AI screening method to speed up drug discovery

Researchers from Tsinghua University and Peking University have unveiled DrugCLIP, a new AI-based screening program that they say could accelerate drug discovery (Science 2026, DOI: 10.1126/science.ads9530).

More than 20,000 human proteins are estimated to have druggable hotspots, making up the so-called “druggable genome.” But researchers have found viable ligands for only a fraction of these proteins. One reason for this is the computing power needed to create large-scale virtual displays in a timely manner. The team, led by Yanyan Lan of Tsinghua University, says its DrugCLIP model addresses this problem by combining contrastive learning and dense retrieval.

“This enables cross-selection campaigns involving more than 10 trillion protein-molecule pairs,” the authors explain by email to C&EN. These pairs, which led to 2 million small molecule hits, were generated from a virtual screen of 10,000 human proteins and 500 million compounds from large virtual libraries.

“Running a screen of this scale with a docking station or other existing methods would take years on a single server,” the researchers say, “whereas our approach achieves it in a day, making previously unfeasible filtering schemes practically feasible.”

The researchers also used the program to identify ligands that were then experimentally validated for the serotonin 2A receptor, norepinephrine transmitter, and thyroid hormone receptor interactor 12. The first two targets were chosen because of their clinical relevance, playing a role in neuropsychiatric processes, and the third was chosen because it previously had no reported inhibitors.

DrugCLIP is available free of charge, but further work will be required to optimize it.

“It is best suited to rapidly narrowing large chemical spaces and prioritizing candidates,” the authors say. “More detailed pose generation, interaction constraints, and structure-level refinement remain important directions for future improvements.” »