Deep learning generates apo RNA conformations with cryptic ligand binding site

RNA plays vital roles in diverse biological processes, thus drawing much attention as potential drug target. Structure Based Drug Design (SBDD) for RNA is a promising approach but available RNA-ligand complex tertiary structures are substantially limited so far. Then, theoretical RNA-ligand docking simulations play central roles. However, success of SBDD highly depends on use of RNA structure sufficiently close to ligand-boundable conformations, which do not necessarily appear in experimentally-resolved and theoretically sampled apo RNA conformations. To overcome such difficulty in SBDD for RNA, we leverage efficiently sampling ligand-boudable apo RNA conformations by using a generative deep learning model (DL). We succeeded to generate HIV-1 Transactivation Response Element (TAR) conformations with a cryptic MV2003 binding cavity without a priori knowledge of the cavity. These conformations bind to MV2003 with binding scores similar to those calculated for experimentally-resolved TAR-MV2003 complexes, illustrating new application of DL to promote SBDD for RNA.