Environ. Sci. Technol. 1994, 28, 459-465
Group Contribution Method for Predicting Probability and Rate of Aerobic
Biodegradation
Robert S. Boethling,t Philip H. Howard,’p* William Meyian,* William Stiteler,* Julie Beauman,* and Mestor Tiradot
Office of Pollution Prevention and Toxics 7406, U.S. Environmental Protection Agency, 401 M Street SW,
Washington, D.C. 20460, and Syracuse Research Corporation, Merrill Lane, Syracuse, New York 13210
Two independent training sets were used to develop four
mathematical models for predicting aerobic biodegradability
from chemical structure. All four of the models
are based on multiple regressions against counts of 36
preselected chemicalsubstructuresplus molecularweight.
Two of the models, based on linear and nonlinear
regressions, calculate the probability of rapid biodegradation
and can be used to classify chemicals as rapidly or
not rapidly biodegradable. The training set for these
models consisted of qualitative summary evaluations of
all available experimental data on biodegradability for
295 chemicals. The other two models allow semi-quantitative
prediction of primary andultimate biodegradation
rates using multiple linear regression. The training set
for these models consisted of estimates of primary and
ultimate biodegradation rates for 200 chemicals, gathered
in a survey of 17 biodegradation experts. The two
probability models correctly classified 90% of the chemicals
in their training set, whereas the two survey models
calculated biodegradation rates for the survey chemicals
with R2r 0.7. These four models are intended for use in
chemical screening and in setting priorities for further
review.
Introduction
Chemical scoring systems for identifying substances of
priority concern have proliferated in concert with environmental
legislation. Much of the early impetus for
developing such systems derived from the need to review
Premanufacture Notifications (PMNs) under Section 5
of the Toxic Substances Control Act (TSCA) and by the
TSCA-mandated screening of existing chemicals by the
US.Interagency Testing Committee. However,virtually
every EPA program is now involved in chemical scoring
in one way or another. Examples include reportable
quantity (RQ) adjustment methodology under the Comprehensive
Environmental Response, Compensation, and
Liability Act (CERCLA; Superfund); the Superfund
Hazard Ranking System;the Officeof Pesticide Programs’
Inerts Ranking Program for “inert” components of pesticide
formulations; and methodologies for listing chemicals
on the Toxics Release Inventory (TRI). This list is
by no means exhaustive.
The characteristics of ranking systems vary, but the
majority include explicit consideration of persistence,
bioconcentration potential, and aquatic and human toxicity.
Persistence is primarily a function of biodegradability
for the majority of organic chemicals released to
soiland water. This creates a problem for priority-setting
exercises, because experimental biodegradation data are
typically either lacking entirely or do not exist in a form
that can be easily incorporated into automated screening
methods.
t U. S. Environmental Protection Agency.
t Syracuse Research Corp.
0013-938X/94/0928-0459$04.50/0 0 1994 American Chemical Society
In responding to this need, we first developed a weightof-evidence
procedure (1) for collecting and evaluating
available data, due to the considerable variability of
biodegradation data. The data and evaluations were then
made available on-line in BIODEG, a component of the
Environmental Fate Data Base (2,3) that now contains
information on more than 800 discrete organics. The
records for each chemical in the BIODEG file constitute
a comprehensive assessment of experimental mixedculture
biodegradation data that exist for the chemical;
pure culture data are not included because they offerlittle
insight into environmental biodegradation rates. Each
test result, whether it reflects biochemicaloxygendemand
(BOD),COzproduction, loss of parent, or something else,
is assigned a qualitative descriptor code such as BR
(biodegrades rapidly) or BSA (biodegrades slowly even
with acclimation). Aspects of biodegradation such as
acclimation, microbial toxicity and temperature are considered
in the evaluation process. A reliability code (3,
one test available; 2, two tests available; and 1, three or
more consistenttests available),whichreflects theamount
and consistency of the available data, is also assigned for
each biodegradation summary code. There are summary
evaluation codesforoverallaerobicbiodegradation,aerobic
biodegradation in screening tests, biological treatment
simulations, grab sample tests with soilor water, and field
studies.
Subsequently,thesummaryevaluation codesfor overall
aerobic biodegradation were used to develop two models
for predicting aerobic biodegradability from chemical
substructures (4). Thesemodels, based on multiple linear
and nonlinear regression against counts of 35 preselected
substructures, calculated the probability of rapid biodegradation
and successfullyclassifiedasrapidly or notrapidly
biodegradable 90% of 264 chemicals in the training set
and 27 chemicals in an independent validation set.
Klopman et al. (5)have also used this data set to develop
apredictive modelbasedon computer-automated structure
evaluation (CASE)methodology,and Gombar andEnslein
(6) have described models for subsets (aliphatic and
aromatic chemicals) of the BIODEG data.
Although the above models are based on carefully
evaluated experimental data,their capabilities arelimited
to classification. To provide a consistent set of data for
quantitative modeling and to determine the feasibility of
a biodegradation expert system,we conducted a survey in
1986 in which 22 biodegradation experts were asked to
estimate rates and products of degradation for 50organic
chemicals (7). A screening-level model for predicting
aerobic biodegradability was developed from the survey
data (8),but the usefulness of that data set was limited
by its small size.
In this paper, we describe four new screening-level
biodegradability models. Two of these represent enhancements
to our previously described (4) linear and
nonlinear BIODEG models; i.e., those based on experiEnviron.
Sci.Technol.,Vol. 28, No. 3, 1994 459
mental data. The other two models are based on data
from a new and greatly expanded survey,in which a panel
of 17 experts estimated rates of primary (loss of parent
chemical identity) and ultimate (essentially, conversion
to C02 and water) degradation under aerobic conditions
in aquatic environments for 200chemicals. These models
permit semi-quantitative prediction of aquatic biodegradationrates.
Theindependent variables forallfourmodels
are a revised set of 36 chemical substructures from the
original linear and nonlinear models (4) plus molecular
weight. With the addition of molecular weight, predictions
are possible for all chemicals even if they do not contain
any of the 36 structural fragments. More importantly,
the successful fitting of a single set of chemical substructures
to both the evaluated (BIODEG) and the survey
datasets affirms the importance of these substructures in
estimating biodegradability.
Methods
Biodegradation Database. Summaryevaluation codes
foroverallaerobic biodegradation, used inthe development
ofthelinear andnonlinear BIODEG models, wereretrieved
from BIODEG, the Evaluated Biodegradation Database.
The design and development of this filewere described in
detail in a previous publication (1)and briefly in the
Introduction. BIODEG is a component of the Environmental
Fate Data Base (EFDB),available on-line or in a
PC-compatible format from Syracuse Research Corp.
(contact P. H. Howard for details).
Linear and Nonlinear Models. The basic approach
in the development of the linear and nonlinear BIODEG
models has been described (4). For this exercise, the
training set and the independent validation set from the
earlier work were combined to yield a new training set of
295 chemicals. Because the approach had already been
validated, it was not considered necessary to keep a
separate validation set. This data set consisted of 186
chemicals that received summary evaluations of “biodegrades
rapidly” and 109 chemicals designated “does not
biodegrade rapidly”. An indicator variable was formed
with chemicalsinthe rapid biodegradation category being
assigned a value of 1and chemicals in the slow biodegradation
category being assigned a value of 0. The
indicator variable wasthen used asthedependent variable
in multiple linear and nonlinear regressions against 37
independent variables. With this definition of the dependent
variable, a regression model estimates the probability
that a chemical is in the “biodegrades rapidly”
group.
In our previously described (4) linear and nonlinear
models, counts of 35structural fragments (i.e.,thenumber
oftimes a substructureoccursinthemolecule)constituted
the independent variables. For this study,severalchanges
were made in the set of independent variables used in the
regression analyses. Two new fragments (-CF3 and
unsubstituted phenyl group,-C6H5)were added, and three
fragments were redefined. The latter include the quaternary
carbon and the tertiary alcohol fragments of the
earlier models (4),now eliminated and replaced with a
single fragment (carbon with four single bonds and no
hydrogens),and the unsubstituted linear alkyl chain LC4,
which now can be used only if it is terminal (Le., -CH2CH2CH2CH3).
Finally, molecular weight was added as a
continuous variable, since it is well-known that as molecular
sizeincreases,biodegradability generallydecreases
(9,10). In general, atoms were used only once; that is, if
an atom is part of one fragment, it cannot be part of
another. Table 1liststhesefragmentsandtheir regressionderived
coefficients.
The linear model was defined as
y1= a0 + alf, +ad2+ ...+a3j36+amM, +e; (1)
where Y;isthe probability that chemicaljwill biodegrade
fast, or for the survey models, the primary or ultimate
biodegradation rate, f , is the number of nth substructure
in jth chemical, a0 is the intercept, a, is the regression
coefficient for nth substructure, M, is the molecular
weight,am isthe regression coefficient forM,, and e, isthe
error term (mean value is zero). Regression coefficients
were estimated by the method of least squares, using the
REG procedureofthe PCversion of the StatisticalAnalysis
System (SAS Institute, Cary, NC). Although the assumption
of homogeneousvariance does not hold whenever the
dependent variable is defined as above, the least-squares
method still results in unbiased estimates.
The logistic equation was used as the basis for the
nonlinear model. This model
exp(a, +a,f, +ad2+ ...a3j 3 6 +a,M,)
1+ exp(ao+alfl +aj 2+...a3j36-ta,M,)
Yj= (2)
estimates the probabilities near 0.0 whenever the linear
combination in the exponent takes large negative values;
near 0.5whenever that linear combination is near 0.0; and
closeto 1.0whenever the linear combination takes a large
positive value. Themaximumlikelihood method wasused
for estimating the coefficients for this model rather than
the method of least squares, because the model is not a
linear functionof the unknown coefficients. Theestimates
were obtained by using the CATMOD procedure of PC-
SAS.
For each of the estimated regression coefficients, a
standard error was computed as well as a test statistic for
evaluating the hypothesis that the true population value
is 0.0. The test statistic followed an asymptotic x2
distribution in the case of the maximum likelihood
estimates (nonlinear model) and anF distribution for the
least-squares estimates (linearmodel). A p value wasalso
calculated for each of the test statistics. These p values
and statisticsarenot included inTable 1topreserveclarity,
but are available from the authors.
The standard errors and the test statistics (or their p
values) were used only as an approximate indication of
the contribution of a particular fragment rather than as
a basis for eliminating the fragment from the model. We
took this approach because the objective was not to
determine the most parsimonious subset of fragments for
predicting biodegradation status but to keep the model as
broadly applicable as possible. As a result, there are
collinearitiesamongsomeof thefragmentsthat could affect
the accuracyof some of thep valuescomputed for the test
statistics.
Biodegradation Survey. Information relating to the
purpose, design and implementation of an earlier survey
of expertknowledgehasbeenpublished (7). For thisstudy,
we developed a larger database of biodegradability estimates
by conducting a second survey in which 17experts
evaluated 200organic chemicals (therewere 50 chemicals
in the first survey). Each expert rated the primary and
480 Environ. Sci. Technol., Vol. 28, No. 3, 1994
Table 1. Structural Fragments and Coefficients
fragment or parameter
equation constant
M W
unsubstituted aromatic (53rings)
phosphate ester
cyanideinitrile (CEN)
aldehyde (CHO)
amide (C(=O)N or C(=S)N)
aromatic (C(=O)OH)
ester (C(=O)OC)
aliphatic OH
aliphatic NH2 or NH
aromatic ether
unsubstituted phenyl group (C6H5). - .
aromatic OH
linear C4 terminal alkyl (CH2CH2CH2CH3)
aliphatic sulfonic acid or salt
caibamate
aliphatic (C(=O)OH)
alkyl substituent on aromatic ring
triazine ring
ketone (CC(=O)C)
aromatic F
aromatic I
polycyclic aromatic hydrocarbon (24rings)
N-nitroso (NN=O)
trifluoromethyl (CF3)
aliphaticether
aromatic NO2
azo group (N=N)
aromatic NH2 or NH
aromatic sulfonicacid or salt
tertiary amine
carbon with 4 single bonds and no H
aromatic C1
pyridine ring
aliphatic C1
aromatic Br
aliphatic Br
freq’
295
2
5
5
4
9
24
23
34
13
11
25
46
44
4
4
33
36
5
12
1
2
6
4
1
11
14
2
32
11
10
9
40
18
12
5
5
BIODEG models survey models
linear coeff nonlinear coeff freq‘
0.748
0.319
0.314
0.307
0.285
0.210
0.177
0.174
0.159
0.154
0.132
0.128
0.116
0.108
0.108
0.080
0.073
0.055
0.0095
0.0068
-0.000476
-0.810
-0.759
-0.657
-0.525
-0.520
-0.305
-0.347
-0.242
-0.234
-0.224
-0.205
-0.184
-0.182
-0.155
-0.111
-0.110
-0.046
Number of compounds in the training set containing the fragment.
3.01
-0.0142
7.191
44.409
4.644
7.180
2.691
2.422
4.080
1.118
1.110
2.248
1.799
0.909
1.844
6.833
1.009
0.643
0.577
-5.725
-0.453
-10.532
-10.003
-10.164
-3.259
-5.670
-3.429
-2.509
-8.219
-1.907
-1.028
-2.223
-1.723
-2.016
-1.638
-1.853
-1.678
-4.443
200
1
6
11
5
13
6
25
18
7
11
22
21
26
4
6
10
36
4
10
1
2
2
1
2
16
13
3
23
8
10
32
27
8
14
4
2
primary coeff
3.848
-0.00144
-0.343
0.465
-0.065
0.197
0.205
0.0078
0.229
0.129
0.043
0.077
0.0049
0.040
0.269
0.177
0.194
0.386
-0.069
-0.058
-0.022
0.135
-0.127
-0.702
0.019
-0.274
-0.0097
-0.108
-0.053
-0.108
0.022
-0.288
-0.153
-0.165
-0.019
-0.101
-0.154
0.035
ultimate coeff
3.199
-0.00221
-0.586
0.154
-0.082
0.022
-0.054
0.088
0.140
0.160
0.024
-0.058
0.022
0.056
0.298
0.193
-0.047
0.365
-0.075
-0.246
-0.023
-0.407
-0.045
-0.799
-0.385
-0.513
-0.0087
-0.170
-0.300
-0.135
0.142
-0.255
-0.212
-0.207
-0.214
-0.173
-0.136
0.029
ultimate biodegradability of each chemical on a semiquantitative
scale,which used thetermshours,days,weeks,
months, and longer than months to indicate the approximatetime
they thought would berequired for the process
to proceed to completion. As the measure of central
tendency, wecalculated an arithmetic mean scorefor each
chemicalafter assigningnumericalscorestothe individual
responses as follows: 5 = hours; 4 = days; 3 = weeks; 2 =
months; 1 = longer. The total number of responses for
each chemical often exceeded 17, since many experts
indicated a range of time by marking more than one term.
The 200 survey chemicals covered a very wide range of
structure and molecular weight, and the majority were
multifunctional. In general, chemicals were selected to
be included in thesurvey forthespecificpurpose of testing
hypotheses regarding the effects of certain substructures
onestimated biodegradability. Someexamplesfollow.To
explore postulated negative influences on estimated biodegradability,
50 of the 200 chemicals were halogenated,
17 had nitro groups, 18 had quaternary carbon atoms
(defined as four single bonds to non-hydrogen atoms), 20
had three or more fused rings, and 35 had nitrogencontaining
heterocycles of various types. With respect to
expected positive influences onestimated biodegradability,
56chemicalswere biologically hydrolyzable or postulated
to be so, and 15chemicals had unsubstituted linear alkyl
chains of C4 or larger. Of the 200 chemicals in the survey
and 295 in the experimental (BIODEG) data set, only 20
were common to both sets.
Survey Models. Multiple linear regressions were
performed usingthemean scoresforprimary and ultimate
biodegradation asdependent variables. The independent
variables were the same as those used in the linear and
nonlinear BIODEG models just described; i.e., counts of
36 structural fragments plus molecular weight. The
fragments and regression-derived coefficients are listed
in Table 1. Regression coefficientswere estimated by the
method of least squares, using the REG procedure of PCSAS.
Primary or ultimate biodegradability is calculated
for any chemicalby summing,forallthe fragmentspresent
inthe chemical,the number of times (ifany)eachfragment
occurstimesitscoefficient,and then addingthe summation
to a constant that was determined for the entire training
set, plus the product of the chemical’s molecular weight
and the M, coefficient. A file that lists the 200 survey
chemicals, the predicted primary and ultimate biodegradation
scores,and the chemicals’ CASregistry numbers
is available from the authors.
Results
Biodegradation Survey. Table 2 contains summary
statistics for the 200 survey chemicals and the experts’
responses. Ethylene glycoldiacetate wasjudged to be the
Environ. Sci. Tschnol., Vol. 28, No. 3, 1994 461
Table 2. Summary Statistics for Survey Data
minimum maximum
parameter mean score chemical score chemical
primary 3.52 2.37 pentahromoethylbenzene 4.57 ethylene glycol diacetate
SD,d 0.84 0.51 ethylene glycol diacetate 1.28 Vat Blue 4; ethylenediaminetetrakkis(methylphosphonic
ultimate 2.60 1.44 pentabromoethylhenzene 3.89 ethylene glycol diacetate
SDd' 0.83 0.58 ethylene glycol diacetate;picloram; 1.15 maleic hydrazide
pri-ulta 0.92 0.43 e-caprolactone 1.75 dacthal
Mw 228.6 53.1 acrylonitrile 697.6 tris-2,3-dibromopropylphosphate
GPrimary degradationscore minus ultimate degradation score for the same chemical.
acid)
pentabromoethylbenzene
most easily degraded chemical and pentabromoethylbenzene
the least degradable for both primary and ultimate
degradation. Thehighest and lowestpossiblemean scores
are 5 and 1,respectively, but no such value was observed
for any of the 200 chemicals. This would have required
the unanimous judgment of all 17 experts that biodegradation
would occur either in hours (=5, by definition)
or longer than months (=l,by definition). Using the
standard deviations of the responses for a given chemical
asa measure of agreement or disagreement, unanimity of
judgment was also greatest for these two chemicals. In
contrast, the largest standard deviations were observed
for Vat Blue 4 and ethylenediaminetetrakis(methy1phosphonic
acid) (primary) and maleic hydrazide (ultimate).
On the average, scores for primary and ultimate degradation
for each chemical differed by almost 1unit (0.92)
relative to the ordinal scale used to assign scores to the
experts' estimates. The largest difference (1.75) was
observed for dacthal, a tetrachlorinated herbicide with
two ester functions that were considered to be relatively
easily hydrolyzed.
Biodegradation Models. Coefficients fitted by the
regressions for all four models are listed in Table 1. With
the BIODEG models, the probability of rapid biodegradation
can be predicted by using the linear or nonlinear
coefficients from Table 1 and either eq 1 or eq 2,
respectively. With the survey models, biodegradability
can be predicted using the coefficients for primary or
ultimate biodegradation in Table 1and eq 1. Toillustrate
a typical estimation of biodegradability, we will explain
the calculations necessary for predicting the ultimate
biodegradability of o-phenylphenol (M, = 170)using the
survey model. Using Table 1and eq 1,we have
Yj= equation constant +(4.00221) (M,) +(0.022)
(one unsubstituted phenyl group, C,H,) +(0.056)
(one aromatic OH group) = 3.199 +(-0.3757) +
0.022 +0.056 = 2.90
Themean scoreforthis chemicalfrom the biodegradation
experts was 3.08. The integer 3corresponded to "weeks"
in the tabulation of the individual survey responses.
Performance of the BIODEG models in classifying
chemicals in their training set is summarized in Table 3.
Each model classified chemicals in the training set with
about90% accuracyoverall,butresults wereslightlybetter
for the nonlinear model. With either model, rapidly
degraded chemicals were classifiedmore accurately than
slowlydegraded chemicals. Thedistributions of residuals
from the survey models are shown in Figure 1,along with
the R2values and percentages of residuals 2 *O.l, *0.3,
462 Enviran. Scl. Technol., VoI. 28. No. 3, 1994
Table 3. Performance of Biodegradability Models in
Classifying Chemicals in Their Respective Training Sets
RIODEG models survey models
pammeter iineaP nonlineaP primaryb ultimatec
total correct 2641296 275 295 165 200 167,200
correct total 89.5 93.2 82.5 83.5
% correct, 97.3 97.3 84.9 93.5
fast biodegradation (181/186) (1811186J ,101 119) (101,108)
slow biodegradation (83 109) (94 109) 64/61) (66 92)
5 correct, 76.1 86.2 79.0 71.7
Fast biodegradationis defined an a predicted probability >OS
for being classified a9 a BR cBiodegrades Rapidly). Fast biodeg.
radation is defined as a biodegradahiiity score 23.5 Fast biodeg.
radation i9 defined as a biodegradahilrtyscore >2.5.
40
iiI 1-1 iI /--I t i t
41
25 - --I IE
5 20I
l l /
, . , . , , , , , , ,
.I .0.8 4.8 0.4 0.2 0 0.2 0.4 0.8 0.8 1
RBIidYai
Figure1. DistributionOf residualsfrombiodegradationsurveymodels.
(A. top) Primary degradation model. (6. botiom) Uitimate degradation
model.
and *0.5 (absolute value). The mean residuals (absolute
value) for primary and ultimate degradation for all 200
chemicals were 0.173 and 0.206,respectively. There were
eight chemicalswith residuals 20.6 in absolute value,and
these are listed in Table 4.
Table 4. Poorly Predicted Survey Chemicalsa
predicted
chemical experts model residual
primary degradation
silvex
2,2,4,4,6,8,8-heptamethylnonane
di-tert-butyldicarbonate
ultimate degradation
e-caprolactone
n-decanal
11-cyanoundecanoicacid
hexachlorophene
ethylene glycol diacetate
di-tert-butyldicarbonate
2.82 3.43
2.43 3.06
4.05 3.23
3.70 3.09
3.80 3.17
3.68 3.01
1.77 1.10
3.89 3.16
3.18 2.29
-0.60
-0.63
0.82
0.61
0.63
0.67
0.67
0.73
0.89
a All survey chemicals with residuals (experts predicted minus
model predicted) 2 Jf0.61.
The primary and ultimate survey models calculate a
biodegradability score rather than a probability of rapid
biodegradation, with integers corresponding to the descriptors
(hours,days,weeks,etc.)used in thesurvey.This
makes direct comparison of performance for thetwo types
of models (BIODEGvssurvey)difficult. Onewaytoenable
such a comparison isto evaluateperformance of the survey
models in classifyingchemicals in the survey training set.
Toaccomplish this, we defined rapid primary degradation
as a biodegradability score of 23.5, corresponding to the
descriptor days-weeks. For ultimate degradation, we
defined rapid biodegradation as a biodegradability score
of >2.5,which corresponds toweeks-months. Usingthese
criteria, performance of the primary and ultimate survey
models as classifiers (Table 3) was somewhat below that
observedforthe BIODEG models,with slightly more than
80% of the survey chemicals classified correctly by each
model. As was true for the BIODEG models, rapidly
degraded chemicals were more accurately classified than
slowly degraded chemicals.
Accuracy of Experts’ Estimates. To assess directly
the accuracy of the experts’ biodegradability estimates,
we retrieved and reviewed experimental data forallsurvey
chemicals that also had water grab sample data in the
BIODEG database (11-34). Our assessments of the
literature data for these 13 chemicals with respect to the
approximate length of time required for complete degradation
(defined here as six half-lives) are summarized
in Table 5. For cornparison, the mean survey scores, our
interpretation of them relative to the biodegradability
descriptors used in the survey,and the calculated (model)
values are also presented. It is evident that the experts’
estimates of biodegradabilityin aquaticenvironments were
generally consistent with existing experimental data.
Biodegradability scorescalculated usingthesurveymodels
(lastcolumn ofTable 5) alsotracked wellwith the experts’
estimates, with mean residuals (absolute value) for these
chemicals of 0.16for primary degradation (n=6)and 0.20
for ultimate degradation (n = 7).
Discussion
Our results demonstrate that a single set of chemical
substructures and molecular weight allow an acceptably
accurate prediction of both experimentally determined
biodegradability, as reflected in the BIODEG evaluation
codes, and experts’ estimates of primary and ultimate
biodegradation rates. This finding lends credence to the
notion that these factors are important determinants of
biodegradability. It also validates expert judgment, as
reflected in the survey data and the models based on it.
The models thus derived have been encoded in an IBMcompatible
PC program (Biodegradation Probability
Program, available from Syracuse Research Corp.) that
predicts the probability of rapid biodegradation and the
timerequired forprimary and ultimate degradation. Only
the chemical’s SMILES notation (35)or CAS registry
number is required as input.
Expertjudgment is alsovalidated by direct comparison
ofsurveyscoresto grab samplebiodegradation data (Table
5). Chlorothalonil seems to be an exception, because the
experts predicted that primary degradation would occur
inweeksto months,whereastheexperimentaldata suggest
days to weeks. But this is an unusual situation since,
according to Davies (36),primary degradation is much
faster than anticipated because the nitriles in chlorothalonil
direct nucleophilic attack to the 4 and 6 positions
on the ring. A lesson to be learned from this is that even
the collective wisdom of experts may be in error when
applied to specific chemical structuresand should not be
considered a substitute for adequate testing.
Our previous models (4) included a library of 35
structural fragments in order to ensure that the models
be as broadly applicable as possible. However, no predictions
could be made for chemicals that did not contain
any of these substructures. With the inclusion of the
molecular weight parameter no structures are excluded,
although the reliability of predictions based on molecular
weight alone is probably fairly low except for chemicals
with very low or very high molecular weights. Among the
five new or redefined substructures listed in Table 1,at
least two also have clear mechanistic significance, since
unsubstituted terminal alkyl groups (represented by the
linear C4fragment) and unsubstituted phenyl groupsboth
provide sites for the initiation of well-known biodegradation
pathways (20,37).
The signs of the coefficients for the fragments and
parameters listed in Table 1are generally consistent with
commonly accepted generalizations regarding effects of
chemical structure on biodegradability. For example,
ester, alcohol,and carboxylicacid groups usually enhance
biodegradability (9,IO),and all have positive signs in all
four models. On the other hand, halogens, nitro groups,
and quaternary carbons are assumed to make a chemical
more resistant to degradation, and allhave negative signs.
However,there arealsoa number of fragmentsforwhich
the signs are not the same in the four models. In some
cases the coefficients are small for all four models, which
suggests that the fragment may not be very important in
determining biodegradability. An example is the ketone
fragment. For other fragments, it may be observed that
the signs of the coefficients are often inconsistent where
the BIODEG and survey training sets contained only a
few chemicals with that fragment. Confidence in those
coefficients is therefore low, but could be raised by
additional testing. Examples of fragments for which few
data are available include the aromatic F, N-nitroso, and
aliphatic Br fragments.
Another phenomenon is that the primary and ultimate
coefficients(surveymodels) are sometimes quite different
in magnitude. This is to be expected. In the case of the
aldehyde, amide, and carbamate fragments, for example,
this suggests that these fragments are considered by
expertsto be likelysitesof initialattack,but without major
Environ. Scl. Technol., Vol. 28, No. 3, 1994 463
Table 5. Comparison of Survey Data to Measured Biodegradability for Survey Chemicals with Water Grab Sample Data
survey literature
chemical scoren intb intC nd U or Pe ref modelf
eicosane
dimethylformamide
cumene
propanil
Acid Orange 6
chlorothalonil
acrylonitrile
o-phenylphenol
diphenyl ether
di-2-cyanoethylether
tert-butylbenzene
hexachlorophene
benzanthracene
4.19
4.09
3.68
3.61
3.45
2.39
3.27
3.08
2.79
2.79
2.62
1.77
1.76
<d
5 d
d-wk
d-wk
d-wk
d-wk
d-wk
wk-mo
wk-mo
wk-mo
wk-mo
t m o
tmo
d-wk; <wk; <mo
d
wk
wk; wk;wk; >mo; >mo
>d
wk;wk
mo
wk-mo; mo
mo
>wk; >mo
>d; >d; wk; >wk; >wk;wk-mo; mo;
d; d-wk; d-wk; wk; wk
d-wk
>mo; >mo; >mo; >mo; >mo; >mo
3
1
1
5
1
5
2
1
1
2
1
2
13
P
P
P
P
P
P
U
U
U
U
U
U
U
11-13
14
15
16-18
19
20
21
22
23
21,24
23
25
26-34
3.98
3.94
3.61
3.40
3.47
2.68
3.00
2.90
2.81
2.76
2.72
1.10
1.89
Observed biodegradability scorefrom survey;value given is for either ultimate or primary degradation,depending on the type of literature
data. * Interpretation of the surveyscoreaccordingto the followingscheme (d = days;wk = weeks; mo = months): 24 = 5d; <4 2 3 = d-wk;
<3 2 2 = wk-mo; <2 = tmo. Interpretation of each study in terms of the approximate time required for complete degradation,defined as
six half-livesfor primary degradation and 60-70% of theoretical for ultimate degradation, in natural water grab samples. Number of studies.
e U = ultimate; P = primary. f Predicted primary or ultimate degradation using the appropriate survey model.
influence on rates of ultimate degradation. Conversely,
triazine rings, azo bonds, and pyridine rings, for example,
seem to be viewed as negative for ultimate but not
necessarily primary degradation.
Closeinspection of theresidualsfromthe surveymodels
suggests several ways in which these models could be
improved. In some cases the solution is obvious. For
example, the experts assumed that di-tert-butyldicarbonate
(Table 4) would be readily hydrolyzed, but our
modelslack a carbonate fragment. Alkyl chainsrepresent
amoresubtleproblem. Thirteen compounds in the survey
had linear alkyl chains of C9 or greater, and 10 of these
had positive residuals for both primary and ultimate
degradation. This suggests that long alkyl chains were
viewed by the experts as having a positive impact on
biodegradability, but that the linear C4 terminal alkyl
fragment does not adequately account for this effect. On
the other hand, compounds with cycloalkane rings (six
survey compounds)and aromatic rings with twonitrogens
(Le., pyrazines, pyrimidines, and pyridazines; six survey
compounds) generally had negative residuals, suggesting
that these groups were considered to increase resistance
to biodegradation. It should be noted that both singleand
three-nitrogen aromatics (i.e.,pyridines and triazines)
are already represented by fragments in our models, and
all but one of the coefficients are negative. In spite of
this, we did not add new fragments for cycloalkane or
two-nitrogenheteroaromatic rings, because our approach
was to require that such chemicals also be adequately
represented in the experimental data (BIODEG)training
set, and they were not. Additional testing will probably
be required to establishan adequate database of measured
values. This kind of analysis shows how the models may
be used to identify chemical classes in need of testing.
There is no doubt that the fragment constant approach
to biodegradability modeling that we have taken is
somewhat simplistic and does not, for example, take into
account the possible interactions among fragments in
multifunctional molecules. Nevertheless, the models
described above meet our goal of providing quantitative
or semi-quantitative estimates of biodegradation rate for
use in chemical ranking schemes, in addition to estimates
of probability of rapid biodegradation.
Literature Cited
(1) Howard, P. H.; Hueber, A. E.; Boethling, R. S. Enuiron.
Toxicol. Chem. 1987, 6, 1.
(2) Howard, P. H.; Sage, G. W.; LaMacchia, A.; Colb, A. J.
Chem. Inf. Comput. Sci. 1982, 22, 38.
(3) Howard, P. H.; Hueber, A. E.; Mulesky, B. C.; Crisman, J.
S.;Meylan,W.;Crosbie,E.;Gray,D.A.;Sage,G.W.;Howard,
K. P.; LaMacchia, A.; Boethling, R.; Troast, R. Enuiron.
Toxicol. Chem. 1986, 5, 977.
(4) Howard, P. H.; Boethling, R. S.; Stiteler, W. M.; Meylan,
W. M.; Hueber, A. E.; Beauman, J. A,; Larosche, M. E.
Enuiron. Toxicol. Chem. 1992, 11, 593.
(5) Klopman, G.;Balthasar, D. M.;Rosenkranz, H. S.Enuiron.
Toxicol. Chem. 1993, 12, 231.
(6) Gombar, V. K.; Enslein, K. In Applied Multivariate
Analysis in SARandEnvironmentalStudies;Devillers,J.,
Karcher, W., Eds.; Kluwer: Boston, MA, 1991;pp 377-414.
(7) Boethling, R. S.; Gregg, B.; Frederick, R.; Gabel, N. W.;
Campbell, S.E.; Sabljic, A. Ecotoxicol. Enuiron. Saf. 1989,
18, 252.
(8) Boethling,R. S.;Sabljic,A.Enuiron.Sci.Technol. 1989,23,
672.
(9) Alexander, M. Biotechnol. Bioeng. 1973, 15, 611.
(10) Scow,K. M. InHandbook of ChemicalPropertyEstimation
Methods; Lyman, W.J., Reehl, W. F., Rosenblatt, D. H.,
Eds.; McGraw-Hill: New York, 1982; pp 9-1-9-85.
(11) Bertrand, J. C.; Esteves, J. L.; Mulyono, M.; Mille, G.
Chemosphere 1986, 15,205.
(12) Matsumoto, G. Water Res. 1983, 17, 1803.
(13) Walker, J. D.; Calomiris, J. J.;Herbert, T. L.; Colwell, R.
R. Mar. Biol. 1976, 34, 1.
(14) Dojlido, J. R. Investigations of Biodegradability and
Toxicity of Organic Compounds,Final Report 1975-1979
Environmental Protection Agency: Cincinnati, OH, 1979;
(15) Walker, J. D.; Colwell,R. R. Prog. Water Technol. 1975, 7,
783.
(16) Call,D.J.,Brooke;L.T.;Kent,R.J.;Knuth,M.C.;Anderson,
C.; Moriarty, C. Arch. Enuiron. Contam. Toxicol. 1983,12,
175.
EPA 60012-79-163.
(17) El-Dib, M. A,; Aly, 0.A. Water Res. 1976, 10, 1055.
(18) Paris, D. F.; Rogers, J. E. Appl. Enuiron. Microbiol. 1986,
51, 221.
(19) Michaels, G. B; Lewis, D. L. Enuiron. Toxicol. Chem. 1986,
5, 161.
(20) Gibson, D. T.;Subramanian, V. In Microbial Degradation
of Organic Compounds; Gibson, D. T., Ed.; Dekker: New
York, 1984; pp 181-252.
464 Envlron. Sci. Technol., Vol. 28, No. 3, 1994
(21)Ludzack, F.J.;Schaffer, R. B.; Bloomhuff, R. N.; Ettinger,
M. B. Proc. 13thInd. WasteConf.,Eng.Ext.Bull.,Purdue
Univ.,Engr, Ext. Ser. PP 1958,13th, 297.
Gonsior, S. J.;Bailey, R. E.;Rhinehart, W. L.; Spence, M.
W. J.Agric. Food Chem. 1984,32,593.
Ludzack,F.cJ.;Ettinger,M. B.Eng.Ext.Ser. (PurdueUniv.)
1963,no. 115, 278.
(24)Cherry, A. B.; Gabaccia, A. J.; Senn, H. W. Sewage Ind.
Wastes 1956,28,1137.
(25)Lee,R. F.;Ryan, C.In Microbial DegradationofPollutants
in Marine Environments; Bourquin, A. W., Pritchard, P.
H., Eds.; Environmental Protection Agency: Gulf Breeze,
(26)Gardner, W. S.;Lee, R. F.;Tenore, K. R.; Smith, L. W.
Water,Air, Soil Pollut. 1978,11,339.
(27)Herbes, S.E.;Schwall,L. R. Appl.Environ.Microbiol. 1978,
35, 306.
(28)Herbes, S.E.; Southworth, G. R.; Schaeffer, D. L.; Griest,
W. G.; Maskarinec, M.P. In TheScientificBasisof Toxicity
Assessment;Witschi,H.,Ed.;Elsevier/North-Holland New
York, 1980;pp 113-128.
FL, 1979;EPA-60019-79-012;pp 443-450.
(29)Herbes, S.E.Appl. Enuiron Microbiol. 1981,41,20.
(30)Hinga, K.R.;Pilson, M. E.Q.;Lee,R.F.;Farrington, J. W.;
Tjessem, K.; Davis, A.C. Environ. Sci. Technol. 1980,14,
1136.
(31)Lee, R. F.Proc. Oil Spill Conf.(APIPubl.)1977,4284,611.
(32)Lee,R. F.;Gardner,W. S.;Anderson,J.W.; Blaylock,J.W.;
(33)Lee, R. F.;Ryan, C. Can.J.Fish Aquat. Sci. 1983,40,86.
(34)Roubal, G.; Atlas, R. M. Appl. Environ. Microbiol. 1978,
(35)Weininger, D.J. Chem. Inf. Comput. Sci. 1988,28,31.
(36)Davies, P.E.Bull.Enuiron.Contam.Toxicol.1988,40,405.
(37)Britton, L. N.In Microbial Degradation of Organic Compounds;
Gibson, D.T., Ed.; Dekker: New York, 1984;pp
Barwell-Clarke, J. Environ. Sci. Technol. 1978,12,832.
35,897.
89-129.
Receivedfor review June 3,1993.Revised manuscript received
October 25,1993.Accepted November I, 1993.@
@ Abstract published in Advance ACS Abstracts,December 15,
1993.
Environ. Sci. Technol., Vol. 28, No. 3. lW4 465