Best choice of mode

[rocketeer1998]

Hi @qiaochen @huangyh09 , thank you for your awesome tool tranSpa which not only imputes missing genes for ST but provides corresponding uncertainty estimation. I'd like to test it on my data. However, I'm a little confused about how to choose the best mode from four choices (TransImpLR, TransImpCls, TransImpSpa, TransImpClsSpa)? According to your manuscript, TransImpLR seems to have more genes with low uncertainty than that of the other three methods (Figure. 2E). Likely, according to your documentation), the CSS of TransImpLR seems always higher than that of the other three methods in each fold. So can I conclude that using TransImpLR is the best choice?

[qiaochen]

Hi, thank you for you interest in our work, hope the tool can be useful to your research.

TransImpLR is a conservative option to start with, while TransImpCls can be a good tractable method when data volume is large and cells can be relatively well clustered e.g. with Leiden method.

Also mentioned in our manuscript is the issue of overestimated spatial patterns (you may hold out a ST gene set and compare Moran's I indices on observed vs imputed genes in this set after training on the rest genes). If there is a need to counteract this overestimation, then the spatially-regularised modes: TransImpSpa and TransImpClsSpa can be the way to pursue.

Overall, TransImpLR can be a strong baseline in terms of CSS scores, while you can also do something like k-fold cross validation to empirically measure the performances of different modes (together with other methods) based on different metrics of interest.

Best,
Chen