Intron retention in health and amyotrophic lateral sclerosis Authors: Chloe Y Wang, Stephanie Taylor, Virenkumar A Pandya, Benjamin E Clarke, Koustav Pal, Tatyana A Shelkovnikova, Yiran Wang, Raphaëlle Luisier, Rickie Patani Published: 2026 ABSTRACT AbstractIntron retention (IR) is the molecular phenomenon by which introns, historically thought to represent non-coding ‘junk’, remain unspliced within pre-mRNA transcripts, resulting in their incorporation into the mature mRNA molecule. While the role of IR is well established in species of plant, fungi, insects and viruses, it remains relatively understudied in mammalian biology. It was previously assumed that IR only played a limited role in downregulating a transcript’s translation potential through downstream initiation of nuclear detention or nonsense mediated decay (NMD). However, recent studies highlight IR’s significantly more complex and dynamic contribution to cellular physiology and disease. In particular, a role for IR is emerging in both health and neurodegenerative diseases, including amyotrophic lateral sclerosis (ALS), a rapidly progressive and invariably fatal disease that renders patients paralysed and unable to eat, speak or breathe. Significant technological advances now permit a comprehensive interrogation of previously unrecognized aspects of RNA metabolism in clinically relevant human cell types.In this review, we focus on the differential role(s) of nuclear and cytoplasmic intron retaining transcripts (nIRTs and cIRTs, respectively), as well as how IRTs may influence subcellular localization of ribonucleoprotein (RNP) complexes, loss of function of bound RNA binding proteins (RBPs) and liquid-liquid phase separation (LLPS) in physiology and disease. Additionally, we discuss the potential of IRTs as independent regulatory elements beyond their protein-coding functions and highlight how artificial intelligence is poised to accelerate discoveries in this area. In the context of IR’s increasing appreciation, we also highlight its potential as a therapeutic target and explore current and future challenges in this burgeoning field. Introduction Neurodegenerative disorders are often taxonomized as being predominantly protein aggregation disorders (e.g. Alzheimer’s disease or AD) or those related to defects in RNA metabolism (e.g. amyotrophic lateral sclerosis or ALS). However, this is potentially a facile dichotomy, given that both processes are fundamentally implicated in different forms of neurodegeneration and this is certainly the case in ALS, a rapidly debilitating and universally fatal neurodegenerative disorder primarily involving motor neurons (reviewed elsewhere).1 The common molecular denominator of these implicated pathogenic processes is the ribonucleoprotein (RNP) complex, composed of RNA and RNA binding proteins (RBPs), which plays a multilayered role in the regulation of gene expression.2 Indeed, it has been two decades since TAR DNA binding protein 43 (TDP-43) protein was discovered as the major constituent of ubiquitinated cytoplasmic inclusions in the vast majority of ALS and about 50% of frontotemporal dementia (FTD) cases.3 This finding galvanized significant research focus towards TDP-43 and its interactome for both mechanistic discovery and therapeutic target development. TDP-43 is an RBP and a DNA binding protein. It predominantly resides in the nucleus and plays key roles in pre-RNA splicing, mRNA stabilization, RNA transport, DNA repair mechanisms and biocondensate formation.4 In physiology, TDP-43 shuttles to the cytoplasm and contributes to the fate of processed RNA targets. TDP-43’s mislocalization, phosphorylation, cleavage and aggregation in the cytoplasm with nuclear depletion is a hallmark feature of ALS motor neurons, and has been recently reviewed comprehensively in this context.5 Given the dearth of life-extending options for ALS patients, a deeper mechanistic understanding of how deregulated RNA metabolism precisely contributes to neurodegeneration remains a key priority, with the potential to yield novel therapeutic opportunity and clinical impact. The canonical information ‘flow’ from DNA through messenger RNA (mRNA) to protein (Fig. 1A) necessitates a series of increasingly understood, albeit complex, events. In general, following transcription, pre-mRNA undergoes 5′-end capping, splicing, 3′-end cleavage and polyadenylation, following which mature mRNA is exported to the cytoplasm, localized, translated, stabilized or degraded. These processes are regulated by RBPs and non-coding RNAs such as microRNAs (miRNAs). The retention of an intronic sequence with mRNA can lead to multiple fates. Canonically spliced mRNA is typically exported to the cytoplasm for translation (A). However, intron retaining transcripts (IRTs) may be stored or ‘poised’ in the nucleus and only spliced following a specific stimulus (B), after which they feed into the canonical pathway (A) for translation. IRTs may also be degraded in the nucleus by the exosome (C) or play a role in cotranscriptional splicing regulation (D). If the IRT is exported to the cytoplasm, it may be degraded by nonsense-mediated mRNA decay (NMD; E), non-canonically translated into a (likely truncated) protein (F) or stabilized within the cytoplasm (G). Created in BioRender. Wang, C. (2026) https://BioRender.com/oxkr523. RBPs are one of the largest groups of proteins: exact estimates of the number of RBPs depends on methodology used, but can confidently be estimated at over 3400, and may be over 4200 when integrating high-throughput RBP detection studies in different human cell types.6 It is likely that further proteins will be identified as RNA binding and regulating, with Cys2-His2 zinc-finger (C2H2-ZNF) transcription factors as recent examples.7 RBPs exhibit high translational efficiency, protein stability and abundance compared with other protein classes.8 In addition, approximately 2000 genes encode miRNAs, with many of these expressed from the non-protein-coding genome.8-10 These studies cumulatively highlight the enormous regulatory complexity of RNAs. As part of the canonical splicing process, approximately 75% of introns are removed cotranscriptionally, i.e. during transcription;11,12 the remainder of splicing occurs post-transcriptionally. In particular, introns that flank alternative exons are sometimes excised post-transcriptionally.13,14 The human genome has the capacity to diversify the output of transcription through a process known as alternative splicing (AS). Intriguingly, transcripts from >90% of human genes undergo AS, where coding and non-coding fragments are alternatively skipped and joined.15,16 In most cases, this splicing out of introns ensures the production of appropriate protein isoforms in the correct cellular environments. In a form of AS however, namely intron retention (IR), the mature, polyadenylated mRNA transcript retains one or more introns. This process has recently gained much attention in mammalian biology. In the context of disease, it is conceivable that changes in the expression of subsets of increased intron-retaining transcripts (IRTs) observed in neurons and other CNS cell types represent either pathogenic or compensatory mechanisms, or have biomarker value, which further reinforces the need to investigate this relatively neglected form of AS to a greater degree. The established roles of intron retention: regulation of gene expression and beyond Historically, IR has been studied extensively in viruses, fungi and plants but its prevalence in mammalian physiology has only been appreciated over the last two decades. Indeed, one study found that transcripts from approximately three-quarters of human genes contain at least one retained intron, with IR occurring in a cell type-dependent manner.17 It was found that IR was broadly functioning to reduce the level of transcripts where and when they are not required.17 Given the sheer number of IRTs, these data have challenged the long-held view that IR merely represents ‘transcriptional noise’. IR is now increasingly recognized to be a regulated process (elaborated in Box 1), influencing gene expression in a variety of ways and this has been reviewed comprehensively elsewhere.18-20 Intronic sequences, like other non-coding regions of mRNA such as the 3′ untranslated region (UTR), have traditionally been studied for their regulatory roles in association with protein-coding transcripts. However, as early as 1994, it was proposed that RNAs derived from introns could have evolved functional roles in the cell.21 Since then, growing evidence suggests that introns and other non-coding sequences, which make up over 90% of pre-mRNA, also possess independent regulatory and catalytic functions. Intronic sequences, for example, can generate stable RNA fragments that persist in nuclear and cytoplasmic compartments. The parallels that exist between intron retention and 3′UTRs are summarized in Box 2. Regulation of intron-retaining transcripts Differential GC content between introns and exons has been posited to be a salient factor in spliceosomal recognition.22 In this model, introns with relatively high GC content use a mode of splicing that involves intron definition rather than the canonical exon definition. Higher GC content may lead to an increase in the complexity of secondary structures, which in turn can impact splice site accessibility. Intron definition refers to the recognition of splice signals within introns. In the context of an ‘intron definition’ mode of splicing, impaired splicing efficiency is more likely to result in an intron-retaining transcript (IRT) rather than skipped exons. Retained introns are generally shorter, flanked by weaker 5′ and/or 3′ splice sites and more evolutionarily conserved than constitutively spliced introns.17,23 It is noteworthy that this ‘IR code’ of cis-features has largely only been considered in the context of whole cell studies rather than compartment-specific interrogation, leaving the issue of whether nuclear versus cytoplasmic IRTs might have divergent IR codes unresolved. Noteworthy parallels between intron retention and 3′-untranslated regions Non-canonical roles for RNAs are precedented. Prior research has, for example, revealed an unexpected function of the non-coding portion of an mRNA transcript independent of its translation potential, hinting that understanding the full range of regulatory functions of mature polyadenylated transcripts, of which intron retaining transcripts (IRTs) are one subset, remains an active area of discovery. One example is Tp53inp2 transcripts, which were shown to regulate axon growth independently of translation in sympathetic neurons, owing to Tp53inp2’s unusually long 3-untranslated region (UTR). In non-neuronal cells however, the transcript is canonically translated, demonstrating cell type-specificity in this role.24 Like intronic sequences, 3′ UTRs can also generate stable RNA fragments that persist in nuclear and cytoplasmic compartments. Stable 3′ UTR cleavage fragments have been detected in neurons,25,26 though their precise functions remain largely unknown.26,27 Beyond gene expression regulation, 3′ UTRs can also contribute to subcellular compartmentalization. 3′ UTRs can promote the formation of cytoplasmic, membraneless organelles,28 creating specialized environments for protein translation with direct functional consequences.29,30 Moreover, the striking similarities among intronic sequences, 5′ UTRs and 3′ UTRs, beyond shared sequence features like enrichment for RNA binding proteins and micro RNA binding sites, highlight the need to consider these regions as interconnected rather than distinct entities. Recognizing their shared properties should inspire broader investigations, where discoveries in one region can guide and stimulate similar explorations in the others, ultimately uncovering new regulatory and functional roles. A little over a decade ago ‘stable intronic sequence RNAs’ or sisRNAs were identified in the Xenopus oocyte and described as being derived from excised intronic sequences.31-35 These were found to be surprisingly stable (up to 48 h), and abundant in the nucleus. More recently, there has been some convergence in nomenclature and the term ‘sisRNAs’ has been revised to encompass both some stable IRTs as well as excised introns, belonging to a broader class of long non-coding RNAs (lncRNAs). SisRNAs contribute to various cellular processes, including host gene regulation, acting as molecular sinks or sponges, and influencing protein translation.32,35,36 Notably however, while stable intron lariats are exported from the nucleus to the cytoplasm in a regulated manner,35 their excessive accumulation is detrimental to the cell.37,38 It is intriguing to contemplate whether these intronic fragments also possess regulatory or catalytic functions in essential biological processes, akin to other non-coding RNAs whose functional repertoires continue to evolve.39,40 Indeed, many lncRNAs originate from intronic regions, and some retain sequence features reminiscent of spliced-out introns.41,42 Thus, the idea that non-coding portions of mRNA such as introns have only limited regulatory roles which relate to their associated proteins is increasingly being challenged, highlighting the need to explore their full potential, paralleling the expanding roles of non-coding RNAs more generally. Many fundamental processes including cell cycle progression,43 cellular differentiation,44 activation of specific cellular substates,45 apoptosis46 and stress responses47 are now known to relate to changes in IR of particular subsets of functionally related transcripts. The repertoire of IR is cell type- and state-specific, and most frequently observed in neural, immune and adipose cells.48,49 The retention of an intronic sequence within mRNA can lead to multiple fates for the transcript (Fig. 1B–G), which is to some degree determined by its localization within the cell. IRTs are often removed or degraded to prevent their translation,50,51 making them largely undetectable by traditional experimental approaches. However, those that persist can localize in the nucleus, referred to as nuclear intron-retaining transcripts (nIRTs), or be stably expressed in the cytoplasm, termed cytoplasmic intron-retaining transcripts (cIRTs), each exhibiting distinct functions and characteristics. Fate mapping experiments have revealed a substantial range in the stability of IRTs, with some taking days to reach their fate.52 Furthermore, transcripts with retained introns are reportedly less stable, suggesting a coupling between IR and degradation in regulating gene expression.45 Notably, there are studies that suggest a more direct link between IRTs and their involvement in transcriptional regulation53,54 and cotranscriptional splicing regulation55 (Fig. 1D). Nuclear functions of IRTs IRTs have been observed to function as nuclearly detained ‘poised’ transcripts that are then post-transcriptionally spliced in response to a specific stimulus (Fig. 1B).23 Transcripts with a high abundance of retained introns seem to be predominantly localized in the nucleus, suggesting that nuclear detention of incompletely spliced transcripts serves as a post-transcriptional mechanism for limiting the expression of mRNAs that would otherwise possess translation potential in the cytoplasm.56 One such example is the nuclearly pooled Cdc-like Kinases 1 and 4 (Clk1/4) IRT, which undergoes stress-specific splicing, enabling its export and translation, which in turn permits phosphorylation of the Serine/Arginine-rich (SR) family of splicing factors.57 Phosphorylated SR factors can then undergo nuclear import and contribute to spliceosome assembly.58 However, such a ‘reservoir’ model, where a precursor pool of IRTs undergoes post-transcriptional splicing, does not apply in all cases and is transcript/context dependent.59 Indeed, IRTs can also serve to simply downregulate expression of the cognate canonically spliced transcript, as alluded to above. The degradation of IRTs in the nucleus might occur through the exosome machinery60 (Fig. 1C). A notable example is the RBP fused in sarcoma (FUS), which forms an autoregulatory loop through intron retention, with transcripts detained in the nucleus to reduce the amount of cytoplasmic mRNA available for translation.61 A recent study considerably extends these findings and demonstrates that FUS IR also fulfils the reservoir model introduced above.62 However, it is noteworthy that this class of IRTs that conform to the ‘reservoir’ model represent distinct, transcript- and context-specific outcomes, and do not represent a universal ′logic' for all IRTs. Cytoplasmic functions of IRTs RNA degradation in the cytoplasm occurs through various decay pathways, including a process termed nonsense-mediated mRNA decay (NMD), which plays a crucial role in maintaining cellular homeostasis and clearing defective RNA species60 (Fig. 1E). This NMD-mediated degradation process is initiated by the presence of a premature termination codon (PTC) in the transcript, often located within the sequence of the retained intron(s). In this way, the presence of IRTs in the cytoplasm can initiate the NMD cascade, ultimately leading to degradation of those IRTs, and further regulation of gene expression.63 AS-NMD coupling represents a conserved regulatory mechanism in physiology. For example, it is paramount in the normal generation of certain haematological constituents including neutrophils, eosinophils and basophils.44 Post-transcriptional splicing and AS-NMD coupling of IRTs are comprehensively reviewed elsewhere.18,20 Although most cIRTs are rapidly degraded, and therefore should not be detected by RNA sequencing, the stable cytoplasmic localization of some intronic sequences in neurons has been reported since 2013.64 Furthermore, we detected that cIRTs undergo tight temporal regulation during human motor neurogenesis.49 In this context, cIRTs can coordinate the regulation of gene expression programmes in a spatiotemporally regulated manner, as demonstrated for mouse cortical neurons undergoing synaptic stimulation.65 Widespread cIRTs may affect or even overwhelm the efficiency of the NMD quality control system, thereby impairing its ability to degrade defective RNA species and leading to accumulation of ′faulty′ transcripts.66 cIRTs are able to evade the NMD surveillance pathway, potentially through concealing the retained introns by protein interactions or altered signals that target particular RNAs to NMD.67 Alternatively, since NMD is a translation-dependent process and stress can induce translational arrest, IRTs targeted by NMD may temporarily evade degradation under stress or specific stimuli, potentially gaining translational capacity in a context-dependent manner (Fig. 1E and F). For example, an IRT of the Arc gene, which contains two introns in its 3′ UTR, localizes to dendrites upon neuronal activation. Following stimulation, the IRT is translated before undergoing NMD.68 Another example of this occurs during axon guidance; specifically, the alternatively spliced isoform of Roundabout Guidance Receptor 3 (Robo3.2) transcript in mice contains a retained intron bearing an in-frame NMD-inducing PTC. However, the transcript can also lead to the production of a ROBO3.2 protein isoform only when crossing ventral midline structures, due to local translation initiated by floor plate signals.69 This spatially regulated control of protein expression allows fine control of commissural axonal trajectory. These examples are reviewed in detail elsewhere.70 Intron retention has also been shown to contribute to regulation of protein diversity in neurons by influencing exon inclusion. For example, by directing inclusion of the stress axis–regulated exon (STREX) to a subset of calcium-activated big potassium channel (BKCa) mRNAs, cIRTs coupled with proposed cytoplasmic splicing serve to increase diversity of BKCa splice variants in hippocampal neurons.71 The controversial possibility of cytoplasmic splicing is addressed in Box 3. Beyond translation or generation of protein diversity, IRTs in the cytoplasm may be stabilized for other potential functions, which are elaborated below (Fig. 1G). The controversy of cytoplasmic splicing The possibility of cytoplasmic splicing remains a highly controversial issue. The spliceosome is divided into the major and minor complex. The former is accepted to exert its function in the nucleus. There is also evidence that the minor complex functions predominantly in the nucleus,72,73 though one study reported a cytoplasmic presence.74 Both major and minor introns have been reported to undergo retention in a functionally important manner.17,75 Notably, both protein and RNA constituents of the spliceosome have been reported in primary rat neuronal projections, where isolated dendrites (physically separated from the nucleus), were reportedly capable of splicing.76 There are examples of cytoplasmic splicing that are well established, including that of IL1b cytoplasmic intron retaining transcripts (cIRTs), which accumulate in pro-platelet projections and persist into maturation, avoiding conventional splicing and NMD. Upon platelet activation, these IL1b cIRTs are spliced in the anuclear platelet and generate their cognate IL1b mRNA for translation.77 Furthermore, the example of pre-mRNA of X-box binding protein 1 (XBP-1) in the unfolded protein response (UPR)78 is noteworthy here. The accumulation of incorrectly folded proteins triggers the UPR, which signals through endoplasmic reticulum transmembrane receptors and activates transcription factor 6 (ATF6), which in turn triggers the downstream expression of (spliced) XBP-1. Activation of inositol requiring kinase 1 (IRE1) during UPR enzymatically catalyzes the splicing of a short intron from XBP-1 mRNA through the endoribonuclease activity of IRE1 itself. The resulting XBP-1 protein is a transcription factor that orchestrates various chaperones and protein degradation factors.79 This mechanism of cytoplasmic splicing was found to be non-canonical (it relies upon tRNA ligase and an endonuclease, and is spliceosome independent)80 and may bear relevance for other cIRTs, though more direct evidence of this is required. Non-canonical roles of intron retention Regulation of nucleocytoplasmic localization of RBPs The predominantly cytoplasmic endoribonuclease RNase L is activated by the innate immune response to degrade host and viral RNA, a response designed to reduce viral gene expression. This natural human defence mechanism can be harnessed to serve as an invaluable paradigm to understand the role of RNA in regulating the compartmental distribution of bound RBPs. Indeed, a recent study has demonstrated that upon RNase L activation and degradation of cytoplasmic RNAs, RBPs translocate from the cytoplasm to the nucleus, likely driven by the relative abundance of intact nuclear RNA. Noting the temporally coincident RNA processing changes in the nucleus, this redistribution of RBPs is likely to be functionally consequential.81 These data suggest that the nucleocytoplasmic localization of RBPs is in part determined by RNA target availability, at least in some circumstances. Therefore, by extension, the processes of transcription, RNA stability, RNA export and compartment-specific RNA decay will each contribute to determining RBP localization. Increasing recognition of cell type-specific RNP expression argues for caution when generalizing from public datasets to more specific experimental contexts. For example, we have previously also reported physiological cIRTs in the cytoplasm of astrocytes (Fig. 2A), which are significantly decreased in ALS astrocytes.82 One possible explanation for this finding presupposes that cIRTs may function to sequester RBPs from the nucleus or to serve as a ‘poised’ reservoir of extra-nuclear RBPs in the event they are needed to respond to a particular stress. Together with the decrease in cIRTs, we also noted an increase in NMD activity, which when taken together raise an intriguing hypothesis supporting the following event line: (i) stress → (ii) cIRT degradation (Fig. 2B) → (iii) liberation of RBPs bound to cIRTs (Fig. 2C) → (iv) nuclear import of these RBPs → (v) enhanced splicing of reactivity related transcripts (Fig. 2D and E) → (vi) nuclear export and translation into reactivity-related proteins (Fig. 2F and G) → (vii) reactive transformation of astrocytes. A similar principle may apply to the liberation of RBPs with localized functions following a stress that activates NMD in subcellular compartments of neurons, for example. Intron retaining transcripts may coordinate nucleocytoplasmic localization of RBPs. Physiological cIRTs exist in the cytoplasm of astrocytes (A), which are significantly decreased in ALS astrocytes.82 cIRTs may function in physiology to ‘store’ RBPs. Working model: (i) ALS ‘stress’ → (ii) cIRT degradation (B) → (iii) liberation of RBPs bound to cIRTs (C) → (iv) nuclear import of these RBPs → (v) enhanced splicing of reactivity-related transcripts (E and F) → (vi) nuclear export and translation into reactivity-related proteins (F and G) → (vii) reactive transformation of astrocytes. ALS = amyotrophic lateral sclerosis; cIRT = cytoplasmic intron retaining transcript; RBP = RNA-binding protein. Created in BioRender. Wang, C. (2026) https://BioRender.com/r4oepav. Regulation of miRNAs Several potential additional roles of IRTs have been speculated.19 For example, it has been suggested that IRTs might regulate non-coding RNAs contained within introns. IRTs could interplay with both the canonical and non-canonical pathways for miRNA processing. By the non-canonical pathway, miRNAs can be produced from intronic by-products of splicing which form lariat structures that are subsequently debranched and processed. As such, IRTs that are degraded could reduce the production of these, or if stable, remove a source of miRNA synthesis. Retained introns could also act in competition with miRNA targets to divert miRNA binding and might contain miRNA response elements. Indeed, miRNA putative binding sites are enriched in retained introns compared with non-retained introns.83 A reduction in small nucleolar RNAs (snoRNAs) could also represent a possible consequence of intron retention, since these are processed from excised introns.84 Although stable cIRTs were reported over a decade ago,64 their possible role(s) remain relatively understudied. Data from 2011 showed that intron sequences are retained in a subset of dendritically targeted mRNAs. The retained introns contained ID elements, a class of Short Interspersed Nuclear Element (SINE) retrotransposon, which were shown to be the dendrite-localizing mechanism.85 Our group studied time-resolved nucleocytoplasmic dynamics during human motor neurogenesis and demonstrated that IRTs are abundant and spatiotemporally regulated during motor neuron development. We found that cIRTs are detected in up to 13% (> 2000) of the expressed genes. Moreover, we identified a particular class of cIRTs, defined by their spatiotemporal expression profile, that do not correlate with reduced expression of their cognate genes but instead exhibit high capacity for RBP and miRNA occupancy, suggesting regulatory roles beyond controlling gene expression49,83 (Supplementary Fig. 1). Putative relationship to liquid-liquid phase separation Noting that not all IRTs are degraded49,64,86 raises the question of how they remain stable in the nucleus or cytoplasm. Although stability may be linked to the activity of degradative machinery in each compartment, there may additionally be other mechanisms at play involving liquid-liquid phase separation (LLPS). LLPS refers to the condensation of biomolecules into liquid-like non-membrane delimited organelles. LLPS encompasses protein:protein, protein:RNA and RNA:RNA interactions and is the molecular principle governing transient, non-membrane delimited, functional compartmentalization of cellular contents. Tightly regulated prion-like polymerization occurs driven by weak, transient molecular interactions between low complexity domains (LCDs) of RBPs and other multivalent protein or RNA interaction domains.87 This is also highly relevant to the aforementioned RNPs, which can phase-separate as part of their functional repertoire.88 Indeed, eukaryotic cells possess a plethora of stress-mitigation strategies, which include large changes in the assembly and/or disassembly of RNPs into ‘granules’ or biocondensates. These processes are likely tightly regulated by prion-like polymerization of RNPs (and/or their constituents).89 A plausible model is that IRTs form complexes with such biocondensates thereby rendering them relatively stable in either the nucleus or the cytoplasm. Indeed, a previous report demonstrated that intronic sequences have a great affinity for stress granules,90 a form of biocondensate. A recent notable study showed that granules may serve more structural roles as molecular ‘plasters’ following organelle damage.91 Given their likely close interaction, one possibility is that IRTs contribute to or further mediate this stress granule phenomenon, though this remains speculative and a hypothesis for future research. We have recently reported that splicing factor proline- and glutamine-rich (SFPQ) and an alternative cytoplasm-predominant isoform with context-specific translation potential both demonstrate markedly increased phase separation when exposed to an intronic sequence derived from an annotated SFPQ intron 9 retention event.92 There are also emerging lines of evidence that intronic sequences can drive LLPS in the nucleus. A notable example is the Serine/Arginine-rich Splicing Factor 7 (SRSF7). By binding to its pre-mRNA, SRSF7 protein promotes the inclusion of a poison cassette exon (PCE), leading to NMD. Conversely, in the context of elevated SRSF7 levels, NMD is inhibited and the translation of two protein ‘halves’, termed split-open reading frames (ORFs), occurs from the bicistronic SRSF7-PCE transcript. One half acts to suppress PCE inclusion, but promotes IR in the flanking introns, which are enriched with SRSF7 binding sites. Following SRSF7 binding at these intronic sites, it oligomerizes, leading to large nuclear bodies formed through LLPS. These phase-separated nuclear bodies in turn sequester SRSF7 nascent transcripts, thereby preventing their export and translation. This mechanism serves to regulate SRSF7 protein levels. Such split-ORFs appear to be a design principle for the regulation of a subset of RBPs in a context-dependent manner.93 A further recent example relates to FUS retained introns 6 and 7, where one intron was shown to have higher phase separation propensity than the other, at least partially caused by differential secondary structure formation of the intronic sequences.62 Modifications to RNA are noteworthy in the context of IRTs. One example is the m6A modification, which can regulate the stability,94 splicing95 and phase separation potential96 of the host transcript. These effects are likely to be influenced by factors such as the site, number and nature of modification, as well as the cellular context. The potential role of IRTs in addressing the unique homeostatic challenges of neurons Neurons have a truly unique and remarkable polarized cytoarchitecture, which correlates with mRNAs exhibiting highly distinctive spatial localization patterns,97,98 representing the molecular substrate for localized translation.99 In some cases, the localization is rather intuitive, for example mRNAs that encode membrane or secretory proteins predominantly localize to the endoplasmic reticulum (ER) membrane. However, the purpose of partitioning for some RNAs within the cell can be more elusive. It is plausible that cIRTs serve to ‘nucleate’ or seed cytoplasmic translationally repressed phase-separated ‘granules’ containing mRNAs, proteins and potentially other cargo during subcellular localization to the dendrites, axon or synapses. Expanding this concept, cIRTs may also functionally sequester cytoplasmic RBPs at specific subcellular sites, poised for liberation and action (Supplementary Fig. 2), for example following a change in phosphorylation status. Such molecular strategies may enable neurons to robustly respond to stressors at sites remote from the soma. Aberrant intron retention and nucleocytoplasmic compartmentalization in ALS The centrality of disordered splicing to ALS pathology has long been apparent given that the best known hallmark of ALS is TDP-43 nuclear to cytoplasmic mislocalization3 and consequent nuclear loss of function, which is associated with usage of non-canonical cryptic splice sites, a disease-related mechanism which might ultimately reduce the abundance of correctly spliced mRNAs for translation.3,100 This is illustrated by the aberrant splicing of the stathmin-2 (STMN2) and UNC13A transcripts triggered by TDP-43 depletion in human neurons.101,102 However, the conspicuous absence of TDP-43 pathology in ∼3% of ALS cases, those caused by mutations in SOD1 or FUS, despite such cases being clinically indistinguishable from the majority of ALS cases, led us to reason that there remain further unidentified unifying mechanisms of disease, which present across patients independent of specific mutations. We therefore analysed RNA sequencing data from human induced pluripotent stem cell (iPSC)-derived motor neurons from ALS patients carrying mutations in valosin containing protein (VCP; an example of TDP-43 proteinopathy positive ALS), SOD1 and FUS genes, and discovered aberrant IR as a molecular feature across these diverse genetic causes of ALS,49,86,103,104 subsequently validated in post-mortem tissue from sporadic ALS.105 Given the multifarious functions of IR in human physiology, particularly in the nervous system as alluded to above, it follows that perturbations in this tightly regulated process might modulate or contribute to pathogenesis. At least as early as 1998, IR received attention in the study of ALS pathomechanisms. It was observed in a seminal study that patients with sporadic ALS exhibited loss of the astroglial glutamate transporter excitatory amino acid transporter 2 (EAAT2) in the spinal cord and motor cortex, and that the presence of partial intron 7 retention transcripts was associated with the loss of EAAT2 protein in sporadic ALS patient tissue.106 However, momentum to pursue this avenue may have been tempered by a subsequent study which found that the aberrant transcripts were present in controls as well as in ALS patients, and furthermore were present in all CNS regions, without an altered ratio of ‘variant’ to ‘normal’ transcripts in ALS patients compared with controls.107 Interestingly, this latter study pointed to the likelihood that an exon-9 skipping EAAT2 transcript variant identified by Lin et al. is physiological, speaking to the relevance of splicing in the nervous system, perhaps particularly in the motor system. This controversy subsequently motivated further study of splicing and intron retention in the motor system. One plausible explanation for the observed IR in ALS is that it seems to temporally correlate with a global downregulation of splicing components, thereby ‘weakening’ the splicing machinery.103 TDP-43 and FUS108 as well as SFPQ109 are important in the function of the spliceosome, and their mislocalization and malfunction in ALS could lead to IR, initiating a positive feedback cycle. Through subsequent studies, we found evidence to support a model whereby the RBPs SFPQ and FUS bind with high occupancy to retained introns and are mislocalized from the nucleus to the cytoplasm.103,110 A putative mechanism underlying this mislocalization of RBPs is that IRTs within the nucleus could sequester and subsequently mislocalize bound RBPs through cytoplasmic export of the complex. However, an alternative scenario is that IRTs and RBPs could move to the cytoplasm independently, subsequently interacting to form an RNP, which may increase the residency time of RBPs in the cytoplasm (Fig. 3). Working hypothesis of intron-retention related pathogenic cascade in ALS. In ALS, our working hypothesis is that widespread aberrant intron retention (A) may disrupt the function of nuclear RNA binding proteins (RBPs) (B). Some intron-retaining transcripts are exported to the cytoplasm (C), where they may sequester RBPs (D), promoting abnormal liquid-liquid phase separation and aggregation. Disruption of RBP homeostasis may lead to global RNA misprocessing, including dysregulation of splicing in other transcripts (E). Together, these molecular disturbances may collectively drive a range of organellar pathophysiology including (but not limited to): mitochondrial impairment (F), oxidative stress (G), lysosomal dysfunction (H) and may ultimately lead to motor neuron degeneration (I). ALS = amyotrophic lateral sclerosis. Created in BioRender. Wang, Y. (2026) https://BioRender.com/0353yur. The aberrant partitioning of RNAs and proteins in ALS may also relate to established perturbations in nucleocytoplasmic transport mechanisms, including the role of nuclear pore components. Indeed, ageing is a major risk factor for ALS and at a cellular level it independently induces changes in nucleocytoplasmic transport. Protein complexes comprising the nuclear pore itself have long half-lives and experience no renewal in postmitotic neurons.111 Therefore, an inevitable consequence is that these cells are characterized by increasing mislocalization between the nuclear and cytoplasmic compartments upon ageing.112 Nuclear Localization Signal (NLS)-containing proteins employ the importin receptors for transport into the nucleus, machinery that is also impaired upon cellular ageing.113 Ageing-related defects in nuclear transport might culminate in the mislocalization and cytoplasmic aggregation of proteins, including RBPs.114 It is also established that aged cells accumulate misfolded, aggregated proteins due to compromised protein homeostasis.115 It is therefore plausible that cytoplasmic aggregation of RBPs drives a secondary ‘functional’ nuclear transport defect. Indeed, cytoplasmic aggregation of TDP-43 C-terminal fragments can induce mislocalization of certain nuclear pore constituents to the same vicinity.116 These pathomechanisms may coexist and occur simultaneously or sequentially. Although this link between ALS and ageing is well recognized (and reviewed elsewhere),117,118 the precise mechanistic common denominators remain crucial to resolve in order to potentially inform therapeutic strategy. Potential pathogenicity of aberrant intron retention in ALS We have previously found that hundreds of IRTs exist; >100 IRTs are exported to the cytoplasm49,86 in the context of widespread bidirectional mislocalization of both RNAs and proteins between the nucleus and the cytoplasm.104 Abnormal cIRTs have been found to be a prominent phenomenon in ALS neurons and are demonstrably increased when compared with healthy controls. Moreover, it has been found that cytoplasmic aberrant IRTs have a strong affinity for RBPs.86 These RBPs were reported to form a tightly interconnected network of interacting proteins enriched in mRNA metabolism functions. Significantly, within this network were RBPs known for their involvement in processing capped intron-containing pre-mRNA, implicating disrupted post-transcriptional splicing in ALS pathogenesis. It is noteworthy that this network includes the RBPs SFPQ, TDP-43 and FUS, which are known to exhibit nuclear-to-cytoplasmic mislocalization in ALS.86 However, nucleocytoplasmic mislocalization in ALS goes far beyond TDP-43, SFPQ and FUS, encompassing hundreds of mRNAs and proteins.104 In ALS iPSC-derived motor neurons carrying mutations in TARDBP and VCP, hundreds of transcripts and proteins exhibited significant nucleocytoplasmic mislocalization. In addition, abnormal nIRTs can aggregate into RNA foci that sequester essential RBPs from their physiological functions, further disrupting global RNA metabolism. A salient example is the binding of RBPs to the well-characterized RNA foci observed in Chromosome 9 Open Reading Frame 72 (C9orf72) mutant ALS, the most common genetic cause for ALS and FTD. Here, the GGGGCC repeat expansion-containing intron 1-retaining RNA species form nuclear aggregates,119 which bind and sequester numerous RBPs including Serine/Arginine-Rich Splicing Factor 2 (SRSF2), Heterogeneous nuclear ribonucleoprotein (hnRNP)A1, hnRNPH/F, Aly/REF export factor (ALYREF),120 Adenosine Deaminase RNA Specific B2 (ADARB2)121 and HnRNPH.122 Interestingly, GGGGCC RNA foci also sequester/co-localize with SFPQ and FUS in cellular models as well as C9orf72 mutation patient brain tissue.123 This is likely to disrupt RNA processing and lead to widespread dysregulation in transcription, splicing and RNA transport. Similar mechanisms have been observed in myotonic dystrophy type 2, where CCUG repeats in the Zinc Finger Protein 9 (ZNF9) intron 1 sequester Muscleblind-like (MBNL) RBPs in the muscle cell, primarily within nuclear foci, affecting target gene expression and muscle pathology.124-126 Indeed, analogous mechanisms may operate across other repeat expansion disorders, although this has not yet been systematically established. There are also reports of aberrant intron retention in functionally diverse genes, converging on DNA damage in the context of spinal muscular atrophy (SMA), a type of motor neuron disorder, through R-loop formation.127 R-loops are transient RNA-DNA hybrids that form cotranscriptionally, displacing the non-template DNA strand and generating susceptibility to DNA damage, especially if stabilized. The presence of G-rich non-template strands can stabilize R-loops by folding into G-quadruplexes and exacerbating double-strand break formation.128 Indeed, DNA damage has been reported as a hallmark feature of ALS and correlates with TDP-43 pathology.129 A further mechanism whereby intron retention may contribute to ALS pathology is as a result of translation across retained introns. Retention of introns 3 and 4 in transcripts of peripherin leads to the production of a novel, truncated protein isoform of peripherin, a neuronal intermediate filament protein.130 Peripherin has previously been implicated in ALS, given its presence in ubiquitinated inclusions, and a recent study suggests peripherin as a novel early diagnostic and prognostic biomarker in ALS.131 This truncated isoform is upregulated in ALS patients compared with controls and associates with disease-related inclusions.130 In summary, IR represents one among several convergent axes, alongside cryptic splicing, RBP mislocalization, RBP aggregation, nucleocytoplasmic transport defects, mitochondrial dysfunction and axonal transport deficits, and the relative contributions of these mechanisms likely vary by genotype and disease stage. Aberrant intron retention and perturbed compartmentalization as therapeutic targets Broadly, IRTs may lead to a diverse range of functional outcomes in ALS, including: (i) suppression of cellular homeostatic mechanisms; and (ii) expression of factors that increase cellular vulnerability. The crucial next step for clarity in this context is to follow a candidate-based approach and examine the necessity and sufficiency of individual IRTs in ALS pathogenesis. Once an aberrant intron retention event is found to be pathogenic, one potential approach to selectively target these transcripts is the use of antisense oligonucleotides (ASOs), which are short stretches of synthetic DNA that hybridize with complementary RNA through Watson–Crick base pairing. There exist an ever-increasing number of potential chemical modifications to the oligonucleotide exist that allow for efficient target binding with predictable and potent functional outcomes including influencing splicing decisions through blocking RBP binding by steric hindrance on otherwise intact mRNAs. These fully modified ASOs, such as ‘MIXmer’ ASOs, are different from their ‘GAPmer’ ASO counterparts in that they do not activate endogenous RNase H and therefore do not indiscriminately lead to transcript degradation, but can exert a more nuanced intervention, for example influencing a specific splicing decision. The ability to predictably manipulate the life cycle of targeted RNAs in this manner has already proven transformational in clinical trials for a US Food and Drug Administration (FDA)-approved ASO, nusinersen, in patients with SMA,132 which has now entered best practice for the management of SMA. Extrapolating these insights to ALS, where disease-modifying therapeutic options remain extremely limited, could have similar ground-breaking clinical impact. Indeed, a GAPmer approach for targeting SOD1 ALS has already yielded encouraging results in clinical trials in the open label extension phase.133 This therapy is FDA and MHRA (Medicines and Healthcare products Regulatory Agency) approved and argues for cautious optimism in adopting antisense approaches for the remaining majority of non-SOD1 ALS patients. It must be noted, however, that while SOD1 and FUS mutations have clear evidence for gain of toxic function amenable to GAPmer approaches, the precise underlying mechanisms are less clear in the remaining 97% of ALS cases, further reinforcing the need for more fundamental discovery. Preventing intron retention using splice-switching ASOs may offer a tractable therapeutic approach as it may then prevent sequestration of their bound RBPs. Translational strategy considerations when targeting ALS-related IRTs include triaging therapeutically tractable events by their significance, abundance, intron length and affinity for binding relevant RBPs. Selectively targeting the IRT isoform should leave the correctly spliced transcripts intact and thereby preserve canonical function. Depending on the approach adopted, some consideration of polytherapy is merited. Perhaps targeting several aberrant transcripts simultaneously may be required to achieve the threshold for cellular rescue. However, recent setbacks in the field underscore the need for cautious framing in this context. For example, the disappointing clinical outcomes of C9orf72 ASOs (NCT03626012, NCT04931862), which primarily targeted sense repeat transcripts, highlight the complexity of RNA-based pathogenesis in ALS and raise the possibility that effective intervention may require simultaneous targeting of multiple RNA species, including both sense and antisense transcripts, or multiple downstream consequences of RNA dysregulation. Compared with TDP-43 loss-focused strategies targeting single cryptic exon events, such as those in STMN2 or UNC13A, IR-targeting ASOs may present an underexplored opportunity across disease contexts both with and without TDP-43 proteinopathy. The fact that aberrant IR occurs in ALS prompts consideration of the downstream processing of intronic sequences. Inhibiting the debranching enzyme Dbr, which usually debranches intronic lariats produced during splicing, relieves toxicity secondary to cytoplasmic TDP-43 (or FUS) accumulation. It is likely that the mechanism behind this is sequestration of TDP-43 by the intronic lariats, preventing deleterious interactions with important RNAs and RBPs.134 However, the impact of lariats (protective or deleterious) is likely to be context dependent with determinants including cell type, cell substate and disease stage. It is possible that a fine balance exists between detrimental ‘trapping’ of RBPs on retained intronic sequences and a compensatory removal of RBPs that are mislocalized and present in excess. This balance could be deregulated in disease leading to a loss of function of the RBPs bound to the retained intronic sequences. It is possible that the compensatory effects in the acute phase evolve into more maladaptive mechanisms if these are sustained over time and/or occur in the disease context. Perturbed cellular compartmentalization of RNA and proteins, as reported in ALS, raises the important issue of enhancing nucleocytoplasmic transport integrity as a therapeutic strategy. Indeed, there are a number of studies overexpressing nuclear importins, exportins or transport partners and demonstrating at least partial aggregation reversal of FUS through interaction with their NLSs.135-137 It is noteworthy that other studies demonstrate that importins such as Kapβ2/Transportin-1 and Importin-α/β1 can suppress or reverse aberrant phase transitions of additional ALS-linked RBPs, including TATA-Box Binding Protein Associated Factor 15 (TAF15), Ewing Sarcoma breakpoint region 1 (EWSR1), hnRNPA1, hnRNPA2 and TDP-43.138,139 Inhibitors of the nuclear export receptor CRM1 (SINE compounds) are also evidently effective in rescuing TDP43-mediated motor phenotypes in different models.140,141 Furthermore, overexpression of nucleoporin POM121 was sufficient to rescue dysfunction of TDP-43 targets.142 Relocalizing RBPs to the nucleus as a therapeutic strategy might also be realized through small molecular pharmacology. For example, arginine methyltransferase inhibitors in the context of FUS mutations mitigate the mislocalization and aggregation of FUS.143 Finally, we have reported that treating human ALS motor neurons with VCP D2 ATPase inhibitors largely corrects the observed RNA and protein mislocalization and additionally rescues cellular phenotypes. Importantly, this relocalization of RNAs and proteins also rescues splicing, including partially reversing aberrant intron retention.104,144 It is noteworthy that existing referenced data concerning VCP D2 ATPase inhibition derive from iPSC-derived motor neurons and short-term in vivo models, and that broader translational generalizability has not yet been established. It is important to recognize that these aforementioned approaches remain at preclinical or at early proof-of-concept stages, with most evidence derived from cellular models and animal systems. Long-term safety, specificity, and efficacy in humans remain unknown and require rigorous evaluation. Enhancing discovery of non-canonical RNA functions using artificial intelligence In recent years, biological research has experienced a paradigm shift, evolving from a traditional hypothesis-driven approach to a sophisticated data-driven methodology. The large volume of sequencing data generated in the last 20 years has motivated the development of a variety of computational techniques to study and characterize the functions and regulations of these RNA sequences.145-148 Initially, the data-driven strategy was heavily dependent on manual feature engineering, relying on established knowledge such as gene models or predefined features for tasks like motif enrichment. However, such methods struggle to capture complex, non-linear patterns, emphasize local over global sequence-context relationships and are constrained by prior knowledge, limiting their adaptability to novel or underrepresented biological contexts, such as aberrant intron retention. The emergence of deep learning has marked a significant leap forward, offering a powerful tool for deciphering signal from otherwise complex and noisy biological datasets, and facilitating the unbiased, automated extraction of deep features such as deep regulatory sequences (intronic splicing silencers or intronic splicing enhancers) within introns. Deep learning algorithms have demonstrated the ability to reduce the dimensionality of RNA features and extract more discriminative patterns for predicting RNA-associated interactions,149-152 uncovering patterns, relationships and the sophisticated statistical properties of RNA molecules. RNA Large Language Models (LLMs) excel at capturing long-range dependencies and achieving high-level abstraction, making them invaluable for studying RNA function. They have been shown to effectively identify RNA-encoded features and improve predictions of RNA interactions.149-152 A key challenge in studying the regulation of non-coding mRNA regions such as retained introns lies in identifying the full repertoire of interacting RBPs, miRNAs and their combined effects on cellular processes in a context-specific manner. In fact, given the context-specific roles of these molecules, it can be posited that the potential functions of RNA molecules are virtually limitless. However, modelling RNA interactions across diverse biological contexts remains difficult for existing algorithms.153 Future computational models should generate context-aware RNA representations, allowing for more precise functional predictions. A recent study exemplified this approach by generating contextualized protein interaction networks, producing 394 760 protein representations across 156 cell types from 24 tissues.153 Similar advances in RNA modelling could greatly enhance our ability to decipher the regulatory landscape of IRTs (and beyond). Thus, deep learning models applied to RNA biology are poised to revolutionize research, opening a new era of discovery in the field. However, additional work is needed to fully explore the potential of these models in capturing the intricacies of the ‘RNA code’.154 Also noteworthy in this context are the limitations of current predictive models, which are typically constrained by training data derived largely from bulk RNA-seq (especially those generated through rRNA-depletion protocols that limit precise identification of intronic reads between mature mRNA and nascent pre-mRNA), a limited array of cell types and incomplete compartment-specific resolution. AI-based predictions of IRTs or regulatory interactions should consequently be interpreted as hypothesis generating until experimentally validated. Concluding remarks and future outlook Intron retention is becoming increasingly recognized as a mechanism to regulate gene expression, and further investigation will be crucial in revealing the extent of the importance of IR in determining RBP function and RBP localization. With interest in studying this form of AS growing, it is important to identify best practice for future studies. Some considerations for RNA sequencing studies such as read depth, library preparation, strandedness, analytical tools and approaches are further considered in Box 4. Beyond such technical considerations to safeguard accuracy of detection, there are further conceptual challenges to overcome. For example, we need to more comprehensively identify the regulators of intron retention beyond cis-features and RBPs to include the influence of epigenetics, epitranscriptomics and transcriptional rate. In order to study the full diversity of IRTs, we need effective tools to block their degradation. Cytoplasmic RNA degradation mechanisms such as NMD can be suppressed genetically [e.g. Up-frameshift 1 (UPF1) knock down] or using small molecules [e.g. SMG1 nonsense-mediated mRNA decay associated PI3K related kinase (SMG1) inhibitors]. The nuclear RNA degradation machinery is less well characterized and requires further study in different cellular and environmental contexts, particularly those of acute versus chronic stress. Technical considerations when studying intron retention The following points are recommendations to consider when studying intron retention using RNA sequencing. (i) Deep sequencing (ideally 100 M reads; at least 50 M reads). A lower sequencing depth will underestimate intron retention.155 (ii) PolyA selection to avoid pre-mRNA contamination. (iii) Strand-specific sequencing to preserve information regarding transcript direction and therefore understanding of overlapping genes and antisense transcripts. (iv) Customized mapping tools: multi-mapping reads and the repetitive sequences within introns would normally lead to these reads being discarded. (v) Ensuring the detection of reads spanning intron-exon junctions to mitigate the risk of erroneously identifying intron-derived non-coding RNA as intron-retaining transcripts (IRTs). (vi) In cases where multiple events within a single transcript are being studied, long read sequencing is advised,156 noting the limitations (relatively low depth, high error rates and prone to detecting shorter transcripts). (vii) Extracting IRTs is possible, particularly if they have high phase separation propensity. In this context, heating/shearing steps are necessary for their efficient isolation.62,157 Future studies will leverage findings from other fields. For example, studies of intron retention in oncology have found that IRTs exhibit diverse effects, encompassing inhibition of tumour suppressor genes, oncoprotein expression and neoantigen formation. IRTs may also have potential biomarker value and roles in cancer therapy effectiveness and stratification.158 Importantly, targeting subsets of functional coherent IRTs seems a tractable therapeutic option across a range of diseases.158,159 Many open questions remain regarding the link between intron retention and disease pathogenesis and are outlined in Box 5. In summary, the roles of intron retention in physiology and disease are coming into focus. The confluence of this burgeoning area of molecular biology with an increasing focus on RNA therapeutics, integrated with artificial intelligence approaches, represents an unprecedented opportunity to finally understand and tackle some of the most complex and devastating neurodegenerative disorders such as ALS. Major open questions in intron retention biology What are the major roles of stable cytoplasmic intron retaining transcripts? Might these function as an extra-nuclear reservoir of RNA binding proteins? What are the major roles of stable cytoplasmic intron retaining transcripts? Might these function as an extra-nuclear reservoir of RNA binding proteins? How might we identify and understand deep regulatory sequences within retained introns? Could artificial intelligence help to predict these sequences? How might we identify and understand deep regulatory sequences within retained introns? Could artificial intelligence help to predict these sequences? What is the extent of context-specific translation of stable intronic sequences? What is the extent of context-specific translation of stable intronic sequences? What is the contribution of stable intronic sequences in liquid-liquid phase separation (LLPS) dynamics and vice versa? What is the contribution of stable intronic sequences in liquid-liquid phase separation (LLPS) dynamics and vice versa? Is aberrant intron retention in disease harmful, compensatory or an interplay of both? Might this vary by disease stage? If aberrant, is targeting one or a small subset of events sufficient to ameliorate disease? Is aberrant intron retention in disease harmful, compensatory or an interplay of both? Might this vary by disease stage? If aberrant, is targeting one or a small subset of events sufficient to ameliorate disease? Might aberrant intron retaining transcripts in disease serve as diagnostic or prognostic biomarkers? Might aberrant intron retaining transcripts in disease serve as diagnostic or prognostic biomarkers? Supplementary Material