Accelerating sequence alignment and homology search algorithms

1
Accelerating sequence alignment and homology
search algorithms

• BioE131/231

2
DP Accelerator Hardware
3
Faster alignment

• Smith-Waterman, Gotoh, etc., guarantee the
  best-scoring alignment
• However, they are too slow to be practical
  (except with expensive h/w)
• Subsequent algorithms attempted to provide
  approximate solutions
• FASTA
• BLAST

4
FASTA (1988)

• Acceleration of the Smith-Waterman algorithm
• Identify strong diagonals in dotplots
• Use exact word hits to find these diagonals
• Rescore best N diagonals with approximate word
  hits
• Try to join together good diagonals
• For sequence-pairs that score well under these
  criteria, do full SW alignment

5
Homology Search

• Good alignments have ungapped diagonal runs
• Such runs contain conserved words
• BLAST lookup table of high-scoring words

6
BLAST (1990)

• Basic Local Alignment Search Tool
• Altschul, Gish, Miller, Myers, Lipman
• Motivation FASTA still too slow
• Idea find fewer, but better, seed hits
• Uses longer word sizes
• Allows approximate word hits in initial search
  phase

7
BLAST terminology

• Segment pair
• Pair of equal-length substrings of sequences X
  and Y
• Locally maximal segment pair
• Segment pair whose score cannot be improved by
  extending it
• Maximum segment pair (MSP)
• Segment pair with highest score of all segment
  pairs between X and Y

8
BLAST outline

• Query sequence, X
• Database of target sequences, Y
• Find all Ys such that
• X and Y have an MSP of score gtS
• Such MSPs are called HSPs
• High-Scoring segment Pairs
• Key choose S large enough so as to minimize
  false positives
• but small enough to get real positives

9
BLAST parameters

• Word length W
• Threshold T
• A hit is a W-length sequence that has score gtT
  when aligned to some W-length sequence in X
• Typically, W is
• 3-5 for amino acids
• 11-12 for DNA
• Choice of T is critical to separate signal from
  noise

10
BLAST algorithm

• Make a list of all length-W words in X
• For each word
• Precalculate score of exact match
• Make a list of nearby words that score gtT
• This greatly reduces the number of words
• From exponential to linear in W, if T is large
  enough

11
BLAST search algorithm

• Search database for hits (matches to precomputed
  word list)
• This can be done very fast, e.g. with a trie

Trie for words i, in, inn, to, tea,
ten
12
Original vs gapped BLAST

• Original BLAST
• For every hit, extend match to locally maximal
  segment (MSP) and check if score gtS
• Gapped BLAST
• Require two hits before you start
• Improves speed by 3X
• Hits must be within distance A
• Use with a lower T threshold

13
BLAST statistics

• Extreme value distribution

14
Masking

• Repetitive or low-complexity sequence generates
  spurious hits
• Hard masking
• Replace low-complexity sequence with Ns
• Soft masking
• Use lower case charactersAGCAGTGatatatatatatatat
  aatatatGAGTCT