UniGene Build Procedure
Clustering is the process of finding subsets of sequences which
belong together within a larger set. This is done by converting
discrete similarity scores to boolean links between sequences. That
is, two sequences are considered linked if their similarity exceeds
a threshold. UniGene clustering proceeds in several stages, with
each stage adding less reliable data to the results of the preceding
stage. This staged clustering affords greater control than a more
egalitarian treatment of all links between sequences. The stages
are:
Screening for contaminants, repeats, and low-complexity sequence
is performed. Low-complexity screening is performed using NCBI's
Dust. Mitochondrial and ribosomal sequences are screened for, as are
vector contaminants and repetitive elements. After screening, a
sequence must contain at least 100 informative bp to be a candidate
for entry into UniGene.
Gene links are found. The set of gene sequences [mRNA or genomic
sequences, many of which are complete CDSs] is compared with itself.
Sequence pairs which are sufficiently similar are linked together to
form initial clusters.
EST to gene links and EST to EST links are added to these
clusters. The set of ESTs is compared with the set of genes using megablast,
and sufficiently similar sequence pairs are added to the clusters.
Any links which would join two distinct clusters from the preceding
stage (that is, join two sets of genes not linked to form one
cluster without the addition of ESTs) are discarded. Any resulting
cluster which does not contain a sequence with a polyadenylation
signal or two labelled 3' ESTs is discarded. Clusters which meet
these criteria are called anchored clusters, since their 3' end is
presumed to be known.
Clone-based edges are added; these ensure that nonoverlapping 5'
and 3' ESTs belong to the same cluster. First, clone based edges
which link at least two 5' ends to a single cluster which contains
at least two 3' ends from the same clones are found. These clone ID
based edges which are duplicated within a cluster are retained even
if they cause clusters from the preceding stage to be merged. Due to
imperfect clone labelling, a single clone-ID based edge is
insufficient to merge two clusters.
ESTs which do not belong to an anchored cluster are
rechecked at a lower level of stringency than in the preceding
passes. An EST which passes this less stringent test is then added
to the cluster which contains the sequence which is the best match
to the EST; it is a guest member.
Clusters of size 1 (that is, clusters which seem to identify
infrequently expressed genes) are compared against the rest of the
sequences in UniGene at a lower level of stringency, and merged with
the cluster containing the most similar sequence.
The resulting clusters are compared with the preceding week's
build and renumbered in an attempt to maintain continuity. Since the
sequences which make up a cluster may change from week to week, and
since the cluster identifier may disappear (typically when two
clusters merge) using the cluster identifier as a reference is
ill-advised. Using the GB accession numbers of the sequences which
comprise the cluster is a safe alternative.
|