A method for classifying antihypertensive peptides by fusing sequence and structure multi-modal features and combining contrastive generative joint optimization

Abstract

The present application relates to a kind of fusion sequence and structural multimodal feature and the classification method of antihypertensive peptide combined with contrast-generative joint optimization, comprising: the sequence characteristic representation and structural characteristic representation of peptide segment are extracted;The sequence characteristic representation and structural characteristic representation are contrast-generative joint optimization multimodal feature enhancement;Using preset Kan-Conv structure and label smoothing joint optimization classification model based on, the enhanced feature representation is classified into peptide.This application realizes the high-precision identification of antihypertensive peptide functional activity, this method innovatively utilizes three aspects of information, such as sequence, structure and generative latent space, fills the blank of ignoring structure-sequence synergistic effect in the existing peptide function prediction field, significantly improves the screening efficiency and prediction reliability of antihypertensive peptide.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics and artificial intelligence, and in particular to a method for classifying antihypertensive peptides that integrates sequence and structural multimodal features and combines contrastive-generative joint optimization. Background Technology

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

[0003] Furthermore, sequence and structural features are often modeled independently in traditional strategies, lacking an end-to-end multi-view collaborative learning framework, making it difficult for models to capture the intrinsic relationship between the two. Meanwhile, while recent deep generative models have achieved significant success in image and small molecule design, they have not yet been fully applied to peptide-level functionalization prediction. Summary of the Invention

[0004] The purpose of this invention is to propose a classification method for antihypertensive peptides that integrates multimodal features of sequence and structure with contrastive-generative joint optimization, thereby addressing the problems existing in the prior art. By fusing multimodal views of peptide sequence information and structural features, and then introducing supervised contrastive learning and latent diffusion model branches for joint optimization, this method achieves precise encoding of high-dimensional features and low-dimensional latent representation of peptide molecules. The end-to-end learning framework constructed in this invention not only significantly improves the accuracy of peptide function prediction but also possesses good generalization ability, making it suitable for the discovery, classification, and design of functional peptide molecules such as antihypertensive peptides.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for classifying antihypertensive peptides that integrates sequence and structural multimodal features and combines contrastive-generative joint optimization includes:

[0007] Sequence and structural feature representations of extracted peptides;

[0008] Multimodal feature enhancement is performed on the sequence feature representation and structural feature representation through contrast-generative joint optimization.

[0009] A pre-defined classification model based on Kan-Conv structure and label smoothing is used to classify peptides based on the enhanced feature representation.

[0010] Optionally, the extracted sequence feature representation and structural feature representation of the peptide include:

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

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

[0013] The peptide is mapped to a three-dimensional structural model, all implicit hydrogen atoms are added to the three-dimensional structural model, and structural data including the three-dimensional coordinate information of all atoms is obtained through a molecular conformation embedding algorithm.

[0014] Atomic coordinates are extracted from the structural data to obtain an atomic coordinate matrix;

[0015] Calculate the Pearson correlation coefficients between the three-dimensional coordinate dimensions based on the atomic coordinate matrix, and obtain the correlation coefficient matrix;

[0016] The correlation coefficient matrix is ​​then subjected to numerical normalization, image transformation, and feature vector generation to obtain a structural feature representation.

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

[0018] Supervised contrastive learning is performed on the sequence feature representation and structural feature representation to obtain contrastive learning-enhanced sequence feature representation and structural feature representation;

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

[0020] Optionally, supervised contrastive learning of the sequence feature representation and structural feature representation includes:

[0021] The sequence feature representation and structural feature representation are randomly perturbed, and the perturbed feature representations are then concatenated sequentially to form a feature matrix;

[0022] Calculate the pairwise dot product similarity matrix between the features;

[0023] Construct a mask matrix with the same scale as the similarity matrix, and multiply the dot product similarity matrix by the element-wise exponentialized mask matrix.

[0024] Summing all non-zero elements in each row of the multiplied matrix yields the denominator of the supervised contrastive loss; where each row of the multiplied matrix represents each augmented sample.

[0025] Based on the definition of positive sample pairs, 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.

[0026] For each augmented sample in the multiplied matrix, the contrastive loss is calculated sample by sample based on the denominator and numerator of the corresponding supervised contrastive loss and averaged to obtain the final supervised contrastive loss.

[0027] Using the final supervised contrastive loss, we obtain the sequence feature representation and structural feature representation enhanced by contrastive learning.

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

[0029] The sequence feature representation and structural feature representation are mapped to the latent space to obtain initial latent variables;

[0030] Noise is injected into the initial latent variables to obtain noisy latent variables;

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

[0032] The initial latent variables are inferred from the predicted noise, and the noise component is removed.

[0033] The recovered latent variables are mapped back to the original dimensions through two fully connected layers to obtain the reconstructed feature vectors.

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

[0035] The enhanced feature representations are dimensionally concatenated to generate merged features of a preset size;

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

[0037] Max pooling layers are used to flatten all channels and spatial dimensions into a long vector, generating a one-dimensional feature vector of a preset size.

[0038] The original classification score is obtained by mapping the one-dimensional feature vector through two fully connected layers.

[0039] The original classification scores are then subjected to a log-softmax operation to obtain the log probability distribution of each sample for the two classes; where the two classes refer to anti-stress peptides and non-stress peptides.

[0040] Based on the logarithmic probability distribution, peptides 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 a preset number of categories and outputs the original classification scores.

[0044] The beneficial effects of this invention are as follows:

[0045] This invention first extracts sequence and structural feature representations of peptides; then, it performs multimodal feature enhancement through contrastive-generative joint optimization on these representations; finally, it uses a pre-defined classification model based on Kan-Conv structure and label smoothing joint optimization to classify the enhanced feature representations into peptides. This invention achieves precise encoding of high-dimensional features and low-dimensional latent representations of peptide molecules by fusing sequence information and structural features of peptides into multiple views, and then introducing supervised contrastive learning and latent diffusion model branches for joint optimization. This innovative approach integrates sequence, structure, and generative latent space information, filling the gap in existing peptide function prediction fields that neglect the structure-sequence synergy effect, and significantly improving the screening efficiency and prediction reliability of antihypertensive peptides. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram 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 The above are comparison charts of AUROC curves and PRROC curves between embodiments of the present invention and other methods.

[0049] Figure 3 This is a comparison of the AUC value distribution of this invention embodiment with other comparative methods;

[0050] Figure 4 This is a comparison of the MCC value distribution of this invention embodiment with other comparative methods. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] This embodiment proposes a classification method for antihypertensive peptides that integrates sequence and structural multimodal features with contrastive-generative joint optimization, including:

[0054] Sequence and structural feature representations of extracted peptides;

[0055] Multimodal feature enhancement is performed on the sequence feature representation and structural feature representation through contrast-generative joint optimization.

[0056] A pre-defined classification model based on Kan-Conv structure and label smoothing is used to classify peptides based on the enhanced feature representation.

[0057] Specifically, this embodiment proposes a method for classifying antihypertensive peptides by fusing sequence and structural multimodal features and combining contrastive-generative joint optimization, for accurate prediction of antihypertensive peptide sequences. DeepATHPFuse first extracts the global context embedding of the peptide sequence using Prot-T5, and then uses RDKit to obtain spatial graph embeddings based on three-dimensional structures. Supervised contrastive learning (SupCon) and conditional latent diffusion model (LDM) are then introduced to jointly optimize the features, simultaneously enhancing class discriminativeness and distribution continuity. The optimized features are then input into a classification network composed of multi-kernel KanConv convolutional layers and fully connected layers, and are jointly supervised by hard-label cross-entropy and KL divergence based on 0.1 label smoothing, thereby effectively suppressing overconfidence and improving generalization and robustness. This embodiment demonstrates high accuracy, high stability, and good scalability in the antihypertensive peptide prediction task, and is suitable for applications such as functional peptide screening and new drug development. The specific implementation process is as follows: Figure 1 As shown.

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

[0059] Furthermore, the extraction of sequence feature representation and structural feature representation of peptides includes:

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

[0061] Furthermore, the structural feature representations generated using RDKit include:

[0062] The peptide is mapped to a three-dimensional structural model, all implicit hydrogen atoms are added to the three-dimensional structural model, and structural data including the three-dimensional coordinate information of all atoms is obtained through a molecular conformation embedding algorithm.

[0063] Atomic coordinates are extracted from the structural data to obtain an atomic coordinate matrix;

[0064] Calculate the Pearson correlation coefficients between the three-dimensional coordinate dimensions based on the atomic coordinate matrix, and obtain the correlation coefficient matrix;

[0065] The correlation coefficient matrix is ​​then subjected to numerical normalization, image transformation, and feature vector generation to obtain a structural feature representation.

[0066] Specifically, in this embodiment, the antihypertensive peptide sequence data is first collected and preprocessed:

[0067] The primary data source for this embodiment is the Food-derived Bioactive Peptide Database (DFBP), which focuses on bioactive peptides extracted from food sources, including those with antihypertensive properties. Additionally, a dataset compiled by Du et al. in pLM4ACE was used, which includes sequences, IC50 values, and original references for all collected peptides. Many existing datasets used in current methods contain a large number of short peptide sequences, especially dipeptides and tripeptides. These short peptides have been observed to have relatively high predictive accuracy. However, despite their good performance metrics, this embodiment argues that their presence may affect the model's generalization ability and practical applicability. Consequently, peptides shorter than 5 residues were removed, and CD-HIT was applied to filter out sequences with a similarity exceeding 90%. Finally, a dataset consisting of 732 positive samples and 1601 negative samples was obtained.

[0068] Sequence feature representation:

[0069] Pre-trained models learn high-level representations of peptide sequences based on large-scale protein datasets, enabling them to deeply capture semantic information and contextual dependencies. In protein research, Prot-T5 is a widely used pre-trained model, based on the Transformer architecture, possessing advantages in handling long sequences and powerful feature modeling capabilities. Utilizing a self-attention mechanism, Prot-T5 can capture global associations between long-range residues in a peptide chain and encode a sequence of length L into an L×1024-dimensional feature matrix, thus comprehensively representing the complex features and interactions of the sequence. The specific implementation process is as follows: Figure 1 As shown.

[0070] Structural feature representation:

[0071] Sequence input and 3D model 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 x, y, and z coordinate values ​​are parsed from each record in turn and collected into three coordinate sets, which are then combined into an atomic coordinate matrix.

[0075] Correlation coefficient matrix calculation and numerical normalization:

[0076] The Pearson correlation coefficient between the 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 transformation and feature vector generation:

[0078] The normalized correlation coefficient matrix is ​​treated as a grayscale image and then interpolated to a uniform size. Subsequently, the pixel values ​​of the magnified image are flattened into a one-dimensional numerical vector in a predetermined order, which serves as the structural correlation coefficient feature output of the peptide.

[0079] Furthermore, the multimodal feature enhancement, which involves contrastive-generative joint optimization of the sequence feature representation and structural feature representation, includes:

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

[0081] Generative learning is performed on the sequence feature representation and structural feature representation to obtain generatively enhanced sequence feature representation and structural feature representation.

[0082] Furthermore, supervised contrastive learning of the sequence feature representation and structural feature representation includes:

[0083] The sequence feature representation and structural feature representation are randomly perturbed, and the perturbed feature representations are then concatenated sequentially to form a feature matrix;

[0084] Calculate the pairwise dot product similarity matrix between the features;

[0085] Construct a mask matrix with the same scale as the similarity matrix, and multiply the dot product similarity matrix by the element-wise exponentialized mask matrix.

[0086] Summing all non-zero elements in each row of the multiplied matrix yields the denominator of the supervised contrastive loss; where each row of the multiplied matrix represents each augmented sample.

[0087] Based on the definition of positive sample pairs, 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 augmented sample in the multiplied matrix, the contrastive loss is calculated sample by sample based on the denominator and numerator of the corresponding supervised contrastive loss and averaged to obtain the final supervised contrastive loss.

[0089] Using the final supervised contrastive loss, we obtain the sequence feature representation and structural feature representation enhanced by contrastive learning.

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

[0091] Input data and view construction:

[0092] Assume a batch contains N peptide sequences, each sequence undergoes two independent random perturbations (or different Dropout / data augmentation strategies) to obtain two sets of augmented features. These two sets of representations are concatenated sequentially 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 referred to as the views. Figure 1 Heshi Figure 2 .

[0093] Pairwise similarity calculation:

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

[0095]

[0096] In this matrix, the value in the i-th row and j-th column represents the phase velocity of the i-th vector and the j-th vector. This matrix is ​​symmetric and quantifies the "closeness" of each pair of augmented representations. Since only relative magnitude is considered subsequently, and the influence of vector length is eliminated in the exponential operation, the features are usually L2 normalized before applying this loss.

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

[0098] To exclude comparisons between each feature vector and itself (i.e., diagonal terms), a mask matrix with the same scale as the similarity matrix is ​​constructed: diagonal elements are set to 0, and off-diagonal elements are set to 1. The similarity matrix is ​​then multiplied element-wise by this mask, i.e.:

[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 augmented sample), sum all non-zero elements to obtain the denominator of the InfoNCE loss:

[0103]

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

[0105] Positive sample pairs are defined as two enhanced views of the same original sample: for the first half of the matrix (view... Figure 1 The positive samples in the i-th row are in the view. Figure 2 The i-th row. For the second half (see...), Figure 2 The positive samples in the i-th row are in the view. Figure 1 The i-th row. Represent this correspondence using the mapping function p(i). Extract the exponential similarity between each enhanced sample and its positive sample, forming the molecule in InfoNCE:

[0106]

[0107] Calculate and average the contrast loss for each sample:

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

[0109]

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

[0111]

[0112] Furthermore, generative learning of the sequence feature representation and structural feature representation includes:

[0113] The sequence feature representation and structural feature representation are mapped to the latent space to obtain initial latent variables;

[0114] Noise is injected into the initial latent variables to obtain noisy latent variables;

[0115] The noisy latent variables are concatenated with the normalized time steps and then input into a preset fully connected network to output a prediction of the noise.

[0116] The initial latent variables are inferred from the predicted noise, and the noise component is removed.

[0117] The recovered latent variables are mapped back to the original dimensions through two fully connected layers to obtain the reconstructed feature vectors.

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

[0119] In the Latent Diffusion Branch, this embodiment uses the following specific parameters to perform generative learning optimization on the multimodal features:

[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 hyperparameter. Initial latent variables:

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

[0123] Diffusion coefficient scheduling:

[0124] Assume a total diffusion step count T = 1000. Initial and final noise variance: β1 = 10. -4 ,β T =0.02. β t In

[10] -4 [0.02] is generated at equal intervals, representing the intensity of the injected noise at step t. The cumulative retention coefficient is calculated. Used to smoothly interpolate the original signal and noise in a closed-form formula.

[0125] Injecting noise into latent variables:

[0126] For each batch of samples (batch size B), randomly select a number of steps. Generate standard normal noise ε. Construct a noisy latent variable using a closed-form formula:

[0127]

[0128] This formula guarantees that, as the limit t→T, z t Approximately pure noise.

[0129] The UNet branch is used 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 containing ReLU, and output the noise prediction. Measure the difference between the predicted noise and the actual noise using mean squared error.

[0131]

[0132] Recovery of noiseless latent variables:

[0133] Formula for inferring initial latent variables using predictive noise: This operation is equivalent to a first-order inverse diffusion approximation, which aims to remove noise components.

[0134] Reconstructing high-dimensional features from latent variables:

[0135] The recovered latent variables are mapped back to the original dimensions through two fully connected layers:

[0136]

[0137] The corresponding reconstruction loss is also expressed using mean squared error:

[0138]

[0139] Total branch loss and weights:

[0140] The noise prediction loss and reconstruction loss are weighted and summed according to their respective weights (typically a ratio of 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 ultimate optimization objective is the sum of the losses from the contrast branch (SupConLoss) and the potential diffusion branch:

[0144] L = L SupCon +L LDM

[0145] In each iteration, only the parameters of the projector, encoder, UNet, and decoder are updated to ensure that the generative and contrastive signals together improve the discriminative and generative properties of the latent representation.

[0146] In summary, the feature extraction in this embodiment is as follows:

[0147] Sequence and structural feature extraction:

[0148] In the feature extraction stage, this method extracts representations from both the linear sequence and 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, average pooling is performed on the residue dimension to obtain a global sequence representation of shape (B, 1024); on the other hand, the same sequence is converted into three-dimensional coordinates using RDKit, the Pearson correlation coefficient matrix between atoms is extracted, and it is flattened after being scaled to 32×32 by grayscale interpolation to obtain a structure vector of the same shape (B, 1024).

[0149] Supervised comparison branch (SupCon):

[0150] For each feature vector, perform two independent Dropout(0.1) operations to obtain two sets of corresponding views. After passing through two fully connected mapping layers and L2 normalization, generate unit-length projected features. In the 2B projected vectors, two perturbations of the same peptide segment are considered positive sample pairs, and all others are negative sample pairs. Calculate the dot product similarity of any two vectors and divide by the temperature factor τ = 0.07, setting the diagonal to zero, and perform an exponential operation. The ratio of the positive sample pair similarity index to the sum of the indices of all negative sample pairs in the same row is taken as the negative logarithm; the average is calculated over all 2B samples to form the supervised contrastive loss.

[0151] Potential Diffusion Branch (LDM):

[0152] Map each fusion vector to d z The latent variables are 1024-dimensional, and the fused features are extended to 1024-dimensional latent vectors through a two-layer fully connected subnetwork. The first layer: a linear mapping projects the feature vector input to 256 dimensions, and applies the SiLU activation function to the projection result. The second layer: a linear mapping maps the 256-dimensional intermediate representation to 1024 dimensions. Random sampling diffusion steps t∈[0,T) (T=1000) are performed, and normal noise of the same dimension is generated. Based on the cumulative retention coefficient... The original latent variables are mixed with noise in a proportional manner to obtain a noisy intermediate state. The noisy latent variables are concatenated with the normalized time step t / T and then input into a "U-net" type fully connected network (multi-layer ReLU) to predict the noise component. The mean square error between the predicted noise and the actual noise is used as the noise reconstruction loss. The predicted noise is used to back-estimate the initial latent variables, and then a symmetrical two-layer fully connected "decoding" network is used to restore the fused vector space, recovering its shape. The mean square 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 with a coefficient (1:0.5) to form the diffusion branch loss.

[0153] Joint optimization and parameter update:

[0154] The supervised comparison loss and the diffusion branch loss are directly added together to obtain the total loss for the current batch. Backpropagation is performed only on the network parameters of the projection head and the diffusion branch, and optimization is performed using Adam (learning rate approximately 0.0001).

[0155] Latent characterization extraction:

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

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

[0158] The enhanced feature representations are dimensionally concatenated to generate merged features of a preset size;

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

[0160] Max pooling layers are used to flatten all channels and spatial dimensions into a long vector, generating a one-dimensional feature vector of a preset size.

[0161] The original classification score is obtained by mapping the one-dimensional feature vector through two fully connected layers.

[0162] The original classification scores are then subjected to a log-softmax operation to obtain the log probability distribution of each sample for the two classes; where the two classes refer to anti-stress peptides and non-stress peptides.

[0163] Based on the logarithmic probability distribution, peptides 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 a preset number of categories and outputs the original classification scores.

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

[0168] (1) Loading and preparing data:

[0169] Mini-batch data is read from the data loader. Each tensor has a shape of (B, 1024), where B is the current batch size; each row corresponds to a 1024-dimensional latent feature of a peptide. The corresponding labels tensor has a shape of (B,), where each element is an integer 0 or 1, representing the category of the peptide. 64 samples are randomly selected each time to form a batch.

[0170] (2) Model structure:

[0171] Network structure:

[0172] First, the two tensors are concatenated along the final "width" dimension to generate a merged feature of size (B, 1, 1, 2048). Five parallel 3×3 adaptive convolutional 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 then performed on each of the five channels, followed by the introduction of non-linearity using the ReLU function while maintaining the size (B, 5, 1, 1024). A 1×2 max-pooling layer with a stride of (2, 2) is then used. All channels are flattened along the spatial dimension into a single long vector, generating a one-dimensional feature vector of size (B, 2560).

[0173] Two fully connected layers: Layer 1: Compresses the 2560-dimensional input to 256 dimensions and applies ReLU to obtain (B, 256). Layer 2: Maps the 256-dimensional hidden representation to 2 classes, outputting the original classification score of size (B, 2). Finally, a log-softmax operation is applied to these scores to obtain the log probability distribution of each sample across the two classes, which can be used for subsequent loss calculations based on cross-entropy or KL divergence.

[0174] Loss function and label smoothing:

[0175] Hard-label cross-entropy: For each sample, take the output probability of the corresponding class, logarithmize and negative it, then average it over the entire batch to obtain the traditional cross-entropy loss L. CE The expression is:

[0176]

[0177] KL divergence after label smoothing: The one-hot hard labels are slightly smoothed by replacing "1" with 0.9 and "0" with 0.1 to obtain the smoothed target distribution. The model output is then subjected to softmax and logarithm calculation. The product of the element-wise differences with the smoothed distribution is summed, and the average is calculated 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 gradients accumulated in the previous round are reset to zero to avoid gradient accumulation. Then, gradients are automatically calculated for all trainable parameters of the convolutional and fully connected layers based on the total loss. Using the Adam algorithm, the network's learning rate is 0.0001, and the parameter values ​​are updated based on the learning rate and first / second-order momentum information.

[0182] (4) Backpropagation and parameter update:

[0183] After each training round, the model is switched to evaluation mode, and the classification accuracy is calculated using the retained validation set to monitor performance changes. Finally, after training is complete, the optimal model parameters are persisted to disk for subsequent prediction or deployment.

[0184] The model of this invention was compared with antihypertensive peptide prediction models and achieved significant performance improvements on multiple evaluation metrics. Evaluation metrics included specificity (SEP), sensitivity (SEN), accuracy (ACC), Matthews correlation coefficient (MCC), F1-score (F1), area under the curve (AUC), and average precise (AP), 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 uses amino acid composition, atomic composition, chemical descriptors, and G-Scales features as input and employs support vector machines for classification. AHPred-DE ​​integrates UniRep pre-trained embeddings, One-Hot encoding, and K-mer frequencies, achieving prediction through a deep ensemble model. LSTM-AHP is a multi-layer LSTM sequence classifier that extracts temporal features from the input sequence and maps the hidden state of the last time step to a dimension through a fully connected layer to complete the classification. PAAP utilizes three types of features—amino acid composition, dipeptide composition, and pseudo-amino acid composition—and performs classification based on random forest. pLM4ACE relies on the sequence embeddings output by the ESM-2 large-scale protein language model, combined with SVM for discrimination. Ensemble-AHTPpred integrates amino acid composition, sequence order descriptors, and secondary structure conformation features, fusing the advantages of each model through weighted voting of random forest, support vector machine, and XGBoost algorithms. AHPNG constructs a deep learning framework combining convolutional neural networks and gated recurrent units based on dipeptide frequencies and amino acid physicochemical properties. All models were validated on a unified, independent test set to ensure the objectivity and fairness of the evaluation results.

[0187] A comprehensive analysis of the performance metrics of DeepATHPFuse compared to seven other mainstream models reveals that DeepATHPFuse exhibits a significant advantage across all metrics. Specifically, the accuracy (ACC) of each model ranges from mAHTPred (0.7901) to DeepATHPFuse (0.8651), with DeepATHPFuse leading significantly. The Matthews correlation coefficient (MCC) of each model falls between mAHTPred (0.4831) and DeepATHPFuse (0.6945), with DeepATHPFuse's value far exceeding the other models, demonstrating a significant advantage in positive and negative sample classification. The F1 score ranges from mAHTPred (0.5952) to DeepATHPFuse (0.7934), indicating that DeepATHPFuse achieves the best balance between precision and recall. Sensitivity (SE) ranges from mAHTPred (0.5051) to AHTpin (0.8452), with DeepATHPFuse showing the highest sensitivity. The e-value reached (0.8231), which is high; the specificity (SP) ranged from AHTpin (0.7252) to mAHTPred (0.9211), with DeepATHPFuse ranking first with (0.8844), achieving effective identification of negative samples; in terms of the area under the ROC curve (AUC) metric, which comprehensively evaluates the model's discriminative ability, the range for each model was from LSTM-AHP (0.8370) to DeepATHPFuse (0.9262), with DeepATHPFuse performing the best and demonstrating excellent robustness to the classification threshold; finally, in terms of the average precision (AP) metric, which measures the model's accuracy under high recall conditions, the range for each model was from mAHTPred (0.5346) to DeepATHPFuse (0.8648), with DeepATHPFuse once again showing the most outstanding performance. Overall, DeepATHPFuse ranked first in all seven evaluation metrics, significantly outperforming all other methods and demonstrating comprehensive performance advantages and robustness in the field of antihypertensive peptide prediction. Specific data are shown in Table 1.

[0188] Table 1 compares this embodiment with other methods on the dataset.

[0189]

[0190] Figure 2 This is a comparison chart of AUROC curves and a schematic diagram of PRROC curves between embodiments of the present invention and other methods; Figure 3 This is a schematic diagram comparing the AUC value distribution of this invention with other comparison methods; Figure 4 This is a schematic diagram comparing the MCC value distribution of an embodiment of the present invention with other comparative methods.

[0191] In summary, this 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 converter) architecture developed by Google, specifically optimized for bioinformatics applications. The model is trained on the UniRef50 dataset, which contains deduplicated protein sequences representing millions of different proteins, providing a comprehensive background for protein sequence analysis. The main advantage of this model lies in its Transformer-based architecture, which captures complex contextual relationships in the input sequence and effectively models long-range dependencies.

[0193] (2) Structural Feature Representation. For the structural information of the antihypertensive peptide, this embodiment fully utilizes the coordinate correlation features calculated by RDKit: First, a hydrogen-containing three-dimensional small molecule model is generated in RDKit based on the peptide's amino acid sequence. Then, the coordinates of all atoms in three-dimensional space are resolved. Next, the Pearson correlation coefficient matrix of these coordinate points in three dimensions is calculated to obtain an original correlation matrix. This matrix is ​​then replaced with NaN and infinity values, scaled proportionally, and bilinearly interpolated to a fixed resolution to generate a grayscale heatmap. Finally, this image is expanded in row-major order into a one-dimensional vector of length 1024, which serves as the structural characterization input for the downstream antihypertensive peptide activity prediction model.

[0194] (3) This embodiment proposes a dual-path multimodal feature enhancement method based on contrastive-generative joint optimization: First, the sequence representation of the peptide is obtained through the large-scale protein language model ProtT5; at the same time, the structural feature vector is directly extracted from the three-dimensional conformation of the peptide using only RDKit; then, these two sets of vectors are fed into their respective projection heads and latent diffusion branches, and supervised contrastive learning and generative noise reconstruction are performed independently, and the contrastive loss and diffusion reconstruction loss on the sequence view and the structural view are calculated respectively; finally, the two types of losses are weighted and fused with adjustable weights, and the parameters of the ProtT5 encoder, each projection head and the diffusion model are updated synchronously through backpropagation, so as to output the enhanced sequence features and structural features respectively.

[0195] (4) This embodiment proposes a classification method for antihypertensive peptides based on joint optimization of single-path KanConv structure and label smoothing: the sequence features and structural features obtained through contrastive-generative joint optimization are concatenated in the width dimension and then fed into a single KanConv convolutional module. This module consists of a KanConv convolutional layer (followed by BatchNorm and ReLU) and a max pooling layer in sequence. The pooling output is then flattened and the final binary classification logits are generated through two fully connected layers. During the training phase, the network uses cross-entropy loss to supervise hard labels and constructs soft labels through smoothing coefficients and uses KL divergence to measure the consistency between the predicted distribution and the soft label distribution. The two are fused into a joint loss in proportion, and the parameters of the convolutional module and fully connected layers are optimized synchronously through backpropagation, thereby significantly improving the discrimination ability and generalization performance between antihypertensive peptides and non-antihypertensive peptides.

[0196] This embodiment leverages Prot-T5's advantages in efficiently capturing global contextual information and accurately representing spatial structural features by fusing Prot-T5 sequence embeddings and RDKit structural embeddings. It employs a combined optimization approach of post-projection head supervised contrast and latent diffusion generative methods, significantly enhancing the discriminative power and information integrity of the composite features. Furthermore, it utilizes the strengths of KanConv convolutional networks in multi-scale neighborhood feature processing and combines a 0.1 label smoothing classification strategy to effectively suppress model overconfidence and improve generalization ability. Its beneficial effects are as follows:

[0197] (1) In this embodiment, the global context features of the sequence obtained by Prot-T5 and the three-dimensional structural spatial adjacency information extracted by RDKit are combined to achieve deep complementarity of multimodal information. Sequence features are good at capturing local functional fragments and long-range dependencies of peptide chains, 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 variations or conformational noise.

[0198] (2) In this embodiment, supervised contrastive learning (SupCon) is used in the contrastive learning stage: random masking or amino acid substitution perturbation is applied to the sequence embedding, and node dropping or edge weight noise perturbation is applied to the structural embedding to generate multi-view positive samples. These samples are then mapped to the contrast space through two layers of MLP projection heads to narrow the distance between similar samples and widen the boundaries between dissimilar samples. In the generative learning stage, a conditional latent diffusion model (LDM) is introduced: noise is added to the features according to linear noise scheduling, and a U-Net noise prediction network combining temporal coding and conditional injection is used for denoising and reconstruction to fill in the high-dimensional low-density distribution regions and enrich the representation of edges and rare samples. The losses of the two methods work synergistically on the same encoder during backpropagation, so that the final output composite features have both the class discrimination enhanced by SupCon and the distribution continuity and information richness brought by LDM, thereby significantly improving the robustness and discrimination accuracy of the model to noise interference, sequence variation and small sample scenarios.

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

[0200] (4) In this embodiment, the hard label cross-entropy loss is used to ensure that the model has a strong ability to distinguish the real category. The soft label distribution constructed based on the smoothing coefficient of 0.1 and constrained by KL divergence loss can effectively alleviate the model's overconfidence and improve the calibration of the output probability. The two are jointly optimized, while maintaining the classification accuracy, introducing smoothing regularization, reducing the risk of overfitting in noisy and low-sample scenarios, accelerating training convergence and enhancing the model's generalization ability and robustness.

[0201] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope 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 contrastive-generative joint optimization, characterized in that, include: Sequence and structural feature representations of extracted peptides; Multimodal feature enhancement is performed on the sequence feature representation and structural feature representation through contrast-generative joint optimization. Multimodal feature enhancements that involve comparative-generative joint optimization include: Supervised contrastive learning aligns sequence feature representations and structural feature representations in a unified embedding space; A latent diffusion branch is introduced, and a generative smooth mapping is learned on the sequence feature representation and the structural feature representation in the latent space; A pre-defined classification model based on Kan-Conv structure and label smoothing is used to classify peptides based on the enhanced feature representation. Using a pre-defined classification model based on Kan-Conv structure and label smoothing for joint optimization, peptide classification is performed on the enhanced feature representation, including: The enhanced feature representations are dimensionally concatenated to generate merged features of a preset size; For the merged features, several parallel adaptive convolutional kernel channels are applied for batch normalization, and then nonlinearity is introduced through the ReLU function. Max pooling layers are used to flatten all channels and spatial dimensions into a long vector, generating a one-dimensional feature vector of a preset size. The original classification score is obtained by mapping the one-dimensional feature vector through two fully connected layers. The original classification scores are then subjected to a log-softmax operation to obtain the log probability distribution of each sample for the two classes; where the two classes refer to anti-stress peptides and non-stress peptides. Based on the logarithmic probability distribution, peptides are classified.

2. The antihypertensive peptide classification method based on the fusion of sequence and structural multimodal features combined with contrast-generative joint optimization as described in claim 1, is characterized in that... Sequence and structural feature representations of extracted peptides include: The feature representation of peptides is extracted based on a pre-defined large-scale protein language model, while the structural feature representation generated by RDKit is utilized.

3. The antihypertensive peptide classification method based on the fusion of sequence and structural multimodal features combined with contrast-generative joint optimization as described in claim 2, is characterized in that... The structural feature representations generated using RDKit include: The peptide is mapped to a three-dimensional structural model, all implicit hydrogen atoms are added to the three-dimensional structural model, and structural data including the three-dimensional coordinate information of all atoms is obtained through a molecular conformation embedding algorithm. Atomic coordinates are extracted from the structural data to obtain an atomic coordinate matrix; Calculate the Pearson correlation coefficients between the three-dimensional coordinate dimensions based on the atomic coordinate matrix, and obtain the correlation coefficient matrix; The correlation coefficient matrix is ​​then subjected to numerical normalization, image transformation, and feature vector generation to obtain a structural feature representation.

4. The antihypertensive peptide classification method based on the fusion of sequence and structural multimodal features combined with contrast-generative joint optimization as described in claim 2, is characterized in that... Multimodal feature enhancement, which involves contrastive-generative joint optimization of the sequence feature representation and structural feature representation, includes: Supervised contrastive learning is performed on the sequence feature representation and structural feature representation to obtain contrastive learning-enhanced sequence feature representation and structural feature representation; Generative learning is performed on the sequence feature representation and structural feature representation to obtain generatively enhanced sequence feature representation and structural feature representation.

5. The antihypertensive peptide classification method based on the fusion of sequence and structural multimodal features combined with contrast-generative joint optimization as described in claim 4, is characterized in that... Supervised contrastive learning of the sequence feature representation and structural feature representation includes: The sequence feature representation and structural feature representation are randomly perturbed, and the perturbed feature representations are then concatenated sequentially to form a feature matrix; Calculate the pairwise dot product similarity matrix between the features; Construct a mask matrix with the same scale as the similarity matrix, and multiply the dot product similarity matrix by the element-wise exponentialized mask matrix. Summing all non-zero elements in each row of the multiplied matrix yields the denominator of the supervised contrastive loss; where each row of the multiplied matrix represents each augmented sample. Based on the definition of positive sample pairs, 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. For each augmented sample in the multiplied matrix, the contrastive loss is calculated sample by sample based on the denominator and numerator of the corresponding supervised contrastive loss and averaged to obtain the final supervised contrastive loss. Using the final supervised contrastive loss, we obtain the sequence feature representation and structural feature representation enhanced by contrastive learning.

6. The antihypertensive peptide classification method based on the fusion of sequence and structural multimodal features combined with contrast-generative joint optimization as described in claim 4, is characterized in that... Generative learning of the sequence feature representation and structural feature representation includes: The sequence feature representation and structural feature representation are mapped to the latent space to obtain initial latent variables; Noise is injected into the initial latent variables to obtain noisy latent variables; The noisy latent variables are concatenated with the normalized time steps and then input into a preset fully connected network to output a prediction of the noise. The initial latent variables are inferred from the predicted noise, and the noise component is removed. The recovered latent variables are mapped back to the original dimensions through two fully connected layers to obtain the reconstructed feature vectors.

7. The antihypertensive peptide classification method based on the fusion of sequence and structural multimodal features combined with contrast-generative joint optimization as described in claim 1, characterized in that, The two-layer fully connected mapping includes: 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 a preset number of categories and outputs the original classification scores.

Images