The 7th IEEE International Conference on E-Health and Bioengineering - EHB 2019

Grigore T. Popa University of Medicine and Pharmacy, Iasi, Romania, November 21-23, 2019

Screening for NBOMe Hallucinogens based on

Artificial Neural Networks and Structural Descriptors
Adelina Ion, Steluta Gosav, Mirela Praisler
Faculty of Sciences and Environment, “Dunarea de Jos” University of Galati
Galati, Romania
Adelina.Ion@ugal.ro, Steluta.Gosav@ugal.ro, Mirela.Praisler@ugal.ro

Abstract— Synthetic hallucinogen trafficking is a global illicit
trade. It represents one of the main dangers to public health
worldwide. NBOMe is a new class of synthetic hallucinogenic
drugs of abuse, which is sold through illicit channels as
alternative to LSD. The most representative member of the
NBOMe class is 25I-NBOMe, a derivative of 2,5-dimethoxy-4-
iodophenetylamine (2C-I). In this study we are presenting and
comparing a series of Artificial Neural Networks (ANNs)
designed to identify NBOMe hallucinogens based on their
structural descriptors. Such a system may automatically predict
the potential toxicity of new NBOMe compounds and thus saves
analytical time and reduces the cost of toxicity studies. For this
purpose, constitutional descriptors and functional groups of the
optimized molecular structures of the main NBOMe
hallucinogens have been determined. Then ANNs have been built
by using only those descriptors found to be the most important
structural descriptors. The efficiency of the ANNS was compared
and the impact of variable selection on ANN performance was
analyzed in detail based on several merit figures.
Keywords— NBOMe hallucinogens; structural descriptors;
constitutional descriptors; functional groups; ANN.
I. INTRODUCTION
Dimethoxyphenyl -N-[ (2-methoxyphenyl) methyl]
ethanolamine derivatives (NBOMes) are a recently discovered
class of potent synthetic hallucinogens that were originally
developed for research purposes as 5-HT2 receptor agonists
[1]. The most representative synthetic derivative of NBOMe
class is 25I-NBOMe, a derivative of 2,5-dimethoxy-4-
iodophenetylamine (2C-I), that is sold online and
through illicit channels as liquid solutions, powders, laced on
edible items, and soaked onto blotter papers [2]. 25I-NBOMe
(IUPAC name: 2-(4-iodo-2,5-dimethoxyphenyl)-N-[(2-
methoxyphenyl)methyl]ethanamine) is a potent illicit
hallucinogen with serotonergic action and
extremely dangerous pharmacological effects, which are
comparable with LSD or psilocin (see Fig. 1).
According to recent clinical toxicology studies, the use of
25I-NBOMe, which is sold under the names N-bomb, N-bome,
Smiles, 25I, Bom-25 or Cimbi-5, has negative effects such as
confusion, agitation, panic and anxiety, visual and auditory
hallucinations, vasoconstriction, increased heart rate and blood
pressure, thought loops, nausea, mydriasis, acute kidney
injury, oxygen desaturation which can lead to death [3-7].
Fig. 1. Molecular structure of the 25I-NBOMe hallucinogenic
phenethylamine
Artificial neural networks (ANN) are statistical learning
algorithms that may be considered the main tools used in
machine learning. Their mathematical model simulates the
structure and functionality of biological neural networks, in
order to find patterns in data or perform classifications. These
computing tools consist of a number of nodes that are
interconnected in a network-like structure and represent
methods for pattern recognition, that is, machine learning
techniques [8-9].
ANNs are trained using algorithms based on the
optimization theory and statistical estimates. They calculate
real gradients by using the backpropagation method.
Subsequently, a gradient descent methodology is applied: the
derivative of the cost function with respect to the network
parameters is determined, which leads to a modification of the
system parameters in a gradient-related direction in order to
determine the best network architecture [10-12].
ANN consists of three types of layers, i.e. an input layer
(containing nodes representing the input variables of the
neural network), one or several hidden layers (consisting of
nodes that have the role of capturing nonlinearity in data) and
an output layer (consisting of nodes representing the
dependent variables yielding the result for given inputs).

978-1-7281-2603-6/19/$31.00 ©2019 IEEE

 The most remarkable aspect is the fact that this
architecture allows ANNs to learn from experience during the
training process and to detect patterns. Hence, ANNs help
solving complex problems, being especially useful in
situations where large volumes of data need to be processed or
when only incomplete information is available [13-16]. In this
paper we are presenting and comparing a series of ANNs
designed to screen for NBOMe hallucinogens based on their
structural descriptors.
II. DATABASE AND METHODS
The database used for this study consists of 160 substances
of forensic interest, which have been divided in two classes.
The first class, named the “NBOMe” class, contains 15
NBOMe hallucinogens, considered positives. The second
class, named the “non-NBOMe” class, includes 145
compounds (negatives), such as narcotics, sympathomimetic
amines, some potent analgesics and their major precursors.
The molecular structures of all the compounds from
database have been represented in 3D coordinates by using the
HyperChem 8.03 software package [17]. The geometries of all
the 160 substances were completely optimized by using the
AM1 semi-empirical quantum method. The Polak-Ribiere
mechanism was applied in order to regulate the geometrical
parameters and to determine the conditions under which the
minimum energy of the molecular system is reached.
A number of 74 structural descriptors (43 constitutional
descriptors and 31 functional groups) were computed for each
of the 160 compounds by using the Dragon 5.5 software
package. The definitions, mathematical formulas and chemical
significance of the molecular descriptors used in this study are
detailed in Todeschini et al. [18].
An ANN was created, with all these structural descriptors,
by using the Easy NN plus software. The most important
descriptors have been identified with this initial network. Then
three other ANNs have been built by using different numbers
of important descriptors in the database. The architecture of all
these ANNs consists of three layers, i.e. the input, hidden and
output layers. The networks have been trained by using the
backpropagation algorithm and, after optimization, the
learning rate was fixed at α =0.70 and the momentum at μ =
0.80. The sigmoid function was chosen as the activation
function for neural networks using backpropagation.
All four ANNs have the same training set, consisting of: 8
NBOMe hallucinogens, 17 non-NBOMe compounds and 135
samples from the validation set. The output layer contains two
nodes, i.e. NBOMe and non-NBOMe. The training process of
the ANN is completed when the convergence is reached, i.e.
the average training error (TE) falls below the target error. In
our case, TE = 0.01 for all networks. The validation method
was the cross-validation method (leave-one-out).
The first network built, named 74_struct_ANN network,
has as input variables 43 constitutional descriptors and 31
functional groups. The 74_struct_ANN network has 12 hidden
nodes and 912 weight connections.
The 74_struct_ANN network was used to evaluate the
importance of the input variables, in order to select the most
relevant ones. The relative importance of an input variable is
given by its contribution in determining the dependent
variables. In other words, the importance of an input variable
is a measure of how each input will influence the next layer in
the network. The absolute importance represents the sum of
the absolute weights of the connections between an input node
and the nodes in the hidden layer.
The 34_imp_struct_ANN neural network has been built by
using as input variables the first 34 most important structural
descriptors of 74_struct_ANN network. It has 12 hidden nodes
and 432 weight connections after optimization.
In order to analyze the dynamics of the classification
accuracy, two other ANNs have been built by varying
(increasing or decreasing by 10) the number of the descriptors
included in the training database. Hence, the third ANN
(44_imp_struct_ANN) has been built by using as input
variables only the first 44 most important structural
descriptors (as determined with 74_struct_ANN). The
architecture of 44_imp_struct_ANN, derived from the
optimization process, consists of 12 hidden nodes and 552
weight connections.
Finally, the first most important 24 structural descriptors,
as determined with the initial network, have been used to build
24_imp_struct_ANN, network that has 13 hidden nodes and
338 weight connections after optimization.
III. RESULTS
The training process of all ANNs ends when convergence
is reached. The root-mean-squared error (RMSE) of training
vs. the number of learning cycles for the 74_struct_ANN
network is presented in Fig. 2.
Fig. 2. The root-mean-squared error (RMSE) of training vs. the number of
learning cycles for the 74_struct_ANN network.
The relative importance, determined for the first 44 most
important descriptors of the ANN built with 74 structural
descriptors is shown in Fig. 3, in descending order.
 the case of 24_imp_struct_ANN network, the learning process
ended after 10 cycles. The architecture of
24_imp_struct_ANN network is illustrated in Fig. 5.

Fig. 5. The architecture of the 24_imp_struct_ANN artificial neural network,

built with the first 24 most important structural descriptors.

After the validation process of each ANN, several figures

of merit have been calculated in order to determine the
Fig. 3. The first 44 variables found to have the highest relative importance by efficiency of each ANN system. For this purpose, the rate of
analyzing the 74_struct_ANN artificial neural network. true positives (TPR), of true negatives (TNR), of false
positives (FNR), the classification rate (CR) and the correct
We should underline that 11 of the most important classification rate (CCR) have been determined for each ANN.
molecular descriptors presented above, are also among the The results obtained for 74_struct_ANN are presented in
most sensitive structural descriptors. For this reason, the Table I.
34_imp_struct_ANN, 44_imp_struct_ANN and 24_
imp_struct_ANN networks have been built only with those TABLE I. CLASIFFICATION ACCURACY INDICATED BY THE VALIDATION
descriptors that are significantly more important than the rest PROCESS FOR THE 74_STRUCT_ANN ARTIFICIAL NEURAL NETWORK
of the descriptors, as is shown in Fig. 4. Network
Validation parameter
74_struct_ANN
TPR (%) 100
TNR (%) 83.33
FNR (%) 0
FPR (%) 16.67
CR (%) 99.38
ACC (%) 84.9

The results obtained for the three ANNs built with the
Fig. 4. The first 11 variables found to have the highest relative sensitivity by most important descriptors are presented in Table II.
analyzing the 74_struct_ANN artificial neural network.

The 34_imp_struct_ANN network contains 21

constitutional descriptors and 13 functional groups as input
variables. A number of 14 learning cycles were necessary in
the learning process. In the case of 44_imp_struct_ANN, 9
learning cycles were needed in the learning process, while in
TABLE II. CLASIFFICATION ACCURACY INDICATED BY THE VALIDATION REFERENCES
PROCESS FOR THE 34_IMP_STRUCT_ANN, 44_IMP_STRUCT_ANN AND 24_
IMP_STRUCT_ANN ARTIFICIAL NEURAL NETWORK
[1] J. Suzuki, M. A. Dekker, E. S. Valenti, et al., “Toxicities associated with
Networks NBOMe ingestion — a novel class of potent hallucinogens: a review of
Validation the literature”, Psychosomatics 56, 2015, pp. 129-39.
34_imp_struct_ 44_imp_struct_ 24_imp_struct_
parameter ANN ANN ANN [2] https://www.deadiversion.usdoj.gov/drug_chem_info/nbome.pdf.
[3] B. V. Dean, S. J. Stellpflug, A.M. Burnett, K. M. J. Engebretsen,“2C or
TPR (%) 100 100 100 not 2C: phenethylamine designer drug review”, Med. Toxicol., 9, 2013,
pp. 172-178.
TNR (%) 88.97 83.33 86.21
[4] S. L. Hill, T. Doris, S. Gurung, S. Katebe, A.Lomas, M. Dunn, P. Blain,
FNR (%) 0 0 0 S. H. L. Thomas, “Severe clinical toxicity associated with analytically
confirmed recreational use of 25I-NBOMe: case series”, Clin.
FPR (%) 11.03 16.67 13.79
Toxicol., 51, 2013, pp. 487-492.
CR (%) 100 99.38 100 [5] A. Kelly, B. Eisenga, B. Riley, B. Judge,“Case series of 25I-NBOMe
exposures with laboratory confirmation”, Clin. Toxicol. (Phila),
ACC (%) 90 84.9 87.5
50, 2012, pp. 702.
[6] J.Suzuki, J.L.Poklis, A.Poklis, “My friend said it was good LSD: a
suicide attempt following analytically confirmed 25I-NBOMe
IV. CONCLUSIONS ingestion”, J. Psychoactive Drugs, 46, 2014, pp. 379–382.
[7] European Monitoring Centre for Drugs and Drug Addiction, EMCDDA–
Europol Joint, “Report on a new psychoactive substance: 25I-NBOMe
The recognition of the positives (NBOMe hallucinogens) (4-iodo-2,5-dimethoxy-N-(2-methoxybenzyl)phenethylamine)”, 2014.
is extremely important in forensic practice, as positives should [8] E. Pasomsub, C. Sukasem, S. Sungkanuparph, B. Kijsirikul, W.
not be missed under any circumstances. Table I and II indicate Chantratita, “The Application of Artificial Neural Networks for
that all ANN systems presented in this study have very good Phenotypic Drug Resistance Prediction: Evaluation and Comparison
sensitivity: all NBOMe hallucinogens are recognized without with Other Interpretation Systems.”, Jpn. J. Infect. Dis., 63, 2010, pp.
87-94.
exception (TPR = 100%, FNR = 0%).
[9] W. Ze, X.-C. Li, W.-X. Zhu, “Prediction of drug bioavailability by
The 74_struct_ANN and 44_imp_struct_ANN networks genetic algorithm and artificial neural network”, Yao Xue Xue Bao, 41,
have the same classification of negatives (TNR = 83.33%). 2006, pp. 1180-1183.
This two networks were able to classify all the samples [10] Y.C. Sun, Y.X. Peng, Y.X. Chen, A.J. Shukla, “Application of artificial
subjected to classification (CR = 99.38%). neural networks in the design of controlled release drug delivery
systems”, Adv. Drug Deliv. Rev., 55, 2003, pp. 1201-1215.
Interestingly, the analysis of the figures of merit [11] E. Byvatov, U. Fechner, J. Sadowski, G. Schneider, “Comparison of
determined for the 34_imp_struct_ANN network indicate that support vector machine and artificial neural network systems for
reducing the number of input variables (selected according to drug/nondrug classification”, J. Chem. Inf. Comput. Sci., 43, 2003, pp.
1882-1889.
their importance) generates a higher selectivity (TNR =
[12] M. Karelson, D. Dobchev, “Using artificial neural networks to predict
88.97%). It increases by 5.64% as compared to the ANNs cell-penetrating compounds”, Expert Opin. Drug Discov., 6, 2011, pp.
presented above. In this case also, the input variable selection 783-796.
has increased the CR to 100%. [13] S. Gosav, M. Praisler, D.O. Dorohoi, G. Popa, “Structure – Activity
Correlations for Illicit Amphetamines Using ANN and Constitutional
Another important observation is that by reducing to less Descriptors”, Talanta, 70, 2006, pp. 922-928.
than half the number of input variables of ANNs created with [14] S. Gosav, M. Praisler, M. L. Birsa, “Principal Component Analysis
only the most important descriptors, the results are visibly Coupled with Artificial Neural Networks—A Combined Technique
improved. The validation process indicates a very good Classifying Small Molecular Structures Using a Concatenated Spectral
selectivity (TNR = 86.21%) for the 24_imp_struct_ANN Database”, Int. J. Mol. Sci., 12, 2011, pp. 6668-6684.
network. The 34_imp_struct_ANN network and [15] S. Gosav, M. Praisler, “Artificial Neural Networks Built for the
24_imp_struct_ANN networks are both very selective, with Recognition of Illicit Amphetamines Using a Concatenated Database”,
Rom. Rep. Phys., 54, 2009, pp. 929-935.
TNR > 86% and CR= 100%. Their ACC differ by 2,5%
[16] S. Gosav, M. Praisler, D.O. Dorohoi, “ANN Expert System Screening
(CCR=90% for 34_imp_struct_ANN and CCR=87.5% for for Illicit Amphetamines using Molecular Descriptors”, J. Mol. Struct.,
24_imp_struct_ANN). 834, 2007, pp. 188-194.
In conclusion, the 34_imp_struct_ANN is the most [17] Hyperchem software, Version 8.0.3., Hyper Co., USA, 2007.
efficient screening system. This network can be used [18] R. Todeschini, V. Consonni, Handbook of molecular descriptors.
Weinheim: Wiley-VCH, 2000.
successfully to predict and estimate the toxicity of novel
compounds having a molecular structure similar to NBOMe
hallucinogens. Its use saves analytical time and diminishes the
cost of toxicity studies.