DrugPred_RNA—A Tool for Structure-Based Druggability Predictions for RNA Binding Sites

Abstract

RNA is an emerging target for drug discovery. However, like for proteins, not all RNA binding sites are equally suited to be addressed with conventional drug-like ligands. To this end, we have developed the structure-based druggability predictor DrugPred_RNA to identify druggable RNA binding sites. Due to the paucity of annotated RNA binding sites, the predictor was trained on protein pockets, albeit using only descriptors that can be calculated for both RNA and protein binding sites. DrugPred_RNA performed well in discriminating druggable from less druggable binding sites for the protein set and delivered predictions for selected RNA binding sites that agreed with manual assignment. In addition, most drug-like ligands contained in an RNA test set were found in pockets predicted to be druggable, further adding confidence to the performance of DrugPred_RNA. The method is robust against conformational and sequence changes in the binding sites and can contribute to direct drug discovery efforts for RNA targets.

Translating from Proteins to Ribonucleic Acids for Ligand‐binding Site Detection

“…Distance‐based methods identify putative ligand‐binding sites by selecting regions with a high similarity to known binding sites, but are prone to suffer from a limited ability to generalize [16–19] . Machine‐learning models such as random forests can learn to recognize comparably complex structural features, but require the extraction of all potential pockets, for example, using geometric filters [13–15] . Deep learning enables direct detection of RNA binding sites.…”

“…For proteins, deep learning-based models have enabled the identification of druggable binding sites based on three-dimensional (3D) structural data. [8][9][10][11][12] For RNA molecules, the existing computational methods for binding site prediction can be categorized into machine-learning-based classifiers, [13][14][15] distance-based metrics, [16][17][18] knowledgebased similarity approaches, [19] and deep learning models. [20] Distance-based methods identify putative ligand-binding sites by selecting regions with a high similarity to known binding sites, but are prone to suffer from a limited ability to generalize.…”

“…[16][17][18][19] Machine-learning models such as random forests can learn to recognize comparably complex structural features, but require the extraction of all potential pockets, for example, using geometric filters. [13][14][15] Deep learning enables direct detection of RNA binding sites. However, limitations arise from the small amount of annotated training data containing the structures of both RNA and their ligands (approx.…”

HARIBOSS: a curated database of RNA-small molecules structures to aid rational drug design

“…DrugPred_RNA is, to the best of our knowledge, the only tool that has been trained and tuned using 3D structure data in order to characterize RNA binding sites [37]. In addition, most of current CADD pipelines have been developed for protein targets and might not be directly applicable to RNA.…”

HARIBOSS: a curated database of RNA-small molecules structures to aid rational drug design

“…Databases that collect all the known compounds binding RNA can be exploited for ligand- [32, 33] and 2D structure-based [34-36] virtual screening. DrugPred_RNA is, to the best of our knowledge, the only tool that has been trained and tuned using 3D structure data in order to characterize RNA binding sites [37]. In addition, most of current CADD pipelines have been developed for protein targets and might not be directly applicable to RNA.…”