I am currently facing challenges regarding model tuning and performance evaluation, particularly because there doesn’t appear to be a tuning function explicitly designed for continuous Y variables in the mixOmics framework. I am also uncertain about the appropriate cutoff for feature selection from each block.
Hello,
Unfortunately we don’t have any performance measure / tuning method implemented for a block.spls at this stage, and even in general the one we implemented for sPLS regression is a bit unsatisfactory (described here: 4 PLS on the liver toxicity study | mixOmics vignette)
However you could probably use this as a first step to define the number of variables to select.
You can then use the functions you propose to measure the final performance.
Kim-Anh
Thank you.
To clarify, are you suggesting that we first perform single-omics analysis using the continuous Y variable through sPLS, determine the number of selected features, and subsequently apply the block.sPLS method to integrate the omics data, setting the keepX parameter based on the results from the single-omics analysis?
Best,
Arif
Hi @mislam6,
Yes, I think pre-tuning the parameters through sPLS would be a good first step. Once you move to block.sPLS you can always refine a bit (based on your evaluation of the outputs).
At this stage of our developments, it would be an exploratory analysis
PS: we have applied such approach already in https://www.biorxiv.org/content/10.1101/2024.01.30.577864v1.full Fig 2 to integrate transcripts, DNA methylation and accessibility features.
Kim-Anh