Leveraging a machine learning derived surrogate phenotype to improve power for genome-wide association studies of partially missing phenotypes in population biobanks

Abstract

While population biobanks have dramatically expanded opportunities for genome-wide association studies (GWAS), these large-scale analyses bring new statistical challenges. A key bottleneck is that phenotypes of interest are often partially missing. For example, phenotypes derived from specialized imaging modalities are often only measured for a subset of the cohort. Fortunately, biobanks contain surrogate phenotype information, in the form of routinely collected clinical data, that can often be leveraged to build machine learning (ML) models that accurately predict missing values of the target phenotype. However, simply imputing the missing values of the target phenotype can invalidate subsequent statistical inference. To address this significant barrier, we introduce SynSurr, an approach that jointly analyzes an incompletely observed target phenotype together with its predicted value from an ML model. As the ML model can combine or synthesize multiple sources of evidence to infer the missing phenotypic values, we refer to its prediction as a ``synthetic surrogate" for the target phenotype. SynSurr estimates the same effect size as a standard GWAS of the target phenotype, but does so with increased power when the synthetic surrogate is correlated with the target phenotype. Unlike classical imputation, SynSurr does not require that the synthetic surrogate is obtained from a correctly specified generative model, only that it is correlated with the target outcome. SynSurr is also computationally feasible at biobank scale and has been implemented in the open source R package SurrogateRegression. In a genome-wide ablation analysis of 2 well-studied traits from the UK Biobank (UKBB), SynSurr consistently recovered more of the associations present in the full sample than standard GWAS using the observed target phenotypes. When applied to 6 incompletely measured body composition phenotypes from the UKBB, SynSurr identified 15.6 times as many genome-wide significant associations than standard GWAS, on average, and did so at 2.9 times the level of significance. These associations were highly enriched for biologically relevant gene sets and overlapped substantially with known body composition associations from the GWAS catalog.

A statistical framework for powerful multi-trait rare variant analysis in large-scale whole-genome sequencing studies

“…Using 61,838 TOPMed lipids samples, it took 8 hours using 250 2.10-GHz computing cores with 12-GB memory for single-variant multi-trait analysis, which is scalable for large WGS/WES datasets. On the other hand, MultiSTAAR could be further extended to allow for dynamic windows with data-adaptive sizes in genetic region analysis 24,44 , to properly leverage synthetic surrogates in the presence of partially missing phenotypes 45 , and to incorporate summary statistics for meta-analysis of multiple WGS/WES studies 46 .…”

Phenotype integration improves power and preserves specificity in biobank-based genetic studies of MDD

“…As such, it is essential to thoroughly validate results with external data. In future work, this could be addressed with multiple imputation 53 or with downstream tests that allow different effect sizes or noise variances between imputed and observed phenotypes 54 .…”

Valid inference for machine learning-assisted GWAS

“…Additionally, POP-GWAS is both user-friendly and computationally efficient. Compared to joint models for primary and surrogate phenotypes 49 , our approach only requires three sets of GWAS summary statistics as input and completes GWAS analysis for millions of SNPs within minutes. We have also made necessary extensions to account for binary phenotypes, sample relatedness, and overlapping samples between summary statistics datasets, making POP-GWAS a versatile tool suitable for broad applications.…”