Leveraging long read RNA-seq to decipher neuronal regulation of alternative polyadenylation

Abstract

Alternative polyadenylation (APA) is a pervasive co/post-transcriptional process affecting >70% of human genes which greatly enhances transcriptome diversity. Here we introduce APALORD (Alternative Polyadenylation Analysis of LOng-ReaDs), a computational framework for APA analysis of long-read (LR) RNA-seq data. For gene level APA analysis, we developed a novel method by comparing distributions of cleavage sites (CSs) between conditions. Applying APALORD to direct RNA-seq (DRS) data from human embryonic stem cells (hESCs) and derived neurons, we identified annotated and novel PASs with high positional accuracy. The transcriptome-wide APA lengthening in neurons was associated with increased usage of stronger PASs enriched for the canonical AAUAAA polyA signal, upstream UGUA and downstream GU/U-rich motifs. PAS strength was negatively correlated with the number of PASs per gene and positively correlated with gene expression levels. Application to Drosophila embryo samples led to a substantial expansion of PAS annotations and uncovered conserved features of APA regulation. A novel, low-abundance class of transcripts that do not map directly to identified PAS was identified in both species and found to display APA regulation trends. Together, APALORD offers a robust framework for high-resolution APA analysis using LR RNA-seq data across species and biological contexts.

Long-read RNA sequencing enables precise mapping of RNA 3′ ends. Here, the authors develop APALORD, a computational framework for high-resolution alternative polyadenylation analysis that uncovers a preference for stronger polyadenylation site selection during neural differentiation.

Introduction

During mRNA processing, the cleavage and polyadenylation (CPA) machinery recognizes specific cis-acting sequence elements surrounding the cleavage site (CS), to collectively define a polyadenylation site (PAS)^1–3. The core element is the hexamer sequence AAUAAA (or its variants), bound by the CPA specificity factor (CPSF)^4,5. Additional elements include the UGUA motif (upstream sequence element, USE) bound by cleavage factor I (CFI)^6,7, and a GU/U- rich downstream sequence element (DSE) bound by cleavage stimulation factor (CSTF)^8,9. Alternative polyadenylation (APA) is a tightly regulated co/post-transcriptional process that generates alternative 3′ ends for genes¹⁰. Most human genes generate more than one 3′UTR mRNA isoform due to APA¹¹. For broadly expressed genes, tissue-specific patterns of APA are found across phyla, with distal PAS selection preferred in the brain and proximal PAS selection preferred in testis^11–13. Cell stress, development, and differentiation regulate APA patterns^14,15. Roles for alternative 3′UTR mRNA isoforms include nervous system development, axon growth, and mating behavior^16–19. APA patterns are disrupted in diseases^20,21 and APA quantitative trait loci contribute to disease heritability²².

While transcriptome-wide APA quantification is feasible with conventional Illumina-based short-read RNA sequencing, it faces trade-offs between gene body coverage and 3′ end accuracy or depth²³. APA analysis using standard, non–3′ end tag-specific short-read library preparations often lacks precise information about 3′ ends. In contrast, 3′ end tag-specific short-read library preparations improve both the accuracy and depth of 3′ end measurements^24–28. However, poor gene body coverage makes it less suitable for widespread application and limited read lengths (typically 50–300 bp) can lead to ambiguous alignment to specific genes or isoforms²⁹. Consequently, short-read APA analysis typically relies on well-annotated PAS references to estimate 3′UTR isoform usage or requires predefined regions to detect APA changes across conditions^30–32. Poor PAS annotation and uneven coverage in short-read data frequently distort gene-level APA analyses.

The advent of long-read (LR) RNA sequencing a decade ago initially aided the discovery of rare isoforms poorly resolved by short-read methods, though its low depth and high error rate limited broader applications^33,34. Recent advances in high-throughput LR RNA-seq have significantly improved its accuracy and depth, establishing it as a powerful tool for RNA profiling^35,36, offering reduced ambiguity and comparable depth to traditional short-read RNA-seq techniques³⁷. LR sequencing captures full-length RNA or cDNA sequences, enabling precise gene and isoform assignment as well as accurate quantification^38–40. Moreover, LR data exhibit high-accuracy 3′ end coverage and more consistent depth over longer regions (>1 kb) compared to short-read data, facilitating data-driven PAS detection and accurate quantification of 3′UTR isoforms. Direct RNA-seq (DRS), which sequences the full length of the native RNA strand from the end of the polyA tail without any additional chemical modifications, offers the most accurate 3′ end information^41,42. DRS data avoids artifacts stemming from PCR amplification and internal polyA priming, which plague the aforementioned APA analysis. Currently, tools for transcriptome-wide APA analysis of LR RNA-seq that fully leverage its high-resolution 3′ end information are lacking.

For genes with more than one PAS, APA can occur in distinct genomic contexts. When alternative PASs reside within the same terminal 3′UTR exon, the resulting isoforms share identical 5′UTR and coding exons, and this mode is called last-exon tandem APA⁴³. Last-exon tandem APA isoforms produce proteins with identical amino acid sequence. When alternative PASs reside in an upstream exon or intron, usage of these PASs generates transcripts with distinct terminal exons compared with isoforms terminating at the canonical last exon. This type of APA is known as alternative last exon (ALE) APA, leading to production of truncated protein isoforms and potentially divergent biological functions^2,44. In tandem APA genes, when there is increased usage of downstream PASs relative to upstream ones, this shift is referred to as 3′UTR lengthening. In ALE APA genes, usage of downstream PASs generally produces longer transcripts, whereas the effect on 3′UTR length is variable. For simplicity, we refer to increased usage of downstream PASs as lengthening and increased usage of upstream PASs as shortening in both last‑exon tandem APA and ALE APA genes. Genes with ALE could also still have multiple PASs within each terminal exon, which would be tandem APA within ALE. Thus, we classify genes with ALE as “mixed APA” genes in this investigation.

Neuronal tissues are characterized by an expanded repertoire of 3′UTR lengthening isoforms compared with other tissues¹³. Transcriptome‑wide analyses have consistently shown a global trend toward lengthening during neuronal differentiation and in mature neurons relative to other cell types^45,46. Neuronal APA regulation is governed by a combination of core CPA factors, including components of the CFIm complex such as CPSF5, whose depletion leads to widespread 3′UTR shortening²¹. In addition, neuron‑enriched RNA‑binding proteins such as the ELAV/Hu family have been shown to promote the usage of distal PASs and contribute to the extended 3′UTR landscape characteristic of neuronal transcriptomes^17,47,48. Dysregulation of APA and its regulators has been implicated in neurodevelopmental processes, neural cell fate transitions, and neurological disease^49,50.

PAS usage (PAU) changes at distal and/or proximal PASs are widely adopted as an indicator of APA trend between conditions in existing APA analysis methods^31,32. PAU for each PAS is defined as the percentage of reads assigned to that PAS among all reads mapped to the host gene. An increase in distal PAS usage (dPAU) is interpreted as APA lengthening trend and has been widely used to reveal transcriptome-wide distal PAS preference in neurons and brain when compared with other tissues. Conversely, increased proximal PAS usage (pPAU) indicates shortening gene and is commonly used to monitor transcriptome-wide elevated utilization of proximal PASs in cancer cells and activated T cells. However, this type of APA analysis would miss APA events occurring at non-distal/proximal sites in genes using more than two PASs. Distal/proximal usage ratio has also been utilized to indicate gene-level APA regulation^51–53, but it cannot identify APA changes that affecting intermediate sites. Another popular concept is to use weighted PAS position (ψ)³⁰ or 3′UTR length^13,46. This type of gene-level APA analysis reflects the general PAS positional shift in APA genes, but they fail to identify events in which PAS usage is redistributed among three or more PASs without altering the mean position or average 3′UTR length. Specifically, shifts in last-exon usage in ALE APA genes do not necessarily result in the change of 3′UTR length, therefore the weighted 3′UTR length method does not apply.

To address the limitations of existing methods, here we propose a 3′ end distribution-based approach leveraging the high-resolution 3′ end information in LR RNA-seq data: comparing the cumulative distribution of 3′ CSs in a gene between two conditions using an adjusted Kolmogorov–Smirnov test (KS test). Based on this gene-level APA analysis framework, we developed APALORD (Alternative Polyadenylation Analysis of Long Reads), an R-based tool for quantifying differentially regulated APA events from LR RNA-seq datasets. APALORD performs statistical analyses at individual PAS level to identify differentially used PASs and detects trends toward shortening or lengthening at the gene level across conditions. Using DRS data from human embryonic stem cells (hESCs) and hESC-derived neurons, we confirmed a global trend of APA-mediated lengthening during neural differentiation. We provide evidence that this trend is driven by increased preference in neurons for selecting stronger PASs that feature the canonical AAUAAA hexamer, an upstream UGUA and downstream GU/U-rich motifs. Because APALORD identifies PASs directly from LR data, independent of PAS annotations, we further demonstrate its utility in a poorly annotated species, Drosophila melanogaster. By performing DRS sequencing of Drosophila embryos, APALORD uncovered thousands of unannotated PASs and revealed a systematic shift of PASs toward non-A CS in the current Ensembl annotation⁵⁴. We also validated the utility of APALORD on long read cDNA datasets from PacBio platform. With the growing adoption of LR RNA-seq, we anticipate APALORD will become a widely used tool for investigating 3′ end transcript dynamics across diverse biological systems.

Results

APALORD computational workflow

The APALORD R package analyzes LR RNA-seq data to provide statistical assessment of differential expression at the individual PAS level and a readout of APA as shortening/lengthening trends at the gene level. LR RNA-seq reads are first aligned to the host genome using minimap2⁵⁵, followed by assignment to transcribed genes using IsoQuant⁴⁰. The output data from IsoQuant is then loaded into R for APA analysis with APALORD (Fig. 1). APALORD requires the same GTF reference file used by minimap2 and IsoQuant to extract general gene information—such as gene name, stop codon position, and strand—but does not rely on existing 3′UTR annotations. Alternatively, reads aligned to the genome in BAM file format can be loaded into APALORD directly and assigned to a host gene using a built-in function based on Bambu³⁹. Using precise 3′ end positions identified by IsoQuant from aligned reads, APALORD identifies PASs, defined as peaks where most reads from a given PAS cluster within a window (this window size is adjustable in APALORD) around the peak. Reads assigned to the host gene are then parsed to a nearby PAS identified by APALORD, while those that could not be parsed are defined as leaky transcripts.

Fig. 1. APALORD workflow.

LR RNA-seq reads are first aligned to the reference genome (with genome sequence and gene annotation as reference) using minimap2, generating BAM files. Aligned reads are then assigned to unique host genes using IsoQuant. APALORD takes these reads assignment results as input; alternatively, BAM files can be loaded directly into APALORD, where reads are assigned to genes with a built-in function based on Bambu. APALORD subsequently identifies PASs from the data and quantifies PAS usage (PAU) within each gene. The framework enables multiple levels of APA analysis, including PAS-level PAU change, gene-level APA changes, and proximal/distal PAU, along with transcriptome-wide and gene-specific visualization.

The APALORD workflow (Fig. 1) includes analysis that employs DEXSeq⁵⁶ to identify differentially used PAS between two conditions (e.g., control vs. experimental), each with at least two replicates. This analysis detects PASs with significant usage differences, regardless of their relative position within the host gene. A second gene-level APA analysis type reports “APA change” by comparing distributions of CSs along the gene body between conditions. To exclude the variance within each PAS, all CSs from reads or transcripts assigned to the same PAS are collapsed to the PAS peak coordinate, which is the most utilized CS within the PAS. PAU change at any PAS alters the cumulative distribution, and the maximum difference between conditions can be used to quantify gene-level APA magnitude, which we call APA change. This constitutes a gene-level APA lengthening (i.e., distal PAS preferential selection when APA change >0.1) or shortening (i.e., proximal PAS preferential selection when APA change < −0.1) trend between conditions. Because collapsing CS coordinates within each PAS produces extensive ties that violate standard KS test assumptions and inflate significance estimate, we implement a tie-adjusted KS test, specifically tailored for the 3′ end coordinate data. The test yields a p-value based on a normal approximation and a signed maximum difference (ranging from −1 to 1) between the two empirical cumulative distribution functions (ECDFs), serving as an indicator of gene-level APA change. These data are visualized transcriptome-wide using a volcano plot (APA change vs. adjusted p-value). APA regulated events can also be profiled at either the distal or proximal PAS of each gene. For a gene of interest, 3′ end distribution of all the reads assigned to the gene are plotted for each condition, and PAU change across conditions are visualized for all identified PASs within the gene using bar plots.

Applying APALORD to identify PASs in hESC neural differentiation DRS data

To demonstrate the performance of APALORD, we differentiated human NGN2-inducible H9 hESCs into neurons^57,58 and conducted DRS on the Nanopore platform. Samples were collected from undifferentiated cells (D0) and after 7 days of neural differentiation (D7) (Fig. 2a). A total of 13,869,924 polyadenylated reads from hESCs were uniquely assigned to 23,821 genes, and 8,623,430 reads from neurons were assigned to 23,563 genes, corresponding to 27,106 genes in total across both conditions. From triplicate datasets for both D0 and D7, we identified 45,119 CSs, each supported by ≥50 polyadenylated transcripts cleaved at the same genomic position within a host gene. Because heterogeneity in the exact CS position is known to occur⁵⁹, we examined the distribution of 3′ end positions surrounding the 45,119 CSs (Supplementary Fig. 1a). We observed that most cleavage events clustered within 50 nt distance of the most frequently used CS, whereas positions farther away (−100 to −44 nt and +49 to +100 nt) together accounted for <1% of transcripts for 99% of the CSs. Given that the canonical AAUAAA polyA signal and its variants defining a PAS are typically located ~20 nt upstream of the CS⁶⁰ and that CSs occur mostly downstream of the polyA signals, we set 20 nt as the minimal distance for assigning reads uniquely to the nearest PAS. Within the ±20 nt window, the most frequently used CS was defined as the PAS peak to represent the location of the identified PAS. To determine the optimal window size, we ran APALORD using bin sizes ranging from 20 to 50 nt. PASs identified from hESCs and hESC-derived neurons DRS data were compared with annotated PASs and classified into three categories: (i) “match” PASs, located at the exact position of an annotated PAS within the same host gene; (ii) “proximity” PASs, located within the specified window of an annotated PAS; and (iii) “novel” PASs, lacking any annotated PAS within the window. Single-nucleotide PAS annotations from GENCODE⁶¹ and PolyA_DB⁶² were used as references. With decreasing bin size, APALORD identified an increasing number of novel PASs, whereas the numbers of “match” and “proximity” PASs remained largely unchanged (Supplementary Fig. 1b). Sequence analysis of novel PASs revealed that the fraction containing an upstream canonical or variant AWUAAA (W = A or U) motif decreased with smaller bin sizes (Supplementary Fig. 1c). Notably, at a 20-nt bin size, a substantial fraction of novel PASs instead exhibited a AWUAAA motif downstream of the PAS peak. Further examination showed that these PASs with a downstream AWUAAA motif generally lacked an upstream AWUAAA motif (Supplementary Fig. 1d), suggesting that they may represent falsely split CSs originating from nearby strong PAS when smaller windows are used to identify PASs. By calculating the proportion of novel PASs containing an upstream AWUAAA hexamer within 50 nt, we found that this fraction approached saturation at a bin size of 50 nt (Supplementary Fig. 1e). Therefore, we selected a 50 nt bin size for all subsequent analyses to prioritize high-confidence PAS identification.

Fig. 2. Performance of APALORD in PAS identification using DRS from hESCs and hESC-derived neurons.

a Schematic of neural differentiation from NGN2-inducible H9 ESCs (D0) to induced neurons (D7). Dox: doxycycline; IM: induction medium; CM, cortical neuron medium. b Classification of PASs identified by APALORD from DRS of hESCs and neurons samples. match: exact alignment with annotated PASs in the combined PolyA_DB and GENCODE reference; proximity: within 50nt of an annotated PAS; novel: outside of match or proximity criteria. c Distance distribution of match and proximity PASs relative to annotated PASs. Colors correspond to PAS classes in (b). d PAS classification stratified expression level. Identified PASs are grouped based on the average expression level (TPM) of transcripts originating from each PAS across samples (0–1, 1–10, 10–100, >100 TPM). TPM transcripts per million. e Usage of identified PASs in D0 and D7 samples. Box plots indicate 25th and 75th percentiles (box bounds), with whiskers extending to the most extreme data point within 1.5× the interquartile range (IQR); the central diamond denotes the median. f Distribution of gene biotypes for PASs, comparing novel PASs with all identified PASs. g Genomic distribution of novel PASs compared with annotated PASs in the DRS data.

After filtering out lowly expressed genes (<5 reads across all sample replicates), APALORD identified 34,081 high-confidence PASs in the dataset (each PAS with overall PAU ≥ 1%, supported by ≥5 reads). For each PAS, CSs within ±50 nt of the PAS peak were collapsed, and the nucleotide position with the highest read depth was selected as the peak to represent the PAS. We compared these PASs with single-nucleotide annotated PASs from GENCODE and PolyA_DB. Of the 34,081 PASs identified, 24,951 (73.2%) were classified as “match” PASs, occurring at the exact genomic position of an annotated PAS within the same host gene. An additional 6911 PASs (20.3%) were classified as “proximity” PASs, located within 50 nt of an annotated PAS, of which 3470 (50.2%) were within 5 nt of the annotated site (Fig. 2b). The distribution of distances between APALORD-detected PAS peaks and nearby annotated PASs further demonstrated the high positional accuracy of APALORD in pinpointing PAS locations (Fig. 2c; Chi-Square test, p = 0). The remaining 2219 PASs (6.5%) were classified as novel PASs (Fig. 2b).

We observed that novel PASs identified in both D0 and D7 samples exhibited lower expression levels compared with annotated PASs (match and proximity, Fig. 2d). Novel PASs also showed relative lower usage (median PAU of 4.61% in hESCs and 4.26% in neurons) compared to match and proximity PASs (Fig. 2e), indicating that novel PASs are generally weaker. These observations align with the expectation that APALORD′s sensitivity, enabled by precise 3′ end detection, facilitates the identification of lowly used weaker PASs. Among the 11,155 genes with at least one identified PAS, 8270 genes (74.1%) utilized two or more PASs, with an average of three PASs per gene. Notably, 5679 genes (over two-thirds of expressed multi-PAS genes) used three or more PASs (Supplementary Fig. 2a), consistent with previous reports^11,63,64. Of the 34,081 PASs identified, 32,284 (94.7%) were derived from protein-coding genes, among which 1325 (3.9%) were classified as novel. In addition, 1521 PASs (4.5%) were identified in long non-coding RNA (lncRNA) genes, with 762 (2.2%) representing novel PASs (Fig. 2f). This highlights that lncRNA 3′ end annotations remain less well characterized compared to protein-coding genes. Regarding the genomic distribution of our PAS annotation, 34.5% of novel PASs resided within annotated 3′UTR exons, followed by 30.0% in introns and 29.4% in 3′ intergenic regions, 2.1% in coding sequences (CDSs) and 0.9% in 5′UTR exons (Fig. 2g). Compared with annotated PASs, novel PASs were enriched in intronic and intergenic regions, which is expected, as LR RNA sequencing enables accurate read-to-gene assignment through long-range connectivity with upstream and downstream gene-specific exons.

PAS-level and gene-level APA analysis during neural differentiation of hESCs

Gene-level APA changes were quantified based on shifts of the cumulative usage of PASs to identify APA-regulated genes during neural differentiation (Fig. 3a). Analysis of genes with multiple PASs (Supplementary Data 1) identified 1256 lengthening genes (677 last-exon tandem APA genes and 579 mixed APA genes) and 79 shortening genes (18 last-exon tandem APA genes and 61 mixed APA genes) in differentiated neurons compared with proliferating cells (Fig. 3b, c; |APA change | > 0.1, adjusted p < 0.05). While lengthening events showed no significant difference between the two types of APA gene (Fisher′s exact test, p = 0.603), shortening events during neural differentiation were significantly enriched in mixed APA genes (Fisher′s exact test, p < 0.00001, odds ratio = 4.12). Among APA-regulated genes, 309 genes expressed transcripts from three PASs, followed by 263 genes with four PASs and 235 genes with two PASs (Supplementary Fig. 2b). Only one single PAS gene exhibited a marginally significant lengthening trend because of leaky transcript distribution (Supplementary Fig. 2c). Gene Ontology analysis of lengthening genes revealed enrichment in biological processes and cellular components related to neural biogenesis, neural activity and synaptic localization (Supplementary Fig. 2d).

Fig. 3. APALORD analysis of APA regulation during hESC neural differentiation.

a Schematic of gene-level APA analysis. APALORD applies the tie-adjusted two-sided Kolmogorov-Smirnov (KS) test to compare the cumulative PAU between two conditions. The signed maximum difference is reported as gene-level APA change. Volcano plots illustrating transcriptome-wide APA changes in two classes of multi-PAS genes: last-exon tandem APA genes (b) and mixed APA genes (c) during hESC neural differentiation. Red, lengthening genes (APA change (D7 – D0) > 0.1, adjusted p < 0.05); blue, shortening genes (APA change (D7 – D0) < −0.1, adjusted p < 0.05). d Differential distal PAS usage (dPAU) in multi-PAS genes during neural differentiation. Red, increased distal PAU (ΔdPAU >10 %, adjusted p < 0.05); blue, decreased dPAU (ΔdPAU < −10 %, adjusted p < 0.05). In (b–d), gene-level APA analysis was performed using the same two-sided test in APALORD, followed by Benjamini–Hochberg false discovery rate (FDR) correction for multiple testing. e Correlation between dPAU change and gene-level APA change in multi-PAS genes. Colors of dots indicate gene-level APA categories (lengthening, shortening or non-regulate) as defined in (b, c). Pearson′s correlation: r = 0.8, p < 2.2e-16. f Overlap of lengthening genes identified by distal PAS analysis and gene-level APA analysis. g Differentially used PASs during neural differentiation. Red, increased PAU (PAU(D7) > PAU(D0), adjusted p < 0.05); blue, decreased PAU (PAU(D7) < PAU(D0), adjusted p < 0.05). PAS-level was performed using the two-sided likelihood ratio test in DEXSeq and p values were adjusted for multiple testing using FDR. h Overlap of genes containing at least one differentially used PAS (g) and genes identified as APA-regulated at the gene-level (b, c). i Representative last-exon tandem APA gene (NAV1). Shown are DRS coverage tracks, cumulative PAU curves, 3′ end density distributions, and PAU changes (D7 – D0) at individual PASs. j Representative mixed APA gene (SYNGR1). Shown are DRS coverage tracks, cumulative PAU curves, 3′ end density distribution and PAU changes (D7 – D0) at individual PASs.

To assess whether gene-level APA-regulated events identified by APALORD are consistent with previously established approaches, we performed APA analysis using dPAU/pPAU changes, which is common in the field^31,65. Focusing on distal PASs, 208 genes exhibited significantly upregulated distal PASs (ΔdPAU>10%, adjusted p < 0.05), whereas 20 genes showed significantly decreased distal PAS usage (ΔdPAU < −10%, adjusted p < 0.05) (Fig. 3d). Gene-level APA changes were strongly correlated with PAU change at distal sites (r = 0.8, Pearson′s correlation, p < 2.2e-16), indicating that a substantial fraction of the gene-level APA changes is driven by dPAU changes (Fig. 3e). Among the 208 lengthening genes identified based on dPAU changes, 189 were also identified as lengthening genes by gene-level APA analysis (Fig. 3f), whereas 18 of 20 shortening genes were consistently identified by both approaches (Supplementary Fig. 3a). In contrast, analysis of proximal PASs identified 62 genes with decreased pPAU (ΔpPAU < −10%, adjusted p < 0.05) and 4 genes with increased pPAU (Supplementary Fig. 3b; ΔpPAU >10%, adjusted p < 0.05). Gene-level APA changes were negatively correlated with PAU change at proximal PASs (Supplementary Fig. 3c, r = −0.65, Pearson′s correlation, p < 2.2e-16), and 59 of 62 lengthening genes and all 4 shortening genes were consistently identified with gene-level APA analysis (Supplementary Fig. 3d, e). Together, these results demonstrate that dPAU/pPAU-based methods identified APA events are mostly covered by the gene-level APA analysis method that we develop in APALORD. For other gene-level APA events that could not be explained by changes at distal/proximal PASs, differential usage of intermediate PASs may contribute to the observed gene-level APA changes. To further evaluate this, we performed PAS-level APA analysis (Supplementary Data 2) and identified 2789 PASs enriched in neurons (log2FoldChange > 0, adjusted p < 0.05) and 2722 PASs enriched in hESCs (log2FoldChange < 0 and adjusted p < 0.05), corresponding to 2926 genes (Fig. 3g). Notably, 1920 (68.8%) of the neuron-enriched PASs were located at distal positions, whereas 1421 (52.2%) hESC-enriched PASs were located at proximal positions (Supplementary Fig. 4a). Among the 1335 APA-regulated genes identified at gene-level analysis (both shortening and lengthening), 1287 (96.4%) contained at least one differentially used PASs (Fig. 3h). These results indicate that gene-level APA events detected by APALORD are supported by significant PAS-level changes.

APALORD provides visualization tools of APA regulation events for individual genes. This is especially helpful given the complexity of APA regulation, which can often include 3 or more different PASs per gene. The NAV1 gene (Neuron Navigator 1) plays a role in axon outgrowth and migration⁶⁶. In our DRS data, APALORD identified that NAV1 undergoes CPA at 4 tandem 3′UTR PASs. NAV1 was classified as exhibiting significant 3′UTR lengthening during neural differentiation, with PAS1 and PAS4 identified as regulated PAS events (Fig. 3i). Specifically, PAS1 usage decreased from 31.2 to 11.5%, while PAS4 usage increased from 8.1 to 37.9%. SYNGR1, which encodes a protein associated with presynaptic vesicles in neuronal cells⁶⁷, utilizes 4 PASs that span multiple exons, and was thus classified as a mixed APA gene. The primary PAU switch occurred between intronic PAS1 and distal PAS4, with PAS1 usage dropping from 70.9 to 18.3% and PAS4 usage rising from 10.7 to 73.8% (Fig. 3j).

Genomic sequence features of regulated PASs during neural differentiation

As both gene-level and PAS-level analyses revealed a global shift toward increased distal PAS usage, we next examined whether these PAS groups exhibit systematic differences in genomic sequence features (Supplementary Data 3). Because the canonical AAUAAA polyA signal, its variants, and associated USEs and DSEs are key determinants of PAS strength, we analyzed the distribution of these fundamental motifs within ±100 nt of the PAS peaks for distal, proximal or single PASs (i.e., the only PAS in single-PAS genes). Single PASs exhibited the highest frequency of the canonical AAUAAA hexamer, whereas distal PASs showed intermediate AAUAAA frequency—lower than single PASs but higher than proximal PASs (Fig. 4a, b). In contrast, the weaker variant AUUAAA exhibited comparable frequencies across all three PAS groups (Supplementary Fig. 4b). The UGUA motif upstream of the polyA signal was more enriched in distal PASs, whereas the UGUA motif located downstream of the polyA signal was most enriched in proximal PASs. In contrast, UGUA motifs were least abundant in single PASs both upstream and downstream of the polyA signal (Fig. 4a). Analysis of GU/U-rich downstream elements using the degenerate motif UKUKUK (K = G or U)⁶⁸ revealed no differential enrichment across PAS groups (Fig. 4a). Additionally, UUU motif upstream of polyA signal was most enriched in distal PASs, whereas AAA motif downstream of the PAS peak was most enriched in proximal PASs; both AAA and UUU elements were least abundant in single PASs (Supplementary Fig. 4b). Given that AAUAAA and upstream UGUA, both known to enhance PAS strength^59,69, were more enriched in distal PASs relative to proximal ones, we wondered whether distal PASs are generally stronger than proximal PASs. Using the PolyaModelsHuman general model⁵⁹ to score PAS strength based on ±100 nt sequence around the PAS peaks, we found the distal PASs exhibited significantly higher strength scores than proximal PASs, although lower than those of single PASs (Supplementary Fig. 5a, two-sided Mann–Whitney U test, p < 2.2e-16).

Fig. 4. APA regulation during neural differentiation is associated with enhanced PAS motif preference.

a Distribution of AAUAAA, UGUA and UKUKUK (K = G or U) motifs in ±100 nt sequences surrounding PAS peaks identified from hESC and neuron DRS, stratified by single-PAS, proximal and distal PASs. b Fraction of PASs containing AAUAAA or AWUAAA polyA signal in the upstream 50 nt region in proximal PASs, distal PASs, and single PAS groups. c Distribution of AAUAAA, AUUAAA, UGUA, and UKUKUK motifs within ±100 nt sequences surrounding PAS peaks in proximal versus distal PASs across APA lengthening, non-regulated and shortening genes during neural differentiation. Fraction of PASs containing AAUAAA (d) or AUUAAA (e) polyA signals within the upstream 50 nt region in proximal versus distal PASs across APA lengthening, non-regulated and shortening genes. f Usage of PASs stratified by polyA signal (AAUAAA, AUUAAA or non-AWUAAA) within the upstream 50 nt region in D0 and D7 samples. g Usage of AAUAAA PASs (from f) stratified by the position of UGUA motifs (upstream, downstream or absent) relative to AAUAAA in D0 and D7 samples. h Usage of AAUAAA PASs (from f) stratified by the position of GU/U-rich motifs (UKUKUK; upstream, downstream or absent) relative to AAUAAA in D0 and D7 samples. In f–h, box plots indicate 25th and 75th percentiles (box bounds), with whiskers extending to the most extreme data point within 1.5× IQR; the central diamond denotes the median. Statistical significance was assessed using the two-sided Mann–Whitney U test, except for paired comparisons between D0 and D7 samples, where a paired, two-sided Wilcoxon signed-rank test was applied.

Given that distal PASs are generally stronger than proximal PASs, we next asked whether the APA lengthening trend observed during neural differentiation reflects a preferential selection of stronger PASs in neurons. In 1231 lengthening genes with multiple PASs, neuron-favored distal PASs were significantly stronger than proximal PASs (Supplementary Fig. 5b). In 6923 APA non-regulated genes, distal PASs were also stronger than proximal PASs, although both were weaker than those observed in lengthening genes (Supplementary Fig. 5b). In contrast, in 79 shortening genes, neuron-favored proximal PASs exhibited slightly lower strength than distal ones (paired two-sided Wilcoxon signed-rank test, p = 0.027) but were stronger than proximal PASs in both lengthening and APA non-regulated genes (Supplementary Fig. 5b; two-sided Mann–Whitney U test, p = 0.043 and p = 0.0078). These observations suggest that APA dynamics during neural differentiation are associated with the preferential usage of relatively stronger PASs. We therefore asked whether this preference is accompanied by enrichment of sequence motifs known to enhance PAS strength. We stratified multi-PAS genes by their APA trend (lengthening, shortening, and non-regulated) during neural differentiation and compared the distribution of key sequence motifs, including the canonical AAUAAA polyA signal, its variant AUUAAA, upstream UGUA motifs, and downstream GU/U-rich elements (UKUKUK), across PAS groups (Fig. 4c). In lengthening genes, neuron-favored distal PASs were enriched for AAUAAA, upstream UGUA, and downstream UKUKUK relative to proximal PASs (Fig. 4c, d). In APA non-regulated genes, distal PASs showed enrichment of AAUAAA polyA signal, but only minimal enrichment of upstream UGUA and no enrichment of downstream UKUKUK motifs. In shortening genes, no difference in AAUAAA usage was observed, whereas the weaker AUUAAA polyA signal was enriched in neuron-disfavored distal PASs (Fig. 4c–e); conversely, neuron-favored proximal PASs showed no enrichment of upstream UGUA motifs but enrichment of downstream UKUKUK motifs (Fig. 4c). Together, these results suggest that neurons preferentially utilize PASs enriched in essential motifs associated with increased PAS strength.

To test this hypothesis directly, we first examined whether PASs containing the canonical AAUAAA signal showed increased usage in neurons. Based on the presence of polyA signal within −50–0 nt upstream of PAS peaks, PASs identified in multi-PAS genes were classified into three groups: AAUAAA PASs (n = 15,599), AUUAAA PASs (n = 5466), and non-AWUAAA PASs (n = 8955). PASs with both AAUAAA and AUUAAA hexamers in this region were excluded from analysis for simplicity. In both hESCs and neurons, AAUAAA PASs showed significantly higher usage than AUUAAA PASs, which in turn were higher than non-AWUAAA PASs (two-sided Mann–Whitney U test, p < 2.2e-16), consistent with their relative strength (Fig. 4f). Importantly, AAUAAA PAS usage was significantly increased in neurons (median PAU from 15.38% in hESCs to 15.63% in neurons, paired two-sided Wilcoxon signed-rank test, p = 1.66e-10), whereas AUUAAA and non-AWUAAA PASs usage decreased (Fig. 4f, paired two-sided Wilcoxon signed-rank test, p = 7.34e-09 and p < 2.2e-16). We next examined the contribution of auxiliary motifs within AAUAAA PASs. PASs were stratified based on the position of the UGUA motifs relative to the AAUAAA hexamer: upstream UGUA (n = 4100), PASs downstream UGUA (n = 2653), and no UGUA motif (n = 9053) within 50nt of the AAUAAA polyA signal (PASs containing both upstream and downstream UGUA motifs were excluded). PASs containing upstream UGUA showed higher usage than those lacking UGUA, whereas downstream UGUA was associated with reduced usage in both hESCs and neurons. Notably, only PASs with upstream UGUA exhibited increased usage in neurons (Fig. 4g; paired two-sided Wilcoxon signed-rank test, p < 2.2e-16). Similarly, analysis of GU/U-rich elements within AAUAAA PASs revealed that PASs containing downstream GU/U-rich motifs exhibited higher usage than those lacking such motifs in both conditions and specifically showed increased usage in neurons (Fig. 4h, paired two-sided Wilcoxon signed-rank test, p = 7.29e-08). These results align with our observations in APA non-regulated genes, in which PASs enriched for AAUAAA lack enrichment of upstream UGUA and downstream GU/U-rich motifs.

Together, these findings support a model in which neural differentiation is accompanied by increased usage of PASs harboring both strong core and auxiliary sequence features, including AAUAAA, upstream UGUA, and downstream GU/U-rich motifs. When such PASs are located distally, genes exhibit APA lengthening; when located proximally, genes exhibit APA shortening; and genes lacking PASs with these features tend to show no significant APA regulation. While the distribution of UGUA motif around the polyA signal did not vary markedly across PASs with different polyA signals (Supplementary Fig. 5c), GU/U-rich elements were preferentially enriched downstream of the PAS peaks in weaker PASs using non-AWUAAA polyA signals (Supplementary Fig. 5d). However, AUUAAA PASs with upstream UGUA did not exhibit increased usage during neural differentiation (Supplementary Fig. 5e, paired two-sided Wilcoxon signed-rank test, p = 0.54), suggesting that the strength of the core polyA signal plays a dominant role. Similarly, although downstream GU/U-rich motifs enhanced PAS usage in AUUAAA PASs, these PASs were downregulated during differentiation (Supplementary Fig. 5f, paired two-sided Wilcoxon signed-rank test, p = 1.03e-04), further supporting the primary importance of the polyA signal in determining PAS selection during neural differentiation.

Leaky transcripts originating from weak cryptic cleavage sites are reduced during hESC neural differentiation

After assigning reads to identified PASs within each host gene, we observed that a subset of polyadenylated reads that were aligned to annotated genes, but could not be assigned to our identified PASs. We refer to these as “leaky” transcripts. For example, three PASs were identified in the WDR5 gene; however, 13 polyadenylated transcripts ($~$1% of total WDR5 transcripts) were located more than 50 nt away from all three PASs and were therefore classified as leaky (Fig. 5a, b). As these leaky transcripts cannot be assigned to any PAS with ≥1% usage, we hypothesized that they originate from weak cryptic CSs whose usage falls below APALORD′s default detection threshold. Therefore, we analyzed genomic sequences flanking their CSs. Strong annotated PASs exhibited characteristic polyA signal, A-rich sequences preceding the PAS peaks, and downstream GU/U-rich elements. In contrast, weaker novel PASs showed reduced enrichment of these features, whereas leaky transcripts lacked recognizable upstream or downstream core polyadenylation motifs (Fig. 5c). We next quantified leaky transcripts at the gene level in both hESCs and neurons and observed a transcriptome-wide reduction in leaky transcript usage during neural differentiation, decreasing from 2.39 to 2.32% (Fig. 5d), a difference that was statistically significant (paired two-sided Wilcoxon signed-rank test, p = 1.307e-06). Next, we compared the usage of leaky transcripts between single-PAS genes (typically harboring stronger PASs) and multi-PAS genes. In both hESCs and neurons, leaky transcripts usage was significantly lower in single-PAS genes compared with multi-PAS genes (Fig. 5e; two-sided Mann–Whitney U test, p < 2.2e-16). Furthermore, leaky transcripts usage decreased significantly during neural differentiation in multi-PAS genes (paired two-sided Wilcoxon signed-rank test, p = 4.579e-09) but not in single PAS genes (paired two-sided Wilcoxon signed-rank test, p = 0.1481). These results suggest that leaky transcripts originate from weak cryptic “PASs” and are suppressed in context where stronger PASs are preferentially utilized, such as in single PAS genes and during neural differentiation.

Fig. 5. Leaky transcripts are suppressed during neural differentiation.

a PolyA tail length and 3′ end positions of all transcripts mapped to the WDR5 gene. b Percentage of transcripts assigned to PAS1, PAS2, PAS3, and leaky transcripts that cannot be attributed to one of the three identified PASs in WDR5. c Sequence logos showing nucleotide composition surrounding PAS peaks (and cleavage site in leaky transcripts). Letter heights are scaled by information content (bits), comparing annotated PASs, novel PASs, and leaky transcripts identified from hESC and neuron DRS data. d Percentage of leaky transcripts per gene in hESC and neuron samples (n = 11120). e Percentage of leaky transcripts per gene in single-PAS (n = 2870) and multi-PAS (n = 8250) genes in both hESC and neuron samples (all genes with >50 reads across samples). In d, e, box plots indicate 25th and 75th percentiles (box bounds), with whiskers extending to the most extreme data point within 1.5× IQR; the central diamond denotes the median. Statistical significance was assessed using the two-sided Mann–Whitney U test, except for paired comparisons between D0 and D7 samples, where a paired, two-sided Wilcoxon signed-rank test was applied.

Differential gene expression is associated with APA regulation during neural differentiation

A substantial proportion of genes undergo differential gene expression (DGE) during neural differentiation⁷⁰. In our DRS dataset, we identified 2401 downregulated and 2967 upregulated genes (Supplementary Fig. 6a; |log2FoldChange| > 1 and adjusted p < 0.05). We classified genes based on both APA trend and DGE status during neural differentiation and observed a significant but weak association between APA trend and DGE (Fig. 6a; Chi-Square test, χ² (4) = 175.31, p < 2.2e-16, Cramér′s V = 0.088). Standardized residual analysis revealed that the association was primarily driven by enrichment of APA lengthening events in upregulated genes (standardized residual = 11.27) and depletion in downregulated genes (−9.15). Conversely, downregulated genes are enriched for APA no-change events (8.91), while upregulated genes were depleted (−11.30). APA shortening events contributed minimally to the observed association, representing <1% of analyzed genes. To further assess the relationship between APA and expression, we examined the correlation between APA change and expression fold change during neural differentiation. A weak but significant positive correlation was observed across all genes (Fig. 6b; r = 0.22, Pearson′s correlation, p < 2.2e-16). Genes without significant APA changes showed an even weaker correlation (r = 0.17, Pearson′s correlation, p < 2.2e-16), whereas shortening genes exhibited a moderate negative correlation (r = −0.55, Pearson′s correlation, p = 1.9e-07) and lengthening genes showed a moderate positive correlation (r = 0.41, Pearson′s correlation, p < 2.2e-16). These results suggest that APA regulated genes tend to exhibit larger expression changes, with both shortening and lengthening genes biased toward upregulated expression during neural differentiation.

Fig. 6. Differential gene expression is associated with APA regulation during neural differentiation.

a Relationship between APA regulation and gene expression level changes during neural differentiation. b Correlation between gene-level APA change and gene expression fold change during neural differentiation. The solid line indicates the linear regression fit, with the shaded area representing the 95% confidence interval. Pearson correlation was assessed using a two-sided test. Dominant PAS strength (c) and gene expression level (d) stratified by the number of PASs per gene. e Dominant PAS strength stratified by APA regulation category (lengthening, shortening, and non-regulated) and the number of PASs per gene. In (c, e), PAS strength is labeled as PolyaStrength Score. In (c–e), box plots indicate 25th and 75th percentiles (box bounds), with whiskers extending to the most extreme data point within 1.5× IQR; the central diamond denotes the median. Statistical significance was assessed using the two-sided Mann–Whitney U test, except for paired comparisons between D0 and D7 samples, where a paired, two-sided Wilcoxon signed-rank test was applied.

Given our observation that both shortening and lengthening genes preferentially utilize stronger PASs harboring motifs that were more favored in neurons compared with hESCs, we wondered whether PAS strength is associated with gene expression levels. Because PASs in single PAS genes are generally stronger than both proximal and distal PASs in multi-PAS genes, we stratified genes by the number of PASs they used. For each multi-PAS gene, the PAS with the highest usage in neurons was defined as the dominant PAS and used to represent the gene-level PAS strength. We observed that genes using more PASs tended to have weaker dominant PASs (Fig. 6c), and lower average gene expression across both hESCs and neurons (Fig. 6d). This trend was consistent when considering all PASs used within each gene (Supplementary Fig. 6b). Furthermore, across genes with varying PAS numbers, upregulated genes consistently exhibited stronger PASs, whereas downregulated genes were associated with weaker PASs, both at the dominant PAS level (Fig. 6e) and across all PASs (Supplementary Fig. 6c). Together, these observations suggest that gene expression levels are linked to the strength of PASs used within each gene, with stronger PASs associated with higher expression, and that this relationship is further reinforced during neural differentiation.

PAS identification and APA analysis in fly embryos using APALORD

While the human genome is supported by well-annotated PASs, PAS annotation in many other organisms remains inaccurate. Because APALORD is designed as an annotation-independent APA analysis tool, we evaluated its performance on LR RNA-seq data from a less well-annotated species. To this end, we analyzed DRS data from Drosophila melanogaster embryos, including early-stage (2–4 h) and late-stage (16–18 h) wild-type embryos. Across these conditions, APALORD identified 20,326 PASs, of which 18,836 PASs were derived from genes with ≥5 transcripts across samples. We compared these PASs with annotated PASs from the Drosophila Ensembl gene annotation (BDGP6). Only 91 PASs (0.5%) matched annotated PASs at the exact same nucleotide position, whereas 8994 PASs (53.4%) were located within 50 nt of annotated PAS and were classified as proximity PASs. The remaining 7751 PASs (46%) were classified as novel PASs (Fig. 7a), indicating a substantial proportion of previously unannotated PASs in the fly genome. Among PASs identified in protein-coding genes, 7443 were novel, and 181 of 282 PASs detected in non-coding RNA genes were also novel (Supplementary Fig. 7a). Notably, among match and proximity PASs, 64.2% were located within 5 nt and 77.8% within 10 nt of an annotated PAS (Fig. 7b). We noticed a systematic offset, with APALORD-identified PASs more frequently located within 10 nt of the annotated PASs rather than at the exact annotated positions. To investigate this discrepancy, we examined the ±10 nt sequences surrounding identified PAS peaks and compared nucleotide composition with the nearby annotated PASs. At PAS peak positions, annotated PASs showed near zero frequency of A, whereas APALORD-identified PASs exhibited a high frequency (>0.9) of A (Fig. 7c), consistent with cleavage pattern observed in human data (Fig. 5c) and previous studies^25,62. In addition, the enrichment of A was shifted downstream of annotated PASs relative to APALORD-identified PASs peaks. This offset is likely introduced by oligodT-primed cDNA-based approaches used in PAS annotation, in which genomic As may be assigned as part of the polyA tail and thus removed, even when there were As that could serve as the template in the corresponding genomic sequences⁵⁹. Together, these results demonstrate that APALORD enables high-resolution PAS identification in species with incomplete annotations and may more accurately capture true CS positions compared with existing annotations.

Fig. 7. PAS identification and APA analysis in Drosophila embryos using APALORD.

a Classification of PASs identified by APALORD from DRS data in Drosophila melanogaster embryos. match: exact alignment with annotated PASs in the fly ENSEMBL reference; proximity: within 50 nt of an annotated PAS; novel: outside of match or proximity criteria. b Distance distribution of match and proximity PASs relative to annotated PASs. Colors correspond to PAS classes in (a). c Nucleotide composition of sequences within 10 nt surrounding identified PAS peaks compared with annotated PASs. d Transcriptome-wide APA changes in late-stage embryos compared with early-stage embryos. Gene-level APA analysis was performed using the same two-sided test in APALORD, followed by FDR correction for multiple testing. e Distribution of the AAUAAA hexamer within ±100 nt sequences surrounding identified PAS peaks in fly embryonic DRS data, including PASs from single-PAS genes and proximal and distal PASs from multi-PAS genes (left). Fraction of PASs containing AAUAAA or AWUAAA polyA signals within the upstream 50nt region across these PAS groups (right). f Usage of PASs stratified by polyA signal type (AAUAAA, AUUAAA or non-AWUAAA) within the upstream 50nt region in early- and late-stage embryos. g Usage of AAUAAA PASs (from f) stratified by UGUA motif position (upstream, downstream or absent) relative to AAUAAA in early- and late-stage embryos. In (f, g), box plots indicate 25th and 75th percentiles (box bounds), with whiskers extending to the most extreme data point within 1.5× IQR; the central diamond denotes the median. Statistical significance was assessed using the two-sided Mann–Whitney U test, except for paired comparisons between 2–4 h and 16–18 h embryos, where a paired, two-sided Wilcoxon signed-rank test was applied.

We next performed gene-level APA analysis using APALORD by comparing late-stage embryos with early-stage embryos (Supplementary Data 4) and identified 455 lengthening and 88 shortening genes during embryonic development (Fig. 7d). This result is consistent with expectations, as late-stage embryos possess a more developed nervous system and have been shown to preferentially express longer 3′UTR isoforms compared with early-stage embryos⁶⁵. A majority (74%) of APA-regulated genes were classified as last-exon tandem APA genes, which accounted for 52.6% of all genes included in the analysis (Supplementary Fig. 7b). We next compared polyA signal usage (Supplementary Data 5) across three PAS groups: 2385 single PASs from single-PAS genes, and 4680 proximal and distal PASs from multi-PAS genes. Consistent with observations in human cells, single PASs exhibited the highest frequency of the canonical AAUAAA polyA signal, followed by distal PASs, which showed higher AAUAAA usage than proximal PASs (Fig. 7e). Because distal PASs enriched for AAUAAA were preferentially used in late-stage embryos, as revealed by the predominance of APA lengthening events, we next examined the PAU changes during embryonic development across PAS groups stratified by polyA signal type (AAUAAA, AUUAAA, and non-AWUAAA). AAUAAA PASs exhibited the highest usage, followed by AUUAAA PASs, whereas non-AWUAA PAS showed the lowest usage in both early- and late-stage embryos (Fig. 7f). During embryonic development, AAUAAA PAS usage increased significantly (paired two-sided Wilcoxon signed-rank test, p < 2.2e-16), whereas non-AWUAAA PASs usage decreased significantly (paired two-sided Wilcoxon signed-rank test, p < 2.2e-16) and no significant change was observed in AUUAAA PASs (Fig. 7f; paired two-sided Wilcoxon signed-rank test, p = 0.26). We further examined the contribution of UGUA motifs within AAUAAA PASs. PASs containing upstream UGUA exhibited higher usage than those containing downstream UGUA or lacking UGUA in both early- and late-stage embryos (Fig. 7g). Notably, during embryonic development, usage of AAUAAA PASs with upstream UGUA or lacking UGUA increased significantly (paired two-sided Wilcoxon signed-rank test, p < 2.2e-16 and p = 9.55e-15), whereas PASs containing downstream UGUA showed a significant but modest increase in usage (paired two-sided Wilcoxon signed-rank test, p = 0.004). Together, these findings revealed a conserved APA lengthening trend driven by enhanced preference of stronger PASs, characterized by the canonical AAUAAA polyA signal and upstream UGUA motifs, during fly embryonic development.

The usage of leaky transcripts as a percentage of reads per gene was significantly higher in early-stage embryos compared with late-stage embryos (Supplementary Fig. 7c, paired two-sided Wilcoxon signed-rank test, p < 2.2e-16). In addition, single-PAS genes, which more frequently used the strongest AAUAAA polyA signal, exhibited significantly fewer leaky transcripts compared with multi-PAS genes across embryonic stages (Supplementary Fig. 7d, two-sided Mann–Whitney U test, p = 2.6e-14 in early-stage embryos and p < 2.2e-16 in late-stage embryos), consistent with our observations in human cells that genes utilizing stronger PASs tend to express less leaky transcripts. Furthermore, leaky transcript usage decreased significantly during embryonic development in both single-PAS and multi-PAS genes (Supplementary Fig. 7d, two-sided Mann–Whitney U test, p < 2.2e-16).

Applying APALORD to PacBio RNA-seq data

We next applied APALORD to cDNA libraries sequenced on the PacBio LR sequencing platform^61,71. Previous studies have reported a global trend of 3′UTR lengthening in brain tissues compared to liver in both mouse and human^13,72. We thus analyzed three dorsolateral prefrontal cortex libraries and two right liver lobe libraries, comprising 3,993,418 reads (1,147,401 from liver and 2,846,017 from cortex). APALORD identified 10,030 PASs across 2,975 genes with at least 5 reads in each sample. 1246 (12.4%) of these PASs matched annotated PASs in the combined PolyA_DB and GENCODE reference, while 6066 (60.5%) were located within proximity of annotated PASs (Supplementary Fig. 8a). In contrast to DRS-based analysis, in which the majority of identified PASs aligned precisely to annotated positions, only 17.0% of PASs identified in the PacBio data matched annotated PASs exactly, whereas 55.3% were positioned within 1–5 nt from annotated PASs (Supplementary Fig. 8b). To investigate this positional offset, we examined nucleotide composition at PAS peak positions. In contrast to DRS data, PacBio-derived PAS peaks showed a low frequency of A at the main CSs (position 0) and a shift of A toward downstream positions (+1 nt position (Supplementary Fig. 8c), similar to our observation in the fly PAS annotation. This observation is expected as the data were oligodT-primed cDNA sequences and genomic As have been removed from the 3′ ends of reads in the PacBio standard data processing pipeline when they were considered as part of the polyA tail⁷¹. Together, these results demonstrate that APALORD can robustly identify PASs from PacBio cDNA data while also revealing platform-specific biases in CS resolution.

We next performed gene-level APA analysis (Supplementary Data 6) and identified 688 lengthening and 199 shortening genes in cortex samples when compared with liver samples (Supplementary Fig. 8d), a trend consistent with previous investigations¹³. Given that oligodT-primed cDNA can introduce internal priming artifacts from genomic A-rich sequences, and that 27.1% of PASs were classified as novel from these low-depth datasets, we assessed the potential contribution of internal priming from genomic A-stretches. In APALORD, PAS peaks flanked by A-stretches (>6As within 10 nt) are labeled as suspicious internal priming artifacts. If polyA trimming was performed, downstream sequences of PAS peaks are examined for A-stretches; otherwise, upstream sequences should be evaluated. Using this criterion, we identified 2796 genes containing PASs potentially affected by internal priming, including 241 (35%) of all lengthening genes and 100 (50.3%) of all shortening genes (Supplementary Fig. 8e). To further assess the robustness of APA analysis, we excluded 265 APA-regulated genes lacking sufficient read support across samples (≥5 reads per sample). Among the remaining 622 APA-regulated genes, 490 (78.8%) contained significantly regulated PASs identified at the PAS-level analysis (fold change >1, adjusted p < 0.05; Supplementary Fig. 8f, Supplementary Data 7). Together, these results suggest that although internal priming may contribute to a subset of PASs identified from cDNA-based LR RNA-seq data, most APA-regulated genes detected by APALORD are supported by PAS-level changes, supporting the robustness of the analysis.

Discussion

To leverage the high-resolution 3′ end information provided in LR RNA-seq data, we developed APALORD, a comprehensive tool for differential APA analysis between conditions. Applying APALORD to DRS data from hESCs and derived neurons, we identified PASs with positional accuracy comparable to existing annotations in the human genome. Notably, APALORD-identified novel PASs are enriched in lncRNA genes, introns and 3′ intergenic regions—regions that are underrepresented in current annotations. In the less well-annotated Drosophila melanogaster genome, APALORD uncovered thousands of novel PASs from embryonic DRS data. APALORD provides a rapid, annotation-independent approach to generate high-confidence PAS maps from DRS or 3′ tagged cDNA LR data in any species. APA analysis methods for single-cell or bulk short-read RNA-seq data often rely on well-annotated PASs⁷³. APALORD-generated PAS annotations will enable more accurate APA analyses across diverse species.

A major difficulty in PAS annotation and APA analysis is how to deal with false 3′ end detection. Internal priming is introduced by widely used oligodT-initiated reverse transcription (RT) in RNA-seq⁷⁴. For various reasons, internal priming generated 3′ ends are easier to be detected in LR RNA-seq reads and thus could be removed. The presence of sequenced polyA tails at the 3′ end of DRS reads greatly mitigates internal priming artifacts. In many current pipelines, adenosine residues at the 3′ ends of reads are removed frequently during data processing, as they are considered as part of the polyA tail instead of the 3′UTR. In PacBio cDNA data, we noticed a systematic shift of PAS peak positions consistent with the removal of terminal As. A similar positional offset was also observed when comparing APALORD-identified PASs with annotated PASs in fly genome. This suggests that such biases are widespread in annotations derived from oligodT-primed cDNA data. Notably, in human cell DRS data, we observed strong enrichment of A at PAS peak positions, particularly for well-supported stronger PASs when compared with novel PASs (Fig. 5c). This pattern suggests that genomic As at CSs might contribute to PAS strength and may have been underrepresented in previous studies due to technical biases. We find a consistent enrichment of A at PAS peaks from DRS data in both fly and human, indicating that this feature may represent a conserved sequence characteristic of functional PASs. These observations also imply that PAS annotations derived from oligodT-primed data, especially in species lacking 3′ end tag-specific sequencing data, might systematically misrepresent true CS positions. Therefore, APALORD provides a reliable approach to generate more accurate PAS reference annotations across species.

Gene-level APA analysis of our human DRS dataset recapitulated a transcriptome-wide APA lengthening trend during neural differentiation, characterized by increased usage of distal PASs. The global APA lengthening trend is associated with increased usage of distal PASs that are enriched for the canonical AAUAAA polyA signal and upstream UGUA motifs. As expected, our analysis confirms the AAUAAA hexamer as the primary determinant of PAS strength⁷⁵, with upstream UGUA and downstream GU/U-rich elements⁶⁹ further enhancing CPA efficiency^59,76. Notably, we observed that the preferential selection of PASs harboring these core and auxiliary sequence features is further enhanced in neurons. Although the mechanisms underlying the enhanced selection of stronger PASs remain to be fully elucidated, our results suggest that APA is coordinately regulated during neural differentiation to favor the usage of stronger PASs, frequently located at distal positions. Interestingly, APA lengthening is more commonly observed in differentiated tissues with low proliferative capacity, such as neurons^11,13, whereas proliferative tissues and rapidly dividing cells (e.g., testis, liver, activated immune cells and cancer cells) tend to exhibit increased usage of proximal PASs^12,13,52,53. This contrast suggests a potential speed versus quality trade-off in PAS selection within multi-PAS genes, whereby rapidly dividing cells may prioritize transcriptional efficiency using proximal, potentially weaker PASs, and differentiated cells favor more accurate and efficient processing at stronger distal PASs.

Using APALORD, we identified a class of low abundance transcripts that cannot be assigned to any nearby identified PASs, which we refer to as leaky transcripts. These transcripts originate from weak cryptic CSs and are observed transcriptome-wide across both human and fly datasets. Their prevalence suggests that such transcripts may be underappreciated in previous APA analyses. Notably, leaky transcript usage is reduced in neurons, where stronger PASs are preferentially utilized, suggesting that selection of stronger PASs in the host genes may suppress cleavage at weak cryptic sites. Conversely, these cryptic sites may become more likely to be used when stronger PASs are less favored or functionally compromised, allowing weaker sites to be utilized under certain conditions. The widespread presence of such cryptic cleavage events suggests that the effective number of PASs utilized by a gene may be more dynamic than previously appreciated and is influenced by cellular environment.

Taken together, APALORD offers a robust framework for identifying PASs from LR RNA-seq data and detecting APA changes across conditions. It enables precise 3′ end annotation independent of existing PAS references and quantifies both PAS-specific changes and cumulative shifts in 3′ end usage along gene bodies. Unlike many existing APA profiling methods, APALORD directly leverages the read-level information within each dataset, enabling reliable APA analysis even in the absence of biological replicates or when sequencing depth is limited. DRS, which captures full-length native RNA molecules without RT or amplification, further enables simultaneous interrogation of nucleotide modifications, polyA tail length and transcript isoform structure^77–79. The growing adoption of LR RNA-seq for integrated analyses of differential gene and isoform expression, alternative splicing, APA, RNA modifications, and rare isoform discovery underscores its power to resolve the interconnected layers of mRNA processing. Within this context, APALORD provides a versatile and scalable framework for dissecting APA regulation, enabling precise characterization of 3′ end dynamics and facilitating a deeper understanding of the role of APA in gene regulation across diverse biological systems.

Methods

Drosophila stock and culture

The Drosophila melanogaster strain w¹¹¹⁸ was used as the wild type for embryos collection. Flies were maintained at 25 °C under 12 h light/12 h dark cycle and fed with standard fly food containing 0.56% agar, 6.71% cornmeal, 1.59% inactivated yeast, 0.92% soy flour, 7.0% corn syrup, 0.44% propionic acid, and 0.15% Tegosept. Embryos were collected on grape juice agar plates and aged as indicated (2–4 h: early-stage or 16–18 h: late-stage). Collected embryos were dechorionated in 50% bleach treatment for 5 min, followed by extensive washing. Total RNA was extracted using Trizol reagent (Invitrogen #15596018) per the manufacturer′s instructions.

Ethics Statement

Experiments using hESCs and derived neurons were approved by the UConn Stem Cell Research Oversight Committee and were performed in accordance with all relevant guidelines and regulations.

Human stem cell culture

H9-NGN2 hESCs were generated by the UCONN Cell and Genome Engineering Core⁸⁰. Stem cells were grown for at least one passage on a 6-well tissue culture plate coated with Matrigel (Corning #CLS354277) in mTeSR Plus media (mTeSR+, StemCell Technologies #100-0276) over 3–4 days to ~60–80% confluency. Cells were thoroughly singularized with 0.5 mL/well Accutase (StemCell Technologies #07920) for 9-10 min at 37 °C and resuspended by pipetting 15–20× with a P1000 in 1 mL mTeSR+/well. The cells were pooled and centrifuged in a 5 mL tube at room temperature in a swinging-bucket rotor at 220 × g for 3 min. The supernatant was removed from the cell pellet, which was resuspended in 1 mL Induction Medium (IM) [DMEM/F12 (ThermoFisher Scientific #11330032), 1x NEAA (ThermoFisher Scientific), 1x GlutaMAX (ThermoFisher Scientific), 100 μg/mL apo-transferrin (MilliporeSigma #T1147), 25 μg/mL insulin (MilliporeSigma #I6634), 100 μM putrescine (MilliporeSigma #5780), 20 nM progesterone (MilliporeSigma #P8783), and 30 nM sodium selenite (MilliporeSigma #S5261)] with 10 μM Y-27632 ROCK inhibitor (RI, Selleckchem #S1049). An aliquot was diluted and counted in a Countess Cell Counter (ThermoFisher Scientific). For each well to be plated for induction, 200,000 cells, 2 mL of IM with RI and 2 μg/mL doxycycline (dox, MilliporeSigma #D5207) were plated on Matrigel plates, distributed evenly, and incubated overnight at 37 °C. The remaining singularized cells were placed on ice, PBS was added to the tube to rinse, and these were centrifuged at 300 × g for 3 min. The resulting pellet was processed fresh from ice or snap-frozen in liquid nitrogen and stored at −80 °C for later processing as D0 samples. Media was replaced on days 1 and 2 with fresh IM and dox. On day 3, cell wells were gently rinsed with PBS and singularized with Accutase as before, pooled and resuspended in 1 mL IM with RI. Cells were counted as before, and replicate wells were plated on 6-well poly-D-lysine- (MilliporeSigma #A-003-E) and laminin- (ThermoFisher Scientific #23017015) coated plates without dox at 500–1000 K/well in 2 mL IM with RI. On day 4, the media was fully and gently replaced with Cortical neuron Medium (CM) [½ DMEM/F12 and ½ Neurobasal media (ThermoFisher # 21103049) with 1x B27 supplement (ThermoFisher Scientific # 12587010), 1x Pen-Strep, 10 ng/mL BDNF (Peprotech # 450-02), 10 ng/mL GDNF (Peprotech #450-10), 10 ng/mL NT-3 (Peprotech #450-03), and 1 μg/mL laminin]. On day 7 induced neurons were collected by removing media, adding 1 mL ice-cold PBS per well, removing cells with a silicone cell scraper, and centrifuging each replicate at 300 × g for 3 min in 1.5 mL tubes (in swinging bucket rotor). The resulting pellets were processed fresh from ice or snap-frozen in liquid nitrogen and stored at −80 °C for later processing as D7 samples.

Nanopore LR RNA-seq library preparation and sequencing

Cell samples were subject to RNA extraction using the PureLink RNA mini Kit (Invitrogen #12183025) and on-column Dnase treatment (PureLink Dnase, Invitrogen #12185010) according to the manufacturer′s instructions. DNA-free total RNA was used for library preparation with the Oxford Nanopore Direct RNA sequencing kit (Oxford Nanopore SQK-RNA004) following the manufacturer′s protocol. In brief, 2–3 μg of total RNA per sample was incubated with the RT adapter for 10–20 min at room temperature, followed by RT using Indura Reverse Transcriptase (New England Biolabs M0681L). The resulting RNA-cDNA complexes were purified using RNAClean XP beads (Beckman Coulter A63987). Purified samples were then ligated to the RNA sequencing adapter by incubation for 20–30 min at room temperature, followed by a second RNAClean XP bead cleanup and elution in 15 µL elution buffer. Libraries were quantified using the Qubit dsDNA HS Assay Kit (Invitrogen Q32854). 40–70 ng of library was loaded onto a PromethION or MinION R10.4.1 flow cell and sequenced for up to 72 h.

LR RNA-seq data analysis

Basecalling of direct RNA-seq reads was performed using Dorado (Oxford Nanopore Technologies) with parameters sup@latest --min-qscore 0 --estimate-poly-. Reads with recognized adapter sequence and polyA tail were retained for downstream analysis. Basecalled reads were aligned against the human genome (hg38) using minimap2 (-ax splice -t 7 -B 3 -O 3,20 --junc-bed, using GENCODE v43 annotation as splice junction guidance). For Drosophila melanogaster samples, reads were aligned to BDGP Release 6 reference genome (Ensembl release 110). For PacBio cDNA datasets, pre-aligned BAM files were obtained and used directly for downstream analysis. Aligned reads were assigned to genes using IsoQuant, retaining only primary alignments (--data_type nanopore/pacbio --gene_quantification unique_only --stranded forward). The resulting output directories were used as input for APALORD to load the data for PAS identification and APA analysis.

APALORD gene-level APA analysis

For gene-level APA analysis, we compared distributions using a tie-adjusted two samples Kolmogorov-Smirnov-type test. The two samples are the 3′ end coordinates at single gene level from two conditions:

$$X=\left\{x_1,\dots ,x_{n_1}\right\},Y=\left\{y_1,\dots ,y_{n_2}\right\}$$ 1

with total sample size N = n₁ + n₂

Empirical cumulative distribution functions (ECDFs) were computed for each group:

$${\widehat{F}}_X\left(t\right)=\frac{1}{n_1}{\sum }_{i=1}^{n_1}\mathbf{1}\left(x_i\le t\right),{\widehat{F}}_Y\left(t\right)=\frac{1}{n_2}{\sum }_{j=1}^{n_2}\mathbf{1}\left(y_j\le t\right)$$ 2

Let $\mathcal{V}$ = $\text{unique}$ ($X\cup Y$) be the set of all observed values. The test statistic $D$ was defined as the maximum absolute difference between ECDFs across all observed values:

$$\begin{array}{c}\hfill D={\max }_{t\in \mathcal{V}}\mid {\widehat{F}}_X\left(t\right)-{\widehat{F}}_Y\left(t\right)\mid \hfill \\ \hfill t^{*}=\arg {\max }_{t\in \mathcal{V}}\mid {\widehat{F}}_X\left(t\right)-{\widehat{F}}_Y\left(t\right)\mid \hfill \\ \hfill \mathrm{\Delta }\left(t^{*}\right)={\widehat{F}}_X\left(t^{*}\right)-{\widehat{F}}_Y\left(t^{*}\right)\hfill \end{array}$$ 3

Here, $\mathrm{\Delta }\left(t^{*}\right)$ is the signed KS statistics, and $t^{*}$ is the value at which the maximum ECDF difference occurs. Let f_k be the pooled empirical frequency of value $v_k\in \mathcal{V}$:

$$f_k=\frac{1}{N}{\sum }_{i=1}^N\mathbf{1}\left(z_i=v_k\right),z=\left(X,Y\right)$$ 4

The tie-adjusted variance term is:

$${\sigma }_{\text{tie}}^2={\sum }_k{f}_k\left(1-f_k\right)$$ 5

Define the effective sample size:

$$n_{\text{eff}}=\frac{n_1{n}_2}{n_1+n_2}$$ 6

The statistic was converted to a Z score using a normal approximation:

$$Z=\frac{D}{\sqrt{{\sigma }_{\text{tie}}^2/n_{\text{eff}}}}$$ 7

Using a normal approximation, a two-sided p value was obtained:

$$p=2\left[1-\mathrm{\Phi }\left(\mid Z\mid \right)\right]$$ 8

where Φ(·) is the standard normal cumulative distribution function. In summary, it was calculated as:

$$Z=\frac{{\max }_{t\in \mathcal{V}}\mid {\widehat{F}}_X\left(t\right)-{\widehat{F}}_Y\left(t\right)\mid }{\sqrt{\left({\sum }_k{f}_k\left(1-f_k\right)\right)\left(\frac{n_1+n_2}{n_1{n}_2}\right)}}$$ 9

The signed ECDF difference Δ(t*) was reported as the gene-level APA change (range from −1 to 1).

Genomic sequence composition and motif feature analysis

For human datasets, genomic sequences surrounding CSs and PAS peaks were extracted using the R package BSgenome.Hsapiens.UCSC.hg38 based on hg38 genome coordinates⁷⁹. Annotated PASs were defined using 3′ ends of transcript isoforms from GENCODE v43 gene annotation and PAS annotations from PolyA_DB v3.2. For Drosophila melanogaster datasets, sequences were extracted using BSgenome.Dmelanogaster.UCSC.dm6 based on the BDGP6 (dm6) genome assembly, and annotated PASs were defined using 3′ ends of transcript isoforms from Ensembl gene annotation (release 110). Motif analysis was performed using the Biostrings package⁸¹ to identify motifs, and motifStack was used to generate sequence logos and quantify position-specific sequence content⁸². For motif distribution analysis, all motif occurrences within each sequence window were counted, and each occurrence was weighted by 1/L, where L is the motif length, to account for positional contribution. For calculation of the fraction of PASs containing a given motif, each motif was counted once per PAS, regardless of multiple occurrences within the same sequence. For the polyA strength of each PAS, the ±100 nt surrounding the peak was used as input for PolyaModelsHuman⁵⁹, the highest score in the output at a position within 50 nt to the peak was used to represent the strength of the given PAS. GO analysis was performed using clusterProfiler package⁸³. Data visualization was carried out using base R and ggplot2. Figures were prepared using Adobe Illustrator.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization, Z.Z. and P.M.; Methodology, Z.Z., D.S. and P.M.; Sample preparation, H.GD. and Z.Z.; Data analysis, Z.Z.; Package development, Z.Z., L.Z., D.S. and P.M.; Writing, Z.Z. and P.M.; Funding acquisition, P.M.; Supervision, P.M.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Funding

This work is supported by NIGMS grant R35GM138319 awarded to P.M.

Data availability

The DRS data generated from human cells and fly embryos in this study have been deposited in SRA under PRJNA1247144. ENCODE PacBio LR RNA-seq data collected from dorsolateral prefrontal cortex (ENCSR205QMF, ENCSR094NFM, and ENCSR463IDK) and right lobe of liver (ENCSR657HJW and ENCSR293MOX) were used for APA analysis.

Code availability

Software is available at GitHub [https://github.com/markandtwin/APALORD]⁸⁴.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76377-8.