Phylogenetic tree information aids supervised learning for predicting protein-protein interaction based on distance matrices
© Craig and Liao; licensee BioMed Central Ltd. 2007
Received: 15 September 2006
Accepted: 09 January 2007
Published: 09 January 2007
Protein-protein interactions are critical for cellular functions. Recently developed computational approaches for predicting protein-protein interactions utilize co-evolutionary information of the interacting partners, e.g., correlations between distance matrices, where each matrix stores the pairwise distances between a protein and its orthologs from a group of reference genomes.
We proposed a novel, simple method to account for some of the intra-matrix correlations in improving the prediction accuracy. Specifically, the phylogenetic species tree of the reference genomes is used as a guide tree for hierarchical clustering of the orthologous proteins. The distances between these clusters, derived from the original pairwise distance matrix using the Neighbor Joining algorithm, form intermediate distance matrices, which are then transformed and concatenated into a super phylogenetic vector. A support vector machine is trained and tested on pairs of proteins, represented as super phylogenetic vectors, whose interactions are known. The performance, measured as ROC score in cross validation experiments, shows significant improvement of our method (ROC score 0.8446) over that of using Pearson correlations (0.6587).
We have shown that the phylogenetic tree can be used as a guide to extract intra-matrix correlations in the distance matrices of orthologous proteins, where these correlations are represented as intermediate distance matrices of the ancestral orthologous proteins. Both the unsupervised and supervised learning paradigms benefit from the explicit inclusion of these intermediate distance matrices, and particularly so in the latter case, which offers a better balance between sensitivity and specificity in the prediction of protein-protein interactions.
Protein-protein interactions play a key role in cellular functions, and thus, to complement the experimental approaches [1, 2], many computational methods have recently been developed in systems biology for predicting whether two proteins interact, based on what is already known about these proteins. One type of data used for prediction is the phylogenetic profile of a protein – a string of ones and zeros encoding respectively the presence and absence of the protein in a group of genomes, conserved operons, gene fusions, etc. [3–6]. The rationale is that interacting proteins tend to co-evolve, and therefore should have similar phylogenetic profiles. Recently, to enhance the prediction accuracy, the focus has been given to using the similarity of phylogenetic trees to infer interactions between receptors and ligands [6–8].
Of particular interest is the so-called mirror tree method by Pazos and Valencia . The mirror tree method predicts protein-protein interactions under the assumption that the interacting proteins show similarity in the molecular phylogenetic protein trees because of the co-evolution caused by the interaction. However, it is difficult to directly evaluate the similarity between a pair of molecular phylogenetic trees. Instead, the mirror tree method compares a pair of distance matrices by calculating the Pearson correlation coefficient for the corresponding elements in the two matrices, and uses the correlation coefficient as a measure to evaluate the extent of co-evolutionary behavior between two proteins.
To address the issue of high rate of false positives with the mirror tree method, recently, Sato et al  suggested that the information about the phylogenetic relationships of the host genomes be excluded by a projection operation, and only the residual information in the distance matrices be used for the calculation of the correlation coefficient between proteins. As a result, significant improvement in prediction specificity was achieved, though at a cost of losing some sensitivity. A similar yet more sophisticated approach is proposed in Pazos et al  to correct the distance matrices based on the phylogenetic tree, which incorporates information on the overall evolutionary histories of the species (i.e., the canonical "tree of life"). In addition to adjusting the distance matrices by excluding the expected background similarity due to the underlying speciation events, this tree of life mirror tree (tol-mirrortree) method can also detect non-canonical evolutionary events, in particular horizontal gene transfers. While both Pazos et al's tol-mirrortree method and Sato et al's projection approach are concerned with – and quite successful at – removing some background from the inter-matrix correlation, like the original mirror tree method they do not directly address the intra-matrix correlations, which can be informative and critical in revealing co-evolution. For example, in some recent related studies, the columns and rows of the distance matrices are reshuffled in an attempt to discover maximal similarity between two matrices in order to predict interaction specificity when paralogs are involved [11–13].
In this work, we propose a novel, simple method to extract the intra-matrix correlational information with reference to the species tree of the host genomes and to represent said information in a way that is conducive to a supervised learning paradigm. We tested our method on the same dataset used in , which consists of interacting proteins from E. coli, where these interactions are experimentally verified. The results from a series of leave-one-out cross validation experiments showed that the prediction accuracy was greatly increased with our data representation method.
The 13 pairs of interacting proteins in E. coli
The 41 source organisms used to find orthologous proteins
Escherichia coli K-12 MG1655
Salmonella typhi CT1 8
Samonella typhimurium LT2
Yersinia pestis CO92
Shigella flexneri 301 (serotype 2a)
Vibrio vulnificus YJ016
Pseudomonas syringae pv. Tomato
Acinetobacter sp. ADP1
Neisseria meningitides MC58 (serogroup B)
Campylobacter jejuni NCTC11168
Bacillus anthracis Ames
Bacillus cereus ATCC 14579
Bacillus lucheniformis ATCC 14580
Staphylococcus aureus N315 (MRSA)
Staphylococcus epidermidis ATCC12228
Rhodopirellula baltica (Pirellula sp.)
Anabaena sp. PCC7120 (Nostoc sp. PCC7120)
The phylogenetic tree for these 41 reference bacterial genomes was built from the 16S rRNA sequences using the neighbor-joining module in PHYLIP package. The 16S rRNA sequences were downloaded from the KEGG/GENES database  and the Ribosomal Database Project-II Release 9 .
Phylogenetic vectors and correlations
The original mirror tree method is proposed by Pazos and Valencia  to infer protein-protein interaction from correlated evolutions. The hypothesis is that two proteins should have a higher chance to share correlated evolutionary history if they interact with each other than if they do not. As the evolutionary history for a protein can be represented as a phylogenetic tree (let's call it a protein tree to distinguish from the species tree), it makes sense to compare the two protein trees to reveal any correlation between their evolutionary history. Instead of comparing two trees directly, which is a highly nontrivial task in terms of both algorithmic implementation and biological interpretation, the mirror tree method uses as a surrogate the distance matrices that store the genetic distance between the protein and its orthologs in a group of genomes. It is from these distance matrices that the proteins trees are typically reconstructed using well known algorithms such as Neighbor-Joining . For two proteins A and B, the mirror tree method compares their distance matrices DA and DB, by examining how the corresponding elements are correlated. Because the distance matrices are symmetric, only the elements in the upper (or lower) triangle of the matrices are needed to calculate the correlation, which is measured as the Pearson correlation coefficient ρ as defined below:
where Ave and Var represent the average and the variance of the elements in the upper triangle of a distance matrix, respectively. To apply this method for prediction, the Pearson correlation coefficient ρ is calculated for all distinct pairs of proteins, and these pairs are then ranked in a non decreasing order of ρ. With a threshold preset on ρ, the pairs with a higher correlation coefficient are predicted to be interacting pairs.
In two recent works [9, 10], the measurement of correlations is refined by excluding the information about the phylogenetic relationships in order to overcome the problem of high rate of false positives reportedly present in the predictions using the mirror tree method. The high rate of false positives is believed to be caused by a high correlation between non-interacting proteins, which can be attributed to some common background shared by the two corresponding distance matrices, because they all are derived from orthologous proteins in the same set of n source organisms. That is to say, these protein trees bear some resemblance to the species tree. Sato et al therefore propose to exclude the species tree resemblance from the distance matrices before comparing them. Specifically, a distance matrix R is computed for the 16S rRNA sequences of these 41 genomes, from which the species tree can be reconstructed. For convenience, all the rows in the upper triangle of this 41 × 41 distance matrix are concatenated, producing a vector of dimension 820, which we refer to as |u16s>. Similarly, all the distance matrices for the protein trees can be transformed into a vector form, of the same dimension 820, which is termed the phylogenetic vector. Let |vi> (i = 1 to 26) be a phylogenetic vector for one of the 26 proteins in the dataset, then the resemblance of |vi> to |u16s> is measured by the projection < u16s|vi > (i.e., the inner product between |vi> and |u16s>), which is then subtracted from |vi>, giving a residue vector |εi> defined as follow:
|εi> = |vi> - |u16s> (<u16s| vi >) (2)
Then, the Pearson correlation coefficient (is calculated for any pair of vectors |εi> and |εj>:
where stands for the k-th component of vector |εi>, Ave and Var represent the average and variance of elements in a vector. It is shown in  that the specificity of predictions using the subtracted vectors is significantly improved, though at a cost of losing sensitivity.
In , phylogenetic trees (protein trees) are first reconstructed from the multiple sequence alignments of orthologous proteins using the neighbor-joining algorithm implemented in ClustalW. The protein distance matrices are then derived from these trees by summing the length of the branches connecting each pair of orthologous proteins, which are represented as tree leaves. New distance matrices for the proteins are obtained by subtracting from each value the distance between the corresponding species in the 16S rRNA distance matrix, termed as R. If we transform the matrices into vectors in the same way as used in , then the element-wise subtraction of 16S rRNA distance matrix R from a distance matrix P for a protein is equivalent to subtraction of two corresponding vectors,
|p'> = |p> - |r> (2')
Super-phylogenetic vectors via TreeSec method
We propose a novel method to utilize the distance matrices and the species tree for predicting protein interactions. There are two major changes to the mirror tree method. First, we augment the phylogenetic vectors with extra bits that encode the topological information of the protein tree with reference to the species tree. Second, in contrast to the unsupervised learning scheme of the mirror tree method, we adopt a supervised learning paradigm, specifically the support vector machines, [20–22] to further tap into the prior knowledge about interacting and non interacting protein pairs. As all proteins are already represented as the phylogenetic vectors of the same dimension, it is convenient to concatenate the two vectors for any pair of proteins and use the concatenated vector to represent the pair. All pairs thus represented are then split into to two subsets – one subset is used for training and the other for testing.
D(α, β) = ∑i∈Cα;j∈Cβ[D(i, j)]/|C α | |C β | (4)
where |C α | and |C β | are the size for the clusters C α and C β respectively. That is, the distance between two clusters C α and C β is equal to the average distance between pairs of orthologous proteins from each cluster. This definition of distance between two clusters is the same as defined in UPGMA during the tree reconstruction . The intermediate distance matrix gives a "snapshot" of the evolutionary history about these orthologous proteins at the time, marked as the tree height, where the section is cut. The snapshot – rather the intermediate matrix derived from it – carries the information about how the hypothetical ancient ancestors at that time are related to one another in terms of evolutionary distance, from the perspective of the protein being studied. Since the matrix is symmetric, only the upper (or lower) triangle is needed, which can be transformed into a vector form and concatenated to the original phylogenetic vector in the mirror tree method. The final representation of a protein is the original phylogenetic vector concatenated with all "snapshot" vectors, which is called super-phylogenetic vector. Figure 3 gives a schematic illustration of the procedure for generating the super-phylogenetic vectors. The number of "snapshots" is a free parameter in our method. One simple way to remove this parameter is to use the full spectrum of sections, i.e., having a section made at each branching point in the tree. The difficulty with this full spectrum approach is that the number of sections is large, and many of the neighboring sections are very similar to one another, and therefore not adding much useful information to the phylogenetic vectors; rather it may inflate the dimension of the resulting super-phylogenetic vectors up to 15,000 or higher, which is beyond the capacity of the classifier used in the study. Instead, a value of 6 is used as the number of snapshots taken in the experiments, which yields a dimension of 2386 for super-phylogenetic vector pairs versus the 1640 for the original phylogenetic vector pairs. One refinement made while generating the super-phylogenetic vector is to first adjust the distances, as defined in neighbor-joining algorithm to remove the molecular clock constraint assumed by the UPGMA:
d(i, j) = D(i, j) - (ri + rj) (5)
ri = ∑k D(i, k)/(|L| - 2) (6)
where |L| is the dimension of the matrix D. Another refinement is to use the same projection procedure introduced in Sato et al's modified mirror tree method Eq(2), only now that |vi> and |u16s> are substituted with the super-phylogenetic vectors. It should be noted that since the background subtraction in tol-mirrortree method and Sato et al's method also utilizes the phylogenetic tree, combining the TreeSec procedure and background subtraction may introduce some redundancy.
The use of the species tree instead of individual protein trees for hierarchical clustering has a twofold effect. One effect is more theoretical; it is to reveal how individual protein trees (embedded in the distance matrices) would differ from the underlying species tree of the host genomes, in the same spirit of subtracting the common background as in [9, 10]. The other effect is more pragmatic; it ensures that the super-phylogenetic vectors thus obtained have the same dimension for all proteins, and therefore can be readily used as input to the support vector machine.
The classifier used here is a support vector machine. As a powerful statistical learning method, support vector machines (SVMs), originally proposed by Vapnik [21, 22], have recently been applied with remarkable success in bioinformatics problems, including remote protein homology detection, microarray gene expression analysis, and protein secondary structure prediction .
There are a couple of reasons to use SVMs. First, the data are already in the vector form, particularly suitable as inputs for SVMs. Second, SVMs have been used to predict protein-protein interaction in previous works [25, 26], though there the different properties of proteins are used. We plan to have a comprehensive study of using SVM on data from different sources, and more importantly, how to combine them for better prediction. Third, SVMs have some inherent advantages over other classifiers, including: 1. quadratic programming to avoid local minima, 2. geometric intuition, 3. lower Vapnik-Chervonenkis dimension leading to better generalization, and 4. amicability with small training samples, which all contribute to its popularity as a classifier adopted in many applications.
The basic idea of SVMs is simple; it is to find a hyperplane that separates two classes of objects, as represented as points in a vector space, with the maximum margin to the boundary lines. Such a hyperplane ensures good generalization and unseen data are then classified according to their location with respect to the hyperplane. The power of SVMs comes partly from the data representation, where an entity, e.g., a pair of proteins, is represented by a set of attributes. However, how those attributes contribute to distinguishing a true positive from a true negative may be quite complex. In other words, the boundary line between the two classes, if depicted in a vector space, can be highly nonlinear. The SVMs method will find a nonlinear mapping that transform the data from the original space, called input space, into a higher dimensional space, called feature space, where the data can be linearly separable.
In general, the mapping can be quite complex and the dimension can be very (even infinitely) high in order for the mapped data to be linearly separable. The trick of SVMs is the use of kernel functions, which define how the dot product between two points in the feature space, which is the only quantity needed to solve the quadratic programming problem for finding the maximum margin hyperplane in the feature space. The use of kernel functions avoids explicit mapping to high dimensional feature space; high dimensionality often poses difficult problems for learning such as over-fitting, thus termed the curse of dimensionality. Polynomial kernel and Gaussian kernel are the two most commonly used generic kernels – linear kernel is not really useful in most cases except when the data are linearly separable. For vectors x and y, Gaussian RBF is defined as
K(x, y) = exp[- (|x-y|2/c)], (7)
and the polynomial kernel is defined as
K(x, y) = [1 + s(x · y)]d, (8)
where c, s and d are parameters adjustable in the software package SVMLight . Both kernels are experimented with the default values for c, s and d, and the Gaussian kernel yielded the best results reported in this paper. Because the polynomial kernel performs significantly worse with the default setting, we also tested with changing the polynomial degree d from the default value (d = 3). We found that the performance is quite sensitive to the degree. Details are given in the next section. Besides using the separation of training and testing as a mechanism to alert us to overfitting, another mechanism built into SVMLight for avoiding overfitting is the use of "soft" margin, i.e., to allow for misclassification for some outlier training data points, and a cap on the number of iterations to stop the optimization process during the training even if the preset error rate is not reached. And we have used SVMLight's default setting for our experiments.
It is worth noting that, overall our method can be viewed as a hybrid that employs in tandem both an explicit mapping, from phylogenetic vectors to super-phylogenetic vectors, and the use of a generic kernel.
Results and discussion
We test our TreeSec method in a series of leave-one-out cross-validation experiments on the data set described above. To prepare an experiment, one of the 13 interacting pairs is selected and reserved as the positive testing example, and 48 non interacting pairs that contain one protein from the positive testing example are reserved as negative testing examples. The rest 325 - 49 = 276 pairs are used as training examples, among which there are 12 interacting pairs as positive training examples. By rotating the positive testing example among the 13 interacting pairs, we can design 13 such leave-one-out cross validation experiments, and the average performance is reported.
Unsupervised (Pearson CC)
Gaussian Kernel SVM
TreeSec × 10
TreeSec × 10 (no NJ)
TreeSec × 10 (random tree)
Polynomial Kernel SVM
TreeSec × 10
TreeSec × 10 (no NJ)
TreeSec × 10 (random tree)
The ROC scores of the mirror tree method and our TreeSec method, with a few variations, are reported in Table 3. Since the mirrortree is an unsupervised learning method, to be fair, we first use TreeSec in an unsupervised learning manner, and compare the two. In this case, since there is no training necessary, for each leave-one-out experiment, only the testing examples are ranked by their Pearson correlation coefficients as if they were scores output from a classifier. The mirror tree method using the phylogenetic vectors prepared via Sato et al's procedure receives a ROC score 0.6587. A slightly higher ROC score (0.6731) is obtained when everything is kept the same except for substituting the phylogenetic vectors with the super-phylogenetic vectors prepared via TreeSec method.
ROC scores for TreeSec × N
Polynomial (d = 2)
The dimension of the super-phylogenetic vectors from TreeSec is obviously higher than that of the phylogenetic vectors in MirrorTree, since the former is derived by concatenating extra bits of information to the latter. Although this may raise concerns with a judicious reader about the fairness for comparing the two approaches if they have different sizes of data, it should not be a problem in our case, because we use the same amount of input data as the Sato et al's approach – the same distance matrices with the same size for proteins and the same distance matrix for species (based on 16S rRNA sequences). The extra bits of information are not really extra; they are the result of how we unravel the information embedded in the input data. In a sense, our method is a hybrid of combining the explicit mapping (to higher dimension) and the use of kernels, which may explain why our method bodes well with the learning task. Nonetheless, care should be taken to not let the increase of dimension go unchecked, as redundancy may arise and lead to overfitting and bad generalization. That is part of the reason that only "snapshots" of evolutionary history are incorporated.
It is not surprising that the neighbor-joining distance adjustment is essential; without it as shown in Table 3 the performance decreases significantly (0.69). To verify that the better performance indeed arises from incorporating the intra-matrix correlations, we run the same experiments on the data prepared using a random species tree, we can see in Table 3 that the ROC scores are consistently low at around 0.5 for cases when a random tree is used.
To further examine the performance, in Figure 4, ROC curves are shown for those ROC scores reported in Table 3 for the unsupervised learning based on Pearson correlation coefficient and for the supervised learning with a Gaussian kernel SVM. Given the Y-axis as the true positives and X-axis as the false positives, the higher a curve means more true positives are identified at cost of a given number of false positives. It is consistent to note that "TreeSec × 10" corresponds to the top curve overall. Also remarkable is the steep slope for the two ROC curves corresponding to the unsupervised learning (CC:MirrorTree and CC:TreeSec) at small false positive rates (X < 0.1). This explains the high specificity for these unsupervised learning based on correlation coefficient, and is consistent with what is reported in [9, 10]. As moving to the right (i.e., when more false positives are made), these two curves quickly lose the upward momentum (i.e., identify fewer true positives), an indication of low sensitivity. Therefore, the supervised learning using the SVMs in these experiments offers a better balance between sensitivity and specificity.
It is worth noting that, the highly skewed learning problem is likely a reflection of situations in the real world, i.e., there are far more negatives than positives. In our case, given n proteins that uniquely interact with only one other member, there are only n/2 positive pairings among the (n2 - n)/2 possible pairings of these n proteins. That is, the interacting network, with nodes representing the proteins and edges representing the interactions, is quite sparse, but our method is still applicable when there are more edges. Because every possible pair of nodes is assigned a score in our method, predictions can be made by going down a list of all pairs that are ranked by their scores in decreasing order. So, regardless the number of the actual edges in the network (sparse or not), the method works, and perfectly so as long as the true interacting pairs are ranked higher than non interacting pairs in the prediction. Indeed, this scheme is also widely used in predicting protein interaction networks in general, both in an unsupervised paradigm such as the original mirror tree method, and in a supervised learning paradigm.
To summarize, in this work we developed a novel, simple method to explore the intra-matrix correlational information embedded in the distance matrices and incorporate such information into a data representation which is conducive to supervised learning. Three methods recognize the importance of the phylogenetic tree, both Sato et al's projection method  and Pazos et al's tol-mirrortree  try to "subtract" from the similarity (measured as correlation coefficients) the effect due to speciation rather than the interaction pressures, whereas our method seeks to "unravel" the intrinsic structure of the distance matrices using the species tree as a guide and then "concatenate" these snapshots of the evolutionary history to the current view (i.e., the original ortholog distance matrices) of the proteins. That is, the main difference between subtracting and adding is that the former is more appropriate for removing background noise so as to reduce false positive and the latter is more appropriate for disentangling intra-matrix correlations so as to aid a supervised learner.
As future work, we will study how the reconciliation between protein trees and species tree can be more explicitly represented and how to associate selection pressure imposed by the interaction to specific evolutionary events, e.g., horizontal gene transfers.
The authors would like to thank the anonymous reviewers for their valuable comments which helped to improve the manuscript significantly.
- Ito T, Chiba T, Ozawa R, Yoshida M, Hattori M, Sakaki Y: A comprehensive two- hybrid analysis to explore the yeast protein interactome. Proc Natl Acad Sci USA 2001, 98: 4569–4574. 10.1073/pnas.061034498PubMed CentralView ArticlePubMedGoogle Scholar
- Uetz P, Giot L, Cagney G, Mansfield TA, Judson RS, Knight JR, Lockshon D, Narayan V, Srinivasan M, Pochart P, Qureshi-Emili A, Li Y, Godwin B, Conover D, Kalbfleisch T, Vijayadamodar G, Yang M, Johnston M, Fields S, Rothberg JM: A comprehensive analysis of protein-protein interactions in Saccharomyces cerevisiae . Nature 2000, 403: 623–627. 10.1038/35001009View ArticlePubMedGoogle Scholar
- Pellegrini M, Marcotte EM, Thompson MJ, Eisenberg D, Yeates TO: Assigning protein functions by comparative genome analysis: Protein phylogenetic profiles. Proc Natl Acad Sci USA 1999, 96: 4285–4288. 10.1073/pnas.96.8.4285PubMed CentralView ArticlePubMedGoogle Scholar
- Marcotte EM, Pellegrini M, Thompson MJ, Yeates TO, Eisenberg D: A combined algorithm for genome-wide prediction of protein function. Nature 1999, 402: 83–86. 10.1038/47048View ArticlePubMedGoogle Scholar
- Enright AJ, Iliopoulos I, Kyrpides NC, Ouzounis CA: Protein interaction maps for complete genome based on gene fusion events. Nature 1999, 403: 86–90.Google Scholar
- Pazos F, Valencia A: Similarity of phylogenetic trees as indicator of protein-protein interaction. Protein Eng 2001, 14: 609–614. 10.1093/protein/14.9.609View ArticlePubMedGoogle Scholar
- Goh CS, Bogan AA, Joachimiak M, Walther D, Cohen FE: Co-evolution of proteins with their interaction partners. J Mol Biol 2000, 299: 283–293. 10.1006/jmbi.2000.3732View ArticlePubMedGoogle Scholar
- Goh C, Cohen FE: Co-evolutionary analysis reveals insights into protein-protein interactions. J Mol Biol 2002, 324: 177–192. 10.1016/S0022-2836(02)01038-0View ArticlePubMedGoogle Scholar
- Sato T, Yamanishi Y, Kanehisa M, Toh H: The inference of protein-protein interactions by co-evolutionary analysis is improved by excluding the information about the phylogenetic relationships. Bioinformatics 2005, 21: 3482–3489. 10.1093/bioinformatics/bti564View ArticlePubMedGoogle Scholar
- Pazos F, Ranea JAG, Juan D, Sternberg MJE: Assessing Protein Co-evolution in the Context of the Tree of Lift Assists in the Prediction of the Interactome. J Mol Biol 2005, 352: 1002–1015. 10.1016/j.jmb.2005.07.005View ArticlePubMedGoogle Scholar
- Gertz J, Elfond G, Shustrova A, Weisingger M, Pellegrini M, Cokus S, Rothschild B: Inferring protein interactions from phylogenetic distance matrices. Bioinformatics 2003, 19: 2039–2045. 10.1093/bioinformatics/btg278View ArticlePubMedGoogle Scholar
- Ramani A, Marcotte EM: Exploiting the co-evolution of interacting proteins to discover interaction specificity. J Mol Biol 2003, 327: 273–284. 10.1016/S0022-2836(03)00114-1View ArticlePubMedGoogle Scholar
- Jothi R, Kann MG, Przytycka TM: Predicting protein-protein interaction by searching evolutionary tree automorphism space. Bioinformatics 2005, 21: i241-i250. 10.1093/bioinformatics/bti1009PubMed CentralView ArticlePubMedGoogle Scholar
- Xenarios I, Rice DW, Salwinski L, Baron MK, Marcotte EM, Eisenberg D: DIP: the Database of Interacting Proteins. Nucleic Acids Res 2000, 28: 289–291. 10.1093/nar/28.1.289PubMed CentralView ArticlePubMedGoogle Scholar
- Kanehisa M, Goto S, Kawashima S, Okuno Y, Hattori M: The KEGG resource for deciphering the genome. Nucleic Acids Res 2004, 32: D277-D280. 10.1093/nar/gkh063PubMed CentralView ArticlePubMedGoogle Scholar
- Felsenstein J: PHYLIP (Phylogeny Inference Package) version 3.6. Distributed by the author. Department of Genome Sciences, University of Washington, Seattle; 2004.Google Scholar
- Jones D, Taylor WR, Thornton JM: The rapid generation of mutation data matrices from protein sequences. Compt Appl Biosci 1992, 8: 275–282.Google Scholar
- Katoh K, Kuma K, Toh H, Miyata T: MAFFT version 5: improvement in accuracy of multiple sequence alignment. Nucleic Acids Res 2005, 33: 511–518. 10.1093/nar/gki198PubMed CentralView ArticlePubMedGoogle Scholar
- Gustafson AM, Allen E, Givan S, Smith D, Carrington JC, Kasschau KD: ASRP: the Arabidopsis Small RNA Project Database. Nucleic Acids Res 2005, 33: D637-D640. 10.1093/nar/gki127PubMed CentralView ArticlePubMedGoogle Scholar
- Durbin R, Eddy S, Krogh A, Mitchison G: Biological Sequence Analysis. Cambridge University Press; 1998.View ArticleGoogle Scholar
- Vapnik V: The Nature of Statistical Learning Theory. Springer; 1995.View ArticleGoogle Scholar
- Vapnik V: Statistical Learning Theory. Wiley; 1998.Google Scholar
- Cristianini N, Shawe-Taylor J: Introduction to Support Vector Machines. Cambridge University Press; 2000.Google Scholar
- Scholkopf B, Tsuda K, Vert J-P: Kernel Methods in Computational Biology. MIT Press; 2004.Google Scholar
- Bock JR, Gough DA: Predicting protein-protein interactions from primary structure. Bioinformatics 2001, 17: 455–460. 10.1093/bioinformatics/17.5.455View ArticlePubMedGoogle Scholar
- Gomez SM, Noble WS, Rzhetsky A: Learning to predict protein-protein interactions from protein sequences. Bioinformatics 2003, 19: 1875–1881. 10.1093/bioinformatics/btg352View ArticlePubMedGoogle Scholar
- Joachims T: Making large-scale SVM learning practical. In Advances in kernel Methods – Support Vector Learning. Edited by: Scholkopf B, Burges C, Smola A. MIT Press; 1999:169–184.Google Scholar
- Gribskov M, Robinson N: Use of receiver operating characteristic analysis to evaluate sequence matching. Computers and Chemistry 1996, 10: 25–33. 10.1016/S0097-8485(96)80004-0View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.