The study presents the DrugCLIP method, which uses deep contrast learning for ultrafast genome-wide virtual screening of potential drugs.[1][3] DrugCLIP places protein binding sites ("protein pockets") and small molecules in a common "hidden" space, in which it can quickly assess their mutual compatibility.[1][3] Compared to classical molecular docking, this method is up to 10 million times faster, while on standard test sets (DUD-E, LIT-PCBA) it also achieves higher accuracy than traditional docking and modern deep learning methods.[1][3] The model has shown good ability to generalize between different chemical types of molecules and protein families and is robust to changes in protein structure.[1][3] The authors performed both computational tests and laboratory validation and confirmed that DrugCLIP can identify potent agonists or inhibitors for multiple target proteins, in some cases using only structures predicted by AlphaFold2.[1][3] The speed of the method enables screening in the range of trillions of protein-molecule pairs, including an open screening of approximately 10,000 human proteins against 500 million small molecules, creating the basis for a genome-wide survey of the "treatable" human proteome.[2][3]