sPLS model validation

1

Hi all,
I read the manual about sPLS with the example here sPLS Liver Toxicity Case Study – mixOmics

My understanding is that this model validates the optimal number of components through cross validated Q2.total and achieved feature selection through lassso penalization applied to the loading vectors ( maximizes the cross-validated correlation between the X- and Y-variates). And my current problem is the total Q2 is always below their suggested threshold 0.00975. Which means its predictivity is low. And I think the reason is there are too many features and only 5 samples.

image
image871×233 54.5 KB

I want to ask can I still use the fitted model to visualize the association between two datasets? Can I still extracted the selected features to explain the association between two datasets? Or does it means I can not use the fitted sPLS anymore?

Best regards,

Xin

2

Hi @xin

Apologies for the delay as we are without a maintainer at the moment.

I dont think it is worth considering the Q2 value if you have only 5 samples, as this is too low for robust estimation. You can still use the outputs of the sPLS on 2 components (for visualisation), and the sparse version to identify potential associations. But be aware that with 5 samples you can only explore and describe the data, not infer or generalise the results.

Kim-Anh