Designing Multiple Simultaneous Seeds for DNA Similarity Search

1
Designing Multiple Simultaneous Seeds for DNA
Similarity Search
Yanni Sun, Jeremy Buhler Washington University
in Saint Louis
2
Outline

• Problem of multi-seed design
• Methods
• Greedy covering algorithm
• Compute conditional match probabilities
• Experiments and results
• Conclusion and future work

3
Sequence Alignment

• Functional regions conserved despite DNA
  mutations over time
• Conserved region can be aligned with high score
• Exact solution DP time complexity O(MN)
• Fast but heuristic solution seeded alignment
  algorithm

4
Seeded Alignment Algorithm

• BLAST is the most popular tool.
• Step 1 word match step 2 extend
  the match to find the high
  similarity pair
  
• TAGGACCTAACC
• GACCACCTTTT

5
Seed and Similarity

• Example of a similarity and a single seed
• tgcagaaatgcagaggca
• tacacaggcaccgaggag
• Similarity 101101000010111100
• Seed 111, weight 3, span 4
• The seed detects/matches this similarity.

6
Seed Choice is Important
1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
Significant alignment
Seed match
7
Seed Design Previous Work

• Traditional seed word (e.g. 11111111111)
• Discontiguous patterns of matching bases
  CR1993 MTL02 111010010100110111
• Our work on single discontiguous seed BKS03

8
Multiple Simultaneous Seeds

• Multiple simultaneous seeds are defined as a set
  of seeds.
• ? seed1, seed2,seed i,, seedn
• ? detects a similarity if at least one of the
  component seeds detects the similarity
• Example
• Simultaneous seeds 111, 111 detect
  similarities 100110100001, 1000010110001,
  1101001011001

9
Multi-seed Design Balance Sensitivity with
Specificity

• SensitivityA / Biologically
• meaningful alignments
• SpecificityA / seed matches
• Increase sensitivity
• Decrease weight of single seed
• Use multiple seeds
• Both methods hurt specificity
• Hypothesis a set of multiple seeds
• has a better tradeoff of sensitivity vs.
  specificity comparing to single seed

biologically meaningful alignments
A
seed matches
10
Our Work Design Multiple Simultaneous Seeds
Efficiently

• Use a new local search method to optimize seed
  set
• Design an efficient algorithm to calculate
  conditional match probability
• Empirical verification that multiple simultaneous
  seeds have better tradeoff of sensitivity vs.
  specificity

11
Multi-seed Design Problem

• Input
• Ungapped alignments sampled from two genomic DNA
  sequences
• Resource constraints of seeds weight, span,
  number
• Goal find a set of seeds ? to maximize the
  detection probability Pr? detects S.
• Pr(? detects S) Pr( (seed1 detects S) or (seed2
  detects S)or (seedn detects S))

12
Outline

• Problem of multi-seed Design
• Methods
• Greedy covering algorithm
• Compute conditional match probabilities
• Experiments and results
• Conclusion and future work

13
Computing Match Probability for Specified Seeds
BKS 03

• Learn a kth-order Markov model from similarities.
  
• Build a DFA that only accepts strings containing
  the given seeds
• Compute the probability that the DFA accepts a
  string chosen randomly from model M by DP.

14
Seek the Locally Optimal Set of Seeds

• Original local search
• Greedy covering algorithm a faster local search
  strategy
• Efficient computation of conditional match
  probability

15
Find Optimal Set of Seeds by Original Local
Search
Seed space with spanlt8,weight3
111, 111 Pr0.70
16
Greedy Covering Algorithm
Similarity space
Design 3 simultaneous seedss1,s2,s3
s1 argmaxxPr(x) s2argmaxx Pr(xs1) s3argmaxx
Pr(xs1,s2)
17
Calculate Conditional Match Probabilities

• Challenge how to calculate the conditional
  probability efficiently ?
• Seeds with small span exact computation via DFAs
• Seeds with large span Monte Carlo

18
Calculate Conditional Match Probability via DFA

• Pr( x ) Pr(x )/ Pr( )
• Build DFA corresponding to x by using
  cross product and complementation of DFA
• Efficiency in the process of local search to
  find optimal single seed x, Pr( ) can be
  precomputed

19
Outline

• Problem of multi-seed design
• Methods
• Greedy covering algorithm
• Compute conditional match probabilities
• Experiments and results
• Conclusion and future work

20
Greedy Covering vs. Original Local Search
Detection probability
21
Greedy Covering is Much Faster

• When n5, on the same hardware platform(P4)
• Greedy covering needs 20 minutes
• The original local search needs 2.4 hours

22
Experimental Setup

• The ungapped alignments are sampled uniformly
  from human and mouse syntenies
• For a specified seed set
• sensitivity the number of significant gapped
  alignments found by our BLAST-like alignment tool
• False positive rate approximated by the number
  of seed matches

23
Results Verify the Hypothesis on Noncoding
Sequences
seed weight number of seeds gapped alignments found (sensitivity) improvement of sensitivity total seed matches (approximation of f.p)
11 1 251941 ---- 1.57x109
10 1 273831 8.7 5.88x109
11 3 292093 15.9 4.56x109
24
Summary of Contributions

• Efficient algorithms to design multiple
  simultaneous seeds at reasonable cost
• Empirical verification multiple simultaneous
  seeds have a better tradeoff between sensitivity
  and specificity

25
Future Work

• Design a better evaluation platform for different
  seeds
• Investigate utility of seeds in multiple sequence
  alignment

26
Acknowledgements

• Dr. Jeremy Buhler (advisor), Ben Westover, Rachel
  Nordgren, Joseph Lancaster and Christopher Swope
• Laboratory for computational genomics in
  Washington University in Saint Louis
• http//www.cse.wustl.edu/jbuhler/mandala