RNA-MoIP: prediction of RNA secondary

RNA-MoIP: prediction of RNA secondary
W440–W444 Nucleic Acids Research, 2017, Vol. 45, Web Server issue
doi: 10.1093/nar/gkx429
Published online 19 May 2017
RNA-MoIP: prediction of RNA secondary structure and
local 3D motifs from sequence data
Jason Yao1 , Vladimir Reinharz2 , François Major3 and Jérôme Waldispühl1,*
School of Computer Science, McGill University, 3480 University Street, Montreal, QC H3A 0E9, Canada,
Department of Computer Science, Ben-Gurion University of the Negev, Beer-Sheva 84105, Israel and 3 Institute for
Research in Immunology and Cancer and Department of Computer Science and Operations Research, Université de
Montréal, Montreal, QC H3C 3J7, Canada
Received March 03, 2017; Revised April 23, 2017; Editorial Decision May 02, 2017; Accepted May 12, 2017
RNA structures are hierarchically organized. The
secondary structure is articulated around sophisticated local three-dimensional (3D) motifs shaping
the full 3D architecture of the molecule. Recent contributions have identified and organized recurrent local 3D motifs, but applications of this knowledge for
predictive purposes is still in its infancy. We recently
developed a computational framework, named RNAMoIP, to reconcile RNA secondary structure and local 3D motif information available in databases. In
this paper, we introduce a web service using our
software for predicting RNA hybrid 2D–3D structures
from sequence data only. Optionally, it can be used
for (i) local 3D motif prediction or (ii) the refinement of
user-defined secondary structures. Importantly, our
web server automatically generates a script for the
MC-Sym software, which can be immediately used to
quickly predict all-atom RNA 3D models. The web
server is available at http://rnamoip.cs.mcgill.ca.
RNA folding is hierarchical. The secondary structures
(Watson–Crick and Wobble base pairs) form rapidly, acting
as a scaffold for the slower formation of three-dimensional
(3D) structures (1,2). This observation has already been successfully used by previous software (3–7), which use secondary structure information to assist in tertiary structure
prediction. Alternate approaches fully based on molecular
dynamics solutions have also been proposed (8). In general, it is worth noting that the accuracy of all prediction
methods significantly decreases as the length of the RNA
sequence increases.
Because of their versatility and reliability, fragment assembly methods have become a popular strategy to predict RNA 3D structures (3,4,6,9). These assemble diverse
fragments of known 3D structures into complete composite
* To
3D structures. Among them, the MC-Pipeline (3), which is
composed of MC-fold for predicting secondary structure
from sequence and MC-Sym for assembling 3D structures
from a sequence and structural constraints, has been successfully used in multiple contexts. These include the refinement of diffraction data and model fitting (10), and model
screening in pharmaceutical applications (11). The present
work aims to boost its performance and ease its utilization.
Our approach builds upon recent advances in RNA 3D
structure analysis. By combining the hierarchical folding
properties with the large public database of experimentally
resolved 3D motifs from RNA3dMotif (12), we have developed a fast and accurate hybrid method, named RNA-MoIP
(RNA Motifs over Integer Programming), which uses an
integer programming framework to predict RNA 2D–3D
structures (i.e. secondary structures in which loop regions
are annotated with local 3D structures) from sequence data
alone (13). This predicted set of candidate 2D–3D structures can then be used as a template in the MC-Sym software
to generate complete 3D structures.
To generate 2D–3D structures, RNA-MoIP uses RNA
secondary structure templates (generated by default with
RNAsubopt, from the ViennaRNA package (14)) in which
it inserts candidate RNA 3D motifs. A distinctive aspect
of RNA-MoIP lies in its ability to use a set of approximate
candidate secondary structures instead of a single highly
reliable template (13). It identifies the most promising secondary structures and eventually improve them by removing incorrectly predicted base pairs. It follows that RNAMoIP can also be used to refine secondary structure predictions.
Currently, RNA-MoIP exists as a command-line script
that accepts as minimum input a primary sequence and a
pool of candidate secondary structures generated by RNAsubopt. However, this implementation requires users to
have supporting software installed on their machine. RNAsubopt, which generates the candidate secondary structures, and the Gurobi Optimizer, which contains a Python
interpreter and integer programming solver for which RNAMoIP is written, are needed. It also requires the manual
whom correspondence should be addressed. Tel: +1 514 398 5018; Fax: +1 514 398 3883; Email: jerome.waldispuhl@mcgill.ca
C The Author(s) 2017. Published by Oxford University Press on behalf of Nucleic Acids Research.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by-nc/4.0/), which
permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact
Downloaded from https://academic.oup.com/nar/article-abstract/45/W1/W440/3835309/RNA-MoIP-prediction-of-RNA-secondary-structure-and
by guest
on 16 September 2017
Nucleic Acids Research, 2017, Vol. 45, Web Server issue W441
transfer of results between RNAsubopt, RNA-MoIP and
MC-Sym. Limitations of the motif database also complicate the process of assembling the motifs suggested by RNAMoIP into an MC-Sym readable format. Missing structural
data (attributable to experimental factors such as crystallography resolution) and gaps in motif coverage make
the preparation of scripts for MC-Sym a time consuming
process. This limited the utility of RNA-MoIP for multisequence processes and made the utility inaccessible to
many users.
The goal of this server was to develop a computational
pipeline to allow users to automatically generate MC-Sym
input scripts with the optimizations provided by RNAMoIP. We developed an end-to-end solution to automate
the process such that a user can directly generate a pool
of predicted RNA tertiary structures (output by MC-Sym)
from any single primary sequence input alone. Additionally,
we implemented tools to visualize the motif insertions and
ease interpretation and selection from the candidate solutions. Their locations in the secondary structure are directly
shown in an annotated schema and an additional interface
allows users to visualize the 3D atomic model of each instance. In the following sections, the RNA-MoIP web server
and its methods are described.
The RNA-MoIP web server is available at http://rnamoip.
cs.mcgill.ca/ and runs on an Ubuntu Server, on a Dell PE
T610 2x Intel Quad core X5570 Xeon Processor, 2.93 GHz
8M Cache, 64 GB Memory (8 × 8 GB), 1333 MHz Dual
Ranked RDIMMs for 15 Processors.
The back-end is written in Python 2.7.3 and bash. For
each task one processor is allocated and the Gurobi optimizer v.6.5.0 (7) API for Python is used to solve the integer
programming equations. Each job is assigned a unique directory on the server and all data is preserved for up to 2
The front-end is designed using the Bootstrap 3.3.7 css
framework. Web pages are generated using Python 2.7.3
standard library cgi module and utilize Javascript.
The minimal input to perform a query is an RNA sequence.
Up to five sequences in the FASTA format can be provided
simultaneously and all results will be presented in the same
An ensemble of additional constraints can be provided
by the users. Those are available under the collapsible
Advanced Options menu. The options are (i) the maximal fraction of base pairs in the secondary structure that
can be removed to accommodate the insertion of motifs
(default 30%), (ii) the largest number of components––or
strands––in a motif that can be inserted (default 3, or three
way junctions, only up to four way junctions is available
now) and (iii) the maximum number of solutions to output
(default 1). This is the number of optimal solutions per secondary structure that are returned. If different secondary
structures achieve solutions with the same score all of them
will be displayed. Additionally, by default, candidate sec-
ondary structures are computed with RNAsubopt, sampling all structures within 3 KCal/mol from the minimal
free energy structure. A stochastic sampling can also be selected, or a set of secondary structures in the dot bracket
notation can be provided by the user.
A name can be assigned to a new job, or a random ID will
be provided, that can be used to jump directly to the working directory. In the top navigation bar, there is a search
menu that users can input either their job name, or the job
ID that is randomly assigned and associated with the created directory. This information is provided to the user immediately after the job has been submitted. In addition, a
general help page is available with explanations for each of
the options, as well as an about page with references and
author contact information.
The main window of the output is shown in Figure 1. Upon
job submission, a working directory is generated that can be
accessed by Job ID or job name (if provided). Each working directory is maintained on the server for a minimum
of 2 weeks. The directory consists of four main elements:
tabs to view the different optimal structures and their solutions, links to the assembled MC-Sym script(s) with an option to automatically submit, a diagram of the secondary
structure selected by RNA-MoIP indicating the position of
the inserted motifs and a table with links to view interactive
3D models of those inserted motifs. It is important to note
that many competing structures can have the same optimal
value in the IP framework (13), in which case each of them
will be displayed.
The working directory is composed of a tabbed main navigation element and a sidebar linking to other directories for
batch (.fasta) uploaded jobs. In the main navigation window, a link allows the script to be automatically submitted
to the MC-Sym server, where the user is then directed to the
MC-Sym control panel for the submitted job. Additionally,
users can view details about the selected secondary structure
and motifs inserted. If a maximum of more than one solution was requested, the solutions can be compared using
the tabbed navigation. For each solution, a secondary structure diagram is created dynamically using VARNA, with
non-canonical interactions annotated in the inserted motifs
(15). This allows for an easy graphical comparison between
candidates in the set of solutions. Motifs are inserted with
sequence-based constraints; thus a set of valid 3D structures
can be assigned to to the same location. Below, users can
view details of the motifs selected by RNA-MoIP, including
the bases it spans and a link to visualize or download the
PDB file of each valid instance. For each inserted motif, a
specific candidate atomic structure can be selected from a
list and visualized in the browser using JSmol. A direct link
to download the PDB file of each instance of the inserted
motif is also available.
The RNA-MoIP web server consists mainly of the RNAMoIP program with a set of new tools easing its use, analysis of the output and interface with MC-Sym for an allatoms prediction. Here we describe its main components:
Downloaded from https://academic.oup.com/nar/article-abstract/45/W1/W440/3835309/RNA-MoIP-prediction-of-RNA-secondary-structure-and
by guest
on 16 September 2017
W442 Nucleic Acids Research, 2017, Vol. 45, Web Server issue
Figure 1. The results screen of the RNA-MoIP web server for a fasta input. Shown is the output for the Escherichia coli thi-box riboswitch 2HOJ. In the
main dashboard, the corrected secondary structure is shown with a 2D secondary structure visualization generated from VARNA. Multiple solutions for
each sequence can be toggled using the tabbed navigation. In the side navigation bar users can access other sequences if submitted in the same .fasta file.
the secondary structure generation, RNA-MoIP, the automatic generation of MC-Sym scripts and the visualization
system. The automatic script generation is a novel approach
to systematize the strategy in (16) which can be quite involved even for small RNAs with few motifs. This new procedure allows a seamless transition from the RNA-MoIP
2D–3D structure to to the full-atom predictions of MC-Sym.
Secondary structures generation
For each sequence, if no secondary structure is provided,
the web server provides two ways to generate a pool of secondary structures. In both cases RNAsubopt is used. RNAsubopt can either list all structures at an energy range from
the minimal free energy structure, or use a stochastic sampling approach. By default, structures are sampled using the
energy range model at 3 KCal/mol. Under this setting, at
least 10 or more RNAs are typically sampled, although the
sample size may drastically increase for longer sequences.
The range can be modified by the user. In the stochastic
model, secondary structures are sampled with probability
equal to their Boltzmann weight in the ensemble. The num-
ber of sampled structures is 25 by default and can be modified by the user.
The size and makeup of the pool of candidates can significantly impact predictions made by the server. We have
previously found that low ranking structures can result in
models with the highest accuracy, measured by RMSD, after refinements by RNA-MoIP (13). This stresses the importance of optimizing efficiency versus accuracy of the model
at the secondary structure level. A large sample is necessary
to provide RNA-MoIP greater flexibility in model selection.
RNA-MoIP is an integer programming framework that refines secondary structure predictions to accommodate the
insertion of RNA 3D motifs by removing base pairs. Motifs are structural units shared across many different RNA
molecules including hairpins, internal loops and k-way
junctions. Between different RNAs, they are expected to
share a similar local 3D structure based on folding interactions at the primary and secondary structure level. In the
RNA-MoIP framework, motifs are inserted based only on
Downloaded from https://academic.oup.com/nar/article-abstract/45/W1/W440/3835309/RNA-MoIP-prediction-of-RNA-secondary-structure-and
by guest
on 16 September 2017
Nucleic Acids Research, 2017, Vol. 45, Web Server issue W443
sequence compatibility. The motifs database is populated
by RNA3dMotif (12), which extracts them from the structures in the Protein Data Bank. Currently, 4695 motifs are
indexed as part of the RNA-MoIP database. We have previously shown that by incorporating publicly accessible 3D
motif data into structural predictions, we can improve the
accuracy and running time of our models compared to MCSym alone.
Given a sequence and secondary structure, RNA-MoIP
performs two functions in parallel: (i) refinement of secondary structures by removing base pairs to accommodate
the insertion of database motifs and (ii) insertion of motifs
from the database.
A fraction of base pairs from the given secondary structure can be removed to accommodate the insertion of motifs. This has a 2-fold advantage of enhancing the flexibility
of the secondary structure to accept motifs, and improving
the accuracy of secondary structure by removing incorrectly
predicted base pairs. The IP framework aims to balance the
cost of removing predicted canonical-base pairs with the insertions of motifs. It additionally prioritizes the insertion
of large motifs instead multiple small ones, aiming to maximize coverage while maintaining the general structure of the
original secondary structure prediction. The IP framework
assigns linear penalties for base-pair deletions and scores
motifs based on the square of the component length.
Formally, RNA-MoIP minimizes the following function,
given an ensemble of base pairs D, an ensemble of motifs
Mot, the total length of a motif Mx and all its occurrences
C. A score is given as follows:
x,1 ⎟
x 2
10 ∗
Du,v −
⎝(|M |) ·
⎠ (1)
x∈Mot j
Each pair of sequence structures gets a score. RNA-MoIP
selects the structures with the minimal score as the most
probable solution. The output provides the secondary structure(s) it has selected from the given input pool, a list of
motifs it has inserted and a list of bases that have been removed.
MC-Sym and input script assembly
To generate full-atom predictions, we offer a direct submission to the MC-Sym web server. MC-Sym requires the instructions to perform the prediction in a specific format,
.mcc, which is automatically generated and provided to the
user. We describe here how the script for MC-Sym is produced.
The script consists of the following main sections:
• Sequence: the RNA sequence
• Library: list of fragments used to assembly the RNA secondary structure. Includes motifs, stems and links
• Backtrack: the order in which fragments are joined. Fragments must be merged contiguously, overlapping with a
previously placed fragment that serves as an ‘anchor’
• Relation: defines dangling at the 5 and 3 ends separately
from the rest of the model
MC-Sym assembles RNA fragments. Each must be individually defined in the library section, and are assembled in
an overlapping fashion. In other words, newly placed fragments must overlap with at least one residue from a previously placed fragment. The only exceptions are the dangling ends which are defined in the Relation section and are
treated independently by MC-Sym. The quality and feasibility of the full-atom prediction is greatly influenced by the
order in which the fragments are merged.
There are three main type of fragments. First stacked
canonical base pairs are used to build the stems. Second
links, which indicate that a certain nucleotide follows another, are used to define nucleotides with less constraints.
Finally motifs, which were previously inserted by RNAMoIP.
To limit the size of the conformational space early on and
to improve the accuracy of the model, the largest fragment
is positioned first. This has the 2-fold advantage of significantly constraining the search space of most nucleotides
from the beginning, and building around a central portion
of the molecule containing a three-way or four-way junction. This is beneficial as these have greater stability than
external regions due to their highly constrained structures.
The rest of the structure is assembled around it using a
depth-first prioritization to constrain free-ends and limit
long-range stem interactions.
Contiguous assembly is a challenge due to the fact that
some motifs may have bases that are undefined in their 3D
structures, as explained in (16). Therefore undefined nucleotides are subsequently merged using links that overlap
the already placed motifs and stems. Because links are unconstrained, they have a very large conformational search
space. The server therefore attempts to limit this type of
fragment by placing them only after all adjacent stems and
motifs have been placed. The stems connecting the motifs
are built from chains of stacked canonical base pairs. The
assembly process proceeds until all residues in the sequence
have been defined by the script.
MC-Sym jobs are run on the IRIC web server. MC-Sym
offers tools to optimize, analyze and retrieve results. Energy minimization is recommended prior to analyzing RNAMoIP output. Clustering using the k-means algorithm can
also identify significant structures out of the complete pool
of results.
Motif insertion visualization
The RNA-MoIP web server provides tools for the visualization of motif insertion locations and 3D structures of inserted motifs. An annotated secondary structure diagram
is dynamically generated by the VARNA visualization applet (15). Using the motif-to-sequence base mapping defined during processing, the secondary structure is annotated and colored based on insertion location. Additionally,
non-canonical binding interactions are annotated using the
Leontis-Westhof symbolic notation. Motifs can contain sophisticated non-canonical interactions that help form the
backbone of the tertiary structure. Therefore, representing
these interactions can allow us to better understand the
folding configuration of the models, and make comparisons
amongst RNA-MoIP candidate outputs.
Downloaded from https://academic.oup.com/nar/article-abstract/45/W1/W440/3835309/RNA-MoIP-prediction-of-RNA-secondary-structure-and
by guest
on 16 September 2017
W444 Nucleic Acids Research, 2017, Vol. 45, Web Server issue
The web server also provides a tool for users to view 3D
models of the RNA motifs. Visualization is performed using
JSmol, a JavaScript framework that allows users to interactively view 3D molecular structures. MC-Sym parses a number of candidate structures for each motif selected. Using
RNA-MoIP, users can visualize and download the PDB files
for each of these candidates directly from their browsers.
This allows for an easy way to visualize and analyze the
structural components used to build the model.
There have been many attempts to computationally solve
the RNA tertiary structure prediction problem. Previous
contributions have used molecular dynamics and short fragment (1–3 nt) assemblies to predict RNA tertiary structures (8,9). The RNA-MoIP pipeline offers a fast and accurate approach to solving the prediction problem using
IP. By combining the available database with the flexibility of the IP model, our solution improves accuracy of
alternative prediction software for RNA sequences of all
lengths (13). Compared to MC-Sym alone, applying RNAMoIP constraints enables us to quickly produce secondary
structures that would otherwise run for an unspecified period of time.
Improvements planned for the RNA-MoIP server include
upgrades to the motif database. New RNA models are constantly being added to the PDB database from which our
motifs are extracted. We hope to build in a way of automatically updating our motif database as new listings are
added, to provide constantly up-to-date results. We later
also hope to allow users to filter the motifs used in their
search, and provide them with the option of adding their
own custom motif database to search on the server. For
the RNA-MoIP framework, an open problem remains optimizing the size of the pool of suboptimal structures used
as input. As mentioned, large pools significantly increase
the runtime of RNA-MoIP, though prior results have shown
that low ranking candidates can offer significant improvements to base-pair accuracy (for molecule 2DU3, RNAMoIP selects the 163rd candidate with a base-pair accuracy
of 91 versus 43% on average). We hope to solve this by clustering similar RNAsubopt structures to reduce the size of
the search pool and look through only meaningful structures.
The RNA-MoIP server was designed with the goal of
making the unique IP-based approach to tertiary structure
prediction available to all users. The server-side processing and optimizations streamline the assembly pipeline to
seamlessly go from any primary sequence directly to the 3D
structure predictions of MC-Sym, and allow for the exploration of the intermediate solutions. The web server was designed with ease-of-use in mind: its minimal design and visualization tools make the server accessible for users of any
background. We believe that it will serve as a valuable tool
for applications requiring both fast and accurate 3D structure predictions, such as RNA structure, function and interaction prediction.
Natural Sciences and Engineering Research Council of
Canada (NSERC) [RGPIN 2015-03786 & RGPAS 47787315 to JW]; Canadian Institutes of Health Research (CHIR)
[CIHR BOP-149429 to JW]; Genome Canada [B/CB 2015
to JW]. Natural Sciences and Engineering Research Council of Canada USRA and Fonds de recherche du Québec
Nature et technologies BRPC fellowships [to JY]. Azrieli
and Fonds de recherche du Québec Nature et technologies
postdoctoral fellowships [to VR]. Canadian Institutes of
Health Research (CIHR) [MT-14604 to FM]; National Institutes of Health (NIH) [R01GM088813 to FM]; and Natural Sciences and Engineering Research Council of Canada
(NSERC) [170165-01 to FM]. Funding for open access
charge: Natural Sciences and Engineering Research Council of Canada (NSERC) [RGPIN 2015-03786 to JW].
Conflict of interest statement. None declared.
1. Tinoco,I. Jr and Bustamante,C. (1999) How RNA folds. J. Mol.
Biol., 293, 271–281.
2. Brion,P. and Westhof,E. (1997) Hierarchy and dynamics of RNA
folding. Annu. Rev. Biophys. Biomol. Struct., 26, 113–137.
3. Parisien,M. and Major,F. (2008) The MC-Fold and MC-Sym pipeline
infers RNA structure from sequence data. Nature, 452, 51–55.
4. Wang,Z. and Xu,J. (2011) A conditional random fields method for
RNA sequence-structure relationship modeling and conformation
sampling. Bioinformatics, 27, i102–i110.
5. Seetin,M.G. and Mathews,D.H. (2011) Automated RNA tertiary
structure prediction from secondary structure and low-resolution
restraints. J. Comput. Chem., 32, 2232–2244.
6. Popenda,M., Szachniuk,M., Antczak,M., Purzycka,K.J.,
Lukasiak,P., Bartol,N., Blazewicz,J. and Adamiak,R.W. (2012)
Automated 3D structure composition for large RNAs. Nucleic Acids
Res., 40, e112.
7. Boniecki,M.J., Lach,G., Dawson,W.K., Tomala,K., Lukasz,P.,
Soltysinski,T., Rother,K.M. and Bujnicki,J.M. (2016) SimRNA: a
coarse-grained method for RNA folding simulations and 3D
structure prediction. Nucleic Acids Res., 44, e63.
8. Sharma,S., Ding,F. and Dokholyan,N.V. (2008) iFoldRNA:
three-dimensional RNA structure prediction and folding.
Bioinformatics, 24, 1951–1952.
9. Das,R. and Baker,D. (2007) Automated de novo prediction of
native-like RNA tertiary structures. Proc. Natl. Acad. Sci. U.S.A.,
104, 14664–14669.
10. Trausch,J.J., Marcano-Velázquez,J.G., Matyjasik,M.M. and
Batey,R.T. (2015) Metal ion-mediated nucleobase recognition by the
ZTP riboswitch. Chem. Biol., 22, 829–837.
11. Palacino,J., Swalley,S.E., Song,C., Cheung,A.K., Shu,L., Zhang,X.,
Van Hoosear,M., Shin,Y., Chin,D.N., Keller,C.G. et al. (2015) SMN2
splice modulators enhance U1-pre-mRNA association and rescue
SMA mice. Nat. Chem. Biol., 11, 511–517.
12. Djelloul,M. and Denise,A. (2008) Automated motif extraction and
classification in RNA tertiary structures. RNA, 14, 2489–2497.
13. Reinharz,V., Major,F. and Waldispühl,J. (2012) Towards 3D structure
prediction of large RNA molecules: an integer programming
framework to insert local 3D motifs in RNA secondary structure.
Bioinformatics, 28, i207–i214.
14. Lorenz,R., Bernhart,S.H., Höner zu Siederdissen,C., Tafer,H.,
Flamm,C., Stadler,P.F. and Hofacker,I.L. (2011) ViennaRNA
Package 2.0. Algorithms Mol. Biol., 6, 26.
15. Darty,K., Denise,A. and Ponty,Y. (2009) VARNA: interactive
drawing and editing of the RNA secondary structure. Bioinformatics,
25, 1974–1975.
16. Waldispühl,J. and Reinharz,V. (2015) Modeling and predicting RNA
three-dimensional structures. RNA Bioinformatics, 2015, 101–121.
Downloaded from https://academic.oup.com/nar/article-abstract/45/W1/W440/3835309/RNA-MoIP-prediction-of-RNA-secondary-structure-and
by guest
on 16 September 2017
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF