Different results from CLI and pySCENIC

[SeZuNo]

Hi everyone,

We are urrently working on our 4th project using SCENIC. So far we have obtained good results, thank you for this amazing tool!

The usual workflow has been to use the CLI tool in a pipeline and after explore the results in a script: binarize, generate regulon tables, RSS, binary UMAPS, both for precalculated UMAP/tSNE in parallel Seurat analysis as well for auc_mtx UMAP, generate case and control comparatives if the experimental design allows and finally the clustermap for regulons vs cell clustering.
Roughly:
From raw counts -> Filter cells -> Format dataset in loom file -> GRN (a)-> CTX (b) -> AUCELL (c) -> Script for SCENIC downstream.

Currently testing the other way proposed in the most recent paper, with the functions:
grnboost2() (a) -> modules_from_adjacencies() -> prune2df() -> df2regulons() (b) -> aucell() (c) -> Script for SCENIC downstream.

1. Step (a) gives the same results both from CLI as well as the script.
2. Step (b) gives different results both in format and content.
   Note about the format of the output:

• ctx from CLI outputs a .csv where different modules (TFs - targets) with their motif ID can be observed together with other 7 Enrichment metrics. Where no direct regulon list can be observed (I'm assuming it is derived by merging all TFs-Targets modules)
• On the other hand, from the sequence of functions to obtain regulons from the adjacencies in the script only the regulon list with importance values is stored (prune2df() function stores the rest of the metrics but this is not exported in the tutorials)

3. Step (c) gives same number or different number of regulons depending on the dataset but altogether different sets of regulons

• Both analysis are masking the dropouts.
• Both analysis are using only 500bpUp_100bpDown database for hg38 v10. (I have asked before about the database selection here: How to choose the ranking database pySCENIC#374 (comment)_ Maybe there's some updates on that?)
• Both analysis are using same annotations table v10.
• Both analysis have been repeated and results are consistent between runs.

So, apparently Step (b) which would be Module generation with Motif enrichment and TF-regulon prediction (cisTarget step) has differences between pySCENIC CLI and pySCENIC script, how come?

Hope it's clear the question,

Thank you so much,

Sebastián Zúñiga.

here's my session_info()

anndata     0.8.0
scanpy      1.9.2
-----
PIL                         9.4.0
attr                        22.2.0
boltons                     NA
cffi                        1.15.1
click                       8.1.3
cloudpickle                 2.2.1
ctxcore                     0.2.0
cycler                      0.10.0
cython_runtime              NA
cytoolz                     0.12.1
dask                        2023.2.0
dateutil                    2.8.2
defusedxml                  0.7.1
dill                        0.3.6
distributed                 2023.2.0
frozendict                  2.3.5
fsspec                      2023.1.0
h5py                        3.8.0
heapdict                    NA
jinja2                      3.1.2
joblib                      1.2.0
kiwisolver                  1.4.4
llvmlite                    0.39.1
locket                      NA
loompy                      3.0.7
markupsafe                  2.1.2
matplotlib                  3.7.0
mpl_toolkits                NA
msgpack                     1.0.4
multiprocessing_on_dill     3.5.0a4
natsort                     8.2.0
numba                       0.56.4
numexpr                     2.8.4
numpy                       1.23.5
numpy_groupies              0.9.20
packaging                   23.0
pandas                      1.5.3
pkg_resources               NA
psutil                      5.9.4
pyarrow                     11.0.0
pycparser                   2.21
pyparsing                   3.0.9
pyscenic                    0.12.1
pytz                        2022.7.1
scipy                       1.10.1
session_info                1.0.0
setuptools                  65.5.1
six                         1.16.0
sklearn                     1.2.1
sortedcontainers            2.4.0
tblib                       1.7.0
threadpoolctl               3.1.0
tlz                         0.12.1
toolz                       0.12.0
tornado                     6.2
tqdm                        4.64.1
typing_extensions           NA
yaml                        6.0
zict                        2.2.0
zoneinfo                    NA
-----
Python 3.10.6 | packaged by conda-forge | (main, Aug 22 2022, 20:36:39) [GCC 10.4.0]
Linux-3.10.0-1160.83.1.el7.x86_64-x86_64-with-glibc2.17