A method for screening oyster taste presenting peptide based on multi-modal umami intensity feature recognition

Abstract

The application provides a kind of oyster taste peptide screening method based on multi-modal umami intensity feature recognition, it is related to the cross technical field of food science and bioinformatics, including: constructing oyster umami peptide multi-modal feature dataset;ACTM model is constructed, ACTM model includes the multi-modal feature fusion layer of integrated amino acid property perception attention mechanism APAA and receptor binding site prediction module RBSP, multi-scale convolution attention layer, umami perception hierarchical modeling Transformer encoder UPLM and dynamic sequence length adaptive prediction layer DSLA;Based on oyster umami peptide multi-modal feature dataset, model training, verification and performance evaluation are carried out;The oyster peptide sequence to be screened is input into the trained ACTM model to identify the umami intensity feature, and the candidate peptide is obtained;The screened candidate peptide is subjected to chemical synthesis, sensory evaluation, electroencephalogram physiological verification and molecular mechanism analysis.The application scheme has the advantages of high oyster taste peptide prediction accuracy, high screening efficiency, strong biological mechanism interpretation and the like.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of food science and bioinformatics, specifically to a method for screening oyster flavor peptides based on multimodal umami intensity feature recognition. Background Technology

[0002] With global population growth and consumption upgrading, the demand for natural functional seasonings in the food industry continues to grow. The efficient utilization of protein resources has become a key technological direction for the sustainable development of the food industry. Transforming proteins into bioactive peptides with specific functions through biotechnology can not only increase the added value of protein resources but also meet consumers' urgent demand for natural and healthy foods. Oysters, as the world's largest-producing marine bivalve mollusks, have an annual output exceeding 5 million tons. Their protein content is as high as 10-15%, and their amino acid composition is rich in umami precursor amino acids such as aspartic acid (12.8%) and glutamic acid (16.2%), making them a typical representative of marine protein resources and an ideal raw material for preparing natural umami seasonings.

[0003] Oyster umami peptides, derived from enzymatic hydrolysis of oyster proteins, possess unique molecular characteristics and biological activities. These peptide molecules specifically bind to T1R1 / T1R3 heterodimeric umami receptors in the oral cavity, activating a G protein-coupled signaling cascade. This transmits taste signals to the cerebral cortex via the facial and glossopharyngeal nerves, resulting in a significant increase in gamma wave power density and ultimately, the perception of umami. Oyster umami peptides generally exhibit an "acid-hydrophobic-basic" sequence pattern. Acidic amino acids are responsible for forming crucial electrostatic interactions with the receptors, hydrophobic amino acids stabilize the peptide-receptor complex conformation, and basic amino acids regulate intermolecular distance and orientation. This precise molecular recognition mechanism gives oyster umami peptides a lower umami perception threshold and a richer flavor profile than traditional MSG.

[0004] Existing computational methods are significantly inadequate in handling the complex biological characteristics of oyster umami peptides. The multi-level biosensory process, from molecular recognition to neural coding, involves different biochemical mechanisms and mathematical modeling requirements. Traditional machine learning methods such as support vector machines and random forests can only establish a direct mapping between input features and output labels, failing to distinguish and model these hierarchical biological processes. Accurate identification of the physicochemical properties of amino acids is crucial for umami prediction, but existing methods treat all amino acids as equivalent symbolic units, ignoring differences in key property parameters such as acidity, hydrophobicity, basicity, and molecular size. This results in an inability to accurately capture typical umami peptide sequence patterns such as "acid-hydrophobic-basic." The length of peptide sequences varies significantly. Short peptides (typically containing 2-5 amino acids) primarily rely on local interactions for function, while long peptides (typically containing 15-20 amino acids) involve more complex long-range synergistic effects. This structural characteristic leads to significant performance fluctuations and processing biases in traditional algorithms using fixed parameters when processing peptide sequences of different lengths. The complex nonlinear relationships between multimodal features are difficult to fully exploit through simple feature concatenation or linear weighting methods, making it impossible for existing fusion strategies to adaptively adjust the weight contributions of different feature modes.

[0005] Standard deep learning architectures also suffer from key shortcomings in biological sequence modeling. While traditional convolutional neural networks can extract local features, their fixed kernel parameters cannot adaptively adjust to the biochemical properties of different amino acids, and they lack the ability to model global sequence dependencies. Transformer models achieve global sequence modeling through self-attention mechanisms, but standard attention computation lacks guidance from prior biological knowledge, failing to distinguish the differentiated roles of different amino acids in umami formation. Furthermore, their single sequence processing approach struggles to effectively integrate multidimensional feature information such as molecular fingerprints and physicochemical descriptors. Existing multimodal learning methods primarily employ a post-fusion strategy, processing features from each modality separately before simple merging. This approach fails to fully utilize the intrinsic molecular-level correlations between different modalities. Deep learning models, when dealing with variable-length biological sequences, typically use fixed-length padding or truncation preprocessing, which results in the loss of crucial sequence length information, affecting the model's accurate modeling of peptide molecules of varying lengths. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a method for screening oyster flavor peptides based on multimodal umami intensity feature recognition, comprising the following steps:

[0007] Construct a multimodal feature dataset of oyster umami peptides;

[0008] Constructing the ACTM model;

[0009] Model training, validation, and performance evaluation were conducted based on a multimodal feature dataset of oyster umami peptides.

[0010] The oyster peptide sequences to be screened are input into the trained ACTM model for umami intensity feature recognition to obtain candidate peptides;

[0011] The selected candidate peptides were chemically synthesized, sensory evaluated, validated by electroencephalography, and their molecular mechanisms were elucidated.

[0012] The method for constructing the ACTM model is as follows:

[0013] A multimodal feature fusion layer integrating the amino acid property sensing attention mechanism (APAA) and the receptor binding site prediction module (RBSP) was constructed to perform feature fusion on the multimodal features of oyster umami peptides.

[0014] A multi-scale convolutional attention layer is constructed to perform local feature extraction and feature enhancement on the features output by the multi-modal feature fusion layer;

[0015] Constructing a hierarchical model of umami perception using a Transformer encoder UPLM to simulate the complete biological process of features from molecular recognition to sensory perception;

[0016] A dynamic sequence length adaptive prediction layer (DSLA) is constructed to predict oyster peptide sequences of different lengths using adaptive feature aggregation and umami intensity grading.

[0017] Optionally, the construction of the oyster umami peptide multimodal feature dataset includes:

[0018] 2856 peptide sequences were collected from the enzymatic digestion of oyster proteins;

[0019] Multimodal feature extraction was performed on each peptide sequence, including ECFP4 molecular fingerprint features, molecular descriptor features and molecular docking binding energy features, forming a comprehensive feature vector of 1024+200+1 dimensions.

[0020] The z-score normalization method is used to normalize all features and eliminate numerical differences between different features;

[0021] The normalized feature data were divided into training, validation, and test datasets in a ratio of 7:2:1.

[0022] Furthermore, the multimodal feature extraction for each peptide sequence includes:

[0023] ECFP4 molecular fingerprint feature extraction: The ECFP4 molecular fingerprint was calculated using the RDKit cheminformatics toolkit, with the radius parameter set to radius=2 and the fingerprint bit length nBits=1024. The amino acid sequence was converted into a SMILES molecular structure representation using the Chem.MolFromSequence() function of the RDKit cheminformatics toolkit. The ECFP4 fingerprint was calculated using the Morgan algorithm to generate a 1024-bit binary fingerprint vector, which was used to characterize the topological structure features of the peptide molecule.

[0024] Molecular descriptor feature calculation: Calculate 50-dimensional basic molecular property descriptors, 80-dimensional molecular topological index descriptors, 35-dimensional molecular geometric descriptors, and 35-dimensional electronic property descriptors;

[0025] Molecular docking binding energy characteristic calculation: Based on the homology modeling method, the known crystal structure of the mGluR1 receptor was used as a template to construct a three-dimensional structural model of the T1R1 / T1R3 umami receptor; the binding free energy of oyster peptide with T1R1 / T1R3 receptor was calculated using AutoDock Vina molecular docking software; 100 independent docking calculations were performed on each peptide-receptor complex, and the average value was taken as the final binding energy characteristic value.

[0026] Optionally, the construction of the multimodal feature fusion layer integrating the amino acid property sensing attention mechanism APAA and the receptor binding site prediction module RBSP includes:

[0027] The amino acid property-sensing attention mechanism (APAA) dynamically adjusts feature weights based on the physicochemical properties of each amino acid in the peptide sequence. The calculation formula is as follows:

[0028]

[0029] Where i represents the position of the i-th amino acid in the peptide sequence, and Pacid(i), Phydro(i), Pbasic(i), and Psize(i) represent the acidity, hydrophobicity, basicity, and molecular size properties of the i-th amino acid, respectively.

[0030] The receptor binding site prediction module (RBSP) is used to predict key binding sites of peptide sequences with the T1R1 / T1R3 receptor. The calculation formula is as follows:

[0031]

[0032] Where ⊕ represents feature concatenation, Ffp, Fdesc, and Fdock represent ECFP4 molecular fingerprint features, molecular descriptor features, and molecular docking binding energy features, respectively, Wreceptor is the receptor binding prediction weight matrix, breceptor is the bias term, and σ is the sigmoid activation function.

[0033] An adaptive weighted fusion mechanism enhanced with biological knowledge is used to combine molecular features with biological knowledge. The fusion formula is as follows:

[0034]

[0035] Where Ffused is a 1024-dimensional biological knowledge-enhanced feature vector, ⊙ represents element-wise multiplication, and α, β, and γ are adaptive weight parameters.

[0036] Furthermore, the construction of the multi-scale convolutional attention layer includes:

[0037] The multi-scale convolutional feature extraction module employs three parallel one-dimensional convolutional branches, using Conv1D layers with kernel sizes of 3, 5, and 7 to extract local features at different scales. Each convolutional branch contains 128 convolutional kernels, and the convolution operation formula is as follows:

[0038]

[0039] Where x represents the input feature vector, k∈{3, 5, 7} is the convolution kernel size, Wk is the weight matrix of the k-th convolution kernel, and bk is the corresponding bias vector. Represents the convolution operator; ReLU is the modified linear unit activation function.

[0040] The channel attention mechanism is used to calculate the importance weight of each feature channel for umami prediction. The calculation formula is as follows:

[0041]

[0042] Where F represents the input feature matrix, GAP is the global average pooling operation, FC is the fully connected layer, and σ is the sigmoid activation function;

[0043] Spatial attention mechanism, used to identify the most critical amino acid sites for umami prediction, is calculated using the following formula:

[0044]

[0045] Wherein, Conv1D is a one-dimensional convolution operation;

[0046] Dual attention feature fusion multiplies the channel attention and spatial attention weights with the original features to obtain an attention-enhanced feature representation. The fusion formula is as follows:

[0047]

[0048] Here, ⊙ represents element-wise multiplication.

[0049] Furthermore, the construction of the umami perception hierarchical modeling Transformer encoder UPLM includes:

[0050] The molecular recognition layer uses the first and second layer encoders to identify key structural features of the peptide, employing a structure-aware self-attention mechanism. The calculation formula is as follows:

[0051]

[0052] Where Q is the query matrix, K is the bond matrix, WQ and WK are the corresponding weight matrices, dk is the bond vector dimension, and Bstructure is the structure bias matrix, which is constructed based on the spatial distance and chemical bond connection relationship between amino acids.

[0053] The receptor activation layer uses the 3rd-4th layer encoder to handle the interaction between peptides and T1R1 / T1R3 receptors, employing a receptor-guided attention mechanism. The calculation formula is as follows:

[0054]

[0055] Among them, Qreceptor and Kreceptor are receptor structural information embedding vectors, which are constructed based on key amino acid residues of the T1R1 / T1R3 receptor;

[0056] The perceptual coding layer employs encoders layers 5 and 6 specifically responsible for converting molecular-level interactions into sensory umami intensity, utilizing a cross-modal perceptual attention mechanism, the formula of which is:

[0057]

[0058] Where Qmolecular is the molecular feature query vector, Ksensory is the sensory feature key vector, and Wperception is the cross-modal projection matrix.

[0059] Furthermore, the construction of the dynamic sequence length adaptive prediction layer (DSLA) includes:

[0060] Adaptive feature aggregation employs a dynamic weight aggregation strategy for peptide sequences of different lengths. The calculation formula is as follows:

[0061]

[0062] Where L is the peptide sequence length, h i Let w be the hidden state vector at position i.i Position weights;

[0063] w i The calculation formula is:

[0064]

[0065] Wherein, Lembedding is the embedding vector for the peptide sequence length, and Wlength and blength are the weight matrix and bias term, respectively;

[0066] Umami intensity grading prediction is transformed into three sub-tasks: weak umami, medium umami, and strong umami. The multi-task prediction formula is as follows:

[0067]

[0068] The calculation formulas for each subtask are as follows:

[0069]

[0070]

[0071]

[0072] Where σ is the sigmoid activation function, and FCweak, FCmedium, and FCstrong are fully connected layers for weak, medium, and strong umami flavors, respectively.

[0073] Adaptive weight fusion calculates the weights of each subtask through an attention mechanism. The weight calculation formula is as follows:

[0074]

[0075] in, Here, Wgrade is the weight vector, Wgrade is the weight matrix, and bgrade is the bias vector.

[0076] Adaptive Dropout regularization dynamically adjusts the dropout probability based on the peptide sequence length. The calculation formula is as follows:

[0077]

[0078] Optionally, the model training, validation, and performance evaluation based on the oyster umami peptide multimodal feature dataset includes:

[0079] Model training configuration: The Adam optimizer is used to optimize model parameters. The initial learning rate is set to 0.001, the training batch size is 32, and the total number of training epochs is set to 200 epochs. Each epoch contains approximately 62 batches. A learning rate decay strategy is set, multiplying the learning rate by 0.1 every 50 epochs to prevent the model from oscillating in the later stages of training. An early stopping mechanism is adopted, stopping training when the loss of the validation dataset has not improved for 10 consecutive epochs to prevent overfitting.

[0080] Composite loss function design: A weighted combination of mean squared error and classification cross-entropy is used, as shown in the formula:

[0081]

[0082]

[0083]

[0084] Where MSE is the mean squared error loss, CrossEntropy is the classification cross-entropy loss, N is the number of samples, and y i For the true score of the i-th sample, For the predicted score, c is the number of categories. The true label for sample i belonging to category j. Let λ be the corresponding predicted probability. 1 and λ 2 These represent the weights of the regression loss and the classification loss, respectively.

[0085] Model validation and tuning: Input the validation dataset into the trained ACTM model, monitor the validation loss to implement early stopping strategy and hyperparameter tuning to prevent the ACTM model from overfitting;

[0086] Model performance evaluation: Input the test dataset into the constructed ACTM model, and use the mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R²), and Pearson correlation coefficient to evaluate the model's predictive performance and obtain the final ACTM model.

[0087] Optionally, the sensory evaluation of the screened candidate peptides includes:

[0088] Sixteen professional sensory evaluators trained in umami recognition were selected, and a sensory evaluation standard of 0-9 points was used.

[0089] The selected candidate peptides were dissolved in purified water at a concentration of 1.0 mmol / L, the pH was adjusted to 7.0, and the temperature was controlled at 25℃.

[0090] Each sample was independently scored by 16 professional sensory evaluators using a blinded, randomly coded test, and the average score was taken as the final sensory score for the candidate peptide.

[0091] The umami perception concentration threshold of candidate peptides was determined by the three-point test method, with concentration gradients of 0.05, 0.10, 0.15, 0.20, and 0.25 mmol / L.

[0092] Optionally, the step of performing electroencephalographic verification on the screened candidate peptides includes:

[0093] 16-32 healthy adults aged 22-35 years were selected as subjects;

[0094] The EEG signals of the subjects tasting the peptide solution were recorded using a 32-lead EEG system. The recording parameters were set as follows: sampling frequency 1000Hz, filter frequency band 0.5-100Hz, and recording duration 30s.

[0095] The acquired EEG signals were preprocessed, including baseline correction, removal of electrooculography artifacts, frequency domain filtering, and independent component analysis.

[0096] The study focused on analyzing the changes in power spectral density in the gamma wave band and comparing the differences in gamma wave power density under candidate peptide stimulation with that of the monosodium glutamate control group. Attached Figure Description

[0097] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0098] Figure 1 This disclosure provides an overall technical roadmap for a method of screening oyster flavor peptides based on multimodal umami intensity feature recognition.

[0099] Figure 2 This is a diagram of the overall architecture of the ACTM model;

[0100] Figure 3 Structural diagram of the amino acid property-sensing attention mechanism (APAA);

[0101] Figure 4 The curve showing the change in the loss function during the training process of the ACTM model. Detailed Implementation

[0102] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0103] like Figure 1 As shown in the embodiments of this disclosure, a method for screening oyster flavor peptides based on multimodal umami intensity feature recognition is provided, including the following steps:

[0104] 1. Construct a multimodal feature dataset of oyster umami peptides

[0105] (1) 2856 peptide sequences were collected from oyster protein enzymatic hydrolysis. The peptide sequence length ranged from 2 to 20 amino acids and the molecular weight ranged from 500 to 2000 Da. These peptide sequences covered different amino acid compositions and sequence patterns, including peptide sequences rich in acidic amino acids (such as aspartic acid D and glutamic acid E), hydrophobic amino acids (such as valine V and tyrosine Y), and basic amino acids (such as arginine R and lysine K), to ensure the diversity and representativeness of the dataset.

[0106] (2) Multimodal feature extraction for each peptide sequence

[0107] ECFP4 molecular fingerprint feature extraction: The ECFP4 molecular fingerprint was calculated using the RDKit cheminformatics toolkit, with the radius parameter set to radius=2 and the fingerprint bit length nBits=1024 to characterize the topological connectivity of the molecule. The amino acid sequence was converted into a SMILES molecular structure representation using the Chem.MolFromSequence() function of the RDKit cheminformatics toolkit to establish the mapping relationship between sequence and structure. The ECFP4 fingerprint was calculated using the Morgan algorithm to generate a 1024-bit binary fingerprint vector to characterize the topological structural features of the peptide molecule.

[0108] Molecular descriptor feature calculations include: basic molecular property descriptor calculations, including physicochemical properties such as molecular weight (MW), LogP value, number of hydrogen bond donors (HBD), number of hydrogen bond acceptors (HBA), number of rotatable bonds (nRotB), and topological polar surface area (TPSA), totaling 50 features; topological index descriptor calculations, including numerical features characterizing molecular topology such as Wiener index, Zagreb index, Randic connectivity index, Kappa shape index, and Balaban index, totaling 80 features; molecular geometry descriptor calculations, including three-dimensional geometric features such as principal inertia, eccentricity, and sphericity, calculated through molecular conformation optimization, totaling 35 features; and electronic property descriptor calculations, including electronic structure features such as highest occupied molecular orbital (HOMO), lowest unoccupied molecular orbital (LUMO), and molecular dipole moment, calculated through quantum chemical methods, totaling 35 features, for a total of 200 molecular descriptors.

[0109] Molecular docking binding energy characteristic calculation: Based on the homology modeling method, a three-dimensional structural model of the T1R1 / T1R3 umami receptor was constructed using the known mGluR1 receptor crystal structure as a template. The binding free energy of oyster peptides with the T1R1 / T1R3 receptor was calculated using AutoDock Vina molecular docking software. The docking parameters were set as follows: search space size_x=20, size_y=20, size_z=20, and energy calculation precision exhaustiveness=8 to ensure the reliability of the docking results. 100 independent docking calculations were performed on each peptide-receptor complex, and the average value was taken as the final binding energy characteristic value. The binding energy value ranged from -5.0 to -10.0 kcal / mol, and the more negative the value, the more stable the binding.

[0110] The 1024-dimensional molecular fingerprint, 200-dimensional molecular descriptor, and 1-dimensional binding energy feature are concatenated to form a 1225-dimensional multimodal feature vector.

[0111] (3) The z-score standardization method is used to normalize all features to eliminate numerical differences between different features.

[0112] (4) The normalized feature data is divided into training dataset, validation dataset and test dataset in a ratio of 7:2:1, resulting in 1999 training samples, 571 validation samples and 286 test samples, ensuring that the training dataset, validation dataset and test dataset maintain a balance in peptide length distribution and amino acid composition.

[0113] 2. Constructing the ACTM model

[0114] ACTM model architecture as follows Figure 2 As shown, the details are as follows:

[0115] (1) Multimodal feature fusion layer

[0116] Amino acid property perception attention mechanism APAA architecture such as Figure 3 As shown, the feature weights are dynamically adjusted based on the physicochemical properties of each amino acid in the peptide sequence.

[0117] For each amino acid position in the peptide sequence, a parameter matrix containing the physicochemical properties of 20 standard amino acids was established. The acidity parameter matrix was set based on the pKa value of the amino acid side chain, with aspartic acid and glutamic acid having high acidity values; the hydrophobicity parameter matrix used the Kyte-Doolittle hydrophobicity index, with hydrophobic amino acids such as valine, leucine, and tyrosine having high hydrophobicity values; the basicity parameter matrix was set according to the basicity strength of the side chain, with arginine, lysine, and histidine having high basicity values; and the molecular size parameter matrix was calculated based on the molecular volume of the amino acid.

[0118] When performing APAA calculations, for a peptide sequence of length L, the amino acid type at position i is obtained by looking up the corresponding four property parameter values ​​in a table. These values ​​are then weighted and summed with four learnable weight matrices, and normalized using the softmax function to obtain the attention weight at that position. This weight reflects the importance contribution of that amino acid position to the overall umami intensity of the peptide molecule. The calculation formula is as follows:

[0119]

[0120] Where i represents the position of the i-th amino acid in the peptide sequence, Pacid(i), Phydro(i), Pbasic(i), and Psize(i) represent the acidity, hydrophobicity, basicity, and molecular size properties of the i-th amino acid, respectively, and Wacid, Whydro, Wbasic, and Wsize are the corresponding learnable weight matrices. This mechanism is mainly based on the biological basis that the chemical properties of amino acids determine protein function. The aim is to enable the model to understand the differentiated contributions of different amino acids to umami production, especially to identify the typical sequence pattern of "acid-hydrophobic-basic" umami peptides.

[0121] The receptor binding site prediction module (RBSP) is used to predict key binding sites of peptide sequences with T1R1 / T1R3 receptors. Based on literature reports and molecular docking analysis, key residues of the T1R1 receptor include ARG30, SER148, and ASP216, while key residues of the T1R3 receptor include GLU301, TRP68, and HIS145. The physicochemical properties (charge distribution, hydrophobicity, hydrogen bonding ability, etc.) of these residues are encoded into high-dimensional vectors as prior knowledge. The aim is to identify amino acid sites in the peptide sequence that form key interactions with taste receptors, providing biological prior knowledge for subsequent attention allocation. The calculation formula is as follows:

[0122]

[0123] Where ⊕ represents feature concatenation, Ffp, Fdesc, and Fdock represent ECFP4 molecular fingerprint features, molecular descriptor features, and molecular docking binding energy features, respectively, Wreceptor is the receptor binding prediction weight matrix, breceptor is the bias term, and σ is the sigmoid activation function.

[0124] By using a fully connected layer containing 512 neurons and a sigmoid activation function, a 512-dimensional receptor binding site prediction vector is output, which represents the probability distribution of each amino acid in the peptide sequence forming a key interaction with the receptor.

[0125] An adaptive weighted fusion mechanism enhanced with biological knowledge combines molecular features with biological knowledge, enabling the feature fusion process to incorporate an understanding of peptide-receptor interactions. This significantly improves the accuracy of identifying umami peptides. The fusion formula is as follows:

[0126]

[0127] Here, Ffused is a 1024-dimensional biological knowledge-enhanced feature vector, ⊙ represents element-wise multiplication, and α, β, and γ are adaptive weight parameters that are dynamically adjusted through a gating mechanism.

[0128] The amino acid property weight vector output by the APAA mechanism is multiplied element-wise with the 1024-dimensional molecular fingerprint to achieve biological knowledge-based molecular fingerprint enhancement. The binding site prediction vector output by the RBSP module is multiplied element-wise with the 200-dimensional molecular descriptor to achieve receptor binding knowledge-guided descriptor enhancement. Molecular docking binding energy is directly preserved. The three enhanced features are linearly combined using adaptive weight parameters α, β, and γ, where the weight parameters are dynamically learned through a neural network with a gating mechanism, ultimately outputting a 1024-dimensional biological knowledge-enhanced feature vector.

[0129] (2) Multi-scale convolutional attention layer

[0130] The multi-scale convolutional feature extraction module employs three parallel one-dimensional convolutional branches, using Conv1D layers with kernel sizes of 3, 5, and 7 to extract local features at different scales. Each convolutional branch contains 128 convolutional kernels, and the convolution operation formula is as follows:

[0131]

[0132] Where x represents the input feature vector, k∈{3, 5, 7} is the convolution kernel size, Wk is the weight matrix of the k-th convolution kernel, and bk is the corresponding bias vector. The formula represents the convolution operator, and ReLU is the modified linear unit activation function. The design principle of this formula is based on the multi-scale feature extraction theory. The purpose is to capture the local feature patterns of peptide sequences at different spatial scales simultaneously by using convolution kernels of different sizes. The small-scale convolution kernel (k=3) can capture the local interaction patterns between adjacent 2-4 amino acids, which is suitable for identifying the local structural features of short peptide segments. The large-scale convolution kernel (k=7) can capture long-distance amino acid correlations, which is suitable for identifying the overall conformational features of peptide molecules. The medium-scale convolution kernel (k=5) can capture medium-distance correlations within a range of 5-7 amino acids, which is suitable for identifying secondary structure fragments such as α-helices or β-sheets. The outputs of the three branches are concatenated in the feature dimension to obtain 384-dimensional multi-scale convolution features.

[0133] The channel attention mechanism is used to calculate the importance weight of each feature channel for umami prediction. The calculation formula is as follows:

[0134]

[0135] Where F represents the input feature matrix, GAP is the global average pooling operation, FC is the fully connected layer, and σ is the sigmoid activation function. The design principle of this formula is based on the attention mechanism theory. The purpose is to extract the global statistical information of each channel through global average pooling, then learn the non-linear dependency relationship between channels through the fully connected layer, and finally normalize the weight values ​​to the [0, 1] interval through the sigmoid function.

[0136] Global average pooling is performed on the 384-dimensional multi-scale features to obtain global statistical information for each feature channel. Then, a non-linear transformation is performed through a fully connected layer containing 192 neurons, followed by restoration to the original dimension through another fully connected layer containing 384 neurons. Finally, a sigmoid activation function is applied to obtain the channel weight vector in the interval [0, 1]. This weight vector represents the importance of each feature channel for umami prediction. A one-dimensional convolution operation with a kernel size of 1 is applied to the 384-dimensional multi-scale features, with the number of output channels set to 1. An attention weight is obtained for each spatial location through a sigmoid activation function. This weight vector represents the importance of each amino acid position in the peptide sequence for umami prediction.

[0137] Spatial attention mechanism, used to identify the most critical amino acid sites for umami prediction, is calculated using the following formula:

[0138]

[0139] Conv1D is a one-dimensional convolution operation. The design principle of this formula is to capture the local correlation of features at different spatial locations through convolution operations and identify the amino acid sites that are most critical for umami prediction.

[0140] Dual attention feature fusion multiplies the channel attention and spatial attention weights with the original features to obtain an attention-enhanced feature representation. The fusion formula is as follows:

[0141]

[0142] Here, ⊙ represents element-wise multiplication. The design principle of this operation is to selectively enhance the original features through dual attention weights, highlighting the feature dimensions and spatial locations that are most important for umami prediction, suppressing irrelevant information, and finally outputting a 384-dimensional multi-scale attention feature vector.

[0143] (3) Umami perception hierarchical modeling Transformer encoder UPLM

[0144] The molecular recognition layer uses the first and second layer encoders to identify key structural features of the peptide, employing a structure-aware self-attention mechanism. The calculation formula is as follows:

[0145]

[0146] Where Q is the query matrix, K is the bond matrix, WQ and WK are the corresponding weight matrices, dk is the bond vector dimension, and Bstructure is the structure bias matrix. It is constructed based on the spatial distance and chemical bond connection relationship between amino acids. The design principle of this layer is based on the spatial geometric constraints of molecular recognition, and the purpose is to enable the model to understand the influence of the three-dimensional conformation of peptide molecules on umami.

[0147] A structure bias matrix based on the spatial distance and chemical bond connections between amino acids is constructed. The element values ​​of this matrix are determined according to the average contact probability and chemical interaction strength of amino acid pairs in the protein structure. In the self-attention computation, the structure bias matrix is ​​added to the attention weight matrix, enabling the model to understand the three-dimensional spatial conformation information of the peptide molecule. Each encoder layer contains eight parallel attention heads, each with a dimension of 64, for a total model dimension of 512. Through the multi-head self-attention mechanism, the model can learn the spatial interaction relationships between amino acids from different perspectives. The feedforward neural network contains two fully connected layers with a hidden layer dimension of 2048 and the ReLU activation function.

[0148] The receptor activation layer uses the 3rd-4th layer encoder to handle the interaction between peptides and T1R1 / T1R3 receptors, employing a receptor-guided attention mechanism. The calculation formula is as follows:

[0149]

[0150] Among them, Qreceptor and Kreceptor are receptor structural information embedding vectors, which are constructed based on key amino acid residues of the T1R1 / T1R3 receptor.

[0151] A receptor structure embedding vector is constructed based on key amino acid residue information of the T1R1 / T1R3 receptor. This vector contains information such as the geometry, charge distribution, and hydrophobicity characteristics of the receptor binding site. During attention computation, the receptor structure information is added to the query matrix and bond matrix to guide the model in learning the specific binding patterns between peptide molecules and receptors. Through the receptor-guided attention mechanism, the model can identify the most critical amino acid positions in the peptide sequence for receptor binding and learn the interaction patterns between these positions and the receptor binding site.

[0152] The perceptual coding layer employs encoders layers 5 and 6 specifically responsible for converting molecular-level interactions into sensory umami intensity, utilizing a cross-modal perceptual attention mechanism, the formula of which is:

[0153]

[0154] Among them, Qmolecular is the molecular feature query vector, Ksensory is the sensory feature key vector, and Wperception is the cross-modal projection matrix. The design principle of this layer is based on the perceptual coding theory in neuroscience. The purpose is to establish a direct mapping relationship between molecular features and sensory scores, so as to realize the transformation from microscopic molecular information to macroscopic perceptual experience.

[0155] A cross-modal projection matrix is ​​constructed to map molecular-level feature information to the sensory perception level. The molecular feature query vector contains structural and receptor-binding information of peptide molecules, while the sensory feature bond vector contains physiological characteristics of human taste perception, such as the activation intensity of taste receptors, neural signal transduction efficiency, and cortical response patterns. Through a cross-modal attention mechanism, the model establishes a direct mapping relationship from microscopic molecular interactions to macroscopic sensory experiences, realizing the conversion of molecular features into sensory ratings.

[0156] Through the three-level UPLM mechanism, the 384-dimensional multi-scale attention features are gradually transformed in three levels: molecular recognition, receptor activation, and sensory encoding, ultimately resulting in a 512-dimensional sensory encoding vector that integrates information from three levels: molecular structure, receptor interaction, and sensory perception.

[0157] (4) Dynamic Sequence Length Adaptive Prediction Layer DSLA

[0158] Adaptive feature aggregation employs a dynamic weight aggregation strategy for peptide sequences of different lengths. The calculation formula is as follows:

[0159]

[0160] Where L is the peptide sequence length, h i Let w be the hidden state vector at position i. i For positional weights.

[0161] w i The calculation formula is:

[0162]

[0163] In this model, Lembedding is the embedding vector for peptide sequence length, and Wlength and blength are the weight matrix and bias term, respectively. The design principle of this mechanism is based on the biological fact that peptide sequences of different lengths have different structure-activity relationships. The aim is to enable the model to adaptively handle peptide sequences in the length range of 2-20 amino acids and avoid the influence of peptide sequence length bias on the prediction results.

[0164] For a peptide sequence of length L, the sequence length information is encoded as a 64-dimensional embedding vector, which is concatenated with the 512-dimensional hidden state vector at each position to obtain a 576-dimensional positional feature vector. An importance weight for each position is calculated using a fully connected layer with 576 input neurons and 1 output neuron, and then normalized using softmax to ensure that all positional weights sum to 1. The hidden state vectors at each position are then weighted and summed using these positional weights to obtain a fixed-dimensional sequence representation vector. This adaptive aggregation method allows the model to dynamically adjust the feature fusion strategy according to the actual length of the peptide sequence, avoiding the information asymmetry problem between long and short sequences.

[0165] Umami intensity grading prediction is divided into three sub-tasks: weak umami (0-3 points), medium umami (3-6 points), and strong umami (6-9 points). The multi-task prediction formula is as follows:

[0166]

[0167] The calculation formulas for each subtask are as follows:

[0168]

[0169]

[0170]

[0171] Where σ is the sigmoid activation function, and FCweak, FCmedium, and FCstrong are fully connected layers for weak, medium, and strong umami, respectively. The design principle of this multi-task prediction strategy is based on the hierarchical characteristics of human taste perception. The aim is to improve the model's discrimination accuracy for different umami intensity ranges and avoid the ambiguity problem in the boundary region of traditional single regression prediction.

[0172] Adaptive weight fusion calculates the weights of the three sub-tasks using an attention mechanism. The weight calculation formula is as follows:

[0173]

[0174] in, Wgrade is the weight vector, bgrade is the weight matrix, and bgrade is the bias vector. The design principle of this mechanism is to dynamically adjust the contribution weights of different umami levels based on the input features.

[0175] The 512-dimensional aggregated feature vector is passed through a fully connected layer containing three output neurons and then activated by softmax to obtain three weight parameters, ensuring that the sum of the weights is 1. The final umami intensity score is obtained by weighted averaging, which maintains the continuity of prediction and improves the discrimination accuracy of different umami levels.

[0176] Adaptive Dropout regularization dynamically adjusts the dropout probability based on the peptide sequence length. The calculation formula is as follows:

[0177]

[0178] The principle of this design is that short peptide sequences (L=2-5) are relatively simple and are discarded with a lower probability (0.1-0.15) to retain more information for prediction, while long peptide sequences (L=15-20) are complex and are discarded with a higher probability (0.25-0.3) to enhance the regularization effect and prevent overfitting. This adaptive mechanism is consistent with the complexity differences of peptide sequences of different lengths.

[0179] 3. Model training, validation, and performance evaluation based on the oyster umami peptide multimodal feature dataset.

[0180] Model training configuration: The Adam optimizer is used to optimize model parameters. The initial learning rate is set to 0.001, the training batch size is 32, and the total number of training epochs is set to 200 epochs. Each epoch contains approximately 62 batches. A learning rate decay strategy is set, multiplying the learning rate by 0.1 every 50 epochs to prevent the model from oscillating in the later stages of training. An early stopping mechanism is adopted, stopping training when the loss of the validation dataset has not improved for 10 consecutive epochs to prevent overfitting.

[0181] Composite loss function design: A weighted combination of mean squared error and classification cross-entropy is used, as shown in the formula:

[0182]

[0183]

[0184]

[0185] Where MSE is the mean squared error loss, CrossEntropy is the classification cross-entropy loss, N is the number of samples, and y i For the true score of the i-th sample, For the predicted score, c is the number of categories. The true label for sample i belonging to category j. Let λ be the corresponding predicted probability. 1 and λ 2 These represent the weights of the regression loss and classification loss, respectively. This composite loss function is based on multi-task learning theory and aims to simultaneously optimize two tasks: continuous value prediction and discrete class classification.

[0186] MSE loss is used to optimize the prediction accuracy of continuous values, while classification cross-entropy loss is used to enhance the model's ability to distinguish between different umami levels. The weight parameters λ1 and λ2 are set to 0.7 and 0.3 respectively, so that the model focuses more on regression accuracy while maintaining sensitivity to class boundaries. The changes of the two loss components are dynamically monitored during training to ensure the balance of training.

[0187] Model validation and tuning: After each training epoch, model performance was evaluated using 571 validation samples. Figure 4 This paper demonstrates the changes in the loss function during the training process of the ACTM model, calculates metrics such as MAE, RMSE, and R² on the validation dataset, and compares them with the metrics on the training dataset to monitor the model's generalization ability. When the performance on the validation dataset begins to decline, an early stopping mechanism is triggered, and the current optimal model parameters are saved. A grid search method is used to fine-tune key hyperparameters, including learning rate (0.001, 0.0001), batch size (16, 32, 64), and Dropout probability (0.1, 0.3, 0.5). The performance of different hyperparameter combinations is evaluated using 5-fold cross-validation, and the parameter configuration with the best validation performance is selected.

[0188] Model Performance Evaluation: The final model was evaluated using 286 test samples. The test dataset was completely independent throughout the training process, ensuring the objectivity of the evaluation results. The mean absolute error (MAE) was 0.26, the root mean square error (RMSE) was 0.35, the coefficient of determination (R²) was 0.92, and the Pearson correlation coefficient was 0.96, indicating that the model has excellent predictive performance. Residual analysis revealed that the model's prediction error was evenly distributed across different umami intensity ranges, without significant systematic bias. For high-umami peptides (score ≥ 7), the prediction accuracy was slightly higher than that for low-umami peptides, meeting the needs of practical applications.

[0189] 4. Input the oyster peptide sequences to be screened into the trained ACTM model for umami intensity feature recognition.

[0190] The trained ACTM model is systematically compared and validated with existing machine learning methods to verify the model's technological advancement.

[0191] Five representative machine learning methods were selected: Support Vector Machine (SVM) using radial basis function kernels, with hyperparameters optimized via grid search; Random Forest (RF) using 100 decision trees with a maximum depth of 10; Gradient Boosting Decision Tree (GBDT) using 100 weak learners with a learning rate of 0.1; Artificial Neural Network (ANN) containing two hidden layers with 256 neurons each; and Long Short-Term Memory Network (LSTM) containing 128 hidden units for processing sequence information. All comparison methods used the same multimodal feature data (molecular fingerprint + molecular descriptor + binding energy), the same dataset partitioning, and the same evaluation metrics to ensure fairness in the comparison. The performance of each model on the test dataset is shown in Table 1.

[0192] Table 1 Model Performance Comparison

[0193] Model MAE RMSE R2 Correlation coefficient ACTM model 0.26 0.35 0.92 0.96 LSTM model 0.32 0.41 0.87 0.93 Basic Neural Networks 0.41 0.55 0.81 0.90 Gradient boosting decision tree 0.49 0.63 0.71 0.84 Random Forest 0.52 0.68 0.69 0.83 Support Vector Machine 0.47 0.61 0.73 0.85

[0194] The ACTM model significantly outperforms the comparative model on all evaluation metrics. Compared with the best comparative method (LSTM), the MAE is improved by 18.8, the RMSE by 14.6, and the R² by 5.7, demonstrating the technical superiority of the model of this invention.

[0195] To verify the contribution of each innovative technology module, ablation experiments were conducted. After removing the APAA mechanism, the model's MAE increased to 0.34 (an increase of 30.8); after removing the UPLM mechanism, the MAE increased to 0.38 (an increase of 46.2); after removing the DSLA mechanism, the MAE increased to 0.31 (an increase of 19.2); and after removing the RBSP module, the MAE increased to 0.29 (an increase of 11.5). The ablation experiment results indicate that the UPLM mechanism contributes the most to the model's performance, followed by the APAA mechanism.

[0196] The trained ACTM model was used to screen oyster umami peptides. Multimodal features of 2856 oyster peptide sequences were input into the trained ACTM model for forward inference calculation. The model output a predicted umami intensity score for each peptide sequence, ranging from 0 to 9. A screening threshold of 7.5 was set, resulting in 157 high-umami candidate peptides with predicted scores ≥7.5, representing 5.5% of the total dataset. Sequence analysis of these 157 candidate peptides revealed that 89% contained acidic amino acids such as aspartic acid (D) or glutamic acid (E), and 76% conformed to an "acid-hydrophobic-basic" sequence pattern, validating the effectiveness of the APAA mechanism in identifying typical characteristics of umami peptides. The k-means clustering algorithm was used to cluster the 157 candidate peptides based on features such as amino acid composition, sequence length, and hydrophobicity index. With a cluster size of 10, 10 peptide groups with different characteristics were obtained. One representative peptide sequence was selected from each cluster center, ultimately identifying 10 candidate peptides for experimental validation. These included DEGVYR, EEAMDR, DEGPYR, DVQMPR, EVQAHR, DDGLPR, AEGPYR, DEGLYR, VEGPYR, and RDEGVYR. These 10 peptide sequences covered different length ranges (5-7 amino acids), different amino acid compositions, and different physicochemical properties, ensuring the representativeness and comprehensiveness of the experimental validation.

[0197] 5. The selected candidate peptides were chemically synthesized, subjected to sensory evaluation, validated by electroencephalography, and their molecular mechanisms were elucidated.

[0198] (1) Chemical synthesis of the screened candidate peptides

[0199] Ten candidate peptides were synthesized using a solid-phase peptide synthesis method. Amino acid raw materials with Fmoc protecting groups were used for synthesis on an automated peptide synthesizer. After synthesis, the peptides were deprotected and cleaved from the resin using trifluoroacetic acid (TFA). The crude peptide products were purified by reversed-phase high-performance liquid chromatography (RP-HPLC) to a purity of over 95%. The molecular weight and structural correctness of the peptides were verified using mass spectrometry.

[0200] (2) Sensory evaluation of the screened candidate peptides, including:

[0201] Sixteen professional sensory evaluators, aged 25-45, with no history of smoking and no taste disorders, were selected to ensure the accuracy and repeatability of the sensory evaluation results.

[0202] A sensory evaluation standard of 0-9 points was adopted, with 0 points indicating no umami, 5 points indicating the umami intensity of the monosodium glutamate reference standard, and 9 points indicating strong umami, to establish a unified evaluation system.

[0203] Ten candidate peptides were dissolved in purified water at a concentration of 1.0 mmol / L, the pH was adjusted to 7.0, and the temperature was controlled at 25℃ to ensure the standardization of the test conditions.

[0204] Each sample was independently scored by 16 professional sensory evaluators using a blinded, randomly coded method. The average value was taken as the final sensory score of the candidate peptides. Results with a standard deviation of less than 0.5 were considered valid. The sensory scores of the candidate peptides are shown in Table 2. As can be seen from Table 2, the DEGVYR peptide obtained the highest sensory score of 8.2±0.3, which was significantly higher than other candidate peptides (P<0.01), and was highly consistent with the model prediction results.

[0205] Table 1 Sensory scores of candidate peptide sequences

[0206] peptide sequence Umami intensity Umami duration Salty bitterness sweet Overall score DEGVYR 9.5±0.1 9.2±0.2 6±0.2 1±0.2 1.5±0.1 8.2 ± 0.3 DEGPYR 8.6±0.3 8.5±0.3 5.5±0.3 1.5±0.3 1.5±0.3 6.8 ± 0.4 AEGPYR 8.5±0.2 8.3±0.4 5±0.2 1.8±0.1 1.8±0.2 5.0 ± 0.2 VEGPYR 7.8±0.3 7.5±0.2 4.5±0.3 2.5±0.3 1.5±0.2 5.5 ± 0.4 EEAMDR 7.5±0.4 7±0.4 7.5±0.1 2±0.2 2±0.2 7.1 ± 0.5 DEGLYR 7±0.2 6.8±0.3 4±0.4 5.5±0.2 1.5±0.2 6.0 ± 0.5 DDGLPR 6.5±0.2 6.2±0.4 7±0.2 4±0.4 2±0.2 7.0 ± 0.3 RDEGVYR 6±0.3 5.5±0.3 3±0.3 8.5±0.2 1±0.3 6.8 ± 0.5 DVQMPR 5.8±0.4 5±0.2 3.5±0.4 7.5±0.3 1.5±0.2 6.3 ± 0.2 EVQAHR 5±0.3 4.5±0.4 2±0.2 9.5±0.4 1±0.2 6.0 ± 0.4

[0207] The umami perception concentration threshold of the candidate peptides was determined by the three-point test method. The concentration gradient was set to 0.05, 0.10, 0.15, 0.20, and 0.25 mmol / L. The umami perception concentration thresholds of the candidate peptides are shown in Table 3. As can be seen from Table 3, the umami perception concentration threshold of DEGVYR peptide is 0.15 mmol / L, which is significantly lower than that of monosodium glutamate (0.3 mol / mL), indicating that this peptide has a stronger umami perception intensity.

[0208] Table 3. Umami Threshold and Gamma Wave Power Spectral Density of Candidate Peptides

[0209] peptide sequence Umami perception concentration threshold (mol / mL) DEGVYR 0.15 EEAMDR 0.45 EVQAHR 1.2 DVQMPR 0.8 DEGPYR 0.25 DDGLPR 0.65 AEGPYR 0.28 VEGPYR 0.35 DEGLYR 0.55 RDEGVYR 0.7

[0210] (3) Perform electroencephalographic validation on the screened candidate peptides, including:

[0211] Thirty-two healthy adults aged 22-35 years were selected as subjects. They had no neurological diseases, did not take any drugs that affect nerve activity, and fasted for 12 hours before the experiment to avoid taste interference and ensure the reliability of EEG signals.

[0212] The EEG signals of the subjects tasting the peptide solution were recorded using a 32-lead EEG system. The recording parameters were set as follows: sampling frequency 1000Hz, filter frequency band 0.5-100Hz, and recording duration 30s.

[0213] The acquired EEG signals were preprocessed, including baseline correction, removal of electrooculogram artifacts, frequency domain filtering, and independent component analysis, to remove interference signals and improve data quality.

[0214] The power spectral density changes in the γ-wave band were analyzed in detail. The power spectrum was calculated using fast Fourier transform, and the integral value of the power spectral density in this band was calculated. The power spectral density of the candidate peptide in the γ-wave band is shown in Table 4. As can be seen from Table 4, the power spectral density of the γ-wave band stimulated by DEGVYR peptide is 15.2±2.1µV² / Hz, which is 17.8% higher than that of the sodium glutamate control group (12.9±1.5µV² / Hz) (P<0.05), indicating that the peptide can induce a stronger brain umami perception response.

[0215] Table 4 Power spectral density of candidate peptides in the γ-frequency band

[0216] peptide sequence γ-wave power spectral density (μV² / Hz) DEGVYR 15.2±2.1 EEAMDR 4.4±0.5 EVQAHR 8.1±0.7 DVQMPR 3.6±0.5 DEGPYR 2.3±0.5 DDGLPR 5.4±0.8 AEGPYR 7.5±0.8 VEGPYR 4.2±0.8 DEGLYR 7.4±0.8 RDEGVYR 3.5±0.8

[0217] (4) Molecular mechanism analysis of the screened candidate peptides.

[0218] The core of the flavor-presenting mechanism of umami peptides lies in their specific binding to the umami receptors T1R1 / T1R3 on the taste buds of the tongue. The activation of these receptors follows the "molecular clamp" model, in which the umami molecule acts as a bridge, binding to both the T1R1 and T1R3 subunits simultaneously, stabilizing their active conformation, thereby triggering neural signals.

[0219] The activity of a peptide is directly determined by its amino acid sequence and structural features. For example, the N-terminus of the DEGVYR peptide, with aspartic acid and glutamate providing negative charges and the C-terminus with arginine providing positive charges, forms a zwitterionic structure that can efficiently bind to the positive and negative charged regions of the receptor. Valine and tyrosine in the sequence optimize spatial matching through hydrophobic interactions; this multi-site synergistic effect results in high sensory scores and strong EEG responses.

[0220] Amino acid substitutions can significantly affect activity. For example, replacing valine with proline in the DEGVYR peptide slightly reduces binding efficiency due to changes in skeletal rigidity. Replacing the critical N-terminal aspartic acid with alanine or valine directly weakens key electrostatic interactions, leading to activity degradation. Replacing glycine with bulky leucine creates steric hindrance, severely interfering with its binding to the receptor.

[0221] Peptides with low sequence similarity or that have been elongated lose activity primarily because they cannot form the correct interactions. For example, adding arginine to the N-terminus of the DEGVYR peptide disrupts the original charge balance and spatial orientation. Peptides containing atypical amino acids such as glutamine and histidine may have structures that better match the binding requirements of bitter taste receptors, or may be completely unable to effectively bind to umami receptors, thus exhibiting weak umami activity.

[0222] The above embodiments demonstrate that the oyster flavor peptide screening method based on multimodal umami intensity feature recognition proposed in this invention has significant advantages such as high prediction accuracy, high screening efficiency, and strong biological mechanism explanatory power, providing important technical support and theoretical guidance for high-throughput screening of marine bioactive peptides and the development of functional foods.

[0223] Although the present invention has been disclosed above with reference to embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for screening oyster flavor peptides based on multimodal umami intensity feature recognition, characterized in that, Includes the following steps: Construct a multimodal feature dataset of oyster umami peptides; Constructing the ACTM model; Model training, validation, and performance evaluation were conducted based on a multimodal feature dataset of oyster umami peptides. The oyster peptide sequences to be screened are input into the trained ACTM model for umami intensity feature recognition to obtain candidate peptides; The selected candidate peptides were chemically synthesized, sensory evaluated, validated by electroencephalography, and their molecular mechanisms were elucidated. The method for constructing the ACTM model is as follows: A multimodal feature fusion layer integrating the amino acid property sensing attention mechanism (APAA) and the receptor binding site prediction module (RBSP) was constructed to perform feature fusion on the multimodal features of oyster umami peptides. A multi-scale convolutional attention layer is constructed to perform local feature extraction and feature enhancement on the features output by the multi-modal feature fusion layer; Constructing a hierarchical model of umami perception using a Transformer encoder UPLM to simulate the complete biological process of features from molecular recognition to sensory perception; A dynamic sequence length adaptive prediction layer (DSLA) is constructed to predict oyster peptide sequences of different lengths using adaptive feature aggregation and umami intensity grading. The multimodal feature fusion layer constructed by integrating the amino acid property sensing attention mechanism (APAA) and the receptor binding site prediction module (RBSP) includes: The amino acid property-sensing attention mechanism (APAA) dynamically adjusts feature weights based on the physicochemical properties of each amino acid in the peptide sequence. The calculation formula is as follows: Where i represents the position of the i-th amino acid in the peptide sequence, and Pacid(i), Phydro(i), Pbasic(i), and Psize(i) represent the acidity, hydrophobicity, basicity, and molecular size properties of the i-th amino acid, respectively. The receptor binding site prediction module (RBSP) is used to predict key binding sites of peptide sequences with the T1R1 / T1R3 receptor. The calculation formula is as follows: Where ⊕ represents feature concatenation, Ffp, Fdesc, and Fdock represent ECFP4 molecular fingerprint features, molecular descriptor features, and molecular docking binding energy features, respectively, Wreceptor is the receptor binding prediction weight matrix, breceptor is the bias term, and σ is the sigmoid activation function. An adaptive weighted fusion mechanism enhanced with biological knowledge is used to combine molecular features with biological knowledge. The fusion formula is as follows: Where Ffused is a 1024-dimensional biological knowledge-enhanced feature vector, ⊙ represents element-wise multiplication, and α, β, and γ are adaptive weight parameters. The construction of the multi-scale convolutional attention layer includes: The multi-scale convolutional feature extraction module employs three parallel one-dimensional convolutional branches, using Conv1D layers with kernel sizes of 3, 5, and 7 to extract local features at different scales. Each convolutional branch contains 128 convolutional kernels, and the convolution operation formula is as follows: Where x represents the input feature vector, k∈{3, 5, 7} is the convolution kernel size, Wk is the weight matrix of the k-th convolution kernel, and bk is the corresponding bias vector. Represents the convolution operator; ReLU is the modified linear unit activation function. The channel attention mechanism is used to calculate the importance weight of each feature channel for umami prediction. The calculation formula is as follows: Where F represents the input feature matrix, GAP is the global average pooling operation, FC is the fully connected layer, and σ is the sigmoid activation function; Spatial attention mechanism, used to identify the most critical amino acid sites for umami prediction, is calculated using the following formula: Wherein, Conv1D is a one-dimensional convolution operation; Dual attention feature fusion multiplies the channel attention and spatial attention weights with the original features to obtain an attention-enhanced feature representation. The fusion formula is as follows: Where ⊙ represents element-wise multiplication; The construction of the umami perception hierarchical modeling Transformer encoder UPLM includes: The molecular recognition layer uses the first and second layer encoders to identify key structural features of the peptide, employing a structure-aware self-attention mechanism. The calculation formula is as follows: Where Q is the query matrix, K is the bond matrix, WQ and WK are the corresponding weight matrices, dk is the bond vector dimension, and Bstructure is the structure bias matrix, which is constructed based on the spatial distance and chemical bond connection relationship between amino acids. The receptor activation layer uses the 3rd-4th layer encoder to handle the interaction between peptides and T1R1 / T1R3 receptors, employing a receptor-guided attention mechanism. The calculation formula is as follows: Among them, Qreceptor and Kreceptor are receptor structural information embedding vectors, which are constructed based on key amino acid residues of the T1R1 / T1R3 receptor; The perceptual coding layer employs encoders layers 5 and 6 specifically responsible for converting molecular-level interactions into sensory umami intensity, utilizing a cross-modal perceptual attention mechanism, the formula of which is: Where Qmolecular is the molecular feature query vector, Ksensory is the sensory feature key vector, and Wperception is the cross-modal projection matrix; The construction of the dynamic sequence length adaptive prediction layer DSLA includes: Adaptive feature aggregation employs a dynamic weight aggregation strategy for peptide sequences of different lengths. The calculation formula is as follows: Where L is the peptide sequence length, h i Let w be the hidden state vector at position i. i Position weights; w i The calculation formula is: Wherein, Lembedding is the embedding vector for the peptide sequence length, and Wlength and blength are the weight matrix and bias term, respectively; Umami intensity grading prediction is transformed into three sub-tasks: weak umami, medium umami, and strong umami. The multi-task prediction formula is as follows: The calculation formulas for each subtask are as follows: Where σ is the sigmoid activation function, and FCweak, FCmedium, and FCstrong are fully connected layers for weak, medium, and strong umami flavors, respectively. Adaptive weight fusion calculates the weights of each subtask through an attention mechanism. The weight calculation formula is as follows: in, Here, Wgrade is the weight vector, Wgrade is the weight matrix, and bgrade is the bias vector. Adaptive Dropout regularization dynamically adjusts the dropout probability based on the peptide sequence length. The calculation formula is as follows:

2. The method for screening oyster flavor peptides based on multimodal umami intensity feature recognition according to claim 1, characterized in that, The construction of the oyster umami peptide multimodal feature dataset includes: 2856 peptide sequences were collected from the enzymatic digestion of oyster proteins; Multimodal feature extraction was performed on each peptide sequence, including ECFP4 molecular fingerprint features, molecular descriptor features and molecular docking binding energy features, forming a comprehensive feature vector of 1024+200+1 dimensions. The z-score normalization method is used to normalize all features and eliminate numerical differences between different features; The normalized feature data were divided into training, validation, and test datasets in a ratio of 7:2:

1.

3. The method for screening oyster flavor peptides based on multimodal umami intensity feature recognition according to claim 2, characterized in that, The multimodal feature extraction for each peptide sequence includes: ECFP4 molecular fingerprint feature extraction: The ECFP4 molecular fingerprint was calculated using the RDKit cheminformatics toolkit, with the radius parameter set to radius=2 and the fingerprint bit length nBits=1024. The amino acid sequence was converted into a SMILES molecular structure representation using the Chem.MolFromSequence() function of the RDKit cheminformatics toolkit. The ECFP4 fingerprint was calculated using the Morgan algorithm to generate a 1024-bit binary fingerprint vector, which was used to characterize the topological structure features of the peptide molecule. Molecular descriptor feature calculation: Calculate 50-dimensional basic molecular property descriptors, 80-dimensional molecular topological index descriptors, 35-dimensional molecular geometric descriptors, and 35-dimensional electronic property descriptors; Molecular docking binding energy characteristic calculation: Based on the homology modeling method, the known crystal structure of the mGluR1 receptor was used as a template to construct a three-dimensional structural model of the T1R1 / T1R3 umami receptor; the binding free energy of oyster peptide with T1R1 / T1R3 receptor was calculated using AutoDock Vina molecular docking software; 100 independent docking calculations were performed on each peptide-receptor complex, and the average value was taken as the final binding energy characteristic value.

4. The method for screening oyster flavor peptides based on multimodal umami intensity feature recognition according to claim 1, characterized in that, The model training, validation, and performance evaluation based on the oyster umami peptide multimodal feature dataset include: Model training configuration: The Adam optimizer is used to optimize model parameters. The initial learning rate is set to 0.001, the training batch size is 32, and the total number of training epochs is set to 200 epochs. Each epoch contains approximately 62 batches. A learning rate decay strategy is set, multiplying the learning rate by 0.1 every 50 epochs to prevent the model from oscillating in the later stages of training. An early stopping mechanism is adopted, stopping training when the loss of the validation dataset has not improved for 10 consecutive epochs to prevent overfitting. Composite loss function design: A weighted combination of mean squared error and classification cross-entropy is used, as shown in the formula: Where MSE is the mean squared error loss, CrossEntropy is the classification cross-entropy loss, N is the number of samples, and y i For the true score of the i-th sample, For the predicted score, c is the number of categories. The true label for sample i belonging to category j. Let λ be the corresponding predicted probability. 1 and λ 2 These represent the weights of the regression loss and the classification loss, respectively. Model validation and tuning: Input the validation dataset into the trained ACTM model, monitor the validation loss to implement early stopping strategy and hyperparameter tuning to prevent the ACTM model from overfitting; Model performance evaluation: Input the test dataset into the constructed ACTM model, and use the mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R²), and Pearson correlation coefficient to evaluate the model's predictive performance and obtain the final ACTM model.

5. The method for screening oyster flavor peptides based on multimodal umami intensity feature recognition according to claim 1, characterized in that, The sensory evaluation of the screened candidate peptides includes: Sixteen professional sensory evaluators trained in umami recognition were selected, and a sensory evaluation standard of 0-9 points was used. The selected candidate peptides were dissolved in purified water at a concentration of 1.0 mmol / L, the pH was adjusted to 7.0, and the temperature was controlled at 25℃. Each sample was independently scored by 16 professional sensory evaluators using a blinded, randomly coded test, and the average score was taken as the final sensory score for the candidate peptide. The umami perception concentration threshold of candidate peptides was determined by the three-point test method, with concentration gradients of 0.05, 0.10, 0.15, 0.20, and 0.25 mmol / L.

6. The method for screening oyster flavor peptides based on multimodal umami intensity feature recognition according to claim 1, characterized in that, The electroencephalometric verification of the screened candidate peptides includes: 16-32 healthy adults aged 22-35 years were selected as subjects; The EEG signals of the subjects tasting the peptide solution were recorded using a 32-lead EEG system. The recording parameters were set as follows: sampling frequency 1000Hz, filter frequency band 0.5-100Hz, and recording duration 30s. The acquired EEG signals were preprocessed, including baseline correction, removal of electrooculography artifacts, frequency domain filtering, and independent component analysis. The study focused on analyzing the changes in power spectral density in the gamma wave band and comparing the differences in gamma wave power density under candidate peptide stimulation with that of the monosodium glutamate control group.

Images