Multiple Sequence Alignment (MSA) 1. Uses of MSA 2. Technical difficulties 1. Select sequences 2. Select objective function 3. Optimize the objective function 1. Exact algorithms 2. Progressive algorithms 3. Iterative algorithms 1. Stochastic 2. Non-stochastic 4. Consistency-based algorithms 3. Tools to view alignments 1. MEGA 2. JALVIEW (PSI- BLAST) Function prediction Fig. from Boris Steipe U. of Toronto Sequence relationshi ps If the MSA is incorrect, the above inferences are
100
Embed
Multiple Sequence Alignment (MSA) 1.Uses of MSA 2.Technical difficulties 1.Select sequences 2.Select objective function 3.Optimize the objective function.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Multiple Sequence Alignment (MSA)
1. Uses of MSA
2. Technical difficulties
1. Select sequences
2. Select objective function
3. Optimize the objective function
1. Exact algorithms
2. Progressive algorithms
3. Iterative algorithms
1. Stochastic
2. Non-stochastic
4. Consistency-based algorithms
3. Tools to view alignments
1. MEGA
2. JALVIEW
(PSI-BLAST)
Function prediction
Fig. from Boris Steipe U. of Toronto Sequence
relationships
If the MSA is incorrect, the above inferences are incorrect!
• a collection of three or more protein (or nucleic acid) sequences that are partially or completely aligned
• homologous residues are aligned in columns across the length of the sequences
• residues are homologous in an evolutionary senseevolutionary sense
• residues are homologous in a structural sensestructural sense
Example: someone is interested in caveolin
Step 1: at NCBI change the pulldown menu to HomoloGeneand enter caveolin in the search box
Step 2: inspect the results. We’ll take the first set of caveolins. Change the Display to Multiple alignment.
Step 3: inspect the multiple alignment. Note that these eight proteins align nicely, although gaps must be included.
Here’s another multiple alignment, Rac:
This insertion could be due to alternative splicing
HomoloGene includes groups of eukaryotic proteins. The site includes links to the proteins, pairwise alignments, and more
Example: 5 alignments of 5 globins
Let’s look at a multiple sequence alignment (MSA) of five globins proteins.five globins proteins. We’ll use five prominent MSA programs: ClustalW, Praline, MUSCLE (used at HomoloGene), ProbCons, and TCoffee. Each program offers unique strengths.
We’ll focus on a histidine (H) residuehistidine (H) residue that has a critical role in binding oxygen in globins, and should be aligned. But often it’s not aligned, and all five programs give different answers.
Our conclusion will be that there is no single best approach to MSA. Dozens of new programs have been introduced in recent years.
ClustalW
Note how the region of a conserved histidine (▼) varies depending on which of five prominent algorithms is used
Praline
Page 194
MUSCLE
Page 194
Probcons
Page 195
TCoffee
Page 195
Multiple sequence alignment: properties
• not necessarily one “correct” alignment of a protein family
• protein sequences evolve...
• ...the corresponding three-dimensional structures of proteins also evolve
• may be impossible to identify amino acid residues that align properly (structurally) throughout a multiple sequence alignment
• for two proteins sharing 30% amino acid identity, about 50% of the individual amino acids are superposable in the two structures
Page 180
Multiple sequence alignment: features
• some aligned residues, such as cysteines that form disulfide bridges, may be highly conserved
• there may be conserved motifs such as a transmembrane domain
• there may be conserved secondary structure features
• there may be regions with consistent patterns of insertions or deletions (indels)
Page 181
Multiple sequence alignment: uses
• MSA is more sensitive than pairwise alignment to detect homologs
• BLAST output can take the form of a MSA, and can reveal conserved residues or motifs
• Population data can be analyzed in a MSA (PopSet)
• A single query can be searched against a database of MSAs (e.g. PFAM)
• Regulatory regions of genes may have consensus sequences identifiable by MSA
[5] Introduction to molecular evolution and phylogeny
Fig. from Boris Steipe Univ. of Toronto
MSA. Technical difficulties
MSA. Technical difficulties
Multiple Sequence Alignment (MSA)
1. Uses of MSA
2. Technical difficulties
1. Select sequences
2. Select objective function
3. Optimize the objective function
1. Exact algorithms
2. Progressive algorithms
3. Iterative algorithms
1. Stochastic
2. Non-stochastic
4. Consistency-based algorithms
3. Tools to view alignments
1. MEGA
2. JALVIEW
Multiple Sequence Alignment (MSA)
1. Uses of MSA
2. Technical difficulties
1. Select sequences
2. Select objective function
3. Optimize the objective function
1. Exact algorithms
2. Progressive algorithms
3. Iterative algorithms
1. Stochastic
2. Non-stochastic
4. Consistency-based algorithms
3. Tools to view alignments
1. MEGA
2. JALVIEW
Objective functions (OF) Define the mathematical objective of the search
A biologically ideal OF should
• Maximize similarity
• Minimize the number of gaps (over their length)
• Retain conserved motifs and patterns
• Retain functionally important alignments
• Recapitulate phylogeny
• Concentrate on alignable regions, not in gapped regions
• Consider the limitations imposed by the 3D structure
Most widely used MSA packages use a simple sum-of-pairs OF
• Define a mathematical optimum
• Use sum-of-pairs and affine gaps
• Use a context-independent Mutation Data Matrix (e.g. Blosum 62)
• Some add weighting proportional to the information in the seq.
It is a non-trivial task to test the biological correctness of an objective function.
Seq.1 AT-AATG
Seq.2 CTGAG-G
Seq.3 ATGAA-G
Sum-of-pairs (SP) Objective Function
Induced pairwise alignment: After the best MSA is obtained, other sequences are removed, spaces facing spaces are removed and a score is calculated using any chosen scoring scheme (distance or similarity).
Sum-of-pairs score: The SP of a MSA is the sum of the scores of all the scores of the induced pairwise global alignments
3
Weighted Sum-of-pairs score: each score can be multiplied by a weight. Weights are often intended to reflect evolutionary distances to induce the MSA to more accurately reflect known evolutionary history, or the information carried by the sequences being aligned.
Sum-of-pairs (SP) Objective Function
Multiple MSA: Depending on the Mutation Data matrix selected (e.g. PAM or BLOSUM) and on the selected gap penalties (opening and extension) different MSA will be obtained. Which one is the correct one?
New Objective functions: less sensitive to gap penalty estimations thanks to the
incorporation of local information
• Segment-to-segment comparisons of the sequences (instead of character-to-
character) without gap penalties is the strategy used by DiAlign. This approach is
efficient where sequences are not globally related but share only local similarities,
(genomic DNA, many protein families) http://bibiserv.techfak.uni-bielefeld.de/dialign/.
• Consistency objective function: (e.g. T-Coffee) The optimal MSA is defined as the
one that agrees the most with all the optimal pair-wise alignments. Given a set of
independent observations the most consistent are often closer to “the truth”.
Seq.1 AT-AATG Seq.1 ATAATG Clustal
Seq.2 CTGAG-G Distance scheme Seq.2 CTGAGG Gap open= 11
ClustalW Blosum62 Gap 11-1Cheaper to open terminal gap than to align C and F
Example of Progressive algorithm• Calculate distances/similarities
between sequences • Construct a tree• Add sequentially, following tree
Multiple sequence alignment: methods
Progressive methods: use a guide tree (related to aphylogenetic tree) to determine how to combine pairwise
alignments one by one to create a multiple alignment.
Examples: CLUSTALW, MUSCLE
Page 185
Multiple sequence alignment: methods
Example of MSA using ClustalW: two data sets
Five distantly related globins (human to plant)
Five closely related beta globins
Obtain your sequences in the FASTA format.
Page 185
Use ClustalW to do a progressive MSA
http://www.ebi.ac.uk/clustalw/ Page 186
Feng-Doolittle MSA occurs in 3 stages
[1] Do a set of global pairwise alignments (Needleman and Wunsch’s dynamic programming
algorithm)
[2] Create a guide tree
[3] Progressively align the sequences
Page 185
Page 186
Progressive MSA stage 1 of 3:generate global pairwise alignments
best score
Number of pairwise alignments needed
For n sequences, (n-1)(n) / 2
For 5 sequences, (4)(5) / 2 = 10
For 200 sequences, (199)(200) / 2 = 19,900
Page 185
Feng-Doolittle stage 2: guide tree
• Convert similarity scores to distance scores
• A tree shows the distance between objects
• Use UPGMA (defined in the phylogeny lecture)
• ClustalW provides a syntax to describe the tree
Page 187
Page 186
Progressive MSA stage 2 of 3:generate a guide tree calculated from the
distance matrix (5 distantly related globins)
Page 188
5 closely related globins
Feng-Doolittle stage 3: progressive alignment
• Make a MSA based on the order in the guide tree
• Start with the two most closely related sequences
• Then add the next closest sequence
• Continue until all sequences are added to the MSA
• Rule: “once a gap, always a gap.”
Page 188
Clustal W alignment of 5 distantly related globins
Fig. 6.3Page 187
Fig. 6.5Page 189
Clustal W alignment of 5 closely related globins
* asterisks indicate identity in a column
Why “once a gap, always a gap”?
• There are many possible ways to make a MSA
• Where gaps are added is a critical question
• Gaps are often added to the first two (closest) sequences
• To change the initial gap choices later on would beto give more weight to distantly related sequences
• To maintain the initial gap choices is to trustthat those gaps are most believable
Page 189
Additional features of ClustalW improveits ability to generate accurate MSAs
• Individual weights are assigned to sequences; very closely related sequences are given less weight,while distantly related sequences are given more weight
• Scoring matrices are varied dependent on the presenceof conserved or divergent sequences, e.g.:
PAM20 80-100% idPAM60 60-80% idPAM120 40-60% idPAM350 0-40% id
• Residue-specific gap penalties are appliedPage 190
See Thompson et al. (1994) for an explanation of the three stages of progressive alignment implemented in ClustalW
Pairwise alignment:Calculate distance matrix
Unrooted neighbor- joining tree
Unrooted neighbor- joining tree
Rooted neighbor-joining tree (guide tree) and sequence weights
Rooted neighbor-joining tree (guide tree) and sequence weights
Progressive alignment: Align following the guide tree
Iterative methods: compute a sub-optimal solution and keep modifying that intelligently using dynamic programming or other methods until the solution converges.
Examples: MUSCLE, IterAlign, Praline, MAFFT
Page 190
MUSCLE: next-generation progressive MSA
[1] Build a draft progressive alignment
Determine pairwise similarity through k-mer counting (not by alignment)
Compute distance (triangular distance) matrix
Construct tree using UPGMA ((Unweighted Pair Group Method with Arithmetic Mean – will be covered later)
Construct draft progressive alignment following tree
Page 191
MUSCLE: next-generation progressive MSA
[2] Improve the progressive alignment
Compute pairwise identity through current MSA
Construct new tree with Kimura distance measures
Compare new and old trees: if improved, repeat this step, if not improved, then we’re done
Page 191
MUSCLE: next-generation progressive MSA
[3] Refinement of the MSA
Split tree in half by deleting one edge
Make profiles of each half of the tree
Re-align the profiles
Accept/reject the new alignment
Page 191
Access to MUSLCE at EBIhttp://www.ebi.ac.uk/muscle/
Iterative approaches: MAFFT
• Uses Fast Fourier Transform to speed up profile alignment
• Uses fast two-stage method for building alignments using k-mer frequencies
• Offers many different scoring and aligning techniques• One of the more accurate programs available• Available as standalone or web interface• Many output formats, including interactive
Consistency-based algorithms: generally use a database of both local high-scoring alignments and long-range global alignments to create a final alignment
These are very powerful, very fast, and very accurate methods
Examples: T-COFFEE, Prrp, DiAlign, ProbCons
Page 192
Consistency-based Algorithms T-Coffee (Consistency Objective Function For alignmEnt Evaluation)
Version 2.00 and higher can mix sequences and structures
Local and global pair-wise alignments can come from different programs and can be redundant
The EL is a position-specific substitution matrix where the score associated with each pair of residues depends on its compatibility with the rest of the library. This library replaces the Mutation data Matrix used in ClustalW.
Pair-wise distances are computedA Neighbor joining tree is estimatedSequences are aligned progressively following the topology of the tree
ProbCons—consistency-based approach
Combines iterative and progressive approaches with a unique probabilistic model.
Uses Hidden Markov Models to calculate probability matrices for matching residues, uses this to construct a guide tree
Progressive alignment hierarchically along guide tree
Post-processing and iterative refinement (a little like MUSCLE)
Page 192
Fig. 5.12Page 158
ProbCons uses an HMM to make alignments
ProbCons—consistency-based approach
Sequence x xi
Sequence y yj
Sequence z zk
If xi aligns with zk
and zk aligns with yj
then xi should align with yj
ProbCons incorporates evidence from multiple sequences to guide the creation of a pairwise alignment.
Page 193
ProbCons output for the same alignment: consistency iteration helps
There are benchmarking multiple alignment datasets that have been aligned painstakingly by hand, by structural similarity, or by extremely time- and memory-intensive automated exact algorithms.
Some programs have interfaces that are more user-friendly than others. And most programs are excellent so it depends on your preference.
If your proteins have 3D structures, use these to help you judge your alignments. For example, try Expresso at http://www.tcoffee.org.
Page 196
[1] Create or obtain a database of protein sequencesfor which the 3D structure is known. Thus we candefine “true” homologs using structural criteria.
[2] Try making multiple sequence alignmentswith many different sets of proteins (very related,very distant, few gaps, many gaps, insertions, outliers).
[3] Compare the answers.
Strategy for assessment of alternativemultiple sequence alignment algorithms
Benchmarking tests suggest that ProbCons, a consistency-based/progressive algorithm, performs the best on the BAliBASE set, although MUSCLE, a progressive alignment package, is an extremely fast and accurate program.
ClustalW is the most popular program. It has a nice interface (especially with ClustalX) and is easy to use. But several programs perform better. There is no one single best program to use, and your answers will certainly differ (especially if you align divergent protein or DNA sequences)
► Hidden Markov models (HMMs) are “states” that describe the probability of having a particular amino acid residue at arranged in a column of a multiple sequence alignment
► HMMs are probabilistic models
► HMMs may give more sensitive alignments than traditional techniques such as progressive alignment
Page 197
Structure of a hidden Markov model (HMM)
main state
insert state
delete state
Fig. 5.12Page 158
Hidden Markov ModelsThe model accommodates the identities, mismatches, insertions, and deletions expected in a group of related proteins. (A) MSA: Each column may include matches and mismatches (red positions), insertions (green positions), and deletions (purple positions). (B) Each column in the model represents the possibility of a match, insert, or delete in each column of the alignment in A. The HMM is a probabilistic representation of the MSA. Sequences can be generated from the HMM by starting at the beginning state labeled BEG and then by following anyone of many pathways from one type of sequence variation to another (states) along the state transition arrows and terminating in the ending state labeled END. Any sequence can be generated by the model and each pathway has a probability associated with it. Each square match state stores an amino acid distribution such that the probability of finding an amino acid depends on the frequency of that amino acid within that match state. Each diamond-shaped insert state produces random amino acid letters for insertions between aligned columns and each circular delete state produces a deletion in the alignment with probability 1. One of many ways of generating the sequence N K Y L T in the above profile is by the sequence BEG ->Ml ->11 ->M2 ->M3 :>M4 ->END. Each transition has an associated probability, and the sum of the probabilities of transitions leaving each state is 1. The average value of a transition would thus be 0.33, since there are three transitions from most states (there are only two from M4 and D4, hence the average from them is 0.5). For example, if a match state contains a uniform distribution across the 20 amino acids, the probability of any amino acid is 0.05. Using these average values of 0.33 or 0.5 for the transition values and 0.05 for the probability of each amino acid in each state, the probability of the above sequence N K Y L T is the product of all of the transition probabilities in the path and the probability that each state will produce the corresponding amino acid in the sequences, or 0.33 X 0.05 X 0.33 X 0.05 X 0.33 X 0.05 X 0.33 X 0.05 X 0.33 X 0.05 X 0.5 = 6.1 X 10 -10. Since these probabilities are very small numbers, probabilities are converted to log odds scores, and the logarithms are added to give the overall probability score. The secret of the HMM is to adjust the transition values and the distributions in each state by training the model with the sequences. The training involves finding every possible pathway through the model that can produce the sequences, counting the number of times each transition is used and which amino acids were required by each match and insert state to produce the sequences. This training procedure leaves a memory of the sequences in the model. As a consequence, the model will be able to give a better prediction of the sequences. Once the model has been adequately trained, of all the possible paths through the model that can generate the sequence N KY L T, the most probable should be the match-insert-3 match combination (as opposed to any other combination of matches, inserts, and deletions). Likewise, the other sequences in the alignment would also be predicted with highest probability as they appear in the alignment; i.e., the last sequence would be predicted with highest probability by the path match-match-delete-match. In this fashion, the trained HMM provides a multiple sequence alignment, such as shown in A. For each sequence, the objective is to infer the sequence of states in the model that generate the sequences. The generated sequence is a Markov chain because the next state is dependent on the current one. Because the actual sequence information is hidden with in the model, the model is described as a hidden Markov model
PFAM (protein family) database is a leading resource for the analysis of protein families
http://pfam.sanger.ac.uk/
Page 198
PFAM HMM for lipocalins:resembles a position-specific scoring matrix
20 amino acids
position
PFAM HMM for lipocalins: GXW motif
G W
20 amino acids
PFAM GCG MSF format
Pfam (protein family) database
PFAM JalView viewer
Alignment Editors
Jalview
• Written in Java
• Input MSF, aligned FASTA
• ClustalW alignment
• Interactive alignment editor
• Multiple color schemes
• Can divide in sub-families
• Produces UPGMA, Neighbor-joining trees and Principal Component Analysis
• Incorporates information from feature Table
• Incorporates structural information
SMART: Simple ModularArchitecture Research Tool(emphasis on cell signaling)
Page 199
[1] Go to NCBI Domains & Structure (left sidebar)[2] Click CDD[3] Enter a text query, or a protein sequence
There are typically few sequences (up to several dozen), each having up to millions of base pairs. Adding more species improves accuracy.
Alignment of divergent sequences often reveals islands of conservation (providing “anchors” for alignment).
Chromosomes are subject to inversions, duplications, deletions, and translocations (often involving millions of base pairs). E.g. human chromosome 2 is derived from the fusion of two acrocentric chromosomes.
There are no benchmark datasets available.
Page 203
Page 205
Multiple alignment of genomic DNA at UCSC50,000 base pairs (at http://genome.ucsc.edu)
Page 205
Note conserved regions: exons and regulatory sites(scale: 50,000 base pairs)
regulatory
Page 205
Multiple alignment of beta globin genescale: 1,800 base pairs
Page 205
Multiple alignment of beta globin genescale: 55 base pairs