what is ligand_tf_matrix for??

[yuzzini]

Hi. Thank you for developing this amazing tool and gathering all the experiment results into an easy accessible form.

I was trying some vignettes,
and found out "ligand_tf_matrix_nsga2r_final_mouse.rds" on zenodo, https://zenodo.org/records/7074291

What is it's difference with ligand_target_matrix?
I thought the tf_matrix would only contain only the interactions between ligand-TF/
while target_matrix(ligand_target_matrix_nsga2r_final_mouse.rds) contains all of the interactions(ligand-genes)

But opening up the ligand_tf_matrix.rds file, I found out that my guess is wrong.
I thought ligand_TF_matrix[ TG, ligand] would be 0, but there are exceptional cases,

for example

ligand_TF_matrix['Xiap','Csf1']
[1] 0.000944018

while Xiap is not TF.

what is ligand_tf_matrix for??

thank you.

[Eisuan]

Thank you for your interest in our model.

The ligand_tf_matrix is the output of the personalized page rank (PPR) algorithm. We used it to estimate signal propagation from ligands to transcription factors. In the original NicheNet paper, it is described as the ligand–gene signalling importance matrix.

To generate this matrix, we first joined the ligand-receptor network and the signalling network. Then, we ran the PPR to compute the probability of signal propagation from ligands to the other components of the signalling graph (such as the TFs). The output of the algorithm naturally displays the probability of transitioning to each node of the signalling graph, according to signal directionality. This explains the presence of probability scores from ligands to intermediate components of the signalling network, and not TFs only.

The ligand_target matrix stores the regulatory potential scores, which is the product of the matrix multiplication between the previously described ligand–gene signalling importance matrix and the weighted adjacency matrix of the integrated gene regulatory network.

Please, refer to the original NicheNet paper for intuitive visualizations of the model. I omitted the details of the model optimization steps in my answer, but you can find all the details of the NicheNet V2 model in our MultiNicheNet preprint.