This multi-center genome-wide association study followed the recommendations of Strengthening the Reporting of Genetic Association Studies (Little, 2009). Study outcomes, data collection, and statistical analyses were established a priori and presented at a multidisciplinary peer-review forum (4 February 2021) prior to data access (Multicenter Perioperative Outcomes Group, 2017). The study was approved by the Institutional Review Boards at both Vanderbilt (191829) and Michigan (HUM00126162). As no patient care interventions were made through conducting the study, patient consent was waived.

Inclusion criteria for the study were all patients who underwent cardiac surgery at Vanderbilt University Medical Center (“Vanderbilt”; Nashville, TN, United States) and Michigan Medicine of the University of Michigan (“Michigan”; Ann Arbor, MI, United States) and had genotype data available. For patients with more than one cardiac procedure, only the first case was considered. Patients were excluded if they had severe preoperative renal dysfunction or preoperative renal failure ¹ or were missing primary anesthesia CPT code data. Additional exclusion criteria were 1) American Society of Anesthesiologists Physical Status Classification VI; 2) procedure duration <45 min; 3) subsequent surgery within 7 days without documentation of a postoperative serum creatinine between cases; 4) implausible body mass index; ² 5) surgery classified as outpatient or 23-hour-admission; 6) or missing primary anesthesia base CPT code data.

The primary outcome was Stage 1, 2, or 3 kidney injury as defined by the Kidney Disease: Improving Global Outcomes (KDIGO) practice guidelines within 7 days following surgery or discharge (whichever occurred first) (Kellum et al., 2012). Baseline preoperative serum creatinine was established using the closest value to the time of surgery (within 1-year), as previously described. ³ (Mathis et al., 2020) The diagnostic criteria for the primary outcome relied on serum creatinine or receipt of dialysis, and did not incorporate urine output, which was not consistently recorded for the entire postoperative surveillance window in our dataset.

The Vanderbilt (BioVU) population had a median age of 62.9 (IQR 55.4–72.0) and was 67% male. Median case duration was 188.7 min (IQR 109.3–254.0). The Michigan Medicine (MGI) cohort had a median age 63.5 years (IQR 55.0–72.0), and was 69% male. Median case duration was 310.0 min (IQR 230.2–407.5). In the Michigan Medicine Cohort, the majority had hypertension (71.8%) and valvular disease (79.7%). Population characteristics for both BioVU and MGI cohorts are shown in Table 1.

Among 1,014 qualifying procedures at Vanderbilt University Medical Center and 478 at Michigan Medicine, 348 (34.3%) and 121 (25.3%) developed cardiac surgery-associated AKI, respectively (Figure 1). Cardiac surgery-associated AKI developed more frequently following longer procedures, and in patients with underlying comorbidities: chronic pulmonary disease and diabetes mellitus (Table 2). Most patients in both cohorts developed KDIGO, Stage 1 AKI. Additional details regarding the breakdown by KDIGO stage can be found in Supplementary Table S1.

Meta-analysis was performed on 6,291,460 SNPs. Among those: 5,364,338 SNPs had R² ≥ 0.8 in both datasets and 927,122 SNPs had R² < 0.8 in at least one dataset. No single nucleotide polymorphisms (SNP) exceeded our pre-determined threshold for genome-wide significance (p < 5 × 10⁻⁸) on any of the three logistic regression models. In the basic regression model, Model 1, which controlled for age, gender, and first four principal components, two SNPs were identified at the suggestive threshold: rs12421245 (p = 3.86 × 10⁻⁷) and rs73131342 (p = 8.76 × 10⁻⁷). The genomic inflation factor (λ) for Model 1 was 0.986. When preoperative serum creatinine was also controlled for (in addition to age, gender, and first four principal components) (Model 2), rs12421245 remained below the suggestive threshold (p = 3.96 × 10⁻⁷), rs73131342 was no longer below the suggestive threshold (p = 3.72 × 10⁻⁶), and rs3847598 decreased to below the suggestive threshold (p = 7.55 × 10⁻⁷). In Model 3, the logistic regression controlling for a full group of clinical covariates, rs12421245 (p = 9.81 × 10⁻⁷) remained suggestive. Additionally, three new SNPs were identified: rs113741905 (p = 1.47 × 10⁻⁷), rs74637005 (p = 9.58 × 10⁻⁷), and rs17438465 (p = 8.74 × 10⁻⁷). Full GWAS results for each of the three models can be found in Table 3. There was no significant heterogeneity between the two centers in any of the reported SNPs (Supplementary Table S2). Manhattan Plots and Quantile-Quantile (QQ) Plots for each of the three Models (meta-analysis) are shown in Supplementary Information (Supplementary Figure S1).

Next, we attempted to determine the gene associated with each SNP identified in the three models. We characterized the function of each variant exceeding the suggestive threshold (Supplementary Table S3). For non-exonic SNPs, the closest gene by chromosome position was selected. Genes identified using this methodology were: ZNF215, CDH26, LMO2/CAPRIN1, RCC2/ARHGEF10L, NFU1, and EVX1/HIBADH (Table 3).

We failed to replicate SNPs from prior unbiased studies of post-surgical acute kidney injury (Supplementary Table S4) (Westphal et al.; Stafford-Smith et al., 2015; Zhao et al., 2017).

When included as a covariate within Model 1, Model 2, and Model 3 of the validation cohort (BioVU), polygenic risk was not significantly associated with post-surgical acute kidney injury (Table 4) (p-value ranged from 0.117 in Model 3 to 0.215 in Model 1). Age was significant in all 3 models and preoperative serum creatinine was significant in both models that it was included as a covariate. In addition to age and preoperative serum creatinine; case duration in minutes (aOR = 1.002, 95% CI 1.000–1.003, p = 0.013), diabetes mellitus (aOR = 2.025, 95% CI 1.320–3.103, p = 0.001), and valvular disease (aOR = 0.558, 95% CI 0.372–0.835, p = 0.005) were included in Model 3 (the full clinical model). Clinical factors associated with cardiac surgery-associated AKI (and those that could not be shown to have a significant association) can be visualized in Figure 2. Full results of the three Models with PRS_AKI included can be found in Table 4. The discrimination was 0.573 (95% CI: 0.537–0.610) for Model 1; 0.661 (95% CI: 0.626–0.697) for Model 2; and 0.681 (95% CI: 0.646–0.715) for Model 3.

The heritability of cardiac surgery-associated AKI was calculated to be <0.001 with standard error of 0.378.

Our study failed to identify SNPs exceeding the pre-specified genome-wide significance level (p < 5 × 10⁻⁸). However, using the a priori defined lower suggestive threshold (p < 1 × 10⁻⁶) identified 6 previously unreported SNPs which may be associated with acute kidney injury following cardiac surgery (Supplementary Table S3). The SNP rs74637005 was located in the exonic region of NFU1, a gene previously shown to be associated with both calcium levels (Sakaue et al., 2021) and elevation in metabolic biomarkers (Martin et al., 2021). Three SNPs were located in intergenic regions (rs3847598, rs113741905, and rs17438465) and two SNPs were intronic (rs12421245 and rs73131342) (Supplementary Table S3). Notably, rs17438465 is located between EVX1 and HIBADH, genes that have been associated with nephrolithiasis and cardiovascular disease (Howles et al., 2019; Kichaev et al., 2019). No genes identified by proximity-based candidate gene analysis had previously been reported to be associated with acute or chronic renal disease.

Furthermore, SNPs exceeding suggestive threshold were interrogated using the HaploReg tool (HaploReg v4.1) (Ward and Kellis, 2012) with key findings highlighted in Supplementary Table S3. Notably, rs3847598 is predicted to be bound by GATA2, a protein already described in AKI (Yu et al., 2017), and rs12421245 is annotated as expression quantitative trait loci of ZNF215 in different tissues. Additionally, rs73131342 alters the Pax8 and Sox motifs. Pax8 is notably a nephric-lineage transcription factor required for renal organogenesis (Tong et al., 2009) potentially involved in regeneration and recovery after acute renal injury (Zivotic et al., 2022) and Sox activation is an early transcriptional response to acute kidney injury (Kumar et al., 2015).

These findings should be considered hypothesis generating and require additional replication and validation.

We failed to replicate any of the variants previously reported to be associated with cardiac surgery-associated acute kidney injury. Differences in phenotype potentially account for this finding. Our study focused on the narrow phenotype of acute kidney injury following cardiac surgery, whereas, the RIPHeart-Study assessed a broad composite of complications occurring following cardiac surgery (although they did adjust for each individual complication) (Westphal et al.). The one SNP in the RIPHeart-Study exceeding genome-wide significance (rs78064607 located in PHLPP2) for AKI was notably not available for inclusion in our meta-analysis and no SNPs in close linkage disequilibrium could be identified to provide a suitable approximation.

We had a similar incidence of cardiac surgery-associated AKI as defined by the KDIGO criteria (31.4% compared with 33.0% in the combined PEGASUS and CATHGEN cohorts) compared to the prior GWAS in cardiac surgery-associated AKI (Stafford-Smith et al., 2015). The GWAS was performed using the continuous variable, percent rise in serum creatinine from baseline level within 10-day following surgery, compared to our KDIGO-based outcome. Zhao and colleagues reported an AKI incidence of 35.8% in the discovery cohort from the Translational Research Investigating Biomarkers Endpoints in AKI (TRIBE) cardiac surgery study and 14.7% (206 cases and 1,406 controls) in the replication cohort (CABG Genomics and Brigham’s Cardiac Surgery Study sub-cohorts) (Zhao et al., 2017). This study defined AKI cases as patients with at least a 0.3-mg/dL or 50% increase in serum creatinine from baseline for at least two consecutive days and non-AKI controls as patients with an increase in serum creatinine not exceeding 25% relative to baseline (Zhao et al., 2017).

The majority of patients in our study developed KDIGO Stage 1 AKI (89% of all AKI in the BioVU cohort and 88% in MGI; Supplementary Table S1). Restricting the analysis to only the patients developing Stage 2 or Stage 3 AKI may reveal a stronger genetic linkage, however, this small sample size precluded this analysis in our 2-center cohort. Furthermore, the incidence in our cohort exceeds that reported in lower risk cardiac surgery cohorts (Hu et al., 2016), suggesting the presence of high-risk cases in the dataset (for example, those involving the aorta, requiring hypothermic circulatory arrest, or inserting ventricular assist devices). In these high-risk cases, the genetic component may contribute very little to predictive accuracy, but for low risk cases (for example, non-emergent, isolated coronary artery bypass graft or isolated valve) genetic susceptibility may play a greater role. Further studies, restricted to lower risk cardiac cases, may reveal populations where genetics exerts larger impact on overall risk profile.

We assess genetic associations between a large number of genetic polymorphisms (BioVU 11,292,079; MGI 10,642,058; Meta-analysis 6,291,460) across a large number of patients (1,492) and surgical procedures. This contrasts to prior genetic studies in AKI which have largely employed the candidate gene approach (Larach et al., 2018). Additionally, we expanded the breadth of our genetic analyses from GWAS to develop a PRS, which quantifies the combined contribution from hundreds of thousands of genetic variants, each of which may individually have a small effect. This single number was then integrated with traditional clinical risk factors in an attempt to better predict AKI.

A major strength of the study was the collaboration between two academic biorepositories (BioVU and MGI). This provided the largest possible cohort for GWAS (BioVU + MGI), and allowed validation of PRS in an independent dataset (MGI = Base Dataset, BioVU = Target Dataset). Our meta-analysis had minimal genomic inflation (λ = 0.986). Meta-analysis of suggestive SNPs also confirmed the directionality of the effect. Another strength of our study was the use of a validated data collection method (covariates) combined with the standardized KDIGO outcome. The major limitation of our study is that we are underpowered to detect the effect of low-frequency variants, exerting only a moderate impact on the relative risk of AKI. Despite this limitation, we assess novel associations and attempt to replicate previously described associations in two independent biobanks with well-curated phenotypic and procedural details. We selected base (MGI) and target (BioVU) cohorts based upon ready availability of the additional clinical features within BioVU that we incorporated into the full clinical model (Model 3). As prediction accuracy of PRS depends mostly on the heritability of the analyzed traits, the number of SNPs, and the size of the discovery sample; our PRS is inherently limited based upon the sample size of our base MGI cohort.

Another inherent limitation of our dataset is the lack of genetic diversity in both biobanks. While data within both BioVU and MGI reflect the patient population served by both institutions and lack of ancestral diversity is commonly encountered problem across many biobanks (Zhou et al., 2019), integration of our data with global biobanks composed of individuals from more diverse ancestries will enable even more equitable, impactful research.

While our evidence suggests that there is limited heritability of postsurgical AKI in the cardiac surgery population, this study only interrogated common genetic variants (minor allelic frequency >1%). Thus, the possibility remains that rare variants contribute risk (or even protection) for some cardiac surgery patients. Our heritability calculation is likely limited by smaller than needed effective sample size, limiting interpretability and generalizability to larger cohorts (Neale Lab, 2017). In GWAS studies with small effective sample sizes (i.e., below 5,000–10,000), statistical artifacts can result in near zero heritability estimates, even in phenotypes with known heritability. Our small sample size and large standard error likely limits the utility of this heritability analysis and further studies with larger sample size are required to better assess heritability of cardiac surgery-associated AKI. Finally, index event bias may be impacting discovery of biologically relevant genetic risk variants, although this is less relevant for risk prediction. Despite these limitations, our results suggest that polygenic risk scores based upon common variants likely have limited predictive utility.