Structural Interaction Fingerprints and Machine Learning for predicting and explaining binding of small molecule ligands to RNA

Abstract

Ribonucleic acids (RNA) play crucial roles in living organisms as they are involved in key processes necessary for proper cell functioning. Some RNA molecules, such as bacterial ribosomes and precursor messenger RNA, are targets of small molecule drugs, while others, e.g., bacterial riboswitches or viral RNA motifs are considered as potential therapeutic targets. Thus, the continuous discovery of new functional RNA increases the demand for developing compounds targeting them and for methods for analyzing RNA-small molecule interactions. We recently developed fingeRNAt - a software for detecting non-covalent bonds formed within complexes of nucleic acids with different types of ligands. The program detects several non-covalent interactions, such as hydrogen and halogen bonds, ionic, Pi, inorganic ion- and water-mediated, lipophilic interactions, and encodes them as computational-friendly Structural Interaction Fingerprint (SIFt). Here we present the application of SIFts accompanied by machine learning methods for binding prediction of small molecules to RNA targets. We show that SIFt-based models outperform the classic, general-purpose scoring functions in virtual screening. We discuss the aid offered by Explainable Artificial Intelligence in the analysis of the binding prediction models, elucidating the decision-making process, and deciphering molecular recognition processes.

RNAmigos2: Fast and accurate structure-based RNA virtual screening with semi-supervised graph learning and large-scale docking data

“…Widely used nucleic-acid-specific docking tools include rDock [46], Autodock VINA [58], and RLDOCK [53]. To enhance the docking procedure, alternative pose scoring functions have been proposed such as AnnapuRNA [51], LigandRNA [42], and very recently FingeRNAt [55], and RLaffinity [54]. These methods offer higher confidence screening results but come at the cost of high computational demands: the time needed to search over poses in a single binding site-ligand pair with a docking software is on the order of minutes.…”

RNA-ligand interaction scoring via data perturbation and augmentation modeling

“…Secondly, for RNA targets of interest with known structures, molecular docking remains the most straightforward virtual screening method, several docking and scoring methods have been developed for RNA-targeting ligands, such as rDock [26], RLDOCK [27], AutoDock Vina [28]. Despite the widespread use of molecular docking, its accuracy is limited due to factors such as force field settings, inaccuracies in scoring functions [29,30], and inadequate sampling of ligand conformations [31]. Thirdly, many studies have begun to utilize advanced deep learning models to study RNA-ligand interactions.…”

How Good Are Current Docking Programs at Nucleic Acid–Ligand Docking? A Comprehensive Evaluation

“…As the core of docking programs, a scoring function (SF) is used to rank the generated poses during docking calculations, which is a very important part of the docking process. As of now, only a limited amount of effort has been devoted to developing accurate SFs specific for NA–ligand interaction. ,,− Among them, SPA-LN, ITScore-NL, LigandRNA, and DrugScoreRNA are traditional knowledge-based SFs that derive the statistical potentials from the crystal structural information. AnnapuRNA, RNAPosers, and the method reported by Szulc et al are machine learning (ML)-based SFs for ligand binding pose prediction targeting NA.…”

“…Additionally, the supporting website for LigandRNA is currently invalid. On another note, the ML-based SFs, including AnnapuRNA, RNAPosers, and the work reported by Szulc et al, 27 are only specifically tailored for RNA targets. Additionally, assessing the performance of these ML-based SFs poses a challenge to the unified test set because different training sets were used by different ML-based SFs.…”