RDR: libratio fitting bias

[Rongtingting]

Bad libraty ratio fitting raise Nan emm_prob_log

GX109_5400 dataset
xconfig = xclone.XCloneConfig(dataset_name = dataset_name, module = "RDR")
xconfig.set_figure_params(xclone= True, fontsize = 18)
xconfig.outdir = out_dir
xconfig.cell_anno_key = "cell_type"
xconfig.ref_celltype = "unclassified"
xconfig.smart_transform = False
xconfig.top_n_marker = 15
xconfig.marker_group_anno_key = "cell_type"
xconfig.xclone_plot= True
xconfig.plot_cell_anno_key = "cell_type"

xconfig.exclude_XY = True
# xconfig.remove_guide_XY = True
# xconfig.guide_qt_lst = [1e-04, 0.96, 0.9999]

xconfig.display()

RDR_Xdata = xclone.model.run_RDR(RDR_adata,
            config_file = xconfig)

raise an error in CNV_optimazation step because of emm_prob_log with np.nan value, which is related to params in NB prob calculation.
In cell CTACGTCTCGGAAACG-1, library_ratio is extremely low (library_alpha is inf), which causes the badly expected layer generated in extra_preprocess

[Rongtingting]

Solve the issue by checking the used depth_key (library_ratio_capped by default) before generating the expected layer.

commit

[Rongtingting]

TODO:

1. use more strict library ratio capping strategy, need to be tested on all datasets.
2. improve llibrary ratio GLM fitting by adding init params, e.g., counts ratio and specific fixed dispersion.
   add start params
   _RDR_libratio.py
   fit_lib_ratio

[Rongtingting]

set total counts ratio as the default library size ratio is recommended.