Genome-Wide Association Studies

Genome-wide association studies (GWAS) test millions of common variants for statistical association with a trait or disease across cohorts of unrelated individuals. The methodology that began as straightforward marginal regression on SNP genotypes has expanded along four axes: variant scope (moving beyond common SNPs to tandem repeats, structural variants, and rare variants that mainstream array-based GWAS misses), phenotype scope (replacing case/control labels with high-dimensional cellular or imaging-derived phenotypes), time-aware modelling (handling diseases that are characterised by an age of onset rather than a binary status), and transferability across ancestry (asking whether a score calibrated in one population retains meaning in another). Modern GWAS papers tend to advance one axis sharply rather than incrementally improving the classical recipe.

Widening the variant scope

Array-based GWAS systematically misses repeat-region variation because short reads cannot resolve tandem repeat length and arrays do not interrogate it. Ziaei Jam et al. (2023) build a deep population reference panel of tandem repeat variation from long-read and high-coverage short-read data, producing a haplotype-resolved catalogue of tandem-repeat alleles that downstream GWAS can impute into existing cohorts. The methodological move — produce a population-scale reference for the variant class your study cannot directly genotype, then impute — is the same pattern that put SNP-GWAS on the map twenty years ago, applied to a class of variants known to drive several neurological diseases that classical GWAS has historically under-explained. Benegas et al. (2023) attack a complementary problem: which variants matter functionally? Their DNA language models are trained self-supervised on reference genomes and used to score each variant’s predicted effect on its sequence context; the scores correlate with experimentally measured variant effects and with disease-association signals from independent GWAS. The methodology unbinds variant-effect prediction from explicit annotation — the model has not been told where exons or transcription factor binding sites are — and yields a single score that can be combined with GWAS summary statistics as a prior.

Better phenotypes and better time models

A GWAS is only as good as the phenotype on the other side of the regression. Tegtmeyer et al. (2024) demonstrate high-dimensional phenotyping of cellular morphology in induced pluripotent stem cells: a Cell Painting-style imaging pipeline produces thousands of morphological features per cell line, and GWAS is run on those features instead of a single trait. The result is a map from genetic variants to cellular phenotypes that bypass the noisy intermediate of clinical labels. Pedersen et al. (2023) work the other end of the phenotype problem with ADuLT, an efficient time-to-event GWAS framework: when the phenotype is “age of onset of disease X”, treating it as a binary case/control loses statistical power and biases effect estimates toward early-onset cases. ADuLT fits a liability-threshold model with explicit time-to-event structure, runs at genome-wide scale, and recovers loci that case/control GWAS misses on the same data.

Cross-ancestry transfer and tissue-specific regulation

Polygenic scores built on European-ancestry cohorts often underperform when transferred to other ancestries, and the gap is large enough to matter clinically. Kurniansyah et al. (2023) evaluate blood-pressure polygenic risk scores across race/ethnic background groups using harmonised data from multiple cohorts and quantify how much of the gap is explained by allele-frequency differences, by ancestry-specific linkage patterns, and by residual environmental confounding. The methodological lesson is sharp: a score’s portability cannot be inferred from its training-set accuracy and must be evaluated in the population it will be deployed in. Shang et al. (2023) address an adjacent gap on the regulatory side, building a high-resolution meQTL map in African Americans through the GENOA study. Their map connects genetic variants to DNA-methylation levels in a non-European cohort and exposes regulatory architectures that European-derived references missed entirely. Open methodological questions cluster around the same axes: how should GWAS infrastructure ingest variant classes (tandem repeats, structural variants, rare variants) that don’t fit the SNP-array model, can self-supervised sequence models replace bespoke functional annotations as the default variant-prior, do time-aware and high-dimensional-phenotype frameworks compose, and what is the right benchmark for the equity of a polygenic score across populations rather than its accuracy in one?