Flavor enhancement of Perinereis aibuhitensis Maillard reaction liquids based on the RSM-BP-GA model

Abstract

The application of Perinereis aibuhitensis is limited by the poor flavor of its hydrolysate. Thus, we aimed to optimize the preparation conditions of P. aibuhitensis Maillard reaction solution (PMS) by using a response surface methodology (RSM) combined with a back-propagation neural network coupled with a genetic algorithm (BP-GA). The optimal conditions were as follows: ribose:glucose, 1:1; reducing sugar addition, 3.21%; pH, 6.5; reaction temperature, 121 °C; and reaction time, 64 min. The volatile flavor components and free amino acid content in PMS (prepared under the optimal conditions) and P. aibuhitensis enzymatic hydrolysis solution were determined. PMS had good flavor, reduced aromatic hydrocarbon and ether contents, and increased aldehyde. The proportions of bitter and sweet amino acids decreased and increased, respectively, which significantly improved taste. This study demonstrates the effectiveness of machine learning-assisted optimization for flavor enhancement and provides a viable strategy for developing high-value seafood seasonings from underutilized marine resources.

Graphical abstract

Highlights

• •Preparation conditions for PMS were optimized using RSM combined with BP-GA.
• •PMS prepared under optimized conditions had reduced undesirable flavor substances.
• •PMS prepared under optimized conditions had more pleasant aroma volatile components.
• •PMS prepared under optimized conditions had significantly improved flavor.
• •Machine learning algorithms can improve the quality of food products.

1. Introduction

Perinereis aibuhitensis is a common marine invertebrate with high yield, nutritional and medicinal value, and distinctive flavor profile (Liu, Wang, et al., 2022). Given its use as fishing bait and in aquaculture, P. aibuhitensis shows potential for environmental remediation (Fang et al., 2017) and bioactive peptide extraction (Jiang et al., 2019). Despite these applications, few studies have focused on developing high-value food products from P. aibuhitensis.

Enzymatic hydrolysis is commonly used for the development of aquatic foods. The Maillard reaction can address issues such as the poor flavor and dark color of enzymatic hydrolysis solutions (Chen & Wang, 2017). In food flavor chemistry, the Maillard reaction, as the core mechanism of non-enzymatic browning reactions, plays a key role in the preparation and flavor construction of seafood seasonings (Siewe et al., 2024). Various flavor compounds with pleasant taste and aroma (Chen et al., 2019) and active substances with antioxidant, anti-inflammatory, and anticancer effects can be generated (Langner & Rzeski, 2014) by controlling Maillard conditions such as reaction temperature, reaction time, and reducing sugar addition; moreover, controlling these conditions can minimize the generation of adverse flavor compounds, thereby improving the yield and quality of seafood seasonings (Siewe et al., 2024).

Precursors of seafood seasonings are usually required to be rich in various amino acids, umami peptides, nucleotides, and other flavor compounds (Bhuiyan, 2026; Cai et al., 2016). Given its high levels of flavor substances, such as certain amino acids, umami peptides, and nucleotides, P. aibuhitensis enzymatic hydrolysis solution (PES) is a suitable precursor for seafood seasonings (Liu, Wang, et al., 2022). P. aibuhitensis Maillard reaction solution (PMS) prepared from PES has good flavor (Guo et al., 2012). Therefore, PES has high potential as a new material for the preparation of seafood seasonings.

Optimizing the Maillard reaction conditions is a complex, multi-variable problem that involves strong non-linear relationships between process parameters and final product quality. Response surface methodology (RSM) is commonly used for such optimization but can exhibit limitations in accurately modeling highly non-linear systems (Qiu et al., 2023). To improve prediction accuracy, machine learning-based approaches have gained increasing attention in food formulation design (Ataguba et al., 2024). Among these, back-propagation (BP) neural networks are well-suited for capturing non-linear patterns, while genetic algorithms (GA) provide robust global search capabilities (Huang et al., 2026). The coupling of BP-GA with RSM enables a more effective exploration of the parameter space, helping to identify optimal processing conditions without becoming trapped in local optima (Chi et al., 2024). This hybrid approach is particularly advantageous in food chemistry contexts where multiple interacting variables must be simultaneously optimized—such as in the Maillard reaction—where traditional RSM may not fully capture the underlying complexity. Although its application may be limited by the need for large sample size, this problem can often be solved by creating virtual samples (Jiang et al., 2025). For example, previous studies have obtained the preparation conditions with the highest degree of hydrolysis for the enzymatic hydrolysis of P. aibuhitensis proteins through this approach (Teng et al., 2025).

In this study, we aimed to (1) explore the optimal preparation conditions for PMS using the RSM-BP-GA model; (2) analyze the reasons for the improvement in taste and flavor changes in PMS (prepared under optimal conditions) by measuring the different compositions of chemical substances and free amino acids between PES and PMS through gas chromatography–mass spectrometry and free amino acid analysis; and (3) provide relevant methods and a theoretical basis for the use of P. aibuhitensis and other marine organisms as aquatic seasonings.

2. Materials and methods

2.1. Materials and reagents

Adult P. aibuhitensis were purchased from Shandong Dongying Zhenyu Aquatic Resources Development Co., Ltd. (Shandong, China). Flavor protease (food-grade, 0.83 kat/kg) was purchased from Shandong Qilu Biotechnology Group Co., Ltd. (Shandong, China). L-cysteine (food-grade), glucose (food-grade), and D-ribose (food-grade) were purchased from Shenzhen Xinrong Biotechnology Co., Ltd. (Guangdong, China). All other reagents were of analytical grade.

2.2. Experimental methods

2.2.1. Preparation of PMS

P. aibuhitensis was subjected to ultrasound-assisted enzymatic hydrolysis. The experimental method and enzymatic hydrolysis parameters were slightly modified based on previous studies (Jiang et al., 2019; Teng et al., 2025). It was homogenized at a solid-to-liquid ratio of 1:3.45. The enzymolysis temperature was 56.8 °C, pH 6.8, flavor protease dosage 1.26% (i.e., 12.6 g/kg), ultrasonic power 360 W, ultrasonic frequency 20 kHz, ultrasonic time 20 min, and enzymolysis time 3 h. After enzymolysis, the enzyme was inactivated in boiling water for 5 min and filtered through a 40-mesh gauze screen for standby to obtain PES. PES was mixed with 2.5% L-cysteine, which enhances flavor and inhibits carcinogenesis, and 3.0% reducing sugar (ribose:glucose = 1:1) and reacted at 120 °C for 60 min at pH 6.5 to obtain PMS.

2.2.2. Single-factor test

According to the steps in Section 2.2.1, single factors including the ribose-to-glucose ratio (3:1, 2:1, 1:1, 1:2, and 1:3), reducing sugar addition (1%, 2%, 3%, 4%, and 5%), reaction temperature (100, 110, 120, 130, and 140 °C), reaction time (20, 40, 60, 80, and 100 min), and pH (5.5, 6.0, 6.5, 7.0, and 7.5) were optimized. Three replicates of each assessment were arranged.

2.2.3. RSM optimization test

Based on the results of the single-factor test, real values were converted to coded values using the Design Expert 13 software (Ruan et al., 2025). The reducing sugar addition, reaction time, and reaction temperature were selected as independent variables (changes in these three factors had a greater impact on the sensory evaluation of PMS than the glucose-to-ribose ratio and pH; Fig. 1B–D), and the sensory score was used as the response value in further RSM optimization tests. The levels and codes of the test factors are listed in Table 1. In total, 17 combinations were generated with three replicates per combinations.

Fig. 1.

Single-factor experiments and feature importance assessments. Single-factor analysis of the reducing sugar ratio (A, ribose: glucose), reducing sugar addition (B), reaction temperature (C), reaction time (D), and reaction pH (E) (n = 3 per group, one-way ANOVA with Tukey's multiple comparison test, P < 0.05).

Table 1.

Response surface factors and levels.

Factors	A Reaction Temperature/°C	B Reaction Duration/min	C Reducing Sugar Addition/%
Level	110	40	2
	120	60	3
	130	80	4

2.2.4. BP neural network model design

The generation of virtual samples is based on errors generated by experiments, that is, due to factors such as raw materials, operations, and instrument equipment. Specifically, they were generated by adding small random noise (±0.1%) to the independent variables of the 17 actual RSM data points to artificially expand the training dataset for the BP neural network, improving its robustness and generalization ability (Zheng et al., 2023). Virtual samples were created for 17 actual samples in the response surface optimization test (four virtual samples were created for each actual sample; specific examples of virtual samples are shown in Table S1), and 85 samples were finally obtained. After normalization, 61 of the data were used for network training, 12 for testing, and 12 for verification (data did not overlap). The reducing sugar addition, reaction time, and reaction temperature were inputs for the BP neural network, and the sensory score was used as the output. A three-layer network structure was adopted. The number of neurons in the hidden layer was determined using a trial-and-error method and the following empirical Eq. (1), in which neural network models were constructed by continuously changing the number of neurons in the hidden layer and compared in terms of performance to find the optimal number of neurons in the hidden layer. The model topology was 3-n-1, as shown in Fig. S1.

$$H=\sqrt{I+O}+\mathrm{a}$$ (1)

where H represents the number of neurons in the hidden layer, I is the number of neurons in the input layer, O is the number of neurons in the output layer, and a is a constant ranging from 1 to 10.

2.2.5. Genetic algorithm for optimization

The genetic algorithm toolbox in MATLAB software was used for optimization (Hunde et al., 2024). There were three variables with upper and lower limits: reaction temperature [109.89 and 130.13 °C], reaction time [39.96 and 80.08 min], and reducing sugar addition [0.01998 and 0.04004]. The optimization process was conducted using a single-point crossover, the mutation probability was 0.05, the evolutionary generations were 100, and the other parameters were maintained at their default settings.

2.3. Sensory score and composition determination

2.3.1. PMS sensory evaluation

Sensory evaluation of PMS was performed as described by Wen et al. (2023). A sensory panel of 10 trained judges (aged 19–26 years, five males and five females) from the Sensory Science Laboratory at China Agricultural University. As the national regulations did not require ethical approval for sensory evaluation studies, the exemption was obtained from the ethics committee. During the research process, we strictly followed the guidelines of the Standard Guide for Protection of Respondents and Informed Consent for Sensory Evaluation Studies (ASTM E3314-21), which entailed requesting for fully voluntary participation without any form of coercion, fully informing participants of all requirements and potential risks of the study, obtaining written/oral informed consent from all participants, not disclosing the data of the participants without informing them, and allowing the participants to withdraw unconditionally at any time during the research process.

These judges, with over 300 h of training in sensory assessment, evaluated the color, aroma, and taste of PMS. Before the formal sensory evaluation, these judges received a 7-day training (twice a day, each lasting an hour). They compared and evaluated the color, aroma, and taste of PMS (at different processing levels) with those of various standard reference materials (such as 200 g of fish meat boiled in 200 mL of water as a reference for fishy smell) (Siewe et al., 2020), common seafood seasonings (such as oyster sauce and fish sauce), and other seafood products (Hu et al., 2021). The formal evaluation process was completed in a standard evaluation room (single room, quiet, clean and tidy, well ventilated, and free of rubbishes). Each judge evaluated each sample randomly and recorded the corresponding results on a five-point scale (1 = very poor, 2 = poor, 3 = acceptable, 4 = good, 5 = excellent). Each evaluation had a 30-min interval, and the sample order was rearranged before re-evaluation (Shi et al., 2024). Each person was tested three times in parallel. The proportions of color, aroma, and taste to the score were 20%, 30%, and 50%, respectively. These proportions were determined based on the importance of color, aroma, and taste in the development of seafood seasonings (Siewe et al., 2024). The sensory evaluation criteria are shown in Table 2.

Table 2.

Sensory evaluation.

Evaluation Criteria	Sensory Description	Score (points)
Color	Grey	1–2
	Light Brown	3
	Reddish Brown	4–5
Aroma	Faint characteristic aroma of PMS, strong fishy and off-odors	1–2
	Moderate characteristic aroma of PMS, moderate fishy and off-odors	3
	Strong characteristic aroma of PMS, faint fishy and off-odors	4–5
Taste	Faint meaty and savory flavor, strong bitterness and burnt taste	1–2
	Moderate meaty and savory flavor, moderate bitterness, and burnt taste	3
	Strong meaty and savory flavor, faint bitterness and burnt taste	4–5

Characteristic aroma of PMS: a rich and unique mellow seafood aroma with a roasted or nutty note.

2.3.2. Determination and analysis of volatile flavor compounds in PES and PMS

The method described by Jeong et al. (2024) was used to identify volatile flavor compounds in PES and PMS (prepared under optimal conditions) (n = 3). The chromatographic column was an HP-INNOWax column (60 m × 250 μm × 0.25 μm). The temperature program was as follows: holding at 40 °C for 5 min, increasing to 250 °C at 5 °C/min, and holding for 10 min. The flow rate of the carrier gas (He) was 1.5 mL/min, and the injection volume was 0.5 μL. The split ratio was 4:1, ion source was electron bombardment, electron energy was 70 EV, ion source temperature was 230 °C, temperature of the quadrupole was 150 °C, and mass scanning range was m/z 35–550. During data analysis, the nist17 database was used for automatic retrieval and manual comparison, and components with a matching degree >80 were counted. The odor activity value (OAV) was calculated using the following equation (Nie et al., 2024):

$$\mathit{OAV}=\frac{c}{\mathit{OT}}$$ (2)

where c represents the absolute concentration of the compound, and OT refers to its olfactory threshold in an aqueous solution.

2.3.3. Determination of free amino acids in PES and PMS

The free amino acid content and taste activity value (TAV) in PES and PMS (prepared under optimal conditions) (n = 3) were determined using an Agilent 1100 liquid chromatograph following the method described by Song et al. (2024). Briefly, 0.5 g of PES and PMS was weighed and placed in a 10 mL centrifuge tube, and 5 mL of 0.01 mol/L hydrochloric acid was added and mixed well. After boiling in a water bath for 30 min, the mixture was centrifuged at 8600 ×g for 10 min. The supernatant was obtained by adding 0.01 mol/L hydrochloric acid to the precipitate and suspending it for 5 min, centrifuging it at 8600 ×g for 10 min, mixing the supernatant, and diluting to 10 mL. The content of free amino acids in the supernatant was measured using membrane filtration.

2.4. Data processing

All measurements were repeated at least three times, and the SPSS 27.0 software (version 27.0; SPSS Inc., Chicago, IL, USA) was used for the t-test and one-way analysis of variance (ANOVA) with Tukey̕’s multiple comparison test. The results are expressed as the mean ± standard error (SE). Statistical significance was set to P < 0.05. Images were plotted using the Origin Software 2021 (OriginLab Corporation, Northampton, Massachusetts, USA), and response surface data were processed using the Design Expert 13 software. The BP-GA models were developed and trained using the MATLAB 2018b software. With reference to Liu, Li, et al. (2022), a flowchart that depicts the determination of parameter values and prediction of the optimal reaction conditions using the RSM-BP-GA model is shown in Fig. S2.

3. Results and discussion

3.1. Single-factor experiment results and analysis

Ribose-to-glucose ratio and reducing sugar addition (Fig. 1A and B): regarding the type of reducing sugar, the color score of PMS decreased with increasing glucose addition ratio. The aroma and taste scores initially increased and then decreased. When the mass ratio of ribose to glucose was 1:1, the total sensory score of PMS was highest, showing a significant difference compared with the other experimental groups (P < 0.05). In addition, the color score of PMS increased with increasing reducing sugar addition, whereas the aroma and taste scores initially increased and then decreased. When reducing sugar addition was in excess, the Maillard reaction product exhibited pronounced scorching bitterness, and the characteristic aroma of PMS (a rich and unique mellow seafood aroma with a roasted or nutty note) was also masked, which decreased the sensory score. Luan et al. (2020) reported similar results for the development of an Alaska Pollock frame seasoning powder. When the reducing sugar addition was 3%, the sensory score of PMS was significantly higher than that of the other experimental groups (P < 0.05).

Reaction temperature (Fig. 1C): the color score of PMS increased with increasing reaction temperature, whereas the aroma and taste scores initially increased and then decreased. When the temperature was increased to 120 °C, the product exhibited the characteristic aroma of PMS, a dark red-brown color, the highest sensory score, and a significant difference compared with the other experimental groups (P < 0.05).

Reaction time (Fig. 1D): as the reaction time was prolonged, the color score of PMS increased, whereas the aroma and taste scores initially increased and then decreased. When the reaction time was 60 min, many Maillard reaction products with good flavor were formed in PMS (Feng et al., 2024), and the sensory scores for aroma and taste increased. PMS had the highest total sensory score, with significant differences compared with the other experimental groups (P < 0.05). The results of Li, Sun, et al. (2023) regarding the optimal production of zanthoxylum seasoning oil also showed that appropriate reaction temperature and time can improve the sensory evaluation of Maillard reaction products.

PH value (Fig. 1E): as the pH was increased, the reddish brown color of PMS deepened and the color score increased, whereas the aroma score initially decreased, increased, and then decreased again, and the taste score initially increased and then decreased. When the pH was 6.5, the total sensory score of PMS was highest.

Based on these results, the reaction conditions selected in this study were as follows: ribose-to-glucose ratio of 1:1, reducing sugar addition of 3%, reaction temperature of 120 °C, reaction time of 60 min, and pH of 6.5.

3.2. RSM optimization test results and analysis

The results of the RSM optimization test and analysis of variance (ANOVA) are shown in Table 3, Table 4, respectively. We correlated the sensory evaluation scores of the Maillard reaction liquid with the reaction temperature (A), reaction temperature (B), and reducing sugar addition (C) in a multivariate regression equation:

$$\text{Sensory score}=4.58–0.19\mathrm{A}–0.23\mathrm{B}+0.07\text{4C}–5.000\times {10}^{-3}\mathrm{AB}–0.038\mathrm{AC}–0.050\mathrm{BCE}–0.75{\mathrm{A}}^2–0.74{\mathrm{B}}^2–0.57{\mathrm{C}}^2$$ (3)

Table 3.

Response surface test analysis scheme and results.

Number	A Temperature/°C	B Time/min	C Reducing sugar addition/%	Sensory score
1	120	80	4	3.09
2	110	60	2	3.44
3	120	60	3	4.61
4	120	40	4	3.62
5	130	60	4	3.09
6	130	40	3	3.15
7	120	60	3	4.57
8	120	40	2	3.35
9	120	60	3	4.56
10	110	60	4	3.54
11	120	80	2	3.02
12	110	40	3	3.54
13	130	60	2	3.04
14	120	60	3	4.55
15	110	80	3	3.04
16	120	60	3	4.62
17	130	80	3	2.63

Table 4.

Analysis of variance results.

Variance Source	Degrees of Freedom	F-value	P-value
Model	9	701.48	<0.0001⁎⁎
A	1	250.11	<0.0001⁎⁎
B	1	367.95	<0.0001⁎⁎
C	1	36.24	0.0005⁎⁎
AB	1	0.083	0.7813
AC	1	4.68	0.0672
BC	1	8.33	0.0235⁎
A2	1	1997.57	<0.0001⁎⁎
B2	1	1906.01	<0.0001⁎⁎
C2	1	1158.39	<0.0001⁎⁎
Residual	7
Misfit	3	1.55	0.3314
Error	4
Total	16
R2		0.9989
RAdj2		0.9975
C.V.%		0.9897

^⁎denotes significant difference, P < 0.05;

^⁎⁎signifies extremely significant difference, P < 0.001.

The F-value of the model was 701.48 (P < 0.001); the P-value of the lack-of-fit term was 0.3314 (P > 0.05); the R_Adj² and R² values were similar (indicating a sufficiently large sample size), and the experiment's coefficient of variation (C.V.%) was 0.9897, indicating reliability, precision, and practical applicability. According to the F-value, the order in which each factor affected the response value was as follows: reaction time (B) > reaction temperature (A) > reducing sugar addition (C). Shen et al. (2021) also found that the reaction time and temperature affect the Maillard reaction greater than the reducing sugar addition.

We formulated a regression equation and drew a response surface graph to investigate the impact of the reducing sugar addition, reaction time, and reaction temperature on the sensory scores of PMS. The steepness of the response surface reflects the extent to which the experimental factors affect the sensory scores. The steeper the curve, the more pronounced the influence on the sensory scores. As shown in Fig. 2, the interaction between the reaction time and reducing sugar addition was the most significant (P < 0.05). The interactions were ranked in decreasing order of significance as follows: BC (reaction time× reducing sugar addition) > AC (reaction temperature × reducing sugar addition) > AB (reaction temperature × reaction time).

Fig. 2.

Response surface optimization experiment results. Three-dimensional response diagrams showing the interaction of various factors: the steepness of the response surface indicates the degree of influence on the sensory score, with a steeper surface correlating to a more pronounced impact. A, B, and C indicate reaction temperature (temperature), reaction time (time), and reducing sugar addition (amount of reducing sugar added).

Based on the RSM results, the optimal conditions selected for PMS were a reaction temperature of 118.70 °C, reaction time of 56.77 min, and reducing sugar content of 3.08%. For practical application, the optimal condition was adjusted to a reaction temperature of 119 °C, reaction time of 57 min, and reducing sugar addition of 3.08%. Three replicate verification experiments were conducted, and the results were averaged under these conditions. The sensory score for PMS was 4.66.

3.3. RSM-BP-GA model optimization results and analysis

The BP neural network has been widely used owing to its strong adaptability and high accuracy in handling non-linear problems. In previous studies, the optimal process parameters of enzymatic hydrolysis of P. aibuhitensis (with the highest degree of hydrolysis) were effectively identified by the RSM-BP-GA model (Teng et al., 2025). In this experiment, we also used this method to determine the optimal preparation conditions of PMS. The range of values for the hidden layer neurons was determined using an empirical formula (Eq. 1 in Section 2.2.4), and the optimal number of neurons was identified using a trial-and-error method (Yuan et al., 2023). The trial-and-error method mainly includes the following steps: 1) Determination of the range of values for hidden layer neurons; 2) Selection of a value for the hidden layer neurons from this range, creation of a neural network, and training of this network using a training set; 3) Evaluation of the performance of the trained network using a validation set, i.e., calculation of the mean squared error; 4) Steps 2 and 3 are then repeated until all values for the hidden layer neurons in the range are traversed; 5) Finally, the values for the hidden layer neurons with the best performance in the validation set are chosen as the final result. Testing revealed that the model performed best with eight hidden layer neurons. The training process is shown in Fig. 3A. Training was stopped at epoch 26 using early stopping, as the best validation performance was achieved at epoch 20, indicating that network convergence was rapid and stable, leading to optimal verification performance. The BP neural network model with a topology of 3–8–1 was used to analyze the sensory score. The degree of fit between the target and output values of the training, validation, testing, and overall sets exceeded 0.999 (Fig. 3B), which indicated that the BP neural network model developed using the aforementioned method exhibited robust stability and predictability. Thus, this model accurately describes the relationship between the test factors and the PMS sensory score.

Fig. 3.

Training process for the neural network model (A); training stopped after 26 iterations, indicating swift and stable convergence with optimal validation performance. Fitting the target and output values in the BP neural network model (B); R values for all datasets exceeded 0.999, attesting the robust stability and predictability of the BP neural network model, which accurately described the relationship between reaction conditions and the PMS sensory score. Optimization process of the genetic algorithm (C); average change in fitness values below 10^-6 termination.

The BP neural network was used following the method described by Zhu et al. (2024) to capture the mapping relationship between the input and output, serving as the fitness function for the GA. The search boundaries were defined based on the ranges of the training samples. After continuous selection, crossover, and mutation operations, iterations were performed until the 91st generation. The average variation of the fitness value was <10^-6 (termination condition), after which iteration stopped (Fig. 3C). Finally, when the reducing sugar amount, reaction time, and reaction temperature were 3.21%, 64.48 min, and 120.80 °C, respectively, the maximum predictive value of PMS sensory score was 4.82.

Considering practical application, the optimal reaction conditions were adjusted to a reducing sugar addition of 3.21%, reaction time of 64 min, and reaction temperature of 121 °C. Three repeated tests were conducted using these conditions to verify the reliability of the predicted results, and the average value was used to determine the actual sensory score. Under these conditions, the actual sensory score of PMS was 4.80, which was higher than that of RSM alone (4.66). The findings illustrate that the BP neural network model can capture the relationship between reaction conditions and sensory scores and that improved Maillard reaction conditions can be obtained by coupling it with a GA to optimize the results of the response surface.

3.4. Composition and analysis of volatile flavor compounds in PMS and PES

The principal component analysis diagram shows that volatile flavor substances in PMS changed significantly than those in PES (Fig. 4A). There were 50 and 62 volatile compounds in PES and PMS, respectively, of which 10 and 22 were unique to PES and PMS, respectively (Fig. 4B). The types of aldehydes and alcohols generated in PMS were significantly more than those in PES. The types of aromatic hydrocarbons and ethers were reduced, similar to the results reported by Fu et al. (2023), who aimed to improve the flavor of tilapia fish head soup through the Maillard reaction. In addition, four ketone substances, which were also generated in PMS, were not detected in PES (Fig. 4C). PES primarily includes 16 key flavor substances (OAV ≥ 1.0), such as dimethyl sulfide (with a strong pungent odor, low sensory threshold) and four modified flavor substances (0.1 ≤ OAV < 1.0) (Fig. 4D). The combination of these key flavor substances and modified flavor substances imparted onion, fishy, and other bad flavors to PES. PMS contained 19 key flavor substances (OAV ≥ 1.0), such as hexanal and heptanal, and seven modified flavor substances, such as 2-nonanone and 2-pentadecanone (0.1 ≤ OAV < 1.0). Aldehydes with desirable flavors, such as hexanal, usually have a low olfactory threshold, which plays an important role in forming aroma during the oxidative degradation of fats from P. aibuhitensis (Dabbaghi et al., 2019). Ketones are primarily derived from fat oxidation, especially 2-ketones, which contribute significantly to meat flavor (Steingass et al., 2015). In addition, PMS has substances with low thresholds, such as pyrazine and heptanol, which contribute greatly to seafood flavor (Li, Ma, et al., 2023). Together, these substances can confer coffee, oil, and nut flavors to PMS. Feng et al. (2021) investigated Maillard reaction products from Shanghai smoked fish and found that the reaction processing produces various aldehydes, alcohols, ketones, and other substances, which contribute significantly to the formation of its unique flavor. After the Maillard reaction in the current study, the OAV value of dimethyl sulfide with an adverse pungent odor in PMS decreased significantly, and the OAV values of hexanal, heptanal, and other substances with good flavors increased significantly (Fig. 4E). These results further indicate that the Maillard reaction can effectively improve the flavor of PES. The specific types, contents, and sensory descriptions of the detected volatile compounds are shown in Table S2.

Fig. 4.

Determination results of volatile flavor components in PES and PMS. PCA of the volatile flavor components in PES and PMS (A). Differences in the number of types of the volatile flavor components (B); differences in the specific types of volatile flavor components between PES and PMS (C); differences in the content of volatile flavor components between PES and PMS (D), and differences in OAV (E). PES and PMS indicate P. aibuhitensis enzymatic hydrolysis solution and P. aibuhitensis Maillard reaction solution, respectively (n = 3 per group, t-test, P < 0.05).

3.5. Flavor characteristics and analysis of free amino acids in PES and PMS

In PMS, the proportion of sweet amino acids relative to the total free amino acids was 56.73%, approximately twice that in PES (Fig. 5A and Table S3). Among them, glycine accounted for 45.50%, which was approximately 15 times higher than the proportion in PES. The significantly increased (P < 0.05) proportion of glycine not only enhanced the sweet taste but also reduced the unpleasant taste of PMS. The proportions of bitter amino acids in PES and PMS were 42.33% and 19.66%, respectively. This result indicates that the Maillard reaction reduced the proportion of bitter amino acids. Qiu et al. (2024) investigated tilapia byproduct hydrolysates and observed that bitter amino acids, whose proportion decreased following the Maillard reaction, are the predominant flavor of amino acids in aquatic enzymatic hydrolysates. TAV is often used to indicate the ratio of the content of flavoring substances to their threshold value; substances with a TAV greater than 1 can significantly contribute to taste (Wang et al., 2024). The greater the TAV of the flavor amino acids, the stronger the flavoring effect and its contribution to taste. In the present study, glycine had the largest TAV in PMS, whereas the TAVs of bitter amino acids, such as methionine and leucine, were high in PES. After the Maillard reaction, the TAVs of bitter amino acids decreased (Fig. 5B).

Fig. 5.

Determination results of free amino acids in PES and PMS. Differences in the free amino acid content between PES and PMS (A), and in TAV (B). PES and PMS indicate P. aibuhitensis enzymatic hydrolysis solution and P. aibuhitensis Maillard reaction solution, respectively (n = 3 per group, t-test, P < 0.05).

3.6. Conclusions, limitations, and future prospects

This study successfully optimized the Maillard reaction conditions for PES using an integrated RSM-BP-GA approach. The model-predicted conditions yielded a PMS with significantly enhanced sensory scores compared with traditional RSM optimization. Flavor analysis confirmed that the Maillard reaction effectively increased pleasant volatile aldehydes and ketones while reducing undesirable sulfur compounds. Additionally, the proportion of sweet amino acids, particularly glycine, rose markedly, contributing to a more balanced and appealing taste profile. These findings validate the utility of machine learning-assisted optimization in food production and highlight the potential of P. aibuhitensis as a promising raw material for premium seafood seasonings.

However, owing to constraints in the experimental design, time, and main objective (i.e., to optimize the preparation conditions of PMS using the RSM-BP-GA algorithm to achieve optimal taste and flavor and explain the underlying mechanisms), this study has some limitations. First, the model's performance is dependent on the quality/range of the initial RSM data. Second, the sensory evaluation, while trained, is still subjective. Third, the study was conducted at the laboratory scale; scaling up may present challenges. Last, functional activities were not the primary focus; therefore, we only measured the partial functional activities (antibacterial and antioxidant abilities) of PMS and compared them with those of PES (see Section 2 of Supplementary_material), and did not study them further.

Future research should focus on enhancing model robustness through expanded datasets or hybrid modeling approaches, integrating instrumental flavor analysis with sensory evaluation to reduce sensory subjectivity, and scaling up the optimized process to assess industrial feasibility. Additionally, systematic evaluation of functional and safety properties should be incorporated into product development. Finally, the RSM-BP-GA model could be extended to other marine resources to further promote sustainable utilization and intelligent manufacturing in aquatic food processing.

Glossary

ANOVA, analysis of variance; BP, back-propagation; BP-GA, back-propagation neural network coupled with a genetic algorithm; OAV, odor activity value; PES, P. aibuhitensis enzymatic hydrolysis solution; PMS, P. aibuhitensis Maillard reaction solution; RSM, response surface methodology; SE, standard error; TAV, taste activity value

CRediT authorship contribution statement

Teng Teng: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Resources, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Yuchun Lu: Writing – review & editing, Visualization, Validation, Software, Resources, Methodology, Investigation, Formal analysis, Conceptualization. Yongkuan Shao: Writing – review & editing, Writing – original draft, Validation, Software, Resources, Methodology, Funding acquisition, Formal analysis. Jiaye Sun: Writing – review & editing, Validation, Software, Resources, Methodology, Formal analysis, Conceptualization. Feng Liu: Writing – review & editing, Validation, Software, Resources, Methodology, Funding acquisition, Formal analysis. Chun-E Liu: Writing – review & editing, Supervision, Resources, Project administration, Methodology, Funding acquisition.

Funding

This work was supported by the Natural Science Foundation of Shandong Province, China (Grant number: ZR2021MC131), Shandong Province Leading Enterprise Project in Aquatic Seedlings Industry (Grant number: Lunong 202351-21), Yantai “Double Hundred Plan” Blue Industry Leading Talent Project (Grant number: Yan Fa Gai 202518-01), Yantai Aquatic Seedlings Industry Enhancement Project (Grant number: Yan Hai Yu 20241230), and Yantai City School-Local Integration Development Project (Grant number: 2024XDRHXMP11).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

Supplementary material

Data availability

Data will be made available on request.