Anticancer peptide sequence classification method based on multi-scale enhancement conditional diffusion model

Abstract

The invention provides an anti-cancer peptide sequence classification method based on a multi-scale enhancement conditional diffusion model, which solves the problems of anti-cancer peptide sequence classification and the like, and comprises the following steps: S1, data collection and preprocessing; s2, performing high-level multi-scale enhanced feature extraction and dimension reduction on the sequence; s3, constructing a conditional diffusion model; s4, model training; s5, generating a classification result of the anticancer peptide sequence; and S6, performing model evaluation. The method has the advantages of accurate classification, high calculation efficiency and the like.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of bioinformatics and computer science, specifically relating to a method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model. Background Technology

[0002] In existing technologies, sequence analysis of anticancer peptides mainly relies on two types of methods: one is the traditional method based on statistical features, which often uses basic features such as amino acid composition and physicochemical properties as input; the other is the representation learning method that relies on deep learning models, such as feature modeling techniques based on pre-trained protein language models (e.g., ProtBERT). While the former has some interpretability, it struggles to characterize the complex context and high-order structural information in the sequence, resulting in limited classification accuracy. The latter, while capable of capturing deeper semantic representations, often has a high dependence on the number of labeled samples and computational resources, especially when dealing with high-dimensional embeddings or large-scale datasets, where training costs and computational overhead increase significantly. Therefore, existing technologies still suffer from problems such as limited classification accuracy, high model training costs, and insufficient generalization ability, making them difficult to efficiently promote in practical biomedical analysis scenarios.

[0003] To address the shortcomings of existing technologies, researchers have conducted long-term explorations and proposed various solutions. For example, Chinese patent literature discloses a method and system for identifying anticancer peptides based on attention mechanisms and multi-granularity hierarchical features [202310677933.3]. First, it learns atomic-level features through transfer learning, revealing potential features that were difficult to discover in previous work. Second, it uses the ChouFasman algorithm to represent secondary structures during amino acid sequence layer extraction, increasing the richness of information. Next, it constructs a hypergraph to solve the problem of existing technologies failing to capture the high-order structural similarity of anticancer peptide sequences, and learns the importance of subsequences to the overall sequence through an attention mechanism. Finally, it fuses multi-granularity hierarchical features, enabling the model to comprehensively understand and describe the characteristics of anticancer peptide sequences.

[0004] The above-mentioned scheme has solved the problem of sequence analysis of anticancer peptides to a certain extent, but it still has many shortcomings, such as insufficient classification accuracy, low computational efficiency and limited feature expression ability. Summary of the Invention

[0005] The purpose of this invention is to address the above-mentioned problems by providing a method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, which offers accurate classification, high computational efficiency, and good feature expression capabilities.

[0006] To achieve the above objectives, the present invention employs the following technical solution: a method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, comprising the following steps: S1: Data collection and preprocessing; S2: High-level multi-scale enhanced feature extraction and dimensionality reduction of sequences; S3: Construct a conditional diffusion model; S4: Model training; S5: Classification results of generated anticancer peptide sequences; S6: Model Evaluation.

[0007] In the above-mentioned method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, step S1 includes the following steps: S11: Data collection, collecting raw data from publicly available anticancer databases and literature, including unique identifiers of peptides, amino acid sequences, functional annotations, and cancer type labels; S12: Data cleaning, removing duplicate records and entries containing missing values, and filtering out illegal amino acid characters; S13: Sequence regularization, adjusting all sequences to a uniform length to form a matrix of equal-length sequences.

[0008] In the above-mentioned method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, step S2 includes the following steps: S21: Extract features using a pre-trained protein language model; S22: Enhance features using MECS modules; S23: Use One-hot encoding to process tags; S24: Principal component analysis (PCA) is used to reduce the dimensionality of high-dimensional features while retaining the main information components.

[0009] In the above-mentioned method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, step S22 includes: Channel attention mechanism: Calculates global information of features, generates channel weight coefficients, and highlights important channel features. The formula is as follows: ; Where pool represents the features obtained through average pooling, max pooling, and median pooling, Conv2d is the convolution operation, and σ is the Sigmoid activation function; Multi-scale convolution: Extracting multi-scale spatial features using different convolution kernels, the formula is as follows: ; Spatial attention mechanism: A spatial attention map is generated based on the results of multi-scale convolution, and the formula is as follows: ; Feature fusion: Applying channel and spatial attention to the input features yields enhanced features, as shown in the formula: .

[0010] In the above-mentioned method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, step S3 includes the following steps: S31: Construct an encoder to map features to the latent space; S32: Conditional input, which converts label information into a conditional vector; S33: Construct the diffusion process, including forward diffusion and backward generation, using the MultiScaleResidualBlock module; S34: Construct a decoder that maps latent features back to the output space.

[0011] In the aforementioned method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, the MultiScaleResidualBlock module in step S33 performs convolution operations on the input data using convolution kernels of different scales. The feature outputs at different scales can be concatenated or fused, and unified through one or more nonlinear mapping structures. A residual connection mechanism is introduced, and the module input and output are weighted or directly superimposed. The calculation formula is as follows: ; Where Gi(x) is the transformation at different scales, and wi is the weight.

[0012] In the above-mentioned method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, step S4 includes the following steps: S41: Dataset preparation, including standardization, SMOTE oversampling, and noise perturbation; S42: Select the optimization algorithm, using stochastic gradient descent (SGD) and momentum; S43: Set the learning rate and batch size; S44: Model tuning and hyperparameter optimization.

[0013] In the above-mentioned method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, the SGD optimization algorithm calculation formula in step S42 is as follows: ; Where 𝜃𝑡 are the model parameters at the current time step 𝑡, 𝜂 is the learning rate, and ∇𝜃𝐿(𝜃𝑡) is the gradient of the loss function 𝐿 with respect to the model parameters 𝜃; The formula for calculating the momentum in stochastic gradient descent is: ; Where 𝑣𝑡 is momentum, 𝛽 is momentum coefficient, ∇𝜃𝐿(𝜃𝑡) is the current gradient, and 𝜂 is the learning rate.

[0014] In the above-mentioned method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, step S5 includes the following steps: S51: The model receives the input protein sequence features and extracts their high-order semantic representation in the encoding module; S52: Label information is incorporated into the feature modeling process through conditional embedding to form a joint representation with category awareness; S53: Indicates that it is mapped into the latent space and enters the diffusion process; S54: In the diffusion modeling stage, the perturbation recovery and structural reconstruction of latent features are achieved through a mechanism that combines forward noise addition with backward stepwise denoising, and discriminative latent representations are generated under conditional constraints. S55: The decoding module maps the expression to the output label space to classify the anticancer peptide sequence.

[0015] In the above-mentioned method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, step S6 includes the following steps: S61: The model is configured for inference or evaluation mode; S62: Obtain the model output results by performing batch predictions on the test set or validation set, and compare and analyze them with the corresponding real labels; S63: Generate classification reports and confusion matrices to visualize the classification results for each category; S64: Save the optimal model parameters based on the evaluation results for subsequent deployment or further application.

[0016] Compared with existing technologies, the advantages of this invention are as follows: By introducing a feature extraction module with multi-scale perception capabilities and combining it with a residual enhancement mechanism, the model can extract local and global information of sequences from different receptive fields, effectively improving the modeling ability of structural hierarchical features and thus enhancing pattern recognition accuracy; by adopting a conditional diffusion modeling process, controlled noise perturbation and stepwise denoising operations are performed in the latent space, enabling the generation process to dynamically respond to conditional information such as category labels, achieving fine control of target classification output and maintaining semantic consistency; the model exhibits good predictive performance under multiple data distribution conditions, not only learning efficiently on training data but also possessing strong generalization ability, making it suitable for accurate discrimination tasks with unseen samples. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the MECS model of the present invention; Figure 3This is a flowchart of the conditional diffusion prediction classification model of the present invention; Figure 4 This is a comparison of four key performance indicators between the present invention and existing methods in the anticancer peptide classification task. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1-4 As shown, a method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model (MECS) is presented. Unlike existing methods that rely on traditional discriminative models or single semantic modeling approaches, this method integrates a conditional diffusion mechanism, a protein language modeling strategy, and a MECS-based multi-scale feature enhancement module. It introduces a label-conditionally controlled diffusion and denoising process into the latent space, enabling the model to extract high-order semantic features from the original protein sequence while performing condition-driven specific classification predictions. Furthermore, the integrated MECS module possesses channel attention and multi-scale convolution mechanisms, effectively capturing local and global structural information in the sequence, further enhancing the discriminative power and stability of feature modeling. At the overall system implementation level, this method not only improves expressive power and generalization performance but also significantly reduces dependence on labeled data and high-computing resources, exhibiting good scalability and engineering deployment potential. The specific steps include: S1: Data collection and preprocessing; S2: High-level multi-scale enhanced feature extraction and dimensionality reduction of sequences; S3: Construct a conditional diffusion model; S4: Model training; S5: Classification results of generated anticancer peptide sequences; S6: Model Evaluation.

[0020] Specifically, step S1 includes the following steps: S11: Data collection. Obtain raw data containing protein sequences and their functional labels from one or more public databases or experimental sources. In this embodiment, the entire dataset contains nine labels and 1680 data entries to form a representative and biologically significant training sample set. Each set of data contains core information such as a unique identifier for the peptide, amino acid sequence, functional annotation, and cancer type label, thereby constructing an anticancer peptide sequence dataset. S12: Data cleaning, removing redundant records and illegal characters, and filtering residues that do not conform to standard amino acid codes (such as B, J, O, U, X, Z) to retain biologically valid sequences; S13: Sequence regularization. During the regularization process, sequences exceeding a set length can be truncated, and insufficient parts can be filled with specific placeholders to construct a sequence matrix of equal length to meet the uniformity requirements of the model input.

[0021] The processed data can be organized into a well-structured and formatted storage format, such as CSV files or other suitable structured file formats, for subsequent tasks such as feature encoding, model training, and result visualization.

[0022] In detail, step S2 includes the following steps: S21: Feature extraction using a pre-trained protein language model. To extract contextual semantic information of amino acid residues in a protein sequence, a pre-trained protein language model, preferably the ProtBERT model, can be introduced. This model can include multiple Transformer network layers to capture contextual dependencies and long-distance interactions between amino acid residues in the sequence. During embedding, the model can format and tokenize the input sequence and generate corresponding vector representations. These vectors can be used to express the feature state of each amino acid and its context. The embedding vectors at the residue level can be aggregated, pooled, or otherwise dimensionality-reduced to construct fixed-length sequence-level feature vectors for downstream tasks, thereby improving the performance of subsequent classification models in terms of prediction accuracy and structural discrimination ability. S22: To further enhance the model's ability to model local structures and global sequence patterns, a multi-scale feature enhancement module is introduced on the basis of embedded features. This module may include multi-scale convolutional network structures and channel attention mechanisms (such as the SE module Squeeze-and-Excitation or the CBAM module ConvolutionalBlockAttentionModule) to strengthen the perception and expression of potential structure-function associations between amino acid residues, thereby improving the model's discrimination accuracy and generalization ability. S23: Use one-hot encoding to process the labels and embed the conditional signal into the diffusion modeling process to achieve label-driven conditional control, thereby improving the model's adaptability and guidance to different categories during the training and inference stages; S24: Given that the output features of the above protein language model have high dimensionality, in order to reduce redundancy, improve training efficiency and reduce computational resource consumption, dimensionality reduction processing can be performed on the embedded features. Principal Component Analysis (PCA) or other dimensionality reduction algorithms can be used to reduce the dimensionality of high-dimensional features, retain the main information components, and achieve effective information compression and feature selection.

[0023] Further, step S22 includes: Channel attention mechanism: Used to weight the features of each channel. By calculating the global information of the input features (average pooling, max pooling, and median pooling), the weighting coefficient of each channel is obtained, thereby adjusting the importance of each channel. The formula is as follows: ; Where pool represents the features obtained through average pooling, max pooling, and median pooling, Conv2d is the convolution operation, and σ is the Sigmoid activation function; Multi-scale convolution: Using different convolution kernels to extract multi-scale spatial features enables the model to capture spatial structure at multiple scales, improving the expressive power of spatial features. The formula is: ; Spatial attention mechanism: A spatial attention map is generated based on the results of multi-scale convolution, focusing on important spatial regions to further enhance the spatial representation of features. The formula is as follows: ; The results of multi-scale convolution operations are summed to generate the final spatial attention.

[0024] Feature fusion: Applying channel and spatial attention to input features yields enhanced features, improving the model's discriminative ability in complex data. The formula is: .

[0025] Furthermore, step S3 includes the following steps: S31: Construct an encoder to perform structural compression and representation embedding processing on the input anticancer peptide sequence features. This module can perform feature transformation operations through one or more neural network sublayers, including fully connected layers or other forms of nonlinear mapping structures, to map the original sequence features to the latent feature space. The mapping process can extract discriminative structural information from the anticancer peptide sequence and transform it into a low-dimensional latent representation vector, thereby improving the feature utilization efficiency of subsequent modules in the generation modeling and classification process. The latent vector output by the encoder (also known as the embedding representation Xembed) can be used as input to the diffusion model for further processing in the diffusion and denoising process, realizing conditional generation and discriminative modeling of the anticancer peptide sequence in the latent space. S32: Conditional input, used to transform label information into a processable conditional vector representation in the latent space, to achieve label-based feature generation and classification guidance; preferably, the label information may include attributes such as category identifiers or functional annotations related to anticancer peptide sequences, and the information can be encoded through a conditional embedding layer; this embedding layer may be one or more neural network sub-modules, capable of mapping labels to a representation compatible with the latent feature space; the conditional vector participates in the diffusion modeling and generation process as an auxiliary input, enabling the model to dynamically adjust the generation path according to different label conditions when performing feature denoising and generation, thereby enhancing the class consistency and semantic relevance of the generation results; by introducing this conditional input mechanism, the diffusion model has condition-aware capabilities, which helps to improve the controllability and output accuracy of the model in multi-class classification or conditional generation tasks; S33: Constructing the diffusion process can include a forward diffusion stage and a backward generation stage. The forward diffusion stage is used to gradually introduce noise components into the original feature representation to break the structural information in the data, thereby promoting the model's learning of the data distribution and noise characteristics. The backward generation stage, guided by a given conditional vector, reconstructs the structural features in the latent space through a stepwise denoising process and generates a discrimination result or target output consistent with the conditional information. The conditional vector is preferably obtained by converting the label information through the embedding module, and can be fused with the latent representation to participate in the diffusion modeling process to achieve condition-guided generation based on category or other attributes, using the MultiScaleResidualBlock module. S34: Construct a decoder to map the latent representations generated by backdiffusion back to the output space to obtain the final classification result; the decoder can consist of one or more fully connected layers to ensure that the latent vectors can be effectively restored to the classification label or other discriminant output.

[0026] Through the synergistic effect of the above modules, the constructed model framework can achieve flexible modeling and discrimination of different categories of conditions while possessing high expressive power, significantly improving the accuracy of anticancer peptide sequence classification, and effectively reducing the computational overhead in the model training and inference process. It is suitable for intelligent analysis tasks of large-scale complex biological sequence data.

[0027] In addition, in step S33, the MultiScaleResidualBlock module uses convolutional kernels of different scales (such as 1×1, 3×3, and 5×5) to perform convolution operations on the input data. Feature outputs at different scales can be concatenated or fused and uniformly processed through one or more nonlinear mapping structures (such as fully connected layers). To improve the deep expressive power of the model and alleviate gradient degradation during training, a residual connection mechanism is introduced. By weighting or directly superimposing the module inputs and outputs, information preservation of the deep structure and stable training are achieved. The calculation formula is as follows: ; Where Gi(x) is the transformation at different scales, and wi is the weight.

[0028] Meanwhile, the training process of the conditional diffusion model in step S4 can be carried out using a supervised learning framework. This training aims to jointly optimize the reconstruction ability of features in the latent space and the classification performance in the output space, thereby improving the overall model's discrimination accuracy and generalization ability. The loss function in the training process can include at least a reconstruction loss for continuous variable regression and a classification loss for multi-class discrimination. The former is the mean squared error loss (MSE), which is used to constrain the denoising and reconstruction effect in the diffusion process; the latter is the cross-entropy loss, which is used to supervise the consistency between the model output and the actual label. Alternatively, other alternative loss functions with discriminative capabilities can be used for combined optimization.

[0029] Regarding optimizer selection, gradient-based optimization algorithms, including but not limited to Adam and stochastic gradient descent (SGD), can be chosen based on the model structure and data scale. To further improve the stability and efficiency of model training, dynamic learning rate scheduling mechanisms or adaptive optimization strategies can be introduced to accelerate convergence and prevent getting trapped in local optima. To alleviate model overfitting and enhance its generalization ability on unknown data, cross-validation techniques and early stopping mechanisms can be combined during training, dynamically adjusting the number of training epochs by monitoring the validation set performance in real time. The training process can be deployed on hardware environments that support graphics acceleration (such as GPUs) and is compatible with distributed training architectures to support efficient processing and modeling tasks on large-scale anticancer peptide sequence datasets, meeting the actual needs for computational efficiency and processing power in industrial bioinformatics applications.

[0030] Step S4 specifically includes the following steps: S41: Dataset Preparation. To improve model training efficiency and classification performance, the extracted feature data is first standardized using StandardScaler to adjust each feature to zero mean and unit variance, ensuring that the data is effectively learned by the model at a uniform scale. Simultaneously, to address the class imbalance problem in the training set, SMOTE oversampling technology is introduced to generate minority class samples to balance the class distribution and improve the model's ability to recognize low-frequency classes. Furthermore, to enhance the model's robustness in practical applications, controlled noise perturbation is artificially introduced during the training phase, enabling the model to maintain stable classification performance even when facing data variations. These multi-dimensional preprocessing strategies collectively construct a more balanced, stable, and generalizable training data foundation. S42: Selecting an optimization algorithm. To effectively train the metric learning model, an appropriate optimization algorithm is used to adjust the model's parameters. Commonly used optimization algorithms include the Adam optimizer and stochastic gradient descent (SGD). In this embodiment, the optimization algorithm is based on stochastic gradient descent (SGD), which is suitable for most machine learning tasks. During model training, the SGD optimizer is selected to update the model's parameters. The SGD optimizer adjusts the model's parameters by calculating the gradient of each batch and adjusting the parameters according to the learning rate. To accelerate convergence and improve the model's training performance, momentum is also used, which helps overcome the problem of local minima and speeds up convergence. S43: Setting the learning rate and batch size. Setting the learning rate and batch size are important hyperparameters when training deep learning models, as they directly affect the efficiency and performance of model training. In this embodiment, the learning rate is set to 0.01, which is a commonly used initial learning rate suitable for many tasks. To avoid instability caused by an excessively large learning rate, a learning rate scheduler is added during training. When the validation loss no longer improves, the learning rate is reduced, thereby helping the model to be fine-tuned in the later stages of training. The batch size is set to 128. This size usually provides good computational efficiency and training stability in many tasks, making full use of parallel computing resources while avoiding the problem of low computational efficiency caused by excessively small batches. S44: Model tuning and hyperparameter optimization. After the initial training of the model, further model tuning and hyperparameter optimization can be carried out to improve its overall performance on the training and validation sets. This process may include, but is not limited to, targeted adjustments and combinations of key elements such as learning rate, batch size, network structure parameters, and regularization strategies. Preferably, a series of regularization techniques (such as random deactivation mechanisms and weight penalty strategies) can be introduced to effectively suppress the risk of overfitting and enhance the model's generalization ability to unseen data. The above optimization strategies can be dynamically executed based on performance feedback, enabling the model to adaptively adjust parameter configurations at different training stages, ultimately achieving the optimal model structure and parameter combination for the current data distribution, thereby improving classification accuracy and robustness.

[0031] As can be seen, in step S42, the calculation formula for the SGD optimization algorithm is: ; Where 𝜃𝑡 are the model parameters at the current time step 𝑡, 𝜂 is the learning rate, and ∇𝜃𝐿(𝜃𝑡) is the gradient of the loss function 𝐿 with respect to the model parameters 𝜃; The formula for calculating the momentum in stochastic gradient descent is: ; Where 𝑣𝑡 is momentum, 𝛽 is momentum coefficient, ∇𝜃𝐿(𝜃𝑡) is the current gradient, and 𝜂 is the learning rate.

[0032] Clearly, step S5 includes the following steps: S51: The model receives the input protein sequence features and extracts their high-order semantic representation in the encoding module; S52: Label information is incorporated into the feature modeling process through conditional embedding to form a joint representation with category awareness; S53: Indicates that it is mapped into the latent space and enters the diffusion process; S54: In the diffusion modeling stage, the perturbation recovery and structural reconstruction of latent features are achieved through a mechanism that combines forward noise addition with backward stepwise denoising, and discriminative latent representations are generated under conditional constraints. S55: The decoding module maps the expression to the output label space to classify the anticancer peptide sequence.

[0033] Preferably, step S6 evaluates the performance of the trained conditional diffusion model on the test dataset to comprehensively measure its effectiveness and generalization ability in the classification task. During the evaluation, the model can be set to inference mode and prediction operations can be performed by batch inputting test samples. To ensure the computational efficiency of the evaluation process, the gradient update mechanism is preferably disabled, and only the forward propagation path is retained for outputting prediction results. Regarding the selection of evaluation metrics, statistical indicators may include, but are not limited to: classification accuracy, recall, F1 score, area under the curve (AUC), and confusion matrix, which reflect the model's recognition ability, discrimination performance, and the degree of distinction between categories under different categories. Specifically, it includes the following steps: S61: The model is configured for inference or evaluation mode; S62: Obtain the model output results by performing batch predictions on the test set or validation set, and compare and analyze them with the corresponding real labels; S63: Generate classification reports and confusion matrices to visualize the classification results for each category; S64: Save the optimal model parameters based on the evaluation results for subsequent deployment or further application.

[0034] In summary, the principle of this embodiment lies in integrating conditional diffusion mechanisms, protein language modeling strategies, and a multi-scale feature enhancement module (MECS) to construct an intelligent anticancer peptide recognition architecture with conditional guidance and multi-scale semantic awareness capabilities. By introducing a label-conditionally controlled diffusion and denoising process into the latent space, the model can extract high-order semantic features from the original protein sequence while performing condition-driven specific classification predictions. Furthermore, the integrated MECS module possesses channel attention and multi-scale convolution mechanisms, effectively capturing local and global structural information in the sequence, further enhancing the discriminative power and stability of feature modeling. At the overall system implementation level, this improves expressive power and generalization performance while significantly reducing dependence on labeled data and high-computing resources, exhibiting good scalability and engineering deployment potential.

[0035] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0036] Although this document makes frequent use of terms such as ProtBERT and MECS, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. A method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model, characterized in that, Includes the following steps: S1: Data collection and preprocessing; S2: High-level multi-scale enhanced feature extraction and dimensionality reduction of sequences; S3: Construct a conditional diffusion model; S4: Model training; S5: Classification results of generated anticancer peptide sequences; S6: Model Evaluation.

2. The method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model according to claim 1, characterized in that, Step S1 includes the following steps: S11: Data collection, collecting raw data from publicly available anticancer databases and literature, including unique identifiers of peptides, amino acid sequences, functional annotations, and cancer type labels; S12: Data cleaning, removing duplicate records and entries containing missing values, and filtering out illegal amino acid characters; S13: Sequence regularization, adjusting all sequences to a uniform length to form a matrix of equal-length sequences.

3. The method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model according to claim 1, characterized in that, Step S2 includes the following steps: S21: Extract features using a pre-trained protein language model; S22: Enhance features using MECS modules; S23: Use One-hot encoding to process tags; S24: Principal component analysis (PCA) is used to reduce the dimensionality of high-dimensional features while retaining the main information components.

4. The method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model according to claim 1, characterized in that, Step S22 includes: Channel attention mechanism: Calculates global information of features, generates channel weight coefficients, and highlights important channel features. The formula is as follows: ； Where pool represents the features obtained through average pooling, max pooling, and median pooling, Conv2d is the convolution operation, and σ is the Sigmoid activation function; Multi-scale convolution: Extracting multi-scale spatial features using different convolution kernels, the formula is as follows: SpatialFeatures= ; Spatial attention mechanism: A spatial attention map is generated based on the results of multi-scale convolution, and the formula is as follows: ； Feature fusion: Applying channel and spatial attention to the input features yields enhanced features, as shown in the formula: 。 5. The method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model according to claim 1, characterized in that, Step S3 includes the following steps: S31: Construct an encoder to map features to the latent space; S32: Conditional input, which converts label information into a conditional vector; S33: Construct the diffusion process, including forward diffusion and backward generation, using the MultiScaleResidualBlock module; S34: Construct a decoder that maps latent features back to the output space.

6. The method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model according to claim 1, characterized in that, In step S33, the MultiScaleResidualBlock module uses convolutional kernels of different scales to perform convolution operations on the input data. Feature outputs at different scales can be concatenated or fused, and unified processing is achieved through one or more nonlinear mapping structures. A residual connection mechanism is introduced, and the module input and output are weighted or directly superimposed. The calculation formula is as follows: ； Where is the transformation of Gi(x) at different scales, w i It's the weight.

7. The method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model according to claim 1, characterized in that, Step S4 includes the following steps: S41: Dataset preparation, including standardization, SMOTE oversampling, and noise perturbation; S42: Select the optimization algorithm, using stochastic gradient descent (SGD) and momentum; S43: Set the learning rate and batch size; S44: Model tuning and hyperparameter optimization.

8. The method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model according to claim 1, characterized in that, In step S42, the SGD optimization algorithm calculation formula is as follows: ； Where 𝜃𝑡 are the model parameters at the current time step 𝑡, 𝜂 is the learning rate, and ∇𝜃𝐿(𝜃𝑡) is the gradient of the loss function 𝐿 with respect to the model parameters 𝜃; The formula for calculating the momentum in stochastic gradient descent is: ； Where 𝑣𝑡 is momentum, 𝛽 is momentum coefficient, ∇𝜃𝐿(𝜃𝑡) is the current gradient, and 𝜂 is the learning rate.

9. The method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model according to claim 1, characterized in that, Step S5 includes the following steps: S51: The model receives the input protein sequence features and extracts their high-order semantic representation in the encoding module; S52: Label information is incorporated into the feature modeling process through conditional embedding to form a joint representation with category awareness; S53: Indicates that it is mapped into the latent space and enters the diffusion process; S54: In the diffusion modeling stage, the perturbation recovery and structural reconstruction of latent features are achieved by combining forward noise addition with backward stepwise denoising, and discriminative latent representations are generated under conditional constraints. S55: The decoding module maps the expression to the output label space to classify the anticancer peptide sequence.

10. The method for classifying anticancer peptide sequences based on a multi-scale enhanced conditional diffusion model according to claim 1, characterized in that, Step S6 includes the following steps: S61: The model is configured for inference or evaluation mode; S62: Obtain the model output results by performing batch predictions on the test set or validation set, and compare and analyze them with the corresponding real labels; S63: Generate classification reports and confusion matrices to visualize the classification results for each category; S64: Save the optimal model parameters based on the evaluation results for subsequent deployment or further application.

Images