Congenic expression of poly-GA but not poly-PR in mice triggers selective neuron loss and interferon responses found in C9orf72 ALS

We generated plasmids for conditional expression of DPRs by intersting GFP-GA₁₇₅ and GFP-PR₁₇₅ genes (encoded using non-repeating alternate codons) downstream of a floxed stop cassette (encoding a puromycin resistance gene followed by SV40 poly-adenylation signal) in the pEX CAG stop-bpA vector and electroporated them together with a PhiC31-Integrase expression plasmid into murine RMCE (recombinase-mediated cassette exchange) embryonic stem cells at the Rosa26 safe harbor [22]. The Rosa26 gene itself was knocked out by the neomycin-bpA gene trap during this process. Targeted integration was screened by neomycin selection and Southern blot. Clones with Cre-inducible DPR expression were injected into blastocytes. We identified mouse lines GAstop and PRstop with germ-line transmission and speed backcrossed them to C57BL/6N background until > 98% purity was confirmed using SNP genotyping (Charles River). Crossing GAstop and PRstop mice with Nestin-Cre [61] and subsequent excision of the stop cassette resulted in DPR expression throughout CNS neurons (ROSA26 GFP-GA₁₇₅^+/−; Nestin-Cre and ROSA26 GFP-PR₁₇₅^+/−; Nestin-Cre). Relevant endpoint criteria for euthanasia according to national regulations were weight loss before 10 weeks of age or stage II seizures consisting of sustained cramping causing the mouse to fall on its side or back. Mice were killed by brief CO₂ exposure, followed by decapitation. Tissue was either stored at − 80 °C for biochemistry analysis or formalin fixed for 24 h for immunohistochemistry analysis. For spinal cord tissue, decalcification using 5% formic acid was performed for 48 h.

Animal handling and animal experiments in this study were performed in accordance with the German animal welfare law and approved by the Government of Upper Bavaria, Germany. Mice were housed in standard cages in a pathogen-free facility in a 12-h light/dark cycle with ad libitum access to food and water.

14-Month-old animals were acclimated to the experimenter for 1 week before training by placing mice on the back of the hand and allowing them to explore for 1 min. Training was performed on Monday, Wednesday and Friday, and animals were acclimated to the test room for 15 min before beginning. The apparatus was cleaned with 70% EtOH and allowed to dry before each run. Mice were placed on the rotarod (RotaRod, TSE Systems) moving at 4 rpm for three 1-min runs per day, each followed by a 10-min break. If mice fell off the rod, they waited at the bottom of the apparatus for the full minute. The following week, mice were tested for their latency to fall off the rotarod as it increased speed up to 60 rpm over 1 min. Each mouse again performed three testing trials each followed by 10-min breaks.

Immunohistochemistry and immunofluorescence was performed on 5-µm paraffin-embedded tissue sections as described previously [37] using the following antibodies: Alyref (ab6141, Abcam 1:50), ChAT (AB144P, Millipore 1:200), GFP (632381, Takara 1:250), Fus (A300-292A, Bethyl Labs 1:100), Hnrnpa1 (sc-32301, Santa Cruz 1:50), Hnrnph (A300-511A, Bethyl Labs 1:500), Iba1 (091-19741, Wako, 1:500 and ab5076, Abcam, 1:500), Map2 (M1406, Sigma 1:250), NeuN (ab177487, Abcam, 1:1000), nucleolin (ab50279, Abcam 1:30), p62 (PM045, MBL 1:1000), poly-GA (1A12, mouse, purified and biotinylated, 1:500) [37], poly-PR (32B3, mouse purified and biotinylated, Helmholtz Zentrum 1:1000) [58], Psmc4 (A303-850A, Bethyl Labs 1:250), Alexa Fluor 594 anti-Sfpq (A301-321A, Bethyl Labs 1:100), TDP-43N-terminal (10782-2-AP, Proteintech 1:500), TDP-43 p409/410 (1D3, gift from Manuela Neumann 1:500), TDP-43 p409/410 (CAC-TIP-PTD-P02, Cosmo 1:500), TDP-43 p409/410 (CAC-TIP-PTD-M01, Cosmo 1:1000), TDP43 p409/410 (3655, gift from Leonard Petrucelli 1:500), TDP-43 pS403/404 (CAC-TIP-PTD-P05, Cosmo 1:500), TDP-43 pT153/Y155 (gift from Yuna Ayala 1:300), TDP-43 C-terminal (12892-1-AP, Proteintech 1:100), TDP-43 non-phospho 404–413 (TAR5 6D6, gift from Manuela Neumann 1:5), TDP-43 non-phospho 404–413 (TAR5 2H4, gift from Manuela Neumann 1:5), TDP-43 N-terminal (10782-2-AP, Proteintech 1:500), TDP-43 p409/410 (1D3, gift from Manuela Neumann 1:500), TDP-43 p409/410 (CAC-TIP-PTD-P02, Cosmo 1:500), TDP-43 p409/410 (CAC-TIP-PTD-M01, Cosmo 1:1000), TDP43 p409/410 (3655, gift from Leonard Petrucelli 1:500), TDP-43 pS403/404 (CAC-TIP-PTD-P05, Cosmo 1:500), TDP-43 pT153/Y155 (gift from Yuna Ayala 1:300), TDP-43 C-terminal (12892-1-AP, Proteintech 1:100), TDP-43 non-phospho 404–413 (TAR5 6D6, gift from Manuela Neumann 1:5), TDP-43 non-phospho 404–413 (TAR5 2H4, gift from Manuela Neumann 1:5), Tubulin β 3 (801208, BioLegend 1:200). Nucleolin immunofluorescence required 1–5 min treatment with 50 µg/mL protease K prior to citrate antigen retrieval. Alexa Fluor 555 Mouse anti-β-Tubulin III (BD Pharmingen 560,339 1:200) and Bungarotoxin Alexa Fluor 647 (Thermo Fisher B35450 1:500) were used to stain motor axons and acetyl-choline receptors at the NMJ, respectively.

Automated quantification of immunofluorescence was performed using CellProfiler (v. 3.0.0) software. For poly-PR quantification, images from sagittal sections approximately 100 µm from the midline were taken covering the entire cortex of each animal. DAPI-stained nuclei and poly-PR aggregates were identified and counted, and the % of neurons with poly-PR inclusions was reported for each animal. For neuron quantification in brain and spinal cord, 10 × scanning images were taken of the entire region on each slide (cortex, hippocampus, or spinal cord). NeuN-positive cells were identified using CellProfiler and divided by the area of the tissue section measured. Because the loss of NeuN can occur in intact neuronal nuclei [65], we also verified that there were no large nuclei with neuronal morphology in the CA fields of the hippocampus stained with DAPI but not NeuN. We used NeuN staining only to assist with semi-automated detection of neurons vs glia in our immunofluorescence staining. Choline acetyltransferase (ChAT)-positive cells with motor neuron morphology in the anterior horn were quantified manually on sections at 1-mm intervals spanning the whole spinal cord. For quantification of the denervated neuromuscular junctions (NMJs), confocal image stacks were recorded using an LSM810 confocal laser scanning system (Carl Zeiss) with 25 × water immersion objective. Innervated or denervated single NMJs in the triangularis sterni and quadriceps muscles were classified by an investigator blinded to genotype.

Blood samples were collected by final blood withdrawal under isoflurane anesthesia by puncturing the retrobulbar vein plexus with a 1-mm-diameter glass capillary (Hirschmann). Blood was collected in EDTA-coated sample tubes (Kabe Labortechnik) and an amount of 50 µl determined by an end-to-end capillary (Kabe Labortechnik) was diluted 1:5 with Cell Pack buffer (Sysmex) and stored at room temperature until analysis. To evaluate peripheral blood cell counts including automated differential white blood counts samples were mixed thoroughly by inverting the tube 3–4 times and analyzed using a Sysmex XT 2000iV hematology analyzer in the “capillary mode” with the predefined mouse settings within 6 h after sample collection and dilution.

Tissues were homogenized in Trizol using homogenizer (Precellys). Total RNA was isolated using Direct-zol-96 RNA isolation kit (ZymoResearch) following the manufacturer’s instructions. RNA quantity and quality were controlled on Agilent 2100 BioAnalyzer using Agilent RNA 6000 Nano Kit.

Reverse transcription was performed at 42 °C for 90 min with a reaction consisting of 25 units Maxima H- enzyme (Thermo Fisher), 2 × Maxima H-Buffer (Thermo Fisher), 2 mM each dNTPs (Thermo Fisher), 4 µM template-switching oligo-(IDT), 4 µM barcoded oligo-dT primers (IDT), and 10 ng of RNA, in a final reaction volume of 10 µL. Following RT, the cDNA was pooled and cleaned using SPRI beads, and then eluted with 17 µL of DNase/RNase-Free Distilled Water (ThermoFisher). Residual primers were digested with Exonuclease I (Thermo Fisher) at 37 °C for 20 min, with a subsequent heat inactivation step at 80 °C for 10 min. Pre-amplification was then performed by adding 25 µL of KAPA HiFi HotStart ReadyMix (2x, KAPABIOSYSTEMS) and 0.33 µM SINGV6 primer (IDT) for a final reaction volume of 50 µL. The PCR was cycled as follows: 98 °C for 3 min for initial denaturation followed by 11 cycles of 98 °C for 15 s, 65 °C for 30 s, 68 °C for 4 min. Final elongation was performed at 72 °C for 10 min.

Following pre-amplification, the sample was cleaned using SPRI beads (1:0.8) and then eluted with 10 µL of DNase/RNase-free distilled water (ThermoFisher). The cDNA was quantified using the Quant-iT PicoGreen dsDNA Assay Kit (Thermo Fisher) and size distributions were checked on high-sensitivity DNA chips (Agilent Bioanalyzer). After passing the quantity and quality controls, 0.8 ng of pre-amplified cDNA was used to construct Nextera XT libraries (Illumina). The library PCR used custom P5 primer (P5NEXTPT5, IDT) to enrich for the 3′ ends. The library was then size selected using a 2% Agarose E-gel (Life Technologies) from 300–800 bp.

qPCR was performed as described previously [69]. qRT-PCR was performed using CFX384 Real‐Time System (Bio‐Rad) with Taqman technology. For GFP expression in the brain, primers and probes were designed (IDT) for the GFP region. Primer 1: GCACAAGCTGGAGTACAACTA, Primer 2: TGTTGTGGCGGATCTTGAA, Probe/56FAM/AGCAGAAGA/ZEN/ACGGCATCAAGGTGA/3IABkFQ/. Signals were normalized to GAPDH levels according to ΔΔCT method and levels are shown relative to control. For GFP expression in the muscle, Taqman probes (ThermoFisher) were used against GFP (Mr04097229_mr) and Atp5a1 (Mm00431960_m1) with Sybr fluor. GFP signals were normalized to Atp5a according to the ΔΔCT method and levels are shown relative to control.

Final libraries were paired-end sequenced on a high output flow cell of a HiSeq 1500 (Illumina), with the following sequencing setup: 16 bases for the cellular barcode and UMI, eight bases for the i7 barcode, and 50 bases for the cDNA read. Libraries were sequenced to a depth of ~ 10 million reads per sample. Raw fastq files were checked using FastQC (v 0.10.1) and then processed using zUMIs (v 0.0.6) with default settings and a provided barcode list. The data were mapped to the mouse genome (mm10) and gene annotations were obtained from Ensembl (GRCh38.89).

The count tables (from tissue mean counts 2.4 M, range 0.225–28 M; from isolated microglia average counts 250 K, range 153–678 K; samples with less than 100 K counts were excluded) were analyzed using the DESeq2 pipeline in R comparing transgenic mice and littermate controls for each region and age separately [36]. The genotype of all mice was confirmed (and corrected) using the GFP-(GA)₁₇₅ and GFP-(PR)₁₇₅ transcript. The number of samples per group was 3 weeks (7 control vs 3 GA-Nes), end-stage (5 control vs 5 GA-Nes), isolated microglia at end-stage (3 control vs 3 GA-Nes), affected (5 control vs 5 PR-Nes), and asymptomatic (3 control vs 5 PR-Nes). All significant hits (p < 0.05 adjusted for multiple testing for each pairwise comparison) are listed in Table S2. All publicly available RNAseq data from TargetALS data were accessed in February 2020. We included all samples from patients with ALS and controls without CNS co-morbidity and with C9orf72-genotype information available. Reads were aligned to GRCh38 using STAR aligner b2.4.2a [12] by the New York Genome Center and further processed by us using SummarizeOverlaps in R. Count tables were analyzed using DESeq2 [36] in a linear model accounting for gender as a confounder in the cerebellum (14 controls, 108 ALS, and 30 C9orf72 ALS) and frontal cortex (16 controls, 90 ALS, and 26 C9orf72 ALS), and differential gene expression was analyzed between the relevant groups. A model accounting for gender and anatomical sub-regions was used for spinal cord (cervical: 13 controls, 86 ALS, 25 C9orf72 ALS; thoracic: eight controls, 44 ALS, 7 C9orf72; lumbar: 11 controls, 77 ALS, 19 C9orf72 ALS) and motor cortex (lateral: 13 controls, 77 ALS, 21 C9orf72 ALS; medial: 15 control, 78 ALS and 18 C9orf72 ALS). The overall LFC for the groups was used for the correlation analysis.

Gene ontology analysis (biological process) was done using the clusterProfiler package in R. Only groups with more than five significant genes were analyzed. Figures show a manual selection of the most significant and least overlapping categories to best represent the key pathways. The full list of significant GO terms is shown in Tables S3 and S5. Concordantly regulated genes were defined by comparing averaged significant up-/down-regulation from all analyzed regions in GA-Nes or PR-Nes mice and ALS patients.

Kmeans clustering was done using default parameters with pheatmap (1.0.12) in R. Pearson’s correlation of gene expression was done using pairwise-complete observations with identical Entrez gene names. We used the log2-fold expression changes of significant genes from the AD and SOD1 mouse models (padj < 0.05) [26], and the p25 mouse model in reference [40] (cZ > 3.0902). For the microglia clusters, we used all genes listed in the supplemental data of reference [19]. The actual number of genes used is indicated in the figure legends.

Statistical analyses other than for RNAseq data (described separately above) were performed using GraphPad Prism 7.02. We used two-tailed t tests to compare two groups, and ANOVAs with Greenhouse–Geisser correction to compare more than two groups. Data are reported in the figure legends with (mean ± standard deviation) and effect size (η², the between-group sum-of-squares divided by the total sum-of-squares). Confidence intervals are reported instead of effect size for non-parametric tests. Data distributions were checked for normality by the Shapiro–Wilk test and homogeneity of variances was checked by the F test or the Brown–Forsythe test. When these were violated, non-parametric tests were used (Mann–Whitney). Individual points for immunoblots represent the average of at least two independent replicates per biological sample. Power analyses were performed with GPower 3.1.

To assess only the contribution of the DPR proteins in the absence of G₄C₂ RNA toxicity, we used non-repeating alternate codons to individually express poly-GA and poly-PR in the absence of (G₄C₂)_n repeats as in previous models [41, 57]. We generated conditional single transgenic DPR expressing mice by targeted insertion of a GFP-GA₁₇₅ or GFP-PR₁₇₅ expression construct driven by the strong CAG promoter with a LoxP-flanked stop cassette at the Rosa26 safe harbor locus (Fig. S1a). To allow CNS-wide neuronal expression we crossed these mice with the well-characterized Nestin-Cre driver line (Fig. S1a, b) [61] creating double transgenic mice (GA-Nes and PR-Nes). GA-Nes and PR-Nes mice were born at Mendelian ratios (23.5% χ² = 2.327, df = 3, p = 0.5075n ≥ 44/group and 23.5% χ² = 6.647, df = 3, p = 0.084n ≥ 24/group, respectively).

We observed progressive motor deficits and early lethality in all GA-Nes mice and a subset of PR-Nes mice. All GA-Nes mice showed abnormal gait and impaired motor function from 5 weeks (S. video 1), and by 6–7 weeks mice had severe muscle weakness but not paralysis (S. video 2). GA-Nes mice were smaller than control littermates beginning around 4 weeks of age, and later even began to lose weight, at which time we were required to euthanize them in accordance with national animal welfare laws. All GA-Nes mice reached this end-stage point before 7 weeks, resulting in a sharp drop in the Kaplan–Meier curve (Fig. 1a), and providing food on the floor of the cage did not improve survival. In contrast, about 40% of PR-Nes mice also failed to thrive and around 4–5 weeks of age developed seizures that were criteria for euthanasia (Fig. 1b). These mice also showed poor coordination reminiscent of ataxia and fell frequently to the side without apparent muscle weakness (Supp. video 3). The remaining PR-Nes mice that did not meet endpoint criteria early on did not develop seizures later in life, gained weight normally and had normal motor function until the end of the study at 68 weeks (Supp. video 4). The rotarod assay revealed no significant motor deficits in these mice even at 14 months (Fig. 1d). Due to these radically different phenotypes, also seen to a lesser extent upon viral poly-PR expression [68], we separated PR-Nes mice into affected (4 weeks) and asymptomatic (4 weeks and 68 weeks) groups for the remaining studies. qPCR analysis showed similar transgene expression at the mRNA levels in the cortex of GA-Nes mice and affected PR-Nes mice, but significantly reduced transgene expression in asymptomatic PR-Nes mice (Fig. 1c). Genomic PCR confirmed that asymptomatic PR-Nes still maintained the 175-repeat length suggesting transcriptional repression of the synthetic construct (Fig. S1c).

To elucidate the differential phenotypes in GA-Nes and PR-Nes mice, we first compared DPR expression patterns in both lines using immunostaining. As expected, in the absence of Cre expression (GAstop), no poly-GA expression was detected (Fig. 1e). Consistent with the pan-neuronal expression of the Nestin-Cre driver, poly-GA formed frequent neuronal cytoplasmic inclusions (NCI) and neuritic aggregates in the vast majority of neurons throughout the CNS of GA-Nes mice (Fig. 1e). Combined fluorescent in situ hybridization (FISH) to label the transgene mRNA and immunostaining of Cnp1, GFAP and Iba1 revealed transgene expression in a subset of oligodendrocytes and astrocytes, but not in microglia (Fig. S1d–f). Importantly, despite similar mRNA levels in GA-Nes and affected PR-Nes mice, poly-PR inclusions were infrequent in the cortex and brainstem (Fig. 1f) and not found in the spinal cord. However, in the hippocampus, most neurons had at least one nuclear inclusion in affected PR-Nes mice, while far fewer hippocampal neurons contained inclusions in asymptomatic PR-Nes mice (68 weeks) (Fig. 1f). Quantification of poly-PR in the cortex showed that inclusions in asymptomatic PR-Nes mice were already reduced at 4 weeks, and this persisted in asymptomatics at 68 weeks (Fig. 1g). By immunoblotting with a GFP antibody, poly-PR (aggregated or soluble) was below the detection level even in affected mice, in contrast to abundant aggregated poly-GA visible in the gel pocket (Fig. S1g). Poly-PR aggregates were barely detectable when 25 times more brain extracts than necessary for poly-GA detection were loaded on a filter trap assay (Fig. S1h). Our findings are consistent with the abundant poly-GA and scarce poly-PR inclusion pathology observed in C9orf72 patients [38, 58], though both are more frequent than observed in patients. Overall, the comparison between these congenic lines shows consistent progressive weakness in poly-GA mice that more closely resembles ALS than the ataxia and seizures found in poly-PR mice, a subset of which escape toxicity completely by reducing transgene expression.

To address the consequences of DPR expression and identify correlates of the motor deficits, we analyzed regional neuron loss and early signs of neurodegeneration and microglial activation in our mouse models. Despite widespread aggregation of poly-GA in the vast majority of CNS neurons, robust microglial infiltration accompanied by mild neuron loss first became apparent in the CA2 of the hippocampus after 3 weeks of age, with nearly complete ablation of the CA2 and thinning of the other CA fields by end-stage (Fig. 2a, b; Fig. S2a). The CA2 is delineated by the end of the compactly layered neurons of the CA1 and the beginning of the larger and more diffusely spaced pyramidal neurons of the CA3. In contrast, both affected and asymptomatic PR-Nes mice showed normal neuron density (Fig. 2c, f, Fig. S2b) and no microglia activation in the hippocampus (Fig. 2d), despite high poly-PR aggregation there (compare Fig. 1f). Surprisingly, infiltration of activated microglia with typical amoeboid morphology in the CA2 region of GA-Nes mice was stronger at 3 weeks of age than at end-stage (Fig. 2b). Quantification at end-stage confirmed significant neuron loss in the whole CA field, even when normalizing to the area of the region in the overall smaller brains (Fig. 2e, Fig. S2a). In contrast, neuron loss was not detected in the cortex of GA-Nes (Fig. 2e, Fig. S2a) or affected PR-Nes mice (Fig. 2f, Fig. S2b).

To address known toxicity mechanisms, we analyzed the subcellular distribution of poly-GA and poly-PR in our mice. Poly-GA co-aggregated with p62/Sqstm1 (Fig. 2g) and the proteasome subunit Psmc4 (Fig. 2h), consistent with cryo-electron tomography data in primary neurons [18]. In PR-Nes mice, poly-PR was restricted to the nucleus and colocalized with heterochromatin (Fig. 2i) and the nucleolus (Fig. 2j) in both affected and asymptomatic mice, as reported in other models [21, 68]. However, despite subcellular localization of both DPRs matching that of other mouse models, only poly-GA led to detectable neuron loss in the hippocampus.

To investigate the muscle weakness in GA-Nes mice, we also examined spinal cord neurons and neuromuscular junctions. While ChAT-positive motor neurons in the anterior horn of the spinal cord were normally abundant in GA-Nes mice at 3 weeks, they were significantly lost by end-stage (Fig. 3a, b). In contrast, the number of calbindin-positive neurons in the dorsal horn was unchanged even at end-stage (Fig. S2c, d), suggesting selective vulnerability of motor neurons in the spinal cord to poly-GA. Consistent with the lack of muscle wasting in PR-Nes mice, no loss of ChAT-positive neurons in the spinal cord was observed (Fig. 3c, d). We further investigated axonal innervation of neuromuscular junctions in GA-Nes mice. While neuromuscular junctions were fully innervated at 3 weeks in GA-Nes mice, by end-stage most α-bungarotoxin-labeled endplates were denervated, as indicated by loss of β-tubulin III immunostaining in both triangularis sterni (whole-mount preparation [4, 27]) and quadriceps muscles (vibratome sections) (Fig. 3e, f). In contrast, no denervation was observed in the triangularis sterni muscle of PR-Nes affected or asymptomatic mice (Fig. 3g).

Despite the clear correlation between motoneuron loss and weakness, we investigated possible peripheral causes of muscle wasting requiring early euthanasia. Hematological parameters, including white blood cell counts, were not significantly different between GA-Nes and controls (Fig. S3a). Systematic macroscopic and microscopic analyses were conducted in the intestines, pancreas, liver, kidney, adrenal gland, spleen, heart, lung, and thymus. Consistent with the known expression profile of Nestin-Cre [11, 13], GFP staining revealed sparse expression of poly-GA in the kidney and pancreas, as well as in neurons of the myenteric plexus (Fig. S4a). However, H + E staining revealed no signs of necrosis or gross pathological abnormality in these or other organs compared to control littermates, other than smaller size consistent with the overall smaller size of GA-Nes mice (Fig. S4b). In addition, the examination of the intact digestive tract at euthanasia showed processed boli present at regular intervals throughout the entire digestive tract in both GA-Nes and controls arguing against impaired peristalsis or intestinal transit (Fig. S4c). Poly-GA aggregates were also found commonly in the quadriceps muscle (Fig. S4d), again consistent with known expression of Nestin-Cre [13] and with the muscle wasting and denervation observed. FISH and qPCR analysis of the quadriceps confirmed that poly-GA was expressed locally in skeletal muscle (Fig. S4e, f), which is particularly interesting given recent observations that poly-GA is commonly found there in human C9orf72 ALS cases [9]. Thus, selective loss of motoneurons and neuromuscular denervation together with potential cell-autonomous effects in the muscle are likely the predominant drivers of progressive muscle wasting in GA-Nes mice.

To elucidate the elusive link between DPR and TDP-43 pathology, we characterized TDP-43 localization in GA-Nes and PR-Nes mice in detail using a panel of antibodies (Fig S5a). In the cortex of control mice, TDP-43 staining with an N-terminal antibody showed the typical nuclear network-like pattern (Fig. 4a). TDP-43 expression seemed more diffuse in GA-Nes mice and to a lesser extent also in PR-Nes mice, but we did not observe nuclear clearance or cytoplasmic TDP-43 aggregates in GA-Nes and PR-Nes lines and did not observe additional cleavage products in the RIPA-insoluble fraction of GA-Nes mice by Western blot (Fig S5b). However, we noticed occasional nuclear TDP-43 aggregates (~ 1% of neurons) specifically in the lateral frontal cortex and hippocampal CA regions of GA-Nes mice (Fig. 4a–c). In contrast, even in neurons with numerous nuclear poly-PR aggregates, no nuclear or cytoplasmic inclusions of TDP-43 were observed in asymptomatic or affected PR-Nes mice (Fig. 4a). We detected these nuclear TDP-43 inclusions in GA-Nes mice with several N- and C-terminal antibodies, and comparison of C-terminal antibodies specific for non-phospho- or phospho-TDP-43 showed that the nuclear aggregates are not phosphorylated at serine 403/404 or serine 409/410 (Fig. S5a).

To further characterize the TDP-43 nuclear pathology in GA-Nes mice, we investigated co-aggregation of other RNA-binding proteins implicated in disease, as well as proteins involved in general RNA trafficking and splicing functions. Unexpectedly, immunostaining revealed strong colocalization of nuclear TDP-43 aggregates with hnRNP H, which has been linked to RNA toxicity in C9orf72 cases, and weak colocalization with hnRNP A1 previously implicated in ALS/FTD [30, 33] (Fig. 4d). Absence of colocalization with the RNA-binding proteins FUS, SFPQ or ALYREF demonstrates the specificity of the interaction (Fig. 4d). Thus, only poly-GA promotes sparse nuclear aggregation of TDP-43 and other hnRNPs implicated in C9orf72 disease in a region-specific manner.

To allow unbiased investigation of disease-related CNS pathomechanisms in our mouse models, we performed RNA sequencing of bulk tissue. For GA-Nes mice, we compared brainstem, cortex, hippocampus and spinal cord at end-stage and at 3 weeks (for the brain regions). For PR-Nes, we analyzed the cortex and hippocampus of 4-week affected and 68-week-old asymptomatic mice. The full expression data are listed in Table S1.

Principal component analysis showed strong separation of the genotypes in different brain regions in GA-Nes end-stage and affected PR-Nes mice, but not in GA-Nes mice at 3 weeks of age, which further excludes developmental effects due to poly-GA expression (Fig. S6a). Correlation analysis revealed overall similar gene expression changes in the different brain regions of GA-Nes mice particularly at the end-stage time point (Fig. S6b). We analyzed differential gene expression for all regions and time points with DESeq2. In total 2351 genes were significantly up- or down-regulated in GA-Nes mice, compared with only 123 genes in PR-Nes mice and 53 in both using our cut-off criteria (adjusted p < 0.05, see Table S2). The genes showing the strongest differential expression in GA-Nes and PR-Nes are shown in Fig. 5a. We detected the strongest changes in the GA-Nes hippocampus, in line with the early pronounced microglial activation and neuron loss visible in this area (compare Fig. 2f). The strongest change in PR-Nes mice is the reduced expression of three 5S rRNA transcripts in the cortex of both affected and asymptomatic mice, which corroborates the known effects of poly-PR in inhibiting protein translation [21]. Among the top differentially expressed genes, only Ccl2 and Cxcl10 were significantly different in both GA-Nes and PR-Nes mice. Gene expression changes in asymptomatic PR-Nes mice are minimal and revealed no obvious mechanistic basis for the transgene repression in asymptomatic compared to affected mice (Table S1).

Among down-regulated transcripts, gene ontology (GO) analysis identified enrichment of genes encoding synaptic proteins (e.g., neurotransmitter receptor subunits and proteins involved in vesicle release) as well as immediate early genes associated with learning and memory in GA-Nes cortex and hippocampus at end-stage, but not in PR-Nes tissues (Fig. 5b, c). Comparison of significant hits (bold), related subunits of GABA, NMDA and dopamine receptors and other key synaptic genes suggests altered synaptic function in GA-Nes mice may result in reduced activation of immediate early genes such as Arc, which is critical for learning and memory [42] (Fig. 5c). Analysis of protein levels in neocortex by western blot showed that the immediate early gene Arc is indeed significantly reduced in the end-stage cortex of GA-Nes mice, but levels of key neuronal proteins β-tubulin-III and PSD-95 were not reduced (Fig. 5d, e). These findings are consistent with the absence of neuron loss in this region by the time of euthanasia (see Fig. 2i). Myelination-associated genes (Klk6, Mal, Pllp, Plp1, Fa2h, Hes5, Aspa) were also downregulated in GA-Nes and PR-Nes mice, likely due to DPR expression in a subset of oligodendrocytes (see Fig. S1f), which has so far not been reported in C9orf72 patients.

Overall, up-regulated transcripts in GA-Nes mice were strongly enriched for immune responses, including cytokine- and chemokine-mediated signaling and interferon-inducible genes such as Ifi205/207 and chemokines in all regions in end-stage GA-Nes but not (or much less) PR-Nes mice (Fig. 5a, Table S3). Thus, poly-GA expression induces pronounced neuroinflammation and may also impair synaptic function.

To examine the relevance of these findings to human disease, we correlated gene expression changes in GA-Nes mice with large public transcriptome data of spinal cord, motor cortex and cerebellum from ALS patients (TargetALS). All regions showed clear expression changes in ALS cases (see full expression data in Table S4). Surprisingly, gene expression changes in all GA-Nes regions correlated better with spinal cord than motor cortex from both C9orf72 and non-C9orf72 ALS patients, and inversely correlated with cerebellum and frontal cortex (Fig. S6c). In contrast, PR-Nes mice showed only weak correlations with human ALS tissue. Comparing gene expression in C9orf72-positive and C9orf72-negative cases (C9vsALS), confirmed the upregulation of homeobox genes in the cerebellum (e.g., HOXA5 [15]) and showed altered expression of genes related to the extracellular matrix in the spinal cord of C9orf72 cases (Table S5 and Fig. 6a). GO analysis of genes concordantly regulated in GA-Nes mice and the C9vsALS patient data also showed enrichment of interferon signaling (e.g., IRF7 and IRF9) and defense response genes (e.g., GRN), while no GO terms were concordantly enriched in both PR-Nes and C9vsALS using our criteria (Fig. 6a). The network of these immune-related genes concordantly upregulated in GA-Nes mice and C9vsALS brains is shown in Fig. 6b.

To further verify whether the pro-inflammatory responses were driven by microglia, we performed RNA sequencing on isolated microglia from end-stage GA-Nes brain. Transgene-specific sequence tags (GA175) further confirm that poly-GA is not expressed in microglia (Fig. S6d). The top differentially expressed genes in GA-Nes microglia include many interferon-stimulated genes (e.g., MX1, Isg15, Oasl1/2 and various Ifi genes) and complement factors (C3 and C4b) (Fig. S6d). Therefore, we further compared our data to different microglial subtypes recently identified by single-cell transcriptome studies of microglia under physiological and pathological conditions [19, 40]. The strongest positive correlations were with cluster 6 from a p25 neurodegeneration model [40] and in a lysolecithin demyelination model (injury 2) [19], two microglia subpopulations with strong induction of interferon-response genes (Fig. 6c). Comparing this with the bulk tissue sequencing, these signatures also correlate moderately in GA-Nes end-stage tissues, but weakly or not at all in PR-Nes (Fig. 6c). Comparing only the signature genes of several published microglia populations shows the greatest overlap with interferon-response microglia (IRM) in a mouse model of Alzheimer's disease [56] (Fig. S6e). Together, these data indicate that poly-GA expression induces a proinflammatory microglial response dominated by interferon signaling.

We confirmed the interferon response in end-stage GA-Nes mice by immunoblotting for the STAT proteins that act as intermediate effectors of interferon signaling. The upregulation of total STAT1 and STAT3 suggests strong microglial proliferation, and the upregulation of phosphorylated STAT3 relative to total in GA-Nes cortex (Fig. 6d, e) is consistent with an interferon response. Stat3 gene expression was also significantly elevated at end-stage in hippocampus and cortex (Table S2). Measuring IFNγ directly by ELISA showed that IFNγ was significantly elevated in end-stage GA-Nes but not affected PR-Nes mice (Fig. 6f, g). Taken together, our data indicate that poly-GA inclusions trigger interferon responses in microglia that are also enriched in human C9orf72 ALS.

Poly-PR is highly toxic in mammalian cells, yeast and drosophila by interfering with stress granule formation, nucleocytoplasmic transport, heterochromatin organization, nucleolar dynamics and protein translation by affecting the liquid-liquid phase separation of other low-complexity proteins [2, 3, 21, 24, 31, 32, 44, 68]. While poly-PR localized to heterochromatin and the nucleolus in both affected and asymptomatic PR-Nes mice as in other models, we observed no neuron loss in PR-Nes animals. In addition, asymptomatic PR-Nes mice seem to escape lethality by suppressing the expression of the synthetic transgene mRNA. Thus, nuclear poly-PR toxicity is dose dependent, consistent with another recent report [20]. In addition, post-transcriptional suppression mechanisms likely exist in vivo, because most neurons outside the hippocampus in affected PR-Nes mice have no detectable poly-PR aggregation. Indeed, we find reduced expression of 5S rRNA in the cortex but not hippocampus in PR-Nes mice that may contribute to reduced poly-PR protein levels in the cortex, and is consistent with poly-GR/PR-mediated inhibition of translation [21, 25, 35, 66].

In contrast, poly-GA shows modest toxicity in rodent primary neurons by gradually sequestering the proteasome and other proteins [18, 41]. We confirmed sequestration of the proteasome in our GA-Nes mice and occasional nuclear aggregation of TDP-43 and other low-complexity RNA-binding proteins suggests poly-GA impairs proteostasis in vivo, as we showed recently in vitro [18, 29]. Surprisingly, using identical expression systems, poly-GA toxicity and aggregation was more severe in the mouse central nervous system and resulted in a more ALS-like phenotype than poly-PR expression. Proteasomal impairment, indicated by partial sequestration of Psmc4 in GA-Nes, may contribute to the strong accumulation of poly-GA compared to poly-PR despite similar mRNA levels in affected PR-Nes mice, consistent with data from cellular models [41]. Overall, direct comparison of our mouse lines indicates a combination of post-transcriptional mechanisms could explain the predominant aggregation of poly-GA compared to poly-PR seen in C9orf72 patients without invoking RAN translation-mediated effects.

In our models, poly-GA led to selective neuron loss in the hippocampus and spinal cord as well as muscle denervation resembling ALS. In contrast, poly-PR expressed at the same mRNA level led to ataxia and seizures without obvious neuron loss or inflammation in the brain or spinal cord. While these findings confirm toxicity of high nuclear poly-PR levels, they do not replicate ALS phenotypes well. However, since national animal welfare regulations required us to euthanize mice after a single stage II seizure (see “Materials and methods”), we cannot exclude that repeated seizures would lead to overt neuron loss in these mice. More ALS-like toxicity of poly-PR has been observed in AAV-based expression models or with homozygous transgenic expression that drives higher poly-PR levels compared to our single-copy transgenic line, leading to weight loss and early lethality in a subset or all mice, respectively [20, 68]. Both of these poly-PR mouse models also showed ataxia-like phenotypes due to Purkinje cell loss, which was not detected in our PR-Nes model (data not shown), and one line showed loss of lower motor neurons and inflammation within the spinal cord [20]. In contrast, in our PR-Nes line, poly-PR inclusions are enriched in the hippocampus, a common focal point of epilepsy and the area where poly-PR aggregates have been reported to be most abundant in human C9orf72 FTD [38, 58]. However, while seizures and epilepsy have been reported in C9orf72 patients, they are not a common feature [5]. The different phenotypes of low and high-expressing PR-Nes mice further support dose-dependent toxicity of poly-PR. However, even asymptomatic PR-Nes mice have more abundant poly-PR inclusions than C9orf72 patients [38, 58]. While the scarcity of poly-PR inclusions in patients has been widely attributed to its high toxicity in vitro [62, 68], the modest transcriptional changes in PR-Nes mice despite Nestin-Cre-driven expression of the poly-PR transgene throughout the CNS under the control of the strong CAG promoter do not support dominant effects of barely detectable poly-PR aggregates in patients. Nevertheless, we cannot rule out that the predominantly cytoplasmic localization of poly-PR in patients [16, 58] and the longer repeat length could still cause more dramatic effects in patients.

Overall, our poly-GA mice model several key aspects of C9orf72 ALS well, including selective loss of motor neurons in the spinal cord, muscle denervation, and an interferon-response microglial signature concordant with C9orf72 ALS. In addition, the sparse nuclear TDP-43 inclusions only in the forebrain are reminiscent of C9orf72 FTD. Together these data provide in vivo evidence for selective vulnerability towards poly-GA that may explain the disconnect between C9orf72-specific pathology and neurodegeneration in patients. However, it is necessary to understand the three limitations of this model to maximize its usefulness. First, our mice show modest poly-GA expression in some cells in kidney, pancreas and GI tract, which has not been reported in C9orf72 patients [1]. While histopathological analysis nevertheless revealed overall normal organ morphology and poly-GA expression in the neurons of the myenteric plexus of GA-Nes mice does not seem to impair peristalsis (based on observation of the intact GI tract), we cannot rule out more subtle functional deficits. Investigating whether DPR inclusions occur in the enteric nervous system of C9orf72 patients could be worthwhile because cachexia is a common but poorly understood feature of sporadic and C9orf72 ALS that often precedes the diagnosis by many years [52]. Second, although the distribution of DPR inclusions is predominantly neuronal a subset of oligodendrocytes and astrocytes do express transgene RNA. Third, GFP-only control littermates would have been ideal to rule out any effects of GFP alone, but this was not possible with our construct design. However, we are not aware that any of the large number of published mouse lines expressing high levels of GFP in the brain have shown phenotypes similar to our GA-Nes mice. Moreover, the largely non-overlapping phenotypes of our GA-Nes and PR-Nes mice argue against non-specific off-target effects due to the widespread expression of DPR proteins or GFP using the Nestin-Cre driver.

An important aspect of our novel poly-GA mouse model is that all mice reached the predefined endpoints according to national animal welfare regulations before 7 weeks, while abundant data from the neuromuscular junctions, motor neurons, and transcript analysis indicate no developmental defects. This consistent and rapid progression with motoneuron loss, denervation and C9orf72-enriched transcriptional changes are unique features of this C9orf72 model and offer an advantage for analyzing poly-GA-targeted therapy in vivo. As in other ALS models and patients, denervation is more pronounced than motor neuron loss arguing for the contribution of dying-back mechanisms due to synaptic toxicity [10]. In addition, poly-GA expression in muscle tissue could contribute to denervation and weakness in GA-Nes mice, and recent studies have uncovered that poly-GA is expressed in the skeletal muscle of C9orf72 patients [9]. Crossing our conditional line with other Cre drivers will allow us to dissect the role of peripheral poly-GA expression in the future.

In addition to DPR inclusions, TDP-43 inclusions are characteristic of sporadic and C9orf72-related and other forms of ALS/FTD and correlate far better with neurodegeneration [37]. However, it is still debated how TDP-43 pathology is triggered by the C9orf72 mutation. The lack of spatial correlation of DPR expression, RNA foci and C9orf72 expression with TDP-43 aggregates in patients suggests that selective vulnerability is critical. This is supported by the striking finding that nuclear TDP-43 aggregates selectively in a subset of neurons in the hippocampus and frontal cortex of GA-Nes mice, despite abundant poly-GA aggregation throughout the brain. Study of the pathomechanism in mice is complicated by differences between mouse and human molecular biology, in particular the processing by caspase 4 that generates the aggregation-prone C-terminal fragments found in cytoplasmic inclusions [64]. Rare cytoplasmic inclusions of TDP-43 have been reported in some (G₄C₂)_n models and viral poly-GA expressing mouse models, but most inclusions in these mice are nuclear [6, 23, 34, 66]. In our congenic mouse models, only poly-GA induced regional intranuclear aggregates of TDP-43 that were not phosphorylated at the disease-specific sites S409/410 or S403/404. We did not observe disease-like proteolytic processing of TDP-43 or cytoplasmic TDP-43 inclusions. Nevertheless, the nuclear TDP-43 aggregates partially colocalize with other RNA-binding proteins hnRNP H identified as an interactor of (G₄C₂)_n RNA, and hnRNP A1 also implicated in ALS/FTD [8, 30, 33]. We now show that poly-GA alone is sufficient to induce aggregation of hnRNP H, in the absence of (G₄C₂)_n RNA. Overall, although these nuclear inclusions are rare at the young age of our mice, our data suggest that poly-GA impairs proteostasis of disease-linked RNA-binding proteins, perhaps mediated by proteasome sequestration [18, 29]. Together with other insults such as defective RNA transport and caspase-mediated cleavage, this may ultimately lead to their cytoplasmic aggregation in patients.

GA-Nes mice show strong activation of microglia throughout the CNS, but particularly in the hippocampus where neurons are lost at an early age. We used RNAseq to elucidate the inflammatory response and conducted a time-course study in multiple CNS regions in GA-Nes and PR-Nes mice. The gene expression profiles are dominated by chemokines and other inflammatory changes in GA-Nes mice with the strongest effects in the degenerating hippocampus. In addition, synaptic genes and several immediate early genes such as Arc, which is crucial for memory formation [42], are downregulated. Together with our observation that cortical neurons are not lost by this age in GA-Nes mice, these data suggest that poly-GA impairs synaptic function, which may contribute to C9orf72 pathogenesis prior to overt neuron loss [54].

Comparing differential gene expression in GA-Nes mice and human ALS cases shows a strong overlap with an inflammatory signature common to both C9orf72 and sporadic ALS. Comparing gene expression in GA-Nes mice and C9orf72-specific changes in human ALS spinal cord highlights the role of enhanced interferon signaling. Interestingly, IFNα levels are higher in the CSF of C9orf72 ALS compared to sporadic ALS [51]. Transcriptome analysis in cortex samples from ALS cases identified a subgroup of patients with high microglia/interferon response, without enrichment in C9orf72 cases, but the spinal cord was not analyzed in that study [60]. Comparing gene expression of isolated GA-Nes microglia to recently published microglia expression signatures from single cell RNAseq experiments [19, 26, 40, 56] highlights the role of pro-inflammatory microglia subpopulations characterized by a strong interferon response. Indeed, many of the differentially expressed genes are known to be interferon-inducible and we show that STAT signaling is enhanced in GA-Nes mice. In addition, we confirmed that IFNγ was increased in GA-Nes but not PR-Nes mice. Moreover, CHIT1, a sensitive CSF biomarker for ALS [50, 59] is also strongly induced by interferon [39] suggesting a sterile interferon response is highly relevant in human ALS. A causal role of interferon in neurodegeneration is supported by inherited interferonopathies such as Aicardi–Goutieres syndrome, associated with cerebral atrophy and psychomotor retardation in children [17]. We show for the first time that poly-GA is sufficient to induce strong interferon responses in vivo, which is seen in the spinal cord of sporadic ALS and enriched in C9orf72 patients. Future experiments to investigate whether inhibiting this pro-inflammatory response will extend the lifespan of GA-Nes animals would help determine whether this may be a therapeutic approach for C9orf72 ALS.