VfoldLA: a web server for loop assembly-based prediction of putative 3D RNA structures

Abstract

RNA three-dimensional (3D) structures are critical for RNA cellular functions. However, structure prediction for large and complex RNAs remains a challenge, which hampers our understanding of RNA structure-function relationship. We here report a new web server, the VfoldLA server (http://rna.physics.missouri.edu/vfoldLA), for the prediction of RNA 3D structures from nucleotide sequences and base-pair information (2D structure). This server is based on the recently developed VfoldLA, a model that classifies the single-stranded loops (junctions) into four different types and according to the loop-helix connections, assembles RNA 3D structures from the loop/junction templates. The VfoldLA web server provides a user-friendly online interface for a fully automated prediction of putative 3D RNA structures using VfoldLA. With a single-RNA or RNA-RNA complex sequence and 2D structure as input, the server generates structure(s) with the JSmol visualization along with a downloadable PDB file. The output result may serve as useful scaffolds for future structure refinement studies.

Graphical Abstract

Introduction

RNA function in many biological processes requires the formation of proper three-dimensional (3D) structures (Eddy, 2001; He and Hannon, 2004; Clancy, 2008). The knowledge about RNA structures helps us not only understand the physical mechanism of RNA functions but also develop RNA structure-based therapeutic design. However, compared with the determination of RNA primary (sequence) (Wang et al., 2009; Chu and Corey, 2012) and secondary (two-dimensional, 2D) structures (Felden, 2007; Ishitani et al., 2008; Butcher and Pyle, 2011; Xu and Culver, 2013), RNA 3D structures are difficult to determine experimentally, highlighting the importance of computational prediction for RNA 3D structures.

Owing to the development of sampling algorithms and accurate energy functions, current computational models can predict structures for short RNAs from the sequence (Das and Baker, 2007; Das et al., 2010; Tan et al., 2006; Sharma et al., 2008; Jonikas et al., 2009; Xia et al., 2010; Pasquali and Derreumaux, 2010; Izzo et al., 2011; Kim et al., 2014; Shi et al., 2014; Kerpedjiev et al., 2015; Boniecki et al., 2016; Zhang and Chen, 2018). For example, guided by a knowledge-based energy function that encodes base stacking and pairing interactions, FARNA/FARFAR (Das and Baker, 2007; Das et al., 2010) predicts RNA 3D structures by assembling short fragments extracted from a single crystal structure. The method can predict structures with atomic resolution for short (< 30 nt) RNAs. iFoldRNA (Sharma et al., 2008) uses a three-bead model to predict RNA 3D structures by applying discrete molecular dynamics (DMD) simulations. Most of the predicted structures among the tested RNAs have an RMSD of < 4 Å to the native RNA structures. The usage of coarsegrained representation and DMD approach significantly reduces the computational complexity and enhances the comformational sampling.

For large RNAs, additional information, such as contact maps for tertiary contacts and base pairing patterns (2D structures), is needed as structural constraints in order to effectively reduce the conformational search space (Parisien and Major, 2008; Zhang et al., 2009; Cao and Chen, 2011; Rother et al., 2011; Reinharz et al., 2012; Popenda et al., 2012; Yamasaki et al., 2012; Zhao et al., 2012; Xu et al., 2014). MC-sym, as the second part of MC-fold and MC-sym pipeline (Parisien and Major, 2008), builds all-atom structures using nucleotide cyclic motif (NCM) fragments, which are not only extracted from known RNA 3D structures but also built on the fly if necessary. Vfold3D (Cao and Chen, 2011; Xu et al., 2014) is a template-based coarse-grained structure prediction model. The model divides a 2D structure into different motifs and assembles the 3D motif templates from known RNA structures to predict 3D structures. With a similar general approach, RNAcomposer (Popenda et al., 2012) treats secondary structural motifs in a more detailed way by using slightly distorted helical conformations and tertiary contacts in secondary structure motifs. On the basis of the machine translation principle, the model enables prediction of large RNA 3D structures with relatively high prediction quality. However, the template-based approaches for RNA 3D structure prediction are severely limited by the low success rate of finding a proper template for a given motif, especially for the motifs with cross-linked contacts.

We have recently developed VfoldLA (Xu and Chen, 2018), a new model for the assembly of RNA 3D structures using loop templates. The model divides a 2D structure into helices and single-stranded loops (junctions), and classifies the loops into four different types. According to the loop-helix connections, the model assembles RNA 3D structures from loop templates. This new template search strategy leads to a much higher success rate of finding appropriate templates than the previous motif-based approach in Vfold3D (Cao and Chen, 2011; Xu et al., 2014), and the computational efficiency is higher than fragment-based approaches such as FARNA/FARFAR (Das and Baker, 2007; Das et al., 2010). The method provides low-resolution predictions for RNA structures at different levels of complexities, such as structures with cross-linked contacts. Here, we introduce a web server that offers a user-friendly interface for a fully automated prediction of putative RNA 3D structures using VfoldLA. With sequence and 2D structure as the input, the VfoldLA server outputs putative 3D structures of RNAs and RNA-RNA complexes, which may serve as useful scaffolds for further structure refinement studies.

Algorithm and methods

Loop template database

We dissect an RNA 2D structure into helices and single-stranded loops. According to the different loop-helix connection modes, we define four types of loops, as shown in Fig. 1A: “helix2” loop for a single strand that connects two helices; “hairpin” loop as a special “helix2” loop whose two ends connect to the same helix; “tail5” and “tail3” loops for segments of unpaired nucleotides at 5’ and 3’ ends, respectively. The length of each loop L is the number of unpaired nucleotides. To account for the helix-loop structural interference, we include the loop-connected terminal base pair(s) in the helix (helices) in the definition of the four loop types (see the red oval shape in Fig. 1A).

Figure 1:

(A) Illustration of four types of loops. Cylinders represent RNA helices. The red oval shape in each helix denotes the terminal base pair attached to unpaired nucleotides (line shown in red). (B) Format of sequences for the different loop types. W-W and N denote terminal base pair and unpaired nucleotide, respectively. For example,W-5’W and W3’-Ware the base pairs connected to the 5’ and the 3’ ends of the loop, respectively. (C) Example of the sequence similarity-based loop score. Q and S are the query sequence and subject loop sequence in database, respectively.

From the RNA 3D structure database in the Protein Data Bank (PDB), VfoldLA extracts the 3D templates for the four types of loops with different loop lengths ☹, along with the corresponding loop sequences (see the sequence format shown in Fig. 1B), and builds a non-redundant loop template database. The detailed method for building the database and the statistics of the current loop template database can be found in Ref (Xu and Chen, 2018). With the increasing PDB database for the known RNA 3D structures, the loop template database will be updated periodically.

Sequence similarity-based loop scoring

The scoring scheme considers the similarity of nucleotides, assuming that less changes in base substitution can result in less changes in the local and nonlocal interactions in the loop and hence less changes in the loop structure. Given a query loop sequence, VfoldLA scores the sequence similarity for a loop template according to the following rule: The score s_i for nucleotide position i in the sequence is equal to 0 for perfect match, 1 for purine/pyrimidine-type match, and 2 for no match; the loop score is defined as the sum over all of the nucleotides in the loop S^loop = ∑_is_i. As shown in Fig. 1C as an example, the query loop is a helix2 loop with four unpaired nucleotides (UCGG) and two helix-terminal base pairs (G-5’C and C3’-G). To avoid possible double counting, we only take into account the sequence similarity for all unpaired nucleotides and one of the two nucleotides in each terminal base pair (within the 5’−3’ loop fragment).

Loop template-directed structure prediction

Unlike the previous models which search for templates based on the whole motif or piece-wise fragments, VfoldLA searches for templates for each single-stranded loop and assembles optimal loop templates (ranked by the loop scores) and helices into 3D structures. As shown in the workflow in Fig. 2 (details in Ref (Xu and Chen, 2018)), the VfoldLA model works with the following steps.

Figure 2:

The workflow of the VfoldLA algorithm.

❿ From the input 2D structure, we first categorize the loops into three types: (I) N−1 “helix2” loops, which are used to assemble (N-helix) helix orientations. We consider volume exclusion between the helices. (II) The rest “helix2” loops for loop closure; (III) All “hairpin”, “tail5”, and “tail3” loops. We account for volume exclusion between the helices in the helix assembly process.

❿ The 3D templates of all the loops are ranked according to the sequence similarity-based scoring scheme. The optimal templates are selected from those with the best loop similarity scores.

❿ The hierarchical loop template assembly, as illustrated in the step-by-step assembly of the three classes of the loops (steps 3, 5, 6) in Fig. 2, allows VfoldLA to predict the global fold (helix orientation) of RNA structures.

❿ The templates for loops I, and III are chosen based on the loop type, loop size, sequence-base loop score. VfoldLA allows a 5 Å heavy-atom RMSD of the helix terminal base pairs as an additional criteria for the selection of templates for the loops II (loop-closure).

Structural refinement

The above structures may contain structural clash and non-ideal bond lengths, bond angles, or bond torsional angles, especially at the helix-loop junction regions. Therefore, the assembled structures need further refinement. To avoid long computational time, the VfoldLA server utilizes the recently developed IsRNA model (Zhang and Chen, 2018), which employs a coarse-grained representation of RNA conformations, knowledge-based interaction potentials, and Replica-Exchange Molecular Dynamics (REMD) simulations, to refine the all-atom structures assembled by the VfoldLA model.

In the IsRNA model, the total force field consists of two types of energy terms: the local interactions, which depend on the local RNA structure, and the nonlocal terms, which describe the interactions between non-neighboring nucleotide residues: E_total = E_bond + E_angle + E_torsion + E_pair + E_ele + E_LJ. Here, E_bond, E_angle, E_torsion are the local energy terms, representing the bond stretching, bending, and torsional torsional energies between three successive bonds, respectively. The nonlocal term E_pair accounts for base-base contacts, including base pairing, base stacking, and base-backbone interactions. The electrostatic term E_ele accounts for backbone-backbone electrostatic interactions. E_LJ describes the excluded volume interaction between any two nonbonded coarse-grained beads. Details of the energy terms, including the functional forms and parameters, are given in Ref (Zhang and Chen, 2018). The IsRNA can resolve most of the structural clash while keeping the overall 3D fold unchanged through small-RMSD conformational sampling.

Results and discussion

Input

As shown in Fig. 3A, upon visiting the webpage, the user is presented with a form to input the sequence and 2D structure for the RNA or RNA-RNA complex (see Tab. 1 for the detailed input format), along with other job information, such as the number of clusters, RMSD cutoff for clustering, job name, and the user’s email address (optional, for sending the output results).

Figure 3:

The user interface of the VfoldLA web server: The web pages for job submission (A) and job information (B).

Table 1:

Examples of sequences and 2D structures for VfoldLA server. Use “-” to separate RNA strands for RNA complex systems.

RNA system	Sequence/2D structure (in dot-bracket format)
Single-RNA	GGAGGUAGUAGGUCGAAAGACCGUUCUACACUCC ((((((((..((((....))))...)))).))))
Two-RNA complex	UACUAACGUAGUA-UACUAACGUAGUA ((((.((((((((-)))).))))))))
Three-RNA complex	GGGCGGCC-GGCUAGACGGUGGGAGAGGC-GCUGGUCCACCCGUGACGCUC ((((.(((-)))...((((((((...(((-)))..))))).)))...))))
With cross-linked contacts	GGAACCGCGAAAGCGGUUCCACGACGAUACUUAUUUCCUUUGAUCGUCGUUAUUACUGGCUUCGGCCACAAAGGAGA ((((((((....))))))))((((((((.......[[[[[..))))))))......((((....))))..]]]]]..

After the job is submitted, the server shows a web page (see Fig. 3B) that displays the basic information of the job, such as the input parameters and the job status. Each submitted job has a unique job ID (marked by blue circle in Fig. 3B), containing the job name provided by the user and a 4-letter unique job suffix randomly generated by the server. User can also check the status of the submitted job by the job ID, as shown the blue arrow in Fig. 3.

Output

The VfoldLA web server terminates jobs when the computational time exceeds 12 hours or the number of assembled structures reaches 100, which may lead to no predictions for some submitted jobs. The server classifies the assembled structures into clusters according to the sequence similarity-based total scores and the input cluster parameters, namely, the RMSD cutoff for clustering (5 Å in this case). The top N clusters are selected and the centroid structures of selected clusters are output as the VfoldLA-predicted putative structures. Here, N is the input number of clusters (5 in this case). As shown in Fig. 4A, along with the job information of input data (Fig. 4B, which matches the information from Fig. 3B), number of assembled structures (100 in this case), and the total number of clusters (38 in this case), the server provides a visualization of the predicted structures in JSmol applet, along with a downloadable PDB file for the predicted (top 5) structures.

Figure 4:

(A) The web page that shows the results. (B) The input parameters.

Computational efficiency

VfoldLA is loop template directed 3D structure prediction model, which dissects whole motifs into loops. The computational time of VfoldLA vaires, depending on the availability of loop templates, the size of input sequence, and the complexity of input 2D structure. The benchmark test (Xu and Chen, 2018) shows that the computational time ranges from minutes to hours. However, due to the large number of combinations for the different templates of different loops, the computation is more efficient for structures containing small number of helices (< 8 helices) and up to 10 strands for RNA complex systems to avoid excessively long computational times (due to the large number of combinations for the different templates of the different loops).

Application examples of the web server

In addition, we use the following three examples with different structural motifs: stem-loop, multi-branched, and cross-linked, to illustrate the usage of the VfoldLA web server for RNA 3D structure prediction. We use the default parameters for the number of clusters (5) and the RMSD cutoff for clustering (5 Å) to submit VfoldLA jobs, and compare the predicted structures with their corresponding native structures.

I. Stem-loop structure: the shortened SRP RNA from E. coli

The shortened SRP RNA from E. Coli (Voigts-Hoffmann et al., 2013) has the sequence of 5’UGUUGGUUCUCCCGCAACGCGGAAGCGUGUGCCGGGAUGUAGCUGGCA3’, and the native 2D structure in dot-bracket of ((((((((.(((((((((((....)))).))).))))...)))))))), which is a stem-loop structure with two bulge loops and one internal loop. VfoldLA generates 100 assembled structures and 9 clusters according to the input clustering parameters. From the top 5 cluster centers, we find that the best prediction, as shown in Fig. 5(I) has the RMSD of 6.2 Å to its native structure (PDB: 4c7o).

Figure 5:

Application examples of the VfoldLA web server: (A) The shortened SRP RNA from E. coli. (B) Glycine riboswitch. (C) The precursor-form HDV ribozyme.

II. Multi-branched structure: glycine riboswitch

The glycine riboswitch (Huang et al., 2010) has the sequence of 5’CUCUGGAGAGAACCGUUUAAUCGGUCGCCGAAGGAGCAAGCUCUGCGCAUAUGCAGAGUGAAACUCUCAGGCAAAAGGACAGAG3’, and the native 2D structure in dot-bracketof (((((......((((......)))).(((....(((... ((((((((....))))))))....)))...))).......))))), which is a multi-branched structure with a three-way junction and two internal loops. VfoldLA generates 100 assembled structures and 46 clusters. From the top 5 cluster centers, we find that the best prediction, as shown in Fig. 5(II) has the RMSD of 7.9 Å to its native structure (PDB: 3owz).

III. Cross-linked structure: the precursor-form HDV ribozyme

The precursor-form HDV ribozyme (Ke et al., 2004) has the sequence of 5’UGGCCGGCAUGGUCCCAGCCUCCUCGCUGGCGCCGGCUGGGCAACACCAUUG CACUCCGGUGGUGAAUGGGAC3’, and the native 2D structure in dot-bracket of .(((((((...[[[[[[(((. [[.....))))))))))]]....((((..........)))).....]]]]]], which contains the cross-linked contacts involving hairpin-junction and internal-tail base pairs. VfoldLA generates 100 assembled structures and 43 clusters. From the top 5 cluster centers, we find that the best prediction, as shown in Fig. 5(III) has the RMSD of 6.8 Å to its native structure (PDB: 1vbx).

For each given set of helix configuration, the VfoldLA model predicts one all-atom structure with the optimal score. This could lead to less accurate predictions for local structures, especially for the base orientations of unpaired nucleotides. RNA solution structures, especially multi-branched junctions, can show multiple helix orientations, and a crystal structure may represent only one of the several structures in flexible ensemble. Therefore, predicting multiple RNA conformations is important for understanding RNA folding and functions, which often involve multiple conformational switches.

Conclusions

Using a loop template-based algorithm, VfoldLA web server predicts putative 3D structures for RNAs from the sequence and 2D structure. The server can model structures of singles RNAs as well as RNA-RNA complexes at the different levels of structural complexities. The predicted structure can provide useful insights into the formation of RNA global fold including helix orientations. Furthermore, the predicted putative folds may be used as low-resolution initial structures for further predictions of high-resolution structures. This multistate (from low- to high-resolution) approach can potentially offer a useful way to accurate prediction of RNA structures, especially for structures with cross-linked contacts such as inter-loop base pairs.

Due to the large number of combinations for the different templates of the loops, the current model can only handle RNAs containing a small number of helices in order to avoid excessively long computational times. To enhance the computational efficiency, future development of the model could use a hybrid approach as the following. First, for motifs with available templates, we build their structures according to their templates. Second, for motifs with no available templates, we build their structures using VfoldLA. Finally, we assemble the different parts together to build the whole structure. Such a hybrid approach can integrate the advantages of the motif- and loop/fragment-based assembly methods and can be computationally efficient.

A new method to significantly enhance the availability of RNA structural templates

A new web interface for template-based RNA 3D structure prediction

A novel efficient algorithm for constructing putative RNA 3D scaffolds