Protein–protein interaction site prediction based on conditional random fields

Abstract

Supplementary data are available at Bioinformatics online.

Predicting interacting and interfacial residues using continuous sequence segments

“…Even at the current stage, the performance is quite satisfactory reaching a 60% level of accurate predictions. This rate is less than the best reported methods [9,16,[18][19][20]22], however, another library-based method described in Ref. [21] has achieved an accuracy level of 24% only.…”

“…This is partly due to the progress made in such experimental high-throughput techniques (see an excellent review [3]) as two-hybrid analysis [4,5], affinity purifications [6] and other methods [7]. These techniques provide large and diverse training sets of protein-protein inter- actions, which allow extensive optimization of the parameters of the algorithms used, which is crucial for success of the machine learning algorithms [8][9][10][11].…”

“…Thus, knowing that two proteins interact, we need computational methods capable of predicting putative interfacial residues, interacting residue pairs or the entire 3D structure of the corresponding protein-protein complex [12][13][14] thereby identifying the role of the individual amino acids on the affinity. To address this necessity neural networks [11,15], conditional random fields [9] or support vector machines [16] algorithms were developed and they showed very promising results. The success is partially due to the continuous growth of the available 3D structures in the Protein Data Bank (PDB) [17] which provides an additional set of training data for these machine learning algorithms.…”

Protein-protein interaction prediction via Collective Matrix Factorization

“…Thus, much effort has been spent on developing informative and effective feature representation methods for PPI prediction. Feature vectors may be extracted based on protein sequences directly or may involve indirect evidences, including domain compositions, motif pairs and related mRNA expression [5], [6], [7], [8], [9], [10], among others. Bock and Gough [13] used SVM method based on compositions of amino acids and physiochemical descriptors.…”

A primal sub-gradient method for structured classification with the averaged sum loss

“…These algorithms, such as conditional random fields (Lafferty et al, 2001), the structured perceptron (Collins, 2002) or structured support vector machines (SSVMs) (Tsochantaridis et al, 2005), are proved to outperform the standard binary and multiclass classifiers, but they are usually more complex to train and require inference during the training procedure. They are applicable to different domains such as natural language processing (Daume, 2006), computer vision (Nowozin and Lampert, 2011), speech recognition (Sas andŻołnierek, 2013) and bioinformatics (Li et al, 2007). Besides easy training for the perceptron algorithm, training the SSVM assumes constrained optimization with possibly exponentially many constraints.…”