TRIPinRNA
Project Overview
TRIPinRNA is a Python-based in silico platform that searches whole transcriptomes for sequences that could fold into intramolecular RNA triple helices. In the human transcriptome only two such structures were well known, the stability elements at the 3′ ends of the lncRNAs MALAT1 and NEAT1, and the study asked whether triple helices are really that rare or had been missed by the search methods used so far.
As co-first author, I co-developed the search code and performed the bioinformatics analysis, and co-wrote the paper. Collaborators carried out the biophysical validation, and the work was co-supervised by Dr. Souvik Maiti.
Transcriptome-Wide Search
Scans every transcript, noncoding and protein-coding, for the three strands of a triple helix
Two Search Modes
With a gap nucleotide in the Hoogsteen strand, as in MALAT1 and NEAT1, or with a continuous strand
Complete Output
Every predicted triplex, with its strand sequences and start and end positions, written to a CSV file
Research Impact
The Biochemistry paper ran TRIPinRNA on the GENCODE human transcriptome (release 44), long noncoding and protein-coding, and on part of the NONCODE lncRNA database, and tested selected predictions by biophysics.
X Chromosome Inactivation
Predicted triplex-forming sequences in six lncRNAs involved in X-chromosome inactivation: XIST, TSIX, JPX, FTX, FIRRE and XACT. One predicted triplex each from JPX, FTX and XACT, made as a synthetic RNA, formed a triple helix in vitro.
Coding Transcripts and UTRs
In protein-coding transcripts more than half of the predicted hits lay in 3′ UTRs. Transcripts with hits had higher GC content in their 5′ UTRs and lower in their 3′ UTRs than average, a bias absent from their coding sequences.
Beyond Existing Tools
Several earlier triplex search programs allow at most 10 nucleotides between strands, too few for MALAT1-like triple helices, and one that allows longer gaps (NeSSie) is too lenient to give a workable number of hits; TRIPinRNA allows 3 to 100. An earlier Infernal-based search found only MALAT1 and NEAT1 in the human genome, and AlphaFold3 did not predict triple helices in TRIPinRNA's hits.
Experimental Validation
Collaborators validated selected predictions with circular dichroism, UV melting and thermal difference spectra, as reported in the publication.
Technical Approach
TRIPinRNA is a sequence-pattern (string-match) search modelled on the MALAT1 and NEAT1 triple helices:
Core Triplex Search
A sliding window looks for a Hoogsteen strand of 7 to 15 nucleotides, a downstream Crick strand that matches it, and a further downstream Watson strand complementary to both, with 3 to 100 nucleotides between the Hoogsteen and Crick strands and between the Crick and Watson strands.
Gap and No-Gap Modes
In “with-gap” mode the Hoogsteen strand carries one gap nucleotide, placed anywhere except the three nucleotides at either end, as in MALAT1 and NEAT1; “without-gap” mode looks for a continuous strand.
Stems and Loops
Once the core triplex is found, the sequences between the strands are checked for an upper and a lower stem of at least 3 base pairs with a loop of at least 3 nucleotides; without a stem, the segment is reported as a loop. No base-composition rule is imposed, so canonical and noncanonical triplexes are both found, and pyrimidine-only and uridine-only hits are filtered afterwards.
Output and Scale
The input is a zipped FASTA file and the output a CSV of every predicted triplex. The transcriptome runs were executed on an HPC cluster under the SLURM scheduler, and the code for both modes is on GitHub.
Key Findings
Thousands of Candidates
In a human transcriptome where essentially only the MALAT1 and NEAT1 triple helices were well known, the without-gap search predicted 10,769 candidate triplex sites in lncRNA transcripts and 45,418 in protein-coding transcripts.
Noncanonical Triplexes Form In Vitro
Heteropurine-pyrimidine triple helices, usually expected to be destabilizing, gave triple-helix signatures in vitro, though they were less thermally stable than canonical ones.
Link to GC Content
Transcripts with predicted triplexes differed in GC content from the transcriptome average: lncRNA transcripts with hits had slightly lower GC (mean 44.5–44.8% against 46.6%), and coding transcripts with hits had higher GC in their 5′ UTRs and lower GC in their 3′ UTRs.
A Canonical Triplex in JPX
A uridine-only triple helix predicted in JPX, the lncRNA that activates XIST, melted in two transitions (about 58 and 75 °C) and gave a CD signature practically identical to that of the MALAT1 triple helix.
Publication
Development of an in silico platform (TRIPinRNA) for the identification of novel RNA intramolecular triple helices and their validation using biophysical techniques
Biochemistry (ACS) 64(1):250-265 (2025). DOI: 10.1021/acs.biochem.4c00334
Summary: TRIPinRNA is a Python tool that scans human long noncoding and protein-coding transcripts for sequence patterns that could fold into intramolecular RNA triple helices. With it, the study predicts both canonical (homopyrimidine) and noncanonical (heteropurine-pyrimidine) triplexes, shows with biophysical tests on selected candidates that both kinds form in vitro, and reports a strong correlation between triplex occurrence and local GC content. The tool can be applied to any organism with an annotated transcriptome.
Future Directions
The paper points to these next steps:
Learning From the Predictions
Using TRIPinRNA's output to train machine-learning models that find triple helices with slightly different features.
Other Organisms
Extending the search to the transcriptomes of other organisms; the code works on any annotated transcriptome.
What Governs a Triplex
Relating loop length, strand length and the presence of a stem to when triple helices form and how stable they are, and asking whether triplexes predicted in coding sequences relate to nonsense-mediated decay.
Structures of Noncanonical Triplexes
Crystal structures of the noncanonical triple helices, which the paper notes are beyond its scope.
Interested in RNA Structure Research?
If you're interested in learning more about TRIPinRNA, collaborating on RNA structure research, or discussing potential applications, I'd be happy to connect.