Introduction
For many years, scientists believed that point mutations
in genes are the genetic switches for somatic and inherited
diseases such as cystic fibrosis, phenylketonuria and
cancer. For this to be the case, disease-associated amino
acid substitutions should occur in functionally important
regions of the protein products of genes. While it has been
shown in specific cases that disease-associated amino acid
substitutions affect protein function, until now few
studies have examined this across many genes. Here we
provide direct evidence that disease-associated point
mutations occur in functionally important regions of the
genome and are not distributed equally across the coding
regions of genes. This work supports recent efforts to
collect disease associated mutational data in databases and
suggests that many of the mutations represented in those
databases are the likely underlying molecular cause of
disease.
Recently there have been a number of commercial and
public projects aimed at collecting and understanding human
genomic variation [ 1 ] . The goal of these projects is to
provide an understanding of how genotype is associated with
disease, how it affects our response to drugs and how it
affects the protein products of genes. Examples of these
projects include the SNP Consortium, the Human Genome
Mutation Database [ 2 ] , many gene specific databases ( [
3 4 ] , for example), and both public and private genome
sequencing efforts [ 5 ] . Much of the data that is being
collected are mutations annotated with their observed
phenotype. Automated annotation methods based on structural
and evolutionary parameters can lead to insight into the
molecular basis of disease.
With more than 4,000,000 identified variations and with
over 20,000 of them annotated with a phenotype, we are
facing the problem of having many uncharacterized
mutations. Algorithms are needed for automatically
annotating these gene variations to gain insight into how
they affect the gene's regulation and/or function of its
protein products. Using many collection technologies,
uncharacterized SNP data is being placed in public
databases such as the Human Genome Mutation Database (over
20,000 entries) [ 2 ] and the National Cancer Institute's
CGAP-GAI (Cancer Genome Anatomy Project Genetic Annotation
Initiative) [ 6 ] . The CGAP-GAI group has identified
10,243 SNPs by examining publicly available EST (Expressed
Sequence Tag) chromatograms.
Software for analyzing unannotated SNPs in known disease
associated genes will be especially useful when previously
unobserved mutations are discovered. Every human has
genotypic differences from the standard genome
approximately every thousand base pairs [ 7 8 9 ] . Given
knowledge of how a genotype differs from the standard, it
is important to be able to predict which of the variations
are likely to be the cause of disease or other phenotypic
differences. Evolutionary information about regulatory and
coding regions of genes can be used to highlight certain
mutations or groups of mutations that are attributable to a
phenotype [ 10 11 12 ] .
Early tools using phylogenetic and structural
information have shown promise in predicting the functional
consequences of a mutation [ 13 ] . These reports predict
that anywhere between 20-36% of non-synonymous SNPs alter
the function of a gene's protein product. In the report by
Chasman and Adams, evolutionary information was predicted
to be a useful component in determining whether a mutation
is deleterious [ 13 14 ] . Disease causing mutations are
also likely structurally perturbing at the protein level [
15 ] . Ng and Henikoff have introduced SIFT, a method for
predicting functional SNPs from a database of unannotated
polymorphisms [ 16 17 ] .
The relationship between disease-associated mutation
positions and evolutionary conservation has been reported
in specific cases. An analysis of the breast and ovarian
cancer susceptibility gene, BRCA1, showed that
disease-associated mutations tend to occur in highly
conserved regions [ 18 ] . An analysis of homologous
sequences in the androgen receptor has shown similar
results [ 10 ] . Keratin 12, KRT12, is associated with
Meesmann Corneal Epithelial Dystrophy (MCD). Reported
mutations often occur in the highly conserved
alpha-helix-initiation motif of rod domain 1A or in the
alpha-helix-termination motif of rod domain 2B [ 19 ] .
Structure based analysis methods have also been used to
analyze Osteogenesis imperfecta associated COL1A1 mutations
and disease-associated P53 mutations (Mooney and Klein,
unpublished), [ 20 ] . Miller and Kumar have reported that
disease-associated mutations are conserved in seven model
genes [ 21 ] .
To determine the degree to which mutation positions
differ evolutionarily from other positions, we have built
alignments of homologous genes for 231 disease-associated
genes. These multiple alignments have then been used to
assess the difference in evolutionary conservation for
positions that are both disease-associated and not
associated. The results show that, in general, positions
with disease-associated mutations are conserved more than
the average position in the alignment. This suggests the
most conserved mutations are likely to be the causative
agents of disease, and our data set identifies these
mutations.
Results and Discussion
Our method compares the negative entropy of
disease-associated columns within an alignment to other
columns in that alignment. The goal of this work is to
build these alignments, map the mutations to them, and show
that disease-associated positions are, in general,
conserved. The analysis was performed on the built
alignments and the results are shown in Table 1.
To collect the mutation data, 231 genes were used for
the analysis. They were chosen because they had a reported
cDNA sequence, disease-associated mutations and homologs in
SWISSPROT. These genes are listed in Table 1. Each
alignment consists of all the homologs in SWISSPROT as
determined by a BLAST search with an e-value threshold of
10e-15. For each alignment the negative entropy for each
column was calculated.
The conservation ratio parameter is defined as the
average negative entropy of analyzable positions with
reported mutations divided by the average negative entropy
of every analyzable position in the gene sequence. Analysis
was performed on 231 genes and 6185 mutations and of those
we found that 84.0% had conservation ratios less than one.
From those, 139 genes had more than ten analyzable
mutations and, of those, 88.0% had conservation ratios less
than one.
Use of evolutionary information is a promising approach
to automated characterization of mutations. These results
show that although conservation alone is not a perfect
predictive measure, there is useful information contained
in sequence alignments containing homologous genes.
Approaches using conservation in a multiple alignment
should work better when associated with other methods such
as structural analysis, population analysis and
experimental data. Knowledge of how the sequence pool
clusters into families may increase the sensitivity of the
method.
Our measured parameter, the conservation ratio, is a
quantity that measures the usefulness of a multiple
alignment for characterizing mutations in a gene sequence.
Knowledge of more mutations in a gene does not necessarily
lower the conservation ratio. We expect that knowledge of
more mutations in a gene will increase the statistical
significance of the conservation ratio. This is the likely
underlying cause of the result showing that genes with 10
or mutations increases are more likely to have a
conservation ratio less than one.
The alignments and BLAST searches are integrated on the
website, http://cancer.stanford.edu/mut-paper/.
Conclusions
In conclusion, there are estimated to be 30,000
non-synonymous differences between an individual and the
draft genome [ 23 5 7 8 9 ] . Determination of which
positions are likely to be disease associated is a
challenging and important problem. The finding that
disease-associated mutations occur in positions of
functional importance supports recent efforts for the
building of methods to predict which positions are likely
to be disease associated [ 14 13 16 17 ] . These methods
are likely to incorporate protein structure, the amino acid
identity of the mutation and phylogenetic information. In
an interesting twist, this observation also suggests that
this data may be useable as a functional genomics tool for
understanding the function of the protein products of genes
on a molecular level. Such a method would use the inherent
functional information contained in a phenotypically
annotated polymorphism to infer functional importance
within a gene.
Methods
Non-synonymous mutations were acquired from the Human
Genome Mutation Database
http://archive.uwcm.ac.uk/uwcm/mg/hgmd0.html. 231 genes
were chosen with known disease-association each having
SWISSPROT homologs, a cDNA sequence and mutations in the
coding region. For each of those genes, all known
non-synonymous mutations were then downloaded with the cDNA
sequence for that gene.
Each cDNA sequence was then translated and placed in a
FASTA formatted file. For each of the resultant files a
BLAST [ 24 ] search was performed against the SwissProt
database. All sequences from the returned hits were then
stored in FASTA format files. For each of the genes that
returned BLAST results with e-value scores smaller than
1e-15, ClustalW [ 25 ] was used to build a sequence
alignment.
For each amino acid in the position of interest, the
negative entropy was determined using the following formula
[ 26 ] :
Where the P
i are the probabilities of finding a
particular amino acid at that position. For this analysis,
gapped positions, "-", were considered independent amino
acids.
For each known mutation, the negative entropy of the
column it occupies was tabulated. The average negative
entropy for each mutation within a gene was compared to the
average entropy of all columns satisfying the criteria for
analysis. Mutations outside of the coding region or
mutations encoding termination codons were discarded.
The list of genes was then sorted by average negative
entropy of the mutations. We then calculated the
conservation entropy, CE, using:
CE = average NE of mutation positions/average NE of all
positions in the gene sequence
