Question about “single-cell object” vs. prior graph (NicheNet)

[wawpaopao]

Hello!Thanks for the great work！
I noticed that in the issue @ #1
"NicheNet (recommended for general use): follow their workflow to derive a gene interaction graph from your single-cell object, or start from their pre-defined prior network."

Could you clarify what “derive a gene interaction graph from your single-cell object” means here?

In the Methods you mention integrating the NicheNet GRN and cellular signaling networks to form the gene–gene prior. That sounds data-agnostic (i.e., not tied to a specific local scRNA-seq dataset). So I’m trying to understand:
1. Is the gene graph you use purely the NicheNet prior (sig ∪ GRN, often unweighted), independent of any local single-cell data, and then only cropped/filtered per sample at training time (e.g., by non-zero genes)?
2. If you do intend the graph to be derived from a user’s single-cell object, could you outline the recommended steps to go from a Seurat/AnnData object to the required adj_t_matrix.csv and dist_t_matrix.csv? (e.g., expression thresholds for keeping nodes, whether to infer edges via co-expression/GRN, how to binarize/threshold, and how to compute shortest paths)

Thanks!

[TianzeCompbio]

Hi,

Thanks for your interest in CellNavi! Regarding your questions:

1. Yes — the usage of NicheNet in the CellNavi paper is exactly as you described.

2. While we reasoned that dataset-specific GRNs might outperform the universal NicheNet prior under certain experimental settings, we did not systematically explore the best practice for constructing such graphs. If you’d like to derive a graph from your own data rather than using the NicheNet prior, I’d recommend starting with a simple pipeline:
   a) Use a well-established GRN inference method such as SCENIC to build the graph.
   b) For binarization, you can start with a simple rule (e.g., set all values > 0 to 1). After binarizing, check the sparsity of the adjacency matrix — if it’s too dense (e.g., >10%) or too sparse (<0.1%), you may want to adjust the threshold accordingly; this is dataset-dependent.
   c) To compute pairwise distances (for the dist_t_matrix.csv), classic algorithms such as Dijkstra’s algorithm can be used. You can find an implementation in NetworkX here, or implement a custom version if needed.

We evaluated several graph choices in the CellNavi paper (see Extended Data Table 1). Among our settings, the NicheNet prior generally performed best. However, other reasonable graphs should still yield interpretable results. If you observe unstable or random outcomes, it’s likely due to other factors such as the choice of fine-tuning data or hyperparameters rather than the graph itself.

Hope this helps!

[wawpaopao]

感谢感谢！