Drug screening data (PRISM) vs. CRISPR screening data

Hi, I have a simple question about the results after drug screening and CRISPR screening.
TL;DR Can the values in those screenings be considered in the same context?

To my knowledge, the CRISPRGeneEffect.csv contains the cell viability value after CRISPR screening, and Repurposing_Public_24Q2_Extended_Primary_Data_Matrix.csv in the PRISM dataset represents the Log2-fold-change viability values (LFC_cb) after drug screening.
I wonder whether they mean the same context (i.e., cell viability) and can be used as the same-level data for the machine learning methods.
In addition, the values in those files seem to have a similar range: about -1.67 ~ 4.85 and -26.389 ~ 8.07, respectively. If I compare them, does the lower value mean a stronger effect on the cell line?

Thank you for reading this question.

[Update]
For example, when I checked the viability value of CRISPR and PRISM (condition: targeting FASN gene in the ACH-000015 cell line by CRISPR KO and GSK2194069), cell viability values were quite different, showing -0.0735 and 0.265813, respectively.
Since I’m unfamiliar with the screening process, it’s hard to understand why they showed different values. Is it due to the difference between ‘knock-out’ and ‘knock-down’?
If I want to use the viability values from CRISPR and PRISM, how can I preprocess them? Can anyone suggest the normalization approach or other methods?

The CRISPR data’s units are “Gene Effect” which is a unit that we defined as relative to the viability effects observed from knocking out a known essential gene and those effects from knocking out a non-essential gene.

The nature of the pooled CRISPR experiment requires a non-trivial model (Chronos) to measure the viability effects specific to each gene, and as a result, the values aren’t really a log fold change.

I imagine you’ll see fairly different distributions for values from PRISM and CRISPR screens, but it is more related to how the data was processed and less about the biological effects of inhibition vs knockout.

I personally don’t know of any instances where people have tried normalizing these to get them onto a comparable scale. I see more often the two datasets being used independently. The most common way I see them analyzed together is when looking for associations between the two, where we use pearson correlation to measure the amount of association.

Thanks,
Phil

1 Like

Thank you so much for your kind and detailed reply.
It was very helpful to understand the differences between CRISPR and PRISM datasets.
As you mentioned, it would be hard to use them at the same level of the true labels for machine/deep learning models.
I’ll try to find a better way to use them for the model.
Once again, thank you for your advice.