UNCOVERseq Exposes Off-Targets Missed by Established Standards

The method, UNCOVERseq, is a cell-based assay built on the familiar principle of capturing a double-stranded DNA tag at nuclease-induced breaks, but rebuilt end to end (see Figure 1). A blocking oligonucleotide removes an artefact that had been consuming 37–67 per cent of sequencing reads in the original GUIDE-seq protocol, and a new alignment and statistical pipeline replaces string matching with a glocal Needleman-Wunsch implementation. Across 48 guides, the reworked workflow nominated a median of 132 off-target sites per guide compared with 23 for GUIDE-seq – between 60 and 4,150 per cent more.

Sequencing depth decides what can be seen

Volume alone proves nothing, so the authors built a benchmarking dataset in which nominated sites were confirmed by ultra-deep amplicon sequencing. Against that dataset, the assay recovered 97.6 per cent of confirmed off-target edits at 78 per cent precision.

The comparison is unflattering for the methods most often cited in regulatory filings (see Figure 2). GUIDE-seq recovered 35.9 per cent of true positives, BLISS 23.1 per cent and DIG-seq 33.0 per cent, while the biochemical in vitro assays SITE-seq and AID-seq reached 83.8 and 80.4 per cent sensitivity but at precision of only 32.1 and 29.7 per cent respectively. In other words, the cell-based assays in common use missed the majority of confirmable events, and the biochemical assays that found them buried the findings in long, largely false-positive lists.

The practical requirements are unusually well specified, which matters for anyone writing an assay into a development plan. UNCOVERseq needs roughly 150,000 genome equivalents of input DNA, and sequencing depth turns out to be the binding constraint: 50,000 reads per sample suffice for high-frequency events, but detecting off-targets at 0.01–0.02 per cent indel frequency requires at least two million. The authors recommend more than 500,000 reads as a minimum. Replication is not optional either – without biological triplicates, 25–40 per cent of the sites the tiering system flagged as high priority were missed or misranked.

To ask whether any of this survives the move into clinically relevant cells, the team screened 192 guides, selected six spanning the full specificity range, and edited CD34+ haematopoietic stem and progenitor cells from a healthy donor using mRNA delivery of high-fidelity S. pyogenes Cas9, adenine and cytosine base editors, and a prime editor. On-target indel editing in the stem cells ranged from 4.4 to 88.4 per cent; intended base conversion reached 16.5–75.9 per cent for the adenine base editor and 0.78–32.9 per cent for the cytosine base editor. Prime editing produced almost no on-target activity without further pegRNA optimisation and was excluded from the off-target analysis.

Between 2.4 and 34.5 per cent of nominated sites converted to confirmed editing in the stem cells, with frequencies from 0.06 to 88 per cent. The conversion rate fell as guide specificity rose, and for the highest-specificity guides the only site confirmed was the on-target – the off-target burden of a well-chosen guide in a primary cell simply drops below what current sequencing can measure. That is a reassuring finding, but it is also a measurement limit rather than a biological zero.

High-fidelity nucleases carry the lighter burden

The most clinically pointed result concerns editor choice. Sites nominated on the basis of double-strand breaks predicted where base editors would act, with rank-order correlations of 0.77 for the adenine and 0.78 for the cytosine editor, so a single nuclease-based nomination can serve both modalities.