I would like to identify cellular pathways (KEGG) with discriminative changes in the average activity (expression levels) of their gene members in a two-condition microarray dataset. The goal is to use the average pathway activity for machine learning (sample classification), not to identify pathways enriched in differentially expressed genes (i.e. not performing a classical gene set enrichment analysis).
How do I score differential activity of cellular pathways in microarray data (not enrichment)?
You would look at downstream genes, which are selective for individual pathways - or genes which have binding sites for transcription factors that sit at the end of your pathway. Depending on the existing literature, and your experiment, and the specific pathway, this can include genes, which are themselves part of the pathway (if feedback). For interpretability you would prefer genes, which have previously been established as makers of a pathway in preceding literature.
I would like to identify cellular pathways (KEGG) with discriminative changes in the average activity (expression levels) of their gene members in a two-condition microarray dataset.
This might be tricky, since the taking the average introduces several assumptions such as: little impact of technical background (expression of most genes of pathway above background), average being representative of the KEGG pathway of interest (and not only of the most abundant pathway members), and enforcing a somewhat arbitrary decision of whether to average log-transformed data (which is justified for many genes, but not genes that don't scale multiplicatively, such as many stress genes), or not log-transforming.
The goal is to use the average pathway activity for machine learning (sample classification),
You can use this to your advantage, and create multiple features per pathway (e.g.: median, mean, variance, non/log-transformed etc..., manually -curated signature genes of pathways), and then having machine learning choose the best features for your classification (e.g.: as would happen if using Random Forest Classifiers)
You can use the PathVar software available here: www.pathvar.embl.de
See also the corresponding publication: http://bioinformatics.oxfordjournals.org/content/28/3/446.long
use the GSVA R package to collapse gene expression data to pathway activity scores. https://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-14-7