After reading your introduction document, I still have questions about PREDICTABILITY. I’m curious if a higher PREDICTABILITY implies higher confidence in the numbers obtained from this CAS9 screening. For example, if I want to choose a potential target in cancer, do I need to consider both lower effect and higher PREDICTABILITY?
I hope to get your answer, thank you very much.
The predictability score is a statement about how well we’d be able to predict the gene effect of a new cell line for which we haven’t performed the CRISPR screen, only knowing it’s characterization data.
The reality is that we don’t actually try to predict the gene effect. However, knowing that we can predict the gene effect from characterization data does give us confidence that the the variation in gene effect is being driven by biological differences between models and not by some technical artifacts.
(Similarly, looking at which features are used to predict the gene effect often provide some hints as to the mechanism that’s driving the variation.)
Now, there are likely to be dependencies for which we cannot predict because there is possibly some property of those models which we have not measured. Or it’s possible that we cannot predict those dependencies well because the variance observed is just experimental noise. I don’t think we can disambiguate between those two cases.
I’d say having a high predictability score is evidence supporting the observed gene effect. It implies higher confidence that the gene effect reported is real.
However, I wouldn’t say the converse is true.
If one has a low predictability score, the gene effect reported could still be accurate. Having a low predictability score is not evidence that the observed gene effect is real or not real.
