Genome-Wide Association Study of Hospitalized Patients and Acute Kidney Injury

Abstract

Acute kidney injury (AKI) is a common and devastating complication of hospitalization. Here, we identified genetic loci associated with AKI in patients hospitalized between 2002–2019 in the Million Veteran Program and data from Vanderbilt University Medical Center’s BioVU. AKI was defined as meeting a modified KDIGO Stage1 or more for two or more consecutive days or kidney replacement therapy. Control individuals were required to have one or more qualifying hospitalization without AKI and no evidence of AKI during any other observed hospitalizations. Genome-wide association studies (GWAS), stratified by race, adjusting for sex, age, baseline estimated glomerular filtration rate (eGFR), and the top ten principal components of ancestry were conducted. Results were meta-analyzed using fixed effects models. In total, there were 54,488 patients with AKI and 138,051 non-AKI individuals included in the study. Two novel loci reached genome-wide significance in the meta-analysis: rs11642015 near the FTO locus on chromosome 16 (obesity traits) (odds ratio 1.07 (95% confidence interval, 1.05–1.09)) and rs4859682 near the SHROOM3 locus on chromosome 4 (glomerular filtration barrier integrity) (odds ratio 0.95 (95% confidence interval, 0.93–0.96)). These loci colocalized with previous studies of kidney function, and genetic correlation indicated significant shared genetic architecture between AKI and eGFR. Notably, the association at the FTO locus was attenuated after adjustment for BMI and diabetes, suggesting that this association may be partially driven by obesity. Both FTO and the SHROOM3 loci showed nominal evidence of replication from diagnostic-code-based summary statistics from UK Biobank, FinnGen, and Biobank Japan. Thus, our large GWA meta-analysis found two loci significantly associated with AKI suggesting genetics may explain some risk for AKI.

Graphical Abstract

Lay Summary

Acute kidney injury (AKI) is a common and harmful condition among hospitalized patients. There are no established treatments for AKI. Understanding if there are genetic risk factors for AKI may help provide clues to the reasons that AKI happens and help develop new treatments. In this study, we examined the association between genetic variants and the risk for AKI in hospitalized patients within the Million Veteran Program and Vanderbilt University Medical Center’s BioVU genetic biobanks. Two novel genetic variants were identified to be associated with the risk of developing AKI. These findings suggest that genetics may explain some of the risk for developing AKI but that the association may be partially explained through other comorbid conditions. Future studies are needed to understand how these risk factors contribute to developing AKI.

Introduction

Acute kidney injury (AKI) commonly complicates hospitalization and strongly associates with short- and long-term morbidity and mortality.^1–6 There are no treatments for this condition, with progress hindered by limited understanding of the determinants of human AKI.⁷

Both indirect and direct evidence suggest a genetic basis for AKI. Examples include heterogeneity in the risk observed in patients with similar risk factors, variations in treatment response in clinical trials,⁸ and an unexplained racial predilection for developing AKI.^{9, 10} These observations have prompted investigations to identify genetic variants that may explain these differences. Early candidate gene studies¹¹ based on mechanistic data have implicated pathways such as inflammation,^{12, 13} oxidative stress,^{14, 15} and endothelial dysfunction.¹⁶ Of few genome-wide association studies(GWAS),^17–20 most have examined homogenous populations, such as patients undergoing surgery. Interpretation has been challenged by smaller sample sizes and limited enrichment for phenotypes likely to reflect parenchymal injury.^{21, 22}

We sought to identify genetic loci associated with AKI among patients with hospitalized AKI. We conducted a large GWAS meta-analysis of persistent AKI among participants of the Million Veteran Program²³ and Vanderbilt University Medical Center’s (VUMC) DNA biobank (BioVU). We also evaluated whether identified loci colocalize with CKD, kidney function loci, and kidney expression quantitative trait loci (eQTLs). Finally, we explored the consistency of genetic loci identified using different phenotyping methods in publicly available data.²⁴

Methods

Study Populations

Million Veteran Program

The Million Veteran’s Program (MVP) is a large cohort of consented participants recruited from 63 Veteran’s Administration (VA) medical facilities.^{23, 25} Recruitment began in 2011 with participants answering baseline and lifestyle questionnaires, providing blood samples, and granting access to medical records. Researchers are provided with de-identified versions of these data and cannot link details with a participants’ identity. All documents and protocols have been approved by the VA Central Institutional Review Board.

BioVU

Data were also collected from Vanderbilt University Medical Center’s BioVU,²⁴ which began in 2004 with more than 329,000 DNA samples as of January 2024. Patient genetic data available for research is linked with the VUMC synthetic derivative (SD), a fully de-identified, date-shifted electronic health record observational data extract. The SD contains 3.5 million individual records collected over 20 years from both home-grown and Epic electronic health record systems. The SD is transformed from multiple source systems and normalized to Observational Medical Outcomes Partnership Common Data Model (OMOP CDM) version 5. This study was approved by VUMC Institutional Review Board #190947 with a waiver of consent due to its retrospective nature.

MVP and BioVU Phenotype Definitions

A similar process was applied for cohort selection in both the MVP and BioVU (Figure 1). All hospitalizations were identified. To be considered for AKI case or control selection, patients had to have at least one qualifying hospitalization. A qualifying index hospitalization was defined as at least one hospitalization while aged ≥ 18, with an outpatient creatinine available in the 7–365 days before hospitalization, and at least one inpatient serum creatinine measurement. Hospitalizations with a history of end-stage renal disease (ESRD; defined as a preadmission outpatient estimated glomerular filtration rate (eGFR) <15 ml/min/1.73 m², chronic dialysis, or kidney transplant) were not considered as qualifying case or control hospitalizations nor were hospitalizations involving nephrectomy or a prior history of glomerulonephritis by administrative codes. AKI case and non-AKI control definitions (below) were applied. We did not impute any missing phenotypic variables or apply transformations. Administrative codes used to define covariate definitions in both cohorts are shown in Supplementary Table S1.

Figure 1. Patient Selection Flowchart.

Inclusion and exclusion steps shown for patient selection in both study cohorts (VA MVP left, VUMC BioVU right). See methods for full description.

AKI Case Definition:

Among those with a qualifying hospitalization, AKI cases were then identified using a modified Kidney Disease Improving Global Outcome (KDIGO) definition.²⁶ To enrich for AKI more likely to involve parenchymal injury, all patients were required to meet a minimum Stage 1 AKI definition of a 0.3 mg/dl or 50% increase from baseline serum creatinine for at least 2 consecutive days (i.e., meet a minimum of Stage 1 AKI for 2 days or more). Stage 2 AKI was defined as a doubling of baseline creatinine, and Stage 3 AKI was defined as a tripling of baseline creatinine or receipt of acute dialysis. Any stage of AKI could qualify as a case so long as there was evidence of meeting a 0.3 mg/dl or 50% increase for ≥2 consecutive days or if the patient received acute KRT. Patients receiving acute KRT could qualify as an AKI case without meeting the persistence definition due to the confounding effects of dialysis on creatinine. Baseline serum creatinine was defined as the mean outpatient serum creatinine value 7–365 days before admission.²⁷ Among multiple qualifying AKI hospitalizations, the first was selected.

Non-AKI Control Definition:

To qualify as a non-AKI control, patients had to be without evidence for AKI for all observed hospitalizations. This strategy was chosen to enrich for patients without historical evidence of AKI as a reference group. Non-AKI controls were first identified as having a qualifying index hospitalization that did not meet AKI criteria (i.e. KDIGO Stage 0). In addition, non-AKI controls could not have evidence of AKI at any other hospitalization defined as having a greater than 0.3 mg/dl or 50% rise from baseline to peak or receiving acute dialysis. For controls, we defined baseline as the lowest of either the mean outpatient baseline creatinine 7–365 days before admission or the lowest inpatient creatinine. The latter was used to reduce the possibility of patients who may have AKI during other nonqualifying hospitalizations without a known outpatient baseline from being included as non-AKI controls.

Genotyping and Quality Control

MVP:

Blood samples drawn from consenting MVP participants were shipped to a central biorepository in Boston, Massachusetts, where DNA was extracted and shipped to two external genotyping centers. A custom Affymetrix Axiom Biobank array designed for the MVP was used, which contained additional content added to provide coverage of African and Hispanic haplotypes, as well as markers for common diseases in the VA population. Standardized quality control pipelines were used to exclude duplicate samples or samples with more heterozygosity than expected. We also excluded related individuals (between 2^nd and 3^rd degree relatives or closer) as measured by KING software.²⁸ Genotypes were pre-phased using EAGLE v2²⁹ and imputed to a reference panel composed of 1000 Genomes phase 3, version 5 and African Genome Resources with Minimac3 software.^{30, 31}

BioVU:

BioVU DNA samples were genotyped on a custom Illumina Multi-Ethnic Genotyping Array (MEGA-ex; Illumina Inc., San Diego, CA, USA). Quality control included excluding samples or variants with missingness rates above 2%. Samples were also excluded if consent had been revoked or if the sample was duplicated. Imputation was performed on the Michigan Imputation Server (MIS) v1.2.4³⁰ using Minimac4 and the Haplotype Reference Consortium (HRC) panel v1.1.³²

Genetic Analyses

For both MVP and BioVU cohorts, GWAS were performed using logistic regression modeling AKI status as a function of additive genotype, age at hospitalization, sex, and baseline renal function (eGFR) and the top ten principal components of ancestry (Supplementary Figure S1a) and stratified by race/ethnicity (harmonized ancestry race and ethnicity [HARE] in MVP,³³ EHR-reported race in BioVU). Analysis in MVP was performed using linear models implemented in PLINK2³⁴ and BioVU analyses used SNPTEST.³⁵ We applied filters to imputed data including requiring imputation quality (r²) scores of >= 0.4, and minor allele frequency filters based on the sample size of each individual analysis set, as well as generated diagnostic Q-Q plots and lambda values to quantify inflation (Supplementary Figure S1b). Secondary analysis was also performed adjusting for BMI and diabetes status. Results were combined using inverse-variance weighted fixed-effects meta-analysis implemented in METAL without correction for genomic control.³⁶ Narrow-sense heritability and inflation due to population stratification were estimated using LD score regression.³⁷ Functional mapping and annotation of genetic associations (FUMA) software³⁸ was used to map variants to genes, identify the number of independent signals at each locus, identify enriched pathways (Gene2Func), identify previous associations from the GWAS catalog, and perform tissue enrichment analyses (MAGMA). LocusZoom was used to create regional plots and identify credible sets.³⁹

Publicly Available Data

We obtained genome-wide summary statistics for AKI as assessed by diagnostic codes for ‘acute renal failure’ from three EHR-based biobank cohorts: BioBank Japan(BBJ),⁴⁰ and meta-analyzed FinnGen⁴¹ and UK Biobank(UKB),⁴² that were made available and previously reported.⁴³ AKI was defined in each cohort using diagnostic codes: BBJ diagnosis (ICD-10 N17), UKB Biobank phecode 585.1, and FinnGen endpoint N14-ACUTERENFAIL (presence of ICD-10 N17 and/or ICD-9 584). We evaluated the top variants from our primary analysis in this dataset in a targeted replication strategy to evaluate consistency of genetic loci nominally associated with AKI across differing phenotyping approaches. Upon finding evidence of association, these complete summary statistics data were then meta-analyzed with the BioVU and MVP sets as described above (analysis adjusted for age, sex, and PCs only to match publicly available datasets), and heterogeneity statistics across cohorts were reported for any significant variants.

Kidney multi-trait genetic correlation and gene expression analyses

Genetic correlation with published GWAS of kidney function traits were also assessed with LD score regression. We performed colocalization analysis with several renal traits including CKD,⁴⁴ eGFR,⁴⁵ blood urea nitrogen⁴³ and urine albumin-creatinine ratio,⁴⁶ as well as with eQTLs from kidney tissue, using the R package ‘coloc’. Colocalization was reported for those with a posterior probability of colocalization of > 0.8.

We also interrogated the Susztak Lab Kidney Biobank and NephQTL2 resources for all SNPs within 500kb of lead SNPs to identify those implicated in gene regulation from human healthy or diseased, respectively, kidney bulk and compartment-specific RNA sequencing or associated with methylation of the same tissues.^{45, 47, 48}

We inferred kidney gene expression associated with AKI using S-PrediXcan.⁴⁹ We used kidney cortex (Genotype-Tissue Expression [GTEx] project v8) gene expression models from predictDB.org with our primary meta-analysis summary statistics and those from the meta-analysis with public code-based phenotype data.^{50, 51}

Results:

There were 1,474,447 and 197,342 hospitalized patients in MVP and BioVU, respectively. After applying selection criteria, there were a total of 54,488 patients who met our definition for KDIGO persistent Stage 1 or worse AKI (49,030 MVP, 5,458 BioVU) and 138,051 non-AKI controls (126,580 in MVP, 11,471 BioVU; Figure 1, Tables 1–2). The BioVU cohort was younger and had a higher proportion of females (Tables 1–2). AKI cases had a higher burden of comorbidities, illness severity, baseline serum creatinine and lower eGFR.

Table 1.

Baseline Characteristics of MVP Study Population at Index Hospitalization

Variable	AKI (MVP)	Non-AKI (MVP)	P value
N	49,030	126,580
Age	66 (59–73)	60 (51–67)	<2.2×10−16
Female Gender (%F)	1,805 (4%)	13,563 (11%)	<2.2×10−16
RACE/Ethnicity
EUR/White	34,557 (70%)	93,623 (74%)
AFR/Black	11,502 (23%)	23,794 (19%)
HIS	2,515 (5%)	7,297 (6%)	<2.2×10−16
ASN	213 (1%)	1,013 (<1%)
NHPI	102 (<1%)	365(<1%)
AIAN	141 (<1%)	486(<1%)
Body Mass Index	29.7 (25.6–34.7)	29.2 (25.7–33.2)	<2.2×10−16
Diabetes	28,143 (57%)	34,052 (27%)	<2.2×10−16
Hypertension	44,874 (92%)	82,349 (65%)	<2.2×10−16
Coronary Artery Disease	25,235 (51%)	29,153 (23%)	<2.2×10−16
Heart Failure	16,574 (34%)	7,965 (6%)	<2.2×10−16
Sepsis	5,741 (12%)	1,065 (<1%)	<2.2×10−16
Mechanical Ventilation	3,693 (8%)	665 (<1%)	<2.2×10−16
Cardiac Surgery	1,788 (4%)	1,096 (1%)	<2.2×10−16
Abdominal/Vascular surgery	6,491 (13%)	6,138 (5%)	<2.2×10−16
Cardiac Angiography	2,609 (5%)	8,533 (7%)	<2.2×10−16
CKD (by admin code)	16,392 (33%)	4,213 (3%)	<2.2×10−16
Baseline Creatinine	1.15 (0.95–1.44)	0.97 (0.85–1.10)	<2.2×10−16
Baseline eGFR	68.6 (51.3–85.9)	85.8 (73.3–96.9)	<2.2×10−16
% <60 ml/min/1.73m2	37%	8%
Peak Creatinine	2.24 (1.73–3.15)	0.99 (0.84–1.10)	<2.2×10−16
Dialysis	1,555 (3%)	NA
Maximum AKI Stage
1	29,705 (61%)	NA
2	10,813 (22%)
3	8,512 (17%)

MVP: Million Veteran Program, BioVU: Vanderbilt Unviersity Medical Center BioVU cohort, EUR: european ancestry, AA: African Ancestry, HIS:Hispanic, ASN: Asian, NHPI: Native Hawaiin and Pacific Islander, AIAN: American Indian and Native Alaskan individuals, CKD chronic kidney disease, eGFR estimated glomerular filtration rate

Table 2.

Baseline Characteristics of BioVU Study Populations at Index Hospitalization

Variable	AKI (BioVU)	Non-AKI (BioVU)	P value
N	5,458	11,471
Age	61 (51–70)	56 (43–67)	<2.2×10−16
Female Gender (%F)	2,469 (45%)	5,853 (51%)	<2.2×10−16
Race
EUR/White	4,693 (86%)	9,939 (87%)	0.18
AFR/Black	765 (14%)	1,532 (13%)
Ethnicity
HL	46 (1%)	158 (1%)
NHL	5,371 (98%)	11,234 (98%)	0.0108
Unknown	41 (1%)	79 (1%)
Body Mass Index	28.7 (24.5–34.2)	29.4 (28.2–33.2)	8.3×10−6
Diabetes	2,028 (37%)	1397 (12%)	<2.2×10−16
Hypertension	3,664 (67%)	4,584 (40%)	<2.2×10−16
Coronary Artery Disease	2,084 (38%)	1649 (14%)	<2.2×10−16
Heart Failure	1,908 (35%)	801 (7%)	<2.2×10−16
Sepsis	868 (16%)	213 (2%)	<2.2×10−16
Mechanical Ventilation	1,320 (24%)	554 (5%)	<2.2×10−16
Cardiac Surgery	259 (5%)	309 (3%)	4.8×10−7
Abdominal/Vascular Surgery	1,772 (32%)	1439 (13%)	<2.2×10−16
Cardiac Angiography	343 (6%)	446 (4%)	1.3×10−6
CKD (by admin code)	1,628 (30%)	203 (2%)	<2.2×10−16
Baseline Creatinine	1.05 (0.83–1.38)	0.84 (0.72–1.00)	<2.2×10−16
Baseline eGFR	70.3 (50.0–92.2)	88.8 (73.2–104.0)	<2.2×10−16
% <60 ml/min/1.73m2	37%	11%
Peak Creatinine	2.2 (1.7–3.2)	0.83 (0.7–1.0)	<2.2×10−16
Dialysis	330 (6.0%)	NA
Maximum AKI Stage
1	2789 (51%)
2	1425 (26%)	NA
3	1262 (23%)

Meta-analyses of GWAS data across population-stratified data from MVP and BioVU (N=192,539 total) and adjusted for age, sex, baseline eGFR, and the top ten principal components of ancestry revealed two loci that reached genome-wide significance (Figure 2, Table 3, Supplementary Table S2a): rs11642015 near FTO (chromosome 16, odds ratio [OR] 1.07 (95% CI, 1.05–1.09), P =1.9×10⁻¹⁵, EAF [effect allele frequency] = 0.36; Supplementary Figure S2a–S2b) and rs4859682 near SHROOM3 (chromosome 4, OR = 0.95 (95% CI, 0.93–0.96), P = 2.3×10⁻⁹, EAF=0.40; Supplementary Figure S2c–S2d). Effects at both loci were largely consistent across groups (heterogeneity P > 0.05; Supplementary Table S2b, Supplementary Figures S2b and S2d).

Figure 2. Manhattan plot displaying SNP association results with AKI in MVP and BioVU adjusted for baseline eGFR.

The red dashed line represents a significant p-value threshold of 5.0×10⁻⁸.

Table 3.

Polymorphisms Associated with AKI

rsID	Nearest Gene	chr	Pos(hg37)	Effect/Other allele	Freq	MVP + BioVU Meta-analysis (adjusted for age, sex, PCs, and baseline eGFR)		MVP + BioVU Meta analysis (adjusted for age, sex, PCs, and baseline eGFR, BMI, and DM)
						Odds Ratio (95% CI)	P value	Odds Ratio (95% CI)	P value
rs11642015	FTO	16	53802494	C/T	0.36	1.07 (1.05–1.09)	1.9×10 −15	1.04 (1.02–1.06)	1.1×10−5
rs4859682	SHROOM3	4	77410318	A/C	0.40	0.95 (0.93–0.96)	2.3×10 −9	0.95 (0.93–0.97)	1.3×10−7

MVP: Million Veteran Program, BioVU: Vanderbilt Biobank, PCs: principal components of ancestry, eGFR: estimated glomerular filtration rate, BMI: body mass index, DM: Diabetes Mellitus

Analysis stratified by population did not identify additional significant loci (Supplementary Tables S3a–S3f). Only the FTO locus was significant in the EUR/White analysis (Supplementary Table S3a; N_cases=39,250, N_controls=103,564; lead SNP rs62048402, P = 3.9×10⁻¹⁷) while a different locus on chromosome 16 was significant in AFR/Black analysis (Supplementary Table S3b; between ADAMTS18 and NUDT7, N_cases=12,267, N_controls=25,326), no other regions were significantly associated in either strata. No other groups had significant variants.

Linkage disequilibrium (LD) score regression analysis indicated that SNP-based heritability for this phenotype is rather low (h₂= 0.031 (95% CI 0.025–0.037)), while the intercept was 1.015 (lambda = 1.074) suggesting adequate control of population stratification.

Due to the significant known association between the FTO locus and obesity and diabetes, we performed a secondary analysis adjusting for both conditions (Table 3, Supplementary Figure S3; Supplementary Table S4). The effect size for rs11642015 (OR 1.04 (95% CI:1.02–1.06, P =1.1×10⁻⁵) remained consistent with prior analyses; however, statistical significance was attenuated due to reduced power caused by missing BMI values (~10%). The association at SHROOM3 also remained consistent in magnitude (rs4859682, OR 0.95 (95% CI: 0.93–0.97)), though no longer genome-wide significant (P =1.3×10⁻⁷).

Kidney multi-trait genetic correlation and gene expression analyses

Genetic correlation with previously published eGFR GWAS results⁴⁵ indicated a significant negative correlation between AKI and eGFR (rg = −0.19; 95% CI: −0.25 – −0.13; P = 1.81×10⁻⁷), while correlations with CKD (rg=0.14), blood urea nitrogen (0.06), and urine albumin-creatinine ratio (0.64) were not significant (0.2<P>0.08; Table 4). Both significant loci (chromosome 4-SHROOM3 and chromosome 16-FTO) also colocalized between AKI-CKD and AKI-eGFR (posterior probability of colocalization > 0.8) at numerous SNPs (Supplementary Table S5).

Table 4.

Genome-Wide Correlation with Kidney Function Traits

phenotype	rg	se	z	P
CKD	0.1353	0.0782	1.7302	0.084
eGFR	−0.1898	0.0364	−5.2182	1.8×10–7
BUN	0.0635	0.0453	1.4014	0.16
UACR	0.642	0.4944	1.2986	0.19

rg: genetic correlation coefficient, se: standard error, z: z test statistic, p: P value

To gain functional information about the genetic signals identified, we performed functional mapping and annotation of our results using FUMA. There were no significantly enriched gene sets among the full summary results using the generalized gene-set analysis of GWAS data (MAGMA) analysis tool, though MAGMA tissue expression analysis identified kidney cortex as the top enriched tissue (P = 0.0054, not significant after correction for multiple tests) from GTEx v8 (Supplementary Figure S4). Kidney medulla was ranked second among the 53 tissues (p = 0.028; all other tissues P > 0.05). Gene2Func analysis of the six genes mapped to the two loci (FTO and CHD9 on chromosome 16, SHROOM3, CCDC158, FAM47E, FAM47E-STBD1) indicated that there was no significant differential expression of these genes in any GTEx tissues, nor were they enriched in any gene sets.

To more precisely identify whether AKI genetic risk is enacted through renal gene regulation, we performed genetically predicted gene expression analysis using the GWAS results (Supplementary Table S6). Among 1,633 genes with predictive models constructed for GTEx kidney cortex tissue from predictdb.org, the top gene identified (p = 0.0026) was the long non-coding RNA CTD-3064H18.2, where increased predicted expression associated with a decreased risk of AKI, though this association was not significant after correction for multiple tests (0.05/1633 = 0.000031). Due to the small numbers of kidney tissue in GTEx for model construction (N=65), we also used larger resources from the Susztak lab and NephQTL2 to identify whether any of our lead significant SNPs were eQTLs in kidney. The lead SNP near SHROOM3 (rs4859682) was associated with expression of SHROOM3 in meta-analyzed bulk kidney tissue (Table 5), however in cell fractionated expression analyses the SNP was associated with expression of CCDC158 in glomeruli. No associations with any SNPs in either locus were observed with expression in tubules.^{45, 47 48} Broader evaluation of all 16,499 SNPs within 500kB of each of our two lead SNPs identified 1176 significant (P < 3×10⁻⁶) eQTLs in glomeruli, 949 in tubules, and in normal bulk kidney tissue after correction for multiple tests (Supplementary Table S7a–S7b). The most significant variant was rs6827617 with NAAA expression in bulk kidney and both kidney compartments from healthy (Supplementary Table S7b) and dysfunctional kidney tissue (Supplementary Table S7c).⁵² Nine genes had at least one significant eQTL in bulk kidney, and eleven in each of the healthy kidney compartments, for a total of fifteen unique gene transcripts. Four of these (AKTIP, RBL2, RP11–44F14.2 and RP11–44F14.8) were located on chromosome 16 near FTO, the remainder were located on chromosome 4. STBD1 was unique to bulk kidney and AKTIP and ART3 were only significant in glomeruli.

Table 5.

eQTL Results from Kidney Tissue Resources.

Gene Name	SNP	Ref/Alt	Beta	Std	P value	SNP_Location(hg37)	Compartment	Direction*
CCDC158	rs13146355	G/A	−0.27	0.054	1.29×10−6	4:77412140	Glomeruli	-
CCDC158	rs10025351	C/T	−0.27	0.055	1.36×10−6	4:77394095	Glomeruli	-
CCDC158	rs4859682	C/A	−0.27	0.055	1.81×10−6	4:77410318	Glomeruli	-
CCDC158	rs28394165	T/C	−0.27	0.055	1.92×10−6	4:77394018	Glomeruli	-
SHROOM3	rs4859682	C/A	0.12	0.023	4.23×10−7	4:77410318	Bulk Kidney	++++
SHROOM3	rs60529470	G/A	0.12	0.024	4.24×10−7	4:77365883	Bulk Kidney	++++
SHROOM3	rs13146355	G/A	0.12	0.023	5.39×10−7	4:77412140	Bulk Kidney	++++
SHROOM3	rs28394165	T/C	0.12	0.023	6.45×10−7	4:77394018	Bulk Kidney	++++
SHROOM3	rs1398018	T/C	0.12	0.024	7.50×10−7	4:77372923	Bulk Kidney	++++
SHROOM3	rs10025351	C/T	0.12	0.023	7.86×10−7	4:77394095	Bulk Kidney	++++
SHROOM3	rs55940751	C/T	0.12	0.024	1.00×10−6	4:77365891	Bulk Kidney	++++
SHROOM3	rs1398016	G/A	0.11	0.024	2.36×10−6	4:77367688	Bulk Kidney	++++
SHROOM3	rs10008637	T/C	0.11	0.023	3.33×10−6	4:77414144	Bulk Kidney	++++
STBD1	rs1398016	G/A	0.19	0.041	6.20×10−6	4:77367688	Bulk Kidney	+?++
STBD1	rs28394165	T/C	0.18	0.041	7.18×10−6	4:77394018	Bulk Kidney	+?++
STBD1	rs1398018	T/C	0.19	0.042	7.55×10−6	4:77372923	Bulk Kidney	+?++
STBD1	rs55940751	C/T	0.19	0.041	7.96×10−6	4:77365891	Bulk Kidney	+?++
STBD1	rs4859682	C/A	0.18	0.041	8.23×10−6	4:77410318	Bulk Kidney	+?++
STBD1	rs10025351	C/T	0.18	0.041	8.48×10−6	4:77394095	Bulk Kidney	+?++
STBD1	rs60529470	G/A	0.18	0.042	1.04×10−5	4:77365883	Bulk Kidney	+?++
STBD1	rs13146355	G/A	0.18	0.041	1.17×10−5	4:77412140	Bulk Kidney	+?++
STBD1	rs10008637	T/C	0.18	0.041	1.57×10−5	4:77414144	Bulk Kidney	+?++

SNP: single nucleotide polymorphism

We also interrogated methylation quantitative trait loci (mQTLs) in kidney tissue within 500kB of each of our lead SNPs. A total of 11,819 pairs were significant (p<1×10⁻⁶; Supplementary Table S8), representing 131 unique CpG sites and 3,071 SNPs. The strongest association was between rs7203644 (near AKTIP on chromosome 16) and cg09728985 (P = 1.13×10⁻⁵¹). Of the numerous significant results, the vast majority (10,191) were on chromosome 4, while 1,673 were near the FTO region on chromosome 16.

Integration with code-based phenotypes from publicly available data.

We identified three large biobanks with AKI phenotyped only by administrative codes to explore potential consistency of our signals despite the differences. We performed the following analyses:

1. Targeted replication of the significant loci using summary statistics for acute renal failure by administrative code from Biobank Japan (BBJ), FinnGen, and UK Biobank (UKB; total 8,160 cases and 648,916 controls).⁴³ Evidence of association was observed at both the FTO locus (rs11642015, P = 0.011; Table 6a, Supplementary Figure S2b) and the SHROOM3 locus (rs4859682, P = 0.021; Supplementary Figure S2d), driven mainly by the two European ancestry datasets (rs11642015, P = 0.0053; rs4859682, P = 0.028). Notably, effect allele frequencies for both lead variants are lower in BBJ than in the European ancestry cohorts (0.20 compared to 0.40).

2. We also performed an exploratory genome-wide meta-analysis across MVP and BioVU results (adjusted only for age, sex and the top principal components of ancestry for consistency as eGFR was not available in the external cohorts) with the publicly available summary statistics from BBJ, FinnGen, and UKB. This analysis contained a total of 62,648 cases and 786,967 controls (Table 6b, Supplementary Table S9, Supplementary Figure S5) and identified two additional significant loci, including rs2065418 on chromosome 11 at MPPED2 [P = 3.5×10⁻⁸, EAF = 0.68; OR 1.04 (1.03–1.06) (Figure 2b; Supplementary Figure S6a–S6b) and rs4134943 near E2F3 on chromosome 6 [P = 2.1×10⁻⁸, EAF = 0.19; OR 0.95 (0.93–0.97) (Figure 2b; Supplementary Figure S6c–S6d). FTO (rs11642015) remained highly significant with the addition of public data (P = 4.4×10⁻²²), while SHROOM3 (rs4859682) did not (P = 1.2×10⁻⁷).

Table 6.

Targeted Replication of SNPs from eGFR-adjusted analysis (MVP+BioVU) within Publicly Available Biobanks (FinnGen, UKBB, BBJ)

rsID	Effect/Other allele	UKB Freq	FinnGen Freq	BBJ Freq	Meta-Analysis (Sakaue et al)		EUR		BBJ
					Odds Ratio (95% CI)	P value	Odds Ratio (95% CI)	P value	Odds Ratio (95% CI)	P value
rs11642015 (FTO)	C/T	0.40	0.42	0.20	1.04 (1.01–1.08)	0.011	1.05 (1.01–1.08)	0.0053	0.93 (0.79–1.09)	0.35
rs4859682 (SHROOM3)	A/C	0.46	0.46	0.22	0.96 (0.93–0.99)	0.021	0.96 (0.93–0.99)	0.028	0.94 (0.81–1.10)	0.46

The increased power for discovery in this meta-analysis also allowed for mapping of genes to expression in kidney tissue (Table 7). Two genes were significant in predicted expression analysis with kidney cortex tissue: CHPT1 (P = 1.1×10⁻⁵) and FGF5 (P = 2.8×10⁻⁵), Increased expression of both genes was associated with a decreased risk of AKI. CHPT1 is located on chromosome 12 and encodes choline phosphotransferase 1. FGF5, located on chromosome 4, codes for fibroblast growth factor 5.

Table 7.

Meta-analysis with Publicly Available Data (MVP + BioVU + FinnGen + UKBB + BBJ)

rsl D	Nearest Gene	chr	pos (hg37)	Effect/Other allele	Freq	MVP + BioVU + Public data
						Odds Ratio (95% CI)	P value
rs11642015	FTO	16	53802494	C/T	0.36	1.07 (1.06–1.08)	4.4×10 −22
rs2065418	MPPED2	11	30422068	G/T	0.68	1.04 (1.03–1.06)	3.5×10 −8
rs4134943	E2F3	6	20483407	C/T	0.19	0.95 (0.93–0.97)	2.1×10 −8

Analysis adjusted only for age, sex, and top 10 principal components of ancestry to match publicly available datasets.

Discussion

In this large GWAS meta-analysis, we found two loci associated with AKI after adjustment for baseline eGFR, rs11642015 near FTO, and rs4859682 near SHROOM3 in our primary study cohorts, the latter of which has previously associated with eGFR and albuminuria.^{44, 53, 54} Neither SNP remained statistically significant after additional adjustment for BMI and diabetes status at index hospitalization.

SHROOM3 is an actin-binding protein that contributes to kidney tubular epithelial development.^{55, 56} Expression has been observed in the glomerulus, podocyte, and tubules,^56–58 with GWAS identifying associations with eGFR and albuminuria.^{44, 53, 54} Alterations in the orientation of tubular epithelium have also been observed in heterozygous SHROOM3 null mice with preserved function under normal conditions.⁵⁹ However, these mice have increased susceptibility to and poor recovery from AKI due to disrupted epithelial redifferentiation and disorganized F-actin.⁶⁰ The latter suggest that under stress, variations in gene expression may influence susceptibility to AKI. Although our identified SNP has also been demonstrated to be associated with eGFR and serum creatinine,^{61, 62} these observations may also help explain its persistent association with AKI after adjustment for baseline eGFR. Notably, the association did not reach statistical significance after further adjustment for diabetes and BMI. Whether the latter is potentially due to a larger incremental risk conferred by diabetes and obesity or a sharing of common pathways in conferring AKI risk is unknown.

The fat mass and obesity-associated gene (FTO) polymorphism (rs11642015) located on chromosome 16 is associated with the development of obesity, type 2 diabetes mellitus, and hypertension,^63–66 known risk factors for AKI. We performed an adjusted analysis for BMI and diabetes status, which resulted in similar effect size that no longer reached statistical significance, suggesting that some of the association may be explained through a predisposition to known risk factors for AKI. Notably, our power was reduced by some missing BMI data, and the lack of more precise measures of body composition beyond BMI may have reduced our ability to better quantify the impact of adiposity on this association. Preclinical models have suggested a direct role for FTO in the pathogenesis of AKI. In one cisplatin-induced model of AKI, reduced FTO expression via meclofenamic acid inhibition increased RNA m⁶A levels, upregulation of p53, and aggravated injury.⁶⁷ A follow-up study also demonstrated enhanced DNA-methylation of FTO in an alcohol-induced model of AKI as associated with increased inflammation.⁶⁸ Lastly, studies have shown that FTO deficiency reduces the fibrogenic response in a ureteral obstruction model, possibly via reduction in TGF-beta stimulation.⁶⁹ Notably, this SNP has not been shown to be in eQTL in kidney tissue, however, it is shown to be associated with expression of FTO in skeletal muscle and IRX3 in pancreas.^70–72

CHPT1 has been observed to be differentially expressed in studies of kidney ischemia/reperfusion injury, with consistent directions as this study.⁷³ CHPT1 acts in the final step of the CDP-choline pathway biosynthesizing choline phospholipids, which are major components of cell membranes. Variants near FGF5 have been previously associated with CKD progression,⁷⁴ while proteomic studies have also implicated it in eGFR.⁷⁵ FGF5 is among several fibroblast growth factors expressed in the developing kidney.⁷⁶ FGF5 treatment has also shown promise in studies of acute lung injury,⁷⁷ which may indicate a potential shared relationship of fibrosis in acute tissue damage.

Several GWAS and targeted gene analyses have been performed in the literature; however, few, if any, potential candidates have successfully replicated externally. Potential reasons include lack of power, misclassification of parenchymal injury due to lower specificity of modest changes in serum creatinine to define AKI, overreliance on administrative codes, and AKI heterogeneity itself. We observed nominal significance (p<0.05) at three SNPs from prior AKI-related GWAS: rs80052123, rs7546189, and rs9945894 (Supplementary Table S10),⁷⁸ however, one of our lead SNPs (rs2065418 near MPPED2 on chromosome 11) in our exploratory meta-analysis including publicly available GWAS data has been previously associated with AKI, severity of illness, and inflammation in acutely ill patients with trauma and other kidney conditions.^79–83 In addition, the variant rs4134943 (chromosome 6), lies between CDKAL1⁸⁴ and E2F3, though functional evidence from hematopoietic cells links the SNP more directly to E2F3. E2F3 encodes a member of a small family of transcription factors we have previously demonstrated to be associated with eGFR and creatinine,^{43, 82} supporting E2F3 as the likely functional gene at this locus.

Despite our sample size, we identified few candidate genes compared to earlier studies. Potential reasons for this include a broader case-mix for AKI (i.e., all hospitalized patients); however, the latter may also increase the probability that candidate genes identify may represent more general susceptibility genes. Other potential reasons include the comparatively large effect of environmental stressors in the pathogenesis of AKI (e.g., bleeding, shock) relative to heritable susceptibility or that the latter may partially manifest more indirectly through other known risk factors for AKI.

Strengths of the study include a large sample size, enrichment for AKI more likely to represent parenchymal injury, and the integration of publicly available data. We also attempted to exclude historical evidence of AKI rather than limit classification to a single hospitalization to reduce potential misclassification of controls. Limitations include a low proportion of female patients within the MVP. It also remains possible that some cases may have had AKI that did not have parenchymal injury despite our persistence criteria. While kidney biopsy is the gold-standard to characterize injury, these are rarely performed in clinical practice and were not available in sufficient quantities for study. Clinical adjudication of the etiology of AKI was also not feasible on this large-scale. Similarly, phenotyping of AKI in publicly available cohorts were limited to administrative codes. We were limited in our ability to adjust for comorbidities in these latter cohorts, and these exploratory results should not be viewed as definitive. Lastly, despite attempts to exclude patients with AKI from our control population, it is possible that AKI could have occurred outside of our ascertainment window, the institutions studied, or in an outpatient setting.

In conclusion, our large GWAS identified two gene regions associated with risk of hospitalized AKI, including a SNP that has been associated with tubular development and structure and another associated with enhanced inflammation and fibrosis in animal models that is also strongly associated with obesity and diabetes. Further studies to delineate the differences between premorbid susceptibility conferred by these candidate variants for AKI and directly conferred risk in the acute setting are warranted.

Supplementary Material

1

Supplementary Table S1. Administrative codes used.

Supplementary Table S2a,S2b. Suggestive GWAS results from cross-ancestry MVP + BioVU meta-analysis

Supplementary Table S3. Population stratified suggestive GWAS results from MVP + BioVU:

S3a: European ancestry individuals

S3b. African ancestry indviduals

S3c. Hispanic individuals

S3d. Asian individuals

S3e. American Indian and Native Alaskan individuals

S3f. Native Hawaiian and Pacific Islanders individuals

Supplementary Table S4. MVP+BioVU cross-ancestry meta-analysis adjusted for body mass index and diabetes status

Supplementary Table S5. Colocalization of top loci with CKD, eGFR, UACR, and BUN summary statistics.

Supplementary Table S6. Predicted gene expression from GTEx v8 Kidney cortex

Supplementary Table S7. Suggestive eQTLs (p<0.05) within 500KB of sentinel SNPs near SHROOM3 and FTO

S7a. Bulk normal kidney meta-analysis

S7b. Microdissected healthy kidney structures

S7c. NephQTL microdissected nephrotic kidney structures

Supplementary Table S8. Suggestive mQTLs (p<0.05) from bulk normal kidney tissue (n=466) with SNPs within 500KB of sentinel SNPs near SHROOM3 and FTO results adjusted for body mass index

Supplementary Table S9. Suggestive results from meta-analysis of MVP and BioVU with publicly available code-based BBJ, UKB and FinnGen

Supplementary Table S10. Replication of previously-reported AKI genetic loci

2

Supplementary Figure S1a-S1b). Population stratification (a) and inflation (b) by analysis subgroup. Principal components (PC) plots for each MVP and BioVU strata showing PC1 by PC2. Top left to right: MVP EUR, MVP AFR, MVP Hispanic, MVP Asian. Bottom left to right: BioVU White, BioVU Black, MVP American Indian or Native Alaskan, MVP Native Hawaiian or Pacific Islander. Quantile-quantile plots for each subgroup analysis and meta-analysis including lambda. Top left to right: MVP EUR, MVP AFR, MVP Hispanic. Middle left to right; BioVU White, BioVU Black, MVP Asian. Bottom left to right: MVP Native Hawaiian or Pacific Islander, MVP American Indian or Alaska Native, Meta-analysis.

Supplementary Figure S2a-S2d. Regional association plots (a, c) and forest plots (b, d) of FTO (a-b) and SHROOM3 (c-d). Lead SNPs for each regional plot (rs11642015, rs4859682, respectively) are indicated by purple circles with neighboring SNPs presented in varying shades reflective of their respective linkage disequilibrium to the lead SNP. Forest plots show effect estimates for each analyzed group for each SNP.

Supplementary Figure S3. Manhattan plot displaying SNP association results with AKI in MVP and BioVU adjusted for baseline eGFR, diabetes, and BMI. The red dashed line represents a significant p-value threshold of 5.0×10⁻⁸.

Supplementary Figure S4. FUMA gene enrichment among GTEx tissues. Kidney cortex was ranked first and kidney medulla third among 53 tissues, both with non-significant enrichment p-values. Methods: To test the (positive) relationship between highly expressed genes in a specific tissue and genetic associations, gene-property analysis is performed using average expression of genes per tissue type as a gene covariate. Gene expression values are log2 transformed average RPKM per tissue type after winsorized at 50 based on GTEx RNA-seq data. Tissue expression analysis is performed for 53 specific tissue types. MAGMA was performed using the result of gene analysis (gene-based P-value) and tested for one side (greater) with conditioning on average expression across all tissue types.

Supplementary Figure S5. Manhattan plot displaying SNP association results with AKI in MVP and BioVU meta-analyzed with publicly available summary statistics from Biobank Japan, FinnGen and UK Biobank. The red dashed line represents a significant p-value threshold of 5.0×10⁻⁸.

Supplementary Figure S6a-S6d. Regional association plots (a, c) and forest plots (b, d) of MPPED2 (a-b) and E2F3 (c-d). Lead SNPs for each regional plot (rs11642015, rs4859682, respectively) are indicated by purple circles with neighboring SNPs presented in varying shades reflective of their respective linkage disequilibrium to the lead SNP. Forest plots show effect estimates for each analyzed group for each SNP.

Table 8.

Predicted Gene Expression From GTEx v8 Kidney Cortex From Meta-analysis With Public Data

Gene	Ensembl ID	Chr.	Position (hg38)	Effect size	P value	Expression variance explained by model	Prediction performance r2	N_SNPs used	N SNPs in Model
CHPT1	ENSG00000111666	12	101696946101744140	−0.36	1.1×10 −5	0.0076	0.11	16	16
FGF5	ENSG00000138675	4	8026659880336680	−0.14	2.8×10 −5	0.019	0.20	12	12
STX2	ENSG00000111450	12	130789599130839266	0.22	6.7×10−5	0.017	0.08	20	20
GTF2I	ENSG00000263001	7	7465766674760692	0.28	9.3×10−5	0.013	0.12	13	13

Data Sharing Statement:

The protocol, statistical code, are available from Drs Siew and Hellwege. Summary statistics will be available in dbGAP. Per VA policy, the summary statistics will be deposited into dbGAP and our phenotypes will be shared through the government-sponsored website phenomics.va.ornl.gov/web. Per the United States Veterans Affairs Policy, primary genotype data/individual-level Veteran is considered highly protected by the US government and are not permitted to leave the Million Veteran Program (MVP) research environment or cross the VA firewall. Further data requests require approval from the Office of Research Development and the Million Veteran Program. The summary statistics for both the baseline-adjusted model and additionally adjusted summary statistics can be located with the following accession numbers at dbGAP: phs001672.v12.p1