To assess the degree to which each perturbagen profiled in L1000 recovered its expected connections to other perturbagens in Touchstone we leveraged annotations compiled from various sources. First, the annotations were used to construct a pairwise binary association matrix for all perturbagens in Touchstone. A pair of perturbagens were considered to be associated if they shared at least one type of annotation. For example, a pair of small-molecules were associated if they shared the same MoA. Similarly a compound and a genetic perturbagen could be associated if they shared the same gene target. We retained 1,902 small-molecule, 994 genetic over-expression, and 1,634 CGS perturbagens after excluding those that had too few (<10) or too many (>3,000) connection pairs. Then for each perturbagen p, we partitioned all associated perturbagen-pairs into a collection of expected connection pairs (Ep) whose members were associated with p and and a collection of background pairs (BP) whose members were not associated with p. Finally, ROC analysis was performed wherein the connectivities between members of Ep were compared to that between members of Bp at different threshold
