The Genom of Homo sapiens.pdf

128 WANG, BUHLER, AND BRENTMETHODSSequences: CFTR study. The sequence of the mouseCFTR region was downloaded from the NISC Web site(http://www.nisc.nih.gov/data/20020612_Target1_0051/mouse_T1.fasta).Sequences: Mouse chromosomes study. The sequencesof mouse chromosomes 11, 17, and 19 weredownloaded from (http://genome.ucsc.edu/golden Path/mmFeb2003/chromosomes/). The chromosomes were dividedinto nonoverlapping 1-Mb segments for both theBLAST and TWINSCAN portions of the analysis. Shotgunreads from Fugu (Takifugu rubripes), chicken (Gallusgallus), dog (Canis familiaris), human (Homosapiens), and rat (Rattus norvegicus) were downloadedfrom the NCBI trace archive (ftp://ftp.ncbi.nih.gov/pub/TraceDB/).Fold coverage calculations. Each 1x database consistsof clipped reads whose total length is equal to the estimatedsize of the source genome. The genome sizes forTetraodon (0.31 Gb), Fugu (0.31 Gb), and rat (2.6 Gb)were estimated as the number of bases in the assembly;the size for human (2.9 Gb) was taken from the mousegenome paper (Waterston et al. 2002), the size for dog(2.4 GB) was taken from the recent shotgun sequencingsurvey paper (Kirkness et al. 2003), and the size forchicken (1.2 Gb) was taken from the proposal to sequencethe chicken genome (McPherson et al. 2002). An alternativedefinition of 1x is the number of raw reads needed toachieve an N-fold redundant assembly, divided by N(Waterston et al. 2002; Flicek et al. 2003). This alternativeyields substantially larger 1x databases because assemblerstypically discard many of the raw input reads.BLAST. To prepare BLAST databases, we masked repeatsin the informant genome sequences with RepeatMasker (Smit and Green, http://ftp.genome.washington.edu/RM/RepeatMasker.html), performed an additionalround of low-complexity masking with nseg (Wootton andFederhen 1996) using default parameters, and removed allstrings of 15 or more consecutive Ns in order to speed processing.All BLAST jobs were run using WUBLAST 2.0,18-Jan-2003, running under x86 Linux. The analysis reportedhere uses the following BLAST parameters: M = 1,N = –1, Q = 5, R = 1, Z = 3,000,000,000, Y = 3,000,000,000,B = 10,000, V = 100, W = 10, X = 30, S = 30, S2 = 30, andgapS2 = 30. The seg and dust filter options were used.Sequence annotation.The annotation of the mouseCFTR region created by the NISC was downloaded fromhttp://www.nisc.nih.gov/data/20020612_Target1_0051/mouse_T1.annot.ff (Genbank accession AE017189). Forthe mouse chromosomes, we downloaded the MGC transcriptsaligned to the genome by BLAT fromhttp://genome.ucsc.edu . This annotation contains 8,547known genes for the entire genome.TWINSCAN. We used TWINSCAN version 1.3. BothTWINSCAN source code and a Web server are availableat http://genes.cse.wustl.edu.Accuracy calculations. The predictions were comparedto the MGC annotations using the Eval software package(Keibler and Brent 2003; http://genes.cse.wustl.edu/eval/).The accuracy measure plotted in Figure 2A is the averageof exact exon sensitivity (the fraction of annotated exonsfor which TWINSCAN predicts both splice sites correctly)and exon specificity (the fraction of exons predicted byTWINSCAN that exactly match annotated exons). Becausethe MGC transcripts include only a fraction of theexons in the mouse genome, specificity measured againstMGC is a systematic underestimate of actual specificity—all predicted exons that are not in the MGC transcripts arecounted as wrong. In order to make the results on the chromosomecomparable to those on the well-annotated CFTRregion, the accuracy measure plotted in Figure 4A is theaverage of exon sensitivity and 3.5 times the exon specificity.DISCUSSIONThe availability of extensive genome sequence from avariety of vertebrate lineages has enabled the first directinvestigation of how evolutionary distance affects comparativegene prediction in the vertebrates. Initially, westudied the greater CFTR region using the sequences ofBACs that were selected for orthology to the humanCFTR region. A previous analysis of this region (Thomaset al. 2003) based on different alignment methods(Schwartz et al. 2003) reported that chicken alignmentscovered a large fraction of coding sequence (CDS) in themammalian CFTR region but only a small fraction of noncodingsequence. Our analysis confirms this and showsthat it has the expected positive effect on the accuracy ofTWINSCAN, a state-of-the-art gene prediction system. Inaddition to improving coding versus noncoding predictions,chicken alignments yielded the greatest accuracy asmeasured by prediction of exact exon boundaries.Application of the same analysis to a much broadersample of the mouse genome (chromosomes 11, 17, and19) told a very different story. The CDS of the greaterCFTR region is exceptionally well conserved betweenthe mammalian and avian lineages, relative to other portionsof the genome. In the broader survey, there was nosudden jump in the percentage of mouse CDS that alignsas one moves from fish comparisons to chicken comparisons.Instead, the aligned percentage seems to increaselinearly with the percent identity in aligned regions as onemoves from fish to chicken to human (Fig. 4). This differencewas reflected directly in the TWINSCAN performance,which peaked at human rather than chicken. Thecurves in Figure 5, as well as previously reported results(Flicek et al. 2003), suggest that using complete, assembledinformant genomes in place of the 5x shotgun coverageis unlikely to have any qualitative effect on the relativeutility of the comparisons. Likewise, changingalignment parameters or algorithms affects the absolutefraction of intron and CDS that aligns in each genomepair, but it does not change the relative values of comparisonsat the distances studied here (data not shown). Usingtranslated alignments (TBLASTX) also has littlequalitative effect and does not improve TWINSCAN per-