SpaGCN: Integrating gene expression, spatial location and histology to identify spatial domains and spatially variable genes by graph convolutional network
MIT License
199
stars
60
forks
source link
Performance without histology better than with histology #80
First of all, thanks for the awesome tool! I (along with @symoon) been investigating using SpaGCN for one of our projects. We're successfully able to reproduce the results from the tutorial notebook, though we've noticed some surprising behavior regarding inclusion of the histology information. In particular, when we train without incorporating histology, we see better agreement with the ground truth spatial domains than training with histology included. We've created a colab notebook to reproduce this phenomenon here.
Have you noticed similar phenomena before? We were surprised to see this, as we assumed performance with histology would be better than without, but we didn't see any discussion of this in the original paper.
EasonBz
commented
1 year ago
Hi Jian,
First of all, thanks for the awesome tool! I (along with @symoon) been investigating using SpaGCN for one of our projects. We're successfully able to reproduce the results from the tutorial notebook, though we've noticed some surprising behavior regarding inclusion of the histology information. In particular, when we train without incorporating histology, we see better agreement with the ground truth spatial domains than training with histology included. We've created a colab notebook to reproduce this phenomenon here.
Have you noticed similar phenomena before? We were surprised to see this, as we assumed performance with histology would be better than without, but we didn't see any discussion of this in the original paper.
Thanks!