Toxicology Mechanisms and Methods, 18:243–250, 2008

Copyright

c Informa Healthcare USA, Inc.
ISSN: 1537-6516 print; 1537-6524 online
DOI: 10.1080/15376510701857189

Combined Computational Metabolite Prediction and

Automated Structure-Based Analysis of Mass
Spectrometric Data
David D. Stranz
Sierra Analytics, Inc., 5815
Stoddard Road, Suite 601,
Modesto, CA 95356, USA
Shichang Miao
Amgen, Inc., Department of
Pharmacokinetics and Drug
Metabolism,1120 Veterans
Blvd., South San Francisco,
CA 94080;
Current address:
ChemoCentryx Inc., 850
Maude Ave., Mountain View,
CA 94043, USA
Scott Campbell and
George Maydwell
Sierra Analytics, Inc., 5815
Stoddard Road, Suite 601,
Modesto, CA 95356, USA
Sean Ekins
ACT LLC, 601 Runnymede Ave,
Jenkintown, PA 19046;
Department of Pharmaceutical
Sciences, University of
Maryland, 20 Penn Street,
Baltimore, MD 21201, USA
Received 21 July 2007;
accepted 28 August 2007.
SE acknowledges GeneGo Inc. Sierra
Analytics acknowledges Rocco
Falchetto of Novartis AG for initial
support for the development of Apex.
Address correspondence to Sean Ekins
M.Sc., Ph.D, D.Sc., ACT LLC, 601
Runnymede Ave, Jenkintown, PA
19046. E-mail: ekinssean@yahoo.com,
sekins@arnoldllc.com
ABSTRACT As high-throughput technologies have developed in the phar-
maceutical industry, the demand for identification of possible metabolites
using predominantly liquid chromatographic/mass spectrometry-mass spec-
trometry/mass spectrometry (LC/MS-MS/MS) for a large number of molecules
in drug discovery has also increased. In parallel, computational technologies
have also been developed to generate predictions for metabolites alongside
methods to predict MS spectra and score the quality of the match with
experimental spectra. The goal of the current study was to generate metabolite
predictions from molecular structure with a software product, MetaDrug. In
vitro microsomal incubations were used to ultimately produce MS data that
could be used to verify the predictions with Apex, which is a new software tool
that can predict the molecular ion spectrum and a fragmentation spectrum,
automating the detailed examination of both MS and MS/MS spectra. For
the test molecule imipramine used to illustrate the combined in vitro/in silico
process proposed, MetaDrug predicts 16 metabolites. Following rat microsomal
incubations with imipramine and analysis of the MSn data using the Apex
software, strong evidence was found for imipramine and five metabolites and
weaker evidence for five additional metabolites. This study suggests a new
approach to streamline MS data analysis using a combination of predictive
computational approaches with software capable of comparing the predicted
metabolite output with empirical data when looking at drug metabolites.
KEYWORDS Apex; Metabolite; Mass Spectroscopy; MetaDrug; QSAR
INTRODUCTION
Research in metabolism is aimed at answering the questions about how the molecule is
metabolized, which enzymes are involved, what are the sites of metabolism, the resulting
metabolites, the rate of metabolism (Austel and Kutter 1983), and whether the metabolites
are biologically active or potentially toxic (Smith and Obach 2005). It is also important to
understand species differences in metabolism and the impact this may have on extrapolation
from preclinical species to man. As there have been increases in the throughput of
experimental in vitro systems using various enzymes, data sets are being generated that
can be used for computational models. As the majority of xenobiotics undergo phase I
metabolism via the cytochrome P450 (P450) enzymes primarily in liver, intestine, and
kidney (Paine et al. 1997), this family of enzymes has been well studied. These enzymes have
therefore been extensively modeled using various computational and structural methods.
The history of methods used for the computational prediction of human drug
metabolism includes several different approaches such as databases, quantitative structure

243
metabolism relationships (QSMRs)/quantitative structure-
activity relationships (QSARs), pharmacophores (de Groot and
Ekins 2002), statistical QSAR approaches (Shen et al. 2003;
Balakin et al. 2004a, 2004b), rule-based approaches, electronic
models (Singh et al. 2003; Korzekwa et al. 2004), homol-
ogy models, and crystal structures with docking approaches
(Jolivette and Ekins 2007). Metabolism knowledge bases (Darvas
1987; Klopman et al. 1994, 1997; Talafous et al. 1994; Boyer
and Zamora 2002; Langowski and Long 2002; Borodina et al.
2003, 2004; Button et al. 2003; Erhardt 2003) combine data
or rules from many different mammalian species and tend to
predict all the metabolic possibilities for a molecule (Nassar and
Talaat 2004), rather than the likely metabolites, thus limiting
their usefulness in drug discovery.
These techniques when used individually have different
levels of success and could be combined to improve predictions,
perhaps by deploying the most appropriate models based on
the molecule neighborhood for the test molecule compared to
training sets. P450–substrate/inhibitor recognition interactions
have been studied extensively and have generally shown the
importance of hydrophobic, hydrogen bonding and ionizable
features for both substrates based on Km data and inhibitors
using Ki , IC50 , and percent inhibition data (Ekins et al.
2001). Molecular models that account for electronic effects
of ligands for P450-mediated metabolism have also been
produced (Jones and Korzekwa 1996; Jones et al. 2002) and
these have combined aliphatic and aromatic oxidation reactions
to generate predictions for metabolic regioselectivities. The
combination of approaches may also balance the strengths and
weaknesses of each approach and hence the introduction of such
hybrid methods is ongoing. For example, a recent technique
called MetaSite (Molecular Discovery, Middlesex, UK) gener-
ates GRID field descriptors used for determining energetically
favorable binding sites on molecules of known structure using
crystal structures or homology models for the P450 enzymes and
the interaction energy descriptors for the molecules evaluated
as substrates (Berellini et al. 2005). A reactivity component is
also used in the MetaSite calculation to produce a probability
for an atom to be metabolized. This approach has been applied
with angiotensin II receptor antagonists predicting the site of
metabolism for CYP2C9 and CYP3A4 (Berellini et al. 2005).
A recent study has expanded the application to CYP2D6,
CYP2C19, and CYP1A2 with 75% to 86% correct predictions
(Cruciani et al. 2005). MetaSite has been compared with docking
of ligands into crystal structures and homology models of
CYP3A4, and MetaSite had 78% overall success compared with
57% for docking (Zhou et al. 2006). MetaSite has a limitation of
not being able to predict the absolute or relative amounts of the
major and minor metabolites as well as the rate of metabolite
formation for a molecule. A second hybrid method, MetaDrug,
includes a manually annotated Oracle database of human
drug metabolism information including xenobiotic reactions,
enzyme substrates, and enzyme inhibitors with kinetic data
(Ekins et al. 2005a; Ekins et al. 2005b; Ekins et al. 2005c).
The MetaDrug database has been used to predict some of
the major metabolic pathways and identify the involvement
of P450s using multiple QSAR models enabling the prediction
of affinity and rate of metabolism for numerous enzymes (Ekins
et al. 2005a, 2006). The user may also upload his or her own
QSAR or QSMR data into the software to offer a further level of
utility. Steps have been undertaken to provide more confidence
in predictions as the QSAR methods also provide Tanimoto
D. D. Stranz et al.
similarity as a measure of similarity to training set molecules.
Structural alerts for likely reactive metabolites (Li 2002; Uetrecht
2003; Williams and Park 2003) are also provided. MetaDrug
has been tested with 66 molecules and captured approximately
79% of first-pass metabolites (Ekins et al. 2006). A more recent
application of the underlying metabolism database has been for
the creation of Kernel-Partial Least Squares (K-PLS) models for
individual metabolic transformations that could be used to rank
suggested metabolites (Embrechts and Ekins 2007).
In the drug discovery process, it is imperative for medicinal
chemists to quickly find out the major metabolic “soft spots”
in the lead molecule structure, so that analog compounds can
be designed and synthesized to address the metabolic stability
problems. The conventional method is to perform in vitro
and/or in vivo metabolic studies and identify the structures
of the metabolites using spectroscopic methods such as LC-
MS/MS and NMR. This is often a costly and time-consuming
process. With considerable emphasis now on increasing the
efficiency of drug discovery, there is interest in using the types
of predictive computational approaches (as described above) in
as many areas as possible, particularly in early drug discovery.
Reliable and rapid metabolite prediction using computational
approaches, if achievable, could become a valuable tool for drug
discovery.
The most powerful tool for analysis of in vivo and in vitro
metabolite samples is mass spectrometry, typically coupled
with liquid chromatographic separation in an online LC/MS
instrument. Mass spectrometry can provide molecular weight
information with high accuracy and sensitivity, and with
tandem MS/MS instruments, information about molecular
structure as well. Confirmation of the likely presence of pre-
dicted metabolites using LC/MS and LC/MS/MS is generally
a straightforward procedure, but requires expert knowledge
in mass spectral interpretation. This manual interpretation is
a bottleneck in the metabolite ID process, severely limiting
throughput. Mass spectrometer manufacturers have responded
to this need by providing some tools for metabolite detection,
but to our knowledge none of them explicitly takes the chemical
structure of the metabolite into account. Thus, there is always
some question where isomers are concerned, and a manual
interpretation is required.
With recent interest in using metabolite prediction software
with mass spectrometry data analysis (Anari et al. 2004; Nassar
and Talaat 2004; Shu et al. 2004; Anari and Baillie 2005; Lin
et al. 2005), we have used MetaDrug along with a second
new software tool, Apex, which can predict the molecular
ion spectrum and a fragmentation spectrum that can in turn
be used to determine which predicted molecules are present
in the analyte. This technology is compatible with all mass
spectrometers and hence the combination of MetaDrug (or
another predictive metabolism tool) and Apex enables the
streamlining of metabolite prediction and mass spectrometry
analysis. In silico metabolite prediction typically generates many
more potential metabolites than are actually observed. This
presents a considerable problem for manual interpretation of
the mass spectral data. The analyst must examine each peak
detected in the liquid chromatogram and compare the mass
spectra under that peak with the spectra expected for the
structures. Quite often, many of the predicted metabolites are
isomers, requiring a detailed examination of both MS and
MS/MS spectra to determine which, if any, are correct. By
automating this tedious process, Apex removes the potential
244
for error. Every structure is compared against every spectrum
in a fast and comprehensive procedure. Predictions containing
tens to thousands of structures are evaluated in a few seconds,
without bias, and metabolites that are actually present in the
sample are easily confirmed. The purpose of the current study
was to illustrate how these independently developed software
tools can be utilized to identify likely metabolites (using
imipramine as an example), and ultimately frees resources for
the identification of other metabolites that are currently not
suggested by available metabolism knowledge bases.
EXPERIMENTAL
Reagents and Chemicals
Rat microsomes were purchased from BD Gentest (Woburn,
MA) and were stored at −80◦ C until use. NADPH and
imipramine were purchased from Sigma (St. Louis, MO) and
Aldrich Chemical Company (Milwaukee, WI), respectively.
In Vitro Incubations
Incubations were conducted in 0.5 mL of 0.1 M phosphate
buffer (pH 7.4) with 10 µM imipramine, 1.0 mg/mL rat liver
microsomes, and 1.3 mM NADPH for 1 h at 37◦ C in open
polypropylene tubes in a shaking water bath. The reaction was
terminated with two volumes of acetonitrile and centrifuged at
4,000 rpm for 20 min. The supernatant was dried down in a
speed vac and the residue was reconstituted in 400 µL water-
methanol (2:1) for LCMS injection.
Instrumentation
The reconstituted sample was examined in the positive
electrospray ionization mode on a Thermo Electron (San
Jose, CA) LCQ instrument, using a data-dependent LC/MSn
acquisition with six scan events in each scan cycle (event
1: MS full scan; events 2 to 3: MS/MS without and with
wide band activation, respectively; Events 4 to 6: MS3 on the
three most abundant MS/MS fragment ions from event 3). In
this experiment, molecular ion (MS) spectra are continuously
recorded during the LC separation. When a molecular ion is
observed with intensity above a predetermined threshold, an
MS/MS fragmentation spectrum of that ion is automatically
generated and recorded (collision energy = 35%; activation Q =
0.25; activation time = 30 ms). Ions observed in the MS/MS
spectrum are then isolated and subjected to further stages of
fragmentation (MS3 ). The result is a combined MS, MS/MS,
and MS3 data set with molecular weight and fragmentation
information on all significant species that can be detected under
the experimental conditions.
For these experiments, 40 µL of sample was injected onto
a Shiseido Capcell PAK C18 AQ 3µ 2 × 100 mm reverse
phase column (Phenomenex, Torrance, CA) with equilibration
at 100% solvent A over 1.5 min, binary gradient elution to 25%
solvent B over 20 min, followed by isocratic elution at 100%
solvent B for 2 min and re-equilibration at 100% solvent A
over 1.5 min for a total run time of 25 min. The flow rate was
400 µL/min throughout. Solvent A was 0.1% formic acid (FA)
in 95% water/acetonitrile (ACN); solvent B was 0.08% FA in
245
ACN. The eluent was ionized in the mass spectrometer using
electrospray ionization. Spectra were recorded in positive ion
mode.
MetaDrug Method for Metabolite
Prediction
The development of MetaDrug (GeneGo Inc, St. Joseph,
MI) has been described in detail previously (Ekins et al. 2005a,
2005d). Approximately 80 human metabolic reaction rules
(Ekins et al. 2006) were included in MetaDrug at the time of
the current study. The molecular structure of imipramine was
used in the mol file format and uploaded into MetaDrug. Using
the metabolite rules coded into the software, likely metabolites
were predicted. These molecules were then exported as an sdf
file.
Data Processing Using Apex
The underlying equivalence between chemical structure,
chemical formula, and molecular weight is fundamental to the
ability of mass spectrometry to distinguish among molecules
based on mass (more precisely, the mass-to-charge ratio, or
m/z, as is measured by the instrument). Apex (Sierra Analytics,
Modesto, CA) is fully structure based and requires a minimum
of user expertise in mass spectral interpretation. The software is
written using the C++ programming language for the Microsoft
Windows operating system. All algorithms for chromatographic
and spectral data processing and structure-spectrum correlation
were developed in-house by Sierra Analytics.
In use, Apex imports the substrate and metabolite structures
contained in the sdf file produced by MetaDrug. From these
structures, MS and MS/MS spectra are predicted for comparison
with the experimental data. The software can take into account
multiple adducts, charge states, oligomers, and neutral losses
when predicting spectral features. For MS/MS, Apex uses heuris-
tic rules to create a straightforward estimate of the fragmentation
of a molecule: from the two-dimensional chemical structure,
find all rings and determine the order of all bonds. Then, for
each acyclic single bond, cleave the molecule into two parts at
that bond, and compute the mass of each part. Additional rules
are applied to predict cross-ring cleavages in saturated rings. The
fragmentation “spectrum” is the set of all such pairs of masses,
where each feature in this “spectrum” has the same intensity.
Since the primary discriminant among candidate structures is
molecular weight, this simulated spectrum is generally sufficient
to distinguish among isomers or isobars by matching one or
more predicted and observed fragment ions.
Next, the experimental MS data set is imported directly from
the MS instrument’s raw format. Apex supports most of the
major vendors and MS data system formats. The spectra are
peak detected and centroided for comparison with the spectra
predicted from the structures. Using a comprehensive scoring
and correlation algorithm, the mass spectral data is searched for
the presence of each of the predicted structures. Each structure
is given a composite score based on the degree of match between
the predicted and experimental spectra. The structures are then
ranked by correlation score. These three basic mass spectral
properties of molecules, their mass, isotopic distribution, and
estimated fragmentation, can be used to derive scores for the
Metabolite Prediction and Mass Spectrometric Data Analysis
TABLE 1 Metabolites of imipramine identified with MetaDrug and Apex
Experimentally
Metabolites Predicted determined Comments

N-demethylation (desipramine) Yes Yes Best match overall; all MS/MS features
matched
Aromatic hydroxylation (e.g., Yes (4 isomers) Yes Moderate score; isomers are
2-hydroxylation, 10-hydroxylation) indistinguishable by MS and MS/MS
Alkyl hydroxylation Yes (3 isomers) Yes High score, but biased by lack of observed
MS/MS fragments
N-oxidation Yes (2 isomers) Yes Oxidation of dimethylamine most likely
2-hydroxy desipramine No∗ N/A
Didesmethylimipramine No∗ N/A
Iminodibenzyl Yes No
∗ These metabolites would be predicted if MetaDrug was used in the sequential mode.
† Apex cannot determine presence of structures de novo. The only structures that can be determined are those imported from the prediction software.

correlation of a molecule against experimental data. Apex uses
each of these properties in several ways to create a set of scores.
Each score is in the range of zero (no match) to one (perfect
match). (1) Similarity score: Using predicted mass and isotope
cluster, compute the “distance” between the predicted features
and those observed in a spectrum. If the spectrum contains
none of the predicted features, the score is zero; if the spectrum
is identical to the prediction, the score is one. A partial match
to the prediction, either by mass or relative isotopic abundance,
yields a score between zero and one. (2) Purity score: Compute
the similarity score, and then weight it by the summed intensity
of the matched spectral features divided by the total intensity
of the spectrum. A high score results when the isotope cluster
is both a good match and is the most intense set of features
in the spectrum; the spectrum is pure for that species. (3)
Molecular ion cluster (MIC) score: Sum the intensities of the
matched spectral features, and then normalize this over all of
the spectra in the data set. (4) MS/MS similarity score: Using the
estimated fragmentation spectrum, match the predicted features
against those observed and weight according to the intensity of
the matched features. The weighting takes into account the
expectation that the major features in the spectrum should be
accounted for by a predicted fragment. (5) In addition, external
signals (such as from an LC/UV or radiochemical detector) can
be incorporated as “scores,” where the signal is normalized to
a 0–1 range. Each of the structure-based scores is computed
for each spectrum in the data set. If a score is plotted for each
spectrum in time order, the result is a score chromatogram. Peaks
appear in the score chromatogram where the underlying spectra
have high score values for that metric. Each individual structure
will have a set of score chromatograms, one for each computed
score.
When a number of potential metabolite structures are
presented to Apex, as is the case with MetaDrug, Apex

FIGURE 1 The integrated in silico/in vitro metabolite prediction protocol using MetaDrug and Apex for predicting and verifying
metabolites with mass spectrometry data.

D. D. Stranz et al. 246

FIGURE 2 Apex identification of imipramine metabolites showing detailed match results for predicted and experimentally observed
metabolites, with the correlation score for each of the potential metabolites.

247 Metabolite Prediction and Mass Spectrometric Data Analysis

computes this set of score chromatograms for every predicted
structure. A typical metabolite prediction result may contain
ten to a hundred or more candidate structures, while a typical
chromatographically separated mass spectrometric data set
might contain a thousand to several thousand individual mass
spectra, and usually several individual scores are combined to
yield the final score for any one molecule. The total number
of score chromatogram points that must be computed is the
product of the number of molecules times the number of
scores times the number of individual mass spectra. This total
will usually be in the hundreds of thousands to millions of
score computations. Apex has been optimized to make this
computational process as fast as possible; a typical calculation
with 100 structures and 1,000 MS and MS/MS spectra takes a
second or two on a 2-GHz Pentium PC.
RESULTS AND DISCUSSION
While there has been some interest in using different metabo-
lite prediction software programs with mass spectrometry data
analysis (Anari et al. 2004; Nassar and Talaat 2004; Shu et al.
2004; Anari and Baillie 2005; Lin et al. 2005), to our knowledge
there has been no detailed discussion of the integration of
such approaches. The current study briefly highlights how two
separate software products could be used to process a molecular
structure, suggest metabolites, and ultimately verify which of
these metabolites are observed from experimental data. While
we would not suggest that our approach is ideal or in fact the
only tool available, we think it is important to bring some of
the difficulties inherent in these approaches to the attention of
the reader.
Application Example
The metabolism of imipramine, a tricyclic antidepressant,
has been characterized extensively in vitro and in vivo in human
and rat (Strandgarden and Gunnarsson 1994; van Breemen et al.
1998). It is known to undergo N-demethylation, hydroxylation,
and N-oxidation. This molecule therefore represents a good
example for evaluating the approach taken for metabolite iden-
tification in the current study. MetaDrug was used to generate
an sdf file containing 15 unique predicted metabolite structures
plus the substrate itself (Table 1). The predicted structures and
experimental MS data were then imported into Apex (Fig. 1).
After selection of appropriate scoring criteria for Apex, the
set of structures were correlated against all of the MS and
MS/MS spectra generated from the rat liver microsomal samples
(Fig. 2). In this example, three scores were computed for each
structure: MS similarity (match between actual and predicted
molecular ion spectrum), isotopic purity (percentage of the
overall spectral intensity due to the structure), and MS/MS
similarity (match between actual and predicted fragmentation
pattern for the structure). After scoring, an overall correlation
for each structure was computed as the product of the three
individual scores and a threshold was applied to eliminate
structures with low correlation. Of the 16 structures, six had high
and five more had moderate correlation scores (Table 1). These
included the known primary metabolites: N-demethylation to
form desipramine, aromatic 2-hydroxylation of imipramine to
form 2-hydroxyimipramine, and imipramine itself. A general
scarcity of fragments in the MS/MS spectra made it difficult to
D. D. Stranz et al.
distinguish among the alkyl-hydroxylated metabolite isomers.
The aromatic-hydroxylated isomers cannot be distinguished due
both to lack of observed fragments and because Apex does
not predict fragmentation across aromatic rings. In both cases,
analysis of bona fide standards of these metabolites would likely
be required for a definitive identification.
CONCLUSION
Software tools for predicting potential metabolites of small
molecule substrates in early drug discovery are important in
guiding lead optimization in drug discovery to produce drug
candidates with desirable metabolic and toxicological proper-
ties. We suggest that the integration of available computational
tools for prediction of metabolites such as MetaDrug with
MS and software such as Apex for combining the outputs of
both represents a method for improving the throughput of
metabolite identification studies and builds on earlier studies
(Anari et al. 2004; Nassar and Talaat 2004; Anari and Baillie
2005). This combined process will likely become a standard
practice in the near future for the pharmaceutical industry. In
fact, we are already seeing the development of tools such as Mass
Frontier that have more predictive capabilities. Future studies
may likely include using additional compounds incubated with
human microsomal or other in vitro systems, extending the
present work performed with rat microsomes. While we can
certainly not claim to have solved the problem of metabolite
prediction, there are many limitations to be aware of. Our
knowledge base or “training set” from which the metabolite
prediction rules have been derived could still be considered
tiny compared to chemistry space. Hence, these methods tend
to overpredict metabolites. The impact of methods to limit the
number of predicted metabolites to those that are most likely
and ranking them (Embrechts and Ekins 2007) will need to
be assessed to provide further confidence in the predictions.
In this previous study K-PLS was used with augmented atom
descriptors to generate models to classify whether a metabolite
is likely to be produced from a particular starting structure alone.
Over 300 molecules, including parent drugs and their primary
and secondary (sequential) metabolites, were used to build
these models corresponding to individual metabolism rules.
Each model was internally validated to assess the capability
to classify other molecules. Receiver operator curve statistics
models for N-dealkylation and O-dealkylation had AUC values
from 0.75 to 0.84 and were able to predict between 61%
and 79% of active molecules upon leave-out testing. Other
efforts at adding prediction confidence or domains such as
Support Vector Machine models for P450s have been described
previously (Ekins et al. 2007). The metabolite prediction
efforts outlined to date and used for validation of software
are relatively simple compared with some of the known
metabolic pathways of drugs and endogenous compounds that
can involve multiple phase I and phase II reactions. The
relative rates of formation of particular metabolites may be
important to predict. Some of these metabolites may even be
inhibitors for other enzymes involved in the same metabolic
pathways and it would be useful to assess or predict this
impact.
As virtually all computational metabolism prediction efforts
have been on human, we should address the considerable
amount of metabolism data for mouse and rat, which may
enable us to understand species differences (Nedelcheva and
248
Gut 1994). Compilation of metabolite rules for different species
will also be important to consider in future software. Four
different rule-based methods have been described in a recent
review and a comparison of the predictive ability of the methods
was limited to seven phenolic compounds (Kulkarni et al.
2005). To be of relevance to the pharmaceutical industry, these
rule-based methods need more extensive testing with diverse
drug-like structures. An in-depth comparison of the different
metabolite prediction algorithms is beyond the scope of this
current study but is also needed in order to ensure the optimal
selection of software components for the integrated in silico/in
vitro metabolite prediction protocol described herein.
REFERENCES
Anari, M. R., and Baillie, T. A. 2005. Bridging cheminformatic metabolite
prediction and tandem mass spectrometry. Drug Discov. Today
10:711–717.
Anari, M. R., Sanchez, R. I., Bakhtiar, R., Franklin, R. B., and Baillie, T. A.
2004. Integration of knowledge-based metabolic predictions with
liquid chromatography data-dependent tandem mass spectrometry
for drug metabolism studies: application to studies on the biotrans-
formation of indinavir. Anal. Chem. 76:823–832.
Austel, V., Kutter, E. 1983. Absorption, Distribution, and Metabolism of
drugs. Quantitative Structure-Activity Relationships of Drugs, ed.
Topliss, J. G., Academic Press, New York, pp. 437–496
Balakin, K. V., Ekins, S., Bugrim, A., Ivanenkov, Y. A., Korolev, D.,
Nikolsky, Y., Ivashchenko, A. A., Savchuk, N. P., and Nikolskaya, T.
2004a. Quantitative structure-metabolism relationship modeling of
the metabolic N-dealkylation rates. Drug Metab. Dispos. 32:1111–
1120.
Balakin, K. V., Ekins, S., Bugrim, A., Ivanenkov, Y. A., Korolev, D.,
Nikolsky, Y., Skorenko, S. A., Ivashchenko, A. A., Savchuk, N.
P., and Nikolskaya, T. 2004b. Kohonen maps for prediction of
binding to human cytochrome P450 3A4. Drug Metab. Dispos.
32:1183–1189.
Berellini, G., Cruciani, G., and Mannhold, R. 2005. Pharmacophore, drug
metabolism, and pharmacokinetics models on non-peptide AT1,
AT2, and AT1/AT2 angiotensin II receptor antagonists. J. Med.
Chem. 48:4389–4399.
Borodina, Y., Rudik, A., Filimonov, D., Kharchevnikova, N., Dmitriev, A.,
Blinova, V., and Poroikov, V. 2004. A new statistical approach to
predicting aromatic hydroxylation sites. Comparison with model-
based approaches. J. Chem. Inf. Comput. Sci. 44:1998–2009.
Borodina, Y., Sadym, A., Filimonov, D., Blinova, V., Dmitriev, A., and
Poroikov, V. 2003. Predicting biotransformation potential from
molecular structure. J. Chem. Inf. Comput. Sci. 43:1636–1646.
Boyer, S., and Zamora, I. 2002. New methods in predictive metabolism.
J. Comp. Aided Mol. Des. 16:403–413.
Button, W. G., Judson, P. N., Long, A., and Vessey, J. D. 2003. Using
absolute and relative reasoning in the prediction of the potential
metabolism of xenobiotics. J. Chem. Inf. Compu. Sci. 43:1371–
1377.
Cruciani, G., Carosati, E., De Boeck, B., Ethirajulu, K., Mackie, C., Howe,
T., and Vianello, R. 2005. MetaSite: understanding metabolism in
human cytochromes from the perspective of the chemist. J. Med.
Chem. 48:6970–6979.
Darvas, F. 1987. MetabolExpert: an expert system for predicting the
metabolism of substances. In QSAR in Environmental Toxicology,
ed. Kaiser, K. L., Reidel Co., Dordrecht, pp. 71–81.
de Groot, M. J., and Ekins, S. 2002. Pharmacophore modeling of
cytochromes P450. Adv. Drug Del. Rev. 54:367–383.
Ekins, S., Andreyev, S., Ryabov, A., Kirillov, E., Rakhmatulin, E. A.,
Sorokina, S., Bugrim, A., and Nikolskaya, T. 2006. A combined
approach to drug metabolism and toxicity assessment. Drug Metab.
Dispos. 34:495–503.
249
Ekins, S., Andreyev, S., Ryabov, A., Kirilov, E., Rakhmatulin, E. A., Bugrim,
A., and Nikolskaya, T. 2005a. Computational prediction of human
drug metabolism. Exp. Opin. Drug Metab. Toxicol. 1:303–324.
Ekins, S., Bugrim, A., Nikolsky, Y., and Nikolskaya, T. 2005b. Systems
Biology: Applications in Drug Discovery. Drug Discovery Handbook,
ed. Gad, S., Wiley, New York, pp. 123–183
Ekins, S., de Groot, M., and Jones, J. P. 2001. Pharmacophore and three
dimensional quantitative structure activity relationship methods
for modeling cytochrome P450 active sites. Drug Metab. Dispos.
29:936–944.
Ekins, S., Embrechts, M. J., Breneman, C. M., Jim, K., Wery J.-P.
2007. Novel applications pf Kernel-partial least squares to mod-
eling a comprehensive array of properties for drug discovery. In
Computational Toxicology: Risk Assessment for Pharmaceutical
and Environmental Chemicals, ed. Ekins, S., Wiley-Interscience,
Hoboken, NJ, pp. 403–432
Ekins, S., Kirillov, E., Rakhmatulin, E., and Nikolskaya, T. 2005c. A novel
method for visualizing nuclear hormone receptor networks relevant
to drug metabolism. Drug Metab. Dispos. 33:474–481.
Ekins, S., Nikolsky, Y., and Nikolskaya, T. 2005d. Techniques: application
of systems biology to absorption, distribution, metabolism, excre-
tion, and toxicity. Trends Pharmacol. Sci. 26:202–209.
Embrechts, M. J., and Ekins, S. 2007. Classification of metabolites with
Kernel-Partial Least Squares (K-PLS). Drug Metab. Dispos. 35:325–
327.
Erhardt, P. W. 2003. A human drug metabolism database: potential roles
in the quantitative predictions of drug metabolism and metabolism-
related drug-drug interactions. Curr. Drug Metab. 4:411–422.
Jolivette, L. J., and Ekins, S. 2007. Methods for predicting human drug
metabolism. Adv. Clin. Chem. 43:131–176.
Jones, J. P., Korzekwa, K. R. 1996. Predicting the Rates and Regioselec-
tivity of Reactions Mediated by the P450 Superfamily, Academic
Press, Inc., New York.
Jones, J. P., Mysinger, M., and Korzekwa, K. R. 2002. Computational
models for cytochrome P450: a predictive electronic model for
aromatic oxidation and hydrogen abstraction. Drug Metab. Dispos.
30:7–12.
Klopman, G., Dimayuga, M., and Talafous, J. 1994. META. 1. A program
for the evaluation of metabolic transformations of chemicals. J.
Chem. Inf. Comput. Sci. 34:1320–1325.
Klopman, G., Tu, M., and Talafous, J. 1997. META. 3. A genetic
algorithm for metabolic transform priorities optimization. J. Chem.
Inf. Comput. Sci. 37:329–334.
Korzekwa, K., Ewing, T. J., Kocher, J. P., Carlson, T. J. 2004. Models
for cytochrome P450-mediated metabolism. In Pharmaceutical
Profiling in Drug Discovery for Lead Selection. Borchardt, R. T, Kerns,
E. H., Lipinski, C. A., Thakker, D. R., and Wang, B., (Eds.). Arlington,
VA: AAPS Press, pp. 69–80
Kulkarni, S. A., Zhu, J., and Blechinger, S. 2005. In silico techniques
for the study and prediction of xenobiotic metabolism: a review.
Xenobiotica 35:955–973.
Langowski, J., and Long, A. 2002. Computer systems for the prediction
of xenobiotic metabolism. Adv. Drug Del. Rev. 54:407–415.
Li, A. P. 2002. A review of the common properties of drugs with idiosyn-
cratic hepatotoxicity and the “multiple determinant hypothesis” for
the manifestation of idiosyncratic drug toxicity. Chem. Biol. Interact.
142:7–23.
Lin, T., Pang, J., Hao, C., Stranz, D., Campbell, S., Mordenti, J., and
Pan, L. 2005. Semi-automated metabolite identification using
information-dependent LC/MS/MS acquisition, in silico metabolite
prediction, and structure based data analysis. Drug Metab. Rev.
37(suppl 2):399.
Nassar, A. E., and Talaat, R. E. 2004. Strategies for dealing with metabolite
elucidation in drug discovery and development. Drug Discov. Today
9:317–327.
Nedelcheva, V., and Gut, I. 1994. P450 in the rat and man: methods
of investigation, substrate specificities and relevance to cancer.
Xenobiotica 24:1151–1175.
Metabolite Prediction and Mass Spectrometric Data Analysis
Paine, M. F., Khalighi, M., Fisher, J. M., Shen, D. D., Kunze, K. L., Marsh,
C. L., Perkins, J. D., and Thummel, K. E. 1997. Characterization
of interintestinal and intraintestinal variations in human CYP3A-
dependent metabolism. J. Pharmacol. Exper. Ther. 283:1552–1562.
Shen, M., Xiao, Y., Golbraikh, A., Gombar, V. K., and Tropsha, A. 2003.
Development and validation of k-nearest neighbour QSPR models
of metabolic stability of drug candidates. J. Med. Chem. 46:3013–
3020.
Shu, Y.-Z., Arora, V., Alberts, J., Belcastro, J., and Skiles, G. 2004. Bridging
metabolite prediction and metabolite identification- the use of
metabolite prediction software for LC/MS data mining of potential
metabolites. Drug Metab. Rev. 36(suppl 1):103.
Singh, S. B., Shen, L. Q., Walker, M. J., and Sheridan, R. P. 2003. A
model for likely sites of CYP3A4-mediated metabolism on drug-like
molecules. J. Med. Chem. 46:1330–1336.
Smith, D. A., and Obach, R. S. 2005. Seeing through the mist: abundance
versus percentage. Commentary on metabolites in safety testing.
Drug Metab. Dispos. 33:1409–1417.
D. D. Stranz et al.
Strandgarden, K., and Gunnarsson, P. O. 1994. Metabolism of
lofepramine and imipramine in liver microsomes from rat and man.
Xenobiotica 24:703–711.
Talafous, J., Sayre, L. M., Mieyal, J. J., and Klopman, G. 1994. META. 2.
A dictionary model of mammalian xenobiotic metabolism. J. Chem.
Inf. Comput. Sci. 34:1326–1333.
Uetrecht, J. 2003. Screening for the potential of a drug candidate to
cause idiosyncratic drug reactions. Drug Discov. Today 8:832–837.
van Breemen, R. B., Nikolic, D., and Bolton, J. L. 1998. Metabolic screening
using on-line ultrafiltration mass spectrometry. Drug Metab. Dispos.
26:85–90.
Williams, D. P., and Park, B. K. 2003. Idiosyncratic toxicity: the role
of toxicophores and bioactivation. Drug Discov. Today 8:1044–
1050.
Zhou, D., Afzelius, L., Grimm, S. W., Andersson, T. B., Zauhar, R. J.,
and Zamora, I. 2006. Comparison of methods for the prediction
of the metabolic sites for CYP3A4-mediated metabolic reactions.
Drug Metab. Dispos. 34:976–983.
250