Reading frame conservation

Following the strategy explained in Methods, we analyzed reading frame compatibility of exons flanking translocation breakpoints, in 583 gene fusions coding for a potential fusion protein in which with both partner genes contribute an annotated protein domain. Interestingly, the final reading frame of the 5′ exon and the starting reading frame of the 3′ exon were compatible in 540 of the fusions analyzed, thus confirming that an in-frame fusion protein was generated in 93% of the cases. In some translocations, the breakpoint fell in the middle of an exon, but even so the reading frame was kept across the fusion. This is illustrated by a set of gene fusions between FIP1L1 (5′ gene) and PDGFRA (3′ gene) in which different exons of FIP1L1 are fused to truncated versions of exon 12 of PDGFRA (Ensembl transcript ENST00000381354). Deletions in this exon always result in a compatible reading frame with the corresponding exons of FIP1L1 (exons 10, 11, 12 and 13 of FIP1L1 transcript ENST00000358575, which end in reading frames +3, +1, +2 and +3 respectively, in version 38.36 of Ensembl), leading to in-frame fusion transcripts in all four configurations.

In the remaining 43 fusions the reading frames of flanking exons were not compatible, so they would not be expected to generate an in-frame fusion protein. In these cases we went back to the original fusion sequence and found that in 31 of these (72%) the reading frame had been restored by various mechanisms such as alternative splicing, insertion of intronic regions, insertion of non-templated nucleotides or deletion of exonic nucleotides. This was particularly common in fusion proteins involving EWSR1, FUS and TAF15, since 48% of such fusions had incompatible reading frames that were corrected by one of these mechanisms (24 out of 50 gene fusions analyzed for these genes). After careful evaluation of the remaining 12 fusions in which reading frames were not compatible (2% of the total 583 fusions), we assumed that a functional protein product cannot be generated in these cases.

The finding that 98% of gene fusions generate transcripts that can be translated as in-frame protein products confirms that reading frame conservation is of great functional importance in oncogenic fusion proteins, since the expected frequency of compatible reading frames between two random exons (assuming equal frequencies of +1, +2 and +3 reading frames) is one third. This is a clear signature of the strong selective pressure in somatic cells that favors those fusion products capable of driving oncogenic transformation, and it has important implications in the discussion about the identity of the factors that govern the position of translocation breakpoints in cancer cells (see below).

Protein domain architectures present in the same fusion protein

A network graph of the genes involved in chromosomal translocations that generate fusion proteins (Supplementary Figure S1) shows three main independent clusters, plus some smaller graphs that are not connected to any of the main components [15], [17]. These were analyzed as explained in the Methods section, in order to create a global network of domain architectures that are brought together to the same fusion proteins in cancer (Figure 2 and Supplementary Text S1). Topological parameters such as degree distribution show that the network of gene fusions and the network of domain architectures are both compatible with scale-free or small-world, but not random, topologies. The number of nodes (114) in the network of domain architectures (Figure 2) is less than half the number of genes rearranged in those translocations (235 nodes in Figure S1), indicating that the same architectures are used in different gene fusion events. Likewise, the network of domain architectures has a smaller diameter (8 vs. 13) and a smaller characteristic path length (3.66 vs. 4.83); as a result, network density in the network of domain architectures is more than twice the density of the network of gene fusions (0.0019 vs. 0.0008). These parameters reflect the lower complexity of the network of domain architectures with respect to the network of gene fusions. A more in-depth discussion of these networks can be found in Supplementary Text S1, Supplementary Figure S2, Supplementary Figure S3, Supplementary Figure S4 and Supplementary Figure S5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Global view of protein domain architectures in oncogenic fusion proteins.

All domain architectures were merged together, resulting in a single large component plus 6 other smaller graphs. Nine nodes with more than 5 neighbors (hubs) are shown in blue. The three main gene fusion networks shown in Supplementary Figure S1 are clearly visible in the hubs corresponding to TK, NUP and COIL/HZ architectures. The size of each node is indicative of its degree (number of neighbors).

https://doi.org/10.1371/journal.pone.0004805.g002

Two interesting features are apparent in the network of domain architectures. First, all the architectures derived from the three main gene fusion networks appear as a large single graph, indicating that some domain architectures are common to all gene networks. Likewise, some of the architectures from the 21 small gene fusion graphs are also included in this large component, so that only 6 small components remain unconnected to the main network of domain architectures. Second, the more connected nodes (hubs with≥5 neighbors, shown in blue in Figure 2) identify the main classes of fusion proteins found in cancer, that is, those involving the tyrosine kinase (TK) domain, the EWS activation domain (EAD), the Runt domain, the ligand binding domain of the nuclear hormone receptor (HRMN), the AT-hook DNA binding domain (HOOK and COIL/HZ), the GLFG repeats (NUP) and coiled coils (COIL). This suggests that the network captures the main biological themes that are presently known to be used by fusion proteins in cancer.

The finding that only certain combinations of protein domains are present in oncogenic fusion proteins, forming a network of non-random topology, implies that such combinations are the result of distinct functional constraints. As mentioned above, this is a signature of the cellular selection pressures that dictate which chromosomal translocations are present in cancer cells.

In order to anticipate how this network will be affected by the discovery of new translocations, we analyzed chromosomal translocations that were published after the beginning of this work (October 2007) and thus not yet included in TICdb at the time [18]–[33]. We collected 17 gene fusions that generate a fusion protein with annotated protein domains, describing 9 new genes and 7 novel domain architectures (two of the new genes contributed an already described architecture). We also observed two novel combinations of previously described domain architectures. Additionally, in one case a 3′ partner gene (TCF3), previously described as 5′ fusion partner, contributes a distinct domain architecture in this new case. The analysis of these new fusions suggests that the network of domain architectures, even though not yet complete, contains most of the architectures used by oncogenic fusion proteins, and grows at a slower rate than the network of gene fusions.

Role of functional selection in the position of translocation breakpoints in cancer cells

The two signatures of functional selection that we have analyzed in fusion transcripts (namely, reading frame conservation and non-random combinations of protein domains) suggest that such forces might be major factors in determining the non-random distribution of translocation breakpoints that is seen in human cancers. In this regard, the widely held view that local sequence factors are responsible for the presence of translocation breakpoints at specific genomic sites relies on the assumption that translocation breakpoints reveal the location of all DSBs generated in those cells. Thus, since translocation breakpoints are non-randomly distributed, the inference is made that DSBs are initially created non-randomly. However, sequence elements responsible for the generation of DSBs (short sequence motifs, topoisomerase II sites, dispersed repeats, intronic transcription initiation sites, cruciform structures, etc) are fairly common throughout the genome, so it is reasonable to suppose that most of the DSBs that are created across the genome during the lifetime of a somatic cell have been properly repaired and that only a small subset of misrepaired DSBs will result in oncogenic fusions and will eventually be found in tumor samples. In this respect, it must be kept in mind that the analysis of tumor samples represents an extreme case of ascertainment bias: by definition, only translocations that have been important for tumor growth will be detected, whereas many other possible translocations that did not provide a proliferative advantage to the cell will not. Some translocations, for example, would be expected to be deleterious to the cell, since two different gene alleles (one allele of each gene) have been inactivated by the breaks, so that cells carrying those translocations will eventually disappear from the tissue. Other rearrangements will be functionally neutral and the resulting fusion gene will not have an advantageous biological function. In the end, the translocations that are found in tumor samples are the result of the clonal expansion of cells that harbor translocations with the potential to promote tumor growth because they create oncogenic fusion genes. The specific breakpoints harbored by these translocations constitute the subset of non-random translocation breakpoints that are found in cancer cells.

This is illustrated in Figure 3, which shows two genes theoretically capable of participating in a reciprocal translocation with oncogenic properties, due to the domains that are present in their respective proteins. Even if the initial DSBs were distributed uniformly across those genes [34], it is obvious that not all possible translocations will create a fusion gene with oncogenic potential. Looking at the position of the regions that code for the necessary protein domains, and considering the reading frames of the various exons involved, it becomes clear that an oncogenic fusion protein will only be generated if breakpoints are located within certain introns. Other possible breakpoint combinations would lead to the loss of an important functional domain in the fusion protein, or to an out-of-frame product, and will not be favoured in tumor samples. A clear implication of this is that the perceived “non-randomness” in the genomic distribution of translocation breakpoints is not necessarily related to the initial localization of DSBs, but could be the result of the selection process by which only a few of those DSBs eventually survive in the cells of a tumor.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Requirements for the emergence of a successful fusion protein.

Three exons of two hypothetical genes are shown (exons A, B and C in blue, exons 1, 2 and 3 in orange). The initial and final reading frame of each exon is shown (+1, +2 or +3). Exons A and B of the top gene code for a protein domain (red bar), whereas exon 3 of the bottom gene codes for another protein domain (green bar). Even if double-strand breaks (DSBs, yellow lightning symbols) were created uniformly across the sequence of both genes, only those breakpoint combinations leading to in-frame fusion proteins that code for both protein domains will display oncogenic potential. As a result, translocation breakpoints found in tumor samples will cluster to specific gene regions (vertical blue arrows).

https://doi.org/10.1371/journal.pone.0004805.g003

Prediction of novel fusion proteins in hematological and solid tumors

If the two signatures identified in this work are important determinants of breakpoint localization, then our results should be useful for the prediction of gene fusions that have not yet been found in tumors. First, the information about which domain architectures are present in the same fusion protein can be used to select all the genes that encode a particular pair of domain architectures. This will predict several gene fusions that are potentially oncogenic. Information about the reading frames of the exons belonging to those genes should identify which specific fusions (if any) are capable of generating an in-frame fusion protein that includes the required combination of protein domains. More importantly, this analysis should also identify which introns are most likely to contain the breakpoints, and thus assist in the design of molecular strategies for the detection of those putative fusion transcripts.

One obvious implication from our work is that many potential gene fusions could generate the same combination of domain architectures, because each architecture is usually encoded by several genes. However, it is generally assumed that the majority of chromosomal translocations responsible for the development of human cancer have already been described [8], [17]. Although some new cases are published every year, most of them report novel breakpoints in previously known gene fusions, or new fusions between genes that had been previously found fused to other partners. It is not clear why many of the potential novel gene fusions have not been detected. A likely explanation is that the genes involved do not meet some of the criteria that are required for a reciprocal translocation to take place, such as proximity within the nuclear space or co-transcription in the same nuclear transcription factories [35]–[39]. Alternatively, some of these novel gene fusions might remain undetected because they were never searched for, since most studies focus on the detection of known translocations. In this regard, it is interesting to consider recent studies in which the genome of various types of cancer cells has been interrogated in an unbiased manner [40]–[43]. In the case of a diploid sample from a leukemia patient, massively parallel sequencing uncovered novel point mutations, but no genomic rearrangements [40]. End Sequence Profiling of cell lines from solid tumors revealed many somatic genomic rearrangements, but only a few of these were gene fusions. For instance, Campbell et al. [41] used massively parallel paired-end sequencing in two lung cancer cell lines and found 22 somatic interchromosomal rearrangements in the NCI-H2171 cell line, but none in NCI-H1770. Of those, only one expressed fusion transcript was identified, although it was predicted to be out-of frame. Raphael et al. [42] found one fusion between HYDIN gene and an anonymous gene in MCF7 metastatic breast carcinoma cell line. Another fusion between SCL12A2 and an expressed sequence tag was found only in high passage MCF7 cells. In this same cell line, in which chromosomal aberrations have been previously described by Spectral Karyotyping (SKY) and array-Comparative Genomic Hybridization (CGH), Hampton et al. [43] found 10 gene fusions using end-sequence profiling with massively parallel sequencing. Of these, only four were found to be expressed, but their oncogenic potential was not directly tested. Considering that these studies were performed on cell lines, the number of novel expressed gene fusions is relatively low.

These recent data are also relevant to the present debate about “driver” and “passenger” mutations in cancer genomes. Due to the inherent instability of the genomes and the clonal nature of the tumorigenic process, many aberrations are expected to be found when cancer genomes are interrogated in an unbiased manner, the majority of which will be passenger aberrations with no functional relevance to the oncogenic process. In this context, there is a great need for new approaches that can distinguish those genomic changes that drive tumor initiation or progression from neutral changes that have been acquired by the clone but have no functional impact. Our results underscore two features that could be useful in this respect. However, due to the nature of the data analyzed in this work, our results are particularly relevant to hematological malignancies, which are overrepresented in our dataset (66% of all fusions studied correspond to hematological cancers). Chromosomal translocations have been difficult to find in solid tumors, particularly in epithelial cancers, due to the presence of complex karyotypes with many aberrations that are difficult to analyze [44]. However, gene fusions with clear oncogenic potential have been recently identified in prostate and lung cancers [45]–[47], suggesting that such rearrangements might be of greater importance for the development of solid tumors than generally thought. It will be interesting to see whether our findings also apply to solid malignancies, as more genomes from primary tumors are sequenced. Even if our findings are not at present entirely applicable to solid tumors, we believe that the two signatures of functional selection will be evident in expressed fusion proteins of most cancer types.

In conclusion, we identify two signatures of functional selection in oncogenic fusion proteins: specific combinations of protein domains and reading frame compatibility. Our results provide experimental support for the view of cancer as an evolutionary process at the cellular level, and highlight the importance of selection in the recurrent emergence of clonal genetic alterations and in the non-random distribution of translocation breakpoints in cancer cells. We also show that building a network of protein domain architectures that are brought together to oncogenic fusion proteins is more informative about the mechanisms involved than the network of gene fusions. The network of domain architectures is less redundant, captures most of the functional processes deregulated in oncogenesis, and will grow at a slower rate in the future. Finally, we propose that this information can be used to assess the functional relevance of novel chromosomal translocations and to predict the position of breakpoints in the genes involved.