Preprint: DrugJEPA screened 96,4 billion molecules in 28 hours, and laboratory tests identified hits for three proteins
Preprint: DrugJEPA screened 96,4 billion molecules in 28 hours, and laboratory tests identified hits for three proteins
On September 7, researchers at Sichuan University published a preprint on DrugJEPA, a model designed to find small molecules that can bind to a specified protein. The model rapidly compares previously computed representations of the protein pocket and the molecule's three-dimensional shape. After screening, the authors purchased or synthesized 40 compounds for each of three targets and tested them in the laboratory.
Drug discovery often begins with a target protein. Chemists look for a small molecule that can alter its function. Conventional virtual screening uses docking, which predicts how a molecule might fit into a protein pocket. This calculation must be performed separately for every protein and molecule pair, so screening tens of billions of candidates requires substantial computing resources.
DrugJEPA converts the description of a protein pocket and the three-dimensional shape of a molecule into compact numerical representations in advance. During screening, it compares these representations and ranks the candidates. In the ZEUS-3D library, 96,4 billion molecules are represented by 450,6 billion conformations, which are alternative three-dimensional shapes of the same molecule. The authors report that the initial pass through this prepared database took 28 hours on four graphics processing units.
The list was then narrowed in several stages. The researchers randomly sampled a limited number of molecules from intermediate sets that were too large to process in full. They grouped similar molecules to retain chemically diverse structures, then docked representatives of each group against several structures of the target. From the top 10 thousand candidates for each target, they manually selected 40 compounds that could be synthesized or purchased for subsequent experiments.
The next stage consisted of laboratory testing. For TYK2, an enzyme involved in immune signaling, 16 of 40 compounds inhibited the selected region of the protein by more than half at 10 micromolar. Six had an IC50, the concentration required to reduce activity by half, below 1 micromolar, and the best candidate had an IC50 of 1,6 nanomolar. For the receptor protein TAAR1, eight of 40 compounds produced more than half of the maximum response in the assay at 100 micromolar. For PUS1, an enzyme that modifies RNA, 12 of 40 compounds inhibited activity by more than half at 50 micromolar. The best candidate had an IC50 of 0,85 micromolar.
Rapid comparison of numerical representations provided the first screening stage. Docking and laboratory assays then reduced the large candidate pool to compounds suitable for further chemical work. Across the three experimental series, this sequence identified active compounds for different protein targets.