Manc-COJO trims false positives in multiancestry genetic studies

A computational strategy for genome-wide association studies can identify nearly independent genetic signals more reliably when people of African and European ancestry are analyzed together, according to a research article in Nature Genetics. The method, called Manc-COJO, combines ancestry-specific GWAS summary statistics with ancestry-specific linkage disequilibrium references, helping researchers distinguish variants that represent separate associations from those tagging the same underlying signal. Simulations found that the approach recovered more causal variants and produced fewer false-positive selections than GCTA-COJO analyses based on European data alone or ad hoc adaptations of that software to multiancestry summary statistics. In analyses of four lipid traits involving about one million participants, Manc-COJO maintained comparable prediction in European cohorts, produced higher point estimates of prediction in African cohorts and selected fewer variants than the European-only comparison.

Conditional and joint analysis, commonly called COJO, is a standard post-GWAS procedure for asking which associated variants remain statistically informative after accounting for variants already in the model. Many variants can tag one causal signal through correlations in linkage disequilibrium, or LD. GWAS meta-analyses increasingly combine people from multiple ancestry groups, but their LD patterns differ, making it difficult for a single reference population to represent all of them. Manc-COJO addresses that problem by treating ancestry-specific GWAS summary statistics and LD references as separate inputs while making joint decisions about which variants enter the final model. The authors developed the method as a proof of principle for inferred European and African ancestry groups, although the software supports an arbitrary number of ancestry groups.

The algorithm extends the forward and backward stepwise selection framework of GCTA-COJO. In each iteration, Manc-COJO conducts an inverse-variance meta-analysis of the ancestry-specific GWAS results and selects the SNP with the smallest meta-analysis P value. It then regresses the selected SNP out of the original phenotype separately within each ancestry group. Those residuals become the phenotypes for conditional GWAS, allowing the software to test whether another variant contributes evidence beyond the variants already selected. Newly selected and previously retained SNPs are fitted jointly, after which SNPs whose joint effects are no longer significant may be removed. These conditional and backward steps continue until the model can neither add nor exclude variants. When an in-sample LD matrix is used, the authors say the procedure is equivalent to applying stepwise-forward and stepwise-backward variable selection to individual-level genotype data.

The implementation is deliberately ancestry-aware. Conditional effect estimates are calculated for each population using its own signed LD correlations, while the selection decision is informed by meta-analysis across ancestries. A variant correlated too strongly with an already selected SNP in any ancestry is excluded across ancestries, reflecting the assumption that the selected SNP is the best available representative of the causal signal in that region. The study used a prespecified genome-wide significance threshold of 5 × 10−8 and a collinearity threshold of squared LD below 0.9. Manc-COJO can take LD matrices as inputs rather than requiring individual-level genotype data from the GWAS cohort, which may make it more practical when direct data access is restricted. The method nevertheless assumes that each LD reference panel closely matches the population from which the corresponding GWAS summary statistics were generated.

Performance was evaluated in 7,600 simulations built from chromosome 22 genotypes in UK Biobank participants. The scenarios covered 38 cross-ancestry, LD-independent genomic blocks, with one, five, 10 or 20 causal variants per block and 50 replicates for each combination. Under idealized conditions, in which causal variants were genotyped, effects were equal across ancestries and LD came from the GWAS samples, multiancestry analyses recovered more causal variants and had greater statistical power than a European-only analysis with the same total sample size. Manc-COJO and a multiancestry GCTA-COJO setup selected similar numbers of causal variants on average, but Manc-COJO produced fewer false-positive discoveries, particularly in simulations involving highly polygenic regions. A European-only model nevertheless achieved good within-European out-of-sample prediction, while showing relatively weak cross-ancestry portability.

The advantage persisted when the simulations departed from those ideal conditions. The researchers varied ancestry sample sizes, allowed causal effects to differ between ancestries, removed some causal variants from the available marker set and used out-of-sample LD references. Performance declined for every method, but Manc-COJO remained more robust than the GCTA-COJO alternatives. In a scenario with 69,010 European participants and 6,901 African participants, the proportion of selected SNPs that were truly causal was 0.80, 0.64 and 0.59 for regions containing five, 10 and 20 causal variants, respectively. Corresponding values for the multiancestry GCTA-COJO analysis were 0.57, 0.50 and 0.48. Even with the strong imbalance toward European participants, adding African data therefore improved the separation of causal signals from false-positive selections in these simulations.

The simulations also exposed a clear limit to using a European LD reference for every ancestry. That shortcut performed reasonably well when European participants outnumbered African participants by ten to one, but became ineffective when representation was approximately balanced. Its false-positive rate rose sharply as the sample moved toward a one-to-one ratio. Likewise, a mixed-ancestry LD reference created from the GWAS participants could work under balanced, well-powered conditions, but produced inflated false discoveries when sample composition or underlying assumptions were unfavorable. The authors note that all COJO models assume Hardy–Weinberg equilibrium in their calculations, which may be violated in mixed-ancestry LD reference panels. They caution that the exact point at which European-only LD becomes unreliable is difficult to define and may vary by chromosomal region and genetic architecture.

Real-data tests used ancestry-specific GWAS summary statistics for high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, total cholesterol and log-transformed triglycerides. The source meta-analysis included approximately 900,000 European and 90,000 African participants. Manc-COJO was compared with a European-only GCTA-COJO analysis of comparable total size and with multiancestry GCTA-COJO analyses using either ancestry-matched or European-only LD references. Independent UK Biobank cohorts comprising 47,066 European and 3,000 African participants were used to evaluate prediction, and the researchers replicated the analysis in the All of Us cohort. In the African UK Biobank validation cohort, Manc-COJO produced higher prediction R² point estimates than the European-only GCTA-COJO model for all four traits: 0.084 versus 0.077 for HDL cholesterol, 0.137 versus 0.124 for LDL cholesterol, 0.0534 versus 0.0274 for log triglycerides and 0.116 versus 0.107 for total cholesterol. The available African validation sample was too small to establish statistically significant differences between the models.

Manc-COJO selected fewer independently associated SNPs in every comparison: 831 rather than 903 for HDL cholesterol, 554 rather than 581 for LDL cholesterol, 655 rather than 705 for log triglycerides and 700 rather than 718 for total cholesterol. Its predictive accuracy in the European validation cohort was nearly identical to that of the European-only model: 0.168 versus 0.170 for HDL cholesterol, 0.158 versus 0.159 for LDL cholesterol, 0.109 versus 0.108 for log triglycerides and 0.145 versus 0.145 for total cholesterol. These observations suggest that the smaller multiancestry model was not sacrificing useful predictive information. The results do not prove that every excluded SNP was false, but they are consistent with the simulations, which found that single-ancestry analyses can overestimate the number of independent associations. The replication in All of Us produced consistent trends, according to the study.

Functional examples provided a more granular view of how ancestry-specific LD can change variant selection. At loci containing previously supported lipid-associated variants, both methods sometimes selected different proxies. Manc-COJO’s choices were more strongly correlated with the functional variants in African samples: the absolute LD values were 0.942 versus 0.918 for rs2277862 in total cholesterol, 0.960 versus 0.476 for rs10889356 in total cholesterol and 1.000 versus 0.899 for rs10872142 in LDL cholesterol. These observations show better tagging of the known variants in the African data, but the small validation cohorts did not provide enough statistical power to establish significant prediction differences. The authors also emphasize that COJO is not a causal fine-mapping method: its outputs are models of variants that are nearly independent and may merely be close to true causal variants. A companion algorithm, Manc-COJO:FIX, allows designated variants to remain in a model during selection so that researchers can search ancestry-specific data for additional associations. In simulations, it recovered an additional 10–20% of ancestry-specific causal variants; in the lipid data, it found none in the African cohort and only small numbers in the European cohort, with negligible changes in prediction. The software was also more than 120 times faster than GCTA-COJO in benchmark analyses, but the results remain limited by imperfect LD references, assumptions about shared effects, restricted ancestry representation and the statistical power available in real cohorts.

Subject of Research: Multiancestry conditional and joint analysis of GWAS summary statistics for identifying independent genetic association signals.

Article Title: Multiancestry conditional and joint analysis applied to GWAS summary statistics

Article References: Wang, X., Wang, Y., Yang, J., Visscher, P. M., Wray, N. R., & Yengo, L. (2026). Multiancestry conditional and joint analysis applied to GWAS summary statistics. Nature Genetics. https://doi.org/10.1038/s41588-026-02799-1

Image Credits: AI Generated

DOI: 10.1038/s41588-026-02799-1

Keywords: Manc-COJO, GWAS, multiancestry genetics, linkage disequilibrium, conditional analysis, joint analysis, false-positive selection, lipid traits, genetic portability, population genomics, Multiancestry, conditional