Network Interactions of Circulating FGF23, HRG-HMGB1, and Cardiac Disease in CKD

Visual Abstract

Abstract

Key Points

• Multitrait analysis of genome-wide association study boosts the statistical power to identify novel genetic traits for fibroblast growth factor 23.

• A functional genomics approach aided network discovery to identify histidine-rich glycoprotein (HRG) and high-mobility group protein box 1 (HMGB1) as key regulators of cardiac disease in CKD.

• Integration of clinical and genetic data enhances the discovery power and is crucial for understanding the genetic underpinnings of mineral bone disorder related to CKD.

Background

Genome-wide association studies (GWAS) have identified numerous genetic loci associated with mineral metabolism markers but have exclusively focused on single-trait analysis. In this study, we performed a multitrait analysis of GWAS (MTAG) of mineral metabolism, exploring overlapping genetic architecture between traits to identify novel genetic associations for fibroblast growth factor 23 (FGF23).

Methods

We applied MTAG to variants common to GWAS of five genetically correlated mineral metabolism markers in participants of European ancestry. We integrated UK Biobank GWAS for blood levels for phosphate, 25-hydroxyvitamin D, and calcium (n=366,484) and Cohorts for Heart and Aging Research in Genetic Epidemiology GWAS for parathyroid hormone (n=29,155) and FGF23 (n=13,716). We then used supervised and unsupervised deep machine learning to identify novel associations between genetic traits and FGF23.

Results

MTAG increased the effective sample size for mineral metabolism markers to n=50,325 for FGF23. After clumping, MTAG identified independent genome-wide significant single-nucleotide polymorphisms for all traits, including 62 loci for FGF23. Many of these loci have not been previously reported in single-trait analyses. Through a functional genomics approach, we identified histidine-rich glycoprotein (HRG) and high-mobility group box 1 (HMGB1) as master regulators of downstream canonical pathways associated with circulating FGF23, and both genes were highly enriched in hypertrophied cardiac tissue of deceased hemodialysis patients. In addition, we found that DNMT3A was associated with uremic toxin, 8-hydroxy-2-deoxyguanosine, a biomarker of DNA damage. In silico gene perturbation analysis revealed that DNMT3A is protective in patients with heart failure caused by hypertrophied or dilated cardiomyopathy.

Conclusions

Our findings highlight the importance of MTAG analysis of mineral metabolism markers to boost the number of genome-wide significant loci for FGF23 to identify novel genetic traits. Functional genomics revealed novel networks that inform unique cellular functions and identified HRG and HMGB1 as key master regulators of FGF23 and cardiovascular disease in CKD.

Introduction

Mineral bone disorder related to CKD (CKD-MBD) leads to impaired skeletal and cardiovascular homeostasis and is associated with higher fracture risk, vascular calcification, and cardiovascular-related morbidity and mortality.¹ Therapeutic strategies to prevent and treat CKD-MBD, including maintenance of mineral marker homeostasis, have not led to meaningful decreases in morbidity and mortality.

We and others have shown that genome-wide association studies (GWAS) in the general population and in CKD have identified common variants associated with circulating mineral metabolism biomarkers (calcium, phosphate, 25[OH]D, parathyroid hormone [PTH], and fibroblast growth factor 23 [FGF23]).^2–6 In addition, Mendelian randomization studies that leverage genetic determinants of a risk factor to determine causality have shown that genetic predictors of FGF23 excess were associated with higher heart failure risk in those with genetically predicted lower GFR.⁷ However, previous studies have exclusively focused on single-trait analysis, leaving gaps in comprehensive understanding of genetic drivers of CKD-MBD. In this study, we performed a multitrait analysis of GWAS (MTAG) for mineral metabolism CKD-MBD markers, exploring overlapping genetic architecture between the traits, to identify novel genetic signatures for FGF23. We then leveraged machine learning methodologies to reprioritize quantitative genetic traits and build relevant biologic networks that identify core disease-associated genes (Figure 1A). Through this omnigenic approach, we have identified previously unknown genetic traits associated with biologic networks at the cellular and tissue levels that play a major role in cardiovascular disease in CKD.

Figure 1.

Schematic representation of study design and MTAG of mineral metabolism. (A) Overall study design. (B) Manhattan plot shows 62 loci from FGF23_MTAG. In these plots, the y axis shows the P values of SNPs in a log-log scale. CHARGE, Cohorts for Heart and Aging Research in Genetic Epidemiology; FGF23, fibroblast growth factor 23; GO, gene ontology; GWAS, genome-wide association study; KEGG, Kyoto Encyclopedia of Genes and Genomes; MTAG, multitrait analysis of genome-wide association study; PTH, parathyroid hormone; SNP, single-nucleotide polymorphism; SVM, support vector machine; TKT, Transketolase.

Methods

MTAG

We conducted MTAG of five genetically correlated mineral metabolism biomarkers using data from large-scale discovery GWAS, as previously described.⁸ We integrated summary-level GWAS data from UK Biobank for blood levels of phosphate, 25(OH)D and calcium (n=366,484), and from individuals of European ancestry in Cohorts for Heart and Aging Research in Genetic Epidemiology (CHARGE) consortium for PTH (n=29,155) and full-length, biologically intact FGF23 (n=13,716).^9,10 We included 25-OH vitamin D3 in the analysis because, despite its primary role reflecting diet and sun exposure, it exhibits significant genetic correlation with FGF23 and other mineral metabolism markers, which is critical for the effectiveness of multitrait GWAS.

MTAG leverages the genetic correlation between related traits to perform joint genome-wide analyses of multiple traits, hence augmenting the available statistical power to detect novel genetic signals for each trait analyzed. As an extension of the inverse-variance meta-analysis in the multitrait setting, the MTAG estimator uses single-trait GWAS summary statistics as inputs and generates trait-specific single-nucleotide polymorphism (SNP) effects and P values. Beyond the advantage of requiring only GWAS summary statistics, the MTAG estimator adequately accounts for potential sample overlap in the GWAS of the different traits included in the analysis using bivariate linkage disequilibrium score regression of each pair of traits. For this study, MTAG associated with circulating FGF23 (MTAG_FGF23) was used to construct network models.

Network-Wide Association Study

Network-Wide Association Study (NetWAS) is a supervised machine learning approach that trains a classifier through grouped disease-associated genes.¹¹ Classifiers locate an optimal hyperplane from the high-dimensional predictor space to separate positive genes from negative genes. Classifiers are constructed using a tissue network relevant to a disease, where the features of a classifier are the edge weights of the labeled examples to all the genes in the network. To create a network-based prioritization of GWAS, a score is assigned to each gene using the distance from the hyperplane. Genes that score higher are more likely to be disease-related. We used MTAG_FGF23 variants as input data after conversion of the associated P value of SNP level to gene level through the versatile gene-based association study web platform.¹² Genes were reprioritized, and the results were ranked according to their scores to create tissue-specific gene clusters for kidney, bone, and heart. Genes with high scores and/or obtained consistent results in sensitivity analyses were considered to be potential causal genes.

Geneformer

Geneformer is a context-aware, attention-based unsupervised deep machine learning model pretrained on approximately 30 million single-cell transcriptomes to understand network dynamics. It can accurately predict dosage-sensitive disease genes and their downstream targets and was successfully applied to identify candidate therapeutic targets for cardiomyopathy.¹³ We used published data from single-cell Drop-seq and DroNc-seq of human hearts with hypertrophied and dilated cardiomyopathy (DCM) explanted at the time of transplantation and deceased donors with nonfailing hearts (Supplemental Methods and Supplemental Table 2).¹⁴ In silico perturbation analysis was performed with Geneformer to determine whether gene deletion/activation shifts cells from hypertrophic cardiomyopathy (HCM/DCM) toward the nonfailing heart state, and we determined the overlap of these data with MTAG_FGF23. We also queried uremic toxin databases to obtain a gene list that overlapped with HCM/DCM and MTAG_FGF23 in the context of CKD.

To ensure methodologic independence and avoid potential error propagation, we conducted four separate machine learning analyses—Ingenuity Pathway Analysis (IPA), NetWAS, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment, and Geneformer—each applied independently to MTAG-derived SNPs. Key findings were validated across multiple platforms, providing complementary yet distinct evidence for biologic relevance.

Statistical Analyses

We performed MTAG analysis of single-trait GWAS of five mineral metabolism biomarkers with a focus on FGF23 as the primary analysis, using the conventional Python command line tool.⁸ The MTAG estimator is an efficient generalized method of moments estimator that generates trait-specific SNP effects by performing a weighted sum of the GWAS estimates while accounting for the correlation in true SNP effects and the correlation in their estimation error (related to phenotypic correlation with sample overlap and correlated biases in SNP effects). The stepwise algorithm estimates a variance-covariance matrix of the SNPs' estimation error (using unadjusted and bivariate linkage disequilibrium score regression), the variance-covariance matrix of the SNP effects (assumed to be homogeneous across all SNPs) and then computes individual SNP effects across all traits using a closed-form solution. Independent genetic signals for FGF23 were obtained by clumping genome-wide significant signals at an r² of 0.01 in 500 kb windows. We probed the credibility of our findings in our work, in part by computing the false discovery rate for each SNP.

Results

MTAG_FGF23

The strength of genetic correlations is important for investigating the pleiotropic effect of genetic variants on traits. Multitrait analyses augment the statistical power to detect novel genetic signals to yield plausible and replicable results. We performed MTAG analysis after combining GWAS of calcium, phosphate, 25(OH)D, PTH, and FGF23 and identified 62 genomic regions for FGF23; 57 loci were novel (not within ±500 kb of previously known loci). The top loci (rs73743401 and rs73188394) were associated with approximately 21% to 23% higher circulating FGF23 and were in proximity to IP6K3, Transketolase (TKT), and FAM162A genes (Figure 1B, Table 1, and Supplemental Table 1).

Table 1.

Top 30 independent genetic loci associated with circulating fibroblast growth factor 23 from multitrait analysis of genome-wide association study analyses

Genetic Variant	Chr	Position	Ref Allele	Effect Allele	MAF	β	SEM	P Valuea	Nearest Gene(s)
rs73188394	3	122129203	A	G	15.4%	21%	0.9	5.3×10−116	TKT, FAM162A
rs73743401	6	33716945	G	T	5.7%	22.5%	1.4	1.5×10−58	IP6K3, ITPR3
rs838717	2	234296444	G	A	44%	−8.8%	0.6	4.9×10−42	DGKD
rs1976403	1	21766453	A	C	41.9%	8.5%	0.6	9.0×10−39	NBPF3
rs11063207	12	4603698	T	C	10%	−12.2%	1.0	1.2×10−36	C12orf4, RAD51AP1
rs6727384	2	97400324	A	G	38.6%	−7.2%	0.6	7.8×10−28	LMAN2L, CNM4
rs7742369	6	34165721	A	G	17.5%	9.2%	0.8	1.0×10−27	HMGA1, C6orf1
rs72786681	10	9318246	T	C	10.3%	11.9%	1.1	2.9×10−27	LINC00709
rs3091842	20	39344272	G	A	6.7%	16.8%	1.5	4.9×10−27	RP4-644L1.2
rs453639	6	132049657	C	A	39.7%	7.1%	0.7	1.3×10−26	ENPP3, CTAGE9
rs4263551	6	74486371	T	C	49.7%	−6.8%	0.6	4.7×10−26	CD109
rs17216707	20	52732362	T	C	22.7%	−8.1%	0.8	4.4×10−23	CYP24A1
rs308032	19	3102748	G	T	21.5%	6.9%	0.7	6.1×10−20	GNA11
rs4782302	16	88524389	A	C	41.5%	−5.9%	0.6	8.4×10−20	ZFPM1, ZNF469
rs12411216	1	155164480	C	A	45.2%	5.7%	0.6	7.1×10−19	MUC1, TRIM46
rs10819178	9	129294976	T	G	35.2%	5.8%	0.7	3.4×10−18	AL356309.1
rs34257685	17	37621053	A	C	26.1%	−6.2%	0.7	3.3×10−17	CDK12, FBXL20
rs12366257	11	13593857	G	A	17.2%	7.5%	0.9	6.9×10−17	BTBD10, PTH
rs4937122	11	126228659	T	G	7.4%	−10.2%	1.2	8.1×10−17	ST3GAL4, DCPS
rs11752385	6	134522487	C	T	39.6%	−5.4%	0.6	1.6×10−16	SGK1
rs7686873	4	40564730	A	G	18.3%	−6.8%	0.8	2.6×10−15	RBM47
rs4718275	7	65220463	T	C	43.5%	5.1%	0.6	2.9×10−15	TPST1, GUSB
rs1858800	16	73024276	C	T	34.8%	5.3%	0.7	3.0×10−15	ZFHX3
rs145032242	15	50995847	C	T	19%	−6.3%	0.8	3.2×10−15	SPPL2A, USP50
rs841572	1	43436051	G	A	40.4%	5.1%	0.6	3.8×10−15	SLC2A1
rs6539287	12	107297650	C	A	47.1%	−5%	0.6	8.2×10−15	CRY1, C12orf23
rs4841132	8	9183596	A	G	7.4%	8.5%	1.1	1.5×10−14	RP11
rs544630567	5	176757962	A	G	34.1%	5.2%	0.7	9.5×10−14	RGS14
rs8025182	15	69601919	C	T	36.9%	−4.9%	0.7	2.2×10−13	PAQR5
rs4744843	9	80298380	C	A	21%	−5.7%	0.8	2.6×10−13	GNAQ

Multitrait analysis of genome-wide association study results for fibroblast growth factor 23, top 30 SNPs displayed. β coefficients represent the multitrait analysis of genome-wide association study–estimated association of each additional effect allele with difference in circulating fibroblast growth factor 23 concentrations. MAF, minor allele frequency; PTH, parathyroid hormone; SNP, single-nucleotide polymorphism; TKT, Transketolase.

^aP values are false discovery rate–adjusted.

NetWAS

Although GWAS has transformed our understanding of complex human diseases such as CKD, only a small fraction of total heritability can be explained by disease-associated variants. To address this “missing heritability,” we used NetWAS to identify the cumulative weak effects of variants, many of which fall below statistical significance. NetWAS integrates tissue-specific networks with GWAS and has the advantage of mitigating literature bias. Thus, candidate genes that are not well represented in the current literature but exhibit strong support for pathogenesis, which can be discovered agnostically.¹⁵ To this end, we tested the omnigenic hypothesis, which describes that genes can affect each other through their tightly interconnected networks, and as such, genes that have little direct bearing on a particular disease may, in aggregate, affect core disease pathways and influence disease risk. Using this approach, we identified significant gene clusters in the kidney, bone, and heart: nine distinct gene clusters in kidney and bone and six gene clusters in heart, including modules that regulate mineral homeostasis (Figure 2A). Given the intricate relationship between these organs in CKD-MBD, we examined for overlap and found 533 genes were common in the NetWAS module discovery network (Figure 2B). Overlapping genes were subjected to further analyses; aryl hydrocarbon receptor, HIF1α, TGFβ, mitochondrial dysfunction, and Nrf2-mediated oxidative stress response were among the top canonical pathways (Figure 2C), and HNF4α was a top upstream regulator (not shown) common to all three organs.

Figure 2.

NetWAS: a tissue-specific network-based functional interpretation of gene variants from MTAG_FGF23. (A) Module discovery network at the tissue level for the kidney, bone, and heart. (B) Venn diagram for overlapping genes in the NetWAS modules. (C) Top canonical pathways common to kidney, bone, and heart tissues for the NetWAS gene clusters. NetWAS, Network-Wide Association Study.

Functional Enrichment Analysis

We further explored the biologic implications of MTAG_FGF23 variants with gene ontology and KEGG enrichment analyses. Gene ontology term “biological processes” was enriched for two key pathways: skeletal development and vitamin D (Supplemental Figure 1), which confirms the biologic relevance of MTAG given FGF23 is a potent regulator of vitamin D metabolism and skeletal mineralization.^16–18 We also queried the top genes from Table 1 with KEGG and found that metabolic pathways for pentose phosphate and phosphatidyl inositol signaling were highly enriched (Supplemental Figures 2 and 3) and were also present in the overlapping gene clusters for kidney, bone, and heart in NetWAS. Interestingly, we found the SNP (rs73188394) that was most strongly associated with FGF23 (Table 1) was in proximity to TKT gene of the pentose phosphate pathway and plays a critical role in energy metabolism and oxidative stress response. Pentose phosphate pathway is highly conserved in evolutionary biology, and metal ions (iron and calcium), which are regulated by FGF23, are vital for the nonenzymatic reactions in this pathway, which also provides substrates for glycolysis. Recently, it was reported that the kidney-specific glycolysis pathway serves as a phosphate sensor and its by-product, glycerol-3-phosphate, upregulates FGF23 production in bone.¹⁹ Given that the pentose phosphate pathway is tightly linked to glycolysis by providing substrates and energy, TKT may play a central role in linking FGF23 to vital cellular functions, including energy metabolism, phosphate utilization, and response to hypoxia.²⁰

Integrated Network Modeling

Sequence models of disease prediction (e.g., GWAS) can describe the molecular effects of gene variants and expressions, but the interpretation of how they lead to disease phenotypes requires an understanding of the dysregulated pathways and processes. Using an integrated network modeling approach, massive collections of omics data can be summarized to create genome-scale functional maps and find novel gene interaction networks specific to a biologic context. To this end, we used a supervised machine learning method that integrates signal from GWAS together with signal in tissue-specific functional networks to reprioritize genes associated with the disease/trait of interest. VDR and TGFβR2 signaling pathways were identified as the top regulator effect networks, and MAPK signaling molecules were common to the top five causal networks (data not shown). These are known pathways activated by FGF23, thereby validating our results. We discovered a novel pathway that was previously unreported involving Coordinated Lysosomal Expression and Regulation (CLEAR) had the highest predicted activation score (z-score 2.24, P value 7.7.E-03; Supplemental Figure 4A). Cardiac hypertrophy, RAR, and estrogen receptor signaling also had high predicted inhibition z-scores. Among the canonical pathways that regulate cell metabolism, pathways for pentose phosphate and phosphatidyl inositol biosynthesis had high predicted activation z-scores. Signaling pathway for FGF6, FGF7, and FGF23 had the highest predicted inhibition z-scores (z-score −2.0, P value 5.4.E-03). We discovered that CLEAR pathway was also among the overlapping gene modules in NetWAS. CLEAR pathway regulates lysosomal function and cellular response to nutrient sensing, and six genes (FNIP1, GBA1, GUSB, ITPR3, PPP2R3A, and RALB) were found to have significant z-scores (Table 2). ITPR3 in proximity to SNP rs73743401 was associated with approximately 22% higher FGF23 (minor allele frequency 0.14, β coefficient 22.0, P value 5×10⁻⁶⁰). Next, we queried for disease states associated with higher z-scores for CLEAR pathway and found that 22 genes were upregulated in heart tissue of deceased hemodialysis patients (Cardiac Aging in CKD [CAIN] cohort) with left ventricular hypertrophy (Supplemental Figure 4B) and proceeded with an in-depth comparison analysis.

Table 2.

Genes associated with multitrait analysis of genome-wide association study associated with circulating fibroblast growth factor 23 in the Coordinated Lysosomal Expression and Regulation signaling pathway

Symbol	Entrez Gene Name	SNP Variant	β Coefficient	P Value	Location	Type(s)
FNIP1	Folliculin interacting protein 1	rs11950815	3.66	1.63E-08	Cytoplasm	Other
GBA1	Glucosylceramidase β 1	rs12411216	5.71	8.69E-12	Cytoplasm	Enzyme
GUSB	Glucuronidase β	rs4718275	5.08	9.49E-38	Cytoplasm	Enzyme
ITPR3	Inositol 1,4,5-trisphosphate receptor type 3	rs73743401	22.45	2.25E-08	Cytoplasm	Ion channel
PPP2R3A	Protein phosphatase two regulatory subunit B″α	rs1393787	5.05	9.49E-38	Nucleus	Phosphatase
RALB	RAS like proto-oncogene B	rs6542581	−3.91	1.28E-20	Cytoplasm	Enzyme

SNP, single-nucleotide polymorphism.

Comparison Analysis of MTAG_FGF23 and Cardiac Disease in Advanced CKD

To further analyze the top signals identified by integrative network modeling, we investigated tissue-specific functional networks and patterns in the heart, an organ system that is directly affected by CKD-MBD and FGF23 excess.²¹ We performed comparison analyses of MTAG_FGF23 with a publicly available left ventricle RNA-seq dataset from dialysis patients collected after death.²² In the CAIN cohort, patients on hemodialysis had left ventricular hypertrophy with significantly greater heart tissue weight, wall thickness, and myocardial fibrosis compared with controls. In the comparison analysis, we found overlapping and distinct pathways (Figure 3). Using this integrative systems approach, we identified the CLEAR pathway as a key regulatory network associated with MTAG_FGF23. This finding was supported by multiple independent analyses, including IPA, NetWAS, and disease state enrichment analysis, which collectively implicated the CLEAR pathway in cardiac pathology in CKD. Specifically, 22 genes from the CLEAR network were found to be significantly upregulated in the CAIN cohort, a dataset derived from deceased hemodialysis patients with left ventricular hypertrophy, further supporting its relevance in CKD-MBD. Among the pathways for disease and functions, organismal death had the highest predicted activation score (Figure 4) and cell migration had the highest inhibition score. We found top genes from MTAG_FGF23 (KPNA1, PARP9, ITPR3, TKT, Rgs14, FGF23, PTH, DGKD, and DNMT3A) overlapped with the organismal death network in the CAIN cohort.

Figure 3.

Comparison analysis of lead genetic loci in MTAG_FGF23 versus differential gene expression from bulk RNA-seq of left ventricular tissue from patients with advanced CKD (CKD-LVH). (A) Canonical pathways and (B) causal networks common and distinct within each dataset. Activation z-scores are represented as orange boxes for predicted activation and blue boxes for inhibition. (C) Genes upregulated (orange) and downregulated (blue) or no change (gray) in the pathway for cardiac hypertrophy. CKD_LVH, CKD-left ventricular hypertrophy; CLEAR, Coordinated Lysosomal Expression and Regulation.

Figure 4.

Comparison analysis of lead genetic loci in MTAG_FGF23 versus differential gene expression from bulk RNA-seq of left ventricular tissue from patients with advanced CKD (CKD-LVH). Pathways for organismal death identified in both datasets are represented in the schematic diagram. Genetic loci are represented in red/pink and in green when associated with higher or lower circulating FGF23, respectively.

Histidine-Rich Glycoprotein and High-Mobility Group Box 1 Identified as Master Regulators of Downstream Canonical Pathways Linked to Both FGF23 and Cardiac Disease in Advanced CKD

We examined MTAG_FGF23 and CAIN for upstream molecules that potentially function as master regulators of disease/trait of interest (i.e., cardiac disease in CKD). We discovered a previously unknown causal network with histidine-rich glycoprotein and high-mobility group protein box 1 (HRG-HMGB1) as the master regulator of downstream canonical pathways common to both datasets (Figure 3B). In disease states, HRG-HMGB1 is proinflammatory and causally associated with kidney and cardiovascular injury.²³ Consistent with the reported literature, HRG-HMGB1 was upstream of major signaling pathways (Figure 5). HRG activation predicts inhibition of HMGB1 to downregulate signaling through TGFβ1, TLR, IL-2, IL-6, NFKB, HIF1α, AKT, and stat5a/b. In addition, we identified several top genes from Table 1 within HRG-HMGB1 networks. Specifically, FAM162a, TKT, PARP9, and DTX3L genes were activated, and FGF23, ST3GAP, USP8, BCAS3, GNAQ, PCGRT, and PPP1R1B were inhibited by HRG-HMGB1.

Figure 5.

Comparison analysis of lead genetic loci in MTAG_FGF23 versus differential gene expression from bulk RNA-seq of left ventricular tissue from patients with advanced CKD (CKD-LVH). Causal network analysis revealed HRG-HMGB1 as a master regulator for downstream canonical pathways in (A) MTAG_FGF23 and (B) CKD_LVH dataset. (C) Magnified view of HRG-HMGB1 network from MTAG_FGF23. HRG-HMGB1, histidine-rich glycoprotein and high-mobility group protein box 1.

Geneformer Analysis

We used the Geneformer deep machine learning and transfer learning biology tool to fine-tune and model cell states, enabling us to distinguish hypertrophic and dilated cardiomyocytes from healthy cardiomyocytes.¹³ In silico gene perturbation strategy was applied to determine whether activation/deletion of genes shifted cell embedding from HCM or DCM to nonfailing heart state (Figure 6A) and the degree of overlap with MTAG_FGF23 (Table 3). Our analysis identified several genes within MTAG_FGF23 that, when activated, shifted the phenotype from HCM or DCM to a nonfailing heart, with six overlapping genes (DNMT3A, SECISBP2L, FAM227B, FGF7, BCAS3, and ZFPM1) (Figure 6B). Similarly, we found five genes that, when deleted, produced the same shift, with one overlapping gene (Figure 6C). Queries within uremic toxin databases revealed several toxins associated with key genes within MTAG_FGF23 and cardiomyopathy. Notably, the uremic solute 8-hydroxy-2-deoxyguanosine (8-OH2dG), a biomarker of DNA damage and oxidative stress, was associated with DNMT3A, which was common to both datasets (Figure 6D).

Figure 6.

Deep machine learning distinguishes hypertrophic and dilated cardiomyocytes from healthy cardiomyocytes. (A) Transfer learning using Geneformer was performed to model cardiac disease to distinguish cardiomyocytes affected by HCM or DCM from NF cardiomyocytes. (B and C) In silico gene perturbation analysis to determine whether gene activation or deletion shifts cell embedding from HCM or DCM toward a NF heart state. Venn diagrams represent genes that are protective against cardiomyopathy and also present in MTAG_FGF23 and HCM or DCM datasets. (D) Venn diagrams represent genes that are present in MTAG_FGF23 and HCM or DCM datasets that are associated with uremic solutes from the uremic toxin database. DCM, dilated cardiomyopathy; HCM, hypertrophic cardiomyopathy; NF, nonfailing.

Table 3.

Cardiac disease modeling for hypertrophic and dilated cardiomyopathy and in silico gene perturbation analysis

Pairing for Fisher Exact Test	Odds Ratio (95% CI)
MTAG genes versus HCM_in silico gene activation	2.5 (1.3 to 4.3)
MTAG genes versus DCM_in silico gene activation	2.6 (1.1 to 5.3)
MTAG genes versus HCM_in silico gene deletion	3.3 (0.7 to 9.9)
MTAG genes versus DCM_in silico gene deletion	3.8 (0.8 to 11.5)

Fine-tuning in Geneformer was performed to distinguish hypertrophic or dilated cardiomyopathy from cardiomyocytes of nonfailing hearts. Disease modeling is performed by in silico perturbation of random genes to identify genes when deleted or activated shifts cell embedding significantly from hypertrophic cardiomyopathy or dilated cardiomyopathy toward the nonfailing heart state. Genes were filtered for nonfailing heart, false discovery rate <0.01, and shift to nonfailing >0. Fisher exact test was performed to determine if there are significant associations with multitrait analysis of genome-wide association study associated with circulating fibroblast growth factor 23 genes. The odds ratio represents the enrichment of multitrait analysis of genome-wide association study associated with circulating fibroblast growth factor 23 genes in the respective gene list from the Geneformer analyses. CI, confidence interval; DCM, dilated cardiomyopathy; HCM, hypertrophic cardiomyopathy; MTAG, multitrait analysis of genome-wide association study.

Discussion

Computational biology has transformed our ability to decode complex diseases such as CKD, enabling the integration of large-scale genetic and functional data to uncover novel insights.²⁴ In this study, we applied advanced machine learning approaches to identify genetic drivers and construct regulatory networks for CKD-MBD. Using MTAG, a novel methodology that integrates GWAS summary statistics across multiple genetically correlated traits, we significantly expanded the number of genome-wide significant loci associated with FGF23. This integration of clinical and genetic data allowed us to enhance discovery power and identify novel loci that are crucial for understanding the genetic underpinnings of CKD-MBD.

Building on these genetic findings, we used a multipronged functional genomics approach to model dynamic and tissue-specific networks. Using NetWAS, we identified kidney-specific, bone-specific, and heart-specific gene clusters implicated in mineral ion homeostasis, as well as integrative pathways linking FGF23 to CKD-MBD. Our analysis uncovered both well-established pathways (VDR, TGFβR2, and MAPK) and novel pathways (CLEAR, pentose phosphate, and phosphatidylinositol). These findings provide critical new insights into how FGF23 contributes to CKD-MBD and related cardiovascular disease.

Our study identified the CLEAR pathway as a novel regulatory network in CKD-MBD through integrative computational analyses. The convergence of multiple independent analytical strategies underscores the robustness of this finding. Notably, the CLEAR pathway was enriched in heart tissue of hemodialysis patients with left ventricular hypertrophy, suggesting a direct role in FGF23-associated cardiovascular pathology. Furthermore, TFEB, the master regulator of CLEAR signaling, has been implicated in cardiac hypertrophy through dysregulation of lysosomal biogenesis. Given that TFEB activity is modulated by calcineurin, and cardiac hypertrophic effects of FGF23 are mediated by phospholipase Cγ-calcineurin–NFAT signaling,²⁵ it is biologically plausible that calcineurin activates CLEAR signaling to play a prominent role in FGF23-induced cardiac hypertrophy, thus expanding our understanding of the molecular interplay between mineral metabolism and cardiac dysfunction. Future studies should explore the mechanistic basis of CLEAR pathway activation in CKD and its potential as a therapeutic target.

Another novel discovery was the association between FGF23 and DNMT3A, a gene identified through multiple independent analyses. DNMT3A is linked to organismal death in the CAIN study and has been implicated in CKD through its regulation of DNA methylation, including enrichment at kidney disease risk loci.^26,27 Moreover, DNMT3A is associated with the uremic toxin 8-OH2dG, a marker of oxidative stress and DNA damage linked to cardiovascular complications.^27–31 Using in silico perturbation with Geneformer, we found that DNMT3A activation protects against HCM and DCM, providing a mechanistic basis for its potential role in mitigating FGF23-induced cardiac disease. Future studies should explore the interplay between DNMT3A, 8-OH2dG, and FGF23 in CKD-related cardiac pathology.

Our findings also emphasize the critical cross-talk between the kidney and heart, key organs in the cardiorenal axis. Fibrosis is a shared end point, and recent single-cell RNA sequencing studies suggest that overlapping pathways drive kidney and cardiac fibrosis.³² Through network analyses, we identified common pathways, including HIF1α, HNF4α, TGFβ, mitochondrial dysfunction, and Nrf2 signaling, that may mediate FGF23-driven injury across the kidney, bone, and heart. HNF4α2 isoform plays a critical role in cell death and pathogenesis of renal osteodystrophy.³³ Notably, both HIF1α and Nrf2 are inhibitors of ferroptosis, a form of iron-dependent cell death implicated in organ injury.^34,35 Given the role of ferroptosis in CKD and cardiovascular disease and well-established associations between FGF23 and HIF1α and Nrf2,^36,37 our findings suggest a plausible link between FGF23, ferroptosis, and end-organ injury.

Finally, we identified HRG-HMGB1 complex as a master regulator of networks associated with both FGF23 and cardiac disease in hemodialysis patients. HMGB1, a multifunctional protein involved in DNA repair and autophagy, is also implicated in cell death and inflammation.²³ Elevated HMGB1 levels have been linked to CKD progression and cardiovascular disease.³⁸ Our results suggest that FGF23-mediated activation of HRG suppresses HMGB1 activity, offering a potential therapeutic target for mitigating FGF23-associated pathologies.

This study has several strengths. MTAG enabled the integration of genetically correlated traits, reducing bias and enhancing power to detect novel loci associated with FGF23. Our multiplatform approach combining MTAG, NetWAS, and in silico perturbation allowed us to uncover previously unrecognized genetic and molecular networks linking CKD-MBD with cardiovascular disease. However, limitations include potential biases in GWAS summary statistics and the need for experimental validation of identified pathways and gene functions. A key concern in multistep informatic analyses is the potential for error propagation if the results are passed sequentially between different methodologies. To mitigate this risk, we performed four independent analyses—IPA, NetWAS, KEGG, and Geneformer—without transferring intermediate outputs between methods. Instead, findings were validated across multiple analytical frameworks, enhancing the robustness of our conclusions. The fact that the same pathways and genes emerged across distinct methodologies (e.g., the CLEAR pathway identified in both NetWAS and IPA, and DNMT3A found in both IPA and Geneformer) provides strong evidence for their biologic significance. Although our approach leverages the power of MTAG to enhance GWAS discovery, we acknowledge the inherent limitations of GWAS-derived datasets. By using independent downstream validation approaches, our study minimizes methodologic biases and enhances confidence in the biologic relevance of our findings. Additionally, the FGF23 GWAS, which served as the primary data for the MTAG method, included participants from the CHARGE consortium, who were not exclusively individuals with CKD or lower eGFR, is a limitation. This broader participant pool may affect the applicability of our findings to the CKD population, as genetic associations with mineral metabolism biomarkers could differ between individuals with CKD and the general population. For example, in our previous genetic study focusing specifically on individuals from the Chronic Renal Insufficiency Cohort Study, we identified genetic variants near the RGS14 and CASR genes that were significantly associated with mineral metabolism markers. Notably, the minor allele of rs4074995 (RGS14) was linked to lower FGF23 levels, as well as a lower prevalence of hyperparathyroidism. This variant was also associated with decreased RGS14 gene expression in kidney tissues. These findings suggest that certain genetic variants may have a more pronounced effect on mineral metabolism in patients with CKD compared with the general population.³ Future research should aim to investigate genetic associations within CKD-specific populations to enhance the relevance of findings for this group.

In conclusion, our study identified novel genetic drivers of CKD-MBD and provided a comprehensive catalog of pathways linking FGF23 to end-organ damage in the kidney, bone, and heart. These findings expand our understanding of the molecular mechanisms underlying CKD-MBD and offer new targets for therapeutic intervention for FGF23-associated cardiovascular disease.

Disclosures

Disclosure forms, as provided by each author, are available with the online version of the article at http://links.lww.com/JSN/F203.

Funding

C. Robinson-Cohen: Division of Diabetes, Endocrinology, and Metabolic Diseases (R01DK122075).

Author Contributions

Conceptualization: Peter A. Friedman, Farzana Perwad, Cassianne Robinson-Cohen, Larry J. Suva.

Data curation: Elvis A. Akwo, Cassianne Robinson-Cohen, Nicholas Vartanian.

Formal analysis: Elvis A. Akwo, Farzana Perwad, Cassianne Robinson-Cohen, Nicholas Vartanian.

Funding acquisition: Cassianne Robinson-Cohen.

Investigation: Farzana Perwad.

Methodology: Elvis A. Akwo, Farzana Perwad, Cassianne Robinson-Cohen, Arushi Singhal, Nicholas Vartanian.

Project administration: Farzana Perwad, Cassianne Robinson-Cohen.

Resources: Cassianne Robinson-Cohen.

Supervision: Farzana Perwad, Cassianne Robinson-Cohen.

Validation: Cassianne Robinson-Cohen.

Visualization: Cassianne Robinson-Cohen, Arushi Singhal.

Writing – original draft: Farzana Perwad, Cassianne Robinson-Cohen, Arushi Singhal.

Data Sharing Statement

Partial restrictions to the data and/or materials apply. UK Biobank data can be accessed by applying through the Access Management System on their website, and approved researchers can use the Research Analysis Platform for cloud-based analysis of the dataset. https://www.ukbiobank.ac.uk/enable-your-research/apply-for-access. https://www.ukbiobank.ac.uk/enable-your-research/research-analysis-platform. https://biobank.ndph.ox.ac.uk/ukb/exinfo.cgi?src=AccessingData. CHARGE consortium data, including GWAS summary statistics, are available via the dbGaP database under accession number phs000930, requiring authorized access approval. https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000930.v3.p1.

Supplemental Material

This article contains the following supplemental material online at http://links.lww.com/JSN/F204, http://links.lww.com/JSN/F205.

Supplemental Methods

Supplemental Table 1. MTAG results for FGF23.

Supplemental Table 2. Study population and source data.

Supplemental Figure 1. Gene ontology and enrichment analysis.

Supplemental Figure 2. KEGG enrichment analysis flow diagram for pentose phosphate pathway.

Supplemental Figure 3. KEGG enrichment analysis flow diagram for phosphatidyl inositol pathway.

Supplemental Figure 4. Canonical pathway analysis of MTAG_FGF23.