Fig. 1.

Overview of proposed approach illustrated with toy example of cell line groups $H_1,H_2,H_3,H_4$, each with genes $G_1,G_2,G_3$ and drugs D₁, D₂. (A) For example cell line group H₃, its gene–drug association patterns are modeled as a bipartite graph of gene vertices and drug vertices, with edges weighted by association strength. The edge weight between gene i and drug j is defined as $a_i{b}_j$, where a_i, b_j are elements of SCCA canonical vectors $a\in {\mathbb{R}}^3$ and $b\in {\mathbb{R}}^2$. The bipartite graph can equivalently be represented by its edge weight matrix. This modeling of gene–drug association patterns is repeated for $H_1,H_2,H_4$. (B) Using the nuclear norm-based dissimilarity measure between bipartite graphs, a matrix of dissimilarities is computed for all the cell line groups, and hierarchical clustering can be applied as an example unsupervised analysis