Stochastic modeling and meta-heuristic multivariate optimization of bioprocess conditions for co-valorization of feather and waste frying oil toward prodigiosin production

Abstract Serratia marcescens strain UCCM 00009 produced a mixture of gelatinase and keratinase to facilitate feather degradation but concomitant production of prodigiosin could make waste feather valorization biotechnologically more attractive. This article describes prodigiosin fermentation through co-valorization of waste feather and waste frying peanut oil by S. marcescens UCCM 00009 for anticancer, antioxidant, and esthetic applications. The stochastic conditions for waste feather degradation (WFD), modeled by multi-objective particle swarm-embedded-neural network optimization (ANN-PSO), revealed a gelatinase/keratinase ratio of 1.71 for optimal prodigiosin production and WFD. Luedeking–Piret kinetics revealed a non-exclusive, non-growth-associated prodigiosin yield of 9.66 g/L from the degradation of 88.55% waste feather within 96 h. The polyethylene glycol (PEG) 6000/Na+ citrate aqueous two-phase system-purified serratiopeptidase demonstrated gelatinolytic and keratinolytic activities that were stable for 240 h at 55 °C and pH 9.0. In vitro evaluations revealed that the prodigiosin inhibited methicillin-resistant Staphylococcus aureus at IC50 of 4.95 µg/mL, the plant-pathogen, Sclerotinia sclerotiorum, at IC50 of 2.58 µg/mL, breast carcinoma at IC50 of 0.60 µg/mL and 2,2-diphenyl-1-picryl-hydrazyl hydrate (DPPH) free-radical at IC50 of 96.63 µg/mL). The pigment also demonstrated commendable textile dyeing potential of fiber and cotton fabrics. The technology promises cost-effective prodigiosin development through sustainable waste feather-waste frying oil co-management.


Introduction
On the global scale, the United States is the leading producer of poultry meat; China ranks as the world's largest producer of eggs and Asia accounts for more than 64% of the global poultry product output. [1] In developing countries such as Nigeria, small-scale, family-based poultry farming has played an increasingly pivotal role as safety net for livelihoods by supplying poultry products locally and controlling approximately 80% of poultry farming thereby reducing poverty and boosting the national economy. [2] Poultry farming generates a lot of waste which includes dead birds, bird heads, feathers, offal, visceral, entrails, process water, waste feed, and excreta. Every component of the waste is degradable by microorganisms and could serve as poultry manure supplying essential micronutrients for plant growth. [1] However, the feather component is largely recalcitrant, with a tendency to accumulate in ecosystems, causing environmental problems owing to poor management. [3] The recalcitrance is caused by the intricate cross-linked nature of keratin, which constitutes 90% of the feather. [4] Chemical and heat treatment, particularly burning, are used in traditional waste feather management practices in Nigeria, to break the tight hydrogen, hydrophobic and disulfide cross-links in keratin. However, apart from its environmental unfriendliness and biotechnological unsavoriness, this method frequently results in the loss of desirable amino acids namely tryptophan, methionine, and lysine, while generating non-desirables including lanthionine and lysinalanine. [5] Feathers could also be processed into animal feeds using conventional methods like steam pressure cooking technique. [4] However, this method is unsustainable because it consumes a great deal of energy and wastes an enormous amount of protein.
Microbial degradation of poultry feathers is an environment-friendly biotechnological alternative for waste feather management and revolves around keratin hydrolytic mechanisms like sulfitolysis, proteolysis, and deamination. Peptidases, including keratinases and gelatinases, are two important enzymes secreted by microorganisms to catalyze peptide bond hydrolysis in keratin to release the huge amount of amino acids trapped therein. [6] The product (hydrolysate) could be harnessed in feed preparation, fertilizer development and as nitrogen source in microbial culture media. [7] Essien et al. [8] reported on keratinolytic dermatophytes, especially from the genera Trichophyton, Microsporum, and Candida while Li [4] summarized waste feather-degrading genera and species with their environmental sources, capabilities, and culture conditions. Gram-positive bacteria namely Bacillus and Streptomyces are top feather degraders [3] while a handful of species from the Gram-negative genera Pseudomonas and Serratia are noteworthy. Frequently reported feather-degrading fungal genera include Aspergillus, Fusarium, Cladosporium, and Microsporum. [9] Serratia species are Gram-negative pink-to-red pigmenting enterobacteria with potentials to elaborate peptidases [10] but only Serratia marcescens strain P3 [11] and Serratia sp. HPC 1383 [12] have been reported to degrade waste feather. Exploitation of microorganisms in waste feather valorization would be economically more meaningful if one or more value-added metabolites were synthesized during the process. [13] While several species and/or strains have demonstrated potential for commercial prodigiosin and prodigiosin-like compound production, [14,15] no report exists, to date, about Serratia species or any other strain that produces prodigiosin on waste feather substrate. Prodigiosin has been heralded as a versatile biological metabolite with significant applications in health care by reason of its antimicrobial, anticancer/antitumor activities and in industry, owing to its anti-oxidant and free-radical scavenging potentials. [14,15] Design of experiment (DoE), in conjunction with response surface methodology (RSM) or artificial neural network (ANN), has been harnessed to optimize conditions toward improved microbial metabolite production. [16,17] A comparison between RSM and ANN has frequently favored ANN because of its ability to handle non-linear stochastic relationships such as observed in microbial production processes. [18] ANN is a biologically-inspired technique that simulates how neurons work in human brain, and tries to solve complicated data issues like classification, regression, and pattern recognition. Its efficiency depends on the choice of weights and biases, and several techniques have been employed to optimize them including back-propagation and feed-forward approaches. [19] In recent times, a number of meta-heuristic algorithms especially ant colony optimization, manta-ray foraging optimization and particle swarm optimization (PSO) are employed to efficiently optimize these weights and biases for ANN. [20][21][22] These hybrid swarm intelligence techniques have proven their efficiency in bioprocess optimizations. [23] Here, we report the enhanced production of prodigiosin by a locally-sourced S. marcescens strain UCCM 00009 during co-valorization of waste feather and waste frying peanut oil using particle swarm-embedded-ANN (ANN-PSO) approach. Evaluation of apparent fermentation kinetics, serratiopeptidase characteristics and prodigiosin applications in health care, food and textile industries are also reported. This is the first report about prodigiosin production on waste feather substrate.

Materials and methods
Reactivation, quality-checking, and phylogeny of producing organism Authors found the producing organism growing as red spots on chicken feather soiled with grease at a poultry waste dump site in Akpabuyo, Nigeria located along Latitude 4 51 0 31 00 N and Longitude 8 27 0 10 00 E. The organism was isolated in pure culture and identified by morphological and biochemical characterizations as Serratia sp. strain SLO3. Later, partial-gene sequencing of 16S rRNA identified the bacterium as a strain of S. marcescens. The bacterium was deposited in University of Calabar Collection of Microorganisms (UCCM) under the code name S. marcescens UCCM 00009 (www.wfcc.info/ccinfo/collection/by_id/ 652.). In the course of this study, gene sequence of the bacterium was retrieved from UCCM and compared, using basic local alignment search tool, with available sequences in the gene bank. Generated sequences were employed to develop a phylogenetic tree in MEGA X software using neighbor-joining method. The pure culture was reactivated in caprillate thallous agar for confirmation of prodigiosin production at 37 C.
Fermentative production of prodigiosin using whole feather as substrate Preparation of waste feather substrate Waste feather was obtained from Okonson Poultry Limited, Akpabuyo, Nigeria along Latitude 4 51 0 11 00 N and Longitude 8 25 0 3 00 E. A portion of feather was washed with sterile distilled water and anionic detergent (TritonX-100), dried at 60 C for 18 h and lipids removed with 1:1 chloroform/ methanol mixture (v/v) under ultra-sound for 30 min. [24] A second portion was washed as above but lipids left intact. Waste feather from both treatments (lipidated and delipidated, respectively) were dried at 60 C overnight, weighed and used for fermentation studies. [25] The better of the two, selected in terms of biomass accumulation measured by the dry cell weight technique [26] was upheld throughout the study.
After fermentation for different time durations (1 À 5 w), amount of feather degraded was determined by gravimetric method. [26] Difference in feather weight at fermentation start and that at fermentation end was expressed as per cent degradation. The red pigment was quantified from cell pellet after washing twice in ultra-pure water and re-suspending in buffer (20 mM Na 2 HPO 4 , pH 7.0, 4 C). The suspension was placed on ice in the same buffer and sonicated at 30 W for 1 min. This was repeated three times and centrifugation of cell suspension performed at 27,190 Â g for 5 min. The supernatant was treated with 2 mL of acidified methanol at pH 3.0 in a 5 mL glass bottle and diluted 1000-fold, sealed in a bottle and stored at 28 ± 2 C for 8 h to attain full dissolution. The sample was then centrifuged at 3214 Â g for 10 min and absorbance measured using UV-Visible Spectrophotometer (DR6000, HACH, Loveland, CO) at 535 nm using acidified methanol as blank. [27] Standard curve was prepared with 95% pure prodigiosin (BioVision Inc, Natick, MA) for determination of prodigiosin concentration.

Regular two-level factorial design
Sixteen input variables including waste feather (1.5 and 3.5% w/v), lipidic carbon (0.5 and 2.5% v/v), nitrogen source (1.5 and 3.5 g/L), inoculum volume (3 and 7% v/v), temperature (36.5 and 40.5 C), pH (6.5 and 8.5), agitation (100 and 200 rpm), Na þ , K þ , Mg 2þ , Ca 2þ , Zn 2þ , Mn 2þ , Fe 2þ , Co 2þ , and Ni 2þ were screened by regular 2-level factorial design (2-LFD) in Design Expert 12 (Stat Ease, Minneapolis, MN) at low and high levels (À1, þ1). The nutritive salts of Na þ , K þ , Mg 2þ , and Ca 2þ were screened at 5 and 15 mM while the biologically-relevant heavy metal cations (Zn 2þ , Mn 2þ , Fe 2þ , Co 2þ , and Ni 2þ ) were screened as 3 and 7 mM. Prodigiosin concentration, determined as previously described (see "One-variable-at-a-time (OVAT) selection of fermentation variables" section) served as outcome variable (Y). The design comprised 21 experimental runs including 5 center points to determine curvature or lack of it. Predictors that had significant main effects at p< 0.05 were included in the first-order model below: whereŷ ¼ predicted concentration of prodigiosin, b 0 ¼ intercept in the restricted region of predictor variables x 1 , x 2 , … … ., k and b i the coefficients of the model predictor terms.
Optimization of waste feather degradation and prodigiosin production Design of experiment using central composite rotatable design Levels of variables with significant main effects on prodigiosin production obtained from "Regular two-level factorial design" section served as center points in a 2 5-1 -half-fractional factorial of a central composite rotatable design (CCRD) of a surface methodology. Each variable was assigned five levels and included 6 center points for replication. Thirty-two CCRD experimental runs were made to obtain raw data for ANN-PSO. Fermentation of waste whole feather was investigated by submerged batch fermentation in 250 mL Erlenmeyer flask containing minimal medium (composition as in "One-variable-at-a-time (OVAT) selection of fermentation variables" section). Flasks were incubated at selected temperature for a selected time period on rotary shaker (300 rpm). Post incubation, cell-free fermentation broth was obtained by centrifuging flask content at 7871 Â g for 10 min and sterilizing by filtration (0.22 mm). Total protein (TP) and biomass concentration (BMC) were quantified according to Bradford [28] and Ekpenyong et al. [26] , respectively. Keratinase and gelatinase activities were determined as described by Mazotto et al. [6] Prodigiosin concentration and WFD were determined as described in "One-variable-at-a-time (OVAT) selection of fermentation variables" section. Data obtained for the six response variables were analyzed by multiple regression and models built by the method of least squares. Response variables were coded as BMC (Y 1 -g/L), TP (Y 2 -mg), gelatinase activity (Y 3 -U/mL), keratinase activity (Y 4 -U/mL), prodigiosin concentration (Y 5 -g/L), and WFD (Y 6 -%). A general second-order polynomial function used to analyze CCRD data is given as Eq. (2): where y indicates Y 1-6 , coefficient of the constant term by b 0 , linear term by b i , quadratic term by b ii and interaction term by b ij of k factors, E ¼ error term. Model performances were evaluated with adjusted r 2 and lack-of-fit p-values.

ANN-PSO multi-objective optimization
The feed-forward multi-layer approach was the selected topology for network training. It operates on the principle of continual learning, handling problems by assigning weights to input data for each neuron and linking with the output using a transfer function. A bias is normally added to curb the problem of continuous non-linearity. Weight and bias values for each layer were first optimized by PSO; a meta-heuristic algorithm inspired by flocking birds, schooling fish and the general swarm intelligence theory before using the optimal solutions to train neural network. [29] The PSO assigns "fitness" values to each particle using functions targeted at maximizing the solution and each particle is recognized by their local and global best positions (x) and velocities (v) which are synonymous with weight and bias as follows: , v i2 , ::::: where D is the number of dimensions.
Equations (5) and (6) were used to optimize velocity and position of particles at each iteration, respectively: where w ¼ coefficient of inertia, R 1 and R 2 values lie between 0 and 1 and are randomly selected, c 1 ¼ individual acceleration, c 2 ¼ social acceleration, P i ¼ individual best position, and P g ¼ swarm or global best position. The flow chart for step-by-step implementation of the ANN-PSO hybrid function in Matlab R2014a (MathWorks Inc., Natick, MA) is presented as Figure 1a. Network data for training, validation, and testing were normalized between À1 and þ1 as given in Eq. (7): The relative importance of each significant variable on process responses (I) was determined by calculating the connecting weights and their accompanying bias using Eq. (8): where I is the response variable and number of input and hidden neurons are, respectively, represented by N i and N h , respectively; W indicated weights while the superscripts i, h, o, respectively, denote input, hidden and output layers and their respective neurons by subscripts k, m and n, respectively. The performance of the network was evaluated by the coefficient of determination, r 2 and mean squared error (mse) using Eqs. (9) and (10): where y and ŷ are the actual and predicted values respectively, and n the number of samples.

Validation experiment
The optimized levels of variables obtained from the swarm intelligence optimization were employed for WFD and prodigiosin production in real-time triplicate experiments. Results obtained for real-life situation were compared with those from the optimization experiments. A difference of less than 5% was sufficient to accept the suggested optimized levels as true and adequate for maximum degradation of waste poultry feather.
Bioreactor kinetics of waste feather degradation and prodigiosin production A 5-L bioreactor with 3.5 L reaction volume was used for WFD. The fermentation medium was composed of significant variables at their optimized levels as adopted in "Validation experiment" section. Other medium components that did not make significant contributions to the valorization process were added as (g/L) MgSO 4 [30] Aliquots of 25 mL were withdrawn at 12 h intervals and analyzed for biosurfactant concentration (BSC) [26] , lipase activity, [31] acyl homoserine lactone (AHL), [32] and residual waste frying oil (RWFO). [33] Determinations of BSC, AHL, and RWFO became necessary to understand the dynamics of the time-related changes that permitted pigmentation of the bacterium during the co-valorization process. TP, gelatinase activity (GNase), keratinase activity (KNase), BMC, WFD, and prodigiosin, PDG concentration were determined as in "Design of experiment using central composite rotatable design" section. Triplicate data were subjected to multivariate analysis of variance (MANOVA), Pearson's bivariate correlation analysis, and principal component analysis (PCA) in SPSS version 20.0 (IBM, Armonk, NY) as described in Ogarekpe et al. [34] The logistic model was employed to fit BMC data to determine maximum specific growth rate, m max while all other response data were fitted to Luedeking-Piret model to determine growth or non-growth-associated kinetics. The logistic model is presented as Eq. (11) while Luedeking-Piret model expressions are as summarized in Eqs. (12) and (13).
where dP/dt and dX/dt are rates of product and biomass formation, respectively; dS/dt is substrate consumption rate; a and c are the growth-associated rate constants; b and k represent the non-growth-associated (stationary phase) rate constants frequently referred to as Luedeking-Piret constants.
Protein purification and molecular weight determination of serratiopeptidase Protein was extracted and purified from a portion of sterile fermentation broth by polymer/salt aqueous two-phase system (ATPS). The two-phase system comprised polyethylene glycol (PEG) of molecular weight 6000 and Na þ citrate salt. [30] A second portion was first subjected to 70% ammonium sulfate precipitation, before ATPS purification and finally molecular exclusion chromatography. Confirmation of protein purity was performed by sodium dodecyl sulfatepolyacrylamide gel electrophoresis (SDS-PAGE) according to the method of Laemmli. [35]  Purification and structural characterization of the red pigment by GC-MS Prodigiosin, obtained as described in "Regular two-level factorial design" section, was re-suspended in methanol and then by means of silica gel and hexane as stationary and mobile phases, respectively, pigment was subjected to column chromatography. The purified product was dried at 45 C and stored in vials each containing 500 mg. Molecular mass and structure of the purified pigment were determined by gas chromatography-mass spectrometry (7000 D Triple Quadrupole GC-MS system, Agilent, CA) in 5 mL of acidified methanol solution. [27] Commercial prodigiosin (see "One-variable-at-a-time (OVAT) selection of fermentation variables" section) was used as standard. PEG served as polar stationary phase and provided separation of prodigiosin with minimal distortion of the chromatographic peaks. The carrier gas was high-purity (99.9999%) helium (NIIKM, Moscow, Russia) with a flow rate of 1.2 mL min À1 . The split (5:1) injection mode had a sample volume of 1 lL. Column thermostat was programmed to hold from 3 min at 40 C to 5 min at 240 C at the 10 C min À1 ramp. The total duration of the analysis was 28 min. The temperatures of the ion source and transfer line were 230 and 240 C, respectively. Detection was performed using electron ionization (70 eV) in SIM and MRM modes. [36] Substrate specificity, activity optima and stability, and inhibitor stability characterizations of serratiopeptidase Substrate specificity of serratiopeptidase was tested by measuring enzyme activity against azocasein, azokeratin, gelatin, and bovine serum albumin. [10,24] Temperature (30 À 65 C) and pH (5 À 11) optima for serratiopeptidase activity and stability were determined according to the procedure by Asitok et al. [30] Purified enzyme was held at specific condition for 30 min before conducting keratinase, gelatinase and caseinase assays. For stability evaluations, purified enzyme was held at optimum temperature or pH for 30-360 min before peptidase assay using azokeratin, gelatin, and azocasein as substrates.
In vitro cytotoxicity assay Different human carcinomas including breast (MCF-7), promyeloid lymphoma (HL-60), lung (HTT-116), and cervical epithelium (HeLa) were treated with prodigiosin at concentrations ranging from 0 to 1175 mg/mL using the 3-(4,5dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT) assay method. Microtiter plates containing 96 wells were seeded with cancer cells at a concentration of 10 5 cells per well and cytotoxic assay conducted as described in Ekpenyong et al. [16] Selective toxicity of the pigment was evaluated against a non-tumor human embryonic cell line, HEK 283 T. Positive control (100% viability) and negative control (0% viability) wells were set up as reaction mixtures without and with the highest prodigiosin concentration, respectively.

Evaluation of industrial applications of prodigiosin
Anti-oxidation potential Free-radical scavenging potential of prodigiosin was investigated according to Mihooliya et al. [38] using L-ascorbic acid (Sigma-Aldrich, St. Louis, MO) as standard. The change in color of 2,2-diphenyl-1-picryl-hydrazyl-hydrate (DPPH) from deep violet to pale yellow was monitored spectrophotometrically at a wavelength of 517 nm for 3 h. The absorbance reduction was calculated as described by Prior et al. [39] Fabric dyeing for textile industry A modification of the procedure of Gulani et al. [40] was used. The fabric materials tested were silk, fiber, and cotton. Briefly, 2 cm 2 of each fabric was immersed in 5 mL 1% (w/v) of acidified methanol solution of prodigiosin in 50 mL beakers and incubated at room temperature (28 ± 2 C) for 48 h. Thereafter, each prodigiosin-treated fabric was reduced to seven small sizes and wash-treated as follows: acid (0.5 M HCl), alkali (0.5 M NaOH), cold water (4 C), hot water (75 C), cold water (4 C) þ detergent (Ariel), hot waterþ detergent, and detergent alone. Wash-treated fabrics were incubated in different beakers for 1 h and examined for pigment retention after sun-drying and ironing.

Results and discussion
The producing bacterium Pure colonies of the producing bacterium on caprylate thallous agar and its phylogenetic tree drawn from 16S rRNA partial-gene sequence are presented as Figure S1a and b, respectively (Supplementary data). The query sequence had 100% homology with S. marcescens strain NPK2 1 20; accession number MN691630.1 in the second clade of the tree.

OVAT screening
As expected, the bacterium accumulated significantly higher biomass on lipidated feather than on de-lipidated substrate ( Figure S2a, Supplementary data). This is corroborated by researches that incorporate lipidic substrates like glycerol and fatty acid esters as co-substrates for prodigiosin production and feather degradation [4] by Serratia species.
Results of OVAT screening are presented in Figure S2b-g (Supplementary data). Highest WFD and prodigiosin production occurred at 45 C and pH 9. Duration of incubation was selected as 96 h when waste frying peanut oil (WFPO) served as co-substrate. The WFPO was characterized to contain (% w/w) 1.86 steric acid, 14.38 volatile fractions, 36.23 oleic acid, 53.25 linoleic acid, 6.87 palmitic acid, saponification value of 73.6, and density of 277.3 kg/m 3 at 30 C. A previous study by Dos Santos et al. [23] reported enhancement of prodigiosin production (119.8 g/kg) on wheat bran by waste soybean oil. Wei and Chen [41] also reported improved prodigiosin-like compound production by oil supplementation during S. marcescens strain SMdeltaR fermentation. It appears that prodigiosin-producing genes in some strains of S. marcenscens may be activated by free fatty acids or glycerol, both of which have been reported to enhance prodigioin production. [4] Highest prodigiosin concentration of 449.6 mg/L was observed when peptone served as nitrogen source, and peptone glycerol medium is frequently relied on as early stage prodigiosin screening medium. [40] An inoculum size lying between 10 8 and 10 10 cfu/mL of bacterium was reported as optimum for WFD toward maximum prodigiosin production. At the end of OVAT screening, a gelatinase/keratinase ratio of 7.10 was obtained from a gelatinase activity of 83.25 ± 6.83 and 11.73 ± 2.17 U of keratinase activity.

2-LFD screening for significant variables
The ANOVA table for prodigiosin regression model revealed that only five variables made significant contributions to prodigiosin production (Table S1, Supplementary data). Waste frying peanut oil made the highest contribution of 22.01% to the model followed by peptone (18.48%), zinc (13.98%), waste feather (8.46%), and temperature (7.57%) The first-order regression model with significant variables is given as Eq. (14) The model was significant at an adjusted r 2 of 0.7503 and significant curvature of p ¼ 0.0004 < 0.05 but the lack-of-fit test was not significant at p ¼ 0.2900 > 0.05 suggesting adequacy of the first-order model to explain data variations in the region of operation. The significant curvature in the model indicated that variable levels from the 2-LFD experiment were already located around optimal region suggesting that CCRD could be conducted without regard to path of steepest ascent experimentation.

CCRD for ANN-PSO modeling and optimization
Data obtained from 32 CCRD experiments were trained with PSO to optimize weights and biases for ANN predictions. The multilayer feed-forward ANN topology and architecture of the bioprocess was 5-10-10-6, with two hidden layers, each with 10 neurons to convey 5 input items to the output layer of six neurons (Figure 1b and Figure S3, Supplementary data). The design matrix, results of the CCRD experiments and ANN-PSO predictions obtained are presented in Table 1 and reveal that highest ANN-PSO predicted prodigiosin concentration of 4219.50 mg/L from degradation of 88.09% of waste feather occurred in run 27. At that point, the valorization process set waste feather to 4.5% (w/v), peptone 5 g/L, waste frying peanut oil 1.5% (v/v), temperature 42.5 C and Zn 2þ to 5 mM. The selection of zinc ions as the only significant metal ion indicates a requirement for the metal for enzyme activity or synthesis. The gelatinase/keratinase ratio was reduced from 7.10 to 1.73 (some 4.1-fold improvement from OVAT screening) indicating a requirement to narrow the activity gap between the component peptidases involved in keratin degradation. The MO-ANN-PSO model r 2 is given in Figure 1c as 0.9712 with mse calculated as 0.2286 which suggests that 97.12% of data variations during the valorization process could be explained by the ANN-PSO model, and with minimal error.
Biologically-inspired metaheuristic optimization algorithms have been reported to give superior results in terms of prediction of conditions targeted at cost optimization, enhanced productivity, and quality which is the major concern of bio-industries today. These results suggest that an optimum balance of gelatinase and keratinase activities of the serratiopeptidase is required for efficient degradation of waste feather as keratinase alone may not achieve significant degradation that would guarantee waste feather management. [4,42] Mazotto et al. [6] reported that higher keratinase (319 U) activity than gelatinase (200 U) in Bacillus species was required to degrade 75% of waste feather. The ANN-PSO model revealed that serratiopeptidase with too high gelatinase activity against too low keratinase activity would degrade less waste feather than one with higher keratinase. The balance of nutrients and environmental conditions at or near operational levels as suggested by the ANN-PSO model may lead to complete degradation of waste feather in the environment through activity gap closure between component peptidases. These results were validated in real time by a yield of 88.55 ± 13.27% WFD, 4598.65 ± 332.22 mg/L prodigiosin and a gelatinase/keratinase ratio of 1.71 from 174.55 ± 36.28 U/mL gelatinase and 102.02 ± 21.73 U/mL keratinase activities. These conditions were accepted as adequate for maximum bioconversion of waste feather into prodigiosin which could be exploited as a feather waste management option.

Bioreactor kinetics of prodigiosin production in CCRD-ANN-PSO optimized medium
Results of the batch kinetics of prodigiosin production in the 5-L bioreactor are presented in Table 2 and summarize  Table 1. Central composite design matrix with experimental (e) and ANN-PSO predicted (p) responses.
S/N   pBMC (g/L) to WDF (%) are experimental (e) and predicted (p) values for biomass concentration, total protein, gelatinase activity, keratinase activity, prodigiosin concentration, and waste feather degradation potential, respectively.
information required for the development of mathematical models for optimized waste feather management, especially the bacterial growth kinetics. Biomass formation data, modeled by the logistic function, revealed a m max of 1.46 h À1 corresponding to a doubling time of 0.48 h. This high growth rate may be attributed to co-utilization of feather with waste frying peanut oil. The growth-associated synthesis of lipase given by Luedeking-Piret model confirms this theory. Lee et al. [43] reported that fatty acid-degrading enzymes were significantly involved in keratin degradation as well. Additionally, production of biosurfactant, which occurred almost predominantly during primary metabolism facilitated solubilization of WFPO on which the bacterium primarily grew. The almost equal growth and non-growth associated production of gelatinase by the bacterium suggests equal involvement of the peptidase in primary and secondary metabolism. The predominantly non-growth associated release of keratinase suggests the presence of a primary substrate specifically required for cellular metabolism but keratin as alternative substrate. The keratinase of S. marcescens has been linked to the serralysin family of proteins which explains the surface activity observed during the valorization process. [11] This is akin to the keratinases of Bacillus that are linked to the subtilysin family of proteins which are also confirmed surface-active compounds. [3] Interestingly, biosurfactant and prodigiosin production by bacteria are physiological functions positively regulated by the microbial cross-talk phenomenon, quorum-sensing. This is a bacterial cell-cell communication meachanism that allows regulation of specific processes in response to changes in population density using an autoinducer like AHL. Investigation of the kinetics of AHL accumulation revealed growth and non-growth associated productions (a ¼0.025 and b ¼ 13.21) suggesting the involvement of the auto-inducer in directing both responses for primary and secondary metabolism. The little growthassociated pigmentation may be required to enhance lag phase accumulation of ATP to drive cellular metabolism toward biomass accumulation. [44] Therefore, S. marcescens may not accumulate sufficient ATP when grown on feather substrate but co-utilization of a suitable substrate could prompt early biomass accumulation and significant release of auto-inducers sufficient for release of the secondary metabolite, prodigiosin.
For multiple response variable studies such as this, MANOVA is more appropriate than the routine ANOVA to explain data variations and bioprocess dynamics. The Wilk's k test (Table 3) revealed that changes in time were responsible for 96% of variations about the responses at F ¼ 69.93, g 2 ¼ 0.96. Details about which response was most influenced by time variations are presented in the tests of between-subjects effects table (Table S2, Supplementary data). The interrelationships between and among response variables were confirmed by PCA to ascertain similar bioprocess source. [45] To establish suitability of data for PCA, the Kaiser-Meyer-Olkin (KMO) and Bartlett's test of sphericity were conducted and a KMO of 0.849, chi-square approximation, v 2 of 296.33 and a determinant of 0.000231 > 0.00001 were obtained. Ogarekpe et al. [34] had earlier reported suitability of data for PCA following a KMO of 0.827 and a determinant greater than 0.00001 while modeling the location of occurrence of hydraulic jump by PCA. The PCA model in this study revealed that the seven responses evaluated could be extracted into two components, PC1 and PC2 with cumulative variance of 91.32% (Table 4).  NA: not applicable; ND: not determined; AHL: acyl homoserine lactone (mg/L); BMC: biomass concentration (g/L); BSC: biosurfactant concentration (cm 2 ); GNase: gelatinase activity (U/mL); KNase: keratinase activity (U/mL); LPase: (lipase activity (U/mL); PDG: prodigiosin concentration (mg/L); TP: total protein (mg/mL); WFD: waste feather degradation (%); WFO: waste frying oil (g/L); r 2 : coefficient of determination; RMSE: root mean squared error; MAE: mean absolute error.
The responses PDG, AHL, BMC and GNase activity, loaded on PC1, were more significantly and positively related and accounted for 75.1% of the variability observed during the valorization process. Keratinase activity and WFD loading on PC2 accounted for 16.2% of process variability and were also positively related.

Characterization of prodigiosin and serratiopeptidase
Molecular mass and structure of the red pigment determined by GC-MS, identified pigment as prodigiosin; a permeable membrane-bound water-insoluble tripyrrole alkaloid with the chemical name 2-methyl-3-pentyl-6methoxy-prodigiosene ( Figure 2a) and a molecular formula of C 20 H 25 N 3 O. [15] Purification protocols for serratiopeptidase revealed that specific activity increased with successive steps of purification but the most between precipitation and polymer-salt ATPS purification (Table S3, Supplementary data). The selected molecular weight of the PEG was 6000 at a concentration of 19%, 11.5% of K 2 HPO 4 and 13.8% NaCl for hydrophobicity with enzyme partitioning to the top phase rich in PEG. [11,46] The ATPS recovered 88.5% of the enzyme at a purification factor of 17.8. However, with a combination of other steps, final peptidase yield and purification factor were 26% and 123.03%, respectively, suggesting that the one-step purification protocol of ATPS may not be sufficient for total purification of the protein. The SDS-PAGE analysis (Figure 2b) and molecular exclusion chromatography ( Figure S4, Supplementary data) showed that the purified protein had a molecular weight of 58.3 kDa from an elution ratio 0.765. Most peptidases have molecular weights <60 kDa [47] but the closest relative of the study peptidase, in terms of size, would be the 53 kDa serratiopeptidase reported by Bach et al. [11] The study serratiopeptidase had caseinase activity (286 U/mL) the most followed by gelatinase (266 U/mL) and keratinase (156 U/mL) activities (Figure 2c). Only 26.65 U/mL activity was observed for albumin. Mazotto et al. [6] reported higher keratinase (319 U/mL) activity than gelatinase (200 U/mL) in Bacillus species peptidases.  Tukey's multiple comparisons post hoc test for one-way ANOVA of substrate specificity studies revealed that significant differences (r 2 ¼ 0.9973, F ¼ 972.1, p < 0.05) existed among different pairs of substrates. Keratinase and gelatinase activities of the serratiopeptidase had their temperature optima at 55 C but caseinase between 45 and 50 C (Figure 2d). At 55 C, keratinase was stable for 240 min while gelatinase lost 7.52% of its activity at 180 min. Significant (p< 0.05) loss of 19.5% gelatinase activity at 55 C occurred at 240 min (Figure 2e). The peptidases of Bacillus species in Mazotto et al. [6] were stable at 45 C. Optimum pH for keratinase activity was found between 9 and10 but gelatinase had a more specific pH optimum of pH 10 at 55 C while optimum caseinase activity occurred between pH 8 and 10 (Figure 2f). Gelatinase and keratinase activities were stable at pH 10, 55 C for 240 min while 42.27% of caseinase activity was lost over that time (Figure 2g). These results suggest that caseinase activity of the serratiopeptidase may not be required for significant feather degradation into prodigiosin under the prevailing bioprocess conditions.
The keratinase activity of serratiopeptidase was completely inhibited by serine protease inhibitors, PMSF and DPFP, but lost 40% of its activity to EDTA. However, residual gelatinase activities after PMSF, DPFP and EDTA pre-exposure were 56.72%, 61.43%, and 0.53% respectively (Figure 2h). This suggests that S. marcescens UCCM 00009 peptidase is a metallo-cysteine protease much like the 32 kDa surfactant-stable a-keratinase of Bacillus pumilus strain K9 [48] Clinical, food, and textile industry applications of prodigiosin The study prodigiosin demonstrated in-vitro antimicrobial activities against MRSA strain ATCC 33591 at a MIC (defined as pigment concentration required to cause 1 log reduction in bacterial concentration) of 1.84 mg/mL and IC 50 of 4.951 mg/ mL ( Figure 3a). P. aeruginosa ATCC 27853 was also very sensitive at MIC ¼ 1.63 mg/mL (IC 50 ¼2.373 mg/mL). Surprisingly, the prodigiosin from S. marcescens reported by Yip et al. [49] did not inhibit P. aeruginosa strain PA 14 which may be attributed to the test strain, its age and probably the pigment concentration. The red pigment also inhibited the plant pathogenic mold, S. sclerotiorum strain UCCM 00134 at 1.46 mg/mL (IC 50 ¼2.578 mg/mL) suggesting its potential for plant-pathogen control. The prodiginines tested by Habash et al. [14] had little activity against a strain of this pathogen. The For Drows 1-3 indicate cotton, wool and silk fabric, respectively, and columns 1-7 indicate washing treatment after dyeing namely acid, alkali, cold water, hot water, detergent (Ariel) þ cold water; detergent þ hot water, and detergent alone, respectively. study red-pigment showed poor activity against the tested strain of C. neoformans and is therefore not recommended for use against the pathogen.
In Figure 3b, a dose-dependent in-vitro cytotoxicity of the pigment against human breast carcinoma (MCF-7) at an IC 50 of 0.602 mg/mL was observed, followed by 2.830 mg/mL against human myeloid lymphoma (HL-60). Safety of the anti-leukemic agent was demonstrated by selective toxicity against the non-tumor cell line, HEK 283 K against which it showed an IC 50 of 81.19 mg/mL. Previous researchers like Lazic et al. [50] have reported this range of toxicity to different carcinomas. A selectivity index of 134.87 was determined for MCF-7 cell line while 28.69 was calculated for HL-60 cell line.
Results of the antioxidant potential of the pigment are presented in Figure 3c and also show a dose-dependent pattern. At 300 mg/mL, ascorbic acid achieved maximum free-radical scavenging activity (100%). However, study prodigiosin, at a similar concentration, demonstrated 95% of the antioxidant potential of vitamin C. Similar results have been reported by Nguyen et al. [51] where 99% DPPH scavenging was reported. Significantly, the IC 50 of the study prodigiosin (96.63 mg/mL) was a lot lower than most reported.
Contrary to the report by Gulani et al. [40] , tested fabrics reacted differently to dye fastness evaluated by different washing conditions (Figure 3d). However, our results agree with the findings of Metwally et al. [52] because the chemical composition of fabric significantly influences dye-fabric interaction especially hydrogen bonding. This results in differences in color intensities after washing, sun-drying and ironing. The rows in Figure 3d indicate fabric type: cotton (1), wool (2), and silk (3) while columns refer to washing conditions. Row 2 (wool fabric) retained dye the most followed by row 1 (cotton) and row 3 the least. This could have been attributed to the fabric quality (texture) but washing conditions revealed interesting trends. Column 1 (acid washing) showed that wool material was most unstable to acid washing followed by silk, and cotton fabric as most stable. Column 2 indicates alkaline washing and all fabric appeared brighter than column 1 but wool fabric was the brightest. All fabric showed similar brightness when washed with cold water (column 3) but silk lost the most color when washed with hot (75 C) water. More pigment was retained in wool material than silk (column 4). Washing fabric with detergent (Ariel) in column 7 did not cause any significant color loss in all three fabric types. Interestingly, silk fabric presented most stability when washed with just detergent. Columns 5 and 6 did not show any significant loss in brightness when wool and cotton fabrics were washed in detergent-containing cold and hot waters respectively. However, color brightness was stronger in wool material than the rest in both columns and silk lost significant color when washed with detergent and hot water (column 6). It appears that successful dyeing of textile is both a function of dye nature, fabric type and washing conditions and the study prodigiosin could be industrially exploited for stable dyeing of some fabrics.
In conclusion, co-valorization of waste feather and waste frying peanut oil by S. marcescens strain UCCM 00009 was successfully improved through combined DoEs and particle swarm embedded neural network optimization (ANN). The bacterium, at a m max of 1.46 h À1 , produced a 58.3 kDa metallo-cysteine peptidase with 62.95% gelatinase and 37.05% keratinase activities to degrade 88% of waste feather producing 9.66 g/L prodigiosin within 96 h. The pigment demonstrated significant antimicrobial activity against multidrug-resistant bacteria, plant pathogenic fungus and selective toxicity against human carcinomas. Food and textile industry applications as well as sustainable waste feather/waste frying oil co-management are recommended.