SiteRadar: Utilizing Graph Machine Learning for Precise Mapping of Protein–Ligand-Binding Sites

Abstract

Identifying ligand-binding sites on the protein surface is a crucial step in the structure-based drug design. Although multiple techniques have been proposed, including those using machine learning algorithms, the existing solutions do not provide significant advantages over nonmachine learning approaches and there is still a big room for improvement. The low ability to identify protein−ligand-binding sites makes available approaches inapplicable to automated drug design. Here, we present SiteRadar, a new algorithm for mapping cavities that are likely to bind a small-molecule ligand. SiteRadar shows higher accuracy in binding site identification compared with FPocket and PUResNet. SiteRadar demonstrates an ability to detect up to 74% of true ligand-binding sites according to the top N + 2 metric and usually covers approximately 80% of ligand atoms. Therefore, SiteRadar can be regarded as a promising solution for implementation into algorithms for automated drug design.

AF2BIND: Predicting ligand-binding sites using the pair representation of AlphaFold2

“…Other prediction approaches either rely on achieving the correct chemical compatibility of the ligand to identify a site (blind docking) [5] [9] [14] or are trained to predict voxels that carve out the space occupied by a known ligand. [12] [18] The myrid unliganded structures now available in AlphaFoldDB [36] and ESM Metagenomic Atlas [24], combined with the predictive power of AF2BIND, offer tantalizing opportunities to discover novel binding sites across the tree of life.…”

Comparative evaluation of methods for the prediction of protein-ligand binding sites

“…Combined approaches or meta-predictors combine multiple methods, or the use of multiple types of data to infer ligand binding sites, e.g., geometric features with sequence conservation [30][31][32][33][34][35][36]. Finally, machine learning methods utilise a wide range of machine learning techniques including random forest, as well as deep, graph, residual, or convolutional neural networks [37][38][39][40][41][42][43][44][45][46][47][48][49][50][51][52][53][54][55]. Machine learning methods comprise the bulk of the methods reviewed in this analysis and are exempli ed by P2Rank [37,39], DeepPocket [47], PUResNet [43,55], GrASP [52], IF-SitePred [53] and VN-EGNN [54].…”

Explainable Graph Neural Networks with Data Augmentation for Predicting pKa of C–H Acids

“…In recent years, graph neural networks (GNNs) have received increasing attention due to powerful learning capabilities for graphs and have been widely applied in predicting molecular properties. The molecular structure of a compound can be represented as a natural graph, where atoms correspond to nodes and chemical bonds correspond to edges. − This method can automatically learn the structural features and patterns of molecules by aggregating messages from neighboring atoms. In 2019, Roszak et al made the first attempt to predict the p K a values of C–H acids in DMSO using graph convolutional neural networks (GCNNs), achieving predictive performance comparable to that of QM models while being faster by several orders of magnitude .…”