Bioresource Technology 101 (2010) 2884–2887

Contents lists available at ScienceDirect

Bioresource Technology
journal homepage: www.elsevier.com/locate/biortech

Short Communication

Artificial neural network modeling and genetic algorithm based medium

optimization for the improved production of marine biosurfactant
C. Sivapathasekaran, Soumen Mukherjee, Arja Ray, Ashish Gupta, Ramkrishna Sen *
Department of Biotechnology, Indian Institute of Technology Kharagpur, West Bengal-721302, India

a r t i c l e i n f o a b s t r a c t

Article history: A nonlinear model describing the relationship between the biosurfactant concentration as a process out-
Received 29 July 2009 put and the critical medium components as the independent variables was developed by artificial neural
Received in revised form 23 September network modeling. The model was optimized for the maximum biosurfactant production by using
2009
genetic algorithm. Based on a single-factor-at-a-time optimization strategy, the critical medium compo-
Accepted 24 September 2009
Available online 14 November 2009
nents were found to be glucose, urea, SrCl2 and MgSO4. The experimental results obtained from a statis-
tical experimental design were used for the modeling and optimization by linking an artificial neural
network (ANN) model with genetic algorithm (GA) in MATLAB. Using the optimized concentration of crit-
Keywords:
Single-factor-at-a-time
ical elements, the biosurfactant yield showed close agreement with the model prediction. An enhance-
Design expert ment in biosurfactant production by approximately 70% was achieved by this optimization procedure.
Neural network modeling Ó 2009 Elsevier Ltd. All rights reserved.
Genetic algorithm

1. Introduction ous engineering fields, this modeling and optimization approach

Biosurfactants are extracellular surface-active molecules pro-
duced by various bacteria including Bacillus sp. (Desai and Banat,
1997). The production of these molecules has gained a consider-
able interest in recent times mainly due to their high surface activ-
ities, heterogeneity and great potential for therapeutic applications
like antimicrobial, antifungal, antioxidant and antiviral agents
(Cameotra and Makkar, 2004). Despite their multifarious advanta-
ges and diverse potential applications, their production on a com-
mercial level could not be achieved due to the low yields in
production process (Mukherjee et al., 2008). One of the primary ap-
proaches applied for obtaining increased yields in fermentative
production is the medium optimization. The most effective method
used for the optimization of factors is the statistical approach (Sen,
1997). There is fairly good amount of available literature on med-
ium optimization using different statistical methods. These include
experimental designs such as Plackett–Burman (Mukherjee et al.,
2008), response surface methodology (Sen, 1997; Mutalik et al.,
2008) and factorial designs (Rodrigues et al., 2006). Other mathe-
matical approaches such as ANN coupled to GA have also been pro-
ven to be powerful tools in optimization and superior to statistical
approaches such as response surface methodology (Pal et al.,
2009). Although ANN-GA has gained popularity for its use in vari-
* Corresponding author. Tel.: +91 3222 283752; fax: +91 3222 278707.
E-mail address: rksen@yahoo.com (R. Sen).
has not been fully exploited to enhance the production of biomol-
ecules such as biosurfactants (Pal et al., 2009).
In this study, an attempt was made to optimize the biosurfac-
tant production by using ANN-GA based nonlinear modeling and
optimization. The critical medium components were identified by
a single-factor-at-a-time strategy. The model was linked with GA
to find the maximum production level and the optimum concen-
trations of the critical medium components that affect significantly
the production process.
2. Methods
2.1. Microorganism, medium, production and quantification
Bacillus circulans MTCC 8281 previously isolated from Andaman
and Nicobar Islands, India was used in this study. The medium
used both for inoculum preparation and biosurfactant production
was a Modified Marine Medium (MMM) containing the following
components in g L 1: glucose 20, urea 3.3, K2HPO4 2.2, SrCl2 0.2,
NaCl 0.1, MgSO4 0.6, FeSO4 0.04 and microsalt solution 500 ll
L 1. The microsalt solution consisted of (g/L): H3BO3 0.56, Na2-
MoO4
2H2O 0.39, KI 0.66, CuSO4
5H2O 0.4, CoCl2
6H2O 0.42,
ZnSO4
7H2O 2.0, EDTA 1.0, NiCl2
6H2O 0.002, MnSO4
4H2O 0.90,
CaCl2 0.04 and KH2PO4 0.14. The fermentation (5% pre-inoculum
was used) was carried out for 28 h at 37 °C on a rotary shaker at
180 rpm. The biosurfactant produced in the fermentation medium
was quantified directly from cell-free supernatant (CFS) by a HPTLC

0960-8524/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.biortech.2009.09.093
 C. Sivapathasekaran et al. / Bioresource Technology 101 (2010) 2884–2887 2885

based densitometric method (Mukherjee et al., 2008) and also cor-
roborated by dry weight method (Sen, 1997).
2.2. Optimization of biosurfactant production
2.2.1. Single-factor-at-a-time experiments and experimental design
The critical media components that influence the biosurfactant
production were identified by single-factor-at-a-time experiments
(Mutalik et al., 2008). The range and levels of the four independent
variables along with the central values for these critical media
components are listed in Table 1. A central composite design
(CCD) was applied for the four factors and the experimental design
(Table 2) was obtained using Design Expert version 7.0 (Stat-Ease
Inc., Minneapolis, USA) (Mutalik et al., 2008).
2.2.2. Artificial neural network linked with genetic algorithm (ANN-
GA) as a modeling and optimization tool
ANN is a biologically inspired computational tool that is widely
used in modeling and simulation of various systems. GA, on the
other hand, is a commonly used globalized optimization technique
Table 1
The boundaries of the experimental domain and the levels of the four variables. Zero
indicates the central value, ( 2) indicates the low level, likewise (+2) indicates the
high level.
that optimizes a given function over a particular range, and is
based on the evolutionary methods of natural selection of the best
individuals in a population (Goldberg, 1989; Imandi et al., 2008).
2.2.2.1. Feed forward back propagation network. Among the various
ANN models, the one of our interest was the feed forward back
propagation network (Pal et al., 2009; Imandi et al., 2008). Briefly,
in the present study four neurons were used in the input layer, six
in the hidden layer and one in the output layer of the network to
model the dependence of the biosurfactant production on the four
media components, i.e., a single hidden layer comprising of six
neurons was used (Fig. 1a). Once the neural network was created,
it was trained to accurately model the given phenomenon by using
the experimental data (Table 2) in MATLAB version 7.1 (Math-
works Inc., Natick, US). The mean standard error (MSE) and the
mean absolute error (MAE) were used as measures of closeness
of the model to the actual system. Once the ANN model was devel-
oped to the desired level of precision, GA was used to optimize and
determine the maximum possible biosurfactant production and
the point in the input space where this maximum was to be ob-
tained. For the optimization using GA, the ANN model was used
as the fitness function, which measures the quality of individuals
in the population.

Critical media component Range of the media component for designing 3. Results and Discussion
experiment (g L 1)
2 1 0 +1 +2 Biosurfactant production was found to be cell growth associ-
Glucose 5 13.75 22.5 31.25 40
ated as reported earlier for other biosurfactant producers (Rodri-
Urea 1 2 3 4 5 gues et al., 2006; Pal et al., 2009). The maximum biosurfactant
SrCl2 0.05 0.1375 0.225 0.3125 0.4 concentration of 2.58 g L 1 was found to be at 28 h during the fer-
MgSO4 0.2 0.4 0.6 0.8 1 mentation in MMM; therefore, all the experiments were performed
up to 28 h.

Table 2 3.1. Single-factor-at-a-time optimization strategy

The 24 full factorial central composite design (CCD) consisting of 30 experiments with
six central experimental runs.
On the basis of single variable at a time experiments, four crit-
Run Glucose Urea SrCl2 MgSO4 Biosurfactant ical medium components glucose, urea, MgSO4 and SrCl2 were
order (g L 1) (g L 1) (g L 1) (g L 1) concentration (g L 1 a
) identified in MMM. These four components in the medium criti-
1 13.75 2.00 0.14 0.40 0.60 cally influenced the cell growth and biosurfactant production as
2 31.25 2.00 0.14 0.40 2.90 evident from Table 3. The optimum concentrations of these critical
3 13.75 4.00 0.14 0.40 1.20 components for enhancing biosurfactant production were deter-
4 31.25 4.00 0.14 0.40 2.60
5 13.75 2.00 0.31 0.40 0.90
mined using the methods of ANN-GA.
6 31.25 2.00 0.31 0.40 1.90
7 13.75 4.00 0.31 0.40 1.20 3.2. Optimization of critical medium components using ANN-GA
8 31.25 4.00 0.31 0.40 1.0
9 13.75 2.00 0.14 0.80 1.60
10 31.25 2.00 0.14 0.80 2.95 The training of the ANN was done for 3000 epochs after which
11 13.75 4.00 0.14 0.80 1.90 the MSE and the MAE were found to be 0.0094 and 0.0066, respec-
12 31.25 4.00 0.14 0.80 2.0 tively. Also, the regression correlation coefficient between the ac-
13 13.75 2.00 0.31 0.80 0.70
tual experimental production values and the ANN-simulated
14 31.25 2.00 0.31 0.80 0.90
15 13.75 4.00 0.31 0.80 1.40
outputs was found to be 0.9996. The average and maximum per-
16 31.25 4.00 0.31 0.80 1.10 centage errors were 0.5% and 2.2%, respectively, which was bet-
17 5.00 3.00 0.23 0.60 0.74 ter than previously reported values in a similar study (Pal et al.,
18 40.00 3.00 0.23 0.60 1.20 2009; Imandi et al., 2008). The best fitness plot for the GA
19 22.50 1.00 0.23 0.60 1.80
(Fig. 1b) maps the gradual convergence of the best fitness values
20 22.50 5.00 0.23 0.60 2.30
21 22.50 3.00 0.05 0.60 1.70 of successive generations towards the final optimum value.
22 22.50 3.00 0.40 0.60 0.80 After the optimization, the glucose concentration required for
23 22.50 3.00 0.23 0.20 1.40 the maximum biosurfactant production showed a marked increase
24 22.50 3.00 0.23 1.00 0.70
from 22 g L 1 to 32 g L 1. This value was very close to the previ-
25 22.50 3.00 0.23 0.60 2.60
26 22.50 3.00 0.23 0.60 2.60
ously reported one (Sen, 1997) and this sharp increase clearly
27 22.50 3.00 0.23 0.60 2.60 showed that the carbon source was the key factor affecting the bio-
28 22.50 3.00 0.23 0.60 2.70 surfactant production (Rodrigues et al., 2006; Pal et al., 2009). The
29 22.50 3.00 0.23 0.60 2.60 optimum concentration of urea in the medium was decreased from
30 22.50 3.00 0.23 0.60 2.70
3.0 to 1.0 g L 1 which meant less amount of urea was required to
a
Experiment performed in triplicate manner. n = 3, average mean value of attain the maximum production. In a previous report urea was
response was coded. used for the production of biosurfactant by Candida (Harrop
2886 C. Sivapathasekaran et al. / Bioresource Technology 101 (2010) 2884–2887

(a) hidden layer

Input layer Wii,jj
X1

Output layer
X2

Results
X3

X4

Best: -4.388 Mean: -4.3281

3 (b)
2 Best fitness
Mean fitness
1
Fitness value

0
-1
-2
-3
-4
-5
0 20 40 60 80 100 120 140 160 180 200
Generation

Fig. 1. (a) Schematic representation of a (4–6–1) neural network (having four neurons in the input layer, six in the hidden layer and one in the output layer). (b)
Representative plots generated from the optimization by GA using MATLAB 7.1. Best and average fitness values with successive generations showed gradual convergence to
the optimum value.

et al., 2003). However, this is the first report showing the use of
urea as nitrogen source for the cultivation of Bacillus sp., for biosur-
factant production, and its optimization. For SrCl2 and MgSO4, the
optimum concentrations were found to be 0.08 g L 1 and 0.84 g
L 1, respectively. SrCl2, a rare salt present in marine environments,
was also used in the medium to enrich the growth of the marine
strain and it is used as a growth stimulant (Zobell, 1941). This is
the first report to indicate its importance in the biosurfactant pro-
duction process. MgSO4, on the other hand, has been reported as a
trace element in the biosynthesis of biosurfactant and a key factor
required for the sporulation of cells (Sen, 1997). There are several
reports describing the optimization study in biosurfactant produc-
tion; however, this is the first report showing the optimization of
critical parameters like urea and SrCl2. The model predicted that
the maximum biosurfactant yield of 4.38 g L 1 could be achieved
by using the aforementioned optimized concentrations of the vari-
ables in the medium.
3.3. Model validation and confirmative studies
successive runs showed very slight variation in the values of media
components as well as biosurfactant production. The results were
confirmed by conducting sets of shake flask experiments using
the optimized media. The experimental results showed the higher
biosurfactant yield of 4.40 ± 0.5 g L 1 (HPTLC method) and
4.35 ± 0.6 g L 1 (dry weight method), respectively. These experi-
mental results were significantly close to the value predicted by
the ANN-GA model.
4. Conclusion
The present study showed the ANN-GA based optimization of
critical media components for the enhanced biosurfactant produc-
tion by B. circulans MTCC 8281 in MMM. ANN was able to model
the desired functional behavior more accurately than previously
reported studies and the optimized values obtained from the appli-
cation of GA predicted an increase in biosurfactant production by
approximately 70% as compared to the un-optimized parent
MMM, which was also validated by experimental study.

The model was validated by running the GA several times using Acknowledgements
different randomly selected initial populations. The results of 10

Table 3
RS and CS acknowledge the Department of Biotechnology (DBT),
Results of single factor at a time experiments – effect of medium components on Government of India for the project grant (BT/PR-6827/AAQ/03/
biomass and biosurfactant production. 263/2005) in marine biotechnology. SM acknowledges CSIR, New
Nutrients absent in MMM Biomass (g L 1 a
) BS (g L 1 a
)
Delhi for his fellowship. Authors acknowledge Palashpriya Das
for strain isolation and her support in this study. Authors acknowl-
Glucose ND ND
edge Roshan Joy Martis, SMST, IIT-Kharagpur for his valuable
Urea ND ND
MgSO4 ND ND support.
SrCl2 1.56 ± 0.20 0.89 ± 0.07
NaCl 3.20 ± 0.10 2.34 ± 0.05 References
K2HPO4 3.21 ± 0.03 2.01 ± 0.04
FeSO4 2.98 ± 0.06 1.96 ± 0.03 Cameotra, S.S., Makkar, R.S., 2004. Recent application of biosurfactant as biological
Microsalts 4.24 ± 0.06 2.47 ± 0.02 and immunological molecules. Curr. Opin. Microbiol. 7, 262–266.
Control (including all nutrients) 4.71 ± 0.06 2.58 ± 0.04 Desai, J.D., Banat, I.M., 1997. Microbial production of surfactants and their
commercial potential. Microbiol. Mol. Biol. Rev. 61, 47–64.
ND – not detected; BS – biosurfactant. Goldberg, D., 1989. Genetic Algorithms in Search, Optimization and Machine
a
Experiments were performed in triplicate (mean ± standard deviation). Learning. Addison-Wesley.
 C. Sivapathasekaran et al. / Bioresource Technology 101 (2010) 2884–2887
Harrop, M.H.V., de Gusmao, N.B., de Campos Takaki, G.M., 2003. New bioemulsifier
produced by Candida Lipolytica using D-glucose and babassu oil as carbon. Braz.
J. Microbiol. 34, 120–123.
Imandi, S.B., Karanam, S.K., Garapati, H.R., 2008. Optimization of fermentation
medium for the production of lipopeptide using artificial neural network and
genetic algorithms. IJNES 2, 105–109.
Mukherjee, S., Das, P., Sivapathasekaran, C., Sen, R., 2008. Enhanced production of
biosurfactant by a marine bacterium on statistical screening of nutritional
parameter. Biochem. Eng. J. 42, 245–260.
Mutalik, S.R., Vaidya, B.K., Joshi, R.M., Desai, K.M., Nene, S.N., 2008. Use of response
surface optimization for the production of biosurfactant from Rhodococcus spp.
MTCC 2574. Bioresour. Technol. 99, 7875–7880.
2887
Pal, M.P., Vaidya, B.K., Desai, K.M., Joshi, R.M., Nene, S.N., Kulkarni, B.D., 2009.
Medium optimization for biosurfactant production by Rhodococcus erythropolis
MTCC 2794: artificial intelligence verses a statistical approach. J. Ind. Microbiol.
Biotechnol. 36, 747–756.
Rodrigues, L., Teixeira, J., Oliveira, R., Mei, H.C., 2006. Response surface optimization
of the medium components for the production of biosurfactants by probiotic
bacteria. Process Biochem. 41, 1–10.
Sen, R., 1997. Response surface optimization of the critical media
components for the production of surfactin. J. Chem. Technol.
Biotechnol. 68, 263–270.
Zobell, C.E., 1941. Studies on marine bacteria. I. The cultural requirements of
heterotrophic aerobes. J. Marine Res. (Sears Foundation) 4, 173–188.