He’s using AI to screen existing compounds for therapeutic potential

Health care researchers have long sought to harness the power of computing to design molecules that bind to specific receptors and can be used as targeted treatments for diseases. Even with the increasing sophistication of artificial intelligence, however, this technological promise has yet to be realized.

Junmei Wang, professor of pharmaceutical sciences in the School of Pharmacy, is tackling the problem from a different perspective.

In a paper published Oct. 28 in the journal Science Advances, Wang outlined Full-Spectrum Generative Lead Discovery (FSGLD), a novel, practical and reusable framework that takes advantage of generative AI to screen for existing compounds with potential for therapeutic development. 

Developing pharmaceuticals using AI isn’t new, but it is laborious and often leads to a dead end. “If you generate a brand-new molecule, that’s actually a limitation,” Wang said. “A lot of compounds can’t be synthesized, and that is a big issue.”

“We use a simpler representation of a molecule: a molecular fingerprint,” Wang said, a digital encoding of a molecule’s key chemical features. To uncover patterns in the fingerprints that might point to useful properties, Wang’s lab used a general adversarial network called DRUG-GAN. This system is made of two components: a generator that creates fake fingerprints and a discriminator that attempts to distinguish the fakes from fingerprints of real models.

This back and forth is the GAN’s training. Over time, the generator learns to generate fingerprints that more closely resemble those of real molecules. Researchers then search chemical databases for existing compounds with similar footprints; the AI-generated pattern leads to the discovery of real molecules that researchers can test — no need to synthesize a new molecule.

In this work, the lab screened for target-specific compounds, or molecules that bind to and act on a receptor related to a particular disease or condition.

They screened for compounds linked to two of the most commonly targeted drug receptors: G protein-coupled receptors (GPCRs) and protein kinase.

“We saw good results,” Wang said. The team identified novel compounds that were different from the molecules used to train the GAN, but that would still bind to the target receptors. 

The promise of FSGLD goes beyond these two receptors; much of its value lies in its flexibility.  

“It’s a reusable framework that connects generative AI with practical compound discovery.”

Wang’s lab is taking advantage of this flexibility to explore two complementary approaches. One, called ADSeqGAN, generates new molecular structures by learning from text-based descriptions of molecules, without requiring a receptor’s three-dimensional structure. Another, currently in development and called DRUG-FF, uses the three-dimensional shape of a receptor and/or a bound molecule to guide the design of new compounds that could fit the binding site.

This work was supported by the National Institutes of Health (R01GM149705) and the National Science Foundation (1955260).