Chemoinformatics in Drug Design. Artificial Neural Networks for

Simultaneous Prediction of Anti-enterococci Activities and

Toxicological Profiles

Alejandro Speck-Planche and M. N. D. S. Cordeiro

REQUIMTE/Department of Chemistry and Biochemistry, University of Porto, 4169-007 Porto, Portugal

Keywords: Artificial Neural Networks, Enterococci, Inhibitors, Toxicity, Topological Indices, mtk-QSBER, BC-3781.

Abstract: Enterococci are dangerous opportunistic pathogens which are responsible of a huge number of nosocomial

infections, displaying intrinsic resistance to many antibiotics. The battle against enterococci by using

antimicrobial chemotherapies will depend on the design of new antibacterial agents with high activity and

low toxicity. Multi-target methodologies focused on quantitative-structure activity relationships (mt-

QSAR), have contributed to rationalize the process of drug discovery, improving the knowledge about the

molecular patterns related with antimicrobial activity. Until know, almost all mt-QSAR models have

considered the study of biological activity or toxicity separately. Here, we developed a unified mtk-QSBER

(multitasking quantitative-structure biological effect relationships) model for simultaneous prediction of

anti-enterococci activity and toxicity on laboratory animal and human immune cells. The mtk-QSBER

model was created by using artificial neural network (ANN) analysis combined with topological indices,

with the aim of classifying compounds as positive (high biological activity and/or low toxicity) or negative

(otherwise) under multiple experimental conditions. The mtk-QSBER model correctly classified more than

90% of the whole dataset (more than 10900 cases). We used the model to predict multiple biological effects

of the investigational drug BC-3781. Results demonstrate that our mtk-QSBER may represent a new

horizon for the discovery of desirable anti-enterococci drugs.

1 INTRODUCTION

The genus Enterococcus is formed by a group of

low-GC Gram-positive, catalase-negative, non-

spore-forming, facultative anaerobic bacteria that

can occur both, as single cocci and in chains (Fisher

and Phillips, 2009). Several species belonging to

Enterococcus spp. are opportunistic pathogens

which constitute the major cause of nosocomial

infections such as bacteremia, bacterial endocarditis,

diverticulitis, meningitis and urinary tract infections

(Ryan and Ray, 2004). The successful elimination of

infections produced by enterococci will depend on

two very important aspects: the efficiency of the

antimicrobial chemotherapies used against the

infection and the safety of the drugs for human

health.

Antimicrobial chemotherapies against

Enterococcus spp. are focused on the use of the β-

lactam antibiotic ampicillin or combination of a cell

wall–active agent (such as ampicillin or

vancomycin) with aminoglycosides (gentamicin,

tobramycin), which may result in synergistic

bactericidal activity against enterococci (Ryan and

Ray, 2004). However, enterococci are intrinsically

resistant to a broad range of antibiotics commonly

used in the hospital setting, which explains in some

way, the high prevalence of these bacteria in

nosocomial infections (Brachman and Abrutyn,

2009). The most alarming aspect in enterococci is

that they are reservoirs for antibiotic resistance

genes, as may be exemplified by their ability to

transfer vancomycin resistance to methicillin-

resistant Staphylococcus aureus (MRSA), for which

vancomycin remains the last therapeutic alternative

(Figure 1). For this reason, there is an increasing

necessity for the search of new, potent and more

efficient antibacterial chemotherapies against

enterococci. On the other hand, when any

antibacterial drug is designed, serious concerns are

expected due to its appearance of toxic effects on

human health. Thus, many trials are carried out on

laboratory animals.

In this sense, Mus musculus and Rattus

458

Speck-Planche A. and Cordeiro M..

Chemoinformatics in Drug Design. Artiﬁcial Neural Networks for Simultaneous Prediction of Anti-enterococci Activities and Toxicological Proﬁles.

DOI: 10.5220/0004542004580465

In Proceedings of the 5th International Joint Conference on Computational Intelligence (NCTA-2013), pages 458-465

ISBN: 978-989-8565-77-8

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

norvegicus are the most valuable species (Hau and

Schapiro, 2011), suffering as consequence of endless

batteries of toxicity tests. At the same time, the study

of the effects of chemicals on human immune

system cells is also very important because these are

the lines of defense of the human body, protecting it

against the entry of any foreign agent (Flaherty,

2012).

Figure 1: Vancomycin: one of the most powerful broad

spectrum antibacterial drugs.

In the last six years, several researchers have

emphasized the use of multi-target for quantitative-

structure activity relationships (mt-QSAR), which

have emerged as very useful tools for rational design

and virtual screening of compounds with dissimilar

biological activities, by considering many different

biological targets (biomolecules, cell lines, tissues,

organisms) (Munteanu et al., 2009; Prado-Prado et

al., 2009; Speck-Planche and Kleandrova, 2012;

Speck-Planche et al., 2012), to the assessment of

pharmacological/toxicological profiles in multiple

assay conditions (Speck-Planche et al., 2013;

Tenorio-Borroto et al., 2012).

Nowadays, no methodology has been reported

for the prediction of anti-enterococci activity and

toxicity at the same time. Furthermore, sometimes,

non-linear modeling by using pattern recognition

methods such as Artificial Neural Networks (ANN)

(Prado-Prado et al., 2010; Tenorio-Borroto et al.,

2012), should be considered in order to find better

relationships between the molecular descriptors

describing the chemical structure of the compounds,

and their biological activities and/or toxicities. Thus,

with the objective to reduce the high costs of

experimentation, in this work we introduce the first

unified multitasking model based on quantitative-

structure biological effect relationships (mtk-

QSBER) and ANN analysis, for the simultaneous

prediction of anti-enterococci activities and

toxicological profiles in multiple assay conditions.

2 MATERIALS AND METHODS

2.1 Topological Indices

Molecular descriptors have served as essential

support for the development and consolidation of

important disciplines such as chemoinformatics

(Oprea, 2005). Among them, topological indices

(TIs) have been very useful to correlate the chemical

structure of compounds with the pharmacological

activity (QSAR) or with the toxicity (QSTR)

(Todeschini and Consonni, 2009). Descriptors like

TIs can be considered as numerical parameters of a

graph which characterize its topology, being graph

invariants, i.e., they will never depend on how the

graph (molecule) will be drawn and/or enumerated

(Todeschini and Consonni, 2009). Then, the

topology of a molecule can be studied in terms of its

size (volume), molecular accessibility, shape,

electronic factors and many other properties. For

development of this work, we selected some of the

classical TIs which include valence connectivity

indices (Kier and Hall, 1986), bond connectivity

indices (Estrada, 1995), and Balaban index

(Balaban, 1982).

2.2 Dataset: Calculation of the

Descriptors and Development of the

mtk-QSBER Model

One of the main factors to take into account for the

development of a predictive model is the use of an

appropriate dataset. In this sense, we extracted a raw

dataset from the large and highly accurate database

CHEMBL (Gaulton et al., 2012), which is available

at http://www.ebi.ac.uk/chembldb/. We retrieved

13073 endpoints of different biological effects

reported for more than 9000 compounds. With the

aim of reducing the uncertainty of the data, we

deleted all the endpoints with missing values or units

of biological effects. After that, our dataset was

formed by N

= 8560 compounds, being some of

them tested against more than one experimental

condition c

. For this reason, the dataset contained

10918 statistical cases. To develop the mtk-QSBER

model, we employed a similar methodology to that

reported for the for the simultaneous modeling of

antituberculosis activity and toxic effects on

laboratory animals (Speck-Planche et al., 2013). As

stated in these previous works, any set of

ChemoinformaticsinDrugDesign.ArtificialNeuralNetworksforSimultaneousPredictionofAnti-enterococciActivities

andToxicologicalProfiles

459

experimental conditions c

by which a compound is

tested, can be expressed as an ontology c

=> (m

, b

, l

). In this ontology, m

represents the measure of

biological effect (anti-enterococci activities or

toxicity). The element b

is referred to different

biological targets such as enterococci, Mus musculus

and Rattus norvegicus, and human immune system

cells (lymphocytes). For all biological targets,

information about different strains was taken into

consideration. The element a

defines specific

information regarding a test, i.e., if an assay is

focused on the study of functional (F) or

pharmacokinectic/pharmacodynamic profiles (A).

The term l

is the level of curation or verification of

the experimental information provided by a

particular test. The elements m

, b

, a

, and l

define

the four conditions which can change in our dataset.

So, we had N = 10918 cases from N

= 8560

compounds mentioned above, where the

experiments were performed using at least one out

of Nm

= 18 measures of biological effects, against

at least one out of Nb

= 131 biological targets, in

one out of Na

= 2 different types of assay

information, with at least one out of Nl

= 3 levels of

curation of the experimental information. In the case

of the element m

, we had diverse measures of

biological effects which were expressed in different

units. For this reason, all values of antibacterial

activity against enterococci were converted to nmol/l

(nM), while all toxicity values associated with

laboratory animals were expressed in umol/kg

(micromoles per kilograms). In both kinds of

conversions, it was necessary to divide the value of

each compound by its molar mass, and after multiply

by a factor (usually 10

or 10

). We realized these

transformations in order to make a better

interpretation of the biological data which permitted

us a more rigorous comparison between the

biological effects of any two compounds, measured

under exactly the same set of conditions c

. Data

associated with cytotoxicity against immune cells,

remained in nM. These transformations together

with the element l

, also contribute significantly to

reduce and control data uncertainty. All cases in our

dataset were assigned to 1 out of 2 possible groups

related with the biological effect of a defined

compound i in a specific condition c

[BE

)]. Then,

any compound was considered as positive [BE

) =

1] when it had high anti-enterococci activity, or any

desirable toxicological profile, otherwise, the

compound was considered as negative [BE

) =

1]. All assignments were realized taking into

account certain cutoff values of biological effects

which are depicted in Table 1. For the whole dataset,

we used a file containing the SMILES of the

compounds/cases. Calculation of TIs using SMILES

was performed with the software MODESLAB

version 1.5 (Estrada and Gutiérrez, 2002-2004). Our

intention is to predict the biological effect of any

compound depending on the molecular structure and

the experimental conditions c

. For this reason if we

use the original TIs calculated above, they will not

discriminate the biological effect for a given

compound by varying the different conditions c

. To

achieve that goal, and inspired by the use of the

moving average approach (MAA) (Hill and Lewicki,

2006), we introduced new sets of molecular

descriptors like TIs which can be defined according

to the following equation:

ΔTI

) = TI

– avgTI

) (1)

In Eq. 1, the descriptor avgTI

) characterizes each

set G of compounds tested under the same

experimental condition c

, being calculated as the

sum of all the TI

values for compounds in a subset

of G, which were considered as positive cases

[BE

) = 1] in the same element of the ontology

(experimental condition) c

. For example, in the case

of the element b

, the descriptor avgTI

) for a set G

of compounds tested against a defined target b

(bacterial strain, immune cell, etc), was calculated as

the average of the TI

by considering only the subset

of G, i.e., those compounds which were considered

as positive [BE

) = 1]. A similar procedure was

carried out for the elements m

, a

, and l

. Anyway,

in Eq. 1, the most important element is the

descriptor ΔTI

), which considers both, the

molecular structure and the experimental conditions

. For this reason, descriptors of the form ΔTI

)

(120 in total) were used to develop the mtk-QSBER

model. These descriptors represent the deviation (in

structural terms) of a compound from the positive

compounds. The CHEMBL codes, SMILES and

other relevant experimental data for all the

compounds used in this work, appear in the

Supplementary Information 1 (Suppl. Inf. 1) file.

Our dataset of 10918 cases was randomly split into

two series: training and prediction sets. The training

set was used to construct the mtk-QSBER model.

This was formed by 8298 cases, with 4217 of

them considered as positive and 4081 negative. The

prediction set was used for validation of the model

and assessment of its predictive power, being

composed by 2620 cases, 1353 positive and 1267

negative cases. Taking into consideration that large

number of molecular descriptors, we used a

combination of the attribute evaluator

CFsSubsetEval and the search algorithms called

IJCCI2013-InternationalJointConferenceonComputationalIntelligence

460

Table 1: Cutoff values for diverse measures of biological effects.

Standard measure

(units)

Biological profile Description Cutoff

(nM) Cytotoxicity Concentration required to reduce cell viability by 50% ≥70000

(umol/kg)

In vivo antibacterial

activity

Concentration required to produce a specific effect in

half of an animal population comprising a test sample

≤14.53

(nM) Antibacterial activity

Concentration required to inhibit the growth of a

microorganism by 50%

≤836.21

(umol/kg)im Toxicity Lethal dose at 50% after intramuscular administration ≥960

(umol/kg)ip Toxicity Lethal dose at 50% after intraperitoneal administration ≥1050

(umol/kg)iv Toxicity Lethal dose at 50% after intravenous administration ≥502.12

(umol/kg)oral Toxicity Lethal dose at 50% after oral administration ≥1110

(umol/kg)sc Toxicity Lethal dose at 50% after subcutaneous administration ≥713.58

MBC (nM) Antibacterial activity Concentration required to kill 100% of microorganisms ≤11040.2

MIC (nM) Antibacterial activity

Lowest concentration that prevents the visible growth of

a microorganism

≤6000

MIC

(nM) Antibacterial activity

Minimum inhibitory concentration required to inhibit

the growth of 50% of microorganisms

≤2112.21

MIC

(nM) Antibacterial activity

Minimum inhibitory concentration required to inhibit

the growth of 90% of microorganisms

≤4982.18

(umol/kg)ip Toxicity

Dose causing a neurological deficit in 50% of

organisms after intraperitoneal administration

≥239.71

(umol/kg)oral Toxicity

Dose causing a neurological deficit in 50% of

organisms after oral administration

≥375.52

Activity/Toxicity Protective index ≥4.9

(umol/kg)ip Toxicity

Dose at which toxicity occurs in 50% of organisms after

intraperitoneal administration

≥395.21

(umol/kg)oral Toxicity

Dose at which toxicity occurs in 50% of organisms after

oral administration

≥632.5

(umol/kg)sc Toxicity

Dose at which toxicity occurs in 50% of organisms after

subcutaneous administration

≥1144.41

Referred to the element m

of the ontology c

Necessary condition for considering a compound as positive.

BestFirst and GeneticSearch, all of them

implemented in the program WEKA version 3.6.9

(Hall et al., 1999-2013). The purpose was to reduce

the dimensionality, i.e., the number of molecular

descriptors. We took into account that at least one

descriptor representing each element of the ontology

, must be selected. To seek the best mtk-QSBER

model, ANN analysis was performed using the

software STATISTICA 6.0 (StatSoft, 2001). In order

to select the most important descriptors, a sensitivity

analysis was performed. In this sense, the neural

network module of STATISTICA has defined a

missing value substitution procedure, which is used

to allow predictions to be made in the absence of

values for one or more input variables (StatSoft,

2001). Thus, to define the sensitivity of a particular

input variable (descriptor) v, each ANN is run on a

defined set of cases (training cases), where a

network error is accumulated (Hill and Lewicki,

2006). After, the network is run again using the

same cases, but this time replacing the observed

values of v with the value estimated by the missing

value procedure. So, a new network error is

accumulated. Taking into consideration that some

information that each network uses, has effectively

been removed (i.e. one of the input variables), it is

logical to expect some deterioration in error to occur

(Hill and Lewicki, 2006). Then, the sensitivity of

any input variable is calculated as the ratio of the

error with missing value substitution to the original

error. The more sensitive the network is to a

particular input variable (descriptor), the greater the

deterioration we can expect, and therefore the

greater the ratio. This procedure used to detect the

relative importance of a variable, is efficiently

implemented in STATISTICA 6.0 (StatSoft, 2001).

We need to emphasize that only the variables with

high sensitivity values (>1) were selected, and we

ensured that at least one variable belonging to each

element of the ontology c

was among the chosen

variables in the final model. The quality and

predictive power of our mtk-QSAR model by

examining some statistical indices such as the

sensitivity (Sens) and specificity (Spec), the

Mathew's correlation coefficient (MCC), and the

areas under the receiver operating characteristic

ChemoinformaticsinDrugDesign.ArtificialNeuralNetworksforSimultaneousPredictionofAnti-enterococciActivities

andToxicologicalProfiles

461

(ROC) curves (Speck-Planche et al., 2012) in both,

training and prediction sets. When the analyst does

not know the system a priori, very sophisticated

methods to seek the best descriptors and optimize of

the neural networks may be needed. However,

taking in mind that the dataset was rigorously

curated, and that descriptors of type ΔTI

) can

phenomenologically explain the structural variation

in the dataset, simple rules for optimizing neural

network can be applied. Thus, the Intelligent

Problem Solver was used to seek the best networks.

This module provides the search for the best models,

and through an internal algorithm, it considers the

maximum number of hidden units, based on the

number of variables and cases (for each type of

network architecture) (StatSoft, 2001).

The first five runs served to determine the best

type of neural network. After, the number of hidden

units of the network selected as the best was

employed as maximum number of hidden units in

five new runs. The process was repeated until a

network had enough small number of hidden units

and the total percentage of correct classification

(accuracy) of cases was ≥90% in both training and

prediction sets.

3 RESULTS AND DISCUSSION

For the selection of the best mtk-QSBER model we

analyzed the different types of ANNs. These were

linear neural network (LNN), probabilistic neural

network (PNN), radial basis function (RBF), and

multilayer perceptron (MLP). We also took into

consideration the principle of parsimony, which

means that the model with the highest statistical

quality, but having as few variables as possible,

should be selected. As depicted in Table 2, the best

mtk-QSBER model found by us was that associated

with the RBF-ANN, which displays the highest

performance in terms of sensitivity and specificity,

with the lowest errors when compared with the other

ANNs. The profile of this ANN is: RBF 5:5-767-

1:1. The symbologies for all the descriptors together

with their corresponding meanings appear

represented in Table 3. Our mtk-QSBER model,

could correctly classify 7740 out of 8298 cases were

correctly classified, for an accuracy of 93.28% in

the training set, while in prediction set, 2395 out of

2620 cases were correctly classified and the

value of accuracy was 91.41%. More details about

the results of classification and predictions can be

found in Table 4 and Supplementary Information 2

(Suppl. Inf. 2) file respectively. All the average

descriptors used in this work, together with the

percentages of correct classification depending on

the elements m

, b

, a

, and l

, can be found in

Supplementary Information 3 (Suppl. Inf. 3) .

The values of areas under ROC curves played an

important role to confirm the quality and the

predictive power of the model. The values of area

under the ROC curve were 0.981 and 0.965 for

training and prediction sets respectively (Figure 2).

These values of area can be interpreted as follows:

value of area 0.981, means that a randomly selected

compound or case from the active group (protein

inhibitor) will have a larger value of probability than

a randomly selected compound or case from the

inactive group, 98.1% of the times. A similar

deduction can be made from the area under the ROC

curve for the case of the prediction set. We are

demonstrating that our mtk-QSBER model is not a

random classifier because the areas under the ROC

curves are clearly different from those obtained by

random classifiers (area = 0.5). By analyzing the

results of Table 4 and the values of areas under the

ROC curves, we can say that our mtk-QSBER model

has excellent quality and predictive power which is

comparable with other reports in the literature

related to the use of the mt-methodologies combined

with ANN analysis (Prado-Prado et al., 2010;

Tenorio-Borroto et al., 2012). The use of classical

TIs permits to obtain simple substructural and

physicochemical information. One important aspect

is that all descriptors employed to construct the mtk-

Table 2: Performance of the different ANNs.

CHARACTERISTICS

Symbology LNN MLP (TLP)

MLP (FLP)

RBF PNN

Profile

5:5-1:1 5:5-8-1:1 5:5-7-10-1:1

5:5-767-1:1

5:5-8298-2-2:1

Training set

Sens (%)

58.88 72.11 77.35

92.98

95.04

Spec (%)

58.61 72.38 77.82

93.58

27.96

Prediction set

Sens (%)

58.17 72.28 77.01

90.76

94.38

Spec (%)

61.01 73.72 78.06

92.11

28.02

Abbreviation for three layer perceptron ANN.

Nomenclature referred to four layer perceptron ANN.

IJCCI2013-InternationalJointConferenceonComputationalIntelligence

462

Table 3: Descriptors used to construct the mtk-QSBER model.

Descriptor Concept

Δ[

)]

Deviation of the vertex connectivity index of order 3 and type path, dependent on the molecular structure and the

measure of biological effect

Δ[

)]

Deviation of the bond connectivity index of order 2 and type path, dependent on the molecular structure and the

biological target

Δ[

)]

Deviation of the bond connectivity index of order 6 and type chain (ring), dependent on the molecular structure and

the biological target

ΔJ(a

) Deviation of the Balaban index, dependent on the molecular structure and the assay information

Δ[

)]

Deviation of the bond connectivity index of order 5 and type chain (ring), dependent on the molecular structure and

the level of curation of the experimental information

QSBER model have the form Δ

). These

descriptors can be considered as measures of the

similarity/dissimilarity of a given compound respect

a group of positive cases depending on the molecular

structure, and a specific element of the ontology c

, b

, a

, or l

). Thus, the descriptor Δ[

)]

encodes information related with the molecular

accessibility in those regions which contain linear

fragments formed by three bonds (Estrada, 2002).

This variable takes into consideration the structure of

the molecule and the measure of biological effect

which was used for that molecule. The variable

Δ[

)], is strongly related with the molecular

volume in linear substructures containing two bonds

(Estrada, 1995).

Figure 2: ROC curves for the mtk-QSBER model.

A similar physicochemical information is

encoded by the descriptor Δ[

)], but with the

difference that only regions formed by six-

membered rings are taken into account. The variable

Δ[

)] as well as Δ[

)] depend on the

chemical structure and the biological target against

which a compound was tested. The variable, ΔJ(a

)

is focused on the global shape (Balaban, 1982),

depending on the structure of the compound and the

assay information. Finally, Δ[

)] will depend on

the molecular structure (considering heteroatoms)

and the level or degree of curation of the

experimental information, and its structural

information will be concerned with the molecular

volume in those regions with five-membered rings.

Any model should be able to predict compounds

which were not used either training or prediction

sets. For this reason, in order to show how our mtk-

QSBER model works, we predict the effects of the

antibiotic BC-3781 against enterococci, as well as

different toxicological profiles under diverse

experimental conditions. BC-3781 is an

investigational drug (Figure 3), which has being

studied as a broad spectrum antibacterial agent due

to its activity against Gram-positive cocci,

Haemophilus influenzae, and many other bacteria

which cause serious skin infections, bacterial

pneumonia or opportunistic infections. BC-3781 has

been obtained by Nabriva Therapeutics (Sader et al.,

2012), a company focused on developing new class

of antibiotics against serious bacterial infections.

Table 4: Results of classification.

Classification

Training set Prediction set

Positive Negative Positive Negative

Total

4217 4081 1353 1267

Correct

3921 3819 1228 1167

Wrong

296 262 125 100

Correct (%)

92.98 93.58 90.76 92.11

Wrong (%)

7.02 6.42 9.24 7.89

Acc (%)

93.28 91.41

MCC

0.866 0.828

Compounds which were correctly classified by the model.

Formally known as sensitivity (Sens) for positive cases and

specificity (Spec) for negative.

Referred to the accuracy as total percentage of correct

classification.

All information regarding this systemic product

can be found at http://www.nabriva.com/.

ChemoinformaticsinDrugDesign.ArtificialNeuralNetworksforSimultaneousPredictionofAnti-enterococciActivities

andToxicologicalProfiles

463

Predictions performed by the mtk-QSBER model are

available in the Supplementary Information 4

(Suppl. Inf. 4) file. We need to emphasize that

predictions were realized against the most important

enterococci, i.e., Enterococcus faecalis and

Enterococcus faecium, which are the principal

bacteria of causing nosocomial infections.

According to the reports available for MIC

and

MIC

values in the webpage of Nabriva

Therapeutics, and the reference 41, BC-3781 may be

used to treat infections caused by Enterococcus

faecium, but not Enterococcus faecalis. Thus,

predictions made by the mtk-QSBER model,

confirm the experimental results. Also, in Suppl. Inf.

4, we performed predictions focused on the

toxicological profiles. According to the different

cutoff values of toxicities reported in Table 1, BC-

3781 can be a very safe antibacterial agent. Our

predictions help to explain why this pleuromutilin

derivative has undergone phase II clinical trials with

positive results. At the same time, we are

demonstrating that our mtk-QSBER model can be

used for virtual screening of toxicologically safe

anti-enterococci agents.

Figure 3: Chemical structure of the promising antibiotic

BC-3781.

4 CONCLUSIONS

Mt-QSAR approaches have emerged as novel and

powerful alternatives in the field of computer-aided

drug design, displaying very good performance for

the modeling of many different biological activities,

against diverse biological targets and experimental

conditions. In our work, we extended the mt-QSAR

concept by constructing an mtk-QSBER model that

allowed us to include not only biological (anti-

enterococci) activity data, but also, toxicological

profiles over several biological entities. Thus, our

mtk-QSBER model was developed to perform

simultaneous prediction of antibacterial activity

against bacteria of the genus Enterococcus spp. and

toxicity of compounds on laboratory animals and

human lymphocytes. The present mtk-QSBER

model confirms the idea that the use of mt-QSAR

methodologies permits to obtain more realistic and

accurate results. The performance of our mtk-

QSBER model, by classifying compounds as

positive or negative from a large and heterogeneous

database of compounds, and depending on dissimilar

measures of biological effects, targets, and

reliabilities of experimental conditions, permits its

use with one essential purpose: discovery of novel,

potent, versatile and safe anti-enterococci drug

candidates.

ACKNOWLEDGEMENTS

A. Speck-Planche acknowledges the Portuguese

Fundação para a Ciência e a Tecnologia (FCT) and

the European Social Found for financial support

(SFRH/BD/77690/2011).

REFERENCES

Balaban, A. T., 1982. Highly discriminating distance-

based topological index. Chemical Physics Letters.

89(5):399–404.

Brachman, P. S., Abrutyn, E., 2009. Bacterial Infections of

Humans: Epidemiology and Control, Springer

Science+Business Media, LLC. New York, NY.

Estrada, E., 1995. Edge adjacency relationship and a novel

topological index related to molecular volume.

Journal of Chemical Information and Computer

Sciences. 35:31–33.

Estrada, E., 2002. Physicochemical Interpretation of

Molecular Connectivity Indices. The Journal of

Physical Chemistry A. 106(39):9085–9091.

Estrada, E., Gutiérrez, Y., 2002-2004. MODESLAB, v1.5.

Santiago de Compostela.

Fisher, K., Phillips, C., 2009. The ecology, epidemiology

and virulence of Enterococcus. Microbiology. 155(Pt

6):1749-1757.

Flaherty, D., 2012. Immunology for Pharmacy, Mosby,

Elsevier Inc. St. Louis, Missouri.

Gaulton, A., Bellis, L. J., Bento, A. P., Chambers, J.,

Davies, M., Hersey, A., Light, Y., McGlinchey, S.,

Michalovich, D., Al-Lazikani, B., Overington, J. P.,

2012. ChEMBL: a large-scale bioactivity database for

drug discovery. Nucleic Acids Research. 40:D1100-

1107.

Hall, M., Frank, E., Holmes, G., Pfahringer, B.,

Reutemann, P., Witten, I. H., 1999-2013. WEKA

3.6.9. Waikato Environment for Knowledge Analysis.

Hamilton.

Hau, J., Schapiro, S. J., 2011. Handbook of Laboratory

Animal Science: Essential Principles and Practices,

CRC Press, Taylor & Francis Group, LLC. Boca

Raton, FL.

Hill, T., Lewicki, P., 2006. STATISTICS Methods and

IJCCI2013-InternationalJointConferenceonComputationalIntelligence

464

Applications. A Comprehensive Reference for Science,

Industry and Data Mining, StatSoft. Tulsa.

Kier, L. B., Hall, L. H., 1986. Molecular connectivity in

structure-activity analysis, Research Studies Press,

Wiley Letchworth, Hertfordshire, England, New York.

Munteanu, C. R., Magalhaes, A. L., Uriarte, E., Gonzalez-

Diaz, H., 2009. Multi-target QPDR classification

model for human breast and colon cancer-related

proteins using star graph topological indices. Journal

Theoretical Biology. 257(2):303-311.

Oprea, T., 2005. Chemoinformatics in Drug Discovery,

WILEY-VCH Verlag GmbH & Co. KGaA.

Weinheim.

Prado-Prado, F. J., Borges, F., Perez-Montoto, L. G.,

Gonzalez-Diaz, H., 2009. Multi-target spectral

moment: QSAR for antifungal drugs vs. different

fungi species. European Journal Medicinal Chemistry.

44(10):4051-4056.

Prado-Prado, F. J., Garcia-Mera, X., Gonzalez-Diaz, H.,

2010. Multi-target spectral moment QSAR versus

ANN for antiparasitic drugs against different parasite

species. Bioorganic and Medicinal Chemistry.

18(6):2225-2231.

Ryan, K. J., Ray, C. G., 2004. Sherris Medical

Microbiology McGraw Hill. Arizona.

Sader, H. S., Biedenbach, D. J., Paukner, S., Ivezic-

Schoenfeld, Z., Jones, R. N., 2012. Antimicrobial

activity of the investigational pleuromutilin compound

BC-3781 tested against Gram-positive organisms

commonly associated with acute bacterial skin and

skin structure infections. Antimicrobial Agents and

Chemotherapy. 56(3):1619-1623.

Speck-Planche, A., Kleandrova, V. V., 2012. In silico

design of multi-target inhibitors for C-C chemokine

receptors using substructural descriptors. Molecular

Diversity. 16(1):183-191.

Speck-Planche, A., Kleandrova, V. V., Cordeiro, M. N. D.

S., 2013. New insights toward the discovery of

antibacterial agents: Multi-tasking QSBER model for

the simultaneous prediction of anti-tuberculosis

activity and toxicological profiles of drugs. European

Journal Pharmaceutical Sciences. 48(4-5):812-818.

Speck-Planche, A., Kleandrova, V. V., Luan, F., Cordeiro,

M. N. D. S., 2012. Rational drug design for anti-

cancer chemotherapy: multi-target QSAR models for

the in silico discovery of anti-colorectal cancer agents.

Bioorganic and Medicinal Chemistry. 20(15):4848-

4855.

StatSoft, 2001. STATISTICA 6.0. Data analysis software

system.

Tenorio-Borroto, E., Penuelas Rivas, C. G., Vasquez

Chagoyan, J. C., Castanedo, N., Prado-Prado, F. J.,

Garcia-Mera, X., Gonzalez-Diaz, H., 2012. ANN

multiplexing model of drugs effect on macrophages;

theoretical and flow cytometry study on the

cytotoxicity of the anti-microbial drug G1 in spleen.

Bioorganic and Medicinal Chemistry. 20(20):6181-

6194.

Todeschini, R., Consonni, V., 2009. Molecular

Descriptors for Chemoinformatics, WILEY-VCH

Verlag GmbH & Co. KGaA. Weinheim.

ChemoinformaticsinDrugDesign.ArtificialNeuralNetworksforSimultaneousPredictionofAnti-enterococciActivities

andToxicologicalProfiles

465