Fig 1.
The SL and SV pairs are retrieved from three sources: CGIdb, BioGRID, and SynLethDB.
Further, these pairs are mapped to the physical protein-protein interaction network, and the network properties are calculated. After identifying the determinant features, MAGICAL is built to classify a given gene pair as SL, SV, or NOT. The NOT dataset is built utilizing three steps: first, retrieval of all proteins from the BioGRID dataset and generating pairwise combinations. Second, removal of the pairs present in CGIdb, BioGRID, and SynLethDB, along with yeast orthologs. Third, Random picking of ~10,000 pairs is utilized for the model building.
Table 1.
List of network properties used as features.
Fig 2.
A stacked barplot to represent the determinant features that contribute to building MAGICAL.
The discriminatory features are identified by 1000 bootstraps and counting the number of times each model chooses a property.
Fig 3.
Boxplots showing A) SL interactions having higher value of shortest path than SVs (p-value < 2.2e-16, KS test). B) SL pairs having a higher value of average neighbor2 compared to SVs (p-value 4.87e-06, KS test). C) SL interactions having higher betweenness values to that of the SVs pairs (p-value < 2.2e-16, KS test). D) SL pairs having higher values of the average triangle, compared to SVs (p-value 1.65e-06, KS test). E) SL pairs having lower adhesion value than SV pairs (p-value < 2.2e-16, KS test).
Fig 4.
A) Stacked barplot representing the number of SL, SV, and NOT pairs correctly predicted by MAGICAL-core. B) The ROC plot depicting the performance of MAGICAL-core model. C) Stacked barplot representing the number of SL, SV, and NOT pairs correctly predicted by MAGICAL-combined. D) The ROC plot depicting the performance of the MAGICAL-combined model. The magenta, cyan, and grey colors represent SL, SV, and NOT.
Fig 5.
A) ROC curve showing an AUC of 0.850 and 0.935 for SLant-SL and MAGICAL prediction, respectively, for the balanced dataset (p-value 0.0003, DeLong test). B) For an unbalanced dataset, an AUC of 0.672 and 0.832 for SLant-SL and MAGICAL is respectively obtained (p-value < 2.2e-12, DeLong test). C) ROC curve showing the prediction accuracy of 0.83 and 0.86 for DepMap and Wang et al., data, respectively. D, E) Kaplan Meier curve showing SL Pairs identified by SLant predicted as SVs by MAGICAL and vice-versa (p-value, 0.15 and 0.052 respectively, log-rank test).
Fig 6.
A) The position and placement of the SL (magenta) and SV (cyan) genes in the protein-protein interaction network. B) SL pairs have a higher value of average GO terms (p-value < 2.2e-16, KS test). C) SV pairs have a higher value of the Jaccard index (p-value 3.198e-12, KS test).