Ultrasonic disruption of Pseudomonas putida for the release of arginine deiminase: Kinetics and predictive models

Accepted Manuscript Ultrasonic disruption of Pseudomonas putida for the release of arginine deiminase: Kinetics and predictive models Mahesh D. Patil, Manoj J. Dev, Sujit Tangadpalliwar, Gopal Patel, Prabha Garg, Yusuf Chisti, Uttam Chand Banerjee PII: DOI: Reference:

S0960-8524(17)30197-9 http://dx.doi.org/10.1016/j.biortech.2017.02.074 BITE 17646

To appear in:

Bioresource Technology

Received Date: Revised Date: Accepted Date:

18 January 2017 16 February 2017 17 February 2017

Please cite this article as: Patil, M.D., Dev, M.J., Tangadpalliwar, S., Patel, G., Garg, P., Chisti, Y., Banerjee, U.C., Ultrasonic disruption of Pseudomonas putida for the release of arginine deiminase: Kinetics and predictive models, Bioresource Technology (2017), doi: http://dx.doi.org/10.1016/j.biortech.2017.02.074

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

1

Ultrasonic disruption of Pseudomonas putida for the release of arginine deiminase: Kinetics and predictive models

Mahesh D.Patil,1 Manoj J. Dev1, Sujit Tangadpalliwar,2 Gopal Patel,1 Prabha Garg,2 Yusuf Chisti3 and Uttam Chand Banerjee1,*

1

Department of Pharmaceutical Technology (Biotechnology), National Institute of

Pharmaceutical Education and Research, Sector-67, S.A.S. Nagar–160062, Punjab, India

2

Department of Pharmacoinformatics, National Institute of Pharmaceutical Education and

Research, Sector-67, S.A.S.Nagar–160062, Punjab, India

3

School of Engineering, Massey University, Private Bag 11 222, Palmerston North, New

Zealand

Email addresses Mahesh D. Patil – [email protected] Manoj J. Dev – [email protected] Gopal Patel – [email protected] Sujit Tangadpalliwar–[email protected] Prabha Garg –[email protected] Yusuf Chisti–[email protected] Uttam Chand Banerjee – [email protected] *Corresponding author: Uttam Chand Banerjee, E-mail: [email protected]

2

Abstract The responses of the ultrasound-mediated disruption of Pseudomonas putida KT2440 were modelled as the function of biomass concentration in the cell suspension; the treatment time of sonication; the duty cycle and the acoustic power of the sonicator. For the experimental data, the response surface (RSM), the artificial neural network (ANN) and the support vector machine (SVM) models were compared for their ability to predict the performance parameters. The satisfactory prediction of the unseen data of the responses implied the proficient generalization capabilities of ANN. The extent of the cell disruption was mainly dependent on the acoustic power and the biomass concentration. The cellmass concentration in the slurry most strongly influenced the ADI and total protein release. Nearly 28 U/mL ADI was released when a biomass concentration of 300 g/L was sonicated for 6 min with an acoustic power of 187.5 W at 40% duty cycle. Cell disruption obeyed first-order kinetics.

Keywords: Pseudomonas putida; ultrasonication; arginine deiminase; artificial neural network; support vector machine

3

1. Introduction Arginine deiminase (ADI; E.C. 3.5.3.6) is a potential enzyme for the treatment of ‘arginineauxotrophic’ tumors (Fultanget al. 2016; Patil et al. 2016a). ADI used in clinical investigations are generally obtained from mycoplasmal sources (Ahn et al. 2014; Fayura et al. 2013). Pseudomonas sp. are the alternative potential sources of ADI. Pseudomonas putida was reported to be the best producer of ADI out of nearly 150 microorganisms screened (Kakimoto et al.1971). Our earlier work had shown that P. putida to be a better producer of ADI than Pseudomonas plecoglossicida CGMCC2039 and Streptococcus faecalis (Patil et al. 2016b). Existing ADI-based drugs (Zhang et al. 2015) make use of mycoplasmal ADI, however, there is a need to explore other potential producers of ADI. Recovery of intracellular ADI requires effective methods of breaking bacterial cells. Power ultrasound of 20–100 kHz frequency is known to be broadly effective in breaking microbial cells (Gao et al. 2014; Iida et al. 2008; Jaeschke et al., 2016). Cell breakage by ultrasound is a result of multiple factors including intense turbulence and pressure changes associated with acoustic cavitation (Doulah 1977; Greenly and Tester, 2015). This work reports an ultrasound-mediated disruption of Pseudomonas putida for the release of ADI. Disruption was assessed in terms of the release of the total intracellular proteins; the release of ADI; and the extent of cell disruption. These disruption indices were characterized with respect to the following operational factors: the concentration of the cells in the suspension being sonicated; the treatment time of sonication; the acoustic power used; and the duty cycle of the treatment. A face-centered central composite design (CCD) of experiments with the above 4-factors measured at 3-levels was used. The experimental data obtained were used to develop response surface, artificial neural network (ANN) and support vector machine (SVM) models for predicting the above noted 3 responses.

4

A response surface models (RSM) accounts for the interactive effects of the factors on the response and therefore, are superior to the data obtained through conventional experiments in which a single factor is varied in a given set (Arun et al., 2016; Patel et al. 2016; Singh 2013). ANN is a data-driven model that can use the input data to predict a response without requiring any assumptions about underlying mechanisms (Maran et al. 2015; Sarve et al. 2015). SVM is another multivariate statistical method for the prediction of a response (Apul et al. 2014) and is a potential alternative to ANN. The models generated by RSM, ANN and SVM were evaluated for their predictive and generalization capabilities. Also, an attempt was made to explain the effect of the process variables of sonication on the kinetics of cell-disruption process, release of ADI and total protein. 2. Materials and Methods 2.1 Microorganism and cultivation conditions The stock cultures of Pseudomonas putida KT2440 were maintained on Luria Bertani (LB) medium and stored at −80 °C. The bacterium was grown aseptically in a 14 L stirred-tank bioreactor (BioFlo 310; New Brunswick Scientific, USA) with a working volume of 8 L as previously reported (Patil et al. 2016b). The broth was harvested after 28 h of fermentation and the cells were recovered by centrifugation (7,000 x g, 20 min, 4°C). The cell pellet was washed twice with phosphate buffer (50 mM, pH 7.0) and kept refrigerated at 4°C for ultrasonic disruption experiments. 2.2 Ultrasonic treatment Bacterial suspensions (15 mL) containing various cellmass concentrations were prepared in phosphate buffer (50 mM, pH 7.0). For each treatment, 30-mL plastic centrifuge tubes containing the suspended cells were subjected to ultrasonic treatment using a horn-type sonicator (Vibra-cell VCX 750; Sonics & Materials Inc., Newton, NJ, USA) at a frequency of 20 kHz. A sonic horn (diameter = 13 mm, length = 136 mm) with a removable tip made up of

5

titanium alloy was used. The tip of the sonic horn was immersed in the slurry to midpoint of the total depth. During sonication, the sample tube containing the suspension was held in an ice bath to prevent a rise in temperature that could be damaging to the released enzymes. A pulsed ultrasound operation with an adjustable duty cycle was used. The acoustic power level of the sonication was controlled by adjusting the amplitude of oscillation of the sonic horn. The experimental factors of sonication time, acoustic power, duty cycle and the biomass concentration were set at their desired values in accordance with the experimental matrix designed by Design ExpertTM 8.0 software (Stat-Ease Inc., Minneapolis, MN, USA). The ADI activity, the total protein release and the fraction of the cells disrupted were measured as explained in the following sections. 2.3 Experimental designs 2.3.1 Response surface method RSM was used to model the dependence of the responses on the process variables. The following experimental factors were used in the central composite: (1) sonication treatment time (A, min); (2) the biomass concentration in the slurry being sonicated (B, g/L); (3) acoustic power of the sonication treatment (C, W); and duty cycle of the sonication treatment (D, %). The actual values of factors, the measured responses and the predicted responses are shown in Table 1. The ranges of the factors were selected based on the preliminary studies and the literature (Feliu et al. 1998; Hua and Thompson 2000; Üstün-Aytekin et al. 2016). The CCD matrix consisted of 30 experiments (Table 1). The response data of the dependent variables were fitted to a second order polynomial equation to generate the contour plots. 2.3.2 Artificial neural network An ANN is a polynomial regression based modelling tool which can model complex nonlinear relationships (Buyukada 2016; Maran et al. 2015; Marchitan et al. 2010, Nair et al. 2016). This study used the Neural Network Toolbox of MATLAB (version 7.8.0;

6

www.mathworks.com) for developing the ANN model. A feed-forward neural network consisting of an input layer, a hidden middle layer and an output layer was used (Fig. 1). The transfer functions used at the hidden layer were tangent sigmoid (tansig) and linear transfer functions (purelin) were used at the output layer (Fig. 1). A back propagation training algorithm was used. Thus the weights of the connections between the nodes of the layers were repetitively adjusted so that the output value closely approached the desired output. Six different back-propagation algorithms (i.e. BFGS quasi-Newton, conjugate gradient with Powell/Beale restarts, Fletcher-Powell conjugate gradient, Polak-Ribiére conjugate gradient, Levenberg-Marquardt, and scaled conjugate gradient back propagation algorithm (trainlm)) were assessed to determine the most suitable algorithm. Normalized values of the four input variables (i.e. sonication time, biomass concentration, acoustic power and duty cycle) were fed to the neural network (Fig. 1). An input was normalized dividing by its maximum value. The output layer had three neurons corresponding to the ADI released, the total protein released and the fraction of cells disrupted (Fig. 1). 2.3.3 Support vector machine (SVM) SVM is a widely used supervised learning algorithm and is an excellent tool for a global optimization (Smola and Scholkopf 2004). SVM is based on the structural risk minimization principle. It attempts to minimize the upper bound of generalization error to increase the likelihood of attaining a global minimum. The sequential minimal optimization (SMO) algorithm available in Weka (Waikato Environment for Knowledge Analysis; www.cs.waikato.ac.nz/ml/weka/) was used to build the model. The experimental data in Table 1 were used to train the SVM model using a polynomial kernel. 2.4 Comparison of the models The prediction ability and goodness of fit of the developed models was assessed by error analyses. For this, root mean square error (RMSE), the standard error of prediction (SEP), the

7

relative percent deviation (RPD) and the determination coefficient (R2) values were estimated (Maran et al. 2015; Patil et al. 2016c; Sarve et al. 2015). 2.5 Analytical methods 2.5.1 ADI activity One unit of ADI activity was defined as the amount of enzyme that converted 1 µmol Larginine to L-citrulline per min at the incubation temperature (Ni et al. 2011). Citrulline formed was measured using the diacetyl monoxime -thiosemicarbazide (DAM) method (Boyde and Rahmatullah 1980). 2.5.2 Total protein quantification The total protein released was determined using the Bradford method (Bradford 1976) using bovine serum albumin as a standard. 2.5.3 Cell disruption The extent of cell disruption was quantified as the fraction of the cells disrupted (Fd), as previously reported (Patil et al. 2016c). 3. Results and Discussion 3.1 Response surface models (RSM) and statistical analysis The response surface models were aimed to separately maximize the responses of the extent of cell disruption, the total protein and ADI release. The actual values of the independent factors, the measured responses and the predicted responses of the 30 experimental runs of the central composite design are shown in Table 1. The RSM predicted responses and the measured data were in a good agreement. The measured responses were fitted to the coded values of the independent factors using following equations:
 = 22.15 − 1.53 + 1.94 − 0.75 + 1.18 − 1.41 − 1.12 − 1.44 − 0.47 − 0.54 + 1.76  − 2.95 + 4.67 − 2.78  − 1.78 

(1)

8


 = 5.01 + 0.65 + 0.91 + 0.25 − 0.22 − 0.17 − 0.19 + 0.074 + 0.056 + 0.079 − 0.11 − 0.64 − 0.13 −   + 0.18 

(2)


 = 71.26 + 22.10 − 0.41 + 13.30 + 1.01 − 2.13 + 0.55 + 0.78 + 0.21 − 1.02 − 2.14 − 15.28 − 0.47 − 8.97  + 0.40

(3)

Where, Y1 is the ADI released (U/mL); Y2 is the total protein released (g/L); Y3 is the extent of the cell disruption (%); A is the sonication time (min); B is the biomass concentration (g/L); C is the acoustic power (W); and D is the duty cycle of the sonicator (%). The adequacies and fitness of the above models (Eq. 1–3) were tested by analysis of variance (ANOVA). The above models were found to be adequate for the specified experimental responses within the experimental space. For ADI release model (Eq. 1), the correlation coefficient (R = 0.962) indicated a good agreement between the experimental data and the model-predicted values. The determination coefficient (R2= 0.9262) indicated that nearly 93% of the variation in ADI release could be attributed to the experimental factors. The adjusted determination coefficient (adjusted R2) adjusts the R2 value for the number of terms in the model relative to the sample size. If the number of model terms is high and the sample size is not sufficiently large, the adjusted R2 would be noticeably smaller than R2 (Maran and Priya 2015; Sarve et al. 2015). The adjusted determination coefficient value was high (adjusted R2 = 0.8572) for the present model (Eq. 1) confirming its high significance. The Fisher’s F-test uses F-values to verify the competence of the factors to describe the variation in the data about its mean. The F-value for the present model was 13.44, suggesting that the model was significant and the probability of the F-value of the model being due to experimental noise was only 0.01%. The ‘adequate precision value’ is an index of the signal-to-noise ratio and the ratio of greater than 4 is a prerequisite for a model to be considered a good fit to the data. Adequate precision value of Eq. (1) was 14.641, suggesting

9

the model to be satisfactory for navigating the design space. A low pure error value (= 1.03) suggested a good reproducibility of the experimental responses. Similarly, the adequacies and fitness of the quadratic models for total protein release (Eq. 2) and the fraction of cells disrupted (Eq. 3) were tested by ANOVA. The determination coefficient values for the models were high (0.9287 and 0.9880, respectively). The corresponding correlation coefficients of 0.9636 and 0.9940 indicated that the values as predicted by the models fitted well to the corresponding experimental data. From the statistical analysis of the quadratic models, it is evident that all the linear terms (A, B, C and D) and quadratic terms were significant for ADI release response. Only two interactive terms (BC and BD) were not significant. For the total protein release response, all the linear terms were significant; however, none of the interactive and quadratic terms, except A2, was significant. For the fraction of cells disrupted response, two linear (i.e. A, C), two interactive (i.e. AB, CD) and two quadratic (i.e. A2, C2) terms were significant. From the P-values of each model term in each of the quadratic models, it was concluded that all the independent factors had statistically significant effects on ADI and total protein release. Furthermore, it could be concluded that sonication time and acoustic power had the most significant effect on the cell fraction disrupted. Moreover, the experimental data were fitted to second order polynomial equations and the response surface plots based on the above models are shown in Figures 2–4. The plots were generated by varying two of the factors while keeping the other two factors at their middle levels. 3.2 Artificial neural network modelling A feed forward back propagation ANN was used in view of its wide applicability to modelling bioprocesses. The earlier mentioned six training functions were assessed (Table 2). The number of neurons for each training function was varied from 1 to 10 to determine the optimum number. For each neuron and unique function, models with 1000 training cycles

10

were built. The prediction capabilities were selected on the basis of R2 and MSE values. Of these models, the MSE value was found to be minimum for a model that used LevenbergMarquardt (LM) training function. The R2 values of the models were not significantly improved once the number of neurons in the hidden layer was increased to beyond 6 (Table 2). Hence, a feed-forward neural network with a LM training function having 4 neurons in input layer, 3 neurons in output layer and 6 neurons at hidden layer was used for modeling (Fig. 1). Purelin transfer function at the output layer was used. For evaluation of the ANN model, the experimental values of the responses were compared with the corresponding predicted values of the responses. The relevant data are shown in Table 1. As shown in Fig. 5, the ANN model with the training dataset had very good R-values of 0.949, 0.981 and 0.996 for ADI and total protein released and the fraction of the cells disrupted (Fd), respectively. The R-value for the entire dataset (i.e. the three datasets combined) was also good (= 0.999). These R-values suggest that the developed ANN model was precise enough for the prediction of performance parameters of the disruption process. 3.3 Support vector machine model The SVM models were developed using the experimental data shown in Table 1. A polynomial kernel was used for model development. A polynomial kernel is a function that represents the similarity of the training sample data in a feature space over polynomials of the original variables (Goldberg et al. 2008). The normalized values of the four input variables were used for development of the models. Python code was written to normalize each input by map min-max function. As SVM operates only on single output; therefore, three different models were developed for the outputs of ADI release, the total protein release and the fraction of the cells disrupted. The final model used an α-value 0.1 and a γ-value of 1.0×10 −12

11

with a cache size of 250007. The R-values of the models were: 0.544 for ADI release model; 0.812 for the total protein release model; and 0.858 for the fraction of cells disrupted model. 3.4 Comparison of predictive and generalization capacities of the models The generalization capacities of the models were evaluated using unseen dataset consisting of 6 experiments (runs 31–36, Table 3) that were not used in development of the models. The experimental and predicted values of the responses for unseen datasets are shown in Table 3. The error analysis parameters were estimated to assess the prediction capabilities of the developed models. The estimated values of these statistical parameters are reported in Table 4 both for the training datasets and the unseen datasets. The prediction capacities of the developed models were also compared. In case of the training dataset, the values for the error parameters for ANN model were lower than the corresponding values for the RSM and SVM models (Table 4, training dataset). As R2 value of the ANN model was higher than the R2 values of the other models, ANN model had a better overall predictive performance for the total protein release response. Similarly, the lower values of the various error analyses parameters for the ANN model attested to its better performance in predicting the percent cell disruption. However, for the ADI release response, the predictions of the RSM model were better than the predictions of the other models (Table 4, training dataset). Based on the unseen data, for ADI release response, the values of the error analyses parameters for the ANN model were lower than the corresponding values of the RSM and SVM models (Table 4, unseen dataset). Therefore, the ANN model was better in comparison to the other models for predicting ADI release. Similarly, the ANN model was superior to the other models for the response of extent of cell disruption. For total protein release response, the RSM and ANN models were comparable in terms of their error parameters and R2; nonetheless, the ANN model may be considered superior to the other models in view of its

12

marginally lower RMSE, MSE, RPD and SEP (Table 4, unseen dataset). Overall, the generalization capabilities of ANN predictions were superior to the other models for all the performance parameters of the ultrasound-mediated cell disruption. 3.5 Kinetics of cell disruption, ADI and total protein release A sonic horn imparts mechanical energy to the cell suspension. This energy is dissipated throughout the suspension and ultimately causes cell disruption and the release of intracellular material (Kapucu et al. 2000). Cell disruption, protein release and enzyme release generally follows first-order disruption kinetics (Doulah 1977) as follows:  



 

=

(4)

In the above equation, Rm is the maximum release response (i.e. maximum amount of protein released, or enzyme released, or complete disruption, R is the release response at time t (s) and K (s−1) is the release rate constant. For disruption of a given batch of cells, the rate constant for ADI release (= KA), the total protein release (= Kp) and fractional cell disruption ( =Kc) can be different. The values of the rate constants were determined as slopes of the plots of ln [Rm/(Rm – R)] versus time in each experiments. The rate constants are shown in Table 5. For relatively short sonication periods, the ADI release (Fig. 2) was most strongly influenced by the concentration of cells in the slurry (Fig. 2B, 2E and 2F). In general, more ADI was released from more concentrated slurries. For example, comparing the 2-min duration Runs 6 and 15 (Table 1), identical in every respect except the concentration of the biomass, the more concentrated slurry (Run 6; biomass concentration = 300 g/L) released more ADI compared to Run 15 (biomass concentration = 100 g/L). The more concentrated slurry also released more total protein. However, the increase in ADI and protein release with increasing biomass concentration was not in direct proportion to the biomass concentration.

13

Thus, Run 6 did not release three times as much ADI as did Run 15, suggesting less efficient disruption at higher biomass concentrations. A reduced efficiency of disruption at a higher initial concentration of biomass may be an effect of viscosity. An increase in viscosity is likely to slow the rate of growth of cavitation bubbles. This would reduce the size of the bubble at implosion and therefore the energy released by the implosion event would be reduced and this in turn will reduce the efficacy of cell disruption. For longer sonication periods, there was clear evidence of denaturation of ADI. Comparing Runs 1 and 15 (Table 1), carried out with identical concentrations of biomass (= 100 g/L), the shorter duration run (Run 15, 2 min duration) released more active ADI compared to the longer duration run (Run 1, 10 min duration), but the longer duration run released more protein in keeping with expectations. Similarly, comparing Runs 6 and 10, identical in all respects except the duration of sonication, the longer duration run (Run 10, 10 min duration) released less ADI relative to Run 6 of 2 min duration, but more protein was released after the longer treatment of Run 10. Further clear evidence of ultrasound-mediated denaturation of ADI with increasing sonication time is seen in Runs 7, 18 and 22 (Table 1). These runs were identical except for the duration of the sonication period. After 2 min of sonication (Run 7, Table 1) the ADI activity was nearly 23 U/mL. After 6 min of sonication, the ADI titer was reduced to 22 (Run 18, Table 1) and was further reduced to 16 U/mL after 10 min of sonication (Run 22, Table 1). In the same three runs, the total protein released continued to increase with increasing duration of the sonic treatment (Table 1). Ultrasound can denature enzymes by various mechanisms, including a direct effect of sonication on the enzyme structure (Chisti 2003) and various indirect effects (Chisti and Moo-Young 1986). Implosion of cavitation bubbles formed in the vicinity of the sonic horn can raise local temperature by 5000 K and pressures to around 1000 MPa (Yu et al. 2014).

14

Deactivation of numerous enzymes by prolonged sonication has been widely reported (Ozbek and Ulgen 2000). Sonication at high power levels is claimed to affect hydrogen bonding and van der Waals interactions within polypeptide chains (Yu et al. 2014). The total protein release (Fig. 3A, 3D and 3E) was also most strongly influenced by the concentration of the cells in the slurry. In ultrasound-mediated disruption of other microorganism, the rates of disruption and product release have been generally found to be independent of the cell concentration in the slurry (Apar and Obek 2008; Singh 2013; Singh et al. 2005). Within the limits tested, the duty cycle had a marginal impact on disruption in terms of protein release. For example, the otherwise identical long duration Runs 20 and 25 carried out duty cycles of 60% and 20%, respectively, released about the same amount of total protein (Table 1). The conditions that maximized the total protein release were: a sonication time of 2 min; an initial biomass concentration in the slurry of 300 g/L; a sonication power of 225 W; and a duty cycle of 20% (Run 21, Table 1). The ADI release was maximized with the following conditions: sonication duration of 2 min; an initial biomass concentration in the slurry of 300 g/L; a sonication power of 225 W; and a duty cycle of 60% (Run 6; Table 1). In the present study, the effect of acoustic power was tested at three levels, i.e. 150, 187.5 and 225 W, corresponding to power intensities of 28.3, 35.3 and 42.4 W/cm2, respectively. (Power intensity = acoustic power/surface area of the tip of the sonic horn.) Of the variables tested, acoustic power was the most influential in affecting the fraction of the cells disrupted (Fig.4B, 4D and 4F). (The p-value for acoustic power was most significant for the fractional cell disruption as a response). For example, for otherwise fixed conditions, an increase in acoustic power from 150 to 225 W increased the fraction of the cells disrupted from 62 to 83% (Runs 1 and 2, Table 1). Increasing acoustic power input increases turbulence in the fluid and the potential to damage cells increases (Doulah 1977; Liu et al.

15

2013). This was consistent with Scanning electron microscope (SEM) images, showing the clear cell damage and breakage of cells as a result of ultrasound. These results were in good agreement with studies on other bacteria. For example, an increasing acoustic power has previously been shown to enhance disruption of bacteria such as Escherichia coli (Ho et al. 2006) and Acetobacter peroxydans (Kapucu et al. 2000). High power levels improve disruption, but can also increase further fragmentation of the cell debris. For example, micronization, or the formation of smaller fragments has been reported in ultrasound mediated disruption of the yeast Saccharomyces cerevisiae (Shynkaryk et al. 2008). Excessively micronized debris is not wanted as it is difficult to remove from the sonicated broth by centrifugation and filtration. Also, high power levels are more likely to damage sensitive enzymes and other macromolecules. A significant p-value and the negative β-coefficient value of acoustic power (C in Eq. 1) for ADI release suggested its significant but negative effect on the ADI release. Ultrasound is generally applied as a pulse; i.e. the sound is cyclically switched on and off. The total time of a cycle is the sum of the periods that the sound is on and off. Duty cycle of sonication is the ratio of the on time to the total cycle time expressed as a percentage. Increase in cell disruption and ADI release was observed with increasing duty cycle from 20 to 60%. This was in a good agreement with the previously reported effects of the duty cycle on cell disruption and protein release (Apar and Obek 2008; Liu et al. 2013). A high value of duty cycle can result in erosion and pitting of the tip of the sonic horn and excessive generation of heat leading to denaturing of sensitive enzymes. Longer off intervals between the sound pulses in a low duty cycle operation allow time for heat to dissipate from the vicinity of the tip of the sonic horn and this helps in reducing protein denaturation (Liu et al. 2013).

16

For the same bacterium as used in the present study, at an initial biomass concentration of 100 g/L, a disruption treatment in a French press was previously reported for ADI release (Patil et al. 2016c). The French press treatment released a maximum ADI amount of 22.5 U/mL. Compared with this, the ADI released after 6 minutes of sonication was higher (26.5 U/mL; Run 13, Table 1). From the same initial concentration of the slurry, the total protein release by sonication was (4.1 g/L; Run 13, Table 1) was 1.5-fold greater than in the release in the French-press (Patil et al. 2016c). The average values of the firstorder disruption rate constants for the sonication-based process were as follows: KA = 0.014 ± 0.007 Sec−1; Kc = 0.009 ± 0.005 Sec −1; and Kp = 0.011 ± 0.006 Sec −1. The standard deviation of Kc was fairly high likely because of a relatively poor correlation between the optical density measurements and the fraction of cells disrupted. Variations in the kinetic constants for the responses implied that ADI release and cell disruption kinetics are the functions of process variables of the ultrasound.

4. Conclusion The ANN model had the best generalization capability of the models tested. The ADI release was maximized by the following processing conditions: an initial biomass concentration of 300 g/L in the cell slurry; a sonication treatment time of 6 min; an acoustic power of 187.5 W; and a duty cycle of 40%. With these conditions the ADI release exceeded 27 U/mL and the total protein release exceeded 5.6 g/L. Variations in the kinetic constants for performance indices implied that the cell disruption and release kinetics is a function of the independent parameters used.

(Supporting information for this manuscript is available from the website of the journal)

17

Acknowledgments MDP and GP gratefully acknowledge Department of Biotechnology (DBT), New Delhi, India, for the award of Senior Research Fellowships. Authors are thankful to Mr. Mukesh Kumar for technical assistance with the bioreactor runs. Authors also gratefully acknowledge Professor Manfred Zinn, Laboratory for Biomaterials, Empa-Swiss Federal Laboratories for Materials Science and Technology, Switzerland for the kind gift of Pseudomonas putida KT2440 strain. Competing interests All authors declare that they have no competing interests. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

References 1.

Ahn, K-Y., Lee, B., Han, K-Y., Song, J-A., Lee, J., 2014. Synthesis of Mycoplasma arginine deiminase in E. coli using stress-responsive proteins. Enzyme Microb. Technol. 63, 46–49.

2.

Apar, D.K., Obek, B., 2008. Protein releasing kinetics of baker’s yeast cells by ultrasound. ChemBiochem. Eng. Q. 22, 113–118.

3.

Apul, O.G., Xuan, P., Luo F., Karanfil, T., 2013. Development of a 3D QSPR model for adsorption of aromatic compounds by carbon nanotubes: comparison of multiple linear regression, artificial neural network and support vector machine. RSC Adv. 3, 23924–23934.

4.

Arun, C., Sivashanmugam, P., 2016. Study on optimization of process parameters for enhancing the multi-hydrolytic enzyme activity in garbage enzyme produced from

18

preconsumer

organic

waste.

Bioresour.

Technol.

http://dx.doi.org/10.1016/j.biortech.2016.12.029 5.

Boyde, T.R.C., Rahmatullah, M., 1980. Optimization of conditions for the calorimetric determination of citrulline using diacetyl monoxime. Anal. Biochem. 107, 424–431.

6.

Bradford, M.M., 1976. A rapid and sensitive method for the quantitation of microgram quantities of protein utilizing the principle of protein-dye binding. Anal. Biochem. 72, 248–254.

7.

Buyukada, M., 2016. Co-combustion of peanut hull and coal blends: Artificial neural networks modeling, particle swarm optimization and Monte Carlo simulation, Bioresour. Technol. 216, 280-286.

8.

Chisti, Y., 2003. Sonobioreactors: using ultrasound for enhanced microbial productivity. Trends Biotechnol. 21, 89–93.

9.

Chisti, Y., Moo-Young, M., 1986. Disruption of microbial cells for intracellular products. Enzyme Microb. Technol. 8, 194–204.

10.

Doulah, M.S., 1977. Mechanism of disintegration of biological cells in ultrasonic cavitation. Biotechnol. Bioeng. 19, 649–660.

11.

Fayura, L.R., Boretsk, Y.R., Pynyaha, Y.V., Wheatley, D.N., Sibirny, A.A., 2013. Improved method for expression and isolation of the Mycoplasma hominis arginine deiminase from the recombinant strain of Escherichia coli. J. Biotechnol. 167, 420– 426.

12.

Feliu, J.X., Cubarsi, R., Villaverde, A., 1998. Optimized release of recombinant proteins by ultrasonication of E. coli cells. Biotechnol. Bioeng. 58, 536–540.

13.

Fultang, L., Vardon, A., De Santo, C., Mussai, F., 2016. Molecular basis and current strategies of therapeutic arginine depletion for cancer. Int. J. Cancer 139, 501–509.

19

14.

Gao, S., Lewis, G.D., Muthupandian, A., Hemar Y., 2014. Inactivation of microorganisms by low-frequency high-power ultrasound: 1. Effect of growth phase and capsule properties of the bacteria. Ultrason. Sonochem. 21, 446–453.

15.

Goldberg, Y., Elhadad, M., 2008. split SVM: fast, space-efficient, non-heuristic, polynomial kernel computation for NLP applications. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics on Human Language Technologies: Short Papers (pp. 237–240). Association for Computational Linguistics.

16.

Greenly, J.M., Tester, J.W., 2015. Ultrasonic cavitation for disruption of microalgae. Bioresour. Technol. 184, 276-279.

17.

Ho, C.W., Chew, T.K., Ling, T.C., Kamaruddin, S., Tan, W.S., Tey, B.T., 2006. Efficient mechanical cell disruption of Escherichia coli by an ultrasonicator and recovery of intracellular hepatitis B core antigen. Process Biochem. 41, 1829–1834.

18.

Hua, I., Thompson, J.E., 2000. Inactivation of Escherichia coli by sonication at discrete ultrasonic frequencies. Wat. Res. 34, 3888–3893.

19.

Iida, Y., Tuziuti, T., Yasui, K., Kozuka, T., Towata, A., 2008. Protein release from yeast cells as an evaluation method of physical effects in ultrasonic field. Ultrason. Sonochem. 15, 995–1000.

20.

Jaeschke, D.P., Rech, R., Marczak, L., Mercali, G. D., 2016. Ultrasound as an alternative technology to extract carotenoids and lipids from Heterochlorella luteoviridis. Bioresour. Technol. http://dx.doi.org/10.1016/j.biortech.2016.11.107

21.

Kakimoto, T., Shibatani, T., Nishimura, N., Chibata, I., 1971. Enzymatic production of L-citrulline by Pseudomonas putida. Appl. Microbiol. 22, 992–999

20

22.

Kapucu, H., Gulsoy, N., Mehmetoglu, U., 2000. Disruption and protein release kinetics by ultrasonication of Acetobacter peroxydans cells. Biochem. Eng. J. 5, 57– 62.

23.

Liu, D., Zeng, X., Sun, D., Han, Z., 2013. Disruption and protein release by ultrasonication of yeast cells. Innov. Food Sci. Emerg. Technol. 18, 132–137.

24.

Maran, J.P., Priya, B., 2015. Comparison of response surface methodology and artificial neural network approach towards efficient ultrasound-assisted biodiesel production from muskmelon oil. Ultrason. Sonochem. 23, 192–200.

25.

Marchitan, N., Cojocaru, C., Mereuta, A., Duca, G., Cretescu, I., Gonta, M., 2010. Modeling and optimization of tartaric acid reactive extraction from aqueous solutions: A comparison between response surface methodology and artificial neural network. Sep. Purific. Technol. 75, 273–285.

26.

Nair, V.V., Dhar, H., Kumar, S., Thalla, A.K., Mukherjee, S., Wong, J.W., 2016. Artificial neural network based modeling to evaluate methane yield from biogas in a laboratory-scale

anaerobic

bioreactor. Bioresour.

Technol.

http://dx.doi.org/10.1016/j.biortech.2016.03.046. 27.

Ni, Y., Liu, Y., Schwaneberg, U., Zhu, L., Li, N., Li, L., Sun, Z., 2011. Rapid evolution of arginine deiminase for improved anti-tumor activity. Appl. Microbiol. Biotechnol. 90, 193–201.

28.

Ozbek, B., Ulgen, K.O., 2000. The stability of enzymes after sonication. Process Biochem. 35, 1037–1043.

29.

Patel, G., Patil, M.D., Soni, S., Khobragade, T.P., Chisti, Y., Banerjee, U.C., 2016. Production of mycophenolic acid by Penicillium brevicompactum—A comparison of two methods of optimization. Biotechnol. Rep. 11, 77–85.

21

30.

Patil, M.D., Bhaumik, J., Babykutty, S., Banerjee, U.C., Fukumura, D., 2016a. Arginine dependence of tumor cells: targeting a chink in cancer’s armor. Oncogene 35, 4957–4972.

31.

Patil, MD, Shinde, K.D., Patel, G., Chisti, Y., Banerjee, U.C., 2016b. Use of response surface method for maximizing the production of arginine deiminase by Pseudomonas putida. Biotechnol. Rep. 10, 29–37.

32.

Patil, M.D., Patel, G., Surywanshi, B., Shaikh, N., Garg, P., Chisti, Y., Banerjee, U.C., 2016c. Disruption of Pseudomonas putida by high pressure homogenization: a comparison of the predictive capacity of three process models for the efficient release of arginine deiminase. AMB Expr. 6, 84.

33.

Sarve, A., Sonawane, S.S., Varma, M.N., 2015. Ultrasound assisted biodiesel production from sesame (Sesamum indicum L.) oil using barium hydroxide as a heterogeneous catalyst: comparative assessment of prediction abilities between response surface methodology (RSM) and artificial neural network (ANN). Ultrason. Sonochem. 26, 218–228.

34.

Shynkaryk, M.V., Lebovka, N.I., Lanoiselle, J.L., Nonus, M., Bedel-Clatour, C., Vorobiev, E., 2008. Electrically-assisted extraction of bio-products using high pressure disruption of yeast cells (Saccharomyces cerevisiae). J. Food Engineer. 92, 189–195.

35.

Singh, R.S., 2013. A comparative study on cell disruption methods for the release of aspartase from E. coli K-12. Ind. J. Exp. Biol. 51, 997–1003.

36.

Singh, R., Banerjee, A., Kaul, P., Barse, B., Banerjee, U.C., 2005. Release of an enantioselective nitrilase from Alcaligenes faecalis MTCC 126: a comparative study. Bioprocess Biosyst. Eng. 27, 415–424.

22

37.

Smola, A.J., Scholkopf, B., 2004. A tutorial on support vector regression. Stat. Comput. 14, 199–222.

38.

Üstün-Aytekin, Ö., Arısoy, S., Aytekin, A., Yıldız, E., 2016. Statistical optimization of cell disruption techniques for releasing intracellular X-prolyl dipeptidyl aminopeptidase from Lactococcus lactis spp. Lactis. Ultrason. Sonochem. 29, 163– 171.

39.

Yadav, R.S., Yadav, M., 2013. Enhanced production of recombinant aspartase of Aeromonas media NFB-5 in a stirred tank reactor. Bioresour. Technol. 145, 217–223.

40.

Yu, Z-L., Zeng, W-C., Zhang, W-H., Liao, X-P., Shi, B., 2014. Effect of ultrasound on the activity and conformation of a-amylase, papain and pepsin. Ultrson. Sonochem. 21, 930–936.

41.

Zhang, L., Liu, M., Jamil, S., Han, R., Xu, G., Ni, Y., 2015. PEGylation and pharmacological characterization of a potential anti-tumor drug, an engineered arginine deiminase originated from Pseudomonas plecoglossicida. Cancer Lett. 357, 346–354.

23

Figure captions

Fig 1 Architecture of the ANN model Fig 2 Response surface plots showing the interactive effects of the following on ADI release: (A) acoustic power and duty cycle; (B) biomass concentration and sonication time; (C) acoustic power and sonication time; (D) sonication time and duty cycle; (E) acoustic power and biomass concentration; and (F) duty cycle and biomass concentration.

Fig 3 Response surface plots showing the interactive effect of the following on total protein release: (A) biomass concentration and sonication time; (B) acoustic power and sonication time; (C) duty cycle and sonication time; (D) acoustic power and biomass concentration; (E) duty cycle and biomass concentration; and (F) duty cycle and acoustic power. Fig 4 Response surface plots showing the interactive effect of the following on percentage of the cells disrupted: (A) biomass concentration and sonication time; (B) acoustic power and sonication time; (C) duty cycle and sonication time; (D) acoustic power and biomass concentration; (E) duty cycle and biomass concentration; and (F) duty cycle and acoustic power. Fig 5 Comparison between ANN-predicted response and the measured responses

24

Fig. 1

25

Fig. 2

26

Fig. 3

27

Fig. 4

28

Figure 5

29

Table 1. Factor values and responses in the various runs

Run

A

B

(min)

(g/L)

C (W)

ADI activity (U/mL)

D (%)

Actual

Total protein release (g/L)

RSM

ANN

SVM

predicted

predicted

predicted

Actual

Cell disruption (%)

RSM

ANN

SVM

predicted

predicted

predicted

Actual

RSM

ANN

SVM

predicted

predicted

Predicted

1

10

100

225

60

18.2

17.9

18.3

17.3

3.2

3.4

3.4

4.0

83.5

85.9

85.0

83.6

2

10

100

150

60

17.3

17.2

18.0

20.4

3.3

3.6

3.1

3.3

62.3

62.9

62.9

62.2

3

2

100

225

20

15.2

13.4

16.5

17.3

2.5

3.0

2.9

3.0

34.5

36.6

35.0

41.3

4

2

300

150

20

24.4

24.0

23.9

24.0

3.8

3.9

3.8

4.4

15.4

12.3

16.3

21.1

5

6

200

187.5

40

21.6

22.2

22.3

20.6

5.0

5.0

5.0

4.2

72.7

71.2

72.6

52.4

6

2

300

225

60

27.8

28.0

27.5

25.6

4.0

4.4

4.2

4.4

35.4

36.6

37.5

42.3

7

2

200

187.5

40

23.0

20.7

22.5

23.0

3.4

3.7

3.4

3.4

30.4

33.9

31.2

30.6

8

2

300

150

60

25.0

24.6

24.5

28.6

3.7

3.7

3.6

3.7

15.5

15.0

15.3

19.8

9

2

100

150

60

17.5

18.0

17.4

25.0

1.6

1.5

1.6

1.6

13.0

14.0

12.6

18.5

10

10

300

225

60

17.8

17.0

18.0

20.9

5.2

5.2

5.4

6.0

77.0

79.2

78.3

85.0

11

6

200

187.5

40

21.5

22.2

22.3

20.6

5.3

5.0

5.0

4.2

69.8

71.3

72.6

52.4

12

6

200

187.5

40

22.6

22.2

22.3

20.6

5.3

5.0

5.0

4.2

72.9

71.3

72.6

52.4

13

6

100

187.5

40

26.5

24.9

23.7

18.8

4.1

4.0

4.1

3.2

72.3

71.2

72.0

51.7

14

6

200

187.5

20

18.3

19.2

20.0

18.3

4.6

5.0

4.8

4.6

66.4

70.7

62.4

53.1

15

2

100

225

60

22.0

23.3

21.2

21.9

2.2

2.0

1.9

2.3

36.5

34.8

34.1

40.0

16

10

100

225

20

12.6

13.8

15.4

12.7

4.4

4.1

4.4

4.7

83.8

84.6

83.8

85.0

17

6

200

225

40

18.1

18.6

17.9

19.1

5.3

5.0

5.0

4.6

80.3

75.6

78.6

63.2

18

6

200

187.5

40

22.0

22.2

22.3

20.6

4.8

5.0

5.0

4.2

75.0

71.3

72.6

52.4

30

19

6

200

187.5

40

21.5

22.2

22.3

20.6

4.7

5.0

5.0

4.2

73.0

71.3

72.6

52.4

20

10

300

150

60

17.0

18.1

18.4

24.0

5.3

5.2

5.2

5.4

58.2

55.4

56.4

63.5

21

2

300

225

20

19.2

20.3

18.8

20.9

5.8

5.1

5.6

5.1

42.8

42.5

42.9

42.6

22

10

200

187.5

40

16.0

17.7

13.0

18.3

5.4

5.0

5.0

5.0

79.5

78.1

82.0

74.2

23

10

300

225

20

16.3

15.0

15.0

16.3

5.1

5.5

5.4

6.8

83.7

82.0

80.9

86.3

24

6

300

187.5

40

27.7

28.8

29.1

22.5

5.6

5.8

5.6

5.2

67.2

70.4

66.1

53.1

25

10

300

150

20

23.6

23.2

23.0

19.4

5.2

5.1

5.0

6.0

47.5

49.5

46.2

64.8

26

6

200

187.5

60

22.1

22.2

22.3

20.6

5.0

5.0

5.0

4.2

70.5

71.3

72.6

52.4

27

6

200

187.5

60

23.0

21.5

21.7

23.0

5.0

4.6

5.0

3.8

74.8

72.7

74.9

51.7

28

2

100

150

20

13.5

15.2

14.6

20.4

2.3

2.0

2.1

2.3

9.1

7.2

10.3

19.8

29

10

100

150

20

21.0

20.1

19.3

15.7

3.9

3.9

3.9

4,0

55.0

53.0

52.8

63.5

30

6

200

150

40

21.2

20.1

21.0

22.2

4.2

4.5

4.6

3.8

42.3

49.0

48.3

41.7

A, sonication time (min); B, biomass concentration (g/L); C, acoustic power (W); D, duty cycle (%)

31

Table 2. Average R2 values for different ANN functions and 1–10 neurons

Algorithm

Number of neurons

1

2

3

4

5

6

7

8

9

10

Training dataset BFGS Quasi-Newton

0.00

0.31

0.45

0.50

0.51

0.74

0.78

0.78

0.78

0.78

Conjugate gradient with Powell/Beale restarts

0.00

0.33

0.48

0.62

0.75

0.84

0.82

0.82

0.85

0.80

Fletcher-Powell conjugate gradient

0.00

0.38

0.44

0.56

0.75

0.65

0.81

0.81

0.83

0.81

Polak-Ribiére conjugate gradient

0.14

0.39

0.47

0.66

0.80

0.69

0.82

0.80

0.85

0.89

Levenberg-Marquardt

0.11

0.38

0.52

0.72

0.80

0.89

0.92

0.93

0.92

0.93

Scaled conjugate gradient

0.07

0.27

0.47

0.66

0.69

0.81

0.88

0.76

0.87

0.82

Test dataset BFGS Quasi-Newton

0.22

0.25

0.54

0.55

0.45

0.74

0.74

0.81

0.72

0.72

Conjugate gradient with Powell/Beale restarts

0.19

0.53

0.41

0.69

0.80

0.80

0.82

0.76

0.79

0.69

32

Fletcher-Powell conjugate gradient

0.32

0.41

0.44

0.65

0.74

0.53

0.77

0.83

0.76

0.73

Polak-Ribiére conjugate gradient

0.17

0.37

0.46

0.67

0.78

0.77

0.73

0.76

0.75

0.83

Levenberg-Marquardt

0.18

0.50

0.48

0.70

0.83

0.80

0.79

0.83

0.81

0.84

Scaled conjugate gradient

0.22

0.59

0.46

0.61

0.75

0.79

0.76

0.86

0.73

0.82

33

Table 3. Experimental and predicted responses for the unseen dataset

Run

A

B (g/L) C (W)

(min)

ADI activity (U/mL)

D (%)

Actual

RSM

ANN

Total protein release (g/L) SVM

Actual

predicted predicted predicted

RSM

ANN

Cell disruption (%) SVM

Actual

predicted predicted predicted

RSM

ANN

SVM

predicted predicted predicted

31

5

100

184

48

23.0

25.2

23.5

17.5

3.9

3.6

3.6

3.5

67.7

63.7

65.2

57.9

32

6.11

230

177

42

24.4

23.3

24.8

17.5

4.6

5.2

5.1

5.5

70.2

67.7

64.4

68.5

33

2.67

300

171

45

25.6

28.9

27.1

24.7

4.8

4.7

4.4

4.3

39.8

36.1

36.3

20.1

34

3.64

200

195

60

21.3

22.5

23.7

23.1

4.3

4.0

4.3

3.7

52.7

55.7

54.6

48.2

35

5.36

295

202

50

26.7

28.1

27.4

19.3

5.1

5.6

5.6

6.0

67.8

70.5

68.9

76.2

36

4.55

100

200

60

23.0

25.2

24.0

19.6

2.8

3.1

3.1

3.2

59.2

60.4

60.1

62.5

A, sonication time (min); B, biomass concentration (g/L); C, acoustic power (W); D, duty cycle (%)

34

Table 4. Error analyses of the three predictive models Statistical parameter

ADI activity (U/mL) RSM

ANN

Total protein release (g/L) SVM

RSM

ANN

Cell disruption (%)

SVM

RSM

ANN

SVM

Training dataset R2

0.926

0.900

0.296

0.929

0.962

0.658

0.988

0.993

0.737

MSE

1.097

1.486

11.581

0.087

0.046

0.470

6.423

3.910

160.194

RMSE

1.048

1.219

3.403

0.296

0.215

0.686

2.534

1.977

12.656

SEP

5.125

5.963

16.646

6.869

4.994

15.942

4.472

3.489

22.333

RPD

4.704

4.988

12.387

6.324

4.204

11.868

5.120

3.255

21.430

Unseen dataset R2

0.675

0.857

0.001

0.818

0.835

0.632

0.888

0.932

0.862

MSE

4.113

1.697

24.679

0.155

0.133

0.429

15.020

9.659

93.148

RMSE

2.028

1.303

4.968

0.394

0.365

0.655

3.876

3.107

9.651

SEP

8.451

5.428

20.700

9.286

8.604

15.416

6.505

5.216

16.198

RPD

7.811

4.767

17.776

8.707

7.309

14.057

6.472

4.586

15.162

R2, determination coefficient; MSE, mean square error; RMSE, root mean square error; SEP, standard error of prediction; RPD, relative percent deviation.

35

Table 5. Rate constants for the ADI release, total protein release and cell disruption Run#

KA

Kc

Kp

Run

KA

Kc

Kp

Run

KA

Kc

Kp

1

0.006

0.008

0.006

15

0.02

0.018

0.019

29

0.009

0.003

0.004

2

0.006

0.003

0.005

16

0.015

0.006

0.004

30

0.01

0.006

0.007

3

0.026

0.015

0.023

17

0.009

0.007

0.008

4

0.027

0.024

0.029

18

0.015

0.006

0.008

5

0.011

0.007

0.008

19

0.012

0.008

0.008

6

0.016

0.015

0.018

20

0.007

0.003

0.005

7

0.032

0.016

0.015

21

0.028

0.019

0.019

8

0.013

0.014

0.014

22

0.011

0.006

0.005

9

0.014

0.012

0.02

23

0.008

0.005

0.006

10

0.011

0.005

0.005

24

0.011

0.009

0.009

11

0.012

0.008

0.01

25

0.009

0.002

0.004

36

12

0.014

0.006

0.006

26

0.014

0.009

0.01

13

0.014

0.006

0.011

27

0.015

0.007

0.007

14

0.013

0.006

0.011

28

0.021

0.012

0.014

KA, Rate constant for ADI release (s−1); Kc, Rate constant for fraction of the cell disrupted (s−1); Kp, Rate constant for total protein released (s−1). #

The experimental conditions were similar as mentioned in Table 1

Highlights


Ultrasound-mediated ADI release from P. putida using machine learning techniques
Satisfactory prediction of the responses implied the proficient capabilities of ANN
ADI release was better than that of high pressure homogenization
Kinetic studies for the performance responses of ultrasound-mediated ADI release