Main

TDP-43, encoded by the TARDBP gene, is an abundant, ubiquitously expressed RNA-binding protein that normally localizes to the nucleus. It has a role in fundamental RNA-processing activities, including RNA transcription, alternative splicing and RNA transport7. A major splicing regulatory function of TDP-43 is to repress the inclusion of cryptic exons during splicing2,8,9,10. Unlike normal conserved exons, these cryptic exons occur in introns and are normally excluded from mature mRNAs. When TDP-43 is depleted from cells, these cryptic exons get spliced into messenger RNAs, often introducing frame shifts and premature termination, or even reduced RNA stability. However, the key cryptic splicing events that are integral to disease pathogenesis remain unknown.

STMN2—which encodes stathmin 2, a regulator of microtubule stability—is the gene whose expression is most significantly reduced when TDP-43 is depleted from neurons3,4. STMN2 harbours a cryptic exon (exon 2a) that is normally excluded from the mature STMN2 mRNA. The first intron of STMN2 contains a TDP-43 binding site. When TDP-43 is lost or its function is impaired, exon2a gets incorporated into the mature mRNA. Exon 2a harbours a stop codon and a polyadenylation signal—this results in truncated STMN2 mRNA and eightfold reduction3 of stathmin 2. Aberrant splicing and reduced stathmin 2 levels seem to be a major feature of sporadic and familial cases of ALS (except those with SOD1 mutations)3,4 and in frontotemporal lobar degeneration (FTLD) due to TDP-43 proteinopathy11 (FTLD-TDP). The discovery of STMN2 cryptic exon splicing in ALS and FTLD-TDP highlights a key mRNA target—we aimed to identify other possible mRNA targets.

To discover cryptic splicing targets regulated by TDP-43 that may also have a role in disease pathogenesis, we used a recently generated RNA sequencing (RNA-seq) dataset12. To generate this dataset, fluorescence-activated cell sorting (FACS) was used to enrich neuronal nuclei with and without TDP-43 from postmortem brain tissue from patients with FTD and ALS (FTD–ALS); RNA-seq was performed to compare the transcriptomic profiles of TDP-43-positive and TDP-43-negative neuronal nuclei. This identified several differentially expressed genes12. We re-analysed the data to identify novel alternative splicing events affected by the loss of nuclear TDP-43. We performed splicing analyses using two pipelines, MAJIQ13 and LeafCutter14, designed to detect novel splicing events (Fig. 1a). We identified 266 alternative splicing events (PΨ > 0.1) > 0.95; where ΔΨ signifies changes of local splicing variations between two conditions) with MAJIQ and 152 with LeafCutter (P < 0.05). There were 66 alternatively spliced genes in common between the two analyses (Fig. 1b), probably because each tool uses different definitions for transcript variations and different criteria to control for false positives (Supplementary Note 1). These genes have at least one region that both tools identified to be alternatively spliced (Supplementary Table 1). Among the alternatively spliced genes identified by both tools were STMN2 and POLDIP3, both of which have been extensively validated as bona fide targets of splicing by TDP-433,4,11,15.

Fig. 1: Nuclear depletion of TDP-43 causes CE inclusion in UNC13A RNA and reduced expression of UNC13A protein.
figure 1

a, Splicing analyses were performed on RNA-sequencing results from TDP-43-positive and TDP-43-negative neuronal nuclei isolated from frontal cortices of seven patients with FTD or FTD–ALS. Some illustrations were created with BioRender.com. b, Sixty-six alternatively spliced genes identified by both MAJIQ (PΨ > 0.1) > 0.95) and LeafCutter (P < 0.05). Genes in blue are previously validated TDP-43 splicing targets3,4,11,15. c, f, RT–qPCR confirmed inclusion of CE in UNC13A mRNA upon TDP-43 depletion in SH-SY5Y cells (n = 5 cell culture experiments for each condition; two sided-Welch two-sample t-test; mean ± s.e.m.) (c) and in 3 independent lines of iPSC-MNs (n = 2 independent cell culture experiments, each with 2 technical replicates for each iPSC-MN) (f). RPLP0 was used to normalize RT–qPCR. Three-way ANOVA; mean ± s.e.m. TDP-43 is also known as TARDBP. d, e, g, h, Immunoblotting for UNC13A and TDP-43 protein levels in SH-SY5Y cells (d; quantified in e) and iPSC-MNs (g; quantified in h) treated with scramble (shScramble) or TDP-43 shRNA (n = 3 independent cell culture experiments for each condition). GAPDH served as a loading control. Two-sided Welch two-sample t-test, mean ± s.e.m. Gel source data are shown in Supplementary Fig. 1a, b.

Source data

UNC13A was one of the genes with the most significant levels of alternative splicing (MAJIQ ΔΨ = 0.779; LeafCutter ΔΨ = 0.8360; P < 0.0001) in neurons with nuclear TDP-43 depletion (Fig. 1b). Depletion of TDP-43 resulted in the inclusion of a 128-bp or a 178-bp cryptic exon between the canonical exons 20 and 21 (hg38; chr19: 17642414–17642541 (128bp); chr19: 17642591–17642414 (178 bp)) (Extended Data Fig. 1a–e, Supplementary Note 2). The two cryptic exons share the same 3′ end but the 178-bp cryptic exon is 50 bp longer than the 128-bp cryptic exon at the 5′ end. The cryptic exons were almost completely absent in wild-type neuronal nuclei (Fig. 1f) and are not present in any of the known human isoforms of UNC13A16. Furthermore, analysis of ultraviolet cross-linking and immunoprecipitation (iCLIP) data for TDP-4317 provides evidence that TDP-43 binds directly to the intron harbouring these cryptic exons (shown by mapped reads) (Extended Data Fig. 1g). Because of the much higher abundance of the 128-bp cryptic exon in the RNA-seq data that we analysed, we focused our analyses on this cryptic exon, which we refer to as CE. Intron 20–21 of UNC13A and the CE sequence are conserved among most primates (Extended Data Fig. 2a, b) but not in mouse (Extended Data Fig. 2c, d), similar to STMN2 and other cryptic splicing targets of TDP-432,3,4. Together, these results suggest that TDP-43 functions to repress the inclusion of a cryptic exon in the UNC13A mRNA.

To determine whether TDP-43 directly regulates this UNC13A cryptic splicing event, we used short hairpin RNA (shRNA) to reduce TDP-43 levels in SH-SY5Y cells and quantitative PCR with reverse transcription (RT–qPCR) to detect CE inclusion in UNC13A transcript. CE was present in cells with TDP-43 depletion but not in cells treated with control shRNA (Fig. 1e). Along with the increase in CE, there was a corresponding decrease in the amount of the canonical UNC13A transcript upon TDP-43 depletion (Fig. 1c). By immunoblotting, we also observed a marked reduction in UNC13A protein in TDP-43-depleted cells (Fig. 1d, e). Reducing levels of TDP-43 in motor neurons derived from induced pluripotent stem cells (iPSC-MNs) (Fig. 1f–h, Extended Data Fig. 3a, b, Supplementary Table 3) and excitatory neurons (i3Ns) derived from human induced pluripotent stem (iPS) cells (Extended Data Fig. 3c) also resulted in CE inclusion and a reduction in UNC13A mRNA and protein. We confirmed insertion of the cryptic exon sequences into the mature transcript by amplicon sequencing of the product of reverse transcription with PCR (RT–PCR) (Extended Data Fig. 3d; see Supplementary Note 2) and demonstrated that the insertions introduce premature stop codons (Extended Data Fig. 3f–i), consistent with the observed decrease in UNC13A protein. UNC13A gene expression is probably regulated at multiple levels beyond simply inclusion of the cryptic exons. Other aspects of the cryptic exon-inclusion event (for example, aberrant peptides produced from it) could cause defects, although we do not yet have evidence that such peptides are produced. Thus, lowering levels of TDP-43 in human cells and neurons causes inclusion of CE in the UNC13A transcript, resulting in decreased UNC13A protein.

UNC13A belongs to a family of genes originally discovered in Caenorhabditis elegans and was named on the basis of the uncoordinated (unc) movements exhibited by animals with mutations in these genes18, owing to deficits in neurotransmitter release. UNC13A encodes a large multidomain protein that is expressed in the nervous system, where it localizes to most synapses in the central nervous system and neuromuscular junctions, and has an essential role in the vesicle priming step, prior to synaptic vesicle fusion19,20,21,22. UNC13A is an essential neuronal protein because mice lacking Unc13a (also called Munc13-1) exhibit functional deficits at glutamatergic synapses, demonstrated by a lack of fusion-competent synaptic vesicles, and die within a few hours of birth21. Our data suggest that depletion of TDP-43 leads to loss of this critical synaptic protein. As well as UNC13A, several other genes encoding synaptic proteins are mis-spliced upon TDP-43 depletion (Fig. 1b). We validated the splicing events in three of these genes (KALRN, RAPGEF6 and SYT7) in induced pluripotent stem (iPS) cell-derived neurons (iNs) using RT–qPCR (Extended Data Fig. 4), providing evidence that disruption of synaptic function could be a major mechanism in the pathogenesis of ALS and FTD.

To extend our analysis of UNC13A cryptic exon inclusion to a larger collection of patient samples, we first analysed a series of 117 frontal cortex brain samples from the Mayo Clinic Brain Bank using RT–qPCR and a pair of primers that detects the shared 3′ end of the cryptic exons. We found a significant increase in UNC13A cryptic exon in the frontal cortices of patients with FTLD-TDP compared with healthy controls (Fig. 2a, Extended Data Fig. 5a). Next, we analysed brain samples from the New York Genome Center (NYGC). We interrogated RNA-seq data from 1,151 tissue samples from 413 individuals (with multiple tissues per individual; see Supplementary Note 3), 330 of whom are patients with ALS or FTD. We detected UNC13A splice variants in nearly 50% of the frontal and temporal cortical tissues donated by patients with neuropathologically confirmed FTLD-TDP (Fig. 2b) and in some of the patients with ALS whose pathology has not been confirmed (Extended Data Fig. 5b). Notably, we did not observe UNC13A splice variants in any of the samples from patients with FTLD associated with FUS (FTLD-FUS) (n = 9) or TAU (FTLD-TAU) (n = 18) or ALS associated with SOD1 (ALS-SOD1) (n = 22), nor in any of the control samples (n = 197) (Fig. 2b). Using the same criteria, we detected the known STMN2 cryptic exon3,4 in tissues from these patients, and the majority of the UNC13A splice variant containing tissues also contained the STMN2 splice variant (Supplementary Table 2). Hyperphosphorylated TDP-43 (pTDP-43) is a key feature of the pathology of these diseases1. We found a strong association between higher levels of pTDP-43 and higher levels of UNC13A CE inclusion in patients with FTLD-TDP (Spearman’s ρ = 0.610, P < 0.0001) (Extended Data Fig. 2c). Thus, UNC13A CE inclusion is a robust and specific facet of pathobiology in TDP-43 proteinopathies.

Fig. 2: UNC13A CE inclusion in human TDP-43 proteinopathies.
figure 2

a, UNC13A CE expression level is increased in the frontal cortices of patients with FTLD-TDP. GAPDH and RPLP0 were used to normalize the RT–qPCR (two-tailed Mann–Whitney test, mean ± 95% confidence interval). The schematic to the right shows the localization of the primer pair (arrows) used for the RT–qPCR assay. Healthy: n = 27; sporadic FTLD-TDP: n = 34; C9ORF72+ FTLD-TDP: n = 47; GRN+ FTLD-TDP: n = 9. b, UNC13A CE is detected in nearly 50% of frontal cortical tissues and temporal cortical tissues from neuropathologically confirmed FTLD-TDP patients in bulk RNA-sequencing from the NYGC ALS Consortium cohort. CE is absent in tissues from healthy controls and patients with FTLD-FUS, FTLD-TAU or ALS-SOD1.

Source data

To visualize the UNC13A CE within single cells in the human brain, we designed custom BaseScope in situ hybridization probes that specifically bind to the exon 20–CE junction (Fig. 3a) or the exon 20–exon 21 junction (Extended Data Fig. 6c). We used these probes for in situ hybridization, combined with immunofluorescence for NeuN (to detect neuronal nuclei) and TDP-43. We stained sections from the medial frontal pole of four patients with FTLD-TDP and three controls (Supplementary Table 4). In neurons showing loss of nuclear TDP-43 and accompanying cytoplasmic TDP-43 inclusions, we observed UNC13A CE-containing mRNA splice variants in the nucleus (Fig. 3b, Extended Data Fig. 6a). We observed between one and four CE-containing mRNA puncta per nucleus. We did not observe puncta in the cytoplasm, perhaps because CE introduces a premature stop codon, which could lead to nonsense-mediated decay. Controls, however, had universally normal nuclear TDP-43 staining and showed no evidence of UNC13A cryptic splicing (Fig. 3b, Extended Data Fig. 6b). Next, we sought to determine whether TDP-43 nuclear depletion is associated with reduced expression of canonical UNC13A mRNA. In control brain tissue, UNC13A mRNA was widely expressed in neurons across cortical layers (Extended Data Fig. 6d, e). In patients, we saw a trend for reduced UNC13A mRNA in neurons showing TDP-43 pathology compared with neighbouring neurons with normal nuclear TDP-43 (Extended Data Fig. 6d, e), consistent with the RT–qPCR data (Fig. 1c, f). These findings suggest that cryptic splicing of UNC13A is absent from controls and, in patients, is seen exclusively in neurons showing depleted nuclear TDP-43.

Fig. 3: UNC13A cryptic splicing is associated with loss of nuclear TDP-43 in patients with FTD and motor neuron disease.
figure 3

a, The design of the UNC13A e20/CE BaseScope probe targeting the alternatively spliced UNC13A transcript. Each Z binds to the transcript independently, and both must be in close proximity for successful signal amplification, ensuring binding specificity. b, BaseScope in situ hybridization using the UNC13A e20/CE probe, combined with immunofluorescence for TDP-43 and NeuN, was performed on sections from the medial frontal pole of patients with FTD and motor neuron disease (FTD–MND) and healthy controls. Representative images illustrate the presence of UNC13A CE (arrowheads) in neurons showing depletion of nuclear TDP-43. Neurons with normal nuclear TDP-43 in patients and controls show no CE signal (arrows). Images are maximum intensity projections of a confocal image z-stack. Scale bar, 10 µm. Images representative of six non-overlapping images from each individual. We optimized UNC13A probes on two cases and two controls in three separate experiments, with similar findings.

UNC13A is one of the top hits for ALS and FTD–ALS in multiple genome-wide association studies5,6,23,24,25,26 (GWAS). Single nucleotide polymorphisms (SNPs) in UNC13A are associated with increased risk of sporadic ALS5 and sporadic FTLD-TDP pathology, especially type B, the subtype associated with FTD–ALS6. In addition to increasing susceptibility to ALS, SNPs in UNC13A are associated with shorter survival in patients with ALS27,28,29,30. But the mechanism by which genetic variation in UNC13A increases risk for ALS and FTD is unknown. Notably, the two most significantly associated SNPs, rs12608932 (A>C) and rs12973192 (C>G), are both located in the same intron that we found harbours CE, with rs12973192 located in CE itself (Fig. 4a, c). This immediately suggested that these SNPs (or other nearby genetic variations tagged by these SNPs) might make UNC13A more vulnerable to CE inclusion upon TDP-43 depletion. To test this, we analysed the percentage of RNA-seq reads that mapped to intron 20–21 that support the inclusion of CE (Extended Data Fig. 7a, b). Among the seven patients included in the initial splicing analysis (Fig. 1a), two out of three who were homozygous (G/G) and the one patient who was heterozygous (C/G) for the risk allele at rs12973192, showed inclusion of CE in almost every UNC13A mRNA that was mapped to intron 20–21. By contrast, the three other patients who were homozygous for the reference allele (C/C) showed much less inclusion of CE (Extended Data Fig. 7a, b). Another way to directly assess the effect of the UNC13A risk alleles on CE inclusion is to measure allele imbalance in RNAs from individuals who are heterozygous for the risk allele. Two of the iPSC-MN lines that we used to detect CE inclusion upon TDP-43 knockdown (iPSC-MN1 and iPSC-MN3; Fig. 1h) are heterozygous (C/G) at rs12973192. We performed amplicon sequencing of the RT–PCR product that spans CE and analysed the allele distribution from these two samples (Extended Data Fig. 3d, e) as well as the one patient sample from the original RNA-seq dataset (Fig. 1a) that is heterozygous (C/G) at rs12973192 (Extended Data Fig. 7b). We found a significant difference between the percentage of risk and reference alleles in the spliced variant, with higher inclusion of the risk allele (Fig. 4b, Extended Data Fig. 7c).

Fig. 4: UNC13A risk haplotype associated with ALS or FTD susceptibility potentiates CE inclusion when TDP-43 is dysfunctional.
figure 4

a, LocusZoom plot showing SNPs associated with ALS or FTD in UNC13A. SNPs are coloured on the basis of levels of linkage equilibrium; SNPs rs12608932 and rs12973192 are in strong linkage disequilibrium (LD). b, There is a higher percentage inclusion of the risk allele (G) at rs12973192 in the UNC13A splice variant (n = 3 biologically independent samples; two-sided paired t-test; mean ± s.e.m.). Quantification in Extended Data Fig. 7c. c, Location of rs56041637 relative to the two known FTD–ALS GWAS hits and UNC13A CE. d, Design of UNC13A CE minigene reporter constructs and location of the primer pair used for RT–PCR. Black (reference alleles) and blue (risk alleles) triangles represent the genetic variants as shown in c. e, Splicing of minigene reporters was assessed in wild-type (WT) and TDP-43−/− HEK 293T cells. In addition to the inclusion of CE (2), some splice variants showed inclusion of one of the other two cryptic splicing products (3 and 4) (Extended Data Figs. 1a–e, 3a, Supplementary Note 2). The risk haplotype-carrying minigene showed an almost complete loss of canonical splicing product (1) and an increase in alternatively spliced products (2, 3 and 4). n = 2 independent cell culture experiments for each condition. f, Top, survival curves of FTLD-TDP patients stratified on the basis of the number of risk haplotypes. Heterozygous (1) and homozygous (2) patients had shorter survival time after disease onset (n = 205, Mayo Clinic Brain Bank; score (log-rank) test, P = 0.004). Dashed lines mark median survival for each genotype. Risk haplotype effect is modelled additively using Cox multivariable analysis adjusted for genetic mutations, sex and age at onset. Bottom, risk table. Summary results of the analysis are shown in Extended Data Fig. 10b.

Source data

Given this evidence for an effect of the risk allele on CE inclusion, we extended our analysis by genotyping patients with FTLD-TDP harbouring CE (n = 86) in the Mayo Clinic Brain Bank dataset for the UNC13A risk alleles at rs12973192 and rs12608932. Because these two SNPs are in high linkage disequilibrium in the European population31, we consider them to be on the same haplotype. Thus, we refer to the haplotype that contains reference alleles at both SNPs as the reference haplotype, and the haplotype that contains risk alleles as the risk haplotype. We excluded the one patient who is homozygous for the reference allele (C/C) at rs12973192 but heterozygous (A/C) at rs12608932. The remaining patients (n = 85) have exactly the same number of risk alleles at both loci, indicating that it’s very likely that they are carriers of the reference haplotype or the risk haplotype. Using a multiple linear regression model, we found a strong positive correlation (β = 0.175, P = 0.0290) between the number of risk haplotypes and the abundance of UNC13A CE inclusion measured by RT–qPCR (Extended Data Fig. 7e). Together, these data suggest that genetic variation in UNC13A that increases risk for ALS and FTD in humans promotes CE inclusion upon nuclear depletion of TDP-43.

GWAS SNPs typically do not cause the trait but rather ‘tag’ other neighbouring genetic variation32. Thus, a major challenge in human genetics is to go from a GWAS hit to identifying the causative genetic variation that increases risk for disease33. A LocusZoom34 plot (Fig. 4a) generated using results from an ALS GWAS35 suggests that the strongest association signal on UNC13A is indeed in the region surrounding the two lead SNPs (rs12973192 and rs12608932). To identify other genetic variants in intron 20–21 that might also cause risk for disease by influencing CE inclusion but were not included in the original GWAS, we analysed genetic variants identified in whole genome sequencing data of 297 ALS patients of European descent (Answer ALS; https://www.answerals.org). We searched for novel genetic variants that could be tagged by the two SNPs by looking for other loci in intron 20–21 that are in linkage disequilibrium with both rs12608932 and rs12973192. We found one that fit these criteria: rs56041637 (Extended Data Fig. 7d). rs56041637 is a CATC-repeat insertion in intron 20–21. Most frequently there are six CATC repeats in this region. In the patient dataset, we observed that patients who are homozygous for the risk alleles at both rs12608932 and rs12973192 tend to have 3 to 5 additional CATC repeats; patients who are homozygous for reference alleles at both rs12608932 and rs12973192 tend to have only 0 to 2 additional repeats (Fig. 4c; CATC is shown as GATG because UNC13A is on the reverse strand). Thus, in addition to the two lead GWAS SNPs (rs12608932 and rs12973192), we now nominate rs56041637 as potentially contributing to risk for disease by making UNC13A more vulnerable to CE inclusion when TDP-43 is depleted from the nucleus.

To directly test whether these three variants in UNC13A—which are part of the risk haplotype—increase CE inclusion upon TDP-43 depletion, we synthesized minigene reporter constructs (Fig. 4d). The reporter uses a bidirectional promoter to co-express eGFP containing a canonical intron and mCherry that is interrupted by UNC13A intron 20–21 from either the reference haplotype or the risk haplotype. Because UNC13A is on the reverse strand, the reference alleles and the risk alleles are the reverse complement of the genotypes reported on dbSNP—for example, in intron 20–21 of UNC13A, the reference allele at rs12973192 is G and the corresponding risk allele is C. We transfected wild-type and TDP-43-knock-out (TDP43−/−) HEK 293T cells36 with each minigene reporter construct. Using RT–PCR, we found that both versions of intron 20–21 were efficiently spliced out in wild-type cells (Fig. 4e, lanes 1–4). However, in TDP43−/− cells there was a decrease in completely intron-free splicing products and a concomitant increase in cryptic splicing products (Fig. 4e, lanes 5–6). Of note, in TDP-43–/– cells transfected with the minigene construct harbouring the risk haplotype in the intron, there was an even greater decrease in complete intron 20–21 splicing, and a proportional increase in cryptic splicing products (Fig. 4e, lanes 7–8; see Supplementary Note 2). The transcript levels of the eGFP control remained constant across all conditions, verifying equal reporter expression levels and the integrity of the splicing machinery independent of TDP-43.

To define the effect of each individual risk variant on splicing, we generated six additional minigene reporters, each carrying a different combination of the reference and risk alleles (individually, two at a time, or all three). Using RT–qPCR, we found that the risk variant in CE (rs12973192) had the strongest effect on CE inclusion in cells lacking TDP-43. The other variants also contributed to mis-splicing, but in a non-additive way, and the largest effect was with the construct harbouring all three risk variants (Extended Data Fig. 8). Expression of full-length TDP-43 rescued the splicing defects, whereas an RNA binding-deficient mutant did not (Extended Data Fig. 9). Together, these results provide direct functional evidence that TDP-43 regulates splicing of UNC13A intron 20–21 and that genetic variants associated with ALS and FTD susceptibility in humans potentiate cryptic exon inclusion when TDP-43 is dysfunctional.

To directly test whether the risk variants affected TDP-43 binding, we performed quantitative electrophoretic mobility shift assays (EMSA). We incubated radioactively labelled RNA probes containing reference or risk versions of sequences within the UNC13A CE or intronic region with recombinant full-length TDP-43. This resulted in a mobility shift, indicating that TDP-43 can bind to these sequences (Extended Data Fig. 10a). TDP-43 binds the reference and risk versions of the CE probe with similar affinity but the intronic risk allele results in a minor reduction in affinity (Extended Data Fig. 10a). TDP-43 had lower affinity for the probe containing the risk allele at rs56041637 compared with the probe containing the reference allele (Fig. 4g). Overall, TDP-43 has a much higher affinity for the intron sequence compared with the exon or repeat sequence. The diminished binding affinity of TDP-43 to risk alleles of the intron and repeat sequence may contribute to the increased cryptic splicing found in ALS and FTD. We note that these in vitro binding results are somewhat different from those reported in the accompanying Article by Brown and colleagues37. Our use of full-length TDP-43, which can be prone to aggregation (although our maltose-binding protein (MBP)-tagged TDP-43 is soluble) but contains additional domains that are important for TDP-43 function, differs from the one used by Brown and colleagues, which contains only the TDP-43 RNA-recognition motifs (RRMs). Future studies will be required to explore how TDP-43 regulates the cryptic splicing of UNC13A and other splicing targets and the effect of different genetic variations on TDP-43 binding in vivo.

To examine whether these SNPs affect survival in patients with FTLD-TDP, we evaluated the association of the risk haplotype with survival time after disease onset using data from the Mayo Clinic Brain Bank (n = 205). Using Cox multivariable analysis adjusting for other factors known to influence survival (genetic mutations, sex and age at onset), the risk haplotype was associated with survival time under an additive model (Fig. 4f). The number of risk haplotypes an individual carries was a strong prognostic factor (Extended Data Fig. 10b). The association remained significant under a dominant model (Extended Data Fig. 10c, d) and a recessive model (Extended Data Fig. 10e, f), indicating that carrying the risk haplotype reduces patient survival time after disease onset, consistent with previous analyses27,28,29,30. Thus, as suggested by previous studies28,30,38, genetic variants in UNC13A that increase cryptic exon inclusion are associated with decreased survival in patients.

Here we have found that TDP-43 regulates a cryptic splicing event in the FTD–ALS risk gene UNC13A. The most significant genetic variants associated with disease risk are located within the intron harbouring the cryptic exon itself. Brain samples from patients with FTLD-TDP carrying these SNPs exhibited more UNC13A CE inclusion than those from patients with FTLD-TDP lacking the risk alleles. These risk alleles seem insufficient to cause CE inclusion because CE is not detected extensively in RNA-seq data from healthy control samples16 (GTEx) and our functional studies indicate that TDP-43 dysfunction is required for substantial CE inclusion. Instead, the UNC13A risk alleles exert a TDP-43 loss-of-function-dependent disease-modifying effect. We propose that UNC13A risk alleles might act as a kind of Achilles’ heel, not causing problems until TDP-43 becomes dysfunctional. The discovery of a novel TDP-43-dependent cryptic splicing event in a bona fide FTD–ALS risk gene opens up a multitude of new directions for validating UNC13A as a biomarker and therapeutic target in ALS and FTD. This cryptic exon inclusion event—similar to that of STMN23,4—is not conserved in mouse, so will require studies in human neuron models to test whether blocking UNC13A cryptic splicing is sufficient to rescue phenotypes associated with loss of TDP-43 function. It is possible that a full rescue of TDP-43 function will require restoration of more than one cryptic splicing target (for example, STMN2, UNC13A and perhaps some of the others (Fig. 1b)). But the human genetics data (Fig. 4f) showing a dose-dependent decrease in survival in individuals carrying UNC13A risk alleles indicate that UNC13A is a key target of TDP-43. We note that UNC13A is more highly expressed in the frontal cortex (transcripts per million (TPM) = 530.2) than in the spinal cord16 (TPM = 35.54). One picture that might emerge is that the cryptic target STMN2 could have a key role in lower motor neurons in the spinal cord, whereas UNC13A could have a key role in the brain. Perhaps some combination of effects could contribute to ALS or FTD. It remains unknown why TDP-43 pathology is associated with ALS, FTLD-TDP, or even other aging-related neuropathological changes39. TDP-43-dysfunction-related cryptic splicing plays out across the diverse regional and neuronal landscape of the human brain. It is tempting to speculate that in addition to STMN2 and UNC13A, there could be specific portfolios of other important cryptic exon splicing events (and genetic variations that increase or decrease susceptibility to some of these events) that contribute to heterogeneity in clinical manifestation of TDP-43 dysfunction.

Methods

All materials used in this study are available upon request.

RNA-seq alignment and splicing analysis

The detailed pipeline v2.0.1 for RNA-seq alignment and splicing analysis is available on https://github.com/emc2cube/Bioinformatics/sh_RNAseq.sh. FASTQ files were downloaded from the Gene Expression Omnibus (GEO) database (GSE126543). Adaptors in FASTQ files were removed using trimmomatic (0.39) (ILLUMINACLIP:TruSeq3-PE.fa:2:30:10 LEADING:3 TRAILING:3 SLIDINGWINDOW:4:15 MINLEN:36). The quality of the resulting files was evaluated using FastQC (v0.11.9). RNA-seq reads were mapped to the human (hg38) using STAR v2.7.3a following ENCODE standard options, read counts were generated using RSEM v1.3.1, and differential expression analysis was performed in R v4.0.2 using the DESeq2 package v1.28.140.

Splicing analysis

MAJIQ

Alternative splicing events were analysed using MAJIQ (2.2) and VOILA13. In brief, uniquely mapped, junction-spanning reads were used by MAJIQ with the following parameters: ‘majiq build -c config–min-intronic-cov 1–simplify’, to construct splice graphs for transcripts by using the UCSC transcriptome annotation (release 82) supplemented with de novo detected junctions. Here, de novo refers to junctions that were not in the UCSC transcriptome annotation but had sufficient evidence in the RNA-seq data (–min-intronic-cov 1). Distinct local splice variations (LSVs) were identified in gene splice graphs, and the MAJIQ quantifier ‘majiq psi’ estimated the fraction of each junction in each LSV, denoted as percent spliced in (PSI or Ψ), in each RNA-seq sample. The changes in each junction’s PSI (ΔPSI or ΔΨ) between the two conditions (TDP-43-positive neuronal nuclei versus TDP-43-negative neuronal nuclei) were calculated by using the command ‘majiq deltapsi’. The gene splice graphs and the posterior distributions of PSI and ΔPSI were visualized using VOILA.

LeafCutter

LeafCutter is available as commit 249fc26 on https://github.com/davidaknowles/leafcutter. Using RNA-seq reads aligned as previously described, reads that span exon–exon junctions and map with a minimum of 6 nt into each exon were extracted from the alignment (bam) files using filter_cs.py with the default settings. Intron clustering was performed using the default settings in leafcutter_cluster.py. Differential excision of the introns between the two conditions (TDP-43-positive neuronal nuclei versus TDP-43-negative neuronal nuclei) were calculated using leafcutter_ds.R.

Sashimi plot

RNA-seq densities along the exons were plotted using the sashimi_plot function included in the MISO package (misopy 0.5.4). In the sashimi plot, introns are scaled down by a factor of 15 and exons are scaled down by a factor of 5. RNA-seq read densities across exons are scaled by the number of mapped reads in the sample and are measured in RPKM units. Slight modifications were made to plot_gene.py and plot_settings.py within the sashimi_plot directory of the MISO package to highlight the RNA-seq density plot. The modified sashimi_plot directory is available at (https://github.com/rosaxma/TDP-43-UNC13A-2021).

Cell culture

SH-SY5Y (ATCC) cells were grown in DMEM/F12 media supplemented with Glutamax (Thermo Scientific), 10% fetal bovine serum and 10% penicillin–streptomycin at 37 °C, 5% CO2. For shRNA treatments, cells were plated on day 0, transduced with shRNA on day 2 followed by media refresh on day 3, and collected for readout (RT–qPCR and immunoblotting) on day 6. HEK 293T TDP-43 knockout cells and parent HEK 293T cells were generated as described36. The cells were cultured in DMEM medium (Gibco 10564011) supplemented with 10% fetal bovine serum, 1% penicillin–streptomycin, 2 mM l-glutamine (Gemini Biosciences) and 1× MEM non-essential amino acids solution (Gibco) at 37 °C, 5% CO2.

iPS cell maintenance and differentiation into iPSC-MNs

iPS cell lines were obtained from public biobanks (GM25256-Corriell Institute; NDS00262, NDS00209-NINDS) and maintained in mTeSR1 media (StemCell Technologies) on Matrigel (Corning). iPS cells were fed daily and split every 4–7 days using ReLeSR (StemCell Technologies) according to the manufacturer’s instructions. Differentiation of iPS cells into motor neurons was carried out as previously described41. In brief, iPS cells were dissociated and placed in ultra-low adhesion flasks (Corning) to form 3D spheroids in media containing DMEMF12/Neurobasal (Thermo Fisher), N2 Supplement (Thermo Fisher), and B-27 Supplement-Xeno free (Thermo Fisher). Small molecules were added to induce neuronal progenitor patterning of the spheroids, (LDN193189, SB-431542, Chir99021), followed by motor neuron induction (with retinoic acid, Smo agonist and DAPT). After 14 days, neuronal spheroids were dissociated with Papain and DNAse (Worthington Biochemical) and plated on poly-d-lysine/laminin coated plates in Neurobasal Medium (Thermo Fisher) containing neurotrophic factors (BDNF, GDNF and CNTF; R&D Systems). For viral transductions, neuronal cultures were incubated for 18 h with media containing lentivirus particles for shScramble, or shTDP-43. Infection efficiency of over 90% was assessed by RFP expression. Neuronal cultures were analysed for RNA and protein 7 days post transduction.

shRNA cloning, lentiviral packaging, and cellular transduction for detecting the UNC13A splice variant

shRNA sequences were originated from the Broad GPP Portal (TDP-43: AGATCTTAAGACTGGTCATTC, scramble: GATATCGCTTCTACTAGTAAG). To clone, complementary oligonucleotides were synthesized to generate 4-nt overhangs, annealed and ligated into pRSITCH (Tet inducible U6) or pRSI16 (constitutive U6) (Cellecta). Ligations were transformed into Stbl3 chemically competent cells (Thermo Scientific) and grown at 30 °C. Large scale plasmid generation was performed using Maxiprep columns (Promega), with purified plasmid used as input for lentiviral packaging with second generation packaging plasmids psPAX2 and pMD2.G (Cellecta), transduced with Lipofectamine 2000 (Invitrogen) in Lenti-X 293T cells (Takara). Viral supernatant was collected at 48 and 72 h post transfection and concentrated using Lenti-X Concentrator (Takara). Viral titer was established by serial dilution in relevant cell lines and readout of percentage of BFP+ cells by flow cytometry, with a dilution achieving a minimum of 80% BFP+ cells selected for experiments.

Immunoblotting

SH-SY5Y cells and iPSCs-MNs were transfected and treated as above before lysis. Cells were lysed in ice-cold RIPA buffer (Sigma-Aldrich R0278) supplemented with a protease inhibitor cocktail (Thermo Fisher 78429) and phosphatase inhibitor (Thermo Fisher 78426). After pelleting lysates at maximum speed on a table-top centrifuge for 15 min at 4 °C, bicinchoninic acid (Invitrogen 23225) assays were conducted to determine protein concentrations. 60 μg (SH-SY5Y) and 30 μg (iPSCs-MNs) protein of each sample was denatured for 10 min at 70 °C in LDS sample buffer (Invitrogen NP0008) containing 2.5% 2-mercaptoethanol (Sigma-Aldrich). These samples were loaded onto 4–12% Bis–Tris gels (Thermo Fisher NP0335BOX) for gel electrophoresis, then transferred onto 0.45-μm nitrocellulose membranes (Bio-Rad 162-0115) at 100 V for 2 h using the wet transfer method (Bio-Rad Mini Trans-Blot Electrophoretic Cell 170-3930). Membranes were blocked in Odyssey Blocking Buffer (LiCOr 927-40010) for 1 h then incubated overnight at room temperature in blocking buffer containing antibodies against UNC13A (1:500, Proteintech 55053-1-AP), TDP-43 (1:1,000, Abnova H00023435-M01), or GAPDH (1:1,000, Cell Signaling Technologies 5174S). Membranes were subsequently incubated in blocking buffer containing horseradish peroxidase (HRP)-conjugated anti-mouse IgG (H+L) (1:2,000, Fisher 62-6520) or HRP-conjugated anti-rabbit IgG (H+L) (1:2,000, Life Technologies 31462) for 1 h. ECL Prime kit (Invitrogen) was used for development of blots, which were imaged using ChemiDox XRS+ System (Bio-Rad). The intensity of bands was quantified using Fiji, and then normalized to the corresponding controls.

RNA extraction, cDNA synthesis and RT–qPCR or RT–PCR for detecting the UNC13A splice variant in iPSC-MNs

Total RNA was extracted using RNeasy Micro kit (Qiagen) per manufacturer’s instructions, with lysate passed through a QIAshredder column (Qiagen) to maximize yield. RNA was quantified by Nanodrop (Thermo Scientific), with 75 ng used for cDNA synthesis with SuperScript IV VILO Master Mix (Thermo Scientific). Quantitative PCR was run with 6 ng cDNA input in a 20 µl reaction using PowerTrack SYBR Green Master Mix (Thermo Scientific) with readout on a QuantStudio 6 Flex using standard cycling parameters (95 °C for 2 min, 40 cycles of 95 °C for 15s and 60 °C for 60 s), followed by standard dissociation (95 °C for 15 s at 1.6 °C s−1, 60 °C for 60 s at 1.6 °C s−1, 95 °C for 15 s at 0.075 °C s−1). ΔΔCt was calculated with the housekeeper gene RPLP0 as control and relevant shScramble as reference; measured Ct values greater than 40 were set to 40 for visualizations. See Supplementary Table 6 for primers.

PCR was conducted with 15 ng cDNA input in a 100 µl reaction using NEBNext Ultra II Q5 Master Mix (New England Biolabs), with the following cycling parameters: initial denaturation: 98 °C for 30 s; 40 cycles: 98 °C for 10 s, 64 °C for 30 s, 72 °C for 20 s; final extension: 72 °C for 2 min. The resulting products were visualized on a 1.5% TAE gel. See Supplementary Table 6 for primers.

Human iPS cell-derived neurons for detecting UNC13A splice variants

cDNA was available from CRISPRi-i3Neuron iPS cells (i3N) generated from our previous publication11, in which TDP-43 is downregulated to about 50%. RT–qPCR was performed using SYBR GreenER qPCR SuperMix (Invitrogen). Samples were run in triplicate, and RT–qPCR reactions were run on a QuantStudio 7 Flex Real-Time PCR System (Applied Biosystems). Relative quantification was determined using the ∆∆Ct method and normalized to the endogenous controls RPLP0 and GAPDH. We normalized relative transcript levels for wild-type UNC13A to that of the neurons treated with control sgRNA (mean set to 1). See Supplementary Table 6 for primers.

Cell culture for validating additional splicing events in iPS cell-derived neurons

We used an induced neuron (iN) system previously established for rapidly differentiating human iPS cells into functional cortical neurons42. In brief, iPS cells (without disease mutation) were cultured using feeder-free conditions on Matrigel (Fisher Scientific CB-40230) using mTeSR1 media (Stemcell Technologies 85850). Cells were transduced with a Tet-On induction system that allows expression of the transcription factor NGN2. Cells were dissociated on day 3 of differentiation and replated on Matrigel-coated tissue culture plates in Neurobasal Medium (Thermo Fisher) containing neurotrophic factors, BDNF and GDNF (R&D Systems) with viral transductions for shScramble or shTDP-43. RNA and protein were extracted 7 days after transduction.

shRNA cloning, lentiviral packaging, and cellular transduction for validating additional splicing events

The lentiviral plasmid targeting TARDBP (Millipore-Sigma TRCN0000016038) and Scramble (CAACAAGATGAAGAGCACCAA) were packaged using third generation packaging plasmids (pMDLg/pRRE, pRSV-Rev, pMD2.G) and transduced with Lipofectamine 3000 (Invitrogen) into HEK 293T cells cultured under standard conditions (DMEM, 10% FBS, penicillin–streptomycin). Viral supernatant was collected at 48 and 72 h post-transfection and concentrated 1:100 using Lenti-X Concentrator (Takara).

RNA extraction, cDNA synthesis and RT–qPCR for validating additional splicing events

Total RNA was extracted using RNeasy Micro kit (Qiagen) and reverse transcribed into cDNA using High-Capacity cDNA Reverse Transcription Kits (Invitrogen). Quantitative PCR was run with 2 ng cDNA input in a 10 µl reaction using PowerTrack SYBR Green Master Mix (Thermo Scientific) with readout on a QuantStudio 6 Flex using standard cycling parameters. ΔΔCt was calculated with RPLP0 or GAPDH as housekeeper gene controls and relevant shScramble transduced condition as reference; measured Ct values greater than 40 were set to 40 for visualizations. See Supplementary Table 6 for primers used for detecting mis-spliced transcripts and normal splicing transcripts, and primers used for internal controls.

Amplicon sequencing of the splice variants

Splice variants in iPSC-MNs were established by PCR amplification from UNC13A exon 19 to exon 21 (UNC13A_19_21 FWD 5′–3′= CAACCTGGACAAGCGAACTG, UNC13A_19_21 RVS 5′-3′= GGGCTGTCTCATCGTAGTAAAC). Resulting products were purified using Wizard SV Gel and PCR Clean-Up columns (Promega) and submitted for NGS (Amplicon EZ, Genewiz). Adaptors in FASTQ files were removed using trimmomatic (0.39) (ILLUMINACLIP:TruSeq3-PE.fa:2:30:10 LEADING:3 TRAILING:3 SLIDINGWINDOW:4:15 MINLEN:36). The quality of the resulting files was then evaluated using FastQC (v0.11.9). The sequencing reads were then mapped to the human (hg38) using STAR v2.7.3a following ENCODE standard options. Uniquely mapped reads were then filtered for using the command ‘samtools view -b -q 255’. The Sashimi Plot were then generated using the sashimi plot function in IGV (2.8.0) with the minimum junction coverage set to 20.

Post-mortem brain tissues for detecting UNC13A splice variant

Post-mortem brain tissues from patients with FTLD-TDP and cognitively normal control individuals were obtained from the Mayo Clinic Florida Brain Bank. Diagnosis was independently ascertained by trained neurologists and neuropathologists upon neurological and pathological examinations, respectively. Written informed consent was given by all participants or authorized family members and all protocols were approved by the Mayo Clinic Institution Review Board and Ethics Committee. Complementary DNA (cDNA) obtained from 500 ng of RNA (RIN ≥ 7.0) from medial frontal cortex was available from a previous study, as well as matching pTDP-43 data from the same samples43. Following standard protocols, RT–qPCR was conducted using SYBR GreenER qPCR SuperMix (Invitrogen) for all samples in triplicates. RT–qPCR reactions were run in a QuantStudio 7 Flex Real-Time PCR System (Applied Biosystems). Relative quantification was determined using the ∆∆Ct method and normalized to the endogenous controls RPLP0 and GAPDH. We normalized relative transcript levels to that of the healthy controls (mean set to 1). See Supplementary Table 6 for primers.

Quantification of UNC13A splice variants in bulk RNA sequencing

RNA-seq data generated by NYGC ALS Consortium cohort were downloaded from the NCBI Gene Expression Omnibus (GEO) database (GSE137810, GSE124439, GSE116622 and GSE153960). We used the 1658 available and quality-controlled samples classified as described11. After pre-processing and aligning the reads to human (hg38) as described previously, we estimated the expression of the full-length UNC13A using RSEM (v1.3.2). PCR duplicates were removed using MarkDuplicates from Picard Tools (2.23.0) using the command ‘MarkDuplicates REMOVE_DUPLICATES=true CREATE_INDEX=true’. We then filtered for uniquely mapped reads using the command ‘samtools view -b -q 255’. Reads that span either exon 19–exon 20 junction, exon 20–CE junction, CE–exon 21 junction or exon 20–exon 21 junction were quantified using bedtools (2.27.1) using the command ‘bedtools intersect -split’. Because of the relatively low level of expression of UNC13A in post-mortem tissues and the heterogeneity of the tissues, it is possible that not all tissues have enough detectable UNC13A for us to detect the splice variants. Since UNC13A contains more than 40 exons and RNA-seq coverages of mRNA transcripts are often not uniformly distributed44, we looked at reads spanning the exon 19–exon 20 junction, which is included in both the canonical isoform variant and the splice variant, and there is a strong correlation (Pearson’s r = 0.99) between the numbers of reads mapped to the exon 19–exon 20 junction and the exon 20–exon 21 junction. We observed that samples that have at least 2 reads spanning either exon 20–CE junction or CE–exon 21 junction have at least either UNC13A TPM = 1.55 or 20 reads spanning exon 19– exon 20 junction. Therefore, we selected the 1,151 samples that had a TPM ≥ 1.55, or at least 20 reads mapped to the exon 19–exon 20 junction as samples suitable for UNC13A splice variant analysis.

In situ hybridization UNC13A CE analysis in postmortem brain samples

Patients and diagnostic neuropathological assessment

Postmortem brain tissue samples used for this study were obtained from the University of California San Francisco (UCSF) Neurodegenerative Disease Brain Bank (Supplementary Table 4). Supplementary Table 4 provides demographic, clinical, and neuropathological information. Consent for brain donation was obtained from subjects or their surrogate decision makers in accordance to the Declaration of Helsinki, and following a procedure approved by the UCSF Committee on Human Research. Brains were cut fresh into 1 cm thick coronal slabs, and alternate slices were fixed in 10% neutral buffered formalin for 72 h. Blocks from the medial frontal pole were dissected from the fixed coronal slabs, cryoprotected in graded sucrose solutions, frozen, and cut into 50 µm thick sections as described previously45. Clinical and neuropathological diagnosis were performed as described previously45. Subjects were selected on the basis of clinical and neuropathological assessment. Patients selected had a primary clinical diagnosis of behavioural variant frontotemporal dementia (bvFTD) with or without amyotrophic lateral sclerosis or motor neuron disease and a neuropathological diagnosis of FTLD-TDP, type B. We excluded subjects if they had a known disease-causing mutation, post-mortem interval ≥ 24 h, Alzheimer’s disease neuropathologic change > low, Thal amyloid phase > 2, Braak neurofibrillary tangle stage > 4, CERAD neuritic plaque density > sparse, and Lewy body disease > brainstem predominant45.

In situ hybridization and immunofluorescence

To detect single RNA molecules, a BaseScope Red Assay kit (ACDBIO, USA) was used. One 50 µm thick fixed frozen tissue section from each subject was used for staining. Experiments were performed under RNase-free conditions as appropriate. Probes that target the transcript of interest, UNC13A, specific to either the mRNA (exon 20–exon 21 junction) or the cryptic exon containing spliced target (exon 20–cryptic exon junction) were used. Positive (Homo Sapiens PPIB) and negative (Escherichia coli DapB) control probes were also included. In situ hybridization was performed based on vendor specifications for the BaseScope Red Assay kit. In brief, frozen tissue sections were washed in PBS and placed under an LED grow light (HTG Supply, LED-6B240) chamber for 48 h at 4 °C to quench tissue autofluorescence. Sections were quickly rinsed in PBS and blocked for endogenous peroxidase activity. Sections were transferred on to slides and dried overnight. Slides were subjected to target retrieval and protease treatment and advanced to ISH. Probes were detected with TSA Plus-Cy3 (Akoya Biosciences), and subjected to immunofluorescence staining with antibodies to TDP-43 (rabbit polyclonal, Proteintech, RRID: AB_615042, dilution 1:4,000, catalogue (cat.) no. 10782-2-AP) and NeuN (Guinea pig polyclonal, Synaptic Systems, dilution 1:500; cat. no. 266004), and counterstained with DAPI (Life Technologies) for nuclei.

Image acquisition and analysis

Z-stack images were captured using a Leica SP8 confocal microscope with an 63× oil immersion objective (1.4 NA). For RNA probes, image capture settings were established during initial acquisition based on PPIB and DAPB signal and remained constant across UNC13A probes and subjects. TDP-43 and NeuN image capture settings were modified based on staining intensity differences between cases. For each case, 6 non-overlapping Z-stack images were captured across cortical layers 2–3. RNA puncta for the UNC13A cryptic exon were quantified using the ‘analyze particle’ plugin in ImageJ. In brief, all images were adjusted for brightness using similar parameters and converted to maximum intensity Z-projections, images were adjusted for auto-threshold (intermodes), and puncta were counted (size: 6-infinity, circularity: 0–1).

Linkage disequilibrium analysis

Recalibrated VCF files of 297 ALS patients of European descent generated by GATK HaplotypeCallers were downloaded from Answer ALS in July 2020 (https://www.answerals.org). VCFtools (0.1.16) were used to filter for sites that are in intron 20–21. The filtered VCF files were merged using BCFtools (1.8). Since there are sites that contain more than 2 alleles, we tested for genotype independence using the chi-squared statistics by using the command ‘vcftools–geno-chisq–min-alleles 2–max-alleles 8’. We found two additional SNPs, rs56041637 (P < 0.0001 with rs12608932, P < 0.0001 with rs12973192), and rs62121687 (P < 0.0001 with rs12608932, P < 0.0001 with rs12973192) that are in linkage disequilibrium with both. However, since rs62121687 was included in a GWAS and has a P-value35 of 0.0186585, we excluded it from further analysis.

Determination of rs12608932 and rs12973192 SNP genotype in human postmortem brain

Genomic DNA (gDNA) was extracted from human frontal cortex using Wizard Genomic DNA Purification Kit (Promega), according to the manufacturer’s instructions. TaqMan SNP genotyping assays were performed on 20 ng of gDNA per assay, using a commercial pre-mixture consisting of a primer pair and VIC or FAM-labelled probes specific for each SNP (cat. no. 4351379, assay ID 43881386_10 for rs12608932 and 11514504_10 for rs12973192, Thermo Fisher Scientific), and run on a QuantStudio 7 Flex Real-Time PCR system (Applied Biosystems), according to the manufacturer’s instructions. The PCR programs were 60 °C for 30 s, 95 °C for 10 min, 40 cycles of 95 °C for 15 s and, 60 °C (rs12973192) or 62.5 °C for 1 min (rs12608932), and 60 °C for 30 s.

Splicing reporter assay

Minigene constructs were designed in silico, synthesized by GenScript and sub-cloned into a vector with the GFP splicing control. HEK 293T TDP-43 knockout cells and the parent HEK 293T cells were seeded into standard P12 tissue culture plates (at 1.6 × 105 cells per well), allowed to adhere overnight, and transfected with the indicated splicing reporter constructs (400 ng per well) using Lipofectamine 3000 transfection reagent (Invitrogen). Each reporter comprised one of the splicing modules (shown in Fig. 4d), which is expressed from a bidirectional promoter. Twenty-four hours after transfection, RNA was extracted from these cells using PureLink RNA Mini Kit (Life Technologies) according to the manufacturer’s protocol, with on-column PureLink DNase (Invitrogen) treatment. The RNA was reverse transcribed into cDNA using the High Capacity cDNA Reverse Transcription Kit (Invitrogen) according to the manufacturers’ instructions. PCRs were performed using OneTaq 2X Master Mix with Standard Buffer (NEB) with the following cycling parameters: denaturation: 94 °C for 30 s; 30 cycles: 94 °C for 20 s, 54 °C for 30 s, 68 °C for 30 s; final extension: 68 °C for 5 min on a Mastercycler Pro (Eppendorf) thermocycler PCR machine. PCR products were separated by electrophoresis on a 1.5% TAE gel and imaged ChemiDox XRS+ System (Bio-Rad). See Supplementary Table 6 for primers.

Assay to assess the effect of variants at rs12973192, rs12608932 and rs56041637 on splicing

Additional minigene constructs shown in Extended Data Fig. 8 were either generated using site-directed mutagenesis (New England Biolabs, E0554S) or synthesized by GenScript, and sub-cloned into the vector with the GFP splicing control. HEK 293T TDP-43 knockout cells and the parent HEK 293T cells were seeded into standard P12 tissue culture plates (at 5 × 105 cells per well), allowed to adhere overnight and transfected with the indicated splicing reporter constructs (400 ng per well) using Lipofectamine 3000 transfection reagent (Invitrogen). Twenty-four hours after transfection, RNA was extracted from these cells using PureLink RNA Mini Kit (Life Technologies) according to the manufacturer’s protocol, with on-column PureLink DNase treatment. The RNA was reverse transcribed into cDNA using the High Capacity cDNA Reverse Transcription Kit (Invitrogen) according to the manufacturers’ instructions. The UNC13A cryptic exon signal was measured using a pair of primers that detect the junction of the CE and the immediately downstream mCherry exon. The splicing of eGFP was measured using a pair of primers that detect the junction of the first and second exons of eGFP. A pair of primers that mapped within the second exon of eGFP was used to measure the transfection efficiency of the splicing reporter construct and was used as a normalizer. ΔΔCt was calculated using the cryptic exon signal level or the splicing of eGFP in the HEK 293T TDP-43 knockout cells expressing the reference haplotype-carrying reporter as reference. See Supplementary Table 6 for primers.

Rescue of UNC13A splicing using TDP-43 overexpression constructs

HEK 293T TDP-43 knockout cells and the parent (wild-type) HEK 293T cells were seeded into standard P12 tissue culture plates (at 5 × 105 cells per well), allowed to adhere overnight and transfected with the splicing reporter construct carrying the reference haplotype (400 ng per well; Fig. 4e) and the indicated TDP-43 overexpression constructs (600 ng per well) using Lipofectamine 3000 transfection reagent (Invitrogen). Twenty-four hours after transfection, RNA was extracted from these cells using PureLink RNA Mini Kit (Life Technologies) according to the manufacturer’s protocol, with on-column PureLink DNase treatment. The RNA was reverse transcribed into cDNA using the High Capacity cDNA Reverse Transcription Kit (Invitrogen) according to the manufacturers’ instructions. Quantitative PCR was run with 8 ng cDNA input in a 10 µl reaction using PowerTrack SYBR Green Master Mix (Thermo Scientific) with readout on a QuantStudio 6 Flex using standard cycling parameters.

The UNC13A cryptic exon signal was measured using a pair of primers that detect the junction of the CE and the mCherry exon immediately downstream of it. A pair of primers that are mapped within the second exon of eGFP was used to measure the transfection efficiency of the splicing reporter construct, and was used as a normalizer. ΔΔCt was calculated using the cryptic exon signal level in the wild-type HEK 293T cells without TDP-43 overexpression constructs as reference. See Supplementary Table 6 for primers.

The expression levels of the overexpression constructs were measured using a pair of primers that detect the second exon of TDP-43. The primers do detect the endogenous TDP-43 but since the HEK 293T TDP-43 knockout cells do not have TDP-43 expression as shown previously36, using the primers do not interfere with the measurement of the expression levels of TDP-43 constructs in the knockout cells. ΔΔCt was calculated using the TDP-43 expression level in the HEK 293T TDP-43 knockout cells with full length TDP-43 overexpression constructs as reference. RPLP0 and GAPDH were used as internal controls. See Supplementary Table 6 for primers.

Generation of pTB UNC13A minigene construct

The pTB UNC13A minigene construct containing the human UNC13A cryptic exon sequence and the nucleotide flanking sequences upstream (50 bp at the of end of intron 19, the entirety of exon 20, and the entirety of intron 20 sequence upstream of the cryptic exon) and downstream (approximately 300-bp intron 20) of the cryptic exon were amplified from human genomic DNA using the following primers: FWD 5′–3′, AGGTCATATGCACTGCTATAGTGGGAAGTTC and RVS 5′–3′, CTTACATATGTAATAACTCAACCACACTTCCATC; and subcloned into the NdeI site of the pTB vector. We have previously used a similar approach to study TDP-43 splicing regulation of other TDP-43 targets46 .

Rescue of UNC13A splicing using the pTB minigene and TDP-43 overexpression constructs

HeLa cells were grown in Opti-MEM I Reduced Serum Medium, GlutaMAX Supplement (Gibco) plus 10% fetal bovine serum (Sigma), and 1% penicillin/streptomycin (Gibco). For double-transfection and knockdown experiments, cells were first transfected with 1.0 µg of pTB UNC13A minigene construct and 1.0 µg of one of the following plasmids: GFP, GFP-TDP-43 or GFP-TDP-43 5FL constructs to express GFP-tagged TDP-43 proteins have been previously described46,47, in serum-free media and using Lipofectamine 2000 following the manufacturer’s instructions (Invitrogen). Four hours following transfection, media was replaced with complete media containing siLentfect (Bio-Rad) and siRNA complexes (AllStars Neg. Control siRNA or siRNA against TARDBP 3′ untranslated region, a region not included in the TDP-43 overexpression constructs) (Qiagen) following the manufacturer’s protocol. Cycloheximide (Sigma) was added at a final concentration of 100 µg ml−1 at 6 h prior to collecting the cells. Then RNA was extracted from the cells using TRIzol Reagent (Zymo Research), following the manufacturer’s instructions. Approximately 1 µg of RNA was converted into cDNA using the High Capacity cDNA Reverse Transcription Kit with RNA inhibitor (Applied Biosystems). The RT–qPCR assay was performed on cDNA (diluted 1:40) with SYBR GreenER qPCR SuperMix (Invitrogen) using QuantStudio7 Flex Real-Time PCR System (Applied Biosystems). All samples were analysed in triplicates. The RT–qPCR program was as follows: 50 °C for 2 min, 95 °C for 10 min, and 40 cycles of 95 °C for 15 s and 60 °C for 1 min. For dissociation curves, a dissociation stage of 95 °C for 15 s, 60 °C for 1 min and 95 °C for 15 s was added at the end of the program. Relative quantification was determined using the ∆∆Ct method and normalized to the endogenous controls RPLP0 and GAPDH. We normalized relative transcript levels for wild-type UNC13A and GFP to that of the control siRNA condition (mean set to 1). See Supplementary Table 6 for primers.

In vitro TDP-43 binding studies

Cloning

The plasmid encoding TDP43 as a C-terminal MBP-tagged protein (TDP43–MBP–His6) was purchased from Addgene (#104480).

Bacterial growth and protein expression

The wild-type TDP-43 expression plasmid was transformed into E. coli One Shot BL21 Star (DE3) cells (ThermoFisher). Transformed E. coli were grown at 37 °C in 1 l of LB media supplemented with 0.2% dextrose and 50 μg ml−1 kanamycin until absorbance at 600 nm reached 0.5–0.6. The culture was then incubated at 4 °C for 30–45 min. TDP-43 expression was induced with 1 mM IPTG for 16 h at 4 °C. Cells were collected by centrifugation.

Recombinant TDP-43 purification

Wild-type TDP-43–MBP was purified as described48. In brief, cell pellets were resuspended in lysis buffer 1 M NaCl, 20 mM Tris (pH 8.0), 10 mM imidazole, 10% glycerol and 2.5 mM 2-mercaptoethanol and supplemented with cOmplete, EDTA-free protease inhibitor cocktail tablets (Roche) then lysed via sonication. Cell lysates were centrifuged at 31,400g at 4 °C for 1 h, filtered, then purified with FPLC using a XK 50/20 column (Cytiva) packed with Ni-NTA agarose beads (Qiagen) which were equilibrated in lysis buffer. TDP-43 was recovered via a 0–80% gradient elution using 1 M NaCl, 20 mM TrisHCl (pH 8.0), 10 mM imidazole, 10% glycerol and 2.5 mM 2-mercaptoethanol as the base buffer and 1 M NaCl, 20 mM TrisHCl (pH 8.0), 500 mM imidazole, 10% glycerol, and 2.5 mM 2-mercaptoethanol as the elution buffer. Eluted protein was concentrated using Amicon Ultra-15 centrifugal filters, MWCO 50 kDa (Millipore), filtered and further purified with size-exclusion chromatography using a 26/600 Superdex 200 pg column (Cytiva) equilibrated with 300 mM NaCl, 20 mM TrisHCl (pH8.0) and 1 mM DTT. The second out of three peaks, as evaluated by absorbance at 280 nm, was collected, spin concentrated as before, aliquoted, flash frozen in liquid N2, and stored at −80 C until further use. Protein concentrations were determined using absorbance at 280 nm (Nanodrop) and purity was determined by running samples on a 4–20% SDS–PAGE gel and visualized with Coomassie stain.

Electrophoresis mobility shift assay

EMSA was used to compare TDP-43 binding to the reference and risk RNA sequences for reference and risk alleles of CE (rs12973192), intron (rs12608932), and repeat sequences (rs56041637) (see Supplementary Table 5). Increasing TDP-43 concentrations ranging from 0–4 mM were incubated with a constant 1 nM concentration of RNA in buffer (50 mM Tris-HCl, pH 7.5, 100 mM KCl, 2 mM MgCl2, 100 mM β−mercaptoethanol, 0.1 mg ml−1 BSA) for 30 min at room temperature. RNA is dual-labelled (Cy3 and Cy5) and contains an 18-nucleotide partial duplex on the 3′ end. Reactions were mixed with loading dye and run on a 6% non-denaturing polyacrylamide gel and imaged using fluorescence mode (Cy5) on a Typhoon scanner. Bound fractions were determined using the Analyze Gel plugin in ImageJ and normalized to the total intensity per lane. Apparent binding affinities were calculated using the ‘Specific binding with Hill slope’ function in Graphpad.

Statistical methods

Survival curves were compared using the coxph function in the survival (3.1.12) R package, which fits a multivariable Cox proportional hazards model that contains sex, reported genetic mutations and age at onset, and performs a score (log-rank) test. Effect sizes are reported as the hazard ratios. Proportional Hazards assumptions were tested using cox.zph function. The survival curves were plotted using ggsurvplot in suvminer (v.0.4.8) R package. Linear mixed effects models were analysed using lmerTest R package (3.1.3). Statistical analyses were performed using R (version 4.0.0), or Prism 8 (GraphPad), which were also used to generate graphs.

Reporting summary

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