In computer-aided drug discovery, structure-based virtual screening (SBVS) predicts protein-ligand interactions to identify promising hits. A key challenge is receptor flexibility, as SBVS often fails to capture diverse activation states. Ensemble screening addresses this by using multiple receptor structures. Protein kinases, major drug targets, adopt DFG-in, DFG-out, or intermediate binding modes depending on the DFG motif and αC-helix position. KinCoRe classifies kinase active sites by locating the activation loop and αC-helix and calculating DFG dihedral angles. However, most protein data bank (PDB) crystal structures are biased toward the DFG-in conformation, especially the BLAminus state, making it difficult for SBVS to identify ligands for diverse conformations. Recent structure prediction methods, such as AlphaFold3, have been applied to generate diverse kinase conformations for ensemble docking, but this approach requires extensive docking computations. To overcome this, we propose KASSPER (kinase active site structure predictor), which uses protein and compound language models. It inputs primary amino acid sequences and ligand information to predict the kinase conformation most likely to bind a given ligand. This allows docking to a single predicted conformation rather than the entire ensemble. Applied to a kinase subset of DUD-E, KASSPER was compared with ensemble screening and showed improved enrichment factors at 1%, 5%, and 10% by 18.9%, 14.0%, and 13.3%, respectively.
Wonkyeong Jang (Sun,) studied this question.