henhaobio
Sign in

PLoS Computational BiologyBioinformaticsPeer-reviewed journal article

MAGMA: Generalized Gene-Set Analysis of GWAS Data

Christiaan de LeeuwiD, Joris M. Mooij, Tom HeskesiD, Daniëlle PosthumaiD

PDFDOICode

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text (17 Apr 2015) · CC BY 4.0

MAGMA couples principal-components regression gene tests with gene-level regression for gene sets, detecting more Crohn's Disease genes and pathways than PLINK while running far faster.

Most gene and gene-set tools summarize per-SNP associations before aggregating, which ties power to local linkage disequilibrium (LD) and weakens detection of multi-marker effects, while permutation-based gene-set tests are slow and leave their null hypotheses implicit. MAGMA instead uses a multiple regression gene model and a parametric gene-level regression for gene sets, yielding faster tests that model LD and can condition on confounders such as gene size and density.

Key results

  1. ShownPrincipal components (PC) regression gene analysis (MAGMA-main) found more Crohn's Disease genes than mean or top SNP statistics in MAGMA, PLINK or VEGAS, driven by a few genes others missed, and stayed ahead of MAGMA's SNP-wise models with stratification covariates or a 10 kilobase window.

    Bonferroni-significant genes: MAGMA-main 10 of 13172 (incl. NOD2, ATG16L1, IL23R) vs VEGAS 6 of 12455, PLINK-top 5, PLINK-avg 4, PLINK-prune 4; at p<0.05, 1203 vs 901 to 973 for SNP-wise models. With 10 PC covariates: 8 vs 4 (MAGMA-mean, MAGMA-top); 10 kb window (16970 genes): 13 vs 7 and 8. · Analysis of CD data—gene analysis; Tables

  2. ShownIn gene-set tests on MSigDB Canonical Pathways, MAGMA's self-contained analysis flagged more family-wise error rate (FWER) significant pathways than PLINK, while among competitive tools only MAGMA and INRICH reached significance; raw-genotype input outperformed SNP p-value input.

    Of 1320 pathways, self-contained FWER hits: MAGMA-main 39, PLINK-prune 27, PLINK-avg 19, MAGMA-pval-1K 4. Competitive: MAGMA-main 1, MAGMA-pval-1K 1, INRICH 3 only at the 0.0001 cut-off (57 sets tested), ALIGATOR and MAGENTA 0 at every cut-off. · Analysis of CD data—gene-set analysis; Tables

  3. SuggestedPhenotype-permutation simulations indicated that MAGMA's gene models and both gene-set tests hold nominal type 1 error; under a polygenic null only the competitive test stayed calibrated, while self-contained error rates were considerably inflated.

    Gene-level type 1 error held at the nominal 0.050 level for PC regression, summary-statistic and SNP-wise models; the polygenic null gave random SNPs effects jointly explaining 50% of phenotypic variance. Exact rates are only in supplementary tables. · Type 1 error rates

  4. ShownBecause its tests rely on known asymptotic distributions rather than permutation, MAGMA completed gene and gene-set analyses of the Crohn's Disease data much faster than raw-data PLINK and than summary-statistic tools VEGAS, ALIGATOR, INRICH and MAGENTA.

    Gene analysis: MAGMA-main 00:00:44 vs VEGAS 03:14:05 and PLINK-avg 11:35:05 (factor 947.8); with 10 covariates 00:00:58 vs over 857:54:57 for PLINK-avg. Gene-set: MAGMA-main 00:01:56 vs MAGENTA 00:24:35, INRICH 01:09:22 and ALIGATOR 02:37:11 (4 cut-offs each), PLINK-avg 44:20:40. · Computational performance; Tables

  5. ShownSelf-contained and competitive tests in one framework differ in null hypothesis: most self-contained hits showed no evidence of exceeding the polygenic background, and disabling default gene size and density covariates added competitive hits, consistent with confounding.

    Only 1 of 39 self-contained FWER-significant sets (Regulation of AMPK activity via LKB1, competitive p = 0.000026) stayed significant competitively. Without size/density correction 3 sets passed FWER, adding ECM receptor interaction (0.000015 vs 0.000094 corrected) and Cell adhesion molecules (0.000011 vs 0.0001). · Gene-set analysis; Analysis of CD data—gene-set analysis; Tables

Methods, data and reproducibility

Gene p-values come from F-tests regressing phenotype on pruned principal components of each gene's SNPs; these are probit-transformed to Z-scores and regressed on an intercept (self-contained), gene-set indicators (competitive) or other gene properties by generalized least squares using gene-gene correlations that capture linkage disequilibrium.

  • Wellcome Trust Case-Control Consortium (WTCCC) Crohn's Disease GWAS genotypes plus both WTCCC control sets, cleaned per the Anderson protocol; 403,227 SNPs · human (Crohn's Disease cases and controls) · n = 1,694 cases and 2,917 controls
  • European samples of 1,000 Genomes, used as the linkage disequilibrium (LD) reference for SNP p-value based gene and gene-set analysis (MAGMA-pval-1K) · human (European reference samples)
  • European samples of HapMap 3, used as the LD reference for VEGAS and the matched MAGMA-pval gene analysis · human (European reference samples)
  • MSigDB Canonical Pathways collection, drawn from multiple gene-set databases, used for all self-contained and competitive gene-set tests · n = 1,320 gene sets
  • Protein-coding gene annotation from dbSNP version 135 SNP positions and NCBI 37.3 gene definitions, within transcription start-stop sites or with a 10 kilobase window · human (annotation of the WTCCC Crohn's Disease data) · n = 13,172 genes (no window); 16,970 genes (10 kb window)
  • Type 1 error simulations: permuted Crohn's Disease phenotypes (global null) and a polygenic null with random SNP effects jointly explaining 50% of phenotypic variance · human (WTCCC Crohn's Disease genotypes)
Methods & tools
MAGMA (gene and gene-set analysis; standalone C++ command-line tool), PLINK (mean and top SNP-statistic gene tests, self-contained gene-set test, --assoc), VEGAS, ALIGATOR, INRICH, MAGENTA, MSigDB Canonical Pathways, dbSNP version 135, NCBI 37.3 gene definitions, Principal components regression with F-test, Generalized least squares gene-level regression, Adaptive permutation (top chi-squared SNP statistic)
Software & environment
Distributed as a standalone command-line application with C++ source code under an open source license; SNP-to-gene annotation used dbSNP version 135 and NCBI 37.3 gene definitions.
Assay conditions & compute
All analyses ran on the Genetic Cluster Computer (part of the Dutch Lisa Cluster); runtimes are reported (e.g., MAGMA-main gene analysis 00:00:44, gene-set analysis 00:01:56) but CPU, core and memory specifications are not.
Reproducibility gaps
No accession or access procedure is given for the WTCCC Crohn's Disease genotypes; the MAGMA release version is not stated; simulation designs, implementation details and exact type 1 error rates are deferred to supplementary files.
Limitations (6)
  • The authors caution that power differences between methods may partly reflect the particular genetic architecture of Crohn's Disease, so extending them to traits with other architectures requires care. · stated by the authors
  • Competitive gene-set tools agreed poorly with each other and across SNP p-value cut-offs, INRICH hits were not corrected for the four thresholds tried, and the causes of this discordance were left to future evaluation. · stated by the authors
  • Self-contained gene-set tests, MAGMA's included, showed considerably inflated type 1 error under a polygenic null, so the authors warn against relying on them for polygenic phenotypes. · stated by the authors
  • Gene and gene-set analysis from SNP summary statistics plus a reference panel had less power than raw-genotype analysis (5 vs 10 Bonferroni-significant genes; 4 vs 39 self-contained gene sets), so the authors recommend raw data whenever available. · stated by the authors
  • Power is inferred from counts of significant hits in a single Crohn's Disease cohort with no complete ground truth or replication, so it is not independently confirmed that the extra MAGMA-main genes and gene sets are true associations. · noted in this summary
  • At strict cut-offs the competitive comparators tested far fewer gene sets than MAGMA (INRICH 57, ALIGATOR 35 at 0.0001 vs 1320), so FWER-significant counts are not on a common footing across tools. · noted in this summary

Read this if you run gene or pathway analyses of GWAS data and want the regression formulation, type 1 error behavior, and power and runtime comparisons underlying MAGMA versus PLINK, VEGAS, ALIGATOR, INRICH and MAGENTA.

Citations
3,651
FWCI
36.43
Field percentile
top 1%

Abstract

By aggregating data for complex traits in a biologically meaningful way, gene and gene-set analysis constitute a valuable addition to single-marker analysis. However, although various methods for gene and gene-set analysis currently exist, they generally suffer from a number of issues. Statistical power for most methods is strongly affected by linkage disequilibrium between markers, multi-marker associations are often hard to detect, and the reliance on permutation to compute p-values tends to make the analysis computationally very expensive. To address these issues we have developed MAGMA, a novel tool for gene and gene-set analysis. The gene analysis is based on a multiple regression model, to provide better statistical performance. The gene-set analysis is built as a separate layer around the gene analysis for additional flexibility. This gene-set analysis also uses a regression structure to allow generalization to analysis of continuous properties of genes and simultaneous analysis of multiple gene sets and other gene properties. Simulations and an analysis of Crohn's Disease data are used to evaluate the performance of MAGMA and to compare it to a number of other gene and gene-set analysis tools. The results show that MAGMA has significantly more power than other tools for both the gene and the gene-set analysis, identifying more genes and gene sets associated with Crohn's Disease while maintaining a correct type 1 error rate. Moreover, the MAGMA analysis of the Crohn's Disease data was found to be considerably faster as well.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISPedroso '15Dutta '21Chu '23Ning '19Sun '19Zhang '17
59% semantic matchGenetics
Inti Inal Pedroso, Michael R Barnes, Anbarasu Lourdusamy+2 more·bioRxiv·Jul 31, 2015