← Back to Home

Generate Molecular Fingerprints, Then Find Existing Compounds: AI Screening Identifies Two Early Leads

A University of Pittsburgh team used generative AI to guide searches for molecules in chemical databases that could be tested experimentally. Tests targeting CB2 and HER4 each identified one active compound, while also showing how far computational predictions remain from drug efficacy.

By SURL BioNews

One challenge in drug discovery is whether molecules proposed by computers can actually make it into the laboratory. Even promising predictions can stall if synthesis proves difficult. A University of Pittsburgh team took another approach: have AI first generate “molecular fingerprints” describing chemical features, then search databases for similar existing compounds, turning abstract designs into research materials that can be obtained and tested.

This drug lead discovery workflow, called FSGLD, was developed by Junmei Wang’s laboratory at the university. The study was published in *Science Advances* on October 28, and the university presented the findings on November 10. It links generative models, database searches, and tiered computational screening to narrow the pool of candidates for early drug research and identify molecules worth testing experimentally.

The core tool, DRUG-GAN, is a generative adversarial network: a generator proposes molecular fingerprints, while a discriminator learns to distinguish them from real molecular data. Fingerprints are digital encodings of chemical features, rather than complete molecular structures. The team therefore used the generated results as search criteria to identify similar compounds in databases including ChEMBL and ZINC, then prioritized them for validation using molecular docking, molecular dynamics simulations, and binding free energy calculations. This allows researchers to use existing molecules at this stage, reducing the need for synthesis from scratch.

The research covered G protein-coupled receptors and protein kinases, with experimental results providing a more rigorous test than computational rankings. In tests targeting cannabinoid receptor type 2, CB2, only ZINC77179032 among nine candidate compounds showed measurable binding activity, with a binding constant Ki of 3.77 micromolar; the other eight showed no detectable binding under the experimental conditions. This result provides a starting point for further research while also underscoring that passing multiple layers of computational screening does not guarantee that a molecule will bind to its target.

Another set of experiments targeted the receptor tyrosine kinase HER4. Researchers tested two candidates, and ZINC33279960 showed kinase inhibition that increased with concentration, with a half-maximal inhibitory concentration, IC50, of 8.615 micromolar. This supports further study of the molecule as an early active compound. However, the CB2 binding assay and the HER4 enzyme inhibition assay measure different phenomena, so the two sets of values cannot be directly ranked by drug efficacy.

Database validation also has limitations that need to be distinguished. The paper notes that the active-compound training data for the CB2 model came from ChEMBL, so retrieving known active molecules from the same database may involve overlapping training data or similar structures. Such results are better understood as showing that the model directs searches toward relevant chemical features; they cannot independently prove that it has discovered entirely new molecules. Selecting compounds from ZINC and testing them experimentally adds another layer of evidence.

The most concrete findings so far are two early hits with micromolar activity, rather than established therapeutic drug candidates. Their activity, selectivity, effects in living organisms, and safety all require further study; the existing results have not demonstrated efficacy in humans. The practical value of FSGLD is that it makes AI-generated leads easier to test experimentally. Whether this can improve the success rate of drug development still requires broader comparison and validation.

References

  1. Phys.org / University of Pittsburgh
  2. Science Advances / PubMed Central
  3. University of Pittsburgh