Figure 1.
Overview of our integrated Bayesian hierarchical system for inferring diverse interaction networks.
An interaction ontology was constructed categorizing gene interaction types. A corresponding Bayesian network was constructed in which each node represents one interaction type. This network's structural parameters, P[parent node label |child node labels], were first determined using prior knowledge from GO [36], KEGG [59], SGD [56], and other curated sources. Second, individual SVM classifiers were trained to predict each interaction type in isolation using heterogeneous data sources. Third, the non-structural Bayesian network parameters, P[true latent node label |SVM output], were filled by relating each observed SVM classifier to a latent interaction type membership node using cross validation. Finally, to generate new predictions, a gene pair's interaction type is first predicted by the SVM classifiers and then hierarchically resolved by finding the most probabilistically consistent set of label assignments corresponding to the latent nodes in our Bayesian network.
Figure 2.
Performance evaluation of inferred networks.
We predicted 30 S. cerevisiae interaction networks, each representing one interaction type. A) To evaluate the overall accuracy of these networks, we withheld ∼30% of the genes in our gold standard as a test set. Performance on this test set averaged an AUC of 0.79 across all interaction types in the ontology; see Text S1 for individual ROCs. B) To specifically assess the accuracy with which interaction directionality was predicted (as opposed to the presence/absence of interactions in part A), we tested the frequency with which an interaction's correct direction was ranked above its incorrect direction in each of the 12 directed interaction networks. These results are uniformly well above random (0.5), supporting our ability to accurately predict both the presence and the directionality of many specific types of protein interactions.
Figure 3.
Examining the mechanisms of protein interactions within the yeast carbon metabolism and cellular transport pathways.
A) Predicted interactions of four specific types combined to assemble B) (arrows in black representing our final predicted pathway interactions) a pathway connecting the transcription factor Adr1 involved in carbon metabolism process to its regulatory input Snf1. This generates two concrete hypotheses suggesting, first, cross-talk between the calmodulin- and Snf1-dependent pathways via Cmk2 phosphorylating Glc7. Second, we also predict coordinated regulation between the glycogen breakdown and glucose utilization pathways through a metabolic interaction between Adr1 and Gph1. C) Previously known and newly predicted interactions in yeast protein transport connecting the plasma membrane, vacuole, golgi and ER. We propose a regulatory competition between the Arf1 and Vsp1 GTPases for Bch1 functionality that is likely regulated by GTP availability, which itself is known to be regulated by protein sorting events in the cell. These predictions also hypothesize that YDL012c may be involved in regulating Vps1 activity.
Figure 4.
Experimental validation of predicted synthetic lethal interactions.
Experimentally tested synthetic lethal hypotheses in the yeast A) DNA topological change and B) regulation of protein biosynthesis processes. A total of 20 gene pairs from our predicted synthetic lethality networks were experimentally tested using the SGA platform [4], [13]. We confirmed 14 of these interactions (70%), 8 in DNA topological change and 6 in protein biosynthesis. Several of the remaining unconfirmed pairs (e.g. GCS1 and SLT2; see main text) show additional evidence of condition-specific synthetic lethality.
Figure 5.
Systems-level analysis of inferred networks.
In all cases, continuously weighted networks were binarized by choosing an edge cutoff three standard deviations above mean, retaining ∼1% of all edges. A) The degree distribution for all 30 of our predicted interactomes agrees strongly with a scale-free network topology. B) Conditional probabilities for a gene to appear in the top 5% of each pair of networks' degree distributions. Similarity indicates that a pair of networks share the same high-connectivity genes and thus represent functional activity carried out by similar sets of proteins. C) Graphlet degree distributions compared using the GDD metric between the 13 leaf interactomes in our interaction ontology. Network pairs with greater similarity demonstrate related local network topologies, suggesting that comparable functional modules might be employed in the two interactomes (e.g. between phosphorylation and synthetic interactions or ubiquitination and post-translational regulation).