A fully differentiable ligand pose optimization framework guided by deep learning and a traditional scoring function

Abstract

The recently reported machine learning- or deep learning-based scoring functions (SFs) have shown exciting performance in predicting protein–ligand binding affinities with fruitful application prospects. However, the differentiation between highly similar ligand conformations, including the native binding pose (the global energy minimum state), remains challenging that could greatly enhance the docking. In this work, we propose a fully differentiable, end-to-end framework for ligand pose optimization based on a hybrid SF called DeepRMSD+Vina combined with a multi-layer perceptron (DeepRMSD) and the traditional AutoDock Vina SF. The DeepRMSD+Vina, which combines (1) the root mean square deviation (RMSD) of the docking pose with respect to the native pose and (2) the AutoDock Vina score, is fully differentiable; thus is capable of optimizing the ligand binding pose to the energy-lowest conformation. Evaluated by the CASF-2016 docking power dataset, the DeepRMSD+Vina reaches a success rate of 94.4%, which outperforms most reported SFs to date. We evaluated the ligand conformation optimization framework in practical molecular docking scenarios (redocking and cross-docking tasks), revealing the high potentialities of this framework in drug design and discovery. Structural analysis shows that this framework has the ability to identify key physical interactions in protein–ligand binding, such as hydrogen-bonding. Our work provides a paradigm for optimizing ligand conformations based on deep learning algorithms. The DeepRMSD+Vina model and the optimization framework are available at GitHub repository https://github.com/zchwang/DeepRMSD-Vina_Optimization.

zPoseScore model for accurate and robust protein-ligand docking pose scoring in CASP15

“…In this work, we demonstrate that our zPoseScore model outperforms both Vinascore and other DL-based scoring methods [20,57,34] in terms of ligand pose ranking and RMSD prediction tasks, based on experimental or predicted protein structures. The use of predicted protein structures and optimized pocket side-chain orientations for docking pose generation and ranking is more practical in the context of the rapidly growing structural proteome study and functional annotations of predicted structures in AlphaFold DB [58].…”

“…The test sets (datasets 3 and 4) are designed to compare the docking pose scoring performance with other methods [34,20]. And to more strictly evaluate the performance of the model, datasets 5 and 6 are designed by collecting the most recent protein-ligand experimental complexes and these complexes are quite different from the complex structures in PDBBind 2019 general set, which is often used as training data in many DL or ML based scoring functions for docking pose scoring [37,31,20,34]. Therefore, datasets 5 and 6 are the most strict testing datasets for docking pose scoring evaluation and performance comparison.…”

“…The selected five poses by clustering and visual inspection thus were further optimized using DeepRMSD [34] guided by DeepRMSD+Vina scoring scheme for no more than 100 steps. The last frames of the optimized poses thus were submitted under group AIchemy_LIG2 (see Figure S1).…”

“…The early models primarily focused on scoring native poses as opposed to docking poses, thus limiting their use in real drug design or virtual screening tasks [30]. Recent advancements in the field of Machine Learning (ML) and Deep Learning (DL) have improved performance on ligand binding pose scoring, leading to more accurate docking pose selection [20,31,32,33] or optimization [34]. For instance, 3D CNNbased pose prediction models [20] have demonstrated the potential of Graph Neural Networks (GNNs) as a viable alternative for docking pose prediction and virtual screening tasks [35,36].…”

Variational graph encoders: a surprisingly effective generalist algorithm for holistic computer-aided drug design

“…In both docking cases a uniform box size of 20 Å was used. Rescoring functions, namely OnionNet-SFCT 60 , RTMScore 61 and DeepRMSD 62 were applied on the QuickVina2.1 docked poses. Fivefold cross-validation as described in section 6.5 was then used to evaluate model efficacy.…”