A Mixture of Feature Experts Approach for

Protein-Protein Interaction Prediction


Yanjun Qi1, Judith Klein-Seetharaman1,2, Ziv Bar-Joseph1

1School of Computer Science, Carnegie Mellon University, Pittsburgh, PA 15213

2Department of Structural Biology, University of Pittsburgh School of Medicine, Pittsburgh, PA 15261





Background: High-throughput methods can directly detect the set of interacting proteins in model species but the results are often incomplete and exhibit high false positive and false negative rates. A number of researchers have recently presented methods for integrating direct and indirect data for predicting interactions. These methods utilize a common classifier for all pairs. However, due to missing data and high redundancy among the features used, different protein pairs may benefit from different features based on the set of attributes available. In addition, in many cases it is hard to directly determine which of the data sources contributed to a prediction. This information is important for biologists using these predications in the design of new experiments.


Results: To address these challenges we propose a Mixture-of-Feature-Experts method for protein-protein interaction prediction. We split

the features into roughly homogeneous sets of feature experts. The individual experts use logistic regression and their scores are combined using another logistic regression. When combining the scores the weighting of each expert depends on the set of input attributes available for that pair. Thus, different experts will have different influence on the prediction depending on the available features.


Conclusions: We applied our method to predict the set of interacting proteins in yeast and human cells. Our method improved upon the best previous methods for this task. In addition, the weighting of the experts provides means to evaluate the prediction based on the high scoring features.


PDF version (PDF)

Online version (Online)



Features (Yeast PPI)


        For the various data sources, each of them has its own representative form. For example, protein sequence is in the form of a character string, which means the order of amino acids as they occur in a polypeptide chain. Gene expression data is usually a vector of expression values across multiple time points for a specific gene. Synthetic lethal data describes that a pair of genes having mutations together would affect the cells inviable or viable. So how could we combine these different forms of data together? We present the converting process briefly in Figure 1. For each data set that represents a certain gene / protein's property, we figured out one natural way to calculate the similarity between two genes / proteins with respect to the specific evidence. For instance, for two proteins' sequence information, we use BlastP sequence alignment E-value as one feature for this protein-protein pair from the protein sequence evidence. For other data sources, similar procedures were pursued to make the features for a protein pair. Concatenating these features together then gave us the feature vector describing a protein-protein pair.

        The yeast features used in the paper were described in detail in the following paper

o       Y. Qi, Z. Bar-Joseph, J. Klein-Seetharaman"Evaluation of different biological data and computational classification methods for use in protein interaction prediction", PROTEINS: Structure, Function, and Bioinformatics. Jan 2006

o       All related data sources and how they were converted into features representing pair of proteins have been described in details in: http://www.cs.cmu.edu/~qyj/papers_sulp/proteins05_pages/features.html

o       These data sets could also be downloaded from:





Features (Human PPI)

        The human PPI features used in the paper were described in detail in the paper

        We share one run of the features and their related pairs @ the following ULR:

o       http://www.cs.cmu.edu/afs/cs.cmu.edu/project/structure-9/PPI/mfe07-HumanGeneral/




Discussion of Number of Feature Experts


        As discussed in the 'Methods' section, the number of feature experts that our heterogeneous data sets are split into could be different to what we used in the paper. The splitting depends on the need of the application and the preference of the biologists who would analyze and/or validate the predictions.

        The methodology we propose, of feature experts, is very general. At the limit we can assign each feature to an individual expert. To test this we carried out new experiment for the human prediction task treating every feature as its own expert. As the following supporting Figure 1 indicate, this does not improve the performance of the algorithm we present, perhaps because it leads to overfitting of the parameters.

        Supporting Figure 1:




        Supporting Table 1:





AUC 50


AUC 100