Frameshift Mutation

RATT is programmed in ‘bash’ and ‘PERL’ and its design is illustrated in Figure 1 and Supplementary Figure S1. First, two sequences are compared using ‘nucmer’ from the MUMmer package (17 (link)) to define sequence regions that share synteny. Those regions are filtered using configurable parameters depending on the type of annotation mapping that is being attempted. Preset parameters are provided for transfers between assembly versions, strains or species (see Supplementary Table S1). To be included, the minimum nucleotide sequence identity between synteny blocks must be 40%. Synteny information is stored as a base range in the query and its associated base range in the reference. However, this information alone is inadequate to map the annotation because insertions or deletions (indels) change the relative distance between mapped synteny blocks. The coordinates are therefore sequentially adjusted across a synteny block by calling indels using ‘show-snp’ from the MUMmer package. Accurately calling indels within repetitive regions presents a particular challenge. Therefore, RATT recalibrates the adjusted coordinates using single nucleotide polymorphisms (SNPs, also called using ‘show-snp’) as unambiguous anchor points within synteny blocks. In transfers between very closely related sequences (e.g. successive assembly versions), SNPs may occur with insufficient frequency to perform this coordinate adjustment. In such cases, RATT modifies the query by inserting a faux SNP every 300 bp to aid in the recalibrating step. The final sequence and transferred annotations remain unaffected.
Figure 1.

Workflow of RATT.

Once the coordinates within synteny blocks have been defined, RATT proceeds to the annotation-mapping step, whereby each feature within a reference EMBL file is associated with new coordinates on the query (Supplementary Figure S1B). A feature is not mapped (and is put in the non-transferred bin file), if it bridges a synteny break and if its coordinate boundaries match different chromosomes, different DNA strands, or if the new mapped distance of its coordinates has increased by more than 20 kb. If a short sequence from the beginning, middle or the end of a feature can be placed within a synteny region, mapping is attempted (see Supplementary Figure S1B). In addition, if the exons of a single gene model map to different gene regions, the model is split and identified in the output file. The bin is an EMBL-format file that can be loaded onto the reference sequence for analysis (see Figure 2, brown colour track). Further outputs include statistics about transferred features or the amount of synteny conserved between the reference and query, as well as Artemis-readable files showing SNPs, indels and regions that lack synteny between the compared sequences, see the example on the sourceforge site.
Figure 2.

Transfer of annotation from the M. tuberculosis strain H37Rv onto the strain F11 sequence, over a deletion. The genomes of H37Rv (upper) and F11 (lower) are shown using the Artemis Comparison Tool (ACT). The source H37Rv annotation (light blue) is directly mapped onto F11 by RATT (green) except for those features corresponding to a region that is unique to the source strain that cannot be transferred and are written to a separate output file (brown).

Although two sequences may be related, differences can occur, such as a change in the start or stop codons of a protein-coding sequence. Therefore, we implemented a correction algorithm in RATT (see Supplementary Figure S1C). Figure 3 shows examples of the correction step. First, the start codon is checked. If it is not present, the upstream sequence is searched for a new start codon (Figure 3A). If a stop codon is found, the first start codon downstream is used. In the absence of any start codon, an error is recorded in the results file. If the sequence between exons has no stop codon and a length divisible by three bases but the splice acceptor or donor sequences are wrong, then the intron is eliminated. Likewise, frameshifts previously introduced into the reference to maintain conceptual translations (for instance, in apparent pseudogenes) will also be removed from coding sequences in the query. RATT will also detect, and attempt to fix, incorrect splice sites. As splice sites are difficult to annotate correctly, RATT only tries to correct a gene model that has one wrong splice site. If one incorrect splice site is detected, the closest alternative splice donor or acceptor is found that, when used, generates no frame shifts. Next, RATT searches for genes or exons with internal stop codons, further than 150 bp from the 3′-end. If the introduction of a frameshift would generate a model without internal stop codons, the model is corrected (Figure 3C). Stop codons are corrected last: if a model has less than five internal stops in its last exon, the model is shortened to the first stop codon (Figure 3B). If the model has no stop codon it is extended downstream until a stop codon is found.
Figure 3.

RATT corrections of transferred annotations. Annotation from H37Rv were transferred onto the F11 sequence (pale blue), corrected (green) and then compared with the existing strain F11 annotation in EMBL (yellow). (A and B) The correction of start and stop codons, respectively. In a more complex mapping situation (C), where all three reading frames are shown for clarity, RATT maps a large single coding sequence (CDS) from H37Rv to a locus within F11 that includes several in-frame stop codons. By inserting a frameshift (i.e. to indicate a pseudogene) the conceptual translation is preserved. This contrasts with two overlapping genes predicted as part of the F11 genome project.

Different criteria can be specified depending on the translation that an organism uses (e.g. such bacterial TTG and GTG start codons) or whether unsual splice sites are used. RATT is programmed in PERL and was tested in UNIX/LINUX environments. The output can be loaded into Artemis/Act. The list and explanation of all the output files can be found at the sourceforge site.

The first part of the IntOGen-mutations pipeline assesses the potential functional impact of somatic mutations detected across the cohort of tumor samples. The Ensembl variant effect predictor^{10 (link)} (VEP, v.70) script and precomputed cache files, downloaded from the Ensembl FTP site (ftp://ftp.ensembl.org/pub/), are used to determine the consequences of somatic mutations in annotated functional elements. The pipeline obtains SIFT^{11 (link)} and PolyPhen2 (ref. 12 (link)) functional impact from VEP. Precomputed MutationAssessor^{13 (link)} functional impacts are obtained from the MutationAssessor Web server (http://www.mutationassessor.org/) during the installation of the pipeline and are queried locally during execution. The transformation of functional impact scores to account for the baseline tolerance of genes to germline mutation (transFIC), described elsewhere^{14 (link)}, has been reimplemented in Python as a module of the IntOGen-mutations pipeline.
The pipeline implements an expression filter to disregard genes that are not expressed across the tumor samples in the cohort. This list of expressed genes is an optional input to the pipeline, which excludes all genes outside the list from the foreground of both OncodriveFM and OncodriveCLUST (see below) while keeping their mutations in the background. In the current release of the IntOGen-mutations Web discovery tool, we have employed as a filter the list of genes expressed across any of the 12 pan-cancer data sets (ref. syn1734155).
The OncodriveFM and OncodriveCLUST approaches, also described elsewhere^{7 (link),9 (link)}, have been reimplemented as IntOGen-mutations pipeline modules and are available as independent programs from two Git-controlled repositories at https://bitbucket.org/bbglab/. Briefly, OncodriveFM receives as input the list of synonymous, nonsynonymous and frameshift-indel mutations and their corresponding SIFT, PolyPhen2 and MutationAssessor scores. Then it assesses whether any gene shows a trend toward the accumulation of mutations with high functional impact as compared to the background distribution of these functional impact scores in all mutations detected across the cohort of tumor samples (FM bias). For each functional impact score included in the pipeline, the method produces an empirical P value that evaluates this FM bias. These three P values are subsequently combined using Fisher's approach to produce one integrated P value for each gene. To account for possible nondependence between the three P values included in the combination, the IntOGen-mutations Web discovery tool considers as significant those with a false discovery rate (FDR) below 0.05.
OncodriveFM also computes an FM bias for pathways. Three z scores are computed in this case to assess the trend of pathways to accumulate mutations with high functional impact. The z scores are combined using Stouffer's approach, and the combined z score is transformed into an integrated P value.
OncodriveCLUST, on the other hand, receives as input two separate lists of mutations: potentially protein-affecting mutations (nonsynonymous, stop and splice site) and silent mutations (synonymous), with their corresponding locations across the proteins' sequences. It then assesses the significance of the trend of potentially protein-affecting mutations to be clustered with respect to a background represented by the homologous trend for silent mutations.
Genes mutated in less than 1% of the samples in projects whose median of mutations per sample was below 100 were not analyzed by OncodriveFM. In projects with higher median of mutations per samples, this threshold was set to 5 samples with mutations. For OncodriveCLUST, the thresholds were 3 and 5 mutated samples, respectively. These and many other parameters of the pipeline are configurable by the user, as explained in its documentation.
In addition to third-party (and in-house) software and data, IntOGen-mutations pipeline installation requires some Python libraries. The most important of these are the numpy and scipy scientific computing libraries and the statsmodels Python statistical library.
The pipeline also relies on other external data files. During pipeline installation, all of the needed external and third-party data files are downloaded and correctly placed, and external libraries are downloaded and compiled, thereby creating a Python environment where the pipeline executes.
The analysis of the 4,623 tumor samples currently included in the IntOGen-mutations Web discovery tool takes approximately 5 h on an eight-core, 12 GB RAM computer.