Abstract

Splicing, including alternative splicing, is a fundamental post-transcriptional mechanism in eukaryotes that generates functional proteins and transcript diversity. Canonical splicing follows well-defined rules, such as sufficient intron and exon lengths, specific splice junction orientations, consensus dinucleotides at donor and acceptor sites, and mediation by the spliceosome. However, certain splicing events deviate from these canonical rules. This review synthesizes multiple forms of non-canonical splicing within a framework that reflects their increasing deviation from canonical mechanisms, including non-canonical splice sites, non-canonical splicing in lncRNAs, microexons, recursive splicing, trans-splicing, and spliceosome-independent splicing. In addition, this review provides a critical analysis of the current state of research for each form of non-canonical splicing and outlines key directions for future investigation. As a case study, we reanalyzed RNA-seq data from mouse neuronal cells to further examine non-canonical splice sites. These analyses show that non-canonical introns tend to be shorter and that many non-canonical junctions retain at least one canonical donor or acceptor dinucleotide, supporting the view that a substantial subset remains compatible with spliceosome-mediated recognition. Together, this review provides a structured perspective on how canonical splicing rules can be relaxed, repurposed, or bypassed across distinct biological contexts.

1 Introduction to RNA splicing

Splicing is a post-transcriptional RNA processing mechanism widely conserved among eukaryotes, in which non-coding introns are excised and coding exons are ligated, generally for translation into functional proteins. Across eukaryotes, splicing occurs in multiple forms, ranging from canonical constitutive and alternative splicing to diverse non-canonical mechanisms (Figure 1, Table 1). Under normal circumstances, the spliceosome recognizes a canonical 5′splice donor, typically a GU dinucleotide at the 5′end of the intron, and a canonical 3′splice acceptor, typically an AG dinucleotide at the 3′end, to define each splicing reaction. Both canonical splice sites and the mediation of the spliceosome are essential components of canonical splicing. Canonical splicing is sometimes completed through a strict adherence to a fixed exon-intron structure, yielding a single, obligatory mature transcript; this process is termed constitutive splicing. Conversely, through alternative combinations of exons within a single gene, known as alternative splicing, this process generates multiple mRNA isoforms that expand proteomic diversity and contribute to gene regulation (Modrek and Lee, 2002; McManus and Graveley, 2011; Kelemen et al., 2013; Marasco and Kornblihtt, 2023) (Figure 2).

An overview flowchart illustrating canonical and non-canonical splicing. Canonical splicing comprises constitutive splicing and alternative splicing, whereas non-canonical splicing encompasses splicing with non-canonical splice sites, non-canonical splicing in long non-coding RNAs (lncRNAs), microexons, recursive splicing, trans-splicing, and spliceosome-independent splicing.

ClassCharacteristicsCanonical constitutive splicingAll exons within a single pre-mRNA are joined sequentially to produce a single transcript isoformCanonical alternative splicingExons within a single pre-mRNA are joined in a selective manner, with one or more exons potentially skipped, thereby generating multiple transcript isoformsNon-canonical splice sitesOne or both nucleotides at the splice junction deviate from the canonical GT-AG consensus, resulting in splice sites that do not conform to the standard motifNon-canonical lncRNA splicingCommonly involves splicing at non-canonical splice sites, most notably GC-AG splice junctionsMicroexon splicingSplicing involving extremely short exons, typically only 3–15 nucleotides in lengthRecursive splicingSplicing of exceptionally long introns that contain multiple internal splice sites in addition to their terminal boundaries, allowing intron removal through a stepwise series of successive splicing reactionsBack splicingA splicing event in which a downstream splice acceptor is joined to an upstream splice donor, often giving rise to circular RNAs (circRNAs)Trans-splicingA splicing process in which a transcript is generated by joining sequences derived from multiple pre-mRNAs of the same gene, different genes, or even different chromosomestRNA splicingAn essential step in tRNA maturation, mediated by a set of enzymes distinct from the spliceosomeIRE1-mediated splicingA spliceosome-independent form of cytoplasmic splicing mediated by IRE1, occurring only in specific target transcripts such as XBP1 or HAC1

Comparison of the canonical and non-canonical splicing types discussed in this review, together with their principal characteristics.

(A) Schematic representation of canonical splicing. In constitutive splicing, all exons within a pre-mRNA are sequentially ligated in their genomic order to generate a single mature transcript. In contrast, alternative splicing selectively includes or skips specific exons, producing multiple RNA isoforms that can be translated into distinct protein variants. (B) Simplified mechanism of canonical splicing. Small nuclear ribonucleoproteins (snRNPs) recognize the 5′ splice donor (GU) and the 3′ splice acceptor (AG) flanking adjacent exons. Multiple ribonucleoprotein components assemble into the spliceosome complex, which excises introns and ligates exons to form mature mRNA.

In eukaryotes, pre-mRNA splicing is carried out by two related ribonucleoprotein machineries: the major (U2-type) spliceosome, which mediates more than 99% of splicing events, and the minor (U12-type) spliceosome, which accounts for less than 1% (Lin et al., 2010; Kretova et al., 2023). The U2-type spliceosome is typically composed of the small nuclear ribonucleoproteins (snRNPs) U1, U2, U4, U5, and U6, whereas the U12-type spliceosome generally consists of U11, U12, U4atac, U5, and U6atac (Bradley and Anczukow, 2023; Kretova et al., 2023). The intron classes processed by these two spliceosomes differ in the conservation and architecture of their recognition signals. U2-type introns generally exhibit less constrained splice-site consensus sequences, which contributes to greater flexibility in splice-site selection and facilitates alternative splicing; they also typically contain a more prominent 3′ polypyrimidine tract (PPT, also known as Py-tract). In contrast, U12-type introns are characterized by more highly conserved recognition elements. Their 3′-end recognition depends more strongly on a highly conserved branch point sequence, and they often lack the canonical 3′ PPT typical of U2-type introns (Wu and Fu, 2015; Akinyi and Frilander, 2021). With respect to terminal splice-site dinucleotides, canonical U2-type splicing predominantly follows the GT–AG rule, whereas U12-type splicing is most associated with the AT–AC motif, although a small subset of U12-type introns also uses GT–AG motif (Turunen et al., 2013; El Marabti et al., 2021).

A canonical splicing event typically follows these principles: introns and exons of sufficient lengths, major canonical donor and acceptor sites (GT–AG) contributed sequentially by the preceding exon and the following exon, and mediation by the spliceosome. However, exceptions to these canonical rules have been increasingly documented. Collectively termed non-canonical splicing, these events do not strictly follow this paradigm. Some remain partially recognized and processed by the spliceosome, whereas others proceed through distinct mechanisms and yield diverse molecular outcomes.

2 Splicing with non-canonical splice sites

While canonical splicing relies on well-defined splice site signals and a conserved exon–intron architecture, accumulating evidence indicates that splice site recognition is not an all-or-nothing process. Instead, the spliceosome exhibits a measurable degree of tolerance to deviations from the major canonical GT–AG consensus, giving rise to splicing events that operate at the boundary of canonical rules. These deviations are most prominently manifested at the level of splice site sequences themselves and represent one of the most fundamental forms of non-canonical splicing. Accordingly, an examination of non-canonical splice sites provides a critical entry point for understanding how canonical splicing principles are relaxed, modified, or repurposed to achieve regulatory flexibility.

Splicing events involving non-canonical splice sites do not conform to the canonical GT–AG motif. Depending on the context, such events may involve a non-canonical donor site, a non-canonical acceptor site, or deviations at both ends (Figure 3). Although splice-site recognition by the spliceosome is dominated by canonical GT–AG signals, multiple lines of evidence indicate that spliceosomal processing is not strictly limited to GT–AG junctions. A subset of non-canonical splice sites can still be recognized and excised when their surrounding sequence context retains features compatible with U2- or U12-type splice-site recognition (Turunen et al., 2013; Parada et al., 2014; Akinyi and Frilander, 2021). In such cases, the terminal dinucleotides deviate from the canonical pattern, but the broader splice-site architecture remains sufficiently similar to known major or minor spliceosomal recognition signals. In addition, some non-canonical splice sites may become functionally more canonical-like through context-dependent mechanisms such as A-to-I RNA editing or may show evolutionary relationships to nearby or ancestral canonical splice sites (Parada et al., 2014). Non-canonical junctions that retain recognizable U2/U12-like consensus features are often described as U2/U12-like non-canonical junctions. By contrast, non-U2/U12-like non-canonical junctions are not readily explained by established major or minor spliceosomal recognition models and may represent a heterogeneous class that includes both mechanistically distinct RNA-processing events and lower-confidence events influenced by technical artifacts (Parada et al., 2014; Sibley et al., 2016).

Schematic illustration of splicing with non-canonical splice sites. Variations in splice sites can be categorized into events involving a non-canonical splice donor only, a non-canonical splice acceptor only, or deviations at both sites. Among these events, some are still tolerated and mediated by the U2/U12 spliceosome, whereas others appear to proceed through diverse and gene-specific mechanisms that remain to be further elucidated.

Non-canonical splice sites are exceedingly rare among all splicing events. In a previous systematic baseline study, splice sites outside the canonical GT–AG class accounted for only about 1% of all events (Parada et al., 2014). Moreover, when the major U2/U12-like non-canonical classes, namely GC–AG and AT–AC, are excluded, the proportion of all remaining minor non-canonical splice-site combinations decreases further to only 0.08% (Parada et al., 2014). Similarly, another study focused on non-canonical 5′ splice sites found that non-canonical 5′ splice junctions represented only 0.4%–1% of all exon junctions (Erkelenz et al., 2018). Taken together, these observations indicate that splicing involving non-canonical splice sites constitutes a minority class rather than a mainstream mode of splicing.

For such rare splicing events, methodological differences can substantially affect prevalence estimates. For example, in the earlier systematic baseline study, after functional relevance filtering, the number of retained non-canonical splice sites was reduced to 28% of the broader set identified by library-guided detection (462 of 1,630), and after further classification into U2/U12-like sites, this number dropped even more sharply, to only 11% of the initially detected candidate set (184 of 1,630) (Parada et al., 2014). Likewise, in the study of non-canonical 5′ splice sites, the raw RNA-seq estimate placed non-canonical 5′ splice junctions at 2.2%, whereas removal of low-usage junctions likely to represent noise reduced the high-confidence proportion to only 0.4%–1% (Erkelenz et al., 2018). These findings show that differences in filtering stringency and functional criteria can materially alter the reported prevalence of non-canonical splice sites across studies; however, even when considered together, published estimates consistently place them within a very small fraction of total splicing events. Moreover, because splicing itself is inherently heterogeneous and dynamic, caution is warranted when applying functionality-based filters to non-canonical splice sites, as overly stringent or overly permissive thresholds may distort their apparent prevalence.

With respect to sequencing modality, the overall distribution of non-canonical splice sites inferred from long-read sequencing has not shown major divergence from estimates derived from short-read sequencing. Nevertheless, the large number of novel splice junctions uncovered by long-read approaches remains an important subject for further investigation. A genome-wide survey of non-canonical splice sites in plants reported that, under long-read sequencing, all non-canonical splice sites together accounted for approximately 1.35%, whereas minor non-canonical splice sites represented only 0.09% (Pucker and Brockington, 2018). In addition, a long-read transcriptomic characterization of mouse neural tissue reported that only 0.1% of all known splice junctions were non-canonical (Tardaguila et al., 2018), and a study of rainbow trout genome annotation likewise reported approximately 0.1% non-canonical junctions (Ali et al., 2021). Collectively, these long-read-based studies are broadly consistent with the view that non-canonical splicing represents a very low-abundance class of splicing events.

At the same time, long-read sequencing has revealed a large number of previously undetected and uncharacterized splice sites, typically classified as novel splice junctions, which appear to have remained relatively underexplored. Many studies report these novel splice junctions as a separate category in long-read RNA-seq analyses; however, because the primary focus of those studies often lies elsewhere, these junctions are not always subjected to further filtering, remapping, or mechanistic analysis (Wright et al., 2022; Humphrey et al., 2025). Determining to what extent such novel junctions correspond to bona fide non-canonical splicing events therefore represents an important emerging challenge. In this regard, the earlier long-read transcriptomic study of mouse neural tissue is particularly informative, as it further analyzed novel splice junctions and found that 31% were classified as non-canonical junctions (Tardaguila et al., 2018). This suggests that long-read sequencing can uncover a substantial number of novel isoforms and potentially informative splice features that may be missed, or at least insufficiently resolved, by short-read approaches. However, an important issue has also emerged: novel-calling in long-read data appears to be highly sensitive to pipeline design and artifact control (Tardaguila et al., 2018; Su et al., 2024), indicating that further improvements in quality-control procedures are still needed. Moreover, there is currently a lack of dedicated long-read-based baseline surveys of non-canonical splice sites, making it difficult to determine whether the large number of novel isoforms identified by long-read sequencing substantially alters the baseline frequencies previously inferred from short-read data.

Non-canonical splice sites can also display tissue-specific expression patterns. In terms of gene regulation, they are often surrounded by an increased density of splicing regulatory elements (SREs), indicating that more complex regulatory architectures may be required for their recognition (Parada et al., 2014; Vuong et al., 2016; Yang et al., 2023). Moreover, certain non-canonical splice sites themselves contribute to gene regulation; for example, some can trigger nonsense-mediated mRNA decay (NMD). In the mouse brain, specific 3′splice sites within the Syngap1 gene, which encodes the Ras GTPase-activating protein SYNGAP1, have been shown to induce NMD, leading to mRNA degradation and haploinsufficiency of SYNGAP1 protein and thereby affecting neuronal function and behavior (Yang et al., 2023). Coincidentally, additional non-canonical splice sites and truncated exon variants have also been reported in Syngap1 (Kilinc et al., 2022). Building on these observations, we re-analyzed RNA-seq data from mouse neuronal cells described in the above studies (Kilinc et al., 2022; Yang et al., 2023) to further characterize the prevalence and sequence features of non-canonical splice sites.

2.1 Sequence enrichment of non-canonical splice sites

A sequence-enrichment analysis was performed on RNA-seq reads from EGFP+ cells sorted from E14.5 Tubb3-EGFP transgenic mice (Yang et al., 2023). Among all detected splice junctions, canonical GT–AG junctions were the most abundant, accounting for approximately 98.9% of the total (Table 2). These were followed by the major U2/U12 spliceosomal non-canonical preference sites, GC–AG and AT–AC, which accounted for approximately 0.83% and 0.12%, respectively, with the remaining fraction comprising other minor non-canonical splice sites (Table 2). The most prevalent splice-site classes and their relative proportions were highly similar to those reported in previous baseline studies (Parada et al., 2014).

Splice siteCount%Splice siteCount%Splice siteCount%GTAG149,78698.9438GCCC20.0013CCGT10.0007GCAG1,2530.8277GCGG20.0013GCGA10.0007ATAC1850.1222ACCG10.0007TGTT10.0007GTTG330.0218AGTG10.0007GGAA10.0007GTGG210.0139CTAG10.0007GGTC10.0007ATAG110.0073GTTC10.0007GGCC10.0007GAAG90.0059TGGA10.0007CTCC10.0007TTAG80.0053TCAG10.0007TGTC10.0007ATAA80.0053ACAC10.0007TTTT10.0007ATAT60.0040GCCT10.0007AGGG10.0007GGAG60.0040AGCC10.0007GGGT10.0007GTAA40.0026TTCT10.0007CTAT10.0007GGAC30.0020GGGC10.0007CTAA10.0007GTCA30.0020CCTT10.0007CAAC10.0007TGTG20.0013TCCC10.0007TCTT10.0007TATC20.0013CCCC10.0007AAGG10.0007GGGA20.0013CCGG10.0007TCGG10.0007AGAG20.0013CTGA10.0007GACA10.0007TAGA20.0013GATG10.0007

Distribution of splice-site dinucleotide combinations among all detected splice junctions. Each splice-site class is defined by the donor and acceptor terminal dinucleotides. The table reports the absolute count and relative frequency (%) of each splice-site combination across all detected junctions.

Examination of the 5′- and 3′-terminal dinucleotides of all non-canonical introns revealed that AG remained by far the most frequently used non-canonical acceptor. Although GT was less frequent among non-canonical donor sites than GC and AT, which were observed mainly in the U2/U12-like GC–AG and AT–AC junctions, GT still occurred more frequently than most other minor donor classes (Figures 4A,B). In other words, GT donor sites and AG acceptor sites remained representative among non-canonical splicing events, which means a proportion of the non-canonical junctions were only non-canonical at one end and retained a canonical dinucleotide at the opposite splice site. These findings further indicate that many mapped non-canonical splice sites still retain at least one canonical donor or acceptor dinucleotide, strongly suggesting that the spliceosome can accommodate a certain degree of non-canonical variation at one or both ends. Although the major U2/U12-like non-canonical classes, GC–AG and AT–AC, accounted for a large fraction of this category, it should also be recognized that this group still includes a subset of single-sided non-canonical splice sites (Table 2), whose underlying splicing mechanisms are also expected to be compatible with spliceosomal recognition. Among splice sites lacking the GT–AG motif, GC and AT emerged as two frequent alternative donor dinucleotides, whereas AC was the most common alternative acceptor (Figures 4C,D). This enrichment is also likely attributable to the frequent occurrence of GC–AG and AT–AC junctions. In addition, TG and GG were also observed at relatively high frequencies as alternative acceptors (Figures 4C,D). Notably, the GG acceptor displayed potential functional links to the non-canonical splicing events reported in SYNGAP1 and to the activation of nonsense-mediated mRNA decay (NMD) (Kilinc et al., 2022; Yang et al., 2023), highlighting a promising direction for future investigation.

4 × 4 dinucleotide heat-map analysis of non-canonical splice sites: (A) donors, (B) acceptors, (C) donors with GT excluded, and (D) acceptors with AG excluded. All non-canonical splice sites were first counted and then converted to frequencies. Rows correspond to the first nucleotide of the splice site and columns to the second. Heat-map colors were scaled using power-law normalization (PowerNorm, gamma = 0.5) to enhance the visual separation of low-frequency minor non-canonical dinucleotide classes. This normalization was applied only to the color scale while the underlying counts/frequencies and the numeric labels remained untransformed.

2.2 Difference in intron length between canonical and non-canonical introns

Within the same sample, introns excised at non-canonical splice sites were significantly shorter than canonical introns, with a median length difference of 415.5 bp (Figure 5; Mann–Whitney U test, p = 3.657 × 10−18). After further stratification of all non-canonical splice introns into U2/U12-like non-canonical and non-U2/U12-like non-canonical subsets, both groups likewise showed significantly shorter introns than canonical introns. Specifically, the median intron length of the U2/U12-like non-canonical subset was 343 bp shorter than that of canonical introns (Figure 5; p = 2.904 × 10−12), whereas the median intron length of the non-U2/U12-like non-canonical subset was 659.5 bp shorter (Figure 5; p = 1.494 × 10−9). These findings strongly suggest that, regardless of whether a non-canonical junction is still likely to be processed by the U2/U12 spliceosome, its associated intron tends to be shorter than those of canonical splice sites. Interestingly, introns in the non-U2/U12-like non-canonical subset were also significantly shorter than those in the U2/U12-like non-canonical subset (Figure 5; p = 2.544 × 10−3). These results indicate that non-U2/U12-like non-canonical introns exhibit a more pronounced shortening relative to canonical introns than do U2/U12-like non-canonical introns, which may reflect underlying differences in their splice-site recognition mechanisms, including possible differences in sequence conservation. Accordingly, a tendency toward shorter introns may serve as a useful indicator in future studies of non-canonical splicing for further identifying and prioritizing non-U2/U12-like splicing events.

(A) Empirical cumulative distribution function (ECDF) plot and (B) box plot showing differences in intron lengths (log10-scaled) between canonical and non-canonical introns (Mann–Whitney U test, p = 3.657 × 10−18), canonical and non-canonical U2/U12-like introns (p = 2.904 × 10−12), canonical and non-canonical non-U2/U12-like introns (p = 1.494 × 10−9), and non-canonical U2/U12-like and non-U2/U12 like (p = 2.544 × 10−3). Individual points shown in the boxplots represent outlier observations only, defined as values lying beyond 1.5 × the interquartile range from the box, rather than all underlying data points.

2.3 Future perspectives of splicing with non-canonical splice sites

Our analysis indicates that a substantial proportion of non-canonical splicing events recognized in mouse neuronal cells involve splicing with only a single non-canonical splice site, remarkably for the acceptor. This suggests that many non-canonical splicing events remain mediated by the U2/U12 spliceosome. At the same time, beyond the major U2/U12-preferred non-canonical junction classes, such as GC–AG and AT–AC, some junctions with a splice-site change at only one end are also likely to be tolerated and successfully processed by the U2/U12 spliceosome. In addition, introns involved in splicing with non-canonical splice sites, especially non-U2/U12-like non-canonical splice sites, are significantly shorter than those associated with canonical splicing events. Given the persistent challenges in accurately identifying diverse forms of non-canonical splicing, a bias toward shorter intron length may serve as an auxiliary criterion for the detection and characterization of non-canonical splicing events. Moreover, as non-canonical splice sites in SYNGAP1 mRNA can potentially trigger NMD, leading to reduced dosage or altered forms of proteins essential for neurodevelopment, the role of aberrant splicing in the pathogenesis of various diseases warrants further investigation.

The continued development of long-read sequencing has revealed many previously unclassified novel splicing isoforms, which are likely to contain substantial numbers of as-yet undiscovered, unclassified, and uncharacterized non-canonical splice sites. Accordingly, systematic investigation of non-canonical splice sites in the context of long-read sequencing, together with the establishment of comprehensive baseline surveys comparable to those developed in the short-read era, represents an important and timely direction for future research.

While the above features outline the general landscape of non-canonical splice sites across transcripts, their distribution and functional relevance are not uniform across RNA classes. Notably, long non-coding RNAs (lncRNAs) exhibit a distinct enrichment of non-canonical splice sites (Khan et al., 2023), providing a permissive context in which spliceosomal tolerance may be amplified and repurposed for regulatory functions.

3 Non-canonical splicing in long non-coding RNAs (lncRNAs)

Compared with protein-coding genes, lncRNAs display a markedly higher frequency of non-canonical splice sites, reflecting reduced coding constraints and increased splicing plasticity (Figure 6). lncRNAs are RNA transcripts longer than 200 nt but lack the potential to be translated into proteins (Mattick et al., 2023). lncRNAs also undergo splicing, and the vast majority are mediated by the spliceosome. However, compared to mRNA, lncRNA exhibits distinct differences in selective pressure, functional constraints, and splicing precision (Derrien et al., 2012; Chen and Kim, 2024). This renders lncRNA a domain where non-canonical splicing occurs more readily yet proves more challenging to define.

A representative and relatively conserved form of non-canonical splicing in lncRNAs. This scenario exemplifies splicing with a non-canonical donor only, in which a non-canonical dinucleotide (GC) replaces the canonical GU at the 5′ splice donor of the first exon in an lncRNA transcript.

3.1 Instability of lncRNA splicing mechanisms

Compared with mRNA splicing, lncRNA splicing is generally characterized by weaker splice site strength, lower conservation of exon-intron architecture, and a markedly higher prevalence of alternative splicing, with intron retention and exon skipping representing common features rather than exceptions in lncRNAs (Ouyang et al., 2022; Khan et al., 2023). Notably, alternative splicing of lncRNAs is particularly prevalent and can regulate selective splicing through multiple mechanisms, including interactions with splicing factors, base pairing with pre-mRNAs, and modulation of chromatin remodeling (Pisignano and Ladomery, 2021; Ouyang et al., 2022). Such variability and instability in splicing provide a broad landscape for the emergence and diversification of non-canonical splicing events.

3.2 Non-canonical splice sites in lncRNAs

The presence of weak splice sites and the high splicing plasticity characteristic of lncRNAs provide a structural basis for non-canonical splicing, particularly for the emergence and utilization of non-canonical splice sites. The occurrence of non-canonical splice sites is markedly higher in lncRNAs than in protein-coding genes (PCGs) (Khan et al., 2023). The GC–AG pattern represents a non-canonical splicing mode that is markedly enriched in lncRNAs. In both human and mouse lncRNAs, GC–AG motifs occur at frequencies approximately two-to four-fold higher than those observed in protein-coding genes and are preferentially enriched within the first intron (Abou Alezz et al., 2020). In splicing events involving non-canonical sites, both donor and acceptor splice sites are significantly weaker, and the associated polypyrimidine tracts are shorter (Abou Alezz et al., 2020; Statello et al., 2021). Such structural features render these sites more susceptible to further alternative splicing and show a strong association with alternative polyadenylation, collectively indicating that non-canonical splice sites play a regulatory role in lncRNA splicing (Abou Alezz et al., 2020).

Non-canonical splicing in lncRNAs remains a relatively infrequent class of RNA-splicing events overall, yet its reported prevalence is consistently higher than that observed in protein-coding genes (PCGs), with no major disagreement across published studies. A previous review summarized that atypical splice sites account for approximately ∼3% of splice sites in lncRNAs, compared with ∼0.8% in PCGs (Khan et al., 2023). Another study similarly reported that non-GT–AG introns comprise approximately 3.4% of lncRNA splice junctions in humans, compared with 1.4% in PCGs; in mice, the corresponding values were 3.4% and 1.2%, respectively (Abou Alezz et al., 2020). Together, these studies support a consistent overall conclusion: although non-canonical splicing in lncRNAs remains a minority phenomenon, its usage is higher than in protein-coding genes.

3.3 Future perspectives of non-canonical splicing in lncRNAs

lncRNAs undergo pervasive alternative splicing, frequently generating aberrant isoforms implicated in oncogenesis (Ouyang et al., 2022; Wang et al., 2022). The extensive splicing variability of lncRNAs, together with their documented roles in tumorigenesis, demonstrates that lncRNAs are far from mere transcriptional noise but instead represent a long-overlooked class of regulatory molecules with substantial functional significance in gene regulation. Beyond the already highly diverse landscape of canonical alternative splicing, lncRNAs tha