AI Virtual Screening Moves Beyond the Computer: New Chemical Starting Points Identified for Four Challenging Targets

The hardest hurdle in virtual drug screening is often not whether computers can rank billions of molecules, but whether the top-ranked molecules actually work once they enter the laboratory. Deep Origin’s latest preprint attempts to answer this question through four prospective screens: the research team selected candidates from approximately 80 billion synthesizable or explorable virtual compounds and identified novel inhibitors of IRAK4, CD73, factor XIa, and IL-17A in biochemical or cellular assays.

The framework consists of DODock and DOScore. DODock first uses a diffusion model to propose possible molecular binding poses, then refines them with a physics-based energy model that describes molecular forces, and finally passes them to a learning model for ranking. DOScore combines the environments surrounding atoms with the three-dimensional shape of the binding pocket to determine which molecules are more likely to bind in practice. The key feature of this design is that predictions can still be constrained by physical rules when no similar proteins or chemical scaffolds can be found in the training data.

According to results disclosed by the company and in the preprint, the IRAK4 screen tested 127 molecules, of which 19 showed activity, for a hit rate of 15%; the most potent had a potency of 158 nanomolar. For factor XIa, 13 active molecules emerged from 299 candidates tested, for a hit rate of 4.3%. For IL-17A, a protein–protein interaction target that is more difficult to address with small molecules, 6 of 194 candidates showed activity. The researchers believe they disrupt the protein interface through a distal allosteric site.

The CD73 figures were the most striking: 56 of 183 molecules tested met the activity criteria set by the study, for a hit rate of 30.6%, and 54 had active concentrations below 100 micromolar. One representative non-nucleotide molecule had a potency of 570 nanomolar in a cellular assay. The company had previously described the same screening program as involving 160 candidates and 48 active molecules below 80 micromolar. The newer preprint uses 183 candidates and a different threshold, possibly reflecting an additional experimental batch or an expanded analysis population; the two sets of results should not be treated directly as entirely identical datasets.

In addition to wet-lab experiments, the study used several structural benchmarks to test the models’ ability to generalize. The company reported that, in the CASF-2016 test using strict data splits, 81.1% of DODock’s predictions fell within 2 angstroms of the crystal structure. In the OpenBind test, which emphasizes novel targets, the top-ranked pose had a success rate of 80%. In another blinded test, the model predicted the binding pose of the PCSK9 inhibitor laroprovstat in an atypical pocket; a crystal structure obtained later differed from the prediction by approximately 1.2 angstroms. These results support the view that the model does more than memorize existing protein–ligand combinations, but they currently still come primarily from analyses by the company developing it.

The real test is only beginning. A “hit” means only that a molecule showed activity at a specified concentration and under specified experimental conditions. Molecules active at tens to hundreds of micromolar, in particular, remain far from becoming selective, safe drugs capable of acting in animals. The hit rates across the four screens also cannot be directly compared without consistent activity thresholds, negative controls, and complete retesting data, nor can they be used to estimate the probability of clinical success.

The work has made its algorithms, hyperparameters, and data-splitting methods public, giving other teams an opportunity to check for data leakage and attempt replication. However, as of publication, the paper had not undergone peer review, and complete pharmacokinetic, toxicity, or animal efficacy data for the candidate molecules had not been presented. If independent research can reproduce such prospective hit rates on unfamiliar targets, the method’s value would extend beyond shortening computational screening to reducing the number of molecules that actually need to be synthesized and sent to the laboratory.