greenelab / biobombe Goto Github PK

Switch labels for panels B and C

Need to write carefully about this point in README (see #90) and especially in the manuscript

Currently the results are being read in for all gene sets. They should be read in once, and then visualized and subset.

Answering question of how much additional information is gained through biobombe serial compression vs. lasso on 200 features

Need to add results generated in #89 to an archived resource

Related with #163

Resources include https://www.nature.com/articles/sdata201733/tables/3 and https://figshare.com/articles/STAR-reads/7613975

I don't think I need to label all facets - probably just A, B, and C is sufficient

Get scores for all top scoring features across k dimension and algorithm for two publicly available datasets.

See if the score is associated with "separation" of target samples

Can also split out "monocyte" vs other in that plot

Currently, A and B are plotted on the same row with two columns. I need to make two rows and 1 column instead

I am adding a new module 7 in #71 - i will need to update the other module numbers (GTEX and TCGA)

The colors in these figures are not adding anything - they are actually a bit confusing

After changes are merged in #125, the function plot_gene_set() will change. I will need to rerun the visualize notebook in the gtex module after the update

Need to update with strip text background color - also should make it so it can be in portrait orientation

Switch panels A and B - also, the two panels currently in A are not in the correct order

Related to #163 as was previously done in GTEx. Also, box plots can be changed to display different correlations with transformed data in both cases as well

Switch panels c and d with a and b

Need to lower case panel labels and move z to k

https://github.com/greenelab/BioBombe/blob/master/8.gtex-interpret/1.visualize-gtex-blood-interpretation.ipynb

A couple figures are redundant - added with different names

Will need to update the author list (and title) on the website once new preprint is posted

related to #181

cc @vincerubinetti

1. Need to predict with top 1 feature
2. also determine which z the features are coming from

We are interested in comparing ensemble VAE performance to ensemble multi-algorithm performance in cancertype and mutation prediction.

Probably good to map compressed features with high weights to their respective genesets

This may alleviate potential confusion between z dimension and z score language

Instead of plotting weight sum per algorithm, plot average absolute value weight per algorithm

Could be interesting to observe the correlation pattern for OV

https://github.com/greenelab/interpret-compression/blob/master/4.analyze-components/figures/TCGA/sample-correlation/sample-type/sample-correlation_OV_TCGA_signal_pearson.png?raw=true

Split out by HGSC subtype assignment

The plot generated here needs an updated y axis label. It should read: "Absolute Rank Enrichment"

Need to track Neutrophils_HPCA_2 and Monocytes_FANTOM_2 genes

SVCCA does not seem to work for the sample activation patterns in our models. I will apply SVCCA to the weight matrices instead to see if the results appear more promising.

I removed the colorblindr dependency in #13 because the package is not currently a conda recipe. Adding back this dependency will require a conda-forge pull request that I will save for a later date.

Currently, the panels are labeled by gene set, they should be lettered by model type

a more complete description of the directory tree structure will help orient a new viewer to the results.

as @ajlee21 pointed out in #56 here

Is it worth creating a lookup table with colors -- HEX code as you've done before?

It will be good to update HEX colors in a table lookup. Also related to #14

related to #108

Should add points representing raw data in panel C. What is the performance and percent zero coefficients?

Add correlation estimates for panels E and F

I have biobombe scores for many datasets by collections - plot z dimension of max feature

Related to #163 and specifically #163 (comment)

Also need to explore gene coefficients in both models

Names of main and supplementary figures need to be updated. Also files should be removed.

I will need to determine how to store these results. They are quite large and there are many of them. I am thinking some sort of figshare or zenodo link

This Figure is large. Panels G and H can be moved to a supplement.

For TCGA figure, should visualize change in ROC across k for cancer-type and mutations separately

will drive home the point that different signal is being detected at different k