Uncovering the rules for protein-protein interactions from yeast genomic data

Abstract

Identifying protein-protein interactions is crucial for understanding cellular functions. Genomic data provides opportunities and challenges in identifying these interactions. We uncover the rules for predicting protein-protein interactions using a frequent pattern tree (FPT) approach modified to generate a minimum set of rules (mFPT), with rule attributes constructed from the interaction features of the yeast genomic data. The mFPT prediction accuracy is benchmarked against other commonly used methods such as Bayesian networks and logistic regressions under various statistical measures. Our study indicates that mFPT outranks other methods in predicting the protein-protein interactions for the database used. We predict a new protein-protein interaction complex whose biological function is related to premRNA splicing and new protein-protein interactions within existing complexes based on the rules generated. Our method is general and can be used to discover the underlying rules for protein-protein interactions, genomic interactions, structure-function relationships, and other fields of research.

Fig. 1.

FPT KS curve of training (A) and testing (B) samples.

Fig. 2.

FPT ROC curves (A and B) for different scales. Red curves represent training samples, and green curves represent testing samples.

Fig. 3.

ROC curve and KS value comparisons for 4 methods. (A and B) ROC curve comparisons of the training (A) and testing (B) samples for 4 methods. (C and D) KS value comparisons of the training (C) and testing (D) samples for 4 methods.

Fig. 4.

Comparisons of correct prediction rate for 4 methods for both training and testing samples

Fig. 5.

Comparisons of correct prediction rate for 5 features and 13 features with both training and testing data samples

Fig. 6.

Predicted complex.

Fig. 7.

Mitochondrial ribosome complex.

Fig. 8.

New complex.