"Wide" PLS with SVD and multiple categorial predictors

Question

I have a 92,000 columns (genes) x 12 rows (samples) table and would like to use PLS to model treatment effects. However, my treatments (site of origin and temperature) are categorical. If I use discriminant analysis, I can only test one at a time. If I use PLS under the Fit Model platform, I can do a full factorial predictor model, BUT there is no SVD (presumably), so the analysis crashes. Is there something akin to SVD in any PLS-like platform that is NOT discriminant analysis? Alternatively, is it possible to have more than one predictor (X) in the discriminant analysis? I attached the table. Ideally, I would have all 92,960 genes as Y and site of origin, temperature, and their interaction as predictors (3 terms). Maybe my computer is just too old (8 GB RAM with 1.7 Ghz processor)! I am wondering if what I am wanting to do is ONLY achievable with JMP Genomics (which, annoyingly, University of Miami refuses to purchase, even after my nagging!).

SHVTS standardized data.jmp

abmayfield · Answer

Good idea. I am trying it now (1 hr of "hanging" at this point). Even if I reproduce your cancer microarray output, which was ~11,000 columns x 230 rows (Using Partial Least Squares: When Ordinary Least Squares Regression Just Won’t Work ), there is a hang with discriminant analysis, but it's not bad (1-2 minutes), and that study had many more subjects. Maybe if I can weed out those (likely to be majority of) genes with essentially no variation across samples (maybe using response screening or predictor screening), I can retry PLS or Gen-reg with the "informative" genes only since I guess you could argue why would you include a gene with no variation (e.g., a housekeeping gene) in such an analysis?

P_Bartell · Answer

One of your options might be one of the penalized regression methods? They don't use SVD, but come at the variable reduction problem from a different computational/variance point of view...not dimensionality reduction which is baked into PLS. Your computer might still hang because of the number of models that will be 'fit' in the background...but worth a try?