Improving the\nReliability of Language Model-Predicted\nStructures as Docking Targets through Geometric Graph Learning

Applying artificial intelligence techniques to flexibly\nmodel the\nbinding between the ligand and protein has attracted extensive interest\nin recent years, but their applicability remains improved. In this\nstudy, we have developed CarsiDock-Flex, a novel two-step flexible\ndocking paradigm that generates binding poses directly from predicted\nstructures. CarsiDock-Flex consists of an equivariant deep learning-based\nmodel termed CarsiInduce to refine ESMFold-predicted protein pockets\nwith the induction of specific ligands and our existing CarsiDock\nalgorithm to redock the ligand into the induced binding pockets. Extensive\nevaluations demonstrate the effectiveness of CarsiInduce, which can\nsuccessfully guide the transition of ESMFold-predicted pockets into\ntheir <i>holo</i>-like conformations for numerous cases,\nthus leading to the superior docking accuracy of CarsiDock-Flex even\non unseen sequences. Overall, our approach offers a novel design for\nflexible modeling of protein–ligand binding poses, paving the\nway for a deeper understanding of protein–ligand interactions\nthat account for protein flexibility.