Blood Metabolic Traits in 610,000 People: The Value of Low-Frequency Variants and Limits of BCAA Targeting
Background: The gap between metabolite associations and disease causation
A study analyzing blood metabolic traits in up to 619,372 people demonstrated the value of low-frequency genetic variants for functional interpretation. At the same time, current genetic evidence did not support the hypothesis that targeting branched-chain amino acid (BCAA) degradation to lower blood BCAA levels would substantially reduce the risk of type 2 diabetes. Published in Nature, the findings show why associations between metabolites and disease cannot be interpreted directly as evidence of a therapeutic target’s effectiveness.
Genome-wide association studies (GWAS) of complex diseases already include more than one million participants, but studies of molecular traits have had comparatively smaller sample sizes. Existing nuclear magnetic resonance (NMR)-based studies of metabolic traits have also focused mainly on common variants. Larger samples increase the chance of identifying low-frequency variants. However, pleiotropy—where the same variant affects multiple traits—can make links to disease more difficult to interpret.
Key findings: Functional clues from low-frequency variants and BCAA analysis
The researchers meta-analyzed data from 185,352 participants in the Estonian Biobank and 434,020 in the UK Biobank. Across 249 NMR-measured metabolic traits, they identified 88,127 common- and low-frequency-variant locus–trait associations spanning 8,398 loci. This count refers to associations between genomic locations and metabolic traits; it does not represent the number of independent disease-causing variants.
Low-frequency variants were defined as those with a minor allele frequency of 0.1–1%. Among variants classified by statistical fine-mapping as having a high probability of being causal candidates, 19.4% were low-frequency variants. Of these low-frequency variants, 23.1% were predicted to cause amino acid substitutions or changes in RNA splicing, compared with 11.5% of common variants. RNA splicing is a process that modifies RNA transcribed from a gene. These computational predictions must be distinguished from experimentally confirmed functional effects, but they can help narrow down the genes and mechanisms to examine in follow-up experiments.
The disease analyses differed depending on the method used. A genome-wide Mendelian randomization (MR) analysis, which uses genetic variants as instruments, found a significant signal between BCAAs and type 2 diabetes. However, a signal combining effects from multiple pathways is difficult to interpret as the causal effect of a specific pathway. The researchers therefore reassessed BCAA degradation using cis-MR, which focuses on variants near target genes.
There were insufficient variants around the target of interest, BCKDK, to provide adequate statistical power for direct analysis. Instead, the researchers analyzed variants near BCAT2, DBT, and PPM1K in the same pathway. They concluded that current genetic evidence does not support the hypothesis that inhibiting BCKDK would substantially lower diabetes risk. This inference comes from genetic data on related pathways and should not be taken as establishing that a drug is clinically ineffective.
Meaning and outlook: Assessing targets while interpreting causality cautiously
The study provides a public resource for interpreting genetic associations with metabolic traits and evaluating candidate drug targets. The researchers made summary statistics available through the GWAS Catalog and provided access to association results and colocalization analyses examining whether different traits share the same genetic signal. The accuracy of individual disease-risk prediction and the performance of clinical screening tests were outside the scope of this study’s validation.
Generalizability is also limited. The sample was predominantly of European ancestry, accounting for 97% of participants, and fine-mapping was limited to the European-ancestry subgroup of the UK Biobank and variants with a minor allele frequency above 0.1%. Measurements were also limited to 249 NMR-based traits, mainly related to lipoproteins.
The authors cautioned against interpreting genome-wide MR results as evidence of direct causal effects because of pleiotropy. In cis-MR, linkage disequilibrium, in which variants are inherited together, and the biology of the metabolic pathway must also be considered. The study design, results, and limitations are described in the publicly available full text.
Nature, Published online: 20 May 2026. DOI: 10.1038/s41586-026-10532-5
Summary: Bypassing the statistical power limitations of traditional genomic cohorts, this landmark study integrates multi-centric omics metadata from the Estonian Biobank and the UK Biobank to construct a massive discovery architecture of over 610,000 individuals. The metaanalysis successfully maps hundreds of novel common and low-frequency locus–metabolic trait associations across human biofluids. Crucially, leveraging large-scale Mendelian Randomization to eliminate environmental confounding and reverse causation, the framework solidifies putative causal links between discrete fatty acid or glucose regulatory pathways and complex cardiovascular or diabetic outcomes, delivering a robust, low-noise computational baseline for programmable disease risk stratification.
Research teams evaluating BCAA-lowering drugs cannot expect a preventive effect solely on the basis of the observational association between blood BCAA levels and diabetes. This analysis shows why links to disease risk need to be assessed for the specific pathways that alter metabolite levels, as well as for metabolite concentrations themselves. With the insufficient statistical power for direct analysis of BCKDK in mind, genetic results from related pathways can inform follow-up functional experiments and the design of drug evaluations. Functional predictions for low-frequency variants and public association data are also useful resources for narrowing down experimental candidates. However, functional predictions require experimental validation. These data alone cannot determine a patient’s risk or treatment effects.