General Recommendations for preprocessing RNA-Seq data about carnival HOT 2 CLOSED

Hello

1. see the tutorial on how to produce CARNIVAL inputs, i.e. estimating TF activity: https://saezlab.github.io/decoupleR/articles/tf_bk.html

We are not using all DE genes as "measurements" in CARNIVAL. CARNIVAL models how sinaling leads to the regulators of differential gene expression, ie. to transcription factors. This is also why our prior knowledge network is a protein-protein interaction network and not a gene regulatory network.

2. About the solvability of problems:
   I cannot think of a case when the CARNIVAL optimization would lead to infeasibility, i.e. when some constraints are contradictory. I guess what you observe is the numerical issues with CBC. We also observed strong limitation with the non-commercial optimizers and thus we recommend Gurobi or cplex that are free in academia.

If you would work with this in a non-academic setting, I would suggest to try to eliminate nodes and edges from the prior knowledge network. We used a couple of strategies in cosmos (https://github.com/saezlab/cosmosR), e.g.
(1) remove nodes from the network that are not expressed in any condition based on the RNAseq data
(2) remove nodes that are not connected with a directed path to the measurement (they are not observable anyways)
(3) remove nodes that are further than N number of steps from the inputs/measurements. E.g. N = 8. The rationale here is that the more steps we do on the network the less likely that our prediction makes any sense.

From these, mostly (3) has the strongest impact on the computational time. Maybe you need to go down to N=5 for CBC.

from carnival.

Thank you for this very clear answer! This is all very helpful.

from carnival.