Integrate Sanger screen data

[jhrcook]

The Sanger performed an CRISPR/Cas9 screen independently from the Broad (DepMap), but the data is available through the DepMap portal. I should use this data for this analysis, too.

These two data sources were integrated in Pacini et al. (2021) and they specifically discuss adjusting for batch and bias. They may even provide an adjusted/harmoized data set that could be worth using directly. Some notes from their paper:

• Broad has 759 cell lines (as of 20Q2) and Sanger has 317 - total 908 unique cell lines with 168 cell lines overlapping.
• They do a substantial number of batch correction steps including ComBat, quantile-quantile normalization, and removal of principal components.
  • These are all linear methods and should be incorporated into the model with a hierarchical structure (y ~ ... + (1|batch|institute))
• Evaluated their pipeline using "the identification of (i) essential and non-essential genes (ii) lineage subtypes (iii) biomarkers of selective dependencies, and (iv) functional relationships."

---

Pacini, Clare, Joshua M. Dempster, Isabella Boyle, Emanuel Gonçalves, Hanna Najgebauer, Emre Karakoc, Dieudonne van der Meer, et al. 2021. “Integrated Cross-Study Datasets of Genetic Dependencies in Cancer.” Nature Communications 12 (1): 1661.

[jhrcook]

I think there are two options for adjusting for batch effect between Broad and Sanger: 1) as a pre-processing step, 2) as a covariate in the model. The paper referenced above experimented with normalizing using various pre-processing steps, but I think I would prefer the latter method. Below is a mock model with the variables adjusting for batch effects:

lfc ~ ... + (1|batch|source) + (0 + sanger|gene)

The first parameter is a standard batch effect covariate but the source (either Sanger or Broad) is added as an additional hierarchical level. The second parameter is a specific batch effect correction for the Sanger per gene. The 0 is there just to explicitly say that this does not introduce a varying intercept per gene as that is already included in the model. Instead, this introduces a varying parameter that is included when the data comes from Sanger, thus representing the effect for the source of the data per gene.

[jhrcook]

As this is due to a biological variable (cell line culture media), I would expect it to also vary per cell line depending on the type of cell line. Therefore, I should experiment with having the covariate vary per gene and lineage when I scale to beyond CRC.

[jhrcook]

When I presented Pacini et al. at Park Lab Cancer Subgroup journal club, Peter and Hu were hesitant about using both data sets, thinking that the batch effects would be more confounding than the increase in data would be informative. Therefore, it may be more trouble than it is worth to use this dataset.