Fig. 4

a Illustration of the two OV datasets used for evaluating INSPIRE. Rows represent genes and columns represent samples. b For k = 91 (left) and k = 82 (right), INSPIRE is compared to WGCNA variants (top) and MGL variants (bottom) in terms of the best cross-validation (CV) negative test log-likelihood (lower is better) across all tested sparsity tuning parameters (λ). c For k = 91, INSPIRE (y-axis) is compared to each of the six competing methods (x-axes) in terms of the best − log₁₀ p from the functional enrichment of the learned modules. Each dot is a KEGG, Reactome, or BioCarta GeneSet, and only the GeneSets with a Bonferroni corrected p <0.05 in at least one of the compared two methods are shown on each plot. For MGL variants and INSPIRE, results from multiple runs are shown. We only considered the GeneSets with sufficiently different significance between the two methods, i.e. | log₁₀ p(INSPIRE) − log₁₀ p(ALTERNATIVE_METHOD)| ≥ δ. δ = 6 here and the results were consistent for varying δ. d For k = 91, INSPIRE (y-axis) is compared to each of the six competing methods (x-axes) in terms of the best − log₁₀ p from the ChEA enrichment of the learned modules. Each dot is for a gene set composed of a TF and its targets, and only the sets with a Bonferroni corrected p <0.05 in at least one of the compared two methods are shown on each plot. For MGL variants and INSPIRE, results from multiple runs are shown. We only considered the TFs with sufficiently different significance between the two methods, i.e. | log₁₀ p(INSPIRE) − log₁₀ p(ALTERNATIVE_METHOD)| ≥ δ. δ = 3 here and the results were consistent for varying δ. Each blue dot corresponds to a TF which sits in the INSPIRE module that is significantly enriched for its targets and each red dot corresponds to a TF which sits in an INSPIRE module different than the one that is significantly enriched for its targets