Pool-hmm: a Python program for estimating the allele frequency spectrum and detecting selective sweeps from next generation sequencing of pooled samples.

0.11 MB·English

by  Boitard, Simon

MolecularEcologyResources(2013)13,337–340 doi:10.1111/1755-0998.12063
Pool-hmm: a Python program for estimating the allele
frequency spectrum and detecting selective sweeps from next
generation sequencing of pooled samples
SIMON BOITARD,* ROBERT KOFLER,† PIERRE FRANC(cid:1)OISE,* DAVID ROBELIN,* CHRISTIAN
€
SCHLOTTERER† andANDREAS FUTSCHIK‡
*LaboratoiredeG(cid:2)en(cid:2)etiqueCellulaire,INRA,24ChemindeBordeRouge,AuzevilleCS52627,CastanetTolosanCedex31326,
France,†Institutfu€rPopulationsgenetik,VetmeduniVienna,Veterin€arplatz1,WienA-1210,Austria,‡InstituteofStatisticsand
OperationsResearch,UniversityofVienna,Universit€atsstrasse5/9,WienA-1010,Austria
Abstract
Due to its cost effectiveness, next generation sequencing of pools of individuals (Pool-Seq) is becoming a popular
strategy for genome-wide estimation of allele frequencies in population samples. As the allele frequency spectrum
provides information about past episodes of selection, Pool-seq is also a promising design for genomic scans for
selection. However, no software tool has yet been developed for selection scans based on Pool-Seq data. We intro-
ducePool-hmm,aPythonprogramfortheestimationofallelefrequenciesandthedetectionofselectivesweepsina
Pool-Seq sample. Pool-hmm includes several options that allow a flexible analysis of Pool-Seq data, and can be
run in parallel on several processors. Source code and documentation for Pool-hmm is freely available at
https://qgsp.jouy.inra.fr/.
Keywords: allele frequency spectrum, hidden Markov models, next generation sequencing, pooled DNA, selective
sweeps.
Received10July2012;revisionreceived26November2012;accepted29November2012
of the reference genome arise from a random sampling
Introduction
amongthe pooledchromosomes,soobservations canbe
Thedetectionofgenomicregionsthatevolvedundernat- redundant. Second, sequencing error probabilities are
ural selection is an important topic in population genet- larger than with classic Sanger sequencing (Luo et al.
ics. The case of hard sweeps, where a new mutant goes 2011)andarevariableamongandwithinreads.
tofixationinapopulationduetostrongdirectionalselec- Recently,Boitardet al.(2012)proposedahiddenMar-
tion, has received particular attention (Kim & Stephan kov model (HMM) for detecting sweeps based on Pool-
2002; Nielsen et al. 2005; Jensen et al. 2007; Boitard et al. Seqdata.Thismethodinvolvescomputingthelikelihood
2009;Alachiotiset al.2012). of the observed read information conditional on allele
Theadventofnextgenerationsequencing(NGS)tech- counts in the pool, for each genome position. Down-
nologies provides a new dimension to such genome stream analyses—estimation of the background allele
scansforselection.Inspiteoftheconsiderablereduction frequency spectrum (AFS) and detection of selective
in the cost of sequencing, the sequencing of individuals sweeps—are then based on these likelihoods. Uncer-
onapopulationscaleremainsexpensive.However,hard tainty concerning the true allele frequencies in the pool,
sweepscanbedetectedusingonlythesampleallelefre- whichmighttypicallybehigherforsiteswithlowcover-
quencies along the genome, and this information can be age or bad quality scores, is thus taken into account in
obtainedbysequencingDNAfromapoolofindividuals the analyses. Possible biases arising from unequal DNA
(Pool-Seq). Although Pool-Seq is considerably cheaper concentration or quality among individuals are not
thanthesequencingofindividuals,therearesomemeth- accounted for by this method, but these effects are
odological challenges associated with the analysis ofthe expected to be limited for large sample sizes (Futschik
resulting data. First, the reads covering a given position et al.2010).
Inthiswork,weproposeaPythonprogram,denoted
Correspondence:SimonBoitard,Fax:[phone]; Pool-hmm, that implements the method of Boitard et al.
E-mail:[e-mail] (2012). The two main applications of this program are
©2013BlackwellPublishingLtd
338 S. BOITARD ET AL.
AFS estimation and detection of selective sweeps, in a frequency estimation. In addition, the estimated allele
given region. These two applications are implemented frequency at genomic positions with low coverage is
independently, so it is possible, for instance, to detect essentiallydeterminedbythepriorAFS.
selective sweeps based on a background AFS that is Our likelihood-based approach for estimating the
specifiedbytheuser.Inaddition,Pool-hmmprovidesan AFS in a region or allele frequencies at each genomic
estimationofallelefrequenciesateachgenomicposition, positionisanalternativetodiscardingbasecallsorgeno-
whichcanbeusedinotherpopulationgeneticssoftware. micpositionsbasedonarbitrarythresholds,andhasthe
These model-based estimations are preferable to na€ıve advantage of using the available information in a more
estimates obtained by computing the ratio of allele continuous way. One important implication is that we
countsataposition,asdiscussedbelow. canestimatewithoutbiastheproportionofsingletonsor
otherlowallelecounts,evenatlow(downto0.5 9 )per
chromosome coverage (Boitard et al. 2012), which is
Inputdata
clearly not possible when thresholding genomic posi-
The main input of the program is the Pool-Seq data, tionsbasedonthenumberofalternativealleles.
which must be provided in the SAMtools pileup format Pool-hmm can be also used to compare the AFS of
(Li et al. 2009). Any alignment file in BAM or SAM for- genomic regions with different annotations (e.g. introns
mat can easily be converted to the pileup format using vs. exons). We provide a script that filters the input
thesamtoolsmpileupcommand(without-uor-goptions), pileup file for any feature present in an annotation .gtf
independently of the software used to align or prepro- file. Pool-hmm then infers the AFS based on the filtered
cessthereads. pileup.
Note that the allele frequencies considered by Pool-
hmmaresampleallelefrequencies(from0ton)andnot
Allelefrequencyestimation
population allele frequencies (from 0 to 1), see for
The method assumes an infinite sites model, where at instance the AFS in Fig. 1. Population and sample
most two alleles can be observed at each genomic posi- frequencies are closely related and essentially provide
tion, the ancestral allele and the derived allele. By the same information, but inference based on coalescent
default, the ancestral allele is considered unknown, and theory,asthederivationsofNielsenet al.(2005)thatare
Pool-hmm focuses on folded (rather thanderived) allele used in our sweep detection model, naturally involve
frequencies.Twootherstrategiesmightalsobespecified sampleallelefrequencies.
(option —ancestral-allele). First, the reference allele pro-
videdinthepileupcanbeconsideredtobetheancestral
Detectionofselectivesweeps
allele. Second, ancestral alleles at each genomic position
can be provided in an additional column of the pileup In the HMM of Boitard et al. (2012), each genomic posi-
file. tion is assumed to have a hidden state, which can take
For a given genomic region, Pool-hmm estimates the oneofthethreefollowingvalues:‘Selection’,forthesites
derived or folded AFS using an expectation maximiza- thatareveryclosetoasweptsite,‘Neutral’,forthesites
tion(EM)algorithm.ThestartingvalueforthisEMisthe thatarefarawayfromanysweptsiteand‘Intermediate’
expected AFS under a model with constant population forthesitesinbetween.Thesethreevaluesareassociated
size and scaled mutation rate h = 4 Nl = 0.005, which withdifferentAFS.The‘Neutral’AFScorrespondstothe
canbemodifiedbytheuser(option–theta). background (whole genome) AFS of the population. It
Pool-hmm can also estimate the derived (or minor) can either be estimated from the Pool-Seq data or pro-
allele frequency at each position in a specified region vided using option —spectrum-file. The ‘Intermediate’
(option—estim). For this estimation, the previously esti- and‘Selection’AFSarethendeducedfromthe‘Neutral’
mated AFS (or any other AFS provided by the user) is AFSusingthederivationsinNielsenet al.(2005),andare
consideredasaprior.Atagivenposition,itiscombined typically more skewed towards low and high allele fre-
withthelikelihoodoftheobservedreadstoobtainapos- quencies. The hidden states form a Markov chain along
terior distribution of the derived (or minor) allele fre- the genome with a per-site probability q of switching
quency. The estimated frequency is the one maximizing state(argument–k).Theobservedvariableateachgeno-
this posterior distribution. This estimation procedure is mic position is a vector summarizing the information
morereliable thanadirect estimation basedon theratio providedbyreadsatthisposition.
ofallelecountsatapositionbecauseitaccountsforaddi- In the HMM described above, a selective sweep is
tional properties of the read data, namely the coverage detected if the hidden state ‘Selection’ is inferred for
and the base qualities at a given position. For instance, a window of sites. Using Pool-hmm, this inference
low-quality base calls have less influence on the allele (option —pred) relies on two different criteria. First, the
©2013BlackwellPublishingLtd
POOL-HMM: A PROGRAM FOR ANALYSING POOLED NGS DATA 339
Fig.1 AFSinaquailsampleofn=20chromosomes,computedfromarandomsampleofgenomicpositions(emptycircles),genomic
positionswithinexons(fullcircles,left panel),genomicpositionswithinsweepwindow1(emptytriangles,rightpanel)orgenomic
positionswithinsweepwindow2(plus,rightpanel).Probabilitiesof0-and20-derivedallelesarenotshownbecausetheyarenotatthe
samescale.Thelargeprobabilityobservedfor19-derivedallelesmaybeduetothemisspecificationoftheancestralalleleatasmall
proportionofsegregatingsites.Sucherrorsareexpectedifthereissharedpolymorphismbetweenquailandchicken.
sequence of hidden states maximizing the likelihood of
Example
the HMM is computed using the Viterbi algorithm (Ra-
biner 1989) and returned in a file with suffix .pred. A WeappliedPool-hmmtoasampleof10quailsthatwere
summaryofthesweepwindowsdetectedfromthisalgo- sequenced ina single pool at 20 9coverage. Reads were
rithm (genomic regions with predicted hidden state aligned against the chicken genome release WUGSC2.1
‘Selection’)isalsoreturnedinafilewithsuffix.stat.Sec- usingglint(Courcelleet al.2008)andconvertedtopileup
ond, the posterior probability of hidden state ‘Selection’ format using samtools (Li et al. 2009). We focused on
is computed for each genomic position using the for- chromosome1((cid:1)200-Mblong,with20millionobserved
ward–backwardalgorithm,andisreturnedinafilewith segregating sites) and conducted the analyses described
suffix.post. above. We used option —a ‘reference’, thereby assuming
thatquailancestralallelesarethosethatarefoundonthe
chicken reference genome. The execution time of all
Parallelization
Pool-hmmcommandswithone,fouroreightprocessors
Analysis of whole genome NGS data can be very time- isgiveninTable 1.Asexpected,itdecreasessignificantly
demanding.TospeeduptheexecutionofPool-hmm,we when the number of processors increases, although not
parallelizedthepartsofthecodewherethelikelihoodof linearly because some parts of the code, as for instance
the observed read information conditional on the num- thequeuemanagement,arenotparallelized.Usingeight
ber of derived alleles in the pool is computed. These cores, a standard analysis involving AFS estimation and
computations represent the largest computational cost sweepdetectiontookabout5 h.Additionalsweepanaly-
(otheralgorithms,suchasEMorViterbi,areveryfastin ses with a different sensitivity (parameter -k) are then
comparison), but they can be performed independently very fast because they can use intermediate results (the
for each genomic position. Taking advantage of this HMM emission probabilities at each observed segregat-
property, our strategy is thus to build a queue of 10 kb ingsite)thatarestoredfromthefirstanalysis.Notealso
blocks that are distributed for analysis to different pro- that AFS estimation is much faster than allele frequency
cessors. The management of this queue and the coordi- estimationbecauseitisbasedononly2%ofthegenomic
nation of all processors are implemented using the positionsofchromosome1(chosenatrandom).Thepro-
Pythonmultiprocessinglibrary. portion of genomic positions used for AFS estimation
This parallelization strategy enables an optimized use canbedefinedusingoption—ratio.
ofmultipleprocessorsonasinglemachine.Whenrunning TheAFSonchromosome1isshowninFig. 1.Forcom-
Pool-hmmonacomputercluster,wealsorecommendcut- parison, Fig. 1 also shows the AFS obtained using only
ting the whole genome data into large regions (typically genomicpositionslocatedwithinexons(asthenumberof
chromosomes)andanalysingtheseregionsindependently thesepositionsismuchsmaller,weusedallofthemrather
ondifferentnodes.Thiscanbecombinedwithparalleliza- thanonly2%).WefilteredthepileupwithPool-hmm,using
tionwithineachnode,asdescribedabove. agtffilecorrespondingtothelatestEnsemblannotationof
©2013BlackwellPublishingLtd
340 S. BOITARD ET AL.
Table1 ExecutiontimeofPool-hmmfortheanalysisofchromosome1inaquailsampleofn=20chromosomes.Resultsareprovided
forseveraltypesofanalysesandforone,fouroreightavailableprocessorsonacomputingcluster.Pool-hmmcommandscorrespond-
ingtotheseanalysesinthecaseofoneavailableprocessorarelistedbelowthetable.
Numberofprocessors AFSestimation* Firstsweepprediction† Additionalsweepprediction‡ Allelefrequencyestimation§
1 6h4min9s 12h21min36s 0h10min14s 31h7min47s
4 1h57min46s 4h32min57s 0h09min21s 7h47min9s
8 1h33min30s 3h15min53s 0h07min06s 4h17min14s
*Pythonpool-hmm.py–input-filequail-n20-a‘reference’–only-spectrum–theta0.005–ratio50.
†Pythonpool-hmm.py–input-filequail-n20-a‘reference’–pred–spectrum-filequail–k0.0000000001.
‡Pythonpool-hmm.py–input-filequail-n20-a‘reference’–pred–emit-file–k0.0000000001.
§Pythonpool-hmm.py–input-filequail-n20-a‘reference’–estim–spectrum-filequail.
thechickenassemblyusedforthealignment(url:ftp://ftp.
Acknowledgements
ensembl.org/pub/release-68/gtf/gallus_gallus/).Exonic
regions have an overall deficit of segregating sites, but We thank ChristineLeterrierfor providingthe quail data, and
apartfromthattheshapeoftheAFSintheseregionsisclose AndreaRauforhercarefulreadingofthemanuscript.Christian
tothatobtainedfromrandomregions. Schlo€ttererissupportedbygrantsoftheAustrianScienceFund
(FWF, P19467). Travels between France and Austria were
Seventy-four sweep windows were detected on chro-
fundedbythePHCAmadeusgrant25154QH.
mosome 1. The .stat file reporting these regions is pro-
vided as supporting information. The evidence for each
sweep window can be assessed using the third column
References
ofthisfile,whichrepresentsthemaximumoftheposter-
ior probability of hidden state ‘Selection’ along the win- AlachiotisN,StamatakisA,PavlidisP(2012)Omegaplus:ascalabletool
dow (in log scale). To illustrate the specificity of the forrapiddetectionofselectivesweepsinwhole-genomedatasets.Bio-
informatics,28,2274–2275.
sweepwindowsdetectedbyourapproach,weestimated
BoitardS,Schlo€ttererC,FutschikA(2009)Detectingselectivesweeps:a
theAFSinthesweepwindowscorrespondingtothetwo newapproachbasedonhiddenMarkovmodels.Genetics,181,1567–
first lines of the .stat file (Fig. 1, right panel). Sweep 1578.
window 1 was characterized by an excess of low- and BoitardS,Schlo€ttererC,NolteV,PandeyR,FutschikA(2012)Detecting
selective sweeps from pooled next generation sequencing samples.
high-frequency alleles, whereas sweep window 2 was
MolecularBiologyandEvolution,29,2177–2186.
characterized by a general deficit of segregating sites. CourcelleE,BeausseY,LetortSetal.(2008)Narcisse:amirrorviewof
Indeed,thedetectionmethodimplementedinPool-hmm conservedsyntenies.NucleicAcidsResearch,36,D485–D490.
makes use of both the density of segregating sites and Futschik A, Schlo€tterer , C (2010) Massively parrallel sequencing of
pooled DNA samples – the next generation of molecular markers.
the allele frequency pattern among segregating sites to Genetics,186,207–218.
distinguish sweep regions from neutral regions. Further JensenJ,ThorntonK,BustamanteC,AquadroC(2007)Ontheutilityof
details on this point and comparisons with alternative linkagedisequilibriumasastatisticforidentifyingtargetsofpositive
approachescanbefoundinBoitardet al.(2009,2012). selectioninnonequilibriumpopulations.Genetics,176,2371–2379.
KimY,StephanW(2002)Detectingalocalsignatureofgenetichitchhik-
ingalongarecombiningchromosome.Genetics,160,765–777.
Li H, Handsaker B, Wysoker A etal. (2009) The sequence alignment/
Conclusion
map(sam)formatandsamtools.Bioinformatics,25,2078–2079.
LuoL,BoerwinkleE,XiongM(2011)Associationstudiesfornext-genera-
Pool-hmmisthefirstsoftwaretoolforPool-Seqdatathat
tionsequencing.GenomeResearch,21,1099–1108.
provides a probabilistic allele frequency estimation and
Nielsen R, Williamson L, Kim Y, Hubisz M, Clark A, Bustamante C
detects selective sweeps on a genomic scale. The imple- (2005) Genomic scans for selective sweeps using SNP data. Genome
mented statistical algorithms account for two important Research,15,1566–1575.
RabinerL(1989)Atutorialonhiddenmarkovmodelsandselectedappli-
features of Pool-Seq data, the random sampling among
cationsinspeechrecognition.ProceedingsoftheIEE,77,257–287.
chromosomes within the pool and sequencing errors.
Pool-hmm includes several options that allow a flexible
analysisofPool-Seqdata. S.B.,D.R.,C.S.andA.F.designedthesoftware.S.B.,R.K.
and P.F. implemented the software. S.B., R.K., A.F. and
C.S.wrotethemanuscript.
Softwareavailability
Source code and documentation for Pool-hmm is freely
available at https://qgp.jouy.inra.fr/. Several test data
setsarealsoprovided.
©2013BlackwellPublishingLtd