RFoldSearch User Guide
RFoldSearch is a web server for the rapid retrieval of RNA tertiary structures. It encodes the local three-dimensional geometry at each nucleotide as a discrete symbol from a learned structural alphabet (3Dr), recasting structure comparison as the alignment of structural-token sequences. This representation enables indexed, large-scale searches against databases of experimentally determined and computationally predicted RNA structures.
This guide describes the preparation and submission of query structures, the configuration of searches, the interpretation of reported metrics, and the inspection of structural alignments. Methodological details are provided in the RFoldSearch manuscript.
1. Query submission
1.1 Input formats
Queries are submitted on the search page by one of three routes:
- File upload: an RNA structure in PDB or mmCIF format.
- Direct input: PDB- or mmCIF-formatted coordinate text pasted into the input field.
- Example query: a pre-loaded RNA structure provided for demonstration (Load example).
1.2 Input requirements
| Criterion | Requirement |
|---|---|
| Molecule type | The file must contain RNA atomic coordinates. |
| Nucleotide validity | A nucleotide is considered valid only if its C3′, C4′ and O4′ atoms are present; other nucleotides are skipped. |
| Minimum length | The query must contain at least 4 valid nucleotides. |
| Multi-chain input | Where multiple RNA chains are present, only the first RNA chain is used. |
| File size | Up to the size shown under the input box, for uploaded and pasted structures alike. |
2. Target databases
RFoldSearch currently supports searches against three structure collections. One or more databases may be selected for a given query.
| Database | Content |
|---|---|
| PDB-RNA | Experimentally determined RNA structures deposited in the Protein Data Bank |
| RNAsolo 2.0 | A non-redundant set of experimentally determined RNA structures |
| NuFold | A large-scale collection of predicted RNA structures. The models carry no sequence, so 3Dr identity is reported in place of sequence identity. |
3. Search modes
Two search modes are available, representing a trade-off between computational cost and ranking accuracy.
| Mode | Procedure |
|---|---|
| RFoldSearch (default) | Structural-token retrieval followed by ranking with a combined token-alignment and geometric score |
| RFoldSearch-TM | Structural-token candidate retrieval followed by coordinate-based reranking with US-align |
The default RFoldSearch mode is recommended for most applications. Upon submission, candidate structures are identified through the indexed structural-token search pipeline and returned as a ranked list for each selected database. By default, up to 50 hits are reported per database.
4. Search parameters
Additional parameters are accessible through the Settings panel.
| Parameter | Definition | Default |
|---|---|---|
| Hits shown | Maximum number of hits reported per database | 50 |
| Candidate depth | Number of candidates retained for downstream alignment and ranking | 200 |
| Min. diagonal k-mers | Minimum number of compatible diagonal k-mer seeds required for a target to be retained as a candidate | 2 |
The default values represent a practical balance between computational efficiency and retrieval sensitivity. Increasing the candidate depth may recover additional, more distantly related structures at greater computational cost. Modifying the seed threshold alters the stringency of candidate selection: higher thresholds increase selectivity, whereas lower thresholds increase sensitivity at the expense of a larger candidate pool.
5. Interpretation of search results
5.1 Reported metrics
Each search returns a ranked table of structural matches. Click a column heading to sort by it, and a row to open the hit.
| Field | Definition |
|---|---|
| Target | Identifier of the matched RNA structure: the PDB entry with its chain, or the RNAcentral id of a predicted model (linked to the entry page) |
| RNA / Organism | RNA description, source organism, and available annotations (for predicted models, the Rfam family when one is assigned) |
| pLDDT | Predicted local structural confidence (reported for predicted structures where available) |
| Seq identity | Nucleotide sequence identity over aligned positions. Databases without sequences report 3Dr identity instead: the fraction of aligned pairs with the same structural token. |
| RMSD (Å) | Root-mean-square deviation of aligned nucleotides after optimal superposition |
| Structure score | Composite score integrating token-alignment and geometric similarity (Section 5.2) |
| TM-score | Length-normalised measure of global structural similarity (Section 5.3) |
| E-value | Statistical significance of the structural-token alignment score (Section 5.5) |
| Query coverage | Fraction of query nucleotides included in the alignment |
| Target coverage | Fraction of target nucleotides included in the alignment |
5.2 Structure score
In the default mode, hits are ranked by a structure score that combines structural-token alignment similarity with geometric agreement:
Rstr = SSW × aTM
where SSW denotes the Smith–Waterman alignment score computed over structural-token sequences, and aTM denotes the TM-score of the token alignment after structural superposition (the TM-score column in the default mode). Higher values of Rstr indicate stronger matches under the RFoldSearch ranking function.
5.3 TM-score
The TM-score quantifies global three-dimensional structural similarity on a scale from 0 to 1, with higher values indicating closer structural correspondence. Because it is normalised by length, it is less sensitive to local deviations than RMSD. RFoldSearch reports the mean of the query- and target-normalised scores: in the default mode for the token alignment, and in RFoldSearch-TM mode the US-align TM-score of the two chains. The hit view shows the same value as the table.
5.4 RMSD
The RMSD quantifies the mean geometric deviation, in ångströms, between aligned nucleotides after optimal superposition. As RMSD is computed only over aligned positions, it should be interpreted in conjunction with alignment coverage: a low RMSD obtained over a short alignment does not by itself indicate extensive structural similarity.
5.5 E-value
The E-value represents the expected number of unrelated database entries that would attain a structural-token alignment score equal to or greater than the observed score under the statistical background model. Lower E-values denote greater statistical significance. Databases that have not yet been calibrated report no E-value.
5.6 Alignment coverage
Query and target coverage together characterise the extent of the detected similarity:
- High query and high target coverage: the alignment encompasses a substantial proportion of both structures, consistent with global structural similarity.
- High query coverage, low target coverage: the query corresponds to a structural region within a larger target RNA, for example a domain or subdomain.
- Low query and target coverage: the similarity is likely confined to a local structural motif.
Coverage should therefore be considered explicitly when distinguishing local motif-level matches from global structural correspondence.
6. Inspection of structural alignments
Selecting a hit opens a detailed view comprising the structural-token alignment and an interactive three-dimensional superposition.
6.1 Structural-token alignment
The alignment displays corresponding nucleotide positions together with their assigned 3Dr structural states. The relationship between each pair of aligned tokens is annotated as follows:
| Symbol | Interpretation |
|---|---|
| | Identical structural tokens |
+ | Favourable structural-token substitution |
. | Other aligned token pair |
These annotations describe similarity in local structural state and are independent of nucleotide sequence identity. Positions count valid nucleotides only; databases without sequences show the target residues as ·.
6.2 Three-dimensional superposition
The structure viewer displays the query and target after superposition, using the following colour scheme:
- Orange: query RNA
- Blue: target RNA
- Pale shading: unaligned regions
Structures may be rotated and zoomed to assess structural correspondence. Selecting a position in either the alignment or the viewer highlights the corresponding nucleotides in both representations.
7. Case study: identification of structurally related RNAs
The following workflow illustrates the use of RFoldSearch to identify candidate structural relatives of an RNA of unknown function.
- Query submission. Upload the query structure in PDB or mmCIF format.
- Database selection. Select RNAsolo 2.0 to search a non-redundant set of experimentally determined structures.
- Mode selection. Use the default RFoldSearch mode.
- Initial screening. Execute the search and examine the highest-ranking hits.
- Metric assessment. Evaluate TM-score, RMSD, query and target coverage, and functional annotations for each candidate.
- Alignment inspection. Examine the structural-token alignments and three-dimensional superpositions of promising candidates.
- Refined ranking. Where more accurate coordinate-based ranking is required, repeat the search in RFoldSearch-TM mode.
- Extended search. Optionally, search the NuFold database to identify related structures among predicted RNA models.
Interpretation. Structural similarity may indicate a conserved RNA architecture even in the absence of detectable sequence similarity. However, structural similarity alone is insufficient to establish common ancestry or shared function, and inferences of this kind should be supported by independent evidence.
8. Data export
Search results may be exported or shared from the results page:
| Option | Description |
|---|---|
| Download TSV | Tab-separated results for downstream statistical analysis |
| JSON | Structured results for programmatic processing |
| Copy link | Persistent link to the search results for sharing |
9. Programmatic access
RFoldSearch provides a REST API for programmatic structural searches, facilitating integration into automated analysis pipelines and high-throughput workflows. Interactive documentation of the endpoints, input formats, request parameters and response schemas is available at /api/docs.
10. Frequently asked questions
Can a search be performed using sequence information alone?
No. RFoldSearch requires a three-dimensional RNA structure as input. Where only a nucleotide sequence is available, a structural model must first be obtained by experimental determination or computational prediction.
Can RFoldSearch detect structural similarity between RNAs with low sequence identity?
Yes. Because RFoldSearch relies on learned representations of local geometry rather than primarily on nucleotide sequence, it can retrieve structurally similar RNAs whose sequences have substantially diverged.
Why may hits with identical sequence identity receive different structure scores?
Sequence identity does not imply conformational identity. Because the structure score reflects both alignment quality and geometric agreement, structures with similar or identical sequences but distinct conformations may receive different scores.
Why is the hit with the lowest RMSD not necessarily ranked first?
RMSD is computed only over aligned positions and does not account for alignment length or token-level alignment quality. A short alignment may yield a low RMSD without representing the strongest overall structural match.
Why are results not ordered by E-value?
Results are ranked by the structure score, which incorporates geometric agreement, whereas the E-value reflects only the statistical significance of the structural-token alignment score. The two orderings may therefore differ.
Which search mode is most appropriate?
The default RFoldSearch mode is recommended for routine and large-scale searches. RFoldSearch-TM is recommended when accurate coordinate-based structural ranking is the primary consideration.
Can predicted structures be used?
Yes. Predicted RNA structures may be submitted as queries, and the NuFold database provides a large collection of predicted structures as search targets. In both cases, the uncertainty inherent in structure prediction should be taken into account when interpreting results; pLDDT values, where available, indicate model confidence.
Does structural similarity imply homology?
Not necessarily. Structural similarity may reflect conserved architecture or a possible evolutionary relationship, but establishing homology requires additional sequence, functional or phylogenetic evidence.
11. Resources
- RFoldSearch web server: https://www.rfoldsearch.com/
- Interactive API documentation: https://www.rfoldsearch.com/api/docs
Technical details of the structural alphabet, alignment algorithm, indexing strategy and statistical calibration are described in the RFoldSearch manuscript and the accompanying source code.
