Federated cross-biobank conditional analysis identifies LDL-C lowering effects of DNAJC13 haploinsufficiency and LDLR regulation

Technical Summary: Federated Cross-Biobank Conditional Analysis for Rare Variant Discovery

Problem Statement
Whole-genome sequencing (WGS) in diverse, population-scale biobanks has revealed the genetic architecture of complex traits, particularly regarding rare and non-coding variants. However, a significant challenge remains in distinguishing true causal signals from artifacts caused by linkage disequilibrium (LD) and complex haplotype structures. Standard conditional analysis methods, which typically rely on reference panels or pre-computed LD matrices, fail to scale effectively for very rare variants. These variants are often observed in only a single biobank or exhibit long-range haplotypes that reference panels cannot accurately capture, leading to a high rate of false-positive associations in meta-analyses. Furthermore, existing approaches struggle to perform conditional analysis across separate biobanks without violating strict data sharing policies regarding individual-level data.

Methodology
To address these limitations, the authors implemented a federated approach for iterative conditional meta-analysis. This methodology allows for the analysis of individual-level genotype and phenotype data across multiple biobanks without the need to centralize the raw data, thereby adhering to data sharing policies. The core of the method involves performing iterative conditioning steps across the federated network to determine conditional independence among variants.

The authors applied this framework to a meta-analysis of low-density lipoprotein cholesterol (LDL-C) levels. The study cohort comprised 614,375 individuals drawn from the UK Biobank and the All of Us Research Program, representing six distinct genetic ancestry groups. The analysis focused on both rare single variants and aggregate tests (gene-based or region-based). To validate the efficacy of their approach, the authors conducted simulations centered on the LDLR locus, a region known for complex haplotype structures, to assess the method's ability to suppress false positives.

Key Results
The application of the federated conditional analysis yielded several critical findings regarding the independence of genetic signals:

• Signal Reduction: After conditioning, the vast majority of initially significant associations were determined to be dependent on other signals. Specifically, only 4.3% of significantly associated rare single variants and 6.9% of significant aggregates remained statistically independent.
• Coding vs. Non-Coding: The proportion of aggregates that remained independent after conditioning was notably higher for coding-based tests compared to non-coding tests, suggesting that non-coding signals are more heavily confounded by LD and haplotype structure.
• False Positive Suppression: Simulations centered on the LDLR locus confirmed that the federated approach effectively suppresses false-positive associations that arise from linkage and haplotype structure artifacts.
• Novel Genetic Insights: The analysis identified specific allelic series associated with reduced LDL-C. Key discoveries included:
  • Loss-of-function (LoF) variants in DNAJC13: These were identified as having a haploinsufficiency effect that lowers LDL-C.
  • Variants in the 3' untranslated region (3' UTR) of LDLR: These variants were also associated with reduced LDL-C levels.

Significance and Claims
The paper posits that federated conditioning is a necessary advancement for multi-ancestry meta-analyses across separate biobanks. By distinguishing independent rare variant signals from the confounding effects of linkage and haplotype structure, this approach refines the genetic architecture of complex traits. The study demonstrates that without such rigorous conditioning, meta-analyses of rare variants are prone to overestimating the number of independent signals. The identification of DNAJC13 haploinsufficiency and specific LDLR regulatory variants serves as a concrete validation of the method's ability to uncover biologically relevant, independent genetic drivers of lipid metabolism that might otherwise be obscured by statistical artifacts.