Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
1
SupplementaryMethods
Sections:1. Barcodedistancematrixcomputation2. AnalysisofsourcesoferrorinMEMOIRoperation,andtheirimpactonreconstruction
accuracy3. SimulationofMEMOIRoperationformissingcellsandsparsetrees4. InferenceoftheswitchingratesbetweenthetwoEsrrbgeneexpressionstates
1. Barcodedistancematrixcomputation
In this sectionwedescribehowwecomputedthematrixofpairwisebarcodedistancesbetweencells
(Figure 3e). These distances are based on the complete barcoded scratchpad profile of each cell,
includingboththetotalnumberoftranscripts,denotedas𝑁,andthefractionofthosetranscriptsthat
colocalizedwithscratchpad,denotedas𝑝.
First,toaccountforstatisticalfluctuationsanderrorsinthedetectedfractionofcolocalizedtranscripts,
weconvertedthemeasuredvalueof𝑝toadistributionofpossiblevaluesforthecolocalizationfraction.
Toaccountformeasurementerrors,weimposedalowerboundof0.04andanupperboundof0.96on
the measured value of 𝑝. To determine the distribution of possible values of 𝑝, we assumed that
fluctuations follow a binomial distribution with mean𝑁𝑝 and variance𝑁𝑝(1 − 𝑝). This assumption
follows from the physical reasoning that the probability of colocalization of one transcript is
independentfromthatofothertranscriptsinthecell.
Thedistributionofpossiblevaluesofthecolocalizationfraction,𝑥,takestheform,
𝐹(𝑥𝑁) = 𝑁𝑥
𝑝*+ 1 − 𝑝 -.* +
Tokeepnumerical computations tractable,weapproximated the factorials in thebinomial coefficient
usinggammafunctions,i.e.𝑁! ≈ 𝛤(𝑁 + 1).
Tocomputethecell-to-celldistanceforasinglebarcode,wequantifiedhowthecolocalizationfraction
distributiondiffersforagivenbarcodebetweentwocells.Wedenotethedistributionofcolocalization
WWW.NATURE.COM/NATURE | 1
SUPPLEMENTARY INFORMATIONdoi:10.1038/nature20777
fractionofbarcode𝑖incell𝑎as𝐹56(𝑥)andthecomparabledistributionincell𝑏as𝐹86(𝑦).Wedefinethe
barcode𝑑6(𝑎, 𝑏)distancebetweenthetwocellsas,
𝑑6 𝑎, 𝑏 = 𝐹56 𝑥 𝐹86 𝑦 |𝑥 − 𝑦|𝑑𝑥𝑑𝑦-
=
-
=
Finally,theoverallcell-to-celldistancebasedonallbarcodeinformationwascalculatedastheaverage
over the distances derived for each barcode weighted for each barcode by the smaller number of
barcode transcripts ineitherof the twocells. Thisensures thatbarcodes thatarehighlyexpressed in
bothcellsareweightedmoreheavily thanbarcodes thatareweaklyexpressed ineitherorbothcells.
Denotingthenumberoftranscriptsofbarcode𝑖 incell𝑎as𝑁56 ,andsimilarlyforcell𝑏,thecell-to-cell
distancematrix,showninFig.3e,thentakestheform,
𝑑 𝑎, 𝑏 = min 𝑁56 , 𝑁86 𝑑6(𝑎, 𝑏)6
2. AnalysisofsourcesoferrorinMEMOIRoperation,andtheirimpactonreconstructionaccuracy
2a.IncorporatingnoiseintoMEMOIRsimulations
To understand what factors impact the overall reconstruction errors in MEMOIR, we simulated the
recording, readout, and reconstruction processes, incorporating relevant sources of stochastic
fluctuations(noise).Sourcesoferrorcanbedividedintothreecategories:(1)noiseinherenttothebasic
mechanismofstochasticscratchpadcollapse,(2)recordingnoise,and(3)readoutnoise(seeExtended
DataFig.6a).
Noise inherent to the system. The first noise category captures fluctuations due to the general
approachofrecording lineage information indiscretestochasticmutagenic(collapse)events,butdoes
not include noise arising from the specific experimental implementation of the system (those noise
sourcesaredescribedbelow).Reconstructionerrorsinthiscategoryresultfromcollapseeventsthatby
chance occur too frequently or too infrequently to provide lineage information, as well as identical
collapse events occurring in different cells (coincidences). For example, if the exact same scratchpad
collapseevent(s)occur independently in twosistercells, the resultingcladeswillbe indistinguishable.
The simulations shown in Figure 3g and Extended Data Figure 8 assume irreversible collapse of an
idealized set of binary scratchpads, occurring at a constant rate and therefore incorporate only this
WWW.NATURE.COM/NATURE | 2
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
source of error. Analysis of these simulations thus shows how the stochasticity of collapse events
impactsreconstructionaccuracyintheabsenceofadditionalsourcesofnoise.
Recording noise. Recording noise accounts for fluctuations in the collapse rate due to stochastic
fluctuationsintheexpressionlevelsofgRNAandCas9.Toaccuratelyrepresentthemagnitudeofthese
effects,wefirstmeasuredthesefluctuationsexperimentally,andthenincorporatedempiricalvaluesfor
thesefluctuationsintothesimulations.
First,wequantifiedgRNAexpressionvariabilitybothwithincolonies(intra-colony)andbetweencolonies
(inter-colony)using the co-expressed fluorescent reporter in individual cells inall of the108MEM-01
colonies.Wedenote the gRNAexpression in cell 𝑖 in colony𝑚 as𝑔6C. (ExtendedData Fig. 6b,c). The
intra-colony noise was computed by first determining the variance of the single-cell measurements
amongstcellsineachindividualcolonyandthenaveragingthatvarianceacrossallcolonies.Namely,
𝜎EF =1𝑁
1𝑛C
𝑔6C −1𝑛C
𝑔6CHI
6J-
FHI
6J-
+
CJ-
Here,𝑁isthetotalnumberofcolonies,and𝑛Cisthenumberofcellsincolony𝑚.
Similarly,the inter-colonynoisewascomputedbytakingthemeanof𝑔6Camongstthecells inagiven
colonyandthencomputingthevarianceofthecolonymeansacrossallcolonies.
𝜏EF =1𝑁
1𝑛C
𝑔6CHI
6J-
− 𝑔
F+
CJ-
where𝑔isthemeangRNAexpressionacrossallcellsinallcolonies,
𝑔 =1𝑁
1𝑛C
𝑔6CHI
6J-
+
CJ-
WWW.NATURE.COM/NATURE | 3
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
Second,wequantifiedvariabilityinCas9expressionusingthesameanalysis,performedonthenumber
ofCas9 transcripts,𝑐, readoutusingsmFISH in the final roundofhybridization.Wedenote the intra-
colony and inter-colony variation in Cas9 expression as 𝜎M and 𝜏M, respectively, and themean Cas9
expressionlevelacrossallcellinallcoloniesas𝑐.
We next incorporated the empirically determined noise measurements for Cas9 and gRNA into the
simulations.Wesettheeffectivecollapserateineachcell,𝜆,proportionaltotheproductofthenumber
ofCas9 transcripts,𝑐, and the gRNAactivity,𝑔:𝜆 = 𝑔𝑐. To incorporate recordingnoise,weassumed
each colony had its own values for 𝑐 and𝑔, drawn from the observed intercolony distributionswith
means𝑐and𝑔andstandarddeviations𝜏M and𝜏Erespectively.Next,weincorporatedtheintra-colony
noise,byaddingfluctuationstovaluesof𝑐and𝑔 ineachcell inagivencolony,withthemagnitudeof
fluctuationssetbytheempiricalintra-colonyvariations,𝜎M and𝜎E.Theprobabilityofcollapseforeach
scratchpadpergeneration,𝑝,wascalculatedfrom𝜆assumingaPoissonprocess,namely,𝑝 = 1 − 𝑒.PQ.
The normalization constant, 𝜂, was selected to ensure an average collapse probability of 0.1 per
scratchpadpergeneration,consistentwiththeempiricallyestimatedaveragecollapserateintheMEM-
01 colonies (seeFig. 2c andMethods).Weverified that the resulting fluctuations in the collapse rate
(from roughly 0.05 to 0.2 per scratchpad per generation) were consistent with the experimentally
observedfluctuationsinthelossofcolocalizationfractionbetweenindividualcellsacrossthe108MEM-
01coloniesusedintheanalysisshowninFigure3.Overall,asshowninFig.3iandExtendedDataFig.6d,
this recording noise produced only a relatively small decrease in the reconstruction accuracy of the
simulatedcolonies.
Readoutnoise.Thiscategoryincludes(a)stochasticexpressionofindividualbarcodedscratchpads,(b)
the effects of multiple integrations of the same barcoded scratchpad type, and (c) errors in smFISH
imagingreadout.Tosimulatetheimpactofreadoutnoiseonreconstructionaccuracy,wefirstobtained
theempiricaldistributionofsingle-celltranscriptcountsofeachofthe13scratchpadsinthecellsinall
108analyzedMEM-01colonies(seeFigurebelow).Thetranscriptcountdistributionsforthebarcoded
scratchpadswerewell-approximatedasgammadistributions(withanaveragemeanof5.4).Thegamma
distributionisexpectedwhenthereisaconstantprobabilityperunittimeofinitiatingatranscriptional
burst,andthenumberofmRNAsproducedperburstfollowsanexponentialdistribution,andhasbeen
observedempiricallytodescribediversestochasticgeneexpressionprocesses43,44.
WWW.NATURE.COM/NATURE | 4
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
Distributionofthenumberoftranscriptsofeachbarcodeinindividualcellsinthe108MEM-01coloniesanalyzedinFigure3.All13detectedbarcodesareshown.BlacklinesarefitstoGammadistributions.
We quantified the intra-colony and inter-colony variability in the expression levels of the barcoded
scratchpadsusing thesameproceduredescribedabove forgRNAandCas9. In this case,variancewas
calculatedforeachbarcodeandthensummedover13barcodespercelltoestimateeachcell’soverall
expression variance. These empirical parameterizations for the transcript count distributions were
incorporatedinthesimulations.
For readout, we first associatedwith each barcode a random number of transcripts drawn from the
gamma distribution associated with that barcode. Second, we incorporated additional noise from
smFISH imaging, based on experimental quantification of the accuracy of smFISH colocalization
detectionrates(Fig.2c).Ifascratchpadwasintact,thentheprobabilityofcolocalization,𝑝M,wassetto
0.87; if a scratchpad was collapsed, the probability of erroneous colocalization, 𝑝S, was set to 0.05.
Theseestimateswereobtainedfromourmeasurementsofcolocalizationfractionsbeforeactivationof
theMEM-01cells(Fig.2c),andareconsistentwithpreviousreportsonsmFISHfidelity16,17.Thenumber
ofscratchpadsthatcolocalizedforeachbarcodewasdrawnfrombinomialdistributionsbasedonthese
colocalizationdetectionprobabilities.Namely,thenumberofcolocalizedtranscripts𝑞fromatotalof𝑇
transcripts,foranintactintegrationofabarcode,wasdrawnfromthedistribution,
0 500
100
200
Cel
l cou
nts
BC: 2
0 50 1000
100
200BC: 9
0 10 20 300
100
200
300BC: 10
0 500
100
200BC: 12
0 500
100
200
300BC: 13
0 10 20 300
100
200
300
Cel
l cou
nts
BC: 14
0 10 200
200
400
BC: 16
0 10 20 300
100
200
300BC: 20
0 5 10 15Transcripts
0
200
400
600BC: 21
0 20 40Transcripts
0
100
200
300BC: 22
0 10 20 30Transcripts
0
100
200
300
Cel
l cou
nts
BC: 23
0 50Transcripts
0
50
100
150BC: 26
0 10 20Transcripts
0
200
400BC: 27
WWW.NATURE.COM/NATURE | 5
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
𝑝6HV5MV 𝑞; 𝑇 =𝑇𝑞𝑝MX 1 − 𝑝M Y.X
Similarly,thenumberofcolocalizedtranscriptsforacollapsedintegrationofabarcodewasdrawnfrom
thedistribution,
𝑝MZ[[5\]S^ 𝑞; 𝑡 =𝑇𝑞𝑝SX 1 − 𝑝S Y.X
Finally,weaccountedformultiple integrationsofthesamebarcodeasfollows: Ifabarcodehadmore
thanoneintegration,thetotalnumberoftranscriptsandthenumberofcolocalizedtranscriptsforthat
barcode were determined by summing the relevant transcripts from each individual integration.
Importantly,thedistributionoftheexpressionlevelofeachindividualintegrationwasselectedsuchthat
the distribution of the total number of transcripts (obtained by taking the convolution of the single-
integration distributions) matched the experimentally observed distribution for that barcode. For
example,ifabarcodehadtwointegrationsandanexpressiondistributioncharacterizedbyameanof10
transcriptsandshapeparameter2(fromthegammafit),theneachintegrationwasassumedtohavean
expressiondistributioncharacterizedbyameanof5transcriptsandshapeparameter1.
This procedure produced an ensemble of simulated colonies, where every cell was assigned a total
number of transcripts and a number of colocalized transcripts for each barcode (same as the
experimentaldata).Theresultingsimulatedcolonieswerethenreconstructedwiththesameneighbor-
joiningalgorithmusedontheexperimentaldata(seeMethods).
Impactof readoutnoiseonoverall reconstructionaccuracy. Figure3iandExtendedDataFigure6e,f
showthatnoiseinscratchpadexpressionandmultipleincorporationsofthesamebarcodedscratchpad
account formostof the reconstructionerror.Bycontrast,othercomponents suchas recordingnoise,
noiseintrinsictotheMEMOIRdesign,andnoisefromsmFISHimagingfidelitycontributedminimallyto
theoverallreconstructionerror.
WWW.NATURE.COM/NATURE | 6
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
Finally, to determine if the addition of fluctuations to the simulations matched the experimentally
observeddecrease in reconstructionaccuracy,wesimulated treeswithall threecomponentsofnoise
for various numbers of integrations per barcode and compared the distribution of reconstruction
accuracytothatobservedforthe108experimentalcolonies.AsshowninExtendedDataFigure6f,the
distribution of reconstruction accuracy obtained from simulations is generally consistent with the
experimentally observed distribution, with an effective value of 2 integration sites for each barcode.
Note that no fitting parameters were used besides the empirically measured fluctuations in the
recording and readout components. Together, these results indicate that this implementation of
MEMOIRbehavedasexpectedofasystemwith13scratchpadsandrealisticsourcesofnoise.Themain
sources of error, the stochastic expression of scratchpad transcripts and variable expression from
multiple integrations of the same barcoded scratchpad types, can be improved in future
implementations.
2b.Potentialsolutionstoaddresssourcesofnoiseandreduceerrors
HerewedescribewaystomitigatethekeysourcesofnoisecurrentlyaffectingMEMOIRreconstruction
accuracy:
● Noiseinherenttothesystem.Becausecollapseeventsarestochastic,similarcollapsepatterns
canarisebychanceindependently,creatingambiguities.Thelikelihoodofsucherrorsdecreases
rapidly as the number and diversity of barcoded scratchpads is increased. Further, tunable
control of Cas9 and gRNA expression or scratchpad affinity (e.g., by varying gRNA target
complementarity)allowsaveragecollapseratetobeoptimizedfordifferentregimes, including
accessingmore generations (ExtendedData Fig. 8b) or low collapse rates for studying sparse
trees(ExtendedDataFig.9).
● Readout noise: Multiple incorporations of barcoded scratchpad types. This is currently a
significant source of noise, as shown in Extended Data Fig. 6, and it is readily addressed as
above,e.g.,bystartingfromalargerandmorediversepoolofbarcodes.
● Readout noise: Stochastic scratchpad expression. Individual scratchpads are read out at the
RNAlevel,andthereforesubjecttofluctuationsintranscription.Thissourceofnoisethatcanbe
addressed by using 1) stronger promoters, 2) site-specific integration in well-expressed loci,
and/or 3) stabilized RNA transcripts (in order to build up signal over longer periods).
Alternatively,barcodeexpressionnoisecanbecircumventedbyemployingotherinsitulabeling
methods,likesingle-moleculeDNAFISH.
WWW.NATURE.COM/NATURE | 7
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
3. SimulationofMEMOIRoperationformissingcellsandsparsetrees
3a.Missingcells
To simulate reconstruction for treeswithmissing cells (ExtendedDataFig. 8c), the forwardalgorithm
described inMethods was used to generate a full binary tree of three generations, which was then
pruned by removing a given number of randomly chosen endpoint cells. To reconstruct the lineage
relationshipsoftheremainingcells,weassignedeachofthepossible315reconstructionsascoregiven
by 𝐻6a − 𝑇6ab6,acF,where𝐻6a is theHammingdistancebetween thebarcodedscratchpadcollapse
patterns of cells 𝑖 and 𝑗, 𝑇6a is the lineage distance between cells 𝑖 and 𝑗 from that particular
reconstruction(1forsisters,2forfirstcousins,and3forsecondcousins),andthesummationrunsover
all pairs of cells. Matrices 𝐻 and 𝑇 were normalized so that their elements summed to one. This
reconstruction score format was useful not only in scoring the topology of possible lineage
arrangements,butalsoinestimatingbranchlengths(e.g.,whethertwocellswerelikelycloselyrelated
sistersormoredistantlyrelatedcousins).Thereconstructionwiththelowestscorewasselectedasthe
reconstructedtreeandusedtocomputethefractionofrelationshipsthatwerecorrectlyidentified(i.e.,
fraction of correct relationships) shown in the heat-maps. If multiple reconstructions had the same
lowestscore,thefractionofcorrectrelationshipswasaveragedoverallofthem.
3b.SimulationofMEMOIRoperationforsparsetrees
Tosimulatesparse trees,weassumedcollapseevents inagiven lineageoccurata specifiedconstant
collapse rate λ (per cell per generation). The sparse regime occurs when λ < 1. Because generating
hundreds of collapse events requires large trees of many generations, we did not simulate the
underlying proliferating binary tree. Instead, we directly generated a tree of collapse events with
variable branch lengths corresponding to the stochastic time intervals between collapse events.
Assuming a constant rate of collapse (a Poisson process),we randomly drewbranch lengths froman
exponentialdistributionwiththetimescalesetbythecollapserate,i.e.meanbranchlength∼1/λ.At
anygivenlevelofthetree,ifthetreehad𝑁branches,thetimeintervaltothenextbranchingeventwas
selectedfromanexponentialdistributionwithrate𝑁𝜆,andthebranchingeventwasrandomlyassigned
tooneofthe𝑁branches.Thebranchingprocesswascontinueduntilthedesirednumberofleaveswas
generated. ExtendedData Figure 9c,d showsexamplesof sparse treeswith variousnumberof leaves
and tree depths (defined as the cumulative number of collapse events since the common ancestor
experienced by each leaf averaged over all the leaves of the tree). For clarity, the approximate
WWW.NATURE.COM/NATURE | 8
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
correspondence between these sparse trees (characterized by number of leaves and tree depth) and
underlyingfullbinarytrees(characterizedbycollapserateandnumberofgenerations)wasdetermined
bysimulatingfullbinarytreesinthesparseregimeandthenfindingparametersthatgeneratedeffective
sparse treeswith thedesirednumberof leavesanddepth (seethecartoon inExtendedDataFig.9a).
The collapse rate and number of generations of the full binary trees corresponding to the example
sparse trees are reported in Extended Data Figure 9c,d. To reconstruct sparse trees, we used the
generatedcollapsepatternattheleavesandaSankoffmaximumparsimonyalgorithm45.Thefractionof
correct partitions identified by the reconstructed trees was calculated using the Robinson-Foulds
distancemetric40.SimulationswereimplementedinMatlab,andPythonusingtheBiopythonpackage46
andETE3toolkit47.
4. InferenceoftheswitchingratesbetweenthetwoEsrrbgeneexpressionstates
Inthissection,wedescribehowweinferredthetransitionratesbetweenthelowandhighexpression
states of Esrrb. In 4a, we compute the occurrence frequencies of pairs of cells in the same gene
expressionstate.In4b,wepresentastraightforwardinferenceofthetransitionratesfromthedecayin
theobservedfrequencyofpairsofcellsinthesameexpressionstate.Finally,in4c,wepresentamore
involvedbutmoreaccurateinferencealgorithm.
4a.OccurrencefrequencyofrelatedpairsofcellsinthesameEsrrbgeneexpressionstateonMEMOIR
trees
Inthissubsection,wedescribehowtheco-occurrencefrequenciesofpairsofrelatedcellsinthesame
Esrrb expression state can be extracted from theMEMOIR reconstructed lineage trees and endpoint
geneexpressionmeasurements.
Tocomputethefrequencies,wefirstassigneachcellaprobabilityofbeinginoneofthetwoEsrrbgene
expressionstatesbasedonitsEsrrbmRNAcount.Todoso,wefitthebimodaldistributionofsingle-cell
Esrrb transcript counts (Figure 4b)with aweighted sum of two negative binomial distributions, each
correspondingtoonegeneexpressionstate.WedenotethedistributioncorrespondingtotheEsrrblow
expression state as𝑃.(𝑚), and the distribution corresponding to the Esrrb high expression state as
𝑃h(𝑚), where𝑚 is the number of transcripts in an individual cell. The bimodal distribution is fit to
𝑓𝑃. 𝑚 + 1 − 𝑓 𝑃h(𝑚),where𝑓isanadditionalfittingparameterthatcorrespondstothepopulation
fractionofcellsinthelowexpressionstate.
WWW.NATURE.COM/NATURE | 9
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
Withthefitparametersdetermined,theprobabilitythatanindividualcellwith𝑛EsrrbmRNAmolecules
isinthelowexpressionstate(E-)is
𝑝j.(𝑛) =𝑓𝑃. 𝑛
𝑓𝑃. 𝑛 + (1 − 𝑓)𝑃h 𝑛
Similarly,theprobabilitythatthecellisinthehighgeneexpressionstate(E+)issimply𝑝jh 𝑛 = 1 −
𝑝j.(𝑛).
Next,wegeneralize thisapproach to jointlyobserving thestatesofapairofcells. If cell1 is in theE-
statewithprobability𝑝j.- andcell2isintheE-statewithprobability𝑝j.F thentheprobabilitythatboth
cells are simultaneously in E- states is𝑝j.- 𝑝j.F . Similarly, the probability of observing cell 1 in the E+
stateandcell2intheE-stateis𝑝jh- 𝑝j.F .Wecandenotetheprobabilityofobservingthetwocellsinany
ofthefourpossiblepairsofstatesbya2×2matrix,
𝑝j.- 𝑝j.F 𝑝jh- 𝑝j.F
𝑝j.- 𝑝jhF 𝑝jh- 𝑝jhF
The entries in thismatrixmust sum to1. Finally, to compute the occurrence frequency of a pair of
sisters cells ineachof the fourpossiblepairsof states,weaverage theabovematrixoverallpairsof
sisterscells.
𝑪 1 =1
𝑁]6]VSm]𝑝j.6 𝑝j.
a 𝑝jh6 𝑝j.a
𝑝j.6 𝑝jha 𝑝jh6 𝑝jh
a6,a
wherethesummationrunsoverall𝑁]6]VSm]pairsofsisterscellsinthe30coloniesusedintheanalysis.
Wenotethatsincetheorderingofthesisterpairisarbitrary,weturntheresultingmatrixsymmetricby
averagingtheoff-diagonalterms.
𝑪 1 isa2x2matrix forsisters.Theargument ‘1’denotesthefact thatsistersare1generationapart.
Similarly, we define𝑪 2 as the frequency of observing a pair of first-cousin cells in a given pair of
WWW.NATURE.COM/NATURE | 10
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
states. 𝑪(2), is determinedby taking the average over all the observedpairs of first-cousin cells. The
occurrence frequencyof sisters, first cousins,andsecondcousins (𝑪(3)) in thesameEsrrbexpression
state(eitherbothE-,orbothE+)isplottedinFigure4d.
Finally,weestimatedthestatisticalerrorontheobservedfrequenciesintwoways(errorbarsinFigure
4d).First,fromthetotalnumberofobservedpairsofcellsandthecomputedfrequencies,weestimated
thecountingerrorassumingamultinomialdistribution.Thatis,foranobservedfrequency𝑓computed
using𝑁pairsofcells,thestatisticalerrorwasestimatedas 𝑓 1 − 𝑓 𝑁.Tounderstandhowmuchof
the similarity observed between states of closely related cells was just due to chance, we randomly
permutedthelineagerelationshipswithineachcolonyandrecomputedthefrequencies.Thestatistical
errorwasestimatedasthevarianceofthefrequenciescomputedover1000iterationsofsuchrandom
permutations. The first error estimate only accounts for the finite size of the total number of
observations.Thesecondestimateaccountsforcorrelationsbetweenthestatesofcloselyrelatedcells
duetochancewiththenumberofdifferentgeneexpressionstates ineachcolonyheldfixed.Thetwo
methods produced similar estimates reflecting the level intra-colony heterogeneity. Because the two
approaches do not capture independent statistical fluctuations, we combined them by taking the
maximumofthetwoestimates.
4b. Inferenceofswitchingdynamics fromreconstructed lineagetreesandendpointgeneexpression
measurements
Here,wedescribetheinferenceoftransitionratesfromthefrequenciesderivedintheprevioussection.
To do so, we first derive the relationship between transitions rates and the expected probability of
observingapairofrelatedcellsinthesamegeneexpressionstate.
We define a minimal model of stochastic and memoryless transitions between the two expression
states.WeassumethattheprobabilityoftransitioningfromtheE-statetotheE+statepergenerationis
constantanddenoteitas𝑇(𝐸h|𝐸.).TheprobabilitythatacellintheE-statewillretainitsstatefrom
one generation to the next is give by 𝑇 𝐸.|𝐸. = 1 − 𝑇(𝐸h|𝐸.). Similarly, the probability of
transitioningoutoftheE+statepergenerationisdenotedas𝑇(𝐸.|𝐸h).Thetransitionprobabilitiescan
becombinedtogetherintoatransitionratematrix,
WWW.NATURE.COM/NATURE | 11
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
𝑻 =𝑇(𝐸.|𝐸.) 𝑇(𝐸.|𝐸h)𝑇(𝐸h|𝐸.) 𝑇(𝐸h|𝐸h)
First, we present a simple approach for inferring the transition rate matrix from the decay in the
observed frequencies of pairs of cells in the same gene expression state going from first-cousins to
second-cousins.Next,wepresentamore involvedderivationthatalsotakes intoaccountthevalueof
the observed frequencies in addition to how they decay from closely related cells to more distant
relatives.
Wecanrelatethefrequencyofobservingsecond-cousinsinthesamestate(eitherbothinE-orbothE+)
to the frequencyof observing the first-cousins in a particular pair of statesby taking into account all
possiblestatetransitionsgoingfromonegenerationtothenext.Itfollows,
𝑪-- 3 = 1 − 𝑇-F 1 − 𝑇-F 𝑪-- 2 + 2𝑇F- 1 − 𝑇-F 𝑪-F 2 + 𝑇F-𝑇F-𝑪FF(2)
𝑪FF 3 = 1 − 𝑇F- 1 − 𝑇F- 𝑪FF 2 + 2𝑇-F 1 − 𝑇F- 𝑪F- 2 + 𝑇-F𝑇-F𝑪--(2)
Wenumericallysolvedtheabovetwonon-linearequationsfor𝑇-Fand𝑇F- fortheobservedvaluesof
𝑪(2) and𝑪(3) (theobserved frequencieswere first corrected formisclassificationsof the two states
caused by measurement uncertainties, see below). To estimate the statistical error on the inferred
transitionsrates,weproducedsimulatedfrequencymatricesbasedontheestimatedstatisticalerrorof
the frequencies (see previous section).We solved for the transition rates for 10000 iterations of the
simulateddataandcomputedthestandarddeviationovertherates.Wefound𝑇-F = 0.10 ± 0.07and
𝑇F- = 0.04 ± 0.03.
4c.Amoreaccurateinferenceofthetransitionrates
Next, to get a more accurate set of inferred transition rates, we incorporated the values of the
frequencies in addition to their decay over the generations in our calculations. To do so, we first
calculatedthejoint-probabilityoffindingapairofsistercells instates𝑖and𝑗usingthetransitionrate
matrix.
𝑪6a 1 = 𝑻 𝑖 𝑠 𝑻 𝑗 𝑠 𝑝(𝑠)]Jjv,jw
WWW.NATURE.COM/NATURE | 12
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
Thesummation inaboveequationrunsover thetwopossiblestatesof theparentcell.Assumingthat
thepopulationfractionsoftheE-andE+cellsdonotchangesignificantlyduringthethreegenerations
spanned by the experiment, 𝑝(𝐸.) and 𝑝 𝐸h can be set approximately equal to the population
fractions derived from fitting the bimodal Esrrb transcript-count distribution (see previous section).
Namely,𝑝 𝐸. = 𝑓 and𝑝 𝐸h = 1 − 𝑓.We experimentally validated the assumption that theEsrrb
transcript-countdistributiondoesnotchangesignificantlyinourexperimentalconditionsoveraperiod
of48hoursorapproximatelyfourgenerations(seeExtendedDataFig.10).
Tokeepthepopulationfractionsconstant,thenumberofcellstransitioningfromtheE-statetotheE+
statemust be equal to the number transitioning in the reverse direction. Hence,𝑇 𝐸h 𝐸. 𝑝 𝐸. =
𝑇 𝐸. 𝐸h 𝑝(𝐸h).Withthissimplification,
𝑪6a 1 = 𝑻 𝑖 𝑠 𝑻 𝑠 𝑗 𝑝 𝑗 = 𝑝 𝑗 𝑻F(𝑖|𝑗)]Jjv,jw
where𝑻F(𝑖|𝑗) denotes the 𝑖, 𝑗thelementof the squareof the transitionmatrix. Similarly, the joint-
probabilityofobservingapairofcousins,andsecond-cousins instates 𝑖and𝑗 is respectivelyequal to
𝑪6a 2 = 𝑝 𝑗 𝑻x(𝑖|𝑗)and𝑪6a 3 = 𝑝(𝑗)𝑻y(𝑖|𝑗).
To infer the state transition rates, we used the measured frequencies and population fractions and
solvedtheaboveequationsfor𝑻.
However, we first need to introduce a correction to the measured frequencies to account for
misclassificationofexpressionstateswhentheyareprobabilisticallyassigned.Namely, it ismorelikely
to misclassify a cell in the E- state as E+ than vice versa because of the asymmetrical shape of the
distributions𝑃±(𝑚) (seethe fits inFigure4b).Tocorrect formisclassifications,wefirstcomputedthe
probabilityofassigningacellinstate𝑗mistakenlytostate𝑖bysimplyintegratingtheprobabilitythata
a cell with 𝑛 Esrrb transcripts is assigned to state 𝑖 (computed in the previous section) over the
probabilitydistributionofobserving𝑛transcriptsinacellinstate𝑗.
𝑸6a = 𝑓𝑃6 𝑛
𝑓𝑃. 𝑛 + 1 − 𝑓 𝑃h 𝑛𝑃a(𝑛)𝑑𝑛
WWW.NATURE.COM/NATURE | 13
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
Forouranalysis,E-cellsaremistakenlyassignedtoE+cellswithprobability0.12,whereasE+cellsare
mistakenlyassignedtotheE-statewithprobability0.04.
Matrix𝑸effectivelyintroducesapparenttransitionsbetweenthetwostatesatthemeasurementstage.
Theseadditionaltransitionscanbeexplicitlyincludedinthecalculationofthejoint-probabilityoffinding
apairofsistercellsinstates𝑖and𝑗,
𝑪6a 1 = 𝑸𝒊𝒌𝑻 𝑘 𝑠 𝑸𝒋𝒍𝑻 𝑙 𝑠 𝑝 𝑠 =𝒍𝒌]Jjv,jw
𝑸𝒊𝒌𝑻 𝑘 𝑠 𝑻 𝑠 𝑙 𝑸𝒍𝒋𝑻 𝑝 𝑙𝒍𝒌]Jjv,jw
= 𝑸𝒊𝒌𝑻𝟐 𝑘 𝑙 𝑝(𝑙)𝑸𝒍𝒋𝑻
𝒍𝒌
However, theapparent transitions representedby𝑸donotcorrespondtoactual transitionsbetween
theEsrrbexpression states.Rather, they representmisclassificationsdue touncertainties in the state
assignment.Tocorrecttheestimate,weremovedthecontributionsoftheseapparenttransitionstothe
measured frequencies, by applying the inverse of matrix 𝑸 to the frequency matrices, 𝑪 =
𝑸.-𝑪 𝑸Y .-.
Finally,toinferthetransitionratesshowninFigure4d,wesolvedforthetransitionratematrix𝑻using
the corrected frequencymatrix for second-cousins , 𝑪(3).𝑪(3) exhibits a lower statistical error than
𝑪(2)and𝑪(1)becausethenumberofpairsofsecondcousinsislargerthanthenumberofpairsoffirst-
cousins and sisters. To validate the inferred rates, we checked that inferring 𝑻 using the corrected
frequencymatricesforfirst-cousins𝑪(2)andsisters𝑪(1)producedconsistenttransitionrates.Finally,
to estimate the statistical error of the inferred rates we used the estimated statistical error in the
measuredfrequencies(seeprevioussection)andproducedsimulatedfrequencymatrices.Theerroron
theinferredrateswasestimatedasthestandarddeviationoftheratesinferredover10000iterationsof
simulateddata.ResultsareshowninFigure4d.
WWW.NATURE.COM/NATURE | 14
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777
WWW.NATURE.COM/NATURE | 15
SUPPLEMENTARY INFORMATIONRESEARCHdoi:10.1038/nature20777