Eddy introduced an accelerated profile HMM search method called the "multiple segment Viterbi" (MSV) algorithm, which significantly improves the speed of HMMER3, a software package for sequence database homology searches. The MSV algorithm computes optimal sum of multiple ungapped local alignment segments using a striped vector-parallel approach, similar to the fast Smith/Waterman alignment method. This allows rapid evaluation of the significance of MSV scores, making it a useful heuristic filter. Additionally, a 20-fold acceleration of the standard profile HMM Forward/Backward algorithms was achieved using a method called "sparse rescaling." These methods are combined in a pipeline where high-scoring MSV hits are passed on for reanalysis with the full HMM Forward/Backward algorithm. HMMER3, which implements this accelerated pipeline, is about as fast as BLAST for protein searches while maintaining the sensitivity of profile HMM methods. Performance benchmarks show that the use of the heuristic MSV filter sacrifices negligible sensitivity compared to unaccelerated profile HMM searches. HMMER3 is substantially more sensitive and 100- to 1000-fold faster than HMMER2. The MSV algorithm bypasses two of BLAST's main heuristics, providing a more sensitive overall heuristic. The MSV model is an ungapped version of HMMER3's multihit local alignment model, and its probabilistic model of multihit ungapped local alignment is achieved by ignoring the match, delete, and insert state transitions of the original profile. The MSV score is essentially analogous to BLAST's "sum score" of one or more ungapped HSPs. The MSV model allows for efficient calculation of scores using vector-parallel techniques, enabling a 16-fold acceleration. The MSV filter is a heuristic acceleration that allows for the rapid calculation of P-values, enabling the use of MSV scores as a tunable and selective sequence filter. The MSV model's scores obey conjectures about the expected Gumbel distribution of probabilistic local alignment scores, allowing for the rapid calculation of P-values. The MSV filter is implemented in HMMER3, which is about 100-fold faster than the previous version of HMMER. The MSV algorithm is implemented in a SIMD vector parallelization approach, which allows for efficient calculation of scores using 128-bit vectors. The MSV algorithm is highly amenable to vector parallelization using commodity SIMD instructions, such as SSE and Altivec/VMX. The MSV algorithm is implemented in HMMER3's source code, and it is about 16-fold faster than standard serial implementations. The MSV algorithm is also used in the Forward/Backward algorithms, which are accelerated using sparse rescaling. The MSV model and its implementation use several features to reduce roundoff error to tolerable limits, including eliminating matchEddy introduced an accelerated profile HMM search method called the "multiple segment Viterbi" (MSV) algorithm, which significantly improves the speed of HMMER3, a software package for sequence database homology searches. The MSV algorithm computes optimal sum of multiple ungapped local alignment segments using a striped vector-parallel approach, similar to the fast Smith/Waterman alignment method. This allows rapid evaluation of the significance of MSV scores, making it a useful heuristic filter. Additionally, a 20-fold acceleration of the standard profile HMM Forward/Backward algorithms was achieved using a method called "sparse rescaling." These methods are combined in a pipeline where high-scoring MSV hits are passed on for reanalysis with the full HMM Forward/Backward algorithm. HMMER3, which implements this accelerated pipeline, is about as fast as BLAST for protein searches while maintaining the sensitivity of profile HMM methods. Performance benchmarks show that the use of the heuristic MSV filter sacrifices negligible sensitivity compared to unaccelerated profile HMM searches. HMMER3 is substantially more sensitive and 100- to 1000-fold faster than HMMER2. The MSV algorithm bypasses two of BLAST's main heuristics, providing a more sensitive overall heuristic. The MSV model is an ungapped version of HMMER3's multihit local alignment model, and its probabilistic model of multihit ungapped local alignment is achieved by ignoring the match, delete, and insert state transitions of the original profile. The MSV score is essentially analogous to BLAST's "sum score" of one or more ungapped HSPs. The MSV model allows for efficient calculation of scores using vector-parallel techniques, enabling a 16-fold acceleration. The MSV filter is a heuristic acceleration that allows for the rapid calculation of P-values, enabling the use of MSV scores as a tunable and selective sequence filter. The MSV model's scores obey conjectures about the expected Gumbel distribution of probabilistic local alignment scores, allowing for the rapid calculation of P-values. The MSV filter is implemented in HMMER3, which is about 100-fold faster than the previous version of HMMER. The MSV algorithm is implemented in a SIMD vector parallelization approach, which allows for efficient calculation of scores using 128-bit vectors. The MSV algorithm is highly amenable to vector parallelization using commodity SIMD instructions, such as SSE and Altivec/VMX. The MSV algorithm is implemented in HMMER3's source code, and it is about 16-fold faster than standard serial implementations. The MSV algorithm is also used in the Forward/Backward algorithms, which are accelerated using sparse rescaling. The MSV model and its implementation use several features to reduce roundoff error to tolerable limits, including eliminating match