Characterizing the expression profile of 3R tau pathology in Pick’s disease

Abstract

Pick’s disease (PiD) is a rare neurodegenerative disorder defined by dementia, frontotemporal lobe atrophy, and 3-repeat tau inclusions. To elucidate PiD pathobiology, we performed the first bulk transcriptomics study on PiD using short- and long-read sequencing on the parietal cortex of 28 PiD and 15 control samples. We identified several significantly differentially expressed genes, with CCL2 displaying the strongest association with 3-repeat tau pathology and increased burden in PiD compared to those in 4-repeat tau progressive supranuclear palsy (PSP) cases. Investigation of co-expressed genes and pathways suggested the involvement of mRNA processing, mitochondrial function, and immune processes in disease pathobiology. Long-read RNA sequencing on a subset of samples (eight PiD and four control) proposed novel, potentially disease-associated transcripts for AZGP1, CD44, HSD11B2, and WIF1, predicted to result in truncated proteins. In conclusion, we observed transcriptomic changes in the parietal cortex of patients with PiD that may inform into clinically relevant biomarkers and therapeutic strategies.

The Pick’s disease brain is characterized by abundant transcriptional changes at the gene and transcript level.

INTRODUCTION

The evolution of RNA sequencing (RNA-seq) platforms has provided important insight into the altered transcriptional landscape of many neurodegenerative disorders. Short-read RNA-seq studies have largely contributed to our understanding of differential gene expression (DGE) in a variety of dementia-causing pathologies, including frontotemporal lobar degeneration (FTLD) and Alzheimer’s disease (AD) (1–4). With the emergence of new analysis pipelines and bioinformatics tools, short-read RNA-seq data have been leveraged to define single-exon cryptic splice junctions, disease-affected pathways, and vulnerable cellular networks (5, 6). While these studies have been instrumental in advancing our understanding of disease pathophysiology, most do not address differential transcript expression in disease. Short-read RNA-seq studies are also unable to capture full-length transcripts and describe the existence of potentially novel disease-associated gene transcripts or altered transcript proportions. In the meanwhile, most transcriptomics studies to date have been focused on the more common dementias, leaving many unanswered questions about the pathobiology of rare dementia-causing diseases, such as Pick’s disease (PiD).

PiD is a rare form of FTLD defined by the presence of 3-repeat (3R) tau inclusions (7, 8). PiD clinically presents as behavioral variant frontotemporal dementia (FTD) but can also present as primary progressive aphasia or corticobasal syndrome as patients will typically suffer from dementia, behavioral changes, and language disturbances (8–12). Symptoms tend to appear at ~57 years of age and last up to 10 years (11). However, as the disease can only be diagnosed postmortem, PiD is used strictly as a neuropathology term (8). PiD is neuropathologically defined by severe atrophy of the frontal and temporal lobes, gliosis, as well as the presence of aggregated tau in structures called Pick bodies and balloon neurons termed Pick cells (7, 13). Due to its neuropathological signature, PiD is classified as a primary tauopathy along with other disorders like progressive supranuclear palsy (PSP) and corticobasal degeneration (CBD) (8, 14).

More specifically, PiD is defined by the aggregation of specific tau protein species in Pick bodies. The tau protein is encoded by the MAPT gene on chromosome 17, a locus that undergoes complex alternative splicing and gives rise to six, main, protein-coding transcripts (15, 16). These transcripts differ in the inclusion/exclusion of MAPT exons 2, 3, and 10, and can be divided into two classes 3R and 4-repeat (4R), based on the exclusion (3R) or inclusion (4R) of exon 10. While 3R and 4R tau isoforms exist in equilibrium in the adult human brain, alterations in the 3R:4R protein ratio have been linked to the pathogenesis of the tauopathies that are further classified on the basis of the identity of the tau isoforms comprising the pathognomonic inclusions of each disease state (14). Hence, the tauopathies can be described as 3R (PiD), 4R (PSP, CBD, globular glial tauopathy, and argyrophilic grain disease) or 3R + 4R (AD and primary age-related tauopathy).

Due to disease rarity along with the lack of available biomarkers and tests for a definitive clinical diagnosis, PiD has remained understudied, and little is known about its pathophysiology. Given the unique nature of PiD as a 3R tauopathy and the importance of understanding full-length transcripts and their proportions in the tauopathies, transcriptomic studies on PiD are imperative to enhance our understanding of disease pathogenesis. Here, we describe the first transcriptomic study performed on PiD brain tissue using short- and long-read sequencing approaches to resolve the gene expression changes that take place in the parietal brain lobe of patients with PiD.

RESULTS

Characterization of the transcriptional profile of PiD in parietal tissue

To characterize the gene expression changes that occur in PiD, we generated short-read bulk transcriptomic data from the parietal cortex of 28 PiD cases and 15 neurologically healthy controls from the Mayo Clinic Florida Brain Bank. Given that the parietal cortex is a region showing characteristic pathology in PiD while remaining relatively spared from extensive atrophy, it allowed for high-quality RNA extractions.

After adjusting for sex, age at death (AAD), and RNA integrity number (RIN) values (Fig. 1A), we identified 12 differentially expressed genes (DEGs) between PiD cases and controls (Fig. 1B). Of those, eight genes were significantly up-regulated and four were significantly down-regulated in the PiD samples (Fig. 1, B and C). Within those, seven genes are protein coding (SST, AZGP1, CCL2, CD44, HSD11B2, and SOCS3), while the rest encode for different forms of RNAs, including small nucleolar RNAs (snoRNAs; SNORA63), small nuclear RNAs (snRNAs; U1), miscellaneous RNAs (misc RNAs; RNY1 and RNY4), and long intergenic noncoding RNAs (lincRNAs; RP11-497H16.9) (Table 1). Notably, we did not observe a significant difference in MAPT expression between PiD cases and control samples. Given that PiD is a primary tauopathy, understanding MAPT expression in the context of 3R tau pathology remains imperative. Our results could indicate more nuanced changes in the levels of specific transcripts that are not best captured by short-read approaches.

Fig. 1. DGE analysis reveals significant changes in the parietal cortex of patients with PiD compared to control samples.

(A) Source of variation analysis depicting the variables used as covariates in the differential expression model and the percentage they contribute to gene expression variation. (B) Volcano plot showing the differentially expressed genes (DEGs) in PiD cases versus control samples. (C) Box plot visualization of the expression of the 12 significant DEGs among the two groups.

Table 1. Significant DEGs in PiD compared to those in control samples identified through short-read bulk RNA-seq.

Table summarizing the significant DEGs identified in the PiD cases compared to those in control samples including statistical summary. Statistical significance was determined using a linear regression model adjusted for AAD, sex, and RIN, and q values were obtained by false discovery rate (FDR) correction. snoRNA, small nucleolar RNA; Misc RNA, miscellaneous RNA; snRNA, small nuclear RNA; lincRNA, long intragenic noncoding RNA; CI, confidence interval; SE, standard error.

Gene ID	Gene name	Chromosome	Start	End	Gene biotype	P value	q value	Log2 fold change (95% CI)	SE
ENSG00000200320	SNORA63	chr3	186,787,300	186,787,431	snoRNA	1.38 × 10−7	2.89E-04	1.97 (1.35 to 2.59)	0.31
ENSG00000108691	CCL2	chr17	34,255,218	34,257,203	Protein coding	8.60 × 10−4	1.44E-02	1.88 (0.83 to 2.93)	0.52
ENSG00000176387	HSD11B2	chr16	67,430,652	67,437,553	Protein coding	3.85 × 10−5	3.36E-03	1.86 (1.05 to 2.67)	0.40
ENSG00000201098	RNY1	chr7	148,987,136	148,987,248	Misc RNA	1.93 × 10−6	8.21E-04	1.77 (1.13 to 2.41)	0.32
ENSG00000252316	RNY4	chr7	148,963,315	148963410	Misc RNA	1.96 × 10−5	2.61E-03	1.74 (1.02 to 2.46)	0.36
ENSG00000184557	SOCS3	chr17	78,356,778	78,360,077	Protein coding	5.49 × 10−3	4.07E-02	1.61 (0.50 to 2.72)	0.55
ENSG00000160862	AZGP1	chr7	99,966,720	99,976,157	Protein coding	6.71 × 10−5	4.24E-03	1.59 (0.87 to 2.31)	0.36
ENSG00000026508	CD44	chr11	35,138,870	35,232,402	Protein coding	3.31 × 10−3	3.03E-02	1.50 (0.53 to 2.47)	0.48
ENSG00000157005	SST	chr3	187,668,906	187,670,399	Protein coding	2.38 × 10−4	7.21E-03	−1.54 (−2.31 to −0.77)	0.38
ENSG00000156076	WIF1	chr12	65,050,626	65,121,566	Protein coding	8.09 × 10−5	4.56E-03	−1.71 (−2.49 to −0.92)	0.39
ENSG00000277918	RNVU1-28	chr1	144,560,666	144,560,829	snRNA	1.21 × 10−3	1.73E-02	−1.78 (−2.80 to −0.75)	0.51
ENSG00000269983	RP11-497H16.9	chr5	70,449,636	70,450,353	lincRNA	3.07 × 10−3	2.92E-02	−1.82 (−2.98 to −0.65)	0.58

Given the nature of bulk transcriptomic approaches, we wanted to explore how altered cell proportions in the disease brain would potentially affect the identified DEGs. Thus, we performed cell-type proportion analysis to estimate cell-type proportion alterations between PiD cases and controls. We did not observe significant differences between PiD cases and control samples for any of the cell types, reinforcing that the driver behind the DEGs that we detected is the PiD pathology (fig. S1A). Further adjustment for cell proportions in the DGE model previously described confirmed the significant DEGs and implicated five additional cell-specific hits (fig. S1, B and C, and table S3).

Reverse transcription polymerase chain reaction (RT-PCR) was used to validate the detected protein-coding DEGs in a subset of sex matched PiD and control samples (n = 15 per group) from the cohort described above, as well as 15 PSP cases that are defined by 4R tau pathology to determine the link of the detected DEGs with neurodegeneration and specific types of tau pathology (table S1). We confirmed that CCL2, CD44, and SOCS3 were overexpressed in PiD compared to those in both control and PSP cases (Fig. 2, A to C). This could indicate a unique association between the expression dysregulation of these genes and 3R tau pathology in the parietal cortex. We also observed significant overexpression of AZGP1 compared to that of control samples, but not PSP, pointing to a potential broader tau pathology effect in the brain (Fig. 2D). We were, however, not able to validate significant overexpression of HSD11B2 or underexpression of WIF1 and SST (Fig. 2, E to G). Yet, even if not statistically significant in these comparisons, PiD cases had higher levels of HSD11B2 expression and lower levels of WIF1 and SST expression compared to those of the control samples. Future studies with increased numbers might assist in increasing the power to confirm the significant association. Furthermore, investigating the presence of different transcripts for these genes will be crucial in designing more accurate gene expression probes based on transcript prevalence.

Fig. 2. DGE validation through RT-PCR and IHC.

RT-PCR in PiD, control, and PSP samples for (A) CCL2, (B) CD44, (C) SOCS3, (D) AZGP1, (E) HSD11B2, (F) SST, and (G) WIF1. Significance was determined with a linear regression model adjusted for AAD. *P < 0.05; **P < 0.01; ***P < 0.001. (H) Representative immunohistochemistry (IHC) image of a PSP case and a PiD case stained with CCL2 antibody (scale bars, 60 μm).

To understand the contribution of the DEGs to PiD pathology, we performed an association study of DEG expression with semiquantitative pathology scores of two of the hallmarks of PiD neuropathology: Pick body burden (3R Tau; RD3 antibody) and astrogliosis [glial fibrillary acidic protein (GFAP) antibody] (fig. S2 and table S1). In linear regression analysis adjusting for AAD and sex, an increased Pick body burden was associated with a significantly higher expression of CCL2 (β: 0.79, P = 0.006), SOCS3 (β: 0.77, P = 0.017), and HSD11B2 (β: 0.35, P = 0.039) (Table 2). These genes are highly involved in the regulation of immune and stress processes in the brain and their association with 3R tau pathology might be indicative of immune and stress processes contributing to PiD pathophysiology (17–19). Given that CCL2 expression showed the strongest association with PiD pathology, we aimed to evaluate its potential link to 3R tau pathology by staining for CCL2 in the parietal cortex of PiD and PSP samples (Fig. 2H and fig. S2). We observed significantly higher CCL2 staining in the PiD cases compared to that in PSP samples (P = 0.01). Overall, our data are suggesting an overexpression of CCL2 linked to 3R tau pathology in the parietal cortex of patients with PiD. Also of note regarding associations between DEG expression and PiD neuropathology was a significant association between increased astrogliosis in the parietal cortex and greater RNVU1-28 gene expression (β: 1.11, P = 0.007). Because RNVU1-28 is down-regulated in PiD cases, this relationship could be indicative of a compensatory mechanism in the brain that requires further in-depth investigation.

Table 2. Associations of neuropathology measures with gene expression of the significant DEGs in the PiD brain.

Overview of the results obtained from the association study of semiquantitative neuropathology measures with DEG expression. β, regression coefficient; CI, confidence interval. β Values, 95% CIs, and P values result from linear regression models that were adjusted for AAD and sex. β Values are interpreted as the change in mean gene expression corresponding to a 1-unit increase in the given score (GFAP or 3R tau).

	Association with GFAP score		Association with 3R tau score
Gene	β (95% CI)	P value	β (95% CI)	P value
RNVU1-28	1.11 (0.35, 1.86)	0.007	0.52 (−0.19, 1.22)	0.14
RP11-497H16.9	−0.60 (−1.66, 0.45)	0.25	−0.25 (−1.12, 0.62)	0.56
RNY4	−0.15 (−0.64, 0.34)	0.52	0.12 (−0.27, 0.51)	0.53
RNY1	−0.33 (−0.86, 0.21)	0.21	0.30 (−0.11, 0.72)	0.14
SNORA63	−0.18 (−0.86, 0.51)	0.60	0.30 (−0.24, 0.83)	0.26
SOCS3	0.26 (−0.64, 1.16)	0.56	0.77 (0.15, 1.38)	0.017
HSD11B2	0.00 (−0.48, 0.48)	>0.99	0.35 (0.02, 0.69)	0.039
AZGP1	−0.28 (−0.93, 0.38)	0.38	0.06 (−0.47, 0.59)	0.82
SST	0.00 (−0.67, 0.66)	0.99	−0.18 (−0.70, 0.34)	0.48
WIF1	−0.62 (−1.29, 0.05)	0.069	−0.48 (−1.02, 0.06)	0.076
CCL2	0.20 (−0.63, 1.02)	0.62	0.79 (0.26, 1.32)	0.006
CD44	0.36 (−0.44, 1.16)	0.36	0.44 (−0.18, 1.05)	0.16

PiD pathology in the parietal cortex is linked to the dysregulation of genes involved in RNA metabolic processes and mitochondrial function

For further insight into dysregulated cellular networks in the parietal cortex of patients with PiD, we performed weighted-gene co-expression analysis (WGCNA) analysis and identified five gene modules (red, green, black, magenta, and pink) that were significantly up-regulated in PiD compared to those in control samples and two significantly down-regulated modules (yellow and brown) (Fig. 3, A and B). The most significant positive association detected was with the magenta module (P = 1 × 10⁻⁶), which contains the largest number of up-regulated DEGs (n = 5) and is enriched in genes involved in RNA metabolic processes (Fig. 3, B and C). The two down-regulated modules, yellow and brown, also contained down-regulated DEGs and are enriched in genes associated with mitochondrial function and the electron transport chain (Fig. 3, D and E). Mitochondrial dysfunction is a known component of aging and neurodegeneration, however, an association with PiD has not been identified to date. While it did not reach statistical significance, the purple module contained a small number of significantly up-regulated DEGs and is enriched in genes associated with the immune response. These data are consistent with the results of the ingenuity pathway analysis (IPA) analysis conducted on the significant DEGs that nominated altered pathways associated with cytokine production and interleukin signaling in PiD compared to those in control samples (table S4).

Fig. 3. Gene network construction using WGCNA.

(A) Gene dendrogram assigning genes to different colored modules. (B) Heatmap of module-trait relationships. (C) Dot plot of the top 10 Gene Ontology (GO) terms from the magenta module. (D) Dot plot of the top 10 GO terms from the yellow module. (E) Dot plot of the top 10 GO terms from the brown module.

Differentially expressed protein-coding genes have greater transcript diversity in PiD cases compared to those in controls

To further assess whether the observed changes in gene expression were a result of total or isoform-specific expression, we used long-read RNA-seq (Iso-seq) in a subset of our cohort (n = 8 PiD cases and n = 4 control samples; tables S1 and S5). Overall, our approach yielded on average 2.8 million reads, capturing ~90,000 transcripts with a mean transcript length of 3400 base pairs (bp) (table S5). First, we investigated the detected transcripts for the seven protein-coding DEGs. In our Iso-seq data, we were able to detect transcripts for all seven genes, six of which showed the same trend of expression in PiD compared to that in control samples (tables S6 and S7). The only exception was HSD11B2, which had the same number of average normalized counts [transcript per million (TPM)] in both control and PiD samples. We identified multiple gene transcripts for the majority of the DEGs, with the exception of SOCS3 and SST for which we only detected a single transcript (Fig. 4, F and G).

Fig. 4. Visual representation of detected transcripts for seven protein-coding DEGs using Iso-seq and their respective abundance in PiD and control samples shown as percentages of total counts.

(A) HSD11B2, (B) AZGP1, (C) CCL2, (D) CD44, (E) WIF1, (F) SOCS3, and (G) SST. The number of reads for each gene transcript is included in table S7, and transcript sequences are available in the Supplementary Materials.

More specifically, we detected four HSD11B2 transcripts, including the annotated, protein-coding transcript, ENST00000326152.6, which was only present in the PiD cases in very small amounts (3.5%), along with three novel, non-annotated transcripts that are produced via combination of known junctions or splice sites (Fig. 4A). While HSD11B2_novel_tx_c is the most common in both control and PiD samples (92.4% in control and 64.3% in PiD), PiD cases are characterized by greater HSD11B2 transcript diversity (Fig. 4A). The HSD11B2_novel_tx_c is a protein-coding transcript, lacking exon 1 and resulting in a shorter 301–amino acid product. On the other hand, HSD11B2_novel_tx_a is characterized by a longer 3′ untranslated region (3′UTR) as well as an intron retention event before exon 2, resulting in a 590–amino acid protein with ~60% similarity to the annotated ENST00000326152.6 transcript. This transcript is seen at a higher percentage in PiD cases (16.7%) compared to that in control samples (3.1%). Similarly, HSD11B2_novel_tx_b makes up 15.4% of HSD11B2 total detected transcripts in PiD samples compared to 4.5% of those in control samples and is most likely resulting in a truncated 21–amino acid protein due to the lack of exon 1 (Supplementary Data).

Additionally, we detected two AZGP1 transcripts: a full splice match (FSM) to the reference (ENST00000292401.9) and a novel, non-annotated transcript (AZGP1_novel_tx_a) that is a result of the combination of known junctions that is more abundant in PiD cases (49%) compared to that in controls (17%) and results in a truncated 230–amino acid protein (Fig. 4B and Supplementary Data).

Similarly, the most abundant CCL2 transcript detected in both PiD and control samples was the annotated protein-coding ENST00000225831.4 transcript (Fig. 4C). However, the PiD cases had a higher percentage (24%) of the noncoding ENST00000624362.2 transcript that is a result of multiple intron retention events compared to the control samples (5%) (table S7 and Supplementary Data).

We additionally detected six CD44 transcripts among the PiD and control samples, with annotated transcript, ENST00000263398.11 producing a 361–amino acid product, being the most abundant in both groups (Fig. 4D). However, the longer, annotated, protein-coding transcript ENST00000434472.6 was lowly expressed in both groups (0.4% in PiD and 1.1% in controls). In our data, we were also able to detect four novel-in-catalog (NIC) transcripts that result from the combination of known junctions and splice sites. Of these, only CD44_novel_tx_c was present in both PiD cases (1%) and controls (2.8%), while the rest were only expressed in the PiD cases. More specifically, CD44_novel_tx_a, which makes up 1.3% of the total detected CD44 transcripts for PiD cases and 0% of the controls, is defined by the presence of five additional exons and a shorter 3′UTR, giving rise to a larger, 580–amino acid protein product. CD44_novel_tx_b, which is only present in PiD samples at a low percentage (0.6%), is defined by the absence of exon 5 and results in a shorter, 284–amino acid protein product. CD44_novel_tx_c is characterized by a truncated, 139-bp, exon 5 giving rise to a shorter, 330–amino acid protein. Last, CD44_novel_tx_d, a transcript only making up a small percentage of the total PiD detected transcripts (0.5%), is lacking exons 2 through 5, and, thus, it is predicted to produce a 161–amino acid product (Supplementary Data).

Last, we detected four transcripts of WIF1 with most abundant in both groups (81.3% in PiD and 77.2% in controls) being the annotated ENST00000286574.9 transcript that encodes for a 379–amino acid protein (Fig. 4E). The second most abundant transcript in both groups (14.6% in PiD and 16.5% in controls) was WIF1_novel_tx_b, which is defined by an intron retention event between exons 8 and 9. This intron retention event introduces a new stop codon, giving rise to a smaller, 319–amino acid protein product. Similarly, WIF1_novel_tx_a, which makes up 3.6% of the PiD total transcripts and 5.4% of the control total transcripts and is characterized by a truncated exon 8, is predicted to code for a 365–amino acid protein (Supplementary Data). WIF1_novel_tx_c is detected at much lower levels in both PiD cases (0.5%) and control samples (0.9%) and is defined by deletion of exon 6, resulting in a 347–amino acid product.

Overall, the novel transcripts detected for the DEGs are predicted to be coding and are more abundant in the PiD cases compared to those in control samples, highlighting the unique transcriptional diversity of PiD compared to that of the healthy adult brain (fig. S3). However, at the global level, PiD and control samples had very similar percentages of splicing events both at the structural category and subcategory level (Fig. 5, A and B), pointing to the distinctive transcriptional dysregulation of the genes that could be contributing to their increased or decreased expression in disease, respectively.

Fig. 5. Splicing event proportions are highly similar between PiD and control samples.

(A) Percentage of total counts of transcripts defined by their structural category. (B) Percentage of total counts of transcripts defined by their subcategory.

To validate the novel transcripts, we investigated our larger cohort of 28 PiD and 15 control samples using the short-read RNA-seq data and the Kallisto software. By incorporating the novel transcript sequences in the Kallisto reference transcriptome, we were able to detect all novel transcripts except for CD44_novel_tx_c. More specifically, we saw significantly higher expression of AZGP1_novel_tx_a, CD44_novel_tx_b, HSD11B2_novel_tx_b, and HSD11B2_novel_tx_c in PiD cases compared to those in control samples and significantly lower expression of WIF1_novel_tx_b in PiD cases, which is consistent to the Iso-seq results (table S6). We additionally used an RT-PCR custom probe to quantify the AZGP1 novel transcript expression, which was the transcript showing the greatest usage in PiD cases compared to the other novel transcripts that we identified. The RT-PCR results were consistent with the Iso-seq data, confirming higher expression of the novel transcript in PiD samples compared to that in control brains. We also observed lower expression of AZGP1_novel_tx_a in PSP cases compared that of to PiD, although it did not reach statistical significance (P = 0.06), which could indicate a 3R tau–specific transcript that needs to be validated in larger, independent cohorts (fig. S4A). Last, we used SWISS-MODEL to create predicted amino acid structures for AZGP1_novel_tx_a and ENST00000292401.9 (fig. S4, B and C). The two–amino acid sequences share the same first 204 amino acids, and the predicted protein sequence for AZGP1_novel_tx_a is missing the class I major histocompatibility complex alpha chain immunoglobulin domain that we believe interferes with protein function (fig. S4C). Overall, we discovered an abundance of novel transcripts in the protein-coding DEGs that are more common in PiD but are defined by shorter or highly different protein products. Thus, the overexpression of these novel transcripts might explain the link between the DEGs and PiD disease processes.

DISCUSSION

PiD is a rare, dementia-causing neurodegenerative disorder that is defined by unique 3R tau pathology in Pick bodies. By studying gene expression changes in PiD, we provided novel mechanistic insight contributing to the pathobiology of the disease and nominated targets for further investigation. Additionally, identifying key aspects of 3R tau pathology can inform into therapeutic efforts for related disorders with tau pathology, e.g., PSP, CBD, and AD.

Here, we present the first transcriptomic study performed on PiD using short- and long-read RNA-seq technologies focusing on global transcriptional changes in the parietal cortex. We identified 12 genes that are significantly differentially expressed between the two conditions. Of note, we did not observe differential expression of MAPT, encoding the tau protein, which may reflect a limitation of short-read sequencing given the variability in transcript expression at this locus. Further targeted long-read studies to fully resolve the role of MAPT splicing in tauopathy are warranted. Candidates identified included genes encoding for snRNAs and snoRNAs such as SNORA63 and RNVU1-28. Past work has implicated different types of noncoding RNAs, including snRNAs and variant U1 snRNAs in the regulation of gene expression and the neurodegenerative process (20). Specifically, studies have shown that tauopathy mouse models as well as AD human brains are characterized by an accumulation of snRNAs and snoRNAs, including many different U1 small nuclear ribonucleoprotein complex components that either are part of the neurofibrillary tangles or partially colocalize with them (21, 22). This accumulation of proteins essential for RNA splicing has been linked to widespread RNA splicing dysregulation and cognitive impairment acceleration in AD, but not in FTD or Parkinson’s disease (22). While we did not detect a widespread splicing dysregulation (Fig. 5), future neuropathological examination and proteomic studies on Pick bodies will be critical.

Through short-read RNA-seq DGE analysis, we additionally identified several protein-coding genes with some having previously reported associations with neurodegeneration. Specifically, we observed an increase in expression of inflammation and immunity related genes CCL2, SOCS3, AZGP1, and CD44. This could indicate an overall dysregulation of the immune response in the brain of patients with PiD linked to disease pathobiology. The CCL2 gene encodes for a chemokine (chemokine C-C motif ligand 2) responsible for regulating monocyte infiltration into the brain following inflammation or infection, and its overexpression has been previously linked with increased tau pathology in tau transgenic mice (23). Furthermore, CCL2 (also known as MCP-1) is a clinically relevant biomarker as increased serum and cerebrospinal fluid (CSF) levels have been observed in patients with mild AD (17). Thus, given the observed positive association of CCL2 expression with increase Pick body burden in our data as well as its increased expression compared to those in both control and PSP samples, we postulate that the CCL2 dysregulation might be linked to 3R tau pathology. We further observed significantly higher CCL2 burden in PiD compared that in to PSP cases with immunohistochemistry (IHC) staining. Joly-Amado et al. proposed a potential mechanism linking CCL2 overexpression with tauopathy through an induction of tau phosphorylation and increase in insoluble tau (23).

Similarly, the SOCS3 gene, which encodes for the suppressor of cytokine signaling 3 protein, is also involved in inflammation and immunity by regulating cytokine signaling and has been shown to be up-regulated in the brain of patients with AD (18, 24). The AZGP1 gene encodes for the Zinc-alpha-2-glycoprotein and has been linked to AD and dementia through a transcriptome-wide association study and a number of CSF biomarker studies; however, its role in the brain remains elusive (25–27). The CD44 gene encodes for a protein that plays an important role in the extracellular matrix and is thus involved in a large range of physiological processes including synaptic plasticity and immunity, mostly studied in cancers thus far (28). Less is known about the potential mechanisms through which these three genes might be acting to contribute to neurodegeneration but the possibility that they are working in synergy with CCL2 will be assessed in future functional studies. We also described a significant up-regulation of the HSD11B2 gene that encodes for a protein that regulates glucocorticoid metabolism in the brain (19).

We observed significant down-regulation of the somatostatin neuropeptide, SST gene, as well as the wnt inhibitory factor 1, WIF1 gene, in PiD compared to those in control samples. Previous studies have shown a similar decrease in SST expression in AD; however, these have been linked to amyloid deposition (29). As PiD cases are not characterized by amyloid deposition, SST could be acting through a distinct mechanism to influence PiD pathogenesis than what has been previously described in the literature. Additionally, the WIF1 protein acts as a wnt pathway inhibitor, and its down-regulation has been previously linked to AD (30, 31). As part of our DGE analysis we also detected genes encoding for misc RNAs RNY1 and RNY4 as well as lincRNA RP11-497H16.9, supporting previous reports that lincRNAs can act as the inflammatory modulators in neurodegenerative diseases by regulating gene expression (32).

Through network analysis, we identified modules enriched in genes related to mRNA processing as positively associated with a PiD diagnosis, as well as modules enriched in mitochondrial function that were negatively correlated with a PiD diagnosis. These two processes, along with immune dysregulation, have been broadly studied in the context of neurodegeneration. Our study is the first to highlight their role in PiD biology. It is likely that unique aspects of these processes are dysregulated in each disease state. Our study begins to elucidate some of these distinct processes by comparing PiD to PSP. However, assessment of our proposed targets in cohorts of AD as well as other 4R tauopathies like CBD or other proteinopathies will be crucial in determining their potential tau-isoform specific associations.

To gain additional insight into the identified protein-coding DEGs and identify specific transcripts contributing to their change in expression, we used Iso-seq data for a subset of our cohort. Iso-seq produces reads that span the entirety of transcripts eliminating the need of read reconstruction and providing us with confident transcript quantification. We were able to detect novel transcripts showing higher expression in PiD compared to that in control cases for AZGP1, CD44, and HSD11B2. These transcripts are commonly associated with truncated or highly altered protein sequences that could interfere with protein functionality and could explain some of the pathway dysregulation that we are detecting. We are thus postulating that specific transcript overexpression in PiD leads to an increase of dysfunctional proteins in the disease state, which, in turn, influences downstream processes like mRNA processing, mitochondrial function, and immune regulation. Additionally, we characterized the transcript usage of HSD11B2 in our samples and observed that the annotated transcript is not highly expressed in either cohort. This could explain the inability of the TaqMan gene expression probe used in our RT-PCR replication to fully capture the gene up-regulation in PiD cases as it was designed to span an exon-exon junction that is rarely seen in our samples. Future studies on the protein products of these novel transcripts and their abundance and structure will provide additional insight into potential disease mechanisms.

There are several limitations of this study that need to be acknowledged. First, given the rarity of PiD, our study is limited by its small sample size, and, therefore, the possibility of a type II error (i.e., a false-negative finding) is important to consider. Along with that, the availability of large, independent, replication cohorts are absent. We hope that, through the creation of the Pick’s Disease International Consortium, we will be able to identify equal sized cohorts that will allow for the independent replication of the detected DEGs as well as novel disease-associated transcripts. We have tried to address some of the replication concerns by assessing the direction of effect of our DEGs across RNA-seq datasets and in our RT-PCR and IHC data. Further considerations come from the long-read RNA-seq technology itself. At this point, we are unable to determine whether some of the splicing dysregulation that we are detecting, especially, the novel transcripts that did not replicate in Kallisto and RT-PCR, is a true result of differential splicing or a capture of pre-mRNA. Our comparison of transcript category and subcategory proportions, revealing no difference between PiD and control samples, increased our confidence that the novel transcripts showing strong association with disease and replicating across analyses and methods are real. Similarly, long-read RNA-seq technologies have generally lower throughput compared to short-read technologies. This can lead to lower detected transcript counts and interpreting the biological relevance of lowly expressed transcripts remains challenging. To overcome some of these challenges, we have used different approaches for validation of transcripts including RT-PCR and Kallisto; however, replication of these findings and trends in an independent cohort will be imperative for shaping our understanding of their potential biological effect in PiD pathophysiology and on tauopathy. Last, it is also important to highlight the limitations of a postmortem study in elucidating specific cause and effect mechanisms. The development of cell or animal models for PiD and 3R tau pathology will inform into the functionality and interactions of the proposed gene targets and their protein products in a living system.

In conclusion, we have described several transcriptional alterations in the parietal cortex of patients with PiD, some of which are potentially specific to 3R tau pathology. The incorporation of both short- and long-read bulk RNA-seq approaches aided in the identification of specific gene transcripts being more abundant in PiD cases, proposing a potential loss of function or gain of toxic function effect. Our study highlights the involvement of pathways and cellular networks previously linked to many other neurodegenerative disorders; however, given the unique nature of PiD pathology, it is possible that different aspects of these pathways are involved with specific pathological conditions. Future functional studies on the proposed targets will be imperative to increase our understanding of their mechanism of action and identify urgently needed therapeutics for the tauopathies.

MATERIALS AND METHODS

Study cohort

We extracted RNA from frozen brain tissue of PiD, PSP, and control samples obtained from the Mayo Clinic Brain Bank for neurodegenerative disorders in Jacksonville, FL. Disease cases were determined postmortem following established diagnostic criteria (11, 33). Controls did not have any clinical evidence of neurological disease or pathology that would results in a diagnosis of a neurodegenerative disorder. Our final cohort was shaped by the selection of samples that yielded the best RNA quality across extractions and included 28 PiD cases, 15 PSP cases, and 15 controls (Table 3; extended demographics in table S1). All participants were unrelated. This study was approved by the Mayo Clinic Institutional Review Board (22-002827), and patient/next of kin provided signed consent for the research study. The work described has been carried out in accordance with The Code of Ethics of the World Medical Association (Declaration of Helsinki) for experiments involving humans.

Table 3. Patient characteristics.

Overview of PiD (n = 28), control (n = 15), and PSP (n = 15) samples used in this study. Continuous variables including age at disease onset, disease duration, and AAD are summarized in years with the sample median and range, while categorical variables sex, APOE genotype, and MAPT haplotype are summarized with number and percentage.

	PiD cases (n = 28)	Controls (n = 15)	PSP cases (n = 15)
Sex (female)	15 (54%)	7(47%)	7(47%)
Age at death (years)	67 (58, 90)	80 (28, 98)	70 (59, 80)
Age at onset (years)	58 (50, 77)	NA	63 (48, 73)
Disease duration (years)	10 (2, 25)	NA	7 (4, 13)
APOE genotype
E2E3	1 (3.5%)	1 (6.7%)	3 (20%)
E2E4	2 (7%)	1 (6.7%)	0 (0%)
E3E3	16 (57%)	10 (66.6%)	9 (60%)
E3E4	8 (29%)	3 (20%)	2 (13.3%)
E4E4	1 (3.5%)	0 (0%)	1 (6.7%)
MAPT haplotype
H1H1	13 (46%)	9 (60%)	11 (73.3%)
H1H2	12 (43%)	6 (40%)	4 (26.7%)
H2H2	3 (11%)	0 (0%)	0 (0%)

RNA extractions

RNA was extracted from parietal lobe tissue of all PiD, control, and PSP samples. While not the primary affected region, the parietal cortex does present with Pick bodies and gliosis, the two hallmarks of PiD pathology, and its study is important in elucidating disease mechanisms without being confounded by the extensive neuronal loss present in the severely affected region of the frontal lobe. RNA extractions were performed using the TRIzol-Chloroform protocol and a Zymo Direct-zol RNA Micro-Prep kit with deoxyribonuclease I treatment (Zymo Research, California, USA). RNA integrity and concentration were assessed with the Agilent 2100 Bioanalyzer RNA Nano Chip (Agilent Technologies, California, USA) and the Qubit 2.0 Fluorometer (Invitrogen, Massachusetts, USA). All samples produced high-quality RNA with an RIN above 5 for short-read RNA-seq and RT-PCR (median, 6) and above 6.5 for long-read RNA-seq (median, 7.1).

Short-read RNA-seq

All library preparation and RNA-seq was performed at the Mayo Clinic Genome Analysis Core in Rochester, MN. For short-read RNA-seq, library preparation was performed using the Illumina TruSeq Stranded Total RNA Library Prep assay. Each sample was sequenced on two lanes of the NovaSeq S2 on 1 Flow cell (Illumina) in a single batch, using paired-end mode resulting in 101-bp paired-end reads.

QC for short-read RNA-seq data

Raw data from the combined cohort underwent extensive preprocessing quality control (QC), genome alignment, and annotation using the MAP-RSeq v3.0 pipeline (34). MAP-RSeq uses a fast, accurate, and splice-aware aligner, STAR, for generating bam files by aligning sequencing reads to the reference human genome build hg38 (35). Gene and exon expression quantification was performed using the Subread package to obtain both raw gene counts and normalized values of fragments per kilobase per million mapped reads (36). Last, FASTQC and MultiQC tools were used for comprehensive QC assessment of the aligned reads (37, 38). On average, we acquired 85,281,596 total reads in our cohort, of which 82,434,517 (97%) successfully mapped to the reference genome (table S2). In total, we acquired expression data for 64,253 genes. This list was further refined to 60,279 genes by removing genes with missing GC content and length information required for future normalization. Principal components analysis did not reveal any outlier samples to be excluded in downstream processing. Read count data were normalized using the R package cqn v1.44.0 (39), taking into consideration the library size, gene length, and GC content of each gene coding region, and the median normalized expression for all genes was calculated. We identified genes with a median normalized expression (in scale of log₂ reads per kilobase of transcript, per million mapped reads) below 2, marked them as low expressing, and excluded them from further analyses. This further refined our gene list down to 16,157 genes. To assess the relationship between DEGs that will be detected from downstream analysis and the potentially altered cell composition in disease brain, we used the BRETIGEA R package (v1.0.3) to analyze relative cell-type proportions in our dataset (40, 41). Source of variation analysis was conducted using the variancePartition R package (v.1.28.9) to assess the contribution of experimental variables to gene expression variation and inform into covariate selection in the DGE model (42).

DGE, pathway, and network analyses for short-read data

For DGE analysis, we used a linear regression model adjusting for sex, AAD, and RIN. To account for multiple testing, P values were adjusted using a false discovery rate (FDR) correction. Significant DEGs were defined by a cutoff of FDR-adjusted q value < 0.05 and log₂ fold change (log₂FC) < −1.5 or log₂FC > 1.5. Supplementary DGE analyses adjusting for altered gene proportions were also performed, and the significant results were compared between models. Volcano plots and box plots of DEGs were created using the ggplot2 R package (v3.4.4).

To identify pathways enriched by detected DEGs, we used IPA (QIAGEN Inc., www.qiagenbio-informatics.com/products/ingenuity-pathway-analysis). The dataset of DEGs (q value < 0.05 and log₂FC < −1.5 or log₂FC > 1.5) detected from the PiD versus control comparison and their expression values were uploaded to the IPA platform, and significant pathways [−log(P value) > 1.3] were identified using the user dataset as reference.

WGCNA was applied to identify gene networks that correlate with a PiD diagnosis (43). Residual expression values after CQN normalization and adjustment for sex, AAD, and RIN were used to build a signed, hybrid, co-expression model with a soft power of 7, a minimum module size of 60 genes, and a minimum height for merging modules of 0.3. Gene modules were then summarized by the first principal component of the scaled module expression profiles (module eigengene) and given a unique color identifier. Genes that did not meet these criteria were assigned to the gray module. For the correlation of modules with PiD, control samples were coded as 0 and PiD samples as 1. The biomaRt R package (v.2.54.1) was used to annotate Ensembl gene IDs with HUGO gene symbols and entrez IDs. The anRichment R package (v1.22) was used for Gene Ontology enrichment analysis (43–45). Scripts used for analysis are available at https://github.com/ORossLab/Picks_Bulk_RNASeq.

Immunohistochemistry

IHC was performed in a subset of PiD, PSP, and control samples for which fixed tissue was available (table S1). Formalin-fixed tissue from the inferior parietal lobe was cut at 5-μm thickness, mounted on Superfrost glass slides, and stained with antibodies against astrocytes (GFAP, 1:5000; antigen retrieval steam), 3R tau (RD3, 1:5000; antigen retrieval formic acid), and CCL2 (CCL2, 1:10.000; antigen retrieval citrate), following the standard protocol for deparaffinization, using 3,3′-diaminobenzidine as a chromogen and counterstained with hematoxylin (46). Glass slides stained for CCL2 were digitized using Aperio ScanScope XT Slide Scanner (Leica Biosystems, Aperio, USA) producing high-resolution digital whole slide images. Fixed-sized annotations were placed in the parietal cortex and white matter and analyzed using a digital positive pixel algorithm. Significance was determined using an unpaired two tailed Student’s t test for the combined gray and white matter measures on GraphPad Prism 10. Sections stained for GFAP or 3R tau were scored by three raters using a multiheaded scope and a semiquantitative four-point scoring scale (0, none; 1, mild; 2, moderate; and 4, severe).

Quantitative pathology associations

To further investigate the relationship between the identified significant DEGs and PiD pathology, we performed an association study. Associations of GFAP score and 3R tau score with gene expression levels were evaluated using linear regression models that were adjusted for AAD and sex. Regression coefficients (denoted as β) and 95% confidence intervals (CIs) were estimated and are interpreted as the change in the mean gene expression level corresponding to each 1-U increase in the given score (GFAP or 3R tau). No adjustment for multiple testing was made in these exploratory analyses, and P values < 0.05 were considered as statistically significant. Related to this, we did not adjust for multiple testing when evaluating these associations owing to our small sample size and limited power to detect associations before any possible multiple testing adjustment, and, therefore, it will be important to validate our significant findings in an independent series. All statistical tests were two sided. Statistical analysis was performed using R Statistical software (version 4.1.2; R Foundation for Statistical Computing, Vienna, Austria).

RT-PCR validation

For validation of our top DEGs from short-read RNA-seq DGE analysis, we used RT-PCR in a sex matched cohort of 15 PiD samples, 15 control samples from the cohort described above, and 15 PSP samples (table S1). High-quality RNA (1 μg) was reversed transcribed with the High-Capacity cDNA Reverse Transcription kit by Applied Biosystems. Generated cDNA from all samples were used for ABI TaqMan Gene expression assays, following the standard protocols using the QuantStudio 7 Flex Real-Time PCR System. Gene expression assays for all protein-coding significant DEGs were performed in quadruplets using available TaqMan probes (CCL2, Hs00234140_m1; SOCS3, Hs02330328_s1; WIF1, Hs00183662_m1; HSD11B2, Hs00388669_m1; AZGP1, Hs00426651_m1; CD44, Hs01075859_m1; and SST, Hs00356144_m1) and normalized to HPRT1 (Hs99999909_m1). Data for each probe were analyzed separately in the QuantStudio software. Relative quantification (RQ = 2^−ΔΔCT) values were used for visualization. Significance was determined using a linear regression model adjusted for AAD.

Long-read RNA-seq and data analysis

Long-read RNA-seq for eight PiD and four control samples from the larger cohort described above was performed at the Mayo Clinic Genome Analysis Core where full-length double-stranded cDNA was prepared from roughly 300 ng of total RNA and the cDNA product was used to generate libraries using standard Iso-seq protocols. Library preparation was performed using the Pacific Biosciences (PacBio) SMRTbell prep kit 2.0. Sequencing was performed on the Sequel II and IIe instruments, and each sample was sequenced on one SMRT cell. We generated highly accurate circular consensus sequencing reads from the raw subreads files for each sample using the pbccs package, keeping reads with at least three passes and a minimal predicted accuracy of 90%. The lima package was used for primer removal, and the isoseq3 package was used for polyA tail and concatamer removal resulting in full-length non-concatemer reads. Reads were aligned to the human reference genome (GRCh38) with the Minimap2 aligner using the flair align module (v2.0.0) (47). Through the correct module of the flair package, we corrected misaligned splice sites using the human_GRCh38.fasta and gencode.v44.chr_patch_hapl_scaff.annotation.gtf files. High-confidence isoforms from corrected reads were generated by the collapse module of flair, and, after collapsing, count information was obtained for each unique transcript variant using the flair quantify module. SQANTI3 (v5.2) was used to apply corrections and filters as well as to add annotations on the basis of the detected splice junctions (48). More specifically, transcripts were classified as FSMs, incomplete splice matches, NIC (e.g., novel combination of annotated splice junctions), or novel not in catalog (e.g., completely novel splice junctions) and further defined by subcategories based on their specific sequence alterations including (3′UTR or 5′UTR differences, intron retention, etc.). Transcript and proportion plots were generated in R using the ggtranscript (v0.99.9) and ggplot2 packages using TPM normalized counts.

Transcript validation

Novel transcripts sequences were extracted from the fasta files generated by SQANTI3. Their sequences were compared to the annotated transcripts to determine discrepancies, and open reading frames were determined using the VectorBuilder.com coding sequence analysis tools. To assess the presence of the junctions defining the novel transcripts identified in the Iso-seq data in our larger short-read RNA-seq cohort, we used Kallisto that uses a pseudoalignment approach to quantify transcript abundances in short-read datasets (49). To detect the novel transcripts of interest, Kallisto was ran with a custom reference file including the sequences of the novel transcripts detected in Iso-seq. For RT-PCR, a gene expression TaqMan probe to determine the expression levels of the novel AZGP1 transcript was designed to target its unique exonic sequence using the Custom TaqMan Assay Design Tool (forward primer, AGGTGAGGGTAACACTGTCTGT; reverse primer, CGATGGGATTTGGAATCTGAAATCTG; probe, CCCAGCTCTGCCTCTT). Last, SWISS-MODEL [SWISS-MODEL (expasy.org)] was used to model the predicted amino acid sequence structures for the AZGP1 novel and annotated transcripts. ProteinPaint [Visualization Community | St. Jude Cloud (stjude.cloud)] was further used to visualize protein domains.

Supplementary Materials

This PDF file includes:

Tables S1 to S8

Figs. S1 to S4

Supplementary data