Fig 8. Construction of disease-specific co-perturbation networks.

Two networks are learned using a Gaussian Markov Random Field network learner. The first network (“disease-control” network) is learned from disease profiles, disease surrogate profiles, control profiles, and surrogate control profiles. Importantly, half the matrix rank is composed of disease and/or disease surrogate profiles and the second half of the matrix ranks is composed of control and/or surrogate control profiles. A second network composed of only control and surrogate controls is also learned (“control-only” network). A final pruning stage subtracts edges from the disease-control network that are also found in the control-only network, outputting the disease-specific network. Disc. Latent Struct. Inference–Discriminative latent structure inference; GMRF–Gaussian Markov Random Field; N(0,1)–the standard normal distribution.