RNA-ligand molecular docking: advances and challenges

Abstract

With rapid advances in computer algorithms and hardware, fast and accurate virtual screening has led to a drastic acceleration in selecting potent small molecules as drug candidates. Computational modeling of RNA-small molecule interactions has become an indispensable tool for RNA-targeted drug discovery. The current models for RNA-ligand binding have mainly focused on the docking-and-scoring method. Accurate docking and scoring should tackle four crucial problems: (1) conformational flexibility of ligand, (2) conformational flexibility of RNA, (3) efficient sampling of binding sites and binding poses, and (4) accurate scoring of different binding modes. Moreover, compared with the problem of protein-ligand docking, predicting ligand binding to RNA, a negatively charged polymer, is further complicated by additional effects such as metal ion effects. Thermodynamic models based on physics-based and knowledge-based scoring functions have shown highly encouraging success in predicting ligand binding poses and binding affinities. Recently, kinetic models for ligand binding have further suggested that including dissociation kinetics (residence time) in ligand docking would result in improved performance in estimating in vivo drug efficacy. More recently, the rise of deep-learning approaches has led to new tools for predicting RNA-small molecule binding. In this review, we present an overview of the recently developed computational methods for RNA-ligand docking and their advantages and disadvantages.

Graphical Abstract

RNA-targeted drug discovery requires the synergy of enhanced sampling and accurate scoring with fast computational speed. The distinct aspects of RNA-ligand docking compared to protein-ligand docking pose unique challenges, which demand a new generation of molecular docking models. This review presents an overview of recently developed RNA-ligand molecular docking methods for RNA-targeted drug discovery.

1. Introduction: targeting RNA with small molecules is a highly promising strategy for drug discovery

Ribonucleic acid (RNA) molecules are transcribed from DNA in the cell nucleus and play a variety of critical roles in gene expression and regulation at the level of transcription and translation. According to their cellular functions, RNA molecules can be categorized into two types: messenger (coding) RNAs (mRNAs) that encode the amino acid sequences and are translated into proteins, and noncoding RNAs (ncRNAs), which, instead of encoding amino acid sequence, serve as enzymatic, structural, and regulatory elements for gene expression. With the coding RNAs occupying only less than 3% of the human genome (1-3), the vast majority of the human genome sequences are transcribed to ncRNAs, such as ribosomal RNAs (rRNAs), microRNAs (miRNAs), small interfering RNAs (siRNAs), small nuclear RNAs (snRNAs), and various riboswitches (4; 5). With the ever increasing discoveries of new RNA structures and cellular functions and the continuous developments of powerful RNA structure determination methods, RNA-based therapeutics are becoming new promising methods to treat human disease. In general, RNA-based therapeutics can be classified into two types. In the first type, therapeutic RNAs—including RNA aptamers, antisense oligonucleotides (ASO), small interfering RNAs (siRNA), and guide RNAs (gRNA)—bind to the target (e.g., RNA transcripts, DNA targets, and protein targets) to inhibit or induce targeted biochemical reactions. This approach has attracted tremendous interest in the field of gene therapy and has been under very active development (6-9). In the second type of RNA-based therapeutics, an RNA molecule serves as the target for drug (small molecule) binding (2; 9-15). This second approach is analogous to protein-targeted drug discovery. However, only ~1.5% of the human genome encodes protein (2; 3; 13; 16-18), and of these protein-encoding genes, only 10-15% are disease-related (2; 3; 13; 19-21). The availability of druggable protein targets is further restricted by the structural and energetic fitness required for high-affinity drug binding. In contrast, genes that are undruggable or difficult to drug by targeting their associated proteins may be inhibited by drugs targeting the corresponding mRNA sequence. Therefore, compared to proteins, RNAs show much broader druggability. Additionally, non-coding RNAs play important roles in most human diseases from cancer to viral infection such as COVID-19. Targeting the large number of non-coding RNAs would open up remarkable new opportunities for drug discovery. For example, antibiotics targeting ribosomal RNA (rRNA), which forms an active site of a bacterial ribosome, effectively inhibit bacterial protein synthesis (22-25). Specific small molecules (ligands) bound to common riboswitches in bacterial cells regulate gene expression through ligand-induced conformational changes of the RNA (26-40). To inhibit viral replication, a potentially effective strategy is to use small molecule as a drug to target viral RNA motifs which are often highly structured (9; 11; 41), such as the HIV transactivation response (TAR) element in the 5’ untranslated region (42; 43), the internal ribosome entry site (IRES) element located in the hepatitis C virus (HCV) genome (44-48), and the influenza A virus RNA promoter (49; 50). Screening small molecule compounds selected from the compound library against an atypical three-stemmed RNA pseudoknot that stimulates −1 programmed ribosomal frameshifting (51; 52) in SARS-Cov RNA genome shows inhibition of the −1 ribosomal frameshifting of SARS-CoV with an IC₅₀ at 210 μM (53; 54). In addition to the above examples, many precursor messenger RNAs and microRNAs have also shown great promise as therapeutic targets (9; 14; 15).

Compared with predicting protein-ligand interactions, which remains a challenging problem, modeling binding interactions between RNA and small ligand molecules presents three unique challenges. First, unlike a protein, RNA is highly charged, with each phosphate group carrying one electronic charge. Thus, RNA folding and ligand binding require the participation of metal ions such as Mg²⁺ and water molecules to stabilize the binding pocket structure of the RNA and to mediate ligand-RNA interactions (55-58). Second, RNA molecules are often quite flexible, capable of folding into multiple stable conformations, and ligand binding often induces structural switches between different conformers or change the structure of an RNA receptor. Compared with protein-ligand binding, ligand binding sites on RNA can be less deep and more polar, solvated, and conformationally flexible (3; 11; 58), which adds further complexity to predicting RNA-small molecule interactions. Third, the fact, that we have a limited number of experimentally determined structures for RNA molecules and RNA-ligand complexes makes pure knowledge-based approaches less effective for RNA-ligand predictions. In this regard, a physics-based approach or a hybrid knowledge-based and physics-based approach can yield unique advantages (57; 59-71).

Although successful computational tools have been developed for protein-ligand binding (72-75), the difference in chemical structure and energetics between RNA and proteins demands new methods for RNA-ligand interactions. These new methods for small molecule selection, shown in Figure 1; are necessary for virtual screening, binding mode identification, and binding affinity prediction of specific RNA targets. A successful computational drug discovery requires the integration of three key components: (i) a method to identify the druggable RNA target, (ii) a computationally efficient, sampling algorithm for RNA conformations, ligand conformers, and ligand binding poses, (iii) accurate scoring functions to assess the RNA-ligand complex structures and evaluate the binding affinity. In this review, we focus on computational challenges in predicting RNA-ligand interactions, with specific emphasis on recent advances.

Figure 1:

Three major applications of an RNA-ligand interaction model. Virtual screening involves docking against, small molecules in a large library and scoring every docked pose. Top-scored selections are treated as the most, promising candidates for putative binders. For a given RNA-ligand pair, computational models for ligand binding pose identification and RNA-ligand binding affinity prediction rely on scoring the possible RNA-ligand complex structures. An ideal scoring function for ligand binding pose identification should have the ability to distinguish the native pose from a large pool of docked decoy poses, while achieving the maximum correlation between the predicted scores and the experimental affinities for different. RNA-ligand pairs.

2. Methods for identifying druggable RNA targets and binding sites

2.1. Identifying druggable target RNAs

The druggability of a particular RNA target depends on the answers to three questions. First, does the inhibition/enhancement of the target RNA function lead to effective control of the disease? Second, is the RNA target accessible for the small molecule binding in the cellular environment? Multiple factors can affect the accessibility of the target RNA, such as the abundance and lifetime of the target RNA in disease-related cellular processes (2; 9; 13). Third, does the target RNA adopt a binding site that enables small molecule binding with high affinity and high specificity? Small molecule targeting the particular RNA with high specificity can reduce the off-target side effects. An effective way to achieve high specificity is to target RNA that is unique in the diseased cells, pathogenic viruses or bacteria, such as riboswitches which are common in bacteria but rare in eukaryotes. Another way is to computationally identify RNA motifs that is able to form unique and high-affinity pockets capable of small molecule binding (2; 13; 76).

Inforna (77; 78) is a template-based method capable of selecting RNA targets according to RNA secondary structure motifs such as hairpins, symmetric and asymmetric internal loops, and bulges. The current Inforna 2.0 template database (77) contains 1936 pairs of known RNA secondary structure motif-ligand bound complexes (78). For a given RNA target, Inforna 2.0 (78) identifies RNA secondary structure motifs and from the template database, for a given motif, finds the corresponding ligand partners with fitness scores (79; 80). The fitness score (79; 80) provides a measure of RNA-ligand binding affinity as well as the selectivity of the RNA motif against many other small molecules (78). RNAs of high selectivity and affinity fitness scores are more likely to be druggable. With the top scored small molecules as lead compounds, chemical similarity screening of compound library gives potential potent binders. Inforna (77; 78) has been proved to be successful in various studies (13), such as identifying small molecules that target oncogenic miRNA precursors (81; 82) and an A bulge in the (iron responsive element) IRE (83) of the SNCA mRNA related to Parkinson’s disease (84).

On the basis of RNA secondary structures, Warner et al. (2) showed that information content (85; 86) can be used to identify druggable RNA motifs for potentially high-specificity and high-potency binding (87). Information content measures the amount of information (in bits) required to specify the sequence and structure complexity of an RNA motif, where motifs with high bits (~30 bits) are more complex and more likely to be unique in the transcriptome (2). In experiments, RNA structural information content is attainable through chemical probing techniques such as selective 2′-hydroxyl acylation analyzed by primer extension (SHAPE) (88-91). Focusing on RNA motifs with sufficient complexity (high information content) can lead us to RNAs with high binding specificity and affinity thus higher druggability. As an example, the binding affinities of GTP (86; 92) and targaprimir-96 (93) both show a strong correlation with the information content of the RNA motifs, where a 10-bit increase in information content results in a 10-fold increase in binding affinity (2).

2.2. Identifying binding sites for a given target RNA

An overall assessment of binding pockets.

A recent statistical analysis demonstrates that many RNAs indeed fold into structures that form pockets amenable to selective small molecule binding (76). To identify potential RNA suitable for ligand binding, Hewitt et al. (76) have evaluated RNA binding pockets using PocketFinder (87) for 1552 structured RNAs and all the proteins in the Protein Data Bank (PDB) (94), where a binding pocket is described by the volume and the solvent exposure of the pocket (buriedness) and the fraction of the pocket considered to be hydrophobic (hydrophobicity). The results suggest that although ligand-bound pockets on RNAs and proteins show overall similar physical properties, RNA pockets are on average less hydrophobic than their protein counterparts (76). Moreover, compared to the unbound pockets of RNA, the ligand-bound pockets are generally larger in volume, more buried, and more hydrophobic (76).

Geometry-based methods.

In search for binding sites based on RNA-ligand shape complementarity, DOCK 6 (65) selects the binding pockets from a negative image of the receptor surface, where each cavity is characterized by a set of overlapping spheres (95). Similarly, rDock (67) applies a two-sphere mapping algorithm to identify the binding sites. Within the defined docking space, large spherical probes are placed on each grid point to rule out superficial and shallow sites. Then, small spherical probes are placed on the remaining unallocated grid points to map the cavities that serve as the possible binding sites (67). Wide and shallow minor grooves of RNA, which can geometrically accommodate a wide range of ligand shapes and serve as non-specific binding sites, are often identified as putative binding pockets and cause false-positive predictions.

Energy-based methods.

Other programs find binding sites by estimating the overall probe-pocket interaction energies, where the probes are virtual atoms and traverse the surface of the receptor. PocketFinder (87) and AutoLigand (96) are such cases and are equipped in ICM (97) and AutoDock (98), respectively. PocketFinder uses a Lennard-Jones (LJ) potential to describe the interactions between probe atoms and receptor atoms, and grid maps generated from the calculated interactions are used to identify the binding sites. AutoLigand uses a similar approach but involves an extra iterative step to identify the optimal binding sites from the grid maps, and it accounts for connection between the neighboring possible pockets.

Network and machine-learning approaches.

By treating RNA-ligand interaction as a network of contacting atoms, network-based approaches have shown great promise in the prediction of the functional sites in RNA-ligand interactions. For example, using inter-nucleotide Euclidean (hamming) distance network for a 3D or 2D structure Rsite (99) and Rsite2 (100) predict the functional sites for RNA-ligand binding from the maximally closely clustered nucleotides. However, since the inter-nucleotide networks in Rsite and Rsite2 do not distinguish the different connection types between the nucleotides, these models often lead to false positive predictions. To distinguish the different connection types, RBind (101; 102) transforms an RNA structure into a graph, where a node and a edge denote a nucleotide and a non-covalent contact between the nucleotides, respectively, and predict the functional sites as regions formed by nucleotides of the maximum closeness. On a test set with 19 RNA-ligand complexes, RBind (average positive predictive value PPV = 0.67) outperforms Rsite (average PPV = 0.42) and Rsite2 (average PPV = 0.40) (101). The result suggests that the different types of inter-nucleotide interactions encoded in the RNA structure provide important information for the prediction of the functional sites. RNAsite (103), a random forest-based model, uses sequence-based and/or structure-based descriptors to predict whether a given nucleotide belongs to the functional sites. In the model, a nucleotide is defined as part of a functional site if it contains an atom within 4 Å to the target ligand. In RNAsite, four different sets of features for each nucleotide are extracted: geometrical features of local surface convexity/concavity (Laplacian norm), topological features of the RNA nucleotide interaction network similar to the one used in RBind (101; 102), nucleotide-specific accessible surface areas, and position-specific evolutionary conservation of the nucleotide calculated from multiple sequence alignment. The model is trained on 60 RNAs with five-fold cross validation and tested on two separate sets with 19 (RB19) and 18 (TE18) RNAs, respectively. By using Mathews correlation coefficient (MCC) and area under the receiver operating characteristic curve (AUC), RNAsite (103) shows better performance compared to Rsite (99), Rsite2 (100) and RBind (101; 102), with 0.253 and 0.567 for MCC, and 0.776 and 0.877 for AUC on TE18 and RB19 sets, respectively. The promising results indicate the necessity to include more independent features as descriptors. Although these models have been trained specifically for RNA-small molecule complexes, further improvements are possible, for example, through a combination with machine learning methods (104-107).

3. Methods for efficient sampling of ligand binding modes with flexible conformations — a major challenge in RNA-ligand docking

Exhaustive sampling of possible RNA-ligand complex structures is challenging due to the flexible nature of RNA and small molecules. Additionally, following the induced-fit effect or the conformational selection mechanism, RNA targets often undergo conformational changes in response to ligand binding (40; 108-110) (see Figure 2a). This leads to the coupling between RNA folding, including cotranscriptional folding, and ligand binding when virtual screening is performed (40; 108-111). A widely used approach to tackle this problem is ensemble docking (42; 43; 55; 112), where a ligand docks to an ensemble of RNA structures. An alternative approach is to sample the conformational changes on the fly in the docking process. Various methods (60; 64-67; 70) have been developed to treat flexible docking. However, in part due to the required computational time for large-scale virtual screening, predicting large conformational changes remains a challenge.

Figure 2:

RNA conformational changes and binding interactions mediated by water molecules and ions. (a) The local structure difference of preQ1 riboswitches between apo (ligand-free) and holo (ligand-bound) states. The structure in orange denotes the apo state (PDB code: 6VUH (113)) and the structure in blue denotes the holo state (PDB code: 3Q50 (114)) with its bound small molecule (PRF) colored in magenta. Upon binding, the small molecule displaces residue A14 (colored in green for both apo and holo states) and causes the local structural transition. (b) Water molecules mediated RNA-ligand interactions. Water molecules form a bridge between small molecule Neomycin B (NEM, magenta) and 16S-rRNA A-site (PDB code: 2ET4 (115)). The isolated red dots denote the oxygen atoms in water molecules. The black dashed lines show the water-mediated hydrogen bonding contacts that promote NEM binding to the RNA receptor. (c) Metal ions in RNA-small molecule interactions. The ligand benfotiamine (BTP, magenta) interacts with residues G60, C77, and G78 of the Thi-box riboswitch through two magnesium ions (green) and the G42-A43 base stack (PDB code: 2HOO (116)). The black solid lines represent the inner sphere metal ion coordination. The polyanionic RNA recognizes the positively charged metal ion complex made up of the monophosphorylated compound and cations.

For ligand docking to RNA targets, such as ribosomal RNAs and riboswitches with reliable binding site information (117; 118), local sampling with a rigorous energy scoring function can often provide accurate predictions for the ligand binding pose. However, for a broad range of therapeutic RNAs including viral genomic RNAs (119; 120), the lack of the binding site information poses a great challenge to drug screening. For RNAs, unlike proteins, we have limited examples of RNA-ligand bound structures and scarce knowledge about the binding sites. This fact highlights the importance of blind docking (vs. local docking), where a small molecule is docked to the entire surface of the receptor without any prior knowledge of the binding site (see Figure 3). Although computational models—including AutoDock Vina (66), GOLD (59), and Glide (61)—or models originally developed for protein-ligand docking, but optimized for RNA targets—such as AutoDock (98), ICM (97), DOCK 6 (65), and FITTED (62)—can be adopted for RNA-ligand docking, methods developed specifically for RNA targets, such as RiboDock (60), rDock (67), MORDOR (64), and RLDOCK (70; 71), have demonstrated advantages for RNA systems. See Table 1 for a list of docking software.

Figure 3:

The difference between local and blind docking. A complex of an aminoglycoside antibiotic, gentamicin (green) and the 16S-rRNA A site of bacterial ribosome is used for illustration (PDB code: 2ET3 (115)). In this example, both docking (local & blind) processes are carried out using the RLDOCK model (70; 71). In local docking, the binding pocket is predefined and the sampling is contained within the red dashed box. The small magenta spheres denote candidate binding sites predicted by RLDOCK. In blind docking, the binding site detection is performed across the whole surface of the RNA. The small yellow and magenta spheres denote the predicted high- and low-probability binding sites, respectively. Two cavities identified by RLDOCK (anchored by yellow spheres) are zoomed out separately.

Table 1:

Docking programs available for RNA. Docking programs without dedicated binding site detection module are shown with a dash.

Docking program	Target	Conformational search algorithm	Binding site prediction
AutoDock Vina (66)	protein	Monte Carlo & quasi-Newton	-
GOLD (59)	protein	genetic algorithm	-
Glide (61)	protein	Monte Carlo	SiteMap (121)
AutoDock (98)	protein/RNA	genetic algorithm	AutoLigand (96)
ICM (97)	protein/RNA	Monte Carlo	PocketFinder (87)
DOCK 6 (65)	protein/RNA	incremental construction	sphgen module (95; 122)
FITTED (62)	protein/RNA	genetic algorithm	-
RiboDock (60)	RNA	genetic algorithm	two-sphere filter
rDock (67)	protein/RNA	genetic algorithm Monte Carlo simplex minimization	two-sphere filter
MORDOR (64)	RNA	molecular dynamics	grid-based systematic search
RLDOCK (70; 71)	RNA	multi-conformer docking	grid-based systematic search

3.1. Sampling of ligand conformations

There are three general ways to treat flexible ligand conformations in docking (123-125): multi-conformer docking, incremental construction, and stochastic optimization. These three strategies differ in their computational speed and the conformational adaptability of the docked ligand to the geometric features of the binding-site.

Multi-conformer docking.

The multi-conformer docking algorithm prepares a conformational ensemble for a ligand (small molecule) and performs rigid docking for each the ligand conformers against the same target (126). For a given binding pocket, this method can be computationally fast if a limited number of conformers are docked. A key determinant for the success of this approach is that the near native conformations of the ligand must be included in the ligand conformer ensemble. Currently, there exist a number of ligand conformer generators, such as OMEGA (127; 128), RDKit (129), and Open Babel (130). These models have been shown in benchmark studies to reproduce reliable conformational ensembles of small molecules within seconds (130; 131). Combining a new molecular dynamics approach and a quantum mechanically-refined ligand-RNA interaction force field, a recently released conformer generator has led to improved accuracy for in silico drug design (132; 133). Its web-based server and the database of bioactive conformational ensembles not only speed up the process of finding experimentally favorable ligand conformations through massive docking but also provide proper initial structures for further optimization (134). In addition, in the docking process, a ligand conformer ensemble is constructed prior to conformational sampling, thus, the conformer ensemble can be appropriately built for the small molecules in question. For example, a ligand conformer ensemble can be generated with bias toward the low-energy states (70; 71) or with maximum diversity in the conformational space.

RNA-Ligand DOCKing (RLDOCK) (70; 71) is a recently developed docking model for flexible ligands using a multi-conformer approach. In RLDOCK, the ligand-binding mode is described by four variables (R, L, A, O), where L denotes the ligand conformer, A denotes the ligand atom placed at (anchor site) R, and O is the 3D rotation angle of L about A (at position R). For each RNA-ligand pair, the RLDOCK algorithm generates an ensemble of flexible ligand conformers and binding poses through the following steps.

1. The algorithm generates an ensemble of viable anchor sites R based on the following two criteria: (a) there should be no steric clash between ligand and RNA atoms and (b) the RNA structural environment around R should form a pocket geometry.

2. Based on the viable anchor sites R generated above, by exhaustively enumerating all the different combinations of (R, L, A, O), RLDOCK samples all the possible ligand binding sites and binding poses. The results are stored for subsequent refinement. Before applying the scoring function to rank order all the binding modes, to accelerate computational speed, RLDOCK first sieves the exhaustive ensemble by removing those with high LJ potential between RNA and ligand.

3. All R sites with low LJ potentials U_LJ(R) (below the threshold) are selected as preferred R sites. Here U_LJ(R) is the minimum LJ potential over all possible (L, A, O) values for a given R.

4. For each preferred R, preferred ligand conformers L are selected from low LJ potentials U_LJ(R, L), the minimum LJ potential over all possible (A, O) values for a given set (R, L).

5. Similarly, for each preferred R and L, preferred ligand atoms A are selected from the low LJ potentials U_LJ(R, L, A), the minimum LJ potential over all possible O values for a given set (R, L, A).

After the above procedure, a preferred (R, L, A, O) ensemble with all the possible orientations (O) of the ligand is generated and subsequently ranked by the scoring function. To speed up the LJ potential calculation, RLDOCK employs a grid-based energy calculation, where each grid stores the LJ energy between RNA and a probe atom on the grid for fast computation of LJ energy for a given binding mode. Through the above procedure, RLDOCK generates millions of possible ligand configurations through exhaustive rotation and translation transformations at each putative binding site for each pre-configured conformer (see Figure 4). Compared with other models, RLDOCK has the unique merit of using complete sampling for ligand conformers and binding modes.

Figure 4:

Illustration of conformational sampling methods used in RLDOCK, using the docking of 2’-deoxyguanosine to 2’-deoxyguanosine riboswitch (PDB code: 3SKL(135)) as an example. An ensemble of different conformers of the 2’-deoxyguanosine (dG) is constructed for flexible docking. The sampling and scoring procedures are shown in order and labeled through A to E. (A) First, the regions of possible anchor sites within the riboswitch, colored in magenta, are determined by the geometric features of the target RNA. (B) Second, with exhaustive sampling of these prepared conformers through translation and rotation around the anchor sites, (C) binding sites (yellow dots) are selected according to Lennard-Jones potential between RNA and ligand atoms. (D) Finally, the sampled ligand conformations associated with the selected binding sites are ranked (E) by a physics-based scoring function.

Incremental construction.

By anchoring rigid fragments through geometric matching and then incrementally building the ligand structure, on-the-fly ligand conformer sampling allows the local environment of the binding pocket to guide the growth of the small molecule. An inherent drawback of this approach is that small errors in the early steps can be amplified throughout the process, especially for large ligands. DOCK 6 (65) adopts this incremental construction strategy for ligand conformational sampling and search (53). Unlike the original greedy algorithm (cluster-based pruning) to sieve the sampled ligand structures followed by clustering and ranking at each step, an improved algorithm, which skips the conformational clustering step in order to retain the original, diverse conformations of the flexible bonds, has led to an increase in the success rate by 10% for the prediction of binding poses (65). The results have demonstrated the importance of maintaining the diversity of ligand conformers.

Stochastic optimization.

The stochastic optimization method searches for binding modes on-the-fly by optimizing flexible torsional angles, orientation, and position of the small molecule. The Monte Carlo (MC) (61; 66) and Genetic algorithms (GA) are the most widely used stochastic optimization algorithms. A combination of different stochastic methods often lead to improved sampling and optimization results. For example, AutoDock Vina (66) adopts a hybrid approach with MC for global optimization and Broyden-Fletcher-Goldfarb-Shanno (BFGS) for local optimization (66). Other examples include ICM (97) and RiboDock (60), which employ MC coupled with simulated annealing. Several protein docking programs, such as GOLD (59), AutoDock (98), and FITTED (62), use GA to sample and search for ligand conformations. Modifications to some of these methods for RNA targets have led to highly promising results. For example, through parameterizing the scoring function (136) or adding a new solvation term to the original scoring function (137), AutoDock can treat RNA-ligand interactions. Similarly, through proper optimization (58), FITTED can be used to predict RNA-ligand docking. The accuracy may be further improved once all possible RNA hydrogen bond donors and acceptors are considered, and after metal ions such as Mg²⁺ and Mn²⁺ are included as part of the receptor (58). Like other stochastic optimization algorithms, a shortcoming of MC and GA is that the optimization process may become trapped in the local minima. This may pose a severe challenge due to the rugged energy landscape of RNA-ligand complexes. The problem can be alleviated through repetitive docking with random placement of the small molecule (ligand) and implementation of algorithms such as tabu search (138; 139) and stochastic tunneling (140) to accelerate the de-trapping from the local minima. By minimizing the likelihood of poses being trapped in a local minimum in the early stages of the conformational search, rDock (67), a model for both nucleic acid and protein docking, employs a GA/MC hybrid method to enhance efficient sampling of ligand binding poses. The GA/MC hybrid method involves three rounds of GA, low-temperature MC, and Simplex minimization, each of which adopts an independent scoring function. An optimized set of scoring functions has been shown to significantly enhance the efficiency of sampling even with an unfavorable initial pose by minimizing the possibility of being trapped in a local minimum during the conformational search.

In summary, while multi-conformer docking provides a fast way to consider ligand flexibility prior to the docking calculation, its performance depends on the quality of the generated conformer ensemble. In contrast, stochastic and incremental sampling approaches can treat ligand flexibility during the docking process. However, such on-the-fly sampling approaches suffer from the problem that a small error in the early steps can be amplified in the later steps, and stochastic approaches suffer from the problem that poses can be potentially trapped in a local minima while docking. In addition, both approaches require additional energy terms to account for intra-ligand interactions. Although the conformational sampling modules in RNA-ligand docking software (57; 67; 70; 71; 141-143) have shown promising results for recovering native or near-native ligand poses, for a flexible RNA that undergoes conformational changes upon ligand binding, a search for a fast and accurate sampling method by combining folding and binding algorithms for both ligand and RNA continues.

3.2. Incorporation of RNA flexibility

It has been shown that for protein-ligand docking, ignoring the protein (receptor) flexibility can cause incorrectly predicted binding modes (144). The problem can be more severe for ligand binding to an RNA, whose structure can be more flexible than a protein. To address this important issue in RNA-ligand docking, several successful approaches have been developed to incorporate RNA conformational changes in the docking algorithm. These approaches can be classified into three types: soft docking, ensemble docking, and fully flexible docking (See Figure 5).

Figure 5:

Different approaches to modeling RNA flexibility in RNA-ligand interactions illustrated using HIV-TAR RNA (PDB: 1ANR (145)) as an example. The orange and blue regions correspond to rigid and flexible portions of RNA, respectively. From left to right, a) bases from the active site are allowed to partially overlap with atoms from ligand through soft potential, b) an ensemble of various RNA conformations is used to perform docking, and c) RNA with full flexibility. Computational efficiency decreases from left to the right.

Soft docking.

A mathematically convenient way for soft-docking is to decrease the energy penalties for steric clashes, thus tolerating some degrees of overlap between RNA and ligand (146). Glide (61) and GOLD (59) offer such options for users. In earlier work, Moitessier et al. (55) have employed this strategy to dock aminoglycosides in ribosomal A-site RNA, where RNA flexibility was considered using a set of soft van der Waals potentials, and the approach led to increased average accuracy. Soft docking is attractive for its convenience of implementation. However, the limited sampling space without the adjustment of backbone prevents its application to large conformational changes.

Ensemble docking.

In ensemble docking, a given ligand docks into an ensemble of RNA conformations or an ensemble-averaged RNA conformation. Ensemble docking is found to be useful in several RNA-targeted studies (42; 43; 55), and has also been proved successful in protein-small molecule docking (147; 148). In an attempt to reproduce the experimentally determined RNA-aminoglycoside complexes, soft docking to an ensemble-averaged RNA structure gives the best performance with an average RMSD of 2.49Å between the predicted and the experimentally measured binding mode (55). Ensemble docking-based virtual screening with the ICM docking model (42) for HIV-TAR (42; 43) has predicted a TAR-targeting compound with high specificity. Furthermore, virtual screening for an experimentally derived TAR conformational ensemble against a ligand library composed of ~100,000 drug-like organic compounds (43) provided an enriched family of TAR-targeting binders. From a practical perspective, the number of receptor conformations used in the ensemble docking is usually limited due to computational feasibility, thus receptor conformation selection can influence the accuracy of the prediction. Using only the conformational ensemble in the lowest-energy basin may not be the optimal strategy as ligand binding can stabilize and selectively enrich the population of conformations in other basins on the energy landscape (149-151). Therefore, ensemble docking should not ignore low-populated ligand-free RNA conformations.

Molecular dynamics.

Molecular dynamics (MD) simulation (152-159) can not only refine conformations of RNA-ligand complexes and generate ligand and RNA structures, but also shed light on the trajectory and the folding/unfolding of possible metastable states for both RNA and ligand in the docking process (43; 46; 53). However, in practice, ligand binding events can occur in the timescale up to seconds (108; 110; 160), and an all-atom MD simulation for the process goes beyond the capacity of available computing power, especially when virtual screening for drug molecules is considered. Powered by advanced sampling techniques, several computational methods have enabled the characterization of RNA conformational changes upon ligand binding (161-165). A non-equilibrium MD simulation (166) and an umbrella-sampling-based MD simulation (167) both have revealed the competitive relationship between the formation of the kissing-loop and the binding of the small molecule. A recent explicit-solvent MD simulation has shown a small molecule-induced stabilization effect in an adenine riboswitch and the ability of the riboswitch in the near-native states to attract small molecules through hydrogen bonding and base-stacking interaction (168). These results demonstrated the unique advantage of MD simulations for the investigation of physical mechanisms in RNA-ligand binding.

Fully flexible RNA.

Molecular dynamics simulation with proper force field can provide reliable sampling of RNA-ligand complex conformations. For example, by applying an RMSD penalty term to the conventional potential energy, MORDOR (64) (MOlecular Recognition with a Driven dynamics OptimizeR), by simulating ligand docking trajectories, can give conformational sampling and show ligand-induced conformational changes for RNA (64). As an application, a MORDOR-based virtual screening has found a small family of binders targeting human telomerase RNAs (hTR) (169). However, further applications of MORDOR are limited by the high computational cost, which can take up to hours for a docking run. To accelerate simulation, Supervised Molecular Dynamics (SuMD) has been proposed to sample the conformations of RNA-drug complexes (170). SuMD accelerates the simulation by applying a tabu-like algorithm to guide the docking when the ligand is far away from the binding site and a conventional MD simulation when the ligand is close to the binding site. The hybrid method enables efficient simulation of the binding process within an affordable timescale. Although the simulated trajectory does not necessarily represent the physical binding process, SuMD may capture possible conformational changes. The reliability of SuMD for RNA-ligand docking is supported by success in predicting binding modes for several pharmaceutically important RNAs (170), where SuMD predicts RNA-ligand docking mode with a minimum RMSD of 0.34Å for the best case.

Similarly, another method based on elastic potential grids was initially proposed for modeling protein flexibility during docking (171) and later extended to RNA targets (112). In this type of method, a 3D grid of the potential field of the initial RNA conformation is calculated in advance using DrugScore^RNA (172). After determining the potential grids, AutoDock (98) is used as a docking engine with precalculated elastic potential grids for docking. Due to its ability to account for RNA flexibility, docking to the deformable potential grids generated from unbound RNA has a much better performance than simply docking to unbound RNA alone. However, one of the limitations of the approach is that it requires a priori knowledge of the available end states of deformation. Moreover, the model cannot treat conformational changes caused by rotational flip motion and 2D structural rearrangements.

In summary, Molecular dynamics-based methods are time-consuming and thus not suitable for large-scale virtual screening. Rigid docking is fast but lacks accuracy. Soft docking and ensemble docking are in the middle ground between the two extremes as they sacrifice the ability of a more complete sampling of conformations in order to reduce the computational time. At the current stage, a versatile approach to accurately treat receptor flexibility awaits to be developed.

4. Accurate scoring functions for RNA-ligand docking: challenges and promises

Selecting a native ligand binding pose from an ensemble of candidates requires a reliable scoring function. There are three different approaches to the development of a scoring function: physics-based approach, knowledge-based approach, and machine-learning approach; see Table 2 and Table 3 for a list of reviewed scoring functions and the summary of the benchmark results, respectively.

Table 2:

Summary of the reviewed scoring functions used in different models for predicting small molecule binding. ^a Some scoring functions optimized for protein may also be used for RNA. ^b Some models contain more than one scoring function, only the default one is listed. ^c The year that the original model was first published.

Category	Model	Targeta	Score typeb	Yearc
Physics-based	MORDOR (64)	RNA	force fields	2008
	DOCK 6 (65)	RNA	force fields	2009
	GOLD (59)	protein	empirical terms	1997
	Glide (61)	protein	empirical terms	2004
	RiboDock (60)	RNA	empirical terms	2004
	AutoDock 4 (63)	protein	empirical terms	2007
	AutoDock Vina (66)	protein	empirical terms	2010
	iMDLScore1 (57) iMDLScore2 (57)	RNA	empirical terms	2012
	rDock (67)	protein nucleic acid	empirical terms	2014
	RLDOCK (70; 71)	RNA	empirical terms	2020
Knowledge-based	DrugScoreRNA (112; 173)	RNA	statistical potentials	2000
	KScore (174)	protein nucleic acid	statistical potentials	2008
	LigandRNA (175)	RNA	statistical potentials	2013
	SPA-LN (176)	nucleic acid	iterative statistical potentials	2017
	ITScore-NL (142)	nucleic acid	iterative statistical potentials	2020
Machine-learning	T-Bind (177)	protein	gradient boosting trees	2018
	RNAPosers (141)	RNA	random forest	2020
	RNAmigos (178)	RNA	graph neural network	2020

Table 3:

Summary of the benchmark results reported in the literature. Results of the different methods tested on the same test set are grouped together for comparison. The first column “Test set” shows the number of test cases and the original references (shown in parentheses) reporting the test results. ^a Performance of affinity prediction is reported in terms of Pearson correlation coefficient, R². Correlation coefficient is calculated between experimental binding affinities and predicted binding affinities. Benchmarks without affinity prediction are shown as dashes. ^b Performance of pose identification is reported as success rate. The criteria for a correct prediction is shown in the (rank, RMSD) format. For example, (1, 2.5Å) means the top-1 prediction that has RMSD less than 2.5 Å to the native pose. Benchmarks without pose identification are shown as dashes. ^c RLDOCK, rDock, rDock_solv, AutoDock Vina use 38 instead of 42 complexes. MORDOR uses 32 instead of 42 complexes. ^d Only several top performing models evaluated in literature (57) are listed for this benchmark dataset. ^e Three outliers 3GX3, 2ESI and 1F1T are excluded in the binding affinity calculation. ^f Near native poses are sampled through rDock reference ligand method (67) and ^g Average and standard deviation from 100 sets of 100 random docking poses out of a pool of 1000 decoy conformations. ^h Native pose is included in pose identification. are included in pose identification. ⁱ RNA-adapted AutoDock scoring function (137) is used. ^j Scoring function is used to guide the docking instead of using default Vina scoring function.

Test set	Scoring function	Docking engine	Affinitya prediction(R2)	Poseb identification(%)
42 (70; 71; 142; 175) complexes			Correlation	(1, 2.0Å)	(3, 2.0Å)
	RLDOCK	RLDOCK	-	55.3c	60.5c
	ITScore-NL	DOCK6	-	50.0	54.7
	LigandRNA+DOCK6	DOCK6	-	47.6	54.8
	rDock_solv	rDock 2014	-	39.5c	55.3c
	DOCK6	DOCK6	-	35.7	45.2
	LigandRNA	DOCK6	-	35.7	42.9
	AutoDock Vina	AutoDock Vina	-	31.6c	44.7c
	DrugScoreRNA	DOCK6	-	31.0	42.9
	rDock	rDock 2014	-	28.9c	47.4c
	MORDOR	MORDOR	-	-	62.5c
34 (57; 142; 176) complexesd			Correlatione	(3, 1.5Å)	(5, 3.0Å
	SPA-LN	rDock 2014	0.36	50.6	76.6
	Gold Fitness	GOLD5.0.1	0.25	42.9	73.2
	ASP	GOLD5.0.1	0.29	42.9	66.1
	rDock_solv	rDock 2006.2	0.18	41.1	73.2
	rDock	rDock 2006.2	0.15	33.9	60.7
	ITScore-NL	-	0.41	-	-
56 (67; 141; 176) complexes			Correlation	-	(1, 2.5Å)
	RNAPosers	rDock 2014	-	-	62.5f
	rDock_solv	rDock 2014	-	-	54±3g
	SPA-LN	rDock 2014	-	-	54±3g,h
	AutoDock Vina	AutoDock Vina	-	-	29±2g
	GlideScore	Glide (v.57111)	-	-	17.8
31 (137; 141; 173)			Correlation	(1, 2.0Å)	(1, 2.5Å)
	RNAPosers	rDock 2014	-	57.1f	61.9f
	DrugScoreRNA	AutoDock 3.0.5j	-	41.9	45.2
	AutoDocki	AutoDock 3.0.5j	-	25.8	35.5
77 (142; 176; 235) complexes			Correlation	(3, 1.5Å)	(5, 3.0Å)
	ITScore-NL	AutoDock 4.2	0.41	71.4h	90.9h
	SPA-LN	rDock 2014	0.33	50.6h	76.6h

4.1. Physics-based methods: physical principles of RNA-ligand binding lead to accurate scoring functions

Force-field approach.

Atom-based physical force fields, originally derived from thermodynamic data and ab initio calculations, have enabled molecular dynamics simulations for nucleic acids-targeted drug discovery (179-183). One of the key issues in physical force field-based computations for molecular docking is the solvent effect. Since the virtual screening of ligands against an RNA target demands high computational efficiency for the docking calculation, implicit solvent models, such as Poisson-Boltzmann surface area (PB/SA) model (184-190) and the Generalized-Born surface area (GB/SA) model (191-199), would be highly promising due to the optimal balance between speed and accuracy. A hybrid force field that combines an implicit solvent model and an all-atom force field can often lead to accurate and efficient simulation of an RNA-ligand binding process. As shown by the success of DOCK 6 (65), generalized Born and Poisson Boltzmann implicit solvent models combined with the AMBER force fields can provide an effective energy model for an RNA-ligand docking system (65). MORDOR (64), which combines an implicit solvent model GBSW (Generalized Born with Simple sWitching) (200) with the CHARMM-27 (201) force fields for the receptor and AMBER force fields (202) for ligand molecules, demonstrated how the hybrid energy function can lead to successful modeling of receptor-ligand binding. By using root-mean-square-distance constraints in energy minimization, the model allows local flexibility of the receptor to accommodate possible conformational changes induced by ligand binding and in the meantime, to guide the ligand to probe the surface of the target RNA.

In summary, physical force field-based scoring functions have the advantage of providing insights into the underlying physical mechanism of RNA-small molecule interaction, however, computational costs and the need for expert knowledge in simulating a specific system hinders the application of these models in large-scale virtual screening for drug discovery.

Empirical energy approach.

Physically, different interactions in an RNA-ligand complex are correlated thus nonadditive. A simplified approach is to evaluate the total energy as a weighted sum of the component interactions such as van der Waals, electrostatic, desolvation and hydrogen-bond interactions:

$$\Delta G=\sum _i{w}_i\cdot \Delta G_i$$ (1)

where the weight coefficients w_i can be fitted by optimizing the success rate of the computational prediction for the training set. The above empirical scoring function has the advantage of high computational efficiency and adaptivity, which makes accurate prediction possible for specific types of RNA targets and ligands. Compared to the more rigorous force-field approach, empirical scoring functions, which often use “softer” energy forms, are more tolerant for minor clashes and suboptimal interactions during docking, thus partially alleviate the problem of incomplete sampling for receptors and ligand conformations. It is important to note that due to the different physical interactions and correlations between the different interactions in protein-ligand and RNA-ligand systems, parameters fitted from protein-ligand docking may not be transferable to RNA-ligand docking.

The semi-empirical free energy function in AutoDock 4 (63), the fully empirical scoring function of AutoDock Vina (66), and other models such as GoldScore in GOLD (59) and GlideScore in Glide (61) have demonstrated success in predicting protein-ligand docking. These docking software packages, not specifically designed for nucleic acids, may not give optimal results for RNA-ligand docking. RNA-specific scoring functions such as iMDLScore1 and iMDLScore2 (57) have optimized the weight coefficients of the scoring terms (63) using multilinear regression (MLR) methods. In a comprehensive evaluation and comparison for eleven other scoring functions, iMDLScore1 and iMDLScore2 (57) have shown better performance in both binding mode and affinity predictions.

Several RNA-ligand docking software packages have incorporated their respective built-in empirical scoring functions (60; 67; 70; 71) specifically for RNA/DNA-small molecule interactions. The scoring function of rDock (67), the successor of the original model RiboDock (60), contains a weighted sum of various intermolecular and intramolecular interaction energies, including the van der Waals potential (vdW), an empirical energy term for attractive and repulsive polar interactions and the desolvation energy. Because virtual drug screening can benefit from our knowledge, such as pharmacophoric points and shape similarity, derived from known RNA-ligand complexes, rDock has added pseudo-energy restraint terms as an empirical bias, such as pharmacophoric restraints. Pharmacophoric restraints used in rDock ensures the generated ligand poses to satisfy the pharmacophores derived from the known RNA-ligand complexes or the hot-spot mapping methods. Applications to the virtual screening against Hsp90, both rDock and Glide show significant improvements with the inclusion of the bias (67). Since rDock has been optimized for RNA docking, it outperforms Vina and Glide for RNA-ligand docking: For a set of 56 RNA-ligand complexes, the top-ranked poses predicted by rDock show a 54 ± 3 % success rate with a 2.5Å RMSD cut-off compared to 29± 2% for Vina and around 17.8% for Glide (67).

RLDOCK (70; 71) trained the weight coefficients for the different interaction terms such as van der Waals (vdW), electrostatic, polar and nonpolar hydration energies, and hydrogen-bond interactions for a set of 30 RNA-small molecule complexes. To enhance computational efficiency. RLDOCK adopts a two-step screening algorithm: In the first step, using a computationally efficient, crude estimation for the Born radii in the electrostatic energy calculation and the solvent-accessible surface area in the hydration energy, the model selects an initial pool of potential binding poses; in the second step, a rigorous scoring function is used to re-rank the binding poses to identify the top-ranked poses. Test on a separate set of 200 RNA-small molecule complexes indicates that the success rate of identifying the native/near-native binding modes increases significantly from the crude scoring function to the more refined scoring function, with 8.3%, 22.2%, and 29.6% for the crude scoring function and 17%, 40.4%, and 49.1% for the more rigorous scoring function within RMSD thresholds 1Å, 2Å, and 3Å, respectively. Considering the fluctuations in the ligand pose, RLDOCK groups similar ligand poses (according to the mutual RMSD) into clusters and rank the clustered poses. With the RLDOCK-ranked ligand poses (clusters), 44.3%, 74.3%, and 82.2% of the top-10 are within 1, 2, and 3Å (RMSD), respectively, to the native pose (70; 71), Furthermore, tested on a previously proposed set of 38 RNA-small molecule complexes (175), RLDOCK has demonstrated a higher success rate compared to other models for recovering the native ligand binding poses within RMSD of 2Å to the native pose. Specifically, RLDOCK shows a success rate of 55.3% (60.5%) for the top-1 (top-3) predicted poses as compared to 28.9% (39.5%), 36.8% (44.7%), 39.5% (47.4%), and 50.0% (57.9%) for DrugScore^RNA(112; 172), DOCK 6(65), LigandRNA(175; 203), and a combined LigandRNA(175; 203) and DOCK 6(65) approach, respectively. Since RLDOCK distinguishes itself from other models by a global, complete sampling of all the possible binding sites and poses, the results above demonstrate the importance of high-quality sampling for ligand poses.

In summary, compared to atomistic force field-based approaches, the empirical energy function methods manage to reduce the computational burden using simple functional forms for RNA-small molecule interactions. However, this approach is subject to two main limitations: (a) neglecting the correlation between different interactions and (b) transferability of the weight coefficients between different RNA-ligand systems. The success of the model depends on the quality of the curated training set, thus the accuracy of the predictions is limited by the lack of available high-quality data for RNA-ligand complexes.

4.2. Knowledge-based scoring functions: statistical potential provides efficient scoring of binding modes

Statistical potential approach based on reference states.

A statistical potential approach uses the inverse Boltzmann law to extract energy-like potential for user-defined interacting pairs from the experimental data:

$$E\propto -\sum _{i\in R}\sum _{j\in L}\text{ln}{\rho }_{ij}$$ (2)

where and ρ_ij is the relative frequency of the user-defined interacting pair (i, j) between the receptor R and ligand L. Before being applied to RNA-ligand docking models (112; 142; 172; 174-176), the statistical potential approach has been demonstrated to be effective for predicting protein-ligand docking (72; 73; 75; 173; 204-216). Different variants of the statistical potential approach have been proposed for RNA-ligand systems since the development of DrugScore^RNA (112; 172). These statistical potential approaches mainly differ in two aspects: the choice of reference state and functional forms of potential energy terms. As early attempts, DrugScore^RNA (112; 172), Kscore (174), and LigandRNA (175) have constructed the reference state by treating all the relevant atoms in the RNA-ligand complex as non-interacting particles, with different atom types differentiated or undifferentiated. In addition to the distance-dependent pairwise potential used in Kscore and DrugScore^RNA, LigandRNA, by taking into consideration the relative orientations between different atom pairs, has added a three-body anisotropic potential. Combined with DOCK 6 (65), this orientation-dependent potential shows a higher success rate than DrugScore^RNA in predicting the native binding modes. Specifically, for a test set consisting of 42 RNA-small molecule complexes, with the 2Å RMSD criteria for a correctly predicted ligand pose, DrugScore^RNA, DOCK 6, and LigandRNA show a success rate of 31.0%, 35.7% and 35.7%, respectively. A DOCK6 and LigandRNA hybrid score scheme further gives the success rate of 47.6%. The results show that the knowledge-based approach can benefit from a more accurate potential that accounts for more detailed information such as distance and angular correlations between the different interactions.

Iterative statistical potential approach.

A major limitation of the above traditional approach is that the reference state ignores the many-body correlations between the different interactions. One way to circumvent this problem is to iteratively refine the energy function until the simulated probability distribution of the different atom pairs agrees with that observed from the experimental data (142; 176; 217-222). Because the simulated distribution is based on sampling over the full energy landscape, an iterative approach can account for both native and nonnative interactions.

SPA-LN (176) is an iterative statistical potential model for predicting nucleic acid-small molecule interactions. Using intrinsic specificity ratio (ISR), a measure of the native vs. nonnative binding modes discriminative power, the energy-like scoring function can account for both affinity and specificity. For binding affinity prediction, using Pearson correlation coefficient between the predicted and the experimentally measured affinity as a measure, for a set of 77 complexes from version 2014 of PDBbind database (223) and a separate set of 34 nucleic acid-small molecule complexes, SPA-LN gives Pearson correlation coefficients 0.58 and 0.60, respectively. For the binding pose prediction, for a test set of 56 nucleic acid-small molecule complexes (67), for the top-scored poses with 2.5Å RMSD cutoff for a pose considered to be native or near-native, SPA-LN (176), rDock (67), AutoDock Vina (66) and Glide (61) give a success rate of 54(±3)%, 54(±3)%, 29(±2)%, and 17.8%, respectively. The performance of SPA-LN suggests the importance of considering not only affinity but also the specificity of RNA-ligand binding.

To capture the stacking and electrostatic interactions in nucleic acids, ITScore-NL (142), an iterative statistical potential approach (219), adds an extra distance-dependent stacking potential term and an electrostatic potential energy term to the scoring function. Stacking potentials were calculated for all carbon-carbon atom pairs involved in stacking interactions between nucleobases and planar aromatic groups of a small molecule. Electrostatic potential was calculated for the polar atom pairs using the Debye-Hückel approximation. The model was compared with other methods in two datasets to validate the performance on native pose recovery and binding affinity prediction. With the same set of 77 nucleic acid-small molecule complexes used in the test of SPA-LN, ITScore-NL achieved a higher Pearson correlation coefficient (R = 0.64) than that shown in SPA-LN (R = 0.58). As for the success rate of native pose recovery, ITScore-NL (142) was able to correctly identify native binding mode of 71.43% (50.64% for SPA-LN (176)) complexes with RMSD cutoff 1.5Å if the top-3 predictions are selected and 90.90% (76.62% for SPA-LN (176)) complexes with RMSD cutoff 3.0Å if the top-5 predictions are selected. Compared to LigandRNA (175) on a 42 RNA-small molecule complexes with only top-scored poses being selected and poses with RMSD cutoff 2.0Å being used, the success rates of LigandRNA (175) and ITScore-NL (142) are 35.7% and 50.0%, respectively. The results indicate the importance of including stacking and electrostatic interactions in RNA-ligand docking.

In summary, compared to atomistic force field methods, the statistical potential approach is associated with a much higher computational efficiency. However, the choice of reference state places obstacles for accurate modeling of RNA-small molecule interactions. Even though iterative approaches have been developed to circumvent the reference state problem, constructing diverse and complete decoy sets for training remains a challenge for RNA-ligand complexes. Furthermore, because data-driven approaches rely on experimentally determined structural data, the success of the models suffers from limited structure data for RNAs and RNA-ligand complexes.

4.3. Machine-learning based scoring method for RNA-ligand docking: an emerging scoring approach with high promise

Machine-learning approach.

With the success of machine-learning methods in various fields, a variety of machine-learning models such as support vector machine (SVM), random forest (RF), neural network (NN), and convolutional neural network (CNN) have been proposed and shown success to predict protein-small molecule interactions (224-229). Machine learning approaches not only have the advantage of utilizing the experimental data and making fast predictions but also with a large number of trainable parameters, can leverage experimental data better than traditional knowledge-based methods. Furthermore, the relation between input features and output results is learned through the training process. Therefore, a machine-learning method can be readily adopted across different types of tasks by simply changing the input features and output format, which can be engineered for the corresponding task. Figure 6 shows a typical workflow of training, validating, and testing a machine-learning model.

Figure 6:

The typical workflow of a machine-learning approach. Training and validation cycle usually needs to be performed many times before the performance on the validation set reaches an acceptable level. After the training-validation cycle, the trained model is used to make predictions on the test set.

Importance of feature engineering.

Although the quality and amount of the training data is vital to the performance of machine-learning methods, input feature engineering, which is often overlooked, is also critical to the success of the model. Generally, input features extracted from structure or geometry-based models for RNA-ligand binding often contain a large amount of detailed structural information, resulting in noise and excessively high dimensions in the parameter space. For example, there are many CNN-based approaches (230-234) that simply extend the 2D image in the original CNN model by treating the binding site as a 3D image. However, this type of 3D image is not rotational invariant hence requires rotational augmentation when used in training and prediction. The extra dimensions added would significantly increase computer time for making a single prediction and for performing large-scale virtual screening for drug discovery. Additionally, many atoms which are shown as pixels in the image may not even contribute to the binding, thus further complicate the learning process. An optimal engineered feature should maximally simplify the input information while capturing the key features that determine RNA-ligand docking results.

T-Bind (177) is a method for protein-small molecule binding affinity prediction. What makes it interesting for its possible application to RNA-ligand binding is not only the machine-learning model but also, more importantly, its feature extraction method. In T-Bind (177), Cang et al. (177) introduces a novel mathematical concept, element specific persistent homology (ESPH) or multicomponent persistent homology, to capture the crucial topological information around the binding site. This feature extraction method offers a new way to embed geometric information into topological invariants and simplify the input features while still capturing the key information. Benchmark tests using PDBbind database (223; 235-239), a comprehensive collection of protein-small molecule binding affinity data together with 3D structures, have yielded Pearson correlation coefficients 0.818 and 0.767 for PDBbind v2007 core set (237) and PDBbind v2013 core set (238), respectively, and the T-Bind outperforms other scoring functions (177). The result shows the merit of this feature extraction method and the promise of applying the same approach to the prediction of RNA-small molecule interactions.

RNAmigos (178) is a machine-learning model designed for the prediction of RNA-small molecule interactions. In RNAmigos (178), the base-pairing network around the binding site is simplified as a connected 2D graph with vertices and edges, where a nucleotide is represented by a vertex and backbone connectivity and base-pairing are represented by the different types of edges. The base-pairing interactions encoded in the 2D graph provide a signature for predicting the fingerprint of the small molecule that wifi most likely bind to the site. Furthermore, RNAmigos (178) shows the versatility of the machine-learning method. The model combines RNA base-pairing information at the binding site (in a 2D graph format) and graph neural network (240) (designed for data with 2D graph structure) to directly predict the fingerprint for the small molecule. RNAmigos encodes the predicted fingerprint as a 166 bit MDL Molecular Access keys (MACCS) (241). The small molecule can be found in a compound library simply by a similarity search against the predicted fingerprint. This procedure circumvents the traditional virtual screening route (docking then scoring) and substantially reduces the time for the search of a putative drug. RNAmigos (178) was trained with a set of 773 RNA-small molecule binding sites associated to 270 unique small molecules. Test results on an enrichment dataset with 176 unique RNA chains against 82 unique ligands have shown that RNAmigos outperforms a template-based method (Inforna 2.0 (77; 78)). Although RNAmigos shows promising results, it has two major limitations. First, RNAmigos requires prior knowledge of the binding site in order to generate a base-pairing network, and a misplaced binding site could led to a degradation in the predictive power. However, accurate determination of the binding site itself can be a challenging task. Second, RNAmigos considers different RNA-small molecule complexes to occur equally likely. Such treatment can cause an effective bias in the training process because the binding affinities of the different RNA-small molecule complexes can span across a large range of values.

As another machine-learning model for RNA-small molecule binding, RNAPosers (141) contains a set of trained pose classifiers that can estimate the “nativeness” of a ligand for a given structure of the RNA and the ligand. The classifiers are based on the random forest method (242) with an ensemble of 1000 decision trees. For a given ligand, RNAPosers takes a pose fingerprint as its input, where the pose fingerprint is calculated as the sum of a Gaussian function multiplied by a cosine damping factor over the RNA-ligand interacting atom pairs. The comparison between RNAPosers and two knowledge-based methods, DrugScore^RNA (112; 173) and SPA-LN (176), shows that RNAPosers (141) is able to yield higher success rates for the prediction of the native binding pose. Specifically, for a set of 31 RNA-small molecule complexes used in DrugScore^RNA (112; 173), RNAPosers gives a success rate of 61.9% as compared to 57.1% for DrugScore^RNA. For another set of 56 RNA-small molecule complexes used as a validation set in SPA-LN (176), RNAPosers gives a success rate of 62.5% compared to 54.0% for SPA-LN. These results show the advantage of the machine learning method over traditional knowledge-based approaches.

Coarse-grained conformational representation, traditionally implemented in RNA folding models, can lead to a unique method for feature engineering. Recently Stefaniak and Bujnicki developed a new machine-learning model, AnnapuRNA (143), specifically for RNA-ligand interactions. The feature engineering algorithm in AnnapuRNA employs a coarse-grained representation of both RNA and ligand to derive RNA-ligand contact statistics. Specifically, RNA structure is coarse-grained with each nucleotide replaced with five beads (243), and a ligand is represented with the concept of pharmacophores (244). Contact statistics collected from the coarse-grained representation with the assumption that the coarse-grained contact statistics can represent the core RNA-ligand interaction data. Five different machine-learning algorithms (Random Forest-RF, k Nearest Neighbors-kNN, Gaussian Naïve Bayes-GNB, Support Vector Machines with RBF kernel-SVM, and Deep-Learning-multi-layer feedforward artificial neural network-DL) have been trained on the coarse-grained statistics and each RNA-ligand complex is evaluated using a scoring function for the nativeness probability of the contacts and the ligand internal energy (202). Benchmark test with a set of 33 RNA-ligand complexes has shown that kNN and DL algorithms give the best results and more extensive tests with 4 docking methods and 9 scoring functions have demonstrated that AnnapuRNA outperformed other programs tested. The results has indicated that the coarse-grained representation combined with the concept of pharmacophores can indeed provide an effective, simplified way of feature engineering for RNA-ligand binding.

In summary, although machine-learning models for protein folding (245-250) and protein-small molecule interactions (72-75) have shown significant success, modeling RNA-small molecule interactions using machine learning is a relatively new adventure. The machine learning approaches have several intrinsic limitations. Because the model involves a large number of trainable parameters, the training process is prone to overfitting, especially for cases with only limited training data. The problem is more important for RNA-small molecule binding due to the lack of a comprehensive and high-quality curated database. Although protein-focused libraries, such as PDB (94), PDBbind (223; 235-239), can contain data for RNA-small molecule complexes, a more comprehensive, dedicated NDB-like database (251; 252) for RNA-ligand complexes is needed.

4.4. Accounting for solvent-mediated interactions

The sugar-phosphate backbone is negatively charged and polar, resulting in an accumulation of water molecules, cations, and water/ion-mediated RNA-ligand interactions. However, most molecular docking models do not explicitly consider the bridging effects of water molecules (Figure 2b) and metal ions (Figure 2c). The neglect of such solvent effects is a notable drawback that can cause inaccurate predictions for RNA-ligand interactions. A viable approach is to use simulations with explicit waters and/or ions to refine RNA structures (53; 55). Then, the positions of important water molecules can be retained for further docking against ligands. In this approach, the results are sensitive to the selection of the important water molecules. Because ligand-RNA interactions are sensitive to the positions and orientations of the water molecules and ions within the cavity space, achieving robust accuracy can be challenging with this approach.

An alternative approach is to predict the binding of water molecules and bound metal ions to the RNA prior to docking (253-260) then treat the predicted bound water molecules and/or ions as part of the receptor for RNA-small molecule docking. The Tightly Binding Ion (TBI) (256; 261; 262) model and the Monte Carlo TBI (MCTBI) model (260; 263) predict the ion distribution around an RNA structure. Through explicit sampling of the discrete ion distributions, TBI and MCTBI go beyond the mean-field Poisson-Boltzmann theory by accounting for the correlation between the different ions. 3D-RISM is another promising model for predicting the distribution of both solvent and ions around a given macromolecule (258). In combination with a force field calibrated by prototype ionophores, 3D-RISM is able to recapitulate the water distribution around guanine quadruplexes by considering the correlation between solvent particles and treating the system as macromolecules in equilibrium with a bulk solvent at constant composition (and chemical potential). For the Oxytricha nova telomeric G-quadruplex structure as a test case, 35% (80%) of the 3D-RISM-predicted water binding modes are within 1Å (2 Å) from the crystallographic modes, indicating that the model may be reliable for treating solvent effects (264). Splash’Em (Solvation Potential Laid around Statistical Hydration on Entire Macromolecules) is a model for predicting bridging water molecules in nucleic acid-ligand complexes. Using the statistical information of water molecules around nucleotides in the PDB (94) and a scoring function containing a hydrogen-bonding potential with both directionality and polarization, Splash’Em has identified 62% of water molecules in nucleic acid-ligand complexes within 1Å (265).

5. Current achievements of RNA-ligand docking models: a performance comparison

Table 3 summarizes the benchmark test results of various computational methods for RNA-ligand docking. The data is adopted from the original publications of the models and the results are grouped according to the test sets. Cautions should be taken when comparing the performance of the different models. First of all, RNAs and ligands in different test systems can have different structural and physical features, thus a direct performance comparison of the different models based on the results from different test systems may not be appropriate. Second, even for methods evaluated based on the same test dataset, the interpretation of the performance comparison can be complicated. For example, for the same benchmark dataset, the decoy poses generated for pose identification can be quite different for the different tests. A robust and objective comparison between the different scoring functions demands a consistent and systematic benchmark test protocol for the generation of the decoy binding poses (57).

Nevertheless, the benchmark test results in Table 3 provides useful insights for the selection of computational methods. First, for pose identification, we find a clear trend that RNA-specific methods consistently outperform those developed for proteins or generic macromolecules. The result shows the importance of considering RNA-specific interactions and structural features for RNA-ligand docking. Second, recent advances in RNA-ligand scoring function have been mainly focused on knowledge-based/machine learning-based approaches (141-143; 176). The knowledge-based/machine learning-based approaches provide equal or better performance (especially for affinity prediction) than the traditional physics-based/empirical approaches (59; 63; 65-67), except for MORDOR (64) and RLDOCK (70; 71). Third, even with the knowledge-based/machine learning-based approaches, the current success rate for affinity prediction is quite low. To improve the prediction accuracy, with the currently limited available data for RNA-ligand binding affinities and complex structures, new physics-based models that can accurately capture RNA-ligand interactions and conformational ensembles would be highly needed.

6. Non-docking based methods for modeling receptor-ligand binding

In addition to the scoring functions discussed above, other physics-based methods can also achieve high accuracy for determining the binding modes and affinities. Some of these methods, however, are not suitable for docking software due to either the expensive computational cost or the technical difficulty of incorporation of the method into a software. Because extensive reviews have been reported for free-energy methods (266), such as Molecular Mechanics/Poisson Boltzmann Surface Area (MM/PBSA) and Molecular Mechanics/Generalized Born Surface Area (MM/GBSA) (267-274), Linear Interaction Energy method (LIE) (275-277) Free-Energy Perturbation (FEP) (277; 278) and Thermodynamic Integration (TI) (272; 279), we here focus on several more recently developed physics-based methods for modeling small molecule binding problems, with the purpose of applying the methods to predict RNA-ligand binding.

Quantum mechanical methods.

Quantum mechanical approaches, which can treat polarization, charge transfer, and many-body effects, are considered to be more accurate than force-fields in molecular mechanics. Methods to model small molecule binding range from less accurate semi-empirical methods such as density functional theory (DFT) to more sophisticated methods, such as second-order Møller-Plesset perturbation theory (MP2), configuration interaction (CT) and strict coupled-cluster calculations (CC) (280). Improvements in computer hardware have substantially reduced the computational time for quantum mechanical calculations. However, the cost is still relatively high, and systems under investigation usually need to be significantly simplified or divided into smaller fragments. Although fragmentation calculation enables the use of quantum mechanics-based methods for computing the energies of large biomolecules, such as proteins (281-283), the method has not been extensively applied to RNA-small molecule interacting systems. The difficulties may stem from long-range electrostatic interactions and many-body quantum mechanical effects in RNAs (283; 284).

Chen et al. (285) have extended molecular fractionation with conjugate caps (MFCC) scheme using quantum mechanical calculations to study DNA/RNA-small molecule interactions. Through the division of the system at each phosphate group, three oligo-nucleic acid interaction systems are decomposed into smaller subsystems, and the calculated interaction energy is found to be in excellent agreement with the results obtained from ab initio calculation for the original full system. In another study, Mlýnský et al. (286) have compared the abilities of various hybrid QM/MM methods, including ab initio, DFT and semi-empirical approaches, to investigate the possible catalytic mechanism for the hairpin ribozyme. By using various hybrid QM/MM methods, Mlýnský et al. have computationally reconstructed potential and free energy surfaces for the catalysis reaction system. Among the tested methods, the activation barriers calculated from spin-component scaled Møller-Plesset (SCS-MP2) method and hybrid MPW1K functional show the best agreement with those derived from the experimental rate constant data. Recently, Bezerra et al. (68) have applied MFCC fragmentation scheme (287-289) within a density functional theory framework to calculate the interaction energy between ribosomal RNA and aminoglycoside hygromycin B. The calculation has revealed the regions where the drug molecule interacts strongly with the ribosome, and the result provides guidance for the improvement of drug-receptor affinity.

Compared to MM/MD based approaches, the quantum-mechanics approaches above are able to carry out more accurate calculations at the expense of expensive computational cost and hence the methods may not be suitable for large-scale virtual screening in drug discovery.

3D-RISM theory.

By treating the distribution of bound ligands around the receptor as a receptor-ligand two-body correlation problem, three-dimensional reference interaction site model (3D-RISM) with Kovalenko-Hirata (KH) closure relation predicts the spatial distribution function of the ligand in the field created by the receptor (290-292). The probable binding sites are identified from the peaks of the ligand spatial distribution function and the binding mode is determined based on the superposition approximation (69). The 3D-RISM/KH approach has three unique advantages. First, it identifies the ligand-binding site from the distribution function for the mixture of solvent and ligand. Because the calculation circumvents the sampling and scoring in the traditional docking process, the 3D-RISM/KH approach avoids the limitations of sampling methods. Second, the theory explicitly accounts for the solvent effects. Therefore, the theory can treat ion and water-mediated interactions in receptor-ligand binding. Third, unlike the traditional continuum models, 3D-RISM/KH calculation can take the hydrogen bond between the solute and solvents into consideration. Several studies (293-298) have employed 3D-RISM/KH theory to predict the binding sites and binding modes of small molecules in proteins. In a recent study, Sugita and coworkers (69) have tested their 3D-RISM/KH approach on 18 different types of proteins and successfully predicted the native binding modes for half of the systems.

However, the 3D-RISM/KH theory is not without limitations. First, there is an upper limit of the query small molecules for a computationally feasible calculation. In practice, a large ligand needs to be divided into fragments which are later re-connected based on the calculated distributions. Second, 3D-RISM/KH takes longer than traditional docking calculation to evaluate tens of thousands of ligands as drug candidates for a given target. Third, the theory is based on correlation functions and may not be able to account for certain discrete configurations and interactions that are important for an accurate prediction.

Kinetic effects.

For many drug molecules, binding equilibrium might not even be reached or maintained in the in vivo system. As a result, the thermodynamic equilibrium binding affinity may not be a proper indicator for the in vivo efficacy of the drug. In contrast, residence time, or the lifetime of the binary drug-target, complex, measured by the inverse of the unbinding rate constant k_off, may be a better indicator for drug efficacy in vivo (299-303). Indeed, a growing amount of evidence points to the direct correlation between the residence time of a drug molecule and its in vivo efficacy (299-309); See Figure 7 for a schematic visualization of the binding process for systems with simple and complex transition and intermediate states.

Figure 7:

A simplified representation of the binding kinetics between the unbound receptor (R), unbound ligand (L) and the bound receptor-ligand complex (RL). (a) The binding kinetics of a system with only one transient state (TS) along the binding reaction coordinate. The figure shows a binding scenario where both receptor and ligand undergo conformational changes in the binding process. The kinetic residence time (i.e., the inverse of the RNA-ligand dissociation constant k_off) depends on the free energy difference (ΔG_off) between the bound state and transient state, while the thermodynamic binding energy (ΔG_bind) is determined by the free energy difference between the unbound state (R+L) and bound state (RL). (b) In practice, often the binding kinetic profile of a system contains multiple transient states (TS) and intermediate states (IS) with a much more complicated kinetic mechanism.

Currently, kinetic models such as unbiased MD, Markov state model, and weighted ensemble and metadynamics (MTD) have demonstrated the success of kinetic modeling in protein-targeted drug discovery (183; 302; 303; 310). Similar kinetic studies for RNA-ligand binding is expected to provide a novel strategy for RNA-targeted drug design (311).

7. Conclusions and future perspectives

The rapidly growing therapeutic interest in RNA-targeted drug discovery causes an increasing demand for computational tools for predicting RNA-ligand interactions. Virtual screening remains an important first step in novel drug design when only the targeted RNA information is available. Various docking (sampling) methods and scoring functions have been developed to accelerate this process and in the meantime, have deepened our understanding of RNA-ligand binding mechanisms. Physics-based and knowledge-based approaches have shown promising success in predicting ligand binding poses and binding affinities. However, powered by advanced algorithms, machine-learning methods, although still in their infancy for RNA-ligand docking, have begun to show highly encouraging improved performance compared to traditional approaches.

Computer-aided drug discovery has come a long way. Past efforts have been mainly protein-centered. New RNA-based therapeutic design and computational methods have emerged. With the growth of the database of known structures and kinetic and thermodynamic measurements (e.g., binding affinity), data-driven methods, especially machine-learning methods, are expected to play a more and more important role. Furthermore, with the appreciation of RNA-ligand binding kinetic effects on in vivo efficacy of the drug, kinetics-based models, although currently have not been fully explored for RNA-drug binding, would be developed at an accelerated pace. With the development of various computational tools developed for RNA-targeted drug discovery, a CASP- and D3R-like (312; 313) community-wide events with blind tests and well-curated benchmark datasets, similar to the benchmarks widely used in the protein-ligand modeling community (216; 314-317), would be much needed.