Recommendations about generating new base editing weights

[RussBainer]

Hi crisprVerse team, thanks again for such an awesome set of tools for the community. I have a somewhat philosophical question that I thought I would surface.

Our group has been characterizing a set of base editors to empirically generate new editing weight matrices, and the team is divided about how to encode these as editing weights for prediction. We have performed experiments quantifying edit rates at a relatively short timepoint (~7 days DPI) in a panel of gRNAs, and the idea is to use these point estimates for prediction of outcomes in a long pooled screen (~3 weeks).

The key question is about standardization of the editing weights across editors, and whether it makes sense to use the raw point editing rates obtained in the experiment, or if the matricies should all be standardized to the unit interval (e.g., lowest observed rate = 0 and highest observed rate = 1). Ideally we would do a timecourse experiment to determine whether the observed edit frequencies are stable, but there isn't bandwidth on the team to do this; do you have any recommendations based on your own experience that you're willing to share?

[Jfortin1]

@RussBainer I don't think there is a right or wrong answer, depends on how you want to interpret the scores; if you think the unnormalized editing weights are good proxies for the true editing probabilities, then there is no need to normalize; for instance if you have a poor base editor that barely edits any of the bases, then normalizing to the unit interval will be misleading.

[RussBainer]

Thanks @Jfortin1 , that is my own assessment as well. As you imply, this is more of a question of the interpretation of the edit prediction score and the stability/accuracy/representativeness of the edit rate experiment.

Again, really love the toolset that you have developed!