Disease gene properties

Proteins derived from disease genes have been found to have properties that distinguish them from all other genes: they are longer [12], more conserved, phylogenetically extended and without close paralog [13]. In addition, when compared against housekeeping genes, they present different patterns of conservation, function and DNA coding lengths [14]. Since inherited disease genes are more likely to be non-essential, one could hypothesize that they arrive later in the evolution of the human species. Surprisingly, Domazet-Loso and Tautz [15] studies showed that non-essential disease genes are of ancient origin. In agreement with previous findings about the disease gene length, their analysis also showed that ancient genes tend to be longer. The authors, confirming what others had reported, found no significant differences in the rates of evolution of disease versus non-disease genes [14]. The question of evolutionary divergence of disease genes, however, remains open, with some findings indicating that there is a higher rate of non-synonymous substitutions than synonymous ones [16] and others the opposite [14, 17].

If disease genes were to evolve at higher speed, could this be just an effect of the weak dominance of these disease genes? To answer this question, Osada et al. [18] compared evolutionary rates and the degree of polymorphism of the dominant and recessive disease genes. They found a higher rate of non-synonymous polymorphisms in recessive genes. In their analysis, the differences in selection intensity are still significant even after taking into account the dominance, suggesting that there are significant differences in the deleterious effect of the dominant and recessive genes.

The interaction network of disease genes has been the subject of many studies [19, 20]. The relevance of protein interactions in diseases and the development of computational tools applied to disease gene identification has been previously discussed by Kann [21]. Protein interaction networks have been studied for Alzheimer's disease [22], ataxias and disorders of Purkinje cell degeneration [23] and for cancer genes [19]. Genes involved in the same disease have been found to form subnetworks [24]. Finding functional modules associated with each disease could reveal important aspects of the disease mechanisms and aid in disease classification [25]. Feldman et al. [26] compared disease genes against essential genes (from mouse orthologs of human genes) in terms of their connectivity and found the two set of genes to be clearly distinct. In their analysis, the authors used a network of interactions derived from the analysis of hundreds of articles obtained with the Gene Ways [27] natural language system. The resulting network includes almost 13 000 physical interactions and 4458 genes. This work exemplifies the impact of text mining in the field of translational bioinformatics for the gathering of data from millions of existing manuscripts.

In addition to the extensive study of the structured regions of the protein and their effect on disease, the study of intrinsically disordered regions of the proteins from disease genes has recently lead to ‘unfoldomics’, or mapping of disordered proteins to human diseases [28, 29]. A number of intrinsically disordered proteins have been shown to be associated with cancer [30], cardiovascular disease [31], diabetes, neurodegenerative diseases [32] and other human diseases [33, 34].

To summarize, not all the properties that characterize disease genes have been probed or can be easily explained. Previous studies require functional, evolutionary and statistical hypotheses to explain the observations about disease genes. Disease genes might need to interact with each other and might also need to be co-expressed as they participate in the same functional pathways. Fewer paralogs within the human genome might explain the inability of the system to compensate for disruptions created when these genes are modified. Longer genes can be explained statistically as they will have more possible sites for mutations. Figure 2 depicts the distribution of lengths of disease and non-disease genes (from OMIM [35] and RefSeq [35], respectively). Based on the two-sample Kolmogorov–Smirnov test statistic, the distribution of lengths for the disease and non-disease genes are significantly different with a P-value of 3.0e-21. In addition, we estimated the number of protein domains (from CDD [36]) of disease genes (from OMIM) to be higher, on average, than non-disease genes (Kolmogorov–Smirnov test statistic with a P-value 1e-6).

Open in a separate window

Figure 2:

Distribution of length of proteins from disease and non-disease genes (black and gray, respectively). Disease genes (from OMIM [99]) are significantly longer than non-disease genes (RefSeq [35]).

Overview of the methods

Methods for gene prioritization rely on the information provided by one or more of the experimental techniques described above. Therefore, the amount and quality of the available experimental data generated by these techniques is a major limitation of the gene-prioritization techniques. For instance, protein–protein interaction-based methods suffer from the incompleteness and low quality of the data currently available for interaction networks in mammals. Another source of uncertainty is the disease mapping information used to train and evaluate the computational methods, for it is of variable resolution and expected to contain large numbers of false positives. Furthermore, gene-prioritization methods have been hampered by the complexity and difficulty in creating functional and disease ontologies. Methods that rely on text mining, also face the difficulties inherited from natural language processing, such as issues related with extracting gene names from the biomedical literature [73].

Prioritizer, developed by Franke et al. [74] is available for download in their site (Table1). Franke et al. [74] studied the effect of using three different gene networks—GO, PPINT and GEXP—to correctly rank the disease genes for a set of 96 disorders with 409 known disease genes. Combining PPINT and GEXP, a better ranking was achieved than what could have been obtained randomly. The method showed considerable improvement (represented by an increase in the area under the ROC curve) when GO was added. The best ranking of disease genes was reached when the three types of data were combined. The authors used the combination of all sources to prioritize genes in artificial susceptibility loci and found a 2.8-fold increase in the chance of detecting disease genes with respect to random selection.

Another tool, PROSPECTR, uses an alternating decision tree which has been trained to differentiate between genes ‘likely to be involved in disease’ and ‘genes unlikely to be involved’ in disease [75]. The method uses gene properties that are characteristic of disease genes (see previous section) to provide each gene with a score, which is a measure of confidence in the classification. In a test set of 675 genes from the Human Gene Mutation Database [76], and 675 picked at random from Ensembl (had no association with disease), PROSPECTR performed with a sensitivity of 0.71 and a specificity of 0.58. SUSPECTS adds an extra layer to PROSPECTR, using GO annotation, protein domain and gene expression data to rank and score each gene [77]. The program scores each gene of the test set based on its relationship to the networks in the training set. Similarly, ENDEAVOUR uses a larger set of over 12 data sets (listed in Table 1) and order statistics to score and rank genes from the test set [78]. The test set of disease genes used to benchmark SUSPECTS was, on average, within the top 13% of the candidates, while results using ENDEAVOUR indicate that disease genes from a test set of 200 genes ranked 13 on average, representing a 7- and 9-fold enrichment over random classifiers for each method, respectively.

GenTrepid uses two methods: common module profiling (CMP), based on similarity of protein-domain composition, and common pathway scanning (CPS), based on common protein interactions and metabolic pathways among disease genes [79]. George et al. [79] found the two methods to be complementary, with the combination of the two approaches yielding the best performance. However, a meta-analysis combining both methods into a consensus would have decreased the performance compared to using both methods independently. When used side by side, the two methods were reported to have a sensitivity of 0.52 and a specificity of 0.97 in a benchmark of 170 genes (29 diseases) representing a 13-fold enrichment in disease genes. In other words, a list of 100 gene candidates could be reduced to 8 with significant cost and time reduction in the posterior experimental analysis of these candidate genes.

CAESAR, developed by Vision and co-workers [80], is a tool primarily based on text mining of disease information mainly from review articles and the integration of other gene data (Table 1). The authors have addressed the challenge of analyzing complex traits, In their study of 18 genes complex human trait susceptibility genes, CAESAR selected 7 of the genes within the top 2% of the ranked genes. From almost 15 000 genes, 16 of the 18 genes were ranked with a median rank of 549.5. This represents a 67-fold average enrichment.

PhenoPred, developed by Radivojac et al. [81] studies the network of interactions and functional relationships of the target protein and detects its local neighborhoods in the network. The authors devised a supervised approach to find local signatures of the disease and find new candidate genes that are not necessarily in close proximity to the known disease genes. PhenoPred can be queried using a gene or a disease name. If the input is a gene, the program will return a list of diseases that the gene could be associated with (based on the network properties of the genes known for that disease). If starting with a disease name, the program retrieves all the genes predicted to be related to the disease. In both cases, PhenoPred provides a similarity score that represents the chance of the gene–disease association to be true. The authors showed that this approach works best when combined with the molecular function of the query gene and physicochemical properties of its protein product.

ToppGene Suite [65] integrates a vast number of genomic data from humans and mice (Table 1). This state-of-the-art resource includes ToppFun and ToppGene methods that can be used for the analysis of gene functional enrichment and for the prioritization of disease gene candidates, respectively. It uses a fuzzy-based similarity measure between the genes in the training and test set based on their semantic annotation. It also derives the probability (P-value) that each annotation is related to the gene in question, using random sampling of the whole genome. The authors analyzed 20 gene–disease associations from five disorders (from recently reported GWAS) and found that ToppGene ranked 19 of 20 candidate genes within the top 20%. The mean rank for ToppGene was 6.8 (excluding diseases that lacked interaction data [65]).

Lastly, CGPRIO, a tool recently developed by Furney et al. [82] is based on gene properties such as length and structure for identifying those features that characterize cancer genes. Based on distinguishing features, a naïve Bayes model is used to classify genes as proto-oncogene or tumor suppressor genes (Table 1).

From the user's perspective, the most desirable features for these methods are:

1. Online availability: all the methods in Table 1 are freely available and can be interactively used within their websites, with the exception of CAESAR and Prioritizer, which are downloadable from their sites as stand-alone applications.

2. Advanced interface for input of training and test sets: the option of a custom-made list of genes for training is highly desirable for biologists (input of ENDEAVOUR and ToppGene). Additional input options such as those found in SUSPECTS and GeneTrepid (including disease name or keywords within the gene description) facilitate the analysis when no customized training set is available.

3. Clear summary of the results with an overall ranking of the genes: some methods like ToppGene and SUSPECTS provide both a measure of the likelihood that the gene is responsible for the disease or associated with the training set and also a clear visualization of the results of ranking the genes with each feature used in the process.

In summary, the methods for disease gene prioritization have led to an improvement in the detection of disease genes and to an increase in our knowledge about the integration of the several data sources for gene function and disease association. However, these methods can only be as accurate as the data they are based upon, which is an important issue, given the low quality of some of the experimental data on which they rely (e.g. protein–protein interaction data is incomplete and unreliable). Producing good ontologies for complex processes and improving the methods for mining and integrating the multisource data are difficult tasks that, unless addressed, will continue to severely limit the progress of gene-prioritization techniques.

Analysis of structural variants

The study of variations of the human genome is not limited to the analysis of SNPs. Other structural variants can also be linked to diseases (see articles [116, 117] and reviews [118, 119]). These structural variants include duplications, inversions and deletions that can currently be identified by array comparative genomic hybridization (aCGH) and paired-end mapping [120–127]. Addressing the need for a common framework of reference for structural variant comparison, Raphael and co-workers proposed a computational approach for the localization of the breakpoints of these modifications [128]. They introduced an algorithm for the identification of data from aCGH and paired-end mapping and provided a framework for comparing structural variants across the different techniques. Advances in the analysis of structural variants will have great implications for the analysis of the human-genome and cancer-genomesequencing projects in the near future.