Optimized Phenotype Definitions Increase GWAS Power

Researchers Michael Zietz, Kathleen LaRow Brown, Undina Gisladottir, and Nicholas P. Tatonetti (Cedars-Sinai Medical Center and Columbia University Irving Medical Center) published MaxGCP (maximum genetic component phenotyping) in PLOS Computational Biology. The method optimizes a phenotype definition by combining multiple observed phenotypes, such as diagnosis codes, into a single linear index that maximizes coheritability (the genetic covariance between two traits, normalized by their phenotypic standard deviations) with a target complex trait. In a real-data analysis of stroke using UK Biobank, the authors report MaxGCP boosted GWAS study power by more than 13 percent compared to conventional, single-code phenotype definitions; the method also improved sensitivity in an Alzheimer's disease analysis, with the strongest gains observed when high-quality genetic covariance estimates were available.

Unlike earlier phenotype-combination approaches, MaxGCP is phenotype-specific, has an exact closed-form solution with linear computational complexity in the number of input features, and does not require manual feature selection. It needs only summary-level genetic covariance estimates (for example, from LD score regression on GWAS summary statistics) and a phenotypic covariance matrix, rather than individual-level genotype data for every input phenotype. This matters because EHR-derived phenotype definitions typically conflate true disease genetics with healthcare-process noise, such as who gets tested, coded, or diagnosed, which dilutes the genetic signal available to association tests.

MaxGCP is released as an open-source Python package (maxgcp, installable via pip install maxgcp, with source at github.com/tatonetti-lab/maxgcp), including a command-line interface that consumes GWAS summary statistic files and LDSC reference panels directly. Genetic epidemiologists and biobank researchers working with EHR-linked cohorts (UK Biobank, All of Us, and similar resources) can apply it to existing single-code phenotype definitions without needing to re-run individual-level genotype pipelines, since it only requires covariance summary statistics as input.

The paper builds on a bioRxiv preprint circulated since mid-2024, and this is the peer-reviewed version of record; watch for independent replication in other biobanks and disease areas beyond stroke and Alzheimer's disease, and for adoption of the method in large-scale GWAS consortia seeking to extract more power from existing EHR cohorts without additional sequencing.