Large-scale imputation models for multi-ancestry proteome-wide association analysis

Chong Wu, Zichen Zhang, Xiaochen Yang, Bingxin ZhaobioRxiv 2023 · DOI

TL;DR

This paper introduces BLISS, a novel framework that enables large-scale proteome-wide association studies (PWAS) using only publicly available summary-level protein quantitative trait loci (pQTL) data. This approach overcomes limitations of traditional methods that require individual-level proteomic data, paving the way for more comprehensive multi-ancestry analyses and drug target discovery.

What Problem Does This Paper Address?

Understanding the link between proteins and complex diseases is crucial for identifying therapeutic targets. Proteome-wide association studies (PWAS) are a powerful tool for this, but traditional methods rely on individual-level proteomic data for model training. This is a significant bottleneck due to data privacy concerns and the increasing availability of summary-level pQTL data from large biobanks. The authors address this by developing a method that can train accurate protein imputation models using only summary statistics, thus democratizing PWAS and enabling multi-ancestry analyses.

Methods

The core innovation is the BLISS (Biomarker expression Level Imputation using Summary Statistics) framework. BLISS trains protein imputation models using only pQTL summary statistics and a linkage disequilibrium (LD) reference panel. Key aspects include:

  • Summary-level training: BLISS avoids the need for individual-level proteomic data by generating pseudo-training, pseudo-tuning, and pseudo-ensemble learning datasets from summary statistics.
  • Penalized regression: A penalized regression framework is used to train models, with parameters tuned by maximizing predictive R-squared in a pseudo-tuning setup.
  • Robust pseudo-ensemble learning: BLISS combines predictions from models trained with varying LD clumping thresholds (r²) to improve robustness.
  • Super learner integration: For non-European ancestries, BLISS integrates European-trained models with smaller, ancestry-specific datasets to improve prediction accuracy.

Datasets

  • pQTL Data:
    • UK Biobank (UKB)
    • deCODE genetics
    • Atherosclerosis Risk in Communities (ARIC) study
  • GWAS Data:
    • Million Veteran Program (MVP)
    • Global Biobank Meta-analysis Initiative (GBMI)
    • FinnGen
    • IEU OpenGWAS
    • Biobank Japan (BBJ)
  • Proteomic Platforms:
    • SomaScan v4
    • Olink Explore 3072

Key Insights and Ideas

  • Summary statistics are sufficient: The study demonstrates that accurate protein imputation models can be built using only summary-level pQTL data, overcoming a major hurdle in PWAS.
  • Multi-ancestry PWAS is feasible: BLISS enables large-scale, multi-ancestry PWAS, revealing ancestry-specific protein-phenotype associations and improving our understanding of disease mechanisms across diverse populations.
  • Platform complementarity: Different proteomic platforms (Olink and SomaScan) capture distinct biological insights, and combining them provides a more comprehensive view of disease pathways.
  • Drug target prioritization: BLISS models show promise in identifying and prioritizing potential drug targets, outperforming other methods in some aspects.

Key Results

  • BLISS models achieved comparable predictive R² to standard PWAS models trained on individual-level data, especially at higher protein expression heritability.
  • The framework successfully developed large-scale European PWAS models covering 5,779 unique proteins.
  • Super learner integration significantly improved predictive performance for non-European ancestries (e.g., a 3.6-fold increase in median predictive R² for Asian models with low initial accuracy).
  • BLISS identified 32,072 significant protein-phenotype associations in the European cohort and 2,063 in the African cohort from MVP data.
  • A comparison with TWAS, MR, and colocalization showed that BLISS identified a substantial number of unique associations and a higher number of therapeutically validated drug targets.
  • Cross-platform comparisons revealed both shared and distinct associations, highlighting the complementary nature of Olink and SomaScan.

Limitations and Future Work

  • Super learner dependency: The current super learner framework for non-European ancestries requires small-scale individual-level pQTL data, deviating from the purely summary-statistic-based nature of BLISS. Future work aims to develop super learning techniques that rely solely on summary statistics.
  • Causality interpretation: PWAS associations do not necessarily establish causality due to complex genomic architecture and potential confounding factors like LD. Careful interpretation is crucial, and PWAS is best used as a hypothesis-generating tool.
  • Proteome coverage: Existing proteomic platforms do not measure the entire proteome, meaning the true causal protein might not always be captured.
  • Asian subpopulation heterogeneity: Due to limited sample sizes, Asian ancestry samples were grouped, potentially not capturing full genetic heterogeneity. Future work will incorporate larger sample sizes and more fine-grained, population-specific models as proteomic data resources expand.

Summary updated

Abstract

Abstract Proteome-wide association studies (PWAS) decode the intricate proteomic landscape of biological mechanisms for complex diseases. Traditional PWAS model training relies heavily on individual-level reference proteomes, restricting its capacity to harness the emerging summary-level protein quantitative trait loci (pQTL) data in the public domain. Here we introduced BLISS, a novel framework to train protein imputation models using only pQTL summary statistics. By leveraging extensive pQTL data from the UK Biobank, deCODE, and ARIC studies, we applied BLISS to develop large-scale European PWAS models covering 5,779 unique proteins. We further extended BLISS to integrate with small-scale non-European individual-level datasets, enabling the development of models tailored to Asian and African ancestries. We validated the performance of BLISS models through a systematic multi-ancestry analysis of over 2,500 phenotypes across five major genetic data resources. The newly identified protein-phenotype associations offer valuable insights into their cross-ancestry transferability, the contributions of different proteomic platforms, and the complementary perspectives provided by distinct genomic mapping approaches relevant to drug discovery. The developed models and data resources are freely available at https://www.gcbhub.org/ .
BibTeX
@misc{wu2023large,
  author = {Chong Wu and Zichen Zhang and Xiaochen Yang and Bingxin Zhao},
  title = {{Large-scale imputation models for multi-ancestry proteome-wide association analysis}},
  year = {2023},
  doi = {10.1101/2023.10.05.561120},
  howpublished = {bioRxiv},
}
Figure