TODOs:

* To find interesting pairs of genes:

  - Look for anti-correlated features used in a model

  - Look for features that appear together in trees more than expected.

* test usage of laplace loss function

* define function to annotate model features (currently done as part of plotting)

* switch to CV using OOB of multiple bootstrapped samples? (however CV is easier than BS to understand)

* cluster features

* add text describibg model parameters (n.trees, int.depth, shrinkage)

* generate table with all value presented in graph

- add training model AUC p-val

- print training model p-val in plot

- predict all CCLE lines

- add lineage barplot in plots

- plot pred features of the same gene of the target feature.

- look for features who don't correlate well with target overall across all models (so should have low Pearson correlation but high relative importance)

- test whether better to color all bars in barplots (by value)

CCLE fixes needed:

- names for Expr$cls

- all mutation features are available for the same number of lines (no?)