Novel Insights of Structure-based Modeling for RNA-targeted Drug Discovery

Abstract

Substantial progress in RNA biology highlights the importance of RNAs (e.g., microRNAs) in diseases and the potential of targeting RNAs for drug discovery. However, the lack of RNA-specific modeling techniques demands for development of new tools for RNA-targeted rational drug design. Herein, we implemented integrated approaches of accurate RNA modeling and virtual screening for RNA inhibitor discovery with the most comprehensive evaluation to date of five docking and 11 scoring methods. For the first time, statistical analysis was heavily employed to assess the significance of our predictions. We found that GOLD:GOLD Fitness and rDock:rDock_solv could accurately predict the RNA ligand poses, and ASP rescoring further improved the ranking of ligand binding poses. Due to the weak correlations (R²<0.3) of existing scoring with experimental binding affinities, we implemented two new RNA-specific scoring functions, iMDLScore1 and iMDLScore2, and obtained better correlations with R²=0.70 and 0.79, respectively. We also proposed a multi-step virtual screening approach and demonstrated that rDock:rDock_solv together with iMDLScore2 rescoring obtained the best enrichment on the flexible RNA targets, whereas GOLD:GOLD Fitness combined with rDock_solv rescoring outperformed other methods for rigid RNAs. This study provided practical strategies for RNA modeling and offered new insights into RNA-small molecule interactions for drug discovery.

Introduction

The recent substantial progress in RNA biology underscores the importance of RNA in normal and aberrant cellular functions. It also highlights the potential of targeting RNA for treatment of a multitude of diseases including bacterial/viral infection ^{1, 2} and cancer ^{3, 4}. RNAs can form well-defined tertiary structures, such as double helices, hairpins, bulges, and pseudoknots, which offer structural basis for designing therapeutic agents. As a matter of fact, some structured RNAs, including bacterial 16S ribosomal RNAs (rRNA), HIV-1 TAR RNAs and small non-coding microRNAs (miRNAs), exhibit attractive structural and functional characteristics similar to proteins ⁵. We have linked miRNAs to different diseases including cancer ^{6, 7}, and recently embarked on the discovery of small molecule inhibitors targeting miRNAs (SMIR) ^{8, 9}. Druggable RNA targets are largely unexplored ¹⁰, and RNA inhibitors, such as aminoglycosides which are for the most well-defined RNA target -- 16S rRNA A-site, usually have poor selectivity or oral bioavailability. There have been several reports in targeting prokaryotic rRNA A-site ^11-13, HIV-1 TAR RNA ^14-16, and riboswitches ^17-19 with small molecules. Researchers are also exploring new generations of drug-like molecules targeting pathogenic or human disease-related RNAs including CUG- or CCUG-repeated mRNA ^20-22, miRNA ^{23, 24}, and internal ribosome entry site (IRES) ^{25, 26}. These studies provided proof of principle that RNAs can be specifically targeted for antiviral or anticancer therapeutic development.

Structure-based modeling techniques, such as molecular docking, have been widely used in the field of protein-targeted drug discovery; however, most of these in silico methodologies were developed for proteins. Mature tools specific for RNAs (e.g. virtual screening for RNA inhibitor identification) are lacking ²⁷. Thus, there is an unmet need to exploit current computational tools and implement new ones for RNA modeling such as fast screening of small molecules against RNA targets. To this end, Li et al. evaluated two docking programs (GOLD and Glide) and concluded that they are helpful in RNA-based drug discovery ²⁸. DOCK6 with implicit solvent models, i.e. GB/SA and PB/SA models, was used to model RNAs and obtained low root mean square deviations (RMSD) ¹⁰. AutoDock4 was also modified to dock RNA targets with flexible grids and achieved some degree of success ^{29, 30}. The knowledge-based DrugScore^RNA was parameterized with 670 RNA-ligand and -protein complexes but only a fair correlation with the experimental binding affinities was observed for a limited set of 15 RNA-ligand complexes ³¹. RiboDock (currently as rDock) has proved to be successful in several drug design cases ^{12, 32, 33}. The latest version of rDock employed a fast weighted solvent accessible surface areas (WSAS) to approximate the solvent accessible areas ³². MORDOR (molecular recognition with a driven dynamics optimizer) ³⁴ was implemented with induced-fit algorithms for flexible RNA docking. However, as MORDOR is computationally expensive, it is not feasible to screen a large chemical database in an efficient manner. Furthermore, we found that the docking parameters widely used in proteins are not always applicable to RNA systems. For instance, the electrostatic interactions between RNA phosphates and ligands can be overestimated ^{10, 27, 35}, while the desolvation term also needs to be modified ²⁹.

To date, due to the challenge of RNA modeling and less developed programs in this field, there have been no reports that provided feasible approaches to address the three critical issues in docking: the accuracy of binding mode prediction, the ranking performance in virtual screening, and the accurate scoring ³⁶. Our study aims to identify cost-effective combinations of existing techniques and develop new docking strategies for RNAs. Therefore, we comprehensively evaluated five popular docking programs, including GOLD 5.0.1 ³⁷, Glide 5.6 ³⁸, Surflex 2.415 ³⁹, AutoDock 4.1 ^{40, 41} and rDock 2006.2 ³², along with 11 scoring functions to explore their capability in RNA docking. Based on our study, we proposed the workflow for structure-based modeling for RNA-targeted drug discovery. As illustrated in Fig. 6, appropriate docking/scoring combinations need to be used to obtain the best virtual screening performance based on the flexibility of binding sites. To maximize the modeling accuracy and improve the enrichment, different modeling strategies can be rationally applied at three critical quality-control steps: near-native pose coverage, pose selection, and hit selection. For the first time, intensive statistical analysis was employed in such studies, and we also implemented RNA-specific scoring functions with the ever largest high-quality RNA-ligand binding affinity dataset. Moreover, our study provided new insights of RNA-small molecule interactions, and this helped us explore the best computational strategies for RNA modeling in drug discovery and development.

Figure 6.

The suggested workflow for structure-based virtual screening for RNA-targeted inhibitor discovery.

Results

Reproduction of experimental binding modes

We first examined whether the current protein-ligand derived docking/scoring software could reproduce the ligand binding poses similar to the experimental structures. Ideally, a “good” RNA docking program should be able to sample all the conformational space and identify at least one near-native pose. Table 1 showed that, if we arbitrarily employed C(5, 3.0Å) (the top 5 ranked pose includes at least one near-native pose with RMSD<3.0Å) to define a successful docking case, GOLD:GOLD Fitness and rDock:rDock_solv outperformed others, both with 73.21% success rate. Additionally, GOLD:ChemScore, GOLD:ASP, Glide:GlideScore(SP), Glide:Emodel(SP) and rDock:rDock obtained more than 50% docking success rate. In contrast, the success rate for Glide:GlideScore(XP), Glide:Emodel(XP), Surflex and AutoDock4.1 (default) were low, ranging from 30.36% to 44.64%. All programs, especially AutoDock4.1 and Surflex, had weak performance (<60%) on flexible and extensively-charged aminoglycosides. When stringent cutoffs such as C(3, 1.5Å) were used, the accuracy decreased but GOLD:GOLD Fitness and rDock:rDock_solv kept as the best (>40%). Results for other programs are available in Table 1.

Table 1.

Successful binding mode reproduction of the 56 RNA-ligand complexes using different docking/scoring combinations with arbitrary cutoffs. The values in the brackets indicated the total number of complexes in that category. The values before the parentheses results satisfying C(5, 3.0Å), and the values in the parentheses were derived with the stringent criterion C(3, 1.5Å) (described in the Materials and Methods). The Glide XP mode could not obtain enough data with the stringent restraints, thus the VUS values and R² were not available.

	GOLD 5.0.1			AutoDock 4.1	Surflex 2.415	Glide 5.6				rDock 2006.2
	GOLD Fitness	ChemScore	ASP	Autodock4.1 Score	Surflex-dock Score	GlideScore (SP)	Emodel (SP)	GlideScore (XP)	Emodel (XP)	rDock	rDock_solv
Aminoglycoside [26]	18 (9)	13 (3)	15 (9)	1 (1)	4 (2)	12 (3)	13 (3)	4 (2)	4 (2)	13 (6)	16 (8)
Small Molecule [30]	23 (15)	17 (10)	22 (15)	16 (9)	21 (13)	18 (13)	18 (13)	16 (11)	15 (11)	21 (13)	25 (15)
X-ray crystal [36]	29 (19)	26 (13)	29 (21)	13 (8)	17 (11)	20 (9)	21 (9)	12 (8)	12 (8)	24 (16)	29 (17)
NMR [20]	12 (5)	4 (0)	8 (3)	4 (2)	8 (4)	10 (7)	10 (7)	8 (5)	7 (5)	10 (3)	12 (6)
Total [56]	41 (24)	30 (13)	37 (24)	17 (10)	25 (15)	30 (16)	31 (16)	20 (13)	19 (13)	34 (19)	41 (23)
Overall Success Rate %	73.21 (42.86)	53.57 (23.21)	66.07 (42.86)	30.36 (17.86)	44.64 (26.79)	53.57 (28.57)	55.36 (28.57)	35.71 (23.21)	33.93 (23.21)	60.71 (33.93)	73.21 (41.07)
VUS %	78.11	65.48	70.17	43.30	55.22	65.41	66.01	NA	NA	63.09	73.13
Score-binding affinity correlation R2 a	0.25	0.03	0.29	0.22	0.05	0.10	0.14	NA	NA	0.15	0.18

^aexcluding 1TOB, 2TOB and 1LVJ when calculating Pearson correlation coefficient

As expected, we observed that the average docking accuracy on crystal structures was higher than that on NMR structures for all of 11 current docking/scoring combinations (58.84% versus 42.27%, p=0.06). Not surprisingly, the pose reproduction performance on small-molecule RNA ligands was remarkably better than that on flexible aminoglycosides with high statistical significance (64.55% versus 39.51%, p<0.01). Among the failed cases (defined as “two or less docking programs are able to reproduce the near-native structure (RMSD<3.0Å) among top 5 scored poses”), five are crystal structures (2O3V, 2BE0, 2FD0, 2PWT and 2Z75) and seven are NMR structures (1UUD, 1LVJ, 1TOB, 1AKX, 1EI2, 1KOD and 1QD3). We found that the current methods were usually less accurate on RNA complexes containing large aminoglycosides (e.g. lividomycin, paromomycin, etc.), weak RNA binders (e.g. arginine and citrulline), or phosphate-containing hydrophilic ligands (glucosamine 6-phosphate). Because the negatively charged moieties can form specific interactions with RNA phosphates in the presence of metal ions acting as the “metal bridge” ⁴², such as 2GDI and 2Z74, we tried docking with consideration of metal ions. As expected, we could significantly improve the pose prediction of the diphosphate tail of thiamine diphosphate in 2GDI when the Mg²⁺ ion was taken into account as part of RNA targets. When compared with rDock:rDock_solv, the GOLD:GOLD Fitness combination achieved better performance for the pose reproduction on aminoglycosides-RNA complexes such as 1J7T, 2FCZ, 2BE0, 1NEM and 2TOB, whereas rDock:rDock_solv was more accurate for drug-like ligands including 2Z74, 2Z75, 1EHT and 1AKX. The detailed results (scores, RMSD and statistics) are available in Supplementary Table S3-S4.

To better demonstrate the relation of the docking accuracy with RMSD and ranking, we illustrated our results with Fig. 1A, in which the heavy-atom RMSD and the ranking of pose were considered at the same time. The volume under the surface (VUS) represents the overall capacity of reproducing the near-native binding modes. It showed that GOLD:GOLD Fitness achieved the best VUS (78.11%), while rDock:rDock_solv was the second best (Table 1 and Fig. 1A). We proposed to employ the contour of 50% success rate to guide the pose selection in RNA docking: if one aims to cover at least one near-native pose (RMSD<3.0Å) in 50% of the ligands, at least the top five poses should be kept when using GOLD:GOLD Fitness. In contrast, we should keep at least top 20 poses to achieve 50% success for Surflex and AutoDock 4.1 (Fig. 1B). From these assessments, we suggest that GOLD:GOLD Fitness and rDock:rDock_solv be the best methods for pose predictions in RNA docking.

Figure 1.

The 3D cumulative success rate from the binding mode reproduction experiments. (A). The cumulative success rate in 3D representation was calculated based on a series of C(X, YÅ). Only scoring functions which obtained the highest VUS values for each docking program were selected for illustration. The contour on the XY (RMSD-Rank) plane represented the 50% (Z=28) success rate (the binding mode can be reproduced for 50% of RNA-ligand complexes); (B). The 50% success contour (Z=28) for all available scoring functions (GlideScore (XP) and Emodel (XP) were not included due to the unavailability of VUS values). (C). The cumulative success rate for 56 RNA-ligand complexes based on the ranking of X-ray/NMR determined poses against 100 decoys. The 50% success line and the corresponding rankings to achieve 50% success were shown as dots.

Rescoring to improve near-native binding pose ranking

To improve the pose ranking accuracy with rescoring, we assessed the scoring function capability of differentiating the ligand crystal/NMR structures from their decoy poses. This was done by investigating two parameters: the ranking of native poses, and the Spearman's correlation between scores and RMSDs.

Generally, an ideal scoring function should rank the crystal/NMR and very near-native poses statistically higher than other decoys. Since GOLD:GOLD Fitness outperformed other docking programs on the coverage of the near-native poses as aforementioned, it was utilized to generate 100 decoys for each target in the hope to cover a wide range of poses, from near-native to decoys. We investigated whether a given scoring function could obtain the highest rankings for experimentally determined ligand poses. Analogous to IC₅₀ (in assessing biological activity), we used 50% success rate to evaluate the performance of different docking/scoring methods. As demonstrated in Fig. 1C, the 50% success rate line (dashed) clustered these scoring functions into three groups: ASP, ChemScore, AutoDock4.1 Score and Emodel (SP) were the first group; the second group included other scoring functions, except rDock which ranked the lowest as the third group. Fig. 1C indicated that GOLD Fitness has 50% of possibility to rank the native ligand conformation within top 10% of the predicted poses, whereas for ASP, ChemScore, AutoDock4.1 Score and Emodel (SP), this value was reduced to top 5%. The native pose ranking performance for different docking/scoring schemes varied with different RNA target structures. For example, most programs performed significantly better for crystal structures than NMR structures (69.14% versus 38.89%, p<0.01) with the top 10 as the cutoff to define a successful ranking case. Surprisingly, ASP was remarkably better in crystal structure ranking, in which only two targets (2O3V and 3DIL) failed, while AutoDock4.1 outperformed others on ranking NMR structures. Taken together, these data suggested that RNA targets with different structural resolutions should be rescored with respective appropriate scoring functions (e.g., ASP or AutoDock4.1) after the initial step of docking with GOLD:GOLD Fitness or rDock:rDock_solv. Detailed results of this ranking study are available in Supplementary Table S5.

The score-RMSD correlation is another parameter related to the ranking capability. Here we investigated this parameter using Spearman's rank correlation because it is known that RMSD is not linearly correlated with docking scores ²⁹. Our study showed that in most cases RMSD and docking scores were positively correlated, as desired, but the correlations varied for different RNA targets. Thus, we grouped the 56 RNA targets based on the strength of correlation for each scoring function (see Materials and Methods). ASP, GlideScore (SP) and Emodel (SP) were the best three scoring functions which had most cases with moderate or strong correlation between RMSD and score (Table 2). rDock, rDock_solv and Surflex-dock scores obtained fair performance, which could derive weak or strong correlations for more than 1/3 of cases. Surprisingly, GOLD Fitness could not achieve satisfactory performance to enrich the near-native ligand conformations (44 cases obtained the weak correlations) (Table 2). Combined with the native pose ranking analysis, these results demonstrated that other scoring functions such as ASP could enrich the near-native poses when applied to decoy poses generated by GOLD:GOLD Fitness.

Table 2.

Score-RMSD Spearman's rank correlations for various scoring functions. The values indicated the number of RNA-ligand complexes fit in each correlation category (Weak: ρ<0.3, Moderate: 0.3≤ρ<0.5, Strong: ρ≥0.5).

	Weak	Moderate	Strong
GOLD Fitness	44	5	7
ChemScore	41	7	8
ASP	33	15	8
GlideScore (SP)	31	15	10
Emodel (SP)	29	14	13
Surflex-dock Score	38	12	6
AutoDock4.1 Score	40	5	11
rDock	35	12	9
rDock_solv	36	12	8

As we have identified ASP as the best scoring function for ranking RNA ligand poses, we wanted to further study whether it could improve the identification of the near-native binding poses generated by GOLD:GOLD Fitness. Supplementary Table S4 showed that when all top 10 poses generated by GOLD:GOLD fitness were rescored by the ASP scoring function, the number of RNA targets satisfying C(5, 3.0Å) increased from 41 to 44, while this number for C(3, 1.5Å) increased from 24 to 30, compared to original GOLD:GOLD Fitness performance. Specifically, we observed that the best RMSD in top 5-scored docking conformations of 2GDI, 2Z74, 2PWT and 1ZZ5 was significantly reduced (below 3.0Å) after ASP rescoring (Supplementary Figure 1 and Supplementary Table S4). In contrast, GOLD:GOLD Fitness alone failed to identified the near-native conformation for these targets. Furthermore, the VUS increased from 78.11% to 79.18%. Compared with the docking accuracy using GOLD:GOLD Fitness alone, the average RMSD for the top-scored conformations was further reduced to 2.61±0.38Å after ASP rescoring (Supplementary Table S4). Combined with native pose ranking and RMSD-score correlation results, our results confirmed that ASP has the best ability for pose ranking, and ASP rescoring can significantly enrich the near-native decoys generated by GOLD:GOLD Fitness for pose reproduction purpose in RNA docking.

Development of new RNA-specific scoring functions for RNA-ligand binding affinity prediction

As most of the scoring functions used to predict RNA-ligand binding affinity were optimized for protein-ligand interactions, it demands for careful evaluation of existing scoring schemes and implementation of new RNA-specific methods. Since the determination of RNA-ligand binding affinities depends on multiple factors (e.g., temperature, assay method, etc.), we carefully curated 45 RNA complexes with reported binding affinities (K_d) (Supplementary Table S2). The score-binding affinity correlations were calculated for the selected scoring functions. As expected, the correlation coefficients (R²) for all evaluated scoring functions were low (R²<0.3). The values were in Table 1, and the score-binding affinity plots for the best three scoring functions (ASP, GOLD Fitness and AutoDock4.1 Score (default)) are available in Supplementary Figure 2.

To improve the result, we developed scoring functions with our large RNA-ligand dataset with available binding affinities. This was done by optimizing AutoDock4.1 scoring terms using multi-linear regression (MLR) methods. During the implementation, only four terms were optimized because 1. The charge-based desolvation energy function employed by AutoDock4.1 was based on Wesson & Eisenberg's model ⁴³, which was entirely trained by protein targets; 2. The solvent accessible surface area calculation in this model was simplified by using inter-atomic contact radius ⁴¹. Upon MLR optimization, the contributions of those scoring terms are 0.1460 for vdW, 0.0745 for hbond, 0.0559 for electrostatic, and 0.3073 for torsions. This new score functions was named as iMDLScore1. Since the training set for iMDLScore1 includes diverse RNA binders, varying from tiny hypoxanthines to large aminoglycosides, we observed a significant contribution of vdW interactions. As illustrated in Fig. 2A, we achieved a much better correlation (R² = 0.70) between docking scores and binding affinities after optimizing the coefficients of AutoDock4.1 scoring terms. When iMDLScore1 was further validated against an external test set consisting of eight complexes, the correlation coefficient between the score and binding affinity was 0.82, and the root-mean-square error (RMSE) of prediction was as low as 4.09kJ/mol.

Figure 2.

Correlation between docking scores and experimental binding affinity for iMDLScore1 and iMDLScore2.

A known challenge in RNA virtual screening is how to enrich the active compounds from a focused library with charged molecules because most RNA binders are potentially positively charged ²⁷. To overcome this problem, we derived a second scoring function, iMDLScore2, with a dataset containing 18 complexes with positively charged ligands. In iMDLScore2 the contribution of the electrostatic term to the docking scores was over 10%. Interestingly, R² and Q² (leave-one-out cross validation R²) for the training set reached 0.79 and 0.62, respectively (Fig. 2B), and R² for the test set was 0.76. Additionally, the RMSE of prediction was comparable to that of iMDLScore1 (4.35kJ/mol). Q², R² and RMSE of prediction indicated the good predictive capability of RNA-ligand binding affinities by iMDLScore2. The new coefficients in iMDLScore2 were 0.1634 (vdW), 0.2436 (hbond), 0.2311 (electrostatic), and 0.2212 (torsion). Obviously, the contributions of electrostatic and hbond were increased. This result indicated that polar non-bonded interactions are critical attributes to accurately predict the relative binding affinity for charged ligands. All results and parameters for the new scoring functions could be found in Table 3. We evaluated the performance of these new scoring functions in the virtual screening section.

Table 3.

Contributions of AutoDock4.1 energetic terms and associated R², Q² and RMSE of prediction.

Parameter	Default	iMDLScore1	iMDLScore2
vdW	0.1662	0.146	0.1634
hbond	0.1209	0.07451	0.2436
electrostatic	0.1406	0.05593	0.2311
desolvation	0.1322	0	0
torsion	0.2983	0.3073	0.2212
Number of complexes as training set	NA	25	18
R2 (training set)	0.22	0.70	0.79
LOO-Q2 (training set)	NA	0.44	0.62
R2 (test set)	NA	0.82	0.76
RMSE of prediction (kJ/mol, test set)	NA	4.09	4.35

Identification of rescoring scheme to improve the cross-docking performance

We employed cross-docking study to explore the best docking/scoring strategy to identify the cognate ligands for their corresponding RNA targets. Among the 56 RNA complexes, we selected a diversity subset of 16 crystal structure and 7 NMR structures for which we obtained good pose reproduction rates. As anticipated, no single docking program could lead to satisfactory enrichment in term of the average rankings of cognate ligands (e.g. average ranking below 5 out of 23). Glide:GlideScore (SP) achieved best average ranking of the cognate ligands (6.70±2.63), whereas GOLD-associated scoring functions had similar performances: 7.22±2.45 for GOLD Fitness, 7.13±3.01 for ChemScore, and 7.17±2.71 for ASP (Fig. 3). Glide:GlideScore (SP) acquired the best recovery AUC of cross-docking (74.66%), but the performance of Glide:Emodel (SP) was the least ideal (AUC=58.61%).

Figure 3.

The average ranking of the cognate ligands and AUC in the cross-docking study. The height of boxes represented the average ranking of the cognate ligands of the 23 RNA targets, while the linespoints are for the AUC. The AUC was calculated from the cumulative number of RNA targets in which the ranking of the cognate ligands was below different cutoffs. The error bars represented the 95% confidence interval from the 23 cross-docking cases. The “*” represented the method of rescoring based on the top 10 predicted poses by GOLD:GOLD Fitness. Ideally, AUC is 517.5 if all cognate ligands are ranked on as the top 1.

To improve the cross-docking performance, we rescored the top 10 poses generated by GOLD:GOLD Fitness with ChemScore, ASP, AutoDock4.1 Score, Emodel (SP), GlideScore (SP) and rDock_solv, because these scoring functions performed well in ranking poses (Fig. 1C). Fig. 3 and Supplementary Table S6 demonstrated that rDock_solv rescoring achieved the best average ranking of cognate ligands (4.2±2.0) and recovery AUC (85.69%). It ranked 16 cognate ligands on top 3; As a comparison, Glide:GlideScore (SP), the best cross-docking program above, could only rank 10 cognate ligands on top 3. When compared with the initial result from GOLD:GOLD Fitness, rDock_solv rescoring improved the ranking of 17 cognate ligands. The ASP rescoring was the second best (after rDock_solv) with average ranking of 4.4±1.9 and recovery AUC=84.53%. It improved the ranking of cognate ligands in 15 cases. Therefore, as expected, we confirmed that rescoring could significantly improve both average ranking of cognate ligands and recovery AUC.

When comparing the average rankings for low-resolution structures with high-resolution ones, we could group these scoring functions into two categories: soft-core scoring and hard-core scoring. For instance, we think the AutoDock4.1 scoring function is relatively “soft”, which obtained better cognate ligand ranking for low-resolution RNA structures that may contain unfavorable structural errors (e.g. clashes) ⁴⁴. In contrast, ASP and rDock_solv statistically achieved significantly better performance (p<0.05) on high-resolution crystal structures, indicating that they are hard-core scoring methods capable of penalizing structural defects (Supplementary Table S6). These characterizations were consistent to the results from our above native pose ranking evaluation, where AutoDock4.1 also achieved better performance for NMR structures while ASP was more accurate for ranking high-resolution crystal structures. Moreover, these findings are critical for selection of proper scoring functions for RNA virtual screening based on the RNA structural flexibility (see below).

Implementation of a two-step scheme for virtual screening

Receptor flexibility remains a challenge in structure-based virtual screening ⁴⁵. This has been partially addressed by applying soft potentials, rotamer libraries, and molecular dynamic simulations during/after docking ⁴⁶. Herein, we aim to identify the most rational combination of docking/scoring/rescoring strategies for RNA virtual screening. We utilized two different RNA targets: the bacterial 16S rRNA A-site and the lysine riboswitch, representing two typical kinds of RNA targets: bulge-containing A-form helix and aptamer, respectively.

We first quantitatively characterize the flexibility 16S rRNA A-site (PDB ID: 1J7T) by comparing the B-factors of the active site residues (4Å around paromomycin) with other non-terminal residues. Supplementary Figure 4A showed that the B-factors of active site were statistically higher than other part of the RNA (p=0.002), indicating that rRNA A-site is a flexible target. Furthermore, normal mode analysis with oGNM ⁴⁷ confirmed this local flexibility because significant fluctuation of the A-site residues (arrow highlighted) could be observed within five lowest-frequency modes (low-frequency motions are expected to have larger contribution to the conformational changes ⁴⁸). In particular, four critical A-site residues (A16, A17, A38 and A39, equivalent to A1492 and A1493 based on E. coli rRNA nucleotide numbering) were predicted to have the largest atomic fluctuations (Supplementary Figure 4B). Our analyses agreed with the finding that great conformational changes occur upon paromomycin binding and that A1492 and A1493 nucleosides can fluctuate between extra- and intra-helical states in molecular dynamic studies ^{49, 50}.

The virtual screening against the 16S rRNA A-site showed that, for 31 aminoglycoside mimetics and 11 aminoglycosides, both GOLD:GOLD Fitness and rDock:rDock_solv achieved extremely excellent virtual screening enrichment against MayBridge drug-like decoys (ROC AUC=1.0, Table 4). However, they could not obtain reasonable results for the Foloppe dataset ¹². rDock:rDock_solv obtained the best enrichment with only AUC=0.61, while for GOLD:GOLD Fitness the AUC is 0.58 (Table 4). Therefore, we concentrated on exploring novel strategies to improve the enrichment of virtual screening of drug-like, positively charged compounds (Foloppe dataset). Surprisingly, GOLD:GOLD Fitness coupled with rDock_solv or ASP rescoring did not improve the results (Fig. 4A and Table 4). However, the enrichment was significantly increased by rescoring either rDock:rDock_solv or GOLD:GOLD Fitness generated poses using our new scoring function iMDLScore2 with AUC=0.74 and 0.69, respectively (Fig. 4A and Table 4). The AUC difference between one-step and two-step virtual screening indicated that, if the binding modes are well covered (e.g., by GOLD:GOLD Fitness or rDock:rDock_solv), rescoring (e.g., by iMDLScore2) could significantly affect the enrichment in virtual screening. As demonstrated in Fig. 4C and Table 4, we also found that the enrichment of virtual screening against 16S rRNA A-site was positively correlated with the weight of polar interactions (W_elec + W_hbond) when iMDLScore2 was used for rescoring. Therefore, we suggested that a soft-core rescoring function that favors electrostatic and hydrogen bond interactions (e.g. iMDLScore2) may be good for a flexible target such as the rRNA A-site.

Table 4.

ROC AUC for different docking and scoring function combinations in RNA virtual screening study.

Initial docking & scoring function	Rescoring function	Bacterial rRNA A-site (1J7T)			Lysine riboswitch (3DIL)
		Aminoglycosides 1	Zhou dataset 1	Foloppe dataset 1	7 known inhibitors 1	7 known inhibitors 2
GOLD:GOLD Fitness	None	1.0	1.0	0.58	0.97	0.82
rDock:rDock_solv	None	1.0	1.0	0.61	0.999	0.86
GOLD:GOLD Fitness	ASP	NA	NA	0.50	0.98	0.51
GOLD:GOLD Fitness	rDock_solv	NA	NA	0.68	0.998	0.86
GOLD:GOLD Fitness	AutoDock4.1 Score	NA	NA	0.64	0.77	NA
GOLD:GOLD Fitness	iMDLScore1	NA	NA	0.58	0.66	NA
GOLD:GOLD Fitness	iMDLScore2	NA	NA	0.69	0.92	0.51
rDock:rDock_solv	AutoDock4.1 Score	NA	NA	0.67	0.46	NA
rDock:rDock_solv	iMDLScore1	NA	NA	0.61	0.33	NA
rDock:rDock_solv	iMDLScore2	NA	NA	0.74	0.81	0.51

¹The decoy set was MayBridge dataset

²The decoy set was seven known lysine analogs inactive to lysine riboswitch

Figure 4.

ROC curves of the virtual screening experiments with various docking/scoring combinations. (A). Virtual screening against the 16S rRNA A-site using the Foloppe dataset. (B). Virtual screening against the lysine riboswitch using 7 known active compounds. (C-D). ROC comparison of the virtual screening performances of AutoDock4.1 and iMDLScore1/iMDLScore2 scoring functions with rRNA A-site (C). and lysine riboswitch (D). GOLD:GOLD Fitness dockings were in thin lines, while rDock:rDock_solv dockings were in thick lines. AutoDock4.1 default scoring function, iMDLScore1 and iMDLScore2 were colored red, blue and black, respectively.

In contrast, the results from virtual screening against rigid and closed lysine riboswitch were different. Based on the crystal structure, the ligand (lysine) is completely enveloped in the rigid binding pocket of lysine riboswitch, and only the small molecules which can sterically fit the pocket can be accommodated ^{51, 52}. B-factor analysis demonstrated that lysine-binding pocket was statistically less flexible than other residues (Supplementary Figure 4A). Normal mode analysis further confirmed the rigidity of this pocket as no significant atomic fluctuation of the active site residues (arrow highlighted) could be observed within five normal modes with the lowest frequencies (Supplementary Figure 4C). Our virtual screening showed that rDock:rDock_solv and GOLD:GOLD Fitness had comparable enrichments (Fig. 4B). GOLD:GOLD Fitness coupled with ASP or rDock_solv rescoring could improve the enrichment. Unfortunately, all AutoDock4.1 related scoring functions (default, iMDLScore1 and iMDLScore2) could not obtained as good enrichment (AUC<0.85) as other rescoring schemes (AUC>0.95). This might be due to the failure of assigning enough penalties to the steric clashes so that the big molecules were docked into the small and closed lysine-binding pocket (Fig. 4B). Additionally, we investigated whether any computational strategies could differentiate the seven known lysine riboswitch inhibitors from the seven experimentally validated lysine-analog decoys (very low chemical diversity compared to the above experiment with MayBridge decoys). Based on this dataset, we found that GOLD:GOLD Fitness combined with rDock_solv rescoring achieved the best enrichment (AUC=0.86) and ranked all seven active compounds on the top eight. As expected, iMDLScore2 rescoring achieved low enrichment (AUC=0.51) (Table 4).

When comparing our iMDLScore1/iMDLScore2 with the original AutoDock4.1 scoring function, we found that rescoring by iMDLScore2, which has the least vdW contribution but most electrostatic and hbond contributions, attained the best enrichment against 16S rRNA A-site with AUC=0.69 (based on GOLD:GOLD Fitness docking) and 0.74 (based on rDock:rDock:solv docking) (Table 4 and Fig. 4C). In contrast, the results from iMDLScore1 rescoring were not ideal. We observed the similar trend (iMDLScore2 (black) > AutoDock4.1 default (blue) > iMDLScore1 (red)) with ROC AUC analysis for the lysine riboswitch, as demonstrated in Table 4 and Fig. 4D. In summary, we identified the best combinations for RNA virtual screening: rDock:rDock_solv coupled with iMDLScore2 rescoring for flat, open and flexible binding sites of RNAs, while GOLD:GOLD Fitness combined with rDock_solv rescoring could be appropriate for closed and rigid RNA targets.

Application of ensemble docking/scoring for structural flexibility of RNAs

RNAs are flexible and ligand binding can significantly alter the backbone dihedral angles, which may in turn change the sugar puckering, narrow grooves, and introduce kinks or intercalation ⁵³. For example, when different ligands bind to HIV-1 TAR RNA, the sugar puckering of A22 can switch from C3′-endo (1UUD) to O4′-endo (1UUI), while U23 switch from C3′-endo (1UUD) to C1′-endo (1UUI). The technique of ensemble docking has been widely used in the field of protein to model flexibility with a limited number of discrete conformations ⁵⁴. The ensemble structures can be obtained from different x-ray crystal structures, NMR models or normal mode analysis ^{16, 55}. To investigate whether ensemble docking could also improve the docking/scoring accuracy in the RNA system, we collected five RNA targets and their alternative conformations in Table 5 for native pose ranking and cross-docking experiments. We found that, when RNA flexibility was modeled by structural ensemble, the results of the native pose ranking were significantly improved using rDock_solv scoring functions for 1LVJ, 1AM0 and 2Z74. As aforementioned, without considering RNA flexibility, we could not obtain reasonable results for these three targets in native pose ranking study. However, with other methods, except ASP, only slight improvement of ranking (<10) were observed. We found that the dihedral angle ε (C4′-C3′-O3′-P_i+1) of U23 (critical for inhibitor binding) in 1LVJ was statistically correlated with the native pose rankings from rDock_solv based on the Spearman's rank correlation (p<0.05) (Supplementary Table S8). It was probably because when ε increased, the phosphate of C24 had a better orientation to interact with the positively charge amine group of the ligands. Such interactions were correctly estimated by the rDock_solv scoring functions and thus obtained better rankings. In the cross-docking experiment, when the RNA flexibility was considered, the rankings of cognate ligand were significantly improved using GOLD:GOLD Fitness coupled with rDock_solv rescoring for both 1LVJ and 1UUD. Unfortunately, the ranking got slightly worse for 1LVJ and 1YKV when GOLD:GOLD Fitness coupled with AutoDock4.1 rescoring was used. Detailed comparative data are available in Supplementary Table S7. Our study implied that the consideration of RNA flexibility by ensemble docking could be beneficial to docking/scoring accuracy with hard-core scoring function (e.g. rDock_solv); however, it also depends on the RNA target and the docking/scoring methods used for modeling.

Table 5.

PDBs used to assess the ensemble docking. Primary RNA targets, which had been studied in native pose ranking and cross-docking, were underlined. Apo structures or the structures containing different ligand with the primary target were given in italic.

RNA	Ligand type	PDB ID
HIV-1 TAR RNA	PMZ (intercalator)	1LVJ (12 conformer models)
AMP aptamer	AMP	1AM0 (8 conformer models)
HIV-1 TAR RNA	P14 (guanidine-containing ligands)	1UUD, 1UUI, 1AKX
Diels-Alder ribozyme	DAI	1YKV, 1YLS, 1YKQ
T. tengcongensis glmS ribozyme	glucose 6-phosphate	2Z74, 2HO7, 2H0Z, 2Z75, 2GCV, 2H0W, 2GCS

Discussion

RNA represents a type of important targets for therapeutic development. In this study, we identified novel strategies for RNA-ligand modeling and virtual screening through a thorough and comprehensive evaluation with statistical analysis heavily used. We provided novel insights into four aspects associated with in silico structure-based molecular design against RNAs. First, GOLD:GOLD Fitness and rDock:rDock_solv, found as the best pose predictors for RNA targets, were appropriate for the initial binding mode generation. However, in order to accurately identify the cognate ligands for RNA targets or improve the virtual screening enrichment, an extra step of rescoring of the predicted binding modes is necessary. Second, we found that ASP and AutoDock4.1 scoring functions, rather than GOLD Fitness, were able to enrich the native or near-native poses. This was based on the observation that ASP rescoring could significantly improve the docking accuracy using GOLD:GOLD Fitness generated poses. Third, soft-core potentials (e.g. iMDLScore1 and iMDLScore2) and hard-core potentials (rDock_solv and ASP) should be properly employed based on the structural resolution and flexibility of the binding sites of RNA targets. Hard-core scoring functions could produce more accurate results for virtual screening using RNA ensemble structures. Finally, implementation of RNA-specific scoring function (e.g. iMDLScore2) improved the virtual screening enrichment as well as the accuracy of RNA-ligand binding affinity prediction. In summary, the suggested workflow for structure-based modeling for RNA-targeted drug discovery was illustrated in Fig. 6.

Consistent to other docking evaluation reports ⁵⁶, we suggest that good performance in binding mode reproduction do not guarantee the success in virtual screening for RNA inhibitor identification. Thereby we proposed a two-step docking/scoring strategy for RNA virtual screening. Although some scoring functions, such as iMDLScore2, can be used in the rescoring (the second step scoring) but are not ideal for the initial pose generation (the first step docking/scoring). We found when iMDLScore2 was used for the initial pose generation, positively charged groups, such as guanidium and amine, were biased to form interactions with the RNA phosphate moieties (Supplementary Figure 3). Based on our evaluation, no existing docking program achieved satisfactory performances on both pose prediction and hit identification. Our two-step strategy performed well by separating conformation-wise pose selection and ligand-wise hit selection using different docking and scoring functions during virtual screening.

The false positives in the selected hits are an overall consequence of incorrectly predicted binding modes and scoring inaccuracy. Here we highlighted the importance of sampling enough conformational space for the final rescoring, as illustrated in Fig. 1B. However, virtual screening involves not only the coverage of near-native conformations for each ligand, but also the enrichment of the true binders. Accordingly, we observed that retaining too many poses for rescoring might accumulate noises. For example, Fig. 5 illustrated the trend that, only when we picked up the top 3 poses based on GOLD:GOLD Fitness docking, could we achieve the best AUC with iMDLScore2 rescoring. The AUC gradually declined if the number of picked top poses were more than 3. The similar trend was also observed for rDock:rDock_solv docking, but the turning point of the number of picked poses was 6. We anticipate that by using GOLD:GOLD Fitness or rDock:rDock_solv as the initial pose generators, retaining top 10 poses for rescoring should be appropriate to cover at least one near-native binding pose without significantly compromising (< 3%) the virtual screening enrichment (Fig. 1B and Fig. 5).

Figure 5.

The ROC AUC (Y axis) against number of candidate poses (X axis) selected for iMDLScore2 rescoring against 16S rRNA A-site. (x025BE)represents the number of picked poses corresponding to the best ROC AUC (turning point).

Similar to protein-ligand interactions which are now considered as non-linear dynamic and cooperative processes ⁴⁶, distal thermodynamic changes may also alter the binding affinity in RNA systems significantly. For example, in SAM-I riboswitch, the mutations on three base pairs (3G×7) only slightly affect the binding modes, but significantly decrease the binding affinity (∼300 fold) (Supplementary Table S2). Existing scoring functions failed to reflect such dramatic change in binding affinity quantitatively according to the slight change of the structure. A better estimation of binding affinity via in silico approach may require refinement of RNA-specific force field, more advanced atom typing, and consideration of structural flexibility. For instance, an RNA-specific WSAS desolvation model was developed instead of inheriting Wesson & Eisenberg's model ^{43, 57}. In addition, the electrostatic potential in RNAs frequently is not dominated by the negatively charged phosphates because these phosphates can be masked by counterions and waters ²⁷. Therefore, we suggest that advanced scoring functions treat phosphate-ligand and base-ligand interactions differently. In particular, the H-bond formed between RNA bases and aromatic moieties should be considered as more favorable than that formed with phosphate. Rewarding short-range electrostatic and π-related interactions such as aromatic stacking, as rDock does ³², with base atoms could also be beneficial.

Finally, unlike proteins, RNA flexible docking is currently limited by the availability of rotamer libraries of nucleosides. Therefore, RNA flexibility are usually addressed by applying soft grids/potentials ⁵⁸, using ensembles with molecular dynamic simulation or various NMR models ¹⁶, or performing post-docking local optimization ³⁴. Since the force field-based optimization algorithms (e.g., MORDOR) are computationally expensive, more studies are needed to reduce the computational cost, probably by either accelerating the force field calculation or alternatively developing novel algorithms for more efficient sampling of ligand-bound conformations. For instance, Rohs et al. utilized Monte Carlo simulation to model flexible DNA-ligand interactions efficiently without any prior binding-site selections ⁵⁹, and this might be applied to RNAs as well. The successful rational design of HIV-1 TAR RNA inhibitors by using structure ensembles from normal mode analysis represented a good example for the ensemble docking ¹⁶. However, the performance of RNA ensemble docking varies depending on targets, scoring functions and other factors, in agreement with our assessment here. Sometimes it is possible to cause more noise rather than signals when flexibility is introduced ⁵⁴. Further exploration of RNA flexible docking/scoring will be necessary when more RNA structural data or more RNA flexibility-oriented docking programs become available.

Materials and Methods

Datasets

Most of the currently published datasets are either too small or lack target diversity ^{10, 28, 29, 31, 32, 34}. Based on these datasets, we compiled our own dataset of high-resolution RNA-ligand complex structures by removing those low-resolution, redundant structures as well as those structures with critical structural defects (e.g. 1AJU). We also included additional structures by mining Protein Data Bank (PDB) and literature to identify RNA-ligand complexes with available binding affinities. This resulted in a unique set of 56 RNA-ligand complex structures with 36 high-resolution (<3.0Å) crystal and 20 NMR structures. Another issue of the published datasets was that over 65% of the ligands were aminoglycosides or low-affinity binders (e.g. spermine) ²⁸. To avoid the potential problems of statistically overestimating the weight of any typical RNA ligands, we reduced the number of aminoglycosides and low-affinity binders, but increased the number of high-affinity small molecules. Our curation encompassed a variety of known RNA targets including: RNA aptamers, prokaryotic and eukaryotic rRNA A-sites, ribozymes, riboswitches, and viral RNAs (TAR RNA, HCV IRES domain, etc.). These RNAs are listed in Supplementary Table S1.

RNA docking and decoy generation

In this article, we used “A:B” to represent the strategy that “docking using A program and B scoring function”. RNA molecules and ligands were prepared using Protein Preparation Wizard in Maestro. Briefly, all water molecules, ions, and cofactors were removed. The hydrogen atoms were minimized in the ligand-bound complexes using default parameters. The average structure of NMR models for each complex was minimized and employed for our studies. All of the phosphates in RNAs were deprotonated. The ligands were protonated/deprotonated using Epik (Schrödinger) ⁶⁰ at PH 7.0. If an RNA has duplicate binding sites/ligands (e.g. 1J7T), the region with the lowest B-factors was used. The ligands were energetically minimized, and molecular docking and rescoring were performed using the similar approaches as previously described ^{28, 29, 32, 61}. Briefly, we employed five docking programs (GOLD 5.0.1, Glide 5.6, Surflex v2.415, AutoDock 4.1 and rDock 2006.2) combined with their native scoring functions to generate 10 poses using the default parameters, but with some modifications to ensure the sufficient conformational sampling. In order to ensure the high diversity and quality of the conformational decoys, we employed GOLD:GOLD Fitness to generate 100 conformational decoys for each RNA target with the tuned genetic algorithm parameters. Detailed docking parameters are available in Supplementary Material.

Reproduction of experimental structures with docking

In this study, both RMSD between experimental structures and predicted docking poses and pose ranking were considered. To simplify the expression, we define C(X, YÅ) as the criterion that “at least one near-native binding mode (<Y Å RMSD) was predicted within the top X poses”. To evaluate the overall ability of docking/scoring programs to reproduce experimental structures, we implemented a new parameter, namely volume under the surface (VUS), to describe the overall performance of binding mode reproduction. As Fig. 1a demonstrated, the surface was based on a series of discrete grids. VUS was calculated as the sum of the volume of all triangular prisms under this surface. Briefly, a series of coordinates were obtained based on their RMSD cutoff (the X dimension), ranking cutoff (the Y dimension), and the number (the Z dimension) of successfully reproduced structures satisfying C(X, YÅ). The spacing of RMSD cutoff was 0.5Å, and it was 1 for the ranking cutoff. The surface was made by connecting any two adjacent points and then partitioned into a series of triangles. Any of these triangles and their projections on the XY plane was used to define the triangular prism unit. Detailed calculation of the volume of each triangular prism unit and VUS were demonstrated in the Supplementary Material. The ideal VUS was calculated as 10(RMSD cutoff)×9(rank cutoff)×56(number of targets). Different from most of previous similar studies, if not all, statistical significance (p-value) was always considered for the analysis throughout this report. Briefly, p-values calculated from student t-test were employed to determine the statistical significance when comparing two different groups.

Docked/native pose ranking

For each ligand, we generated 100 decoys to the corresponding RNA target. Therefore, together with the crystal/NMR ligand structure, we obtained 101 RMSD-docking score data points for each cognate RNA-ligand pair. For native pose ranking study, we scored these 101 poses using different docking programs as aforementioned. The ranking of native poses for 56 targets were calculated, and the recovery curves were made as the ranking cutoffs (X axis) against the cumulative number of targets (Y axis) in which the ranking of native pose cutoff was smaller than the ranking cutoff. The Spearman's rank correlation coefficient was used to evaluate the ranking capability. To make the docking scores positively correlated with RMSD (the higher the scores, the higher the RMSD), we used the negative value of GOLD Fitness, ChemScore, ASP and Surflex-dock scores. If a pose was assigned a score with the absolute value more than 1000 (outliers), this RMSD-score pair will be excluded. The Spearman's rank correlation coefficient (ρ) was computed from Equation (1), where r_RMSD,i and r_score,i are the ranks of the RMSD and score for the pose i, and we took the average of the ranks for the tied values. $r_{\mathit{\text{RMSD}}}^{\mathit{\text{avg}}}$ and $r_{\mathit{\text{score}}}^{\mathit{\text{avg}}}$ are the average ranks of RMSD and score for 101 poses. We classified the resulted 56 ρ values (calculated from 56 RNA-ligand complexes) for each scoring function into three groups based on the widely-used criteria: weak correlation: ρ<0.3, moderate correlation: 0.3≤ρ<0.5, strong correlation: ρ≥0.5.

$$\rho =\frac{{\sum }_i(r_{\mathit{\text{RMSD}},i}-r_{\mathit{\text{RMSD}}}^{\mathit{\text{avg}}})(r_{\mathit{\text{score}},i}-r_{\mathit{\text{score}}}^{\mathit{\text{avg}}})}{\sqrt{{\sum }_i{(r_{\mathit{\text{RMSD}},i}-r_{\mathit{\text{RMSD}}}^{\mathit{\text{avg}}})}^2{\sum }_i{(r_{\mathit{\text{score}},i}-r_{\mathit{\text{score}}}^{\mathit{\text{avg}}})}^2}}$$ Equation (1)

Cross-docking

The dataset used for cross-docking was a subset of our overall collection. In order to improve the robustness of cross-docking, we only choose 23 complexes with unique RNAs and a limited number of non-selective ligands (e.g. arginine and aminoglycoside). The RNA-ligand complexes selected for cross-docking are available in Supplementary Table S6. Cross-docking was performed by docking all 23 ligands in this subset to each RNA targets using the methods described in “RNA docking and decoy generation”. Rescoring was performed based on the top 10 scored poses for each docking method. Cross-docking performance was evaluated by average ranking of cognate ligand calculated from these 23 RNA targets and AUC of the recovery curve based on the average ranking of cognate ligands (similar to the recovery curve used in “Docking pose ranking” section).

Virtual Screening for RNA inhibitor identification

Two different targets were assessed, namely bacterial 16S rRNA A-site (representing open and flexible binding site) and lysine riboswitch (representing closed and rigid binding site). Bacterial 16S rRNA A-site and lysine riboswitch structures were obtained from PDB as 1J7T and 3DIL, respectively. We collected 75 known rRNA inhibitors including 34 drug-like small molecules from the Foloppe dataset ¹² and 31 aminoglycoside mimetics from the Zhou dataset ¹³. Additionally, we obtained 11 aminoglycoside inhibitors which have the crystal structures in complex with the bacterial rRNA A-site (1J7T, 1YRJ, 2F4T, 2ET8, 1LC4, 1MWL, 2BE0, 2G5Q, 2ESI, 2PWT and 1BYJ). For virtual screening against lysine riboswitch, we collected 14 compounds including 7 known inhibitors and 7 experimentally validated inactives ¹⁷. In order to avoid artificial enrichment ⁶², a focused library containing 942 drug-like and positively charged decoys was generated from MayBridge database. We assumed this randomly constructed decoy library does not include or include very few active compounds as previous studies did ²⁸. The receiver operating characteristic (ROC) curve along with the area under the curve (AUC) was used to assess the overall performance (e.g., recovery rate, enrichment, etc.) of our protocols.

Docking score-binding affinity correlation

Although it is known that current docking scores are usually poorly correlated with experimental binding affinity ⁵⁶, we are still interested to evaluate the performance of our protocols on RNA targets and try to learn how to improve the scoring. To this end, dissociation constant (K_d) values were carefully collected from literature (values and references are available in Supplementary Table S2), and we compared them with PDBBind database (http://www.pdbbind-cn.org/) and other reports/databases to ensure the consistency of data. If the collected K_d values for each RNA-ligand pair are within 10-fold difference, the average values were used. Of note, we used 2μM as the K_d of gentamicin C1a-rRNA A-site complex (1BYJ) for studies because this is the K_d under room temperature, instead of 0.01μM (K_d under 4°C) ⁶³. Additionally, K_d for neomycinB-HIV-1 TAR RNA complex (1QD3) should be 5.9±4μM. The K_d values used in some previous studies were actually from the U24C mutant ^{29, 32, 64}. The binding free energy were calculated as ΔG = RTln(K_d) under 300K. The Pearson correlation coefficients (R²) between these docking scores and their corresponding binding affinities were obtained. Three common outliers, 1LVJ, 1TOB and 2TOB, were excluded to avoid noise, as they contained many unfavorable steric clashes.

RNA-specific scoring function optimization

The weak correlation between docking scores and binding affinity might be because most of the current scoring functions were derived from protein-ligand interactions. To implement new RNA-specific scoring functions, we optimized the energetic coefficients in AutoDock4.1 scoring function using our largest ever RNA-ligand dataset with known experimental binding affinities. This empirical scoring function was shown as Equation (2) ⁴¹. The parameters (A, B, C, D, S, V) were obtained from default AutoDock4 scoring function, as previously described ⁴¹. With our RNA-ligand dataset, we derived a new set of coefficients of five terms (W_vdW, W_hbond, W_elec, W_sol and W_tors) with multiple linear regression methods. Besides R², we calculated leave-one-out (LOO) cross-validation correlation coefficients (Q²) and validated against an external test set consisting of eight complexes to evaluate the predictive power of our new schemes. The PDBs employed to train and validate iMDLScore1 and iMDLScore2 are listed in Supplementary Table S2.

$$\Delta G_{\mathit{\text{bind}}}=W_{\mathit{\text{vdw}}}\sum _{i,j}(\frac{A_{ij}}{r_{ij}^{12}}-\frac{B_{ij}}{r_{ij}^6})+W_{\mathit{\text{hbond}}}\sum _{i,j}\xi (t)(\frac{C_{ij}}{r_{ij}^{12}}-\frac{D_{ij}}{r_{ij}^{10}})+W_{\mathit{\text{elec}}}\sum _{i,j}\frac{q_i{q}_j}{\varepsilon (r_{ij})r_{ij}}+W_{\mathit{\text{sol}}}\sum _{i,j}(S_i{V}_j+S_j{V}_i)e^{\frac{-r_{ij}^2}{2{\sigma }^2}}+W_{\mathit{\text{tors}}}{N}_{\mathit{\text{tors}}}$$ Equation (2)