We have attempted to perform predictions over nine common phage genes using SynFPS. These are major head, major tail, tape measure, prohead protease, integrase, terminase, portal, holin and lysin genes. They were selected on the basis of regular existence – they encode necessary functions not provided by their hosts, including structural and assembly genes, as well as lysis genes 16. These genes were searched against the annotation database using regular expression patterns defined in Table 1. Manual modifications of the search results have been conducted to remove ambiguous entries.

Table 2 indicates the amount of genes that can be detected if sequence alignment (BLAST) alone was used. The K-Means clustering result based on these genes can be found in Supplementary Material (see Additional file 1).

We perform leave-one-out (LOO) cross validation to evaluate the prediction performances for these genes. For each gene function, we run the cross validation in each cluster individually over a discrete range of values of the kernel parameter – σ for Gaussian RBF kernel 17. The σ value that gives the best accuracy is chosen and is used for all future predictions for that function. The prediction accuracies shown in Table 3 are the averages of cross validation results across all the clusters.

K-fold cross validation may also be used to evaluate the prediction performances and it is expected that accuracies are lower with a smaller K value. For instance, the prediction accuracy for Terminase is 79.8% for K = 4 and 62.3% for K = 2. However, LOO is more suited to our overall purpose – one primary objective of the cross validation is to find out the near optimal σ value for the gene class to perform future predictions. Since most clusters contain only a very small portion of genomes that require genuine prediction, they are best simulated by LOO, where only one genome is taken out for prediction testing at a time.

The prediction accuracies are averaged at ~80%. The 100% prediction accuracy of lysin can be explained by the strong context relationship between lysin and holin. Since the presence of a lysin is always accompanied by the presence of a holin immediately beside it 18, SynFPS can easily identify the lysin gene if it already knows the position of the holin. However, the converse is not true: the identification of holin genes may not depend upon the presence of lysin. Consequently, the prediction accuracy for holin is not as high.

These prediction accuracies reflect the sensitivity of the system (true positives/(true positives + false negatives)). The specificity of the system (true negatives/(true negatives + false positives)) on the other hand is always larger the sensitivity because of two system features. Firstly, we allow only a single positive prediction for each genome (see Methods). Thus, the number of false negatives is always the same as the number of false positives, implying that the specificities always scale together with the sensitivities. Secondly, the number of negative training data (hence true negatives) is always larger than the number of positive training data (hence true positives), and consequently Specificity > Sensitivity. One reason for using LOO cross validation accuracies to evaluate the system is the lack of benchmark for our problem. However, it may be noteworthy that other genomic-context based methods for the prediction of functional elements have similar reported accuracies ranging from 72% to 80% 6.

We have examined the effect of the K-Means adaptive threshold t on the prediction accuracies. The value of t ∈ (0,1] implicitly specifies the maximum tolerable distance between any two genomes within a cluster. As a result, as t → 0, there are as many clusters as the number of genomes, and as t → 1, there is only one cluster. Both of these cases do not provide useful information for prediction. Since there is no analytical method to find out a good value for t, we have run SynFPS over a range of values from t = 0.05 to t = 0.3. Values outside this range generate either too many or too few clusters (average number of genomes per cluster < 2 or number of clusters < 3 respectively). Using different t values lead to a different amount of genomes that are covered by the automated prediction (a.k.a. prediction coverage). Genomes within the "coverage" are those for which SynFPS has made a classification decision; the remaining genomes are discarded or ignored by SynFPS. Here are examples of genomes not in coverage:

• genomes not containing the gene being predicted (discarded during cross validation only)

• genomes that is in a cluster on their own

• within a cluster, if there are fewer than two genomes that contain the gene being predicted, then all the genomes are discarded

• genomes with genomic context different to the consensus of the group may be discarded

Figure 3 shows the plot of prediction accuracies versus prediction coverage. The highest coverage values for all gene functions are about 20–25%, achieved by using a t value ~0.1. The results indicate that we can obtain a higher accuracy by lowering the coverage. However, the ultimate purpose of the system is to make genuine predictions over the genomes that lack identification of the genes being predicted. Lowering the coverage can lead to ignorance of many of these genomes. Consequently, one must find a balance between the accuracy and the coverage according to the intended task.

Using the maximum coverage and the σ values optimized by LOO cross validation, we have generated predictions over genomes within which certain gene functions were not already detected. The outcome of SynFPS is to identify which genes within those genomes correspond to the functions of our interest. The prediction outcomes are listed in Table 4.

Three genes that we have predicted functions for have no existing functional annotation in the database (marked uncharacterised in Table 4). Seven genes in Table 4 exhibit sequence similarity to their reference genes, suggesting that their predicted functions are supported by both sequence similarity and the genomic context information embedded in our system, such as gene order conservation and positional coupling. For other genes that show no sequence similarity (a total of 12 of them in Table 4), the predicted functions are only evident by the genomic context. It is noteworthy that sequence alignment based methods would have failed in finding correspondences to these genes. Other prediction results have complemented existing annotations in the database in cases where they do exist, and therefore support the validity of our approach.

We present a novel method called Synteny-based Function Prediction System (SynFPS) capable of predicting gene functions among completed genomes based on the conservation of gene order (synteny) and the conservation of gene-to-gene distance. An overview of SynFPS is shown in Figure 1. The genome annotation database as shown in the figure defines the scope of analysis for the system. In our work, it consists of 296 phage genomes retrieved from GenBank (see Additional file 1).

SynFPS runs on Windows and is publicly available. It was developed in C# and requires the free Microsoft .NET Framework 2.0 to run. Bioperl 1.4 23 is needed for data retrieval from public databases. Workstations with a single CPU of ~3.0 GHz and 1 GB of RAM are sufficient for reasonable performance over a collection of ~300 phages.

The system begins by identifying in the database a collection of genes that correspond to a set of user-specified gene functions. Instead of using sequence similarity as in many other methods 124512, SynFPS identifies functionally similar genes using regular expressions 24. For example, to search for genes that encode the major head proteins of phages, one possible regular expression pattern is "(?<!minor)\b(head|capsid) protein". With this pattern, we are including genes that have been annotated with "head protein" or "capsid protein" except those with the prefix term "minor". The use of regular expression is aimed at tackling annotation discrepancies among coding sequences in databases that do not have vocabulary control. The regular expression syntax used in SynFPS follows the syntax defined for the .NET Framework 25.

Once a regular expression pattern is given, the system searches against the annotation data of all the genomes that have been supplied to the program. By default, it will identify coding sequence (CDS) regions in each of the genome and then try to match the patterns against their annotated features such as "product", "function" and "note". The set of annotated features that the search will perform over is customisable by the users. The search results can be visually displayed, where the genomes and matching genes are illustrated. The display is interactive in which annotations can be viewed and search results can be modified via manual addition and removal of genes.

Although genome annotation processes are often assisted by sequence alignment, many annotations are prepared manually by biologists who conducted research on the genomes. Therefore, the set of sequences found by annotation search could embrace functionally similar genes that show no sequence similarity. In the results section, we provide an assessment on sequence alignment in relation to regular expression search.

The annotation search process leads to a mapping of genes across the genomes. This mapping provides the necessary information for a context based clustering. Let G = {g₁, g₂,..., g_n} be the set of all gene functions where g is a symbol representing a function and n is the total number of functional classes identified. Let m be the number of genomes in the database. We define X_k ⊆ G, k = 1,2,..., m to be the set of genes detected in genome k and C_kl = C(X_k, X_l) = X_k ∩ X_l to be the common set of genes between genomes k and l. The genomic-context distance between two genomes k and l is defined as:

D k l = ∑ g i , g j ∈ C k l ; i < j [ d k ( g i , g j ) − d l ( g i , g j ) ] | C k l | + p ( | X k ∪ X l | − | C k l | )       ( 1 ) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBamXvP5wqSXMqHnxAJn0BKvguHDwzZbqegyvzYrwyUfgarqqtubsr4rNCHbGeaGqiA8vkIkVAFgIELiFeLkFeLk=iY=Hhbbf9v8qqaqFr0xc9pk0xbba9q8WqFfeaY=biLkVcLq=JHqVepeea0=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@8F3A@

where d_k(g_i, g_j) = location of g_j - location of g_i in genome k, |s| denotes the size of a set s and p is a parameter to penalize the genomes not sharing the same set of genes. The summation term dictates the conservation of gene order as well as the conservation of gene-to-gene distances between the two genomes. The second term dictates gene co-occurrence.

We represent each genome k by a vector of distance values: F_k = [D_k1, D_k2,...,D_km] and then we perform K-Means clustering over the set S = {F_k | k = 1,..., m}. We implemented an adaptive technique such that the number of clusters grows incrementally until the size of the largest cluster is smaller than a specified threshold. The threshold t ∈ (0,1] describes the fractional size of the Euclidean space spanned by S. Each resulting cluster contains genomes with high resemblance in gene distribution. Alternative adaptive clustering methods include dynamic self-organizing maps 2627.

The clusters of genomes are analysed separately and individually in the last stage of the system. For each cluster, we use the information of the previously identified genes to predict the functions of other genes that exhibit similar context. This is achieved by extracting a set of genes from the cluster and converting them into positive and negative training data for a discriminative classification. Positive data are formed by the group of genes previously identified by the system during the match of regular expression plus any manually added genes, with each gene function representing one class. Negative data comprise the genes that are neighbours to the positive genes. The size of neighbourhood is determined by the statistics of the gene locations in that particular cluster. We use 99% confidence interval on the gene locations of each class to determine the range in which neighbour genes are to be included. This interval also determines the set of candidate genes on which function predictions are performed (see Figure 2). The discriminative classification is carried out by a Support Vector Machine (SVM) 28, which has been reported with superior results in a variety of biological applications 293031. For each gene function, the SVM produces a binary result on each candidate gene indicating whether or not the gene belongs to that function class. Since the number of gene functions is specified by the user and is not likely to cover every possible function, only a subset of the candidate genes – those with positive results – will eventually be assigned with predicted functions.

To enhance prediction accuracy, we force a unique positive prediction in every genome within a cluster. This is based on an assumption that all pairs of genomes within a cluster would have a one-to-one mapping of genes (gene correspondence). The decision values generated by SVM depict the relative positiveness of each candidate gene. Consequently, the gene with the strongest decision value will be chosen as the positive prediction.

In order to apply SVM, each gene is converted into a numeric vector capturing the following features: composition, normalized van der Waals volume, hydrophobicity, polarity 3032, pairwise similarity scores against other genes in the database 29, relative position and gene size. To compute the "relative position", the system first finds the gene class which has the most conserved distance to the gene under current prediction. For example, as demonstrated in Figure 2, if we are making predictions over class A, then class B will be chosen as the reference for computing the relative positions because the distances between class B genes and class A genes are the most conserved. The relative position of a gene in class A is then computed as the distance between itself and the class B gene in the corresponding genome.

The pairwise similarity scores have been observed to improve classification accuracies. These scores represent the distance between a gene and every other gene in the database 29. However, it should be emphasized that while these sequence similarity scores enhance the strength of the feature vectors, the system does not rely upon similarity significances to detect gene correspondence.