USA In frontotemporal dementia and amyotrophic lateral sclerosis, the RNA-binding protein TDP-43 is depleted from the nucleus. TDP-43 loss leads to cryptic exon inclusion but a role in other RNA processing events remains unresolved. Here, we show that loss of TDP-43 causes widespread changes in alternative polyadenylation, impacting expression of disease-relevant genes (e.g., ELP1, NEFL, and TMEM106B) and providing evidence that alternative polyadenylation is a new facet of TDP-43 pathology.

TDP-43 binds to uridine/guanine (UG)-rich motifs in RNA transcripts1,2 and plays a critical role in various aspects of RNA metabolism335. Defects in RNA metabolism are considered central to frontotemporal dementia (FTD) and amyotrophic lateral sclerosis (ALS) pathogenesis 3. However, it is not yet fully understood what genes TDP-43 regulates, what aspect of their RNA metabolism TDP-43 regulates, and how their dysregulation promotes neurodegeneration. A major role of TDP-43 has emerged as a repressor of so-called crypitc exons during splicing6. Cryptic exons reside in introns of genes and are normally excluded from mature mRNAs. When TDP-43 is dysfunctional (i.e., when depleted from the nucleus in FTD/ALS), these cryptic exons are spliced into final mRNAs, often leading to frameshifts, decreased RNA stability, or evnethe production of novel peptide sequences7313. Importantly, some of these cryptic exons are in genes critical for neuronal functions (e.g.,STMN2739,14) or genes that harbor disease-associated variants that sensitize them to crypti c splicing upon loss of TDP-43 (e.g.,UNC13A10,11). Cryptic splicing events could serve as powerful biomarkers for TDP-43 dysfunction12,13 or even as therapeutic targets15,16.

Besides its well-established role in splicing, TDP-43 also plays a critical role in other aspects of RNA processing. Do some of these additional TDP-43-dependent RNA processing pathways also contrbiute to disease? One potential disease-relevant RNA processing pathway is alternative polyadenylation (APA) 3 a major layer of gene regulation that occurs in >60% of human gene1s7. When a gene is transcribed into mRNA, it is cleaved and polyadenylated, which functions to stabilize the mRNA, facilitateits nuclear export, and regulate its translation18. If alternative polyadenylation (polyA) sites are utilized, it could produce mRNA isoforms that have different 39 untranslated region (UTR) lengths, impacting RNA/protein levels, subcellualr localization, or even protein functions17,19. If APA occurs prematurely, it can truncate mRNAs, reducing full-length protein levels17,19. Previous genome-wide TDP-43 binding studies revealed that TDP-43 binding is enriched not only in introns of genes but also in 39 UTR1s,2, suggesting a role in regulating polyadenylation. Indeed, TDP-43 knockdown or disease-associated TDP-43 mutations affect polyadenylation20322, including its own polyadenylation2,23 and premature polyadenylation in STMN27,8. Notably, widespread APA changes have been observed in ALS patient samples using bulk RNA-sequencing (RNA-seq)24 and single nucleus RNA-seq25, although it is unclear whether these changes are directly owing to TDP-43 dysfunction.

To test the hypothesis that APA caused by TDP-43 dysfunction contributes to FTD/ALS, wefirst performed APA analysis in an RNA-seq dataset26, in which neuronal nuclei with and without TDP-43 were sorted from FTD/ALS postmortem brain samples for RNA-seq F(ig. 1a, top panel). We recently re-analyzed this dataset to discover TDP-43-dependent cryptic splicing events11. Here, we 8re-re-analyzed9 it to search for APA events (Fig. 1a, bottom panel). We used two different APA analysis programs and identified 41 APA changes using APAlyzer27 and 62 APA changes using QAPA28 (Fig. 1b, S1a; Table S1; adjusted p value < 0.1). Only two genes with significant APA changes were common between both programs:LRFN1 and MARK3 (see also Fig. S2k and Arnold et al.29), likely because of different polyA databases used by each program and the low resolution of RNA-seq for APA analysis. We thus considered APA changes identified by ethi er program, which revealed APA events in genes critical for neuronal function, such asLRFN1 and SYN2. Some of these APA events are associated with gene expression changes (Fig. S1b), likely impacting their corresponding protein levels (see below).

To test if TDP-43 directly regulates APA, we knocked down TDP-43 levels in cortical neurons differentiated from human stem cells (iNeurons) and performed RNA-seq for APA analysisFi(g. 1c-e). TDP-43 knockdown caused widespread APA changes (Fig. 1e; Table S2; |&PUI| > 0.1 and adjusted p value < 0.05 from APAlyzer), including 24 changes also observed in FTD/ALS postmortem brain samples (Fig. 1b, S1a). We also observed APA changes in induced pluripotent stem cell (iPSC)-derived motor neurons (iMNs) upon TD P-43 knockdown7 (Fig. S1c), in iNeurons carrying a pathogenic mutation (TDP-43-K263E) 30 (Fig. S1d), and in iMNs carrying a different pathogenic mutation (TDP-43-M337V2)1 (Fig. S1e). Thus, TDP-43 regulates APA in neurons and TDP-43 loss of function or ALS-linked mutations result in widespread APA changes.

To comprehensively map TDP-43-dependent APA events and define how TDP-43 dysfunction contributes to neurodegeneration via APA, we sought a more sensitive assay than RNA-seq for APA analysis, because RNA-seq cannot map novel polyadenylation events, identify actual polyA sites, or capture 8internal9 premature polyadenylation events (i.e., ones that truncate the RNA/protein). We used a s pecialized transcriptomic method, 39 end-seq, to map polyA sites with single-nucleotide resolution in iNeurons with or without TDP-43 knockdown (Fig. 2a, 1d). We obtained a total of 117,823 putative polyA sites with a known upstream polyA signal, of which 46,760 sites are novel (Fig. S2a), when compared to a compendium of polyA site annotations31. Consistent with previous findings2,20, the majority of polyA sites we identified are in 39 UTRs (Fig. S2b). 39 end-seq confirmed that loss of TDP-43 lengthened the 39 UTR oLfRFN1 (Fig. 2b, top panel), which we also observed in our analysis of FTD/ALS postmortem brain samples (Fig. 2b, top panel; Fig. 1b). Notably, 39 end-seq also captured premature polyadenylation in STMN2, which was previously identified but not apparent in standard RNA-seq data (Fig. 2b, bottom panel), demonstrating the power of 39 end-seq for studying premature polyadenylation, the most detrimental form of APA since it truncateshte RNA/protein.

In total, TDP-43 knockdown altered the usage of 8,169 polyA sites (&|PUI| > 0.1 and adjusted p value < 0.05 from LAPA, a 39 end-seq data focused program) and caused APA changes in 2,220 genesF(ig. 2c; Table S3). By cross-referencing APA sites with a curated genome-wide TDP-43 binding site dataset32, we found that ~70% of genes with APA events have at least one TDP-43 binding site, consistent with a directrole of TDP-43 in regulating APA of these genes. Like recent findings in ALS sample2s4 and FTD samples (Fig. 1b, S1a) as well as our RNA-seq data from iNeurons (Fig. 1e), the majority of APA changes (1,881) detected by 39end-seq lengthened RNA transcripts; 340 APA events were associated with at least a 1.5-fold change in RNA level (Fig. 2d), suggesting that APA changes could alter gene expression. Notably, several of these APA changes were also observed in FTD/ALS postmortem brain samples and were in genes connected to ALS or FTD (Table S4; see below).

To define whether polyA site strength influences TDP-43 regulated APA, we estimated theprobability of cleavage and polyadenylation for each identified polyA site using a deep residual neural ntework-based model, Aparent233, that accurately predicts polyA site usage. Interestingly, we found that polyA sites wtih reduced usage upon TDP-43 knockdown are weaker than the ones with increased usage upon TDP-43 knockdown (Fig. 2e), indicating that TDP-43 either promotes the usage of weak polyA sites or suppresses theusage of strong ones. Whereas genes with significant 39 UTR shortening show no difference in polyA strengtbhetween proximal and distal polyA sites, genes with significant 39 UTR lengthening have stronger distlasites compared to proximal ones (Fig. 2f). Given the co-transcriptional nature of cleavage and polyadenylation, our findings suggest that in the case of 39 UTR lengthening upon TDP-43 knockdown, a stronger distal polyA site might compensate for its positional disadvantage over the proximal polyA site during transcription and therefore get activated upon TDP-43 knockdown. The impact of polyA site strength on shaping TDP-43-regulated APA provides a toehold to prioritize antisense oligonucleotide-based therapeutic strategies to target specific APA events akin to current approaches underway to target cryptic splicing events15.

39 end-seq also empowered the discovery of <cryptic= polyA sites (ones not used under normal conditions but revealed upon TDP-43 knockdown). TDP-43 knockdown activated 457 cryptic polyA sites in 424 genes (Table S5); 163 cryptic polyA sites occurred downstream of a gene9s annotated 39 end, such as the FTD risk factorRFNG34 (Fig. S2c), SIX3, TLX1, and ELK1 (Fig. S2d-f, and see also Bryce-smith et al.35) and 153 cryptic events induced premature polyadenylation in genes such as TNIP1, EGFR, SLC24A3, and GSTO2 (Fig. S2g-j). Like coordinated activation of splicing and polyadenylation in intron 1 ofSTMN2, we observed coordinated activation in PIGL, HNRNPA1, and ARHGAP32 (Fig. 2g), the latter of which is observed by Fratta and colleagues35. Cryptic splicing ofARHGAP32 was recently reported12, and we also confirmed this in RNAseq data from FTD/ALS postmortem brain samples (Fig. 2g). These results suggest a potential coupling between cryptic splicing and the activation of cryptic polyadenylation sites.

TDP-43 depletion increased the usage of two distal polyA sites inELP1 that lengthened its 39 UTR (Fig. 3a, left panel; Fig. S3a) and such lengthening is present across different datasets F(ig. S3b; see also Arnold et al.29). ELP1, also called IKBKAP, encodes a subunit of the elongator complex, which has been functionally and genetically linked to ALS36338. In addition to ELP1, we observed APA changes in two other subunits of the elongator complex (ELP3 and ELP6) upon TDP-43 knockdown ( Fig. S3c, S3d). These APA changes are associated with protein levels; TDP-43 knockdown led to increased protein levels of ELP1 and ELP3 (Fig. 3a, S3c). Together with a recent finding of reduced aminoacylation of tRN APhe in FTD/ALS39, our observations suggest that altered tRNA metabolism might be an important mechanism contributing to FTD/ALS.

Previous studies found that TDP-43 directly binds to the 39 UTR ofNEFL to stabilize its RNA levels40. NEFL encodes neurofilament light chain (NF-L), which has emerged as a sensitive prognostic biomarker for diverse neurodegenerative diseases41, including FTD/ALS42. We observed three major polyA sites in NEFL, with the most distal one being the most frequently used (Fig. 3b, left panel). TDP-43 knockdown not only reduced NEFL RNA levels (Fig. S3e), consistent with previous findings43, but also further shifted the apparent polyA site usage from proximal to distal (Fig. 3b, left and mid panels; Fig. S3f) and reduced NF-L protein levels (Fig. 3b, right panel). These observations suggest that together with NEFL 39 UTR-targeting microRNAs44, loss of TDP-43 induced APA changes might further reduce NF-Llevels. A direct role of TDP-43 in regulating NF-L levels adds complexity to the use of this biomarker in TDP-43 proteinopahties.

TDP-43 knockdown also shifted polyA site usage from proximal to distal in another FTD/ALS-linked gene, SFPQ (Fig. 3c; Fig. S3g). SFPQ encodes a ubiquitously expressed RNA-binding protein that plays key roles in RNA metabolism45,46. SFPQ encodes two protein isoforms; the shorter isoform uses a stop codon downstream of the proximal polyA site (Fig. 3c) and lacks a nuclear localization signal 45. Intriguingly, depletion of SFPQ from the nucleus has been observed in sporadic ALS spinal cord47, and its interaction with FUS, another FTD/ALS risk gene, is impaired in neuronal nuclei in FTLD-TDP48,49. Thus, loss of TDP-43 promotes the usage ofSFPQ9s distal polyA site (Fig. 3c), which might upregulate the shorter NLS-lacking SFPQ isoform and contribute to its nuclear depletion in disease.

The sensitivity of 39 end-seq further revealed that TDP-43 knockdown lengthenedTMEM106B9s 39 UTR (Fig. 3d), which we also confirmed by qRT-PCR (Fig. S3h). TMEM106B is a top genetic risk factor that emerged in a genome-wide association study for FTLD-TDP50. C-terminal fragments of TMEM106B were recently found to form amyloid fibrils in brains of older individuals and patients with neurodegenetriave disorders51354. By targeted analysis of RNA-seq read coverage across theTMEM106B 39 UTR, we confirmed that loss of TDP-43 is associated with a longerTMEM106B 39 UTR in FTD/ALS postmortem brain samples (Fig. 3e). We also analyzed the TMEM106B 39UTR in a series of 83 frontal cortex brain samples from the Mayo Clinic Brain Bank using qRT-PCR and found a significant increase in longerTMEM106B 39 UTRs in the frontal cortices of patients with FTLD-TDP compared with healthy controls (Fig. 3f), indicating that the increased 39 UTR length ofTMEM106B might be functionally relevant in FTD/ALS.

Because TMEM106B protein levels are altered in disease50,55, we next asked whether TDP-43 knockdown affected TMEM106B protein levels. Using a condition that has been shown previously to perserve TMEM106B dimers on SDS-PAGE55 (see also Fig. S3i), we found that TDP-43 knockdown did not affect levels of TMEM106B monomers (~42 KDa) but reduced dimer levels (Fig. 3g). A recent proteomics study also detected decreased TMEM106B levels caused by TDP-43 knockdown12. TDP-43 knockdown only modestly reduced TMEM106B RNA levels (Fig. S3j), suggesting that the longer 39 UTR might affect TMEM106B protein levels via translation. To test this hypothesis, we cloned short and long versions of the TMEM106B 39 UTR into a dual-luciferase reporter. We mutated the proximal polyA site in the long 39 UTR-containing reporter to prevent its usage and confirmed the production of the intended long 39 UTRF(ig. S3k). The long 39 UTRcontaining reporter had significantly lower luciferase activity and higher RNA levels thanhte short 39 UTRcontaining reporter (Fig. 3h), providing evidence that the loss of TDP-43-induced longerTMEM106B 39 UTR reduces translation efficiency. But how could this reduce the formation of TMEM106B dimersT?MEM106B APA might impact dimer levels by influencing the subcellular localization ofTMEM106B mRNA and its ability as a scaffold for protein-protein interactions, as has been shown for other APA even1t7s. Recent studies discovered an Alu element insertion in the TMEM106B 39 UTR, in perfect linkage with the top FTLD-TDP risk allele at this locus56358. Given the proximity of these variants toTMEM106B9s distal polyA site, future work will explore if and how these disease-associated genetic variants impact APA.

In the present study, we applied 39 end-seq to comprehensively map neuronal polyadenylation on a transcriptomic scale with base-pair resolution and discovered that TDP-43 dysfunction causeswidespread APA changes, some of which are in genes critical for neuronal function and in disease-associated gene.sThese changes are highlighted by coupled cryptic splicing and premature polyadenylation inARHGAP32 (Fig. 2g), changes observed in multiple subunits of the elongator complex (Fig. 3a, S3c, S3d), and 39 UTR lengthening in NEFL, SFPQ, and TMEM106B (Fig. 3b-d). This study is accompanied by two independent manuscripts also presenting widespread APA changes associated with TDP-43 dysfunction in FTD/ALS (Fratta and colleagues35; La Spada and colleagues29). Despite using different approaches, each of the three studies observed common APA changes (Fig. 2g, S2d-f, S2k-m), underscoring the importance of APA changes during neurodegeneration.

Our application of 39 end-seq complements and extends RNA-seq based APA analysis because it enables de novo identification of polyadenylation events with base-pair resolution, revealing complex and highly esnsitive APA changes (e.g., NEFL, ELP3, and TMEM106B). Together, the findings of these three studies provide evidence that TDP-43 pathology contributes to disease pathogenesis through not only cryptic splicingbut now also changes in alternative polyadenylation. level changes on the y-axis, illustrating that APA changes could alter RNA levels. Genes with significant APA changes are plotted; genes with significant RNA level changes are labeled in red when favoring distal polyA sitesor blue when favoring proximal polyA sites.e, Cumulative plots show that for TDP-43 regulated APA sites, distal polyA sites are stronger than proximal polyA sites. f, Cumulative plots show that for genes with signifciant 39 UTR lengthening upon TDP-43 KD, their distal polyA sites are stronger than their proximal ones. PolyA site scores (in panels e and f) were calculated using Aparent2, expressed as log odds ratio, and reflect polyA site strengths.g, TDP-43 KD activated cryptic splicing and premature polyadenylation in intron 12 ofARHGAP32. The <polyA DB= track marks annotated polyA sites 32 and the <TDP-43 sites= track marks the observed TDP-43 binding . Gene structure is shown on the bottom. 59 43 knockdown (KD) samples were used. b, Pie chart shows the distribution of 39 end-seq identified polyA sites. Note that <intergenic= represents polyA sites not associated with annotated 39 UTRs, exons, or introns. c-f, Examples of the use of an unannotated distal polyA site upon TDP-43 KD. g-j, Examples of premature polyadenylation upon TDP-43KD. k, Example of complex usage change of multiple polyA istes upon TDP-43 KD. l. Example of increased proximal polyA usage upon TDP-43 KD. m, Example of 39 UTR shortening upon TDP-43 KD. Details, as in Fig. 2g.

TMEM106B (h), confirmed by qRT-PCR.i, WB shows that as the temperature increases, a ~75 KDa band collapses to a 42 KDa band. Cell lysates were incubated at respective temperatures for 10 min before electrophoresis.j, Bar plots show that TDP-43 KD modestly reduced TMEM106B RNA levels. The adjusted p value was calculated in DEseq2. k, Bar plots show that the reporter with the long TMEM106B 39 UTR produced the long 39 UTR, as designed. RNA levels were measured by qRT-PCR. Unless stated otherwise, p-values were calculated by Student9s t test. ns (not significant),p > 0.05; *, p f 0.05; **, p f 0.01; ***, p f 0.001; ****, p f 0.0001.

Table S1: Alternative polyadenylation changes detected in Liu et al. 2019 Table S2: Significant APA changes in 39 UTRs upon TDP-43 knockdown detected by RNA-seq Table S3: Significant APA changes upon TDP-43 knockdown detected by 39 end-seq Table S7: Summary of patient data

We thank members of the Gitler lab and the Petrucelli lab for helpful discussions and commentosn the manuscript. We thank Abigail Song and Maylin Fu for help with experiments. We thank Ziwei Chen for hleping with Aparent2. Y.Z. is supported by a postdoctoral scholar award from The Phil and Penny Knight Initiative for Brain Resilience at the Wu Tsai Neurosciences Institute, Stanford University, and afellowship grant from the Larry L. Hillblom foundation. T.A. is supported by NIH (2T32AG047126-06A1) and a fellowship from the Takeda Science Foundation. C.G. is supported by Milton Safenowitz Postdoctoral Fellowship Program. Work in A.D.G. is supported by NIH (grants R35NS097263, U54NS123743, R01AG064690) and Target ALS. A.D.G. is a Chan Zuckerberg Biohub 3 San Francisco Investigator. Work in L.P. is supported by NIH (U54NS123743, R35NS097273, P01NS084974) and Target ALS. Work in M.P. is supported by NIH (U54NS123743, RF1NS120992) and Target ALS. Nat Neurosci 14, 4523458 (2011). from loss of TDP-43.Nat Neurosci 14, 4593468 (2011).

The sequencing data generated in this study will be available at GEO. The data and code supporting the findings of this study are available from the corresponding authors upon reasonable request. The following publicly available data are used in this study: GSE126542, GSE121569, GSE147544, GSE196144, and ERP126666.