Materials
All chemicals used in this study were of analytical grade and purchased from Sigma-Aldrich and Fisher scientific, unless stated otherwise. CG used was kindly provided by Greenergy, UK. The CG contained glycerol (72.8%), non-glycerine material (soaps, fatty acids, esters, salts, other organic byproducts) (5.7%), methanol (2.0%), water (12.2%) and ash (9.6%). The xylose (26.4 g/L) rich SCB hydrolysate was obtained from Nova Pangea Technologies, UK.
Microorganism, culture maintenance and inoculum preparation
The current study made use of Y. lipolytica Po1t (Ura+, Leu+) derived from wild-type strain W29 (ATCC20460). The Y. lipolytica strain was preserved in 20 % glycerol (v/v) at - 80 °C and maintained on a petri dish containing YPD agar medium (1% yeast extract, 2% Peptone, 2% Dextrose and 2% Agar) at pH 7.0 and 30 °C. The seed culture was grown in a 250 mL Erlenmeyer flask containing 50 mL YPD broth. The final pH of the medium prior to sterilization was adjusted to 7.0. Cultivation was carried out for 24 h at 30 °C on a rotary shaker at an agitation speed of 250 RPM.
Submerged cultivations in shake flask
The fermentation medium had the following composition: (g/L) PG/CG/glucose, 20; xylose, 20; yeast nitrogen base (YNB), 1.7; NH4Cl, 1.5. The medium was prepared in 50 mM phosphate buffer. The initial pH was adjusted to 6.8 before inoculation by using 5N NaOH. The submerged cultivations were carried out in 500 mL shake flasks containing 100 mL working volume. The flasks were inoculated with fresh inoculum at OD600 (optical density) of 0.1 and kept at 30 °C under constant shaking at 250 RPM on a rotary shaker (Excella 24, New Brunswick).
Central composite design (CCD) and artificial neural network linked genetic algorithm (ANN-GA) for media optimization
The CCD was carried out, with the view of optimizing the variables and to give insight over the combined effect of four variables (xylose, YNB, NH4Cl and phosphate buffer) at constant glycerol concentration on maximizing the production of xylitol concentration. Design-Expert software (version 7.0) was used to develop CCD for four independent variables and five levels (Table 1). The total number of experiments (N) was based on Equation (1) (see Supplementary Files)
where k is the number of independent variables. The experiment comprised 2 axial points and 6 replicates for centre points for the evaluation of pure error. The second-order polynomial for predicting the optimal levels was expressed according to the Equation (2). (see Supplementary Files)
where, Yi is the Predicted response; β0 βi, βij, βii are constant and regression coefficients of the model, Xi, Xj represent the independent variables in coded values and ε represents the error.
To further optimise the media components, the artificial neural network (ANN) methodology was adapted. ANN is biological inspired model, which mimics neural system and tends to optimize non-linear systems. Multi-layer perceptron method was incorporated, and training of the network was based on feed-forward back propagation method. The network architecture consisted of four input layers (xylose, YNB, NH4Cl, phosphate buffer), eight hidden layers and one output layer representing xylitol concentration. In the feed-forward training system, the data was channelized from input to output via., hidden layer, which was connected by parameters such as weights (w) and biases (b). Transfer functions such as tan sigmoid (f1: tansig) and Pure linear (f2: purelin) were situated between hidden and output layer, respectively. Tansig sums up weighted input including the biases, and the purelin carried out the linearization function for the output. The predicted output function is represented by the Equation (3) (see Supplementary Files)
where Yp is the predicted response, wo, bo and wH, bH are weights and biases of the output and hidden layer, respectively. The network training was carried out by adapting Levenberg-Marquardt (LM) backpropagation algorithm, which calculates error function based on the difference between actual output and predicted output. The algorithm was trained repeatedly until subsequent minimisation in the error between the input and output layer is met [65]. Mean squared error (MSE) was used to calculate error function using Equation (4). (see Supplementary Files)
where, Ya is the actual output, Yp is the predicted output and N is the number of data points. The simulation of the network was carried out by in built neural network toolbox of MATLAB (version 2010a).
Genetic algorithm (GA) is a heuristic method used to determine the global optimal solution for a non-linear problem and are independent of initial values; GA is often coupled with ANN to achieve precise optimization values. GA follows four steps to find a global solution. In the first step, initialization of the solution for the population takes place followed by fitness computation. The selected individual based on the fitness computation then undergoes crossing over and mutation, creating a new set of individuals [66,67]. This process is repeated until a global optimum value is achieved.
The trained neural network model was used as a fitness function to further optimise the input space. The schematic representation of ANN-GA algorithm for optimisation of medium components to maximize xylitol production was shown in Fig S1 (supplementary data). The objective function of GA is given by Equation 5: (see Supplementary Files)
where f is the objective function (ANN model), x denotes input vector, w denotes corresponding weight vector, Y refers to the xylitol experimental yield, X denotes operating conditions, P denotes number of input variables, xiL & xiU are lower and upper bounds of xi fitness of each candidate solution.
Model validation under shake flask conditions
The integration of CCD and ANN-GA predicted some crucial parameters and their concentrations, which could give optimum xylitol yields. Therefore, it became essential to validate the predicted values at shake flask level based on the global optimum values obtained by ANN-GA training. Shake flask studies were conducted with PG as co-substrate. However, simultaneously the efficacy of Y. lipolytica Po1t (Ura+ Leu+) using co-substrate combinations namely PG + xylose rich SCB hydrolysate and CG + pure xylose was also evaluated to assess the tolerance, utilization and biovalorizaion ability of the said strain for carbon sources derived from renewable feedstock.
Batch cultivation in bioreactor
The batch experiments were performed in a 2.5 L bioreactor (Electrolab Bioreactors, UK) with 1.0 L working volume. The inoculum was prepared using optimised media and the optimum values of media components were as follows (g/L): PG/CG, 20; xylose, 55; YNB, 5.0; NH4Cl, 3.94; phosphate buffer, 132.5 mM. The starting pH was 6.8 and not controlled during the fermentation. The temperature and agitation speed were controlled at 30 ⁰C and 250 RPM, respectively, while the aeration rate was maintained at 2.0 L/min for initial 48 h and then changed to 1.0 L/min for the rest of fermentation period.
Biotransformation by resting cells
For active cells, Y. lipolytica was grown on optimised medium with PG in 500 mL flasks containing with 20% working volume. The temperature, pH and agitation speed were maintained at 30°C, 6.8 and 250 RPM, respectively. For the second stage (biotransformation), the cells were harvested in the late exponential period (after 48 h) when the OD600 was somewhere between 20-25. Immediately after, the culture was centrifuged at 2800 x g for 10 min, and the resulting pellet was washed with ice-cold 100 mM phosphate buffer (pH 7.0). The cells were resuspended in a bioconversion medium containing xylose (30, 70 and 100 g/L) in phosphate buffer (100 mM). The bioconversion experiments were carried out at 30 °C with freshly prepared biomass.
Downstream processing of xylitol
The purification protocol for xylitol was performed according to Rivas et al., (2006) [68]. The 100 ml of spent fermentation broth was subjected to centrifugation at 20000 x g to separate the cells and the clarified broth was treated with 5% activated charcoal. The charcoal treated broth was precipitated by adding four volume of absolute ethanol and incubated at 4°C for 1h. After 1h, the precipitates were removed by centrifuging the mixture at 4000 x g for 10 min. The supernatant was vacuum concentrated at 40 °C. The concentrated sample and ethanol were mixed at a ratio of 1:4 and incubated at -20 °C with slight agitation (50 RPM) until crystals were observed. To improve the crystallization about 1 g/L of xylitol was mixed with the concentrated sample.
Analytical methods
The samples were withdrawn periodically and analysed for OD, pH, residual glycerol/glucose, xylose and xylitol. Cell growth was quantified by measuring the optical density at 600 nm wavelength in a 1 mm-path-length cuvette using a double beam spectrophotometer (Jenway 6310, UK). One unit of absorbance at 600 nm corresponded to a cell dry weight (CDW) of 0.21 g/L. The concentrations of glycerol, glucose, xylose and xylitol were measured by high performance liquid chromatography (Agilent Technologies 1200 series, USA). The supernatants, obtained by centrifugation of the culture samples at 10,000 x g for 10 min, were filtered through a 0.22 µm PVDF membrane (Sartorious, Germany) and eluted using Rezex ROA-Organic Acid H+ (Phenomenex, USA) column at 60 °C attached with refractive index detector (RID). The mobile phase and flow rate were 0.5 mM H2SO4 and 0.4 mL/min, respectively. All measurements were conducted in triplicates and the values were averaged. The standard deviation was not more than 10 %.