The study team applied Genomic SEM to published GWAS summary statistics for fasting insulin, the insulin sensitivity index (ISI), triglycerides (TG), and waist-to-hip ratio (WHR) to capture a shared genetic component related to insulin resistance. The single-factor structural equation captured a biologically coherent pattern: fasting insulin, TG, and WHR loaded positively on the latent factor, while ISI loaded negatively. Model fit was described as acceptable but not optimal using conventional fit indices; the authors reported CFI and SRMR benchmarks alongside chi-square statistics within the Genomic SEM framework. The latent factor was explicitly framed as a genetic phenotype representing shared metabolic susceptibility and not as a direct clinical diagnostic measure.
Genetic covariance and SNP-based heritability were estimated by multivariable LD score regression prior to SNP-level analyses. The input summary statistics were limited to European-ancestry GWAS, underwent uniform quality control, and excluded MHC-region variants. The study team noted that sample overlap and differences in sample sizes were considered during Genomic SEM estimation, consistent with recommended practice for cross-phenotype genomic modeling.
SNP-level association analysis of the latent factor revealed detectable SNP heritability and identified 3,067 genome-wide significant candidate SNPs based on the genome-wide threshold used in the analysis. The authors retained Q_SNP values for each locus as a complementary diagnostic of SNP-level heterogeneity relative to the common-factor model; non-significant Q_SNP values were not interpreted as proof of homogeneity.
Functional mapping and annotation using the FUMA platform defined 100 genomic risk loci from the Genomic SEM results. Gene-level association analysis using MAGMA identified 1,530 genes meeting FDR significance, with a subset surviving Bonferroni correction and subsequently used in downstream pathway and enrichment tests. The investigators compared Genomic SEM lead SNPs against the univariate GWAS of each input trait to flag lead loci that emerged specifically from the multivariable modeling strategy; lead SNPs more than 1 Mb from univariate loci were considered potentially novel according to the study’s prespecified approach.
To narrow credible variant sets, the authors applied two Bayesian fine-mapping methods, SuSiE and FINEMAP, within ±250-kb windows around lead SNPs using the 1000 Genomes European LD reference. Cross-method concordance informed prioritization: variants supported by both methods and showing higher posterior inclusion probabilities were prioritized for locus-level interpretation. Fine-mapped regions included loci near PPARG, RSPO3, SLC38A11, SLC39A8, BCL2, GDF5, and SPATA5. The manuscript makes clear that the prioritized missense substitutions later explored structurally were not GWAS lead variants nor were they prioritized as causal by the fine-mapping; the fine-mapping evidence therefore does not constitute variant-level causal proof for those substitutions.
The study team integrated expression reference models with the Genomic SEM GWAS via TWAS using FUSION and GTEx v8 cross-tissue sCCA weights. Multiple-testing correction was applied across the tested gene–panel models with a stringent Bonferroni threshold. Statistically significant TWAS signals were subjected to FOCUS for Bayesian gene-level fine-mapping to estimate gene posterior inclusion probabilities and construct credible gene sets. High-confidence candidate functional genes were then prioritized by synthesizing TWAS significance, FOCUS outputs, FUMA mapping, and MAGMA gene-level results.
This integrative pipeline highlighted genes with plausible roles in adipocyte biology, glucose–lipid handling, and insulin secretion and regulation. Among the prioritized genes mentioned in the source were PPARG, MLXIPL, TCF7L2, FAM13A, CCDC92, and SLC22A3. The authors also queried the GWAS Catalog to explore pleiotropy of lead SNPs with other complex traits.
Enrichment analyses pointed to tissue and pathway contexts for the shared genetic signal. Results implicated fat distribution and cardiometabolic trait categories, with enrichment in adipose tissue and mesenchymal cell lineages. The investigators used MAGMA and FUMA for gene-set and pathway-level enrichment, MendelVar for Disease and Human Phenotype Ontology assessments, and CELLECT/CELLEX plus stratified LDSC for cell-type specificity using Tabula Muris single-cell reference data. Partitioned heritability analyses further assessed the contribution of functional genomic annotations to the latent factor’s SNP heritability. The authors treated nominally significant cell-type signals that did not survive FDR correction as suggestive only.
A separate exploratory effort combined AlphaFold3 structural predictions, ThermoMPNN stability predictions, and molecular dynamics simulations to probe three ultra-rare missense substitutions nominated from candidate genes: BCL2-W188G, GDF5-C400Y, and RSPO3-C129Y. The study noted that the substitutions were selected based on functional-impact annotations and modeling feasibility rather than on being GWAS lead or fine-mapped causal variants. The structural assessments suggested potential local perturbations in protein conformation or stability for these ultra-rare changes, but the authors explicitly stated that these computational results do not establish mechanisms for the GWAS associations and require experimental validation.
The authors acknowledged several limits inherent to their approach. The latent factor represents a shared genetic phenotype and is not a clinical diagnostic; the input GWAS were restricted to European-ancestry datasets; and the single-factor model showed acceptable but not optimal fit, which may reflect incomplete capture of phenotypic heterogeneity. Fine-mapped variants and structural predictions do not constitute causal evidence and require follow-up functional experiments. The manuscript calls for further experimental validation of prioritized genes and for broader ancestry representation in future genetic studies to clarify generalizability and causal mechanisms.