Equivariant Flexible Modeling of the Protein–Ligand Binding Pose with Geometric Deep Learning

Abstract

Flexible modeling of the protein−ligand complex structure is a fundamental challenge for in silico drug development.Recent studies have improved commonly used docking tools by incorporating extra-deep learning-based steps. However, such strategies limit their accuracy and efficiency because they retain massive sampling pressure and lack consideration for flexible biomolecular changes. In this study, we propose FlexPose, a geometric graph network capable of direct flexible modeling of complex structures in Euclidean space without the following conventional sampling and scoring strategies. Our model adopts two key designs: scalar-vector dual feature representation and SE(3)-equivariant network, to manage dynamic structural changes, as well as two strategies: conformation-aware pretraining and weakly supervised learning, to boost model generalizability in unseen chemical space. Benefiting from these paradigms, our model dramatically outperforms all tested popular docking tools and recently advanced deep learning methods, especially in tasks involving protein conformation changes. We further investigate the impact of protein and ligand similarity on the model performance with two conformation-aware strategies. Moreover, FlexPose provides an affinity estimation and model confidence for postanalysis.

An expanded view of ligandability in the allosteric enzyme PTP1B from computational reanalysis of large-scale crystallographic data

“…The 59 new ligand-bound structures of PTP1B we report here, obtained from computational reanalysis of existing experimental fragment screening data, constitute many-fold more structural information than is seen in the vast majority of publications about protein-ligand interactions. This is useful, as the rapidly emerging set of deep learning methods for protein-ligand docking and ligand design benefit from larger training sets of experimental protein-ligand structures [33][34][35] ; crystallographic fragment screening and improved hit detection could benefit these endeavors. On a larger scale, pharmaceutical companies house private databases that collectively contain many thousands of protein-ligand crystal structures; although making these available to developers of deep learning methods would pose a significant logistical challenge, the potential benefits to society are substantial…”

“…Importantly, so do ligands bound to proteins 3,38,39 . The potential impact of this widespread bias in the training data for current deep learning methods for modeling protein-ligand interactions [33][34][35] or protein structures more generally [40][41][42] remains to be explored.…”

Rationally Designed Novel Antimicrobial Peptides Targeting Chitin Synthase for Combating Soybean Phytophthora Blight

“…Our approach to designing anti-phytophthora peptide compounds could hold significant implications for the discovery and optimization of future antimicrobial peptides. The use of computational methods for effective drug design is an irreversible trend as it facilitates a comprehensive understanding of the structure and binding mode between target proteins and ligands, which is crucial for optimizing antimicrobial peptides [ 79 , 80 , 81 ]. Our design theory may guide others by directing the saturation mutagenesis of antimicrobial peptides, focusing on both electronic effects and spatial structures.…”

Advancing Ligand Docking through Deep Learning: Challenges and Prospects in Virtual Screening

“…However, when dealing with scenarios where only the unbound protein structure (apo) is available or when protein structures are predicted using AlphaFold2, a significant drop in performance is observed for KarmaDock and other DLLD methods. Since then, a series of flexible docking methodologies , and several de novo LD methods − (Figure D) have been developed. These methods are able to generate ligand binding poses and modify protein conformations or generate the conformations for proteins and ligands from sequences at the cost of speed.…”

“…The majority of existing DLLD methodologies ,− ,, including KarmaDock suffer from inadequate generalization capabilities, resulting in excellent performance on samples resembling the training set but decreased effectiveness or the generation of physically implausible conformations on out-of-distribution samples. Therefore, enhancing the generalization capacity of the DLLD approaches is crucial.…”