Evaluating a Polygenic Risk Score for Hepatocellular Carcinoma in Multiethnic Cirrhosis Patients in Texas

ABSTRACT

Background and Aims

Genome‐wide association studies (GWAS) that were previously conducted in European and Asian cohorts have identified numerous genetic risk factors for hepatocellular carcinoma (HCC) and its precursors and risk factors. A polygenic risk score (PRS) composed of these key SNPs for HCC may be useful for risk stratification in clinical settings for U.S. patients with cirrhosis.

Methods

Using genetic and clinical data from the Texas HCC Consortium cohort, we evaluated a weighted PRS constructed from 11 independent SNPs previously shown to be associated with advanced liver disease (PNPLA3–TM6SF2–GCKR–HSD17B13–GATAD2A–TRIB2–PPP1R3B–CPN1–ERLIN1–APOE–MBOAT7). We used Fine‐Gray competing risk models to estimate hazard ratios (HR) and 95% confidence intervals (CI) to test the association between the PRS and HCC.

Results

The 1888 patient cohort included non‐Hispanic White (NHW; 50.4%), non‐Hispanic Black (NHB; 18.5%), and Hispanic (28.6%) patients, and 116 developed incident HCC. Each unit increase in the PRS was associated with increasing risk of HCC (HR = 2.20; 95% CI (1.53, 3.18)). The 11‐SNP PRS was strongly associated with risk of HCC in NHW patients (HR = 2.53, 95% CI (1.65, 3.89)) but not among NHB and Hispanic patients. Similarly, the magnitude of the PRS effect on HCC risk differed by cirrhosis etiology with the strongest association for cirrhosis due to Hepatitis C (HR = 3.81, 95% CI (1.64, 8.82)).

Conclusions

Our study shows that PRS is generally promising for HCC risk prediction; however, the current PRS that was derived from previous GWAS of primarily NHW and Asian patients is not robust over different groups, limiting their use for risk stratification in diverse populations with different etiologies.

Abbreviations

A1

test allele

AF

allele frequency

AFP

alpha‐fetoprotein

ALD

alcohol‐related liver disease

ALT

alanine aminotransferase

BMI

body mass index

Chr

chromosome

CI

confidence intervals

C‐statistic

concordance index

EMRs

electronic medical records

GWAS

genome‐wide association studies

HBV

hepatitis B virus

HCC

hepatocellular carcinoma

HCV

hepatitis C virus

HRs

hazard ratios

MASLD

metabolic dysfunction‐associated steatotic liver disease

MetALD

metabolic dysfunction and alcohol‐related liver disease

NHB

non‐Hispanic Black

NHW

non‐Hispanic White

NRI

net reclassification improvement

PRS

polygenic risk score

SNP

single‐nucleotide polymorphisms

THCCC

Texas Hepatocellular Carcinoma Consortium Cohort

THCC‐RI

Texas HCC Risk Index

1. Introduction

Hepatocellular carcinoma (HCC) is a primary cause of cancer‐related mortality worldwide [1, 2]. In the United States, HCC‐associated mortality rates have been increasing since the 1970s [3]. Texas has high HCC rates, and it was previously found that Hispanic populations in Texas had the highest HCC incidence rates in the U.S. [4]. Up to 95% of HCC cases develop in patients diagnosed with cirrhosis [5]. In the U.S., in recent years, the causes of cirrhosis and HCC have changed such that most cases now arise in patients with metabolic dysfunction‐associated steatotic liver disease (MASLD), alcohol‐related liver disease (ALD), or MetALD (metabolic dysfunction and alcohol‐related liver disease), rather than chronic hepatitis B virus (HBV) and hepatitis C virus (HCV) infections [6, 7].

Despite advancements in surveillance, diagnosis and treatment, most HCC cases are still diagnosed in patients with advanced disease, and overall 5‐year relative survival rates remain around 20% [8]. Thus, better risk stratification is still needed for improved prevention and early detection of HCC in populations at high risk of developing HCC. Risk stratification methods have been developed to identify patients at increased risk of developing HCC to some success [9]. However, these current models are only modestly reliable to predict and discriminate among cirrhosis patients who have varied risk for disease progression.

Genome‐wide association studies (GWAS) have previously identified several single‐nucleotide polymorphisms (SNP) that are associated with HCC, as well as its precursor conditions and risk factors. Many of the earlier GWAS and meta‐analyses of hepatocellular carcinoma identified variants that confer risk of hepatitis‐associated HCC [10, 11, 12, 13, 14, 15]. Subsequent GWAS have focused on risk variants associated with lifestyle‐ and chronic disease‐related HCC, as well [16, 17, 18, 19, 20, 21]. More recently, studies have followed GWAS with functional studies to identify risk variants for HCC and provide further evidence of these associations [18, 22].

Previous polygenic risk scores (PRS) comprised of hepatic fat SNPs (PNPLA3–TM6SF2–GCKR–MBOAT7) were predictive of HCC risk in MASLD patients of European ancestry in Italy and the United Kingdom and in a multiethnic cohort of patients with cirrhosis in the U.S. [23, 24]. This hepatic fat PRS was reported to have moderate discriminatory ability for risk of HCC in a Singapore Chinese cohort [18]. We previously showed that these same 4 SNPs in a PRS offer similar discriminatory ability for HCC risk among our racially/ethnically diverse cohort of patients with cirrhosis [24]. Nonetheless, given the strong heritable component of hepatic fat, these previous 4‐SNP PRS may not adequately capture heritable hepatic fat or HCC risk. Expanding the PRS to include more SNPs and explain more heritability of hepatic fat‐related HCC has the potential to improve discrimination and risk stratification for HCC in cirrhosis patients. However, since most GWAS have been conducted in people of European or East Asian ancestry, the generalizability of results to diverse populations, including African American and Hispanic populations in the U.S., is unclear [16, 17, 25].

Therefore, we evaluated the utility of an expanded PRS that included 11 independent and well‐established SNPs associated with susceptibility to hepatic fat [26]. Using data from a multiethnic cohort of patients diagnosed with cirrhosis, we evaluated the association of this 11‐SNP PRS with HCC risk in our full cohort and within subgroups defined by etiology, race and ethnicity, and sex [9].

2. Materials and Methods

2.1. Study Population

We studied data obtained on 1888 patients with cirrhosis from the Texas Hepatocellular Carcinoma Consortium Cohort (THCCC). The details of the THCCC have been previously reported [27]. Briefly, the THCCC identified eligible patients with cirrhosis and obtained informed consent during a clinic visit at one of the study sites in Texas (Michael E DeBakey VA Medical Center in Houston, Baylor‐St. Luke's Medical Center in Houston, University of Texas‐Southwestern, Parkland Health & Hospital System, Baylor University Medical Center‐ Dallas, Baylor All Saints Fort Worth, Doctor's Hospital at Renaissance in McAllen, and the Texas Liver Institute in San Antonio). Patients with previous HCC, hepatic decompensation, or current non‐melanoma primary cancers were excluded. Study participants completed surveys of demographic and epidemiological information at enrollment, and blood samples were obtained within 8 weeks of the baseline study visit. The participants' electronic medical records (EMRs) were also abstracted for relevant data including information on cirrhosis‐related diagnoses, liver disease treatments and complications, liver imaging, liver pathology, and laboratory results within 12 months of enrollment. Study participants received imaging surveillance for HCC (ultrasound, CT, or MRI) as per the treating physician. Participants were seen every 6–12 months in the clinic for routine liver disease follow‐up, which provided study follow‐up information as well.

2.2. Outcome

The primary outcome of interest was the development of primary incident HCC following study enrollment. Review of the study participants' EMRs was conducted at 3‐month intervals to capture data on incident HCC diagnosis, liver transplantation information, and death dates. HCC was defined as per the American Association for the Study of Liver Diseases criteria, including histological or radiological diagnosis using characteristic tumor features (arterial enhancement and delayed washout) on triphasic CT or MRI (Liver Imaging Reporting and Data System [LI‐RADS] 5) or suspicious lesions (LI‐RADS 4) that underwent multidisciplinary tumor board review, treated as HCC, and adjudicated centrally [24].

2.3. Clinical and Epidemiological Risk Factors

Clinical and epidemiological measures were obtained from a combination of clinician‐recorded information in the EMR and through the study survey. We obtained details about cirrhosis etiology, HCC diagnosis, liver disease treatment, and co‐morbidities from the EMR. The study participants completed the survey at study enrollment to provide their self‐reported race and ethnicity using a constructed variable, sex, history of alcohol consumption, and smoking status. We measured the height and weight of the participants to calculate BMI (body mass index, kg/m²).

2.4. Genotyping and Genetic Methods

Germline DNA was purified from whole blood samples using the QIAamp DNA Blood Mini kit (Qiagen, Valencia, CA), and the Axiom Precision Medicine Research Array was used to genotype the samples. The genotype data were aligned to human genome assembly GRCh37, and the variants were restricted to biallelic, autosomal SNPs with genotyping rates ≥ 98% that did not violate Hardy Weinberg equilibrium (p < 10⁻⁶). We used KING as implemented in Plink 1.9 to verify that the patients were not first‐degree relatives, and the patient genotyping rates were ≥ 95%. Imputation was not performed for this data. These standard QC measures were performed with PLINK 1.9 and PLINK 2.0 (www.cog‐genomics.org/plink/) and bcftools software [28, 29].

2.5. SNP Selection for the PRS Analysis

We evaluated a PRS comprised of 11 independent and well‐established SNPs associated with susceptibility to hepatic fat and MASLD, and the test SNP (A1) was selected to be consistent with the PRS analysis reported by Wang et al. [26]. The 11‐SNP PRS included those from our previous 4‐SNP PRS (rs738409 in PNPLA3, rs58542926 in TM6SF2, rs1260326 in GCKR, and rs641738 in MBOAT7) and 7 other well‐established MASLD SNPs (rs13118664 in HSD17B13, rs4808199 in GATAD2A, rs2954021 in TRIB2, rs4240624 in PPP1R3B, rs10883437 in CPN1, rs10883451 in ERLIN1, and rs429358 in APOE) [24]. We derived a weighted version of the PRS using internally derived weights, as determined by the beta coefficients from analyses of each SNP's association with the outcome within the study population. Specifically, we counted the number of test alleles (0, 1, or 2) for each of the 11 SNPs, multiplied the allele counts for each SNP by the internal derived weights as described above, and finally summed these products to form a weighted PRS. We modeled the weighted PRS as a continuous variable, as well as categorized into tertiles.

2.6. Statistical Methods

We summarized the study participants' characteristics using descriptive statistics. Fine‐Gray competing risks regression models, using diagnosis date as the origin time and HCC as the primary outcome, were used to evaluate the association between the PRS and the risk of HCC. Liver transplantation and death were treated as competing events in the models, and patients who were lost to follow‐up were censored after their last known contact. We reported hazard ratios (HRs) with corresponding 95% confidence intervals (CIs). Stratified analyses were performed by etiology of cirrhosis, self‐reported race/ethnicity, and sex to explore subgroup‐specific associations.

To address potential population stratification, we performed a supervised principal components analysis and included the top three principal components (PCs) in the models as covariates. We combined the cohort data with a reference panel, specifically the 1000 Genomes‐Phase 3 data, and pruned the overlapping SNPs using linkage disequilibrium to obtain a smaller, ancestry‐informative dataset [30]. Using this combined dataset, we performed principal components analysis as implemented in Plink 1.9 to produce ancestry‐informative eigenvectors (PCs). The top three PCs were included as covariates in the models to address population stratification [31].

We evaluated the discriminatory ability of the PRS using the concordance index (C‐statistic/index). To assess the added ability of the PRS, we compared the C‐index for the PRS alone, the clinical risk model alone, and the clinical model and PRS combined. Model discrimination was assessed using the 3‐year competing‐risk C‐index with death or liver transplant treated as a competing event. An optimism‐corrected C‐index was estimated using 1000 bootstrap resamples. The clinical model was our previously derived Texas HCC Risk Index (THCC‐RI), which we applied to our study population by using the previously determined parameter estimates (i.e., model coefficients) [9]. The THCC‐RI model includes terms for age, sex, smoking history, alcohol history, BMI, etiology, alpha‐fetoprotein (AFP), albumin, alanine aminotransferase (ALT), and platelet levels.

To assess internal stability of the models and address potential overfitting, we performed nonparametric bootstrap resampling with 1000 resamples. In each resample, participants were sampled with replacement from the analytic cohort, and the Fine–Gray competing‐risk models were refitted for the overall cohort. Bootstrap standard errors and percentile‐based 95% confidence intervals were calculated from the distribution of estimates. To assess the clinical utility of the PRS, we calculated the continuous net reclassification improvement (NRI) at 3 years using inverse probability of censoring weighting. Predicted HCC risks were estimated from Fine–Gray models with and without the genetic risk variable, treating death and liver transplantation as competing events. Continuous NRI was then calculated using inverse probability of censoring weighting, with bootstrap resampling used to estimate 95% confidence intervals. Positive values indicated improved reclassification. Statistical analyses were conducted with SAS 9.04 (SAS Institute Inc., Cary, North Carolina), StataMP 18 (StataCorp LLC, College Station, Texas), and PLINK 1.9 and 2.0 software [28].

3. Results

Among 1888 patients with cirrhosis, 116 developed incident HCC during 6618.65 person‐years of follow‐up (annual HCC incidence rate: 1.75%, 95% CI (1.45%, 2.10%)). Selected characteristics of the cohort are shown in Table 1. The mean age was 59.6 years (standard deviation, 10.0 years), 64.6% of patients were male, 50.4% NHW, 28.6% Hispanic, and 18.5% NHB. The etiology for cirrhosis was MetALD in 39.5%, MASLD for 33.1%, and active HCV for 13.4% of patients.

TABLE 1.

Descriptive statistics of the cohort.

Characteristics	Overall (N = 1888)
Age in years, mean (std.)	59.6 (10.0)
Age in years, n (%)
< 55	480 (25.4)
55–< 65	780 (41.3)
65+	628 (33.3)
Gender, n (%)
Female	669 (35.4)
Male	1219 (64.6)
Race and ethnicity, n (%)
NH‐white	951 (50.4)
NH‐black	349 (18.5)
Hispanic	539 (28.6)
Others	49 (2.6)
Smoking, n (%)
Never	767 (40.6)
Former	724 (38.4)
Current	389 (20.6)
Missing	8 (0.4)
Alcohol drinking, n (%)
Never	580 (30.7)
Past, no heavy	373 (19.8)
Past, heavy	634 (33.6)
Current, no heavy	123 (6.5)
Current, heavy	118 (6.3)
Declined	52 (2.8)
Missing	8 (0.4)
Hypertension, n (%)
No	927 (49.1)
Yes	961 (50.9)
Dyslipidemia, n (%)
No	1241 (65.7)
Yes	647 (34.3)
Diabetes, n (%)
No	1065 (56.4)
Yes	823 (43.6)
Body mass index (kg/m2) (BMI), n (%)
< 30	925 (49.0)
≥ 30	958 (50.7)
Missing	5 (0.3)
Etiology, n (%)
HCV active	252 (13.4)
MASLD only	625 (33.1)
ALD only	42 (2.2)
MetALD	745 (39.5)
Others a	224 (11.9)
Lab values b
a‐Fetoprotein	5.93 (9.6)
Albumin	3.77 (0.6)
Alanine aminotransferase	39.98 (37.2)
Platelet levels (103/μL)b	137.07 (70.9)

^a HBV, hereditary hemochromatosis, Wilson's disease, alpha‐1 antitrypsin deficiency, autoimmune hepatitis, or unspecified ALD are to be defined.

^b Lab values are in the format: mean (standard deviation).

The weighted PRS was comprised of 11 risk SNPs that were previously shown to be associated with hepatic fat (Table 2). In the overall cohort, only one of the SNPs, rs58542926 (TM6SF2) was significantly associated with HCC risk (p = 0.007). However, among NHW cirrhosis patients, only rs13118664 (HSD17B13) was associated with HCC risk (HR = 1.67, 95% CI (1.09, 2.55)), and this association was attenuated in the overall, multiethnic cohort (HR = 1.44, 95% CI (1.00, 2.07)). This SNP was not significantly associated with HCC risk in NHB or Hispanic patients, nor were any of the other SNPs in the PRS.

TABLE 2.

Association between 11 SNPs in PRS and HCC risk, adjusted for first 3 PCs.

SNP (gene)	Location (GRCh37) Chr: position (A1)	HR (95% CI)	p	HR (95% CI) NHW	HR (95% CI) NHB	HR (95% CI) hispanic	AF cases	AF controls
rs1260326 (GCKR)	2: 27730940 (T)	0.82 (0.63–1.07)	0.14	0.78 (0.57–1.09)	0.95 (0.42–2.15)	0.80 (0.46–1.40)	0.43	0.36
rs13118664 (HSD17B13)	4: 88239609 (A)	1.44 (1.00–2.07)	0.05	1.67 (1.09–2.55)	0.95 (0.21–4.25)	0.98 (0.36–2.68)	0.79	0.85
rs4240624 (PPP1R3B)	8: 9184231 (G)	0.97 (0.71–1.34)	0.86	0.95 (0.64–1.41)	1.09 (0.44–2.69)	0.98 (0.52–1.83)	0.19	0.18
rs2954021 (TRIB2)	8: 126482077 (A)	0.99 (0.77–1.28)	0.96	1.13 (0.80–1.58)	0.82 (0.39–1.74)	0.81 (0.50–1.29)	0.52	0.51
rs10883437 (CPN1)	10: 101795361 (T)	1.01 (0.79–1.28)	0.95	0.98 (0.70–1.37)	0.79 (0.42–1.47)	1.13 (0.73–1.75)	0.56	0.56
rs10883451 (ERLIN1)	10: 101924418 (T)	0.93 (0.70–1.24)	0.61	0.95 (0.67–1.35)	2.00 (0.87–4.60)	0.64 (0.34–1.21)	0.63	0.62
rs58542926 (TM6SF2)	19: 19379549 (T)	1.67 (1.15–2.44)	0.007	1.52 (0.91–2.53)	1.24 (0.33–4.68)	1.98 (0.97–4.05)	0.12	0.07
rs4808199 (GATAD2A)	19: 19545099 (A)	1.15 (0.84–1.59)	0.38	1.32 (0.89–1.97)	0.88 (0.29–2.61)	0.85 (0.46–1.58)	0.26	0.23
rs429358 (APOE)	19: 45411941 (T)	0.74 (0.50–1.10)	0.14	0.57 (0.33–1.00)	0.90 (0.41–1.96)	0.98 (0.45–2.14)	0.89	0.85
rs641738 (MBOAT7)	19: 54676763 (T)	0.85 (0.65–1.11)	0.22	0.77 (0.54–1.10)	0.39 (0.16–0.93)	1.20 (0.74–1.92)	0.35	0.39
rs738409 (PNPLA3)	22: 44324727 (G)	1.19 (0.91–1.55)	0.21	1.22 (0.86–1.74)	1.79 (0.89–3.58)	1.15 (0.67–1.97)	0.45	0.39

Abbreviations: A1, test allele; AF, allele frequency; Chr, chromosome; HR, hazard ratio; NHB, non‐Hispanic Black; NHW, non‐Hispanic White; PCs, principal components.

The weighted PRS was significantly associated with HCC risk (Table 3). Specifically, a one‐unit increase in the weighted PRS was associated with over 2‐fold increased HCC risk (HR = 2.20; 95% CI, 1.53, 3.18). When modeled in tertiles, we found a dose–response relationship with the weighted PRS. Compared to cirrhosis patients in the lowest PRS tertile, those in the middle and highest tertiles had 1.79‐fold (95% CI, 1.03, 3.11) and 2.29‐fold (95% CI, 1.33, 3.94) higher risk of HCC, respectively. To address the potential for overfitting, we included the effect sizes and 95% confidence intervals from the nonparametric bootstrap resampling for the overall model with the PRS (continuous and tertiles) in Table 3. We found that the results from this internal validation analysis were largely consistent with those from the model‐based results.

TABLE 3.

Subgroup analysis of weighted PRS by etiology, race/ethnicity, sex and first 3 PCs.

HR (95% CI) by etiology	Overall (n = 1888)	Active HCV (n = 252)	MASLD only (n = 625)	MetALD (n = 745)	ALD only (n = 42)	Others a (n = 224)
Weighted PRS Continuous Bootstrap b	2.20 (1.53, 3.18) 2.20 (1.53, 3.22) b	3.81 (1.64, 8.82)	2.58 (1.55, 4.31)	1.98 (0.96, 4.07)	0.11 (0.01, 1.06)	0.81 (0.20, 3.33)
Low tertile	Ref	Ref	Ref	Ref	Ref	Ref
Middle tertile Bootstrap b	1.79 (1.03, 3.11) 1.79 (1.04, 3.26) b	4.18 (1.06, 16.47)	3.02 (1.05, 8.66)	1.06 (0.44, 2.59)	2.07 (0.19, 23.04)	1.00 (0.20, 5.07)
High tertile Bootstrap b	2.29 (1.33, 3.94) 2.29 (1.37, 4.17) b	3.95 (0.87, 18.05)	3.98 (1.48, 10.69)	1.63 (0.69, 3.86)	N/A	1.28 (0.26, 6.40)

HR (95% CI) by race/ethnicity	Overall (n = 1888)	NH‐white (n = 951)	NH‐black (n = 349)	Hispanic (n = 539)	Others (n = 49)
Weighted PRS continuous	2.20 (1.53, 3.18)	2.53 (1.65, 3.89)	1.99 (0.70, 5.63)	1.54 (0.61, 3.86)	2.08 (0.02, 183.54)
Low tertile	Ref	Ref	Ref	Ref	Ref
Middle tertile	1.79 (1.03, 3.11)	2.59 (1.21, 5.55)	3.15 (1.09, 9.13)	0.58 (0.20, 1.70)	N/A
High tertile	2.29 (1.33, 3.94)	3.26 (1.55, 6.87)	0.98 (0.11, 8.57)	1.16 (0.44, 3.02)	0.29 (0.00, 17.67)

HR (95% CI) by sex	Overall (n = 1888)	Female (n = 669)	Male (n = 1219)
Weighted PRS continuous	2.20 (1.53, 3.18)	2.21 (0.98, 4.97)	2.08 (1.40, 3.10)
Low tertile	Ref	Ref	Ref
Middle tertile	1.79 (1.03, 3.11)	3.09 (0.73, 13.07)	1.58 (0.85, 2.92)
High tertile	2.29 (1.33, 3.94)	3.34 (0.79, 14.11)	2.05 (1.14, 3.68)

Note: Tertile cutoff is based on all patients. Model‐based 95% CI and p‐value are from the original full‐sample model.

Abbreviations: CI, confidence interval; HR, hazard ratio; NHB, non‐Hispanic Black; NHW, non‐Hispanic White; PCs, principal components.

^a HBV, Hereditary hemochromatosis, Wilson's disease, alpha‐1 antitrypsin deficiency, autoimmune hepatitis, or unspecified ALD.

^b Bootstrap 95% CI: 2.5–97.5th percentiles of 1000 bootstrap replicates (unrestricted sampling with replacement).

3.1. Subgroup Analyses

We evaluated the association of the PRS with HCC risk in subgroups stratified by cirrhosis etiology, race/ethnicity groups, and sex (Table 3). The weighted 11‐SNP PRS (continuous) model was significantly associated with increased risk of HCC in cirrhosis patients with active HCV (HR = 3.81, 95% CI (1.64, 8.82)) and MASLD only (HR = 2.58, 95% CI (1.55, 4.31)). Within race/ethnicity groups, the weighted PRS was significantly associated with increased risk of HCC among NHW patients (HR = 2.53; 95% CI (1.65, 3.89)) but not among NHB or Hispanic patients. The weighted PRS was statistically significantly associated with 2‐fold increased risk of HCC only among males (HR = 2.08, 95% CI (1.40, 3.10)).

3.2. Model Discrimination

The C‐statistic for the weighted PRS model alone was 0.62 (95% CI 0.55, 0.68), while the C‐statistic was 0.69 (95% CI 0.63, 0.75) for the THCC‐RI clinical risk factor model alone (Table 4) [9]. Considering these models together, the combined PRS + clinical model had better discriminatory ability than the clinical model alone (C‐statistic 0.73, 95% (0.67, 0.79)). All three models had better discriminatory ability among NHW patients compared with all patients. The C‐statistic for the THCC‐RI clinical model among NHW patients was 0.76 (95% CI (0.70, 0.83)) compared to 0.69 in the overall cohort. The C‐statistic for the PRS only model among NHW was 0.64 (95% CI (0.56, 0.73)) compared to 0.62 in the overall cohort. For the combined PRS + clinical model among NHW, the C‐statistic was 0.78 (95% CI (0.71, 0.86)) compared to 0.73 in the overall cohort. We were unable to examine this in the other racial/ethnic subgroups due to small numbers of events.

TABLE 4.

Performance of predictive models using weighted PRS.

	Model: clinical only a C‐index (95% CI)	Model: PRS C‐index (95% CI)	Model: PRS + clinical a C‐index (95% CI)	Model: PRS + clinical a optimism‐corrected C‐index (95% CI)
Overall
Weighted PRS continuous (0, 1, 2)	0.69 (0.63, 0.75)	0.62 (0.55, 0.68)	0.73 (0.67, 0.79)	0.71 (0.66, 0.77)
Weighted PRS tertiles	0.69 (0.63, 0.75)	0.61 (0.55, 0.66)	0.72 (0.67, 0.78)	0.71 (0.66, 0.77)
non‐Hispanic White
Weighted PRS continuous (0, 1, 2)	0.76 (0.70, 0.83)	0.64 (0.56, 0.73)	0.78 (0.71, 0.86)	0.77 (0.70, 0.83)
Weighted PRS tertiles	0.76 (0.70, 0.83)	0.61 (0.54, 0.69)	0.78 (0.70, 0.85)	0.77 (0.70, 0.84)

^a THCC‐RI model (age, gender, smoking history, alcohol history, BMI, etiology, alpha‐fetoprotein, albumin, alanine aminotransferase, and platelet levels).

Adding the genetic risk score improved 3‐year HCC risk reclassification. The continuous PRS yielded an NRI of 0.265 (95% CI (0.007, 0.515); p = 0.041), and it was likely driven by improved classification of non‐HCC participants (NRI− = 0.249; p < 0.001); there was little improvement among HCC cases (NRI+ = 0.016; p = 0.897) (Table 5). The tertile risk score produced a similar total NRI (0.265; 95% CI (0.047, 0.461); p = 0.012), reflecting better classification of HCC cases (NRI+ = 0.580; p < 0.001) but worse classification of non‐HCC participants (NRI− = −0.315; p < 0.001).

TABLE 5.

Censoring‐adjusted continuous NRI for 3‐year HCC risk prediction.

Model	Component	Estimate (95% CI)	Z	p
Weighted PRS continuous (0, 1, 2)
Total NRI	Overall reclassification	0.265 (0.007, 0.515)	2.05	0.041
NRI+ (events)	HCC cases reclassified upwards	0.016 (−0.226, 0.267)	0.13	0.897
NRI− (non‐events)	Non‐HCC reclassified downwards	0.249 (0.200, 0.299)	9.82	< 0.001
Weighted PRS tertiles
Total NRI	Overall reclassification	0.265 (0.047, 0.461)	2.52	0.012
NRI+ (events)	HCC cases reclassified upwards	0.580 (0.374, 0.768)	5.77	< 0.001
NRI− (non‐events)	Non‐HCC reclassified downwards	−0.315 (−0.368, −0.265)	−11.95	< 0.001

Abbreviations: CI, confidence interval; HCC, hepatocellular carcinoma; NRI, net reclassification improvement.

4. Discussion

Using genetic variants associated with hepatic fat, we have evaluated the association of an 11‐SNP PRS for HCC in a diverse cohort of cirrhosis patients. Compared to those in the lowest tertile of the PRS, cirrhosis patients in the highest tertile (i.e., having the highest genetic risk) had over 2‐fold increased risk for HCC. Furthermore, the PRS was strongly associated with HCC risk among patients with cirrhosis etiology of active HCV and moderately associated with MASLD and MetALD, compared to ALD and other etiologies. The PRS was also only statistically significantly associated with HCC risk among NHW patients, and not among NHB and Hispanic patients, which is consistent with one of the most important criticisms of PRS‐ that they have limited generalizability to non‐European ancestry populations.

Although we also demonstrated that the 11‐SNP PRS performed better for HCC risk discrimination than our previous 4‐SNP PRS, the improvement in predictive ability was modest [24]. Our previous study of HCC risk prediction using PRS employed a 4‐SNP PRS (PNPLA3, TM6SF2, GCKR, and MBOAT7) with a C‐statistic of 0.58 (95% CI (0.52, 0.53)) [24]. In the present study, we have included the 4 hepatic fat SNPs from the original PRS, as well as 7 additional SNPs identified for hepatic fat in previous studies [26]. These additional 7 SNPs improved the robustness of the PRS to identify cirrhosis patients at risk of developing HCC, and the resulting 11‐SNP PRS only model had a C‐statistic of 0.61 (95% CI (0.55, 0.67)) in the same population. Furthermore, we have shown that the inclusion of a PRS of genetic variants for hepatic fat provides additional predictive ability over that provided by our clinical risk stratification model (THCC‐RI model) alone [9]. Specifically, the clinical THCC‐RI model's C‐statistic was 0.71 (95% CI (0.66, 0.76)) and it increased to 0.73 (95% CI (0.68, 0.78)) by adding PRS, which shows modest improvement in discrimination ability.

Objectively, the increase in C‐statistic from the clinical only model to the PRS + clinical model reflects the added value of a targeted genetic model to clinical risk stratification. The PRS model was added to an established clinical model that includes key metabolic risk factors, including BMI and alcohol history; therefore we were able to account for key potential confounding factors. The findings from the net reclassification improvement analysis demonstrated that adding the genetic PRS to the clinical model improved the 3‐year risk classification. This is consistent with previous studies that reported that the inclusion of genetic risk factors into clinical stratification models improved the discriminatory ability of the models for HCC [32, 33]. However, the modest improvements in C‐statistic and the stronger association of the PRS among subgroups of our study cohort demonstrate that the PRS was not robust over different patient groups; consequently, the PRS would have limited utility as a risk stratification tool in its present configuration and should be interpreted with caution.

Instead, the study findings support the need to include ancestry‐specific variants and weights in PRS models, as well as the need to conduct discovery GWAS in heterogenous populations to identify these variants. The SNPs selected for PRS are established as HCC risk variants from literature; however, we have found that many of these SNPs have ancestry‐specific effects (Table 2). These variants were largely identified from GWAS of HCC conducted in European and Asian populations, and they may not include some causal variants in Hispanic and NHB populations, thereby limiting the generalizability of these findings and risk stratification models to different populations. Furthermore, the risk alleles for well‐established HCC SNPs vary in frequency by population, which affects the relevance of the PRS across groups [34, 35]. The weighted PRS can be improved additionally by incorporating ancestry‐specific weights to adjust for the variability in SNP effect sizes by ancestry group. These crucial modifications can aid in developing more accurate PRS for risk assessment in the clinic, particularly for underserved populations.

One of the main limitations of this study is the possible risk of overfitting the PRS + clinical model, as the THCC‐RI model was developed in the same population. Without a validation set, it is difficult to assess the accuracy of the findings; however, the lack of a comparable validation set may be due to the limited number of studies of this population. To address the potential for overfitting in evaluating PRS performance, we conducted internal validation using nonparametric bootstrap resampling. The results of this analysis were generally consistent with the original estimates, supporting the stability of our findings. However, external validation in independent and diverse cirrhosis cohorts is needed and these findings must be evaluated with caution at present. As there remains concern about overestimating the discrimination ability of the PRS due to overfitting, our future directions include establishing an independent cohort to test the validity of these results. In addition, an independent cohort would allow us to re‐estimate the coefficients that were established for the THCC‐RI model. Furthermore, our future studies would include conducting discovery GWAS in multi‐ethnic populations. We intend to use the results of these studies to develop and adopt genetic ancestry‐specific estimates for PRS to improve their clinical utility.

Similarly, overfitting the model due to population stratification in the model is another limitation of this study. Although we adjusted for the first three PCs from a supervised principal components analysis, population stratification may persist and influence the validity of the effect estimates, potentially producing spurious associations and missing true genetic effects. Other limitations of this study are related to sample size and power. The analysis may have limited power to detect modest genetic associations, particularly in a multiethnic cohort with rich genetic diversity and risk SNPs that may vary by ancestry. Furthermore, the subgroup analyses produced several wide confidence intervals, emphasizing the uncertainty of the findings and these results should be considered with caution. These limitations emphasize the need for additional cirrhosis and HCC study cohorts from diverse patient populations to obtain more valid effect estimates for PRS that can be used to improve patient outcomes.

Although the current study includes germline genetic variants, PRS can be expanded to include transcriptomic, proteomic, and other risk variants for HCC as discovery expands via other “omics” studies [36, 37]. A recent study combined polygenic risk scores and polygenic transcriptome risk scores (PTRS) to create a summary polygenic risk score that had improved predictive accuracy compared to the separate polygenic scores [38]. The study also reported that PTRS improve generalizability of risk scores compared to PRS when evaluated in multi‐ancestry cohorts.

In conclusion, our findings show that the PRS was strongly associated with HCC risk in patients with cirrhosis. However, the magnitude of association was not consistent across subgroups of patients and, in its current form, offered minimal discriminatory ability over prior clinical risk stratification models for HCC. Future studies are needed to refine the PRS, including using ancestry‐specific weights, before it can be considered for HCC risk stratification in patients with cirrhosis.

Author Contributions

Nia Choi: conceptualization, methodology, formal analysis, writing – original draft, writing – review and editing. Ashwin Rao: data curation. Priya B. Shetty: conceptualization, formal analysis, writing – original draft, writing – review and editing, methodology. Saira Khaderi: data curation. Hao T. Duong: formal analysis. Christopher I. Amos: formal analysis. Michelle Luster: data curation. Jinyoung Byun: formal analysis. Amit G. Singal: data curation. Abeer Al‐Sarraj: data curation. Hashem B. El‐Serag: conceptualization, methodology, data curation, formal analysis, writing – original draft, writing – review and editing. Aaron P. Thrift: conceptualization, methodology, formal analysis, writing – original draft, writing – review and editing. Fasiha Kanwal: data curation. Sumeet K. Asrani: data curation.

Funding

This work was supported by Cancer Prevention and Research Institute of Texas, RP150587, RP220119, RP190641, RP200537, RP250355, RP200554. National Cancer Institute, P01 CA263025, P01 CA263025‐03S1, U01 CA230997, R01 CA186566, R01 CA274528, R03 CA262911, P30 CA125123, R01 CA256977, U01 CA271887. National Institute of Diabetes and Digestive and Kidney Diseases, P30 DK 56338. Center for Innovations in Quality, Effectiveness and Safety, CIN 13‐413. National Institute of Environmental Health Sciences, P30ES030285.

Ethics Statement

This study involves human participants, and this research was approved by the Baylor College of Medicine (BCM) IRB under ID number H‐40241. The BCM IRB is organized, operates, and is registered with the US Office for Human Research Protections according to the regulations codified in the United States Code of Federal Regulations at 45 CFR 46 and 21 CFR 56. The BCM IRB operates under the BCM Federal Wide Assurance No. 00000286, as well as those of hospitals and institutions affiliated with the College. Participants gave informed consent to participate in the study before taking part. All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards.

Conflicts of Interest

Dr. Singal consults for Bayer, Fujifilm Medical Sciences, Roche, Abbott, Exact Sciences, Helio Genomics, Glycotest, DELFI, Curve BioSciences, and IMCare. All other authors declare no conflicts of interest.

Data Availability Statement

The data from this study are available for IRB‐approved projects on request from the corresponding authors.