lncRNA-mRNA Interactions and RNA Structure Tools
Long non-coding RNAs (lncRNAs) regulate gene expression by binding proteins, DNA and other RNAs. Predicting which mRNAs a lncRNA can pair with, and how RNA folds, is a common bioinformatics task, but the tools give many false positives if used carelessly. This guide covers definitions, the biology of RNA-RNA interaction, study design, how thermodynamic tools such as RNAfold, RNAplex, RNAup and IntaRNA work and differ, how to build a prediction pipeline, how to validate with experimental data (CLIP, RIC-seq, SHAPE), and how to troubleshoot.
1. What are lncRNAs and why study their interactions?
Long non-coding RNAs (lncRNAs) are transcripts longer than 200 nucleotides with little or no protein-coding potential. They are transcribed by RNA polymerase II, often spliced and polyadenylated, and show tissue-specific, low-level expression.
Classes by genomic position
| Class | Position | Example |
|---|---|---|
| lincRNA | Between protein-coding genes | HOTAIR, MALAT1 |
| Antisense | Opposite strand of a gene | XIST partner TSIX, ANRIL |
| Intronic | Inside an intron | Many regulatory transcripts |
| Sense overlapping | Overlapping exons on the same strand | Various |
| Enhancer RNA (eRNA) | Transcribed from enhancers | Short, often unspliced |
| Circular RNA (circRNA) | Back-spliced circles | CDR1as |
What do lncRNAs do?
- Guide or scaffold chromatin-modifying complexes (HOTAIR with PRC2 and LSD1).
- Decoy that sequesters proteins or microRNAs (competing endogenous RNA, ceRNA).
- Signal of transcription events in a specific cell state.
- Base-pair with other RNAs, for example antisense lncRNAs that regulate mRNA stability, translation or splicing of the sense gene (BACE1-AS stabilizes BACE1 mRNA).
- Structural roles in nuclear bodies (NEAT1 in paraspeckles).
Why predict lncRNA-mRNA interactions?
A lncRNA that pairs with an mRNA may change its stability, translation, splicing or localization. Prediction narrows thousands of possible partners to a testable list before costly experiments such as RNA pull-down or CRISPR.
2. Core concepts: RNA secondary structure and RNA-RNA pairing
RNA folds by forming base pairs: Watson-Crick pairs (A-U, G-C) and wobble pairs (G-U). A secondary structure is the set of paired bases, drawn with stems, hairpin loops, internal loops, bulges and multiloops.
- Minimum free energy (MFE): the most stable structure predicted by thermodynamic nearest-neighbor rules (Turner parameters). Energy is in kcal/mol; more negative means more stable.
- Partition function: instead of one structure it computes the probability of every base pair across the whole ensemble of structures, giving base-pair probabilities and the ensemble free energy. This is more robust than the single MFE structure.
- Accessibility: the probability that a region is unpaired. Interaction between two RNAs is possible only if the interacting regions are accessible, or if the energy gain of pairing exceeds the energy cost of opening local structure.
- Interaction energy: total energy = hybridization energy of the intermolecular duplex + the energy needed to unfold the intramolecular structure (opening energy).
- Pseudoknots and kissing loops are not handled by standard MFE algorithms and need specialized tools (for example pKiss, IPknot).
Key limitations
Thermodynamic models were calibrated on short, isolated RNAs in vitro. In cells, proteins, modifications, co-transcriptional folding, cell compartment and RNA concentration change what actually forms. Long RNAs (several thousand nucleotides) fold into local structures more reliably than into one global structure. Therefore predictions are hypotheses, not evidence.
3. Study design
- Define the question. Examples: which mRNAs can a lncRNA of interest pair with; does an antisense lncRNA pair with its sense mRNA; which regions of a lncRNA are structured.
- Choose the candidate set. Use co-expression (correlation across samples), genomic neighborhood (cis targets within a window), differential expression after knockdown, or a literature list. Testing all of the transcriptome gives a large multiple-testing burden.
- Choose the sequence regions. Use the full transcript, or separate 5' UTR, CDS and 3' UTR, because pairing sites differ in their functional meaning and structure.
- Use the right isoforms. Take the transcript expressed in your tissue (from your own RNA-seq assembly with StringTie, or Ensembl and GENCODE annotations).
- Add a null model. Compare the predicted energy with that of dinucleotide-shuffled or length-matched random sequences, to get an empirical p-value.
- Plan validation. Decide in advance which experimental data you will use (see below).
Where to get sequences and expression
- GENCODE or Ensembl for annotated transcripts, NONCODE and LNCipedia for lncRNA catalogs, RNAcentral as a general ncRNA database.
- Your own RNA-seq, assembled with HISAT2 and StringTie or aligned with STAR, to define expressed isoforms and filter candidates by expression.
- Coding potential tools (CPC2, CPAT, LGC, RNAsamba) to confirm that a transcript is non-coding.
4. Tools compared
Secondary structure prediction (single RNA)
| Tool | Method | Output | Strengths | Limits |
|---|---|---|---|---|
| RNAfold (ViennaRNA) | Thermodynamic MFE and partition function | Structure in dot-bracket, base-pair probabilities, dot plot | Standard, fast, local folding with RNALfold and windows, constraints supported | No pseudoknots; accuracy drops for long RNAs |
| RNAstructure (Mathews lab) | Thermodynamic, with SHAPE constraints | Structures, probabilities | Good support for experimental constraints | Slower for very long sequences |
| Mfold and UNAFold | Thermodynamic | Suboptimal structures | Classic, web server | Older interface |
| CONTRAfold | Machine learning (probabilistic model) | Structures | Competitive accuracy | Fewer features |
| LinearFold | Linear-time approximations | Structures | Very fast for long RNAs, up to genomes | Approximation of MFE |
| RNAalifold and LocARNA | Comparative, uses alignments | Consensus structure | Conservation improves accuracy | Needs homologous sequences |
| Infernal and Rfam | Covariance models | Family annotation | Identify known RNA families | Only for known families |
| SPOT-RNA, UFold, MXfold2 | Deep learning | Structures including pseudoknots | Can capture non-canonical pairs | Generalization to new families is uneven |
RNA-RNA interaction prediction
| Tool | Approach | Typical use | Notes |
|---|---|---|---|
| RNAcofold (ViennaRNA) | Joint folding of two RNAs with a linker | Short RNAs, dimers | Does not scale to long transcripts |
| RNAplex | Fast duplex search with simple accessibility | Screening a short RNA (sRNA, siRNA) against many targets | Very fast, less precise |
| RNAup | Interaction energy with accessibility (opening energy) | Medium-size RNAs | More accurate, slower |
| IntaRNA | Accessibility-aware interaction with seed constraints, heuristics for long RNAs | General purpose, lncRNA-mRNA and sRNA-mRNA | Widely used, flexible, supports multiple result reporting |
| RNAhybrid | Minimum free energy hybridization, no intramolecular structure | miRNA target search | Ignores accessibility |
| LncRRIsearch and web servers | Precomputed lncRNA-RNA interactions (human, mouse) based on RIblast | Quick lookup | Limited to the annotated set |
| RIblast | Seed-and-extend search using accessibility, built for large scale | Transcriptome-wide lncRNA-RNA screens | Fast, designed for lncRNA data |
| LncTar, catRAPID, lncPro | Statistical or physics-based for RNA-protein or RNA-RNA | Candidate ranking | Check species and training data |
| TargetScan, miRanda, miRDB | miRNA-target prediction | ceRNA analysis (lncRNA as sponge) | Different task: miRNA seed matching |
For a typical lncRNA-mRNA study, IntaRNA or RIblast are practical first choices; RNAplex and RNAhybrid are fast filters; RNAup is a more careful second step for selected pairs.
5. Running the analysis
Prepare sequences
# Extract transcript sequences from annotation and genome
gffread annotation.gtf -g genome.fa -w transcripts.fa
# Pull your lncRNA and candidate mRNAs
seqtk subseq transcripts.fa lncRNA.ids > lncRNA.fa
seqtk subseq transcripts.fa mRNA.ids > mRNA.faPredict structure with ViennaRNA
# MFE and partition function, with base-pair probability matrix and a drawing
RNAfold -p --noLP < lncRNA_fragment.fa > fold.out
# Local folding for long RNAs: only pairs within a span of 150 nt
RNALfold -L 150 < lncRNA.fa > local.out
# Visual structure
RNAplot < fold.outUse RNAfold on windows (for example 1000 nt with 500 nt step) for long lncRNAs, or RNALfold with a maximum span, because global MFE structures of long RNAs are unreliable.
Predict the interaction with IntaRNA
# One lncRNA against many mRNAs, best interaction per pair
IntaRNA -q lncRNA.fa -t mRNA.fa \
--outMode C --outNumber 1 \
--threads 8 \
--out intarna.csv
# Add a seed requirement and report the energy components
IntaRNA -q lncRNA.fa -t mRNA.fa --seedBP 7 --outMode C \
--outCsvCols id1,start1,end1,id2,start2,end2,E,E_hybrid,ED1,ED2,subseqDP \
--out intarna.csvImportant columns: E is the total interaction energy (the lower, the stronger), E_hybrid the duplex energy, and ED1 and ED2 the energy needed to open each RNA. A strongly negative E_hybrid with a large ED can result in a weak total interaction, which is exactly what accessibility corrects.
Fast screen with RNAplex and RIblast
# RNAplex: duplex energy between a short query and many targets
RNAplex -q query.fa -t targets.fa -e -10 > rnaplex.out
# RIblast: build a database once, then search
RIblast db -i mRNA.fa -o mRNA_db
RIblast ris -i lncRNA.fa -d mRNA_db -o ris_out.txtStatistical significance with a null model
Energy scores depend on length and GC content, so raw energies are not comparable between pairs.
- Shuffle the mRNA sequence preserving dinucleotide frequency (
ushuffleorseqkit shuffleis not enough; use dinucleotide shuffling such as ushuffle). - Run IntaRNA on 100 to 1000 shuffled sequences per target.
- Compute an empirical p-value as the fraction of shuffles with an energy at least as strong as the real one.
- Correct for multiple testing (Benjamini-Hochberg).
Integrate with expression data
A real interaction should be supported by biology:
- Co-expression of the lncRNA and candidate mRNA across samples (positive or negative correlation).
- Change in the mRNA after lncRNA knockdown or overexpression (differential expression).
- Localization: both must be in the same compartment (nuclear lncRNAs rarely pair with cytoplasmic mRNAs). Use cell fractionation RNA-seq or databases such as RNALocate and lncATLAS.
- Conservation of the binding site across species.
6. Experimental validation and data to integrate
| Method | What it shows |
|---|---|
| RIC-seq, PARIS, SPLASH, LIGR-seq, MARIO | Transcriptome-wide RNA-RNA contacts and duplexes in cells (proximity ligation and sequencing) |
| CLIP-seq, iCLIP, eCLIP | Protein-RNA binding sites (useful when a lncRNA acts via a protein) |
| CHIRP-seq, CHART, RAP | Where a lncRNA binds chromatin or RNA |
| SHAPE-MaP, DMS-MaPseq, icSHAPE | Single-nucleotide RNA structure probing in vitro and in cells |
| RNA pull-down with antisense probes | Proteins and RNAs that co-purify with a lncRNA |
| Luciferase reporter with mutated binding site | Direct test of a predicted pairing region |
| Compensatory mutations | Mutate one RNA to break pairing, then the partner to restore it; the strongest functional evidence |
| smFISH and RNAscope | Co-localization in cells |
Use structure probing as a constraint. RNAfold and RNAstructure can accept SHAPE reactivities as soft constraints, which improves structure accuracy. For interactions, compare predicted sites with RIC-seq or PARIS contact maps where available for your cell type.
7. Analyzing predicted structures
- Plot the base-pair probability dot plot and the MFE structure, and prefer regions where the MFE structure agrees with high-probability pairs.
- Compute the positional entropy (low entropy means well-defined structure).
- Calculate accessibility profiles with
RNAplfold(probability of being unpaired in a window), which feed interaction tools. - Compare the lncRNA structure with conservation and with Rfam families. A conserved structure suggests function, but many lncRNAs show weak sequence conservation and may still be functional.
- Visualize with VARNA, forna, RNAcanvas or R2DT, which draw publication-quality structure diagrams.
8. Special cases
Antisense lncRNAs and natural antisense transcripts
The sense-antisense pair is a perfect complement over the overlap, so full complementarity is guaranteed. The relevant question is what the overlap region does (RNA masking, RNase A or Dicer processing, transcriptional interference). Check strand-specific RNA-seq to confirm both transcripts are expressed.
lncRNAs as microRNA sponges (ceRNA)
Predict miRNA binding sites with TargetScan or miRanda, then check co-expression of the lncRNA, miRNA and mRNA, and consider the stoichiometry. Many reported ceRNA effects require high lncRNA abundance relative to miRNA, which is often not met; see Thomson and Dinger (2016) for a critical discussion.
Bacterial small RNAs
The same tools apply: small regulatory RNAs (sRNAs) bind mRNA near the ribosome binding site, often with the help of Hfq. IntaRNA was originally developed for these cases and works well with short seed regions.
Plant lncRNAs and target mimics
Plant miRNAs pair with near-perfect complementarity, so tools such as psRNATarget are preferable for target and mimic prediction.
9. Troubleshooting: symptoms, causes and fixes
| Symptom | Likely cause | What to do |
|---|---|---|
| Thousands of significant interactions | Energy threshold too permissive, no null model | Use dinucleotide-shuffle p-values, require a seed, apply FDR |
| Tool runs out of memory or never finishes | Very long RNAs (above about 10 kb) | Split into windows, use IntaRNA heuristics (--qIntLenMax, --tIntLenMax), or RIblast |
| Stable interaction predicted but no effect in the cell | Different compartments, protein occupancy, non-expressed isoform | Check localization, expression of the correct isoform and RNA-binding proteins |
| Structure changes with sequence window | Global MFE of long RNA is unstable | Use local folding or sliding windows, compare ensemble probabilities |
| Predicted structure contradicts SHAPE data | Model ignores modifications or proteins | Add SHAPE as constraints, test alternative structures |
| lncRNA turns out to code for a peptide | Misclassified transcript | Re-check with CPC2 and ribosome profiling or a proteomics support |
| Unequal results between IntaRNA and RNAplex | Different energy models and accessibility handling | Treat overlap as higher confidence, follow up with RNAup |
| Zero hits for a known interacting pair | Wrong strand, wrong isoform, or seed requirement too strict | Verify orientation (use --outMode with the query as given, check reverse complement), relax the seed |
| Energy values not comparable among targets | Energy depends on length and GC | Normalize by length or use the empirical p-value from shuffles |
| Annotation mismatch when mapping IDs | Different Ensembl or GENCODE versions | Use one annotation version consistently |
| Pseudoknot is expected but missing | Standard algorithms exclude pseudoknots | Use pKiss or IPknot, or experimental structure data |
10. Best-practice checklist
- Define the biological question and the candidate set before running tools.
- Confirm the transcripts are non-coding, expressed and in the same compartment.
- Use the correct isoforms and split long RNAs into regions or windows.
- Use an accessibility-aware tool (IntaRNA, RNAup, RIblast) rather than only duplex energy.
- Estimate significance with shuffled sequences and correct for multiple testing.
- Combine with co-expression, knockdown data and conservation.
- Validate with proximity-ligation data or compensatory mutation experiments.
- Report software versions, parameters and sequence sources.
11. Frequently asked questions
Is a low free energy enough to claim an interaction? No. It indicates that pairing is thermodynamically possible, not that it happens in cells.
Which tool should I start with? IntaRNA for general lncRNA-mRNA and sRNA-mRNA work, with RNAfold for the structures of individual RNAs.
Can deep learning replace thermodynamic models? Deep-learning predictors can capture patterns beyond Turner rules but generalize unevenly to new RNA families. Use them as complementary evidence.
How long can the RNAs be? RNAfold handles several thousand nucleotides but accuracy falls with length. For interactions, windows or heuristic tools are needed beyond a few kilobases.
What is the difference between a lncRNA binding an mRNA and binding a protein? RNA-RNA pairing is sequence complementarity predicted with the tools above. RNA-protein binding depends on motifs and structure and is studied with CLIP and tools such as catRAPID.
12. References
- Lorenz R et al. (2011). ViennaRNA Package 2.0. Algorithms for Molecular Biology 6:26.
- Mann M, Wright PR, Backofen R (2017). IntaRNA 2.0: enhanced and customizable prediction of RNA-RNA interactions. Nucleic Acids Research 45:W435-W439.
- Busch A, Richter AS, Backofen R (2008). IntaRNA: efficient prediction of bacterial sRNA targets incorporating target site accessibility and seed regions. Bioinformatics 24:2849-2856.
- Tafer H, Hofacker IL (2008). RNAplex: a fast tool for RNA-RNA interaction search. Bioinformatics 24:2657-2663.
- Muckstein U et al. (2006). Thermodynamics of RNA-RNA binding. Bioinformatics 22:1177-1182.
- Rehmsmeier M et al. (2004). Fast and effective prediction of microRNA/target duplexes. RNA 10:1507-1517.
- Fukunaga T, Hamada M (2017). RIblast: an ultrafast RNA-RNA interaction prediction system based on a seed-and-extension approach. Bioinformatics 33:2666-2674.
- Fukunaga T et al. (2019). LncRRIsearch: a web server for lncRNA-RNA interaction prediction integrated with tissue-specific expression and subcellular localization data. Frontiers in Genetics 10:462.
- Mathews DH (2014). RNA secondary structure analysis using RNAstructure. Current Protocols in Bioinformatics 46:12.6.1-12.6.25.
- Huang L et al. (2019). LinearFold: linear-time approximate RNA folding by 5'-to-3' dynamic programming and beam search. Bioinformatics 35:i295-i304.
- Spitale RC, Incarnato D (2023). Probing the dynamic RNA structurome and its functions. Nature Reviews Genetics 24:178-196.
- Cai Z et al. (2020). RIC-seq for global in situ profiling of RNA-RNA spatial interactions. Nature 582:432-437.
- Lu Z et al. (2016). RNA duplex map in living cells reveals higher-order transcriptome structure. Cell 165:1267-1279.
- Statello L et al. (2021). Gene regulation by long non-coding RNAs and its biological functions. Nature Reviews Molecular Cell Biology 22:96-118.
- Thomson DW, Dinger ME (2016). Endogenous microRNA sponges: evidence and controversy. Nature Reviews Genetics 17:272-283.
- Kang YJ et al. (2017). CPC2: a fast and accurate coding potential calculator based on sequence intrinsic features. Nucleic Acids Research 45:W12-W16.
- Pertea M et al. (2015). StringTie enables improved reconstruction of a transcriptome from RNA-seq reads. Nature Biotechnology 33:290-295.
💬 Commentaires