Unravelling genetic susceptibility and causal factors in liver health using MRI quantification of inflammation, fat and iron in the liver

Abstract

Background

Liver steatosis, fibroinflammation, and iron overload, are growing global health concerns, yet the genetic architecture and causal pathways linking liver pathology to systemic disease remain incompletely understood.

Methods

We analysed MRI-derived liver traits—corrected T1 (cT1), proton density fat fraction (PDFF), and liver iron—in 37,626 UK Biobank participants. Genome-wide (GWAS), transcriptome-wide (TWAS), and cis-protein Mendelian randomisation (cis-MR) analyses were used to identify genes and proteins influencing these traits. We applied two-sample MR to assess bidirectional causal relationships with metabolic and vascular traits and used Multi-Trait Analysis of GWAS (MTAG) to enhance discovery by leveraging genetic correlations.

Results

GWAS identified 18 loci for cT1, 15 for PDFF, and 5 for liver iron, including six not previously reported. TWAS, cis-MR, and proteome-wide analyses prioritised genes (e.g., FADS1, GPAM, MBOAT7, RAD51C) and proteins (e.g., RAB2B, GPN1, GSTM4) with putative mechanistic roles. Fine-mapping refined several signals (GSTM1, TMPRSS6) to single-variant credible sets. Cell-type enrichment revealed distinct tissue contributions: hepatocytes and intestinal mucosa for cT1, adipose tissue for PDFF, and gastrointestinal tissues for liver iron. MR suggested causal effects of higher liver PDFF and cT1 on obesity-related traits, and inverse genetic associations between liver iron and coronary artery disease. MTAG identified seven additional loci (three for cT1, four for PDFF) not previously reported.

Conclusions

This integrative imaging-genetics study reveals 13 potentially novel genes and several protein candidates implicated in hepatic steatosis, inflammation, and iron homeostasis. These findings enhance understanding of liver disease biology and may help identify new targets for early detection or treatment.

Impact and implication

This large imaging-genetics study in over 37,000 people identifies genetic and protein factors linked to liver fat, fibroinflammation, and iron levels. It shows that higher liver fat and inflammation are associated with increased cardiometabolic risk, while higher liver iron appears inversely linked to risk of heart disease. These findings highlight molecular targets such as FADS1, FUT2, TMC4, RAB2B, and GPN1, which could inform future efforts to improve early detection or treatment of liver disease and its complications in people with obesity or metabolic syndrome.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40246-026-00913-2.

Introduction

Chronic liver diseases represent a growing global health concern, encompassing a spectrum of conditions such as metabolic dysfunction-associated steatotic liver disease (MASLD, formerly NAFLD) [1, 2], metabolic associated steatohepatitis (MASH), and iron overload syndromes like hereditary haemochromatosis. These diseases affect hundreds of millions of individuals worldwide, with MASLD alone estimated to affect over 25% of the global adult population. If left unrecognised or untreated, these conditions can silently progress to advanced liver fibrosis, cirrhosis, and hepatocellular carcinoma [3]. The rising prevalence of obesity, sedentary lifestyles, and cardiometabolic conditions underscores the urgency of improving early detection and mechanistic understanding of liver pathology [4, 5].

For clinicians and researchers alike, accurate assessment of liver fat accumulation, fibroinflammation, and iron deposition is essential for risk stratification, treatment decisions, and advancing pathophysiological insight [6]. However, traditional diagnostic tools have limitations: liver biopsy, while considered the gold standard, is invasive and unsuitable for large-scale use; circulating liver enzymes are insensitive and nonspecific [2, 7]. Magnetic resonance imaging (MRI) now provides robust, non-invasive quantification of liver tissue properties. Proton density fat fraction (PDFF) measures hepatic fat content; corrected T1 (cT1) reflects fibroinflammatory activity via extracellular fluid; and R2*-based MRI enables quantification of liver iron [8–11]. Together, these imaging biomarkers allow detailed characterisation of liver health at scale.

Parallel to advances in imaging, genome-wide association studies (GWAS) have identified genetic variants associated with individual liver traits [12]. Key discoveries include PNPLA3 and TM6SF2 for PDFF and HFE and TMPRSS6 for iron overload [13, 14]. Yet most previous studies have examined these traits in isolation, often relying on smaller sample sizes, or limited to single-omic or univariate approaches. Recent studies using UK Biobank MRI data have made important contributions, but opportunities remain to comprehensively dissect the shared and distinct genetic architecture across liver PDFF, cT1, and iron [13, 15–17]. Moreover, causal inference methods such as Mendelian randomisation (MR) have often been applied separately and in limited scope, without integrating post-GWAS functional analyses across transcriptomic and proteomic layers [18–20]. In particular, multivariate methods such as multi-trait analysis of GWAS (MTAG), and integrative tools such as transcriptome- and proteome-wide association studies, can improve power for discovery and enhance biological interpretation.

Here, we report an integrated analysis of MRI-derived liver traits in 37,626 participants from the UK Biobank. We applied GWAS, MTAG, Bayesian fine-mapping, and functional annotation to map genetic influences on liver cT1, PDFF, and liver iron. To prioritise biologically relevant genes and proteins, we incorporated transcriptomic, proteomic, and cell-type enrichment analyses, and used MR to assess potential causal effects on 17 cardiometabolic and vascular outcomes. This approach enables a detailed exploration of the genetic architecture and systemic consequences of common liver traits and highlights putative molecular targets relevant to early detection and prevention of liver-related complications.

Methods

UK biobank participants

This study was conducted using data from the UK Biobank (application number 9914), a prospective cohort of over 500,000 individuals aged 37–73 years at recruitment (99.5% aged 40–69 years), between 2006 and 2010 [21]. All participants provided written informed consent and ethical approval was granted by the North West Multi-Centre Research Ethics Committee (ref: 11/NW/0382).

Liver imaging acquisition and processing

We utilised quantitative abdominal MRI to assess three clinically relevant liver traits in UK Biobank participants: proton density fat fraction (PDFF, %) for steatosis, hepatic iron concentration (mg/g) and corrected T1 (cT1, ms) as a marker of fibroinflammatory activity. Imaging was conducted at the UK Biobank Imaging Centre in Cheadle, UK, using a Siemens 1.5 T Magnetom Aera scanner, as previously described [13]. Participants were selected based on proximity to imaging sites. PDFF and iron were estimated using a multi-echo spoiled gradient-echo (mGRE) sequence, while T1 times were obtained via a Shortened Modified Look-Locker Inversion Recovery (ShMOLLI) sequence. Both sequences captured a single axial slice at the level of the porta hepatis, acquired during a breath-hold without contrast. This slice-based method has shown strong concordance with histological measures of liver fibrosis, fat, and iron [11]. The slice at the level of the porta hepatis was selected because it provides a large, homogeneous cross-section of liver parenchyma while avoiding major vessels and has been used consistently in prior validation studies [70]. Previous work has demonstrated strong agreement between single-slice cT1 measurements at this level with histology and predict liver-related outcomes [10, 11]. Potential systematic variability related to body types and respiratory variation is mitigated by standardized breath-hold acquisition, fixed anatomical landmarking, and the use of normalized quantitative MRI metrics rather than relative signal intensities. Residual variability is expected to contribute random noise rather than systematic bias at the population level. The deep learning segmentation approach was previously published in Irving et al. 2017 [22].

Images were processed using LiverMultiScan® Discover 4.0 (Perspectum Ltd, Oxford, UK), which generated parametric maps of PDFF, T2*, and T1. Liver segmentation was performed using the LiverMultiScan (Perspectum Ltd, Oxford, UK) software. The software generates quantitative T2 ∗ , cT1 and PDFF maps and applies an automated liver delineation using a previously published U-net deep-learning model [22], which excludes the major hepatic vessels. The median T2 ∗ value within the segmentation is converted to liver iron concentration and combined with the ShMOLLI T1 data to compute the corrected T1 (cT1) map [11]. R2* values were converted to hepatic iron concentrations, and T1 values were corrected for iron to derive cT1. PDFF was computed as the fat signal fraction relative to the total liver signal. We applied rigorous quality control to exclude scans with artefacts, flawed protocols, mispositioned slices, segmentation failures, or features (e.g., cysts or lesions) that impaired parenchymal quantification. After filtering, 37,651 scans passed quality control. Among these, 37,626 had valid PDFF data, 37,294 had liver iron measurements, and 31,540 had cT1 estimates. All participants included in the final analysis had genotype data and were classified as White European ancestry based on principal component analysis.

Genetic association analysis

Detailed protocols for participant genotyping, data collection, and quality control have been previously described [23]. Rigorous quality control procedures were applied to exclude individuals with high genotype missingness, excess heterozygosity, or discrepancies between reported and genetically inferred sex. Participants with a missingness rate > 5% across variants that passed quality control were also excluded. Genotype imputation was performed using two reference panels: the Haplotype Reference Consortium and a combined UK10K/1000 Genomes panel [24, 25]. Only variants with an imputation quality score > 0.3 and a minor allele frequency (MAF) > 1% were retained for analysis.

To identify individuals of White European ancestry, principal components (PCs) derived from the 1000 Genomes reference panel were projected onto the UK Biobank dataset. K-means clustering was then applied to classify participants into ancestry groups, and only those aligning with the European cluster were included in the genetic analyses.

Following image and genotype quality control, liver cT1, liver PDFF, and liver iron measures were available for 31,540, 37,626, and 37,294, individuals, respectively. Genome-wide association analyses were conducted using BOLT-LMM v2.3.4 [26], which models both polygenic and relatedness structures through a linear mixed model approach. The genetic relationship matrix was derived from common (MAF > 5%) genotyped variants passing quality control on both genotyping platforms. Phenotypic traits were inverse-normal transformed prior to analysis, and models were adjusted for age, sex, imaging centre, genotyping array, and principal components for ancestry. Sensitivity analyses adjusting for body mass index (BMI), as well as sex-stratified GWAS, were conducted to explore the robustness of findings and potential sex-specific effects.

Independent loci were defined using a ± 500 kb window around the lead variant and annotated with the nearest protein-coding gene using Ensembl [27]. Conditional analysis was performed using the COnditional and JOint analysis (COJO) implemented in GCTA to identify additional independent signals within loci [28]. A randomly selected subset of 10,000 unrelated White British individuals served as the linkage disequilibrium (LD) reference panel [21]. Secondary signals were retained if they reached genome-wide significance (P < 5 × 10^-8) after conditioning on the lead variant.

To further prioritise likely causal variants, we performed Bayesian statistical fine-mapping for each genome-wide significant locus leveraging the approximate Bayes’ factor (ABF) method [29]—a single-causal-variant, LD-free fine-mapping method that does not require an LD panel. We chose this approach as it is computationally inexpensive and can be implemented on GWAS summary statistics, without requiring access to individual-level genotype data and a matched reference panel. Future work may incorporate alternative fine-mapping approaches to validate and refine these results; for example, methods such as FINEMAP or SuSiE could be applied to further refine candidate causal variants and assess the presence of multiple causal signals per locus. The analysis was performed using GWAS summary statistics for the selected traits via the easyfinemap pipeline (version 0.4.4; https://jianhua-wang.github.io/easyfinemap/) with default settings for continuous traits without any reference panel. Briefly, we constructed 95% credible sets (CSs) by calculating Bayes factors and posterior probabilities (PP) for all variants within each region. Variants were ranked by descending PP, and cumulative PPs were summed until the 95% threshold was reached. These credible sets highlight variants most likely to be functionally responsible for the observed genetic associations. We performed GCTA-COJO and fine-mapping independently because they address related but distinct questions. GCTA-COJO was used to reveal secondary signals in loci with complex LD, while fine-mapping was performed to construct 95% credible sets to identify the most likely causal variants at each locus, using the original (unconditioned) marginal summary statistics without a LD reference. This comprehensive framework allowed us to robustly identify novel genetic associations for MRI-derived liver phenotypes and provided a high-resolution map of potential causal variants underpinning liver traits.

Genetic correlation

We employed linkage disequilibrium score regression (LDSC) [30] to estimate the genetic correlation between liver imaging-derived phenotypes—proton density fat fraction (PDFF), liver iron, and cT1—and a range of metabolic, cardiovascular, and liver-related traits (supplementary Table 1). LDSC quantifies the shared genetic architecture between two traits by regressing the product of their GWAS Z-scores (Z₁Z₂) against LD scores across the genome. The resulting regression slope provides an estimate of genetic correlation, reflecting the degree of genome-wide pleiotropy between the traits.

We conducted the analysis using the publicly available LDSC software (https://github.com/bulik/ldsc), applying pre-computed LD scores derived from European-ancestry samples in the 1000 Genomes reference panel [30]. The analysis was limited to approximately 1.2 million high-quality HapMap3 SNPs, which are known to be reliably imputed across diverse GWAS datasets [31]. GWAS summary statistics were formatted using the munge_sumstats.py script, and regression models were run using ldsc.py. To control for multiple comparisons, we applied Bonferroni correction, considering results to be statistically significant at a threshold of P < 3.13 × 10⁻³ (0.05/16). This approach enabled a rigorous assessment of the shared genetic basis between liver traits and key clinical outcomes.

Two-sample univariable Mendelian randomization (MR)

To assess potential causal relationships between genetically predicted liver traits (cT1, PDFF, and liver iron) and a range of metabolic and clinical outcomes (supplementary Table 1), we conducted two-sample univariable Mendelian randomisation (MR) analyses [32]. We selected independent single nucleotide polymorphisms (SNPs) strongly associated with each liver trait (P < 5 × 10⁻⁸), excluding variants with minor allele frequency (MAF) ≤ 5% or F-statistics < 10 to minimise weak instrument bias [33]. Instrument strength was assessed using the F‑statistic, calculated as F = β²/SE², where β denotes the effect estimate for each SNP on the exposure and SE its corresponding standard error. For cT1, F-statistics ranged from 30.05 to 1346.75 (average F-statistics = 149.25). For PDFF, F-statistics ranged from 30.16 to 806.35 and a mean average F-statistic of 169.81. For liver iron, F-statistics ranged from 35.09 to 469.02, with an average F-statistic of 137.04. All minimum F-statistics exceeded the conventional threshold of 10, indicating a relatively low risk of weak instrument bias. To ensure independence among SNPs, we applied linkage disequilibrium (LD) clumping (r² < 0.001), retaining the variant with the smallest P-value within each locus using the 1000 genomes European data (version 3) as reference panel [25].

Causal estimates for individual SNPs were derived using the Wald ratio method (SNP-outcome effect divided by SNP-exposure effect). These were then aggregated using an inverse-variance weighted (IVW) random-effects model [34], which served as our primary analytic approach. To test the robustness of our findings and account for potential horizontal pleiotropy, we applied weighted median and MR-Egger regression (35) as sensitivity analyses. In addition, we performed the Mendelian Randomization Pleiotropy RESidual Sum and Outlier (MR-PRESSO) test [36] to mitigate potential bias from horizontal pleiotropy. When MR-PRESSO identified possible outliers, we conducted outlier-correction tests, and reported the outlier-corrected causal estimates. Bonferroni correction was used to adjust for multiple comparisons across five sets of MR tests, with statistical significance set at P < 2.94 × 10⁻³ (0.05/17).

cis-Mendelian randomisation (cis-MR)

We performed cis-MR to identify circulating plasma proteins that may mediate the effects of genetic variation on liver phenotypes. We leveraged protein quantitative trait locus (pQTL) data from the UK Biobank Pharma Proteomics Project (UKB-PPP; 2,031 proteins) [37] and deCODE Genetics (1,803 proteins) [38]. For each protein, we selected cis-acting variants that were within ± 1 megabase (Mb) of the transcription start site, were genome-wide significant (P < 5 × 10⁻⁸) and were not in LD with each other (r² < 0.001). Steiger filtering was applied to ensure directional consistency [39], retaining only instruments that explained more variance in protein levels than in the liver trait. Causal effects were estimated using the Wald ratio method (for < 2 SNPs) or IVW (for ≥ 2 SNPs). To distinguish true causal effects from confounding due to LD, we performed a colocalisation analysis using the COLOC package in R [40]. Statistical significance was defined as an FDR-adjusted P-value < 0.05. Associations with a posterior probability of shared causal variant (PPH4) ≥ 0.70 were considered strong evidence of colocalization while those with PPH4 ≥ 0.50 were considered suggestive evidence of colocalization.

Transcriptome-wide association study and genetic colocalization

We conducted TWAS [41] to identify genes whose cis-regulated expression levels are associated with liver cT1, PDFF, and iron. TWAS was performed using pre-computed expression weights from the GTEx v8 resource [42] and GWAS summary statistics for the liver traits, implemented via the FUSION software framework [41]. Expression-phenotype associations were inferred using a reference panel with paired expression and genotype data.

We focused on ten tissues most relevant to liver biology and metabolism from the GTEx v8 multi-tissue expression datasets, including adipose subcutaneous (N = 581), adipose visceral omentum (N = 469), colon sigmoid (N = 318), kidney cortex (N = 73), liver (N = 208), pancreas (N = 305), small intestine terminal ileum (N = 174), spleen (N = 227), stomach (N = 324), and whole blood (N = 670; http://gusevlab.org/projects/fusion/). Statistical significance was defined as an FDR-adjusted P-value < 0.05. For all nominally significant associations (P < 0.05), we conducted a Bayesian colocalisation analysis to determine whether GWAS and expression quantitative trait locus (eQTL) signals shared a causal variant [40]. We considered a posterior probability for shared causal variant (PPH4) ≥ 0.70 as strong evidence of colocalisation.

Summary data-based Mendelian randomisation (SMR)

To further assess whether genetic associations with liver traits were mediated through gene expression, we employed the SMR method [43], combined with the heterogeneity in dependent instruments (HEIDI) test. SMR was applied using liver tissue eQTL data from GTEx v8 [42], integrating SNPs within 1 Mb of transcription start sites and GWAS summary statistics from our BOLT-LMM analyses. The top eQTL (P < 5 × 10⁻⁸) for each gene was selected as the instrument. HEIDI tests were used to assess the likelihood that associations were due to a shared causal variant rather than LD (P > 0.05 indicating non-heterogeneity). Analyses were restricted to cis-regions by default. Multiple testing was addressed via FDR correction (FDR < 0.05).

Tissue and cell-type enrichment analyses

To identify tissues and cell types most relevant to the genetic regulation of liver traits, we applied stratified LD score regression (stratified LDSC) using the –h2-cts flag within the LDSC framework [44]. This approach quantifies the enrichment of trait heritability in genes highly expressed in specific tissues or cell types. We utilised gene expression datasets from GTEx [45] and the Franke lab, which includes RNA-seq profiles from 205 tissues and cell types [44]. Enrichment was considered significant based on partitioned heritability estimates after correcting for multiple testing (FDR < 0.05).

Multi-trait analysis of GWAS (MTAG)

Given the well-established genetic correlations between hepatic traits and cardiometabolic diseases, we applied MTAG to improve locus discovery for liver phenotypes. MTAG is a generalised meta-analytic framework that leverages the shared genetic architecture across traits to boost power for identifying trait-specific associations, while accounting for sample overlap and incomplete genetic correlation [46]. By incorporating shared genetic signal with cardiometabolic traits/diseases (e.g., T2D/CAD), MTAG can uncover loci that have weak effects on the liver trait alone but are part of a shared biological pathway with the cardiometabolic disease.

MTAG was run with default parameters, treating PDFF and cT1 as primary traits of interest. To ensure robust inference, we applied strict filtering to prioritise pleiotropic signals likely to reflect true shared biology rather than statistical artefacts. Variants were retained only if they met all of the following criteria:

1. Genome-wide significance in MTAG (pMTAG  P < 5 × 10⁻⁸);

2. Suggestive association in the original liver GWAS (P < 5 × 10⁻⁶) with consistent effect direction; and

3. Evidence of association in the corresponding cardiometabolic trait GWAS (e.g., T2D or CAD).

The post-hoc filtering criteria were deliberately chosen to be stringent to prioritize robust and biologically plausible MTAG associations. Requiring (P < 5 × 10⁻⁶) the original GWAS ensured at least moderate single-trait evidence, reducing the risk of spurious MTAG-driven signals, while confirmation of association in the cardiometabolic trait GWAS supported shared biological relevance and minimized false positives. To evaluate potential bias due to violations of the MTAG homogeneity assumption (i.e., variants with null effects on liver traits but strong effects on other traits), we estimated the maximum false discovery rate (MaxFDR), which was 0.03 for PDFF and 0.04 for cT1—well within acceptable limits, indicating robustness of findings.

Significant variants (P < 5 × 10⁻⁸) were grouped into loci based on a ± 1 Mb window and annotated to the nearest protein-coding gene. To avoid artefactual inflation from MTAG-induced correlations, colocalisation analyses were performed using the original single-trait GWAS summary statistics for both the liver trait and cardiometabolic outcome, using the R package coloc as described above. Details of GWAS datasets used are provided in Supplementary Table 1.

Results

Characteristics of the liver MRI cohort

The imaging cohort comprised 37,626 participants (18,007 men and 19,619 women). The median age was 66 years for men (interquartile range [IQR]: 59–71) and 64 years for women (IQR: 58–69). Liver cT1 values showed a median of 701 ms (IQR: 669–708) in men and 685 ms (IQR: 565–720) in women. Elevated cT1 values (> 800 ms), indicative of potential liver fibroinflammatory activity, were observed in 6.4% of men (n = 941) and 4.2% of women (n = 665). Liver PDFF, had a median of 3.64% (IQR: 2.53–6.55) in men and 2.59% (IQR: 1.95–4.33) in women. A substantial proportion of participants (35% of men and 21% of women) exceeded the 5% threshold for hepatic steatosis. Median liver iron concentrations were slightly higher in men (1.24 mg/g, IQR: 1.14–1.39) than in women (1.20 mg/g, IQR: 1.11–1.33). Elevated liver iron levels (defined as > 1.8 mg/g) were present in 4.3% of men and 2.3% of women. These results suggest men have higher median levels meaning increased susceptibility to hepatic inflammation, steatosis, and iron overload.

Liver cT1

Supplementary Fig. 1 provides an overview of the study design and summarizes the main results.

We identified 15 loci associated with cT1, including 12 previously reported (e.g., PNPLA3, TM6SF2, FADS1, HFE) and three newly implicated loci: CXXC4, TFCP2, and TSPAN5 (Table 1, Fig. 1, Supplementary Fig. 2–3). The most striking effect was observed for a variant in SLC39A8 (rs13107325), where individuals carrying the minor allele showed substantially higher cT1 levels (an increase of 0.55 standard deviations (SD); P = 1 × 10⁻³⁰¹). These associations remained consistent after accounting for BMI, and we did not observe differences between men and women (supplementary Table 2).

Table 1.

Loci significantly associated with liver traits

Trait	GWAS model	rsID	CHR	POS	EA	NEA	EAF	BETA	SE	P	N	GENE
cT1	Univariate GWAS	rs759359281	1	220,100,497	CA	C	0.944	− 0.126	0.017	1.10E−13	31,540	SLC30A10
		rs62325476	4	99,662,849	A	G	0.735	− 0.048	0.009	1.80E−08	31,540	TSPAN5
		rs13107325	4	103,188,709	C	T	0.930	− 0.555	0.015	1.00E−301	31,540	SLC39A8
		rs17269772	4	105,480,260	T	C	0.978	− 0.152	0.027	8.10E−09	31,540	CXXC4
		rs79220007	6	26,098,474	T	C	0.929	0.141	0.015	3.00E−20	31,540	HFE
		rs13191659	6	27,001,055	C	G	0.916	0.081	0.014	1.10E−08	31,540	HIST1H2BJ
		rs1495743	8	18,273,300	G	C	0.222	− 0.057	0.009	2.70E−10	31,540	NAT2
		rs532436	9	136,149,830	G	A	0.816	− 0.076	0.010	9.00E−15	31,540	ABO
		rs751229834	11	61,575,807	CA	C	0.690	− 0.047	0.008	2.60E−08	31,540	FADS1,FADS2
		rs59372312	12	51,511,269	A	G	0.933	0.093	0.016	1.70E−09	31,540	TFCP2
		rs111723834	14	24,572,932	G	A	0.984	− 0.370	0.031	7.80E−35	31,540	NRL,PCK2
		rs58542926	19	19,379,549	C	T	0.926	− 0.149	0.015	1.10E−23	31,540	TM6SF2
		rs516246	19	49,206,172	C	T	0.490	− 0.049	0.008	2.40E−10	31,540	FUT2
		rs6000553	22	37,469,192	A	G	0.465	0.071	0.008	9.50E−21	31,540	TMPRSS6
		rs738409	22	44,324,727	C	G	0.786	− 0.112	0.009	1.10E−33	31,540	PNPLA3
	MTAG	rs72662001a	1	39,967,280	T	C	0.776	− 0.060	0.007	7.94E−20	68,293	BMP8A
		rs7012637b	8	9,173,209	G	A	0.528	− 0.044	0.008	1.64E−08	33,765	PPP1R3B
		rs3890483a	19	7,220,596	G	T	0.560	− 0.030	0.006	3.63E−08	68,293	INSR
PDFF	Univariate GWAS	rs140584594	1	110,232,983	A	G	0.268	− 0.046	0.008	6.50E−09	37,651	GSTM1,GSTM2
		rs2642438	1	220,970,028	A	G	0.296	− 0.054	0.008	1.80E−11	37,651	MARC1
		rs1260326	2	27,730,940	T	C	0.394	0.061	0.007	7.30E−19	37,651	GCKR
		rs1229984	4	100,239,319	T	C	0.026	− 0.162	0.022	1.60E−13	37,651	ADH1B
		rs28601761	8	126,500,031	C	G	0.580	0.061	0.007	5.50E−18	37,651	TRIB1
		rs7096937	10	113,950,418	T	C	0.270	0.058	0.008	5.30E−13	37,651	GPAM
		rs56252442	19	18,229,208	G	T	0.748	− 0.046	0.008	1.70E−08	37,651	MAST3
		rs58542926	19	19,379,549	C	T	0.926	− 0.304	0.013	3.30E−116	37,651	TM6SF2
		rs429358	19	45,411,941	T	C	0.849	0.120	0.010	7.10E−36	37,651	APOE
		rs11668882	19	54,675,097	T	C	0.561	− 0.039	0.007	1.10E−08	37,651	TMC4
		rs738408	22	44,324,730	C	T	0.786	− 0.240	0.008	4.50E−179	37,651	PNPLA3
	MTAG	rs13389219a	2	165,528,876	C	T	0.605	0.044	0.005	6.96E−19	87,794	COBLL1
		rs28650790b	5	55,861,464	C	T	0.808	− 0.053	0.009	9.36E−10	44,116	AC022431.2
		rs9687846a	5	55,861,894	G	A	0.799	− 0.052	0.006	4.63E−18	87,794	AC022431.2
		rs741782b	17	18,019,712	T	C	0.695	− 0.045	0.007	1.24E−09	44,116	MYO15A
		rs2315007a	20	62,343,845	T	C	0.336	− 0.029	0.005	1.17E−08	87,794	ZGPAT
Iron	Univariate GWAS	rs4667296	2	190,517,931	G	A	0.228	− 0.052	0.009	2.00E−09	37,294	ASNSD1
		rs1800562	6	26,093,141	G	A	0.927	− 0.305	0.014	7.70E−105	37,294	HFE
		17:56,428,612	17	56,428,612	CAAA	C	0.526	0.047	0.008	1.60E−09	37,294	SUPT4H1
		rs2645490	17	57,878,509	G	A	0.785	0.052	0.009	5.40E−09	37,294	VMP1
		rs855791	22	37,462,936	A	G	0.436	− 0.104	0.007	1.20E−46	37,294	TMPRSS6

rsID reference SNP identifier, CHR chromosome, POS base-pair position (GRCh build 37), EA effect allele, NEA non-effect allele, EAF effect-allele frequency, BETA estimated effect size (per effect allele), SE standard error of the effect estimate, P p-value for association, N sample size, GENE nearest gene, Variants shown in bold represent associations not previously reported for these liver traits

^a,bGenetic variants discovered through multi-trait analysis (MTAG) of liver traits either with type 2 diabetes^a or coronary artery disease^b, ^a Loci came from bivariate MTAG analysis with T2D; ^b Loci came from bivariate MTAG analysis with CAD

Fig. 1.

Manhattan plots of genome-wide association studies (GWAS) for liver traits. A. liver cT1, B. liver PDFF, and C. liver iron. The horizontal red line indicates the genome-wide significance threshold (P < 5 × 10⁻⁸). Nearby genes have been annotated for each lead SNP. Loci highlighted in red represent novel associations not previously reported for these liver traits

We used fine-mapping techniques to narrow down the list of likely causal variants at several loci, including SLC39A8 and TMPRSS6, to a single candidate variant with high confidence (≥ 95% probability; supplementary Table 3). Enrichment analysis suggested that the strongest genetic signals were active in hepatocytes and intestinal tissues (supplementary Fig. 2). TWAS, which links genetic variation to gene expression in specific tissues, identified 58 unique genes, including FUT2, FADS1, ABO, and NTN5—all of which showed strong evidence that the same genetic variant influences both gene expression and cT1 (Fig. 2). These findings were further supported by SMR-HEIDI (Fig. 3). We also explored whether blood protein levels could mediate these genetic effects. Using protein-level cis-MR, we found that higher circulating levels of ABO and ADH1B proteins were likely to contribute causally to increased cT1, with strong statistical support that the same variants affect both the protein and the liver trait (Fig. 4).

Fig. 2.

Transcriptome-wide association studies (TWAS) and colocalisation analyses for liver traits. Heatmaps display tissue-specific associations for: A. liver cT1, B. liver PDFF, and C. liver iron. Only tissue-specific expression-phenotype associations that were FDR significant (P-TWAS < 0.05) and with strong evidence of sharing the same causal variant (PPH4 ≥ 70%) were highlighted in the heatmap. The colour of the cell represents the direction of the TWAS Z-score

Fig. 3.

SMR and HEIDI analyses highlight putative causal genes for liver traits. Heatmaps display significant gene–trait associations identified through SMR and HEIDI tests for: A. liver cT1, B. liver PDFF, and C. liver iron. Only tissue-specific expression-phenotype associations that were FDR-adjusted P_SMR < 0.05 and P_HEIDI ≥ 0.01 were highlighted in the heatmap. The colour of the cell represents the direction of the SMR beta estimate

Fig. 4.

cis-Mendelian randomization (cis-MR) analyses of circulating plasma proteins (from Olink and SomaScan (deCODE) platforms) on liver traits. Top panel: forest plots showing the effect of proteins on PDFF. Bottom panel: Forest plots showing the effect on cT1. Only proteins meeting a 5% FDR threshold and a suggestive Bayesian colocalization posterior probability (H4 > 50%) are displayed. Only proteins with PP.H4 > 0.7 were considered to have strong evidence of colocalization

Two-sample MR analyses revealed that genetically higher alanine transaminase (ALT), BMI, WHR, triglycerides, and genetic liability to type 2 diabetes and NAFLD were associated with higher cT1 levels (Fig. 5). In reverse MR, higher genetically predicted cT1 was associated with lower HDL-cholesterol and higher BMI (Fig. 5). Detailed estimates for all these analyses of liver cT1 are provided in supplementary tables 4–11.

Fig. 5.

Genetic correlation and Mendelian randomisation estimates between liver traits and various cardiometabolic risk factors. A forest plot of genetic correlation (rg) results; B forest plot of MR estimates for the effect of liver traits on outcomes; and C. forest plot of MR estimates for the effect of exposures on liver traits. Genetic correlations and MR estimates that reach the Bonferroni significance level (3.13 × 10⁻³ for genetic correlations and P < 2.94 × 10⁻³ for MR estimates) were marked with an asterisk. CI confidence intervals

Liver PDFF

The GWAS for PDFF confirmed associations with established loci (e.g. PNPLA3, TM6SF2, GCKR, APOE) and revealed one novel region near GSTM1 (rs140584594; β = 0.046 SD, P = 6.5 × 10⁻⁹; Table 1, Fig. 1, supplementary Fig. 3A). The strongest effects were seen for missense variants in TM6SF2 (rs58542926; β = 0.31 SD; p = 6.8 × 10⁻¹¹⁵) and PNPLA3 (rs738408; β = 0.26 SD; p = 4.6 × 10⁻¹⁵⁵). These variants were also significantly associated with cT1, showing consistent directions of effect. Adjustment for BMI generally strengthened the associations, and no sex-specific differences were observed (supplementary Table 2). Conditional analysis revealed independent additional signals near TM6SF2 and APOE, and fine-mapping identified putative causal variants (posterior probability ≥ 95%) at several loci, including APOE, SUGP1, and GSTM1/2 (supplementary Table 12).

Enrichment analysis highlighted the liver and subcutaneous adipose tissue as the primary cell types involved in the genetic regulation of liver PDFF (supplementary Fig. 2). TWAS identified 63 unique genes (Fig. 2). Of note, lower expression of C21orf33 and the long non-coding RNA ENSG00000226824 were consistently associated with higher liver PDFF across all tested tissues. SMR-HEIDI analysis supported 10 unique genes, with overlap observed for TMC4 between SMR and TWAS results (Fig. 3). Using cis-MR, we identified five plasma proteins—RAB2B, APOE, GSTM4, GPN1, and GCKR—as likely causal contributors to liver PDFF, supported by strong colocalisation evidence (PPH4 ≥ 70%; Fig. 4).

Two-sample MR provided robust evidence that genetically predicted increases in BMI, WHR, ALT, gamma-glutamyl transferase (GGT), triglycerides, and liability to type 2 diabetes were causally associated with higher liver PDFF (Fig. 5). In the reverse direction, higher liver PDFF was genetically associated with elevated ALT, increased risk of NAFLD, and higher diastolic blood pressure (Fig. 5). Detailed estimates for all these analyses of liver PDFF are provided in supplementary tables 13–18 (Table 2).

Table 2.

Summary of evidence grading across TWAS, cis-MR, and SMR–HEIDI analyses for candidate loci

rsID reference SNP identifier, GWAS genome-wide association study, BETA estimated effect size (per effect allele), P p-value for association, TWAS Transcriptome-Wide Association Study, SMR Summary-data-based Mendelian Randomization, HEIDI Heterogeneity in Dependent Instruments, MR Mendelian randomisation

Liver iron

The GWAS for liver iron identified five associated loci, including three previously reported regions (ASNSD1, HFE, TMPRSS6) as well as two newly implicated loci near VMP1 and SUPT4H1 (Table 1, Fig. 1, supplementary Fig. 3B). The most prominent signal was observed at the HFE locus (rs1800562; 0.31 SD increase per A allele; p = 7.7 × 10⁻¹⁰⁵), a well-known variant associated with hereditary haemochromatosis. These associations were unaffected by adjustment for BMI and were consistent across sexes (supplementary Table 2).

Fine-mapping confirmed rs855791 (TMPRSS6) as the likely causal variants (posterior probability ≥ 95%) (Supplementary Table 3). Enrichment analysis suggested that the most relevant tissues for liver iron biology include the liver itself, as well as the small intestine and terminal ileum (supplementary Fig. 2). TWAS identified 28 unique genes whose genetically predicted expression levels were associated with liver iron with strong evidence for inverse associations of RAD51C expression with liver iron across most tissues (Fig. 2). SMR-HEIDI further confirmed 45 genes, with overlapping evidence for genes such as SNX32, RAD51C, and ASNSD1 across TWAS and SMR (Fig. 3). In contrast to liver cT1 and PDFF, protein-level cis-MR did not identify any circulating proteins with a likely causal role in regulating liver iron.

Two-sample MR demonstrated that genetically higher ALT and GGT were associated with higher liver iron (Fig. 5). In the reverse direction, higher liver iron was linked to lower coronary artery disease risk and lower LDL-C. Although the direction of effect is not straightforward to interpret biologically. We therefore treat this result as a statistically supported genetic association rather than evidence of a protective biological mechanism. Detailed estimates for all these analyses of liver iron are provided in supplementary tables 19–24.

Integrated evidence from GWAS, TWAS, cis-MR, and SMR–HEIDI

Several loci were supported by multiple complementary lines of evidence, strengthening their candidacy for future functional investigation. Notably, ABO was supported by TWAS, SMR and cis-protein MR for CT1. FADS1/2 and FUT2 were supported by both TWAS and SMR for cT1, while ASNSD1 showed support from TWAS and SMR for liver iron. In addition to these multi-method-supported loci, several signals were supported by a single analytical approach. HFE and PNPLA3 were supported by SMR for cT1, GCKR by cis-protein MR for PDFF, ADH1B, GPAM, MAST3, and TMC4 by TWAS for PDFF, and APOE by cis-protein MR for PDFF. For liver iron, SUPT4H1 and TMPRSS6 were supported by SMR. Together, these findings highlight loci with convergent evidence across methods as well as additional candidates identified through individual approaches.

Multi-trait analysis identifies novel loci for liver cT1 and PDFF

To increase discovery power beyond single-trait GWAS, we applied MTAG by incorporating genetic summary data from type 2 diabetes, coronary artery disease, and stroke, which showed strong genetic correlations with both liver cT1 and PDFF in our study. Only disease outcomes with LDSC genetic correlation z > 4 were included to ensure robust genetic overlap with liver traits. This analysis identified seven additional loci reaching genome-wide significance (P < 5 × 10⁻⁸ including variants near or within the genes BMP8A, INSR, and PPP1R3B associated with liver cT1 and COBLL1, ZGPAT, MYO15A, AC022431.2, and PPP1R3B associated with liver PDFF (Table 1, supplementary Fig. 3C–F).

To evaluate whether these associations reflect shared causal variants between liver traits and cardiometabolic disease outcomes, we performed Bayesian colocalisation. We found strong evidence of colocalisation (PPH4 > 70%) at COBLL1, BMP8A, MYO15A, and ZGPAT, suggesting that a single variant may influence both liver traits and cardiometabolic disease outcomes at these loci. Colocalization results for liver traits and various outcomes are presented in supplementary Table 25.

In follow-up analyses, we also examined the functional relevance of the lead MTAG variants using transcriptomic, metabolomic, and proteomic datasets in the two largest plasma proteogenomic studies [37, 38], 47. For example, rs2315007, the lead variant near ZGPAT showed consistently significant effect on liver traits through TWAS ARFRP1 in pancreas (β = –0.017, p = 1.21 × 10⁻⁷) and LIME1 in subcutaneous adipose (β = –0.017, p = 1.21 × 10⁻⁵) with strong evidence of colocalization with liver PDFF. This variant acts as a metabolite QTL (mQTL) for indolepropionate (IPA; β = –0.017, p = 1.21 × 10⁻⁵), a microbiome-derived metabolite with hepatoprotective properties [48], and showed strong evidence of colocalization with liver PDFF (PPH4 = 0.98, supplementary Fig. 4). The same variant is a cis-protein QTL (cis-pQTL) for TNFRSF6B (Decoy Receptor 3; p < 1 × 10⁻³⁰⁰ (Fig. 6 and supplementary Fig. 5) and a trans-pQTL for FGFR4. TNFRSF6B and PDFF share a single causal variant in this region as shown by strong colocalization evidence (rs2315007, PPH4 = 0.92; supplementary Fig. 4; supplementary Table 26).

Fig. 6.

Molecular QTL and disease associations for rs2315007-T near ZGPAT. A Association of the rs2315007-T allele with GWAS-PDFF, MTAG-PDFF in the current study, indolepropionate from (48), and plasma TNFRSF6B and FGFR4 levels from (37). B Stacked regional association plot for the lead MTAG-PDFF variant rs2315007, which colocalizes with the plasma TNFRSF6B cis-pQTL and T2D risk signal in the European population. The color hue represents the strength of r2 indicating the LD structure, as shown in the legend

Other variants were also associated with circulating protein abundance relevant to both liver and cardiometabolic traits (supplementary Fig. 6). For liver cT1, rs7012637 (PPP1R3B) showed associations with multiple proteins, including CREB3L4, PYGL, BCHE, LEPR, COLGALT1, and CTSB. For liver PDFF, rs13389219 (COBLL1) was associated with PLA2R1, LY75, and leptin (LEP), and rs9687846 was associated with IGFBP2, LPL, and LDLR, suggesting this locus may influence hepatic fat accumulation through modulation of lipid metabolism and insulin sensitivity pathways.

Discussion

This study presents a comprehensive investigation of liver health, leveraging advanced MRI-based phenotyping, large-scale genomic analysis in over 37,000 UK Biobank participants. We simultaneously assess three complementary liver traits including liver inflammation (cT1), steatosis (PDFF), and iron content, mapping both shared and distinct components of their genetic architecture. We expand our understanding of liver health beyond traditional liver function tests and histological proxies, identifying distinct and shared genetic mechanisms across these key hepatic traits. Finally, our comprehensive post-GWAS analyses, leveraging large-scale transcriptomic and proteomic databases, pinpointed genes and proteins associated with these traits and provided novel insights into the biological mechanisms underlying the GWAS signals.

Genetic architecture and biological insights

Our GWAS confirmed established loci and identified novel associations (e.g., not reported in specific prior large-scale GWAS of the liver MRI traits as in ref 13, 15–17) for each trait. For liver cT1, which reflects fibroinflammatory activity, we identified three new loci, including CXXC4, TFCP2, and TSPAN5. Fine-mapping analyses pinpointed a high-confidence causal variant in SLC39A8, a gene previously linked to metal ion transport and inflammation, reinforcing its potential role in hepatic fibrogenesis. Similarly, TMPRSS6, a known regulator of hepcidin and iron homeostasis, was fine-mapped to a single likely causal variant, highlighting its relevance to both cT1 and liver iron biology. In liver PDFF analyses, we refined known associations (e.g. PNPLA3, TM6SF2, GCKR) [12], and identified novel signals near GSTM1/GSTM2, genes previously associated with liver enzyme levels and metabolic traits [49, 50]. These findings suggest an expanded role for glutathione metabolism and detoxification pathways in hepatic steatosis. Additionally, high-confidence associations were detected at MARC1, GPAM, and TRIB1, which may influence triglyceride synthesis and liver fat accumulation. For liver iron, novel signals were detected near ASNSD1, SUPT4H1, and VMP1. ASNSD1 variants have previously been linked to circulating hepcidin and transferrin receptor levels, suggesting a role in systemic iron regulation and hepatic storage [51].

The MTAG analyses identified seven additional loci. For example, liver cT1 results included PPP1R3B, a regulator of hepatic glycogen storage, and BMP8A, involved in metabolic thermogenesis, highlighting links between liver inflammation and systemic energy metabolism. The ZGPAT locus (rs2315007) for liver PDFF emerged as particularly noteworthy. This variant was associated with PDFF and colocalised with a range of molecular QTLs across transcriptomic (ARFRP1 and LIME1), metabolomic (indolepropionate (IPA)), and proteomic (TNFRSF6B and FGFR4) datasets. Prior research has demonstrated that IPA is a potent anti-non-alcoholic steatohepatitis (NASH) metabolite with hepatoprotective properties [52]. Both TNFRSF6B and FGFR4 are implicated in liver and immune signalling pathways. TNFRSF6B is a soluble receptor that can neutralise the activity of TNFSF members including TL1A (TNFSF15), which has been identified as a susceptibility gene for primary biliary cirrhosis [53, 54]. FGFR4, part of the hepatic FGF19–FGFR4/β-Klotho axis, regulates bile acid synthesis and lipid metabolism, and its activation may reduce hepatic fat accumulation and suppress gluconeogenic gene expression through an insulin-independent pathway [55]. Rodent studies have shown that FGFR4 overexpression can prevent hyperlipidaemia, insulin resistance, and fatty liver [56].

TWAS and SMR-HEIDI analyses identified several genes with strong evidence of trait-specific expression effects, many of which have experimentally validated roles in hepatocytes. For cT1, FUT2, FADS1 and NTN5 were highlighted. FADS1 encodes a desaturase highly expressed in hepatocytes and experimentally shown to regulate hepatic lipid composition and inflammation [57, 58]. For PDFF, the MBOAT7/TMC4 locus stood out and was supported by both transcriptomic approaches. MBOAT7 is a hepatocyte-expressed acyltransferase whose loss reduces phosphatidylinositol remodelling and promotes steatosis in cellular and mouse models [59, 60], consistent with its observed liver-specific TWAS/SMR signal. For liver iron, RAD51C has validated functions in DNA damage responses whose encoded proteins has been shown to promote Hepatitis B virus replication in mice hepatocytes [61]. Several loci prioritised through TWAS and SMR also mapped closely to GWAS sentinel variants. For example, GPAM, which encodes the glycerol-3-phosphate acyltransferase 1 (GPAT1) protein and is central to glycerolipid synthesis in hepatocytes [62], was prioritised by SMR in sigmoid colon tissue, and previous research has shown that overexpression of mitochondrial GPAT in rat hepatocytes leads to decreased fatty acid oxidation and increased glycerolipid biosynthesis [62]. Overall, the overlap between TWAS and SMR provides strengthened confidence in these genes as plausible effector transcripts.

TWAS further identified biologically compelling genes that did not reach genome-wide significance in GWAS but showed strong tissue-specific expression associations, including GUSB and SLC25A19 (PDFF), and LRRFIP2, CSF3 (cT1) and SNX32 (iron). LRRFIP2 functions as a macrophage regulator of the NLRP3 inflammasome [63], and its prioritisation for cT1—an inflammation-related measure—is biologically consistent [63, 64]. CSF3, a key cytokine controlling granulocyte development and macrophage signalling [65], has demonstrated roles in hepatic inflammatory responses [66]. These findings demonstrate how transcriptomic integration can highlight effector genes involved in lipid metabolism, inflammation and immune–liver cross-talk, even in the absence of strong GWAS signals.

Protein-level MR from large-scale proteomic datasets highlighted biologically plausible proteins—such as ABO, ADH1B, GCKR, GPN1, GSTM4, and RAB2B—as potential upstream drivers of liver phenotype variability. Among these, the novel associations of GPN1 and RAB2B with liver PDFF are worth highlighting. GPN1 encodes a guanosine triphosphatase enzyme, and RNAi knockdown experiments have shown that it may impact fat accumulation in the liver independently of ethanol exposure in Drosophila models [67]. RAB2B, a member of the Ras protein family, was recently reported to be associated with alcohol-associated hepatitis in an exome-wide association study [68]. These findings suggest a potential role for RAB2B in the molecular mechanisms connecting lipid metabolism and liver inflammation in alcohol-related liver disease. These proteins may serve as future biomarkers or therapeutic targets.

Causal inference and metabolic interplay

MR analyses demonstrated that liver phenotypes are not merely passive markers of metabolic dysfunction but participate in bidirectional, and at times divergent, causal relationships. Genetically predicted liver cT1 levels were positively associated with BMI and inversely associated with HDL-C. These results underscore the role of fibroinflammation in the pathophysiology linking adiposity and vascular risk. Liver PDFF was causally influenced by obesity-related traits, including BMI, waist circumference, and triglycerides. In turn, elevated liver PDFF was associated with higher ALT, NAFLD risk, and diastolic blood pressure. These findings reinforce the utility of liver PDFF as an early biomarker of systemic metabolic risk. For liver iron, genetic instruments indicated inverse associations in reverse MR, whereby higher genetically predicted liver iron was linked to lower CAD risk and reduced LDL-C. We emphasise that this finding should not be interpreted as evidence of a clinically protective effect of hepatic iron accumulation. Rather, it likely reflects pleiotropic effects of genetically regulated iron homeostasis pathways—particularly at loci such as HFE and TMPRSS6—on lipid metabolism, inflammation, or erythropoiesis, rather than a direct causal benefit of increased liver iron per se [69]. In addition, cohort selection, survivor bias, and the cross-sectional nature of MRI-derived liver iron may contribute to non-intuitive directionality. As with all MR findings in complex traits, these results should be viewed as hypothesis-generating and interpreted cautiously pending replication and mechanistic validation.

Clinical translation and risk stratification

From a translational standpoint, MRI-derived traits such as cT1 and PDFF, when combined with genetic data, may enable improved risk stratification through polygenic risk scores or integrated imaging–genomics prediction models. Such approaches could refine patient selection for monitoring, early intervention, or trial enrolment in MASLD and related conditions. Several of our findings point to genes and proteins that could serve as potential targets for future pharmacologic modulation. For example, FADS1 is a biologically and chemically tractable target with active and experimental ligands based on the PHAROS database, suggesting potential for future pharmacologic modulation, while GSTM4 interacts with approved and experimental compounds (cisplatin and POM2-C-HMBP), highlighting its functional relevance and potential as a complementary circulating biomarker. The observed sex differences in liver composition further suggest that future models should incorporate sex-specific thresholds or interactions to improve predictive performance and clinical utility. In current clinical practice, MASLD screening relies on serum scores (such as FIB-4), ultrasound, and selective use of elastography or MRI. Our findings suggest a future workflow where MRI-derived cT1 and PDFF serve as quantitative end points in at-risk individuals (for example, those with type 2 diabetes or obesity), and genetic information helps identify people with a lifelong high burden of steatosis or fibroinflammation who may benefit from earlier imaging or closer follow-up. Variant-informed or polygenic scores could also be integrated into existing cardiometabolic risk calculators. These applications remain exploratory and will require evaluation of clinical utility and cost-effectiveness.

Strengths and limitations

Key strengths of our study include the use of high-resolution, quantitative MRI phenotypes; robust quality control; and multiple layers of post-GWAS validation, including fine-mapping, transcriptomics, and proteomics. Furthermore, our use of large-scale causal inference frameworks provides mechanistic insights into upstream and downstream consequences of liver trait variation.

Nevertheless, several limitations merit consideration. First, while MRI-derived metrics such as cT1, PDFF, and liver iron are validated non-invasive measures of fibroinflammation, steatosis, and iron content, they do not capture all histological features with the same granularity as liver biopsy. These imaging biomarkers reflect continuous tissue properties rather than discrete histological classifications and should be interpreted within that context.

Second, our analyses were restricted to individuals of European ancestry, which limits generalisability to other populations and prevents formal assessment of ancestry-specific effects. Differences in allele frequencies, LD structure, and environmental exposures across ancestries may affect both locus discovery and the transferability of polygenic risk scores, highlighting the need for future trans-ethnic replication to improve generalizability. As MRI-based liver phenotyping becomes available in more diverse cohorts, it will be important to replicate these findings, perform trans-ethnic fine-mapping, and test whether polygenic scores for cT1, PDFF and liver iron retain predictive performance across ancestries. Such efforts will be essential to ensure that genetically informed risk stratification for liver disease does not exacerbate existing health disparities. Third, our TWAS colocalization analysis (COLOC in FUSION) assumes a single causal variant and therefore cannot capture multiple independent signals; however, the consistently low PPH3 values across prioritized hits suggest the probability of different causal variants between traits is unlikely. Methods accommodating multiple causal variants could be used in future studies. Finally, although MR provides a powerful framework for causal inference, horizontal pleiotropy cannot be entirely excluded.

Conclusion

Our findings provide a high-resolution map of the genetic landscape underlying three clinically relevant liver traits—cT1, PDFF, and liver iron—and their causal interrelations with cardiometabolic health. The identification of new loci, causal risk factors, and functionally implicated genes and proteins opens new avenues for early detection and personalised intervention in liver disease. Future efforts should extend these findings to diverse ancestries and investigate the translational potential of key molecular mediators identified in this study.

Author contributions

H.Y. and A.D. conceived and designed the study and jointly supervised the overall project. H.Y. conducted the GWAS analyses. D.M., J.H., and A.A. performed the post-GWAS analyses. D.M., A.D., and H.Y. contributed to the interpretation of the analyses and to the development of the analytical strategy and study direction. D.M. and H.Y. drafted the manuscript. M.P., A.F., A.D., and R.B. critically reviewed the manuscript and provided expert input. M.P., A.F., and R.B. contributed to the derivation and processing of the MRI-based liver phenotypes. H.Y. is the guarantor of this work and had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Funding

This study was supported by financial support from Diabetes UK (Grant 23/0006598; PI: Hanieh Yaghootkar).

Data availability

Full data will be returned to UK Biobank and made publicly available via application (amsportal.ukbiobank.ac.uk).

Declarations

Ethics approval and consent to participate

This research was conducted using the UK Biobank resource under application number [9914]. UK Biobank has obtained research ethics approval from the North West Multi-centre Research Ethics Committee (REC reference for UK Biobank is 11/NW/0382). All participants provided written informed consent, and the study was conducted in accordance with the principles of the Declaration of Helsinki.

Competing interests

The authors declare no competing interests.