Fig 1.
A schematic visualising how SLant’s source data is collated from STRING and the Gene Ontology Consortium, preprocessed so that this source data can be directed joined with BioGRID data for labeling and processed to create the final training set.
Feature generation was completed using R, the R igraph library and GoSemSim, a Bioconductor package.
Table 1.
Names and descriptions of the node-wise and pairwise network parameters and GO term features used in Slant.
Fig 2.
A set of violin plots illustrating the value distributions for each feature in our human training set grouped into SSL and non-SSL classes.
The features were derived from 411 SSL and 411 non-SSL gene pairs (see S6 Table). Feature distributions that show greater variance between SSL and non-SSL gene pair classes, for example the shortest path feature, often provide improved predictive power in classifiers.
Fig 3.
a. Human protein-protein interaction network with clustered communities generated by a spin glass random walk. Nodes and edges are coloured by their source community cluster as per the legend provided in Fig 3B. b. Community cluster connection graph where the weight of each connection corresponds to how many SSL interacting pairs begin and end at each community. We observe the largest count of SSL interactions occurring between cluster 9, notably associated with transcription regulation and DNA damage response GO terms and cluster 15, associated with MAPK cascade, cell proliferation and gene expression GO terms.
Table 2.
Cross validation ROC AUC scores for each organism from both in-species and cross species SSL models.
The best score for each species model is highlighted in green. Models are displayed vertically in rows with the consensus model displayed at the bottom of the table and the results for those models are displayed in columns with the consensus results highlighted in blue.
Fig 4.
Cross-species ROC AUC scores for each models classification performance on our human SSL interaction validation set.
An additional curve for our consensus predictions was added separately based on the performance of the consensus validation set.
Fig 5.
Carcinogenic survival assay results charting survival of PBRM1 / BAF180 knock-out cell lines with concentration intervals of the PARP inhibitor Olaparib, the POLA inhibitor Erocalciferol and the ABL inhibitor Dasatanib.
These results suggest PBRM1 mutant cells may be more sensitive to both the PARP and ABL1 inhibitors while gaining some resistance to POLA1 inhibition. Error bars measure standard error of measurement. All drug intervals are measured in mM.