Previous Next Top of Manual Home Page

PSI-BLAST Benchmark

Explanation & Summary

In out benchmark be focus on remote homology. Long alignments (>300 residues) with less than 20% sequence identity defines remote homology (the opposite is close homology). The benchmark is based on SCOP (Structural Classification Of Proteins). In SCOP proteins that share a common function and have a closely related three dimensional structure are put into the same superfamily. Although these proteins sometimes do not share obvious sequence similarity a superfamily represents homologues (proteins diverged from a common ancestor).

The result of our benchmark are shown below. In the table sequences are chains (single or multi-domain proteins). For sequences 'correctly assigned' means there is at least one domain correctly assigned regardless whether there are also false assignments in the same sequence. Accordingly 'false' in sequences context means there is at least one false domain assignment in the sequence. An 'over assignment' is e.g. when two domains are assigned to a region in the query that represents only one domain. A 'Correct' assignment is when the domain of the query and the assignment (target) are of the same superfamily. 'Residues' is the sum of domain length.

Accurary of genome assignment

No. in SCOP genome 1254 1621 263,863
No. correctly assigned 503 652 84,827
Coverage of correct assignments(%) 40 40 32
No. of false positive assignments 13 16 1,985
No. of over assignments 2 2 163

The pie-chart summarises the results (bases on residues). 'missing' are assignments that could have been found by PSI-BLAST (because there are homologues proteins in the target database). The ration of missing to true positives is 2.1 and can be used to estimate the fraction of undetected remote homologues in real genomes.

pie-chart of benchmark results

Ratio of detected to undetected remote homologues is 2.1

Previous Next Top of Manual Home Page

Copyright © 1999-2002 Cancer Research UK
All Rights Reserved, disclaimer
Comments to author: a.mueller@cancer.org.uk
Generated: Thu Jun 27, 2002