Preclustering is a technique used to partition the input data into groups small enough for the assembly program to process. Even on powerful computers, an assembler such as CAP3 or PHRAP can not effectively run with more than 20,000 input sequences. Either the memory requirements are too large or the runtime is unacceptably large.

By preclustering, we reduce the input size into disjoint groups of sequences which are not at all similar to any of the sequences in other groups. This limits the work on the assembler by excluding sequences which are obviously not transcripts from the same gene. Thus, the assembly program is used to decide (and assembly into contigs) the number of unique transcripts in a "cluster" of similar ESTs, preclustering is used to partition the input set into disjoint clusters of similar sequences which are small enough to allow the assembler to run efficiently.

Transitive Closure Clustering

There are many general methods of clustering data. For purposes of partitioning data into disjoint sets for unigene assembly, we use a simple method which we call "transitive closure clustering." The same methodology has been described elsewhere as "single-linkage clustering."

Pairwise scores are found for all pairs of sequences. If the score for a pair of sequences is higher than some given threshold, the pair is considered linked. If A is linked to B, and B is linked to C, then A, B, and C are clustered together, even if A is not considered linked with C. Hence, the linkage relationship is transitive, and a cluster is found by finding the transitive closure of the linkage relationship.

In context of unigene assembly, this effectively yields disjoint clusters of sequences for which no sequence in a given cluster has a detectable coarse overlap with any sequence in any other cluster. Thus, there is no possibility for contig assembly of two sequences which are in different clusters, so the exclusion does not in theory alter the outcome of the assembly step. Since the preclustering pairwise comparisons are much more efficient coarse approximations than the assembler's full alignments, the overall runtime and resource consumption of the unigene build becomes manageable.