PLS-DA and N-integration for DNAme and proteomics data

Analysis
2

Hi @Chloe.J ,

Apologies for the very delayed answer. We’ve had a change in the team. I am only answering in case it might benefit others!

Here are a few thoughts:

1. Singular matrix on plsda

With ~10 samples and 890,000 probes, the residual matrices become singular very quickly, as you experienced. I recommend you

  • filter the features by variance. Keep the top 5,000 (maximum 10,000) most variable probes. A bimodal variance distribution is common for methylation you want to keep the right-hand tail (highly variable probes). Random filtering discards any important signal so I don’t recommend this approach.

  • reduce ncomp. With N around 10, more than 2–3 components will exhaust the residual matrices.

Similar recommendations for DIABLO.

2. Cross validation when n = 9

Given that only 9 samples are shared, choose leave-one-out cross-validation in perf(). The error rate is fairly high, but not surprising given n = 9, I assume divided into several sample groups. This analysis should remain exploratory.

3. plotDiablo

I was not able to see your plot. A negative correlation between the block components is quite unusual. perhaps suggesting that the two data sets share less common discriminative signal. Try a higher design value (e.g. 0.5) or first inspect the between-block correlation with a non-sparse pls() and calculate the correlation between the X and Y components from PLS to understand the correlation structure.

Related posts:

Kim-Anh