How-to guides

Which tool do I need?

Typical path:

Walkthroughs

Examples use ZNF426 (chr19; a long, nearly perfect dsRNA in its 3′ UTR), MAVS (chr20) and SON (chr21, the IGV default).

Reference: IGV Viewer

The IGV Viewer is igv.js with the dsRNAome tracks preloaded and a filter sidebar added. Standard igv.js behaviour (track menus from the gear icon, drag to reorder tracks, right-click options) all works.

Tracks and zoom limits

Every feature track stops drawing beyond a maximum window so that dense data do not stall the page. When a track is blank, check the zoom first.

Filter sidebar

The sidebar hides arcs from both dsRNA tracks. Nothing changes until you click Apply Filters; Reset returns every control to its default and re-applies. Every active condition must pass, except that the three model sliders can be combined with OR.

The sidebar can be hidden with the ❮ button or Filter Tracks in the menu bar; the IGV panel widens to fill the space.

Menu bar

Tracks

Local File... loads a track from your disk; URL... loads one from a web address. Indexed formats (.bed.gz + .tbi, bigWig, bigBed, BAM + .bai) stream only the region in view.

Share

Link to the current view (locus + tracks + filters), embeddable iframe, QR code.

Share Coordinates

A plain locus link in a prompt box for quick copying.

Save Image

SVG (vector) or PNG of the whole panel.

Help

The igv.js user guide, this guide, and the About box with version numbers.

Swap strands ⇅

Mirror the arc layout: reverse-strand arcs on top opening upward, forward-strand arcs at the bottom opening downward. For genes on the minus strand. Recorded in share links as flip=1.

Filter Tracks

Show or hide the filter sidebar.

Track Selection

Checklist of all built-in tracks with Select All / Unselect All / Apply.

Reference: Filter dsRNAs (table)

A Shiny app over a SQLite database of all 5,134,754 predictions. Filters are combined with AND; within the gene box, the tokens are ORed.

Search

Gene name(s) or Ensembl ID(s)

Symbols match either arm's gene exactly (case-insensitive). ENSG IDs (version suffix allowed) match the i-arm's gene ID. Separate several with commas, semicolons, spaces or new lines.

Locus

chr4, chr4:1,000,000 or chr4:1,000,000-2,000,000. Keeps dsRNAs whose i-arm starts at or after the start and ends at or before the end. Unparseable text is ignored without a message, so check the Summary tab's "Filters applied" list.

Strand

Any, +, or −.

Filters

Tabs

Data Table

Essentials or All columns; client-side search, sort and per-column filters over the loaded rows; "View on IGV" and "View Structure" links per row.

Summary

Match count and percentage of the database, the filters applied, and counts by editing status, repetitive status, and the two crossed.

Density Plots

Arm length, longest helix, percent paired, loop length and folding energy, split by editing status and repetitive status. Drawn from a random sample of up to 20,000 matches, so distributions are representative even when the table is capped.

Box Plots

The same five variables as box plots by editing status, with n in the labels.

Download dialog

Format: CSV, compressed CSV, Parquet, BEDPE (default), compressed BEDPE. Columns: one checkbox per database column, preselected to the Essentials set; quick links Essentials / All / None. BEDPE ignores the column picker: columns 1–10 are standard BEDPE (chr1, start1, end1, chr2, start2, end2, name, score, strand1, strand2) and 11–38 carry every attribute, with no header row. Large downloads can take 30 seconds or more; a progress bar shows the stage.

Reference: dsRNA Browser

A single-page app that loads its database into the browser once (cached for the session), then shows a filterable table and a detail view for the selected dsRNA. The database is the ~108,000 dsRNAs whose arms overlap exons; 13,789 of them have probing data that passes the paper's power filter (≥ 0.5), and those are the ones whose ROC-AUC values are worth reading closely.

Table columns

Click any header to sort. The sort and page size persist between visits. The default order is Best Reactivity (powered): among experiment-arms with detection power ≥ 0.5 (the paper's power filter), the validation with the highest lower 95% confidence bound on its ROC-AUC. Ranking by the bound rather than the ROC itself keeps a perfect 1.000 from a handful of nucleotides from outranking a well-measured 0.87. A green left border marks rows where one experiment has both arms powered with ROC > 0.7; a red border marks both arms powered with ROC < 0.5 (well-measured, and the duplex is contradicted).

Filters

Gene Search

Substring match on the gene name or Ensembl ID of either arm.

dsRNA ID

Exact numeric ID.

Max CI Width

Keep rows whose best-ROC confidence interval is at most this wide (rows without a CI are kept).

Reactivity Data

All / Both Arms / Any Arm / No Reactivity Data.

Min Length, Min BP %, Min i Edits, Min j Edits

Numeric minimums.

Power-filtered only (≥0.5)

At least one experiment-arm with detection power ≥ 0.5; about 15% of the table.

Conserved only

All phastCons scores > 0.5.

High Confidence ML Models

GTEx ≥ 0.251, Stability ≥ 0.247, Probing ≥ 0.4574, or Probing ≥ 0.33 (medium, includes GRIA2-like structures). Ticked boxes are ANDed.

When any filter is active a red banner shows "showing X of Y entries" with a Clear All Filters button.

Detail view, top to bottom

1. Select Reactivity Experiment: dropdown; best experiment starred and preselected.
2. Reactivity legend: low → high colour bar, grey = no data, green outline = editing site.
3. Predicted dsRNA Structure with Key Structural Properties (arm lengths, RNAduplex-trimmed lengths, loop, energy, longest helix, % paired) and the per-arm RNA Structure Probing Metrics.
4. Genomic Context: location, "Open in IGV Browser", and an embedded IGV view of the dsRNA ±100 nt with its base pairs and GENCODE genes.
5. RNA Editing Sites (when present): table of REDIportal sites with sample and tissue counts; link to REDIportal.
6. Local Folding (RNAfold): each arm folded alone, with MFE energy and, when an experiment is selected, the same probing metrics.
7. RNA Structure Probing Summary: dsRNA structure vs RNAfold ROC per arm; ★ marks the better fit.

Interpreting scores

A prediction in this resource means dsRNAscan found an inverted repeat that can form a duplex of at least 25 bp at ≥ 70% pairing. Everything else is a layer of evidence on top of that. The layers are independent, so a candidate supported by two or three of them is far more likely to be a real, cellular dsRNA than one supported by none.

The three machine-learning scores

The IGV Viewer and Filter dsRNAs round the cutoffs to two decimals; the dsRNA Browser shows the exact values. They select the same dsRNAs. 2,458,476 predictions pass at least one model's cutoff (the paper's high-confidence set) and 1,509,138 pass all three. The GTEx and Stability models agree on 1.7 million; the Stability model adds about 471,000 that the GTEx model misses (mostly dsRNAs with little GTEx expression) and the GTEx model adds about 58,000 (Figure 3C of the paper). Both subsets are on the Downloads page as BEDPE files.

What a high score does not mean. The GTEx and Stability models were trained on ADAR editing, so a high score says "this looks like the dsRNAs ADAR edits", not "this is immunogenic" or "this is functional". A dsRNA with no expression in GTEx tissues can score low on the GTEx model and high on the Stability model; that is a difference in expression, not in structure. Use the Stability model for tissues or conditions GTEx does not cover.

Experimental layers

Editing sites

REDIportal v3 A-to-I sites (9,642 GTEx samples, 54 tissues; 9,681 TCGA samples). Sites on both arms are the strongest signal that the two arms pair in vivo. Sites on one arm only may come from a different partner. Sites on the opposite strand still count (the "unstranded" columns): the same inverted repeat forms the same duplex on either strand, so editing of one strand is taken as evidence for the mirror-image prediction on the other, which is unedited only because that strand is not expressed. See the glossary.

Probing ROC-AUC

From DMS, SHAPE and related experiments: how well reactivity separates predicted-unpaired from predicted-paired nucleotides. Before any filtering, most of the 189,000 arm-level AUCs sit near 0.5; applying the paper's power filter (≥ 0.5) keeps 35,000 of them and the distribution splits into supported (> 0.6) and contradicted (< 0.4) structures. Above 0.7 with a CI width under 0.1 on both arms is strong agreement. Always read the n: a ROC of 1.0 from six nucleotides means little; a power-filtered validation has a median of 56 paired and 17 unpaired nucleotides with data.

Conservation

phastCons 100-way (vertebrates) and 17-way (primates), per arm. Most of the dsRNAome is Alu-derived and primate-specific, so a conserved dsRNA (both arms > 0.5) is unusual and worth a closer look.

Repeat status

Over 90% of predictions involve a repeat, mostly inverted Alu pairs. Non-repetitive dsRNAs are rarer and more often conserved.

Rules of thumb

• Immune-sensor candidates: long arms (≥ 300 nt), ≥ 90–96% paired, strongly negative energy. Then check the Probing Model and, in the dsRNA Browser, whether any experiment supports the duplex.
• ADAR substrates: GTEx or Stability high confidence plus editing on both arms.
• Regulatory candidates in a specific gene: look at where the arms fall (3′ UTR, intron, exon) in the IGV gene track, and whether an eCLIP peak overlaps an arm.
• Prioritising a long list: sort by the number of models calling high confidence (the "# Models High-Conf" column under All columns), then by editing counts.

Glossary

dsRNAscan

The pipeline that produced the predictions: overlapping 10 kb windows scanned for inverted repeats that RNAduplex folds into a duplex of ≥ 25 bp at ≥ 70% pairing. Code at github.com/Bass-Lab/dsRNAscan.

i-arm / j-arm

The two halves of the inverted repeat. i is upstream (lower genomic coordinate); j is downstream. On the minus strand the j-arm is transcribed first.

Loop

The genomic distance between the arms. It can be a few nucleotides or many kilobases; long loops are typical of intronic Alu pairs.

Percent paired

Fraction of nucleotides in the predicted duplex that are base-paired (including G-U).

Longest helix

The longest run of consecutive base pairs with no bulge or mismatch.

Energy

RNAduplex free energy of the duplex in kcal/mol. More negative = more stable. Scales with length, so compare energies between dsRNAs of similar size.

RNAduplex length

Arm lengths after RNAduplex trims unpaired ends; shown in the dsRNA Browser next to the original lengths.

Edited (paper definition)

A dsRNA is called edited when more than one REDIportal site falls on each arm, counting sites on either strand. 1,236,563 of the 5.1 million predictions are edited by this definition, over 99% of them in repetitive elements.

Stranded vs unstranded editing sites

Stranded sites are REDIportal sites on the same strand as the dsRNA; unstranded sites are on the opposite strand. Both count as evidence, for this reason: an inverted repeat is complementary on both strands, so every prediction on the plus strand has a mirror-image prediction on the minus strand. If one strand is transcribed and clearly edited, the opposite-strand prediction would fold the same way and be edited too, were that strand expressed. Its lack of same-strand sites reflects expression, not structure, so the resource credits it with the editing seen on the other strand.

Intermolecular editing

Editing clusters that cannot be explained by a duplex within one transcript and are explained by pairing between two different transcripts.

ROC-AUC

Area under the receiver operating characteristic curve. Here: the probability that a randomly chosen predicted-unpaired nucleotide has higher reactivity than a randomly chosen predicted-paired one.

CI width

Width of the 95% confidence interval on the ROC-AUC (Hanley & McNeil 1982). Narrow = many data points.

Detection power

Estimated power to detect the observed paired/unpaired reactivity difference given the sample size.

phastCons

Per-nucleotide probability of being in a conserved element, 0–1, averaged over an arm.

FORNA

The ViennaRNA structure-drawing tool used in the dsRNA Browser and linked from IGV and the table.

BEDPE

Paired-end BED: one line per dsRNA with both arms' coordinates (0-based, half-open), the format bedtools pairtobed reads.

FAQ and troubleshooting

Citing and contact

If you use the dsRNAome in published work, please cite:

Andrews RJ and Bass BL. dsRNAscan maps human dsRNAome revealing conservation, intermolecular dsRNA, and determinants of ADAR dependency.

Questions, bug reports and feature requests: ryan.andrews@biochem.utah.edu, bbass@biochem.utah.edu.

The original user guide published with the paper (Document S1) is available as a PDF. It describes an earlier version of the IGV page; this page is the current guide.