A One-Shot Lossless Algorithm for Cross-Cohort Learning in Mixed-Outcomes Analysis

Ruowang Li, Luke Benz, Rui Duan, Joshua C. Denny, Håkon Håkonarson, Jonathan D. Mosley, Jordan W. Smoller, Wei‐Qi Wei, Thomas Lumley, Marylyn D. Ritchie, Jason H. Moore, Yong ChenmedRxiv 2024 · DOI

TL;DR

Researchers have developed mixWAS, a "one-shot" lossless algorithm designed to identify genetic variants associated with multiple traits (pleiotropy) across distributed databases. By integrating summary statistics rather than raw patient data, mixWAS bypasses privacy hurdles while maintaining the statistical power of a massive, pooled dataset, successfully identifying thousands of robust genetic associations across US and UK biobanks.

What Problem Does This Paper Address?

Genome-wide association studies (GWAS) have traditionally focused on "one gene, one trait." However, biology is rarely that simple. Many genetic variants exhibit pleiotropy, meaning a single mutation influences multiple, seemingly unrelated phenotypes (e.g., a variant affecting both BMI and cholesterol). Identifying these multi-phenotype associations (MPA) is key to understanding the shared architecture of complex diseases.

The challenge is twofold:

  1. Data Silos: Electronic Health Record (EHR) data is often locked in different hospital systems. Privacy regulations make pooling raw patient data into one central "mega-database" difficult or impossible.
  2. Statistical Noise: Current methods like Phenome-wide association studies (PheWAS) suffer from heavy "multiple-testing penalties"—the more traits you look at, the harder it is to distinguish a real signal from random noise. Furthermore, existing meta-analysis tools struggle with "mixed" data (a combination of continuous measurements like blood pressure and binary outcomes like a diabetes diagnosis).

Methods

The authors introduce mixWAS, a distributed algorithm that is lossless—meaning it produces the exact same results as if you had pooled all the raw data, but without ever moving a single patient record.

Key technical innovations include:

  • Composite Likelihood Framework: It models the relationship between a genetic variant and multiple traits simultaneously, accounting for correlations between those traits without needing to model the full, complex joint distribution.
  • One-Shot Communication: Each site (hospital/biobank) calculates local summary statistics (score vectors and variance matrices) and sends them to a central analyst. This happens only once, making it highly efficient.
  • Omnibus Testing: mixWAS combines two different statistical tests using the ACAT method. One test is optimized for "dense" signals (where a variant affects many traits), and the other for "sparse" signals (where it affects only one or two). This makes the tool robust regardless of how the genetic architecture is structured.

Datasets

The study evaluated mixWAS using three distinct approaches:

  • Simulated Data: Used to benchmark power and accuracy across five hypothetical sites with varying levels of trait correlation and missing data.
  • Discovery Set (eMERGE): Real-world data from seven US-based EHR sites (including Vanderbilt, Mayo Clinic, and Mass General Brigham), totaling nearly 60,000 unrelated individuals of European ancestry.
  • Validation Set (UK Biobank): A massive, independent dataset of ~500,000 individuals from the UK used to replicate the findings.

Key Insights and Ideas

  • The Power of Correlation: Unlike standard PheWAS, which treats traits as independent, mixWAS accounts for the correlation between phenotypes. The authors found that mixWAS gains the most power when the genetic effect of a variant goes against the environmental correlation (e.g., a gene that lowers LDL but is found in a population where LDL is generally high due to diet).
  • Heterogeneity Handling: EHR data is messy. Different sites have different missing data patterns and different ages of disease onset. mixWAS was designed to handle this site-level "noise" effectively.
  • Efficiency vs. Privacy: The paper proves that we don't need to sacrifice statistical rigor for privacy. Summary statistics are sufficient to reach the "Oracle" level of performance (the theoretical maximum power).

Key Results

  • Superior Power: In simulations, mixWAS consistently outperformed traditional meta-analysis and PheWAS methods, particularly in sparse signal scenarios.
  • High Replication Rate: Applying mixWAS to the eMERGE network identified 4,534 significant MPAs involving blood lipids, BMI, and circulatory diseases. An incredible 97.7% (4,428) of these were successfully validated in the independent UK Biobank data.
  • Discovery Boost: In the UK Biobank validation, mixWAS detected 18.9% more trait-SNP associations than traditional PheWAS. This suggests that many real genetic links are currently being missed by standard tools due to overly conservative statistical penalties.
  • Biological Relevance: Functional annotation showed these variants are heavily enriched in pathways for cholesterol metabolism and lipoprotein function, providing a "sanity check" for the results.

Limitations and Future Work

  • Missing Data Assumptions: The algorithm assumes that data is "Missing Completely at Random" (MCAR). If data is missing because of the disease state itself (Missing Not at Random), the results could be biased.
  • The "Null" Phenotype Problem: While mixWAS is powerful, its strength can be diluted if too many "null" phenotypes (traits totally unrelated to the gene) are included in the analysis.
  • Communication Overhead: While much more efficient than moving raw data, mixWAS requires sharing more complex summary statistics than a simple meta-analysis. The authors note, however, that these costs are still orders of magnitude lower than full data transmission.
  • Future Directions: The authors suggest scaling the approach to even larger "phenomes" and exploring its utility in non-European populations to ensure global genetic equity.

Summary updated

Abstract

In cross-cohort studies, integrating diverse datasets, such as electronic health records (EHRs), is both essential and challenging due to cohort-specific variations, distributed data storage, and data privacy concerns. Traditional methods often require data pooling or complex data harmonization, which can reduce efficiency and limit the scope of cross-cohort learning. We introduce mixWAS, a one-shot, lossless algorithm that efficiently integrates distributed EHR datasets via summary statistics. Unlike existing approaches, mixWAS preserves cohort-specific covariate associations and supports simultaneous mixed-outcome analyses. Simulations demonstrate that mixWAS outperforms conventional methods in accuracy and efficiency across various scenarios. Applied to EHR data from seven cohorts in the US, mixWAS identified 4,530 significant cross-cohort genetic associations among traits such as blood lipids, BMI, and circulatory diseases. Validation with an independent UK EHR dataset confirmed 97.7% of these associations, underscoring the algorithm's robustness. By enabling lossless cross-cohort integration, mixWAS improves the precision of multi-outcome analyses and expands the potential for actionable insights in healthcare research.
BibTeX
@misc{li2024one,
  author = {Ruowang Li and Luke Benz and Rui Duan and Joshua C. Denny and Håkon Håkonarson and Jonathan D. Mosley and Jordan W. Smoller and Wei‐Qi Wei and Thomas Lumley and Marylyn D. Ritchie and Jason H. Moore and Yong Chen},
  title = {{A One-Shot Lossless Algorithm for Cross-Cohort Learning in Mixed-Outcomes Analysis}},
  year = {2024},
  doi = {10.1101/2024.01.09.24301073},
  howpublished = {medRxiv},
}
Figure