Databases and Systems

fragment. In some cases, a same gene has been independently sequenced by different
groups, or has been sequenced several time from different individuals to study
polymorphism. In principle, one should find in databases only one entry for each
gene, and if it is polymorphic, then allelic variations should be described in the
annotations. But in practice, all redundant sequences are entered in databases, and
there is no merging of partial overlapping sequences.
This redundancy is very problematic, not only because it gives a confuse view of
the status of these redundant sequences (are they identical? splicing or allelic variant
of a same gene? paralogous genes?), but also because it can considerably bias the
results of statistical analyses.
In HOVERGEN, we systematically compare all CDSs between each other (with
BLASTN [ 13]) to try to identify those that correspond to a same gene. As previously
discussed [6], the problem with that approach is that two redundant CDSs may show
some differences due to polymorphism, sequencing errors, or annotation errors.
Taking into account published estimates of sequence polymorphism [ 15], and
sequencing error rates [16-18], we decided to consider as redundant all CDS that
share more than 99% identity (at the DNA level), and have less than 3 bases of
difference in length. Using these criteria, we detected 21% of redundancy among the
63,520 vertebrate CDSs available in GenBank (release 101, June 1997). This level of
redundancy is remarkably high, and it is thus necessary to take it into account when
doing statistical analyses on sequence databases.
Redundancy is not eliminated from HOVERGEN because each entry may be
associated to useful information. Rather, redundancy is explicitly declared, using a
new qualifier ('redundanc y_ref') that is included in sequence annotations. This
qualifier is unique for each set of redundant CDSs. Thus, this information can easily
be used to eliminate redundancy when required.
It is important to note that two homologous genes resulting from a recent
duplication, speciation or conversion may be more than 99% identical (e.g. human
1 and a-2 globin genes). Thus, declaration of redundancy in HOVERGEN should not
be taken into account when one wants to study recent evolutionary events (< 4
million years) [6].
Classification of sequences into families of homologous genes
Sequence selection and similarity search. All available CDSs are classified, except
partial CDSs shorter than 300 nt (about 25% of all CDSs). When several redundant
CDS are available, only one is analyzed for the classification. CDSs are translated
into proteins and compared between each others with BLASTP [13], using the PAM
120 score matrix. The threshold score to report similarity (S parameter in BLASTP)
is set according to proteins length (L): S=150 for L 170 aa, S=L-20 for L