A new machine learning algorithm, LESTAT (LEngth and STructure-based sequence Alignment Tool) has been developed for detecting protein homologs having low-sequence identity. LESTAT is an iterative profile-based method that runs without reliance on a predefined library and incorporates several novel features that enhance its ability to identify remote sequences. To overcome the inherent bias associated with a single starting model, LESTAT utilizes three structural homologs to create a profile consisting of structurally conserved positions and block separation distances. Subsequent profiles are refined iteratively using sequence information obtained from previous cycles. Additionally, the refinement process incorporates a "lock-in" feature to retain the high-scoring sequences involved in previous alignments for subsequent model building and an enhancement factor to complement the weighting scheme used to build the position specific scoring matrix. A comparison of the performance of LESTAT against PSI-BLAST for seven systems reveals that LESTAT exhibits increased sensitivity and specificity over PSI-BLAST in six of these systems, based on the number of true homologs detected and the number of families these homologs covered. Notably, many of the hits identified are unique to each method, presumably resulting from the distinct differences in the two approaches. Taken together, these findings suggest that LESTAT is a useful complementary method to PSI-BLAST in the detection of distant homologs.
2007 Wiley-Liss, Inc.