Identifiability and inference of pathway motifs by epistasis analysis

The accuracy of genetic network inference is limited by the assumptions used to determine if one hypothetical model is better than another in explaining experimental observations. Most previous work on epistasis analysis-in which one attempts to infer pathway relationships by determining equivalences among traits following mutations-has been based on Boolean or linear models. Here, we delineate the ultimate limits of epistasis-based inference by systematically surveying all two-gene network motifs and use symbolic algebra with arbitrary regulation functions to examine trait equivalences. Our analysis divides the motifs into equivalence classes, where different genetic perturbations result in indistinguishable experimental outcomes. We demonstrate that this partitioning can reveal important information about network architecture, and show, using simulated data, that it greatly improves the accuracy of genetic network inference methods. Because of the minimal assumptions involved, equivalence partitioning has broad applicability for gene network inference.