Running pySCENIC for disease vs. control comparison with isoform-level TF dysfunction

[donghyeon321]

Hi,

I am studying a neurological disease using single-cell RNA-seq data and have a few questions about the best strategy for running pySCENIC in my context.

Background: I am interested in not just TF expression levels, but specifically which isoforms of a TF are expressed — functional vs. non-functional isoforms. In the disease group, some TFs show isoform switching where the non-functional isoform becomes dominant.

Question 1: Merging conditions for regulon activity comparison To compare regulon activity between disease and control, I am currently merging disease and control cells by cell type and running pySCENIC on the combined expression matrix. This way I can compare AUC scores for the same regulon between conditions. However, if a TF predominantly expresses non-functional isoforms in disease cells, would this disrupt co-expression patterns with its target genes — potentially preventing GRN formation for that TF altogether? If so, the TF would be absent from the regulon results, making disease vs. control comparison impossible.

Question 2: Running pySCENIC separately per condition Would it be more appropriate to run pySCENIC separately for disease and control (per cell type) and then compare the resulting regulons? Or does this introduce other issues (e.g., different regulon compositions making direct AUC comparison invalid)? Additionally, we are interested in a specific scenario: if a TF successfully forms a GRN in the control group but fails to form a GRN in the disease group — potentially because non-functional isoforms are predominantly expressed in disease cells, disrupting co-expression with target genes — would running pySCENIC separately per condition be a valid approach to capture this kind of isoform-driven GRN collapse? In other words, can the absence of a regulon in the disease condition (while present in control) be interpreted as a biologically meaningful result reflecting isoform-level TF dysfunction?

Question 3: Activating vs. repressing targets Does pySCENIC only identify activating TF-target relationships, or does it also capture repressing relationships? I noticed the 'activating' tag in the motif enrichment output — does this mean repressor regulons are excluded from the final results?

Question 4: Raw count vs. log-normalized input I have seen previous discussions mentioning that GRNBoost2 does not require transformation since it uses a Random Forest-based approach. However, I am unsure whether using raw counts vs. log-normalized counts makes a meaningful difference in practice — especially for the cisTarget step which relies on correlation-based module filtering. Could you clarify the recommended input format, and whether the choice between raw and log-normalized counts leads to substantially different regulon results?

Thank you for your time.