RNA editing sites no longer rest on A-to-G alone: a two-enzyme assay plus InoDREAM yields 590,249 high-confidence inosine sites in HEK293T cells
Synopsis
Locating RNA inosine sites has long relied on A-to-G base transitions in sequencing, which are hard to separate from DNA mutations and sequencing errors. REX-seq couples EndoV cleavage at inosines with XRN-1 degradation of fragments bearing 5′ phosphate ends, yielding cleavage signatures while preserving gene-expression quantification; in HEK293T cells it identified 590,249 high-confidence inosine sites, 84,645 of them not previously reported. Trained on these sites, InoDREAM reached an AUROC of 0.965 and average precision of 0.964 on a held-out test set, against 0.921 and 0.890 for EditPredict.
Interpretation
RNA inosine sites can now be read from two directions at once, a cleavage signature and an expression measurement, instead of relying only on A-to-G base transitions. Earlier EndoV-based enrichment methods altered the original stoichiometric ratios of editing sites, making absolute quantification impossible, and lost gene-expression profile information during enrichment. In HEK293T cells, REX-seq identified 590,249 high-confidence inosine sites, 84.2% of which overlapped REDIportal records; expression levels of inosine-containing genes quantified by REX-seq correlated with conventional RNA-seq at r = 0.99.
The two-enzyme method has EndoV cut downstream of an inosine, after which XRN-1 removes the downstream fragment carrying a 5′ phosphate. EndoV cleavage generates 5′ phosphate termini, and XRN-1 is a 5′-monophosphate-dependent 5′→3′ exoribonuclease that degrades only RNA fragments bearing such termini. With a 60-nucleotide synthetic RNA containing inosine, EndoV cleaved specifically at the inosine site and XRN-1 degraded only the downstream 5′-phosphorylated fragment, while RNA strands lacking inosine and carrying 5′-OH termini stayed intact.
InoDREAM uses the high-confidence REX-seq sites as training data and predicts from a 201-nucleotide sequence window whether a given adenosine will be edited. The earlier EditPredict was trained on about 4.67 million computationally derived editing sites from REDIportal, most lacking experimental validation. Trained on the same REX-seq high-confidence sites and evaluated on a held-out test set, InoDREAM reached an AUROC of 0.965 and average precision of 0.964 versus 0.921 and 0.890 for EditPredict; in a 51-bp context InoDREAM reached an AUC of 0.953 and AP of 0.950 versus 0.591 and 0.550 for ATTIC.
Perspective
The demonstrated scope is HEK293T cell mRNA with a human brain tissue comparison, plus model evaluation on REDIportal multi-tissue sites; for low-abundance novel sites, the authors state that site-specific orthogonal validation such as targeted amplicon sequencing or RT-PCR remains the gold standard for confirmation.
XRN-1 digestion introduces a mild 5′ bias confined to roughly the first 50 bp of clustered editing regions, with Δ below 0.1, and editing levels beyond 50 bp agree closely with RNA-seq; the cis-element enhancer and silencer motifs and the RBP-binding motifs are model-derived and lack direct experimental validation; siRNA knockdown of MSI1 and PTBP1 produced only marginal global editing changes (mean Δ of –0.049 and –0.014).