InteractionGraphNet: A Novel and Efficient Deep Graph Representation Learning Framework for Accurate Protein–Ligand Interaction Predictions

Abstract

Accurate quantification of protein–ligand interactions remains a key challenge to structure-based drug design. However, traditional machine learning (ML)-based methods based on handcrafted descriptors, one-dimensional protein sequences, and/or two-dimensional graph representations limit their capability to learn the generalized molecular interactions in 3D space. Here, we proposed a novel deep graph representation learning framework named InteractionGraphNet (IGN) to learn the protein–ligand interactions from the 3D structures of protein–ligand complexes. In IGN, two independent graph convolution modules were stacked to sequentially learn the intramolecular and intermolecular interactions, and the learned intermolecular interactions can be efficiently used for subsequent tasks. Extensive binding affinity prediction, large-scale structure-based virtual screening, and pose prediction experiments demonstrated that IGN achieved better or competitive performance against other state-of-the-art ML-based baselines and docking programs. More importantly, such state-of-the-art performance was proven from the successful learning of the key features in protein–ligand interactions instead of just memorizing certain biased patterns from data.

Accurate Protein-Ligand Complex Structure Prediction using Geometric Deep Learning

“…LigPose also performed better than recent hybrid deep learning methods on the core set. Many researches in this őeld focused on building hybrid methods [13,15,14,11,43,16]. However, as we demonstrated in experiments (Fig.…”

“…4c, without selective poses, the success rates of all methods are decreased compared with those in CASF-2016, ranging from 62.1% to 82.5%. LigPose score -Core showed a greatly improved success rate (77.5%) compared with Smina (67.7%), and was also comparable to the best performing scoring function (i.e., DeepBSP (79.7%) [11]). Moreover, LigPose (82.5%) outperformed DeepBSP and LigPose score -Core with the notable higher success rates of 2.8% and 5%, respectively, LigPose indicating that the end-to-end learning paradigm of LigPose is more effective in real applications without pre-designing the pose candidates.…”

“…To conduct a fair comparison with the deep learning-based methods, we performed LigPose on the core set (N = 285), a well-designed but smaller subset of PDBbind. Here we summarized the success rates of LigPose with diversiőed methods, including docking tools and state-of-the-art deep learningbased methods [14,11,12] in Fig. 2e, where LigPose obtained the success rate of 82.5%, which is higher than the best-performed method DeepBSP (= 79.7%), indicating the high performance of the proposed end-to-end paradigm.…”

“…A.2.4 for details). The comparisons of LigPose affnity with several recent structure-based deep learning methods [37,39,35,36,41,38,40] Pearson R of 0.835. Compared to the methods that use native structures, LigPose affnity is also comparable or better than the best performing method PointTransformer (MAE: 0.91, RMSE: 1.19, Pearson R: 0.852), indicating its strong ability to learn native-like atom interactions.…”

“…However, popular docking tools generated the native-like poses with accuracies from approximately 40% to 60% in terms of success rate [10], which were far from satisfactory. Since deep learning has impacted the őeld of drug development, many researchers attempted to build hybrid methods by regarding it as a more expressive scoring function to rank the poses sampled by conventional docking tools [11,12,13,14,15,16]. Nevertheless, the performance of these hybrid methods has a constrained upper bound, as limited by the sampling time and space of docking tools [15].…”