search

meyer-lab / DDMC

Clusters phosphoproteomics data based on a combination of the sequence information and abundance changes over conditions.
https://asmlab.org
1 stars 1 forks source link

Wrong normalization axis on infiltration data #358

Closed mcreixell closed 3 years ago

mcreixell commented 3 years ago

I found odd that the PCA analysis of figure 5 showed that most of the variation was in PC1 but the loadings barely showed any differences:

Screen Shot 2021-02-17 at 12 22 14 PM

That's how I realized that similarly to what was happening with the clusters as we were discussing on slack, I was zscoring this data by sample instead of by cell type... I fixed this but sadly now the multitask lasso model is no longer predictive (find updated figure below, r2=0.25). I tried PLSR too and doesn't work either.

fig5

However, I tried predicting whether a tumor is "cold-enriched" or "hot-enriched" which is also part of the data generated by xCell on Gillete et al. (Table S5) and the performance is decent:

ROC_TIIC

mcreixell commented 3 years ago

This is the updated figure 5 predicting hot vs cold tumors:

figM5

mcreixell commented 3 years ago

I'm not sure if you get notifications if you're not tagged so just in case. @aarmey

aarmey commented 3 years ago

Did this data need normalization in the first place? It sounds like you can predict the degree of infiltration, but not the relative makeup of those immune cells. That's fine.

mcreixell commented 3 years ago

Yes, the raw xCell estimates are not normalized. They're pretty large numbers. I'm closing this then.