Single-cell libraries already hold RNA-end evidence: scPolyA turns polyadenylation sites into comparable cell and tissue measurements
Synopsis
Conventional 3′-end single-cell and spatial transcriptomes were built to count gene expression, yet a small subset of reads carrying untemplated poly(A) tails marks where RNAs actually end. scPolyA uses those tail reads to discover polyadenylation sites de novo, then sets UMI-counting intervals from total-read coverage and reports per-cell or per-spot terminal-site usage, 3′ UTR length and intronic polyadenylation. Across eight single-cell datasets, 63.3% of its sites fell within 25 nt of PolyA_DB 3, above polyApipe, scAPAtrap, Sierra and SCAPE and close to DeepPASS-filtered SCAPTURE.
Interpretation
Routine 3′-end single-cell and spatial transcriptomes can be read as cell- and spot-resolved maps of RNA-end choice rather than only as gene-abundance measurements. These libraries were analysed predominantly as measurements of gene abundance, and polyadenylation work relied on reference atlases or dedicated callers that treated site boundaries and molecule-counting intervals separately. Across eight single-cell datasets spanning four protocol families, two species and six tissues, a median 63.3% of scPolyA sites lay within 25 nt of PolyA_DB 3, versus 23.9% for polyApipe, 26.6% for scAPAtrap, 32.4% for Sierra, 35.0% for SCAPE and 59.4% for DeepPASS-filtered SCAPTURE.
Site discovery and molecule counting are linked: poly(A)-tail reads anchor candidate ends, and total-read coverage sets the width of the interval counted for each end. Fixed upstream windows, such as polyApipe's 250-nt default, and coverage-accumulation regions assign different UMIs to the same alternative end. Across eight spatial sections, a median 89.9% of scPolyA sites lay within 25 nt of PolyA_DB 3, versus 71.3% for polyApipe, 55.6% for Sierra and 36.5% for scAPAtrap; in a Slide-seqV2 section where recognisable tail reads were about 0.03% of alignments, scPolyA identified 1,853 sites against 1,217 for polyApipe.
The companion deep-learning module DeepCPS adds evidence when tail-read support is sparse: one model scores how closely a candidate resembles reference sites, the other predicts the cleavage nucleotide from sequence. Ranking by tail-read count alone can exclude genuine sparsely supported sites, while relaxing the threshold admits many downstream A-rich candidates that resemble internal priming. On held-out chromosomes DeepCPS reached a scoring AUROC of 0.940 (95% bootstrap interval 0.937–0.942), against 0.909 for DeepPASS and 0.832 for DeepPASTA; at 4,115 reference-supported sites with at least ten tail reads it matched the observed modal read endpoint exactly at 62.1%, against 21.5% for APARENT.
Perspective
The demonstrated scope is barcoded, UMI-tagged 3′-end single-cell and spatial alignments, with the main benchmark using read-based discovery with DeepCPS filtering disabled; the cutaneous squamous cell carcinoma analysis uses 47,061 cells from ten patients and spatial sections from six patients, with regions defined as marker-enriched rather than purified cell types.
Reference-atlas agreement measures concordance with existing annotations, so an unmatched candidate is not necessarily unsupported; usage precision depends mainly on the number of observed molecules, and relaxing the tail-read threshold expands the site catalogue far more than the count of genes passing a terminal-UMI depth screen. Tissue-level 3′ UTR differences shift with the reference cell population, and gene-level agreement between single-cell and spatial modalities is modest.