000 001 002 003 004 005 006 007 008 009 010 011 012 013 014

015 016 017 018 019 020 021 022 023 024 025 026 027 028 029 030 031 032 033 034 035 036 037 038 039 040 041 042 043 044 045 046 047 048 049 050 051 052 053

Scalable imputation of genetic data with a discrete fragmentation-coagulation process

Anonymous Author(s) Afliation Address email

Abstract
We present a Bayesian nonparametric model for genetic sequence data in which a set of genetic sequences is modelled using a Markov model of partitions. The partitions at consecutive locations in the genome are related by their clusters rst splitting and then merging. Our model can be thought of as a discrete time analogue of continuous time fragmentation-coagulation processes [Teh et al 2011], preserving the important properties of projectivity, exchangeability and reversibility, while being more scalable. We apply this model to the problem of genotype imputation, showing improved computational efciency while maintaining the same accuracies as in [Teh et al 2011].

Introduction

The increasing availability of genetic data (for example, from the Thousand Genomes project [1]) and the importance of genetics in scientic and medical applications requires the development of scalable and accurate models for genetic sequences which are informed by genetic processes. Although standard models like the coalescent with recombination [2] are accurate, they suffer from intractable posterior computations. To address this, various hidden Markov model (HMM) based approaches have been proposed in the literature as more scalable alternatives (e.g. [3, 4]). Due to gene conversion and recombination processes, genetic sequences exhibit local mosaic-like structure, where sequences are composed of prototypical segments called haplotypes [5]. Locally, these prototypical segments are shared by a cluster of sequences: each sequence in the cluster is described well by a haplotype that is specic to the location of the cluster. An example of such a structure found by the DFCP is shown in Figure 1. HMMs can capture this structure by having each latent state correspond to one of the haplotypes [3]. Unfortunately, this leads to symmetries in the posterior distribution arising from the nonidentiability of the state labels [6]. Furthermore, current state-of-the art HMM methods often involve costly model selection procedures in order to choose the number of latent states. A continuous fragmentation-coagulation process (CFCP) has recently been proposed for modelling local mosaic structure in genetic sequences [7]. CFCPs are nonparametric models dened directly on unlabelled partitions, and thus avoid costly model selection and also the label switching problem [8]. Inference algorithms for CFCPS that scale linearly in the number of sequences and the length of the sequences have been developed previously [7], but because CFCPs are continuous time Markov jump processes, the computational overhead needed to model the latent continuous time jump events can preclude scalability to large datasets. In this work, we present an alternative fragmentation and coagulation process dened on a discrete grid (DFCP) which provides the advantages of the original continuous time CFCP while being more scalable. The DFCP describes location dependent unlabelled partitions such that at each location on the chromosome the clusters will split into multiple clusters which then merge to form the clusters 1

054 055 056 057 058 059 060 061 062 063 064 065 066 067 068 069 070 071 072 073 074 075 076 077 078 079 080 081 082 083 084 085 086 087 088 089 090 091 092 093 094 095 096 097 098 099 100 101 102 103 104 105 106 107
3 5 33 4 7

5 6

10

2 2 7 5 7 1 2 7 6 5 36 35 3 13

10

1 64 19 3 9 10 20 2 3 1 67 2 3 1 11 10 11 1 68 1 12 1 8 15 1 6 6 6 6 2 2 18 20 37 34 42 41 48 53 17 3 5 11

Figure 1: A haplotype structure discovered by DFCP on data from SeattleSNPs [9]. The data consists of single nucleotide polymorphisms (SNPs) on the ABCA3 gene typed by SeattleSNPs. Horizontal axis represents SNP locations along the gene. Each node is a cluster of sequences at the SNP, with number denoting the size of the cluster and the grey level lling denoting the proportion of sequences with the major (more common) allele. Each path of a different colour is a prototypical segment (haplotype). A hotspot [10] would correspond to the complex pattern in the middle. This haplotype structure is induced by a single MCMC sample at convergence of the chain. at the next location. Like the CFCP, the DFCP avoids the label switching problem by dening a probability distribution directly on the space of unlabelled partitions. The structure of the splitting and merging of clusters across the chromosome then forms a mosaic structure of haplotypes. Figure 1 gives an example of the structure discovered by DFCP. We describe the DFCP and its properties in section 2, and a forward-backward inference algorithm in section 3. Section 4 reports some experimental results showing good performance on imputation tasks, and in section 5 we conclude.

The discrete fragmentation coagulation process

In Humans, most of the bases on a chromosome are the same for all individuals in a population. Genetic variations arise through mutations such as single nucleotide polymorphisms (SNPs), which are locations in the genome where a single base was altered by a mutation at some time in the ancestry of the chromosome. At each SNP location, a particular chromosome has one of usually two possible bases (referred to as the major and minor allele). Consequently, SNP data for a chromosome can be modelled as a binary sequence, with each entry indicating which of the two bases is present at that location. In this paper we consider SNP data consisting of n binary sequences x = (xi )n , i=1 where each sequence xi = (xit )T is of length T and corresponds to the T SNPs on a segment of t=1 a chromosome in an individual, with the tth entry xit denoting the two possible alleles. We will model these sequences using a discrete time fragmentation-coagulation process (DFCP). This models the sequence values at a SNP location t using a latent partition t of the sequences with each cluster in the partition corresponding to a haplotype. The DFCP models the sequence of partitions using a (discrete time) Markov chain as follows: Starting with t , we rst fragment each cluster in t into smaller clusters, forming a ner partition t , then coagulating the clusters in t to nally form the clusters of t+1 . In the remainder of this section, we will rst give some background theory on partitions, fragmentations and coagulations. Then we will describe the DFCP as a Markov chain over partitions and the likelihood model used to relate the sequence of partitions to the observed sequences x.

2.1

Random partitions, fragmentations and coagulations

A partition of a set S is a clustering of S into non-overlapping non-empty subsets of S whose union is all of S. The Chinese restaurant process (CRP) forms a canonical family of distributions on 2

108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161

partitions. A random partition of a set S is said to follow the law CRP(S, , ) if: Pr() = [ + ]#1 [ + 1]#S1 1
a

[1 ]#a1 1

(1)

where [x]n = (x)(x + d) . . . (x + (n 1)d) is Kramps symbol and > , [0, 1) are d the concentration and discount parameters respectively [11]. A CRP can also be described by the following analogy: customers (elements of S) enter a Chinese restaurant and choose to sit at tables (clusters in ). The rst customer chooses any table. Subsequently, the i-th customer sits at a previously chosen table a with probability proportional to #a where #a is the number of customers sitting there and at some unoccupied table with probability proportional to + # where # is the number of tables sat at by previous customers. Fragmentation and coagulation operators are random partition-valued functions of partitions. The fragmentation F RAG(, , ) of a partition is formed by partitioning further each cluster a of according to CRP(a, , ) and then taking the union of the resulting partitions, yielding a partition of S that is ner than . Conversely, the coagulation C OAG(, , ) of is formed by partitioning the set of clusters of (i.e., the set itself) according to CRP(, , ) and then replacing each cluster with the union of its elements, yielding a partition that is coarser than . The fragmentation and coagulation operators are linked through the following theorem: Theorem 1. [12] Let S be a set and let A1 , B1 , A2 , B2 be random partitions of S such that: A1 CRP(S, 2 , 1 2 ), B2 CRP(S, 2 , 2 ), B1 |A1 F RAG(A1 , 1 2 , 2 ), A2 |B2 C OAG(B2 , , 1 ).

Then, for all partitions A and B of the set S with B a renement of A, Pr(A1 = A, B1 = B) = Pr(A2 = A, B2 = B). 2.2 The discrete fragmentation and coagulation process

The DFCP is parameterized by a concentration > 0 and rates (Rt )T 1 with Rt [0, 1). Under t=1 the DFCP, the marginal distribution of the partition t is CRP(S, , 0) and so controls the number of clusters that are found at each location. The rate parameter Rt controls the strength of dependence between t and t+1 , with Rt = 0 implying that t = t+1 and Rt 1 implying independence. Given and (Rt )T 1 , the DFCP on a set of sequences indexed by S = {1, . . . , n} is described t=1 by the following Markov chain. First we draw a partition 1 CRP(S, , 0). This CRP describes the clustering of S at t = 1. Subsequently, we draw t |t from F RAG(t , 0, Rt ), which fragments each of the clusters in t into smaller clusters, and then t+1 |t from C OAG(t , /Rt , 0), which coagulates clusters in t into larger clusters in t+1 . Each t has invariant marginal distribution CRP(S, , 0) while each t is distributed as CRP(S, , Rt ). This can be seen by applying Theorem 1 with the substitution 1 = 0, 2 = Rt , = /Rt . In population genetics the CRP appears as (and was predated by) Ewens sampling formula [13], a counting formula for the number of alleles appearing in a population, observed at a given location. Over a short segment of the chromosome where recombination rates are low, haplotypes behave like alleles and so a CRP prior on the number of haplotypes at a location is sensible. Further, since fragmentation and coagulation operators are dened in terms of CRPs which are projective and exchangeable, the Markov chain is projective and exchangeable in S as well. Projectivity and exchangeability are desirable properties for Bayesian nonparametric models because they imply that the marginal distribution of a given data item does not depend on the total number of other data items or on the order in which the other data items are indexed. In genetics, this captures the fact that usually only a small subset of a population is observed. Finally, the theorem also shows that conditioned on t+1 , t has distribution F RAG(t+1 , 0, Rt ) while t |t has distribution C OAG(t , /Rt , 0). In other words, the Markov chain is reversible with the reversed chain having the same distribution. Replication of chromosomes is directional and so the statistics for genetic processes along chromosomes are not expected to be reversible. The extent to which the directionality of replication affects the statistics of genetic data is not currently known, and many population genetics models are reversible for simplicity. 3

162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215
1 1 2 2 T 1 T

xi1

xi2

xiT i

1 t |t t+1 |t xit |ait ta |t t

1a
a 1

2a
a 2

T a
a T

CRP(S, , 0), F RAG(t , 0, Rt ), for t = 2, . . . , T , C OAG(t , /Rt , 0), Bernoulli(tait ), for t = 1, . . . , T , Beta(mt t , mt (1 t )), Beta(.5t , .5t ). (2)

Figure 2: Left: Graphical model for the discrete fragmentation coagulation process. Hyperparameters are not shown.

2.3

Likelihood model for sequence observations

Given the sequence of partitions (t )T , we model the observations in each cluster at each location t t=1 independently. For each cluster a t let ta be a parameter governing a likelihood model L(|ta ). For each sequence i, let ait t be the cluster in t containing i. We assume that conditioned on the partitions and parameters, the observations xit are independent, with Pr(xit = |ait = a) = L(|ta ). Since we are dealing with SNP data with binary labels, we take L(|ta ) to be a Bernoulli likelihood model with mean ta [0, 1]. To complete the model specication, we place on ta a Beta(mt t , mt (1 t )) prior (the t are marginally independent) with mass mt and mean t , with t itself having a Beta(t /2, t /2) prior. All mass hyperparameters are given an uninformative prior with p(m) m1 . We place uninformative priors on and Rt as well: p() 1 and 1 p(Rt ) Rt . Note that the prior gives more mass to values of Rt close to 0, since we expect the partitions of consecutive locations to be relatively similar to form the mosaic haplotype structure. The graphical model is shown in Figure 2. The evolution of the Markov chain over partitions (t )T can be visualized as a set of intersecting t=1 and diverging paths like those in Figure 1. Each path is a cluster (e.g. a) of sequences whose membership is stable over a number of SNP locations (e.g. t1 . . . , t2 ). This means that the sequences i a have entries generated as xit L(|ta ) for t1 t t2 . Since these sequences share the same set of parameters (ta )t2 1 , we expect the sequences to have similar local structure, i.e. to resemble t=t a prototypical haplotype. 2.4 Relationship with the continuous fragmentation-coagulation process

The continuous time version of the fragmentation-coagulation process [7], which we will refer to as the CFCP, is a partition valued continuous time Markov jump process. The CFCP is a pure jump process and can be dened in terms of its rates for various jump events. There are two types of events in the CFCP: fragmentation events, in which a single cluster a is split into two clusters b and c at a rate of R(#b)(#c)/(#a), and coagulation events in which two clusters b and c merge to form one cluster a at a rate of R/. As was shown in [7] the CFCP can be realised as the continuous time limit of the DFCP. Consider a DFCP with concentration and constant dependence parameter R. Then as 0 it is easy to see that the probability that the coagulation and fragmentation operations at a specic time step t induce no change in the partition structure t approaches 1. Conversely, the probability that these operations are non-trivial and involve exactly two clusters scales as O(), while all other events scale as O(2 ) or higher. Thus if we rescale the time steps as t t, then the expected number of nontrivial events involving two clusters over a nite time interval approaches a nite positive number, while that for all higher order events approaches zero. These are exactly the events in the CFCP. Further, the exact rates of the CFCP events can be derived as 0 with more careful calculations.

Inference with the discrete fragmentation coagulation process

We derive a Gibbs sampler for posterior simulation in the DFCP by making use of the exchangeability of the process. In particular, each iteration updates the trajectory of cluster assignments of one sequence i through the partition structure. To arrive at the updates, we rst derive the condi4

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

tional distribution of the ith trajectory given the others, which can be shown to be a Markov chain. Coupled with the likelihood terms, we then use a backwards-ltering/forwards-sampling algorithm to obtain a new trajectory for sequence i. In this section, we derive the conditional distribution of trajectory i and also the posterior distributions of the parameters Rt , which we will update using slice sampling [14]. 3.1 Conditional probabilities for the trajectory of sequence i

i We will refer to the projection of the partitions t and t onto S\i by t and i respectively. Let t at (respectively bt ) be the cluster assignment of sequence i at location t in t (respectively t ). We will denote adding sequence i to a cluster by itself in t (respectively t ) by at = (respectively i bt = ). Otherwise, the sequence i joins an existing cluster in t (i ), which corresponds to t i i i at t (bt t ). Thus the state spaces of at and bt are respectively t {} and i {}. t

Starting at t = 1, since the initial distribution is 1 CRP(S, , 0), the conditional cluster assignment of the sequence i in 1 is given by the CRP probabilities from (1):
i Pr(at = a|1 ) = i #a/(n 1 + ) if a t , /(n 1 + ) if a = .

(3)

To nd the conditional distribution of bt given at , we use the denition of the fragmentation operation as independent CRP partitions of each cluster in t . If at = , then the sequence i is in its own cluster in t and so it will be in its own cluster after fragmenting. Thus, bt = with probability i 1. If at = a t then bt must be one of the clusters in t into which a fragments. This can be a singleton cluster, in which case bt = , or it can be some cluster in i . We will refer to this set of t clusters in i by Ft (a). Since a is fragmented according to CRP(a, 0, R), when the i-th sequence t is added to this CRP it chooses to join b Ft (a) with probability proportional to (#b R) and it chooses to be in its own cluster with probability proportional to R#Ft (a). Normalizing these probabilities yields the following joint distribution: i (#b Rt )/#a if a t , b Ft (a), i Rt #Ft (a)/#a if a t , b = , i i Pr(bt = b|at = a, t , t ) = (4) 1 if a = b = , 0 otherwise. Similarly, to nd the conditional distribution of at+1 given bt = b we use the denition of the coagulation operation. If b = , then the sequence i is not in its own cluster in i and so it must t i follow the rest of the sequences in b to the unique a t+1 such that b a (i.e., b coagulates with other clusters to form a). We will refer to the set of clusters that coagulate to form a by Ct (a). If b = then the sequence i is in its own cluster in i and so we can imagine it being the last t customer added to the coagulating CRP(t , /Rt , 0) of the clusters of t . Hence the probability i that sequence i joins a cluster a t+1 is proportional to #Ct (a) while the probability that it is in i its own cluster in t+1 is proportional to /Rt . This yields the following joint probability: i 1 if a t+1 , b Ct (a), i i Rt #Ct (a)/( + Rt #t ) if a t+1 , b = , i Pr(at+1 = a|bt = b, t+1 , i ) = (5) t /( + Rt #i ) if a = b = , t 0 otherwise. 3.2 Message passing and sampling for the sequences of the DFCP

Once the conditional probabilities are dened, it is straightforward to derive messages that allow us to conduct backwards-ltering/forwards-sampling to resample the trajectory of sequence i in the DFCP. This provides an exact Gibbs update for the trajectory of the sequence conditioned on those of all other sequences. The messages we will dene are the conditional distribution of all the data seen after a given location in the sequence conditioned on the cluster assignment of sequence i at that location. The messages are dened as follows: mt (b) F
i mt (a) = Pr(xi,t+1 , . . . , xiT |at = a, t:T , i ). C t:T

(6) (7)

= Pr(xi,t+1 , . . . , xiT |bt = 5

i b, t:T , i ). t:T

270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323

We dene the last messages to be mT (a) = 1. These messages are computed as follows: C mt (b) = F mt (a) C =
bi {} t
i at+1 {}

i mt+1 (a)L(xi,t+1 |at+1 = a) Pr(at+1 = a|bt = b, t+1 , i ) . t C Coagulation probabilities from (5).

(8)

mt (b) Pr(bt F

= b|at =

i a, t i ) . t

(9)

Fragmentation probabilities from (4).

As the fragmentation and coagulation conditional probabilities are only supported for clusters a, b such that b a, these sums can be expanded so that only non-zero terms are summed over. For simplicity we do not provide the expanded forms here. Note that we marginalize out the cluster parameters t+1,a in (8) using the conjugacy between beta and Bernoulli distributions. Given these computations it is easy to dene backwards messages using the reversibility of the process. The backwards messages can be used to compute marginal probabilities of the observation as in the forward-backward algorithm. To sample from the posterior distribution of the trajectory for sequence i conditioned on the other trajectories and the data, we use the Markov property for the chain a1 , b1 , . . . , aT , bT and the denition of the messages. Starting at location 1, we have:
i Pr(a1 = a|xi , 1:T , i ) 1:T i i Pr(a1 = a|1 ) Pr(xi1 |a1 = a) Pr(xi2 , . . . , xiT |a1 = a, 1:T , i ), 1:T i = Pr(a1 = a|1 ) L(x1 |a1 = a)m1 (a). C CRP probabilities from (1).

(10)

For subsequent bt and at+1 for locations t = 1, . . . , T 1,

i i Pr(bt = b|at = a, xi , 1:T , 1:T ) i i Pr(bt = b|at = a, t , i ) Pr(xi,t+1 , . . . , xiT |bt = b, t:T , i ), t t:T i = Pr(bt = b|at = a, t , i ) mt (b). t F Fragmentation probabilities from (4). i Pr(at = a|bt1 = b, xi , 1:T , i ) 1:T i i Pr(at = a|bt1 = b, t , i ) Pr(xit |at = a) Pr(xi,t+1 , . . . , xiT |at = a, t:T , i ), t1 t:T i = Pr(at = a|bt1 = b, t , i ) L(xit |at = a)mt (a). t C Coagulation probability from (5).

(11)

(12)

The complexity of this update is O(KT ) where K is the average number of clusters in each i of the partitions 1:T , i . This complexity class is the same as that for the continuous time 1:T fragmentation-coagulation process and also for other related HMM methods such as fastPHASE. However since there is no overhead associated with modelling latent continuous time events, we expect inference in the DFCP to be more efcient than for the continuous time version. 3.3 Parameter updates

We use slice sampling [14] to update the and Rt parameters conditioned on the partition structure of the DFCP. Using Bayes rule, the denition (2) of the DFCP, and the identity [a]n = bn (a/b + b n)/(a/b), the posterior probability of and Rt given the partitions 1:T , 1:T follows. Pr(|, ) Pr() Pr(1 |, R1 ) Pr(1 |1 , , R1 ) Pr(T |T 1 , , RT 1 ), 1 () T + ( + n)
T t=1

#t

T 1 t=1

(/Rt ) . (/Rt + #t )

(13)

Pr(Rt |, , ) Pr(Rt ) Pr(t |t , , Rt ) Pr(t+1 |t , , Rt ),

1 #t #t #t+1 +1 (/Rt )(1 Rt )#t R Rt t (#t + /Rt )

bt

(#b Rt ).

(14)

324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377

1.00

0.994

0.99 a c c ura c y (proportion c orre c t) 0.993

accuracy

0.98

0.992

0.97

0.96

BEAGLE fastPHASE DFCP CFCP

0.25 0.50 held out sites 0.75

0.991

CFCP DFCP
0.990 0.0 0.2 0.4 0.6 0.8 1.0 runtim e (hours ) 1.2 1.4 1.6

0.95

Figure 3: Allele imputation for X chromosomes from the Thousand Genomes project. Left: Accuracy for prediction of held out alleles for the continuous (CFCP) and discrete (DFCP) versions of fragmentation-coagulation processes and for popular methods BEAGLE and fastPHASE. Right: Runtime versus accuracy for the continuous (CFCP) and discrete (DFCP) versions of the fragmentation-coagulation process. Each point reports the runtimes and the accuracies after a number (between 1 and 200) of samples have been collected. The runtimes do not start at 0 due to the initial 100 burn-in iterations.

Experiments

To examine the accuracy and scalability of the DFCP we conducted allele imputation experiments on SNP data from the Thousand Genomes project. We considered SNPs typed in 524 male X chromosomes from the March 2012 release. We chose 20 random intervals each containing 500 consecutive SNPs. The average length of these regions was 1.0 105 base pairs long. In three conditions we held out nested sets of 25%, 50% and 75% of the alleles uniformly over all pairs of sites and chromosomes, and used fastPHASE [3], BEAGLE [15], CFCP [7] and our DFCP to predict the held out alleles from the observed alleles. For these datasets the mean at-chance accuracy which would be found by predicting the major alleles is 0.945. For the 22 Human chromosomes that come in pairs, observed alleles are not ordered (this is called unphased data). We used X chromosome data from males because the accuracy of allele prediction can be evaluated based on the genetic sequence of a single chromosome (this is called phased data). Furthermore, as the X chromosome undergoes recombination in females it is a good model for the other 22 chromosomes. For more details about phasing we refer the reader to [16]. The software for fastPHASE, BEAGLE and the CFCP was obtained from their respective authors. BEAGLE and fastPHASE were run using their default settings. For all conditions, both CFCP and DFCP were run for 300 iterations and with the rst 100 iterations discarded as burn-in. There were 10 parameter updates per iteration. The accuracies for the DFCP and CFCP were computed by thresholding the empirical marginal probabilities of the held out alleles at 0.5 and reporting the proportion of correctly called alleles. Accuracies may be improved by choosing a threshold based on empirical estimates of the calibration of the model [17]. Accuracies for genotype imputation are displayed in Figure 4(left). We see that fastPHASE and BEAGLE outperformed CFCP and DFCP at low held-out levels, but their edge reduces with increasing held-out levels, with the CFCP and DFCP methods outperforming when 75% of the alleles were held out. The software for fastPHASE failed to run properly with high levels of held-out data (the diagnostics produced by the software did not make sense), so its predictive accuracies (which were at chance) were not reported here. The DFCP displayed performance comparable to the CFCP in all conditions. In Figure 4(right) we plot the accuracies of both CFCP and DFCP for a single trial at 50% held-out level, versus the amount of time taken. The behaviour exhibited by the algorithms in this single trial is typical. We see that DFCP is signicantly faster than CFCP. To further explore the computational capabilities of DFCP versus CFCP, we simulated datasets consisting of 1000 to 9000 sequences using the software ms [18] which is based on the coalescence with recombination model. We applied both CFCP and DFCP to the datasets. We report the runtimes per iteration per sequence of DFCP and CFCP on the datasets in Figure 4. While the DFCP 7

runtime per iteration (log seconds)

378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431

10

10

10

CFCP DFCP
10 1000
1

2000

3000

4000 5000 6000 number of sequences

7000

8000

9000

Figure 4: Runtimes per iteration per sequence of DFCP and CFCP on simulated datasets consisting of large numbers of sequences. managed to nish all 200 iterations in the allotted time, CFCP only managed to run a few iterations before the allotted time, which is why its curve has signicantly higher variance. All the MCMC was performed using the same computer and without computing unnecessary marginal statistics. We believe the signicantly slower performance of CFCPs displayed here is due to a combination of memory allocation and garbage collection issues due to the large numbers of fragmentation, coagulation and virtual events getting created and disposed of by the software, as well as the code itself being slowed down since the runtime is linear in the number of events as well. Our DFCP code was not signicantly optimized and so we expect that further improvements in performance are possible.

Discussion

In this paper we have presented a discrete time fragmentation coagulation process. The DFCP is a partition-valued Markov chain, where partitions evolve during each iteration via a fragmentation operation followed by a coagulation operation. The sequence of partitions mimics the mosaic haplotype structure observed in genetic sequences, and we applied the DFCP to an allele prediction task on X chromosome data from the Thousand Genomes Project, showing good accuracies and lower runtimes than the continuous-time fragmentation-coagulation process of [7]. Although the asymptotic computation cost of inference in the DFCP is the same as for the CFCP, we have found that both runtime and memory requirements of the DFCP are lower than for the CFCP. We are currently further optimizing our DFCP code. We expect DFCP to be scalable to modern genomic scale datasets. We will also extend the DFCP approach to unphased data as well, for solving both imputation and phasing problems. In future work we will explore better calling methods to improve imputation accuracies. We will also investigate why the performance of the fragmentation coagulation does not degrade as severely as that of other methods on sparsely observed data. Another avenue of future research is to understand how other genetic processes can be incorporated into the DFCP framework, including population structure and gene conversion. Although haplotype structure is a local property, the Markov assumption does not hold in real genetic data. This feature could be reected in hierarchical models based on DFCPs.

Acknowledgements

References
[1] The 1000 Genomes Project Consortium. A map of human genome variation from population-scale sequencing. Nature, 467:10611073, 2010. [2] R. R. Hudson. Properties of a neutral allele model with intragenic recombination. Theoretical Population Biology, 23(2):183 201, 1983.

432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485

[3] P. Scheet and M. Stephens. A fast and exible statistical model for large-scale population genotype data: Applications to inferring missing genotypes and haplotypic phase. The American Journal of Human Genetics, 78(4):629 644, 2006. [4] J. Marchini, B. Howie, S. Myers, G. McVean, and P. Donnelly. A new multipoint method for genome-wide association studies by imputation of genotypes. Nature Genetics, 39(7):906913, 2007. [5] M. J. Daly, J. D. Rioux, S. F. Schaffner, T. J. Hudson, and R. S. Lander. High-resolution haplotype structure in the human genome. Nature Genetics, 29:229232, 2001. [6] M. Stephens. Dealing with label switching in mixture models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 62(4):795809, 2000. [7] Y. W. Teh, C. Blundell, and L. T. Elliott. Modelling genetic variations using fragmentation-coagulation processes. In Advances in neural information processing systems, volume 24, 2011. [8] A. Jasra, C. C. Holmes, and D. A. Stephens. Markov chain Monte Carlo methods and the label switching problem in Bayesian mixture modeling. Statistical Science, 20(1):5067, 2005. [9] NHLBI Program for Genomic Applications. SeattleSNPs. June 2011. http://pga.gs.washington.edu. [10] L. Bottolo S. Myers, C. Freeman, G. McVean, and P. Donnelly. A ne-scale map of recombination rates and hotspots across th ehuman genome. Science, 310(5746):321324. [11] J. Pitman. Combinatorial stochastic processes. Springer-Verlag, 2006. [12] J. Pitman. Coalescents with multiple collisions. Annals of Probability, 27:18701902, 1999. [13] W. J. Ewens. The sampling theory of selectively neutral alleles. Theoretical Population Biology, 3:87 112, 1972. [14] R. M. Neal. Slice sampling. Annals of Statistics, 31:705767, 2003. [15] B. L. Browning and S. R. Browning. A unied approach to genotype imputation and haplotype-phase inference for large data sets of trios and unrelated individuals. American Journal of Human Genetics, 84:210223, 2009. [16] Jonathan Marchini, David Cutler, Nick Patterson, Matthew Stephens, Eleazar Eskin, Eran Halperin, Shin Lin, Zhaohui S. Qin, Heather M. Munro, GonAalo R. Abecasis, and Peter Donnelly. A comparison of phasing algorithms for trios and unrelated individuals. The American Journal of Human Genetics, 78(3):437 450, 2006. [17] B. N. Howie, P. Donnelly, and J. Marchini. A exible and accurate genotype imputation method for the next generation of genome-wide association studies. PLoS Genetics, 6, 2009. [18] R. R. Hudson. Generating samples under a wright-sher neutral model of genetic variation. Bioinfomatics, 18:337338.