Computing Bi-Clusters for Microarray Analysis

Transcription

1 December 21, 2006

2 General Bi-clustering Problem Similarity of Rows (1-5) Cheng and Churchs model General Bi-clustering Problem Input: a n m matrix A. Output: a sub-matrix A P,Q of A such that the rows of A P,Q are similar. That is, all the rows are identical. Why sub-matrix? A subset of genes are co-regulated and co-expressed under specific conditions. It is interesting to find the subsets of genes and conditions.

3 General Bi-clustering Problem Similarity of Rows (1-5) Cheng and Churchs model Similarity of Rows (1-5) 1. All rows are identical All the elements in a row are identical (the same as 1 if we treat columns as rows)

4 General Bi-clustering Problem Similarity of Rows (1-5) Cheng and Churchs model Similarity of Rows (1-5) 3. The curves for all rows are similar (additive) a i,j a i,k = c(j, k) for i = 1, 2,..., m. Case 3 is equivalent to case 2 (thus also case 1) if we construct a new matrix a i,j = a i,j a i,p for a fixed p indicate a row.

5 General Bi-clustering Problem Similarity of Rows (1-5) Cheng and Churchs model Similarity of Rows (1-5) 4. The curves for all rows are similar (multiplicative) a 1,1 a 1,2 a 1,3... a 1,m c 1 a 1,1 c 1 a 1,2 c 1 a 1,3... c 1 a 1,m c 2 a 1,1 c 2 a 1,2 c 2 a 1,3... c 2 a 1,m... c n a 1,1 c n a 1,2 c n a 1,3... c n a 1,m Transfer to case 2 (thus case 1) by taking log and subtraction. Case 3 and Case 4 are called bi-clusters with coherent values.

6 General Bi-clustering Problem Similarity of Rows (1-5) Cheng and Churchs model Similarity of Rows (1-5) 5. The curves for all rows are similar (multiplicative and additive) a i,j = c i a k,j + d i Transfer to case 2 (thus case 1) by subtraction of a fixed row (row i), taking log and subtraction of row i again. The basic model: All the rows in the sub-matrix are identical.

7 General Bi-clustering Problem Similarity of Rows (1-5) Cheng and Churchs model Cheng and Churchs model The model introduced a similarity score called the mean squared residue score H to measure the coherence of the rows and columns in the submatrix. where H(P, Q) = 1 P Q a i,q = 1 Q j Q i P,j Q a i,j, a P,j = 1 P (a i,j a i,q a P,j + a P,Q ) 2 i P a i,j, a P,Q = 1 P Q i P,j Q If there is no error, H(P, Q)=0 for case 1, 2 and 3. A lot of heuristics (programs) have been produced. a i,j.

8 Consensus Sub-matrix Problem Bottleneck Sub-matrix Problem NP-Hardness Results Consensus Sub-matrix Problem Bottleneck Sub-matrix Problem

9 Consensus Sub-matrix Problem Bottleneck Sub-matrix Problem NP-Hardness Results Consensus Sub-matrix Problem Input: a n m matrix A, integers l and k. Output: a sub-matrix A P,Q of A with l rows and k columns and a consensus row z (of k elements) such that r i P d(r i Q, z) is minimized. Here d(, ) is the Hamming distance.

10 Consensus Sub-matrix Problem Bottleneck Sub-matrix Problem NP-Hardness Results Bottleneck Sub-matrix Problem Input: a n m matrix A, integers l and k. Output: a sub-matrix A P,Q of A with l rows and k columns and a consensus row z (of k elements) such that for any r i in P d(r i Q, z) d and d is minimized Here d(, ) is the Hamming distance.

11 Consensus Sub-matrix Problem Bottleneck Sub-matrix Problem NP-Hardness Results NP-Hardness Results Theorem 1: Both consensus sub-matrix and bottleneck sub-matrix problems are NP-hard. Proof: We use a reduction from maximum edge bipartite problem.

12 for Consensus Sub-matrix Problem Input: a n m matrix A, integers l and k. Output: a sub-matrix A P,Q of A with l rows and k columns and a consensus row z (of k elements) such that r i P d(r i Q, z) is minimized. Here d(, ) is the Hamming distance.

13 Assumptions: H opt = p i P opt d(x pi Qopt, z opt) = O(kl), H opt c = kl and Q opt = k = O(n), k c = n. : We use a random sampling technique to randomly select O(logm) columns in Q opt, enumerate all possible vectors of length O(logm) for those columns. At some moment, we know O(logm) bits of r opt and we can use the partial z opt to select the l rows which are closest to z opt in those O(logm) bits. After that we can construct a consensus vector r as follows: for each column, choose the (majority) letter that appears the most in each of the l letters in the l selected rows. Then for each of the n columns, we can calculate the number of mismatches between the majority letter and the l letters in the l selected rows. By selecting the best k columns, we can get a good solution.

14 How to randomly select O(logm) columns in Q opt while Q opt is unknown? Our new idea is to randomly select a set B of (c + 1)logm columns from A and enumerate all size logm subsets of B in time O(m c+1) which is polynomial in terms of the input size O(mn). We can show that with high probability, we can get a set of logm columns randomly selected from Q opt.

15 Algorithm 1 for The Consensus Submatrix Problem Input: one m n matrix A, integers l and k, and ɛ > 0 Output: a size l subset P of rows, a size k subset Q of columns and a length k consensus vector z Step 1: randomly select a set B of (c + 1)( 4 log m + 1) columns from A. ɛ 2 (1.1) for every size 4 log m ɛ 2 subset R of B do (1.2) for every z R Σ R do (a) Select the best l rows P = {p 1,..., p l } that minimize d(z R, x i R ). (b) for each column j do Compute f (j) = l i=1 d(s j, a pi,j), where s j is the majority element of the l rows in P in column j. Select the best k columns Q = {q 1,..., q k } with minimum value f (j) and let z(q) = s q1 s q2... s qk. (c) Calculate H = l i=1 d(xp i Q, z) of this solution. Step 2: Output P, Q and z with minimum H.

16 Lemma 1: With probability at most m 2 ɛ 2 c 2 (c+1), no subset R of size 4 log m used in Step 1 of Algorithm 1 satisfies ɛ 2 R Q opt. Lemma 2: Assume R = 4 log m and R Q ɛ 2 opt. Let ρ = k R. With probability at most m 1, there is a row x i in X satisfying d(z opt, x i Qopt ) ɛk ρ > d(z opt R, x i R ). With probability at most m 1 3, there is a row x i in X satisfying d(z opt R, x i R ) > d(z opt, x i Qopt ) + ɛk. ρ

17 Lemma 3: When R Q opt and z R = z opt R, with probability at most 2m 1 3, the set of rows P = {p 1,..., p l } selected in Step 1 (a) of Algorithm 1 satisfies l i=1 d(z opt, x pi Qopt ) > H opt + 2ɛkl. Theorem 2: For any δ > 0, with probability at least 2 1 m 8c δ 2 c 2 (c+1) 2m 1 3, Algorithm 1 will output a solution with consensus score at most (1 + δ)h opt in O(nm O( 1 δ 2 ) ) time.

18 Thanks for Bottleneck Sub-matrix Problem Input: a n m matrix A, integers l and k. Output: a sub-matrix A P,Q of A with l rows and k columns and a consensus row z (of k elements) such that for any r i in P d(r i Q, z) d and d is minimized Here d(, ) is the Hamming distance.

19 Thanks Assumptions: d opt = MAX pi P opt d(x pi Qopt, z opt ) = O(k), d opt c = k and Q opt = k = O(n), k c = n. : (1) Use random sampling technique to know O(logm) bits of z opt and select l best rows like Algorithm 1. (2) Use linear programming and randomized rounding technique to select k columns in the matrix.

20 Thanks Linear programming Given a set of rows P = {p 1,..., p l }, we want to find a set of k columns Q and vector z such that bottleneck score is minimized. min d; Σ n y i,j = k, i=1 Σ j=1 j=1 y i,j 1, i = 1, 2,..., n, Σ n χ(π j, x ps,i)y i,j d, s = 1, 2,..., l. i=1 j=1 y i,j = 1 if and only if column i is in Q and the corresponding bit in z is π j. Here, for any a, b Σ, χ(a, b) = 0 if a = b and χ(a, b) = 1 if a b.

21 Thanks Randomized rounding To achieve two goals: (1) Select k columns, where k k δd opt. (2) Get integers values for y i,j such that the distance (restricted on the k selected columns) between any row in P and the center vector thus obtained is at most (1 + γ)d opt. Here δ > 0 and γ > 0 are two parameters used to control the errors.

22 Thanks Lemma 4: When nγ 2 3(cc ) 2 2 log m, for any γ, δ > 0, with probability at most exp( nδ2 ) + m 1, the rounding result 2(cc ) 2 y = {y 1,1,..., y 1, Σ,..., y n,1,..., y n, Σ } does not satisfy at least one of the following inequalities, Σ n ( y i,j) > k δd opt, i=1 j=1 and for every row x ps (s = 1, 2,..., l), Σ n ( χ(π j, x ps,i)y i,j) < d + γd opt. i=1 j=1

23 Thanks Algorithm 2 for The bottleneck Sub-matrix Problem Input: one matrix A Σ m n, integer l, k, a row z Σ n and small numbers ɛ > 0, γ > 0 and δ > 0. Output: a size l subset P of rows, a size k subset Q of columns and a length k consensus vector z. nγ 2 3(cc ) 2 if 2 log m then try all size k subset Q of the n columns and all z of length k to solve the problem. if > 2 log m then nγ 2 3(cc ) 2 Step 1: randomly select a set B of size subset R of B do 4 log m ɛ 2 4(c+1) log m columns from A. for every ɛ 2 for every z R Σ R do (a) Select the best l rows P = {p 1,..., p l } that minimize d(z R, x i R ). (b)solve the optimization problem by linear programming and randomized rounding to get Q and z. Step 2: Output P,Q and z with minimum bottleneck score d.

24 Thanks Lemma 5: When R Q opt and z R = z opt R, with probability at most 2m 1 3, the set of rows P = {p 1,..., p l } obtained in Step 1(a) of Algorithm 2 satisfies d(z opt, x pi Qopt ) > d opt + 2ɛk for some row x pi (1 i l). Theorem 3: With probability at least 1 m 2 ɛ 2 c 2 (c+1) 2m 1 3 exp( nδ2 ) m 1, Algorithm 2 2(cc ) 2 runs in time O(n O(1) m O( 1 ɛ γ 2 ) ) and obtains a solution with bottleneck score at most (1 + 2c ɛ + γ + δ)d opt for any fixed ɛ, γ, δ > 0.

25 Thanks Thanks Acknowledgements This work is fully supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China [Project No. CityU 1070/02E]. This work is collaborated with Dr. Lusheng Wang and Xiaowen Liu in City University of Hong Kong, Hong Kong, China.

26 Thanks Let X 1, X 2,..., X n be n independent random 0-1 variables, where X i takes 1 with probability p i, 0 < p i < 1. Let X = n i=1 X i, and µ = E[X ]. Then for any 0 < ɛ 1, Pr(X > µ + ɛ n) < e 1 3 nɛ2, Pr(X < µ ɛ n) e 1 2 nɛ2.