## Results: 5

1.

Recall, or fraction of known positives predicted by the system, is plotted on the X-axis (log scale); precision, or fraction of predictions that are in the training set, is plotted on the Y axis. The Bayesian integration PRC shows greater area under the curve than the correlation PRC, especially in the left-most, highest-confidence regime.

2.

Eigenvectors of the transcription (left) and metabolite (right) concentration data sets were calculated. Each eigenvector is composed of a characteristic response under carbon starvation (“carbon”, in light gray circles) and under nitrogen starvation (“nitrogen”, in dark gray triangles). For each gene eigenvector on the left, the corresponding metabolite eigenvector is plotted on the right. The corresponding eigenvectors correlated significantly, with p-values of (A) 6.9×10

^{−3}, (B) 2.2×10^{−4}, and (C) 2.1×10^{−2}. For the transcript data, the percent information explained by each eigenvector was (A) 46%, (B) 13%, and (C) 9%. For the metabolite data, the percent information explained was (A) 33%, (B) 20%, and (C) 10%.3.

(A) Schematic of reactions involving FBP and

*VID24*. The conversion of FBP to hexose phosphate is catalyzed by fructose-1,6-bisphosphatase (*FBP1*). Vid24p destroys this enzyme by targeting it to the vacuole for destruction. (B) Scatterplot showing the relationship between*VID24*and FBP abundances over carbon starvation (“carbon” in the legend) and nitrogen starvation (“nitrogen” in the legend). As in , lines represent linear best-fit curves, calculated separately for each condition (solid lines) or over both conditions (dashed line).*VID24*and FBP are inversely correlated under carbon starvation (light gray), but positively correlated under nitrogen starvation (dark gray), as anticipated for a gene interacting with a glycolytic metabolite.*VID24*was in the top 3% of predictions made for FBP.4.

(A) Overview of Bayesian integration procedure. Transcript and metabolite data were used to compute correlations between genes and metabolites over time under different experimental conditions. These correlations, along with a set of positive and negative examples obtained from KEGG, were used to train a Bayesian network. (B) Structure of the Bayesian network. This four-node network states that the variables corresponding to gene–metabolite correlations observed under either nitrogen or carbon starvation depend on the class of the metabolite involved, and whether or not a functional relationship between the gene and metabolite exists. The rounded boxes by each node represent the possible values that the nodes can take. (C) Conditional probability distributions learned from the experimental data. The parameters of the Bayesian network were computed from the experimental data and the set of positive and negative examples of gene–metabolite functional interactions. The light gray line gives the probability (y-axis) that, given no functional relationship, one would observe a given correlation (x-axis); the dark gray line gives the corresponding probability if given a true functional relationship instead. (D) Conditional probability distributions represented as log-odds scores. The sign of the bar corresponds to whether observing a certain strength and direction of correlation is more likely for a true functional relationship (positive) or for no functional relationship (negative), while the magnitude of the bar corresponds to how much more likely this is.

5.

Metabolite and gene transcript concentration changes are represented as ratios of measurements from starved cells to measurements from unstarved cells. The responses observed under carbon starvation (light gray circles, “Carbon” in legend) are labeled distinctly from the responses under nitrogen starvation (dark gray triangles, “Nitrogen” in legend), but are plotted on the same axes. Solid light gray and dark gray lines are linear best-fits for the responses observed under carbon and nitrogen starvation, respectively; the dashed line is a linear best-fit curve for all data. (A–E) Scatterplots of metabolites from the glycolysis and pentose-phosphate pathway metabolic class versus related genes show an inverse relationship under carbon starvation, but a positive correlation under nitrogen starvation. The dashed line shows that this relationship would be obscured by computing correlation across all data points.

*ILV2*catalyzes the first step of isoleucine and valine biosynthesis from pyruvate;*ARO3*catalyzes the first step in aromatic amino acid biosynthesis from PEP and erythrose-4-phosphate;*ALD6*, which also plays a key role in redox metabolism, is involved in the creation of cytosolic acetyl-CoA from pyruvate;*GLK1*phosphorylates glucose to glucose-6-phosphate; and*PGM2*catalyzes the interconversion of glucose-1-phosphate and glucose-6-phosphate. (F–H) Scatterplots of metabolites from the amino acid metabolic class versus related genes, in contrast, show positive correlation in both carbon and nitrogen starvation. Even in this case, however, computing correlation across both conditions can lead to an underestimation of the extent of the relationship (e.g., (H) threonine vs.*THR4*, where although and , ).*HTS1*charges (i.e. aminoacylates) the histidinyl-tRNA;*MET6*catalyzes the formation of methionine from homocysteine; and*THR4*converts phosphohomoserine to threonine.