What algorithms exist for measuring gene coexpression in transcriptional space?
My objective is to obtain gene coexpression information that is not influenced by overrepresented cell types and is also sensitive to genes at either end of the transcriptional space occupancy continuum (i.e. housekeeping genes vs. cell type marker genes). These types of genetic interactions are poorly recovered (if at all) by traditional correlation-based algorithms such as WGCNA.
As an inelegant thought experiment to solve this problem, consider projection of single cell transcriptomes onto n-dimensional space using n principal components. Now define a volume of space filled with pixels of size p containing at least c cells and the average expression of genes across all cells in that pixel. p and c are parameters that can be optimized by k-fold cross validation against robustness of resulting gene similarity matrices. Assume zero-inflation is not an issue, because cells in this dataset are massively integrated from large experiments. The probability of mutual coexpression of any gene pair A and B with overlap (A∈B) is given by the simple probability of ((A∈B)/A)*((A∈B)/B).
To be clear, I am not asking how to define axes of transcriptional space. I am asking how to best measure gene coexpression in that space independent of cell density and dependent only on spatial scales and averaged gene expression throughout that space.
Any insights, conceptual or technical, are appreciated!