Figure 1.
Bayesian Markov Random Fields analysis (BMRF) for protein function prediction in a nutshell.
A. The topology of the interaction network is given. B. Functional annotations of proteins using a set of Gene Ontology terms. C. A partially annotated network. D–E. BMRF analysis.
Figure 2.
AUC scores for 90 GO terms, where the performances of the BMRF, MRF-Deng, LK and KLR was evaluated.
Figure 3.
Performance comparison for 90 GO terms, using the Area Under the ROC Curve (AUC).
The points above the diagonal denote improved performance of BMRF against A. MRF-Deng B. LK C. KLR. BMRF performs better for the majority of the tests compared to MRF-Deng and LK. KLR performs slightly better, but it is difficult to be applied in large datasets.
Figure 4.
Comparison of parameter estimation and prediction performance between BMRF and MRF-Deng for the GO term “ homeostatic process”.
A–B. In BMRF the parameters and
are sampled closeby to the true parameter values, in contrast to MRF-Deng where the parameters are estimated using only the annotated part of the network and lead to overestimated values. C. Both methods estimate the intercept reasonably well. D. ROC curves for the prediction performance of the two methods.The AUC value for BMRF is 0.79 and for MRF-Deng is 0.71.
Figure 5.
Running times for KLR and BMRF.
The horizontal axis represents the size of the network and the vertical the time (in seconds) needed by each method. The computations were performed using the same hardware i.e. a Pentium 4 with dual core processor with 4GB of RAM and Linux operating system. The crosses denote the network size where the running times were evaluated. For BMRF the running time grows linearly with the network size while for KLR it grows polynomially.
Table 1.
Manually evaluated predictions of protein functions.
Figure 6.
Number of unannotated proteins and number of interactions against Purification Enrichment (PE) score.
The numbers are divided by their values for PE = 0 (i.e. the network without any cutoff that contains the full set of proteins and edges). The validation network was constructed using PE = 3.19 as suggested by [24].