Journal of Antimicrobial Chemotherapy (2004) 53, 65–73
DOI: 10.1093/jac/dkh014
Advance Access publication 25 November 2003
New agents active against Mycobacterium avium complex selected by
molecular topology: a virtual screening method
Ángeles García-García1, Jorge Gálvez1*, Jesus-Vicente de Julián-Ortiz2, Ramón García-Domenech1,
Carlos Muñoz3, Remedios Guna1,3 and Rafael Borrás3
1Unidad
de Investigación de Diseño de Fármacos y Conectividad Molecular, Departamento de Química Física,
Facultad de Farmacia, Universitat de València, 46100 Burjassot, Valencia; 2Institut de Química Computacional,
Universitat de Girona, 17071 Girona; 3Departamento de Microbiología, Facultad de Medicina y Odontología,
Universitat de València, Av. Blasco Ibáñez 15, 46010 Valencia, Spain
Received 27 March 2003; returned 18 July 2003; revised 25 September 2003; accepted 3 October 2003
Objectives: In order to select new drugs and to predict their in vitro activity against Mycobacterium avium
complex (MAC), new quantitative structure–activity relationship (QSAR) models were developed.
Methods: The activities against MAC of 29 structurally heterogeneous drugs were examined by means of
linear discriminant analysis (LDA) and multilinear regression analysis (MLRA) by using topological indices
(TI) as structural descriptors. In vitro antimycobacterial activities were determined by a broth microdilution
method with 7H9 medium.
Results: The topological model obtained successfully classifies over 80% of compounds as active or inactive;
consequently, it was applied in the search for new molecules active against MAC. From among the selected
candidates demonstrating in vitro activity, aflatoxin B1, benzalkonium chloride and pentamidine stand out,
with MIC50s between 4 and 32 mg/L.
Conclusion: The method described in this work is able to select molecules active against MAC.
Keywords: MAC, QSAR studies, topological indices
nitroimidazopyrans, pyridones, lipophilic analogues of isoniazid,3
thiazine derivates4 and carboxamidrazone analogues,5 showing
antimycobacterial activity are being developed. The identification of
novel targets requires the characterization of mycobacterium-specific
biochemical pathways, but the rational design of new antimycobacterial
agents is complex and many metabolic processes are unknown.
Most of the drug design methods currently available demand a
previous knowledge of the mechanism of action involved. However,
extramechanistic approaches are increasingly being used to design
new drugs. Concretely, molecular topology6 has proved a useful formalism to find quantitative structure–activity relationships (QSAR).
One of the most interesting advantages of molecular topology is the
straightforward calculation of the topological descriptors. In this
method, each structure is assimilated to a hydrogen-depleted graph
where the atoms are represented by vertices and the bonds by edges;
the connectivity between atoms is represented in topological matrices,
which can be either distance or adjacency. Mathematical manipulation of such matrices provides different sets of numbers called topological indices (TI),7–9 which characterize each molecule at different
Introduction
Mycobacterium avium complex (MAC) is a group of opportunistic
bacterial pathogens related to Mycobacterium tuberculosis that are
found almost everywhere in our environment, including food, water
and soil. The emergence of natural resistant MAC clinical isolates in
immunodeficient patients,1 and the spread of multidrug-resistant
M. tuberculosis in several world areas, emphasize the need for new
antimycobacterial drugs.
Alternative antimicrobial treatment has been employed in the
therapy of infections caused by these organisms. Agents such as fluoroquinolones, novel macrolides or rifamycin derivates, have been introduced for the control of multidrug-resistant or latent tuberculosis
infections,2 but experience shows that the use of high-cost treatment
in developing countries is limited, and, on the other hand, an accelerated
increase in resistance is observed. Thus, more extensive investigations are needed. Many other new drugs are being evaluated to obtain
efficient treatments and entirely novel classes of compounds such
as oxazolidinones, nitroimidazoles, riminophenazines, ketolides,
..................................................................................................................................................................................................................................................................
*Corresponding author. Tel: +34-963-544-891; Fax: +34-963-544-892; E-mail: [email protected]
...................................................................................................................................................................................................................................................................
65
JAC vol.53 no.1 © The British Society for Antimicrobial Chemotherapy 2003; all rights reserved.
Á. García-García et al.
descriptive structural levels.6,10 If well-chosen, these topological
descriptors can be used for the selection and design of new analgesics,11 bronchodilators,12 antihistamines,13 antivirals,14 antibacterials,15 antifungals,16 etc., many of which can be considered as lead
drugs. In a recent paper, we have developed a study of prediction of
quinolone activity against M. avium by molecular topology and
virtual computational screening.17
The aim of this study was to develop new QSAR models, based on
TI, statistical linear discriminant analysis (LDA) and multilinear
regression analysis (MLRA), in order to select new drugs and to predict their in vitro activity, expressed as MIC90 (lowest concentration
of an antimicrobial that inhibits 90% of the different strains of bacteria), against M. avium complex.
interval of values of a given function, the expectancy of activity is
Ea=a/(i+100), where a is the percentage of active compounds in the interval and i is the corresponding percentage of inactive compounds within
the same interval. The expectancy of inactivity is defined likewise as
Ei=i/(a+100). This representation provides good visualization of the
regions of minimum overlap and helps to select regions in which the
probability of finding active compounds is optimal.
Multilinear regression analysis, MLRA. The property, MIC90 against
MAC expressed in mg/L and µmol/L, as well as their respective logarithms, were correlated versus TI in order to select the best regression
equation.
The MLRA was carried out with the 9R module of the BMDP
program, which estimates regression equations for the best subsets of
predictor variables and provides detailed residual analysis by using the
Furnival-Wilson algorithm.27 Equations with minimal Mallows Cp
parameter28 were initially chosen.
The stability of the equation selected was evaluated through a crossvalidation by the leave-one-out algorithm.29 To do this, one compound of
the set is extracted, and the model is recalculated using the remaining
N–1 compounds as the training set. The property is then predicted for the
removed element. This process is repeated for all the compounds of the
set so obtaining a prediction for each one. This procedure also aids in the
detection of outlying points.
In order to examine the possible existence of fortuitous regressions,
the randomization test was adopted in this paper.30 Thus, the values of the
property of each compound are randomly permuted and linearly correlated
with the aforementioned descriptors. This process is repeated as many
times as compounds are in the set. The usual way to represent the results
of a randomization test is by plotting the correlation coefficients versus
predicted ones, r2 (squared multiple correlation) and r2cv (squared multiple
correlation by cross-validation) respectively.
Virtual screening. The equations obtained and the corresponding intervals constitute a model for the filtering of databases. If a molecule has its
output values from the equations within the thresholds, it is selected as
potentially active. Otherwise it is discarded. In this work, a homemade
database of approx. 20 000 compounds was used, extracted from the
Merck Index and the Sigma–Aldrich databases.
Materials and methods
Statistical treatment
The search for new compounds showing antimycobacterial activity was
carried out using the following steps:
Structural descriptors. A group of compounds with known anti-MAC
activity was selected from several sources.17–23 Each compound was
characterized by a set of 62 TI24 calculated using the program DesMol
(Unidad de Investigación de Diseño de Fármacos y Conectividad Molecular, Dept. de Química Física, Facultad de Farmacia, Universitat de
València, Spain). The TI used were 32 connectivity RandiM-Kier-Hall
type indices, mχt,6 and differences and quotients, mDt and mCt, between
them; 20 topological charge indices, Gm and Jm,25 and 10 other discrete
invariants.24
Linear discriminant analysis (LDA). The objective of LDA is to find
linear functions able to discriminate between two or more categories. In
our case, two groups of compounds were considered for analysis. The
first included 22 anti-MAC drugs, and 35 presumably inactive compounds made up the second. These compounds constituted the training
set. LDA was carried out with the BMDP package, module 7M (W. J.
Dixon, BMDP Statistical Software, University of California, Berkeley,
CA, USA). The selection of the descriptors was based on the FisherSnedecor F parameter. Stepwise, the variable that adds the most to the
separation of the groups is entered into (or the variable that adds the least
is removed from) the discriminant function (DF). The classification
criterion was the maximal posterior classification probability. It is basically proportional to the Mahalanobis distance from the data points to the
centroid. This is the point that represents the means for all variables in the
multivariate space defined by the group. The quality of the DF is evaluated by the Wilks parameter (λ) and the percentage of correct classifications into each group (discriminant ability). The independent variables
in this study were TI, and the discrimination property was the presence or
absence of anti-MAC activity.
The validation of the selected DF was carried out both by an internal
leave-n-out test, in which the program randomly chooses and pulls out a
number of compounds and uses them to evaluate the DFs obtained with
the remainder, and by an external test set of non-previously used data. A
randomness test, in which the classification variable was randomly reassigned and discriminant functions with the same number of variables as
DF were obtained, was also carried out in order to compare these equations with DF and detect possible chance effects.
The equation obtained by LDA must be able to predict the existence of
the antimycobacterial activity for a given structure and, consequently,
must be useful to select new candidate drugs. In order to choose intervals
of values derived from this equation in which the active compounds
showed their optimal values, a pharmacological activity distribution
diagram (PDD), was used.26 This is a histogram-like plot of the calculated
DF in which expectancies appear on the ordinate axis. For an arbitrary
Microbiological study
Thirty-two M. avium complex isolates from respiratory and non-respiratory
clinical samples from the Hospital Clínico Universitario de Valencia,
Spain, were used in this study. All isolates were identified by DNA
hybridization probes (Accuprobe, Gen Probe Inc., San Diego, CA,
USA). Susceptibility to conventional antimycobacterial drugs (azithromycin, ciprofloxacin, clarithromycin, clofazimine, ethambutol, moxifloxacin and rifabutin) and the selected drugs was carried out by a
microdilution method.31
The organisms were grown in modified Middlebrook 7H9 broth supplemented with 10% OADC (oleic acid/albumin/dextrose/catalase)
enrichment (Difco Laboratories) for 7 days at 37°C. The inoculum size
was obtained by dilution of a culture suspension in 7H9 broth to yield an
absorbance equivalent to that of a 0.5 McFarland Standard for M. avium
complex isolates.
Antimicrobial susceptibility tests were carried out in 96-well microplates using serial two-fold microdilution in 7H9 broth. Initial drug dilutions were prepared in deionized water or, if not soluble, dimethyl
sulphoxide. Subsequent two-fold dilutions were carried out in 150 µL of
modified 7H9 broth in the microplates to provide a final test range of
128 to 0.125 mg/L. Ten microlitres of a suspension of mycobacteria were
added to the wells. Plates were covered with Parafilm ‘M’ (Laboratory
Film, American National Can), and incubated for 5 days at 37°C. Starting
at 5 days of incubation, 20 µL of resazurin (Sigma 2127) with a concen-
66
New agents against Mycobacterium avium complex
Table 1. Classification by DF (Equation 1) for each compound
Compound
DF
Probability of
activity (active group)
or inactivity
Predicted
(inactive group)
activitya
Active group
Azithromycin
Capreomycin
Ciprofloxacin
Clarithromycin
Dihydrostreptomycin
Enviomycin
Ethambutol
Isoniazid
Kanamycin
Moxifloxacin
Neomycin B
Ofloxacin
Rifabutin
Rifampicin
Streptomycin
Streptonicozid
Subathizone
Thiacetazone
Tiocarlide
Tuberin
Verazide
Viomycin
1.83
2.38
1.71
1.63
1.51
2.90
0.69
–0.34
2.05
1.32
3.25
1.33
2.06
2.85
1.02
2.54
–1.73
–0.11
1.00
0.29
0.97
2.80
0.975
0.993
0.968
0.961
0.949
0.998
0.730
0.190
0.985
0.924
0.999
0.924
0.986
0.998
0.856
0.995
0.009
0.290
0.849
0.510
0.839
0.997
+
+
+
+
+
+
+
–
+
+
+
+
+
+
+
+
–
–
+
NC
+
+
Inactive group
Acifran
Acipimox
Acronine
Aldicarb
Alpidem
–1.72
–0.53
–0.36
–3.06
–0.18
0.991
0.869
0.816
1.000
0.744
–
–
–
–
–
Compound
Alprazolam
Allicin
Aminopromazine
Aminothiazole
Amitraz
Amsacrine
Antrafenine
Azacosterol
Azapicyl
Azaserine
Beclobrate
Benzoctamine
Benzoic acid
Bixin
Brilliant Blue
Bromazepam
Buspirone
Butibufen
Camazepam
Captodiamine
Carmofur
Carmustine
Carnitine
Carprofen
Clofibrate
Ornithin
Paraoxon
Piroxicam
Prazepam
Theofibrate
DF
–0.83
–0.07
–0.80
–0.17
–1.42
–0.32
–1.41
–1.51
–0.20
–0.52
–1.02
–2.51
–0.68
0.39
–0.57
–0.02
–0.16
–2.20
–0.54
1.17
–1.36
–1.96
–0.99
–0.92
–1.63
–1.43
–1.53
–1.57
–0.46
–0.85
Probability of
activity (active group)
or inactivity
Predicted
(inactive group)
activitya
0.931
0.691
0.927
0.739
0.982
0.800
0.981
0.985
0.753
0.865
0.954
0.999
0.905
0.427
0.880
0.667
0.735
0.997
0.873
0.106
0.979
0.995
0.951
0.943
0.989
0.983
0.986
0.987
0.848
0.934
–
–
–
–
–
–
–
–
–
–
–
–
–
NC
–
–
–
–
–
+
–
–
–
–
–
–
–
–
–
–
aWhen the probability of activity or probability of inactivity is >0.60, the compound is classified as active (+) or inactive (–), respectively. For any other case,
the compound is non-classified (NC).
cluster sub-fragments, which decrease the possibility of being active;
the topological charge-transfer indices G1v and J3v, which are measures
of the contribution of molecular topological structure to the charge
transfer at topological distance 1 and 3, respectively; the quotient
index 4Cc = 4χc/4χcv, where 4χc and 4χcv are, respectively, single and
valence RandiM-Kier-Hall indices of order 4 and cluster-type. It is
related to the four-cluster fragment electronic densities and its magnitude diminishes the activity profile. The dominant term in Equation 1
is the contribution of G1v, which is the only one that it is positive
and can be considered as the total molecular electronic density. Its
maximum value corresponds to the viomycins, neomycin B and
rifampicin that also exhibit the greater DF value. Conversely isoniazid
and tuberin, having the lowest values of G1v also present low DF,
even negative for isoniazid. Charge transfers through heteroatoms
separated by three bonds are limiting in this equation. So, small molecules such as azaserine and carnitine are inactive.
Table 1 shows the results of the classification for each one of the
compounds included in the LDA. As can be seen in Table 1, the linear
equation gave good results since most compounds are classified with
tration of 250 mg/L was added to the wells, and the microplates were reincubated at 37°C for an additional period of 48 h. MIC50 (lowest
concentration of an antimicrobial that inhibits 50% of the different strains
of bacteria) and MIC90 values were determined as the lowest concentrations of the compounds yielding no visible changes from blue to pink.32
Results and discussion
Linear discrimination analysis
The compounds included in the LDA are a structurally heterogeneous
set of drugs (ranging from simple structures, such as benzoic acid, up
to much more complex structures, such as aminoglycosides). The following equation for DF, obtained by stepwise LDA, classified the
compounds as active against MAC if DF > 0 and inactive if DF < 0:
DF = 0.47 – 1.26 3χcv + 0.22 G1v – 14.53 J3v – 0.22 4Cc
N = 57; λ = 0.42; F = 17.86
Equation 1
The topological descriptors selected in this equation were: the
valence RandiM-Kier-Hall index 3χcv, related to the presence of three-
67
Á. García-García et al.
Table 2. Results of the internal validation for DF (Equation 1)
Training group
Test group
Run no.
λ
(+)
(–)
(+)
(–)
1
2
3
4
5
Average
0.27
0.43
0.43
0.51
0.35
–
88% (15/2)
80% (16/4)
84% (16/3)
80% (12/3)
90% (17/2)
84%
100% (0/27)
97% (1/28)
96% (1/27)
92% (2/24)
100% (0/29)
97%
60% (3/2)
100% (2/0)
67% (2/1)
100% (7/0)
33% (1/2)
72%
75% (2/6)
100% (0/6)
100% (0/7)
100% (0/9)
83% (1/5)
92%
DF
0.42
86%(19/3)
94%(2/33)
No
No
(a/b) = number of (+) compounds/number of (–) compounds.
Table 3. Results of the external validation for DF (Equation 1)
Compound
DF
Probability of
activity (active
group) or inactivity
(inactive group)
Predicted
activity
Active group
Clofazimin
Imipenem
Sparfloxacin
Tobramycin
Gatifloxacin
1.98
–0.31
0.85
1.95
1.28
0.98
0.20
0.80
0.98
0.92
+
–
+
+
+
Inactive group
Altretamine
Antipyrine
Benorylate
Canthaxanthin
Chloropal
Dichlone
Etifoxine
Feprazone
Genite
Glucosamine
–2.99
–0.99
1.20
–3.42
1.00
–1.68
–1.77
–0.96
–1.45
0.26
1.00
0.95
0.10
1.00
0.15
0.99
0.99
0.95
0.98
0.51
–
–
+
–
+
–
–
–
–
NC
Figure 1. PDD in the training group obtained by DF (Equation 1). E, expectancy
of activity or inactivity; solid line, active group; dashed line, inactive group. As
can be seen, the optimal DF range to find active compounds is 0 < DF < 4.
Table 5. Statistical parameters of the best equations obtained by
MLRA
P (MIC90)
N
NTI
r2
SEE
Cp
F
NC, non-classified.
mg/L
log mg/L
µmol/L
log µmol/L
29
29
29
29
6
5
6
3
0.78
0.36
0.64
0.40
2.18
0.48
13.26
0.50
–3.76
4.10
3.36
3.14
14.50
3.59
9.04
7.33
Table 4. Results of the randomness test for DF (Equation 1)
N, number of molecules correlated; NTI, number of TIs in the equation; SEE,
standard error of estimates.
Run
Selected variables
λ
% Success
actives
% Success
inactives
DF
1
2
3
4
5
3χ v, G v, J v, 4C
c
1
3
c
4χ , 4χ v, G , J
c
pc
2 3
4χ v, J v, 1D, 4D
c
1
pc
1D, 1C, 4C , PR0
p
G2, J1, J2v, 2D
4χ v, J v, 4C , PR2
p
2
pc
0.42
0.70
0.63
0.84
0.56
0.70
86
67
85
67
68
73
94
76
74
76
88
71
P value
<0.00001
<0.0109
<0.00001
<0.0006
a posterior probability of over 80%. Under this framework, the error
percentage in the active set was about 14%, whereas in the inactive it
was about 6%.
The validation on an internal set is illustrated in Table 2. Five runs
were carried out. A number of compounds, ranging from 8 to 16, were
randomly extracted from a training to a test set. Wilks λ values for
each equation are displayed. The percentages of correct classification
for training and test sets, for active and inactive compounds, are
shown. The number of compounds classified as active (+) or inactive
68
New agents against Mycobacterium avium complex
Table 6. Prediction of MIC90-MAC (mg/L) by P (Equation 2)
Compound
5-Bromosalicylhydroxamic acid
Amikacin
Azithromycin
Benzoylpas
Capreomycin
Clarithromycin
Clofazimine
Cycloserine
Dihydrostreptomycin
Enviomycin
Ethambutol
Ethionamide
Imipenem
Kanamycin
Morphazinamide
Moxifloxacin
Neomycin B
Ofloxacin
Protionamide
Rifabutin
Rifampicin
Sparfloxacin
Streptonicozid
Subathizone
Tobramycin
Tubercidin
Tuberin
Verazide
Viomycin
Figure 2. Results obtained from the stability (a) and randomness analysis (b) for
P (Equation 2). (a) Plot of the residuals obtained versus residuals (cross-validation) for P (Equation 2); (b) Plot of the randomness analysis for P (Equation 2).
(–) appears parenthetical, where (a/b) = number of (+) compounds/
number of (–) compounds. Average values are also shown, as well as
the reference equation performance.
The results for the training and test groups are within the same
range. The average percentage of success obtained with the training
group was 84% for active (+) and 97% for inactive (–). For the test, it
was 72% and 92%, respectively. The results were similar to the ones
obtained with DF equation, which points out the validity of Equation 1.
The LDA results on the external validation test are shown in Table 3.
For the active set, we found a misclassified compound, namely imipenem. The same occurs in the opposite group, where chloropal and
benorylate are misclassified.
Table 4 shows the results obtained in the randomness test for DF.
Five runs were carried out with the 22 active compounds and the
35 inactive compounds from the original sets. A number of randomly
chosen compounds, between five and eight, were changed category
in each run and new equations with four variables were obtained. As
can be seen in Table 4, the new equations obtained showed greater
λ and a lower classification ability, which indicates that Equation 1
models a true structural pattern.
Figure 1 shows the PDD obtained from DF. As can be seen, the
optimal DF range to find active compounds must be redefined
between 0 and 4. Compounds lying between –0.5 and 0 were considered non-classified (+/–) because many active compounds can be
found in this interval, although the proportion of inactive compounds
is much greater.
MICObsa
MICCalcb
10
12
5
10
5
4
2
16
5
3
8
10
4
5
15
2
8
2
10
1
1
4
5
5
8
16
10
10
8
8.11
6.05
4.92
10.59
5.1
2.46
1.41
15.86
4.58
6.37
10.55
9.32
6.03
7.63
13.21
4.49
6.15
2.86
9.29
1.57
0.34
3.78
4.71
5.6
7.58
12.86
10.8
11.09
6.15
MICCalc CVc
7.71
5.28
3.97
10.64
5.13
1.77
1.22
15.8
4.46
6.93
10.8
9.19
6.22
8.01
12.89
4.79
5.82
2.97
9.16
1.91
–0.31
3.73
4.65
6.87
7.52
12.29
10.89
11.28
5.8
aFrom
Refs. 17–23.
Equation 2.
cFrom cross-validation (CV) study.
bFrom
regression equation obtained corresponded to MIC90 expressed in
mg/L (Table 5). The activity expressed in micromolar concentration
was not so well correlated. This can be explained because some of the
descriptors used (connectivity indices, G charge indices and discrete
invariants) are obtained by summations of fragment contributions.
Thus, they are not absolutely independent from the molecular mass.
The use of molar magnitudes introduces a redundant factor. In relation to the logarithmic correlations, they usually give better results
with variables that span several magnitude orders, which is not the
case.
The selected equation and its statistical parameters were:
P = –0.59 + 43.94 4χvc–2.91G4v + 7.50G5 +
9.29 2C – 2.43 4Cc – 0.93 PR3
Equation 2
N = 29; r2 = 0.78; r2cv=0.69; SEE = 2.18; Cp = –3.76; F = 14.5;
p < 0.00001
where P is correlated activity (MIC90); N, number of cases; SEE,
standard error of estimates and p is significance. Six topological
descriptors are used in this function: the valence RandiM-Kier-Hall
index 4χvc related to the presence of four-cluster sub-fragments; the
charge indices G4v and G5, related to the total contribution of the
molecular topological structure to the charge transfers at topological
Multilinear regression analysis
After carrying out correlations with the MIC90 values (in mg/L and
µmol/L) and with their logarithms versus TI (Table 5), the best linear
69
Á. García-García et al.
Table 7. Application of the MAC model to the database
Compound
Azobencene
Aflatoxin B1
Benzalkonium chloride
Cefoperazone
Ceftriaxone
Chlortetracycline
Citrazinic acid
Cyanuric acid
Cytarabine
Cytidinsilane
Demeclocycline
DNOC
DOCA
Doxycycline
Erythrosine
Furandicarboxylate
Gallic acid
Indomethacin
Mefenamic acid
Oxytetracycline
Paromomycin
Pentamidine
Phenylsilane
Probenecid
Reserpine
Ribavirin
Riboflavin
TEPP
Theobromine
Tributyl phosphate
Trifluoperazine
Trimethylurea
Veratrine
aMolecules
DF
Classification DF
interval 0 < DF < 4
P value
–0.45
0.73
0.05
–0.67
–1.21
–2.94
–0.74
–2.88
–0.15
–2.15
–2.74
–0.26
–1.04
–3.00
–3.40
–0.52
–0.16
–0.04
–0.76
–2.58
3.35
1.37
–1.08
–2.14
3.38
–0.24
–0.08
–2.64
–2.60
–0.01
0.27
–3.10
–2.33
–
+
+
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
+
+
–
–
+
–
–
–
–
–
+
–
–
4.27
6.76
16.88
7.47
15.24
18.31
12.65
11.50
12.89
12.25
14.45
14.88
28.42
8.94
–2.92
14.17
12.94
6.79
9.76
15.85
6.58
15.44
10.59
2.44
0.41
13.43
5.41
–5.46
7.85
11.36
7.82
12.65
23.58
Classification P
interval 0 < P < 20
+
+
+
+
+
+
+
+
+
+
+
+
–
+
–
+
+
+
+
+
+
+
+
+
+
+
+
–
+
+
+
+
–
Predicted activity
by MAC modela
+/–
+
+
–
–
–
–
–
+/–
–
–
+/–
–
–
–
–
+/–
+/–
–
–
+
+
–
–
+
+/–
+/–
–
–
+/–
+
–
–
classified as ‘+/–’ pass through –0.5 < DF < 4 and 0 < P < 20.
Table 9. Experimental MICs (mg/L) of the selected compounds
against MAC
Table 8. Experimental MICs (mg/L) of the conventional
drugs against MAC
Compound
MIC range
Azithromycin
Ciprofloxacin
Clarithromycin
Clofazimine
Ethambutol
Moxifloxacin
Rifabutin
0.5–32
0.5–16
<0.125–2
0.25–8
4–16
0.25–2
<0.125–8
MIC50
8
4
1
1
8
0.5
1
MIC90
16
8
2
2
8
1
2
70
Compound
MIC range
MIC50
MIC90
Azobencene
Aflatoxin B1
Benzalkonium chloride
Cytarabine
DNOC
Gallic acid
Indomethacin
Paromomycin
Pentamidine
Reserpine
Ribavirin
Riboflavin
Tributyl phosphate
Trifluoperazine
128–>128
32–128
2–8
>128
128–>128
128–>128
>128
4–64
8–32
16–>128
32–128
>128
>128
4–16
128
32
4
>128
128
128
>128
8
16
64
64
>128
>128
8
>128
128
4
>128
>128
>128
>128
16
32
>128
128
>128
>128
8
New agents against Mycobacterium avium complex
Figure 3. Structures of the selected compounds.
distances 4 and 5; the quotient indices 2C = 2χ/2χv and 4Cc = 4χc/4χcv
and the discrete invariant PR3 (pairs of branches separated by three
bonds). Two terms are dominant in Equation 2, those corresponding
to G5 and 2C. The first is related to electronic density and is partially
compensated by the G4v term that is of the same type. Fluoroquinolones
give great values for the balance between these terms. Likewise, the
term with 2C, related to electronic transfers through two bonds, is
partially compensated by 4Cc in the case of aminoglycosides, rifabutin
and macrolides because these are the only members of the training set
that display carbons bonded to four different non-hydrogen atoms.
The observed and calculated P for each compound and the P
obtained in the cross-validation study are shown in Table 6. The prediction success of P was very satisfactory if we consider that we are
working with a non-logarithmic property, and it is also significant
that it can distinguish between low and high values of P.
The results of the cross-validation analysis (leave-one-out) are
outlined in column 4 of Table 6.
The results obtained from the stability and randomness analysis
for P (Equation 2) are shown in Figure 2. Figure 2(a) shows the stability
of the regression equation for cross-validated-residuals versus residuals, whilst Figure 2(b) illustrates the randomness analysis results.
Regressions showing r2 < 0.5 were obtained by random reassigning
P values.
It is considered that good drug candidates should display a calculated P (MIC90) lower than 20 mg/L, in accordance with that proposed
by Heifets in 1996.33 This author considers that compounds with
MIC90 values between 4 and 16 mg/L should be considered moderately active drugs, and those with values lower than 4 mg/L should be
considered as active drugs. Furthermore, some conventional drugs
such as cycloserine and capreomycin, present an interval of susceptibility in vitro between 5 and 20 mg/L.
71
Á. García-García et al.
Topological mathematical model and virtual screening
Concluding remarks
The model to filter potential active molecules for database searches
against MAC is constituted by the equations DF (Equation 1) and
P (Equation 2) with their corresponding intervals: 0 < DF < 4 and
0 < P < 20.
After building a home-made database, a virtual screening with the
MAC model to select potentially active molecules was carried out.
Table 7 shows the molecules found, their values for DF (Equation 1)
and P (Equation 2) and the classification results.
Those molecules classified as ‘+/–’ pass through –0.5 < DF < 4
and 0 < P < 20 because, as can be seen in PDD (Figure 1), in the interval –0.5 to 0, active compounds could be found.
The most active compounds were benzalkonium chloride, trifluoperazine, paromomycin and pentamidine, with MIC90 values of 4, 8, 16
and 32 mg/L, respectively. Their respective predicted values were 17,
8, 7 and 15 mg/L.
The great structural heterogeneity of the selected compounds
must be noted although some of them, paromomycin, reserpine and
trifluoperazine, show described antimycobacterial activity. For these
compounds, we have checked that our theoretical calculations of
activity are comparable to those obtained in vitro and to those
reported by other groups. On the other hand, trifluoperazine had been
described as an inhibitor of the in vitro growth of multidrug-resistant
M. tuberculosis. However, no reference of activity against MAC
appears in the literature.
The fact that the method described in this work is able to select
molecules that have been identified by other means enhances the
validity of our approach. These results confirm other previously published work, so indicating the usefulness of molecular topology as a
potent tool to identify new drugs, especially new leads. It also overcomes the need for previous knowledge on the drug mechanism of
action.
Microbiological study
The results of the susceptibility test against MAC for conventional
drugs are shown in Table 8, and the experimental results for the new
compounds are illustrated in Table 9. The two tables show the MIC
range (variation between sensitive and non-sensitive strains) in
column 2, the MIC50 in column 3 and the MIC90 in column 4.
According to the criterion suggested by Heifets33 to interpret
MICs against M. avium complex of conventional drugs, the set of
isolates included in our study represents no specific resistance profile
and the method used here is reproducible and comparable with the
standardized proportion method (Table 8).
As can be seen in Table 9, all of the selected compounds classified
as ‘+/–’ by the model were inactive against MAC except ribavirin.
Those classified as ‘+’ exhibited antimycobacterial activity against
the tested strains, with MIC50 ranging from 4 to 64 mg/L (Table 9).
These results confirm the validity of the model.
Particularly active were benzalkonium chloride, paromomycin,
pentamidine and trifluoperazine with MIC50 values ≤16 mg/L.
Benzalkonium chloride is a quaternary ammonium compound
(QAC) with activity over bacterial membranes and activity against
Gram-positive bacteria such as Staphylococcus aureus isolates. In an
experimental disinfection assay, QACs are mycobacteriostatic
agents even at low concentrations.34 Our results indicate the possibility
of increasing the susceptibility of mycobacteria to other agents for the
treatment of cutaneous processes. Paromomycin is an oligosaccharidetype antibiotic with demonstrated antibacterial35 and antiamoebic
activity, but it is not used as a primary antimycobacterial drug. Pentamidine, a dibenzamidine derivate, is an antimicrobial agent with
activity against protozoa and Pneumocystis carinii; it inhibits aerobic
glycolysis. Trifluoperazine is a phenotiazine employed in the treatment of psychosis, which has demonstrated inhibition of in vitro
growth of multidrug-resistant M. tuberculosis36 by a calmodulin
antagonist mechanism.
Less active but still showing significant inhibition results,
were aflatoxin B1 (MIC50 = 32 mg/L), reserpine and ribavirin (MIC50
= 64 mg/L for both). Reserpine, an antihypertensive drug, demonstrated an inhibitory effect in Gram-positives by an efflux inhibition
mechanism. A probable drug efflux protein has been characterized in
M. tuberculosis and other mycobacteria.37 Aflatoxin B1, a secondary
fungal metabolite, and ribavirin, the first synthetic non-interferoninducing broad-spectrum antiviral nucleoside, are two molecules
with limited applications as antibacterial drugs as a result of important
toxicological problems.
The structures of the selected compounds are presented in Figure 3.
The wide structural diversity of the selected compounds is noteworthy,
since they could eventually be considered as new leads in this field.
Acknowledgements
We acknowledge financial support for this research by Generalitat
Valenciana (GV2001–047) and the Spanish Ministry of Science and
Technology (SAF2000–0223-C03–02). J. V. de Julián Ortiz
acknowledges his grant (EX2001–19851827) from the Spanish
Ministry of Education.
References
1. Yajko, D. M., Nassos, P. S. & Hadley, W. K. (1987). Broth microdilution testing of susceptibilities to 30 antimicrobial agents of Mycobacterium avium strains from patients with acquired immune deficiency
syndrome. Antimicrobial Agents and Chemotherapy 31, 1579–84.
2. Hoffner, S. E., Gezelius, L. & Olsson-Liljequist, B. (1997). In vitro
activity of fluorinated quinolones and macrolides against drug-resistant
Mycobacterium tuberculosis. Journal of Antimicrobial Chemotherapy 40,
885–8.
3. Tomioka, H. (2002). Prospects for development of new antituberculous drug. Kekkaku 77, 573–84.
4. Koketsu, M., Tanaka, K., Takenaka, Y. et al. (2002). Synthesis of
1,3-thiazine derivates and their evaluation as potential antimycobacterial
agents. European Journal of Pharmaceutical Sciences 15, 307–10.
5. Sandbhor, U., Padhye, S., Billington, D. et al. (2002). Metal complexes of carboxamidrazone analogs as antitubercular agents. Synthesis,
X-ray crystal-structures, spectroscopic properties and antimycobacterial
activity against Mycobacterium tuberculosis H(37). Journal of Inorganic
Biochemistry 90, 127–36.
6. Kier, L. B. & Hall, L. H. (1986). Molecular Connectivity in Structure–
Activity Analysis, pp. 225–46. John Wiley & Sons, Letchworth, UK.
7. Balaban, A. T. (1983). Topological indices based on topological
distances in molecular graphs. Pure and Applied Chemistry 55, 199–206.
8. Estrada, E., Peña, A. & García-Domenech, R. (1998). Designing
sedative/hypnotic compounds from a novel substructural graph-theoretical
approach. Journal of Computer-Aided Molecular Design 12, 583–95.
9. Kier, L. B., Murray, W. J., RandiM, M. et al. (1976). Molecular
Connectivity V: connectivity series concept applied to density. Journal of
Pharmaceutical Sciences 65, 1226–30.
10. Balaban, A. T. (1982). Highly discriminating distance based topological index. Chemical Physics Letters 89, 399–404.
72
New agents against Mycobacterium avium complex
11. Gálvez, J., García-Domenech, R., de Julián-Ortiz, J. V. et al. (1994).
Topological approach to analgesia. Journal of Chemical Information and
Computer Sciences 34, 1198–203.
12. Ríos-Santamarina, I., García-Domenech, R., Gálvez, J. et al.
(1998). New bronchodilators selected by molecular topology. Bioorganic
and Medicinal Chemistry Letters 8, 477–82.
13. Casabán, E., Antón-Fos, G. M., Gálvez, J. et al. (1999). Search of
new antihistaminic compounds by molecular connectivity. Quantitative
Structure-Activity Relationships 18, 35–42.
14. de Julián-Ortiz, J. V., Gálvez, J., Muñoz-Collado, C. et al. (1999).
Virtual combinatorial syntheses and computational screening of new
potential anti-herpes compounds. Journal of Medicinal Chemistry 42,
3308–14.
15. de Gregorio-Alapont, C., García-Domenech, R., Gálvez, J. et al.
(2000). Molecular topology: a useful tool for the search of new antibacterials. Bioorganic and Medicinal Chemistry Letters 10, 2033–6.
16. García-Domenech, R., Catalá, A. I., García-García, A. et al.
(2002). QSAR by molecular topology of 2,4-dihydroxythiobenzanilides: a
virtual screening approach to optimize the antifungal activity. Indian Journal
of Chemistry 41B, 2376–84.
17. Gozalbes, R., Brun-Pascaud, M., García-Domenech, R. et al.
(2000). Prediction of quinolone activity against Mycobacterium avium by
molecular topology and virtual computational screening. Antimicrobial
Agents and Chemotherapy 44, 2764–70.
18. Burke, A. & Cunha, M. D. (1988). Aminoglucósidos: papel actual
en la terapia antimicrobiana. Pharmacotherapy 8, Suppl. 6, 334–50.
19. Fernandes, P. B., Hardy, D. J., McDaniel, D. et al. (1989). In vitro
and in vivo activities of clarithromycin against Mycobacterium avium.
Antimicrobial Agents and Chemotherapy 33, Suppl. 9, 1531–4.
20. Perronne, C., Gikas, A., Truffot-Pernot, C. et al. (1991). Activities
of sparfloxacin, azithromycin, temafloxacin, and rifapentine compared
with that of clarithromycin against multiplication of M. avium complex
within human macrophages. Antimicrobial Agents and Chemotherapy
35, 1356–9.
21. Jacobs, M. R. (1995). Activity of quinolones against mycobacteria.
Drugs 49, Suppl. 2, 67–75.
22. García-Sánchez, J. E., López, R. & Prieto, J. (1999). Antimicrobianos en medicina, pp. 449–55. Sociedad Española de Quimioterapia.
Prous Science, S.A.
23. Gillespie, S. H. & Billington, O. (1999). Activity of moxifloxacin
against mycobacteria. Journal of Antimicrobial Chemotherapy 44, 393–5.
24. Gálvez, J., García-Domenech, R., de Julián-Ortiz, J. V. et al.
(1995). Topological approach to drug design. Journal of Chemical Information and Computer Sciences 35, 272–84.
25. Gálvez, J., García-Domenech, R., Salabert, M. T. et al. (1994).
Charge indices. New topological descriptors. Journal of Chemical Information and Computer Sciences 34, 520–5.
26. Gálvez, J., García-Domenech, R., de Gregorio-Alapont, C. et al.
(1996). Pharmacological distribution diagrams: a tool for the new drug
design. Journal of Molecular Graphics and Modelling 14, 272–6.
27. Furnival, G. M. & Wilson, R. W. (1974). Regressions by leaps and
bounds. Technometrics 16, 499–511.
28. Hocking, R. R. (1972). Criteria for selection of a subset regression:
which one should be used? Technometrics 14, Suppl. 4, 967–70.
29. Allen, D. M. (1974). The relationship between variable selection
and data augmentation and a method for prediction. Technometrics 16,
125–7.
30. Wold, S. & Eriksson, L. (1995). Statistical validation of QSAR
results. In Chemometric Methods in Molecular Design (Van de Waterbeemd, H., Ed), vol. 2, pp. 309–18. VCH, New York.
31. Heifets, L. (1996). Susceptibility testing of Mycobacterium avium
complex isolates. Antimicrobial Agents and Chemotherapy 40, 1759–67.
32. Franzblau, S. G., Witzig, R. S., McLaughlin, J. C. et al. (1998).
Rapid low technology MIC determination with clinical Mycobacterium isolates by using the microplate alamar blue assay. Journal of Clinical
Microbiology 36, 362–6.
33. Heifets, L. B. (1996). Clinical mycobacteriology. Drug susceptibility
testing. Clinics in Laboratory Medicine 16, Suppl. 3, 641–56.
34. Rikimaru, T., Kondo, M., Kondo, S. et al. (2000). Efficacy of
common antiseptics against mycobacteria. International Journal of
Tuberculosis and Lung Disease 4, 570–6.
35. Donald, P. R., Sirget, F. A., Kanyok, T. P. et al. (2000). Early bactericidal activity of paromomycin (aminosidine) in patients with smearpositive pulmonary tuberculosis. Antimicrobial Agents and Chemotherapy 44, Suppl. 12, 3285–7.
36. Gadre, D. V. & Talwar, V. (1999). In vitro susceptibility testing of
Mycobacterium tuberculosis strains to trifluoperazine. Journal of Chemotherapy 11, Suppl. 3, 203–6.
37. Zhang, Y., Permar, S. & Sun, Z. (2002). Conditions that may affect
the results of susceptibility testing of Mycobacterium tuberculosis to
pyrazinamide. Journal of Medical Microbiology 51, Suppl. 1, 42–9.
73