Multi-modal umami peptide recognition method and system based on molecular graph and sequence characteristics

Abstract

This invention discloses a multimodal umami peptide identification method and system based on molecular graphs and sequence features, belonging to the interdisciplinary field of food bioinformatics and artificial intelligence. This invention extracts amino acid sequence features of peptides through a protein pre-trained language model, and simultaneously constructs a two-dimensional molecular graph based on the peptide's SMILES characterization to extract atomic-level structural features. Standard one-dimensional convolutional layers and dilated convolutional layers in a sequence encoder are used to capture local features and long-range dependency features, respectively, and atomic local structural features are extracted through a two-dimensional convolutional layer in a structural encoder. Subsequently, the extracted three types of features are pooled and concatenated, and an enhanced residual module is introduced to discriminate the peptide identification results. This invention overcomes the limitations of traditional methods that rely on single sequence features, and significantly improves the accuracy and interpretability of umami peptide identification by deeply fusing sequence semantics and molecular structural information.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of food bioinformatics and artificial intelligence, specifically relating to a method and system for identifying multimodal umami peptides based on molecular graphs and sequence features. Background Technology

[0002] Traditional methods for screening umami peptides, which rely on enzymatic hydrolysis and sensory evaluation, are time-consuming, costly, and have low throughput, making them unsuitable for large-scale functional peptide resource mining. Therefore, developing efficient and accurate computational prediction models to achieve rapid and intelligent identification of umami peptides has become a research hotspot in the field of food bioinformatics.

[0003] In recent years, machine learning and deep learning technologies have been widely applied to the functional prediction of bioactive peptides. Early studies mainly employed traditional classification algorithms such as Support Vector Machines (SVM) and Random Forests (RF), combined with handcrafted features such as amino acid composition (AAC), dipeptide composition (DPC), and physicochemical properties (e.g., hydrophobicity, charge, molecular weight) for modeling. While these methods achieved some success on small datasets, feature engineering relied on expert experience, had limited expressive power, and struggled to capture complex patterns and long-range dependencies in sequences. To further improve prediction performance, researchers began introducing deep neural networks, such as Convolutional Neural Networks (CNNs) for extracting local sequence patterns, and Recurrent Neural Networks (RNNs) and their variants LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Units) for modeling sequence order information. These methods have improved the accuracy of umami peptide identification to some extent, but remain limited to single-modality sequence analysis; single sequence features cannot comprehensively characterize the peptide prediction mechanism.

[0004] Peptides are essentially proteins (short chain form) based on their amino acid sequences; however, protein function is closely related to its structure. In umami peptide recognition research, the lack of structural information makes it difficult for models to fully capture the molecular mechanisms related to umami. Traditional umami peptide recognition methods rely heavily on sequence features such as amino acid composition and sequence motifs, neglecting the crucial influence of peptide molecular structure (such as interatomic interactions, chemical bond distribution, and local chemical environment) on umami expression. For example, the carboxyl structure of aspartic acid (D) and glutamic acid (E), as well as the polarity / hydrophobicity of amino acid side chains, can all directly relate to the intensity of umami perception by affecting the binding affinity and hydrogen bond formation ability of umami peptides to T1R1 / T1R3 umami receptors in the human body.

[0005] In summary, existing methods for identifying umami peptides have certain limitations. Single feature extraction methods are insufficient to obtain richer information; furthermore, most peptide identification models utilize three-dimensional structural information, but umami peptides are short peptides, making it difficult to accurately predict their three-dimensional structure. The lack of targeted design for short peptides significantly limits the performance and interpretability of these models. Summary of the Invention

[0006] To overcome the shortcomings of existing methods, this invention provides a multimodal umami peptide identification method and system based on molecular graphs and sequence features. This invention addresses the problem of some models relying on single feature extraction methods and resulting in insufficiently comprehensive and detailed feature representation by employing a multimodal feature fusion strategy. It innovatively utilizes the fusion of SMILES molecular graphs and sequence features to enrich the representation information. Furthermore, it employs gated linear units with added Gaussian error linear unit activation functions for feature fusion, achieving high-precision and interpretable prediction of umami activity.

[0007] This invention provides a method for identifying multimodal umami peptides based on molecular graphs and sequence features, comprising the following steps:

[0008] Step S1: Extract amino acid sequence features of peptides using a protein pre-trained language model; construct a two-dimensional molecular map based on the SMILES characterization of peptides; and construct atomic-level structural features based on the two-dimensional molecular map.

[0009] Step S2: Input the amino acid sequence features into a sequence encoder for processing to obtain splicing features;

[0010] Step S3: Input the atomic-level structural features into the structural encoder for processing to obtain adaptive average pooling features;

[0011] Step S4: Concatenate the spliced ​​features with the adaptive average pooling features to obtain combined features;

[0012] Step S5: Input the combined features into the enhanced residual module to determine whether it is an umami peptide.

[0013] Furthermore, in step S1, the protein pre-training language model is the esm2_t33_650M_UR50D model from the ESM-2 series.

[0014] Furthermore, in step S1, the amino acid sequence feature is obtained by inputting the amino acid sequence of the peptide into a protein pre-trained language model to obtain a 1280-dimensional vector representation of each amino acid residue in the last hidden state of the peptide's amino acid sequence.

[0015] Furthermore, the step of constructing atomic-level structural features in step S1 includes:

[0016] A two-dimensional molecular map is generated based on the amino acid sequence of peptides characterized by SMILES, where each amino acid corresponds to a single atom and the edges correspond to the chemical bonds between atoms.

[0017] For each atom, 21-dimensional comprehensive features are extracted as node features, specifically covering atom type, degree, total number of hydrogen atoms, implicit valence state, aromaticity, ring size, hybridization type, and Gasteg charge;

[0018] Construct atomic-level structural features.

[0019] Furthermore, in step S2, the sequence encoder includes a standard one-dimensional convolutional channel and a dilated convolutional channel; within the standard one-dimensional convolutional channel, a standard one-dimensional convolutional layer is used to extract local features in the amino acid sequence, and the local features are max-pooled to obtain max-pooled features.

[0020] Within the dilated convolutional channel, long-range dependency features in the amino acid sequence are extracted using a dilated convolutional layer; the long-range dependency features are then subjected to adaptive max pooling to obtain adaptive max pooling features; the max pooling features and the adaptive max pooling features are then concatenated to obtain concatenated features.

[0021] Furthermore, in step S3, the structure encoder includes a two-dimensional convolutional channel. Within the two-dimensional convolutional channel, the atomic-level structural features are further optimized through multiple two-dimensional convolutional layers to obtain atomic local structural features. The atomic local structural features are then subjected to adaptive average pooling to obtain adaptive average pooling features.

[0022] Furthermore, the two-dimensional convolutional layer contains at least two different convolutional kernel sizes. The outputs at each scale are adaptively averaged and concatenated along the channel dimension before being passed through a linear layer to obtain a structure vector.

[0023] Furthermore, during the extraction of atomic-level structural features in step S1, the atomic set of each residue is zero-filled according to the preset maximum number of atoms M, and a tensor of size L×M×21 is constructed, where L is the number of residues and 21 is the atomic attribute dimension.

[0024] Furthermore, in step S5, the enhanced residual module includes a gated enhanced fusion unit, layer normalization, and residual connection. The input features of the enhanced residual module pass through layer normalization and the gated enhanced fusion unit in one path, and through layer normalization in the other. The two features are then connected via residual connection to obtain the output features. The calculation process of the gated enhanced fusion unit is expressed by the following formula:

[0025]

[0026] in, Let W be the activation function of the gated augmented fusion unit, x be the input feature, and W be the activation function of the gated augmented fusion unit. o To output the weight matrix of the projection, W is the activation function of the Gaussian error linear unit. v W gThese are the weight matrices for the value path and the gated path, respectively. v b g The path weight matrix W is the value. v The weight matrix W of the gated path g The corresponding bias vector, It is the Sigmoid activation function. b represents element-wise multiplication; o This is the bias vector for the output projection.

[0027] A multimodal umami peptide identification system based on molecular graphs and sequence features is applied to the multimodal umami peptide identification method based on molecular graphs and sequence features described in steps S1 to S5, comprising:

[0028] The sequence feature extraction module is used to extract amino acid sequence features of peptides using a pre-trained protein language model.

[0029] The structural feature extraction module is used to construct a two-dimensional molecular map based on peptide SMILES characterization and extract atomic-level structural features from the two-dimensional molecular map.

[0030] The sequence encoder receives the amino acid sequence features and includes two parallel channels:

[0031] The first channel uses a standard one-dimensional convolutional layer to extract local features from the amino acid sequence;

[0032] The second channel uses a dilated convolutional layer to extract long-range dependency features in the amino acid sequence;

[0033] The structural encoder receives the atomic-level structural features and further optimizes them through multiple two-dimensional convolutional layers to obtain atomic local structural features.

[0034] The feature fusion module is used for:

[0035] Within the sequence encoder, the local features are max-pooled to obtain max-pooled features, and the long-distance dependency features are adaptively max-pooled to obtain adaptive max-pooled features.

[0036] Within the structural encoder, the atomic local structural features are adaptively averaged pooled to obtain adaptive averaged pooled features;

[0037] The max pooling feature and the adaptive max pooling feature are concatenated and then combined with the adaptive average pooling feature to form a combined feature.

[0038] An enhanced residual module receives the combined features and outputs the umami peptide discrimination result through channel-level selection and nonlinear transformation.

[0039] The beneficial effects of this invention are:

[0040] Compared with existing methods for identifying umami peptides, this invention proposes a multi-view representation learning framework for amino acid sequences of single peptides by fusing two-dimensional molecular graphs and amino acid sequence features into a dual-modal feature representation. This effectively solves the problem of limited expressive power caused by the reliance on handcrafted features in traditional methods, as well as the deficiency that a single representation cannot fully characterize the functional properties of peptide molecules. By introducing gating mechanisms, layer normalization, and residual connections to fuse features, the method of this invention enhances its ability to capture both amino acid sequence features and atomic-level structural features, thereby improving training stability and generalization performance.

[0041] By fusing and nonlinearly transforming amino acid sequence features with atomic-level structural features, fine-grained, multi-level characterization and learning of umami peptide molecules are achieved. This is equivalent to "simultaneously seeing its atomic structure and biological sequence semantics" at the feature level, significantly enhancing the ability of the method to identify key functional groups and active sites. This method can achieve end-to-end umami activity prediction without the need for manually designed physicochemical parameters or structural prediction.

[0042] Experimental results show that the method of the present invention outperforms other comparative methods on an independent test set. Attached Figure Description

[0043] Figure 1 This is a flowchart of the multimodal umami peptide identification method based on molecular graphs and sequence features of the present invention;

[0044] Figure 2 This is a schematic diagram of the sequence encoder of the present invention;

[0045] Figure 3 This is a schematic diagram of the structural encoder of the present invention;

[0046] Figure 4 This is a schematic diagram of the enhanced residual module of the present invention. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0048] like Figure 1 As shown in the figure, this embodiment provides a multimodal umami peptide identification method based on molecular graphs and sequence features, which includes the following steps:

[0049] Step S1: Extract amino acid sequence features of peptides using a protein pre-trained language model; construct a two-dimensional molecular map based on the SMILES characterization of peptides; and construct atomic-level structural features based on the two-dimensional molecular map.

[0050] Step S2: Input the amino acid sequence features into a sequence encoder for processing to obtain splicing features;

[0051] Step S3: Input the atomic-level structural features into the structural encoder for processing to obtain adaptive average pooling features;

[0052] Step S4: Concatenate the spliced ​​features with the adaptive average pooling features to obtain combined features;

[0053] Step S5: Input the combined features into the enhanced residual module to determine whether it is an umami peptide.

[0054] In this embodiment, the amino acid sequence of the peptide is represented as a 1280-dimensional vector using the esm2_t33_650M_UR50D model from the ESM-2 series of protein pre-trained language models. The esm2_t33_650M_UR50D model contains 33 layers of Transformer encoders with a total of 650 million parameters. Simultaneously, by combining the molecular structure and physicochemical properties of the peptide's amino acid sequence, the amino acid sequence is converted into a digital representation. The amino acid sequence of each peptide is first converted into its corresponding SMILES characterization, generating a two-dimensional molecular map for each amino acid.

[0055] ;

[0056] Where S is the amino acid sequence of the peptide, a1 represents the first amino acid, and a k M1 represents the molecular map of the k-th amino acid, where k is the total number of amino acids. k This represents the molecular map of the k-th amino acid. For each atom in the two-dimensional molecular map, a comprehensive set of 21 features is extracted to represent its structure and physicochemical background, resulting in a 21-dimensional vector representation.

[0057] In this embodiment, umami peptide-related data were collected from four existing prediction models: UMPreg_FRL (umami peptide prediction - federated reinforcement learning model), Umami_YYDS (a chemometrics-based umami / bitter peptide judgment webpage system), Umami_MRNN (umami peptide multilayer recurrent neural network model), and Umami_DRLF (umami peptide deep reinforcement learning framework) to construct positive and negative sample datasets. Sequences containing non-standard amino acids were removed, and after deduplication and standardization preprocessing, 310 high-quality amino acid sequences of bitter peptides were obtained as the negative sample set. The positive sample set contained 279 validated amino acid sequences of umami peptides. To ensure the balance of the classification model training, 279 sequences were randomly selected from the 310 negative samples and combined with the positive samples to form the final balanced dataset for subsequent model training and evaluation.

[0058] To further mine the local and long-range structural information contained in the amino acid sequence features, in the standard one-dimensional convolutional channel of the sequence encoder, the amino acid sequence features capture the local interactions between adjacent amino acids in the peptide's amino acid sequence through a standard one-dimensional convolutional layer. The standard one-dimensional convolutional layer has a kernel size of 3, 64 channels, a stride of 1, and uses zero padding (padding=1) to maintain the amino acid sequence length. After the convolution operation, batch normalization and the ReLU activation function are applied to obtain the local features. (R is the set of real numbers, L is the number of amino acid residues). In the dilated convolutional channels of the sequence encoder, amino acid sequence features are modeled through dilated convolutional layers to model the functional dependencies between distant amino acids. A dilated convolutional layer (3 kernels, 256 output channels) is used, but the dilation rate is set to 2, meaning the receptive field on the input amino acid sequence is 5, allowing it to cover a wider sequence range and thus capture long-range synergistic effects (such as spatially adjacent residues after folding) without increasing the number of parameters. Batch normalization and the ReLU activation function are applied after the convolution operation to obtain long-range dependency features. For local features and long-distance dependency features Perform max pooling and adaptive max pooling separately, then concatenate them along the feature dimension to obtain the concatenated feature. This generates a 320-dimensional vector, which serves as the high-level sequence feature of the convolutional pathway.

[0059] To adapt the two-dimensional molecular diagram to the two-dimensional convolutional layer, the atomic set of each residue is aligned and zero-padded according to a preset maximum number of atoms M, thereby constructing a "residue × atom × attribute" tensor T of size L×M×21. S (Atomic-level structural features), where L is the number of residues and 21 is the atomic attribute dimension. Optionally, a normalized neighborhood propagation is first performed on the atomic-level structural features to integrate the initial features of each atom with its local chemical environment information, thereby obtaining a more expressive and structure-aware representation, providing higher-quality input for subsequent models. ,in ,in, To aggregate the normalized features of each atom and its first-order neighboring atoms, Â is the normalized adjacency matrix, F is the eigenvector of the atom, D is the degree matrix, A is the adjacency matrix, and I is the identity matrix.

[0060] Subsequently, within the structural encoder, tensor T S (Atomic-level structural features) are used as input to a multi-scale two-dimensional convolutional layer. The convolutional output is aggregated by adaptive average pooling, concatenated in the channel dimension, and passed through a linear layer to obtain a structure vector.

[0061] like Figure 3 As shown, in the structural encoder, atomic-level structural features are passed through multiple two-dimensional convolutional layers to capture local and global dependencies between atomic-level features. For the th... The output of a convolutional layer is defined as follows:

[0062] ;

[0063] in, Let f be the output feature of the i-th convolutional layer, and T be the size of the product filter. S It is a tensor representing the atomic feature matrix in a two-dimensional molecular diagram. Conv2D represents a two-dimensional convolution operation, where the convolution kernel size is set to 3×3, and ReLU is used as an activation function to increase nonlinearity.

[0064] In this embodiment, to alleviate the gradient vanishing problem that may occur in deep networks and to promote the effective propagation of local atomic structural features, a residual connection is introduced after the two-dimensional convolutional layer. If the tensor input to the two-dimensional convolutional layer is T... S (i.e., atomic-level structural features), then the output after two-dimensional convolution and batch normalization, plus the original input (or its appropriate transformation), forms the output:

[0065] ;

[0066] in, W is the output of the original input, which is the batch-normalized output of the i-th convolutional layer. proj It is a learnable weight matrix used to adjust dimensionality matching (when the number of input and output channels is inconsistent). This operation allows the UmamiScan (the method of this invention) model to directly utilize input information, thereby improving training efficiency and UmamiScan model performance.

[0067] Within the sequence encoder, the local amino acid sequence features T obtained after passing through a standard one-dimensional convolutional layer seq In this embodiment, max pooling (MaxPool1D) is used to obtain a fixed-size output:

[0068] ;

[0069] in, The output of max pooling is the max pooled feature. This is for max pooling.

[0070] Within the sequence encoder, the long-range dependency features T obtained through dilated convolution layers... dilatedThen, the Adaptive Maxpool1D operation is used to capture the most representative features:

[0071] ;

[0072] in, This is the adaptive max-pooling feature output after adaptive max-pooling. This is for adaptive max pooling.

[0073] Within the structural encoder, the atomic local structural features T are extracted by a two-dimensional convolutional network. conv Then, use the Adaptive Avgpool2D operation to convert it into a fixed-length vector:

[0074] ;

[0075] in, This is the adaptive average pooling feature output after adaptive average pooling. Adaptive average pooling.

[0076] The above maximum pooling feature T pool,seq Adaptive max pooling feature T pool,dilated and adaptive average pooling feature T pool,conv Concatenate the features to form a combined feature representation:

[0077] ;

[0078] in, For splicing features, As a combination feature, This is for splicing operations.

[0079] like Figure 4 As shown, to improve training stability and representation capability, an enhanced residual module (denoted as GEFUBlock) is constructed, using GEFU as the nonlinear activation unit and combining layer normalization and residual connections. The enhanced residual module (GEFUBlock) performs channel-level selection of combined features, and performs nonlinear transformation and feature selection through GEFU. GEFU is defined as:

[0080] ;

[0081] in, Let be the activation function of the gated augmented fusion unit, and x be the input feature of GEFU. It is the Sigmoid activation function. The activation function is the Gaussian Error Linear Unit (GELU). W represents element-wise multiplication; g W v These are the weight matrices for the gated path and the value path, respectively. g b v For W g W v The corresponding bias vector; W o and b o For the weight matrix and bias vector of the output projection, when the hidden dimension is equal to the output dimension, W o It can degenerate into an identity mapping.

[0082] The calculation process for GEFU is as follows:

[0083] ;

[0084] in: Let y be the gated enhancement fusion unit activation function, x be the input feature, v be the output feature after linear transformation, and g be the output feature after introducing the sigmoid activation function. Here, element-wise multiplication of v and g is used to obtain y. int y is obtained after applying the Gaussian error linear unit activation function, and finally z is obtained by linear transformation of y. It is the Sigmoid activation function. The activation function is the Gaussian Error Linear Unit (GELU). This represents Hadamard (element-by-element) multiplication, d h To hide the dimension, d out For output dimensions, These are the weight matrices for the value path and the gated path, respectively. For W v， W g Its corresponding bias vector, W o For the output projection matrix: when If the projection is linear, it degenerates into an identity mapping.

[0085] For input Enhanced residual module output The calculation is as follows:

[0086] ;

[0087] Where x is the input feature, To prevent overfitting, the input x is obtained by layer normalization. To generate intermediate output features through the gated enhancement fusion unit, r is the value added to the residual term to obtain the final output feature. For a learnable weight matrix, when the dimension d of the input x is... in and the dimension d of the GEFU branch outputout When they are different, use Multiply by x (plus the bias) ), linearly transform x to d out dimension.

[0088] The first enhanced residual module has input dimension d. in =576, output dimension d out =h; The second enhanced residual module has input and output dimensions of h; A Dropout layer is set between the two modules. To mitigate overfitting, using the combined feature X as the input feature, we obtain the following sequentially:

[0089] ;

[0090] Here, x1 represents the intermediate feature representation extracted and dimensionality reduced by the first enhanced residual module, and the subscripts "576→h" and "h→h" indicate the input and output dimensions of the residual block. Finally, the discriminative features obtained are those enhanced by gating and residual reinforcement. .

[0091] Discriminant features enhanced by gating and residuals Inputting the data into a fully connected discriminant head yields the logit for binary classification:

[0092] ;

[0093] Where m is the logarithmic probability. Let b be the weight vector and b be the bias peak.

[0094] Mapping the log-odds m to the posterior probability of umami peptides:

[0095] ;

[0096] During the inference phase, a threshold decision is used:

[0097] ;

[0098] in, For binary classification prediction values, For the preset threshold (preferred value) ).

[0099] During training, the loss function is a binary cross-entropy loss function with log odds, defined as:

[0100] ;

[0101] in, This is the true label of the nth sample (1 for umami peptides, 0 for non-umami peptides). The raw log odds output by the UmamiScan model (without Sigmoid activation). For the Sigmoid function, The number of samples in the batch is denoted as 'batch size'. The loss function automatically applies the Sigmoid function to the log-odds ratio and calculates the loss, improving numerical stability.

[0102] Furthermore, the Adam optimizer is used to update the parameters of the UmamiScan model. The parameter update rules of the Adam optimizer are as follows:

[0103] ;

[0104] Where t is the step size, m t and v t These are the first and second moment estimates of the gradient, respectively. These are the first and second decay coefficients of the exponent. and The first and second moment estimates of the gradient after bias correction more accurately reflect the statistical information of the true gradient. For learning rate, For model parameters, To prevent small constants from being divided by zero.

[0105] By minimizing the loss function, the model's predicted probability distribution approximates the true label distribution, thereby achieving high-precision identification of umami peptide activity.

[0106] The gated enhancement fusion unit proposed in this invention enhances the UmamiScan model's ability to capture nonlinear feature combinations by introducing Gaussian error linear unit activation after the gated output. Experiments show that it outperforms the traditional MLP (Multilayer Perceptron) and standard GLU (Gated Linear Unit) structures in the umami peptide recognition task.

[0107] To verify the effectiveness of this method, in this embodiment, the balanced dataset of 558 collected samples is divided into training set, validation set and test set in a ratio of 7:1:2.

[0108] First, the UmamiScan model is trained on the training set. Then, the best-performing model parameters on the validation set are selected by adjusting the hyperparameters. Finally, the best-performing model parameters are tested on an independent test set to observe performance. Table 1 shows the performance of the method described in this invention and other methods on the independent test set. It can be seen that the method of this invention outperforms the other four methods in accuracy, F1 score, Sn sensitivity, and Matthews correlation coefficient (MCC). Furthermore, compared with other methods, the improvement rates of the method of this invention in accuracy, F1 score, sensitivity, and MCC are 2.68%–6.25%, 3%–6.41%, 1.79%–25%, and 5.43%–12.63%, respectively. In conclusion, the method of this invention significantly outperforms other advanced methods in the identification of umami peptides. This indicates that the method of the present invention achieves high accuracy and consistency in peptide function prediction. Among them, UmamiScan is the method of the present invention, Umami-gcForest is the umami deep forest model, Umami_MRNN is the umami peptide multilayer recurrent neural network model, Umami_YYDS is the stoichiometric umami / bitter peptide judgment webpage system, and UMPred_FRL umami peptide prediction - federated reinforcement learning model.

[0109] Table 1 shows the comparison with other models on the test set.

[0110]

[0111] In this implementation, the present invention fuses sequence features obtained from a protein pre-trained language model with atomic-level structural features corresponding to amino acids in each sequence. This dual-channel fusion allows for better learning of features not only from the sequence but also from the structure, increasing the model's generalization ability. To gain a deeper understanding of the different features and the roles of each component in the model and to verify their contributions, the present invention designed a series of ablation experiments. These experiments not only utilize the complete model framework and a controlled variable strategy of "changing only one component at a time" to systematically evaluate the impact of key modules on the final prediction performance, but also compare the performance of single-feature models and dual-feature models, further demonstrating the effectiveness of this innovative approach of fusing sequence and atomic features.

[0112] As shown in Table 2, the sequence encoder's sequence channels pass through two different neural network channels. With only the sequence channels, if only the dilated convolutional layers are retained (i.e., the standard one-dimensional convolutional layers are removed), all evaluation metrics of the proposed method are below 90%. Compared to the complete sequence channels, accuracy (ACC), F1 score, sensitivity (Sn), specificity (Sp), and MCC all decrease by approximately 3% to 7%, indicating that the standard one-dimensional convolutional layers play a crucial role in feature extraction. However, if the dilated convolutional layers are removed (retaining the standard one-dimensional convolution), the ACC is 91.07%, F1 is 91.38%, Sn is 94.64%, Sp is 87.50%, and MCC is 0.8235. While this overall performance is better than the case where the standard one-dimensional convolutional layers are removed, it is still significantly lower than the complete sequence channels (ACC 91.96%, F1 91.89%, MCC 0.8394). This demonstrates that dilated convolution can capture longer-distance sequence dependencies, compensating for the limited receptive field of the standard one-dimensional convolutional layers. In summary, standard one-dimensional convolutional layers and dilated convolutional layers each have their advantages in sequence feature extraction. Their synergistic fusion can comprehensively capture multi-scale local and long-range sequence patterns, thereby significantly improving the model's predictive performance.

[0113] When the sequence channels are removed, leaving only the structural encoder's structural channels, the model performance significantly degrades. As shown in Table 2, after removing the 2D convolutional layers, the model accuracy (ACC) drops to 83.93%, the F1 score is 0.8475, the sensitivity (Sn) is 89.29%, the specificity (Sp) is 78.57%, and the MCC is only 0.6825. When adaptive average pooling is replaced with adaptive max pooling, all metrics further decrease to the same level (ACC, F1, Sn, and Sp are all 83.93%, and MCC is 0.6786). Compared to the complete structural channels (ACC 84.82%, F1 83.17%, MCC 0.7103), these results indicate that the 2D convolutional layers play a crucial role in capturing spatial topological features and neighborhood dependencies, while adaptive average pooling helps maintain the stable convergence of global structural information.

[0114] Furthermore, in the fusion module, if the enhanced residual module (GEFUBlock) is removed, the ACC of the method of the present invention is 91.96%, F1 is 92.04%, and MCC is 0.8394. Compared with the method of the present invention (UmamiScan) (ACC is 93.75%, F1 is 93.91%, and MCC is 0.8763), all indicators have decreased, indicating that the enhanced residual module (GEFUBlock) mechanism can effectively adjust the weight distribution of different channel features and improve the recognition ability of multimodal information fusion.

[0115] In summary, sequence features, atomic-level structural features, and their fusion via linear layers all play crucial roles in the umami peptide recognition task. Specifically, the sequence channel is responsible for capturing multi-scale sequence patterns, the structural channel characterizes spatial conformation and interaction features, and the linear layer fusion module adaptively integrates global semantic and local structural information, thereby significantly improving the overall predictive performance and generalization ability of the model.

[0116] Table 2 Comparison of the effects of different module combinations

[0117]

[0118] In summary, the method of this invention obtains the feature representation of peptides by combining a protein pre-trained language model and atomic structure, integrating context-aware sequence representation and atomic-level structural features, thereby improving the accuracy of the model in predicting umami peptides. This multimodal method uses SMILES two-dimensional molecular graphs to supplement structural and chemical environment information, embeds large-scale sequence semantic priors into the protein pre-trained language model, and achieves end-to-end discrimination through learnable gating / attention fusion. It fundamentally alleviates problems such as relying solely on sequences and manual features, while also possessing better robustness for implementation and design guidance value.

[0119] Another embodiment of the present invention provides a multimodal umami peptide recognition system based on molecular graphs and sequence features, comprising:

[0120] The sequence feature extraction module is used to extract amino acid sequence features of peptides using a pre-trained protein language model.

[0121] The structural feature extraction module is used to construct a two-dimensional molecular map based on peptide SMILES characterization and extract atomic-level structural features from the two-dimensional molecular map.

[0122] The sequence encoder receives the amino acid sequence features and includes two parallel channels:

[0123] The first channel uses a standard one-dimensional convolutional layer to extract local features from the amino acid sequence;

[0124] The second channel uses a dilated convolutional layer to extract long-range dependency features in the amino acid sequence;

[0125] The structural encoder receives the atomic-level structural features and further optimizes them through multiple two-dimensional convolutional layers to obtain atomic local structural features.

[0126] The feature fusion module is used for:

[0127] Within the sequence encoder, the local features are max-pooled to obtain max-pooled features, and the long-distance dependency features are adaptively max-pooled to obtain adaptive max-pooled features.

[0128] Within the structural encoder, the atomic local structural features are adaptively averaged pooled to obtain adaptive averaged pooled features;

[0129] The max pooling feature and the adaptive max pooling feature are concatenated and then combined with the adaptive average pooling feature to form a combined feature.

[0130] An enhanced residual module receives the combined features and outputs the umami peptide discrimination result through channel-level selection and nonlinear transformation.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying multimodal umami peptides based on molecular graphs and sequence features, characterized in that, Includes the following steps: Step S1: Extract amino acid sequence features of peptides using a protein pre-trained language model; A two-dimensional molecular map was constructed based on the SMILES characterization of peptides, and atomic-level structural features were constructed based on the two-dimensional molecular map. Step S2: Input the amino acid sequence features into a sequence encoder for processing to obtain splicing features; Step S3: Input the atomic-level structural features into the structural encoder for processing to obtain adaptive average pooling features; Step S4: Concatenate the spliced ​​features with the adaptive average pooling features to obtain combined features; Step S5: Input the combined features into the enhanced residual module to determine whether it is an umami peptide; In step S5, the enhanced residual module includes a gated enhanced fusion unit, layer normalization, and residual connection. The input features of the enhanced residual module pass through the layer normalization and gated enhanced fusion unit in one path and the layer normalization in the other path. The two features are then connected through the residual to obtain the output features. The calculation process of the gated enhancement fusion unit is expressed by the following formula: ； in, Let W be the activation function of the gated augmented fusion unit, x be the input feature, and W be the activation function of the gated augmented fusion unit. o To output the weight matrix of the projection, W is the activation function of the Gaussian error linear unit. v W g These are the weight matrices for the value path and the gated path, respectively. v b g The path weight matrix W is the value. v The weight matrix W of the gated path g The corresponding bias vector, It is the Sigmoid activation function. b represents element-wise multiplication; o This is the bias vector for the output projection.

2. The method for identifying multimodal umami peptides based on molecular graphs and sequence features according to claim 1, characterized in that, In step S1, the protein pre-training language model is the esm2_t33_650M_UR50D model from the ESM-2 series.

3. The method for identifying multimodal umami peptides based on molecular graphs and sequence features according to claim 2, characterized in that, In step S1, the amino acid sequence feature is obtained by inputting the amino acid sequence of the peptide into a protein pre-training language model to obtain a 1280-dimensional vector representation of each amino acid residue in the last hidden state of the peptide's amino acid sequence.

4. The method for identifying multimodal umami peptides based on molecular graphs and sequence features according to claim 1, characterized in that, The steps for constructing atomic-level structural features in step S1 include: A two-dimensional molecular map is generated based on the amino acid sequence of peptides characterized by SMILES, where each amino acid corresponds to a single atom and the edges correspond to the chemical bonds between atoms. For each atom, 21-dimensional comprehensive features are extracted as node features, specifically covering atom type, degree, total number of hydrogen atoms, implicit valence state, aromaticity, ring size, hybridization type, and Gasteg charge; Construct atomic-level structural features.

5. The method for identifying multimodal umami peptides based on molecular graphs and sequence features according to claim 1, characterized in that, In step S2, the sequence encoder includes a standard one-dimensional convolutional channel and a dilated convolutional channel. Within the standard one-dimensional convolutional channel, a standard one-dimensional convolutional layer is used to extract local features from the amino acid sequence, and the local features are max-pooled to obtain max-pooled features. Within the dilated convolutional channel, a dilated convolutional layer is used to extract long-range dependency features from the amino acid sequence; the long-range dependency features are then subjected to adaptive max pooling to obtain adaptive max pooling features; the max pooling features and the adaptive max pooling features are then concatenated to obtain concatenated features.

6. The method for identifying multimodal umami peptides based on molecular graphs and sequence features according to claim 1, characterized in that, In step S3, the structure encoder includes a two-dimensional convolutional channel. Within the two-dimensional convolutional channel, multiple two-dimensional convolutional layers are used to further optimize the atomic-level structural features to obtain atomic local structural features. The atomic local structural features are then subjected to adaptive average pooling to obtain adaptive average pooling features.

7. The method for identifying multimodal umami peptides based on molecular graphs and sequence features according to claim 6, characterized in that, The two-dimensional convolutional layer contains at least two different kernel sizes. The outputs at each scale are adaptively averaged and concatenated along the channel dimension before being passed through a linear layer to obtain a structure vector.

8. The method for identifying multimodal umami peptides based on molecular graphs and sequence features according to claim 1, characterized in that, In step S1, when extracting atomic-level structural features, the atomic set of each residue is zero-filled according to the preset maximum number of atoms M, and a tensor of size L×M×21 is constructed, where L is the number of residues and 21 is the atomic attribute dimension.

9. A multimodal umami peptide recognition system based on molecular graphs and sequence features, applied to the multimodal umami peptide recognition method based on molecular graphs and sequence features as described in claim 1, characterized in that, include: The sequence feature extraction module is used to extract amino acid sequence features of peptides using a pre-trained protein language model. The structural feature extraction module is used to construct a two-dimensional molecular map based on peptide SMILES characterization and extract atomic-level structural features from the two-dimensional molecular map. The sequence encoder receives the amino acid sequence features and includes two parallel channels: The first channel uses a standard one-dimensional convolutional layer to extract local features from the amino acid sequence; The second channel uses a dilated convolutional layer to extract long-range dependency features in the amino acid sequence; The structural encoder receives the atomic-level structural features and further optimizes them through multiple two-dimensional convolutional layers to obtain atomic local structural features. The feature fusion module is used for: Within the sequence encoder, the local features are max-pooled to obtain max-pooled features, and the long-distance dependency features are adaptively max-pooled to obtain adaptive max-pooled features. Within the structural encoder, the atomic local structural features are adaptively averaged pooled to obtain adaptive averaged pooled features; The max pooling feature and the adaptive max pooling feature are concatenated and then combined with the adaptive average pooling feature to form a combined feature. An enhanced residual module receives the combined features and outputs the umami peptide discrimination result through channel-level selection and nonlinear transformation.

Images