Fig. 2.

Selection of classifiers using the phosphoinositide 3-kinase-related kinase (PIKK) family of kinases as an example. (A) ANNs are trained for individual domains, subfamilies, and families of domains; by contrast, the PSSMs are initially assigned to the specific domain with which the in vitro assay was performed. (B) As some PSSMs (for example, the one for ATM) may be better used as classifiers for a subfamily of closely related kinases (for example, ATM/ATR), we backtrack all PSSMs toward the root of the tree. (C) We eliminate families that contain domains that are highly dissimilar from each other (for example, the PIKK family and the ATM/ATR/mTOR subfamily), in order not to describe highly divergent domains with the same ANNs and PSSMs (see Methods). (D) Whenever possible, we benchmark the ANNs and PSSMs and discard classifiers that do not perform significantly better than random expectation. (E) A nonredundant set of classifiers is selected that maximizes the average AROC across all kinases, SH2 domains, or PTB domains. (F) For the PIKK family of kinases, this procedure selects the ANNs for the ATM/ATR subfamily, mTOR, and DNA-dependent protein kinase (DNAPK) to be the best combination of classifiers. See fig. S3 for an overview of the current selection of classifiers.