Supplementary Figure 1: Integrating the entire data compendium with hierarchy-aware tissue-specific knowledge generates networks that better capture tissue-specific interactions than limiting the integration to tissue-specific data (P = 1.3 × 10−9).
From: Understanding multicellular function and disease with human tissue-specific networks
For each tissue, two networks—one integrating the entire data compendium and the other integrating only tissue-labeled data—were generated, and their performance was measured using area under the receiver operator curve (AUC) on the basis of cross-validation. The scatterplot shows that the performance for 64 tissues (points) with tissue-labeled data (x axis) and all data (y axis), with 62 of 64 performing better with all data (above the diagonal line; P = 3.2 × 10−12). The remaining 80 tissues did not have sufficient tissue-specific data (fewer than 5 data sets) available to perform a tissue-restricted integration. The performance of our Bayesian integration for these tissues is shown on the disconnected axis.
