This is an incompletely framed but I think very promising analysis approach we could explore and recommend to the world. There’s probably a dissertation in here, but I think we could do something meaningful in a few days.
Background:
Say you have a differential expression question you're interested in, e.g. "how do tumor cells respond to T-cell proximity?" You could run a giant model across all the tumor cells in your study and get the average tumor cell's response. But that approach has two big problems:
- It is subject to confounding from all manner of other spatially-varying factors. Biology is so complex it's all but futile to specify the full, correct model, and until you do so, confounding will always threaten your DE results.
- It would miss all the ways the relationship in question varies across the tumor. Presumably there are conditions where tumor cells respond strongly to T-cell proximity, and conditions where they barely register it, or respond with different pathways. A thorough analysis would detail these conditions separately.
Proposal: spatially-stratified DE:
In many cases - especially in tumors, where heterogeneity is so pronounced - we could run differential expression in tiny patches across the span of the tissue(s), obtaining DE results across dozens or hundreds of distinct spatial regions. These patches can then be input to meta-analysis techniques. Another benefit is that this permits analysis across massive datasets, since cells only need to analyzed in small groups.
I see two basic approaches:
Enticing problems to solve along the way:
- What's the best way to define patches for DE? Ideally they'll have variability in your predictor variable but be pretty low-variance in potential confounders.
- Cluster genes by their DE patterns across patches: can we find gene modules with correlated DE results?
- How do you perform metaanalyses that can discover predictors of DE relationships? E.g., say tumor cells only express B2M near T-cells when they're at the tumor-stroma interface... can we define methods to detect that kind of relationship?
Nomenclature:
- index cells i, genes j, patches k
- call gene expression y_{i,j}
- call the predictor variable(s) X_{i}
- call a patch's estimate and SE from DE beta_k and SE_k
- call Z_k some vector of relevant patch attributes (e.g. the number of each cell type in the spatial neighborhood around the patch)
- then there's probably some model where beta_k = f(Z_k)
This is an incompletely framed but I think very promising analysis approach we could explore and recommend to the world. There’s probably a dissertation in here, but I think we could do something meaningful in a few days.
Background:
Say you have a differential expression question you're interested in, e.g. "how do tumor cells respond to T-cell proximity?" You could run a giant model across all the tumor cells in your study and get the average tumor cell's response. But that approach has two big problems:
Proposal: spatially-stratified DE:
In many cases - especially in tumors, where heterogeneity is so pronounced - we could run differential expression in tiny patches across the span of the tissue(s), obtaining DE results across dozens or hundreds of distinct spatial regions. These patches can then be input to meta-analysis techniques. Another benefit is that this permits analysis across massive datasets, since cells only need to analyzed in small groups.
I see two basic approaches:
Enticing problems to solve along the way:
Nomenclature: