Algorithms for Regulatory Motif Discovery

1
Algorithms for Regulatory Motif Discovery

• Xiaohui Xie
• University of California, Irvine

2
Positional weight matrix representation
Lambda cI/cro binding sites
A 9 9 94 25 1 71 1 1 1 9 17 32 9 17 1 48 1 71 63
C 17 17 1 25 94 9 86 55 9 40 71 9 1 1 1 1 1 1 9
G 9 1 1 1 1 1 1 9 71 40 9 55 86 9 94 25 1 17 17
T 63 71 1 48 1 17 9 32 17 9 1 1 1 71 1 25 94 9 9
Sequence Logo
3
Probabilistic model for motif discovery

• Input a set of sequence S1, S2 ,SN, each of
  which is of length w, i.e. Siw for all i.
• Two models
• Motif model P(y?)
• Background model P(y?)
• Hidden variables z1, z2 ,zN, where zi 1 if
  Si is a motif site and 0 otherwise.
• P(Sizi1) P(Si?) and P(Sizi0) P(Si?)
• P(zi1Si) P(Sizi1)P(zi1)/P(Sizi1)P(zi1)
  P(Sizi0)P(zi0)

4
Probabilistic model for motif discovery

• Input a set of sequences S1, S2 ,SN, each of
  which is of length w, i.e. Siw for all i.
• Parameter estimation problem
• Find the ? and ? that maximize P(S1, S2 ,SN ?,
  ?)
• If zi 1 for all i, then ?ij cij/N where cij is
  the number of letter j occurring at position i in
  the set of sequences.

5
String matching
6
Exact String Matching

• Input Two strings T1n and P1m, containing
  symbols from alphabet ?.
• Example
• ? A,C,G,T
• T112 CAGTACATCGAT
• P1..3 AGT
• Goal find all shifts 0s n-m such that
  Ts1sm P

7
Simple Algorithm

• for s ? 0 to n-m
• Match ? 1
• for j ? 1 to m
• if Tsj?Pj then
• Match ? 0
• exit loop
• if Match1 then output s

8
Analysis

• Running time of the simple algorithm
• Worst-case O(nm)
• Average-case (random text) O(n) (expectation)
• Ts time spend on checking shift s
• (the number of comparisons until 1st mismatch)
• ETs lt 2 (why)
• ESsTs SsETs O(n)

9
Worst-case

• Is it possible to achieve O(n) for any input ?
• Knuth-Morris-Pratt77 deterministic
• Karp-Rabin81 randomized
• Digression what is the difference between
• Algorithm that is fast on a random input (as seen
  on the previous slide)
• Randomized algorithm (as in the rest of this
  lecture)

10
Karp-Rabin Algorithm

• Idea semi-numerical approach
• Consider all m-mers
• T1m, T2m1, , Tm-n1n
• Map each Ts1sm into a number ts
• Map the pattern P1m into a number p
• Report the m-mers that map to the same value as p
• Problem how to map all m-mers in O(n) time ?

11
Implementation

• Attempt I
• Assume S0,1
• (for A,C,G,T convert A00, C01, G10, T11)
• Think about each Ts1sm as a number in binary
  representation, i.e.,
• tsTs12m-1Ts22m-2Tsm20
• Find a fast way of computing ts1 given ts
• Output all s such that ts is equal to the number
  p represented by P

12
Magic formula

• How to transform
• tsTs12m-1Ts22m-2Tsm20
• into
• ts1Ts22m-1Ts32m-2Tsm120 ?
• Three steps
• Subtract Ts12m-1
• Multiply by 2 (i.e., shift the bits by one
  position)
• Add Tsm120
• Therefore ts1 (ts- Ts12m-1)2 Tsm120

13
Algorithm

• ts1 (ts- Ts12m-1)2 Tsm120
• Can compute ts1 from ts using 3 arithmetic
  operations
• Therefore, we can compute all t0,t1,,tn-m using
  O(n) arithmetic operations
• We can compute a number corresponding to P using
  O(m) arithmetic operations
• Are we done ?

14
Problem

• To get O(n) time, we would need to perform each
  arithmetic operation in O(1) time
• However, the arguments are m-bit long !
• If m large, it is unreasonable to assume that
  operations on such big numbers can be done in
  O(1) time
• We need to reduce the number range to something
  more managable

15
Hashing

• We will instead compute
• tsTs12m-1Ts22m-2Tsm20 mod q where
  q is an appropriate prime number
• One can still compute ts1 from ts
• ts1 (ts- Ts12m-1)2Tsm120 mod q
• If q is not large, we can compute all ts (and
  p) in O(n) time

16
Problem

• Unfortunately, we can have false positives, i.e.,
  Ts1sm?P but ts mod q p mod q
• Our approach
• Use a random q
• Show that the probability of a false positive is
  small
• ? randomized algorithm

17
Aligning two sequences

• Longest common substring (LCS) problem (no gaps)
• Input Two strings T1n and P1m, nm
• Goal Largest k such that
• Ti1ik Pj1jk
• for some i,j
• How can we solve this problem efficiently ?
• Hint How can we check if LCS has length k ?

18
Checking for a common m-mer