Pitt researchers use AI-generated fingerprints to find existing compounds for drug research

University of Pittsburgh researchers are using generative AI to find existing compounds with potential for drug development, rather than invent molecules that may be impossible to make. In a November 10 account, Pittwire describes a framework that identified compounds unlike its training examples that could still bind the biological targets being screened.

Junmei Wang, a professor of pharmaceutical sciences at Pitt’s School of Pharmacy, outlined the approach in a Science Advances paper published October 28. Called Full-Spectrum Generative Lead Discovery, or FSGLD, it connects AI-generated chemical patterns with compounds researchers can find in databases and take forward for testing.

The starting problem is practical, not just computational. Generating a molecular design does not necessarily give researchers something they can make. Wang said many generated compounds cannot be synthesized. His team’s approach instead uses the generated output as a guide to finding real molecules, avoiding the need to synthesize a new molecule for that discovery step.

Generate the pattern, then search for a match

The system works with molecular fingerprints: digital encodings of a molecule’s key chemical features. These are simpler representations of molecules, rather than complete new molecular structures. Wang’s lab uses a system called DRUG-GAN to learn patterns in those fingerprints that might point toward useful properties.

DRUG-GAN has two parts with opposing jobs. A generator produces artificial fingerprints, while a discriminator tries to distinguish them from fingerprints of real molecules. That back-and-forth provides the training process. Over time, the generator learns to produce fingerprints that more closely resemble those of real compounds.

Different compounds, selected targets

The team screened for compounds associated with G protein-coupled receptors, usually shortened to GPCRs, and protein kinase targets. Pitt describes these as two commonly targeted groups in drug research. The goal was target-specific discovery: finding molecules that bind to and act on a biological target related to a disease or condition.

The team reported finding compounds different from the molecules used to train the GAN that could nevertheless bind the selected targets. The finding concerns potential leads for therapeutic development, not demonstrated treatments.

Wang describes FSGLD as a reusable framework that can be adapted to different compound-discovery approaches, rather than a method limited to these targets. The work received support from the National Institutes of Health and the National Science Foundation.