Whole-Genome Sequencing Analysis Reveals New Susceptibility Loci and Structural Variants Associated with Progressive Supranuclear Palsy

Abstract

Progressive supranuclear palsy (PSP) is a neurodegenerative disease characterized by the accumulation of aggregated tau proteins in astrocytes, neurons, and oligodendrocytes. We performed whole genome sequencing (WGS) and conducted association analysis for single nucleotide variants (SNVs), small insertions/deletions (indels), and structural variants (SVs) in a cohort of 1,718 individuals with PSP and 2,944 control subjects. Analysis of common SNVs and indels confirmed known genetic loci at MAPT, MOBP, STX6, SLCO1A2, DUSP10, and SP1, and also uncovered novel signals in APOE, FCHO1/MAP1S, KIF13A, TRIM24, TNXB, and ELOVL1. In contrast to Alzheimer’s disease (AD), we observed the APOE ε2 allele to be the risk allele and the ε4 allele to be protective, a pattern similar to the association pattern observed in age-related macular degeneration (AMD) but the opposite observed for Alzheimer’s disease (AD). Analysis of rare SNVs and indels identified significant association in ZNF592 and further gene network analysis identified a module of neuronal genes dysregulated in PSP. We also observed seven common SVs associated with PSP on the H1/H2 haplotype region (17q21.31) and in a few other loci: IGH, PCMT1, CYP2A13, and SMCP. Particularly, in the H1/H2 haplotype region, there is a burden of rare deletions and duplications (P = 6.73×10⁻³) in PSP. Through WGS, we significantly refine our understanding of the genetic basis of PSP, providing new targets for exploring disease mechanisms and therapeutic interventions.

Introduction

Progressive supranuclear palsy (PSP) is a neurodegenerative disease that is pathologically defined by the accumulation of aggregated tau protein in multiple cortical and subcortical regions, especially involving the basal ganglia, dentate nucleus of the cerebellum midbrain¹. An isoform of tau harboring 4 repeats of microtubule-binding domain (4R-tau) is particularly prominent in these tau aggregates². Clinical manifestations of PSP include a range of phenotypes, including the initially described and most common, PSP-Richardson syndrome that presents with multiple features, including postural instability, vertical supranuclear palsy, and frontal dementia. However, there are several other phenotypes, such as PSP-Parkinsonism, PSP-Frontotemporal dementia, PSP-freezing of gait, PSP-speech and language disturbances, etc³. Presentation of these phenotypes varies widely depending on the distribution and severity of the pathology^4,5.

Currently, the most recognized genetic risk locus for PSP is at the H1/H2 haplotype region covering MAPT gene at chromosome 17q21.31⁶, where individuals carrying the common H1 haplotype are more likely to develop PSP with an estimated odds ratio (OR) of 5.6⁷. Previous studies usually ascribed the observed association in the H1/H2 haplotype to MAPT^6,8,9. However, recent functional dissection of this region using multiple parallel reporter assays coupled to CRISPRi demonstrated multiple risk genes in the area in addition to MAPT, including KANSL1 and PLEKMHL1¹⁰. Genome-wide association studies (GWASs) in PSP have identified common variants in STX6, EIF2AK3, MOBP, SLCO1A2, DUSP10, RUNX2, and LRRK2 with moderate effect size^7,11–13. In addition, variants in TRIM11 were identified as a genetic modifier of the PSP phenotype when comparing PSP with Richardson syndrome to PSP without Richardson syndrome¹⁴.

To date, no comprehensive analysis of single nucleotide variants (SNVs), small insertions and deletions (indels), and structural variants (SVs) in PSP by whole genome sequencing has been conducted. To gain a more comprehensive understanding of the genetic underpinnings of PSP, we performed whole genome sequencing (WGS) and analyzed SNVs, indels and SVs. As a result, we not only validated previously reported genes but also unveiled new loci that provide novel insights into the genetic basis of PSP.

Results

Common SNVs and indels associated with PSP

We conducted whole genome sequencing at 30x coverage (Methods) in 4,662 European-ancestry samples (1,718 individuals with PSP of which 1,441 were autopsy confirmed and 277 were clinically diagnosed and 2,944 control subjects, Table 1). We successfully replicated the association of known loci at MAPT, MOBP and STX6^7,11,12 and identified a novel signal in APOE with a genome-wide significance of P < 5 × 10⁻⁸ (Fig. 1, Fig. S1, Table 2, Table S1). Furthermore, eight loci showed suggestive significance (5 × 10 ⁻⁸ < P < 1 × 10⁻⁶), including two loci reported genome-wide significant (SLCO1A2 and DUSP10) and one locus (SP1) reported suggestive significant in previous studies^11,12, as well as five new loci in FCHO1/MAP1S, KIF13A, TRIM24, ELOVL1 and TNXB.

Table 1.

Characteristics of study participants.

	PSP (n = 1,718)		Control (n = 2,944)
	Autopsy Confirmed (n = 1,441)	Clinical Diagnosed (n = 277)
Female	625 (43%)	129 (46%)	1,775 (60%)
Age, y (SD)	68.38 (8.22)	65.72 (7.68)	81.19 (6.01)
APOE ε4 a
ε4 carriers	350 (24%)	57 (21%)	905 (32%)
Non-ε4 carriers	1,085 (75%)	216 (78%)	1,913 (65%)
Data missing	6 (0.42%)	4 (1%)	126 (4%)
APOE ε2b
ε2 carriers	234 (16%)	36 (13%)	220 (8%)
Non-ε2 carriers	1,193 (83%)	238 (86%)	2,522 (86%)
Data missing	14 (1%)	3 (1%)	202 (7%)
H2 c
H2 carriers	158 (11%)	27 (10%)	1,182 (40%)
Non-H2 carriers	1,283 (89%)	250 (90%)	1,761 (60)
Data missing	0 (0%)	0 (0%)	1 (0.03%)

^aAPOE ε4 is represented by the genotypes of rs429358-C.

^bAPOE ε2 is represented by the genotypes of rs7412-T.

^cH2 haplotype is determined by the genotypes of rs8070723-G.

SD, standard deviation.

Fig. 1: Manhattan plot of SNVs/indels for PSP.

Loci with a suggestive or genome-wide significant signal are annotated (novel loci in red and known loci in black). Variants with a P- value below 1 × 10⁻¹⁴ are not shown. The red horizontal line represents genome-wide significance level (5 × 10⁻⁸). The blue horizontal line represents suggestive significance level (1 × 10⁻⁶).

Table 2.

Genome-wide and suggestive significant loci.

SNV	Chr	Position	Ref	Alt	AF (Alt)	β (Alt)	P	Gene	eQTL/sQTL
Genome-wide Significance (P < 5 × 10 −8)
rs62057121	17	45823394	G	A	0.15	−1.32	7.45 × 10−78	MAPT	LRRC37A4Pc*
rs4420638	19	44919689	A	G	0.20	−0.57	2.91 × 10−19	APOE	TOMM40b
rs7412	19	44908822	C	T	0.06	0.87	9.57 × 10−16	APOE
rs11708828	3	39458158	C	T	0.46	−0.35	7.04 × 10−12	MOBP	PRSAc
rs10753232	1	180980990	C	T	0.44	0.31	6.79 × 10−10	STX6	STX6a*
Suggestive Significance (P < 1 × 10 −6)
rs56251816	19	17750888	A	G	0.22	0.35	6.57 × 10−08	FCHO1/MAP1S
rs12817984	12	53410523	T	G	0.16	−0.37	8.91 × 10−08	SP1	SP1a*
rs4712314	6	17833813	G	T	0.51	0.27	2.37 × 10−07	KIF13A
rs74651308	12	21323155	G	A	0.07	0.51	2.86 × 10−07	SLCO1A2
rs111593852	7	138449166	C	T	0.02	0.87	3.75 × 10−07	TRIM24
rs367364	6	32052169	C	T	0.13	−0.37	7.07 × 10−07	TNXB	CYP21A1Pc*
rs839764	1	43367703	T	A	0.41	0.27	7.94 × 10−07	ELOVL1	TIE1a*
rs 1202665 9	1	221976623	G	A	0.21	0.31	9.48 × 10−07	DUSP10

Chr, chromosome; Ref, reference allele; Alt, alternative allele; AF, allele frequency.

^*Represents the SNV regulates multiple genes, and the gene with the smallest P-value was shown here (eQTL/sQTL for the brain region was obtained through GTEx).

^aSNVs with significant eQTL hits.

^bSNVs with significant sQTL hits.

^cSNVs with both eQTL and sQTL hits.

MAPT, MOBP and STX6

In the MAPT region, a multitude of SNVs and indels in high linkage disequilibrium (LD) with the H1/H2 haplotype remains the most significant association with PSP (Fig. S2A). From our analysis, the prominent signal within the MAPT region is rs62057121 (P = 7.45 × 10⁻⁷⁸, β = −1.32, MAF = 0.15). Fine mapping (Methods) suggests that rs242561 (P = 4.49 × 10⁻⁷⁴, β = −1.23, MAF = 0.16) is likely to be a causal SNV underling the statistical significance. The SNP rs242561 is located in an enhancer region, containing an antioxidant response element that binds with NRF2/sMAF protein complex. The T allele of rs242561 showed a stronger binding affinity for NRF2/sMAF in ChIP-seq analysis, therefore inducing a significantly higher transactivation of the MAPT gene¹⁵. rs242561 and rs62057151 were both in high LD (r² > 0.9) with H1/H2 (defined by the 238 bp deletion in MAPT intron 9) and represented the same association signal as the H1/H2. However, in previous studies^16,17, the H1c tagging SNV (rs242557) inside the H1/H2 region was found to be significant when conditioning on H1/H2. We confirmed that rs242557 was genome-wide significant after adjusting for H1/H2 (P = 3.68 × 10⁻¹⁵, β = 0.39, MAF = 0.42) though in weak LD with H1/H2 (r² = 0.14). To pinpoint the causal genes underlying the association in H1/H2 requires additional functional study. For example, Cooper et al.¹⁰ analyzed transcriptional regulatory activity of SNVs and suggested PLEKHM1 and KANSL1 were probable causal genes in H1/H2 besides MAPT. In MOBP (rs11708828, P = 7.04 × 10⁻¹², β = −0.35, MAF = 0.46, Fig. S2B) and STX6 (rs10753232, P = 6.79 × 10⁻¹⁰, β = 0.31, MAF = 0.44, Fig. S2C), the associated variants were of high allele frequency and exhibited moderate effect size.

APOE and risk of PSP

One newly identified significant locus from our analysis is the well-known Alzheimer’s Disease (AD) risk gene, APOE. We observed a significant association between the APOE ε2 haplotype and an elevated risk of PSP (P = 9.57 × 10⁻¹⁶, β = 0.87, MAF = 0.06, Table 3, Fig. S3B). The APOE ε2 haplotype is encoded by rs429358-T and rs4712-T, which is considered a protective allele in AD. The increased risk of APOE ε2 in PSP has been previously reported in a Japanese cohort, albeit with a relatively small sample size¹⁸. Furthermore, Zhao et al.¹⁹ confirmed that APOE ε2 is linked to increased tau pathology in the brains of individuals with PSP and reported a higher frequency of homozygosity of APOE ε2 in PSP with an odds ratio of 4.41. Consistent with these findings, our dataset exhibited a higher frequency of homozygosity of rs7412-T in PSP, yielding an odds ratio of 3.91.

Table 3.

Allele Frequency of APOE ε4 SNV (rs429358) 230 and ε2 SNV (rs7412)

Studies	rs429358		rs7412
	AF (Case)	AF (Control)	AF (Case)	AF (Control)
PSP WGS (This study)	0.1279	0.1742	0.0844	0.0414
PSP GWAS25	0.1159	0.1366	0.0826	0.0794
1000 Genomes Project26		0.1512		0.0771
ExAC European (non-Finnish)27		0.2078		0.1060
gnomAD V4 European (non-Finnish)28		0.1506		0.0783
TOPMed Freeze 8 NFE (Non-Finnish European)		0.1501		0.0752
ADSP R3 Non-Hispanic White29	0.3139 (AD as cases)	0.1803	0.0244 (AD as cases)	0.0406

For APOE ε4 allele, contrary to its association with AD, we observed that rs429358-C exhibits a protective effect against PSP (P = 5.71 × 10⁻¹⁸, β = −0.60, MAF = 0.16, Table 3). The lead SNV demonstrating this protective association from our analysis is rs4420638 (P = 2.91 × 10⁻¹⁹, β = −0.57, MAF = 0.20, Fig. S3A), which is in LD (r² = 0.74) with rs429358. In a previous PSP GWAS conducted by Hoglinger et al.⁷, another APOE ε4 tagging SNV (rs2075650, r² = 0.52 with rs429358) was also found to be diminished (MAF_case = 0.11 and MAF_control = 0.15) in PSP, although not reaching significance (P = 1.28 × 10⁻⁵). Notably, in our analysis, rs2075650 reached genome-wide significance (P = 3.39 × 10⁻¹³, β = −0.51, MAF = 0.15). APOE ε4 or ε2 displayed an independent effect for PSP risk without a significant epistatic interaction with H1/H2 haplotype (P > 0.05) (Fig. S4).

Given that our dataset included external controls from ADSP collected for Alzheimer’s disease studies, there were a potential selection biases for APOE ε4 and ε2 in controls. To address this concern, we broke down the allele frequencies of APOE ε4 and ε2 by cohorts (Table S2) and indicated cohorts with potential selection bias. The association analysis excluding these cohorts (Methods) shows the ε2 SNV (rs7412, P = 1.23 × 10⁻¹², β = 0.70, MAF = 0.06) remained genome-wide significant and ε4 SNV (rs429358, P = 0.02, β = −0.16, MAF = 0.14) was nominal significant (Table S3, Table S4).

Suggestive significant loci

Eight loci were suggestive of significance in our analysis of which three, SLCO1A2, DUSP10, and SP1, were previously reported^11,12. In SLCO1A2, the lead SNV rs74651308 (P = 2.86 × 10⁻⁷, β = 0.51, MAF = 0.07, Fig. S5A) is intronic and in LD (r² = 0.98) with missense SNV rs11568563 (P = 1.45 × 10⁻⁶, β = 0.47, MAF = 0.07), which was reported in a previous study¹¹. About 250 kb upstream of DUSP10 lies the previously reported SNV rs6687758¹¹ (P = 3.36 × 10⁻⁶, β = 0.29, MAF = 0.21), which is in LD (r² = 0.98) with the lead SNV rs12026659 in our analysis (P = 9.48 × 10⁻⁷, β = 0.31, MAF = 0.21, Fig. S5B). In SP1, the reported indel rs147124286¹² (P = 4.39 × 10⁻⁷, β = −0.35, MAF = 0.16) is in LD (r² = 0.995) with the lead SNV rs12817984 (P = 8.91 × 10⁻⁸, β = −0.37, MAF = 0.16, Fig. S5C). Notably, disruption of a transcriptional network centered on SP1 by causal variants has been implicated previously in PSP¹⁰.

Five newly discovered suggestive loci are in FCHO1/MAP1S, KIF13A, TRIM24, TNXB, and ELOVL1. Within FCHO1/MAP1S, the most significant signal (rs56251816, P = 6.57 × 10⁻⁸, β = 0.35, MAF = 0.22, Fig. S6A) is in the intron of FCHO1. rs56251816 is a significant expression quantitative trait locus (eQTL) for both FCHO1 and MAP1S (13 kb upstream of FCHO1) in the Genotype-Tissue Expression (GTEx) project²⁰. MAP1S encodes a microtubule associated protein that is involved in microtubule bundle formation, aggregation of mitochondria and autophagy²¹, and therefore, is more relevant than FCHO1 regarding PSP. KIF13A, which encodes a microtubule-based motor protein was also suggestive of significance (rs4712314, P = 2.37 × 10⁻⁷, β = 0.27, AF = 0.51, Fig. S6B). The significance in genes involved in microtubule-based processes, such as MAPT, MAP1S and KIF13A, implicates the neuronal cytoskeleton as a convergent aspect of PSP etiology.

Other variants with suggestive association evidence include TRIM24 (rs111593852, P = 3.75 × 10⁻⁷, β = 0.87, MAF = 0.02, Fig. S7A). TRIM24 is involved in transcriptional initiation and shows differential expression in individuals with Parkinson disease^22,23. Another suggestive locus is TNXB, located in the major histocompatibility complex (MHC) region on chromosome 6, with the lead SNV rs367364 (P = 7.07 × 10⁻⁷, β = −0.37, MAF = 0.13, Fig. S7B). Finally, ELOVL1 yields suggestive evidence of association (rs839764, P = 7.94 × 10⁻⁷, β = 0.27, MAF = 0.41, Fig. S7C). This gene encodes an enzyme that elongates fatty acids and can cause a neurological disorder with ichthyotic keratoderma, spasticity, hypomyelination and dysmorphic features²⁴. Furthermore, we found a few SNV/indels that reached genome-wide or suggestive significance without other supporting variants in LD (Fig.S1, Table S1). These signals could be due to sequencing errors and need further experimental validation.

Rare SNVs/indels and network analysis

The heritability of PSP for common SNVs and indels (MAF > 0.01) was estimated to be 20%, while common plus rare SNVs/indels was estimated to be 23% from our analysis using GCTA-LDMS³⁰ (Methods). Therefore, we performed aggregated tests for rare SNVs and indels, and identified ZNF592 (SKAT-O FDR=0.043, burden test FDR=0.041) with an of OR = 1.08 (95% CI: 1.008–1.16) (Fig. 2, Table 4, Table S5) for protein truncating or damaging missense variants (Methods). There was no genomic inflation with a λ=1.07 (Fig. 2). Risk in ZNF592 was imparted by 16 unique variants, with one splice donor and 15 damaging missense variants (Table S5). ZNF592 has not been previously associated with PSP but showed moderate RNA expression in the cerebellum compared to other tissues from GTEx (Fig. S8). There were no significant genes identified when evaluating protein-truncating variants (PTVs) only or when restricting to loss of function intolerant genes.

Fig. 2: Association analysis of rare SNVs/indels.

A. Manhattan plot for genes with protein truncating variants or damaging missense variants. B. Q-Q plot of gene P-values with protein truncating variants or damaging missense variants.

Table 4.

Association analysis of ZNF592 and the C1 module.

Gene	Variants	Total MAC	Case MAC	Control MAC	Fraction Case	Fraction Control	OR (95% CI)	SKAT-O		Burden
								FDR	P	FDR	P
ZNF592	16	19	8	11	0.0023	0.0018	1.08 (1.01–1.16)	0.044	7.60×10−6	0.041	7.30×10−06
Module	Variants	Total MAC	Case MAC	Control MAC	Fraction Case	Fraction Control	OR (95% CI)	Permutation test	P	Permutation test	P
C1	180	234	101	133	0.029	0.022	1.31 (1.01–1.70)	0.19	0.048	0.078	0.006

Considering that genes do not operate along, but rather within signaling pathways and networks, we and others have shown that better understanding of disease mechanisms can be achieved through gene network analysis^31–33. Therefore, we scrutinized rare variants within a network framework, focusing on co-expression network analysis performed in PSP post mortem brain that had previously identified a brain co-expression module, C1, which was conserved at the protein interaction level and enriched for common variants in PSP³⁴. We found this C1 neuronal module was significantly enriched with PSP rare variants (P = 0.006, OR [95% CI] = 1.31 [1.01–1.70], Table 4; Table S3). Genes from the C1 module were more likely to be loss of function intolerant compared to the background of all brain expressed genes (Methods) (Fig. S8). To ensure that this was association not spurious, we performed permutation testing using random gene modules of brain expressed genes with the same number of genes as C1. The C1 module remains significant (Permutation P = 0.078). Exploring GTEx, we found that C1 genes are highly expressed in brain tissues including the cerebellum, frontal cortex, and basal ganglia (Fig. S8), consistent with regions affected in this disorder.

SVs associated with PSP

Seven high-confident SVs achieved genome-wide significance with PSP (Table 5, Fig. S9), including three deletions tagging the H2 haplotype. The most significant signal is a 238 bp deletion in MAPT intron 9 (Fig. S10A, chr17:46009357–46009595, P = 3.14 × 10⁻⁵⁰, AF = 0.16) that has been reported on the H2 haplotype^35,36 and is in LD (r² = 0.99) with the lead SNV, rs62057121 (chr17:45823394, P = 7.45 × 10⁻⁷⁸, β = −1.32, MAF = 0.15), in the MAPT region. Adding to this, two other deletions, one spanning 314 bp (Fig. S10B, chr17:46146541–46146855, AF = 0.19) and the other covering 323 bp (Fig. S10C, chr17:46099028–46099351, AF = 0.22), both are Alu elements and in LD (r² > 0.8) with the top signal (the 238 bp deletion). This observation indicates that transposable elements may play an important role in the evolution of H1/H2 haplotype structure.

Table 5.

Significant structural variants from association analysis (P < 5 × 10⁻⁸).

Name	N	AF	beta	P	AF (case)	AF (control)	Odds Ratio	Fisher’s P	Gene
chrl 7:46009357–46009595 :DEL*	4357	0.16	−1.22	3.14×10−50	0.054	0.23	0.19	5.80×10−118	MAPT
chrl 7:46146541–46146855 :DEL*	3697	0.19	−1.12	2.13×10−39	0.079	0.25	0.26	1.58×10−83	KANSL1
chrl 7:46099028–46099351 :DEL*	3699	0.22	−1.07	3.88×10−37	0.11	0.28	0.33	2.05×10−66	KANSL1
chr14:105864208–105916743:DEL	4378	0.010	−1.53	4.74×10−14	0.0053	0.014	0.39	1.33×10−04	IGH
chr6:149762615–149763234:DEL	3811	0.55	0.50	8.60×10−12	0.75	0.42	4.19	6.00×10−182	PCMT1
chr19:41102802–41104285 :DEL	2921	0.17	0.64	7.46×10−09	0.21	0.14	1.59	5.95×10−11	CYP2A13
chr1:152880979–152880979:INS	2872	0.74	0.67	2.37×10−08	0.79	0.71	1.62	1.46×10−13	SMCP

^*Represents SVs with DNA samples available and PCR validated

Beyond the identified SVs in the H1/H2 region, we uncovered a significant deletion (chr14:105864208–105916743, P = 4.74 × 10⁻¹⁴, AF = 0.01) within the immunoglobulin heavy locus (IGH), which is a complex SV region (Fig. S11) related to antigen recognition. Moreover, a 619 bp deletion (chr6:149762615–149763234, P = 8.60 × 10⁻¹², AF = 0.55; Fig. S10D) in PCMT1 displayed increased risk of PSP with an odds ratio of 4.19. The odds ratio increased to 8.38 when comparing 1,244 individuals with homozygous deletions in PCMT1 with the rest of sample set. PCMT1 encodes a type class of protein carboxyl methyltransferase enzyme that is highly expressed in the brain³⁷ and is able to ameliorate Aβ_25–35 induced neuronal apoptosis^38,39. Additionally, we found a deletion between CYP2F1 and CYP2A13 (chr19:41102802–41104285, AF = 0.17) and an insertion in SMCP (chr1:152880979–152880979, AF = 0.74) which were also significant (Table 5). The 1.5 kb deletion (chr19:41102802–41104285) almost completely overlaps the SINE-VNTR-Alus (SVA) transposon region annotated by RepeatMasker⁴⁰.

SVs in H1/H2 haplotype region

The H1/H2 region stands out as the pivotal genetic risk factor for PSP^17,41. The H2 haplotype exhibits a reduced odds ratio of 0.19, as we observed the allele frequency of the 238 bp H2-tagging deletion is 23% in PSP and only 5% in control (P < 2.2 × 10⁻¹⁶). Moreover, our analysis pointed out five common (MAF > 0.01) and 12 rare deletions and duplications in the region (Table 6), ranging from 88 bp to 47 kb. Additionally, one common and four rare high-confidence insertions were reported in the region.

Table 6.

High-confident structural variants in the H1/H2 haplotype region

Name	Size	N	AF	AF (PSP)	AF (Control)	Gene	Annotation
chrl 7:46099028–46099351 :DELa*	323	3,699	0.24	0.11	0.28	KANSL1	intron
chr17:46146541–46146855:DELa*	314	3,697	0.21	0.08	0.25	KANSL1	intron
chrl 7:46237619–46238142: DELa	523	3,686	0.19	0.09	0.22	MAPK8IP1P1	intergenic
chr17:46009357–46009595:DELa*	238	4,357	0.19	0.05	0.23	MAPT	intron
chr17:46277789–46282210:DEL	4,421	4,233	0.12	0.03	0.15	ARL17B	intron
chr17:46113802–46113802:INS	311	2,464	0.31	0.32	0.32	KANSL1	intron
Name	Size	N	N (Carriers)	N (PSP)	N (Control)	Gene	Annotation
chr17:46811121–46811289:DELa	168	2,614	36	15	21	WNT3	intron
chr17:45847702–45851880:DELa	4,178	4,427	31	17	14	MAPT-AS1	splicing
chr17:46837153–46839088:DELa	1,935	4,415	12	8	4	WNT9B	intron
chr17:45918825–45920861:DELa	2,036	4,422	1	0	1	MAPT	intron
chr17: 45 916681 –45920693 :DEL	4,012	4,430	3	0	3	MAPT	intron
chr17:45570198–45572012:DEL	1,814	4,243	3	2	1	AC091132.4	intron
chr17:45334194–45381549:DELa	47,355	4,430	1	0	1	AC003070.2	transcript ablation
chr17:45311955–45312258:DEL	303	4,365	2	0	2	MAP3K14	intron
chr17:45894637–45914976:DUPa	20,339	4,260	1	1	0	MAPT-AS1	transcript amplification
chr17:45993882–45993970:DELa	88	4,283	1	1	0	MAPT	splicing
chr17:45665996–45666370:DELa	374	4,412	1	1	0	LINC02210-CRHR1	TFBS ablation
chr17:45879141–45881180:DEL	2,039	4,431	1	1	0	MAPT-AS1	intron
chr17:45741582–45741582:INS	315	4,420	10	4	6	LINC02210-CRHR1	intergenic
chr17:45929579–45929579:INS	453	3,025	5	1	4	MAPT	intron
chr17:46754483–46754483 :INS	330	3,692	12	2	10	NSF	intron

AF, allele frequency; N, number of individuals with non-missing genotypes.

^*High-quality SVs that were included in association analysis.

^aRepresents SVs with DNA samples available and PCR validated.

Of the five common deletions and duplications (Fig. S12), three show genome-wide significant association with the disease (Table 2); four are located in regions with transposable elements (SVA, L1, or Alu) and in LD (r² from 0.63 to 0.92) with the 238 bp H2-tagging deletion (Methods). This further highlights the important role of transposable elements in shaping the landscape of H1/H2 region.

Among the 12 rare deletions and duplications (Fig. S13), five are located in potentially functional regions, such as splice sites, exons, and transcription factor binding sites (Table 6). Particularly, one deletion (chr17:45993882–45993970) in exon 9 of MAPT was identified in a PSP patient, adding to previous reports of exonic deletions in the MAPT in frontotemporal dementia, such as deletion of exon 10⁴² and exons 6–9⁴³ in MAPT. Using the SKAT-O test (N = 4,432), the 12 rare CNVs displayed a significantly higher burden in PSP than controls (P = 0.01, OR = 1.64).–

Discussion

Through comprehensive analysis of whole genome sequence, we identified SNVs, indels and SVs contributing to the risk of PSP. For common SNVs, previously reported regions, including MAPT, MOBP, STX6, SLCO1A2, DUSP10, and SP1^7,11,12 were replicated in our analysis and novel loci in APOE, FCHO1/MAP1S, KIF13A, TRIM24, ELOVL1, and TNXB were discovered. EIF2AK3 which was significantly associated with PSP in a previous GWAS¹⁷ did not reach significance in our study. The SNV with the lowest P around EIF2AK3 was rs13003510 (P = 8.30 × 10⁻⁵, β = 0.22, MAF = 0.3).

The APOE ε4 haplotype was of particular interest as it is a common risk factor for AD, explaining more than a 1/3 of population attributable risk^44,45. Typically, individuals with one copy of the APOE ε4 allele (rs429358-C and rs4712-G) have approximately a threefold increased risk of developing AD, while those with two copies of the allele have an approximately a 12-fold increase in risk⁴⁶. In striking contrast, the ε4 tagging allele rs429358 was protective in PSP and the ε2 tagging allele rs7412 was deleterious. This observation is particularly intriguing since both AD and PSP have intracellular aggregated tau as a prominent neuropathologic feature. Notably, both ε2 allele and ε4 allele have been associated with tau pathology burden in the brain of mice models^19,47, which raises the question of distinct tau species in 4R-PSP versus 3R-4R-AD. It is also notable that the ε2 allele is also associated with increased risk for age-related macular degeneration (AMD), and the ε4 allele was associated with decreased risk^48,49. These results demonstrate that the same variant may have opposite effects in different degenerative diseases. This is especially important, given the advent of gene editing as a therapeutic modality, and programs focused on changing APOE ε4 to ε2. Although this therapy would likely decrease risk for AD, our results indicate that it would increase risk for PSP, in addition to AMD. From this standpoint, caution is warranted in germ-line genome editing until the broad spectrum of phenotypes associated with human genetic variation is understood.

Burden association tests are an highly valuable for addressing sample size limitations in analyzing rare variats⁵⁰. Indeed, burden testing allowed us to identify ZNF592, a classical C2H2 zinc finger protein (ZNF)^51,52, as a candidate risk gene. ZNF proteins have been causative or strongly associated with large numbers of neurodevelopmental disease^53,54 and neurodegenerative disease including Parkinson’s disease⁵⁵ and Alzheimer’s disease^56,57. ZNF592 was initially thought to be responsible for autosomal recessive spinocerebellar ataxia 5 from a consanguineous family with neurodevelopmental delay including cerebellar ataxia and intellectual disability due to a homozygous G1046R substitution⁵⁸. However, further analysis of this family identified WDR73 to be the most likely causative gene, consistent with Galloway-Mowat syndrome, although ZNF592 may have contributed to the phenotype⁵⁹.

We also extended classical gene-based burden analysis to consider rare risk burden in the context of a gene set defined by co-expression networks^34,60. We leveraged combined previous proteomic and transcriptomic analysis of post-mortem brain from patients afflicted with PSP, and showed that rare variants enrich in the C1 neuronal module, which was the same module enriched with common variants³⁴. This, along with our recent work identifying a neuronally-enriched transcription factor network centered around SP1 disrupted by PSP common genetic risk, suggests that although PSP neuropathologically is defined by tufted astrocytes and oligodendroglial coiled bodies^61–63, initial causal drivers of PSP appear to be primarily neuronal.

In analysis of SVs, we found deletions in PCMT1 and IGH were significantly associated with PSP. The IGH deletions are in a complex region on chromosome 14 that encodes immunoglobins recognizing foreign antigens. The size of the IGH deletion varies across individuals (Fig. S9). In addition, the IGH deletions can be accompanied by other deletions, duplications, and inversions (Fig. S9). These combined make the experimental validation of the deletion challenging. The PCMT1 deletion is common (AF = 0.55) with an odds ratio of 8.38 for PSP in homozygous individuals.

There were limitations to this study. Not all PSP were pathologically confirmed, although pathological confirmation was available in a significant subset (of the 1,718 PSP individuals, 1,441 were autopsy-confirmed and 277 were clinically-diagnosed). Additionally, the majority of control samples in this study were from ADSP and were initially collected as controls for AD studies. As ADSP is a dataset composed of multiple cohorts from diverse sources, it is imperative to ensure that any observed allele frequency differences between controls and cases can be attributed to the disease itself rather than sample selection biases arising from technical artifacts or batch effects. To mitigate the risk of false reports, we meticulously examined the allele frequencies of both cases and controls, especially in relation to novel and significant signals.

This work represents an important first step; future work is necessary to further delineate the rare genetic risk in PSP harbored in coding and noncoding regions. These results may come to fruition as additional genomic analytical methods are developed, sample size increased, and orthogonal genomic data are integrated. While PSP is rare, it is the most common primary tauopathy, and studying this disease is critical to understanding common pathological mechanisms across tauopathies. Further work to include individuals with diverse ancestry background will also improve our understanding of genetic architecture of the disease.

Methods

Study subjects

We performed WGS at 30x coverage (Table S7) for 1,834 PSP cases and 128 controls from the PSP-NIH-CurePSP-Tau, PSP-CurePSP-Tau, PSP-UCLA, and AMPAD-MAYO cohorts included in ADSP (ng00067) and used 3,008 controls from ADSP⁶⁴. Control subjects were self-identified as non-Hispanic white. WGS data is available on NIAGADS⁶⁵. We removed related subjects (IBD>0.25), five clinically diagnosed PSP who were not found to have PSP on autopsy, and non-Europeans (subjects that were eight standard deviations away from the 1000 Genomes Project European samples^26,66 using the first six principal components), resulting in 1,718 individuals with PSP and 2,944 control subjects. Of the 1,718 PSP individuals, 1,441 were autopsy-confirmed and 277 were clinically-diagnosed (Table 1).

Given that our dataset included external controls from ADSP collected for Alzheimer’s disease studies, there was a potential selection biases for APOE ε4 and ε2 in controls. We broke down the allele frequencies of APOE ε4 and ε2 by cohorts (Table S2) and reviewed the study design of each cohort. The ADSP-FUS1-APOEextremes study used an age extremes sampling approach stratified by APOE genotype, comparing younger onset AD cases against older cognitively normal controls: the controls were APOE ε4/ε4 controls with age-at-last-assessment ≥ 75 years, APOE ε3/ε4 controls with age-at-last-assessment ≥ 80 years, or APOE ε3/ε3 controls with age-at-last-assessment ≥ 85 years⁶⁴. The ADSP-FUS1-StEPAD1 study aims to identify and characterize novel genetic variants that promote resilience to AD pathology in the presence of the APOE4 allele: controls from ADSP-FUS1-StEPAD1 were protected APOE4 carriers have normal cognition at older age⁶⁴. The CacheCounty study selects “AD resilient individuals” and define them as individuals who are at least 75 years old, cognitively normal, and carry at least one APOE ε4 allele⁶⁷.

Common SNVs/indels analysis

Only biallelic variants were included in common SNVs/indels analysis. Variants were removed if they were monomorphic, did not pass variant quality score recalibration (VQSR), had an average read depth ≥ 500, or if all calls have DP<10 & GQ<20. Individual calls with a DP<10 or GQ<20 were set to missing. Indels were left aligned using the GRCh38 reference^68,69. Common variants (MAF > 0.01) with a missing rate < 0.1, 0.25 < ABHet < 0.75 and HWE (in control) > 1 × 10⁻⁵ were kept for analysis, leaving 7,945,112 SNVs/indels for analysis. Genetic relatedness matrix was obtained using KING⁷⁰. Principal components were obtained by PC-AiR⁷¹ which accounts for sample relatedness. Linear mixed model implemented in R Genesis⁷² were used for association. Sex and PC1–5 were adjusted in the linear mixed model. Age was not adjusted as more than half (1,159 of 1,718) of PSP cases had age missing. After association, variants with a P < 1 × 10⁻⁶ were reported. For SNVs/indels without supporting evidence from nearby SNVs/indels in LD, we removed possible spurious calls with FS (Phred-scaled P-value using Fisher’s exact test to detect strand bias) > 4, VQSLOD (Log odds of being a true variant versus being false under the trained gaussian mixture model) < 15, or located in regions of genome showing discrepancy from Telomere-to-Telomere Consortium⁷³ and GRCh38. Fine-mapping of the H1/H2 region were analyzed using SuSie⁷⁴. We ran the analysis several times assuming the number of maximum causal variants were from 2 to 10. The only variant (rs242561) robust to the choice of maximum causal variants was reported. To avoid potential confounding effects (particularly for APOE alleles), we also performed association analysis (Table S4) for suggestive and genome-wide significant signals when excluding subjects from the three cohorts with selection bias against APOE alleles along with cohorts with less than 10 subjects (NACC-Genentech, FASe-Families-WGS, KnightADRC-WGS).

Rare SNVs/indels analysis

Multi-allelic variants were split into biallelic variants. Variants where ALT=*, representing a spanning deletion, were removed. Biallelic and multiallelic variants were concatenated, and duplicated variants were removed. Variants were removed if they were monomorphic, did not pass VQSR, had an average read depth ≥ 500, or if all calls have DP<10 & GQ<20. Individual calls with a DP<10 or GQ<20 were set to missing. Indels were left aligned using the GRCh38 reference^68,69. Then, variants with a missing rate > 0.1 or a P_HWE < 1 × 10⁻⁷ in controls were removed, resulting in 91,863,622 variants. We calculated the heritability of PSP using GCTA-LDMS³⁰ for common SNVs/indels (MAF > 0.01) and common plus rare SNVs/indels. A prevalence of 5 PSP cases per 100,000 individuals (0.00005) was used in the GCTA-LDMS analysis.

For aggregated tests of rare variants, we considered rare protein truncating variants (PTVs) and PTVs/damaging missense variants. Variant were annotated with ANNOVAR (version 2020–06-07)⁷⁵ and Variant Effect Predictor (VEP, version 104.3)⁷⁶. PTVs were in protein coding genes (Ensembl,version 104)⁷⁷ and had VEP consequence as stop gained, splice acceptor, splice donor or frameshift. Damaging missense variants were in protein coding genes (Ensembl version 104)⁷⁷ and had a VEP consequence as missense, CADD score ≥ 15, and PolyPhen-2 HDIV of probably damaging. Rare variants were selected based on a MAF < 0.01% from gnomAD and a MAF < 1% in our dataset. The number of alternative allele variants in protein truncating variants (PTV) and PTV/damaging missense variants was similar across sequencing centers and when evaluated for loss of function intolerant genes (observed/expected score upper confidence interval < 0.35⁷⁸) (Fig. S14)

After LD clumping with a r² cutoff of 0.2, we applied the bigsnpr R package to perform PCA using variants with MAF > 1%. We tested if genes with PTVs or PTVs/missense variants were associated with PSP using the sequence kernel association test-optimized (SKAT-O)⁷⁹ (SKAT R package version 2.0.1)⁸⁰. We used a linear kernel and weighed each variant by the maximum external database MAF where lower MAF would have higher weight. We normalized variant MAFs, where $MA{F}_{norm}=MA{F}_{ext}/MA{F}_{max}$, where $MA{F}_{ext}$ is the external database MAF from gnomAD⁷⁸ and $MA{F}_{max}=0.0001$. The variant weight is defined by the $MA{F}_{norm}$ on the β (1,4) distribution. Thus, variant weight is high at very low $MA{F}_{norm}$ and spread across the range of 1 to 4. Covariates included sex, PC1–3, and H1/H2 haplotype. P-values were FDR corrected for the number of genes with a total minor allele count (MAC) ≥ 10. In addition to SKAT-O, we performed gene burden testing (SKATBinary method=‘burden’). As SKAT-O does not calculate an odds ratio, we calculated the odds ratio of significant genes using logistic regression with the same covariates as SKAT-O and burden testing, and the same variant weights. We also considered only PTVs or PTVs/missense variants in loss of function intolerant genes (observed/expected score upper confidence interval < 0.35⁷⁸).

We evaluated the C1 module, a gene set, which was previously shown to be composed of neuronal genes and enriched for common variants in PSP³⁴. We performed a permutation test (N=1000) of random gene set modules from brain expressed genes that contained the same number of genes as C1. From the human protein atlas (www.proteinatlas.org)⁸¹, brain expressed genes were defined as the union of unique proteins from the cerebral cortex, basal ganglia and midbrain (N=15,638). We calculated SKAT-O P-values from these random gene modules to determine the null distribution. We calculated the unadjusted odds ratio of significant genes or gene sets by summing the number of alternate alleles in the gene set among the total number alleles in cases and controls。

Normalized quantification (TPM) gene expression across tissues was obtained from Genotype-Tissue Expression (GTEx)⁸². The expression of ZNF592 and C1 module (summarized as an eigengene⁸³) were plotted.

SV detection and filtering

For each sample, SVs were called by Manta⁸⁴ (v1.6.0) and Smoove⁸⁵ (v0.2.5) with default parameters. Calls from Manta and Smoove were merged by Svimmer ⁸⁶ to generate a union of two call sets for a sample. Then, all individual sample VCF files were merged together by Svimmer as input to Graphtyper2 (v2.7.3)⁸⁶ for joint genotyping. SV calls after joint-genotyping are comparable across the samples, therefore, can be used directly in genome-wide association analysis⁸⁶. A subset of SV calls was defined as high-quality calls⁸⁶. Details of SV calling pipeline were in our previous study⁸⁷.

There are regions in the human genome that tend to have anomalous, or high signal in WGS experiments⁸⁸. SVs that reside in those regions can be unreliable and should be reported. Specifically, we compiled problematic regions in the genome from the following sources: (1) the ENCODE blacklist: a comprehensive set of regions that could result in erroneous signal⁸⁹; (2) the 1000 Genome masks: regions of the genome that are more or less accessible to next generation sequencing methods using short reads; (3) the set of assembly gaps defined by UCSC; (4) the set of segmental duplications defined by UCUC; (5) the low-complexity regions, satellite sequences and simple repeats defined by RepeatMasker⁴⁰. For each individual SV reported, Samplot⁹⁰ or IGV⁹¹ were used to keep only high-confident CNVs and inversions that are supported by read depth or split reads; for insertions, we kept high-confident insertions that are high-quality and not in the masked regions.

SV analysis

For SV association, more strict sample filtering was applied: outlier samples with too many (larger than median + 4*MAD) CNV/insertion calls or too little (smaller than median - 4*MAD) high-quality CNV/insertion calls were removed. There were 4,432 samples (1,703 cases and 2,729 controls) remaining for PSP SV association analysis. Due to more false positives being picked up, the genomic inflation would be high (λ = 1.89, Fig. S9) if all SVs were included in the analysis. Therefore, we restricted our analysis to high-quality SVs only, making the genomic inflation drop to 1.27 (Fig. S9). The 14,792 high-quality common SVs (MAF > 0.1) with call rate > 0.5 were included in the analysis. Mixed model implemented in R Genesis⁷² were used for association. Sex, PCR information, SV PCs 1–5, and SNV PCs 1–5 were adjusted in the mixed model. After association, we manually inspect deletions, duplications, and inversions by Samplot⁹⁰ or IGV⁹¹ to keep only those with support from read depth, split read or insert size. For insertions, those not on masked regions were reported.

For SVs inside the H1/H2 region, all SVs those that are not high-quality are included. Then, we removed SVs with missing rate > 0.5 and manual inspect deletions, duplications, and inversions by Samplot⁹⁰ or IGV⁹¹ to keep only those with support from read depth, split read or insert size. For insertions, those high-quality ones not on masked regions were kept for analysis. LD between SVs was calculated using PLINK (V1.90 beta)⁹².Rare SV burden on H1/H2 region was evaluated by SKAT-O⁷⁹ adjusting for gender and PCs 1–5. As SKAT-O does not calculate an odds ratio, we calculated the odds ratio using logistic regression with the same covariates.

Availability of data and materials

NIAGADS Data Sharing Service (https://dss.niagads.org/)

https://github.com/whtop/PSP-Whole-Genome-Sequencing-Analysis