differentialabundance
differentialabundance copied to clipboard
nf-core
Support for linear mixed effects models via DREAM
Description of feature
In #325/#339, an attempt was made to add support for limma mixed models. However, I believe this approach is not ideal for the following reasons
- Limma supports only one random effect variable
- It requires to manually combine multiple experimental factors into a single variable, therefore making it cumbersome to use
Alternatively, I propose to add support for a linear mixed effects model that generalizes to an arbitrary number of random effects and allows to specify random effect variables using Wilkinson formulas, e.g. ~ timepoint + (1 | patient_id).
An obvious candidate would be DREAM which is a generalization of limma/voom for random effects. It's certainly a good choice for gene expression data as it also does the empirical bayes shrinkage of gene dispersions.
However, there's also other models that could be condidered, e.g. MRMM, or plain lme4 models.
Depends on #362 for the Wilkinson formulae Mixed effects models are also a way of addressing technical replicates -> fixes https://github.com/nf-core/differentialabundance/issues/220
CC @apeltzer @tschwarzl @nschcolnicov @atrigila @alanmmobbs93
IMO DREAM is the prime candidate for mixed effect models on expression data.
- [ ] create module for DREAM. It should work farily similar as the limma module
- [ ] add DREAM to differential subworkflow.
@nschcolnicov @atrigila @grst @apeltzer The basic module and a preliminary position within the workflow is completed. I'll work on defining some parameters that can be used to customize the modeling in Dream (like limma and deseq2). Later, we need to update the yml structure to decide how/when to call this module with the formula/contrasts definition. I think it can be used in parallel with deseq2 when there's no random variables associated as well. Just to open debate
