A 'Silent' DNA Tag Can Change RNA Processing

Reading a person's DNA is now fast. Explaining what each change means is not. Many changes land in a bin called variants of uncertain significance, which means nobody knows whether they matter. Some matter because they disturb the gene's RNA message instead of the protein it encodes. A study published in Nature Communications on October 15, 2026 offers a new way to measure that, plus a warning about a common lab shortcut. ^[1]

The finding

Silent tags can disturb RNA

A gene is copied into RNA, and the cell then splices that RNA, cutting out unwanted stretches and joining the rest. A synonymous change swaps one DNA letter for another without altering the protein's amino acids, so it was long assumed to be harmless. The authors point out that such changes can still affect how RNA is processed or how long it lasts. ^[1]

Two earlier approaches to measuring this each have limits, according to the authors. Reporter assays test variants on a separate piece of DNA, away from the gene's natural surroundings. Saturation genome editing works inside the real gene, but relies on a repair route that works poorly in many cell types and needs a separate donor DNA pool for each 200-letter stretch. ^[1] Saturation genome editing has been used successfully to classify variants in the cancer gene BRCA1. ^[3]

Prime editing offers a third route. It uses a modified Cas9 protein fused to a reverse transcriptase (an enzyme that copies RNA into DNA). A guide called a pegRNA both points to the target site and carries the new sequence to write there, with no double-strand break and no donor template. ^[2]

Checking the tag as well as the edit

A variant's effect on RNA is part of its story, and cPRIME gives a way to measure it in the gene's own setting. The study also shows that a tag meant to improve accuracy can add its own artifact, and that a quick computational check can flag it. ^[1]

The evidence

Measuring splicing with cPRIME

How cPRIME works. The new method is called cPRIME, short for prime editing with cDNA readout. The team used two leukemia cell lines and edited NF1, a gene frequently mutated in cancer. After weeks of growth they collected genomic DNA and RNA, converted the RNA into cDNA (a DNA copy of the RNA), and sequenced the edited exon from both. Dividing a variant's read share in cDNA by its share in genomic DNA gives the cPRIME score. A score near 1 means the RNA looks like the DNA. A lower score means RNA carrying that variant is under-represented. ^[1]

The tag problem. Prime edits are rare, and a single-letter edit looks a lot like a sequencing mistake. So the team added a silent "marker" change next to each edit, a trick several groups developed. Background for single-letter edits in unedited cells was 0.037% (± 0.041) in KBM7 cells and 0.045% (± 0.049) in K562 cells. The edit-plus-marker combination had a background 1,280 times lower in KBM7 and 1,120 times lower in K562. With markers, nearly 90% of variants were quantifiable above background, versus about 10% without. ^[1] An earlier study found that silent changes near an edit can improve prime editing by evading the cell's mismatch repair machinery, so markers do double duty. ^[4]

First results. In a library of 98 guides aimed at one exon (exon 27), 69% of intended edits in KBM7 and 76% in K562 passed a detection threshold of 10 reads per million. Edits that create a premature stop signal had average cPRIME scores of 0.19 and 0.30, consistent with the cell destroying faulty messages through nonsense-mediated decay. Most other variants scored near 1, with three exceptions. ^[1]

The surprise. Those three non-stop variants scored below 1 when paired with their markers. The team ran a deep learning splice predictor called Pangolin, which scores how much a change strengthens or weakens a splice site. ^[5] All three combinations had predicted splice-gain scores above 0.5, suggesting a new splice site. Removing the marker cut the predicted splice-gain score from 0.57 to 0 for one variant and from 0.75 to 0.22 for another, and from 0.6 to 0.46 for the third. Only the third still lowered cDNA reads without its marker, and less than before. One variant alone showed no drop, and neither did the same edit paired with two different marker letters. ^[1] So specific edit-and-marker pairings can create splice signals that neither piece creates alone.

Scaling up. A second library of 341 guides covering 178 intended edits across six exons had its markers checked with Pangolin during design. Scores in the two cell lines agreed with a correlation of 0.91, and exon 27 scores matched the first screen at 0.96. The share of edits detected above threshold ranged from 7.3% to 38% across exons in KBM7 cells, with about 35,000 cells per guide. A second predictor, SpliceAI, agreed closely with Pangolin (correlation 0.93). ^[1] SpliceAI is a deep neural network that predicts splice junctions from sequence. ^[6]

Clinical check. Against ClinVar labels, cPRIME scores separated 3 pathogenic variants from 17 benign ones perfectly (AUC 1.0) in both cell lines. These variants were a selected set, enriched for possible splice-active sites, not a representative sample of ClinVar. The authors call this a proof of principle, not a performance benchmark. A statistical grouping method sorted variants into RNA-neutral, intermediate and depleted classes: 46, 17 and 7 variants in KBM7, and 44, 11 and 7 in K562. ^[1]

Editing efficiency is low and uneven, which limits how many variants can be measured. A low score shows that a particular RNA form is depleted, not why. The screen reads one targeted exon at a time, and results come from one gene in two cell lines. We did not independently replicate any of this. ^[1]

Future directions

The authors name isoform-specific RT-PCR and long-read transcript sequencing as ways to understand low scores. Experiments that block RNA decay could help separate transcript loss from altered splicing. For variants with raised scores, their suggested m6A mapping and RNA-stability assays could test the proposed mechanism, which the screen's RNA abundance readout cannot resolve alone. ^[1]

Marker design is another practical starting point. The authors recommend screening synonymous markers with a splice predictor, as they did for the multi-exon library. Follow-up work could examine how that filter changes the eligible marker set, while keeping the distinction between a predicted splice effect and an experimentally measured one. ^[1]