generated from scverse/scvi-tools-skeleton
-
Notifications
You must be signed in to change notification settings - Fork 7
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Sebastian Birk
committed
Oct 15, 2024
1 parent
49f7ba0
commit 96857f9
Showing
4 changed files
with
28 additions
and
2 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,14 @@ | ||
# User guide | ||
|
||
## Hyperparameter selection | ||
|
||
We conducted various ablation experiments on both simulated and real spatial transcriptomics data to evaluate important model design choices and hyperparameters. The detailed results and interpretations can be found in the NicheCompass manuscript. | ||
|
||
In summary, we recommend the following: | ||
- Regarding the loss, we observed that finding a balance between gene expression and edge reconstruction is a key element for good niche identification (NID) and GP recovery (GPR) performance, while regularization of de novo GP weights is essential for GPR. The loss weights have been specified in the NicheCompass package accordingly and in most cases do not need to be changed by the user. | ||
- With respect to size of the KNN neighborhood graph, a smaller number of neighbors is more efficient in NID and de novo GP detection while a larger number of neighbors can facilitate GPR of prior GPs. Here, we recommend users to specify a neighborhood size based on the expected range of interactions in the tissue. Empirically, we observed a neighborhood size between 4 and 12 to work well. | ||
- The inclusion of de novo GPs is crucial for GPR. However, the number of de novo GPs should not be too high as important genes might be split across multiple GPs. The default in the NicheCompass package is 100 and we do not recommend to increase this number. | ||
- GP pruning can slighlty improve GPR and NID while reducing the embedding size of the model; we therefore recommend users to use the default setting of weak GP pruning. | ||
- We recommend that users define prior GPs solely based on the biology that they are interested in (as opposed to including as many prior GPs as possible). | ||
- We recommend users to use a GATv2 encoder layer (as opposed to a GCNConv encoder layer) unless performance is a bottleneck or niche characterization is not a priority and the data has single-cell resolution. | ||
- Since the use of prior GPs can significantly improve NID compared to a scenario without prior GPs, we recommend users to use the default set of prior GPs even if interpretability is not a main objective. |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters