Statistical Analysis of the Ames Salmonella/Microsome Test Author(s): Barry H.

Margolin, Norman Kaplan, Errol Zeiger Source: Proceedings of the National Academy of Sciences of the United States of America, Vol. 78, No. 6, [Part 2: Biological Sciences] (Jun., 1981), pp. 3779-3783 Published by: National Academy of Sciences Stable URL: http://www.jstor.org/stable/10921 . Accessed: 29/04/2011 19:18
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at . http://www.jstor.org/action/showPublisher?publisherCode=nas. . Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

National Academy of Sciences is collaborating with JSTOR to digitize, preserve and extend access to Proceedings of the National Academy of Sciences of the United States of America.

http://www.jstor.org

Proc.Natl.Acad.Sci. USA Vol. 78, No. 6, pp. 3779-3783, June 1981 Genetics

Statistical analysis of the Ames Salmonella/microsome test

(Ames test/mutagenesis/negative BARRY H. MARGOLIN*, binomial model/statistical modeling/toxicity) NORMAN KAPLAN*, AND ERROL ZEIGERt *Biometry Branch tLaboratory Molecular and of Genetics,National Instituteof Environmental HealthSciences,Research 27709 Triangle Park,NorthCarolina

Communicated GeraldN. Wogan,March17, 1981 by ABSTRACT A family of statisticalmodels for analysisof Ames Salmonella/microsometest data is constructedthat considers mutation and toxicity as competing risks and allows hyper-Poisson variability.These models have a parameter that can be employed as a mutagenicindexbecause it approximatesthe slope at zero dose of a dose-response curve adjustedfor toxicity.A secondparameter quantifiesplate-to-platevariabilityand lends itself to the study of factors affecting internal test reproducibility. The detection of aberrant plate counts is also addressed. This methodology is illustrated with data from a Salmonellatest. The Ames Salmonella/microsome test (1) currently holds a preeminent position among the various tests available to the genetic toxicologist for investigation of a chemical's mutagenicity. Indeed, in many quarters, a positive Salmonella test result is presumptive evidence for carcinogenicity. Although various authors (2-5) have proposed analyses of data from this test, each treatment has been deficient in some way: a Poisson distribution model for the sampling variability of the data is assumed, the possibility of toxicity is disregarded or treated in an ad hoc manner, or the multigenerational aspect of the test is ignored. There is substantial empirical evidence that the Poisson model, nearly universal in its adoption as the sampling distribution of revertants per plate from a Salmonella test, lacks the flexibility to adequately describe the variability in a set of plate counts. This can lead to false conclusions concerning the outcome of the assay. An alternative sampling distribution is proposed that incorporates a measure of experimental reproducibility able to reflect dependence of internal variability among a set of replicate plates on the specific laboratory, technician, strain, or any other factor influencing the results from a single test. Toxicity denotes any genetic or cellular phenomenon precluding reversion to histidine prototrophy or expression of the his' phenotype. Dose-response curves from a test frequently exhibit departures from monotonicity because of toxicity. A biologically motivated family of mathematical models is constructed to relate plate counts, on average, to their respective doses of the test chemical. This family considers both mutation and toxicity to be dose-dependent competing risks that confront each Salmonella bacterium. These models are limited in their abilities to resolve ambiguities caused by a limited understanding of the behavior of a chemical once it is placed on a test plate.

(ii) Eachmicrobeplacedon a particular plate (andits lineage) experiencesthe same environment any other microbe(and as its lineage)platedon the sameplate;thisenvironment, denoted by f, is a function the test compound its dose, the amount of and of histidinepresent,andsundryotherfactors.(Wii) targeted The number microbes of placedon a plate,denotedby No,is largee.g., 108.(iv)The probability, P(6),that any single platedmicrobe,throughits descendants,gives rise to a visiblerevertant colonyis small,say i0-5 to 10-9, andis not alteredby the number of microbesplated. These fourassumptions implythat the numberof revertant coloniesobservedon a plate with environment denotedby ~, a with meanparamX4,has approximately Poissondistribution eter NOP(6) (6)-i.e., for any nonnegativeinteger k, the probabilitythatX, equalsk is given by pr{X,= k}= [NOP()]kexp[-NOP(6]/k!,

[1]

in whichexp[t] = et. For a set of r replicateplates with observed counts Xf1, assumed(e.g., refs. 2, 3, and5) that . .,Xf, it is commonly Xv, these r counts are adequatelydescribedas a randomsample fromEq. 1. This does not followsolely fromthe fourassumpthe tions abovebut rests on an additional assumption: r repliand catedplatessharesufficiently similarenvironments plated are numbersof microbes,so thatthe realizedmeanparameters essentiallyconstantacrossreplicates.If one accepts assumpcan tionsi-iv, the validityof the additional assumption be evaltest uatedempirically througha standard of whetherobserved countsXi1,X2,.k..,X, behaveas a random samplefromEq. 1. To investigatethis, fourlaboratories, withoutbeing told the intendedpurpose,were requestedto produce20 replicate plate with Salmonella strain countsundersolventcontrolconditions TAIOO. strainwas selectedbecauseit is oftencited as the This most sensitiveof the Salmonella strains.The numberof plates was chosento be thatroundnumbernearestthe totalnumber of platesinvolvedin a single strainexperimentconsistingof a zerodose or solventcontrolandfive or sixchemical doses,with three replicateplates per dose; this experimental design was a recentmutagenicity conferencerecommendation The re(7). sultantcounts, orderedby magnitude,togetherwith various summary statistics,are in Table 1. The statisticT, evaluatedfor each dataset is
T=
i=l1

A CRITIQUE OF THE POISSON MODEL

The Poisson distribution has been employed extensively in statistical modeling of microbial data. Its applicability to Salmonella test data appears to be a consequence of the following set of assumptions. (i) All microbes, whether on the same plate or not, behave in a stochastic manner independently of each other. Thepublication costsof thisarticle weredefrayed partby pagecharge in payment.This article must thereforebe hereby marked"advertisement" accordance in with 18 U. S. C. ?1734solelyto indicatethisfact. 3779

(X_ -X)2/X,

[2]

in whichXe = (1/r)EX0.The statisticT, is proportional the to ratioof samplevariance samplemeanandis the usualstatistic to usedto test whethera random sampleof size r exhibitsbehavior consistentwith Eq. 1. If the underlyingdistribution suffiis ciently neara Poissondistribution, then the distribution T, of is well-approximated a x2 distribution with (r - 1) degrees by

3780

Genetics: Margolin al. et

Proc. Natl. Acad. Sci. USA78 (1981) plate-to-plate variability and, thereby, validlyweigh the evidence for mutagenicity. AN ALTERNATIVESAMPLINGDISTRIBUTION It is ourconjecture laboratories that occasionally exhibiting hyper-Poissonvariability a particular in experimentare unable thatdayto maintain nearconstancy across replicates the plate of environment becauseof variability pipettingthe microbes, in the agar,the S9 liverhomogenate, the test compound. and For eachplate, the Poissonmeanwouldthen be a stochastic quantity, havingits own samplingdistribution It followsthat a G. set of replicate platecountsarea random samplefroma mixture of Poissonsrepresentedby
00

Table 1. Twentyreplicatedsolventcontrolplate countswith SalmonellaTA100fromfourlaboratories Laboratory* 1 81, 86 93, 97 99, 104 105, 110 112, 112 113, 114 115, 117 118, 122 131, 135 155, 183 Mean 115.10 Variance 540.62
Te

2 63, 64 65,68 69,70 72,73 75, 80 82, 83 83, 84 84, 85 90, 91 t, t 76.72 81.15 17.98 0.0 0.4

3 74, 77 81, 85 87, 88 89, 90 93, 97 98, 99 102, 103 105, 108 108, 110 111, 124 96.45 163.10 32.13 0.6 0.5

4Ht 168, 171 174, 175 185, 189 190,191 195, 197 198, 198 203, 205 205, 207 210, 214 216, 218 195.45 226.58 22.03 0.1 0.2

4Dt 132, 134 144, 145 145, 145 157,158 158, 161 162, 164 166, 169 174, 177 201, 208 208, 219 166.35 631.29 72.10 1.5 0.7

pr{X = k} = (k!)-lt

Te-TdG(,r).

[3]

89.24 2.7

The distributionG must be that of a nonnegativerandom variableand is probablyunimodal,the targetedmean determiningthe mode if no systematicbias is involved.The distributionG that affords greatestmathematical tractability that is associated with a gammadensity:
g(t;,uc) = [(,c)C'r(c1)]1tC1 exp[-t/,c], [4]

(X 10-2)

SE (c)(x 10-2) 1.1

in which ,u > 0, c > 0. With this choice, Eq. 3 reducesto a negativebinomialdistribution:

pr{Xf = kIA,c}= (k + c1)[A/(A + c-1)]k

ascendingorderwithin each laboratory pairedsolely forconcisenessof presentation. and t Laboratory conducted experiment 4 this twice,oncewith water(4H) and once with dimethylsulfoxide (4D) as the solvent.The same individualdidthe twotests buton different days.All otherlaboratories used water. t Contaminated plate, not counted. Offreedom (8). For 20 replicates, this would imply that Tf exceeds 30.1 once in 20, 36.2 once in 100, and 43.8 once in 1000. Table 1 clearly demonstrates that these data are not Poisson distributedin all laboratories all times. Zeigeret al. (9) and at Vollmar (cited in ref. 10) have previously reported this finding. Although Vollmar stated that a Poisson distribution can be assumed up to about 150 revertants per plate, the data in Table 1 contradict this conclusion. The main practicalimplication of Table 1 is that the variability of any statistic used to quantify or judge the evidence for mutagenicity may be substantially underestimated if one disregards the possibility of hyper-Poisson variability. To the extent that the results in Table 1 typify laboratoryfunctioning, an underestimate of the variance by a factor of 2-4 is not rare. This, in turn, would elevate the real false-positive rate, and possibly even the real false-negative rate, and would yield confidence limits for parameters of interest whose coverage rates are overstated. The degree of elevation of the two error rates is dependent on the statistical procedure used. A later example will demonstrate how variance underestimation affects model parameter estimates. For present illustrative purposes, consider a one-tailed test for mutagenicity based on a statistic that is approximately normally distributed with mean 0 and variance pa-2 under an assumption of no mutagenicity. If this statistic is treated as if its variance is &,2 then the real false-positive rates for this test when nominally 5% and 1% are respectively 12% and 5%if p = 2 and 21%and 12%if p = 4. The possible presence of hyper-Poisson variability underscores the need for replicate plates; only by this means can one properly quantify internal

* Revertantcoloniesper plate listed in

x [c'A/A + c1)]c-',

[5]

in whichk is a nonnegative has integer.Thisdistribution mean ,uandvariance + c,u),whichis greaterthanthe mean.The ,u(l next sectiondiscussesmodeling,uas a function applieddose. of The negativebinomialcan be evaluatedat c = 0 througha limit argument obtainthe Poissondistribution; to hence, large valuesof c are evidence for departures fromPoissonsampling behavior.Asidefromunimodality tractability, gamma the and rests on the historical role of the resultant densityassumption to distribution animportant as alternative the negativebinomial Poissonmodel in microbiology; use for this purposedates its with backat leastto Gossett[Student](11),whowasconcerned countingyeast cells usinga haemacytometer. c The parameter reflectsthe intrinsicprecisionof an experon imenterin producing replicated plateenvironments a given day, despite the varyingtest chemicaldose. Perfectreproducis ibilityof replicatedplate environments representedby c = 0. Variability replicatedplate environments represented is in the by c > 0: the greaterthe variability, largerthe valueof c. likelihoodestimatec of c and Table 1 includesthe maximum its estimatedstandard error(12)for each of the five datasets. variable random Becausethe variance a negativebinomial of is a functionof both ,uandc, one is unableto estimatethe standarderrorof an observation solely from its model-fittedor y for predictedvaluey, somethingthatis obtainable Poissondiswithinan extributeddatasimplyas y'2. If c is nearlyconstant errorfor y can be estimatedby periment, then the standard observation Thus, an individual maybe checked [y(l + 5g)]"2. for empirically aberrance computingits absolutestandardby ized residual:
IY-

l/[!?(1

12.

[6]

An illustration this diagnostic of checkoccurslaterin the text.

et Genetics: Margolin al. MODELING DOSE RESPONSE behavior Salmonella Poissonsampling for The lackof universal of test data is one complication the statistical analysisof such resultsfromthe frequently Additional experiments. complexity of nonmonotonic relationship revertants plateto test chemper that ical dose, a relationship increasesto a peak and then decreases,sometimesto a valuebelow the solventcontrolvalues froma nondecreasing and occasionally zero. This departure to responseis commonlyascribedto microbial toxicity,both genetic and cellular,causedby the test chemical.Were the curtest procedureto include a rently recommendedSalmonella the measured numberof microbes parallel assaythataccurately to succumbing toxicityon the treatedplates, one would then be able to modifystatistical analysesto adjustfor the effectsof is toxicity.However,thisdualexperiment notcommon practice becauseof cost and convenience. An effectiveway to disentanglethe competingrisksof mutationand toxicityin order to drawinferencesregardingthe formeris to model, as NOPD, whereNOis the averagenumber of microbesplacedon a plate, and PD is the probability a that single platedmicrobeyields a revertantcolonywhen exposed to dose D of a test chemical.Haynesand Eckardt havere(13) and cently come to a similarconclusion have proposedmodels theirmodelsareessentially the forPD incorporating of toxicity; form:
PD =

Proc. Natl. Acad. Sci. USA78 (1981)

3781

after his- to his' if it is a memberof the first m generations if of plating,probability PM(O)of mutating it is a memberof the thereafter.Both next k-m generations,and a zero probability haveprobability and auxotrophs prototrophs pTD) of succumbsince of if ingto toxicity theyaremembers the firstt generations thereafter. platingand zero probability behaviorassumedin viii is intendedas a The step-function that to approximation probabilities varyin a comparsimonious mutaplex fashionwith generationtime. Becausemeasurable and tioneffectively ceaseswithhistidinedepletion(ifnotbefore) toxicity(if present)is not necessarilyso because measurable restricted,it has been assumedin viii that m ' k, and m 5 t c oo. In fact, for the single-hit assumptionssubsequently adopted,settingm = k introducesnegligibleerrorin the paestimates.Heret = o is intendedto indicatelong-lastrameter ing toxicity(e.g., the entire test time of 48 or 72 hr). The asabovepermitthe followingmodeling. sumptions that Let PD(r,t) denotethe probability a platedmicrobeexposed to dose D of the test chemicalgives rise to a revertant and is colonywhen mutation presentform generations toxicity define QD(s) to be the probability for t generations.Similarly exposedto doseD become thatanymicrobe its descendents and extinctbecauseof toxicitywithin s generations,countingthe of one. Then uponconsideration platedmicrobeas generation the possiblefirstgenerationevents, one obtainsthe following recursion:
= PD(M,t) PM(D)[1 PT(D)][1 QD(t - 1)] + [1- PM(D)]
_

{1 - exp[-HM(D)]}exp[-HAD)],

[7]

in whichHM(D)and HT(D)are typicallylow-order polynomials and representrespectivelythe averagenumberof mutagenic andtoxic"hits" microbewhen dose D is appliedto a plate. per hit" The simplestandmostcommoncase of Eq. 7 is the "single or linearmodel, where
PD= {1 - exp[-(a + /3D)]}exp[-yD],

x [1 - pWD)]{1 - [1 - PD(m-1,t-1)]2}.

[9]

also The quantityQD(s) satisfiesa recursion.If 1 ' s ' t, then + QD(S)= pT(D) [1 -pTD)]Q'(s-1);
= otherwiseQD(S) 0 if s = 0, or QD(S) = QD(t) if s > t. Useful are valuesof QD(s) QD(l) = PT(D)for all t 2 1, + QD(2) pT(D) [1 - PT(D)]P2 (D) forall t - 2, =

a > O,f : O, y? O.

[8]

Modelsof the formin Eq. 7 have been investigated disand cussedextensivelyin the fieldof radiation mutagenesis may but be inappropriate chemicalmutagenesis.Radiation an infor is stantaneous process;in response to a single burst, microbes either succumbto toxicity,mutatewith regardto a specified genetic trait,or surviveunchanged with regardto thattrait.If a platedmicrobeescapesthe toxiceffectsof radiation mubut tateswith respectto the specifiedtrait,it is essentiallyassured of givingrise to a mutantcolony. For this process,the plated microbes the onlyones to,runthe twinrisksof mutation are and toxicity.In the Salmonella plate assay,however,variousalternativesarepossibledepending uponthe testchemical.Ifit loses its mutagenicand toxiccapabilities rapidly,say in a single microbialgeneration,then the processclosely resemblesthat of on single-burst radiation; the other hand, if it is active for an extendedperiod of time, then the plated microbesand their descendants at riskof both mutationand toxicity.Because are the activityof a chemicalon a plateis not generally known,this uncertainty the arguesforbroadening classof modelsin Eq. 7. To this end, in additionto assumptions and the negative i-iv binomialsamplingassumption, followingare assumed.(v) the The probability platingpreexisting of mutantsor of experiencmutationfromprototrophy auxotrophy negliing forward to is gible. (vi) Expression occurswithinthe firstgeneration following mutagenesis. Sufficient (vii) histidineis placedon eachplate to permitauxotrophic growththroughk generations, where k is typically three to six. (viii) Mutation toxicity stochastand are icallyindependentprocesses.A microbeon a plate with dose D of the test chemicalhas probability pM(D) of mutating from

and min{l, pT(D)/[1-pT(D)]}fort- . BecausePD(r,v) very smallforall r -- m andv < t, Eq. 9 may is be approximated by
QD(Oc) =

PD(r,t)

PM(D)[1- PT(D)][1- QD (t

1)]

+ 2[1 - pM(D)][1 p7D)]PD(m-I-1). If m = k, Eq. 10 admitsthe followingsolution:

PD(m,t)= pM(D)[l - pT(D)]{ , (2[1-pM(D)]
j=0

[10]

X [1 - p7(D)]Y X [1-

_(t-L_j)]}.

[11]

membersof the familyof models Fourmodelsareboundary in Eq. 11 subjectto m ' min (k,t);becausethey straddlethe family,these modelsgive goodevidenceof the variedbehaviors possiblein Eq. 11: = I. Form = k = 1 = t: PD(1,1) pM(D)[1-PT(D)]. This is Eq. 7 with pM(D)
=
=

1 -exp[-HM(D)]

and pT(D)

1 -exp[-HT(D)]. , to= :

II. Forrnm=k=

3782

et Genetics: Margolin al.

= pM(D)[1 pT(D)][l
QD@()] D

Proc.Natl. Acad. Sci. USA78 (1981)

= PM(D) [12TD)]

PD(1,)

~~~[1-p7(D)]

Table3. Parameterestimatesfromnegativebinomialsingle-hit modelsI-V forquinolinedata Model I (m = 1,

t = 1) (X10-8)

i.e., mutationfor one generationbut long-lasting toxicity; [x]+ = max(O,x). III. For m = k = t > 1: PD(k,k)in Eq. 11 admitsno further and are both simplification; mutagenicity toxicity histidinelimitedprocesses. = IV. Form = k > 1, t = 00; PD(k,oo) PD(1, 0o)Yj>- (2[1-pM(D)] but [1-PT(D)])i. Here mutationis histidine-limited, toxicity, if present, is long lasting. For the remainderof the discussion,simplifying single-hit assumptions mutation toxicityareadopted,andm is set for and to k. Theseplus earlierassumptions to the following: lead pM(D) = 1 - exp[-(a + /3D)], a > 0, 83 - 0, and pT(D)= 1 in exp[- yD], y 2 0. The key parameter thiscase is /3, because the chemicalunder study is mutagenicif and only if /3 > 0; it is the only indexof mutagenicity date thatbothderivesfrom to mechanisticmodeling of the Ames test and is adjustedfor toxicity. One distinctionamongthe modelsis immediately apparent. The durationof toxicityon a plate can have a markedimpact on the formofPDand,hence, ,u. Forexample,modelI describes survivalby the exponential decayfunction exp[- 'yD], whereas model II depicts it as the "shoulder-shaped" function [2 exp(yD)]+,reachingzero at a finite dose D = y-fln 2 for y > 0. Thevisibleeffectsof toxicityareusually apparent at high only doses. If the positivedoses are selected to be equispacedon a logarithmic scale,thentheremaybe littleinformation regarding the formof the survivalfunction.As the examplein the next sectionillustrates, modelsI andII mayfit the dataequallywell, asjudgedby the evaluatedlogarithmic likelihoods, the difyet /3 ferencesin their estimatesof the parameter can frequently exceed that which is explainableby samplingvariability.Of greaterconcernare differencesin estimatesof /3that models II and IV can exhibitbecausethe duration mutagenicity of on a plate is unknown.The examplebelow illustrates point. this APPLICATIONAND IMPLICATIONS OF THE MODELS was as Quinoline studiedby W. Speck(unpublished) partof the Environmental Test DevelopmentProgram the at Mutagenesis National Instituteof Environmental HealthSciences.The data in Table2 wereobtained froma testwithSalmonella strain TA98 and inducedrat liver homogenateS9. Dimethylsulfoxide was the solventcontrol,and each dose was replicatedthree times. ModelsI-IV, togetherwiththe single-hit negativebinomial and were fitted to these databy maximum assumptions, likelihood techniques(14).Parameter estimatesof a, /3,y, andc andtheir estimatedstandard errorsare given in Table3. A fifth model (V)thatassumesm = k = 6 and t = oo includedas a further is check on model sensitivity. The fitted values of these five models (Table4) show substantialagreementin fittingthe datadespite the differences in Table Quinoline (TA98, liverhomogenate 2. data rat S9) Revertant colonies increasing with quinoline*
0 15 21 29 10 16 18 21 33 16 26 33 100 27 41 60 333 33 38 41 1000 20 27 42

II (m = 1,
t = x)

III (m = 3,
t = 3)

IV (m = 3,
t = 0)

V (m = 6,
t = x0)

10-9) SE(&)(x 3(x10-10) SE(,)(x 10-11)

i X10-4)

SE(')(x 10-5)
e(x 10-2) SE(6)(x10-2)

20.3 26.1 21.5 62.7 21.2 35.3 5.36 2.98

22.0 26.4 10.7 32.4 5.72 4.01 6.53 3.37

2.99 3.73 2.38 8.17 5.71 15.2 5.45 3.01

3.04 3.74 2.14 5.90 5.14 5.07 6.02 3.20

0.325 0.413 0.318 0.898 3.79 5.61 5.49 3.02

model form and parameter estimates. The five logarithmic likelihoods obtained were -63.20, -64.36, -63.73, -63.86, and -63.34, respectively. Estimates of the mutagenic index 's vary from 2.15 x 10-9 to 3.18 x 10-11. Models I and II differ by a factor of 2 in this estimate, which is solely attributable to differences in their modeling of toxicity. The wide variation in estimates of /3 among models II, IV, and V, all of which assume long-lasting toxicity, can be elucidated. If toxicity is essentially negligible, then because pM(D)is generally very small one can closely approximate Eq. 11 by:
PD(m,t) = (2m - I)pM(D) = (2m - 1){1 - exp[-(a + 83D)]}. [12]

Consequently, it is virtually impossible to distinguish empirically between the models associated with PD(m,t)and PD(l,t), since their fits to the data will be nearly the same. However, ,(m)) their parameter estimates for a and /3, denoted by (&(m), and (&'(),8( )), respectively, will differ by nearly a factor of (2m
1)-i.e., &? and ,') (2m - 1)&(m) (2 - 1) "m). The re-

spective standard errors of these estimates will differ by the same factor. The estimates of ,3 in Table 3 exhibit comparable multipliers in the presence of mild toxicity and moderate sam5 I'v and I _ 34 3v pling variability-e.g., 'II Therefore, in the presence of negligible toxicity and sufficient histidine for more than one generation, if it is assumed that mutagenic activity persists for m* generations but in reality it persists for m generations, then the estimate ,B actually estimates approximately (2' - 1)/(2m - 1) times the real /3. If sufficient histidine for four or five generations is supplied to each plate, then the uncertainty regarding the mutagenic index /3can be nearly an order of magnitude. This is far greater than the uncertainty evidenced by the estimated standarderrorsin Table 3, which must be interpreted as uncertainties conditional on the validity of the corresponding models. This uncertainty will affect any quantitative analysis that attempts to measure mutagenic "potency"because the number of microbes at risk of muTable4. Fittedvalues for quinolinedata Quinoline, ,ugper plate 0 10 33 100 333 1000 I 20.3 22.0 25.6 33.9 45.5 28.4 II 22.0 22.9 25.0 30.8 45.5 29.4 Model III IV 20.9 22.3 25.2 32.6 46.4 28.9 21.3 22.5 25.1 31.9 45.7 29.1 V 20.5 22.1 25.4 33.4 45.5 28.6 Averaged plate counts 21.7 18.3 25.0 42.7 37.3 29.7

* Revertant coloniesperplatearelisted in ascending within order each

dosegroup ofquinoline plate). per Three plates weretested per (pug

dosage.

Genetics: Margolin al. et tation is not known. Nevertheless, it is conceivablethat for certainpurposesone can acceptan orderof magnitude unof one certaintyin estimating/3. Alternatively, can theoretically reducethis uncertainty supplying by only enoughhistidinefor a small numberof generations.For identifyingmutagens,as of opposed to the potentiallymore useful quantification the the test strengthof mutagenicity, naturalone-tailedstatistical in thattreatsS = /3/SE,3 a standard as normalrandom variable the absence of mutagenicity much less affectedby the unis certaintyof the numberof microbesat risk of mutation.The valuesof S for the five modelsconsideredhere are 3.43, 3.30, 2.91, 3.62, and 3.54, all of which indicate that quinolineis mutagenic. The distinctionbetween the negativebinomialand Poisson sampling modelscanbe inferred fromTable3 andthe following valuesof f3 ? SE;,obtainedformodelsI-V, respectively, under Poissondistributionsampling:(2.04 ? 0.363) x 10-9, (9.67 ? 1.69) x 10-10, (2.24 ? 0.466) x 10-'?, (1.97 ? 0.321) x 10-10, and (2.99 ? 0.512) x 10-11.The estimatesof ,l for the dataunderthe Poissonandnegativebinomial quinoline models exhibitsmalldifferences anyone of the five modelsconsidfor ered, but their respectivestandard errorsunder the negative binomial 1.75 modelareapproximately timeslargerthanunder the Poissonmodel. This is intuitivelyreasonable because the two sampling modelsas used here differnot in expectedvalues but in variances. Asmentioned earlier,one canuse Eq. 6 to checkforevidence of aberrant observations that might undulyinfluencethe estimated 3. For the dataon quinoline,modelsI-V all pointto the observed60 revertant coloniesat a dose of 100 Ag per plate as the candidate designation a rogueobservation. values for as The of Eq. 6 for the readingof 60 under each of the five negative binomialmodels are 2.67, 3.03, 2.88, 2.91, and 2.73, respecevidenceforaberrance in tively;there is marginal-to-moderate the readingof 60, dependinguponthe modelemployedin the of analysis.The implication this findingcan be investigated by refittingeach model to the quinolinedataset with the observationof 60 deleted. The new values of , plus or minustheir standard errorsfor the five modelsare (1.55 ? 0.447) x 10-9, (8.58 ? 2.27) x 10-10, (1.78 ? 0.584) x 10-10,(1.67 ? 0.427) x lo-10, and(2.35 ? 0.653) x 10-11, respectively,whichdoes not alterthe conclusion thatquinolineis mutagenic. The major impactof omitting the 60-revertants readingis to reduce c, which reflects lack of fit and lack of reproducibility, both of which are reduced when the most discrepantobservationis omittedfromthe modeling.The valuesof c upondeletionof the observationof 60 are 0.0212, 0.0249, 0.0215, 0.0231, and

Proc. Natl. Acad. Sci. USA78 (1981)

3783

0.0215, respectively.Each model still leads to the conclusion that the sampling distribution not Poisson. is DISCUSSION Themodelsconstructed discussedhere areapproximations and to the biochemistry genetics of the Ames Salmonella and test that ignorepossiblecomplexitiesof pharmacokinetics, threshold phenomena,and expressiondelay. This familyof models providesa methodology analysisof mutagenicity for datathat does not presupposeor exclude either Poissonvariability or toxicityor rely on subjective judgmentsof densityof the backgroundlawnto determinewhethertoxicityis present.The pain rameters these modelsservedifferent purposes.The parameter a can be employed to check the credibilityof solvent controldata vis-a-vishistoricalvalues. The parameter rep,B resents a mutagenicindex that estimates the instantaneous slope at zero dose of the observeddose responseadjustedfor toxicity.Finally, the parameterc serves to quantifyplate-toplatevariability withina test andcanbe used as a meansto implement qualitycontrolfor a test protocolthat is rapidlyapin a proaching level of massproduction manylaboratories. Theassistance K. Risko allnumerical of with computations is gratefully acknowledged.
1. Ames, B. N., McCann, &Yamasaki, (1975)Mutat.Res. 31, J. E. 347-364. 2. Katz,A. J. (1979)Mutat.Res. 64, 61-77. 3. Weinstein,D. & Lewinson,T. M. (1978)Mutat.Res. 51, 433434. 4. Venitt,S. & Crofton-Sleigh, (1979)Mutat.Res. 68, 107-116. C. 5. Stead,A. G., Hasselblad, Creason, P. &Claxton, (1981) V., J. L. Mutat.Res. 85, 13-27. 6. Feller, W. (1967)An Introduction Probability to Theoryand Its Applications (Wiley,New York), 3rd. Ed., Vol. 1. 7. de Serres,F. J. & Shelby, M. D. (1979)Environ.Mutag.1, 8792. 8. Snedecor,G. W. & Cochran,W. G. (1967)StatisticalMethods (IowaStateUniv. Press,Ames, 10), 6th Ed. 9. Zeiger,E., Chhabra, S. & Margolin, H. (1979)Mutat.Res. R. B. 64, 379-389. 10. Seiler, J. P., Mattern,I. E., Green, M. H. L. & Anderson,D. (1980)Mutat.Res. 74, 71-75. 11. Gossett,W. S. [Student](1907)Biometrika 351-360. 5, 12. Fisher,R. A. (1941) Ann. Eugenics11, 182-187. 13. Haynes,R. H. &Eckardt, (1980)in Chemical F. PrinMutagens, ciples and Methods TheirDetection,eds. Hollaender,A. & for de Serres,F. J. (Plenum,New York), Vol. 6, pp. 271-307. 14. Rao, C. R. (1973)LinearStatisticalInferenceand Its Applications (Wiley,New York), Ed. 2nd