#5 - Dynamic Programming 8/29/07 Required Reading BCB 444/544 (before lecture) Mon Aug 27 - for Lecture #4 Pairwise Sequence Alignment • Chp 3 - pp 31-41 Lecture 5 Wed Aug 29 - for Lecture #5 Dynamic Programming • Eddy: What is Dynamic Programming? 2004 Nature Biotechnol 22:909 Dynamic Programming Thurs Aug 30 - Lab #2: Databases, ISU Resources & Pairwise Sequence Alignment #5_Aug29 Fri Aug 31 - for Lecture #6 Scoring Matrices and Alignment Statistics • Chp 3 - pp 41-49 BCB 444/544 F07 ISU Review: Dobbs #5 - Dynamic Programming 8/29/07 1 8/29/07 2 3 Major types of electronic databases: Introduction to Biological Databases 1. Flat files - simple text files • no organization to facilitate retrieval What is a Database? Types of Databases Biological Databases Pitfalls of Biological Databases Information Retrieval from Biological Databases BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 2. Relational - data organized as tables ("relations") • shared features among tables allows rapid search 3. Object-oriented - data organized as "objects" • objects associated hierarchically 3 Examples of Biological Databases BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 4 Examples of Biological Databases 1- Primary 2- Secondary • DNA sequences • Protein sequences • GenBank - USA • Swiss-Prot, TreEMBL, PIR • European Molecular Biology Lab - EMBL • these recently combined into UniProt • DNA Data Bank of Japan - DDBJ 3- Specialized • Structures (Protein, DNA, RNA) • Species-specific (or "taxonomic" specific) • PDB - Protein Data Bank • Dobbs #5 - Dynamic Programming Types of Databases Chp 2- Biological Databases • Xiong: Chp 2 • • • • • BCB 444/544 F07 ISU • Flybase, WormBase, AceDB, PlantDB NDB - Nucleic Acid Data Bank • Molecule-specific, disease-specific See: http://www.oxfordjournals.org/nar/database/c/ BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming BCB 444/544 Fall 07 Dobbs 8/29/07 5 BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 6 1 #5 - Dynamic Programming 8/29/07 SUMMARY: #2- Biological Databases Chp 3- Sequence Alignment SECTION II BEWARE! Xiong: Chp 3 Pairwise Sequence Alignment • • • • • • Who was that Icelandic fellow? BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 7 Evolutionary Basis Sequence Homology versus Sequence Similarity Sequence Similarity versus Sequence Identity Methods Scoring Matrices Statistical Significance of Sequence Alignment Adapted from Brown and Caragea, 2007, with some slides from: Altman, Fernandez-Baca, Batzoglou, Craven, Hunter, Page. BCB 444/544 F07 ISU "Sequence comparison lies at the heart of bioinformatics Homology = similarity due to descent from a common evolutionary ancestor Pairwise sequence alignment is fundamental; it used to: But, • Search for common patterns of characters • Establish pair-wise correspondence between related sequences • Database searching (e.g., BLAST) • Multiple sequence alignment (MSA) We can infer homology from similarity (can't prove it!) Dobbs #5 - Dynamic Programming 8/29/07 9 aligned residues with similar physicochemical properties (e.g., size, hydrophobicity, charge) IMPORTANT: • Sequence homology: • result of gene duplication events • proteins may (or may not) have similar functions (e.g., human α-globin & human β-globin) • An inference about a common ancestral relationship, drawn when two sequences share a high enough degree of sequence similarity • Homology is qualitative A is the parent gene Speciation leads to B & C Duplication leads to C’ BCB 444/544 F07 ISU • Sequence similarity: • The direct result of observation from a sequence alignment • Similarity is quantitative; can be described using percentages B and C are Orthologous C and C’ are Paralogous Dobbs #5 - Dynamic Programming BCB 444/544 Fall 07 Dobbs 10 evolutionary ancestry • Paralogs - "similar genes" within a species; C' 8/29/07 • Similar sequences - sequences that have a high percentage of • result of common ancestry • corresponding proteins have "same" functions (e.g., human α-globin & mouse α-globin) C Dobbs #5 - Dynamic Programming • Homologous sequences - sequences that share a common 2 types of homologous sequences: • Orthologs - "same genes" in different species; B BCB 444/544 F07 ISU Sequence Homology vs Similarity Orthologs vs Paralogs Duplication HOMOLOGY ≠ SIMILARITY When 2 sequences share a sufficiently high degree of sequence similarity (or identity), we may infer that they are homologous Pairwise sequence alignment is basis for: Speciation 8 For us: Sequence comparison is important for drawing functional & evolutionary inferences re: new genes/proteins A 8/29/07 Homology has a very specific meaning in evolutionary & computational biology - & term is often used incorrectly Jin Xiong BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming Homology Motivation for Sequence Alignment analysis." SEQUENCE ALIGNMENT 8/29/07 11 BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 12 2 #5 - Dynamic Programming 8/29/07 Sequence Similarity vs Identity What is Sequence Alignment? For nucleotide sequences (DNA & RNA), sequence similarity and identity have the "same" meaning: • Two DNA sequences can share a high degree of sequence identity (or similarity) -- means the same thing • Drena's opinion: Always use "identity" when making quantitative comparisons re: DNA or RNA sequences (to avoid confusion!) For protein sequences, sequence similarity and identity have different meanings: Given 2 sequences of letters, and a scoring scheme for evaluating matching letters, find an optimal pairing of letters in one sequence to letters of other sequence. Align: 1: THIS IS A RATHER LONGER SENTENCE THAN THE NEXT. 2: THIS IS A SHORT SENTENCE. 1: THIS IS A RATHER LONGER SENTENCE THAN THE NEXT. 2: THIS IS A ######SHORT###SENTENCE##############. • Identity = % of exact matches between two aligned sequences • Similarity = % of aligned residues that share similar characteristics (e.g, physicochemical characteristics, OR 1: THIS IS A RATHER LONGER SENTENCE THAN THE NEXT. 2: THIS IS A ##SHORT###SENT#EN###CE##############. structural propsensities, evolutionary profiles) Is one of these alignments "optimal"? Which is better? • Drena's opinion: Always use "identity" when making quantitative comparisons re: protein sequences (to avoid confusion!) BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 13 BCB 444/544 F07 ISU Goal of Sequence Alignment • 2 sequences • Scoring system for evaluating match (or mismatch) of two characters • Penalty function for gaps in sequences 4 letter alphabet (+ gap) Find: Optimal pairing of sequences that: • Retains the order of characters • Introduces gaps where needed • Maximizes total score 20 letter alphabet (+ gap) RKVA-GMA RKIAVAMA BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 15 Types of Sequence Variation BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 16 Gaps Indels of various sizes can occur in one sequence relative to the other e.g., corresponding to a shortening of the polypeptide chain in a protein • Sequences can diverge from a common ancestor through various types of mutations: • Substitutions • Insertions • Deletions 14 Given: TTGACAC TTTACAC • Proteins 8/29/07 Statement of Problem Find the best pairing of 2 sequences, such that there is maximum correspondence between residues • DNA Dobbs #5 - Dynamic Programming ACGA → AGGA ACGA → ACCGA ACGA → AGA • Insertions or deletions ("indels") result in gaps in alignments • Substitutions result in mismatches • No change? match BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming BCB 444/544 Fall 07 Dobbs 8/29/07 17 BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 18 3 #5 - Dynamic Programming 8/29/07 Avoiding Random Alignments with a Scoring Function Not All Mismatches are the Same • Introducing too many gaps generates nonsense alignments: s--e-----qu---en--ce sometimesquipsentice e.g., Ser & Thr are more similar than Trp & Ala • Need to distinguish between alignments that occur due to homology and those that occur by chance • Define a scoring function that accounts for mismatches and gaps Scoring Function (F): Match: Mismatch: Gap: • Some amino acids are more "exchangeable" than others (physicochemical properties are similar) + w - x - y • Substitution matrix can be used to introduce "mismatch costs" for handling different types of substitutions e.g. +1 0 -1 • Mismatch costs are not usually used in aligning DNA or RNA sequences, because no substitution is "better" than any other (in general) F = w(#matches) + x(#mismatches) + y(#gaps) BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 19 BCB 444/544 F07 ISU Substitution Matrix Dobbs #5 - Dynamic Programming 8/29/07 20 8/29/07 22 8/29/07 24 Methods • • • • s(a,b) corresponds to score of aligning character a with character b Match scores are often calculated based on frequency of mutations in very similar sequences (more details later) Global and Local Alignment Alignment Algorithms Dot Matrix Method Dynamic Programming Method • Gap penalities • DP for Global Alignment • DP for Local Alignment • Scoring Matrices • Amino acid scoring matrices • PAM • BLOSUM • Comparisons between PAM & BLOSUM • Statistical Significance of Sequence Alignment BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 21 Global vs Local Alignment BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming Global vs Local Alignment - example S = CTGTCGCTGCACG T = TGCCGTG Global alignment • Finds best possible alignment across entire length of 2 sequences Global alignment • Aligned sequences assumed to be generally similar over entire length Local alignment CTGTCGCTGCACG Local alignment CTGTCGCTGCACG -TGCCG-TG---- -TG-C-C-G--TG • Finds local regions with highest similarity between 2 sequences • Aligns these without regard for rest of sequence CTGTCGCTGCACG • Sequences are not assumed to be similar over entire length BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming BCB 444/544 Fall 07 Dobbs -TGCCG-T----G 8/29/07 23 Which is better? BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 4 #5 - Dynamic Programming 8/29/07 Global vs Local Alignment Which should be used when? Alignment Algorithms Both are important but it is critical to use right method for a given task! 3 major methods for pairwise sequence alignment: Global alignment: 1. Dot matrix analysis • Good for: aligning closely related sequences of similar length • Not good for: divergent sequences or sequences with different lengths 2. Dynamic programming 3. Word or k-tuple methods (later, in Chp 4) Local Alignment: • Good for: searching for conserved patterns (domains or motifs) in DNA or protein sequences • Not good for: generating an alignment of closely related sequences Global and local alignments are fundamentally similar; they differ only in optimization strategy used to align similar residues BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 25 • For proteins, usually use more sophisticated scoring schemes than "identical match" • Diagonal lines indicate areas of match 26 • Diagonal lines of dots indicate regions of similarity between 2 sequences • Reverse diagonals (perpendicular to diagonal) indicate inversions A C A C G • What do similar patterns mean when comparing a sequence with itself (reverse complement)? • e.g.: Reverse diagonals crossing diagonals (X's) indicate palindromes Exploring Dot Plots Dobbs #5 - Dynamic Programming 8/29/07 27 BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 28 Strengths & Weakneses of Dot Plots Dot Matrix Variations Compare 2 sequences Strengths: • Fast and easy • Allows direct visual identification of regions of similarity • Repeats, inversions, etc. are readily apparent • Displays all possible matches • Identify matching regions • Identities for DNA seqs • Similarities for protein seqs Compare sequence with itself • Identify repeated regions • Identify inverted repeats • Identify palindromes Weaknesses: • Doesn't generate full alignment - user must "connect the diagonals" • No statistical assessment of quality of alignment (score) • Impractical and noisy for long sequences • Difficult to scale up to muliple alignment For long sequences? • Too many dots! Noisy! • Instead of per "residue," plot one dot per "window" of n matching residues to reduce noise BCB 444/544 F07 ISU 8/29/07 When comparing 2 sequences: A C G C G • Contiguous diagonal lines reveal alignment; "breaks" = gaps (indels) BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming Interpretation of Dot Plots Dot Matrix Method (Dot Plots) • Place 1 sequence along top row of matrix • Place 2nd sequence along left column of matrix • Plot a dot each time there is a match between an element of row sequence and an element of column sequence BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming BCB 444/544 Fall 07 Dobbs 8/29/07 29 BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 30 5 #5 - Dynamic Programming 8/29/07 Dynamic Programming Global alignment: Scoring For Pairwise sequence alignment CTGTCG-CTGCACG Idea: Display one sequence above another with spaces inserted in both to reveal similarity -TGC-CG-TG---Reward for matches: α Mismatch penalty: β Space/gap penalty: γ A: C A T - T C A - C | | | | | B: C - T C G C A G C Score = αw – βx - γy w = #matches BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 31 C - T T G G T C C – G C – G 10 2 5 C – BCB 444/544 F07 ISU y = #spaces Dobbs #5 - Dynamic Programming 8/29/07 32 Optimum Alignment Global alignment: Scoring Reward for matches: Mismatch penalty: Space/gap penalty: x = #mismatches T T • Score of an alignment is a measure of its quality G G • Optimum alignment problem: Given a pair of sequences X and Y, find an alignment (global or local) with maximum score C - -5 10 10 -2 -5 -2 -5 -5 10 10 -5 Total = 11 BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 33 Alignment algorithms Dobbs #5 - Dynamic Programming 8/29/07 34 Dynamic Programming (DP) • As computer science concept - formalized in early 1950's by Bellman at RAND Corporation • Global: Needleman-Wunsch • Local: Smith-Waterman “ Frequently, however, there are only a polynomial number of subproblems… If we keep track of the solution to each subproblem solved, and simply look up the answer when needed, we obtain a polynomial-time algorithm. “ • Both NW and SW use dynamic programming • Variations: ----Aho, Hopcroft, Ullman • Gap penalty functions • Scoring matrices BCB 444/544 F07 ISU BCB 444/544 F07 ISU • Reported to biologists for sequence alignment problems by Needleman & Wunsch, 1969 Dobbs #5 - Dynamic Programming BCB 444/544 Fall 07 Dobbs 8/29/07 35 BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 36 6 #5 - Dynamic Programming 8/29/07 Key Idea Problem Formulation and Notations Given two sequences (strings) • X = x 1x 2 …xN of length N Score of the best possible alignment that ends at a given pair of positions (i,j) in two sequences is the score of the best alignment previous to those two positions PLUS the score for aligning those two positions • Y = y1y2 …yM of length M x = AGC N=3 y = AAAC M=4 Construct a matrix with (N+1) x (M+1) elements, where S ( i,j) = score of best alignment of x[1..i]=x1 x 2…xi with y[1..j]=y1y2…yj x1 Next best alignment = previous best + local best x2 x3 y1 S(2,3) = score of best alignment y2 of AG (x1x2) to AAA (y1y2y3) y3 y4 BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 37 BCB 444/544 F07 ISU Dynamic Programming 4 Components: Dobbs #5 - Dynamic Programming 8/29/07 38 Global Alignment: Algorithm 1. Recursive definition for optimal score x = Prefix of length i of x 1.. i 2. Matrix for storing optimal scores of subproblems y = Prefix of length j of y 3. Bottom-up approach for filling the matrix, by solving smallest subproblems first 1.. j S(i, j) = Score of optimal alignment of x and y ! 4. Traceback of path through matrix to recover the optimal alignment(s) that gave the optimal score 1..i 1..j ! BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 39 BCB 444/544 F07 ISU Calculating Score of Optimum Alignment 0 0 1 Initial conditions: S(0, j) = " j # $ Recursive definition: ! For 1 ≤ i ≤ n, 1 ≤ j ≤ m: Dobbs #5 - Dynamic Programming 1 N S(i-1,j-1) S(i-1,j) S(i,j-1) S(i,j) S(N,M) Recursion % S(i "1, j "1) + # (x i , y j ) ' S(i, j) = max& S(i "1, j) + $ ' S(i, j "1) + $ ( 8/29/07 40 S(0,0)=0 M % S(i "1, j "1) + # (Si ,T j ) ' S(i, j) = max& S(i "1, j) " $ ' S(i, j "1) " $ ( BCB 444/544 F07 ISU 8/29/07 Computing the best current score S(i,j) satisfies the following relationships: S(i,0) = "i # $ Dobbs #5 - Dynamic Programming 41 BCB 444/544 F07 ISU Initialization S(i,0) = "i # $ S(0, j) = " j # $ Dobbs #5 - Dynamic Programming 8/29/07 42 ! ! BCB 444/544 Fall 07 Dobbs ! 7 #5 - Dynamic Programming 8/29/07 What happens at the last step in the alignment of x[1..i] to y[1..j]? DP Implementationn - 3 steps: 1. Construct sequence vs sequence matrix and fill in from [0,0] to [N,M], the best possible scores for alignments including the residues at [i,j]. Also, keep track of dependencies of scores (in a pointer matrix). 1 of 3 cases: xi aligns to a gap yj aligns to a gap x1 x2 . . . xi-1 xi x1 x2 . . . xi-1 xi x1 x2 . . . xi y1 y2 . . . yj-1 yj y1 y2 . . . yj xi aligns to yj — S(i-1,j-1) + σ(xi,yj) S(i-1,j) BCB 444/544 F07 ISU 2. For a global alignment of the sequences, find the score S(N,M) — y1 y2 . . . yj-1 yj +γ S(i,j-1) Dobbs #5 - Dynamic Programming 3. Trace back through pointer matrix to get the optimal alignment. Do this position by position to retrieve alignment of all residues of sequences, including gaps (i.e., repeat alignment calculations in reverse order, following path back through matrix, starting at from position with highest score. +γ 8/29/07 43 BCB 444/544 F07 ISU Example λ Case 1: Line up x i with y j x: C y: C A - T T i-1 C A C A j -1 T T Case 2: Line up x i with space x: C y: C A - T T T T C C A A i λ C G j -5 C A -10 i-1 G i C - j Case 3: Line up y j with space x: C y: C A - T T T T i A C A j -1 C C BCB 444/544 F07 ISU 0 G j T -15 T -20 C A -25 C -35 C T C Dobbs #5 - Dynamic Programming G C A G 8/29/07 44 C -5 -10 -15 -20 -25 -30 -35 -40 10 5 -30 +10 for match, -2 for mismatch, -5 for space Dobbs #5 - Dynamic Programming 8/29/07 45 BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 46 Affine Gap Penalty Functions λ C T C G C A G C λ 0 -5 -1 0 -1 5 -2 0 -2 5 -3 0 -3 5 -4 0 C A -5 10 5 0 -5 -1 0 -1 5 -2 0 -2 5 -1 0 5 8 3 -2 -7 0 -5 -1 0 T -1 5 0 15 10 5 0 -5 -2 -7 T -2 0 -5 13 8 3 -2 -7 -4 C A -2 5 -1 0 5 20 15 18 13 8 3 -3 0 -1 5 0 15 18 13 28 23 18 C -3 5 -2 0 -5 10 13 28 23 26 33 Gap penalty = h + gk where * 10 * k = length of gap h = gap opening penalty g = gap continuation penalty Can also be solved in O(nm) time using dynamic programming Traceback can yield both optimal alignments BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming BCB 444/544 Fall 07 Dobbs 8/29/07 47 BCB 444/544 F07 ISU Dobbs #5 - Dynamic Programming 8/29/07 48 8