They act as matchmakers, pressing a target protein against cereblon, part of the cell's own recycling system, which then tags the target for disposal. The Baylor team, led by its Center for NextGen Therapeutics, first used high-throughput proteomics, a way of measuring thousands of proteins at once, to find compounds that lowered VAV1 levels.

Finding a compound that works is only half of the problem, because chemists also need to know where the glue grips. Normally that requires an experimental structure of the three-way complex of glue, VAV1 and cereblon, and none existed. GluePlex filled that gap by chaining three computational steps: AI prediction of likely contact hotspots, physics-based docking, and AI co-folding, which predicts how the pieces fit together.

The model pointed to an unusual contact point, a short loop in VAV1 called the RT loop. Many cereblon glues instead rely on a different, better-known feature called a G loop. Lab experiments that mutated the RT loop backed up the prediction. Chemists then refined a series of compounds, using calculations on the predicted structures to help rank which designs held the complex together best. The improved compound, NGT-201-18, reduced VAV1 and dampened T-cell activation in primary human T cells, according to Baylor and the paper. Those were cells in a dish, not patients.

The glues are not perfectly selective yet. VAV1 was the main protein degraded, but some compounds also removed LIMD1, a protein with a standard G loop, which shows that a drug built to delete one protein can catch others. Baylor and the authors also disclose a provisional patent on the NGT-201 series.

A separate research group has since reached a matching conclusion about the VAV1 loop using experiment-based methods, which independently supports the biology. What the Baylor work adds is a way for AI to map a hard-to-predict contact before anyone had solved the structure, and the GluePlex code is public on GitHub for other teams to try. The compounds themselves remain early research, a long way from a medicine.