1. Introduction
Antibiotics are a significant class of metabolites produced by microorganisms, animals, or plants. The biological activities of antibiotics, such as cytotoxicity, bacteriostatic, antimalarial, and anti-parasitic properties, make them widely used in medicine, agriculture, animal husbandry, the food industry, and other fields [
1]. In recent years, the exploitation of antibiotic resources has gradually shifted from soil to the ocean [
2]. Typically,
Streptomyces parvus is considered as a potential source of biologically active compounds that has been explored widely for drug development [
3]. For example, ZM-1 (identified as holomycin) was isolated from
S. parvus 33 and found to have strong antibacterial activity against plant pathogenic fungi [
4].
S. parvus NEAE-95 produced an anti-neoplastic agent, L-asparaginase, which was used in acute lymphoblastic leukemia treatment [
5]. LYRM03, isolated from
S. parvus HCCB10043, showed higher potent inhibitory activity against aminopeptidase N for cancer therapy than bestatin [
6]. Eumelanin pigment, purified from
S. parvus BSB49, could be utilized for pharmaceutic and cosmetic product development [
7]. Silver nitrate nanomaterials from
S. parvus Al-Dhabi-91 were a suitable active substance for treating infectious disease [
8].
Vicenistatin, a 20-membered macrolactam core with an amino-sugar vicenisamine, was first isolated from marine microorganisms,
S. halstedii HC34, and showed great potential to be used as an antitumor drug. It not only exhibits cytotoxicity against COLO205 and HL-60 cells, but also has antitumor activity toward Co-3 cells [
9]. Recently, a vicenistatin analogue (4′-N-demethyl-vicenistatin), which showed better antitumor activity and reduced cytotoxicity than vicenistatin, was isolated from
S. parvus SCSIO Mla-L010/Δ
vicG (a disruptant of the N-methyltransferase gene). It was characterized as an macrolactam antibiotic and impressive antitumor drug. In comparison to vicenistatin, 4′-N-demethyl-vicenistatin exhibited good antimicrobial activities, including methicillin-resistant
Staphylococcus aureus, methicillin-resistant
Staphylococcus epidermidis,
Micrococcus luteus, and
Bacillus subtilis, with low cytotoxicity [
10]. However, the low product concentration of 4′-N-demethyl-vicenistatin (4 mg/L) was the main limiting factor for subsequent medicinal property evaluation and applications [
11].
The compositions and concentration of a fermentation medium play a crucial role in the growth of microorganisms and the formation of secondary metabolites. Thus, it is essential to identify and optimize the important components (such as carbon sources, nitrogen sources, and inorganic salts) in the medium for efficient 4′-N-demethyl-vicenistatin production. For fermentation medium optimization, the one-factor-at-a-time (OFAT) method is a basic strategy and is frequently used, but it does not consider the interactions among various factors [
12]. In contrast, response surface methodology (RSM) can elucidate the interactions among individual factors. Prior to employing RSM, the Plackett–Burman design (PBD) was often conducted to identify the factors that exert a significant influence on the outcomes [
13,
14]. Apart from the above classical methods, machine learning tools such as the artificial-neural-network-genetic-algorithm (ANN-GA) have been confirmed to possess a better predictive capability, especially for complex and nonlinear processes like biological fermentation [
15,
16]. However, there is no related report concerning the systematic study of the effects of medium compositions, nor the model for optimizing the fermentation medium to enhance 4′-N-demethyl-vicenistatin production.
The aim of this study is to optimize the fermentation medium in order to achieve efficient 4′-N-demethyl-vicenistatin production using S. parvus SCSIO Mla-L010/ΔvicG with the statistical design of experiments. First, the medium compositions were screened by OFAT and PBD. Then, the comparative performance of RSM and ANN-GA for modeling and optimizing the medium compositions was conducted. Finally, the predictive ability of RSM and ANN-GA was experimentally confirmed and achieved a 226% and 283% improvement in 4′-N-demethyl-vicenistatin production, respectively.
3. Discussion
The optimized medium formulation (12 g/L cassava starch, 17 g/L glycerol, 34 g/L seawater salt, 7.5 g/L soybean meal, 7.5 g/L ammonium citrate, 30 mg/L FeSO
4·7H
2O, 2 g/L calcium carbonate) achieved through the implementation of OFAT, CCD, and ANN-GA in this study resulted in a remarkable 38-fold increase in yield compared to the previously reported value of approximately 5 mg/L [
11]. Moreover, 4′-N-demethyl-vicenistatin exhibited excellent antibacterial activity and displayed promising potential for patenting purposes [
11]. Consequently, this enhanced production can serve as a solid foundation for subsequent derivatization experiments and other research and development endeavors.
The focus of this discussion will be on the following two key aspects: the optimization strategy for medium formulation and the selection of research methods.
The first aspect pertains to the optimization of the medium formulation strategy. This study revealed that the composition and proportion of carbon sources (glycerol and cassava starch), in conjunction with seawater salt concentration, exerted a significant influence on secondary metabolite production (4′-N-demethyl-vicenistatin) by
S. parvus SCSIO Mla-L010/Δ
vicG. The
p-values of glycerol, cassava starch, and seawater salt in
Table 1 were all found to be less than 0.05, indicating the significant impact of their concentration changes on the production of 4′-N-demethyl-vicenistatin. In comparison to the formula (15 g/L glycerol, 15 g/L soluble starch, 30 g/L seawater salt) utilized in the AM3 medium, the optimized formula (17 g/L glycerol, 12 g/L cassava starch, 34 g/L seawater salt) predicted by the ANN-GA model resulted in an increased yield, from 0.0502 g/L to 0.1921 g/L (
Table 6). The composition and concentration of the medium can exert a significant influence on microbial fermentation production [
45,
46,
47]. This is particularly applicable to microbial secondary metabolic processes, where the composition and concentration of the carbon sources in the medium can significantly impact the fermentation of secondary metabolites, such as pigment fermentation and antibiotic production [
48,
49,
50]. The choice of carbon source can exert a significant influence on the biosynthesis of specific secondary metabolites, owing to its capacity to inhibit gene expression and repress enzyme activity [
51]. For example, the production of cephamycin C, a β-lactam antibiotic synthesized by
Streptomyces clavuligerus, encounters challenges in the presence of glycerol, due to its suppressive impact on the activity of cephamycin C synthetase and expandase enzymes [
52]. Glucose exerts a suppressive effect on
afsR2 mRNA synthesis, encoding a global regulatory protein responsible for facilitating secondary metabolite biosynthesis in
Streptomyces lividans, thereby resulting in the inhibition of actinorhodin (a polyketide) production [
53]. While glycerol enhances cellular growth and internal ATP levels, it hinders the synthesis of spiramycin. Spiramycin, a potent macrolide antibiotic derived from
Streptomyces ambofaciens, is commonly prescribed for the treatment of toxoplasmosis. The presence of glucose and glycerol adversely impacts the production of spiramycin [
54]. Therefore, in order to exert better control over the concentration and type of carbon source, numerous researchers opt for feeding carbon into the fed-batch fermentation process. For example, the enhanced production of secondary metabolites in
Inonotus obliquus, a traditional medicinal fungus utilized for cancer and other ailments, by 65% is conducted employing a glucose-fed batch dissolved oxygen (DO) control strategy. This approach entails supplementation with 10 g/L glucose upon reaching a residual sugar concentration of 10 g/L while maintaining the DO level at 50% [
55]. The production of alkaline amylase in
B. subtilis 168 mut-16# strain was significantly enhanced to 591.4 U/mL by optimizing the agitation speed and supplementing with hydrolyzed starch during the 10th hour of fermentation [
56]. The
Streptomyces graminearus F3-4 strain was employed in a fed-batch fermentation process, resulting in the production of epsilon-PL reaching a maximum concentration of 13.5 g/L. This remarkable production was achieved by increasing the initial glucose concentration from 50 to 85 g/L. To maintain optimal conditions, the supplementary mixture was manually introduced into the broth when the glucose concentration dropped to 0.5%, ensuring its final concentration reached 1.5% [
57].
Additionally, employing a diverse range of carbon sources in the culture medium is a judicious approach. This strategy entails harnessing both readily metabolized carbon sources (monosaccharides) and sustained-release carbon sources (polysaccharides). By capitalizing on
Streptomyces’ capacity to hydrolyze polysaccharides such as amylase, the microorganisms can initially utilize easily accessible carbon sources for growth and subsequently synthesize their own hydrolases to metabolize and utilize complex polysaccharides. As presented in
Table 6, the RSM model formulation incorporates a concentration of 22 g/L of readily metabolized carbon sources (glycerol) and 4 g/L of sustained-release carbon sources (cassava starch), whereas the ANN-GA model formulation comprises 12 g/L of readily metabolized carbon sources (glycerol) and 17 g/L of sustained-release carbon sources (cassava starch). By optimizing the concentration and composition of carbon sources, the production of the ANN-GA model exhibited a significant increase from 0.1637 g/L (RSM model) to 0.1921 g/L (ANN-GA model). In the absence of feeding extra carbons, the microorganisms autonomously regulate the concentration and composition of sugars in the culture medium, thereby circumventing some of the potential impact on the secondary metabolism induced by carbon sources [
58,
59,
60]. For example, the production of cold-adapted beta amylase from
Streptomyces was enhanced by Cotârlet et al. through the utilization of a medium formulation comprising glycerol and starch, with an optimized ratio [
61,
62]. Smaoui [
63], Al-Ansari [
64], and Ni [
65] et al. opted for a medium formulation comprising glucose and starch and meticulously optimized the proportions of these constituents to enhance the yields. However, different
Streptomyces species may exhibit diverse metabolic mechanisms for utilizing multiple carbon sources. For instance,
Streptomyces albulus M-Z18 demonstrates the ability to efficiently utilize both glucose and glycerol in a manner unaffected by the presence of glucose. Zeng et al. successfully enhanced ε-poly-L-lysine productivity from
S.albulus M-Z18 by supplementing the growth medium with a combination of glucose and glycerol [
66]. Additionally, glycerol exerts a beneficial effect on the biosynthesis of antibacterial compounds in certain strains of Streptomyces [
67,
68]. In conclusion, there exists a scientific foundation for
S. parvus SCSIO Mla-L010/Δ
vicG to achieve enhanced production through the utilization of a medium with an optimized ratio of glycerol to starch. However, further investigations into the carbon metabolism and polyketide synthase (PKS) mechanisms of
S. parvus SCSIO Mla-L010/Δ
vicG are imperative for production improvement.
The strategy used to enhance the yield by manipulating the seawater salt concentration is predicated on the principle that microorganisms maintain equilibrium between internal and external osmotic pressures through endogenous synthesis, environmental absorption, and other mechanisms for counteracting external osmotic pressure [
69,
70]. In particular, marine microorganisms, such as
Streptomyces, isolated from the sea often exhibit significant alterations in their secondary metabolism due to variations in seawater salt concentration. For instance, Sanjivkumar, Selvaraj, and Manivasagan et al., respectively, optimized the sea salt proportion in the medium to enhance the production of chitinase in
Streptomyces olivaceus MSU3 [
71], antibiotics in
Streptomyces sp. CMSTAAHAL-3 [
72], and α-amylase in
Streptomyces sp. MBRC-82 [
73]. In this study,
S. parvus SCSIO Mla-L010/Δ
vicG exhibited a remarkable tolerance to high osmotic pressure (twice that of the NaCl concentration in seawater) and demonstrated the production of 4′-N-demethyl-vicenistatin. However, the synthesis pathway of 4′-N-demethyl-vicenistatin was found to be obstructed under high osmotic pressure. Surprisingly, through harnessing the synergistic interaction between the sea salt and the carbon source in the medium, augmenting the sea salt concentration by 4 g/L significantly amplified its production.
The second aspect pertains to the selection of research methods. In this study, the ANN-GA model demonstrated a superior performance over the RSM model in optimizing the production of 4′-N-demethyl-vicenistatin by
S. parvus SCSIO Mla-L010/Δ
vicG, both in terms of example verification and overall model performance. It has been widely acknowledged that researchers in the field of microbial fermentation optimization commonly employ conventional statistical methods, including the OFAT approach [
74], orthogonal experiments [
75], PBD [
76,
77,
78], and RSM [
14,
79,
80], for optimization purposes. For instance, the RSM, whether it is the Box–Behnken design (BBD) or CCD, is generally applicable when there are no more than four experimental factors. Consequently, by designing a lesser number of experimental groups, one can effectively explore the influence of factor interactions on the response value while employing multivariate quadratic polynomials to represent the constructed model [
81]. However, the limitations of simple binary equations for accurately elucidating the intricate mechanisms involved in microbial fermentation necessitate the utilization of ANN, which possess the capability to construct more sophisticated models and generally outperform RSM when confronted with such challenges [
39,
82,
83,
84,
85].
The ANN is a computational model composed of interconnected nodes, also known as neurons, wherein each node represents an activation function that determines a specific output. The interconnections between the nodes are represented by weighted values called weights, which serve as the memory component of the ANN. Consequently, the network’s output is contingent upon these connections, weight values, and activation functions [
86]. The ANN model has the following advantages: Firstly, it possesses the capability of autonomous learning. The autonomous learning function is particularly crucial for prediction. It is anticipated that ANN will offer extensive economic forecasts, market predictions, and benefit projections, thereby rendering its application prospects highly promising. Secondly, ANN exhibits an associative storage function, which can be achieved through its own feedback networks. Thirdly, it demonstrates high-speed computational abilities to rapidly identify optimal solutions. Solving complex problems often necessitates extensive calculations; however, by employing a specific feedback artificial neural network, optimal solutions can be swiftly obtained [
87]. Thanks to the development of AI technologies, the rise of ChatGPT, and the advent of Python, non-computer domain researchers now have the opportunity to use ANN to solve problems more conveniently. Furthermore, artificial neural networks (ANN) exhibit exceptional performance in some specific domains, such as the modeling of drinking water quality [
88], thermal analysis [
89], the design of model predictive control system [
90], pattern recognition [
91], and photovoltaic fault detection and diagnosis [
92], among others. This also encompasses applications in biology [
93,
94,
95].
The application of ANN in biological fields, particularly when combined with RSM, has demonstrated remarkable outcomes, even with a limited number of samples. However, caution should be exercised due to the potential risk of inaccurate fitting arising from the constrained sample size. Consequently, to enhance the generalization capability of ANN models, some researchers opt for incorporating k-fold cross-validation during the training process [
96,
97]. Additionally, the training process of an ANN model necessitates the meticulous configuration of numerous parameters. Inadequate parameter settings can lead to either overfitting or underfitting, thereby compromising the model’s predictive capacity. Relevant experiments for determining these parameters are typically devised by researchers either employing OFAT designs or establishing parameter values based on their own expertise [
98,
99]. In recent years, researchers have increasingly employed GA, particle swarm optimization (PSO), artificial bee colony (ABC), and the backtracking search algorithm (BSA), in conjunction with ANN models, to enhance the efficiency of optimizing training parameters for ANN models, enabling them to address progressively intricate problems more effectively [
100].
In this study, after employing k-fold cross-validation and OFAT to optimize the training parameters, an ANN model (3-6-1) was ultimately utilized to investigate the impact of variations in cassava starch, glycerol, and seawater salt concentrations on 4′-N-demethyl-vicenistatin production. Based on the knowledge obtained here, it can be understood that this ANN computational model encompasses 21 functions with diverse weights, facilitating the more precise prediction of outcomes compared to the ternary quadratic polynomial of the RSM model. Therefore, as illustrated in
Table 6, the ANN model exhibits a superior performance compared to the RSM model, as evidenced by its lower error value (ANN-GA: 1.90% vs. RSM: 11.40%), reduced APD value (ANN-GA: 5.88 vs. RSM: 10.58), and decreased RMSE value (ANN-GA: 0.0051 vs. RSM: 0.0056). Furthermore, as depicted in
Figure 5, the experimental results align more closely with those expected from the ANN-GA model during validation.
The utilization of RSM-ANN in optimizing the medium formulation may consequently yield superior outcomes compared to those achieved solely through RSM.