DeLeakage preprint journal club slides

Transcript

1. Correcting spatial transcriptomics data affected by a prevalent transcript leakage problem across platforms, species, and tissues Kevin Yip Lab Director of the Center of Data Science and A.I. Sanford Burnham Prebys Medical Discovery Institute Journal club 7/8/26
2. • Key observation: across spatial transcriptomics datasets, transcripts were detected in cell-types that they are not expected to be expressed. • Paper claims to show this is due to ‘transcript leakage’, transcripts from neighboring cells essentially spread out spatially • Genes vary in their leakage properties, necessitating a leakage- specific computational method
3. • Specifically in non-neurons, looking at Slc17a7 (VGLUT1) expression • Specificity of marker gene expression is notably higher in single cell data, more fuzzy in the spatial data Figure 1
4. • Identified pairs of mutually exclusive expressed genes across pairs of cell-types from the companion scRNA- seq data to the MERFISH dataset • Showing transcript co-detection is a general feature across all cells, so not likely to be driven by incorrect annotations of the spatial data Figure 1
5. Figure 1 • As you get further away from a particular cell-type, you have less captured expression of that cell- types marker gene in spatial cells • ‘Endogenous’ is marker expression from the expected cell-type, looking at the exclusively expressed markers from the single-cell data
6. • Not a cell segmentation issue, tried three additional cell segmentation approaches and fuzzy spatial marker gene expression was maintained across the approaches Figure 1
7. • Focusing on the Astro- Ependymal cells • Those close to Ex neurons have high expression of Ex marker • Same for GABA and Oligos • Spatial variation of potential ‘contaminating’ genes follows the spatial pattern of potential ‘contaminating’ cell-types, rather than any spatial structure of the cell-type of interest Figure 1
8. Figure 1 summary • Exclusive cell-type expression observed in single-cell data is weaker in spatial data, off-target marker genes are detected at much higher rates • This relationship also seems dependent on spatial proximity • Seems pretty convincing to me
9. Figure 2 • Not a MERFISH specific phenomenon
10. Figure 2
11. Figure 3: DeLeakage approach • Decompose transcripts identified per cell into ‘endogenous’ and ‘leaked’ categories, based on a diffusion process • Each gene has its own diffusion parameters • Produce leakage-corrected gene counts per cell • The exact math is difficult to parse through, but sounds like they are making reasonable assumptions • Interestingly, also accounts for cell size, assuming larger cells contribute more to leakage • You don’t need cell-type labels, but I think you do need to supply the expected number of cell- types
12. Figure 3 Comparing DeLeakage to conventional deconvolution, where deconvolution is essentially a mixture model of different cell- types. DeLeakage is a decomposition of endogenous vs contaminating transcripts
13. Figure 3 Key parameters estimated from the data • Size of each cell • Mean expression of each gene in each cell-type • Latent cell-type label of each cell • Proportion of each cell-type (the probability of each cell-type essentially) • The diffusion coefficient of each gene • And distance- dependent decay
14. Figure 3 How are parameters estimated? They start with a bunch of priors, And then iterate through with a Markov chain Monte Carlo algorithm, but I have no clue what the objective function is or how ‘optimization’ is determined. Not even really sure what these panels are trying to communicate, figure legend is not useful. Also, is K (number of cell-types) estimated or provided as a hyperparameter? Unclear
15. Figure 3
16. Figure 4: testing with simulated data Ground truth simulated data Inferred data from DeLeakage Seems oddly perfect…
17. Figure 4 K-O: Residuals between inferred and actual expression across cell- types P-T: Cosine similarity measures between cell- types
18. Figure 5: testing with real data
19. Figure 5 Gene distribution parameters inferred from DeLeakage Cosine similarity again Jaccard index of mutually exclusive gene pairs. Globally decreased with DeLeakage
20. Figure 6: downstream benefits • Better clustering, supposedly, but this is a weird analysis • ARI is a metric that quantifies the consistency between two clusterings and they use the original author provided annotations as the ground truth • But wouldn’t the original annotations be impacted by this transcript leakage? • Wish they did a finer resolution cluster comparison
21. Available as a python package
22. General summary and thoughts • Observation of transcript leakage • Compelling argument, likely a real thing. • Will likely impact cell-clustering, cell-cell communication. • Also combinatorial neuronal populations, this might be an issue • Approach: model endogenous versus leaked transcripts and correct the counts • Conceptually makes sense, a bit lost on the actual implementation of the model • If we were to apply this, how do we check how well it works? • Results in paper • Everything seems pretty straightforward, difficult to perform comparisons since this is a novel idea. I’m always cautious when simulated results are near perfect, almost always implies circularity or overfitting • They claim genes vary in their dispersion characteristics, but I don’t think they actually show this empirically, just through their modeling results. • I would like to see an assessment of higher resolution cell-type annotations, which should be available for the brain datasets they used.