Figure 1.
Images of a genome-wide cellular RNAi knockdown screen (the screening data was derived from the Mitocheck project, www.mitocheck.org) were segmented and their features extracted to compile pairwise phenotype descriptors for a large set of gene pairs. These descriptors were used to train a machine learning system to discriminate activating and inhibiting PPIs taken from a reference. The performance was evaluated using cross-validation. The trained SVM models were used to predict the effects of uncharacterized PPIs. In addition, the SVM models were used to estimate similarity of the effects of proteins for all combinations of protein pairs in the network. Subsequently, this Effect Similarity Rate (ESR) was exemplarily used for clustering of functionally related protein sub-networks.
Figure 2.
Characterization of phenotypic similarity by linear discrimination.
(a-c) Images of cells in which sfrp1, dvl2 or fzd7 were knocked down, respectively. (d) First two principal components (PC 1 and PC 2) of the features for cells with knockdown of sfrp1 and dvl2. (e) First two principal components of the features for cells with knockdown of dvl2 and fzd7. Dotted lines sketch a linear separation.
Figure 3.
a) Receiver Operating Characteristics curves for the predictions of activation.
Cross-validation results for all pathways combined (AUC = 0.75, dashed line) and when training and validation was done for each set of pathways separately (AUC = 0.82, solid line). b) Histogram of the votes for activating PPIs (green) and inhibiting PPIs (blue) when training and validation was done for each set of major signaling pathways separately. The thresholds for 80% confidence were set at 920 and 88 votes for activation and inhibition, respectively (dashed lines).
Table 1.
Pairs of Pfam domain sets showing significant* enrichment of predicted interactions.
Figure 4.
Clustering of the chemokine receptors and their interactors.
Figure 5.
Querying and visualizing inhibitory and activating interactions with HIPPIE.
a) Querying HIPPIE with the IL2RB, a receptor that transduces IL2 signals in immune response, reveals 35 interaction partners from which 24 are associated with an effect prediction (HIPPIE indicates activations by arrows and inhibitions by bars). 11 of these effects are also found in KEGG and are correctly reproduced by our approach. 2 interactions for which we predict an effect are listed in KEGG but have no effect assigned there (IL2, SOS1). 11 predicted effects are not annotated at all as interactions in KEGG (FYN, HGS, IL15, IL2RG, IRS1, LCK, PIK3R1, RAF1, STAT1, STAT5A, SYK). Many of these interaction partners are, however, organized in cytokine-related pathways and, thus, demonstrate the potential of our approach to not just reproduce KEGG annotations but also to detect novel and meaningful interaction effects. b) We uploaded interactions for which we could predict an activating or inhibiting effect within the CCR-subset and its direct interactors to HIPPIE (blue edges). Using the default HIPPIE output-options, HIPPIE extended the query set of interactions with additional interactions between the input proteins. The newly added interactions are colored in grey. We also enabled the option from the HIPPIE menu to display interaction directions as defined by KEGG. Arrows are unidirectional if the directions were known (from KEGG) and bidirectional otherwise. HIPPIE uses diamond-shaped (grey) arrows to indicate that an interaction has associated a direction but no effect. We manually highlighted CCRs in pink, JAKs in blue, SOCS in yellow and G-proteins in green. The depicted predictions for activation and inhibition are correct (except for the interactions between GNGT1 and GNB3, and between SOCS1 and CXCR4): SOCS are inhibiting JAKS, CCRs, G-proteins, all other protein pairs are activating. It is to note that in our representation, edges with two arrows (bidirectional edges) indicate that the direction of the effect is not known. They do not indicate a bi-direction of an interaction of e.g. a simple feedback loop (in which A activates B which in turn activates A).