1.6

1.6 (June 1994) is a major overhaul from the previous releases. All reported bugs and problems have been fixed. Several important parts of the code have been reimplemented; most notably, the HMM alignment algorithms themselves.

Changes since 1.5:

• Large speed increases in all the alignment algorithms. Expect 50% average speed gains in searching, training, and alignment; some applications, such as hmmsw, show a three-fold speed gain on my benchmarks. A database search with a 150-position HMM of SWIR4 (about 50,000 proteins) will now take less than an hour on an R4000 Indigo.
• The program hmmfs, ``HMM fragment search'', is added. hmmfs implements correct multiple-hit Smith/Waterman searching for repeated motifs such as Alus or immunoglobulin superfamily domains (for instance). An optimal non-overlapping set of hits is reported; all reported problems with too-long tracebacks and the ad hoc overlap-filtering code are now fixed (except in hmmls, where the problems temporarily remain). hmmsw now reverts to single best hit Smith/Waterman, and is best for protein database searches. The scores reported by hmmsw and hmmfs are (finally) theoretically justifiable, because both programs now use full probabilistic models appropriate to their tasks; the need to artificially set high cutoffs for hmmsw is gone.
• The code is usefully portable to the Alliant for the first time. (It was portable before, but it ran at 0.01 speed for arcane reasons.)
• The long-standing simulated annealing bug has been fixed. Simulated annealing [hmmt] and suboptimal alignments [hmms] now work as advertised.
• The program hmm-convert is added, for converting saved models to/from compact binary and architecture-independent ASCII formats. hmmb and hmmt save in compact binary format by default; use hmm-convert if you want to examine or modify the files yourself, or if you need to move models to different machine architectures.
• There's been a general cleanup of the code. Several rarely used and/or experimental options have been removed or changed. Double-check the usage of each program to see if your favorite option has been affected.