For a different snRNA-seq neural dataset, we are hoping to derive the disease trajectory by evaluating the continuous pseudo-progression scores (CPS). In the pre-print of the SEA-AD Alzheimer’s data the CPS has been derived from the QNP scores, using HALO. We intend to derive the trajectory from gene-expression data. The short block under the sub-heading “Continuous Pseudo-Progression Score” isn’t entirely clear to me. Like, I want to know the exact process from the data in Supplementary Table 2 to the actual CPS. Is there somewhere I can read about deriving the CPS in greater detail? So that I can hopefully understand if we can modulate it to fit our purpose or not.
We would be happy to provide detail on the process to construct CPS from quantitative neuropathology. Which parts are unclear / do you have specific questions?
More broadly, I’m not sure if modeling pseudo-progression using orthogonal quantitative neuropathology would follow the same process as modeling from transcriptomics data for 2 reasons:
(1) Cell types are not affected in the same ways at the same points in the disease process (e.g. Microglia may have an earlier and different response than a Neuron would with the same level of pathological burden in the tissue). This means you likely have to compute cell type-specific pseudo-progressions, which then become very challenging to relate to one another (we explored this space using approaches like the two other articles you cite).
(2) If you construct a pseudo-progression from your transcriptomics data you then need to be very careful when testing for changes along it to avoid artificially inflating your discovery power by double-dipping (https://stat.uw.edu/seminars/double-dipping-problems-and-solutions-application-single-cell-rna-sequencing-data).
Are you able to derive orthogonal measures of pathology? If not, another option could be to create a pseudo-progression based on cell type abundances present in the dataset and then test for molecular changes (e.g. gene expression changes) along that. It still might artificially increase your discovery power, but it would be worth exploring further.
I wanted to bump this discussion thread to see if you had published the exact code/methods used to generate the pseudo-progression score.
I looked through the AllenInstitute/SEA-AD_2024 repo but wasn’t able to find anything. I also had a bit of trouble following the methods in the publication, as I don’t have much experience with Bayesian modeling.