registration using gene expression information

[common-chenlab]

I'm trying to get a better understanding of the objective function that is used to determine how well the image at a given point in space matches and how it could be used to take into account expression of multiple genes to obtain a more accurrate fit. Currently, most tutorials seem to be working with a single channel image of using cell density or intensity value as input (ie. a 2D matrix). However, looking at the code, I see that inputs could take a tensor. In principle, would it be possible to supply a "multi-channel" image wherein each channel corresponds to the spatial map for N genes to calculate the match? Would this work out of the box or would changes need to be made?

[kpclifton]

Hi,

Thanks for your inquiry about STalign and its applications. Yes, it should be possible to align with gene expression information because the source and target inputs to the objective functions are tensors and not limited to certain dimensions. During the development of STalign, one of my co-authors aligned with principle components of gene expression data so I believe that what you are describing with multi-channels for each gene is also possible with STalign. We just haven't shared this analysis because if we use gene expression info to make the alignment then gene expression correspondence is not an orthogonal method to test alignment performance. You are welcome explore and provide us with feedback if you identify limitations of STalign on aligning images with gene expression information.

Best,
Kalen

[common-chenlab]

We tried to apply STalign using the gene expression to improve our alignment. We used genes that label blood vessels because we can readily see how the blood vessels line up. There seems to be issues with the alignment that we’re trying to understand. 1. How dense or sparse can the image information be to get an accurate alignment? In many of your examples, images are quite dense, but here the blood vessel patterns are more sparse 2. How can you control degree of warping allowed to achieve the correct transformation? In this example, the correct solution requires that the image be warped more than potentially what the current settings are allowing it to do. 

[common-chenlab]

[common-chenlab]

Maybe to clarify my question and the guidance I'm looking for, I see that there are a number of parameters that can be edited ( epV, sigmaM, sigmaB, sigmaA,sigmaR, sigmaP) but I don't have the intuition for how to iteratively adjust the parameters, which ones will have the biggest effect, and in what order should one try to adjust the parameter. In tutorial, I see each tutorial has different parameters values. It would be helpful to know what why those values were chosen and or the logic for how you arrived at those values. For example, if you were to start with data that you have not work before, what would be first value that you try to change, then would would be the next, etc etc?

[cathalgking]

@kpclifton Can I use STalign for Visium data and a matched TIFF (mass spec data)? For the mass spec data I also have a table with m/z values and x-y co-ordinates.