Method for classifying antihypertensive peptides by fusing sequence and structure multi-modal features and combining contrast-generative combined optimization

Abstract

The invention relates to a method for classifying antihypertensive peptides by fusing sequence and structure multi-modal features and combining contrast-generative combined optimization. The method comprises the following steps: extracting sequence feature representation and structure feature representation of peptide fragments; performing multi-modal feature enhancement of comparison-generative joint optimization on the sequence feature representation and the structural feature representation; and performing peptide classification on the enhanced feature representation by using a preset Kan-Conv structure and tag smooth joint optimization classification model. According to the method, high-precision recognition of the functional activity of the antihypertensive peptide is achieved, information in the three aspects of the sequence, the structure and the generative potential space is creatively and comprehensively utilized, the blank that the structure-sequence synergistic effect is ignored in the existing peptide function prediction field is filled, and the screening efficiency and prediction reliability of the antihypertensive peptide are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics and artificial intelligence intersection technology, and in particular to a method for classifying antihypertensive peptides that integrates sequence and structural multimodal features and combines comparative-generative joint optimization. Background Art

[0002] In the field of discovery and design of antihypertensive peptides, traditional computational prediction methods mainly rely on manual features of amino acid sequences or simply use sequence language models. However, these methods often ignore the folding conformation and dynamic interaction information of peptide molecules in three-dimensional space. The functional activity of peptides depends not only on the linear sequence but also on its spatial structure. On the other hand, although existing structure prediction tools can provide high-precision peptide tertiary structure, existing downstream prediction models rarely deeply integrate this structural information with sequence information, resulting in limited prediction of key functions such as peptide-receptor binding sites, cell membrane penetration ability, and enzyme inhibitory activity.

[0003] Furthermore, traditional strategies often model sequence and structural features independently, lacking an end-to-end multi-view collaborative learning framework, making it difficult for models to capture the inherent connections between the two. Meanwhile, while recent deep generative models have achieved significant success in image and small molecule design, they have yet to be fully applied to peptide-level functionalization prediction. Summary of the Invention

[0004] The purpose of this invention is to propose a method for classifying antihypertensive peptides that integrates multimodal sequence and structural features with contrastive-generative joint optimization to address the aforementioned challenges of the prior art. By integrating peptide sequence information and structural features across multiple views and integrating supervised contrastive learning with a latent diffusion model for joint optimization, this method achieves a refined encoding of high-dimensional features and a low-dimensional latent representation of peptide molecules. The end-to-end learning framework constructed by this invention not only significantly improves the accuracy of peptide function prediction but also possesses excellent generalization capabilities, making it suitable for the discovery, classification, and design of functional peptide molecules, such as antihypertensive peptides.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A classification method for antihypertensive peptides that integrates sequence and structural multimodal features and combines comparative-generative joint optimization, including:

[0007] Extract sequence feature representation and structural feature representation of peptide segments;

[0008] Performing multimodal feature enhancement on the sequence feature representation and the structural feature representation by contrastive-generative joint optimization;

[0009] The enhanced feature representation is used to perform peptide classification using a preset classification model based on the joint optimization of Kan-Conv structure and label smoothing.

[0010] Optionally, extracting the sequence feature representation and structural feature representation of the peptide segment includes:

[0011] The feature representation of peptides is extracted based on a preset large-scale protein language model, and the structural feature representation generated by RDKit is also utilized.

[0012] Optionally, the structural feature representation generated using RDKit includes:

[0013] Mapping the peptide segment into a three-dimensional structural model, filling all implicit hydrogen atoms in the three-dimensional structural model, and obtaining structural data including the three-dimensional coordinate information of all atoms through a molecular conformation embedding algorithm;

[0014] Extracting atomic coordinates from the structural data to obtain an atomic coordinate matrix;

[0015] Calculating the Pearson correlation coefficient between the three-dimensional coordinate dimensions based on the atomic coordinate matrix to obtain a correlation coefficient matrix;

[0016] The correlation coefficient matrix is ​​numerically normalized, image-converted, and feature vectors are generated in sequence to obtain a structural feature representation.

[0017] Optionally, performing contrastive-generative joint optimization of the multimodal feature enhancement on the sequence feature representation and the structural feature representation includes:

[0018] Performing supervised contrastive learning on the sequence feature representation and the structural feature representation to obtain a sequence feature representation and a structural feature representation enhanced by contrastive learning;

[0019] Generative learning is performed on the sequence feature representation and the structural feature representation to obtain the sequence feature representation and the structural feature representation enhanced by generative learning.

[0020] Optionally, performing supervised comparative learning on the sequence feature representation and the structural feature representation includes:

[0021] Randomly perturb the sequence feature representation and the structural feature representation, and sequentially concatenate the perturbed feature representations up and down to form a feature matrix;

[0022] Calculate the dot product similarity matrix between each pair in the feature matrix;

[0023] Construct a mask matrix of the same scale as the similarity matrix, and multiply the dot product similarity matrix by the mask matrix after element-by-element exponentiation.

[0024] Sum all non-zero elements in each row of the multiplied matrix to obtain the denominator of the supervised contrast loss; where each row in the multiplied matrix refers to each enhanced sample;

[0025] According to the definition of the positive sample pair, the exponential similarity between each enhanced sample and the positive sample in the multiplied matrix is ​​extracted to form the numerator of the supervised contrast loss;

[0026] For each enhanced sample in the multiplied matrix, the contrast loss is calculated sample by sample according to the denominator and numerator of the corresponding supervised contrast loss and averaged to obtain the final supervised contrast loss;

[0027] The final supervised contrastive loss is used to obtain the sequence feature representation and structural feature representation enhanced by contrastive learning.

[0028] Optionally, generative learning of the sequence feature representation and the structural feature representation includes:

[0029] Mapping the sequence feature representation and the structural feature representation to a latent space to obtain initial latent variables;

[0030] Injecting noise into the initial latent variable to obtain a noisy latent variable;

[0031] The noisy latent variables are concatenated with the normalized time steps and then fed into the preset fully connected network to output the prediction of the noise.

[0032] Use the predicted noise to infer the initial latent variables and remove the noise components;

[0033] The recovered latent variables are mapped back to the original dimension through two layers of full connection to obtain the reconstructed feature vector.

[0034] Optionally, using a preset classification model based on a Kan-Conv structure and label smoothing joint optimization, peptide classification is performed on the enhanced feature representation, including:

[0035] Perform dimension splicing on the enhanced feature representation to generate merged features of preset size;

[0036] For the merged features, several parallel adaptive convolution kernel channels are applied to perform batch normalization respectively, and then nonlinearity is introduced through the ReLU function.

[0037] Use the max pooling layer to flatten all channels and spatial dimensions into a long vector, generating a one-dimensional feature vector of a preset size;

[0038] The one-dimensional feature vector is mapped through two layers of full connection to obtain the original classification score;

[0039] Applying a log-softmax operation to the raw classification scores yields a logarithmic probability distribution of two classes for each sample; wherein the two classes refer to anti-stress peptides and non-anti-stress peptides;

[0040] Based on the logarithmic probability distribution, the peptide segments are classified.

[0041] Optionally, the two-layer fully connected mapping includes:

[0042] First layer: compress the one-dimensional feature vector of a preset size and apply ReLU to obtain the compressed hidden representation;

[0043] The second layer maps the compressed hidden representation to the preset number of categories and outputs the original classification score.

[0044] The beneficial effects of the present invention are:

[0045] The present invention first extracts the sequence feature representation and structural feature representation of the peptide segment; then performs contrast-generative joint optimization multimodal feature enhancement on the sequence feature representation and structural feature representation; finally, the enhanced feature representation is used to classify peptides using a preset Kan-Conv structure-based and label smoothing joint optimization classification model. The present invention achieves fine encoding and low-dimensional potential representation of high-dimensional features of peptide molecules by fusing the sequence information and structural features of the peptide segment through multiple views, and on this basis, introduces supervised contrast learning and latent diffusion model branches for joint optimization. The present invention innovatively integrates the three aspects of sequence, structure and generative latent space information, filling the gap in the existing field of peptide function prediction that ignores the structure-sequence synergistic effect, and significantly improves the screening efficiency and prediction reliability of anti-hypertensive peptides. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 This is a flow chart of a method for classifying antihypertensive peptides that integrates sequence and structural multimodal features and combines comparative-generative joint optimization according to an embodiment of the present invention;

[0048] Figure 2 AUROC curve comparison diagram and PRROC curve comparison diagram of the embodiment of the present invention and other methods;

[0049] Figure 3 Comparison of AUC values ​​between the embodiments of the present invention and other comparative methods;

[0050] Figure 4 The figure shows the MCC value distribution comparison between the embodiment of the present invention and other comparison methods. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] This example proposes a method for classifying antihypertensive peptides that integrates sequence and structural multimodal features and combines comparative-generative joint optimization, including:

[0054] Extract sequence feature representation and structural feature representation of peptide segments;

[0055] Performing multimodal feature enhancement on the sequence feature representation and the structural feature representation by contrastive-generative joint optimization;

[0056] The enhanced feature representation is used to perform peptide classification using a preset classification model based on the joint optimization of Kan-Conv structure and label smoothing.

[0057] Specifically, this embodiment proposes an antihypertensive peptide classification method that integrates sequence and structural multimodal features and combines contrast-generative joint optimization for accurate prediction of antihypertensive peptide sequences. DeepATHPFuse first extracts the global context embedding of the peptide sequence through Prot-T5, and uses RDKit to obtain the spatial graph embedding based on the three-dimensional structure, and introduces supervised contrastive learning (SupCon) and conditional latent diffusion model (LDM) to jointly optimize the features to simultaneously enhance category discriminability and distribution continuity; the optimized features are then input into the classification network composed of a multi-core KanConv convolutional layer and a fully connected layer, and hard label cross entropy and KL divergence based on 0.1 label smoothing are used for joint supervision, thereby effectively suppressing overconfidence and improving generalization and robustness. This embodiment shows high accuracy, high stability and good scalability in the antihypertensive peptide prediction task, and is suitable for application scenarios such as functional peptide screening and new drug development. The specific implementation process is Figure 1 shown.

[0058] The core of this example includes: (1) extracting sequence representations based on a large-scale protein language model (ProtT5) and utilizing peptide structural features generated by RDKit; (2) aligning sequence and structural views in a unified embedding space through supervised contrastive learning (SupConLoss) to enhance multi-viewing of the same peptide sample. Figure 1 Consistency; (3) a latent diffusion model is introduced to learn a generative smooth mapping in the latent space, improving the model's robustness to noise and small variations; (4) the above multimodal fusion representation is used for downstream activity classification prediction, and through a carefully designed cross-entropy and label-smoothed KL divergence joint optimization, high-precision identification of the functional activity of antihypertensive peptides is achieved. This method innovatively integrates sequence, structure, and generative latent space information, filling the gap in the existing field of peptide function prediction that ignores structure-sequence synergy, significantly improving the screening efficiency and prediction reliability of antihypertensive peptides.

[0059] Furthermore, extracting the sequence feature representation and structural feature representation of the peptide segment includes:

[0060] The feature representation of peptides is extracted based on a preset large-scale protein language model, and the structural feature representation generated by RDKit is also utilized.

[0061] Furthermore, the structural feature representation generated by RDKit includes:

[0062] Mapping the peptide segment into a three-dimensional structural model, filling all implicit hydrogen atoms in the three-dimensional structural model, and obtaining structural data including the three-dimensional coordinate information of all atoms through a molecular conformation embedding algorithm;

[0063] Extracting atomic coordinates from the structural data to obtain an atomic coordinate matrix;

[0064] Calculating the Pearson correlation coefficient between three-dimensional coordinate dimensions based on the atomic coordinate matrix to obtain a correlation coefficient matrix;

[0065] The correlation coefficient matrix is ​​numerically normalized, image-converted, and feature vectors are generated in sequence to obtain a structural feature representation.

[0066] Specifically, in this embodiment, antihypertensive peptide sequence data collection and preprocessing are first performed:

[0067] The primary data source for this example was the Food-Derived Bioactive Peptide Database (DFBP), which focuses on bioactive peptides extracted from food sources, including those with antihypertensive properties. In addition, a dataset compiled by Du et al. in pLM4ACE was used, which includes the sequences, IC50 values, and original references of all collected peptides. Among existing datasets employed by current methods, many contain a large number of short peptide sequences, particularly di- and tripeptides. It has been observed that these short peptides have relatively high prediction accuracy. However, despite the favorable performance metrics of these short peptides, this example considers that their presence may affect the generalization ability and practical applicability of the model. Consequently, peptides shorter than five residues were removed, and CD-HIT was applied to filter out sequences with a similarity exceeding 90%. Ultimately, a dataset consisting of 732 positive samples and 1601 negative samples was obtained.

[0068] Sequence feature representation:

[0069] The pre-trained model relies on large-scale protein datasets to learn high-level representations of peptide sequences, and can deeply capture the semantic information and contextual dependencies of the sequence. In the field of protein research, Prot-T5 is a widely used pre-trained model. It is based on the Transformer architecture and has the advantages of processing long sequences and powerful feature modeling capabilities. With the help of the self-attention mechanism, Prot-T5 can capture the global correlation between long-range residues in the peptide chain and encode a sequence of length L into a feature matrix of L×1024 dimensions, thereby comprehensively characterizing the complex features and interactions of the sequence. The specific implementation process is as follows: Figure 1 shown.

[0070] Structural features indicate:

[0071] Sequence input and 3D conformation generation:

[0072] First, the peptide sequence to be processed is mapped into a three-dimensional structural model using a molecular modeling tool; all implicit hydrogen atoms are added to the model, and structural data containing the three-dimensional coordinate information of all atoms is obtained through a molecular conformation embedding algorithm.

[0073] Atomic coordinate extraction:

[0074] The structural data is parsed row by row to filter out all atomic records; the three coordinate values ​​of x, y, and z are parsed from each record in turn, and are respectively collected into three coordinate sets, and finally combined into an atomic coordinate matrix.

[0075] Correlation coefficient matrix calculation and numerical normalization:

[0076] The Pearson correlation coefficient between three-dimensional coordinate dimensions is calculated based on the atomic coordinate matrix to obtain a three-by-three correlation coefficient matrix; invalid values that may appear in the matrix are replaced to ensure that all elements are finite values.

[0077] Image conversion and feature vector generation:

[0078] The normalized correlation coefficient matrix is regarded as a grayscale image, and interpolation magnification of a uniform size is performed. Subsequently, the pixel values of the magnified image are flattened into a one-dimensional numerical vector in a predetermined order, and the vector is output as the structure correlation coefficient feature of the peptide segment.

[0079] Further, the multi-modal feature enhancement of the sequence feature representation and the structure feature representation by contrastive generative joint optimization includes:

[0080] Supervised contrastive learning is performed on the sequence feature representation and the structure feature representation to obtain the sequence feature representation and the structure feature representation enhanced by contrastive learning;

[0081] Generative learning is performed on the sequence feature representation and the structure feature representation to obtain the sequence feature representation and the structure feature representation enhanced by generative learning.

[0082] Further, the supervised contrastive learning on the sequence feature representation and the structure feature representation includes:

[0083] Random perturbation is performed on the sequence feature representation and the structure feature representation, and the perturbed feature representations are sequentially concatenated to form a feature matrix;

[0084] A dot product similarity matrix between each other in the feature matrix is calculated;

[0085] A mask matrix of the same size as the similarity matrix is constructed, and the dot product similarity matrix is element-wise exponentiated and multiplied by the mask matrix;

[0086] The sum of all non-zero elements in each row of the multiplied matrix is calculated to obtain the denominator of the supervised contrastive loss; wherein each row of the multiplied matrix refers to each enhanced sample;

[0087] According to the definition of the positive sample pair, the exponential similarity between each enhanced sample and the positive sample in the multiplied matrix is extracted to form the numerator of the supervised contrastive loss;

[0088] For each enhanced sample in the multiplied matrix, the contrastive loss is calculated sample by sample according to the denominator and the numerator of the corresponding supervised contrastive loss, and then averaged to obtain the final supervised contrastive loss;

[0089] The sequence feature representation and the structure feature representation enhanced by contrastive learning are obtained by using the final supervised contrastive loss.

[0090] Specifically, in this embodiment, supervised contrastive learning includes:

[0091] Input data and view construction:

[0092] Assume that there are N peptide sequences in a batch, and each sequence is subjected to two independent random perturbations (or different Dropout / data augmentation strategies) to obtain two sets of enhanced features. These two sets of representations are concatenated in order to form a feature matrix of size 2N×D, where D is the feature dimension output by the projection head. In this embodiment, they are respectively referred to as visual Figure 1 Heshi Figure 2 .

[0093] Pairwise similarity calculation:

[0094] For all 2N concatenated vectors, calculate the dot product similarity between each pair and divide it by the temperature coefficient τ (the default value is 0.07) to obtain a 2N×2N symmetric similarity matrix:

[0095]

[0096] In this matrix, the value in row i and column j represents the phase velocity between the i-th and j-th vectors. This matrix is ​​symmetric and quantifies the "closeness" of each pair of enhanced representations. Since we only care about relative magnitude and eliminate the effect of vector length in exponential operations, we typically perform L2 normalization on the features before applying this loss.

[0097] Self-contrast masking and negative sample indexing:

[0098] In order to exclude the comparison between each eigenvector and itself (i.e., diagonal items), a mask matrix with the same scale as the similarity matrix is ​​constructed: the diagonal is set to 0 and the non-diagonal is set to 1. The similarity matrix is ​​indexed element by element and then multiplied by the mask, that is:

[0099]

[0100] At this point, all non-zero elements in each row of the matrix represent the exponential similarity between the sample and “all other samples” (including other perturbation views in the same batch) as negative samples.

[0101] Normalized denominator:

[0102] For each row (i.e., each enhanced sample), all non-zero elements are summed to obtain the denominator of the InfoNCE loss:

[0103]

[0104] Determination of positive sample pairs and calculation of molecules:

[0105] A positive sample pair is defined as two augmented views of the same original sample: for the first half of the matrix (view Figure 1 ), the positive sample in the i-th row is in the i-th row of view Figure 2 . For the second half (view Figure 2 ), the positive sample in the i-th row is in the i-th row of view Figure 1 . This correspondence is denoted by the mapping function p(i). Extract the index similarity between each augmented sample and its positive sample to form the numerator in InfoNCE:

[0106]

[0107] The contrastive loss is calculated sample by sample and averaged:

[0108] For the i-th augmented sample, the InfoNCE loss is calculated as the negative logarithm of the ratio of the positive sample similarity to all negative sample similarities:

[0109]

[0110] The final supervised contrastive loss is obtained by averaging the loss of all 2N augmented representations:

[0111]

[0112] Further, the generative learning on the sequence feature representation and the structure feature representation comprises:

[0113] Mapping the sequence feature representation and the structure feature representation to a latent space to obtain an initial latent variable;

[0114] Injecting noise into the initial latent variable to obtain a noisy latent variable;

[0115] Concatenating the noisy latent variable and the normalized time step and inputting them into a preset fully connected network to output a prediction of the noise;

[0116] Using the predicted noise to back-propagate the initial latent variable to remove the noise component;

[0117] Mapping the recovered latent variable back to the original dimension through two fully connected layers to obtain a reconstructed feature vector.

[0118] Specifically, the specific steps of the generative learning branch in the embodiment include:

[0119] In the latent diffusion branch, the multi-modal feature is optimized by generative learning with the following specific parameters in the embodiment:

[0120] Input features and network structure:

[0121] Input features D: Sequence embeddings from Prot-T5 and structural embeddings from RDKit. Latent dimension L: set to 1024 via hyperparameters. Initial latent variables:

[0122] z0=W (2) (SiLU(W (1) h+b (1) ))+b (2)

[0123] Diffusion coefficient scheduling:

[0124] Assume the total number of diffusion steps T = 1000. Initial and final noise variance: β1 = 10 -4 ,β T =0.02. β t In [10 -4 , 0.02], which represents the intensity of the noise injected in the tth step. The cumulative retention coefficient is calculated Used to smoothly interpolate the original signal and noise in a closed-form formula.

[0125] Inject noise into the latent variable:

[0126] For each batch of samples (batch size is B), randomly select the number of steps Generate standard normal noise ε. Closed-form construction of noisy latent variables:

[0127]

[0128] This formula ensures that in the limit t→T, z t Almost pure noise.

[0129] UNet branch for noise prediction:

[0130] The noisy latent variable z t Concatenate with the normalized time step τ = t / T∈[0,1] to construct a 64+1 dimensional vector. Input a three-layer fully connected U-net network, all with ReLUs, and output a prediction of the noise. The difference between the predicted noise and the actual noise is measured using the mean squared error:

[0131]

[0132] Recover the noise-free latent variable:

[0133] The formula for inferring the initial latent variables using prediction noise is: This operation is equivalent to a first-order inverse diffusion approximation and aims to remove noise components.

[0134] Latent variables reconstruct high-dimensional features:

[0135] The recovered latent variables are mapped back to the original dimension through two layers of full connection:

[0136]

[0137] The corresponding reconstruction loss also uses mean square error:

[0138]

[0139] Total branch loss and weight:

[0140] The noise prediction loss and the reconstruction loss are weighted and summed according to the weights (typical ratio is about 1:0.5) to obtain the final loss of the diffusion branch:

[0141] L LDM =L noise +0.5L recon .

[0142] Joint Optimization with Contrastive Learning:

[0143] The final optimization target is the sum of the contrast branch (SupConLoss) and the potential diffusion branch loss:

[0144] L=L SupCon +L LDM

[0145] In each iteration, only the parameters of the projection head, encoder, UNet, and decoder are updated to ensure that the generative and contrastive signals jointly improve the discriminability and generativeness of the latent representation.

[0146] In summary, the feature extraction of this embodiment:

[0147] Sequence and structure feature extraction:

[0148] In the feature extraction stage, this method extracts representations from both the linear sequence and the three-dimensional structure of the peptide in parallel: on the one hand, the amino acid sequence is input into the pre-trained Prot-T5 model, and after multi-layer self-attention calculation, the residue dimension is average pooled to obtain a global sequence representation with a shape of (B, 1024); on the other hand, RDKit is used to convert the same sequence into three-dimensional coordinates, extract the Pearson correlation coefficient matrix between atoms, and flatten it after scaling to 32×32 through grayscale interpolation to obtain a structure vector with the same shape of (B, 1024).

[0149] Supervised comparison branch (SupCon):

[0150] Each feature vector is subjected to two independent dropouts (0.1) to obtain two sets of corresponding views. After passing through two layers of fully connected mapping networks and L2 normalization, unit-length projection features are generated. In the 2B projection vectors, two perturbations of the same peptide are considered as positive sample pairs, and all others are negative sample pairs. Calculate the dot product similarity between any two vectors and divide it by the temperature factor τ = 0.07, set the diagonal to zero, and perform an exponential operation. Take the negative logarithm of the ratio of the similarity index of the positive sample pair to the sum of the indices of all negative sample pairs in the same row; average all 2B samples to form the supervised contrast loss.

[0151] Latent Diffusion Branch (LDM):

[0152] Map each fusion vector to d z =1024-dimensional latent variables, and expand the fused features to a 1024-dimensional latent vector through a two-layer fully connected sub-network. First layer: linear mapping projects the feature vector input to 256 dimensions and applies the SiLU activation function to the projection result. Second layer: linear mapping maps the 256-dimensional intermediate representation to 1024 dimensions. Randomly sample the diffusion steps t∈[0,T)(T=1000) and generate normal noise of the same dimension. According to the cumulative retention coefficient The original latent variable is mixed with noise in proportion to obtain a noisy intermediate state. After concatenating the noisy latent variable with the normalized time step t / T, it is fed into a "U-net" fully connected network (multi-layer ReLU) to predict the noise component. The mean squared error between the predicted noise and the true noise is used as the noise reconstruction loss. The predicted noise is used to infer the initial latent variable estimate, which is then restored to the fused vector space through a symmetrical two-layer fully connected "decoding" network. The shape is restored, and the mean squared error between the decoded vector and the original fused vector is used as the reconstruction loss. The noise reconstruction loss and the vector reconstruction loss are weighted and summed by a coefficient (1:0.5) to form the diffusion branch loss.

[0153] Joint optimization and parameter update:

[0154] The supervised contrast loss is directly added to the diffusion branch loss to get the total loss for the current batch. Only the projection head and diffusion branch network parameters are backpropagated and optimized with Adam (learning rate is about 0.0001).

[0155] Latent representation extraction:

[0156] After the training, the (N, 1024)-dimensional latent variables output by the encoder were taken as the final latent features for subsequent antihypertensive activity prediction or other machine learning tasks.

[0157] Furthermore, peptide classification is performed on the enhanced feature representation using a preset Kan-Conv structure and label smoothing joint optimization classification model, including:

[0158] Perform dimension splicing on the enhanced feature representation to generate merged features of preset size;

[0159] For the merged features, several parallel adaptive convolution kernel channels are applied to perform batch normalization respectively, and then nonlinearity is introduced through the ReLU function.

[0160] Use the max pooling layer to flatten all channels and spatial dimensions into a long vector, generating a one-dimensional feature vector of a preset size;

[0161] The one-dimensional feature vector is mapped through two layers of full connection to obtain the original classification score;

[0162] Applying a log-softmax operation to the raw classification scores yields a logarithmic probability distribution of two classes for each sample; wherein the two classes refer to anti-stress peptides and non-anti-stress peptides;

[0163] Based on the logarithmic probability distribution, the peptide segments are classified.

[0164] Furthermore, the two-layer fully connected mapping includes:

[0165] First layer: compress the one-dimensional feature vector of a preset size and apply ReLU to obtain the compressed hidden representation;

[0166] The second layer maps the compressed hidden representation to the preset number of categories and outputs the original classification score.

[0167] Specifically, in this embodiment, the classification model is optimized jointly based on the Kan-Conv structure and label smoothing:

[0168] (1) Loading and preparing data:

[0169] Read mini-batches of data from the data loader. Each tensor has a shape of (B, 1024), where B is the batch size. Each row corresponds to a 1024-dimensional latent feature for a peptide. The corresponding labels tensor has a shape of (B,), where each element is an integer 0 or 1, indicating the peptide's class. Each batch is composed of 64 randomly sampled samples.

[0170] (2) Model structure:

[0171] Network structure:

[0172] First, these two tensors are concatenated along the final "width" dimension, generating a merged feature of size (B, 1, 1, 2048). Five parallel 3×3 adaptive convolution kernels (KANConv) are applied to the merged feature with a stride of (2, 2) and edge padding of 1, resulting in an output tensor of (B, 5, 1, 1024). Batch normalization is performed on each of the five channels, followed by a ReLU nonlinearity, maintaining the size (B, 5, 1, 1024). A max pooling layer of size 1×2 with a stride of (2, 2) is used. All channels and spatial dimensions are flattened into a single long vector, generating a one-dimensional feature vector of size (B, 2560).

[0173] Two-layer fully connected mapping: The first layer compresses the 2560-dimensional input to 256 dimensions and applies ReLU to it, resulting in (B, 256). The second layer maps the 256-dimensional hidden representation to 2 categories, outputting the raw classification scores of size (B, 2). Finally, a log-softmax operation is applied to these scores to obtain the logarithmic probability distribution of each sample corresponding to the two categories, which is used in the subsequent loss calculation based on cross entropy or KL divergence.

[0174] Loss function and label smoothing:

[0175] Hard label cross entropy: Take the output probability of the corresponding category for each sample, log it and take the negative, then average it over the entire batch to get the traditional cross entropy loss L CE , the expression is:

[0176]

[0177] KL divergence after label smoothing: Slightly smooth the one-hot hard label: replace "1" with 0.9 and "0" with 0.1 to obtain the smoothed target distribution. Softmax the model output and take the logarithm. Multiply the product with the element-wise difference of the smoothed distribution and then average it over the batch to obtain the KL divergence loss L. KL , the expression is:

[0178]

[0179] Total loss: L = L CE +L KL .

[0180] (3) Backpropagation and parameter update

[0181] First, all accumulated gradients from the previous round are reset to zero to prevent gradient accumulation. Gradients are automatically calculated for all trainable parameters in the convolutional and fully connected layers based on the total loss. The Adam algorithm is used with a learning rate of 0.0001, and parameter values ​​are updated based on the learning rate and first-order and second-order momentum information.

[0182] (4) Backpropagation and parameter update:

[0183] After each round of training, the model is switched to evaluation mode, and the classification accuracy is calculated using the reserved validation set to monitor performance changes. After the final training is completed, the model parameters at the optimal moment are persisted to the disk for subsequent prediction or deployment.

[0184] The model of the application is compared with the anti-hypertensive peptide prediction model, and significant performance improvement is achieved in multiple evaluation indicators. The evaluation indicators include specificity (Specificity, SEP), sensitivity (Sensitivity, SEN), accuracy (Accuracy, ACC), Matthews correlation coefficient (Matthews correlation coefficient, MCC), F1-score (F1), area under the curve (AUC) and average accuracy (Average precise, AP), which are defined as follows.

[0185]

[0186] To comprehensively evaluate the performance of DeepATHPFuse, we compared it with seven mainstream methods on the same training set: AHTpin takes amino acid composition, atomic composition, chemical descriptors and G-Scales features as input, and uses support vector machine classification. AHPred-DE integrates UniRep pre-trained embedding, One-Hot encoding and K-mer frequency, and realizes prediction through deep ensemble model. LSTM-AHP is a multi-layer LSTM sequence classifier that extracts time sequence features from the input sequence and maps the hidden state of the last time step through a fully connected layer to a dimension to complete classification. PAAP uses amino acid composition, dipeptide composition and pseudo-amino acid composition features, and classifies based on random forest. pLM4ACE relies on sequence embedding output by ESM-2 large-scale protein language model, and cooperates with SVM to complete discrimination. Ensemble-AHTPpred integrates amino acid composition, sequence order descriptor and secondary structure conformation features, and through the weighted voting of random forest, support vector machine and XGBoost three algorithms, it combines the advantages of each model. AHPNG is based on dipeptide frequency and amino acid physicochemical properties, and builds a deep learning framework combining convolutional neural network and gated recurrent unit. All models are verified for performance on a unified independent test set to ensure the objectivity and fairness of the evaluation results.

[0187] A comprehensive analysis of the performance indicators of DeepATHPFuse and the other seven mainstream models shows that DeepATHPFuse has obvious advantages in all indicators. Specifically, the accuracy (ACC) of each model ranges from mAHTPred (0.7901) to DeepATHPFuse (0.8651), with DeepATHPFuse significantly ahead; the Matthews correlation coefficient (MCC) of each model ranges from mAHTPred (0.4831) to DeepATHPFuse (0.6945), and the value of DeepATHPFuse far exceeds that of other models, showing a significant advantage in positive and negative sample classification; the performance of F1 score ranges from mAHTPred (0.5952) to DeepATHPFuse (0.7934), indicating that DeepATHPFuse is the best in balancing precision and recall; the sensitivity (SE) ranges from mAHTPred (0.5051) to AHTpin (0.8452), with DeepATHPFuse e reached (0.8231), which is at a high level; the specificity (SP) ranged from AHTpin (0.7252) to mAHTPred (0.9211), and DeepATHPFuse ranked at the top with a performance of (0.8844), achieving effective recognition of negative samples; in the area under the ROC curve (AUC) indicator for comprehensively evaluating the model's discrimination ability, the range of each model was LSTM-AHP (0.8370) to DeepATHPFuse (0.9262), and DeepATHPFuse performed best, reflecting excellent classification threshold robustness; finally, in the average precision (AP) indicator that measures the model accuracy under high recall conditions, the value range of each model was from mAHTPred (0.5346) to DeepATHPFuse (0.8648), and DeepATHPFuse once again performed the best. Overall, DeepATHPFuse ranked first in all seven evaluation indicators, significantly surpassing all other methods, demonstrating comprehensive performance advantages and robustness in the field of antihypertensive peptide prediction. Specific data are shown in Table 1:

[0188] Table 1 Comparison of this embodiment with other methods on the data set

[0189]

[0190] Figure 2 A comparison diagram of the AUROC curve and a schematic diagram of the PRROC curve of the embodiment of the present invention and other methods are shown; Figure 3 AUC value distribution comparison diagram of the embodiment of the present invention and other comparison methods; Figure 4 Schematic diagram comparing the MCC value distribution of the embodiment of the present invention and other comparison methods.

[0191] In summary, the present embodiment proposes:

[0192] (1) Sequence feature representation. Prot-t5_xl_uniref50 is a high-level protein sequence encoding model based on the T5 (text-to-text transfer transformer) architecture developed by Google, which is specifically optimized for bioinformatics applications. The model is trained on the UniRef50 dataset, which contains de-duplicated protein sequences representing millions of different proteins, providing a comprehensive background for protein sequence analysis. The main advantage of this model is its Transformer-based architecture, which can capture complex contextual relationships in the input sequence and effectively model long-range dependencies.

[0193] (2) Structure feature representation. For the structural information of antihypertensive peptides, the present embodiment completely adopts the coordinate correlation features calculated by RDKit: first, a three-dimensional small molecule model containing hydrogen is generated in RDKit according to the amino acid sequence of the peptide, and then the coordinates of all atoms in three-dimensional space are parsed. Then calculate the Pearson correlation coefficient matrix of these coordinate points in three dimensions to get an original correlation matrix. Replace NaN and infinite values in the matrix, then scale it proportionally and bilinearly interpolate to a fixed resolution to generate a grayscale heat map. Finally, this image is unfolded in row-major order into a one-dimensional vector of length 1024, which is used as the structural representation input for the downstream antihypertensive peptide activity prediction model.

[0194] (3) The present embodiment proposes a dual-path multi-modal feature enhancement method based on contrastive-generative joint optimization: first, obtain the sequence representation of the peptide segment through the large-scale protein language model ProtT5; At the same time, only use RDKit to directly extract the structure feature vector from the three-dimensional conformation of the peptide; Then send these two groups of vectors into their respective projection heads and latent diffusion branches, independently perform supervised contrastive learning and generative noise reconstruction, and calculate the contrastive loss and diffusion reconstruction loss in sequence view and structure view respectively; Finally, the two types of losses are weighted and fused with adjustable weights, and the ProtT5 encoder, projection head and diffusion model parameters are updated synchronously through backpropagation, so as to output the enhanced sequence features and structure features respectively.

[0195] (4) This embodiment proposes a method for classifying antihypertensive peptides based on a single-pass KanConv structure and label smoothing joint optimization: the sequence features and structural features obtained through contrastive-generative joint optimization are spliced ​​in the width dimension and then sent to a single KanConv convolution module. The module consists of a KanConv convolution layer (followed by BatchNorm and ReLU) and a maximum pooling layer in sequence, followed by flattening the pooling output, and generating the final binary classification logits through two layers of fully connected layers. During the training phase, the network uses cross-entropy loss to supervise the hard labels, constructs soft labels through smoothing coefficients, and measures the consistency between the predicted distribution and the soft label distribution using KL divergence. The two are proportionally fused into a joint loss, and the convolution module and fully connected layer parameters are simultaneously optimized through backpropagation, thereby significantly improving the discrimination ability and generalization performance of antihypertensive peptides and non-antihypertensive peptides.

[0196] This example combines Prot-T5 sequence embedding with RDKit structure embedding, leveraging Prot-T5's advantages in efficiently capturing global contextual information and accurately expressing spatial structural features. It also employs post-projection supervised comparison and latent diffusion generative optimization to significantly enhance the discriminability and information integrity of composite features. Furthermore, based on the KanConv convolutional network's expertise in multi-scale neighborhood feature processing, combined with a 0.1 label smoothing classification strategy, it effectively suppresses model overconfidence and improves generalization. Its beneficial effects are as follows:

[0197] (1) This example combines the global contextual features of the sequence obtained by Prot-T5 with the three-dimensional structural spatial neighboring information extracted by RDKit, achieving deep complementarity of multimodal information. Sequence features are good at capturing local functional fragments and long-range dependencies of the peptide chain, while structural features accurately reflect the spatial distribution and physicochemical interactions between residues. The combination of the two can effectively improve the ability to identify key active sites and enhance robustness to sequence variation or conformational noise.

[0198] (2) This embodiment adopts supervised contrastive learning (SupCon) in the contrastive learning link: random masking or amino acid substitution perturbation is performed on sequence embeddings, node dropping or edge weight noise perturbation is performed on structural embeddings, multi-view positive samples are generated, and mapped to the contrast space through a two-layer MLP projection head to shrink the distance between similar samples and widen the boundaries between heterogeneous samples; the conditional latent diffusion model (LDM) is introduced in the generative learning link: noise is added to the features according to linear noise scheduling, and the U-Net noise prediction network combining temporal coding and conditional injection is used for denoising and reconstruction to fill the representation of low-density areas of high-dimensional distribution, rich edges and rare samples. The losses of the two methods act synergistically on the same encoder during back propagation, so that the final output composite features have both the category discrimination enhanced by SupCon and the distribution continuity and information richness brought by LDM, thereby significantly improving the model's robustness and discrimination accuracy to noise interference, sequence variation and small sample scenarios.

[0199] (3) This embodiment uses the KanConv network. This network adopts a parallel multi-core design, which can simultaneously extract sequence and structural features at different scales, and approximate complex high-dimensional mappings with very few parameters based on the Kolmogorov-Arnold theory; its multi-branch adaptive fusion mechanism highlights the interactions and spatial patterns of key residues, has a natural inhibitory ability for noise and distortion channels, and avoids overfitting of a single convolution kernel; at the same time, this structure maintains efficient parallel computing while ensuring strong nonlinear expression, thereby significantly improving the robustness, discrimination performance and computational efficiency of the antihypertensive peptide prediction model in sequence variation and small sample environments.

[0200] (4) This embodiment uses hard label cross entropy loss to ensure that the model has strong discrimination ability for real categories, while the soft label distribution constructed based on the smoothing coefficient 0.1 and constrained by KL divergence loss can effectively alleviate the overconfidence of the model and improve the calibration of the output probability; the two are jointly optimized, while maintaining classification accuracy, introducing smoothing regularization, reducing the risk of overfitting in noise and few-sample scenarios, accelerating training convergence and enhancing the generalization ability and robustness of the model.

[0201] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for classifying antihypertensive peptides that integrates sequence and structural multimodal features and combines comparative-generative joint optimization, characterized in that: include: Extract sequence feature representation and structural feature representation of peptide segments; Performing multimodal feature enhancement on the sequence feature representation and the structural feature representation by contrastive-generative joint optimization; The enhanced feature representation is used to perform peptide classification using a preset classification model based on the joint optimization of Kan-Conv structure and label smoothing.

2. The antihypertensive peptide classification method according to claim 1, characterized in that: The sequence feature representation and structural feature representation of extracted peptides include: The feature representation of peptides is extracted based on a preset large-scale protein language model, and the structural feature representation generated by RDKit is also utilized.

3. The antihypertensive peptide classification method according to claim 2, characterized in that: The structural feature representations generated using RDKit include: Mapping the peptide segment into a three-dimensional structural model, filling all implicit hydrogen atoms in the three-dimensional structural model, and obtaining structural data including the three-dimensional coordinate information of all atoms through a molecular conformation embedding algorithm; Extracting atomic coordinates from the structural data to obtain an atomic coordinate matrix; Calculating the Pearson correlation coefficient between the three-dimensional coordinate dimensions based on the atomic coordinate matrix to obtain a correlation coefficient matrix; The correlation coefficient matrix is ​​numerically normalized, image-converted, and feature vectors are generated in sequence to obtain a structural feature representation.

4. The antihypertensive peptide classification method according to claim 2, characterized in that: The multimodal feature enhancement of the sequence feature representation and the structural feature representation by performing contrastive-generative joint optimization includes: Performing supervised contrastive learning on the sequence feature representation and the structural feature representation to obtain a sequence feature representation and a structural feature representation enhanced by contrastive learning; Generative learning is performed on the sequence feature representation and the structural feature representation to obtain the sequence feature representation and the structural feature representation enhanced by generative learning.

5. The antihypertensive peptide classification method according to claim 4, characterized in that: The supervised comparative learning of the sequence feature representation and the structural feature representation includes: Randomly perturb the sequence feature representation and the structural feature representation, and sequentially concatenate the perturbed feature representations up and down to form a feature matrix; Calculate the dot product similarity matrix between each pair in the feature matrix; Construct a mask matrix of the same scale as the similarity matrix, and multiply the dot product similarity matrix by the mask matrix after element-by-element exponentiation. Sum all non-zero elements in each row of the multiplied matrix to obtain the denominator of the supervised contrast loss; where each row in the multiplied matrix refers to each enhanced sample; According to the definition of the positive sample pair, the exponential similarity between each enhanced sample and the positive sample in the multiplied matrix is ​​extracted to form the numerator of the supervised contrast loss; For each enhanced sample in the multiplied matrix, the contrast loss is calculated sample by sample according to the denominator and numerator of the corresponding supervised contrast loss and averaged to obtain the final supervised contrast loss; The final supervised contrastive loss is used to obtain the sequence feature representation and structural feature representation enhanced by contrastive learning.

6. The antihypertensive peptide classification method according to claim 4, characterized in that: Generative learning of the sequence feature representation and the structural feature representation includes: Mapping the sequence feature representation and the structural feature representation to a latent space to obtain initial latent variables; Injecting noise into the initial latent variable to obtain a noisy latent variable; The noisy latent variables are concatenated with the normalized time steps and then fed into the preset fully connected network to output the prediction of the noise. Use the predicted noise to infer the initial latent variables and remove the noise components; The recovered latent variables are mapped back to the original dimension through two layers of full connection to obtain the reconstructed feature vector.

7. The antihypertensive peptide classification method according to claim 4, characterized in that: Using the preset Kan-Conv structure and label smoothing joint optimization classification model, peptide classification is performed on the enhanced feature representation, including: Perform dimension splicing on the enhanced feature representation to generate merged features of preset size; For the merged features, several parallel adaptive convolution kernel channels are applied to perform batch normalization respectively, and then nonlinearity is introduced through the ReLU function. Use the max pooling layer to flatten all channels and spatial dimensions into a long vector, generating a one-dimensional feature vector of a preset size; The one-dimensional feature vector is mapped through two layers of full connection to obtain the original classification score; Applying a log-softmax operation to the raw classification scores yields a logarithmic probability distribution of two classes for each sample; wherein the two classes refer to anti-stress peptides and non-anti-stress peptides; Based on the logarithmic probability distribution, the peptide segments are classified.

8. The antihypertensive peptide classification method according to claim 7, which combines sequence and structure multimodal features with comparative-generative joint optimization, is characterized in that: Through two layers of fully connected mapping including: First layer: compress the one-dimensional feature vector of a preset size and apply ReLU to obtain the compressed hidden representation; The second layer maps the compressed hidden representation to the preset number of categories and outputs the original classification score.

Images