This document summarizes the results of a multi-year blind trial study evaluating the computer-assisted structure elucidation software ACD/Structure Elucidator. Over the study period from 2003-2011, 112 structure determination challenges were submitted globally and analyzed by the software. The challenges varied in complexity and data quality. The software was able to correctly determine the structure in most cases (89%), but had difficulties with some early versions that have since been improved. The blind trial process helped identify areas for software development and increased the program's ability to solve complex problems.
2. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 2 of 22
http://www.jcheminf.com/content/4/1/5
We present the results of a unique, multiyear world- submitter was not confident enough to verify the most
wide blind trial study on a CASE application, namely probable structure delivered by the program or did not
ACD/Structure Elucidator (StrucEluc) [5,6], and the respond back to confirm. In most of these challenges,
necessary evolution of the CASE technology as various this structure is considered proprietary and acknowled-
complex challenges were encountered. StrucEluc is an ging its correctness with an outside source could breech
artificial intelligence system that can interpret data from company policies. A total of 100 challenges fell into one
a variety of spectral datasets including 1D and 2D NMR, of these two categories.
MS, IR, etc. Based on the restrictions imposed by the set For the Incorrect category (colored in red), the struc-
of spectral data, all possible atomic combinations are ture generated by the software was not in agreement
worked out to ensure that no plausible candidate with the proposed structure of the submitter. For Stru-
escapes consideration [7]. In addition, a general view- cEluc version 5, three of the seven trials consisted of
point is presented regarding the inherent trends in the unknowns larger than 1000 Da, thus surpassing the size
complex nature of the data associated with each limitations of the software. This limitation in place at
challenge. the time of the analysis has since been removed. The
remaining four trials did not want to share their pro-
Results and discussion posed candidates. For the last category, Data Rejected
1. Categorizing the Global Challenges (colored in grey), the required data were inadequate for
In 2003 a worldwide challenge [8] was initiated with the analysis due to poor instrument practice, exhibited
intent of testing and showcasing the performance of the extremely poor S/N or contained indiscernible artefacts
CASE expert software system StrucEluc. Originally or impurities, etc. A total of 12 challenges fell into one
intended as a single-blind trial, a scientist was requested to of these two categories.
submit spectral data for an organic compound while with- The StrucEluc software failed to generate a structure
holding the structural skeleton so as to not bias the opera- corresponding to that expected by the submitter only
tor of the software. The software would be used to with versions 5 and 6, released in 2003 and 2004
generate one or more candidate structures consistent with respectively. Reviewing the data showed that the soft-
the spectral data, the results would be reported to the ware lacked several features that prevented the software
scientist and they would confirm validity of the analysis. from successfully elucidating the structures. This
As of January 2011, a total of 112 official challenges included library searches using chemical shifts and
had been received from a variety of institutions includ- handling ambiguous assignments for COSY and HMBC
ing academic (50%), pharmaceutical (42%), industrial correlations. The ongoing challenging of the system
(5%) and government (3%) institutions. The global using hundreds of real problems helped to direct the
responses segmented into the following regions: North development of the system as it is impossible to ima-
America at 47%, Europe at 30%, Asia at 18% and the gine all difficulties a priori. The limitations were discov-
remaining continents at 5%. ered during the process of problem solving and the
Each challenge provided a variety of degrees of com- software was improved incrementally over time to over-
plexity and expertise in the elucidation of unknown come them.
compounds. For ten of the 112 challenges (9%), the As the StrucEluc software was developed to accommo-
structures were unknown to both the submitter and the date specific nuances associated with an elucidation, the
analyst and the double-blind trials were highly valued number of submitted challenges also increased, together
and scientifically interesting to both parties. In addition, with the number of correct structures. The popularity of
a separate set of five challenges (4%) were submitted to the challenge attracted the attention of a new group of
ascertain and validate the submitter’s proposed struc- chemists, specifically Ph.D. students, seeking out
ture; this ensured that additional candidates were not answers to structure elucidation problems that could be
overlooked. included into their thesis. For version 12, four out of the
Figure 1 illustrates the evolution of the StrucEluc soft- five challenges were submitted by students requiring
ware in respect to the number of challenges received assistance in their thesis work. Unfortunately, for one of
when the various incremental versions were available. the problems only a 13 C NMR spectrum was received
The challenges were divided into four results categories: and a library search resulted in no direct hits; the chal-
Double Agreement, Single Agreement, Incorrect and lenge proceeded no further as additional data was not
Data Rejected. The Double Agreement category (colored made available. The remaining challenge was rejected
in green) indicates that the proposed structure was due to poorly collected 1H NMR, 13C NMR and 1H-13C
agreed upon and validated by both the submitter and HSQC spectra and inconsistencies among the data. In
software. This also includes the double-blind trials. The all five cases we offered guidance regarding how to col-
Single Agreement (colored in blue) indicates that the lect better data but these particular challenges did not
3. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 3 of 22
http://www.jcheminf.com/content/4/1/5
Figure 1 A summary of the performance of the StrucEluc system on data for 112 blind trials submitted during the years 2003 - 2011.
The results represent challenges performed using an incremental version. For example, a total of 7 challenges were analyzed with version 5. For
version 6, 12 new and different challenges were analyzed.
progress further. The submitters also declined to have There are a number of factors that contribute to the
their data showcased. successful elucidation of a structure using a CASE sys-
In one particular example the submitter presented tem. Experience has shown that time invested upfront
twenty-four tabular 13C chemical shifts with a molecular offers an improved probability of a successful result. The
formula (MF) of C20HwNxOySz where w, x, y and z are amount of time invested in collecting a diverse range of
used to obscure the numerical values for the MF. No data and of high enough signal-to-noise is important.
further clarification was made by the submitter despite a Also, the care with which data is processed, the time
request for further information. This data was insuffi- invested in peak picking and the piecing together of frag-
cient to proceed with an analysis, because in such cases ments to complete a proposed structure(s) through a
the number of structures that can correspond to the structure generation process all contribute to a successful
available data is hardly constrained. result (Figure 2) [11]. When data under analysis present
The incremental analyses and successes of the system complicated and numerous possibilities to consider, then
were a means by which to build confidence in the gen- CASE systems present an alternative approach [3,12].
eral applicability of a CASE application to assist che-
mists. Each iterative development utilized new strategies 2. Data Processing and Dereplication
to accommodate the diverse nature of the challenge For the 112 challenges discussed in this publication,
data [9,10]. encompassing both Double and Single agreement
4. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 4 of 22
http://www.jcheminf.com/content/4/1/5
Collect/
Process/
Organize Data
Publish/
Report/
Present/ Dereplication
Store Structure
Elucidation
Peak
Verification
Matching
Fragment
Assembly/
Structure
Generation
Figure 2 An arrangement of the common tasks incorporated into a complete structure elucidation workflow [11]. A structure
elucidation encompasses several tasks including data collection, library searching and fragment assembly.
analyses and listed in Table 1, the average processing submitter can quickly assess the top candidates. There
time was determined to be around 84 minutes (~1.4 is a clear advantage of elucidating with a software tool
hours). This includes the time spent on processing the over attempting an elucidation by hand as this ensures
data (e.g. adjusting the window functions, Fourier trans- that every potential isomer is assessed.
formation, phasing, peak picking, assessing impurities, In order to initiate a structure elucidation challenge a
etc.). After processing of the NMR data, dereplication is minimal set of data is required from the submitter.
the first step and consumes only about 3 minutes. The Additional data was willingly accepted (see the Experi-
spectral library used for dereplication comprises of more mental section for more details). In those cases where
than 19,000,000 records with structure information and there may be sample limitations and experiments may
assigned 13C chemical shift values. The computational take a long time to acquire, for example, a 1 H- 13 C
time spent performing structure generation averages just HMBC may take weeks to acquire [11], dereplication
over 25 minutes and, in this time period, an average of was nevertheless feasible.
2639 structures are generated by the software excluding Dereplication is a quick and effective pre-screening
duplicate structures with differing NMR assignments. It approach for the identification of an unknown com-
is necessary to keep in mind that the computational pound. There are several advantages to searching across
time and the number of candidate structures strongly a database or library of known structures when a set of
depend on the uniqueness of the initial information. data is available. These include saving time, energy,
The input of additional data can reduce the computa- instrument time and ultimately this of course equates to
tional time and the number of potential candidates saving money. The ultimate goal is to determine
quite dramatically. The generated candidates are ranked whether a compound is novel or not. If a compound is
according to the deviation between the predicted 13C not found in the database then dereplication can at least
chemical shifts and the experimental shifts so that the help to identify potential classes of chemical compounds
Table 1 Ranges for the calculation times and structures generated across the challenges.
Processing Time (min.) Library Search Time (min.)a Generation Time (min.)a Structures Generated
Minimum 0 (Tabular) 1 1 1
Maximum 245 10 240 100000
Average 84 3 26 2639
a
The calculation was performed using desktop PCs operating at processor speeds of 200 MHz to 2 GHz. For example, structure calculations for version 6.0 were
conducted on a Pentium III 1 GHz system equipped with 512 MB RAM and using the Microsoft Windows 2000 operating system.
5. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 5 of 22
http://www.jcheminf.com/content/4/1/5
similar to the unknown on the basis of the heuristic rule 3. Structure Generation
that “similar structures have similar spectra”. In StrucE- Tables 2, 3 and 4 summarize the results of eight single-
luc the searches can be performed with a MF, monoiso- blind challenges presented in detail in Figures 9, 10, 11,
topic mass, or 13C NMR chemical shifts. 12, 13, 14, 15 and 16. These test sets examine the eluci-
Nine percent of the submitted challenges were solved dation of chemical structures varying in mass from 190
simply with a library search through two available data- to 721 Da. The majority of challenges could be solved
bases, an internal library of ~ 400,000 records and the using typical data extracted from 1H, 1H-13C HSQC and
PubChem library [13] at ~ 19,000,000 records for which HMBC NMR. Multiplicity-edited HSQC data were used
chemical shifts were pre-calculated. The search process when available and were obviously preferred. In some
involved taking the 13 C chemical shifts from the 1D trials, the data from the 1H-1H COSY, TOCSY, NOESY
NMR data or extracting it from the 2D NMR data and and ROESY experiments were not required to solve the
searching for compounds matching the chemical shifts. unknown. In some cases, the data from these experi-
A series of random examples of compounds identified ments reduced the generation time from hours to min-
by searching the internal and PubChem libraries are utes and assisted in the final stage of verifying the
presented in Figures 3, 4, 5, 6 and 7. The compounds consistency for the final candidate. In one case, the sub-
vary in the degree of complexity, size and nature of the mitter supplied spectral data in a table form, which was
compound including synthetic and natural products. It manually entered.
should be noted that these searches consume very little Example 3 exhibits a large number of candidates and a
time, only a few minutes. long generation time due to the high number of het-
In 3% of the challenges, an internal fragment library eroatoms, 12, without any correlating NMR data, a
(~2,000,000 records) was searched and fragment infor- number of atoms without defined hybridization states
mation was utilized to complete the elucidation. Not all and a number of ambiguous correlations. These obser-
challenges were searched through the fragment library vations have been discussed previously [5,12,14].
(vide infra). Figure 8 illustrates an example of a chal- It should be noted that the references for the publica-
lenge where the fragment shown in red was found from tions listed beside Figures 9, 10, 11, 12, 13, 14, 15 and
a 13C chemical shift search of the fragment library. Such 16 were obtained from the submitter after the elucida-
a fragment dereplication approach can assist with the tion was performed, and presented herein as a source of
elucidation of novel compounds with similar scaffolds to spectral information. A number of publications have
known compounds and thus reduce time spent on already reported the use of StrucEluc for the purpose of
structure generation. validating their proposed structure [15,16].
Exp. 13C DB 13C Shift Difference of
# Shift (ppm) (ppm) Exp. - DB (ppm) XHn
1 56.4 56.2 0.2 CH3
6 OH 2 60.0 60.6 -0.6 CH3
8 13 3 91.5 91.3 0.2 CH
4 102.6 102.8 -0.2 CH
O 3 O 7 6
H3C 12 11 14 8 5 105.1 105.7 -0.6 C
1
6 115.9 115.2 0.7 CH
H3C 9 5 4
7 121.0 120.9 0.1 C
2
O 10 15
8 128.4 128.0 0.4 CH
OH O 9 131.9 133.4 -1.5 C
10 151.7 152.2 -0.5 C
Ref J. Nat. Prod./1997/60/638 11 152.0 153.0 -1.0 C
Trivial Name cirsimaritin
C17 H14 O6 12 158.6 159.2 -0.6 C
13 161.3 161.0 0.3 C
14 164.0 163.6 0.4 C
15 182.1 181.7 0.4 C
13
Figure 3 Cirsimirtin (C17H14O6), a compound identified through dereplication using experimental C chemical shifts. The search was
performed using the internal library.
6. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 6 of 22
http://www.jcheminf.com/content/4/1/5
Exp. 13C DB 13C Shift Difference of
# Shift (ppm) (ppm) Exp. - DB (ppm) XHn
1 14.2 14.1 0.1 CH3
OH 2 18.3 18.1 0.2 CH2
3 18.5 18.3 0.2 CH2
20
4 21.9 21.7 0.2 CH3
17 16
5 22.9 22.7 0.2 CH2
19 18 6 27.1 27.0 0.1 CH3
5 21 7 33.2 33.0 0.2 C
CH3 8 33.7 33.5 0.2 CH3
1
10 13 O 9 38.4 38.5 -0.1 C
2 9 15
10 40.1 39.9 0.2 CH2
CH3
11 14 12 6 11 40.7 40.5 0.2 CH2
7 3 12 41.9 41.8 0.1 CH2
13 49.6 49.4 0.2 CH
H3C CH3 14 55.3 55.1 0.2 CH
8 4
15 75.3 75.2 0.1 C
Ref 01 Helv. Chim. Acta/1983/66/1672 16 113.9 113.8 0.1 CH
Ref 02 J. Org. Chem./1980/45/1435
Ref 03 Tetrahedron/1992/48/6667 17 114.8 114.8 0.0 CH
Trivial Name 01 8-epi-chromazonarol 18 117.5 117.3 0.2 CH
C21 H30 O2 19 123.7 123.3 0.4 C
20 148.6 148.2 0.4 C
21 148.7 148.5 0.2 C
13
Figure 4 8-epi-chromazonarol (C21H30O2), a compound identified through dereplication using experimental C chemical shifts. The
search was performed using the internal library.
4. Handling Spectral Data number of candidate structures, information regarding a
The bar graph in Figure 17 summarizes the various fragment or starting material was helpful in reducing
types of datasets received during this research to exam- the generation time by establishing a portion of the
ine the performance of the StrucEluc CASE program. structure and thereby reducing the number of potential
For most of the challenges, we received the minimal candidates. This has been discussed in detail elsewhere
required data as dictated by the guidelines of the chal- [12,17].
lenge (see Experimental section for more details). A Other experiments such as 1H-1H TOCSY, NOE dif-
molecular weight, molecular formula, mass spectrum ference, UV/Vis and IR spectra were not used during
(MS), user fragment information and/or starting mate- the CASE elucidation process. Nevertheless the data
rial was provided for about 85, 68, 26, 7, and 5% of the were not necessarily superfluous but could still be uti-
challenges, respectively. In three cases no information lized for candidate verification purposes. In those exam-
was provided regarding MF, MW, MS, user fragments ples where a 13C chemical shift search was deemed to
or starting materials. All three challenges were neverthe- be successful, the 1H NMR spectra were not required.
less solved and subsequently verified by the submitters. This equated to 8% of the cases.
Two cases were solved through a library search using As a result of the analyses reported in this work it was
the 13C chemical shifts. In the third case, an sp carbon possible to determine what pieces of spectral data were
was suggested by the software based on a 13C chemical required to perform a computer-assisted structure eluci-
shift present in the spectrum; the submitter had not dation and what data could be ignored without loss of
considered this option. fidelity in the results. This type of information can be
Figure 18 summarizes the experiments used to per- useful in future experimental design for gathering data
form the elucidations using StrucEluc. In challenges for an unknown. Figure 19 summarizes the minimal sets
where spectral data such as 13 C NMR and 1 H- 15 N of spectral data employed in a challenge. The combina-
HMBC were available those data were utilized in all tion of spectral 1H/13C/HMQC/HMBC/COSY data were
cases. In complex challenges that produced a large used in 33% of the challenges while only 15%
7. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 7 of 22
http://www.jcheminf.com/content/4/1/5
Exp. 13C DB 13C Difference of
# Shift (ppm) Shift (ppm) Exp. - DB (ppm) XHn
O 1 20.2 20.8 -0.6 CH3
2 21.1 21.5 -0.3 CH3
20 OH 3 29.3 21.9 7.4 CH3
12
6 4 44.6 44.8 -0.2 C
O
18 21 5 60.0 57.8 2.2 CH
5
H3C 6 87.8 81.5 6.4 C
1 9
4 OH 7 108.0 107.9 0.2 CH
H3C 17
3 8 115.2 112.4 2.9 CH
16 9 119.9 119.5 0.4 C
O 19
11 10 128.0 127.0 1.1 C
O
14 11 131.9 127.9 4.0 C
7 O
10 12 132.2 134.6 -2.4 CH
13 13 134.4 135.6 -1.2 C
H3C 15 14 140.5 142.2 -1.7 C
2 8
OH 15 146.2 146.3 0.0 C
16 148.8 150.3 -1.5 C
17 155.0 152.3 2.8 C
PubChem search using 13C NMR 18 175.0 177.6 -2.5 C
CID# 9843671
C17 H14 O6 19 175.5 177.8 -2.3 C
20 195.2 194.6 0.6 C
21 200.0 199.9 0.1 C
Figure 5 CID#9843671 (C17H14O6), a compound identified through dereplication using experimental 13C chemical shifts. The search was
performed using the PubChem library [13] containing 13C chemical shifts predicted using ACD/CNMR Predictor [21]. Chemical shift differences
greater than 2 ppm are highlighted in red.
represented the 1 H/HMQC/HMBC combination. In All unknowns were organic compounds typically con-
most challenges, long-range heteronuclear 2D NMR taining C, H, O and N but also included atoms such as
data were useful in reducing the number of potential S, Br, Cl, F, and Na. Additional file 1 details the com-
candidates. When there were more types of data plexity of the molecular formulae. The challenges
included in a dataset associated with a submitted chal- become more complex when N and S atoms in particu-
lenge then more time was required for standard spectral lar are present as these atoms can exist in multiple
processing of these additional data (i.e. Fourier transfor- valence states and thus increase the number of potential
mation, phasing, peak-picking, assessing impurities, etc.). candidates to be considered [1].
Two key parameters representing an optimal CASE Figure 21 relates the distribution of the molecular
system are: 1) the time required to perform a successful weights across the frequency of each challenge. The can-
elucidation relative to the time it would consume to didates range from the smallest challenge at 149 Da to
perform the analysis manually and 2) the diverse range the biggest at 1256 Da with the average hovering around
of candidates that can be investigated that would not be 419 Da.
feasible if the analysis was attempted manually. The number of heavy atoms (excluding hydrogen
atoms) contained within a MF ranges varies mainly
5. Categorizing the Candidates from 10 to 90 atoms. The total number of heteroatoms
The histogram in Figure 20 represents the distribution of range from 1 to 26. The Ring and Double Bond Equiva-
structures relative to the number of skeletal atoms. A large lence (RDBE) ranges from 1 to 35. As these structure
portion of the compounds are within 31 to 90 atoms. Pre- properties increase in number then the elucidation
vious work by Elyashberg et al. [14] focused on the range becomes more complex. In general of course, a higher
of 20 to 50 skeletal atoms and had only 2 examples over 80 mass relates to more atoms and the spectral data will be
atoms. The elucidations performed in this work included more challenging to interpret. This is a generality as
over 20 challenges for unknowns containing over 80 atoms. clearly a high mass compound can have a simple
8. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 8 of 22
http://www.jcheminf.com/content/4/1/5
Exp. 13C DB 13C Difference of
# Shift (ppm) Shift (ppm) Exp. - DB (ppm) XHn
1 16.0 15.8 0.2 CH3
2 17.9 17.6 0.2 CH3
3 18.1 18.0 0.0 CH3
4 18.6 18.5 0.0 CH3
31 5 19.2 19.1 0.1 CH3
32
H3C 32
6 19.8 19.6 0.2 CH3
25
33 7 19.8 19.8 0.0 CH3
O
33 8 20.3 20.2 0.1 CH3
34
42 9 25.1 24.6 0.5 CH2
18
27 10 25.1 24.7 0.4 CH2
30
11 27.5 27.3 0.2 CH
36 N
12 28.0 27.7 0.3 CH
37
O
13 28.7 28.4 0.4 CH2
O
29 14 29.0 28.5 0.4 CH2
O
14 15 CH3 15 29.3 28.9 0.5 CH
9 7
24
39 16 30.8 30.7 0.1 CH3
20 H3C 17 31.4 31.1 0.3 CH
O 1
N 18 35.1 34.9 0.2 CH2
13
19 43.0 43.0 0.1 CH3
38
23 20 46.9 46.4 0.5 CH2
10
O 21 48.0 47.8 0.2 CH2
N
21 22 53.9 53.6 0.3 CH
35 O
23 58.4 58.0 0.3 CH
H3C
4 24 58.7 58.3 0.5 CH
26
11
25 58.9 58.3 0.6 CH3
CH3
N 16 26 59.4 59.2 0.2 CH
H3C H3C
5
41
19
CH3
27 60.3 59.9 0.4 CH
O N 19 28 76.9 76.5 0.3 CH
22 NH
28 29 78.0 77.8 0.2 CH
40
17 CH3 30 94.9 94.7 0.2 CH
H3C 12
8
3 O 31 127.2 127.0 0.3 CH
CH3
6 H3C 32 128.4 128.1 0.2 CH
2
33 130.4 130.0 0.4 CH
34 134.6 134.2 0.5 C
35 169.0 169.1 -0.1 C
36 169.5 169.3 0.2 C
37 169.8 169.5 0.3 C
Ref J. Org. Chem./1989/54/6005
C45 H68 O9 N6 38 170.6 170.8 -0.2 C
39 171.6 171.4 0.1 C
40 172.2 171.8 0.4 C
41 173.1 173.0 0.2 C
42 178.7 178.2 0.6 C
13
Figure 6 C45H68N5O9, a compound identified through dereplication using experimental C chemical shifts. The search was performed
using the internal library.
spectrum: consider the buckminsterfullerene, C60, that molecules can be elucidated quickly if they are rich in
has a single peak in the 13 C NMR spectrum but the hydrogen atoms as the number of 2D NMR correlations
structural interpretation of the peak was not a simple will be high and, assuming there is not too much over-
problem. It is important to note that very large complex lap elucidation may in fact be rather simple. An increase
9. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 9 of 22
http://www.jcheminf.com/content/4/1/5
Exp. 13C DB 13C Difference of
# Shift (ppm) Shift (ppm) Exp. - DB XHn
1 7.4 9.1 -1.7 CH3
CH3
2 32.0 30.7 1.3 CH3
3 3 32.0 30.7 1.3 CH3
H3C CH3
2 5 4
4 32.0 30.7 1.3 CH3
6 OH 5 32.3 30.9 1.4 C
H 13 9
6 37.0 36.9 0.1 CH2
8 O 10 O
O 7 41.3 41.8 -0.5 CH
15 12 11 18
O 20 19 8 41.4 42.0 -0.6 CH2
14 16 O 17 O
7 9 49.5 50.3 -0.8 CH
O 10 67.8 65.3 2.5 CH
H3C OH H
1
11 69.2 66.1 3.1 C
12 69.2 69.2 0.0 C
PubChem Search using 13C NMR 13 86.8 83.1 3.7 CH
CID# 9909368 14 86.9 84.4 2.5 C
Trivial Name Ginkgolide A 15 88.8 87.0 1.8 CH
C20 H24 O9
16 101.5 96.8 4.7 C
17 110.8 109.2 1.6 CH
18 172.4 171.9 0.5 C
19 175.2 172.3 2.9 C
20 178.1 176.1 2.0 C
Figure 7 Ginkgolide A (C20H24O9), a compound identified through dereplication using experimental 13C chemical shifts. The search was
performed using the PubChem library [13] containing 13C chemical shifts predicted using ACD/CNMR Predictor [21]. Chemical shift differences
greater than 2 ppm are highlighted in red.
in the number of heteroatoms is demonstrated in the the challenges the molecular formula included mixed het-
candidate structure as more combinations of positioning eroatoms. If the unknown contains mixed heteroatoms
of the atoms. Higher RDBE values lead to complex ring with an exchangeable proton such as OH and NH, the
systems and/or an abundance of quaternary carbon number of possibilities increases since an exchangeable pro-
atoms and a deficiency in protons. ton can exist on either the oxygen or nitrogen atoms. If for
There are numerous attributes of complexity for the instance there are two X-H bonds and the molecule con-
elucidation of an unknown using a CASE system. It is tains two oxygen and two nitrogen atoms then the follow-
certainly not the complexity of a molecule to the human ing combinations are possible: OH/OH, NH/NH or OH/
eye as many complex structures can be elucidated very NH. It is important to note that IR and Raman data can
efficiently by a CASE program that might initially seem assist in distinguishing NH and OH groups. Over half the
intractable based on visual inspection. The degree of challenges consisted of mixed heteroatoms.
complexity is affected by the number of protons in the For 7% of the challenges, the submitted data corre-
molecule, from which single and multi-bond correla- sponded to a sodium salt. As with the mixed heteroa-
tions are generated in the 2D NMR spectra. A deficit of toms, the number of sites of ionization and association
hydrogen atoms causes the greatest challenge as the with the sodium ion increases the potential candidates
number of direct and long-range correlations to use in for elucidation of an unknown.
the CASE analysis will be reduced. The level of ambigu- With 6% of the challenges exhibiting some form of
ity in terms of the quantity, diversity and nonstandard structural symmetry, such as an inversion centre or a C2
lengths of the long-range correlations is a major chal- axis, there is a higher incidence of coincident chemical
lenge [1,17]. A large number of candidates and a large shifts. The increase in ambiguity results in longer struc-
generation time results from the interpretation and ana- ture generation time. The problem of generating sym-
lysis of data complicated by these issues. metric structures has been partly solved within the
The complexity of the problem is further compounded by software recently [9] and we expect that symmetry will
the presence of mixed heteroatoms (excluding C atoms), soon be used to facilitate the acceleration of structure
the presence of a salt and molecular symmetry. In 50% of generation.
10. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 10 of 22
http://www.jcheminf.com/content/4/1/5
Exp. 13C DB 13C Difference of
# Shift (ppm) Shift (ppm) Exp. - DB XHn
1 14.9 15.7 -0.8 CH3
2 16.6 16.2 0.4 CH3
O
3 22.8 22.7 0.1 CH3
4 27.4 27.4 0.0 CH2
5 30.7 30.6 0.1 CH2
O OH 6 32.1 32.1 0.0 CH2
CH3 OH 7 34.0 34.0 0.1 CH
1
HO 9
14 8 35.8 35.4 0.5 CH
13 12
9 36.1 36.0 0.1 CH2
CH3 7 CH3
18 3 15 10 2 10 43.7 43.6 0.1 CH
6 11 48.3 47.9 0.4 C
17 11 8
F 12 49.2 48.3 0.8 C
21 19 4 13 71.3 71.2 0.2 CH
O 16 5 14 89.9 90.6 -0.8 C
15 100.7 100.9 -0.2 C
16 124.5 124.4 0.1 CH
Searching a Database for Fragment 17 129.3 129.3 0.1 CH
Information 18 152.9 152.9 0.0 CH
C22 H27 O9 F1
19 166.8 166.9 -0.1 C
20 170.1 - - C
21 186.1 185.9 0.2 C
22 200.5 - - C
Figure 8 C22H27O9F1, an example of a single-blind challenge elucidated using StrucEluc version 7. The fragment, shown in red, was
retrieved via a fragment-based dereplication using the internal fragment library. The 13C chemical shifts are listed in the right panel.
StrucEluc attempts to generate a set of candidate between the experimental 13C chemical shifts and the
13
structures consistent with the data. In many cases a C shifts predicted using incremental and artificial
pool of candidates is generated and a rank-ordering of neural network algorithms [2], as well as a HOSE code
the candidates in terms of their agreement with the [18] based approach [1,17]. A deviation closer to zero
experimental data is required in order to simplify user signifies a better correspondence. Chemical shifts can be
review. As discussed in detail elsewhere [1,17] a number generated for 1H, 13C, 15N, 31P and19F nuclei using var-
of approaches are available including the comparison of ious algorithmic approaches and rank-ordering can be
experimental with predicted NMR spectra as well as performed based on each of the predicted nuclei as well
comparison with mass spectral fragmentation data. The as by favored algorithm. The reader is encouraged to
candidates can be ranked, for example, by the deviation read the references [17,19] for details and examples.
The average 13C deviation for the top ranked struc-
tures is 2.2 ppm with a standard deviation of 3.1 ppm.
13
Table 2 Results of 8 single-blind trials. C NMR shift prediction is chosen for the primary
Computer-assisted Molecular Formula MW (Da) RDBE ranking as the predictions are less affected by solvent
Example 1 C10H10O2N2 190.2 7 than 1H NMR predictions. Based on the results of this
Example 2 C13H10O5 246.2 9 work we have adjusted our benchmark deviations for
Example 3 C31H47NO11 609.7 9 future elucidations when separating good candidates
Example 4 C30H42O9 532.6 10 from poor ones.
Example 5 C15H16O4 260.3 8 For Table 5, the average structure rank includes two
Example 6 C36H60N6O7S 721.0 10 challenges where the correct structures were ranked at
Example 7 C24H34O9 465.5 8 positions 28 and 80. In both cases, the lists of candi-
Example 8 C25H25NO7 451.5 13 dates were very close in the 13C deviation and the sub-
Molecular formulae, molecular weights and RDBE values of compounds.
mitters did not consider the proposed structure listed in
11. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 11 of 22
http://www.jcheminf.com/content/4/1/5
Table 3 Results of 8 single-blind trials.
1D NMR Data 2D NMR Data (Total Correlations/Ambiguous Correlations) Data not used
1 13
Example 1 H, C HSQC, HMBC (10/2) COSY
1 13
Example 2 H, C, DEPT135 HSQC, HMBC (20/5) COSY
1 13
Example 3 H, C, DEPT135 HETCOR, HMBC(86/5), COSY (31/4) ROESY
1 13
Example 4 H, C HSQC, HMBC(74/17), COSY (14/1) -
1 13
Example 5 H, C HSQC, HMBC (23/4), COSY (25/12) -
1 13
Example 6 H, C HSQC-DEPT, HMBC(67/0), COSY (31/0) NOESY
1
Example 7 H HSQC, HMBC(47/0) COSY, TOCSY
1 13
Example 8 H, C HSQC, HMBC(20/14), COSY(2/2) -
Spectral data of compounds.
the first position. Excluding these two challenges, the modest size) was symmetric. To overcome this difficulty,
average structure ranking lists the correct structure in the algorithm was reworked in such a manner to detect
first place. the presence of molecular symmetry from a logical ana-
lysis of the NMR spectral data and to perform structure
6. Dealing with the Problem of Molecular Symmetry generation taking into account the molecular symmetry.
Version to version StrucEluc has continued to be incre- During the process of algorithm improvement the per-
mentally improved to accommodate the nuances of formance was continuously tested using a particular set
complex and challenging data and experiences obtained of structures. One of these representative compounds
from solving hundreds of problems. The application of uses the experimental data borrowed from the work of
the software over the decade since initial development Tsuda et al. [20]. The structure for Dendridine A
has helped to characterize a wide variety of analyzed (C20H20Br2N4O2) exhibits a C2 axis (see Figure 22).
structures and associated spectral data. While this publi- Figure 23 shows that the 2D NMR data produced 1H-
1
cation cannot exhaustively examine the incremental H COSY (blue lines) and 1H-13C HMBC (green lines)
design and algorithm changes which have occurred correlations and only one pair of CH 2 groups were
from version to version, and for that the reader is defined by the program as having no heteroatom neigh-
referred to our myriad of publications and review arti- bours. This indicates that all other carbon atoms may
cles. However, an example of the impact of one algo- be connected with N, O or Br atoms, and it can be con-
rithm enhancement on the performance of the software cluded that a great number of structures may appear
does warrant mention. For many years it was observed during the structure generation process.
that the algorithm for structure generation from 2D The step-by-step progress in improvements regarding
NMR data failed to solve a problem in a reasonable the performance of StrucEluc in dealing with symmetry
time if the molecule under investigation (even of a is illustrated in Table 6. The table shows the initial
Table 4 Results of 8 single-blind trials.
Position of Accepted Number of Structures Spectral Processing Generation Time d13C HOSE stdd 13C
Candidate Generated Time (min.) (min.) (ppm) (ppm)
Example 1 389 135 <1 2.76 3.30
1
Example 1 50 Tabular <1 1.73 2.23
2
Example 1 9339 180 30 2.10 2.74
3
Example 1 388 180 6 2.20 3.05
4
Example 1 116 60 <1 1.44 2.52
5
Example 1 9872 240 8 1.02 1.91
6
Example 1 127 120 <1 0.84 1.07
7
Example 1 3224 120 60 1.82 2.08
8
Numerical data. In most cases, raw experimental NMR data files were submitted. For example 2, the processed data were submitted in the form of a table.
12. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 12 of 22
http://www.jcheminf.com/content/4/1/5
HC CH 110.77
CH3 92.22 HC 110.77
121.52
59.01 109.80 109.80
HC C 124.72 NH
C 121.52 C 134.39
134.39
118.67 124.72 NH
92.22
CH NH O 118.67
173.28 173.28 160.07
C NH
160.07
O
O H3C
O 59.01
Figure 9 Example 1: a single-blind challenge [22] elucidated using StrucEluc version 7. The molecular connectivity diagram (MCD) in the
left panel consolidates the data from the MF and the spectral data into a single diagram. The blue, green and black lines represent the
connectivities extracted from a 1H-1H COSY, 1H-13C HMBC spectra and user fragments, respectively. The dashed lines indicate ambiguity in the
assignment of the correlations. The carbon atom colors dictate the hybridization state of the atom: blue, pink and black represent sp3, sp2 and
sp/sp2/sp3, respectively. The right panel exhibits the most probable candidate with assigned 13C chemical shifts.
difficulty of generating symmetric molecules with the StrucEluc versions reduced both the output file size and
version available in 2005 and the incremental improve- the time associated with structure generation. Many
ment in the results as a result of adjusting the algorithm algorithmic improvements were introduced over the life-
in 2006. Further improvements in performance between time of the software but such examples have become
H2C C CH
H3C 62.88 HC 106.80 109.27
19.69 104.50
HO 104.50 O
151.01 151.27 62.88
CH C C
113.85 151.01 141.38 106.80 110.81 O
C C 151.27 109.27 155.97 175.38
110.81 141.38 HO
O 113.85
C O O
164.97
C 164.97 C
155.97 175.38 CH3
19.69
O H H
O O
Figure 10 Example 2: a single-blind challenge [23] elucidated using StrucEluc version 7.
13. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 13 of 22
http://www.jcheminf.com/content/4/1/5
CH3 H3C CH2 CH2 CH2 CH3
7.72 21.47 29.04 32.92 37.23 21.47
142.61
H2 C
111.61 37.23
H2 C CH H2C CH CH HC
28.04 41.85 111.61 70.61 80.50 81.27 81.27 OH
H3 C
HC 56.26O 71.81
80.77
175.10
O
HN
H3 C H3C CH3 CH
9.35
H3C 22.24 C 55.11 H3C 71.81 CH
12.56 38.42 78.34 29.04 OH
56.26 72.06
O 72.68 70.61
HC CH C C H3 C O 38.42CH3
72.68 HC 99.95 C 113.95 C 142.61 C 55.11
80.50 12.56
78.34 100.08 139.88 162.17
CH3
HO 32.92 22.24
C C O O CH3 72.06
163.23 C 175.10 N O O 9.35
171.24 HO 113.95 28.04 41.85
163.23 139.88 80.77 CH3
O O O O 7.72
O O O H 99.95 100.08 O
162.17 171.24
OH O
H H H
H H
Figure 11 Example 3: a single-blind challenge [24] elucidated using StrucEluc version 9.
CH2
H3C CH3 CH2 25.16 CH2 H3C
H3C 11.95 H3C 18.43 22.34 27.88 8.59 27.88
8.59 13.98
O
O 59.80 173.58
CH2
H2C CH2 CH 41.57 CH 206.46 CH3
H2C 34.27 H2C 36.15 39.41 48.30 CH3 O 52.47
31.77 34.37 41.57 18.43
C 39.41 73.59
C H2C CH 98.73 CH
H3C 59.80 CH 73.59 75.29 110.55
52.47 67.28 48.30 O
H3C
11.95 112.31
C
C C C 170.52 C 34.27 O OH
C 112.32 C 149.92 162.73 173.58
H3C
13.98
112.31 142.49 149.92 170.52
22.34 67.28
O 112.32 98.73 O
31.77 O
O O O O
C O 25.16 75.29 142.49 162.73
206.46
36.15 34.37 110.55 OH
O H
O HO
Figure 12 Example 4: a single-blind challenge [16] elucidated using StrucEluc version 9.
14. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 14 of 22
http://www.jcheminf.com/content/4/1/5
H2C C CH2 142.97 O
22.05 37.16
H3C CH2 62.87
15.98 29.77
110.03 141.16
120.12
HC C C
110.03 120.12 CH3
HC HC 132.57
29.77 15.98 80.67
80.67 110.23
22.05 37.16 O
CH C C
141.16 157.21 136.69 157.21 165.77
HC CH 165.77
136.69 142.97 132.57 110.23 O
62.87
O O H
O O HO
Figure 13 Example 5: a single-blind challenge [25] elucidated using version 9.
very useful for emphasizing the impact of particular 7. Spectral Purity
algorithmic enhancements as well as helping to isolate Spectral purity is an important criterion for a successful
classes of structural challenges requiring focused efforts. and relatively pain-free elucidation. Datasets that exhibit
The details regarding the symmetry handling will be dis- poor signal-to-noise, poor signal resolution, unexpected
cussed in detail in a separate publication. impurities, mixtures and/or artefacts tend to produce
CH2 CH3 24.46
H3C CH3 CH3 22.14 CH2 23.51 29.63 O
H3C
18.76
H3C 17.74 H C
3
20.40 20.89 23.06 46.79 26.20 CH3
13.99 18.76 61.00 171.59 20.40
CH2 CH2 N N 58.77
O
H2C CH2 CH2 28.64 CH2 28.93 169.94 H O
168.81
HC 24.46 HC 28.36 28.47 28.68
24.22 26.20 H3C 47.98 O
17.74
NH
CH2 CH 170.37
CH2 HC CH2 43.65 CH2 47.98 NH 52.24
H2C 31.29 H C 41.02 NH
2 41.36 46.79 170.53
29.63 40.41 O 28.36
43.65
C C 50.39 NH
41.02 S
CH CH CH 167.70 C 169.94
40.41
HC 52.24 HC 61.00 61.67 168.81 167.70
61.67
50.39 58.77
24.22 O
NH NH H3C 203.81
C C N NH
20.89
CH3
O
41.36
C 170.53 C 203.81 23.51
170.37 171.59 23.06
O O 28.47
NH O O O 28.64
HN S 28.93
31.29
28.68
22.14
O CH3
O 13.99
Figure 14 Example 6: a single-blind challenge [26] elucidated using StrucEluc version 12.
15. Moser et al. Journal of Cheminformatics 2012, 4:5 Page 15 of 22
http://www.jcheminf.com/content/4/1/5
O O O
C CH C
68.12 170.39 H3C
22.48
CH2 O 25.88 43.67
H3C
22.48 43.67
HC H3C O CH3
25.88 22.55 20.45
O 68.12 136.59
CH3
22.55 O 27.86 123.77
O HC O 170.39O
67.43 64.69 43.25 67.43
C
172.94 H3C
O H 21.15 O
48.68 O
H3C 47.30
O 6.95
78.81
84.68
C H3C H3C C O 78.48
H3C 20.45 CH3 21.15 CH2 43.25
6.95 21.10 27.86
172.94 OH
C CH CH H3C
H2C 48.68H2C 78.48 CH 84.68 CH
47.30 64.69 78.81 123.77
21.10 O
C O
136.59
Figure 15 Example 7: a single-blind challenge [27] elucidated using StrucEluc version 12.
longer generation times, higher numbers of candidates, data is acquired, a range of datasets varying in spectral
and in some cases, prevent any sensible candidates [5]. purity were received. Datasets deemed to be of too low
Since submitters vary in their laboratory procedures in a quality were rejected and requests for better data col-
regards to how samples are prepared and how the NMR lection by the client were issued.
HC CH CH3 126.86
127.04 127.00 24.53 127.00 127.00
HC CH CH CH
126.53 126.85 126.68 126.86 126.68 126.68
C 133.93
111.33
66.74
C CH C CH H3 C
133.67 127.04 133.93 127.00 24.64 O
CH3
HC CH 53.76 O 24.53
126.53 126.68 155.98 73.25 111.33
N 128.38
O CH3
80.56 O 24.64
CH2 CH CH 128.70
HC HC CH 128.38 O
53.76 2 66.21 66.74 73.25 C 70.15
168.49
70.15 80.56 O
O
C N O 66.21
HC 155.98 C O O 126.53 133.67
128.70 168.49
127.04
126.53
O O 126.85
O O 127.04
Figure 16 Example 8: a single-blind challenge [28] elucidated using StrucEluc version 12.