| 214 | | Often one is interested in how a treatment or condition alters the gene expression profile in a selected cell type. In the desirable case where multiple biological replicates are present for each condition, K. D. Zimmerman, M. A. Espeland and Carl D. Langefeld have recently [https://www.nature.com/articles/s41467-021-21038-1 highlighted] the importance of properly taking account of the correlation present in such hierarchically structured data. One strategy, the so-called pseudobulk approach, is to aggregate counts across cells from the same biological sample or subject. Mixed-effects modeling, where sample is treated as a random effect, is another strategy. The code below uses mixed effects modeling within [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-015-0844-5 MAST] and has been adapted from [https://github.com/kdzimm/PseudoreplicationPaper K. D. Zimmerman et al.] |
| | 214 | Often one is interested in how a treatment or condition alters the gene expression profile in a selected cell type. In the desirable case where multiple biological replicates are present for each condition, K. D. Zimmerman, M. A. Espeland and Carl D. Langefeld have recently [https://www.nature.com/articles/s41467-021-21038-1 highlighted] the importance of properly taking account of the correlation present in such hierarchically structured data. One strategy, the so-called pseudobulk approach, is to aggregate counts across cells from the same biological sample or subject. Mixed-effects modeling, where sample is treated as a random effect, is another strategy. The code below uses mixed effects modeling within [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-015-0844-5 MAST] and has been adapted from [https://github.com/kdzimm/PseudoreplicationPaper/blob/master/Type_1_Error/Type%201%20-%20MAST%20RE.Rmd K. D. Zimmerman et al.]. Note that the additional complexity and potential benefit of these mixed-effects models are accompanied by computational expense: fitting these models to thousands of genes in thousands of cells can be slow. A vignette outlining how to use MAST for differential expression in the more traditional fixed-effect mode (i.e. ''without'' including any random effects) can be found [https://www.bioconductor.org/packages/release/bioc/vignettes/MAST/inst/doc/MAITAnalysis.html | here]. |
| 258 | | # In the model below, "Treat" is a categorical fixed effect, while "Sample" |
| 259 | | # is a categorical random effect with intercept varying by sample. |
| | 258 | # In the model below, "Treat" is a categorical fixed effect, while "Sample" is a categorical random effect with intercept |
| | 259 | # varying by sample. |
| | 260 | # |
| | 261 | # A note on the nAGQ=0 argument for fitting the discrete component of MAST's hurdle model: |
| | 262 | # |
| | 263 | # Fitting the model involves optimizing an objective function, the Laplace approximation to the deviance, with respect to the |
| | 264 | # parameters. The Laplace approximation to the deviance requires determining the conditional modes of the random effects. |
| | 265 | # These are the values that maximize the conditional density of the random effects, given the model parameters and the data. |
| | 266 | # This is done using Penalized Iteratively Reweighted Least Squares (PIRLS). In most cases PIRLS is fast and stable. It is simply |
| | 267 | # a penalized version of the IRLS algorithm used in fitting GLMs. |
| | 268 | # |
| | 269 | # The distinction between the "fast" (nAGQ=0) and "slow" (nAGQ=1) algorithms in lme4 (to which glmer belongs) |
| | 270 | # is whether the fixed-effects parameters are optimized in PIRLS or a nonlinear optimizer. |
| | 271 | |
