The Genom of Homo sapiens.pdf

HUMAN-SPECIFIC EVOLUTIONARY CHANGES 481scripts, but generated twice as many human–mouse alignmentsthat passed quality control.The multiple alignment program ClustalW, run withdefault parameters from the ClustalX package v1.83 α(Thompson et al. 1997), was used to align human, chimp,and mouse coding sequences. Mouse–human andchimp–human alignments were independently examinedfor the introduction of alignment gaps that would result ina frame-shifted human protein. Although these alignmentgaps could represent real differences between twospecies, other causes (incorrect base calls, annotation, orortholog inference) are more likely, especially betweenhuman vs. mouse alignments. Only alignments that producedeither zero insertion/deletions, or those that hadgaps whose length was a multiple of three bases were analyzedfurther. The alignment failure rate for chimp–humanand mouse–human pairs was 3.3% and 31%, respectively(Table 1). The alignments passing quality controlwere converted into Phylip format (Felsenstein 1981) andare available at http://panther.celera.com/appleraHCM_alignments/index.jsp for download.EVOLUTIONARY MODELSA commonly used measure to identify genes undergoingadaptive protein evolution involves comparing the ratioof nonsynonymous to synonymous substitution ratesfor each gene (d N /d S ). If selection pressures favor proteinswith altered amino acid sequence, then nonsynonymouschanges are favored at the nucleotide level, and d N /d S willbe greater than 1. However, since humans and chimpshave a relatively recent common ancestor, the overall sequencedivergence is only about 1.2% (Chen and Li2001), and coding regions show less than half this level ofdivergence (Shi et al. 2003). This results in wide variationfrom gene to gene in the absolute count of synonymousnucleotide changes, and low values of d S in the denominatorresult in large variance in d N /d S ratios. Requiringthat the predicted protein sequence must have diverged atmore than one residue from the most parsimonious ancestralsequence can partly control this variance. Therewere 363 human genes (4.7% of the total) that had morethan one amino acid difference and a d N /d S > 1. Formalstatistical models to test for significant departure from aneutral evolutionary model were also applied to go beyondthis ad hoc description of the genes.A model specifying sequence divergence with parametersfitted by maximum likelihood (Felsenstein 1981) canbe applied to multispecies sequence alignments and anevolutionary tree that describes the ancestral history ofthose sequences. This approach can be extended by allowingseparate parameters specifying rates of synonymousand nonsynonymous substitution (Goldman andYang 1994; Muse and Gaut 1994), variability amongamino acid residues in their degree of constraint, and lineage-specificdifferences in the divergence rates (Yangand Nielsen 2002).In the first of two evolutionary models applied to theset of 7,645 alignments, we applied a classical test of thenull hypothesis of d N /d S = 1 in the human lineage (NielsenAd hoccriterion(363)106 02616395and Yang 1998; Yang 2002). This test may be rejected ifd N /d S > 1, showing evidence of positive selection, ord N /d S < 1, showing strong conservation of the gene in thehuman lineage. The neutral null hypothesis of model 1was rejected by 72 genes (0.94%) at p < 0.001, 414(5.4%) at p < 0.01, and 1216 genes (15.9%) at p < 0.05.There were 6 genes (0.08%) with p < 0.05 and d N /d S > 1.The second formal model applied to test for positiveselection is modified from Yang and Nielsen (2002) andallows variation in the d N /d S ratio among lineages andamong sites at the same time (see also Yang and Swanson2002). In this method (Model 2), a likelihood ratio test ofthe hypothesis of neutrality is performed by comparingthe likelihood values for two hypotheses. Under the nullhypothesis, it is assumed that all sites are either neutral(d N /d S = 1) or evolve under negative selection (d N /d S < 1).Under the alternative hypothesis, some of the sites are allowedto evolve by positive selection in the human lineageonly. The neutral null hypothesis of Model 2 was rejectedby 28 genes (0.37%) at p < 0.001, 178 genes(2.3%) at p < 0.01, and 667 genes (8.7%) at p < 0.05.The overlap between these three sets of genes is high(Fig. 1), but differences reflect the different attributes ofthe data that the tests consider. For example, small genesor genes with few substitutions may be flagged by the adhoc criterion, but not attain statistical significance by theevolutionary models. Importantly, Model 2 can detectcases where a portion of the protein (perhaps a protein domain)is undergoing positive selection, but the overalld N /d S may not be elevated, resulting in those genes beingmissed by the ad hoc criterion and by Model 1. For thisreason, the remainder of the analysis considers onlyModel 2 test results.THE IMPACT OF LOCAL SEQUENCECOMPOSITIONGenome sequence composition such as GC content,gene density, repeat density, and local recombination ratecan influence patterns and rates of sequence divergence(Hellmann et al. 2003; Webster et al. 2003). Before we go151195Model 1P < 0.05(1216)Model 2P < 0.05(667)Figure 1. Overlap of human genes identified as exhibiting a signatureof positive selection by the ad hoc criterion, a model testingfor departure of d N /d S from 1 (Model 1) and a model testingfor excess nonsynonymous substitution within a domain of theprotein in the human lineage only (Model 2).