Constrained azeotropic optimization of extraction system components for the safe and efficient recovery of a desired metabolite (e.g., 3-demethylated colchicine)

Abstract

This study suggests an efficient strategy for the maximal extraction of human consumable metabolites in a safe and non-toxic solvent system for the cases where the best extraction solvent belongs to the restricted class (ICH, Quality Guidelines). An optimum composition of extraction system components (chloroform : acetone : ethyl alcohol) was determined for the maximum and chloroform free recovery of 3-demethylated colchicine (3-DMC) from a fermentation broth. A feed forward backprop artificial neural network (ANN) was trained and validated to establish a mathematical relationship between the inputs (volumes of extraction system components) and outputs (concentration of 3-DMC). Inputs to the ANN were kept under strict constraints, always satisfying the azeotropic ratio for chloroform : acetone with an excess of ethyl alcohol forming a complex ternary azeotrope. A genetic algorithm (GA) with linear equality matrices defining ternary azeotropic constraints was implemented to perform an exhaustive search for the global optimum composition of the extraction system components with the validated ANN as the selection function. Since, this strategy only allows for the azeotropic ratio, the global optimum is a typical azeotrope with maximum 3-DMC recovery. The purpose of the azeotrope is to safely distil out all the chloroform after extraction, by performing an azeotropic distillation at a lower temperature (63.2 °C) leaving only an excess of ethyl alcohol, which boils at a higher temperature (78 °C). The final optimized volumes of chloroform: 63.055 mL, acetone: 44.138 and ethanol: 92.807 mL resulted in a maximum recovery of 6.5 g L⁻¹ of 3-DMC. This approach was also tested successfully for the energy efficient recovery of betulinic acid in a Class 3 solvent.

---

Introduction

Azeotropic distillation (AD) is a method of separating the components of a mixture which have a tendency to form azeotropes.¹ AD has been commonly used for the dehydration of ethanol.² It has been used to break azeotropes thereby separating benzene and cyclohexane, methanol and acetone, isopropyl alcohol, water mixtures, etc.^3–5 Purification of azeotropic mixtures using AD (using entrainers) results in a high degree of purity (≥99.5%) of the separated mixture/system components.^4,6,7 Kindly refer to Treybal (1980)¹ and/or Moore (1962)⁸ for more explanation of the types and characterization of azeotropes and AD. AD has been used for enhanced bioethanol dehydration,⁹ purification, recovery of enzymic peptides,¹⁰ and bio-recycling of hexanes from laboratory waste mixtures of hexanes and ethyl acetate.¹¹ Also, AD has been used to perform repeated dehydrations for the purification of monosaccharide fatty acid esters synthesized via a chemo-enzymic process,¹² and for the removal of byproducts in the lipase catalysed glucose fatty acid ester synthesis.¹³ In all these cases the AD process was successfully designed to maximize the recovery of the respective case specific molecules.

There are several statistical and machine learning methods available for establishing a mathematical relationship between dependent and multiple independent variables. However, machine learning methods provide an edge by being open ended and flexible, they can be easily applied in the cases where little or no prior information is available regarding the underlying function. Their disadvantage lies in being black box techniques, where the actual meaning of the coefficients is not clear or mixed. However, machine learning techniques such as artificial neural networks (ANNs) prove to be most appropriate for the cases where the concern is not the actual underlying function but its outcomes (for a given class of inputs). ANNs have been successfully used to model several biological behaviours including fermentation conditions and process designs.^14–18 In the present study, an evolutionary optimization technique, a genetic algorithm (GA) was used to optimize the azeotropic ratio of solvent system components. The GA is a method for solving optimization problems based on natural selection.^19,20 The reason for selecting a GA as an optimization technique in the present study is its robustness for use even in discontinuous functions as its selection function and its ability to generate the desired constrained population within the given bounds.²⁰ GAs have been used to optimize several biochemical and process design functions to obtain maximum or minimum amounts of the desired or undesired molecules, respectively.^{15,18,21–24}

Due to selectivity and higher yields, the biotransformation process is gaining substantial importance in the pharmaceutical sector.^25–27 Some of the major problems associated with the biotransformation process are low substrate solubility in aqueous media, higher residual substrate concentrations, product toxicity, feedback inhibition effects, etc.^28–30 One of the important applications of the biotransformation process is the conversion of colchicine into pharmacologically active colchicine derivatives i.e., 3-demethylated colchicine (3-DMC), thiocolchicoside, and colchicoside.^31,32 These colchicine derivatives have greater therapeutic importance as anti-inflammatory and anti-tumor drugs in the treatment of various forms of leukemia and solid tumors.^29,33–36 Few studies on the B. megaterium/recombinant E. coli derived microbial biotransformation of colchicine have been published in the last few years and have demonstrated the regiospecific demethylation of colchicine molecules in the tropolone ring at the C-3 position through the P450 BM-3 (product of CYP102A1 gene) mono-oxygenase enzyme.^29,35,37,38 Hence, optimization of an extraction process has become vital for obtaining higher recoveries of 3-DMC, which ensures the economic feasibility of the microbial biotransformation process. Earlier, we reported solvent extraction of 3-DMC using a classical approach and found higher recoveries in chloroform as it was proven to be the best solvent for 3-DMC under optimum conditions.³⁸ However, complete removal or minimization of chloroform is suggested by the International Conference on Harmonization of Technical Requirements for Registration of Pharmaceuticals for Human Use [Q3C Impurities, Residual Solvents – Q3C Tables and List, Revision 2 (2012)] Guidance for Industry, Guidance for Industry, International Conference on Harmonisation – Quality, 1997, as chloroform belongs to the Class 2 solvents, so its use is strictly restricted (http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM073395.pdf).³⁹

The process of AD is used to break azeotropes by adding an entrainer that produces a new minimum boiling azeotrope. A minimum boiling azeotrope has a lower boiling point than any of its solvent components.¹ This lower boiling azeotrope will safely distil out at or above the azeotropic boiling point.¹ In the present study an optimized minimum boiling azeotrope will be generated with a Class 2 ICH restricted class solvent by adding Class 3 solvents in excess as the entrainer thereby distilling out all the Class 2 solvents in an energy efficient manner at lower temperatures. The present report is a novel attempt towards optimizing the azeotropic combination of solvent system components using artificial intelligence and evolutionary optimization methods for maximizing the non-toxic (chloroform free) recovery of a desired metabolite (i.e., 3-DMC in this case). The schematic representation of the entire process of constrained azeotropic optimization of the extraction system components for 3-DMC recovery is given in Fig. 1. The suggested strategy was also tested and validated successfully for the energy efficient optimized extraction of betulinic acid from an optimized fermentation broth in an ICH Class 3 solvent (i.e., methyl acetate).

Fig. 1 Schematic representation of the metabolite (3-DMC, in this case) recovery process using constrained azeotropic optimization of extraction system components.

Experimental

Microbial culture and fermentation conditions

Microbial inoculation and fermentation conditions were maintained as published earlier.^35,38 Briefly, the shake flask experiments were carried out to obtain the pre-culture using 100 mL and 250 mL Erlenmeyer flasks containing 15 mL and 25 mL of medium having 0.1 g L⁻¹ colchicine, respectively. After inoculation, the flasks were incubated over night at 28 °C at 200 rpm. Twenty mL of the above culture broth was then inoculated in a stirrer tank reactor (5 L STR, working volume 3 L; Sartorious Inc., Germany) containing 10 g L⁻¹ colchicine to obtain a large amount of fermented culture broth containing biotransformed 3-DMC. The cultivation time for the growth was 72 h. Operational conditions were employed as reported earlier.^35,38

Extraction and solvent system conditions

The fermentation broth recovered from the STR was subjected to microfiltration using a 0.22 μm ceramic cartridge in order to remove the microbial cell debris and other undissolved material. The clear culture filtrate containing dissolved 3-DMC was then subjected to extraction (1 : 1 v/v) under optimum process conditions (i.e., pH 10, temperature 50 °C and extraction time 120 min) as suggested by Dubey et al.³⁸

All the azeotropic Solvent Systems (SSs) mentioned in Table 1 have an azeotropic composition for each component of the system. For all non-azeotrope SS mentioned in Table 1, the components of the system were maintained in equal proportions (1 : 1 for binary; B and 1 : 1 : 1 for ternary; T). SS no. 1 is pure chloroform. Each SS was then subjected to extraction under the conditions mentioned above.

Table 1 Comparison of the chloroform containing solvent systems

SS no.	Mix	Component 1	Component 2		Component 3		Azeotrope (if any)		3-DMC recovery (g L^−1)
		BP (61 °C)	Solvent	BP (°C)	Solvent	BP (°C)	Type	Az BP (°C)
0	L	Chloroform	—	—	—	—	Non-azeotrope	—	4.2 ± 0.2
1	B	Chloroform	Acetone	56.1	—	—	Maximum boiling	64.7	4.0 ± 0.3
2	B	Chloroform	Ethanol	78.3	—	—	Minimum boiling@20 psig	59.4	3.4 ± 0.3
3	B	Chloroform	Water	100	—	—	Minimum boiling	53.3	3.6 ± 0.2
4	B	Chloroform	Ethyl acetate	77.1	—	—	Non-azeotrope	—	3.2 ± 0.2
5	T	Chloroform	Ethanol	78.3	Ethyl acetate	77.1	Non-azeotrope	—	3.9 ± 0.3
6	T	Chloroform	Water	100	Acetone	56.1	Saddle	60.2	3.5 ± 0.2
7	T	Chloroform	Acetone	56.1	Ethanol	78.3	Saddle	63.2	4.3 ± 0.2
8	T	Chloroform	Water	100	Ethanol	78.3	Saddle	78	3.9 ± 0.3

---

Table 2 shows the azeotropic Test Solvent System (TSS) for chloroform, acetone and ethanol. Each TSS contained chloroform and acetone in an exact ternary azeotropic ratio while ethanol was maintained in excess of the ternary azeotropic ratio. Each TSS was of 200 mL only. The TSS no. 24–30 were the same, this was done to minimize the degree of error. Each TSS was subjected to extraction under similar conditions as mentioned before. Acetone and ethanol present in the TSSs are miscible with water, direct extraction of 3-DMC from the aqueous fermentation broth (FB) with the TSS will disturb the azeotropic composition of the TSS. Since the extraction potential of the TSS (SS no. 7) is almost similar to pure chloroform (Table 1) 3-DMC is primarily extracted with chloroform (volume as per Table 2, per 200 mL of FB) which forms a separate immiscible bottom organic layer (containing 3-DMC) which is then collected separately and mixed with the corresponding azeotropic ratio of acetone and an excess of ethanol (as mentioned in Table 2) forming a complex ternary azeotrope. The extracted yield of 3-DMC (column 5, Table 2) served as the target and the volumes of chloroform, acetone and ethanol in column no. 2–4 of Table 2 served as the inputs for the training, validation and testing of the ANN. All the experiments were performed in triplicate and the averages are reported.

Table 2 Azeotropic test solvent system (Tr: training TSS; V: validation TSS; T: testing)

TSS no.	Chloroform (mL)	Acetone (mL)	Ethanol (mL)	3-DMC recovery (g L^−1)	ANN predicted recovery (g L^−1)
1T	10.00	7.00	183.00	1.5	1.5471
2Tr	14.00	9.80	176.20	1.6	1.7092
3T	18.00	12.60	169.40	1.9	1.9835
4V	22.00	15.40	162.60	2.3	2.3501
5Tr	26.00	18.20	155.80	2.7	2.7502
6T	30.00	21.00	149.00	3.2	3.1320
7Tr	34.00	23.80	142.20	3.6	3.5637
8Tr	38.00	26.60	135.40	4.2	4.1999
9Tr	42.00	29.40	128.60	5.0	4.9808
10Tr	46.00	32.20	121.80	5.5	5.5093
11V	50.00	35.00	115.00	5.7	5.5726
12Tr	54.00	37.80	108.20	5.8	5.6844
13Tr	58.00	40.60	101.40	6.0	6.1091
14T	62.00	43.40	94.60	6.1	6.1591
15Tr	66.00	46.20	87.80	6.1	6.1059
16V	70.00	49.00	81.00	5.8	5.6466
17Tr	74.00	51.80	74.20	5.0	5.0355
18V	78.00	54.60	67.40	4.8	4.7217
19Tr	82.00	57.40	60.60	4.6	4.5715
20Tr	86.00	60.20	53.80	4.3	4.4806
21Tr	90.00	63.00	47.00	4.4	4.4154
22Tr	94.00	65.80	40.20	4.3	4.3666
23T	98.00	68.60	33.40	4.4	4.3291
24Tr	54.00	37.80	108.20	5.8	5.6844
25T	54.00	37.8	108.2	5.7	5.6844
26V	54.00	37.8	108.2	5.6	5.6844
27Tr	54.00	37.8	108.2	5.5	5.6844
28Tr	54.00	37.8	108.2	5.9	5.6844
29T	54.00	37.8	108.2	5.5	5.6844
30Tr	54.00	37.8	108.2	5.6	5.6844

---

Quantification of 3-DMC

The amount of colchicine derivatives extracted from the fermentation broth (using a particular TSS or SS) was determined using high-performance liquid chromatography (HPLC; Shimadzu, Model LC20, Class VP) having the following parameters and specifications: a C-18 column, column oven temperature 35 °C, λ_max 245 nm, 2 mL min⁻¹ flow rate, and a UV-Vis detector. Isocratic elution was performed with a methanol–water mobile-phase system (6 : 4, v/v). The sample volume injected for analysis was kept at 20 mL per injection. The efficiency of the employed method was verified through reproducibility after repeated injection loads of the same sample. The linearity of the results was checked by making 10 different dilutions ranging from 0.005 to 0.05 g mL⁻¹ and loading in triplicate onto the HPLC column.

Artificial neural network design and training

A feed forward backprop artificial neural network (ANN) was setup and tested for its efficiency to represent a relationship between the test solvent systems (rows, Table 2) and 3-DMC recovery (Table 2). Numbers of hidden layers were optimized during several attempts to represent targets from the training data. The rows of Table 2 (n = 30) were divided into three sets of input data: training (n = 20), validation (n = 5) and testing (n = 5) (refer to Table 1 for exact training, validation and test rows). The ANN was set up on a MATLAB® platform with the following conditions: training algorithm: Levenberg–Marquardt; total allowed epochs: 1000; transfer function: tansig and purelin; the rest of the parameters were kept as default. Numbers of hidden layers were optimized during multiple attempts.

Genetic algorithm based optimization

The GA was implemented on a MATLAB® platform to perform an exhaustive search of a constrained extraction solvent system composition resulting in maximum 3-DMC recovery. Constraints were maintained using equality matrices and lower/upper bounds of the population in each generation. The GA was applied with equality matrices to determine the optimum extraction solvent system composition. The validated and tested ANN served as the selection function (@sann). The GA optimization was performed with a population size of 250 with the following parameters: crossover fraction, 1; elite count, 2; migration direction, forward; migration interval, 20; migration fraction, 0.2; generations 150; stall gen limit, 50; creation Fcn, @gacreationuniform; fitness scaling Fcn, rank wise; gaproblem.fitnessfcn:@aggr used in the MATLAB® implementation of the GA for optimization. Since the GA implementation of MATLAB® is designed to minimize, the output of the selection function (@sann) was made negative by multiplying by −1. Refer to ESI:† Annexure I for the detailed GA optionset.

Results and discussion

Complete removal or minimization of the use of chloroform in pharmaceutical products is advisable according to the International Conference on Harmonization of Technical Requirements for Registration of Pharmaceuticals for Human Use [Q3C Impurities, Residual Solvents – Q3C Tables and List, Revision 2 (2012) (http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM073395.pdf)] industrial guidelines,³⁹ which makes recommendations as to what amounts of residual solvents are considered safe in pharmaceutical preparations for the U.S. Department of Health and Human Services, Food and Drug Administration (the concentration limit of chloroform in pharmaceutical products is 600 ppm but still not advisable to be used). As chloroform is a Class 2 solvent, its use is restricted or very limited. Contrastingly, solvents in Class 3 (acetone, ethanol etc.) may be considered as less toxic and of lesser risk to human health. Class 3 contains no solvent identified as a human health hazard at levels normally accepted in pharmaceutical preparations.

Since chloroform was found to be the most suitable solvent for the efficient extraction of 3-DMC,³⁸ the challenge was to design a solvent system having a high extraction efficiency with the least toxicity and set with energy efficient down-stream processing (DSP). The proposed solvent system should have the following properties.

(1) The extraction solvent system should be equally or more efficient than pure chloroform.

(2) The DSP of the extraction solvent system should be able to completely remove any organic solvent classified as Class 2 (by ICH) present in the solvent system.

(3) The DSP of the extraction solvent system should be energy efficient.

Chloroform containing solvent systems appear to be superior extractants as compared to others.³⁸ Hence, it was decided to design a chloroform containing solvent system which must qualify the above mentioned three criteria. Several binary and ternary solvent systems were designed and tested on the above three criteria and for their ability to extract 3-DMC from a fermentation broth. Only ICH Class 3 solvents (solvents with less toxic potential for humans as per the ICH classification) were used in combination with chloroform to form either binary or ternary complexes (Table 1). The extraction efficiency of each SS mentioned in Table 1 was revealed by its 3-DMC recovery. None of the SSs appear to be better than pure chloroform. SS no. 1 and 7 appeared to be equally efficient as pure chloroform. Az BP represented in Table 1 is the boiling point at which the azeotrope boils off simultaneously. Binary solvent system number 1 (SS no. 1; chloroform and acetone) and ternary SS no. 7 (chloroform, acetone and ethanol) appeared to be equally good solvent systems for the extraction of 3-DMC. However, SS no. 1 forms a maximum boiling azeotrope (Az BP: 64.7 °C). Combinations of organic solvents (system components) which may result in a maximum boiling azeotrope (MBA) are not preferred for pure single solvent recovery.¹ Even if the initial ratio of the solvent system components was non azeotropic, the ratio of components in the residual volume (liquid phase while heat concentrating/boiling a MBA during DSP) will tend towards the maximum boiling azeotropic ratio during the boiling/distillation process,¹ thereby making it impossible to separate one component from another. Therefore, once the ratio of chloroform and acetone in the residual volume for SS no. 1 reaches the maximum boiling azeotropic ratio, chloroform and acetone will be boiling off together (equal proportions in the liquid and gas phases), it will become impossible to obtain chloroform free acetone containing extracted 3-DMC. Whereas, SS no. 7 forms a saddle azeotrope, i.e., the Az BP lies somewhere below the boiling point of any one component [in this case, ethanol].¹ SS no. 7 does contain chloroform and acetone but formation of a maximum boiling azeotrope (in the residual volume due to heat concentrating/boiling during DSP) was avoided by the addition of ethanol. Here ethanol acts as an entrainer in excess. The entrainer (separating agent: ethanol) upsets any previously established azeotrope (in this case chloroform : acetone) by disturbing molecular interaction and hence affecting their relative volatility.¹ When this solvent system, carrying an excess of entrainer (ethanol), reaches the Az BP during DSP (concentrating 3-DMC by evaporating the solvent at higher temperatures) all the chloroform and acetone along with a little ethanol (as per the ternary azeotropic volumetric ratio) will distil out safely leaving an excess of pure ethanol (ICH Class 3 solvent) as the residue. Since the residue contains only pure ethanol (≥99.5%) one can safely maintain 3-DMC at the desired concentration levels (in ethanol). This is similar to azeotropic distillation used for dehydrating ethanol/water mixtures using entrainers, such as, toluene or cyclohexane etc.^2,40 The lower the Az BP is, the less the heating requirement is to remove the chloroform completely from the solvent system thereby making DSP of 3-DMC more energy efficient. SS no. 7 had quite a low Az BP (63.2 °C) which characterized it to be fairly energy efficient. Further low heating during DSP shields 3-DMC from substantial denaturation during recovery.

On account of the above discussed properties and also of 3-DMC recovery being a non-linear function of chloroform volume (Fig. 2), SS no. 7 was selected for constrained azeotropic optimization via an artificial neural network and genetic algorithm resulting in maximizing the recovery of 3-DMC from a filtered fermentation broth. The design of the experiments is represented in Table 2. Each row of Table 2 forms a ternary azeotrope wherein the chloroform and acetone are in exact azeotropic ratios and ethanol is in excess. The extraction potential for each TSS is present in column 5 of Table 2. The predictions for the tested and validated ANNs are shown in column 6 of Table 2 (Fig. 3A). The ANN achieved an acceptable degree of efficiency (mean error 0.0048) after 10 iterations in 4 epochs only (Fig. 3B). TSSs subjected to successful ANN training, validation and testing have been marked as Tr, V and T, respectively in Table 2. The regression coefficients for training, validation and testing were fairly close to 1 (Fig. 3C). Also, the overall regression of the input test solvent system composition for the network targets, i.e., 3-DMC recovery was 0.99753. This classifies that the ANN was fairly efficient for representing the relationship between the composition of the extraction solvent system components and 3-DMC recovery. The number of hidden layers required to achieve said efficiency of the ANN was optimized during multiple training and testing attempts. Fig. 3D represents the final network in a MATLAB® notation. Kindly refer to ESI:† Annexure I for network layer weights and biases.

Fig. 2 Variation in 3-DMC recovery as a function of chloroform volume in the solvent system.

Fig. 3 (A) Comparison of the observed and ANN predicted 3-DMC recovery: the ANN achieved an acceptable degree of efficiency (mean error 0.0048) after 10 iterations in 4 epochs only (Fig. 4). TSSs subjected to successful ANN training, validation and testing are marked Tr, V and T, respectively in Table 2. (B) ANN performance during training, validation and testing.
(C) ANN regression efficiency. (D) ANN structure (W: layer weights; b: layer bias).

If the optimum volumetric azeotropic combination of chloroform, acetone and ethanol is known one can easily calculate the critical chloroform free residual volume (V_Cr) from the volumetric azeotropic ratio of chloroform, acetone and ethanol in an azeotropic ternary complex. For safety and indemnity achieving the residual volume ≤ 70% of the calculated V_Cr assures ∼100% chloroform free recovery (V_cfr: chloroform free residual volume) of the metabolite, in this case, i.e., 3-DMC.

V_Cr = 200 − chloroform optimum volume (mL) − acetone optimum volume (mL) − volume of ethanol as per ternary azeotropic volumetric ratio (V_E) (1)

V_E = 0.299 × chloroform optimum volume (mL) (2)

V_cfr = 0.7V_Cr (3)

A Genetic algorithm (GA) was employed to perform an exhaustive global search of solvent system composition to achieve maximum 3-DMC recovery. Only azeotropic combinations of chloroform, acetone and ethanol (allowed in excess) were to be screened for maximum and non-toxic recovery of 3-DMC. A constrain dependent population generation function was selected that generated only azeotropic populations of a given size (in this case 250) where all the individuals satisfied the constraints provided by equality matrices (ESI:† Annexure I).

The population generation of the GA must be constrained so as to meet the following criteria for all the individuals generated within a particular generation.

(1) Only azeotropic ratios of chloroform : acetone : ethanol (in excess) are permitted. This constraint was applied through the equality matrices.

(2) The total volume of TSS was always kept equal to 200 mL. This constrain was applied through the equality matrices.

(3) The components of the solvent system were kept within the respective range.

Component	Lower bound (mL)	Higher bound (mL)
Chloroform	10	98
Acetone	7	68.6
Ethanol	33.4	183.00

This was applied by supplying population bounds in the GA optionset.

The major objective of this optimization was not only to determine the optimum azeotropic combination of the TSS components (chloroform, acetone and excess of ethanol) resulting in maximum 3-DMC recovery but also to minimize the amount of chloroform in the azeotrope without any concession on 3-DMC recovery. It is possible that different azeotropic combinations of chloroform with acetone and an excess of ethanol may result in similar (maybe maximum/higher) recovery of 3-DMC and the GA may return that combination (maybe the azeotrope with high chloroform) as optimum. This problem was successfully handled by introducing a new selection function (@snl) which included the validated and tested ANN along with an extra code for minimizing chloroform. This extra code remembers the volume of chloroform in the azeotrope for the maximum 3-DMC recovery encountered during the GA optimization. If during the GA optimization an azeotropic combination was found resulting in an equally good 3-DMC recovery this code will compare the corresponding volume of chloroform with the previously stored volume of chloroform. If the volume of chloroform in the recently encountered azeotropic combination is less than the previously stored value this extra code will accept this azeotropic combination by returning a more favorable value to the GA and will also store this volume of chloroform in its memory and let the GA move freely as per its evolutionary direction while selecting this as an elite individual. However if the volume of chloroform in the recently encountered azeotropic combination is more than the previously stored value, the said extra code will reject this azeotropic combination by returning a non-favorable value to the GA and hence keeping the previously elite individual (lower chloroform volume azeotrope) as such. This ensures that the azeotrope with the lowest chloroform was selected without compromising on 3-DMC recovery.

Recovery of 3-DMC is not a linear function of chloroform volume as is evident from the cubic polynomial fitting (Fig. 2) of the observed values from Table 2 (column 5 vs. 2). An exhaustive search of the optimum azeotrope (ethanol in excess) was performed using the GA, where the tested and validated network along with an extra code (refer to Discussion section) acted as the selection function, mapped the given azeotrope to its 3-DMC recovery. The results of the GA optimization as per the GA optionset (ESI:† Annexure I) are shown in Fig. 4. The GA predicted the best fit value or the in silico maximum of 3-DMC to be 6.4297 g L⁻¹ Fig. 4 (negative sign explained in material and methods) at an optimum chloroform: 63.055 mL, acetone: 44.138 and ethanol: 92.807 (total volume = 200 mL). This optimum combination of solvent system components was tested experimentally in triplicate, the 3-DMC recovery was found to be within ±2.5% of the GA predicted maximum. V_Cr and V_cfr calculated as per eqn (1) and (3) for GA optimum are ∼74 and ∼52 mL, respectively.

Fig. 4 GA predicted optimum (above: GA progress towards maximum 3-DMC recovery; below: best/optimum combination of (1) chloroform; (2) acetone and (3) ethanol for maximum 3-DMC recovery).

The residual volume V_cfr was subjected to purge and trap GC/MS analysis (TSQ 8000 Evo Triple Quadrupole GC-MS/MS; Thermofisher Scientific Inc.) for the determination of chloroform. A negligible level of chloroform was observed in the chloroform free residual volume V_cfr (well below the prescribed ICH threshold). Reducing the residual volume to V_cfr (52 mL, during DSP) assures almost 100% chloroform free recovery of 3-DMC in ethanol. Almost a ∼33% increase in 3-DMC recovery was observed (as compared to un-optimized chloroform). The improved recovery may be attributed to the proposed novel constrained azeotropic optimization procedure involving the ANN along with evolutionary optimization (i.e., GA). The proposed methodology was also tested successfully for the energy efficient optimized extraction of betulinic acid from an optimized fermentation broth. Betulinic acid was finally extracted into methyl acetate (Class 3 solvent) by azeotropically distilling out n-hexane and methanol from a minimum boiling ternary azeotrope (refer ESI:† Annexure II).

Conclusions

We propose an efficient method to obtain the desired metabolite(s) in a nontoxic solvent system, where the most suitable solvent for the extraction of the desired metabolite(s) belongs to the restricted class. This strategy encompasses an energy efficient method to obtain the global optimum combination of solvent system components resulting in a non-toxic and enhanced recovery of the desired metabolite. In this study, the azeotropic behavior of the binary and ternary solvent systems in amalgamation with artificial learning and evolutionary optimization techniques was employed to distill out all the chloroform after extraction. The proposed strategy explains the method to supply the inputs to an ANN under strict constraints, always satisfying the azeotropic ratio while optimizing the network via a GA. Overall, the study explains how to obtain a global optimum within the (constrained) azeotropic population of a ternary solvent extraction system. This extraction process can be applied to other metabolites having similar extraction patterns.

Conflict of interest

The authors declare no conflict of interest exists.

Acknowledgements

The authors are grateful to Jamia Millia Islamia (New Delhi), MD University (Rohtak, Haryana) India, and the Deanship of Scientific Research, Jazan University, Saudi Arabia, for providing the necessary facilities for this research work.

References

1. R. E. Treybal, Mass-transfer operations, McGraw-Hill, New York, 2nd edn, 1968, vol. 3.
2. S. Young, J. Chem. Soc., Trans., 1902, 81, 707–717.
3. E. Hilmen, Separation of azeotropic mixtures: tools for analysis and studies on batch distillation operation, PhD Thesis, Norwegian University of Science and Technology, Dept. of Chemical Engineering, 2000, http://www.nt.ntnu.no/users/skoge/publications/thesis/2000_hilmen/Thesis_Hilmen.pdf, accessed 03 March 2015.
4. S. Kumar, N. Singh and R. Prasad, Renewable Sustainable Energy Rev., 2010, 14(7), 1830–1844.
5. R. W. Rousseau and R. F. James, Handbook of separation process technology, Wiley-IEEE, 1987, pp. 261–262, ISBN 0-471-89558-X.
6. M. F. Doherty and J. P. Knapp, Distillation, Azeotropic and Extractive, in Kirk-Othmer Encyclopedia of Chemical Technology, John Wiley & Sons, New York, 2004.
7. V. Julka, M. Chiplunkar and L. O’Young, Chem. Eng. Prog., 2009, 105(3), 47–53.
8. W. J. Moore, Physical Chemistry, Prentice-Hall, 3rd edn, 1962, pp. 140–142.
9. A. A. Kiss and D. J.-P. C. Suszwalak, Sep. Purif. Technol., 2012, 86, 70–78.
10. P. Vossenberg, H. H. Beeftink, M. A. Cohen Stuart and J. Tramper, Biocatal. Biotransform., 2013, 31(5), 255–268.
11. L. H. Andrade, P. B. D. Vitta, R. S. Utsunomiya, I. H. Schoenlein-Crusius, A. L. M. Porto and J. V. Comasseto, Appl. Catal., B, 2005, 59(3), 197–203.
12. G. Fregapane, D. B. Sarney and E. N. Vulfson, Biocatal. Biotransform., 1994, 11(1), 9–18.
13. F. Ganske and U. T. Bornscheuer, Org. Lett., 2005, 7(14), 3097–3098.
14. F. L. Ezequiel, L. Hannes and W. B. Dirk, Process Biochem., 2006, 41, 2200–2206.
15. M. A. Haider, K. Pakshirajan, A. Singh and S. Chaudhry, Appl. Biochem. Biotechnol., 2008, 144(3), 225–235.
16. S. Khan, V. Bhakuni, V. Praveen, R. Tewari, C. K. M. Tripathi and V. D. Gupta, Appl. Microbiol. Biotechnol., 2011, 89(1), 99–108.
17. C. S. Rao, T. Sathish, M. Mahalaxmi, G. S. Laxmi, R. S. Rao and R. S. Prakasham, J. Appl. Microbiol., 2007, 104(3), 889–898.
18. V. Singh, M. Khan, S. Khan and C. K. M. Tripathi, Appl. Microbiol. Biotechnol., 2009, 82(2), 379–385.
19. D. R. Baughman and Y. A. Liu, Neural Networks in Bioprocessing and Chemical Engineering, Academic Press, San Diego, 1992.
20. D. E. Goldberg, Genetic algorithms in search, optimization and machine learning, Addison-Wesley Longman Publishing, 1st edn, 1989.
21. F. Baishan, C. Hongwen, X. Xiaolan, W. A. Ning and H. Zongding, Process Biochem., 2003, 38, 979–985.
22. J. J. S. Cheema, N. V. Sankpal, S. S. Tambe and B. D. Kulkarni, Biotechnol. Prog., 2002, 18, 1356–1365.
23. S. Khan, M. Adnan, S. Haque, M. Lohani, M. Khan and C. K. M. Tripathi, Cell Biochem. Biophys., 2014, 68(1), 133–142.
24. Y. Nagata and K. H. Chu, Biotechnol. Lett., 2003, 25(21), 1837–1842.
25. R. Bernhardt, J. Biotechnol., 2006, 124, 128–145.
26. S. Pflug, S. M. Richter and V. B. Urlacher, J. Biotechnol., 2007, 129, 481–488.
27. A. D. Rathbone and N. C. Bruce, Curr. Opin. Microbiol., 2002, 5, 272–281.
28. J. A. M. Bont, Trends Biotechnol., 1998, 16, 493–499.
29. K. K. Dubey, S. Haque, A. Jawed, B. P. Singh and B. K. Behera, Process Biochem., 2010, 45, 1036–1042.
30. C. T. Marshall and J. M. Woodley, Biotechnology, 1995, 13, 1072–1078.
31. O. Boye and A. Brossi, Tropolonic colchicine alkaloids and allocongeners, in The Alkaloids, Chemistry and Pharmacology, ed. A. Brossi and G. Cordell, Academic Press, New York and London, 1992, vol. 41.
32. A. Brossi, J. Med. Chem., 1990, 33, 2311–2319.
33. M. Rosner, H. G. Capraro, A. E. Jacobson, L. Atwell, A. Brossi, M. A. Iorio, T. H. Williams, R. H. Sik and C. F. Chignell, J. Med. Chem., 1981, 24, 257–261.
34. T. Tateishi, P. Soucek, Y. Caraco, F. P. Guengerich and A. J. J. Wood, Biochem. Pharmacol., 1997, 53, 111–116.
35. K. K. Dubey, A. R. Ray and B. K. Behera, Process Biochem., 2008, 43, 251–257.
36. K. K. Dubey, A. Jawed and S. Haque, Process Biochem., 2013, 48(8), 1151–1158.
37. A. Poulev, E. Bombardelli, C. Ponzone and M. H. Zenk, J. Ferment. Bioeng., 1995, 79, 33–38.
38. K. K. Dubey, A. Jawed and S. Haque, Eng. Life Sci., 2011, 11(6), 598–606.
39. Q3C Impurities: Residual Solvents [Q3C Tables and List, Revision 2, 2012], Guidance for Industry, International Conference on Harmonisation – Quality, 1997, http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM073395.pdf, accessed on 03 March 2015.
40. V. Gomis, A. Foant, R. Pedraza and M. D. Saquete, Fluid Phase Equilib., 2005, 235(1), 7–10.

---

Footnotes

† Electronic supplementary information (ESI) available: Annexure I. GA constrain and optionset and network layer weights and biases. Annexure II. Example of constrained azeotropic optimization of extraction system components for safe and efficient recovery of betulinic acid. See DOI: 10.1039/c5ra26608d

‡ Current address: Research and Scientific Studies Unit, College of Nursing & Allied Health Sciences, Jazan University, Jazan-45142, Saudi Arabia.

---