Across all models, are certain cell painting features more explanatory than others?

[gwaybio]

Exploring model coefficients across all models, what does this distribution look like?

[gwaybio]

This paper might be an important resource Zahedi et al. 2018.

I haven't done the analysis listed above, but anecdotally, I have seen a bunch of mito features pop up with high weights.

[gwaybio]

Comparing coefficients across all 70 cell health models using real and shuffled models. The model coefficient sum is much higher in real data models. An, on average, it looks like the Mito channel is the highest compared to all other labeled channels.

Remaining Todo

• I imagine that this result will be different depending on the actual cell health variable. Stratify the type of variable and re plot

[gwaybio]

cc @AnneCarpenter @shntnu

[gwaybio]

In 9a33ac0, I compare feature performances across cell lines.

Interpretation

Not surprisingly, training with real data shows lower cell line specific MSE across features. These values are relatively consistent across cell lines as well, although it does appear that HCC44 has the lowest overall MSE.

The F statistic is tracking the ratio of between group variance over within group variance. So high values will map to features that have high performance differences across cell line. Low values indicate features that are predicted consistently across cell lines. There are some features that are predicted well across cell lines, and some that are predicted with higher variance. If the feature is predicted poorly in HCC44, it tends to have a high F stat. Not surprisingly, the F statistics are higher in shuffled data.

[gwaybio]

In #81, I add two additional visualizations:

Note that the axes represent the total sum of each individual coefficient across all models.

Top 50 Features

All Features

Summary

• Feature weights are much higher in real vs. shuffled data
• Top features don't really participate in many models
  • This is sort of expected since feature selection is not applied a priori (so there are lots of redundant features here) and the regression models are elastic net
• Not sure how to interpret coefficients with high weights!