alexcoul / multiomics_integration Goto Github PK

Use cells' masks and morphological descriptors

• Study the effect of selecting high variability genes on detection of small phenotypes

• Try DE genes specific that are cell types specific

Elastic Net is great to select variables, but is classically limited to linear models.
For sure there must be generalization to other non linear models like kernel SVC or multilayers perceptrons.

And look at the effect of retrieving phenotypes contribution to the computed signatures.

Zhu et al:

To validate our predictions, we first checked the expression of known marker genes and compared the average gene expression profiles between scRNAseq and seqFISH data

If we look at enough genes, aren't we sure to find at least one that validates our domain by being specific to it?

Use criteria like AIC, BIC or KIC

Subtract phenotype contributions to detect only area related genes (but first check for clusters convexity!)

They actually used

clf = svm.LinearSVC(class_weight="balanced", dual=False, C=C, \
	verbose=debug, max_iter=10000, tol=1e-4)

Aggregate statistics of more distant neighbors

First look at gene expression statistics and then maybe transform data

For the automatic version I'll need a clustering model that defines boudaries (like SVC) to be able to check in which cluster region the interpolated points fall into.

Use lower weights to compute the aggregated statistics for more distant neighbors