Matrix eQTL: Ultra fast eQTL analysis via large matrix operations

Andrey A. Shabalin,

Department of Biostatistics, University of North Carolina at Chapel Hill

... just shortened my computing time from a year to a couple of days
Adaikalavan Ramasamy, Ph.D., King's College London

Key features Designed for eQTL analysis of large datasets. Performs testing for each transcript-SNP pair. Ultra fast without loss of precision. Equally fast for models with covariates. Supports Linear additive and ANOVA models. Includes tests for genotype-covariate interaction. Covariates to account for sex, population structure, surrogate variables, etc. Correlated and heteroskedastic errors. Correction for multiple testing using FDR . Separate FDR for cis- and trans- eQTLs (more info). Matlab and R implementations. Performance analyzing CF dataset with 573,337 SNPs and 22,011 transcripts over 840 samples. Tested on a quad-core PC, using additive linear models with zero and with 10 covariates. Method No covar. 10 covar. Matrix eQTL, Matlab 11.8 11.8 minutes Matrix eQTL, Rev R 14.6 14.6 minutes Matrix eQTL, R+GOTO 19.4 19.4 minutes Plink 9.4 583.3 days Merlin 19.6 20.0 days R/qtl 1.0 4.7 days snpMatrix 3.2 5.1 days eMap 17.8 N/A days FastMap 10.3 N/A hours Info: Details of the testing procedure. Note: Matrix eQTL results match those from other software.

Manuscript in Bioinformatics (2012) R code Matlab code Manual / FAQ   Sample dataset of CF data size (7zip archive ). Questions, comments, concerns? Contact me: Andrey A. Shabalin. Support Matrix eQTL

Update history

• 2012, Feb 12 — Fixed an incompatibility with R 2.12.0. (It does not have function simplify2array yet.) Version 1.5.2.
• 2012, Feb 12 — Fixed an incompatibility with some versions of R. Version 1.5.1.
• 2012, Feb 5 — MAJOR UPDATE (R version only). Use update.packages() to update the package to version 1.5.0.
  • Added new mode: useModel = modelLINEAR_CROSS to test for SNP-covariate interaction.
  • Added option to plot histogram of p-values for (all/local/distant) tests (including not significant).
  • Greatly improved performance of cis-eQTL analysis, added the option for cis-only analysis.
  • Memory consumption is reduced by 50% when the genotype matrix is integer.
  • Added methods to SlicedData class, including SaveFile, CreateFromMatrix, RowReorder, and FindRow.
  • Added generic methods for SlicedData, including colnames, dim, length, ncol, nrow, rownames, show, and summary.
  • The code is redesigned for easier addition of new models.
  • Some bugs fixed.
  
  
• 2011, Dec 9 — Updated R/qtl timing in the table. The previous tests used a suboptimal parameter (method = 'mr' instead of method = 'hk'). My apologies to all R/qtl fans.
• 2011, Oct 6 — Major Update. Use update.packages() to update the package in R.
  • Added FDR calculation.
  • Memory required for ANOVA model is reduced.
  • Added option of separate treatment of cis- (local) and trans- (distant) eQTLs. (Matrix_eQTL_engine_cis.r, R only for now.)
  • The eQTL function does not change the input matrices now (snps, gene, and cvrt).
  • Fixed R code to avoid the ``Design Flaws in R #1 — Reversing Sequences'' .
  • Replaced my implementation of Gram–Schmidt process with a simple call of QR factorization.
  • Replaced !is.finite with is.na.
  • Several minor bugs fixed.
• 2011, Aug 14 — Added a column with the test statistic to the output file (t-statistics show the direction of the effect).
• 2011, Aug 14 — Constant genes and those colinear to the covariates are now removed from the analysis.
• 2011, Aug 14 — Fixed mistreatment of missing values in the R version (is.nan is replaced with !is.finite).
• 2011, Jul 7 — R code is now available as MatrixEQTL package through CRAN.
• 2011, Jul 7 — Improved performance of SlicedData.ResliceCombined in Matlab.
• 2011, Jun 21 — Added support for input files with no row and/or column labels.
• 2011, Jun 19 — Added various bound checks, corrected treatment of pvOutputThreshold = 1.
• 2011, Jun 5 — Added large sample dataset (7zip archive, 1.9mb packed, 1.15gb unpacked).
• 2011, Jun 4 — Fasten creation of ANOVA dummies in R code.
• 2011, Jun 4 — Reduced memory consumption in both implementations.
• 2011, Jun 1 — Due to a performance bug in R, switched from using list to environment for data storage.
• 2011, May 31 — Fixed slow data preprocessing (for large data) in R code.
• 2011, May 30 — Introduced R code for Matrix eQTL.
• 2011, May 26 — Introduced Matlab code for Matrix eQTL.