Suffix Trees Come of Age in Bioinformatics

1
Suffix Trees Come of Age in Bioinformatics

• Algorithms, Applications and Implementations
• Dan Gusfield, U.C. Davis

2
AABAAB
1 2 3 4 5 6
S
Suffix Tree
B
A
A
A
B

B
A
3


B
A
5
6
A

4
B
A

A
2
B
Every suffix of S is encoded by a root to leaf
walk. Every substring encoded by a walk from the
root.

1
3
AABAAB
1 2 3 4 5 6
SUFFIX ARRAY
B
A
A
A
B

B
A
3


B
A
5
6
A

4
B
A
4 1 5 2 3 6

A
2
The suffixes listed in lexicographic order. Found
by lexicographic DFS
B

1
4
Basic Facts about Suffix Trees

• Suffix tree for a string of length n can be built
  in O(n) time.
• Basic implementations need about 25 bytes per
  character.
• Numerous applications in string algorithms
  achieving significant (sometimes amazing)
  speedups over naïve algorithms O(m) time
  substring search for a string of length m in a
  string of length n, regardless of n.

5
Suffix Tree/Array Construction

• Weiner 1973
• McCreight 1974
• Ukkonen 1996
• Farach-Colton 1999
• Manber-Myers 1991 (Suffix-Arrays)
• O(n log n time) with small space. O(n) time
  possible as on prior slide.

6
aabbaa
B
A
B
A
A


B
B
6
A
3
A
B

A
A
5
B

A

B
4
A
2
A

1
7
Bioinformatics Applications pre 1997

• A small number of applications
• Limited by space requirements of the tree, small
  computer memories, complication of the
  algorithms, poor locality of reference

8
Post 1997

• Much has changed, and Suffix trees and their
  relatives (particularly suffix arrays) are more
  widely applied in bioinformatics
• 20- 40 new publications with substantial
  connection to bioinformatics, post 1997

9
New Results since 1997

• Fundamental Algorithms Farach-Colton
  construction simpler algorithm for
  least-common-ancestor.
• Implementation improvements (space) for suffix
  trees and arrays
• New variants of suffix trees affix trees,
  virtual suffix trees
• New algorithms for tandem repeats

10
New Results since 1997

• Approximate repeats and large-scale repeat
  structure
• Fast lookups in databases
• Substring frequencies
• Motif and pattern discovery
• Hybrid dynamic programming
• Oligo and probe construction

11
New Results since 1997

• Genome fragment assembly and resequencing
• (multiple) whole genome comparison
• Large-scale sequence comparison in resequencing
  projects
• Misc. clever applications
• Related less-used data structures

12
Suffix trees collect together repeated substring
prefixes
Example Speeding up the use of Position Specific
Scoring Matrices. Accelerating protein
classification ISMB 2000 position 1 2 3 4 5
A 3 1 4 2 1 T 2 4 4 3 2
C 1 3 5 6 2 G 4 2 4 6 2
AACTGAACTG.AACTG
PSSM
AACTG
Walk around the tree to depth 5
Starting locations of AACTG
13
Whole Genome Alignment with MUMs

• Delcher Salzberg NAR 1999
• A MUM is a Maximal Unique Matching substring in
  two strings S and S
• MUMMER finds all MUMs selects and aligns
  non-overlapping pairs of MUMs and then recurses
  in the regions between adjacent selected MUMs.
• Key issue finding MUMs efficiently

14
Spotting MUMs with a suffix tree
7
Example S AATCCGTG. S
GATCCGTA
20 So ATCCGT is a MUM if it only occurs in
these two places
root
Path spelling ATCCGT
Exactly two sibling leaves
S7
S20
15
Finding MUMs in a suffix tree

• Build a suffix tree for the concatenation of S
  and S, noting at each leaf whether the suffix is
  from the S part or the S part. Also note the
  prior character in S or S for that position.
• Do a DFS traversal of the S.T. to note at each
  node how many leaves below it are from S and how
  many from S.
• A node corresponds to a MUM if and only if the
  count there is 1 and 1, and the two prior
  characters are different.
• For total string length n, the time used is O(n).

16
Reducing the space needed for MUM findingMUMMER
II - NAR spring 2002
root
Path to a node v, spelling xB, where B is a string
Path to a node v, spelling B
v
v
Suffix link
17
Space reduction when finding MUMs

• Given strings S and S, build a suffix tree T,
  with suffix links , but for only the smaller of S
  and S, say S.
• Using T, find for each position i in S, the
  longest substring starting at i which occurs in
  S exactly once. Mark the end location in T of
  that match. Each such match is a candidate MUM.

18
Finding MUMs in less space

• To find the matches, walk S through T, using
  suffix links to reduce search time.
• After S has been completely processed, every
  marked position in T that is visited only once
  and has no marked position below it, corresponds
  to a MUM.
• So all MUMs can be found in O(n) time, but in
  much reduced space.

19
Finding Tandem Repeats
Stoye and Gusfield, Theoretical Computer Science,
2001
TATAACTAACTAAGATT..
For a string of length n, a naïve algorithm might
use (on the order of) n3 operations to find all
T.R.s
20
Example Finding Tandem Repeats in Strings
TATAACTAACTAAGATT.
Every Tandem Repeat is part of a chain of tandem
repeats ending with a Branching Tandem
Repeat. To extend a chain, examine the first
character of the T.R. and the character after the
end of the T.R. If they are the same, a new T.R.
exists on place to the right.
21
Example Finding Tandem Repeats in Strings
TATAACTAACTAAGATT.
Branching Tandem Repeats
Every Tandem Repeat is part of a chain of tandem
repeats ending with a Branching Tandem Repeat.
22
Example Finding Tandem Repeats in Strings
TATAACTAACTAAGATT.
Branching Tandem Repeats
Every Tandem Repeat is part of a chain of tandem
repeats ending with a Branching Tandem Repeat.
23
Example Finding Tandem Repeats in Strings
TATAACTAACTAAGATT.
Branching Tandem Repeat
Every Tandem Repeat is part of a chain of tandem
repeats ending with a Branching Tandem
Repeat. So to find all T.R.s find all
Branching T.Rs Use a suffix tree to efficiently
find Branching T.R.s
24
Finding Branching T.R.s

• Every Branching T.R. ends at a node of the suffix
  tree so examine each node.

root
String to here is length 8
leaves 5 , 13 and 21 are below v
v
5
13 5 8
21
25
Basic Algorithm

• Build the S.T. for string S process it to find
  depth of each node, and do a DFS numbering of the
  nodes so that ancestry of two nodes can be
  determined in constant time. (Classic property of
  DFS)
• Check each internal node to see if it is at the
  first or second part of a tandem repeat.

26
How to check

• From a node v of depth d(v), walk the subtree to
  collect the leaf numbers below v.
• If leaf k is collected, check if v is an ancestor
  of leaf kd(v), and the next characters k1 and
  kd(v) 1 differ. If so, then the string to v is
  the first part of a Branching T.R.
• Alternatively, focus on k-d(v) to see if the
  string to v is the second part of a Branching
  T.R.

27
AABAAB
1 2 3 4 5 6
S
Suffix Tree
B
A
A
A
B

B
A
3


B
A
5
6
A

4
B
A

A
2
B

1
28
Finding Tandem Repeats

• This approach gives O(n2) time to find all
  branching tandem repeats.The time can be reduced
  to O(n log n) with a classic trick.
• Related more complex result In O(n) time, one
  can mark in the suffix tree, the endpoints of all
  tandem repeat species. Gusfield and Stoye 2000
  UCD techreport
• Uses the fact (Frankel et al 1999) that there
  are only 2n tandem repeat species complex
  proof.
• Example aaaaaaaaaa has only five tandem repeat
  species, but 25 tandem repeat occurrences

29
Space issues again

• Use virtual suffix tree or extended suffix
  array to reduce space
• Def for positions i and j in string S, LCP(i,j)
  is the length of the longest matching substring
  starting at positions i and j in S.

30
AABAAB
1 2 3 4 5 6
SUFFIX ARRAY With LCP or string depth information
B
A
A
A
B

B
A
3


B
A
5
6
A

4
B
A

4 1 5 2 3 6 3 1 2 0 1
A
2
Node depths at time of descent after a backup
B

LCP of neighbors
1
31
Extended Suffix Array

• String S, the suffix array A for S, and the LCP
  array for A, completely determine the suffix tree
  T for S. This is the extended suffix array for
  S.
• The suffix array and the LCP array can be found
  in O(n log n) time and small space, without an
  explicit suffix tree.
• For many efficient algorithms using an explicit
  suffix tree, it is possible to simulate the
  algorithm using only the extended suffix array,
  greatly reducing space usage without increasing
  time.
• S. Kurtz WABI 2002

32
String Barcoding Oligo Construction

• Uncovering Optimal Virus Signatures

Sam Rash and Dan Gusfield RECOMB April 2002
33
Motivation

• Need for rapid virus detection
• Given
• unknown virus
• database known viruses
• Problem
• identify unknown virus quickly
• Ideal solution
• have sequence of
• viruses in database
• unknown virus
• Solution
• use BLAST (or any sequence similarity
  program/algorithm)

34
Motivation

• Real World
• only have sequence for pathogens in database
• not possible to quickly sequence an unknown virus
• can test for presence small (lt 50 bp) strings in
  unknown virus
• substring tests
• Another Idea
• String Barcoding
• use substring tests to uniquely identify each
  virus in the database
• acquire unique barcode for each virus in database

35
Implementation

• Basic Idea Formulate problem as an Integer
  Linear Program (ILP)
• Enumerate some useful set of substrings from S
• variable in ILP for each substring
• Constraint for each pair of strings in S
• means that at least one substring will be chosen
  to distinguish each pair
• Objective Function
• Minimize sum of variables in ILP

36
Problem Definition

• Formal Definition
• given
• set of strings S
• goal
• find set of strings S, the testing set
• wlog, for each s1,s2 in S, there exists at least
  one u in S where u is a substring of only s1
• u is a signature substring
• minimize S
• result
• barcode for each element on S

37
Idea

• strings
• 1. cagtgc
• 2. cagttc
• 3. catgga
• Each node in the suffix tree has a corresponding
  set of string IDs below it

Figure 1.1 - suffix tree for set of strings
cagtgc, cagttc, and catgga
38
Idea

• strings
• 1. cagtgc
• 2. cagttc
• 3. catgga
• Each node in the suffix tree has a corresponding
  set of string IDs below it

Figure 1.1 - suffix tree for set of strings
cagtgc, cagttc, and catgga
39
Implementation Suffix Trees

• root-edge walk
• Creates string
• appears in exactly the strings that label the
  node at which it ends
• 2 root-edge walks ending on the same edge
• Both strings created by the walk
• occur in exactly the same set of original
  strings
• Can use ether string

example - a root edge walk
40
Practical Implementation

• If two substrings occur in exactly the same set
  of original strings, only one need be considered
• Use strings from suffix tree for each uniquely
  labeled node
• Build ILP as discussed
• Solve ILP using CPLEX
• Acquire barcode and signatures for each original
  string
• signature is the set of substring tests occurring
  in a string

41
Implementation Example
minimize V18 V22 V11 V17 V8 objective
function st V18 V22 V11 V17 V8 gt 2 this
is the theoretical minimum V18 V17 V8 gt
1 constraint to cover pair 1,2 V22 V11 V8
gt 1 constraint to cover pair 1,3 V18 V22
V11 V17 gt 1 constraint to cover pair
2,3 binaries all variables are 0/1 V18 V22
V11 V17 V8 end
Figure 1.4 - barcodes
Figure 1.5 - signatures
42
Implementation Extensions

• minimum and maximum lengths on signature
  substrings
• acquire barcodes/signatures for only a subset of
  input strings (wrt to whole set)
• minimum string edit distance between chosen
  signature substrings
• redundancy
• require r signature substrings to differentiate
  each pair
• adds a higher level of confidence that signatures
  remain valid even with mutations

43
Results

• Works quickly on most moderately sized datasets
  (especially when redundancy gt 2)
• dataset properties
• 50k virus genomes taken from NCBI (Genbank)
• 50-150 virus genomes
• average length of each sequence 1000 characters
• total input size ranged from approximately 50,000
  150,000 characters
• increasing dataset size scaled approximately
  linearly
• reach 25 gap (at most 1/3 more than optimum) in
  just a few minutes
• reach small gap (often lt 1) in 4 hours

44
Summary

• Numerous applications of Suffix Trees and
  relatives in Bioinformatics
• More yet to be found
• Suffix tree and array software at my UCD website.