Reading a book such as this brings home how much BLAST-now in its teenage years-has grown, and provides an occasion for fond reflection. BLAST was born in the first months of 1989 at the National Center for Biotechnology Information (NCBI). The Center had been created at the National Institutes of Health in November 1988, by an act of the U.S. Congress, to foster the development of a field that then had no widely accepted name, but which has since come to be known as "Bioinformatics." In early 1989, David Lipman, my post-doctoral advisor, who at the time was perhaps best known as a codeveloper of the FASTA program, was appointed director of NCBI. On the first of March we moved into new offices at the National Library of Medicine.The NCBI was small, but had large ambitions, and already a number of friends. Several of these well-wishers made it a point to drop by for a visit. Gene Myers, a computer scientist then at Arizona, arrived during a week in which Science was hyping a special-purpose computer chip for sequence comparison. He and David, software partisans both, were unimpressed and over dinner resolved to do better. Their original idea was to find not subtle sequence similarities, but fairly obvious ones, and to do it in a flash. Gene pursued a rigorous approach at first, but David, with a fine Darwinian wisdom, was willing to settle for imperfection. If one were to gamble, what kind of match could one expect a strong alignment to contain? Detailed algorithmic and code development on BLAST by Webb Miller-later to be joined by Warren Gish-had hardly begun before Sam Karlin, a Stanford mathematician, came calling. I had approached him a few months earlier with a conjecture concerning the asymptotic behavior of optimal ungapped local sequence alignments. Since then, he had spun this conjecture into a beautiful theory. Now, for the first time, rigorous statistics were available for alignment scoring systems of more than academic interest, and the essential nature of amino acid substitution matrices also began to come into clear focus. This theory dovetailed perfectly with the work that had just started on BLAST: both informing the selection of its algorithmic parameters, and yielding units for the alignment scores produced.
Although David chose BLAST's name as a bit of a pun on "FASTA" (it was only later that I realized "BLAST" to be an acronym), the new program was never intended to vie with the earlier one. Rather, the idea was to turn the "threshold parameter" way up, to find undoubted homologies before you take more than one sip of coffee. It surprised us all when BLAST started returning most weak similarities as well. Thus was born a sort of friendly competition with Bill Pearson's and David's earlier creation. From the start, BLAST had two major advantages to FASTA and one major disadvantage. In the plus column, BLAST was indeed much the faster, and it also boasted Sam's new statistics, which turned raw scores into E-values. However, BLAST could only produce ungapped local alignments, thereby often eliding large regions of similarity and sometimes completely missing weak alignments that FASTA, in its most sensitive but slowest mode, was able to find. These points of comparative advantage were healthy for both programs. In time, FASTA fit its scores to the extreme value distribution, yielding reliable statistical evaluations of its output. And by the mid '90s, Warren Gish's WU-BLAST from Washington University, and NCBI's BLAST releases, introduced gapped alignments, using differing algorithmic strategies. The result, at least for protein sequence comparisons, is that BLAST and FASTA have converged in many important ways, although there still remain significant differences.
The programs comprehended by the name "BLAST" have multiplied astonishingly in the nearly 15 years since the first one was conceived. Learning the best way to use these various programs for research can be a challenge, and this book is a significant aid.While BLAST's developers have done their best to make the programs' default behavior the most generally applicable, a sophisticated user still has many issues to consider.
To achieve speed, BLAST is a heuristic program. It isn't guaranteed to find every local alignment that passes its reporting criteria, and there are an array of parameters that control the shortcuts it takes.With the introduction of gapped alignments, the programs' complexity increased, as did the number of parameters that influence BLAST's tradeoff of speed and sensitivity. In a certain sense, however, these mechanics are the least important for a user to understand because, except for the occasional appearance or disappearance of a weak similarity, they don't greatly effect the programs' output. Perhaps of more importance is an understanding of attendant matters that are relevant to the effective use of any local alignment search method, such as the filtering of "low-complexity" sequence regions, the proper choice of scoring systems, and the correct interpretation of statistical significance. This book deals with these and many other matters, and nicely combines theoretical considerations with practical advice informed by these considerations.
The BLAST programs have been the fruit of much hard work by scores of talented programmers and scientists. This work continues, linking BLAST output to other databases, improving alignment formatting options, refining the types of queries that may be performed. Newer offshoots, such as PSI-BLAST for protein profile searches, also continue under development, and BLAST is thus a moving and a growing target. This book should prove a valuable guide for those wishing to use the programs to best effect.
June 26, 2003