Approximate matching

Ben Langmead

Department of Computer Science

You are free to use these slides. If you do, please sign the
guestbook (www.langmead-lab.org/teaching-materials), or email
me ([email protected]) and tell me briefly how youre
using them. For original Keynote files, email me.

Read alignment requires approximate matching

Read
CTCAAACTCCTGACCTTTGGTGATCCACCCGCCTNGGCCTTC

Reference
GATCACAGGTCTATCACCCTATTAACCACTCACGGGAGCTCTCCATGCATTTGGTATTTT
CGTCTGGGGGGTATGCACGCGATAGCATTGCGAGACGCTGGAGCCGGAGCACCCTATGTC
GCAGTATCTGTCTTTGATTCCTGCCTCATCCTATTATTTATCGCACCTACGTTCAATATT
ACAGGCGAACATACTTACTAAAGTGTGTTAATTAATTAATGCTTGTAGGACATAATAATA
ACAATTGAATGTCTGCACAGCCACTTTCCACACAGACATCATAACAAAAAATTTCCACCA
AACCCCCCCTCCCCCGCTTCTGGCCACAGCACTTAAACACATCTCTGCCAAACCCCAAAA
ACAAAGAACCCTAACACCAGCCTAACCAGATTTCAAATTTTATCTTTTGGCGGTATGCAC
TTTTAACAGTCACCCCCCAACTAACACATTATTTTCCCCTCCCACTCCCATACTACTAAT
CTCATCAATACAACCCCCGCCCATCCTACCCAGCACACACACACCGCTGCTAACCCCATA
CCCCGAACCAACCAAACCCCAAAGACACCCCCCACAGTTTATGTAGCTTACCTCCTCAAA
GCAATACACTGACCCGCTCAAACTCCTGGATTTTGGATCCACCCAGCGCCTTGGCCTAAA
CTAGCCTTTCTATTAGCTCTTAGTAAGATTACACATGCAAGCATCCCCGTTCCAGTGAGT
TCACCCTCTAAATCACCACGATCAAAAGGAACAAGCATCAAGCACGCAGCAATGCAGCTC
AAAACGCTTAGCCTAGCCACACCCCCACGGGAAACAGCAGTGATTAACCTTTAGCAATAA
ACGAAAGTTTAACTAAGCTATACTAACCCCAGGGTTGGTCAATTTCGTGCCAGCCACCGC
GGTCACACGATTAACCCAAGTCAATAGAAGCCGGCGTAAAGAGTGTTTTAGATCACCCCC
TCCCCAATAAAGCTAAAACTCACCTGAGTTGTAAAAAACTCCAGTTGACACAAAATAGAC
TACGAAAGTGGCTTTAACATATCTGAACACACAATAGCTAAGACCCAAACTGGGATTAGA
TACCCCACTATGCTTAGCCCTAAACCTCAACAGTTAAATCAACAAAACTGCTCGCCAGAA
CACTACGAGCCACAGCTTAAAACTCAAAGGACCTGGCGGTGCTTCATATCCCTCTAGAGG
AGCCTGTTCTGTAATCGATAAACCCCGATCAACCTCACCACCTCTTGCTCAGCCTATATA
CCGCCATCTTCAGCAAACCCTGATGAAGGCTACAAAGTAAGCGCAAGTACCCACGTAAAG
ACGTTAGGTCAAGGTGTAGCCCATGAGGTGGCAAGAAATGGGCTACATTTTCTACCCCAG
AAAACTACGATAGCCCTTATGAAACTTAAGGGTCGAAGGTGGATTTAGCAGTAAACTAAG
AGTAGAGTGCTTAGTTGAACAGGGCCCTGAAGCGCGTACACACCGCCCGTCACCCTCCTC
AAGTATACTTCAAAGGACATTTAACTAAAACCCCTACGCATTTATATAGAGGAGACAAGT
CGTAACCTCAAACTCCTGCCTTTGGTGATCCACCCGCCTTGGCCTACCTGCATAATGAAG
AAGCACCCAACTTACACTTAGGAGATTTCAACTTAACTTGACCGCTCTGAGCTAAACCTA
GCCCCAAACCCACTCCACCTTACTACCAGACAACCTTAGCCAAACCATTTACCCAAATAA
AGTATAGGCGATAGAAATTGAAACCTGGCGCAATAGATATAGTACCGCAAGGGAAAGATG
AAAAATTATAACCAAGCATAATATAGCAAGGACTAACCCCTATACCTTCTGCATAATGAA
TTAACTAGAAATAACTTTGCAAGGAGAGCCAAAGCTAAGACCCCCGAAACCAGACGAGCT

Sequence dierences occur

because of...
1. Sequencing error
2. Natural variation

Approximate string matching

Looking for places where a P matches T with up to a certain number of
mismatches or edits. Each such place is an approximate match.
A mismatch is a single-character substitution:
T: G G A A A A A G A G G T A G C G G C G T T T A A C A G T A G
||| |||||
P:
GTAACGGCG

An edit is a single-character substitution or gap (insertion or deletion):

T: G G A A A A A G A G G T A G C G G C G T T T A A C A G T A G
||| |||||
P:
GTAACGGCG
Gap in T

T: G G A A A A A G A G G T A G C - G C G T T T A A C A G T A G
||||| |||
P:
GTAGCGGCG
T: G G A A A A A G A G G T A G C G G C G T T T A A C A G T A G
|| ||||||
P:
GT-GCGGCG
Gap in P

Hamming and edit distance

For two same-length strings X and Y, hamming distance is the minimum
number of single-character substitutions needed to turn one into the other:
X: G A G G T A G C G G C G T T T A A C
| |||| ||| |||||||
Y: G T G G T A A C G G G G T T T A A C

Hamming distance = 3

Edit distance (Levenshtein distance): minimum number of edits required to

turn one into the other:
X: T G G C C G C G C A A A A A C A G C
|| |||||||||| ||||
Y: T G A C C G C G C A A A A C
-A
CG
AC
GC
-C
AG
CC
GC
X: G C G T A T G C G G C T A A
|| |||||||||| |||
Y: G C T
-A
TT
AG
TC
GG
CGC
GT
CA
TT
AA
TC
AG
CC
GC

Edit distance = 2

Edit distance = 2

Approximate string matching

Adapting the naive algorithm to do approximate string matching within
configurable Hamming distance:
def naiveApproximate(p, t, maxHammingDistance=1):
occurrences = []
for i in xrange(0, len(t) - len(p) + 1): # for all alignments
nmm = 0
for j in xrange(0, len(p)): # for all characters
if t[i+j] != p[j]: # does it match?
nmm += 1 # mismatch
if nmm > maxHammingDistance:
break # exceeded maximum distance
if nmm <= maxHammingDistance:
# approximate match; return pair where first element is the
# offset of the match and second is the Hamming distance
occurrences.append((i, nmm))
return occurrences

Instead of stopping upon first mismatch, stop when maximum

distance is exceeded

Python example: http://bit.ly/CG_NaiveApprox

Approximate string matching

How to make Boyer-Moore and index-assisted exact matching approximate?
Helpful fact: Split P into non-empty non-overlapping substrings u and v. If P
occurrs in T with 1 edit, either u or v must match exactly.

P
u

v
...or here. Cant go anywhere else!

Either the edit goes here...

More generally: Let p1, p2, ..., pk+1 be a partitioning of P into k+1 nonoverlapping non-empty substrings. If P occurrs in T with up to k edits, then
at least one of p1, p2, ..., pk+1 must match exactly.

P
p1

p2

p3

p4

...

pk+1

k edits can aect as many as k of these, but not all

Approximate string matching

These rules provides a bridge from the exact-matching methods weve
studied so far, and approximate string matching.
P
p1

p2

p3

...

p4

pk+1

k edits can overlap as many as k of these, but not all

Use an exact matching algorithm to find exact matches for p1, p2, ..., pk+1.
Look for a longer approximate match in the vicinity of the exact match.
check

p1

p2

check

p3

p4
Exact
match

p5

Approximate string matching

def bmApproximate(p, t, k, alph="ACGT"):
""" Use the pigeonhole principle together with Boyer-Moore to find
approximate matches with up to a specified number of mismatches. """
if len(p) < k+1:
raise RuntimeError("Pattern too short (%d) for given k (%d)" % (len(p), k))
ps = partition(p, k+1) # split p into list of k+1 non-empty, non-overlapping substrings
off = 0 # offset into p of current partition
occurrences = set() # note we might see the same occurrence >1 time
for pi in ps: # for each partition
bm_prep = BMPreprocessing(pi, alph=alph) # BM preprocess the partition
for hit in bm_prep.match(t)[0]:
if hit - off < 0: continue # pattern falls off left end of T?
if hit + len(p) - off > len(t): continue # falls off right end?
# Count mismatches to left and right of the matching partition
nmm = 0
for i in range(0, off) + range(off+len(pi), len(p)):
if t[hit-off+i] != p[i]:
nmm += 1
if nmm > k: break # exceeded maximum # mismatches
if nmm <= k:
occurrences.add(hit-off) # approximate match
off += len(pi) # Update offset of current partition
return sorted(list(occurrences))

Full example: http://bit.ly/CG_BoyerMooreApprox

Approximate Boyer-Moore performance

Boyer-Moore, 1 mismatch
with pigeonhole

Boyer-Moore, exact

Boyer-Moore, 2 mismatches
with pigeonhole

# character wall clock

# character wall clock
# character wall clock
#
matches
#
matches
# matches
comparisons
comparisons
comparisons
time
time
time

P: tomorrow
T: Shakespeares
complete works

786 K

1.91s

17

3.05 M

7.73 s

24

6.98 M

16.83 s

382

32.5 M

67.21 s

336

107 M

209 s

1,045

171 M

328 s

2,798

P: 50 nt string
from Alu repeat*
T: Human
reference (hg19)
chromosome 1

* GCGCGGTGGCTCACGCCTGTAATCCCAGCACTTTGGGAGGCCGAGGCGGG

Approximate string matching: more principles

Let p1, p2, ..., pk+1 be a partitioning of P into k+1 non-overlapping non-empty
substrings. If P occurrs in T with up to k edits, then at least one of p1, p2, ..., pk+1
must match exactly.

P
p1

p2

p3

p4

...

pk+1

New principle:
Let p1, p2, ..., pj be a partitioning of P into j non-overlapping non-empty
substrings. If P occurs with up to k edits, then at least one of p1, p2, ..., pj must
occur with floor(k / j) edits.

Review: approximate matching principles

Let j = k + 1

Non-overlapping substrings
General

Pigeonhole principle
p1, p2, ..., pj is a partitioning of P. If P occurs
with k edits, at least one partition
matches with floor(k / j) edits.

Specific

Pigeonhole principle with j = k + 1

p1, p2, ..., pk+1 is a partitioning of P. If P occurrs
in T with k edits, at least one partition
matches exactly.

Why?
Smallest value s.t. floor(k / j) = 0
Why make floor(k / j) = 0?
So we can use exact matching
Why is smaller j good?
Yields fewer, longer partitions
Why are long partitions good?
Makes exact-matching filter more
specific, minimizing # candidates

Approximate string matching: more principles

We partitioned P into non-overlapping substrings

p1

p2

p3

p4

...

pk+1

P
Consider overlapping substrings
P3

P6

P2
P1

Pn-l+1

...

P5
P4

Pn-l
Pn-l-1

Approximate string matching: more principles

q

x
x
x
x

...

n-q+1
of these

n
Say substrings are length q. There are n - q + 1 such substrings.
Worst case: 1 edit to P changes up to q substrings
Minimum # of length-q substrings unedited after k edits?
q-gram lemma: if P occurs in T with up to k edits, alignment
must contain t exact matches of length q, where t n - q + 1 - kq

n - q + 1 - kq

Approximate string matching: more principles

If P occurs in T with up to k edits, alignment contains an exact match of

length q, where q floor(n / (k + 1))
Derived by solving this for q:

n - q + 1 - kq 1

Exact matching filter: find matches of length floor(n / (k + 1)) between

T and any substring of P. Check vicinity for full match.

Approximate matching principles

Non-overlapping substrings

q-gram lemma

General

Pigeonhole principle
p1, p2, ..., pj is a partitioning of P. If P occurs
with k edits, at least one partition matches
with floor(k / j) edits.

Pigeonhole principle with j = k + 1

Specific

Overlapping substrings

p1, p2, ..., pk+1 is a partitioning of P. If P occurrs

in T with k edits, at least one partition
matches exactly.

If P occurs with k edits, alignment

contains t exact matches of length q,
where t n - q + 1 - kq

q-gram lemma with t = 1

If P occurs with k edits, alignment
contains an exact match of length q
where q floor(n / (k + 1))

...

...
P
P

Sensitivity
Sensitivity = fraction of true approximate matches discovered by the algorithm
Lossless algorithm finds all of them, lossy algorithm doesnt necessarily
Weve seen lossless algorithms. Most everyday tools are lossy. Lossy algorithms
are often much speedier & still acceptably sensitive (e.g. BLAST, BLAT, Bowtie).

...
P

Example lossy algorithm: pick q > floor(n / (k + 1))