A step-by-step protocol for the Concept Signature Analysis

 Nat Biotechnol. 2009 Nov;27(11):1005-11 [Pubmed]

The principle of the Concept Signature Analysis

To examine the key functional characteristics that assist in new cancer gene discovery, we performed a multi-dimensional characterization of cancer-related gene fusions and point mutations. Placing the array of cancer genes in the context of a compilation of “molecular concepts”, including molecular interactions, gene annotations and pathways revealed the “signature concepts” defining the genes driving cancer initiation and progression (Figure 1). This prompted us to generalize this finding to develop a method that could filter non-specific genetic aberrations in cancer. We hypothesize that the “signature molecular concepts” frequently found in cancer genes may be used to define biologically meaningful genes underlying cancer, similar to signature genes defining certain phenotypes. Using such information, we developed an innovative concept signature (ConSig) technology that nominates biologically important genetic aberrations from high-throughput data by assessing their association with molecular concepts characteristic of cancer genes.

Figure 1. Signature molecular concepts for fusion and point mutation genes in cancer.

 A step-by-step manual for Concept Signature Analysis

Compile molecular concept database.

In this study, we compiled 28,963 molecular concepts from the Gene Ontology database, the Reactome database, the Kyoto Encyclopedia of Genes and Genomes (KEGG), Biocarta, the HPRD database, and the Entrez Gene conserved domain database (Table 1). To remove redundancy, in the processing of gene ontologies, the genes that appeared in the child ontologies were subtracted from the parents to avoid duplicate representation. The compendia of non-redundant molecular concepts used in this study can be downloaded from this link.

Table 1. The compendia of molecular concepts were compiled from 6 public databases

Compile cancer-causal gene database.

Here we use the Fusion gene list in the Mitelman database (2008) provided by Dr. Mitelman and mutation gene list extracted from the Cancer Gene Census (2008) as an example. The compiled cancer causal gene lists can be downloaded here (in NCBI Entrez Gene ID) -
[Fusion gene list; Point mutation gene list]

 Mapping the fusion or point mutation gene lists against the compendia of molecular concepts. Click the following link for the per script of this step: [Sample perl script 1].

Calculate the fusion and mutation ConSig-score for all known human genes.

Computationally, let k be the number of concepts associated with a specified gene. Let n_i represent the number of total genes and x_i represent the number of fusion or mutation genes participating in a given concept i, i=1,…,k. The ConSig-score then integrates a signal measure of fusion or mutation genes participating in concept i (x_i/n_i^0.5) over all possible i, with the incorporation of normalization factor for k using the formula in Figure 2.

Figure 2. The algorithm for the concept signature analysis

With this computation, if a gene has high probability to be involved in gene fusions or mutations, the fusion/mutation ConSig-score will be high respectively; thus the radius in the two-dimensional ConSig-score plot for fusions and mutations will correlate with the role of tested genes in cancer. To eliminate the bias from the gene itself in the overlap, the seeding genes were subtracted from the signature concepts during the calculation of their own ConSig score.

Download link: the perl script for this step [Sample perl script 2].

Calculate d- and r- ConSig score for all human genes.

Plotting the fusion vs. mutation ConSig-scores produced a striking segregation of known fusion genes from mutation genes (Figure 3a). The distinction line (D-line), y=1.67x, was determined by testing optimal separation capacity (Figure 3b), which separates 85% of mutation genes from 80% of fusion genes. In this setting, the radius to the zero point is defined as the radial ConSig-score of a gene (rConSig-Score), which indicates the strength of association with signature concepts of both fusion and mutation genes, thus implies the functional relevance of candidate genes in cancer. The distance vector from the node to the D-line, which illustrates a distinction between fusion and mutation genes, is defined as the distinction ConSig Score (dConSig-Score).

Figure 3. Plotting the fusion and mutation ConSig score for all human genes (a) and determining the D line (b).

Download link: the perl script for this step [Sample perl script 3]

Testing the performance of the ConSig score ranking.

Rating all human genes by the rConSig-Score will produce substantial enrichment of established cancer genes in top-scoring genes. Replacing the fusion or mutation gene sets with random gene sets produced no enrichment of the randomly selected genes, thus validating the significance of this observation.

The pre-computed d- and r-ConSig score for all human genes can be downloaded from this link (v2008).

 The application of the Concept Signature Analysis

The discovery of driving gene fusions in cancer.

The ConSig technology preferentially identifies biologically important genes in cancer. This is particularly useful in the analysis of a large number of putative chimeras generated by next generation sequencing data to filter secondary fusions. Of note, in this application, the main theme is to evaluate the biological relevance of putative chimeras, in stead of distinguishing fusions and mutations, whereby the radial ConSig will be more informative. This is especially important for evaluating the genes involved in both fusion and point mutations (mixed type cancer genes), for example, a fusion involving EGFR gene will be considered as biologically important because of the prior knowledge of EGFR point mutation in cancer. Moreover, the 3’ fusion partners display more distinctive signature concepts than the 5’ partners, therefore the ConSig technology will be more discriminative in evaluating 3’ genes. In practice, we usually first rate the 3’ partners of fusion chimeras by rConSig scores, and then rate the 5’ partners by rConSig score to supplement this analysis.

To demonstrate the application of the above principles, we applied the ConSig technology to benchmark the fusion candidates detected by paired-end transcriptome sequencing, which were then assessed for recurrent chromosomal aberrations using high-quality copy number data based on the fusion breakpoint principle. While analysis of the transcriptome data from 12 lung cancer cell lines generated 530 putative chimeras, the ConSig score was able to identify the known EML4-ALK fusion as the top-ranked candidate in the H2228 lung cancer cell line. In addition, we found further evidence of a R3HDM2-NFE2 fusion in H1792 cell line, which results in overexpression of wild-type NFE2, and promotes cell proliferation and invasion. Moreover, through analysis of SNP arrays and lung TMAs, we find that chromosomal rearrangements at the NFE2 locus are recurrent in a small subset of patient tumors, suggesting that NFE2 fusion may contribute to a new class of lung cancer biology.

The discovery of cancer-causal point mutations.

It is notable that in our ConSig analysis, the point mutation genes demonstrated more distinctive concept signatures, so that the separation of mutation genes from the rest of human genome is even better than fusion genes. Therefore, the ConSig technology can be applied to the deep sequencing data to reveal the driver point mutations in cancer.

The discovery of over- or under-expressed cancer genes.

The over- or under-expressed genes in cancer identified from large-scale gene expression analysis can also be ranked by ConSig score to reveal the most biologically meaningful genes, thus nominates the most interesting candidate for biological studies.

The discovery of amplified or deleted cancer genes.

The same approach also applies to the amplified or deleted cancer genes. For example, applying the ConSig analysis to the TCGA array CGH data for glioblastoma revealed EGFR as top amplified oncogene candidate in the chr 7 amplicon. 

Figure 4. Revealing the primary target of Chr 7 amplicon in glioblastoma.