Skip to main content
. 2022 Oct 11;23:210. doi: 10.1186/s13059-022-02761-4

Table 1.

Summary of the 15 methods we compare based on simulation studies, including Ideal, Unadjusted, and 13 variants of PCA, SVA, PEER, and HCP (Additional file 1: Section S4). Out of the 15 methods, we select a few representative methods (Section 5.2) for detailed comparison in Simulation Design 2, the abbreviations of which are shown in (D). Y denotes the gene expression matrix, Yresid denotes the residual matrix outputted by PEER, X1 denotes the known covariate matrix, and X2 denotes the hidden covariate matrix. In Line 3, PCA is run on Y directly; in Line 4, PCA is run after the effects of X1 are regressed out from Y (Additional file 1: Section S4). The addition signs in (C) denote column concatenation. “filtered” means that we filter out the known covariates that are captured well by the inferred covariates (unadjusted R20.9); this filtering is only needed when the hidden variable inference method in (A) does not explicitly take the known covariates into account

Inference method Method Response, covariates Method abbr. (if selected)
(A) (B) (C) (D)
1 Ideal Y, X1 + X2 Ideal
2 Unadjusted Y, X1 Unadjusted
3 PCA_direct PCA_direct_screeK Y, X1 (filtered) + top PCs PCA
4 PCA_resid PCA_resid_screeK Y, X1 + top PCs
5 SVA_trueK SVA_trueK Y, X1 + SVs
6 SVA_BE SVA_BE Y, X1 + SVs SVA
7 PEER_noCov_trueK PEER_noCov_trueK_factors Y, X1 (filtered) + PEER factors
8 PEER_noCov_trueK PEER_noCov_trueK_residuals Yresid, NULL
9 PEER_noCov_largeK PEER_noCov_largeK_factors Y, X1 (filtered) + PEER factors
10 PEER_noCov_largeK PEER_noCov_largeK_residuals Yresid, NULL
11 PEER_withCov_trueK PEER_withCov_trueK_factors Y, X1 + PEER factors PEER, true K, factors
12 PEER_withCov_trueK PEER_withCov_trueK_residuals Yresid, NULL
13 PEER_withCov_largeK PEER_withCov_largeK_factors Y, X1 + PEER factors
14 PEER_withCov_largeK PEER_withCov_largeK_residuals Yresid, NULL PEER, large K, residuals
15 HCP_trueK HCP_trueK Y, X1 + HCPs HCP