Rare coding variation and stroke heterogeneity in Saudi Arabia: an exome‑wide association study across severity, etiology, vascular territory, and early‑onset disease Authors: Fahad Alkhamis, Majed M Alabdali, Abdullah S Alamri, Rudaynah Alali, Alawi H Habara, Mohammed S Akhtar, Shahad F Alkhamis, Mercy Rophina, Abdullah M Alkhudair, Chittibabu Vatte, Brendan Keating, Amein K Al-Ali Published: 2026 ABSTRACT AbstractBackgroundIschemic stroke in Saudi Arabia arises in a highly consanguineous population with a distinctive genetic architecture, likely enriching rare coding variants that influence stroke risk. Yet the contribution of these variants to stroke susceptibility, age at onset, and subtype patterns in this setting remains incompletely defined.MethodsWe analyzed whole-exome sequencing data from 514 stroke patients in a case only study design. Gene-based rare variant burden analyses was performed using SAIGE-GENE+ burden and SKAT-O tests across predefined phenotype contrasts within the cohort, stroke severity (modified Rankin Scale, mild < 3 vs. severe ≥ 3 ,), age at onset (early-onset < 25 years vs. late-onset > 45 years; mid-life onset < 45 vs. > 45), etiological subtypes (TOAST classification), and vascular imaging patterns (intracranial vs. extracranial).Post-association functional annotations included GTEx expression, gnomAD constraint metrics, and draggability insights.ResultsGene-level associations at a suggestive threshold (p < 0.005) identified several candidates including HSP90AB1, PRR23A, and LRRC42 (severity and age-at-onset); POGZ and SMIM34 (age-at-onset); and COL9A3, DCP1B, and ADGRV1 (imaging and etiology subtypes).The highlighted genes showed varying expression across brain and vascular tissues and intolerance to loss-of-function variation. Notably, HSP90AB1, a molecular chaperone highly expressed in the brain and vasculature, has small-molecule inhibitors, supporting its potential relevance to stroke biology.ConclusionsOur findings identify exploratory candidate gene-level signals across clinically defined ischemic stroke phenotypes based on case-only within-cohort comparisons, in underrepresented population. These results should be considered hypothesis-generating and require replication and functional validation before biological or clinical inferences can be made. Supplementary InformationThe online version contains supplementary material available at 10.1186/s12883-026-04935-0.Keywords: Stroke, Exome sequencing, Variation, Rare coding, Ischemic, Saudi Background Stroke is a leading cause of death and disability worldwide, with ischemic stroke accounting for the majority of cases. Each year, more than 13 million people experience a stroke, and nearly 5.5 million dies as a consequence [1]. Although stroke is often regarded as a disease of older adults in high-income countries, its burden falls disproportionately on low- and middle-income countries, where younger patients are more frequently affected and outcomes are often worse [2]. In the Kingdom of Saudi Arabia, this pattern is evident, stroke represents a growing public health challenge, shaped not only by high prevalence of conventional cardiovascular risk factors but also by unique demographic and genetic features, including high rates of consanguinity that increase the burden of heritable disease [3, 4]. Genetic factors are recognized as central to the development of both cerebrovascular and cardiovascular disease [5, 6]. Twin and family studies suggest substantial heritability for ischemic stroke, while genome-wide association studies (GWAS) have uncovered hundreds of risk loci across diverse stroke subtypes [7]. However, most genomic studies to date have focused on European ancestry cohorts, limiting insights into Arab and Middle Eastern populations, which remain significantly underrepresented. This lack of diversity restricts the transferability of risk models and may obscure variants enriched in these populations. Saudi cohorts, in particular, provide unique opportunities for discovery due to their population structure and high rates of consanguinity, which increase the likelihood of identifying rare, high-impact variants. Experience from studies of myocardial infarction (MI), a related cardiovascular phenotype, illustrates this point. Candidate gene studies in Saudis have implicated PCSK9, LRP8, and HNF1A in MI susceptibility [8–10]. A Saudi-specific GWAS revealed novel signals, including a variant in SHISA5, which is enriched in local populations [10]. A recent systematic review confirmed roles for PCSK9, CETP, and CDKN2B-AS1 in MI risk among Saudis [8–10]. Together, these findings demonstrate that while many European-derived loci replicate, Saudi populations harbour distinct genetic signatures, underscoring the value of population-specific studies. Rare variant studies are particularly informative in this context. Unlike common variants typically detected by GWAS, rare coding variants often exert larger biological effects and may underlie disease heterogeneity. In consanguineous populations such as Saudi Arabia, these alleles are more likely to be observed at higher frequencies, enhancing their detectability. The present study differs from prior Saudi exome-based stroke studies in both analytic focus and phenotypic scope. Rather than examining ischemic stroke as a single broad phenotype, we evaluated rare protein-altering coding variants using gene-based aggregation analyses across clinically relevant subphenotypes, including severity, TOAST-defined etiology, vascular territory, and early-onset disease. Another critical dimension of stroke in Saudi Arabia is its age distribution. Pediatric and early-onset stroke (< 45 years) accounts for a substantial proportion of cases. These patients often face lifelong disability, and the etiologies and genetic contributions to their disease may differ from those of later-onset stroke. Comparing the determinants of early- versus late-onset stroke, therefore, offers an opportunity to disentangle distinct biological pathways and refine risk stratification. Taken together, these observations highlight the importance of characterizing the genetic architecture of ischemic stroke in Saudi populations, with a particular focus on rare variants, disease subtypes, and age of onset. Previous studies from our group have applied whole-exome sequencing to stroke cohorts in Saudi Arabia, including candidate gene analyses and polygenic risk score evaluation [11]. However, systematic gene-based rare variant burden testing across the exome in this population remains limited in this population. The aim of this study, was to identify genes associated with phenotypic heterogeneity among ischemic stroke cases using a case-only rare variant analysis across clinically defined subgroups in a Saudi cohort. We evaluated the contribution of rare coding variants across multiple clinically defined stroke phenotypes, including stroke severity, etiological subtype, vascular territory, and age at onset. By integrating rare variant burden analyses with phenotype-stratified comparisons, our goal was to identify exploratory candidate gene-level signals associated with stroke heterogeneity and to generate hypotheses for future studies of cerebrovascular disease in an understudied population. Understanding genetic risk in these underrepresented populations is essential for improving the generalizability of genomic discoveries and advancing equitable precision medicine. Materials and methods Study Population and phenotype definitions We analyzed ischemic stroke patients recruited through Saudi hospital networks, as previously described [11]. Whole-exome sequencing and initial variant processing for this cohort were reported in that study. The present work performs additional gene-based analyses focused on phenotypic heterogeneity within the stroke case-only cohort. Clinical data included ischemic stroke etiology, vascular imaging findings, and age at onset. All diagnoses were confirmed by neurologists based on clinical presentation and neuroimaging. Patients with incomplete phenotypic data for a given comparison were excluded from that specific analysis. Because the available dataset consisted of stroke cases only, we conducted multiple predefined phenotype-stratified comparisons within the cohort. Stroke severity was defined using the modified Rankin Scale, comparing severe (mRS ≥ 3) cases against mild (mRS < 3) cases. The modified Rankin Scale was dichotomized into mild versus severe outcomes, reflecting commonly used clinical thresholds distinguishing functional independence from disability in stroke outcome studies. Hemorrhagic versus ischemic stroke status was analyzed as a sensitivity contrast, acknowledging the small hemorrhagic sample size. For ischemic stroke, etiological subtypes were assigned using the TOAST classification system and included large artery atherosclerosis, small vessel occlusion, and cardioembolic stroke [12]. Vascular imaging was used to classify cases into predominantly intracranial versus extracranial involvement. Age at onset was stratified into pediatric/early-onset (≤ 25 years), early adult (26–45 years), and late-onset (> 45 years), with the main comparison defined as early-onset (≤ 25 years) versus late-onset (> 45 years). Ethical approval was granted by the Imam Abdulrahman bin Faisal University Institutional Review Board (IRB #2025-01-0207), and the study was carried out in line with the Declaration of Helsinki and Good Clinical Practice guidelines. Informed consent to participate was obtained from all participants included in the study. Written informed consent was obtained using an English consent form with a verified Arabic translation. For children or adults or who could not provide consent because of stroke-related cognitive or physical impairment, consent was provided by a legal guardian or an appropriate representative (next of kin) on the participant’s behalf. No waiver of consent was used for this study. Sequencing and primary variant calling Peripheral blood was collected in EDTA tubes and stored at − 20 °C until processing. Genomic DNA was extracted using the Gentra Puregene Blood Kit (Qiagen, USA) following the manufacturer’s instructions. Whole-exome sequencing (WES) alignment, variant calling, and primary quality control procedures were performed as reported in Alkhamis et al. (2023) [11]. In the present study, we focus on downstream annotation, variant grouping, and gene-based analyses. Variant annotation and grouping for gene-based testing Variants annotation and grouping for rare-variant, association analyses were performed using a Snakemake-based workflow. All analyses were conducted on genome build GRCh38, with gene definition from GENCODE v39. For each variant we reported chromosome, genomic position, reference allele, and alternate allele, to enable reproducibility and cross referencing across datasets even when frequency reporting thresholds precluded detailed allele count release. Annotations were generated with the Ensembl Variant Effect Predictor (VEP) v105, and predicted loss-of-function calls were refined using LOFTEE (v1.04_GRCh38), supplemented by LOFTEE v1.04_GRCh38 for loss-of-function (LoF) prediction [13]. Transcript selection followed the MANE Select transcript wherever available; for genes without a MANE Select transcript, the Ensembl ‘canonical’ transcript for protein-coding genes was used. Consequences were reported relative to the canonical transcript, and LOFTEE annotations were provided for all putative LoF variants, distinguishing high-confidence (HC), low-confidence (LC), and flagged variants. Allele frequency annotation included overall and population-specific frequencies from gnomAD v3.1.2, with reporting of popmax frequencies and filtering based on broad ancestry categories. Internal allele counts (AC), allele numbers (AN), allele frequencies (AF), and the number of homozygotes were also tabulated, stratified by ancestry where possible. In silico deleteriousness scores were appended for all protein-coding variants, including CADD v1.6, PHRED-scaled scores, REVEL scores from dbNSFP v4.3, and SpliceAI v1.3 scores [14–17]. SpliceAI was run using a GPU-enabled fork to accelerate computation, and the maximum donor/acceptor gain/loss score across all categories was retained as the representative splicing annotation. Annotation was performed against the updated GENCODE v39 transcript set for GRCh38 protein-coding genes. Variant quality summary metrics were calculated, which included mean sequencing depth across variant sites, as well as mean allele balance for heterozygous and homozygous genotypes. All code for variant annotation and grouping was implemented in Snakemake with reproducible command logs and environment files, including the use of the Ensembl database v105 (homo_sapiens_core_114_38). For gene-based testing, each variant was assigned to a single mutually exclusive category to avoid double-counting. Given the number of genes and phenotype contrasts evaluated, these analyses were considered exploratory, and suggestive findings were interpreted cautiously. Priority was assigned in descending order. First, high-confidence predicted loss-of-function (pLoF) variants were defined using LOFTEE HC stop-gain, frameshift, and canonical splice-disrupting variants. Second, damaging missense or protein-altering variants were identified as those not already classified as HC-pLoF but with at least one deleteriousness criterion, including a missense, start-loss, stop-loss, or in-frame indel with either REVEL ≥ 0.773 or CADD ≥ 28.1, or a SpliceAI score ≥ 0.2 [17]. These thresholds were prespecified prior to analysis and correspond to calibrated pathogenicity evidence levels derived from probabilistic evaluation of computational predictors within the ACMG/AMP variant interpretation framework [18]. Third, low-confidence LoF variants were retained if classified as LC by LOFTEE. Fourth, other missense or protein-altering variants not meeting the damaging thresholds above were assigned to a separate category. Finally, synonymous variants with SpliceAI DS < 0.2 were annotated as a negative control category. Synonymous variants were included in a separate aggregation mask to provide a reference class of coding variants generally considered functionally neutral, enabling sensitivity analyses and the evaluation of potential background signal in the burden-testing framework. Variants with a maximum allele frequency > 1% in gnomAD (across any broad ancestry group) were removed from all categories. This categorization ensured that each variant was assigned to only one functional group, with precedence rules applied when annotations overlapped. Gene-based rare variant burden association testing We conducted gene-based rare variants association test using SAIGE-GENE (v1.5.0.2) [19], which allows for robust modeling of unbalanced case–subgroup comparisons while accounting for relatedness and population structure. Both burden tests and SKAT-O tests were applied for each gene under each annotation mask. Variants were weighted by their frequency using a beta distribution, assigning higher weights to rarer alleles. Covariates included sex and the first 10 principal components (PCs) for all models. For severity, ischemic versus hemorrhagic, and vascular imaging analyses, we additionally adjusted for age at stroke. For age-at-onset comparisons, age was not included as a covariate, as it defines the phenotype. For etiological subtypes, both pairwise binary comparisons and multinomial models were evaluated. Because gene-based tests were performed across the exome (~ 17,000 genes), an exome-wide significance threshold was defined using Bonferroni correction (p < 2.9 × 10⁻⁶). Given the limited sample size and the exploratory nature of phenotype-stratified analyses, we additionally reported suggestive associations at p < 0.005 to highlight candidate genes for further investigation. Because multiple phenotype contrasts were examined, results meeting the suggestive threshold (p < 0.005) should be interpreted cautiously and considered exploratory. Findings meeting this suggestive threshold were interpreted as hypothesis-generating and requiring independent replication. Gene annotation and constraint metrics We annotated each gene with constraint metrics and tissue expression profiles. Constraint metrics (pLI, LOEUF) were obtained from the gnomAD gene constraint release (v2.1.1 or v4), and merged into our gene list [20]. Genes with pLI > 0.9 or LOEUF < 0.35 were considered under strong functional constraint. Tissue-specific expression levels were obtained from the GTEx (v8) (Genotype-Tissue Expression (GTEx), n.d.) median tissue expression file, and for each gene, we recorded the tissue with the highest expression among brain and artery. These annotations were used to generate plots of constraint scores (pLI, LOEUF) stratified by variant grouping and expression heatmaps of tissue specificity. Results Study population We analyzed a total of 514 stroke patients (193 females, 294 males, and 27 not available). All reported results derive from within-case phenotype-stratified comparisons among patients with ischemic stroke rather than from comparisons between cases and unaffected controls. Baseline characteristics of the study cohort, including age at stroke onset, sex distribution, stroke type, severity (as measured by the Modified Rankin Scale), vascular imaging patterns, and etiological subtypes, are summarized in Table 1. The majority had ischemic stroke (n = 313, 60.9%), while n = 5 (1%) had hemorrhagic stroke. Stroke severity was available for all patients, with 385 (75%) classified as mild (mRS < 3) and 129 (25.1%) as severe (mRS ≥ 3). Table 1Baseline demographic and clinical characteristics of the study populationCharacteristicsCount (N)Percentage (%)Gender Male29457.2% Female19337.5% NA275.3%Age Group ≤ 25 yrs122.3% 26–45 yrs12123.5% > 45 yrs34467% NA377.2Stroke Severity (MRS) Mild (MRS < 3)38575% Severe (MRS ≥ 3)12925.1%Stroke Etiology-based phenotypes Small artery25249% Large artery8015.6% Cardioembolic5610.9% Undetermined336.4% NA9318.1%Imaging-based phenotypes Intracranial19237.4% Extracranial6512.6% Others5510.7% NA20239.3%Distribution of sex, age groups, stroke severity, stroke etiology–based phenotypes, and imaging-based phenotypes in the study cohort. Values are presented as counts (N) and percentages (%). Stroke severity was classified using the Modified Rankin Scale (MRS) as mild (MRS < 3) or severe (MRS ≥ 3). Stroke etiology was categorized as small artery, large artery, cardioembolic, or undetermined. Imaging-based phenotypes were classified as intracranial, extracranial, or other locations Baseline demographic and clinical characteristics of the study population Distribution of sex, age groups, stroke severity, stroke etiology–based phenotypes, and imaging-based phenotypes in the study cohort. Values are presented as counts (N) and percentages (%). Stroke severity was classified using the Modified Rankin Scale (MRS) as mild (MRS < 3) or severe (MRS ≥ 3). Stroke etiology was categorized as small artery, large artery, cardioembolic, or undetermined. Imaging-based phenotypes were classified as intracranial, extracranial, or other locations Age at stroke onset and early-onset subgroup characteristics Age at stroke onset displayed a wide distribution, including a substantial proportion of early-onset cases (≤ 25 years), consistent with the younger population structure of Saudi Arabia. For subgroup analyses, we compared early-onset (≤ 25 years) and late-onset (> 45 years) individuals to investigate differences in clinical characteristics. (Supplementary Figure S1 for the study cohort schematics.) Clinical feature comparison by age group We examined stroke severity, etiology, and vascular imaging features between early-onset and late-onset groups (Table 2). The counts shown in Table 2 do not sum to the overall totals in parentheses, as the overall totals also include participants aged 26–45 years and, where applicable, cases with missing or unavailable data, while the statistical comparisons were restricted to the early-onset (≤ 25 years) and late-onset (> 45 years) groups. Stroke severity showed no significant difference between age groups (Fisher’s Exact p = 1.00). Although small artery stroke was the predominant phenotype in the early-onset group, the distribution of stroke etiology-based phenotypes was not significantly different between the early-onset and late-onset groups. Table 2Comparison of clinical features between early-onset (<25 years) and late-onset (>45 years) groupsClinical characteristicsEarly-Onset (≤25 years, n=12)Late-Onset (>45 years, n=344)P-value Chi-square Fisher’s Exact Stroke Severity (MRS)Mild (MRS < 3) (n=385)92550.98621.00Severe (MRS ≥ 3) (n=129)386Stroke Etiology-based phenotypesSmall artery (n=25281650.08770.1325Large artery (n=80)067Cardioembolic (n=56)045Undetermined (n=33)026Imaging-based phenotypesIntracranial (n=192)51290.93251.00Extracranial (n=65)248Pairwise Fisher’s Exact Test for Stroke EtiologySmall ArteryLarge ArteryCardioembolicUndeterminedLarge Artery0.1101-1.001.00Cardioembolic0.21081.00--Undetermined0.60001.001.00-Totals in parentheses represent all available cases, including the intermediate age group (26–45 years) and missing/unavailable data where applicable; statistical comparisons were performed only between the early-onset and late-onset groups Comparison of clinical features between early-onset (<25 years) and late-onset (>45 years) groups Totals in parentheses represent all available cases, including the intermediate age group (26–45 years) and missing/unavailable data where applicable; statistical comparisons were performed only between the early-onset and late-onset groups Vascular imaging subtypes (intracranial vs. extracranial) did not differ significantly by age at onset (p > 0.8). (Supplementary Figures S1 for the distribution of ischemic stroke etiologies and vascular imaging involvement). Population structure and principal component analysis Principal Component Analysis (PCA) was conducted using high-quality common variants to assess ancestry and control for population structure. The first two principal components (PC1 and PC2) revealed a broadly homogeneous genetic cluster consistent with Middle Eastern ancestry, with no major outliers (Supplementary Figure S2a). Sex and stroke subtype distributions were balanced across PCs, indicating minimal confounding from population stratification. PC1 and PC2 explained 7.31% and 6.19% of the total variance, respectively. Although four individuals were flagged as outliers on PCs 1–4 (|z| > 6), further inspection did not indicate strong ancestry divergence. A scree plot of the top 20 PCs (Supplementary Figure S2b) showed elbow points at PC4 and PC14, suggesting two potential thresholds for substructure correction. Accordingly, we conducted sensitivity analyses using both 4 and 14 PCs for covariate adjustment in gene-based association models to ensure robustness of findings to population stratification. Variant preprocessing, annotation, and distribution of rare coding variation We analyzed whole-exome variant data from 514 individuals, comprising 135,818 unique variant sites and over 68 million genotype calls. The mean sequencing depth was 29.23× across all genotypes, indicating high-quality coverage. Among these, 328,517 heterozygous calls showed a mean allele balance of 0.442, closely approximating the expected 0.5 for true heterozygotes. Similarly, 16,075 homozygous alternate calls had a mean allele balance of 0.988, reflecting near-complete alternate allele representation. These quality metrics support high-confidence genotype calls suitable for downstream rare variant analyses. Post-quality control, variants were categorized by functional consequence using LOFTEE annotations, prioritizing high-confidence predicted loss-of-function (pLoF) variants, followed by missense and synonymous variants. We further assessed the minor allele frequency (MAF) spectrum across the cohort and reference allele frequencies from gnomAD populations. A comprehensive overview of variant consequence categories, frequency distribution, and gnomAD ancestry representation is shown in Fig. 1 and Supplementary Table S1. A Pie chart showing the distribution of variant functional categories based on LOFTEE annotations (pLoF, missense, synonymous, etc.). B Bar chart of minor allele frequency (MAF) distribution among all post-QC variants. C Distribution of variant allele frequencies across gnomAD reference populations Gene-based association analyses After genotype QC there were 132,949 markers in the PLINK file and 514 samples with genotype data; 480 samples had non-missing phenotypes and covariates and were retained for model fitting (35 samples lacked phenotype/covariate data and were excluded) [21]. We applied SAIGE-GENE to perform gene-based rare variant burden testing across six phenotypic contrasts relevant to stroke heterogeneity [19]. All analyses used a maximum minor allele frequency (MAF) threshold of 1% and included pLOF, damaging missense, and synonymous variants. Population structure was adjusted using either 4 or 14 principal components (PCs), with consistent top signals across both models. Results using 14 PCs are reported as the primary analysis, as this approach yielded improved correction for substructure and slightly more robust associations. A suggestive significance threshold of p < 0.005 was used throughout. Manhattan plots for each analysis are presented in Supplementary Figure S3. The distribution of burden-test P values across phenotype-specific analyses is shown in Supplementary Figure S4. Genomic inflation assessment using Quantile-quantile (QQ) plots showed adequate calibration for the age-of-onset and stroke severity gene-based analyses (λGC = 0.941 and 1.029, respectively), with deviation from expectation largely confined to the tail of the distribution (Supplementary Figure S5). Stroke severity (severe vs. mild) association analyses We investigated gene-level associations between mild (mRS < 3) and severe (mRS ≥ 3) stroke outcomes. Among 17,253 genes tested, two loci demonstrated suggestive significance (p < = 0.003). Rare variant burden in HSP90AB1 (p = 0.0036), a heat shock protein involved in cellular stress responses, showed a suggestive association with severe stroke in this exploratory analysis. A second signal was observed at PRR23A (p = 0.0011), a gene with limited annotation and unknown function in cerebrovascular biology. These associations remained consistent after adjustment for both 4 and 14 PCs. Age at onset (early vs. late) association analyses We conducted two separate analyses stratifying by age at stroke onset in stroke case-only cohort. In the first comparison, between pediatric onset ( onset ≤ 25 years) and adult onset (> 45 years), LRRC42 reached suggestive significance (p = 0.0031). A second comparison contrasting older adults (> 45 years) versus younger adults (< 45 years) onset strokes identified SMIM34 (p = 0.0025), POGZ (p = 0.0017), and MAPK8IP1 (p = 0.0042) as candidate genes. POGZ, a chromatin regulator, has been previously implicated in neurodevelopmental disorders. MAPK8IP1 encodes a JNK-interacting protein with potential relevance to neuronal inflammation. SMIM34, a small integral membrane protein, has no known prior association with stroke; however, its enrichment in early-onset cases may represent a hypothesis for further investigation, but no biological interpretation can be made at this stage. Vascular imaging (intracranial vs. extracranial) association analyses To examine genetic contributions to vascular disease distribution, we compared individuals with imaging-confirmed intracranial versus extracranial disease. Among 981 genes tested, suggestive associations were observed for COL9A3 (p = 0.0014), a collagen subunit involved in extracellular matrix integrity, and DCP1B (p = 0.0021), an mRNA decapping enzyme with potential regulatory roles in cellular stress responses. Etiological subtypes association analyses We further explored genetic variation across ischemic stroke etiologies using TOAST classification. In the comparison of small-artery occlusion versus large-artery atherosclerosis, ADGRV1 emerged as a candidate gene (p = 0.0043). This adhesion G-protein-coupled receptor is expressed in neural and vascular tissues and may influence microvascular remodelling. Comparisons between small-artery and cardioembolic strokes, as well as between large-artery and cardioembolic subtypes, did not yield significant findings, possibly due to limited power or phenotypic overlap. Across all phenotypic contrasts, a total of ten gene-level associations reached suggestive significance (p < 0.005), while none reached the exome-wide significance threshold after correction for multiple testing. These include both biologically plausible candidates (e.g., HSP90AB1, COL9A3, POGZ) and uncharacterized loci (e.g., SMIM34, PRR23A) that may represent previously unreported contributors to stroke heterogeneity. None of the associations exceeded the exome-wide significance threshold after correction for multiple testing. These results should be interpreted as exploratory and hypothesis-generating pending replication in independent cohorts (Table 3). Table 3Summary of gene-based association results obtained from SAIGE-GENE across different phenotypic categories in stroke analysisPhenotypeGeneVariant GroupP-valueStroke Severity ( mRS < 3, mRS ≥ 3 vs. ) HSP90AB1 pLoF; other_missense; damaging missense; synonymous; Cauchy0.0036 PRR23A pLoF; other_missense; damaging missense; synonymous; Cauchy0.0011Age of Onset ≤ 25 vs. > 45 (Pediatric vs. late Onset) LRRC42 pLoF; other_missense; damaging missense; synonymous; Cauchy0.0031Age of Onset < 45 vs. > 45 (Young-adult onset vs. late onset) SMIM34 pLoF; damaging_missense; other_missense; synonymous; Cauchy0.0025 POGZ pLoF; damaging_missense; other missense; synonymous; Cauchy0.0017 MAPK8IP1 pLoF; damaging_missense; other missense; synonymous; Cauchy0.0042Imaging (Intracranial vs. Extracranial) COL9A3 pLoF; damaging_missense; other missense; synonymous; Cauchy0.0014 DCP1B pLoF; damaging_missense; other missense; synonymous; Cauchy0.0021 Etiology (Small Artery vs. Large Artery) ADGRV1 pLoF; damaging_missense; other missense; synonymous; Cauchy 0.0043 Results are presented for five phenotypes of interest: Stroke Severity (mRS < 3 (mild) vs mRS ≥ 3 (severe)), Age of Onset (Pediatric vs Late Onset, Young-Adult vs Late Onset), Imaging (Intracranial vs Extracranial), and Etiology (Small Artery vs Large Artery). Variant groups represent the aggregation masks tested within each gene and included predicted loss-of-function (pLoF), damaging missense, other missense, and synonymous variants combined using the Cauchy framework. Synonymous variants were included to allow broader coding-variation and sensitivity analyses; however, interpretation was prioritized for masks enriched for predicted deleterious variants. P-values shown are gene-level association p-values Summary of gene-based association results obtained from SAIGE-GENE across different phenotypic categories in stroke analysis Results are presented for five phenotypes of interest: Stroke Severity (mRS < 3 (mild) vs mRS ≥ 3 (severe)), Age of Onset (Pediatric vs Late Onset, Young-Adult vs Late Onset), Imaging (Intracranial vs Extracranial), and Etiology (Small Artery vs Large Artery). Variant groups represent the aggregation masks tested within each gene and included predicted loss-of-function (pLoF), damaging missense, other missense, and synonymous variants combined using the Cauchy framework. Synonymous variants were included to allow broader coding-variation and sensitivity analyses; however, interpretation was prioritized for masks enriched for predicted deleterious variants. P-values shown are gene-level association p-values Gene annotation and constraint metrics To better contextualize our gene-based rare variant association findings, we annotated prioritized genes with tissue-specific expression data from GTEx and functional constraint metrics from gnomAD, as shown in Supplementary Table S3. Among the genes reaching suggestive significance, several demonstrated strong biological plausibility based on expression in neurologically or vascularly relevant tissues and intolerance to loss-of-function variation. HSP90AB1, a molecular chaperone with known roles in proteostasis and stress response, exhibited high expression across the brain (median TPM: 927.4), artery (median TPM: 1069), and heart (median TPM: 1887), and was highly constrained (pLI = 1.0; LOEUF = 0.35). MAPK8IP1, associated with axonal transport and signaling, also showed high brain expression (278.3 TPM) and moderate constraint (pLI = 1.0; LOEUF = 0.38). POGZ, previously implicated in neurodevelopmental disorders, was expressed in brain (124.9 TPM) and among the most constrained genes in our list (pLI = 1.0; LOEUF = 0.20). In contrast, genes such as PRR23A and SMIM34, though statistically associated with stroke-related phenotypes, showed little expression in brain or vascular tissues and lacked or had limited constraint annotations. This may suggest a less direct functional role or warrant cautious interpretation. Expression patterns also varied across vascular phenotypes: for example, DCP1B showed mild arterial expression (17.5 TPM) with modest constraint (LOEUF = 0.94), and COL9A3, a collagen-encoding gene, was notably expressed in the brain (106.6 TPM) but less constrained (LOEUF = 1.06). ADGRV1, associated with auditory and neural development, also showed moderate brain expression (12.2 TPM) and partial constraint (LOEUF = 0.64). Full constraint and expression metrics for all prioritized genes are presented in Supplementary Table S2. Discussion In this study, we leveraged rare variant burden testing to explore the genetic underpinnings of ischemic stroke in a Saudi Arabian cohort. Rather than using a traditional case–control design, our approach was aimed at identifying genetic differences across clinical subtypes of stroke. Our study extends previous sequencing efforts in Saudi stroke cohorts by applying gene-based rare-variant aggregation tests across the exome and examining multiple clinical stroke phenotypes, including severity, age at onset, vascular territory, and etiological subtype, in an understudied population. By integrating phenotype-stratified analyses with gene-based rare variant burden testing, this study provides complementary insights beyond prior sequencing studies conducted in Saudi stroke cohorts. We identified multiple genes nominally associated with clinically defined stroke-phenotypes at a suggestive threshold (p < 0.005). For stroke severity, HSP90AB1 and PRR23A showed consistent associations across variant groupings, as assessed by burden and Cauchy combination tests. Among the most consistent findings was HSP90AB1, associated with stroke severity across multiple variant groupings [22]. This gene encodes a ubiquitously expressed molecular chaperone involved in proteostasis and cellular stress responses. HSP90AB1 has been explored as a therapeutic target in other disease contexts, providing an additional layer of support for its potential contribution to stroke pathophysiology [23]. Importantly, it was also one of the most highly expressed genes across brain and vascular tissues and showed strong intolerance to loss-of-function variation, highlighting its importance in maintaining cellular homeostasis [24, 25]. Taken together, these findings point to HSP90AB1 as a compelling candidate for future functional validation. To explore age-related genetic signatures, we defined two age-at-onset comparisons: pediatric-onset vs. late-onset stroke (< 25 vs. > 45 years) and young-adult vs. late-onset stroke (< 45 vs. > 45 years). In the pediatric-to-late comparison, LRRC42 was identified with suggestive association, whereas in the broader young-adult-to-late comparison, three genes emerged: SMIM34, MAPK8IP1, and POGZ. Notably, POGZ has been implicated in neurodevelopmental syndromes, and in our study, exhibited high expression in neural and vascular tissues and strong evolutionary constraint [26]. These findings suggest potential age-specific genetic mechanisms in stroke and support the need for further age-stratified genomic studies, particularly in populations with a high prevalence of early-onset disease. Additional gene-level associations were observed for imaging-defined stroke types (intracranial vs. extracranial) and TOAST-defined etiologies. For example, COL9A3 and DCP1B were associated with vascular imaging patterns, while ADGRV1 was associated with small-artery vs. large-artery stroke. These findings warrant follow-up in larger datasets, to determine whether they reflect distinct biological pathways underlying ischemic stroke heterogeneity. To further contextualize these associations, we annotated each gene for tissue-specific expression using GTEx data and for intolerance to loss-of-function variation using gnomAD metrics (pLI and LOEUF). Several genes, including HSP90AB1, MAPK8IP1, and POGZ, showed high expression in relevant tissues and low LOEUF scores, indicating selective pressure against disruption. The combination of expression, constraint, and gene function strengthens the biological plausibility of these signals, even in the absence of replication. Crucially, this study highlights the value of studying genetic variation in underrepresented populations. The Saudi Arabian cohort examined here includes a relatively young demographic, high prevalence of consanguinity, and unique environmental exposures, all of which may contribute to genetic architectures distinct from those seen in primarily European-ancestry stroke studies. By leveraging population-specific data and careful phenotypic stratification, we demonstrate that meaningful insights can be derived even from modestly sized cohorts, particularly when combined with rare variant aggregation approaches. Despite the promise of these findings, several limitations must be acknowledged. Although whole-exome sequencing enabled broad interrogation of coding variation, the modest sample size restricted statistical power and precluded discovery-level conclusions. This limitation was particularly relevant for phenotype-stratified comparisons, where some groups, such as pediatric-onset stroke and certain etiological subtypes defined by the TOAST classification, contained relatively small numbers of individuals. For example, there was a marked imbalance in sample sizes between early-onset and late-onset groups. Such subgroup imbalance may reduce power to detect true rare-variant associations and increase uncertainty in effect estimates. In rare variant analyses, a finding may sometimes be influenced by only a few individuals carrying a variant. Larger studies with more balanced subgroup sizes will be needed to confirm the reliability of these findings. Furthermore, the absence of an independent replication cohort means that these findings should be interpreted as exploratory and hypothesis-generating. Future studies with larger sample sizes may also benefit from modeling stroke severity using ordinal regression approaches to more fully capture the information contained within the modified Rankin Scale. Our results therefore require replication in independent Saudi and non-Saudi cohorts, and meta-analytic efforts across consortia with compatible phenotyping will be crucial. Nonetheless, this study provides an exploratory survey of gene-based rare variant signals across ischemic stroke subphenotypes in a Saudi cohort. The associations reported here should be interpreted as hypothesis-generating rather than definitive, particularly given the sample size, multiple phenotype comparisons, and the use of a suggestive significance threshold. In addition, the study illustrates how combining large-scale sequencing, rigorous gene-based statistics, and detailed clinical characterization can uncover plausible biological pathways that shape stroke presentation and outcome in populations that have historically been underrepresented in genomic research. As genomic resources expand and collaborative networks grow, our findings can be integrated into larger efforts aimed at refining stroke risk prediction, tailoring preventive strategies, and identifying novel therapeutic targets informed by population-specific genetic architectures. Future work should prioritize replication in independent cohorts, including additional Saudi and Middle Eastern stroke samples, as well as larger multi-center or multi-ethnic consortia with harmonized phenotype definitions. Such efforts will be important for validating the observed exploratory signals, improving power for rare-variant analyses, and assessing generalizability across populations. Functional follow-up and integration with complementary molecular and clinical data will also be important for clarifying the biological relevance of prioritized genes. Conclusions Overall, this study demonstrates the utility of a case-only phenotype-stratified rare variant framework for investigating the genetic basis of clinical heterogeneity in ischemic stroke. The candidate genes and pathways identified here provide a starting point for deeper mechanistic investigations and for future large-scale, multi-ethnic collaborations. With continued efforts to expand sample sizes, validate findings, and integrate genetic data with longitudinal clinical and imaging phenotypes, this line of work has the potential to inform more tailored strategies for stroke prediction, intervention, and recovery in Middle Eastern and other diverse populations. Supplementary Information Supplementary Material 1: Supplementary Figure S1. Clinical and imaging characteristics stratified by stroke severity. (A) Age distribution of participants stratified by stroke severity, shown as overlaid histograms for mild stroke (mRS < 3) and severe stroke (mRS ≥ 3). (B) Etiologic distribution of ischemic stroke subtypes, presented by sex across the categories small artery, large artery, cardioembolic, and undetermined. (C) Distribution of imaging findings/types according to stroke severity, comparing the number of participants classified as extracranial, intracranial, none, or other in the mild and severe stroke groups. Supplementary Material 2: Supplementary Figure S2. Principal component analysis of the study cohort. A, Scatter plot of principal components 1 and 2 (PC1 vs PC2), with individual participants colored by ischemic stroke subtype (small-artery, large-artery, cardioembolic, multiple, or unknown). The broadly overlapping cluster indicates a largely homogeneous genetic background consistent with Middle Eastern ancestry and no major outliers by subtype. B, Scree plot showing the percentage of variance explained by the top 20 principal components. Elbow points at approximately PC4 and PC14 suggest potential thresholds for inclusion of PCs as covariates in rare variant association analyses. Supplementary Material 3: Supplementary Figure S3. Manhattan plots of gene-based rare variant burden tests across stroke phenotypes. For each analysis, genes are plotted along the X-axis by genomic position (grouped by chromosome), and the –log10(p-value) of the burden test is shown on the Y-axis. Panels depict results for: overall stroke severity, age at stroke onset (young adult onset <45 years vs late onset >45 years), pediatric/early-onset stroke (≤25 years vs late onset >45 years), and ischemic stroke etiologic subtype contrasts (small-artery vs large-artery, small-artery vs cardioembolic, and large-artery vs cardioembolic). The horizontal dashed line denotes the suggestive significance threshold used in the study (p < 0.005). No gene reached exome-wide significance after correction for multiple testing. Supplementary Material 4: Supplementary Figure S4. Histograms of burden test P-value distributions across ischemic stroke phenotypes. Each panel shows the frequency distribution of P values obtained from burden analyses performed for distinct ischemic stroke phenotypes or subgroup comparisons, including stroke severity, age of onset, pediatric age of onset, and etiologic subtype combinations involving small artery, large artery, and cardioembolic stroke. The x-axis denotes the P value and the y-axis denotes the number of tests. Overall, the distributions are approximately uniform, with no substantial excess of low P values. Supplementary Material 5: Supplementary Figure S5. QQ plots of gene-based association tests across ischemic stroke phenotypes. Quantile-quantile plots comparing observed and expected −log⁡10(P)−log10(P) values from gene-based association analyses of stroke severity, imaging type, age at onset and pediatric onset. The red diagonal indicates the null expectation. Across analyses, the distributions are generally well calibrated, with modest tail deviation in some phenotypes consistent with limited enrichment of low P values. Genomic control inflation factors (λGC)(λGC) are shown in each panel. Supplementary Material 6: Supplementary Table S1. Burden test results for stroke severity (mRS < 3 vs mRS ≥ 3 ). Supplementary Table 1 shows gene-based rare variant burden test results comparing individuals with severe stroke (modified Rankin Scale mRS ≥ 3) to those with non-severe stroke (mRS < 3). A total of 293 unique genes were evaluated. For each gene and variant grouping, the table lists the gene/region name (Region), the variant grouping used in the burden test (Group; for example, other missense, damaging missense; other missense; pLoF; synonymous), the maximum minor allele frequency threshold applied (max_MAF), and the resulting burden test P value (Pvalue_Burden). Genome-wide significance was defined as P = 0.0002, and suggestive significance was defined as P = 0.005. Rows labeled “Cauchy” represent combined P values from multiple variant groupings using a Cauchy combination test. Supplementary Table S2. Expression and constraint metrics for genes implicated in stroke severity. Supplementary Table 2 summarizes tissue-specific expression and gene constraint metrics for genes showed a suggestive association in the stroke severity analysis. The table lists, for each gene (Gene) and variant grouping used in the burden tests (Variant_Grouping), the median transcripts-per-million (TPM) expression in brain tissue (Median TPM (Brain)), artery tissue (Median TPM (Artery)), and heart tissue (Median TPM (Heart)) derived from reference transcriptomic datasets. Gene intolerance to loss-of-function variants is described by the probability of loss-of-function intolerance (pLI) and the loss-of-function observed/expected upper bound fraction (LOEUF), obtained from population sequencing resources such as gnomAD. Higher pLI and lower LOEUF indicate stronger intolerance to loss-of-function variation. Supplementary Table S3. Replication of gene expression and constraint metrics in an independent dataset. Supplementary Table 3 presents an independent replication set of tissue-specific expression and gene constraint metrics for the genes evaluated in the stroke severity burden analysis. As in Supplementary Table 2, the table includes the gene symbol (Gene), the variant grouping used in the rare variant burden tests (Variant_Grouping), median TPM expression in brain, artery, and heart tissues (Median TPM (Brain), Median TPM (Artery), Median TPM (Heart)), and the corresponding gene constraint measures (pLI and LOEUF). These data provide external validation of gene expression patterns and intolerance to loss-of-function variation for genes showing suggestive or genome-wide significant association with stroke severity. Acknowledgements The authors would like to thank the Deanship of Scientific Research at Imam Abdulrahman bin Faisal University, Saudi Arab for providing financial support for this research (#2016-158-Med). We are also grateful to the nurses, technical staff, and students for their efforts, commitment, and dedication. Authors’ contributions Conceptualization, R.A., N.K., B.J. and A.K.; Methodology, A.H., C.V., C.C., M.V., B.J. and A.K.; Software, M.V., B.J., C.V. and A.K.; Validation, R.A., N.K., M.M., A.A., M.H., F.S., M.A., S.A., H.A. and F.A.; Formal Analysis, A.H., C.V., M.V., C.C., B.J. and A.K.; Investigation, R.A., N.K., M.M., A.A., M.H., F.S., M.A., S.A., H.A. and F.A.; Resources, R.A., N.K., M.M., A.A., M.H., F.S., M.A., S.A., H.A. and F.A.; Data Curation, A.H., C.V., C.C., B.J. and A.K.; Writing – Original Draft Preparation, R.A., M.M., M.A., M.V., B.J. and A.K.; Writing – Review & Editing, R.A., N.K., A.H., M.A., A.A., C.V., C.C., S.A., H.A., M.H., F.S., M.M., M.V., B.J. and A.K.; Visualization, R.A.; Supervision, A.K. and R.A.; Project Administration, R.A., N.K., A.A., S.A. and M.A.; Funding Acquisition, R.A. and A.K. All authors have read and agreed to the published version of the manuscript. All authors have read and agreed to the published version of the manuscript. Funding This research was funded by the Deanship of Scientific Research at Imam Abdulrahman bin Faisal University via grant number 2016-158-Med. Data availability The datasets generated during the current study are available in the European Variation Archive (EVA) repository (https://www.ebi.ac.uk/ebisearch/search? query=PRJEB59227&submit=Search&db=allebi&requestFrom=global-masthead), with accession number PRJEB59227. All requests for data can be sent to the corresponding author (AKA) and verified academic investigators will be granted full access. Declarations Ethics approval and consent to participate Ethical approval for the study was obtained from the Imam Abdulrahman bin Faisal University Institutional Review Board (IRB) committees (IRB# 2025-01-0207).The study was conducted in accordance with the principles of the Declaration of Helsinki and Good Clinical Practice guidelines. Informed consent to participate was obtained from all participants included in the study. Written informed consent was obtained using an English consent form with a verified Arabic translation. For adult participants who were unable to provide informed consent due to stroke-related cognitive and/or physical impairment, informed consent to participate was obtained from a legal guardian or an appropriate representative (next of kin) on the participant’s behalf. No waiver of consent was granted or applied. Consent for publication All authors have give their consent for publication of this manuscript. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1: Supplementary Figure S1. Clinical and imaging characteristics stratified by stroke severity. (A) Age distribution of participants stratified by stroke severity, shown as overlaid histograms for mild stroke (mRS < 3) and severe stroke (mRS ≥ 3). (B) Etiologic distribution of ischemic stroke subtypes, presented by sex across the categories small artery, large artery, cardioembolic, and undetermined. (C) Distribution of imaging findings/types according to stroke severity, comparing the number of participants classified as extracranial, intracranial, none, or other in the mild and severe stroke groups. Supplementary Material 2: Supplementary Figure S2. Principal component analysis of the study cohort. A, Scatter plot of principal components 1 and 2 (PC1 vs PC2), with individual participants colored by ischemic stroke subtype (small-artery, large-artery, cardioembolic, multiple, or unknown). The broadly overlapping cluster indicates a largely homogeneous genetic background consistent with Middle Eastern ancestry and no major outliers by subtype. B, Scree plot showing the percentage of variance explained by the top 20 principal components. Elbow points at approximately PC4 and PC14 suggest potential thresholds for inclusion of PCs as covariates in rare variant association analyses. Supplementary Material 3: Supplementary Figure S3. Manhattan plots of gene-based rare variant burden tests across stroke phenotypes. For each analysis, genes are plotted along the X-axis by genomic position (grouped by chromosome), and the –log10(p-value) of the burden test is shown on the Y-axis. Panels depict results for: overall stroke severity, age at stroke onset (young adult onset <45 years vs late onset >45 years), pediatric/early-onset stroke (≤25 years vs late onset >45 years), and ischemic stroke etiologic subtype contrasts (small-artery vs large-artery, small-artery vs cardioembolic, and large-artery vs cardioembolic). The horizontal dashed line denotes the suggestive significance threshold used in the study (p < 0.005). No gene reached exome-wide significance after correction for multiple testing. Supplementary Material 4: Supplementary Figure S4. Histograms of burden test P-value distributions across ischemic stroke phenotypes. Each panel shows the frequency distribution of P values obtained from burden analyses performed for distinct ischemic stroke phenotypes or subgroup comparisons, including stroke severity, age of onset, pediatric age of onset, and etiologic subtype combinations involving small artery, large artery, and cardioembolic stroke. The x-axis denotes the P value and the y-axis denotes the number of tests. Overall, the distributions are approximately uniform, with no substantial excess of low P values. Supplementary Material 5: Supplementary Figure S5. QQ plots of gene-based association tests across ischemic stroke phenotypes. Quantile-quantile plots comparing observed and expected −log⁡10(P)−log10(P) values from gene-based association analyses of stroke severity, imaging type, age at onset and pediatric onset. The red diagonal indicates the null expectation. Across analyses, the distributions are generally well calibrated, with modest tail deviation in some phenotypes consistent with limited enrichment of low P values. Genomic control inflation factors (λGC)(λGC) are shown in each panel. Supplementary Material 6: Supplementary Table S1. Burden test results for stroke severity (mRS < 3 vs mRS ≥ 3 ). Supplementary Table 1 shows gene-based rare variant burden test results comparing individuals with severe stroke (modified Rankin Scale mRS ≥ 3) to those with non-severe stroke (mRS < 3). A total of 293 unique genes were evaluated. For each gene and variant grouping, the table lists the gene/region name (Region), the variant grouping used in the burden test (Group; for example, other missense, damaging missense; other missense; pLoF; synonymous), the maximum minor allele frequency threshold applied (max_MAF), and the resulting burden test P value (Pvalue_Burden). Genome-wide significance was defined as P = 0.0002, and suggestive significance was defined as P = 0.005. Rows labeled “Cauchy” represent combined P values from multiple variant groupings using a Cauchy combination test. Supplementary Table S2. Expression and constraint metrics for genes implicated in stroke severity. Supplementary Table 2 summarizes tissue-specific expression and gene constraint metrics for genes showed a suggestive association in the stroke severity analysis. The table lists, for each gene (Gene) and variant grouping used in the burden tests (Variant_Grouping), the median transcripts-per-million (TPM) expression in brain tissue (Median TPM (Brain)), artery tissue (Median TPM (Artery)), and heart tissue (Median TPM (Heart)) derived from reference transcriptomic datasets. Gene intolerance to loss-of-function variants is described by the probability of loss-of-function intolerance (pLI) and the loss-of-function observed/expected upper bound fraction (LOEUF), obtained from population sequencing resources such as gnomAD. Higher pLI and lower LOEUF indicate stronger intolerance to loss-of-function variation. Supplementary Table S3. Replication of gene expression and constraint metrics in an independent dataset. Supplementary Table 3 presents an independent replication set of tissue-specific expression and gene constraint metrics for the genes evaluated in the stroke severity burden analysis. As in Supplementary Table 2, the table includes the gene symbol (Gene), the variant grouping used in the rare variant burden tests (Variant_Grouping), median TPM expression in brain, artery, and heart tissues (Median TPM (Brain), Median TPM (Artery), Median TPM (Heart)), and the corresponding gene constraint measures (pLI and LOEUF). These data provide external validation of gene expression patterns and intolerance to loss-of-function variation for genes showing suggestive or genome-wide significant association with stroke severity. Data Availability Statement The datasets generated during the current study are available in the European Variation Archive (EVA) repository (https://www.ebi.ac.uk/ebisearch/search? query=PRJEB59227&submit=Search&db=allebi&requestFrom=global-masthead), with accession number PRJEB59227. All requests for data can be sent to the corresponding author (AKA) and verified academic investigators will be granted full access.