Bioinformatics Algorithms and Data Structures Chapter 11 Core

Bioinformatics Algorithms and Data Structures Chapter 11: Core String Edits Lecturer: Dr. Rose Slides by: Dr. Rose January 30 & February 4, 2003 UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Core String Edits • This chapter introduces inexact matching – Inexact matching is used to compute similarity. – Sequences similarity is a key concept. – Sequence similarity implies • Structural similarity • Functional similarity – We will consider a dynamic programming approach to inexact matching. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Edit Distance • One measure of similarity between two strings is their edit distance. • This is a measure of the number of operations required to transform the first string into the other. • Single character operations: – Deletion of a character in the first string – Insertion of a character in the first string – Substitution of a character from the second character into the second string – Match a character in the first string with a character of the second. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Edit Distance Example from textbook: transform vintner to writers vintner replace v with w wintner insert r after w wrintner match i wrintner delete n writner match t writner delete n writer match e writer match r writer insert s writers UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Edit Distance Let S = {I, D, R, M} be the edit alphabet Defn. An edit transcript of two strings is a string over S describing a transformation of one string into another. Defn. The edit distance between two strings is defined as the minimum number of edit operations needed to transform the first into the second. Matches are not included in the count. Edit distance is also called Levenshtein distance. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Edit Distance Defn. An optimal transcript is an edit transcript with the minimal number of edit operations for transforming one string into another. Note: optimal transcripts may not be unique. Defn. The edit distance problem entails computing the edit distance between two strings along with an optimal transcript. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

String Alignment Defn. A global alignment of strings S 1 and S 2 is obtained by: 1. Inserting dashes/spaces into or at the ends of S 1 and S 2. 2. Aligning the two strings s. t. each character/space in either string is opposite a unique character/space in the other string. Example 1: S 1 = qacdbd S 2 = qawxb q a c - d b d q a w x - b UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

String Alignment Example 2: S 1 = vintner S 2 = writers v - i n t n e r w r i - t - e r s • Mathematically, string alignment and edit transcripts are equivalent. • From a modeling perspective they are not equivalent. • Edit transcripts express the idea of mutational changes. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming • Observation 1: There are many possible ways to transform one string into another. Observation 2: This is like the knapsack problem Recall: dynamic programming is used to solve knapsacklike problems. Defn. Let D(i, j) denote the edit distance of S 1[1. . i] and S 2[1. . j]. • • • – That is, D(i, j) is the minimum number of edit ops needed to transform the first i characters of S 1 into the first j characters of S 2. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming • • Notice that we can solve D(i, j) for all combination of lengths of prefixes of S 1 and S 2. Examples: D(0, 0), . . , D(0, j), D(1, 0), . . , D(1, j), … D(i, j) Dynamic programming is a divide and conquer method. The three parts to dynamic programming are: – The recurrence relation – Tabular computation – Traceback UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming • • • The recurrence relation expresses the recursive relation between a problem and smaller instances of the problem. For any recursive relation, the base condition(s) must be specified. Base conditions for D(i, j) are: – D(i, 0) = i Q: Why is this true? What does it mean in terms of edit ops? – D(0, j) = j Q: Why is this true? What does it mean in terms of edit ops? UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming • • • The general recurrence is given by: D(i, j) = min[D(i - 1, j) + 1, D(i, j - 1) + 1, D(i - 1, j - 1) + t (i, j) ] Here t (i, j) = 1 if S 1(i) S 2(j), o/w t (i, j) = 0. Proof of correctness on Pages 218 -219 Basic argument: D(i, j) must be one of : 1. D(i - 1, j) + 1 2. D(i, j - 1) + 1 3. D(i - 1, j - 1) + t (i, j) There are NO other ways of creating S 2[1. . j] from S 1[1. . i]. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming Q: How do we use the recurrence relation to efficiently compute D(i, j) ? Wrong Answer: simply use recursion. Q: Why is the wrong answer? A: recursion results in inefficient duplication of computations for subproblems. Q: How much duplication? A: Exponential duplication! Example: Fibonacci numbers UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming Example: Fibonacci numbers f(n) = f(n - 1) + f(n - 2) Base conditions: f(0) = 0, f(1) = 1 UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming Note: In calculating D(n, m), there are only (n + 1) (m + 1) unique combinations of i and j. Clearly an exponential number of computations is NOT required. Soln: instead of going top-down with recursion, go bottom-up. Compute each combination only once. • • • – – – Decide on a data structure to hold intermediate results. Start from base conditions. These are the smallest D(i, j) values and are already defined. Compute D(i, j) for larger values of i and j. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming • Example: Fibonacci numbers • Decide on a data structure: simple array • Start from base conditions: f(0) = 0, f(1) = 1 • Compute f(i) for larger values of i. From bottom up. • Each f(i) is computed only once! UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming • Q: What kind of data structure should we use for edit distance? 1. Has to be a random access data structure. 2. Has to support the dimensionality of the problem. • • D(i, j) is two-dimensional: S 1 and S 2. We will use a two-dimensional array, i. e. , a table. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming Example: edit distance from vintner to writers. Fill in the base condition values. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming • • • Q: How do we fill in the other values? A: use the recurrence: D(i, j) = min[D(i - 1, j) + 1, D(i, j - 1) + 1, D(i - 1, j - 1) + t (i, j) ] where t (i, j) = 1 if S 1(i) S 2(j), o/w t (i, j) = 0. We can first compute D(1, 1) because we have D(0, 0), D(0, 1), and D(1, 0) – • D(1, 1) = min[ 1+1, 0+1] = 1 Then we have all the values needed to compute in turn D(1, 2), D(1, 3), . . , D(1, m) UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming First compute D(1, 1) because we have D(0, 0), D(0, 1), and D(1, 0) Then compute in turn D(1, 2), D(1, 3), . . , D(1, m) UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming Fill in subsequent values, row by row, from left to right. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming Alternatively, first compute D(1, 1) from D(0, 0), D(0, 1), and D(1, 0) Then compute in turn D(2, 1), D(3, 1), . . , D(n, 1) UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming Fill in subsequent values, column by column, from top to bottom. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming • Filling each cell entails a constant number of operations. – Cell (i, j) depends only on characters S 1(i) and S 2(j) and cells (i - 1, j - 1), (i, j - 1), and (i - 1, j). • There are O(nm) cells in the table • Consequently, we can compute the edit distance D(n, m) in O(nm) time by computing the table in O(nm). UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming • Having computed the table we know the value of the optimal edit transcript. • Q: How do we extract the optimal edit transcript from the table? • A: One way would be to establish pointers from each cell, to predecessor cell(s) from which its value was derived, i. e, – If D(i, j) = D(i - 1, j) + 1 add a pointer from (i, j) to (i - 1, j) – If D(i, j) = D(i, j - 1) + 1 add a pointer from (i, j) to (i, j - 1) – If D(i, j) = D(i - 1, j - 1) + t(i, j) add a pointer from (i, j) to (i - 1, j - 1) UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming • We can recover an optimal edit sequence simply by following any path from (n, m) to (0, 0) • The interpretation of the path links are: – A horizontal link , (i, j) (i, j-1), corresponds to an insertion of character S 2(j) into S 1. – A vertical link, (i, j) (i-1, j), corresponds to a deletion of S 1(i) from S 1. – A diagonal link, (i, j) (i-1, j-1), corresponds to a match S 1(i) = S 2(j) and a substitution if S 1(i) S 2(j) UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming An optimal edit path. What edit transcript does this path correspond to? S, S, S, M, D, M, M, I UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming Another optimal edit path. What edit transcript does this path correspond to? I, S, M, D, M, M, I UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming The third possible optimal edit path. What edit transcript does this path correspond to? S, I, M, D, M, M, I UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming • Alternatively we can interpret any path from (n, m) to (0, 0) as an alignment of S 1 and S 2. • The interpretation of the path links are: – A horizontal link , (i, j) (i, j-1), corresponds to an insertion of a space/dash into S 1. – A vertical link, (i, j) (i-1, j), corresponds to an insertion of a space/dash into S 2. – A diagonal link, (i, j) (i-1, j-1), corresponds to a match if S 1(i) = S 2(j) or a mismatch if S 1(i) S 2(j) UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming Possible optimal path. What alignment does this optimal path correspond to? w r i t - e r s v i n t n e r - UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming A second possible optimal path. What alignment does this optimal path correspond to? w r i - t - e r s v - i n t n e r - UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Dynamic Programming A third possible optimal path. What alignment does this optimal path correspond to? w r i - t - e r s - v i n t n e r - UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology

Summary • Any path from (n, m) to (0, 0) corresponds to an optimal edit sequence and an optimal alignment • We can recover all optimal edit sequences and alignments simply by extracting all paths from (n, m) to (0, 0) • The correspondence between paths and edit sequences is one-to-one. • The correspondence between paths and alignments is one-to-one. UNIVERSITY OF SOUTH CAROLINA College of Engineering & Information Technology