Predicting gene function using few positive examples and unlabeled ones

Background A large amount of functional genomic data have provided enough knowledge in predicting gene function computationally, which uses known functional annotations and relationship between unknown genes and known ones to map unknown genes to GO functional terms. The prediction procedure is usually formulated as binary classification problem. Training binary classifier needs both positive examples and negative ones that have almost the same size. However, from various annotation database, we can only obtain few positive genes annotation for most offunctional terms, that is, there are only few positive examples for training classifier, which makes predicting directly gene function infeasible. Results We propose a novel approach SPE_RNE to train classifier for each functional term. Firstly, positive examples set is enlarged by creating synthetic positive examples. Secondly, representative negative examples are selected by training SVM(support vector machine) iteratively to move classification hyperplane to a appropriate place. Lastly, an optimal SVM classifier are trained by using grid search technique. On combined kernel ofYeast protein sequence, microarray expression, protein-protein interaction and GO functional annotation data, we compare SPE_RNE with other three typical methods in three classical performance measures recall R, precise P and their combination F: twoclass considers all unlabeled genes as negative examples, twoclassbal selects randomly same number negative examples from unlabeled gene, PSoL selects a negative examples set that are far from positive examples and far from each other. Conclusions In test data and unknown genes data, we compute average and variant of measure F. The experiments showthat our approach has better generalized performance and practical prediction capacity. In addition, our method can also be used for other organisms such as human.


Background
One of the important challenges in the post-genome era is determining the functional role of all genes in the cell although about one-third of the genes have been annotated and deposited in database such GO(gene ontology) [1]. With the recent invention of several large-scale experimental methods, a wealth of functional genomic data was accumulated, including sequence, micro-array expression profile and protein-protein interaction data.
These large data-sets have fueled an interest in computational approaches to gene function prediction, which promises to harness the information present in these large collections of data to automatically deduce accurate gene annotations [2,3]. Furthermore, many works have shown that integration of different kinds of data sources can considerably improve prediction results [4,5]. GO is a widely-used set of functional terms with which some genes are annotated, we also call functional terms as functional classes in related to classification problem from machine learning. GO functional annotation associates each gene or gene product to some functional terms. For an unknown gene, predicting its functions will assign some GO functional terms to it, which is called multi-label classification problem in machine learning community. The mainstream approach is to transform it into a binary classification task for each functional class, which focuses on training a classifier such as SVM (support vector machine) with some labeled positive and negative examples. However, the available information from the annotation databases, such as GO [1], is only about positive examples, i.e. for a functional class, we only know which gene is assigned to it, but we are not sure that a gene has no this function except for too few genes. As a result, when training classifier for a functional class, we can only obtain  [6]. We call these two methods twoclass and twoclassbal algorithm respectively.  [6,7]. These methods eliminate the impact of imbalanced problem, but, only few negative examples can be selected, as a result, the classifier is trained on a small training set and easily leads to over-fitting. When we use GO annotation, many functional classes have few annotated genes, which will lead to a lower prediction accuracy and need to be solved [9]. In this paper, aiming at both imbalance and over-fitting problem for genes function prediction with only few positive examples and unlabeled examples, we propose a novel strategy for predicting genes function using SVM. Firstly, we create some synthetic positive examples with few negative noises to enlarge positive examples set P. Secondly, we extract a representative negative example set RN from unlabeled genes U using SVM iteratively. Finally, an optimal SVM classifier with RBF (Radial Basis Function) is trained by using Grid-search technique. This method is called SPE_RNE(Learning classifier by Synthetic Positive Examples and Representative Negative Examples).

Experiment setting Data sets
Gene annotation We used gene ontology and corresponding gene function association of Yeast [10] released in April 2007. Gene association file contained 5,873 genes ,the number of known and unknown genes is 3,796 and 2,077 respectively. We up-propagate the gene annotation along GO hierarchical structure and obtained a reduced GO which has only 99 GO terms under guidance of biological experts. To compare the algorithm performance, we divide them into four groups according to number of annotated genes as shown in Table 1. There are 53 functional classes with annotated genes less than 60 among total 99 terms. Protein sequence The protein sequence of all of the Yeast genes were downloaded from SGD [10].We applied the Smith-Waterman pairwise sequence alignment algorithm [11] to these sequences. Each protein is represented as a vector of Smith-Waterman log Evalues, and computed with respect to all 5,873 Yeast genes. A 5873*5873 similar matrix is obtained. Microarray expression profile Microarray datasets are real-valued matrices measuring gene expression levels under different experimental conditions. We use gene expression microarray data from the Stanford Microarray Database(SMD) [12] containing results from several publications, providing a total of 294 real-valued features for all 5,873 genes. Microarray entries typically include missing values due to experimental imperfections. We estimate such entries using the widely accepted KNNimpute algorithm [13] with default k value. Then, we computed similarity between two genes using Gauss kernel with g = 2. The second 5873*5873 gene similar matrix is generated. Protein-protein interaction We downloaded the protein-protein interaction data from BioGRID2.0.30 [14]. Protein-protein interaction data is described as a graph in which nodes denote protein and edges denote interaction and diffusion kernel [15] with diffusion constant b = 2 is used to measure the similarity between two proteins. Each gene is also represented as a vector of similarity with respect to all 5,873 genes. The third gene similar matrix is computed. Several previous researches have shown that integrating various genomic data to predict gene function can improve prediction accuracy [4,5]. In this paper, we add three similar matrices and obtain a sum matrix. It is noticeable that each matrix should be centralized and normalized to eliminate the effect from major data before adding them [4]. While training SVM classifier, this pre-computed kernel matrix is used.

Experiment setting and evaluation
We used LIBSVM [16] to implement SPE-RNE and related algorithms two class SVM twoclass, two class balanced SVM (twoclassbal)and PSoL in matlab. First, we divided 3796 known genes into training set and validation set, after training SVM classifier on training set, the generalized performance of algorithms were compared on validation set. Widely-accepted measures, including precision rate P, recall rate R and their combination F1, are used. Their definitions are as follows: where TP,FP and FN denote the number of true positive, false positive and false negative respectively. Then, using 2,077 unknown genes released in April 2007 as test examples, we predict their functions and evaluate ROC (Receiver operating characteristics) score with gene function association released in December 2008 as annotation standard.

Performance comparison on known genes
For each functional class, 3796 genes are in two categories: genes assigned to this functional class and unlabeled genes. we randomly select 20 percent from these two categories as validation set, the others are training set. SVM classifiers are learned on training set and used to predict genes functions on validation set to evaluate generalized performance of algorithms.
When  Table 2 shows the number of NaN for each functional group.
As shown by table 2, twoclass has the most serious imbalance and PSoL has more serious imbalance, but our method SPE-RNE, like twoclassbal, doesn't suffer from imbalanced problem at all because we select reasonable quantity of negative examples after enlarging the positive examples set. In addition, functional classes with few annotated genes have more serious imbalance.
For twoclassbal algorithm, while serious imbalance does not occur, the over-fitting may arise to affect prediction performance due to few training examples. To evaluate the algorithm fairly, we set F1 = NaN to F1 = 0. For each algorithm and functional group, the means and variances of F1 are listed in Table 3 and Table 4 respectively.
In table 3, although twoclassbal has not class imbalance problem, but it has worst performance because it has too few training examples, which easily causes overfitting. For functional classes with few annotated genes, the existed algorithms, like twoclass, twoclassbal, PSoL have lower F1 values, our algorithm significantly improves the F1 values in this case. For functional classes with more annotated genes, our algorithm has better performance too. While only few training examples are used to learn SVM classifier, over-fitting problem may occur and make algorithms unstable. We compute the variances of F1 for each functional group Table 2 The number of functional class whose F1 = NaN for four algorithms and four functional groups  Table 2 shows the number of NaN for each functional group.  to evaluate the stability of algorithm. Table 4 shows that our algorithm has good stability.

Predicting performance on unknown genes
Since April 2007, some of 2077 unknown genes have been annotated with some functions. We consider these 2077 genes as test examples and use trained SVM classifier to predict function for them. The GO function association released in December 2008 is regarded as complete annotation, that is, for each functional class, if a gene is assigned to it, the label is set to 1, otherwise -1. For each algorithm, the ROC score, which is area under ROC, is evaluated as comparison measures. In previous section, we use F1 to evaluate algorithm performance because we think that GO function association released in April 2007 is incomplete. The ROC scores are listed in Table 5.
For group 1 and 2, our algorithm significantly improve the ROC score, which illustrates better prediction performance for unknown genes.  Figure 1 for each group. In each group, our algorithm can recall more positive genes on average.

Predicting result on unknown genes
We list predicted functional classes for ten genes with most predicted functional terms in Table 6, these genes were unknown in April 2007, but they were annotated with one or multiple functional classes in December 2008.

Conclusions
In this paper, We propose a novel approach to predict- For SPE-RNE, the validation on known gene data set shows its best F1 value and good stability. Prediction on unknown genes set illustrates its higher ROC scores and better prediction performance than several classic algorithms. All the algorithms run in a sum matrix which is obtained by adding simply several similarity matrixes from heterogeneous data sources, which may loss some information. How to integrate effectively these heterogeneous data to predict gene function is our next research subject in future. In addition, our method can also be used for other organisms such as human.    ROC score, that is area under ROC curve, is a widely-accepted performance measure for prediction classification problem. The larger ROC score is, the better the performance of classification algorithm is.

Proof
Let the classification hyperplane be According to Vapnik's theory minimizing empiric risk [19], following two inequalities are correct with probability close to 1: That is,  e = ae + (1 -a)e′ is a positive example with probability close to 1. Therefore, the synthetic example  e is probably positive example and the enlarged set of positive examples,  P has few negative noise. In fact, from the point of view of algebra,  e is convex combination of e and e′, and from the point of view of geometry,  e is a random point in line segment from e to e′. We have following algorithm 1 for creating synthetic examples.
In our experiments, all the functional classes are divided into four groups according to number of annotated genes. We enlarge positive example set with different times for different groups and set k = 10, which is shown in Table 7. Particularly, we did not make Table 6 The ten genes with the most predicted functions and their predicted functions  Table 6 lists ten genes with top predicted functional classes and their predicted functional terms. The first column is gene name and other columns are functional classes predicted correctly by algorithm SPE-RNE.
synthetic positive examples for functional classes with more than 300 annotated genes. Step 1, identifying a reliable initial negative examples set. For gene vectors, only distance-based similarity can be used and the most dissimilar genes to positive examples are assumed as reliable negative examples. In PSoL, initial negative set N met two conditions:  all elements in N are most dissimilar to positive example set  .  elements in N are far from each other. Since this problem is NP-hard, an approximate solution was used.
In our algorithm, one-class SVM [20] is utilized to extract efficiently initial negative examples. Give a percentage of negative examples, such as 10 percent, it can draw an initial decision boundary to cover most of the positive and unlabeled examples. The data points not covered by the decision boundary can be regarded as negative example points because these data points are far from the positive set in the feature space.
Step 2, learning SVM iteratively to move classification hyperplane to an appropriate place. In this iteration process, a set of classifier, C S and corresponding set of negative example set, N S are obtained. In i th iteration, we obtain not only SVM classifier C i but also negative example set N i+1 to be used to train next SVM N i+1 = N pred ∪ N SVs (5) where N pred is the most reliable negative examples predicted in current iteration and N SVs is negative support vector set of C i . To avoid class imbalance, the size of N pred is set as N pred ≤ m|  P | and many experiments shown that m = 3 is suitable. In addition, only the negative support vectors of C i are selected as representatives of previous negative training examples. In this iteration process, |U| becomes smaller and smaller. When only few unlabeled examples are remained, the N i may has more false positive examples and the classifier may become bad, therefore, we stop iterating. According to a large number of experiments, the stopping criteria is set as: This process is inspired by [21] and can be intuitively demonstrated using figure 2, 3, 4, 5. In figure 2 Step 3, selecting the best representative negative set. The representative negative set should have best classification performance, therefore, we use each SVM from Cs to classify a validation set V that is selected randomly form  P and U with 10 percent of total respectively at the start of algorithm. The discrimination ability of the trained classifiers is evaluated with F1 . Accordingly, the negative set corresponding to the best classifier is returned as the representative negative samples RN. In stead PSoL selects final classifier to classify remaining examples, our algorithm selects best classifier according to classification performance on validation set and corresponding negative set is regarded as representative

Algorithm 2
Algorithm selecting representative negative examples 1: function SELNEGEXPS(  P , U) 2: randomly selecting 10 percent of  P and U respectively as validation set V 3:  ′ P =  P -V, U′ = U -V; 4: identifying the initial reliable negative examples 5: Training one-class SVM classifier C 0 based on  ′ P and U′; 6: Classify U′ using C 0 . The predicted negative set N 1 is used as the initial negative training set 7: U′ = U′ -N  [22] is used to search the optimal parameters c and g and an optimal SVM classifier can be successfully obtained.