That’s what happened when Genentech applied Cerella to their kinase screening programs
The Challenge
Poor kinase selectivity can cause unexpected safety failures, or interfere with our understanding of mechanisms of action. Therefore, kinase selectivity screening is important. But with hundreds of possible kinases to measure against, which experiments are most valuable?
The Solution
By building an imputation model with Cerella’s deep learning methods, we were able to fill in Genentech’s very sparse data set (less than 4% of all possible measurements were present). By providing uncertainty values, we could see with confidence which predicted values were most likely to be accurate.
In addition, Cerella indicated which measurements within the test set were most informative for imputing missing values, based on the complex interrelationships between the kinases and variability in particular assays. The vast majority of assays were deemed not necessary for the accurate prediction of all endpoints.