ARTICLES
Structure Elucidator: A Versatile Expert System for Molecular Structure Elucidation
from 1D and 2D NMR Data and Molecular Fragments
Mikhail E. Elyashberg,§ Kirill A. Blinov,§ Antony J. Williams,*,† Sergey G. Molodtsov,‡
Gary E. Martin,| and Eduard R. Martirosian§
Advanced Chemistry Development, Moscow Department, 6 Akademik Bakulev Street, Moscow 117513,
Russian Federation, Advanced Chemistry Development, Inc., 90 Adelaide Street West, Suite 600,
Toronto, Ontario, M5H 3V9 Canada, Novosibirsk Institute of Organic Chemistry, Siberian Branch of Russian
Academy of Science, Lavrentiev Avenue 9, Novosibirsk 630090, Russia, and Rapid Structure Characterization
Group, Global Research and Development, Pfizer, Kalamazoo, Michigan 49001-0199
Received May 28, 2003
StrucEluc is an expert system that allows the computer-assisted elucidation of chemical structures based on
the inputs of a series of spectral data including 1D and 2D NMR and mass spectra. The system has been
enabled to allow a chemist to utilize fragments stored in a fragment database as well as user-defined fragments
submitted by the chemist in the structure elucidation process. The association of fragments in this way has
been shown to dramatically speed up the process of structure generation from 2D NMR data and has helped
to minimize or eliminate the need for user intervention thereby further enabling the vision of automated
elucidation. The use of fragments has frequently transformed very difficult 2D NMR elucidation challenges
into easily solvable tasks. A strategy to utilize molecular fragments has been developed and optimized
based on specific challenging examples. This strategy will be described here using real world examples.
Experience gained by solving more than 150 structure elucidation problems from a variety of literature
sources is also reviewed in this work.
1. INTRODUCTION
Two generations of the StrucEluc expert system (ES) have
been described previously.1,2 The first generation system,
StrucEluc-1,1 was developed for the structure elucidation of
organic molecules from 1D 13C NMR spectra and was found
to be successful for molecules in the mass range of 150-
300 and containing up to 20-25 skeletal atoms. The second
system (StrucEluc-2)2 is capable of elucidating the chemical
structure of large molecules, up to a mass of 1285 amu to
date and 90 skeletal atoms; typically in the area of natural
product structure elucidation, this task is mainly accom-
plished from 2D NMR spectral data. In general, the system
has been designed to elucidate structures containing up to
250 skeletal atoms.
The capabilities of the StrucEluc-2 system in terms of
general utility as a tool for the structure elucidation of natural
products have been demonstrated previously.3 In this study,
StrucEluc-2 was applied to the structure determination of
60 recently isolated natural products. The experimental data
were obtained from a series of original articles published
mainly in 2000. The strategy of structure elucidation was
discussed in detail.
It should be emphasized that problems described in ref 3
were solved using only heteronuclear (HMQC, HMBC, or
COLOC) and homonuclear (H-H COSY) 2D NMR cor-
relations and without using any additional structural informa-
tion. In this mode of program operating, referred to as the
common mode, the system creates connectivities from the
spectral data and generates all possible structures in ac-
cordance with default settings for the number of intervening
bonds between corresponding skeletal atoms and with atom
properties including the state of hybridization and the
possibility of neighboring heteroatoms being taken into
account.
Further investigations challenged StrucEluc-2 with prob-
lems that could not be solved or proved to be very time-
consuming due to a lack of information in the 2D NMR data.
In these cases it proved necessary to introduce additional
structural information, if available, to facilitate the elucidation
process. In the real world, it is common for a chemist or
spectroscopist faced with the need to elucidate a structure
to have prior knowledge of reaction components in a
synthesis, knowledge of the class of compounds that may
have been isolated, or even hypothetical structures for
validation rather than full elucidation.
The hypothesis that the utilization of molecular fragments
found from the system knowledge base1 or potential sub-
structures proposed by the chemist would be helpful to
circumvent the difficulties was tested. The fragment ap-
proach has been used in a number of first generation expert
* Corresponding author phone: (919)570-0217; fax: (425)790-3749;
e-mail: tony@acdlabs.com.
† Advanced Chemistry Development, Inc.
‡ Siberian Branch of Russian Academy of Science.
§ Advanced Chemistry Development.
| Pfizer.
771J. Chem. Inf. Comput. Sci. 2004, 44, 771-792
10.1021/ci0341060 CCC: $27.50 © 2004 American Chemical Society
Published on Web 03/23/2004

systems. It is based on correlation tables containing sub-
structures and their associated characteristic intervals of
spectral feature, e.g., chemical shift variation (for example,
see refs 4-6). In contrast, systems including both SpecSolv7
and StrucEluc-11 employ databases containing substructures
and their associated 13C NMR subspectra. At present, the
StrucEluc database contains over 215 000 chemical structures
and more than one million substructures. The database
continues to grow as further literature data is added. The
value of including substructures directly into the elucidation
process is that a fragment, considered as a macroatom, can
absorb a significant number of the skeletal atoms and leads
to a reduction in the complexity of the problem. This results
in the acceleration of the structure generation process, which
is typically the most time-consuming stage of the structure
elucidation process.
Nevertheless, even in the case when 2D NMR data are
employed, the utilization of molecular fragments is hampered
by the fact that all carbon atoms existing in a fragment
utilized in solving the problem must be supplied with
chemical shifts. Moreover, the values of these chemical shifts
must be as close as possible to the observed values for the
atoms of the corresponding fragments in the experimental
13C NMR spectrum of the unknown under study. In addition,
the accommodation of one or more fragments within a set
of connectivities derived from the 2D NMR data is a problem
that requires the development of new algorithms. To our
knowledge, no attempts have been made to investigate the
possibility of using molecular fragments for structure genera-
tion from 2D NMR connectivities.
In this work, attention has been focused on methods that
allow the utilization of fragments stored in both the 13C NMR
database (found fragments, FF) and those introduced by the
user (user fragments, UF) in combination with 2D NMR data.
To develop an appropriate strategy and refine the methodol-
ogy for the utilization of fragments, a series of challenging
tasks related to the chemistry of natural products has been
examined.
During this work, more than 150 structural problems were
solved in order to develop and test fragment utilization-based
methods to elucidate the chemical structures. Advantages and
disadvantages inherent to this approach were identified. The
utilization of fragments provides the following main advan-
tages: (1) some problems that cannot be solved in the
common mode are solvable when fragments are utilized in
the process; (2) commonly user intervention or the necessity
of structural assumptions is minimal; and (3) generally the
time for structure generation is drastically reduced when
compared with the common mode.
2. FUNCTIONAL SCHEME OF THE SYSTEM
Many abbreviations will be introduced in this section and
will be used for the sake of brevity and ease of communica-
tion. For reference purposes they have been included at the
end of this manuscript and are listed prior to the references
section. The flowchart represented in Scheme 1 provides an
overview of the system. As shown, the system consists of
three interrelated branches, each having access to the system
database. The knowledge base (KB) of StrucEluc consists
of three components: (1) a library of 215 000 molecular
structures and their associated 13C NMR spectra; (2) a
fragment library (FL) containing more than 1 million
fragments with their corresponding 13C NMR subspectra; and
(3) a Library of Spectrum-to-Structure Correlations (LSC)
comprising the most widespread functional groups and their
accompanying characteristics in both their NMR and IR
spectra. The FL is created from the full molecular structures
stored in the KB using proprietary fragmentation algo-
rithms.
Depending on the initial data available and the complexity
of the molecule being analyzed, the system offers a wide
range of methods for solving a problem. These methods are
outlined briefly below.
2.1. System Operation with 1D NMR Spectra. 2.1.1.
“Standard” Mode. In this mode (see Scheme 2), similar to
that first described in ref 7, a13C NMR spectrum is used as
the main experimental data input for the elucidation process.
The chemical shifts, multiplicity, and intensity (number of
carbon atoms) of signals are specified. Quite frequently, an
experimental 13C NMR spectrum is sufficient to elucidate
the structure of a molecule of up to 20-25 skeletal atoms.
However, IR and 1H NMR spectra, if available as supple-
mentary data, can certainly assist with the solution of the
problem as explained later. In principle, the program is
capable of working without either a molecular mass (M) or
molecular formula (MF).
The program starts with a search of the entire 13C NMR
spectrum in the KB. If the spectrum is found in the database,
the corresponding structure is displayed; otherwise, the search
of the FL is carried out. As a result of the search in the FL,
the program displays L fragments that have corresponding
subspectra that do not contradict the experimental spectrum.
The generated fragments are then ranked in order of
decreasing number of carbon atoms. If 1H NMR and IR
spectra are available, the program can filter the selected
substructures with the help of the LSC library, which can
noticeably reduce the number of selected fragments and move
the “good” fragments to the beginning of the list. The
possibility of merging the fragments by overlapping common
atoms is checked with the intent of assembling these
fragments into a final structure. After the merging of
fragments, the 13C NMR subspectra of the integrated frag-
ment are predicted to help identify whether the fragments
have been merged correctly.
If the result of checking all possible fragment combinations
is the detection of a structure having no free bonds, then
13C NMR spectrum prediction and signal assignment are
performed. Each structure is then verified in accordance with
all available spectral data, general rules of structural chem-
istry, and any constraints imposed by the chemist. If several
structures are generated, then they are ranked in order of
increasing d value which is calculated as the sum of
deviations found for each 13C NMR signal divided by the
total number of signals in the spectrum. If an attempt to
assemble a structure fails, then the program automatically
switches to the “classical” mode of the ES that requires the
availability of the MF. The X-PERT4 system serves as an
example of this “classical” approach.
2.1.2. The “Classical” Approach. The “classical” ap-
proach for the StrucEluc system (see Scheme 2) starts with
a molecular formula calculation and the formation of sets
of fragments. The methods have been described previously
in ref 8. The first l basis fragments, (l ) 10-30), l e L, are
772 J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 ELYASHBERG ET AL.

chosen from the list of selected fragments. They will be the
largest fragments.
The subspectrum of the first basis fragment is compared
with the experimental spectrum. The chemical shifts that
were not originally identified in the experimental spectrum
are associated. The first basis fragment is then combined
with combinations of smaller fragments to provide complete
spectral interpretation. The sets from the second, third, etc.
basis fragments are similarly formed. Each set is then
checked for correspondence with the suggested MF. A
chemist may add sets of fragments and apply appropriate
restrictions common to most “classical” expert systems:
GOODLIST, BADLIST, minimum and maximum ring cycle
sizes, multiplicities of bonds, etc., see refs 4-6. The program
is capable of displaying a “generalized portrait” of the
analyzed molecule. This overview is a distribution of
functional groups included in the Typical Functional Group
Library (TFGL) by the frequency of their occurrence in the
fragments selected from the FL.1 Groups that are completely
absent and those that occur most frequently are correspond-
ingly used to automatically form the BADLIST and
GOODLIST.
All of the generated structures are checked for their
correspondence with the available spectra as well as with
the rules of organic chemistry and stereochemistry. Finally,
to choose the most probable structure, 13C NMR spectra of
Scheme 1. General Scheme of the StrucEluc System
A VERSATILE EXPERT SYSTEM J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 773

candidate structures are predicted, and the structures are
ranked in order of increasing average deviations of the
predicted spectra from the experimental data. The method
that finds the most probable structure is described below in
section 2.2.4. It has been shown that the system operating
in this 1D NMR mode can generally elucidate structures of
medium size, up to 20-25 skeletal atoms.
2.2. Molecular Structure Elucidation from 2D NMR
Spectra. 2.2.1. The “Common” 2D NMR Mode. If an
analyzed molecule is large and is related to newly isolated
complex natural products, it is most probable that both the
“standard” and “classical” approaches will fail to elucidate
a final structure. Experience has shown reliable elucidation
of the structure of large molecules (containing > 20-25
skeletal atoms) is generally impossible without employing
2D NMR data.
A third module of the StrucEluc system (see Scheme 3)
is based on a number of programs developed for elucidating
a molecular structure from a combination of 2D NMR
spectra. The most typical combination providing a basis for
structure determination includes H-H COSY, HSQC/
HMQC, and HMBC. The StrucEluc system presently oper-
ates with the following 2D NMR methods: H-H COSY,
HSQC/HMQC, HMBC, ROESY, NOESY, and INAD-
EQUATE. Other methods can be also used by the system
through a flexible procedure that allows input and processing
of experimental 2D NMR data. In addition to spectral data
the program also needs the molecular formula to proceed in
this mode.
First, the coordinates of the 2D cross-peaks should be
determined and entered into the program. This procedure can
be carried out either automatically or manually. The program
is presently capable of reading data represented in the
primary vendor formats of Varian, Bruker, and JEOL. The
program forms peak tables where signal intensities are also
displayed. These tables are converted into tables indicating
carbon atom connectivities. To provide analysis and editing
of the initial data, all CH3, CH2, CH, and C groups as well
as heteroatoms and hydrogens attached to heteroatoms are
displayed on the screen with their respective 1H and 13C
chemical shifts. The resulting presentation corresponds to a
molecular connectivity diagram (MCD) and an example is
shown in Figure 2. HMBC and COSY connectivities are
shown as subgraphs (fragments) connecting carbon atoms
and/or carbon and nitrogen atoms when the corresponding
1H-1H COSY and 15N HMBC data are available. Different
possible distances between the connected atoms are marked
on the MCDs with specific colors. The program analyzes
the 13C and 1H chemical shifts of the CHn groups (n ) 0-3)
and automatically sets parameters showing the allowed
hybridization and possible heteroatom neighborhood for each
carbon atom. To carry out this procedure, special Atom
Property Correlation Tables (APCT) are derived from
fragments stored in the FL. In the latest version of the
StrucEluc system, the relationship to neighboring heteroat-
oms is marked as “forbidden” (fb), “at least one”, “at least
two”, “at least three”, “four” and “not defined” (nd). The
possible states of atom hybridization are designated as sp3,
sp2, sp, not sp, and “not defined”. Those descriptors for the
carbon atoms allow the system to analyze 2D NMR data
and to efficiently apply restrictions during the process of
structure generation. The chemist can then edit these
parameters using other available information. For example,
if the molecule belongs to the CHNO class and the carbon
atom C(175.5) is marked as sp2 at least two, the system will
only generate O-CdO and N-CdO fragments on the basis
of that atom. The chemist is also offered the opportunity to
draw bonds of any multiplicity between the atoms and to
set some functional groups (for instance CdO, O-CdO,
O-H, etc.). In so doing, the structural information evident
from the 1H NMR and IR spectra can be successfully used.
The lengths of the COSY and HMBC connectivities are taken
as default to be 1 and 1-2 bonds correspondingly, thus
indicating the number of bonds between skeletal atoms. We
call these “standard length” connectivities. Since the con-
nectivity tables contain information that initially will be used
during the structure generation process, they should be
examined for consistency. For these purposes, specific criteria
allowing the detection of the presence of contradictions
between some atom pairs have been employed. As a rule,
the main cause of contradictions is in setting the distances
between intervening nuclei that are shorter than those in the
molecule under investigation. The method of detecting the
Scheme 2. A Flowchart of “Standard” and “Classic” Modes Operating
on the Basis of 1D 13C NMR Spectra
774 J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 ELYASHBERG ET AL.

presence of contradictions and their subsequent removal is
described in ref 9.
The data formed at this stage are used as the input data
for the 2D structure generator developed in this work.
Structures are generated under the constraints determined
from the MCD diagrams and any additional constraints that
may be introduced by the chemist. The structure generator
is based on mathematical algorithms described in refs 10-
12. Generated structures are verified using the approaches
discussed previously: filtering with LSC, GOODLIST,
BADLIST, etc. 13C NMR spectrum prediction is performed
for all structures included in the output file, and the structures
are ranked in order of increasing d value using the method
described in detail in section 2.2.4.
The system can also be used for verification of the
proposed structure and the associated 13C and 1H NMR signal
assignments by validation of all available two-dimensional
NMR spectra. If any contradictions are found, for instance
if the distance between a pair of carbon atoms in a proposed
structure is greater than that postulated from a particular
connectivity, the program displays a textual message detailing
the probable cause(s) of the conflict(s) and showing the
connectivities in graphical form.
2.2.2. Application of Fragments in Combination with
2D NMR Data. As previously reported,3 computer-assisted
structure elucidation using 2D NMR data is quite efficient
for the structures of complex natural compounds. However,
if the structural restrictions imposed by the MCD are not
Scheme 3. A Flowchart Describing the Common and Fragment Mode Approaches Operating on the Basis of 2D MNR Data
A VERSATILE EXPERT SYSTEM J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 775

sufficient for the generation of a reasonable number of
possible structures within an appropriate time, it is to be
expected that the utilization of molecular fragments can
greatly facilitate solution of the problem solution. Commonly
appropriate fragments to aid in the solution of a problem
can be found in the knowledge base. The main advantage
of these fragments is that all fragment carbon atoms are
supplied with the 13C NMR assignments obtained from the
full structures that were used for creation of the fragment
DB.
The first step of the process is a fragment search of the
KB (see Scheme 3). As a result, a set of L found fragments
is selected. The next step is the creation of the MCDs using
the found fragments (FF). For this purpose, either all FFs or
any selected number of them can be incorporated by the
operator. The main idea of the algorithm that implements
this procedure is as follows. Prior to creation of the MCD,
the chemist defines the number of fragments, l (l e L), that
will be used in this procedure and sets an error, E, that
defines the maximum difference allowed between the chemi-
Figure 1. The first 20 fragments obtained in the found fragment list during the elucidation of compound 1.
776 J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 ELYASHBERG ET AL.

cal shifts of the fragment carbons and the corresponding
values observed in the experimental spectrum under study.
It is important to note that both parameters, l and E, are
closely interrelated and choosing the most efficient values
may be a matter of trial and error.
The 13C NMR subspectrum of each fragment is compared
with all experimental chemical shifts. The number of
hydrogen atoms attached to a carbon atom is taken into
account during this process. Consider a fragment contains f
carbon atoms and an arbitrary atom Ci of the fragment has
a chemical shift äi (i ) 1  f) and multiplicity mi. Suppose
that the experimental chemical shifts äi1, äi2,... äiq,... äip meet
conditions jäi - äiqj e E and mi ) miq. Then all possibilities
of substituting the di for the experimental values äi1, äi2,...
äiq,... äip must be verified.
If the conditions jäi - äiqj e E and mi ) miq hold for all
f carbon atoms, then the given fragment is recognized as a
candidate for inclusion in the process of creating the MCD.
If this condition does not hold, then the fragment is excluded
from consideration. It is possible that one experimental
chemical shift äiq can also substitute chemical shifts assigned
to several carbon atoms within the fragment. All rearrange-
ments of the experimental chemical shifts within the corre-
sponding carbon atoms of the fragment should be considered.
The chemical shift distribution of carbon atoms that produce
a conceivable assignment of a given fragment carbon atom
has to be verified. During the verification process, the
program checks whether the carbon atom assignments
correspond to the experimental chemical shift correlations
comprising the skeletal atoms making up the fragment. The
fragments that survive the test are then included in the set
of prospectiVe fragments.
The more skeletal atoms that are “absorbed” by the
fragments the shorter the process of structure elucidation.
With this in mind, the algorithm that combines the prospec-
tive fragments within one molecular connectivity diagram
was developed. To realize this procedure, all possible
combinations of prospective fragments are searched, and only
combinations that are in agreement with the experimental
2D NMR correlations are chosen. The fragment combinations
that pass this examination form a set of prospectiVe fragment
combinations. These fragments are then “projected” onto the
MCDs together with any remaining free atoms. The user can
then visually analyze these MCD diagrams.
The total number of MCDs, nMCD, depends on the
following parameters which are defined by the user: l,
number of found fragments which will be used for the
creation of MCDs (l e L); nf, the minimal number of
fragments that must be present in each MCD; q, the
minimum percentage of all skeletal atoms that must be
absorbed by the fragments present in each MCD.
As noted above, in the general case, the more atoms from
the fragments consumed by the MCD, the greater the
likelihood that the process of structure generation from a
given MCD will be more time efficient.
The speed of structure generation depends on the size of
the molecular fragments. If the number of small fragments
composing the MCD is large enough, then this will speed
up structure generation. Generation is also much faster even
when the MCD is comprised of only a small number of large
fragments. Depending on the size of the molecule being
analyzed and the size of fragments placed at the beginning
of the ranked list of found fragments, the nf value is usually
defined as a number from 1 to 4. The most efficient results
are obtained if q is significant, generally 40-60%.
The conclusion of all further verification procedures is a
check of all produced MCDs for contradictions. The program
offers an option that deletes all MCDs that are recognized
as contradictory. The diagrams remaining after checking can
be used in the structure generation process. The user has
the opportunity to omit the connectivity verification because
contradictory MCDs will in any case be detected and rejected
in the process of structure generation. Moreover, for the
process of structure generation, the user can select one or
more MCDs that are attractive to the user who may have
prior knowledge of a particular structure class or target
structure. To alleviate having to choose a preferable MCD,
they are automatically ranked in order of the increasing
number of free carbons. In this way, it is possible to select
a series of appropriate MCDs, starting from that ranked first.
The number of MCDs produced from a given set of
fragments can be rather large with the nMCD value sometimes
being >1000. To allow editing of a large set of MCDs, we
developed a procedure which allows the transfer of all
changes made in one MCD to all the MCDs contained within
the set. In particular, it is possible to specify options which
transfer atom coordinates, the display zoom factor, atom
properties, manually drawn connectivities, and connectivities
automatically modified during the process of checking the
MCD for the presence of contradictions.9 This procedure
essentially alleviates using prior information in the fragment
mode of structure elucidation.
In the process of analyzing a novel compound, it is entirely
possible that there will be no fragments in the database that
will reduce the magnitude of the challenge. It is natural in
such cases to expect that the introduction of user-defined
fragments may help to form the MCDs. The main qualitative
difference between a found fragment (FF) and a user
fragment (UF) is that the FF already contains carbon atoms
with assigned chemical shifts, while the carbon atoms of the
UF have no carbon chemical shift assignments. Two ways
have been suggested to introduce user fragments into the
program:
Calculate the carbon chemical shifts of the fragment using
the “accurate” method (see section 2.2.4.);
Figure 2. Molecular connectivity diagram created for compound
1. 13C and 1H NMR chemical shifts of carbon and hydrogen atoms
are shown in brackets.
A VERSATILE EXPERT SYSTEM J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 777

Search the KB for fragments that comprise the user
fragment.
It is likely that fragments from at least one of the two
sources would be available for use by the program.
2.2.3. Choice of “E” Value. In the process of MCD
creation from fragments, the E value is of great importance
since it markedly influences the result of applying the
fragments. There are a number of principles governing the
selection of the E value. As a rule, the smaller the value of
E then the smaller the number of molecular connectivity
diagrams, nMCD, created from found fragments. The advan-
tage of a small number of MCDs is of course that it can
reduce the time for structure generation, tg. At the same time,
tg is also a function of the fragment dimensions. Larger
fragments generally shorten the structure generation process.
However, if a fragment is large and correspondingly contains
many assigned carbon atoms, then as a consequence it is
not as likely that all carbon atoms, especially the terminal
ones, of a large fragment will fit the experimental shifts,
thereby satisfying a narrow interval for ( E. The program
automatically sets the E value for terminal atoms equal to
12 ppm to account for this issue.
Large fragments are the most useful but to utilize them in
the structure elucidation process a large E value is necessary.
A large E value can correspondingly, in turn, increase the
nMCD value. The optimal approach would be to set a large
enough E value and select only MCDs containing large
fragments for the structure generation. This principle there-
fore justifies manual (user) or automatic rejection of MCDs
containing small fragments. With testing, it has been shown
that the optimal program parameter controlling the minimum
number of carbon atoms in the fragment used for MCD
creating should be set to a value of 5. Unfortunately, it is
impossible to determine an optimal E that is valid for a
diverse range of problems. The value of E should be
optimized for each task by gradually increasing the E value
starting from 0.5 ppm.
There is one more situation that requires relatively large
E values. Experience has shown (see section 4.1.2) that the
program sometimes accepts incorrect fragments as legitimate
when the E value is small. Obviously if the magnitude of
the error is sufficiently large, the program will select
fragments of increased structural diversity.
2.2.4. Selection of the Preferable Structure. The StrucE-
luc system provides a three-step procedure for identifying
the most probable structure in the output file.
First Step. 13C NMR spectra are predicted for all generated
structures using an incremental method, our so-called “fast”
method, and dF values, the average deviation of an experi-
mental 13C NMR spectrum (or the chemical shifts derived
from corresponding projections of 2D NMR spectra), versus
predicted chemical shifts, are calculated.
Second Step. Duplicate structures in the output file are
deleted. Among the generated structures there are usually
duplicates that differ from each other only in terms of the
assignment of the chemical shifts to different carbon atoms.
If this possibility is not appropriately considered when
deleting isomorphic structures, then the structure with the
correct assignment of the chemical shifts could conceivably
be the deleted isomorphic structure. To avoid this eventuality,
the system executes a special procedure for duplicate
removal. For each duplicate family only the structure that
has the minimum dF value is retained in the file as “the best
representative” of the family. After duplicates are removed,
the structures are then ranked by the dF value and sorted in
ascending order. The smallest dF value indicates the best
match between the experimental and calculated spectra, and
this structure will therefore be the first in the list. Experience
shows that for the fast calculation of 13C NMR spectra and
their subsequent ranking places usually the correct structure
is the first or second one in the list. Only in rare instances
will the correct structure be listed below fifth place. Such a
preliminary ranking of the big resulting files can help to reject
hundreds and thousands of structures that are known to be
unsuitable.
Third Step. During the third stage, more accurate 13C NMR
spectra are calculated for the first 10-25 structures of the
ranked file. Accurate predictions are performed using a
database of fragments with their corresponding assigned
subspectra. The description of each nuclear environment is
defined using the HOSE code approach13 (Hierarchical
Ordering of Spherical Environments). The average deviation
values between the experimental and calculated values (dA)
are found, and the structures are again rank-ordered. Sub-
sequent ranking dramatically increases the probability of
moving the correct structure to the first position in the list.
For additional control over the correct choice of the output
structure, the accurate proton chemical shifts can be predicted
and displayed together with the corresponding deviation
value dH. If the experimental 1H NMR spectrum is not
recorded, the chemical shifts are automatically found from
2D NMR projections. For proton NMR prediction, the
predicted proton-proton couplings are enhanced by three-
dimensional optimization of the structure. The position of
the correct structure in the file determines its rank depending
on the type of ranking parameter, i.e., dA, dF, or dH
correspondingly. The “rates” of the correct structure in the
ranked file are denoted as rA, rF, and rH. If the correct
structure is the first in the list ranked by dA values, then rA
) 1. As a rule, the final structural ranking is carried out
according to increasing dA values, while magnitudes of the
dF and dH parameters serve as secondary aids for estimating
reliability of the correct structure selection. When the
structures at the beginning of the ranked file possess different
structural elements, prediction of the MS match factor (mi,
where i is the position of a structure in the ranked file) may
also be useful for the confirmation of the preferable structure.
The system utilizes a routine that is capable of calculating
the percentage of peaks in the experimental MS spectrum
that can be interpreted on the basis of a given structure. The
calculation of the MS match factor is relatively time-
consuming, so it is used only in cases when the difference
¢(2-1) ) dA(2) - dA(1) is small. Here dA(1) and dA(2)
represent the deviations corresponding to the first and second
structures in the ranked file.
In ambiguous cases, it may be useful to display the
calculated 1H NMR spectra because of the complexity of
some multiplet patterns. Also, to facilitate structure analysis
in the output file, the StrucEluc system is supplied with a
feature that calculates structural similarity coefficients.14 In
this way, if the investigator has an idea of the class of
structure under investigation this structure can be used as
an input to allow rank ordering relative to the structural
similarity of the results file.
778 J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 ELYASHBERG ET AL.

3. EXAMPLES OF APPLYING DIFFERENT SYSTEM
MODES
To demonstrate the main features of the approaches
suggested above, a number of examples will be cited.
Consider the elucidation of the chemical structure from the
1D and 2D NMR spectra of a rather simple molecule, 2,4,4′-
trihydroxydihydrochalcone (1), recently isolated and char-
acterized by Gonzales and coauthors.15 This example allows
demonstration of all methods of structure elucidation cur-
rently available in StrucEluc. A molecular formula of
C15H14O4 was deduced from the high-resolution MS spec-
trum. The NMR spectral data (1H and 13C NMR, DEPT,
HMBC) obtained in ref 15 and used here are represented in
Table 1. For the explanations of the analysis to be clear, the
target structures of the examples described in this article will
be displayed, if necessary, with already assigned carbon
chemical shifts.
3.1. Examination of 1D NMR Spectra. 3.1.1. “Standard”
Mode. When the Search Fragments by CNMR Spectrum
command was performed (see Scheme 2), the program
picked out L ) 180 fragments and ranked them according
to decreasing molecular formula. The first 20 fragments from
the list are shown in Figure 1. Note that all carbon atoms of
the fragments stored in the database are supplied with 13C
NMR assignments taken from the full structures that were
used to create the fragment DB.
The comparison of fragments with structure 1 shows that
fragments #1 and #20 exist in the molecule under investiga-
tion and have one overlapping group which is a CH2.
Standard structure generation accompanied by structure
filtering with spectral libraries was performed. This is the
application of the first system branch. The program combined
the found fragments using at least one oVerlapping atom.
As a result, the program produced only one structure
coinciding with compound 1 with a structure generation time
of tg ) 45 s (in this article, all calculations are performed
on an PC Intel Celeron operating at 500 MHz, Windows
98, RAM 128 Mb). The single structure was actually
generated in 4 s, while the remaining time was consumed
by attempts to assemble other conceivable structures from
the first 30 fragments, this number being the default for the
assembly process. A single correct structure was obtained
from the 1D 13C NMR spectrum in fully automatic mode.
3.1.2. “Classical” Mode. An attempt was made to apply
the second system branch to the data utilizing the algorithms
in the so-called “classical” generation mode (see Scheme 2)
based on combining nonoVerlapping fragments and free
atoms. According to the methodology described previously,
the program automatically selected the first 20 found
fragments (l ) 20 is the default setting in this case) from
the list and then created 223 fragment sets. Classical structure
generation was performed, and the spectral libraries and
internal BADLIST were used for structure filtering during
the generation process. As a result, the number of structures
generated, k, gave 112 resultant molecules, while the
generation time, tg, was 1 min 30 s. After the removal of
identical structures, the number of structures was reduced
to 73. This is denoted by the representation of k ) 112 f
73 and will be used throughout this article. The 13C NMR
prediction and structure ranking allowed the program to
reliably distinguish the correct structure as the first structure.
The difference between the first and second structure is given
by the difference between dA(1) ) 1.64 ppm and dA(2) )
3.98 ppm which gives ¢(2-1) ) dA(2) - dA(1) ) 2.34 ppm.
Further details regarding the process of evaluation of the
structure reliability will be discussed in section 7.
3.2. Application of 2D NMR Data Analysis. 3.2.1. The
“Common” 2D NMR Mode. The possibility of elucidating
an unknown structure from the 2D NMR data in an
automated or semiautomated manner is certainly attractive.
In this case, the user would not need to make any assump-
tions that would serve as constraints for the system. However,
to demonstrate the role of different structural constraints the
two common mode methods (see Scheme 3) of solving a
given problem will be considered: without user intervention
and with the influence of the user by inclusion of additional
information.
3.2.1.1. Solution Without Any User Assumptions. As
described earlier, the program automatically created and
displayed a molecular connectivity diagram (MCD). The
MCD reflecting the HMBC data from Table 1 is shown in
Figure 2.
The atom properties were automatically determined for
all carbon atoms by the program using the APCT table (see
section 2.2.1). In this example, for the carbon atoms of the
benzene rings the parameters of hybridization and atom
neighborhood with attached heteroatoms were automatically
set to sp2/not defined and for the carbon with a chemical
shift of 198.8 ppm atom the values were set to sp2/at least
one. As a result, the number of structures generated and
filtered was k ) 32 f 4, with tg ) 26 min. The relatively
large tg value is explained by utilization of HMBC data only
since the COSY spectrum was not cited by the authors of
ref 15.
The structures were ranked using the methodology de-
scribed earlier. The first structure was characterized by dA-
(1) ) 1.51 ppm, dF(1) ) 1.74 ppm, and dH(1) ) 0.42 ppm
and was identical to compound 1. The correctness of the
Table 1. NMR Data for Compound 1 in Pyridine-d5
atom number äC, multiplicity äH HMBC
1 119.68, s 7.02, 6.79
2 157.82, s 7.30
3 104.22, d 7.02 6.79
4 158.66, s 7.30
5 107.02, d 6.79 7.02
6 131.42, d 7.30
7 129.59, s 7.16
8 131.27, d 8.17
9 116.14,d 7.16
10 163.54, s 8.17
11 116.14, d 7.16
12 131.27, d 8.17
13 36.68, t 3.54 3.43
14 26.12, t 3.43 7.30, 3.54
15 198.80, s 8.17, 3.43
A VERSATILE EXPERT SYSTEM J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 779

4. SOLVING 2D NMR PROBLEMS USING FRAGMENTS
FROM THE SYSTEM KB
During the process of testing the StrucEluc-2 system, a
number of problems have been encountered where the
program failed to solve using the common mode, despite
the absence of any contradictions in the 2D NMR data. These
cases will be used as examples to illustrate the strategy of
utilizing fragments for solving these challenging problems.
Experimental data were obtained from literature articles
where the structures analyzed were absent from the full
structure library of the system.
4.1. Lyngbyabellin A. An attempt to automatically
determine the structure of lyngbyabellin A (2), a novel
cytotoxic compound isolated and investigated most recently
by Luesch et al.,17 was not successful. Compound 2 has a
molecular formula of C29H40Cl2N4O7S2 and a molecular mass
of M ) 690. In this example, 2D NMR data obtained from
COSY, HMQC, and HMBC experiments as well as a list of
IR vibrations were obtained from the original work.17
4.1.1. “Common” Mode. The molecule under investiga-
tion is fairly complex. It contains four types of non-carbon
skeletal atoms, N, O, S, and Cl to give a total of 15
heteroatoms and two thiazole rings. The presence of two
sulfur atoms implies atom properties of sp3/nd, sp2/nd. Sulfur
can exist in different valence states and as a result can
introduce a large number of possible isomers during structure
generation.
Several attempts were made to solve this problem. Many
different forms of user intervention were given, including
specifying atom properties and drawing CdO bonds. The
structure generation process was aborted after generating
more than 150 000 structures within a 24-h period. With this
in mind, it was concluded that only the application of the
fragment approach could be helpful for solving the problem.
4.1.2. Utilization of Fragments from the KB. For this
example, the number of found fragments was fairly large:
L ) 7427. Initial viewing of the found fragments indicated
that the fragments listed at the start of the file had 13C
chemical shifts that were very close to those observed
experimentally. For the first fragment, the general pattern
of both the experimental and predicted spectra appeared
similar when they were visually compared (see Figure 4).
In addition, this fragment contained 19 carbon atoms and a
total of 28 skeletal atoms. It was decided that these
coincidences were likely not serendipitous. It was natural to
suggest that this and related fragments could be related to
the structure under investigation.
With these assumptions, the process of creating connec-
tivities from the first 25 fragments (l ) 25) was initiated
using q ) 60%, nf ) 1, and an initial E value of 0.5 ppm.
The program created no MCD for E values lying in the
interval 0.5-9 ppm. 144 diagrams were created using E )
10 ppm. The absence of any vibrational frequency in the
region near 2600 cm-1 as well as the 1H NMR spectrum
suggested that an SH group be introduced into the user
Figure 4. Fragment #1 in the found fragments list. The experimental 13C NMR spectrum of compound 2 is displayed in the upper window
and the fragment subspectrum in the lower window.
A VERSATILE EXPERT SYSTEM J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 781

BADLIST. Structure generation was performed, but no
structures were produced from these MCDs.
Successively increasing the E value to E ) 14 ppm
produced 272 MCDs when 128 new MCDs were produced.
This indicates that 272 different distributions of the carbon
assignments are possible at this given error value (E). Only
32 MCDs survived a check for contradictions. Eight of these
contained the first fragment, and the others were composed
of two fragments. Structure generation was performed using
a single constraintsfour-membered rings, which are uncom-
mon in natural products, were forbidden. The results gave k
) 20 f 20 and tg ) 2 min. The correct structure has
minimum deviations for all spectral predictions: dA(1) )
3.08 ppm, dF(1) ) 3.84 ppm, and dH(1) ) 0.28 ppm, while
rA ) rF ) rH ) 1 and the difference ¢(2-1) ) dA(2) - dA(1)
) 1.3 ppm. The ¢(2-1) ) 1.3 ppm value indicates the
reliability of the solution as described in detail in section 7.
When structure generation was repeated without any
limitation on ring cycle sizes, 25 structures were generated,
and the spectrum prediction again endorsed the priority of
the correct structure. To determine the similarity of the
generated structures to the real structure, we calculated
Tanimoto coefficients for structural similarity.14 The struc-
tures were ranked in decreasing Csim values, and the first
three are shown in Figure 5. The structure most similar to
the correct one gave Csim ) 0.96 and was the second in the
ranked output file in accordance with the spectral data. Figure
5 directly demonstrates the high resolving power of the
method based on accurate 13C NMR spectrum prediction.
It was concluded that a visual comparison of the 13C NMR
spectrum of large fragments ranked by the program at the
top of the list of FFs with the experimental spectrum can
aid in the choice of fragments suitable for creating the MCDs.
At present, this approach may be the only possible way to
solve a difficult problem without making risky assumptions.
4.2. Paradisin C. Reference 18 details the isolation and
identification of a new natural compound, paradisin C (3).
The elucidation was based on high-resolution FAB-MS data
which gave a molecular formula of C42H46O11 and using a
series of 2D NMR spectra (COSY, HMQC, and HMBC).
An attempt to solve the problem in the “common” mode
failed. The structure generation process continued for over
24 h and was aborted at that time. Searching through the
knowledge base using the 13C NMR spectrum as the input
gave 2946 fragments. As in the previous example it was
determined based on visual comparison that the carbon
chemical shifts of the first fragments in the list were close
to the experimental values. This provided a basis to set the
conditions for forming the MCDs which would lead to the
consumption of the “free” skeletal atoms. MCDs were
formed using l ) 500. With E ) 5 ppm, q ) 40%, and nf )
3, the program created n(MCD) ) 384 MCDs. The process of
structure generation without any additional restrictions
resulted in k ) 15, tg ) 2 min 30 s, dA(1) ) 2.10 ppm, dF-
(1) ) 2.47 ppm, dH(1) ) 0.14, ¢(2-1) ) 0.76 ppm, and rA )
rF ) rH ) 1.
These examples indicate that using fragments found in the
database allows the user to solve problems that seemed to
be impossible for the system to cope within the “common”
mode.
5. UTILIZATION OF BOTH USER AND FOUND
FRAGMENTS
Using fragments from the KB often leads to solution of
the problem, but unfortunately the KB is restricted to
published data and may not include suitable fragments
Figure 5. The elucidated structures most similar to the correct structure. Those displayed are selected from the beginning of the ranked
answer file and ranked in decreasing Csim values.
782 J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 ELYASHBERG ET AL.

specific to proprietary chemistries. In this section, two
examples related to the structure elucidation of peptides will
be considered that could not be solved by utilizing fragments
found in the KB. In this case, it was demonstrated that these
problems could be solved with the help of additional data
including user fragments.
5.1. Apramide G. Luesch et al.19 extracted and identified
a new metabolite apramide G (4) whose molecular mass was
determined as M ) 976 and a molecular formula of C52-
H80N8O8S. In the publication,19 tables of data are listed that
contain 1H and 13C NMR spectra and 2D NMR correlations
from COSY, HMQC, and HMBC. Various methods of
solving this problem were investigated. An initial attempt
to solve the problem in the “common” mode indicated that
the processing time would be excessive. The structure
contains a thiazole ring and therefore a sulfur atom with all
of its valence issues. The structure also contains a CtC
group. If the maximum bond multiplicity is set to 3 during
the structure generation process, a large number of structures
result and the generation time correspondingly increases. For
the first attempt at elucidation the structural data obtained
by the authors from the spectrum analysis were used.
5.1.1. “Common” Mode. Checking the MCD revealed that
there were no nonstandard connectivities. The authors had
found that for the fragment CH2-CtCH, one COSY
correlation corresponded to 4JHH. This was taken into account
when creating the MCD. Attempts were made to solve the
problem in the “common” mode using the data obtained by
the authors:19 the molecule contained six NCH3 groups (6
singlets in the 1H NMR spectrum) and one CH2-CtCH
group (a triplet in the 1H NMR spectrum at 1.77 ppm with
a coupling constant of 4JHH ) 2.4 Hz provided evidence for
the presence of the terminal acetylene group). A correlation
between äH(1.77 ppm) and äC(68.9 ppm) suggested sp
hybridization. Instead of the nd/nd property, this atom and
the quaternary äC(84.3 ppm) were manually attributed with
the sp/fb parameter.
After 16 h of operation in the “common” mode, the
program did not generate any structures that complied with
the structural and spectral filters. This result was not
unexpected. The formal number of double bonds in the
structure under investigation is 17, and the molecule is proton
deficient. However, the main difficulty here is the size ofthe
molecule, 69 skeletal atoms, and the fact that the molecule
contains a significant number of heteroatoms including a
sulfur atom.
5.1.2. Utilization of User Fragments. The authors19
identified from the experimental data the presence of a
thiazole ring. This is supported by 2 doublets in the 1H NMR
spectrum at 6.56 and 7.45 ppm (J ) 3.2 Hz) though this
could just as easily have been a furan ring based on these
data. With this information, a thiazole fragment, 5, was
entered as a User Fragment. The chemical shifts shown in
structure 5 were calculated using the accurate method and
MCD creation was initiated.
With E ) 2.5 ppm, one MCD was created containing this
fragment. Bonds were manually included from the six CH3
groups to the nitrogen atoms according to the informa-
tion provided by the authors. The properties of these C
atoms were adjusted to sp hybridization. These restric-
tions were still insufficient to solve the problem, and the
generation process was aborted after 2 h. It was clear that
the process would be time-consuming since over 7000
structures were generated in that time. An attempt was
therefore made to use fragments from the system knowledge
base.
5.1.3. Utilization of Found Fragments. After a fragment
search in the database the number of selected fragments was
L ) 13 934. To prune the fragments that are unlikely and to
move prospective fragments to the top of the list, OH and
NH groups were placed into the BADLIST since there were
no absorptions in the IR spectrum above 3000 cm-1. The
fragments were then filtered with simultaneous exclusion of
ionic structures since the system does not yet allow ionic
structures. The result gave L ) 13 934 f 5121, and the
number of fragments was reduced by a factor of almost 3.
Visual investigation of fragments at the start of the list
revealed that there were some fragments that showed good
correspondence with the experimental spectrum.
The first 10 fragments were selected from the list to form
the MCDs. To confirm whether the problem could be solved
with only the found fragments, the thiazole ring was deleted
from the User Fragments list before starting the MCD
generation process. The following options were used for
creation of the MCDs: l ) 10, E ) 2.5, nf ) 1, and q )
30%. The program created 12 MCDs, and each contained
fragment 6 but with different distributions of the chemical
shifts. In structure 4 this fragment is highlighted with blue
bold lines. Compare the chemical shifts of the structure and
the fragment and obvious similarities exist.
A VERSATILE EXPERT SYSTEM J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 783

Before structure generation, atom properties were adjusted
in the first MCD and automatically transferred to other
diagrams. After 3 h of structure generation, the process was
stopped for the same reasons as described in the previous
step. To solve the problem in a reasonable time, user
fragments, UF, in combination with found fragments, FF,
were considered.
5.1.4. Combining User and Found Fragments. With
settings of E ) 2.5, nf ) 2, and q ) 30%, the program
created 12 MCDs, each containing the user fragment 5 and
fragment 6. Structures were then generated with one restric-
tion only, that is a ring cycle size of Rc ) 5-6. The results
are k ) 4991 f 2143, tg ) 31 min, dA(1) ) 1.13 ppm, dF-
(1) ) 2.35 ppm, dH(1) ) 0.27, ¢(2-1) ) 0.41 ppm, and rA )
rF ) rH ) 1. The first two structures are shown in Figure 6.
The second structure differs from the first only by the
position of the substituent on the benzene ring. It can be
concluded that the usage of both user fragments and found
fragments in the database can be combined to solve the
problem. The unresolved question is obviously how much
information can or should be extracted by a scientist to feed
into the program initially?
5.2. Hetcochlorin. A similar approach allowed the elu-
cidation of hetcochlorin (7), a novel natural product of
molecular formula C27H34Cl2N2O9S2 isolated and character-
ized in the report by Marquez and co-workers.20 The spectral
data included HMBC data with 53 connectivities and COSY
with 4 connectivities.
The molecule contains 2 thiazole rings and based on earlier
experience it was evident that the structure could likely only
be elucidated with the help of user fragments and fragments
found in the database. The authors20 detected the presence
of two thiazole rings and four carbonyls from the 1D 1H
and 13C NMR spectra.
A search through the knowledge base gave L ) 3401.
Visual comparison of the chemical shifts of the carbon atoms
of the first fragment in the list, substructure 8, with the
experimental values indicated they were similar. This is
evident by comparing structure 7 and fragment 8 as
highlighted by blue lines.
Figure 6. The two top structures from the answer file (Apramide G).
784 J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 ELYASHBERG ET AL.

To complete the investigation, two MCDs were obtained
from this fragment with E ) 12 ppm. Attempts to generate
structures from these MCDs failed. When the number of
generated structures exceeded 8000, the structure generation
process was aborted. It was evident that the usage of user
fragments would again be necessary. Since fragment 8
already contained one thiazole ring, an additional thiazole
ring was introduced prior to creation of the MCDs. The
program created 20 MCDs with settings defined as E ) 12
ppm, q ) 50%, and nf ) 2. Structure generation without
any restrictions resulted in the following: k ) 8, tg ) 15 s,
dA(1) ) 2.73 ppm, dF(1) ) 3.54 ppm, dH(1) ) 0.29, ¢(2-1)
) 0.60 ppm, and rA ) rF ) 1, rH ) 3. This approach based
on the usage of fragments from various sources allowed the
elucidation of a complex molecule when the utilization of
both the common mode and fragments from the knowledge
base failed.
6. UTILIZATION OF THE USER DATABASE
Currently, about 20 million compounds have been identi-
fied and well over a quarter million new compounds are
synthesized or isolated each year. It is possible that many
such compounds will have no analogues in the knowledge
bases of expert systems. As a result, it is not always possible
to find fragments in the database that will help to elucidate
the structure of a compound from a new class. Besides, it is
common that an investigator is unable to specify a fragment
that may be present in the structure under examination. Our
investigation has shown that if the methods described above
are ineffective, then the creation of a user database could
enable a solution. The StrucEluc system provides algorithms
and capabilities to create user databases and thereby allow
searching for fragments of related compounds. In particular,
even if only one compound of a similar structure is known,
it can be successfully used for the creation of a user database.
With the help of user databases, the system can easily be
adjusted for the elucidation of compound classes that are
commonly investigated by a given laboratory. Investigations
have shown that the failure to utilize library fragments was
most frequently due to the following issues: (a) the fragments
appropriate for a given problem are missing in the knowledge
base; (b) appropriate fragments are found but the number of
possible permutations of the carbon atom assignments in
these fragments is so huge that the structure generation
process is too long. In previous reports,21,22 the structure
elucidation of the three alkaloids from the cryptolepine series,
cryptolepicarboline (9), cryptospirolepine (10), and 5,5′-
dimethyl-5′H-10,11′-biindolo[3,2-b]quinolin-11(5H)-one (11),
was considered. The structures of these natural products
which failed to be elucidated both in the common and found
fragment modes are shown below.
These molecules are relatively large, highly unsaturated,
and have large fragments (displayed in bold) that contain
no hydrogen atoms, thereby preventing access to structural
information through COSY correlations. In particular, for
structures 9-11 COSY correlations are observed for protons
contained within the benzene rings only. Cryptospirolepine
10 has an especially complex structure that consists of two
planar fragments lying in perpendicular planes joined via a
spiro carbon, while each fragment makes up a system of
conjugated bonds. Long-range “nonstandard” correlations
(nJCH, nJHH, n g 4) inside these planar fragments are likely.
These factors as well some unusual spectral properties all
contribute to a very challenging elucidation process for
cryptospirolepine. The structure determination of compound
11, a degradation product of cryptospirolepine that contains
a large “silent” fragment encompassing 19 skeletal atoms,
is also a complex analytical problem, one whose solution
has been comprehensively described previously.21
As mentioned earlier, the StrucEluc system contains
algorithms and program features that allow the user to create
their own knowledge base focused on chemical classes that
are of particular interest to a given researcher. In this way,
fragments are excised from these structures using proprietary
algorithms. The carbon atoms of the fragments and molecules
are attributed as usual to the corresponding chemical shifts
in the 13C NMR subspectra. The system can therefore be
adjusted to account for the elucidation of new classes of
compounds.
A VERSATILE EXPERT SYSTEM J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 785

6.1.3. Cryptoquindolinone. For the elucidation of this
structure, the raw spectral data were processed and input into
the program using the processing tools available in the ACD/
Labs SpecManager software.31 A 1D 13C NMR spectrum was
not available as is very common in natural product analysis
where quantities of material are limited. The 13C shift inputs
were thus created from the HSQC and HMBC spectra. 18
peaks were identified in the HSQC spectrum (2 CH3 and 16
CH), and 13 peaks were extracted from the HMBC data to
give 31 peaks. According to the molecular formula, the
molecule contains 32 carbon atoms. It was concluded that
one quaternary carbon atom did not show an HMBC peak
and one was added to the spectrum with a chemical shift of
130 ppm, an estimated value to resonate in the middle of
the aromatic interval. The number of peaks in the HMBC
spectra acquired in standard and phase-sensitive mode was
different, 32 and 45 peaks, respectively. To avoid contradic-
tions, the extra peaks observed in the second HMBC
experiment were attributed to a range of potential couplings
and allowed to be 2-4JCH.
Searching for the 13C NMR spectrum in the user fragment
library resulted in 101 fragments. 3144 MCDs were created
with E ) 2 ppm and qfr ) 4 after about 25 min. Checking
for contradictions reduced the number of MCDs to 1376.
Structure generation was performed with all spectral filters
switched off and a cycle size constraint of Rc ) 5-7. A
carbonyl group was added to the GOODLIST. The result of
structure generation gave k ) 7850 f 75 and t ) 30 min.
As the structures were ranked by dA values, the target
structure was moved to the first position. The first three
structures of the ranked file are shown in Figure 11.
The difference in the deviation dA(2) - dA(1) ) 0.5 ppm
is small. However, the increase in deviation of both dH and
dF from structure to structure suggested that the structure of
compound 11 was correctly identified.
It is common for an experienced spectroscopist to detect
molecular fragments simply by visual analysis of 1D and
2D NMR data. This ability is based on experience, knowl-
edge, and insight of a highly qualified researcher, and the
structural information “extracted” from a problem by this
route is invaluable. Providing spectroscopists with software
tools that enable the assembly of the molecular structure in
an interactive mode while allowing them to modify their
hypotheses is of obvious value. This approach provides a
synergistic effect.
Figure 9. Twelve 1,2-Ar-containing fragments sorted out of the 60 fragments extracted from the User Fragment Library.
Figure 10. One of MCDs created during cryptospirolepine
elucidation.
Figure 11. First three structures of the ranked file obtained as the result of identification of compound 11.
A VERSATILE EXPERT SYSTEM J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 787

The ability of the StrucEluc system to act as an assistant
to the elucidation process was tested for this example. Visual
analysis of the molecular connectivity diagram of compound
11 allows the expert to observe three 1,2-Ar fragments and
suggest the presence of a fourth. HMBC connectivities
identified the connection of the two aromatic fragments via
the N-CH3 group and binding of another N-CH3 group to
the third benzene fragment. Several other bonds were drawn
manually; in particular, the carbon atom resonating at 167.1
ppm was connected by a double bond to the oxygen atom.
The resulting diagram is shown below in Figure 12:
The results of structure generation from this manually
created MCD with the aforementioned options were as
follows: k ) 550 f 85 and t ) 1 min. The correct structure
was identified from the other structures since ¢(2-1) ) 2 ppm.
The application of the spectroscopist’s insight had a ben-
eficial effect and was successful without the user fragment
library thereby saving a considerable amount of time. This
example illustrates that a highly qualified expert capable of
determining very complex structures on his own can provide
dramatic increases in the speed of the elucidation process
and can improve the reliability of a system such as StrucEluc
when the researcher’s insights can be incorporated into the
structure solution protocol.
6.1.4. Quindolinocryptotackieine. The same approach has
been applied to the solution of a particularly challenging
problem that remained unresolved for over a decade.32 The
structure of an unknown alkaloid 20, quindolinocryptotackie-
ine, also belonging to the cryptolepine family was elucidated.
NMR data were collected starting in 1991, and a solution
remained elusive, in part due to extensive spectral overlap
in both the 1H and 13C spectra. Manual analysis did not
present a conclusive structure, while computer-assisted
elucidation produced a series of conceivable structures from
the single structure consistent with all of the spectral data
that was selected from the list of computer-generated
structures. To the best of our knowledge, this is the first case
of the application of an expert system to determine the
molecular structure of a complex natural product which was
not amenable to identification by competent spectroscop-
ists.
6.2. An r,â-Unsaturated Pyruvate. The StrucEluc system
has also been employed for the automated structure elucida-
tion of two unexpected reaction products in a reaction of an
R,â-unsaturated pyruvate. The corresponding procedures and
results are described by Sharman et al.33
7. DISCUSSION AND CONCLUSIONS
During the development and testing of StrucEluc, over 150
structure elucidation problems involving natural products
have been solved using their 2D NMR spectra as inputs.
Natural products provide some of the most challenging
structural diversity due to the prevalence of complex
heterocycles and proton-deficient molecules. Experimental
data published in recent years in the Journal of Natural
Products as well as raw spectra obtained in collaboration
with various laboratories have been analyzed. The presence
of tabulated data containing COSY, HMQC (HSQC), and
HMBC data, and the availability of the molecular formula
and/or molecular ion data served as the criteria for selecting
problems from the literature. Statistical processing of the
results generated by the program revealed a number of
commonalities which describe the ability of the system to
solve problems.
Certainly one important system characteristic is the size
of the molecules that can be identified. The distribution of
problems solved (in %) with the number of skeletal atoms
contained in the molecules is presented in Figure 13. Only
molecules having 20 or more skeletal atoms were taken in
consideration, in total more than 130 problems. Figure 13
Figure 12. MCD of cryptoquindolinone built in the interactive
mode.
Figure 13. Distribution of problems solved (in %) with the number
of skeletal atoms of molecules.
788 J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 ELYASHBERG ET AL.

shows that the program can recognize structures of fairly
sizable molecules. The largest of these molecules was n )
90, M ) 1284, C62H92O28. Spectral data were obtained from
ref 34. The structure was solved using the common mode
with COSY and HMBC input data only and a constraint of
Rc ) 6. The results gave k ) 5140 f 66, t ) 4 h 32
min, and rA ) rF ) 1 and the structure obtained is shown
below:
In 60 cases, almost 50% of the problems studied, the
size of the molecules exceeded 30 skeletal atoms. With
these statistics we challenge the opinion of Steinbeck35 who
insists that a deterministic system is not capable of elucidat-
ing the structure of molecules with more than 30 skeletal
atoms.
When 2D NMR data is used in combination with found
fragments and/or user fragments, two crucial stages can be
distinguished: the formation of molecular connectivity
diagrams and structure generation. The time consumed for
the process of MCD creation is usually measured by seconds
and minutes. If MCDs containing appropriate fragments are
not created, then the problem cannot be solved within
reasonable amounts of time. Structure generation needs to
generate an output file that is not empty, and the processing
time for structure generation should not be too long. If the
generation time is acceptable, then checking the different
structural hypotheses is rarely an issue.
An examination of the distribution of problems in regards
to the time required for structure generation showed that in
the overwhelming majority of cases (75%), structure genera-
tion was completed in less than one minute. As a comparison
of structure generation speeds, the StrucEluc and COCON
systems have been compared using materials published by
Junker et al.36 They described the structure elucidation of
ascomicyne, 22, (C43H69NO12) using COSY and HMBC
spectral data:
In the common mode, with default atom properties set,
spectral filtering and application of APCT during structure
generation, StrucEluc generates only one correct structure
within 2 s. This result was obtained using a Celeron processor
at a clock speed of 300 MHz. When atom hybridizations
were set, but checks by heteroatom neighboring and spectral
filter were switched off, the results gave k ) 1926, tg )
56s, rA ) rF ) 1, dA(1) ) 0.95 ppm, and ¢(2-1) ) 0.64 ppm.
Spectral filtering of the output file reduced the output file:
k ) 1926 f 321. Under similar conditions, the COCON
program produced the following results: k ) 350, tg ) 48
min, r ) 2, d(1) ) 3.51 ppm, and d(2) ) 3.56 ppm. These
results were obtained using an SGI R10000 running at a
clock speed of 195 MHz. Under essentially the same
conditions, the StrucEluc structure generator runs markedly
faster, and, in this example, the correct structure selection
proved to be more successful (r ) 1).
This work emphasizes that only expert systems based on
the application of 2D NMR data allows the elucidation of
complex molecules and, as evidenced in this work, in
particular natural products. In this regard we argue with
Meiler et al.37 who suggest using the 1D 13C NMR spectrum
as an alternative to 2D NMR. That group suggests37 that 1D
13C NMR could be successfully used when a genetic
algorithm for structure generation is employed. The authors
argue37 that one of the benefits of 1D 13C NMR is the fact
that “the time-consuming determination of connectivity
information... by 2D NMR spectroscopy is replaced by the
much easier and rapidly obtainable chemical shift value”.
While this may be true for small organic molecules where
there is an abundance of sample and where the 13C NMR
spectra are not crowded, this is certainly not true in many
structural elucidation problems. In particular, when only
small quantities of natural products, metabolites, forensic
samples, etc. are isolated from their respective source, the
possibility of efficiently obtaining a 1D 13C NMR spectrum
is minimal even with recent advances in hardware technol-
ogy. The problems presented in the article37 are limited to
molecules where the number of skeletal atoms is less than
20. Generally speaking, the amount of structural information
described within a heteronuclear shift correlation 2D NMR
spectrum is much higher than that extracted from a 1D 13C
NMR. It appears that the genetic algorithm is unable to
compensate for a lack of structural information especially
in regards to larger molecules.
For choosing the most probable structure in StrucEluc, the
procedure consists of four steps which are fulfilled along
with removing duplicate structures: (1) fast prediction of
13C NMR spectra for all structures using the fast increment
method; (2) preliminary ranking of the structures in order
of increasing dF Value; (3) accurate prediction of 13C NMR
spectra for the first 10-20 structures within a ranked file; and
(4) rank ordering based on increasing dA values. Statistical
analysis of these results shows that for a reliability estimation
of the most probable structure in the output file, the value
¢(2-1) ) dA (2) - dA (1) can be used. If ¢(2-1) g 1 ppm, the
first structure of the output file ranked in order of increasing
dA Value is, as a rule, correct. Using this approach, it is pos-
sible to distinguish the correct structure even in those cases
when the structural file contains thousands of structures. As
the procedure of preferable structure selection is by necessity
routine, it is completely automated in the StrucEluc system.
A VERSATILE EXPERT SYSTEM J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 789

During this work, the calculation of MS match factors,
mi,38 helped to confirm the validity of the most probable
structure. The structure of tasiamide (C42H67N7O10) was
elucidated using 2D NMR data from the article by Williams
et al.39 The top ranked structures had very close dA devia-
tions: dA (1) ) 1.48, dA (2) ) 1.64, dA (3) ) 1.91 ppm, and
¢(2-1) ) 0.16 ppm. The priority of the correct structure was
confirmed by the following MS match factors calculated for
the ranked structures: m1 ) 97%, m2 ) 78%, and m3 )
48%.
The possibility of utilizing substructural fragments during
the elucidation process using 2D NMR data has been
described. The methodology and corresponding algorithms
have been developed to employ both KB fragments as well
as those introduced by the user. The value of this approach
has been corroborated by solving a number of problems.
As a result of testing this approach, it has been concluded
that the utilization of found and user fragments is profitable
in the following situations:
When the number of experimentally available 2D NMR
correlations is markedly less than the number of theoretical
correlations for a given structure. This situation can occur
due to unfortunate experimental parameter choices and/or
sample induced limitations such as a low S/N ratio.
When the number of 2D correlations is small due to a
deficit in the number of hydrogen atoms and/or specific
stereochemical factors.
When a molecule under investigation is so large and the
spectral data so complex that even rich 2D NMR data can
lead to a long structure generation time.
The application of found fragments frequently allows the
solution of a problem either with minimum user intervention
or in a fully automated fashion. The precision of the accurate
method of 13C NMR spectrum prediction is more than enough
to provide high quality user fragments to the program.
Moreover, the user-defined fragments can be utilized in the
creation of molecular connectivity diagrams from found
fragments containing the given user fragments.
This investigation has allowed the identification of a
number of criteria for estimating the appropriateness of a
user fragment. A user fragment, UF, should be selected such
that the UF would represent a terminal or external part of a
molecule. If a fragment belongs to the internal core of a
molecule, the accurate 13C NMR spectrum calculation turns
out to be too approximate. In this case, when the discrepancy
between calculated and experimental chemical shifts is
considerable, the program is usually not capable of accepting
the fragment for the creation of molecular connectivity
diagrams. Moreover, the introduction of such UF’s that have
two or more free bonds from a terminal atom should be
avoided. In this case, the difference between the experimental
and calculated chemical shifts generally exceeds the default
value of 12 ppm adopted for the terminal atoms in StrucEluc-
2.
The application of found fragments requires the following
three operations: (1) searching for the fragments in the
database; (2) creating a set of molecular connectivity
diagrams from the found fragments; and (3) checking the
diagrams for contradictions. All of these procedures can take
some time, and in some cases the total time required for
problem solving may be of the same order or even longer
than the common mode approach. Occasionally the most
time-consuming procedure can be the process of choosing
the valid E value by gradually increasing its magnitude. It
should be emphasized that when fragments are used, the
structure generation is performed most frequently not from
a single MCD, as in the case of common mode, but from a
set of MCDs. Therefore, it is necessary to resort to the
fragment approach only when the common mode fails or a
significant number of assumptions are required to solve the
problem in a reasonable time.
The procedure of creating MCDs from found fragments
can be made more effective by employing a user defined
BADLIST for filtering the list of FFs. For example, the
absence of X-H groups, where X is a heteroatom, for a
molecule under analysis can generally be determined from
either NMR or IR spectra. These groups can then be included
into the BADLIST. All found fragments that have the
mentioned groups may then be rejected by filtering the FFs
file with the BADLIST. As a result, incorrect fragments are
excluded from the process of forming the MCDs. Depending
on the nature of the molecular formula, the number of found
fragments, L, can be very large. For a compound with
molecular formula C52H80N8O6S (section 5, Apramide G) the
number of found fragments turned out to be 14 000.
Application of found fragment filtering with a reasonably
formed BADLIST usually leads to a considerable decrease
in the number of FFs, occasionally by 30-40%.
This study has shown that there is a need for utilizing the
fragment approach in expert systems based on 2D NMR data.
In particular, the StrucEluc-2 program has already been
successfully applied to the automated structure elucidation
of a series of cryptolepine derivatives, the identification of
degradants, e.g., 11, of the complex alkaloid cryptospirol-
epine (10) using NMR cryoprobe technology,21,22,30 deter-
mination of products obtained in a reaction of an R,â-
unsaturated pyruvate,31 and in a series of investigations not
yet published.
Despite the promising results reported here, continued
development and optimization of the system is necessary.
The program needs to be enhanced to allow new 2D NMR
techniques that are presently in development to be used.
There is also a necessity to enhance the system with new
algorithms to allow the identification of stereochemical
information from recently developed 2D NMR experiments.
The ability of the algorithms utilized to detect the presence
of contradictions in 2D NMR data also needs to be further
refined to make the process still more robust. We have
already demonstrated that in 90% of the problems that were
extremely challenging due to the presence of contradictions,
the latter were automatically revealed by the program.9
List of Abbreviations: KB, knowledge base, containing
full structure library and fragment library; FL, fragment
library; LSC, library of spectrum-to-structure correlations;
TFGL, typical functional group library; APCT, atom property
correlation table; MCD, molecular connectivity diagram; FF,
fragment found in FL from the 1D 13C NMR spectrum; UF,
user fragments involved by the investigator in a structure
elucidation process; MF, molecular formula; E, the maxi-
mum difference allowed between the chemical shifts of the
fragment carbons and the corresponding values observed in
the experimental spectrum under study; L, number of FFs
chosen as a result of fragment search in FL from the 13C
NMR spectrum; l, number of found fragments which will
790 J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 ELYASHBERG ET AL.

be used for the creation of MCDs (l e L); nf , the minimal
number of fragments that must be present in each MCD; q,
the minimum percentage of all skeletal atoms that must be
absorbed by the fragments present in each MCD; nMCD,
number of MCDs created using fragments; k, number of
structures generated; tg, structure generation time; dA, dF,
average deviations of predicted 13C NMR chemical shifts
from the experimental ones calculated by “accurate” (A) and
“fast” (F) methods correspondingly; dH, average deviation
of predicted 1H NMR chemical shifts from the experimental
ones; rA, rF, rH, positions of the correct structure in an output
structural file ranked by corresponding average deviations.
ACKNOWLEDGMENT
The authors are grateful to Sherry Peach for a thorough
review and proofreading of this manuscript.
REFERENCES AND NOTES
(1) Elyashberg, M. E.; Blinov, K. A.; Martirosian, E. R. A new approach
to computer-aided molecular structure elucidation: the expert system
Structure Elucidator. Autom. Inf. Manage. 1999, 34, 15-30.
(2) Blinov, K. A.; Elyashberg, M. E.; Molodtsov, S. G.; Williams, A. J.;
Martirosian, E. R. An expert system for automated structure elucidation
utilizing 1H-1H, 13C-1H and 15N-1H 2D NMR correlations. Fresenius
J. Anal. Chem. 2001, 369, 709-714.
(3) Elyashberg, M. E.; Blinov, K. A.; Williams, A. J.; Molodtsov, S. G.;
Martirosian, E. R. Application of a new expert system for the structure
elucidation of natural products from their 1D and 2D NMR data. J.
Nat. Prod. 2002, 65, 693-703.
(4) Elyashberg, M. E.; Martirosian, E. R.; Karasev, Yu. Z.; Thiele, H.;
Somberg, H. X-PERT: a user-friendly expert system for the molecular
structure elucidation by spectral methods. Anal. Chim. Acta 1997, 337,
265-286. Expert systems as a tool for the molecular structure
elucidation by spectral methods. Strategies of solution to the problems.
Anal. Chim. Acta 1997, 348, 443-463.
(5) Funatsu, K.; Miyabayashi, N.; Sasaki, S. Further development of
structure generation in the automated structure elucidation system,
CHEMICS. J. Chem. Inf. Comput. Sci. 1988, 28, 18-28.
(6) Christie, B. D.; Munk,M. E. Structure elucidation by reduction - a
new strategy for computer-assisted Structure Elucidation. J. Chem.
Inf. Comput. Sci. 1988, 28, 87-93.
(7) Will, M.; Fachinger, W.; Richert, J. R. Fully automated structure
elucidationsa spectroscopist’s dream comes true. J. Chem. Inf.
Comput. Sci. 1996, 36, 221-227.
(8) Elyashberg, M. E.; Karasev, Yu. Z.; Martirosian, E. R. Spectroscopic
determination of elemental composition of organic compounds with
the aid of the X-PERT system. Anal. Chim. Acta 1999, 388, 353-
363.
(9) Molodtsov, S. G.; Elyashberg, M. E.; Blinov, K. A.; Williams, A. J.;
Martin, G. M. Detection and removal of contradictions in the 2D NMR
data during structure elucidation by the StrucEluc expert system. J.
Chem. Inf. Comput. Sci. Submitted for publication.
(10) Molodtsov, S. G. Computer-aided generation of molecular graphs.
Commun. Math. Chem. (MATCH) 1994, 302, 213-224.
(11) Molodtsov, S. G. Generation of molecular graphs with a given set of
nonoverlapping fragments. Commun. Math. Chem. (MATCH) 1994,
30, 203-212.
(12) Molodtsov, S. G. The generation of molecular graphs with obligatory,
forbidden and desirable fragments. Commun. Math. Chem. (MATCH)
1998, 37, 157-162.
(13) Bremser, W. HOSE - a novel substructure code. Anal. Chim. Acta
1978, 103, 355-365.
(14) Willett, P.; Barnard, J.; Downs, G. J. Chemical similarity searching.
Chem. Inf. Comput. Sci. 1998, 38, 983-996.
(15) Gonzales, A. G.; Leon, F.; Sanchez-Pinto, L.; Pardon, J. I.; Bermejo,
J. Phenolic compounds of dragon’s blood from Dracaena draco. J.
Nat. Prod. 2000, 63, 1297-1297.
(16) Gribov, L. A.; Elyashberg, M. E.; Serov, V. V. On the solution of the
one classical problem in vibrational spectroscopy. J. Mol. Struct. 1978,
50, 371-387.
(17) Luesch, H.; Yoshida, W. Y.; Moore, R. E.; Paul, V. J.; Mooberry, S.
L. Isolation, structure determination, and biological activity of
lingbyabellin A from the marine cyanobacterium Lyngbya majuscula.
J. Nat. Prod. 2000, 63, 611-615.
(18) Ohta, T.; Maruyama, T.; Nagahashi, M.; Miyamoto, Y.; Hosoi, S.;
Kiuchi, F.; Yamazoe, Y.; Tsukamoto, S. Paradisin C: a new CYP3A4
inhibitor from grapefruit juice. Tetrahedron 2002, 58, 6631-
6635.
(19) Luesch, H.; Yoshida, W. Y.; Moore, R. E.; Paul, V. J. J. Apramides
A-G, novel lipopeptides from the marine cyanobacterium Lyngbya
majuscula. J. Nat. Prod. 2000, 63, 1106-1112.
(20) Marquez, B.; Watts, K. S.; Yokochi, A.; Roberts, M. A.; Verdier-
Pinard, P.; Jimenez, J. I.; Hamel, E.; Scheuer, P. J.; Gerwick, W. H.
Structure and absolute stereochemistry of hectochlorin, a potent
stimulator of actin assembly. J. Nat. Prod. 2002, 65, 866-871.
(21) Martin, G. E.; Hadden, C. E.; Kaluzny, B. D.; Russell, D. J.; Stiemsma,
B. A.; Thamann, T. J.; Crouch, R. C.; Blinov, K. A.; Elyashberg, M.
E.; Williams, A. J.; Martirosian, E. R.; Schiff, P. L., Jr. Identification
of degradants of a complex alkaloid using NMR cryptoprobe technol-
ogy and ACD/Structure Elucidator. J. Heterocycl. Chem. 2002, 39,
1241-1250.
(22) Blinov, K. A.; Carlson, D.; Elyashberg, M. E.; Martin, G. E.;
Martirosian, E. R.; Molodtsov, S. G.; Williams, A. J. Computer assisted
structure elucidation of natural products with limited 2D NMR data:
application of the StrucEluc system. Magn. Reson. Chem. 2003, 41,
359-372.
(23) Ablordeppey, S. Y.; Hufford, C. D.; Borne, R. F.; Dwumu-Badu, D.
1H NMR and 13C NMR assignment of cryptolepine, a 3: 4-benz-ä-
carboline derivative isolated from Cryptolepis sanguinolenta. Planta
1990, 56(4), 416-417.
(24) Hadden, C. E.; Duholkc, W. K.; Guido, J. E.; Robins, R. H.; Martin,
G. E.; Sharaf, M. H. M.; Schiff, P. L., Jr. Structural characterization
of components of a mixture - characterization of a facile oxidation
product of cryptolepinone in situ. J. Heterocycl. Chem. 1999, 36, 525-
531.
(25) Spitzer, T. D.; Crouch, R. C.; Martin, G. E.; Sharaf, M. H. M.; Schiff,
P. L., Jr.; Tackie, A. N.; Boye, G. L. Total assignment of the proton
and carbon NMR spectra of the alkaloid quindoline - utilization
of HMQC-TOCSY to indirectly establish protonated carbon-proto-
nated carbon connectivities. J. Heterocycl. Chem. 1991, 28, 2065-
2070.
(26) Sharaf, M. H. M.; Schiff, P. J., Jr.; Martin, G. E.; Phoebe, C. H., Jr.;
Tackie, A. N. Two new indoloquinoline alkaloids from Cryptolepis
sanguinolenta: cryptosanguinolentine and cryptotackieine. J. Hetero-
cycl. Chem. 1996, 33, 239-243.
(27) Sharaf, M. H. M.; Schiff, P. J., Jr.; Tackie, A. N.; Phoebe, C. H., Jr.;
Davies, A.; Andrews, C. W.; Crouch, R. C.; Martin, G. E. Isolation
and elucidation of the structure of homocryptolepinone. J. Heterocycl.
Chem. 1995, 32, 1631-1636.
(28) Sharaf, M. H. M.; Schiff, P. J., Jr.; Tackie, A. N.; Phoebe, C. H., Jr.;
Johnson, R. L.; Minick, D.; Andrews, C. W.; Crouch, R. C.; Martin,
G. E. The isolation and structure determination of cryptomisrine, a
novel indolo[3,2-b]quinoline dimeric alkaloid from Cryptolepis san-
guinolenta. J. Heterocycl. Chem. 1996, 33, 789-797.
(29) Crouch, R. C.; Davies, A.; Spitzer, T. D.; Martin, G. E.; Sharaf, M.
H. M.; Schiff, P. J., Jr.; Phoebe, C. H., Jr.; Tackie, A. N. Elucidation
of the structure of quindolinone, a minor alkaloid of Cryptolepis
sanguinolenta: submilligram 1H-13C and 1H-15N heteronuclear shift
correlation experiments using micro inverse-detection. J. Heterocycl.
Chem. 1995, 32, 1077-1080.
(30) Tackie, A. N.; Boye, G. L.; Sharaf, M. H. M.; Schiff, P. L.; Crouch,
R. C.; Spitzer, T. D.; Johnson, R. L.; Dunn, J.; Minick, D.; Martin,
G. E. Cryptospirolepine, a unique spiro-nonacyclic alkaloid isolated
from Cryptolepis sanguinolenta. J. Nat. Prod. 1993, 54, 653-
670.
(31) ACD/SpecManager program. This program is capable of reading data
presented in practically all formats common in NMR spectroscopy.
http://www.acdlabs.com/products/spec_lab/exp_spectra/.
(32) Blinov, K. A.; Elyashberg, M. E.; Martirosian, E. R.; Molodtsov, S.
G.; Williams, A. J.; Sharaf, M. M. H.; Schiff, P. L., Jr.; Crouch, R.
C.; Martin, G. E.; Hadden, C. E.; Guido, J. E. Quindolinocrypto-
tackieine: the elucidation of a novel indoloquinoline alkaloid structure
through the use of computer-assisted structure elucidation and 2D-
NMR. Magn. Reson. Chem. 2003, 41, 577-584.
(33) Sharman, G. J.; Jones, I. C.; Parnell, M. J.; Willis, M.; Carlson, D.
V.; Williams, A.; Elyashberg, M. E.; Blinov, K. A. and Molodtsov,
S. G. Automated structure elucidation of two unexpected reaction
products in a reaction of R,â-unsaturated pyruvate. Magn. Reson.
Chem. In press.
(34) Jayasuriya, H.; Lingham, R. B.; Graham, P.; Quamina, D.; Herranz,
L.; Genilloud, O.; Gagliardi, M.; Danzeizen, R.; Tomassini, J. E.; Zink,
D. L.; Guan, Z.; Singh, S. B. Durhamycin A, a potent inhibitor of
HIV tat transactivation. J. Nat. Prod. 2002, 65, 1091-1095.
A VERSATILE EXPERT SYSTEM J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 791

(35) Steinbeck, C. J. SENECA: a platform-independent, distributed, and
parallel system for computer-assisted structure elucidation in organic
chemistry. Chem. Inf. Comput. Sci. 2001, 41, 1500-1507.
(36) Junker, J.; Maier, W.; Lindel, T.; Koeck, M. Computer-assisted
constitutional assignment of large molecules: COCON analysis of
ascomicyn. Org. Lett. 1999, 1, 737-740.
(37) Meiler, J.; Will, M. C. Automated structure elucidation of organic
molecules from 13C NMR spectra using genetic algorithms and neural
networks. J. Chem. Inf. Comput. Sci. 2001, 41, 1535-1546.
(38) Williams, A. J.; Lee, M. S.; Lashin, V. An integrated desktop mass
spectrometry. Processing and molecular structure management system.
Spectroscopy 2001, 16, 38-49.
(39) Williams, P. G.; Yoshida, W. Y.; Moore, R. E.; Paul, V. J. Tasiamide,
a cytotoxic peptide from the marine cyanobacterium Symploca sp. J.
Nat. Prod. 2002, 65, 1336-1339.
CI0341060
792 J. Chem. Inf. Comput. Sci., Vol. 44, No. 3, 2004 ELYASHBERG ET AL.