erice - structured pattern detection and exploitation deterministic algorithms

Erice - Structured Pattern Detection and Exploitation

Deterministic Algorithms

Outline

• Structured patterns - weird things happen - chance or necessity?

• Structured pattern detection - suffix trees, viruses and integer linear programming

• ??more on suffix arrays - computing LCP in linear time; finding close repeats

• detecting and exploiting patterns of recombination in binary (SNP) sequences.

• Adding in the complication of gene conversion.

Kmer frequency

• Kmers in a string S over all K

• Ex. S = abxabcxab

K = 2, the Kmers are: ab, bx, xa, ab, bc, cx, xa, ab

five distinct 2-mers, ab and ax repeat

K = 3: abx, bxa, xab, abc, bcx, xab

five distinct 3-mers, xab repeats

K = 4: abxa, bxab, xabc, abcx, bcxa, cxab

six distinct 4-mers, no repeats

K = 5,6,7,8,9 have 5,4,3,2,1 distinct Kmers, no repeats

Weird (non-obvious) Patterns?

• K* = Maximum K such that some Kmer repeats in S.

• K’ = K where the number of distinct Kmers is maximum.

• D = number of distinct Kmers for K = K’• Observations from Data: K’ = K* + 1, and D + K* = |S| Chance or necessity? - Deterministic Necessity!

What would a statistical approach conclude?

String Barcoding

Uncovering Optimal Virus Signatures

Sam Rash, Dan Gusfield

University of California, Davis.

Motivation

• Need for rapid virus detection– Given

• unknown virus

• database known viruses

– Problem• identify unknown virus quickly based on a small set

of substrings.

Motivation

– Real World• only have sequence for pathogens in database

– not possible to quickly sequence an unknown virus

• can test for presence small (<= 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

Problem Definition

• Formal Definition– given

• set of strings S

– goal• find set of strings S’, the testing set

– such that 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

Figure 1.5 - signatures

cagtgc {“tg”}

cagttc

catgga {“tg”, “atgga”}

Example

Problem Complexity

• Complexity– unknown if NP-hard when size of any u in S’ is

unbounded

– Max-Length String Barcoding• additional parameter k, a maximum length of any u in S’

• this variant is NP-Hard

• reduction from Minimum Testing Set (Garey, Johnson, 1979)

• means all real world uses have to deal with NP-hard result

Implementation

• Basic Idea: Formulate problem as an 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

Implementation

– Key point: complexity of ILP primarily a function of the number of variables

– reducing number of candidate substring tests reduces the number of variables in ILP

– how to reduce?• Key to our method: suffix trees• finds minimum cardinality set of “useful” substrings

for use as candidate signature substrings

Implementation: Suffix Trees

• Key Properties of Suffix Tree build for set of strings S– tree with character sequences labeling edges– nodes labeled with a subset of original string

IDs– every substring of original input set appears as

a root-edge walk exactly once• root-node walk is considered root-edge walk into

node’s in-edge from parent

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 walkc

a

g

t

t

a

c g

a

g

t t

g t t c

c g a

Implementation: Solving

• 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

Implementation: Example

• strings:1. cagtgc

2. cagttc

3. catgga

• Each node in the suffix tree has a corresponding set of string IDs below it v1 - {1,2,3} v2 - {1,2,3} v3 - {3} v4 - {1} v5 - {3}

v6 - {1,2} v7 - {2} v8 - {1} v9 - {1,2,3} v10 - {1,2,3}

v11 - {1,2} v12 - {1} v13 - {2} v14 - {3} v15 - {1,2,3}

v16 - {2} v17 - {2} v18 - {1,3} v19 - {1} v20 - {3}

v21 - {1,2,3} v22 - {3} v23 - {2} v24 - {1,2} v25 - {1}

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

Figure 1.2 - table of string labels for each node in suffix tree from figure 1.1

Implementation: ExampleminimizeV18 + V22 + V11 + V17 + V8 #objective functionstV18 + V22 + V11 + V17 + V8 >= 2 #this is the theoretical minimumV18 + V17 + V8 >= 1 #constraint to cover pair 1,2V22 + V11 + V8 >= 1 #constraint to cover pair 1,3V18 + V22 + V11 + V17 >= 1 #constraint to cover pair 2,3binaries #all variables are 0/1V18 V22 V11 V17 V8end

Figure 1.3 - ILP constructed for suffix tree in figure 1.1 using no additional constraints (length, etc)

Figure 1.4 - barcodes

Figure 1.5 - signatures

tg (V18) atgga (V22)

cagtgc 1 0

cagttc 0 0

catgga 1 1

cagtgc {“tg”}

cagttc

catgga {“tg”, “atgga”}

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

Results: Summary

• Works quickly on most moderately sized datasets (especially when redundancy >= 2)– dataset properties

• ~50k virus genomes taken from NCBI (Genbank)• 50-150 virus genomes• average length of each genome ~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 < 1%) in 4 hours

Results: Summary

– increasing redundancy greatly decreases run time and % gap at 4 hours in all cases tested

Figure 2.1 - effect of redundancy on avg 25% gap Figure 2.2 - effect of redundancy on avg gap at 4 hours

Conclusion

• Practical sized testing sets obtained on reasonable sized input datasets– testing set consisting of 50 – 270 substring tests on input sets of

~100 genomes– works well with reactions that have high number of assays

(substring tests) per reaction• GeneChip – 400 assays per reaction

• Redundancy– Good concept in theory– Reduces solution space and hence computation time– GeneChip makes higher number of assays needed cost-effective

Recognizing Patterns of Historical Recombination

Dan Gusfield

UC DavisDifferent parts of this work are joint with Satish Eddhu, CharlesLangley, Dean Hickerson, Yun Song, Yufeng Wu.

10100 01011

5

10101

The first 4 sites come from P (Prefix) and the sitesfrom 5 onward come from S (Suffix).

P S

Sequence Recombination

A recombination of P and S at recombination point 5.

Single crossover recombination

Network with Recombination Deriving a Set of Sequences

00000

1

2

4

3

510100

1000001011

00010

01010

12345given set101001000001011010100001010101

10101

5

P

S

Only one mutation perposition is allowed.

The biological ProblemGiven a set of binary sequences derived by one mutation perposition and possibly many recombinations, find the positionswhere clusters of historical recombinations likely occurred. These are called recombination hotspots.

Applications: 1) Insight into the mechanics of recombination Science article October 14, 2005, and Nature article thisweek: in humans and chimps most recombinations occur in hotspots, but in different places in humans compared to chimps.

2) Association mapping: A major strategy being developed forfinding genes that influence disease - the whole strategy relieson the historical effects of recombination.

Two Approaches

• Stochastic models of recombination and mutation - maximum likelihood - very intensive computations.

• Deterministic approaches based on minimizing the number of needed recombinations, or bounding that number.

Regions where the minimum number is large, or where close bounds on the minimum are large, indicate regions of hotspots. Science article.

Reconstructing the Evolution of Binary Bio-Sequences

• Perfect Phylogeny (tree) model• Phylogenetic Networks (DAG) with recombination• Phylogenetic Networks with disjoint cycles: Galled-

Trees• Phylogenetic Networks with unconstrained cycles:

Blobbed-Trees• Combinatorial Structure and Efficient Algorithms• Efficiently Computed Lower and Upper bounds on

the number of recombinations needed

The Perfect Phylogeny Model for binary sequences

00000

1

2

4

3

510100

1000001011

00010

01010

12345sitesAncestral sequence

Extant sequences at the leaves

Site mutations on edgesThe tree derives the set M:1010010000010110101000010

Classic NASC: Arrange the sequences in a matrix. Then (with no duplicate columns), the sequences can be generated on a unique perfect phylogeny if and only if no two columns (sites) contain all four pairs:

0,0 and 0,1 and 1,0 and 1,1

This is the 4-Gamete Test

When can a set of sequences be derived on a perfect phylogeny?

A richer model

00000

1

2

4

3

510100

1000001011

00010

01010

12345101001000001011010100001010101 added

pair 4, 5 fails the threegamete-test. The sites 4, 5``conflict”.

Real sequence histories often involve recombination.

Network with Recombination

00000

1

2

4

3

510100

1000001011

00010

01010

12345101001000001011010100001010101 new

10101

The previous tree with onerecombination event now derivesall the sequences.

5

P

S

Elements of a Phylogenetic Network (single crossover

recombination)• Directed acyclic graph. • Integers from 1 to m written on the edges. Each integer

written only once. These represent mutations.• A choice of ancestral sequence at the root.• Every non-root node is labeled by a sequence obtained

from its parent(s) and any edge label on the edge into it.• A node with two edges into it is a ``recombination node”,

with a recombination point r. One parent is P and one is S.• The network derives the sequences that label the leaves.

A Phylogenetic Network

00000

52

3

3

4S

p

PS

1

4

a:00010

b:10010

c:00100

10010

01100

d:10100

e:01100

00101

01101

f:01101

g:00101

00100

00010

Minimizing recombinations

• Any set M of sequences can be generated by a phylogenetic network with enough recombinations, and one mutation per site. This is not interesting or useful.

• However, the number of (observable) recombinations is small in realistic sets of sequences. ``Observable” depends on n and m relative to the number of recombinations.

• problem: given a set of sequences M, find a phylogenetic network generating M, minimizing the number of recombinations (Hein’s problem).

Minimization is NP-hard

The problem of finding a phylogenetic network that creates a given set of sequences M, and minimizes the number of recombinations, is NP-hard. (Wang et al 2000) (Semple 2004)

Wang et al. explored the problem of finding a phylogenetic network where the recombination cycles are required to be node disjoint, if possible.

They gave a sufficient but not a necessary condition to recognize cases when this is possible. O(nm + n^4) time.

We can solve the minimization problem in polynomial time,when node disjoint recombination cycles are possible.

Recombination Cycles

• In a Phylogenetic Network, with a recombination node x, if we trace two paths backwards from x, then the paths will eventually meet.

• The cycle specified by those two paths is called a ``recombination cycle”.

Galled-Trees

A recombination cycle in a phylogenetic network is called a “gall” if it shares no node with any other recombination cycle.

A phylogenetic network is called a “galled-tree” if every recombination cycle is a gall.

4

1

3

2 5

a: 00010

b: 10010

d: 10100

c: 00100

e: 01100

f: 01101

g: 00101

A galled-tree generatingthe sequences generatedby the prior network.

3

4

p s

ps

Sales pitch for Galled-TreesGalled-trees represent a small deviation from true trees.

There are sufficient applications where it is plausible that a galled tree exists that generates the sequences.

Observable recombinations tend to be recent; block structure of human DNA; recombination is sparse, so the true history of observable

recombinations may be a galled-tree.

The number of recombinations is never more than m/2. Moreover,when M can be derived on a galled-tree, the number of recombinations used is the minimum number over any phylogenetic network, even if multiple cross-overs at a recombination event are counted as a single recombination.

A galled-tree for M is ``almost unique” - implications for reconstructing thecorrect history.

Old (Aug. 2003) Results

• O(nm + n^3)-time algorithm to determine whether or not M can be derived on a galled-tree with all-0 ancestral sequence.

• Proof that the galled-tree produced by the algorithm is a “nearly-unique” solution.

• Proof that the galled-tree (if one exists) produced by the algorithm minimizes the number of recombinations used, over all phylogenetic-networks with all-0 ancestral sequence.

New work

We derive the galled-tree results in a more general setting that addresses unconstrained

recombination cycles and multiplecrossover recombination. This also solves theproblem of finding the ``most tree-like” network when a perfect phylogeny is not possible. In this algorithm, no ancestral

sequence is known in advance.

Blobbed-trees: generalizing galled-trees

• In a phylogenetic network a maximal set of intersecting cycles is called a blob.

• Contracting each blob results in a directed, rooted tree, otherwise one of the “blobs” was not maximal.

• So every phylogenetic network can be viewed as a directed tree of blobs - a blobbed-tree.

The blobs are the non-tree-like parts of the network.

Ugly tanglednetwork insidethe blob.

Every network is a tree of blobs. How do the tree partsand the blobs relate?

How can we exploitthis relationship?

Incompatible Sites

A pair of sites (columns) of M that fail the

4-gametes test are said to be incompatible.

A site that is not in such a pair is compatible.

0 0 0 1 01 0 0 1 00 0 1 0 01 0 1 0 00 1 1 0 00 1 1 0 10 0 1 0 1

1 2 3 4 5abcdefg

1 3

4

2 5

Two nodes are connected iff the pairof sites are incompatible, i.e, fail the 4-gamete test.

Incompatibility Graph

M

THE MAIN TOOL: We represent the pairwise incompatibilities in a incompatibility graph.

The connected components of G(M) are very informative

• The number of non-trivial connected components is a lower-bound on the number of recombinations needed in any network (Bafna, Bansal; Gusfield, Hickerson).

• When each blob is a single-cycle (galled-tree case) all the incompatible sites in a blob must come from a single connected component C, and that blob must contain all the sites from C. Compatible sites need not be inside any blob. (Gusfield et al 2003-5)

Simple Fact

If sites two sites i and j are incompatible, then the sites must be together on some recombination cycle whose recombination point is between the two sites i and j.

(This is a general fact for all phylogenetic networks.)

Ex: In the prior example, sites 1, 3 are incompatible, as are 1, 4; as are 2, 5.

A Phylogenetic Network

00000

52

3

3

4S

p

PS

1

4

a:00010

b:10010

c:00100

10010

01100

d:10100

e:01100

00101

01101

f:01101

g:00101

00100

00010

Simple Consequence of the simple fact

All sites on the same (non-trivial) connected component of the incompatibility graph

must be on the same blob in any blobbed-tree.Follows by transitivity.

So we can’t subdivide a blob into a tree-like structure if it only contains sites from a single connected component of the incompatibility graph.

Key Result about Galls: For galls, the converse of the simple

consequence is also true.Two sites that are in different (non-trivial) connected

components cannot be placed on the same gall inany phylogenetic network for M.

Hence, in a galled-tree T for M each gall contains all and only the sites of one (non-trivial) connected component of the incompatibility graph. All compatible sites can be put on edges outside of the galls.

This is the key to the galled-tree solution.

4

1

3

2 5

a: 00010

b: 10010

d: 10100

c: 00100

e: 01100

f: 01101

g: 00101

A galled-tree generatingthe sequences generatedby the prior network.

2

4

p s

ps

1 3

4

2 5

Incompatibility Graph

Motivated by the one-one correspondence between gallsand non-trivial connected components, we ask:

To what extent does this one-one correspondence holdin general blobbed-trees, i.e. with no constraints onhow recombination cycles interweave?

The Decomposition Theorem (Recomb 2005)

For any set of sequences M, there is a blobbed-tree T(M) that derives M, where each blob contains all and only the sites in one non-trivial connected component of G(M). The compatible sites can always be put on edges outside of any blob.

A blobbed-tree with this structure is called fully-decomposed.

Optimality of Galled-Trees

Theorem: (G,H,B,B) The minimum number of recombination nodesin any phylogenetic network for M is at least the number of non-trivial connected components of the incompatibility graph.

Hence, when the sequences on each blob on T(M) can be generatedwith a single recombination node, the blobbed-tree minimizesthe number of recombination nodes over all phylogeneticnetworks and all choices of ancestral sequence. This solvesthe root-unknown galled-tree problem in polynomial time.Code is on the web.

The number of arrangements on agall (all-0 ancestral sequence)

By analysing the algorithm to layout the sites on a gall (not discussed here), one canprove that the number of arrangements of any gall isat most three, and this happens only if the gall has two sites.

If the gall has more than two sites, then the number ofarrangements is at most two.

If the gall has four or more sites, with at least two siteson each side of the recombination point (not the side ofthe gall) then the arrangement is forced and unique.

Computing close bounds on the minimum number of

recombinationsDan Gusfield UCD

Y. Song, Y. F. Wu, D. Gusfield (ISMB2005)

D. Gusfield, D. Hickerson (Dis. Appl. Math 2005)

D. Gusfield, V. Bansal (Recomb 2005)

The grandfather of all lower bounds - HK 1985

• Arrange the nodes of the incompatibility graph on the line in order that the sites appear in the sequence. This bound requires a linear order.

• The HK bound is the minimum number of vertical lines needed to cut every edge in the incompatibility graph. Weak bound, but widely used - not only to bound the number of recombinations, but also to suggest their locations.

Justification for HK

If two sites are incompatible, there must have been some recombination where the crossover point is between the two sites.

1 2 3 4 5

HK Lower Bound

1 2 3 4 5

HK Lower Bound = 1

More general view of HK

Given a set of intervals on the line, and for each interval I, a number N(I), define the composite problem: Find the minimum number of vertical lines so that every interval I intersects at least N(I) of the vertical lines.

In HK, each incompatibility defines an interval I where N(I) = 1.

The composite problem is easy to solve by a left-to-right myopicplacement of vertical lines: Sort the intervals by right end-point;Process the intervals left to right in that order; when the rightendpoint of an interval I is reached, place there (if needed) additional vertical so that N(I) lines intersect I.

This general approach is called the Composite Method(Simon Myers 2002).

If each N(I) is a ``local” lower bound on the number ofrecombinations needed in interval I, then the solution tothe composite problem is a valid lower bound for thefull sequences. The resulting bound is called the compositebound given the local bounds.

Haplotype Bound (Simon Myers)

• Rh = Number of distinct sequences (rows) - Number of distinct sites (columns) -1 (folklore)

• Before computing Rh, remove any site that is compatible with all other sites. A valid lower bound results - generally increases the bound.

• Generally really bad bound, often negative, when used on large intervals, but Very Good when used as local bounds in the Composite Interval Method, and other methods.

Composite Interval Method using RH bounds

Compute Rh separately for each of the C(m,2) intervals of the m sites; let N(I) = Rh(I) be the local lower bound for interval I. Then compute the composite bound using these local bounds.

Polynomial time and gives bounds that often double HK in our simulations.

Composite Subset Method (Myers)

• Let S be subset of sites, and Rh(S) be the haplotype bound for subset S. If the leftmost site in S is L and the rightmost site in S is R, then use Rh(S) as a local bound N(I) for interval I = [S,L].

• Compute Rh(S) on many subsets, and then solve the composite problem to find a composite bound.

RecMin (Myers)

• World Champion Lower Bound Program (until now). Often RecMin gives a bound three times as large as HK.

• Computes Rh on subsets of sites, but limits the size and the span of the subsets. Default parameters are s = 6, w = 15 (s = size, w = span).

• Generally, impractical to set s = w = m, so generally one doesn’t know if increasing the parameters would increase the bound. (example: Myers bound of 70 on the LPL data).

Optimal RecMin Bound (ORB)

• The Optimal RecMin Bound is the lower bound that RecMin would produce if both parameters were set to their maximum values (s = w= m).

• In general, RecMin cannot compute (in practical time) the ORB.

• Practical computation of the ORB is our first contribution.

Computing the ORB

• Gross Idea: For each interval I, use ILP to find a subset of sites S that maximizes Rh(S) over all subsets in interval I. Call the result Opt(I).

• Set N(I) = Opt(I), and compute the composite bound using those local bounds.

• The composite bound is the ORB. -- the result one would get by using all 2^m subsets in RecMin, with s = w = m.

We have moved from doing an exponential number of simplecomputations (computing Rh for each subset), to solving a quadratic number of (possibly expensive) ILPs.

Is this a good trade-off in practice? Our experience - very much so!

How to compute Opt(I) by ILPCreate one variable Xi for each row i; one variable Yj for each column j in interval I. All variables are 0/1 variables.

Define S(i,i’) as the set of columns where rows i and i’ have different values. Each column in S(i,i’) is a witness that rows i and i’ differ.

For each pair of rows (i,i’), create the constraint: Xi + Xi’ <= 1 + ∑ [Yj: jin S(i,i’)]

Objective Function: Maximize ∑ Xi - ∑ Yj -1

Alternate way to compute Opt(I) by ILP

Create one variable Yj for each column jin interval I. All variables are 0/1 variables.

S(i,i’) as before.

For each pair of rows (i,i’) create the constraint: 1 <= ∑ [Yj: jin S(i,i’)]

Objective Function: Maximize N - ∑(Yj) -1

Finds the smallest number of columns to distinguish the rows.

First remove any duplicate rows. Let N be the number of rowsremaining.

Two critical tricks

1) Use the second ILP formulation to compute Opt(I). It solves much faster than the first (why?)

2) Reduce the number of intervals where Opt(I) is computed:

If the solution to Opt(I) uses columns thatonly span interval L, then there is noneed to find Opt(I’) in any interval I’ containing L. Each ILP computationdirectly spawns at most 4 other ILPs. Apply this idea recursively.

I

L

With the second trick we need to find Opt(I) for only 0.5% - 35% of all the C(m,2) intervals (empirical result). Surprisingly fast in practice (with either the GNU solver or CPLEX).

Bounds Higher Than the Optimal RecMin Bound

• Often the ORB underestimates the true minimum, e.g. Kreitman’s ADH data: 6 v. 7

• How to derive sharper bound? Idea: In the composite method, check if each local bound N(I) = Opt(I) is tight, and if not, increase N(I) by one.

• Small increases in local bounds can increase the composite bound.

Bounds Sharper Than Optimal RecMin Bound

• A set of sequences M is called self-derivable if there is a network generating M where the sequence at every node (leaf and intermediate) is in M.

• Observation: The haplotype bound for a set of sequences is tight if and only if the sequences are self-derivable.

• So for each interval I where Opt(I) is computed, we check self-derivability of the sequences induced by columns S*, where S* are the columns specified by Opt(I). Increase N(I) by 1 if the sequences are not self-derivable.

Algorithm for Self-Derivability

• Solution is easy when there are no mutations --only recombinations are allowed and one initial pair of sequences is chosen as ``reached”.

• Two reached sequences S1 and S2 can reach a third sequence S3, if S3 can be created by recombining S1 and S2.

• Do BFS to see if all sequences can be reached by successive application of the ``reach operation”.

• Clearly polynomial time and can be optimized with suffix-trees etc. (Kececiouglu, Gusfield)

Self-derivability Test with mutations allowed

• For each site i, construct set MUT(i) of sequence pairs (S1, S2) in M where S1 and S2 differ at only site.

• Try each sequence in M as root (which is the only reached sequence initially).

• For each root, enumerate all ways of choosing exactly one ordered pair of sequences (Sp, Sq) from each MUT(i). Sp is allowed to ``reach” Sq.

• Run the prior self-derivability algorithm with these new permitted reach relations, to test if all sequences in M can be reached. If so, then M is self-derivable, otherwise it is not.

Checking if N(I) should be increased by two

If the set of sequences is not self-derivable,we can test if adding one new sequence makes

it self-derivable.the number of candidate sequences is

polynomial and for each one added to M we check self-derivability. N(I) should be increased by two if none of these sets of sequences is self-derivable.

Program HapBound

• HapBound –S. Checks each Opt(I) subset for self-derivability. Increase N(I) by 1 or 2 if possible. This often beats ORB and is still practical for most datasets.

• HapBound –M. Explicitly examine each interval directly for self-derivability. Increase local bound if possible. Derives lower bound of 7 for Kreitman’s ADH data in this mode.

HapBound vs. RecMin on LPL from Clark et al.

Program Lower Bound Time

RecMin (default) 59 3s

RecMin –s 25 –w 25 75 7944s

RecMin –s 48 –w 48 No result 5 days

HapBound 75 31s

HapBound -S 78 1643s

2 Ghz PC

Example where RecMin has difficulity in Finding Optimal Bound

on a 25 by 376 Data MatrixProgram Bound Time

RecMin default 36 1s

RecMin –s 30 –w 30 42 3m 25s

RecMin –s 35 –w 35 43 24m 2s

RecMin –s 40 –w 40 43 2h 9m 4s

RecMin –s 45 –w 45 43 10h 20m 59s

HapBound 44 2m 59s

HapBound -S 48 39m 30s

Frequency of HapBound –S Bound Sharper Than Optimal

RecMin Boundms param. Rho=1 Rho=5 Rho=10 Rho=20

Theta=1 0.0% 0.4% 0.5% 1.5%

Theta=5 0.7% 4.0% 10.4% 27.0%

Theta=10 1.4% 9.2% 17.8% 40.4%

Theta=20 1.4% 10.5% 27.8% 45.4%

For every ms parameters, 1000 data sets are used.

Computing Upper Bounds

• The method is an adaptation of the ``history” lower bound (Myers).

• A non-informative column is one with fewer than two 0’s or fewer than two 1’s.

Single History Computation

1) Set W = 02) Collapse identical rows together, and remove non-

informative columns. Repeat until neither is possible.3) Let A be the data at this point. If A is empty, stop, else

remove some row r from A, and set W = W + 1. Go to step 2).

Note that the choice of r is arbitrary in Step 3), so the resulting W can vary.

History Lower Bound

Theorem (Myers) Let W* be the minimum W obtained from all possible single history computations.

Then W* is a valid lower bound on the number of recombinations needed.

Naïve time: theta(n!) (RecMin), but can bereduced to theta(2^n) (Bafna, Bansal).

Converting the History Lower Bound to an Upper Bound

• Given a set of rows A and a single row r, define w(r | A - r) as the minimum number of recombinations needed to create r from A-r (well defined in our application).

• w(r | A-r) can be computed in linear time by a greedy-type algorithm.

Upper Bound Computation

1) Set W = 02) Collapse identical rows together, and remove non-

informative columns. Repeat until neither is possible.3) Let A be the data at this point. If A is empty, stop, else

remove some row r from A, and set W = W + W(r | A-r). Go to step 2).

Note that the choice of r is arbitrary in Step 3), so the resulting W can vary.

This is the Single History Computation, with a change instep 3).

Note, even a single execution of the upper bound computation gives a valid upper bound, and a way to construct a network. This is in contrast to the History Bound which requires finding the minimum W over all histories.

However, we also would like to find the lowest upper bound possible with this approach. This can be done in O(2^n) time by DP. In practice, we can use branch and bound to find the lowest upper bound, but we have also found that branching on the best local choice, or randomizing gives good results.

Branch and Bound

(Branching) In Step 3) choose r to minimize w(r | A-r) + L(A-r), where L(A-r) is some

fast lower bound on the number of recombinations needed for the set A-r. Even HK is good for this purpose.

(Bounding) Let C be the min for an full solution found so far; If W + L(A) >= C, then backtrack.

Kreitman’s 1983 ADH Data

• 11 sequences, 43 segregating sites

• Both HapBound and SHRUB took only a fraction of a second to analyze this data.

• Both produced 7 for the number of detected recombination events

Therefore, independently of all other methods, our lower and upper bound methods together imply that 7 is the minimum number of recombination events.

A minimal ARG for Kreitman’s data

QuickTime™ and aTIFF (LZW) decompressor

are needed to see this picture.

SHRUB produces code that can be input to an open source program to display the constructed ARG

The Human LPL Data (Nickerson et al. 1998)





Our new lower and upper bounds

Optimal RecMin Bounds

(We ignored insertion/deletion, unphased sites, and sites with missing data.)

(88 Sequences, 88 sites)

Study on simulated data: Exact-Match frequency for varying parameters

= Scaled mutation rate = Scaled recombination rate

• n = Number of sequences



n = 25

For < 5, our lower and upper bounds match more than 95% of the time.

Used Hudson’s MS to generate1000 simulated datasets for each pair of and



n = 15

Exact-Match frequency for varying number of sequences

Match frequency does not depend on n as much as it does on or



A closer look at the deviation

Average ratio of lower bound to upper bound when they do not match



For n = 25:

The numerical difference between lower and upper bounds grows as or increases, but their ratio is more stable.

Multiple Crossover Recombination

4-crossovers

2-crossovers = ``gene conversion”

Extensions to Gene Conversion

• ``Gene Conversion” is short two cross-over recombination that occurs in meiosis.

• The extent of gene-conversion is only now being understood, due to prior lack of fine-scale molecular data, and lack of algorithmic tools.

• Gene Conversion may be the Achilles heel of fine-scale association (LD) mapping methods.

New Results

• Both the lower bound and upper bound methods have been extended to incorporate gene conversion as well as single-crossover recombination.

• Allowing gene-conversion can reduce the total number of nodes in the network, and also the associated bounds. But …

Distinguishing gene conversion from recombination

For a given set of sequences, let B be thebound (lower or upper) when only recombination is

allowed, and let BC be the bound when gene-conversion is also allowed. Define D = B - BC.

We expect that D will generally be larger when sequences are generated using gene-conversion compared to when they are generated with recombination only.

In such studies, we have shown that we can use statistics like D to correctly determine the amount of gene-conversion used to generate the sequences.

Take-home message

• The upper and lower bound algorithms cannot ``make-up” gene-conversions. The bounds reflect the extent of gene-conversion in the true generation of the sequences.

Papers and Software

wwwcsif.cs.ucdavis.edu/~gusfield/

erice - structured pattern detection and exploitation deterministic algorithms

Documents