Interpretation of project_cell_annotation output about tangram

[lewlin]

Ciao edward and apologizes for the delay in responding (holidays and such).

project_cell_annotation simply transfers any annotation from the scRNAseq data to space. Most times, we use it to transfer to the cell type probabilities stored in tangram_ct_pred. These values are useful on a relative scale, not absolute. As you notice, when we plot them, the plotting function standardizes to 0-1 scale irrespective of the scale of each annotation (cell types). We wanted Tangram to do that.

The meaning of plotting tangram_ct_pred in space is to tell you where your cell type is likely to be within the ROI,

[Nh-Zhu]

Hello, how do I get the most likely cell type for each voxel from tangram_ct_pred. Thanks!

[lewlin]

The simplest way is to fetch the most probable cell for each voxel (which assumes one cell per voxel).

To get a more accurate histogram, you can sample one cell from the probability distribution of each voxels: in this way, you do not get only the most probable cell, as you sample proportionally to the cell probability.

[cuicathy]

The simplest way is to fetch the most probable cell for each voxel (which assumes one cell per voxel).

Thanks for your clarification. If each voxel only has one cell, do you mean that we can use np.argmax(adata_MxIF.obsm["tangram_ct_pred"].values, axis=-1) to assign a label for each voxel with the class with the largest probability? Thanks!

[lewlin]

Short answer: yes.
Long answer: yes but be careful.

Let's say that you have 4 voxels (containing 1 cell each), each of them is 75% prob. of being a cancer cell and 25% of being an immune cell.

How many immune cell do you expect? I would say 1 - but using argmax will give you all cancer cells.

Despite the above argument, we still use argmax to do assignment. But if you asked us to build an histogram of cell types, we would instead draw a random binary label with 75%/25% probability to get a more accurate estimate. We don't do this for deconvolution b/c we have no way to know "which" of the four is an immune cell. For more advanced cases, we compute cell type statistics by argmax vs by random sampling from probability, and compare that to estimate Tangram uncertainty in deconvolution.

The "right" way of doing is to have a more robust method for deconvolution, which perhaps even works w/o using segmentation. Well - that's Tangram2 which we are currently building :)