ReverseScreen.ai: Pharmacophore-Guided Reverse Screening Across the Growing Co-Complex Proteome

Pharmaceutical firms frequently employ forward screening to discover prospective ligands for biological targets. Reverse screening is also gaining popularity, enabling proactive detection of adverse side effects and opportunities for repurposing existing drugs. However, screening every known receptor site with a molecule is computationally expensive.

In this study, researchers introduce a method that streamlines this process by indexing every co-crystallized ligand in the Protein Data Bank using their three-dimensional pharmacophore fingerprint, along with the UniProt accessions against which they were solved. A query molecule is fingerprinted in just 4 milliseconds, and the nearest neighbors are retrieved in 40 milliseconds.

For a given molecule, its five most similar indexed ligands are pooled, revealing an average of 21.1 distinct proteins. When tested on 3,000 molecules not included in the training set, 48.8% of the time, the molecule's known target was found within this pool. Docking these 21.1 proteins takes only 108 seconds, compared to 40.5 hours for the entire panel.

By indexing structures only based on their ligand-protein association rather than pocket-grade coordinates, the method covers 28,579 target sites, up from 3,670 previously. The technique was validated using two 2025 clinical molecules, with orforglipron identifying the GLP-1 receptor first among the 27,797 screened and daraxonrasib finding the KRAS and cyclophilin A tri-complex fourth.

When run in batches, the pipeline can screen approximately 96,000 molecules per hour on a single core, scaling to over 2.3 million per day, making the retrieval step feasible even with a vast virtual library. The method is accessible at reversescreen.ai, permitting users to screen molecules against the current index, view the retrieved sites, and dock them individually. Additionally, the index can be downloaded for confidential screening in a local setting.