A deep learning-based antibacterial peptide recognition and multi-attribute prediction method

Abstract

This invention discloses a deep learning-based method for antimicrobial peptide identification and multi-attribute prediction, relating to the field of bioinformatics. The method includes the following steps: obtaining the original amino acid sequence and activity tag; performing standardized preprocessing on the original amino acid sequence to generate a standardized short peptide sequence; performing three-dimensional conformation prediction and encoding on the standardized short peptide sequence, and calculating physicochemical features using a five-view physicochemical prior mechanism to construct a structural contact graph, a semantic embedding matrix, and a physicochemical feature matrix; concatenating and optimizing the physicochemical feature matrix and the semantic embedding matrix along the channel dimension to generate a robust residue node representation matrix; based on the robust residue node representation matrix and the structural contact graph, constructing a hybrid graph of main chain sequence edges and structural contact edges; training a prediction model for the original amino acid sequence to output the antimicrobial peptide identification probability result and multi-attribute activity prediction result of the original amino acid sequence.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method for identifying and predicting multiple attributes of antimicrobial peptides based on deep learning. Background Technology

[0002] Antibiotic resistance continues to spread in clinical settings and the environment. Coupled with the complexity of resistance mechanisms and cross-host proliferation, this undermines the effectiveness of traditional anti-infective therapies, evolving into a global public health crisis. Against this backdrop, antimicrobial peptides (AMPs), as core effector molecules of the innate immune system, possess unique short peptide structures, cationic properties, and amphiphilicity. They can exert broad-spectrum antimicrobial activity through multiple pathways, including disrupting membrane integrity, inhibiting biomembranes, and modulating immunity. Compared to traditional small-molecule antibiotics, antimicrobial peptides often have a lower risk of inducing resistance, thus being considered promising anti-infective drug candidates. However, the activity and safety of antimicrobial peptides cannot be fully characterized by a single functional tag; their scope and efficacy are the result of the synergistic effects of sequence characteristics, physicochemical properties (such as net charge and hydrophobicity), and three-dimensional conformation. Therefore, achieving joint characterization of function and properties is crucial for assessing activity and guiding sequence optimization. Given the high cost, long cycle, and inability to cover the vast sequence-structure combination space of traditional wet experimental screening, there is an urgent need to construct scalable computational screening systems to support large-scale candidate molecule discovery and multidimensional property prediction. This has become a key link in promoting the transition of antimicrobial peptides from basic mechanism research to drug development and application.

[0003] Current multi-attribute research often employs hierarchical or cascaded strategies (such as iAMP-2L and MLAMP) to address class imbalance. With the development of deep models, multi-attribute prediction is evolving towards a more unified and deeply integrated framework. TransImbAMP, combined with Transformer, improves the prediction performance of common functions; iAMP-CA2L introduces novel sequence-based image representation; iAMPCN, DMAMP, and deep-AMPpred integrate antimicrobial peptide recognition and multi-activity prediction within a unified model framework; furthermore, task-specific models such as TriNet and iAFPs-Mv-BiTCN demonstrate outstanding performance in predicting specific functions. Despite increasingly sophisticated architectural designs, mainstream methods still generally employ a "parallel independent head" strategy in their prediction mechanisms, assuming that attributes are independent of each other. This neglects the intrinsic biological dependencies and co-occurrence patterns between activities and lacks explicit joint modeling of activity tag associations and prior knowledge. Therefore, in scenarios with long-tailed activities and complex data distributions, the generalization ability and robustness of these models still have significant room for improvement. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a deep learning-based method for antimicrobial peptide identification and multi-attribute prediction. It proposes a unified deep learning framework, CLMP-AMP (Constrained Learning with MultimodalPresentation for AMP identification and multi-attribute prediction), to achieve integrated modeling for antimicrobial peptide identification and multi-attribute prediction. Figure 1 As shown, in the classification stage, this invention constructs a multimodal graph representation architecture, co-encoding residue-level PLM contextual representations, physicochemical priors, and structural constraints, and introduces a structure-aware graph attention mechanism to enhance the characterization of cross-site interactions and improve discrimination reliability. In the multi-attribute prediction stage, this invention employs a multi-task learning strategy based on shared encoding to explicitly capture co-occurrence and association patterns between attributes, combined with lightweight fine-tuning for single attributes, thereby significantly improving multi-attribute prediction performance. Overall, the CLMP-AMP model achieves deep integration of structure-aware multimodal recognition and tag-association-driven multi-activity prediction within a unified framework, providing a systematic computational solution for high-throughput screening and multidimensional functional characterization of antimicrobial peptides.

[0005] This invention provides a deep learning-based method for antimicrobial peptide identification and multi-attribute prediction, comprising the following steps:

[0006] S10: Obtain the original amino acid sequence and active tag, and perform standardized preprocessing on the original amino acid sequence to generate a standardized short peptide sequence;

[0007] S20: Perform three-dimensional conformation prediction and encoding on the standardized short peptide sequence, and calculate physicochemical features through a five-view physicochemical prior mechanism to construct a structural contact diagram, a semantic embedding matrix and a physicochemical feature matrix;

[0008] S30: The physicochemical feature matrix and the semantic embedding matrix are concatenated and optimized along the channel dimension to generate a robust residue node representation matrix;

[0009] S40: Based on the robust residue node characterization matrix and the structural contact graph, a hybrid graph of main chain sequence edges and structural contact edges is constructed. After classification head operation, an attribute correlation graph is constructed in combination with the active tag and cross-attribute association is performed. A prediction model of the original amino acid sequence is obtained through training.

[0010] S50: Input the robust residue node characterization matrix into the prediction model and output the antimicrobial peptide recognition probability result and multi-attribute activity prediction result of the original amino acid sequence.

[0011] In summary, the present invention has at least the following beneficial effects:

[0012] This invention first fuses residue embeddings extracted from the pre-trained protein language model ProtT5 with multi-perspective physicochemical prior information through a characterization fusion module. Then, a bidirectional long short-term memory network and a self-attention module are used to encode sequence dependencies, resulting in a unified residue node representation. Based on these node features, the identification branch constructs a hybrid graph topology composed of sequence adjacency edges and structural contact edges, and introduces a structure-aware graph attention mechanism to aggregate multimodal residue graph features, thereby achieving more reliable antimicrobial peptide identification. The multi-attribute prediction branch adopts a two-stage strategy of association-driven and decoupled fine-tuning. In stage one, graph Laplacian regularization is used to constrain attribute co-occurrence relationships to a shared parameter space to learn cross-attribute sharing patterns. In stage two, a lightweight attribute-specific predictor combined with a gating fusion mechanism dynamically integrates global shared information and task-specific evidence, thereby improving the accuracy and robustness of multi-attribute prediction.

[0013] 2. The CLMP-AMP model of this invention outperforms existing benchmark methods in key metrics such as accuracy (ACC), specificity (SPE), Matthews correlation coefficient (MCC), and area under the POC curve (AUC). The CLMP-AMP model not only demonstrates robust antimicrobial peptide recognition capabilities but also achieves competitive overall performance in multi-attribute prediction. In summary, the CLMP-AMP model successfully constructs a computational framework integrating recognition and multi-attribute prediction, providing a more reliable computational tool for high-throughput screening of antimicrobial peptides. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is the overall framework diagram of the CLMP-AMP model in this invention;

[0016] Figure 2 This is a rain cloud diagram of the peptide length distribution in the antimicrobial peptide dataset of this invention;

[0017] Figure 3 This is a diagram showing the overall amino acid distribution of the antimicrobial peptides and non-antimicrobial peptides in this invention.

[0018] Figure 4 This is a distribution diagram of amino acids in the five main functional categories of this invention;

[0019] Figure 5 This is the ROC curve of the comparative classification performance in this invention;

[0020] Figure 6 This is a PRE curve graph comparing the classification performance in this invention;

[0021] Figure 7 This is a distribution map of the original features visualized by t-SNE in this invention;

[0022] Figure 8 This is a distribution diagram of the features learned by the CLMP-AMP model in this invention, visualized using t-SNE.

[0023] Figure 9 This is the ROC curve of the predictive performance of the anti-mammalian cell functional activity in this invention;

[0024] Figure 10 This is the ROC curve of the predictive performance of the antibacterial functional activity in this invention;

[0025] Figure 11 This is the ROC curve of the predictive performance of the anticancer functional activity in this invention;

[0026] Figure 12 This is the ROC curve of the predictive performance of the antifungal functional activity in this invention;

[0027] Figure 13 This is the ROC curve of the predictive performance of the anti-Gram-negative cell functional activity in this invention;

[0028] Figure 14 This is the ROC curve of the predictive performance of the anti-Gram-positive cell functional activity in this invention;

[0029] Figure 15 This is the ROC curve of the predictive performance of the anti-biomembrane cell functional activity in this invention;

[0030] Figure 16 This is the ROC curve of the predictive performance of antiviral cell functional activity in this invention;

[0031] Figure 17 This is a heatmap for visual analysis of attribute correlation in this invention;

[0032] Figure 18 This is a visualization analysis heatmap of the sequence IHFKWRRWKFHI in this invention;

[0033] Figure 19 This is a visualization analysis heatmap of the sequence RFLVCWKQKIWGKARPSMCTRRARF in this invention;

[0034] Figure 20This is a visualization analysis heatmap of the sequence IPCGESCVWIPCISGMFGCSCKDKVCYS in this invention. Detailed Implementation

[0035] 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.

[0036] The following is in conjunction with the appendix Figures 1 to 20 The present invention will be described in further detail below.

[0037] This invention provides a deep learning-based method for antimicrobial peptide identification and multi-attribute prediction, comprising the following steps:

[0038] Step 1: Dataset Construction and Preprocessing.

[0039] This invention uses the publicly available benchmark dataset provided by Diff-AMP, whose core sequences are from the CAMP database. As shown in Table 1, two complementary subsets were constructed according to different uses: Dataset 1 contains 8225 antimicrobial peptide sequences, such as... Figure 2 As shown, dataset 1 is used to characterize the statistical regularities of positive samples in terms of sequence length and amino acid composition; dataset 2 is a balanced binary classification dataset filtered by the CD-HIT tool, containing 4134 positive samples of antimicrobial peptides and 4134 negative samples of non-antimicrobial peptides, used for training and evaluating the CLMP-AMP model, and the differences between positive and negative samples in the composition of 20 standard amino acids are shown in the figure. Figure 3 As shown.

[0040] For the multi-attribute prediction task, this invention uses the benchmark dataset provided by iAMPCN and performs final evaluation using an independent external test set published in the same study. This benchmark covers 22 functional and pharmacological activities, containing approximately 49,115 positive samples and approximately 195,525 negative samples filtered from the UniProt database; the sample size distribution for each attribute is shown in Table 1, and the differences in amino acid composition among the five main functional activities are as follows: Figure 4 As shown. The label construction follows the principles of iAMPCN and AMPfun: if a sequence is labeled with a certain activity from any source, then that activity dimension is marked as positive; otherwise, it is marked as negative.

[0041] It should be noted that Diff-AMP comes from Wang R, Wang T, Zhuo L, et al. Diff-AMP: tailored designed antimicrobial peptide framework with all-in-one generation, identification, prediction and optimization [J]. Briefings in Bioinformatics, 2024, 25(2). iAMPCN comes from Xu J, Li F, Li C, et al. iAMPCN: a deep-learning approach for identifying antimicrobial peptides and their functional activities[J]. Briefings in Bioinformatics, 2023, 24(4): bbad240. iAMPCN and AMPfun are derived from Chung CR, Kuo TR, Wu LC, et al. Characterization and identification of antimicrobialpeptides with different functional activities[J]. Briefings in bioinformatics, 2020, 21(3):1098-1114.

[0042] Table 1: Statistical information of peptide sequences in binary classification and multi-attribute prediction datasets

[0043]

[0044] To adapt to the two-stage training process, this invention merges and removes duplicate annotations of the same sequence in the training set across different activity dimensions, constructing a multi-label vector y∈{0,1}. TT represents the number of attributes, used for attribute association-guided training in Phase 1, while Phase 2 constructs a corresponding activity-specific subset for each attribute to achieve decoupled fine-tuning. To reduce bias introduced by data heterogeneity, all amino acid sequences are uniformly converted to uppercase single-letter codes, and samples containing non-standard amino acids (B / J / O / U / X / Z) are removed. To ensure the stability of short peptide characterization, sequence length is limited to 10–200 residues; ProtT5 embedding and physicochemical features are standardized with zero mean and unit variance based on training set statistics. Furthermore, this invention uses AlphaFold2 to predict the three-dimensional conformation of the cleaned sequence and constructs a contact map to generate structural contact edges, providing structural information for subsequent structure-aware modeling. Regarding data partitioning and evaluation, the antimicrobial peptide recognition task mainly employs stratified 10-fold cross-validation, retaining the fixed partitioning of Diff-AMP for ablation controls; the multi-activity prediction task follows the iAMPCN data partitioning protocol and is task-adapted; finally, evaluation is performed on an independent test set.

[0045] It should be noted that AlphaFold2 is derived from Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold[J].nature, 2021,596(7873):583-589. ProtT5 is a protein language model based on the T5 architecture in the ProtTrans model family.

[0046] Step 2: Learning the characterization of residue-level multimodal sequences.

[0047] The bioactivity of antimicrobial peptides is essentially the result of the synergistic effects of deep evolutionary constraints, local physicochemical motifs, and specific residue arrangements. To capture multi-scale information such as local patterns, sequence dependencies, and global interactions, this invention constructs a hierarchical multimodal coding module. This module first aligns PLM semantic features with multi-view physicochemical priors at the residue level through view-specific local projection; then, it explicitly models long-range dependencies across sites using BiLSTM and gating mechanisms, generating context-aware node representations, providing a feature foundation for subsequent structure-aware reasoning.

[0048] While single-protein language models (PLMs) can effectively capture residue-level contextual semantics, their explicit coverage of biophysical property details is often limited. In contrast, physicochemical and evolutionary priors are interpretable but lack the ability to model deep contextual dependencies. To achieve a complementary advantage, this invention constructs a five-view prior enhancement descriptor stack and employs a differentiated local projection strategy to preprocess shallow priors, mitigating the distribution problems caused by direct fusion of heterogeneous features. Given a batch size N and a sequence length L, the residue-level input is defined as:

[0049] ;

[0050] in, Let R represent the total input matrix before residue-level multimodal fusion, R represent the real number field, N represent the number of amino acid sequence samples input in one model training iteration, L represent the number of amino acid residues contained in a single peptide sequence, and D represent the total input matrix before residue fusion. in This represents the total input dimension after concatenating all features of U, A, P, C, S, and O. The amino acid residue embedding matrix generated by ProtT5 provides a macroscopic evolutionary semantic prototype; while A, P, C, S, and O represent the physicochemical scale (AAlndex) feature matrix, the pseudo-amino acid composition (PAAC) feature matrix, the physicochemical principal component (PC6) feature matrix, the substitution preference (BLOSUM62) feature matrix, and the identity encoding (one-hot) feature matrix, respectively.

[0051] Considering that directly concatenating low-order priors and high-order semantic embeddings may cause distribution mismatch, this invention adopts a differentiated local projection strategy. It maintains the original form of the ProtT5 embedding U, and applies dedicated one-dimensional convolution and layer normalization to each prior view to achieve local motif aggregation and dimensional alignment.

[0052] ;

[0053] Where M represents the original feature matrix of any physicochemical prior view. This represents the output feature matrix of M after local projection transformation. Represents view-specific convolution operators used to aggregate local features and unify dimensions; This indicates the number of channels after the view is convolved; Representation layer normalization is used to stabilize the feature scale of different views.

[0054] Subsequently, the prior features, after projection adaptation, are combined with the original embedding. By concatenating along the channel dimension, a semantically aligned multimodal fusion representation matrix is ​​constructed. :

[0055] ;

[0056] in, Indicates the channel dimension after fusion. This represents the projected prior feature matrices of the five physicochemical classes. This strategy, while preserving the deep semantic structure of the PLM, maps physicochemical and evolutionary signals to a more compatible latent space, providing more complete and consistent input features for subsequent sequence modeling.

[0057] To capture long program column dependencies, this invention will fuse features. Input a Bidirectional Long Short-Term Memory (BiLSTM) network. BiLSTM generates hidden states containing rich sequence semantics through bidirectional recursion. This bidirectional modeling enhances the CLMP-AMP model's ability to characterize long-range dependencies.

[0058] However, the fused multimodal representation matrix may still contain redundant or noisy channels. Therefore, this invention introduces an adaptive channel gating mechanism to reweight the BiLSTM output at the channel level. Specifically, the gating mechanism generates channel-level gating vectors. And on Recalibrate to obtain the gated feature matrix. :

[0059] ;

[0060] ;

[0061] in, This represents the feature matrix output by the BiLSTM network. This represents a channel-level gated vector, which acts as a gate in the feature / channel dimension and is broadcast to... Shape alignment for element-wise scaling; The weight matrix represents the gating mechanism. This represents the bias vector of the gating mechanism. Represents element-wise product. This is the Sigmoid activation function. The introduction of the gating mechanism effectively optimizes the channel weights of the features, improving the feature representation quality of the BiLSTM output.

[0062] To further improve the robustness of features, this invention enhances the robustness of gated features. Residual connections and layer normalization are introduced. Residual connections directly add the input to the output, which helps alleviate the gradient vanishing problem in deep networks; layer normalization normalizes the features of different channels to keep their scale consistent, thereby further improving the stability of the CLMP-AMP model and finally obtaining a robust node representation matrix. :

[0063] ;

[0064] The term "residual connection operator" is represented by , and "d" represents the final channel dimension of the residue node representation. By fusing BiLSTM, channel gating mechanisms, and residual connections with layer normalization, this module not only effectively models long sequence dependencies and suppresses redundant information but also improves the representativeness and accuracy of features. These optimized features provide a more accurate and robust input foundation for subsequent structure-aware graph attention networks, thereby enhancing the overall performance of the CLMP-AMP model in the antimicrobial peptide activity prediction task.

[0065] Step 3: Structure-aware recognition of antimicrobial peptides.

[0066] While multimodal sequence encoders can capture residue-level contextual semantics and physicochemical priors, the bioactivity of antimicrobial peptides is often dominated by three-dimensional folded topology, and relying solely on sequence neighborhoods is insufficient to fully characterize nonlocal spatial interactions. To address this, this invention constructs a sequence-structure hybrid graph based on a unified residue node representation and proposes a geometrically constrained structure-aware graph attention network. By injecting structural contact and sequence distance priors as inductive biases into the attention computation, long-range interaction modeling can be strengthened, thereby improving the reliability of antimicrobial peptide recognition.

[0067] To integrate the sequence continuity and spatial folding constraints of a peptide chain, for a given set of residue nodes... The construction sequence-structure hybrid diagram G of this invention:

[0068] ;

[0069] in, The set of edges representing the main chain sequence is used to ensure the basic connectivity of the graph and promote the smooth diffusion of local sequence information; i represents the position number of any residue node in the peptide chain. This represents the set of contact edges of the structure, derived from the contact map of the AlphaFold2 predicted structure, connected by conditions such that the spatial distance is less than 1 / 2. Non-adjacent residue pairs are explicitly introduced into the diagram to create spatial connections across sequence dimensions, thereby providing a propagable path for nonlocal structural actions.

[0070] To further characterize the heterogeneity of interactions between residues, this invention assigns a dual attribute to each edge in the hybrid graph. One attribute is the encoding of the sequence dimension, r. ij The sequence index difference |i−j| is discretely bucketed to distinguish short-range, medium-range, and long-range interaction patterns, thus explicitly injecting the sequence distance prior into the graph reasoning process. Secondly, the structural dimension w is encoded. ij ∈{0,1}, used to identify whether an edge originates from a spatial contact that satisfies a specific threshold. Here, j represents the position number of another residue node that forms an edge connection with i.

[0071] Based on the above encoding, this invention introduces a learnable scaling parameter, enabling the CLMP-AMP model to adaptively adjust the strength of structural information injection in the attention mechanism, thereby reducing sensitivity to prediction distance errors while preserving structural priors. Furthermore, this invention also designs a differentiated DropEdge regularization strategy, that is, for sequential edges... Imposing a higher random dropout rate causes the CLMP-AMP model to rely more on structural edges during information transmission. By performing cross-site information aggregation, the CLMP-AMP model is prevented from overfitting simple main chain sequence patterns and ignoring complex three-dimensional structural signals, thereby enhancing the robust perception of three-dimensional folding information.

[0072] Based on the constructed hybrid graph G, this invention proposes a geometrically constrained graph attention mechanism, aiming to encode the sequence dimension. Encoding with structural dimensions This is transformed into a learnable inductive bias to dynamically constrain the feature aggregation process. Let the input feature vector of any residue node n of the mixture graph G in the l-th layer of the graph attention network be... The present invention first performs linear projection to unify the feature space:

[0073] ;

[0074] ;

[0075] in, Indicates by The eigenvectors after linear projection Indicates the first The layer can learn the transformation matrix. and These represent the input and output channel dimensions, respectively.

[0076] For any edge in the hybrid graph G This invention constructs a scoring function that integrates semantic relevance and structural prior to calculate the nonnormalized attention coefficient. :

[0077] ;

[0078] Where m represents the neighboring nodes of n, Indicates by The eigenvectors after linear projection This represents a leaky ReLU activation function. Represents the attention parameter vector. This represents a learnable parameter used to adaptively adjust the overall contribution of structural information to attention logits; Represents structural dimension encoding. This represents the learnable bias corresponding to the sequence distance bins, used to encode the differences in the functional contribution of different sequence spans.

[0079] Thus, the CLMP-AMP model, while maintaining semantic attention dominance, incorporates the geometric priors of structural edges and sequential edges into the same scoring space, achieving unified coupling of semantic and topological information.

[0080] Subsequently, feature aggregation is performed using Softmax normalization and multi-head attention mechanism, and the next layer representation is generated through residual connections and nonlinear activation. :

[0081] ;

[0082] ;

[0083] ;

[0084] in, Indicates attention weights. Let represent the natural exponential function, and u represent all neighboring nodes of node n. Let n represent the neighborhood set of node n. This represents the feature vector obtained by aggregating the k-th attention at node n. Indicates the number of heads of attention. Indicates the l-th layer, the th The linear transformation matrix of the head. This represents the aggregated features of k attention heads. Splice along the channel dimension.

[0085] By stacking the aforementioned Structure-Aware GAT layers, this invention obtains the final node feature representation. It simultaneously encodes sequence semantics and dual geometric constraints, forming a structured representation for discrimination.

[0086] To obtain graph-level discriminant vectors from node-level representations, this invention employs a CLS active readout mechanism based on learnable queries to initialize a learnable virtual vector. As a global query, the residue features of the entire image are used. As keys and values, a graph-level representation is obtained through multi-head cross-attention aggregation:

[0087] ;

[0088] in, Graph-level vector representation This represents a multi-head cross-attention mechanism. This mechanism enables the CLMP-AMP model to actively aggregate key structural motifs most relevant to antibacterial activity and compress them into a global representation.

[0089] It should be noted that in the standard format of multi-head attention mechanisms, You need to enter the query, key, and value in sequence. This formula uses two... This is because in a global readout mechanism based on learnable queries, the query is represented by a separate virtual vector. The key (K) used to calculate attention similarity and the value (V) used for feature weighting aggregation both come from the same set of final node features. Therefore, two consecutive writes are required. The positions of the keys and values ​​are respectively assigned to satisfy the input specifications for multi-head cross-attention.

[0090] Finally, the probability of the antimicrobial peptide is output through the classification head, and then compared with the binary cross-entropy. Regular expression joint optimization:

[0091] ;

[0092] in, This represents the total loss function of the CLMP-AMP model. Indicates the first The predicted probability of a sample. Indicates the true label, This represents the set of learnable parameters of the CLMP-AMP model. Represents the regularity coefficient. Represents the parameter set Norm squared.

[0093] In summary, through an integrated design of hybrid topology-geometric constraint reasoning-active aggregation, the CLMP-AMP model achieves deep coupling between sequence semantics and three-dimensional topology, providing a more stable and interpretable structured basis for antimicrobial peptide recognition.

[0094] Step 4: Attribute association-guided multi-attribute prediction.

[0095] Antimicrobial peptides possess a series of multifunctional properties, and different attributes often exhibit co-occurrence dependencies driven by common physicochemical mechanisms and structural motifs. If multi-attribute prediction is simply treated as an independent parallel binary classification task, the correlations between activities cannot be effectively utilized, especially in long-tailed activity scenarios. Therefore, this invention proposes a two-stage framework consisting of association-guided training and decoupling fine-tuning.

[0096] Phase 1 injects the activity association topology into the shared representation space, encouraging strong-correlated activities to exhibit geometric consistency in decision parameters; Phase 2 performs lightweight independent fine-tuning on each activity on top of the shared encoder, achieving a controllable transition from global association learning to local specialization.

[0097] Phase 1 aims to learn cross-attribute sharing patterns based on residue-level sequence representations and explicitly constrain attribute co-occurrence structures to the task head parameter space. To extract multi-activity features from single sequence representations, this invention extends the single-query CLS readout used in classification tasks to multi-head query pooling.

[0098] Given residue features Introducing a learnable query matrix , serving as a multi-attribute semantic anchor point. For samples MHQPooling calculates the first Query headers and positions attention weights The residue value vectors are then weighted and aggregated; subsequently, the pooling results of each query header are concatenated along the channel dimension to obtain a shared sequence-level representation. .at last, via nonlinear mapping Forming a shared representation and output through the linear task header The predicted probabilities of each attribute:

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] ;

[0104] in, This represents the dimension of each query. The vector feature representing the e-th residue of sample b. Represents the key projection matrix. This represents the f-th query vector. Indicates the scaling factor. . The projection matrix represents the values. g represents the raw logits vector output by the task header, which is directly used to represent the prediction strength of the CLMP-AMP model for each attribute; Let B represent the task header weight matrix, and let B represent the bias term of the linear layer of the task header. This represents the predicted probability of each attribute.

[0105] Furthermore, to explicitly incorporate attribute co-occurrence relationships into the CLMP-AMP model, this invention constructs a label correlation graph and applies graph Laplacian regularization to the task header parameters. Specifically, the correlation coefficients between attributes are calculated based on the training set label matrix, and then sparsified to obtain the adjacency matrix T, which is subsequently used to construct the graph Laplacian matrix. Subsequently, this topological association is constrained to the task classification head weight matrix. superior:

[0106] ;

[0107] Where D represents the degree matrix and P represents the graph Laplacian matrix. This represents the graph Laplace regularization term. Represents the matrix trace operator. express The transpose of the matrix, Represents the adjacency matrix A of the th element. Line number The elements of the column.

[0108] This constraint encourages the classification weights of strongly correlated attributes to be closer in the parameter space, thereby achieving structured sharing and providing transferable discrimination patterns for long-tail attributes. Furthermore, to balance the learning difficulty and class imbalance among different activities, this invention employs a homoscedastic uncertainty weighting strategy for each attribute. Introducing learnable noise parameters To recalibrate the loss weights. The total loss for Phase 1 is defined as:

[0109] ;

[0110] in, This represents the total loss function for the first stage of multi-attribute prediction. Let represent the learnable noise parameter of the t-th attribute. This represents the binary cross-entropy loss of the t-th attribute. This indicates the strength of the control association constraints. The objective function improves multi-task fitting by explicitly encoding the active association topology in the shared representation and multi-task head.

[0111] Phase two employs a freeze-decoupling strategy. Building upon the shared encoder learned in Phase one that establishes attribute associations, a lightweight branch is trained for each target attribute to adapt to the long-tailed distribution and imbalanced characteristics of each attribute. For the target attribute... This invention constructs a lightweight attribute-specific original branch, directly using original residue-level features as input, and extracts local sequence features with high discriminative power for the current attribute through attribute-specific MHQPooling. .

[0112] Since different samples have varying degrees of dependence on shared knowledge and specific knowledge, simple linear concatenation is insufficient to effectively fuse the two. Therefore, this invention proposes an adaptive gating fusion mechanism: firstly, shared representations are compressed through bottlenecks to avoid over-reconstruction, and then gating coefficients are calculated. Adaptive modulation is applied to the information flow of the Raw branch:

[0113] ;

[0114] in, Represents the learnable weight matrix. This represents the learnable bias vector of the gating fusion mechanism. Represents the output features of the Raw branch The channel dimension.

[0115] Final characterization It is obtained by concatenating the compressed shared features with the gated modulated raw features:

[0116] ;

[0117] in, This represents the bottleneck compression operator.

[0118] when When sufficient to distinguish the current sample Suppressing the Raw branch reduces the risk of overfitting to small sample attributes; conversely... To enhance the specificity of the Raw branch and adjust the decision boundary, this invention considers that some long-tail attributes may exhibit extreme class imbalances. Relying solely on weighted loss may cause gradient oscillations and affect training stability. Therefore, this invention employs a resampling-based distribution balancing strategy in the fine-tuning stage to construct a more balanced local optimization space, and uses binary cross-entropy loss and L2 regularization to optimize the parameters of the active specific branch.

[0119] ;

[0120] in, This represents the loss function for the second stage of multi-attribute prediction. This represents the binary cross-entropy loss. Indicates the first The true label of each attribute Indicates the first The predicted probability of each attribute. This represents the learnable parameters of the Raw branch and the fusion header. This represents the regularization coefficient.

[0121] Through the two-stage design described above, the CLMP-AMP model learns the transferable attribute association structure in stage one and performs lightweight adaptation for each attribute in stage two, thereby significantly improving the overall performance of multi-activity prediction.

[0122] Validation of the predictive function of antimicrobial peptide recognition:

[0123] To evaluate the robustness of the CLMP-AMP model in the antimicrobial peptide recognition task, this invention performed hierarchical 10-fold cross-validation on a balanced benchmark dataset, as shown in Table 2. Experimental results show that the CLMP-AMP model maintains consistently low variance across folds and exhibits a stable convergence trend. This demonstrates that by deeply integrating sequence semantics and structural features, the CLMP-AMP model can effectively capture key discrimination patterns that distinguish between antimicrobial and non-antimicrobial peptides.

[0124] Table 2: Performance of the CLMP-AMP model on Dataset 2 based on 10-fold cross-validation

[0125]

[0126] Subsequently, this invention systematically compares the CLMP-AMP model with seven representative tools, as shown in Table 3. The comparison methods cover classic deep learning and machine learning models, including AMPScannerV2 based on CNN-RNN proposed by Veltri et al., AmPEPpy based on random forest, and MACREL, which is for genomic / metagenomic antimicrobial peptide screening and uses feature-driven classification. It also includes recent methods utilizing pre-trained protein language models or generative frameworks, such as deep-AMPpred using ESM-2 embedding, AMPlify based on attention mechanisms, and Diff-AMP based on diffusion models. Furthermore, this invention compares multi-perspective methods that integrate structural information or molecular dynamics features, including PGAT-ABPp based on graph attention and molecular dynamics-assisted strategies proposed by Cao et al. Overall, the CLMP-AMP model achieved the highest precision PRE (0.9243) and specificity SPE (0.9330), and outperformed all comparative methods in MCC (0.7560) and F1 (0.8677); at the same time, its ACC (0.8754) and AUC (0.9273) also remained at a high level, demonstrating a more balanced overall performance.

[0127] Table 3: Comparison with benchmarks of different models

[0128]

[0129] Furthermore, the ROC and PRE curves further validate the discriminative ability of the CLMP-AMP model, such as... Figure 4As shown in Figure 5, the CLMP-AMP model exhibits a steeper upward trend in the ROC space and covers a larger area under the curve in the PRE space. Compared to the baseline model, the CLMP-AMP model demonstrates a competitive advantage in the trade-off between precision (PRE) and recall (SNE), further confirming its effectiveness and reliability in complex biological sequence classification tasks.

[0130] To evaluate the feature extraction and representation learning capabilities of the CLMP-AMP model, this invention employs t-SNE to project sample features into a low-dimensional space for visualization. This invention systematically compares the distribution differences between the original input feature space and the deep latent representation space encoded by the CLMP-AMP model. For example... Figure 7 As shown in Figure 8, in the original feature space, the sample distributions of antimicrobial peptides and non-antimicrobial peptides show a high degree of overlap, making it difficult to form clear separable boundaries. This indicates that relying solely on shallow statistical descriptors is insufficient to capture complex patterns in biological sequences and provide effective discrimination signals.

[0131] In contrast, the latent space encoded by the CLMP-AMP model exhibits a significantly improved distribution pattern. For example... Figure 6 As shown, positive and negative samples achieve more efficient aggregation in the latent space, forming independent clusters with identifiable boundaries. This optimization of distribution characteristics is mainly attributed to the CLMP-AMP model's joint encoding of evolutionary semantics and structure awareness, enabling it to extract more discriminative deep representations from multimodal inputs. This evolution from feature overlap to spatial separation intuitively demonstrates that the CLMP-AMP model can capture key biophysical determinants related to the bioactivity of antimicrobial peptides, thus providing an interpretable basis for its classification performance.

[0132] To verify the effectiveness of each core component in the CLMP-AMP model framework and its contribution to overall performance, this invention designed a systematic ablation experiment, as shown in Table 4. First, at the feature representation level, removing the ProtT5 embedding caused the most significant performance degradation in the CLMP-AMP model, indicating that the deep contextual semantics implied by large-scale evolutionary information are the foundation for achieving high-precision recognition. Removing the five-view physicochemical features also led to a decline in MCC. Explicitly expressed biophysical properties (such as charge distribution and hydrophobic / hydrophobic patterns) can effectively supplement implicit semantic representations, and both together enhance the feature space's ability to characterize key biophysical determinants.

[0133] Table 4: Ablation experiments of different components in the CLMP-AMP model

[0134]

[0135] Secondly, regarding the topology of graph attention networks, this invention removes structural edges and main chain sequential edges, respectively. Experimental data shows that both result in performance loss. Therefore, a single sequence adjacency or spatial contact is insufficient to fully capture features, while the proposed sequential-structural hybrid topology can more effectively aggregate key functional motifs and improve discrimination performance within the folded conformational space.

[0136] Finally, by replacing the ProtT5 model with the ESM-2 model, the results showed that none of the metrics reached optimal levels. This may be attributed to the fact that the specific encoding mechanism of ProtT5 exhibited better compatibility and feature robustness under the multimodal fusion architecture of this invention. In summary, the complete CLMP-AMP model demonstrated the best performance across all evaluation dimensions, indicating that a positive synergistic effect was formed among the components of the CLMP-AMP model, jointly driving the significant improvement in the predictive performance of the CLMP-AMP model.

[0137] Performance comparison of antimicrobial peptide multi-attribute prediction:

[0138] To comprehensively evaluate the performance of the CLMP-AMP model in multi-attribute prediction tasks, this invention systematically benchmarks the CLMP-AMP model against six representative multi-attribute prediction methods currently in the field. These comparison methods cover various paradigms from traditional machine learning to deep learning. Specifically, AMAP employs a hierarchical multi-label framework, combining amino acid composition and physicochemical properties for prediction; AMPfun achieves multi-class functional prediction through feature selection and combining various machine learning classifiers (such as random forest, SVM, etc.); dbAMP constructs a database platform based on random forest for integrating antimicrobial peptide identification with functional and physicochemical annotations; iAMP-CA2L introduces image representation based on cellular automata and combines it with a CNN–BiLSTM architecture; iAMP-RAAC focuses on simplifying amino acid cluster features; and AMPDiscover provides a general prediction process based on unpaired features. Furthermore, this invention further incorporates the high-performance deep model iAMPCN as the main reference baseline.

[0139] The experiments were conducted on an independent test set, and detailed performance metrics are shown in Table 5. Overall results show that the CLMP-AMP model achieves competitive performance in most functional categories, particularly excelling in the MCC and AUC metrics, which reflect overall discriminative power. Figure 9-16As shown, this invention plotted ROC curves for eight representative activities. The results indicate that the CLMP-AMP model maintained a superior ROC curve shape across different attribute tasks, with its AUC (e.g., 0.8822 for antiviral and 0.7568 for antifungal) generally higher than the baseline method. This stable performance on common activities stems from the CLMP-AMP model's explicit modeling of co-occurrence relationships between activities: it can capture the potential dependencies between different biological activities and form mutually reinforcing representations.

[0140] Table 5: Performance comparison of different models on independent test datasets

[0141]

[0142] Table 5: Performance comparison of different models on independent test datasets (continued from Table 1)

[0143]

[0144] Table 5: Performance comparison of different models on independent test datasets (continued from Table 2)

[0145]

[0146] Table 5: Performance comparison of different models on independent test datasets (continued from Table 3)

[0147]

[0148] Table 5: Performance comparison of different models on independent test datasets (continued from Table 4)

[0149]

[0150] Furthermore, for long-tailed attributes with scarce samples (such as chemotaxis), the CLMP-AMP model improved the MCC to 0.7877 and AUC to 0.9388, while iAMPCN achieved 0.5774 and 0.8163 respectively. Performance improvements were also observed for other long-tailed activities (such as antiparasitic and insecticidal). Overall, the CLMP-AMP model significantly improved the predictive reliability of long-tailed activities while maintaining high-frequency activity prediction performance, demonstrating stronger robustness in challenging multi-label scenarios.

[0151] To investigate whether the CLMP-AMP model effectively captures and encodes potential biological attribute associations, this invention analyzes the multi-attribute weights learned by the CLMP-AMP model and constructs, for example... Figure 17The comparison heatmap is shown. The results show that the topological structure induced by the weights of the CLMP-AMP model is generally consistent with the label co-occurrence statistics, indicating that Laplace regularization can transform statistical label correlation into geometric proximity in the parameter space.

[0152] At the microscopic level, compared to the diffuse nature of the original statistical data, the parameter heatmap of the CLMP-AMP model exhibits a clearer and more compact topological structure. Specifically, the boundaries of strongly correlated functional clusters are significantly sharpened, and spurious correlation noise lacking biophysical support is effectively suppressed. This difference strongly demonstrates that deep networks possess inherent denoising and correlation reconstruction capabilities during feature learning, indicating that the CLMP-AMP model is not a simple fit to label co-occurrence statistics, but rather successfully extracts potential biological dependencies. This geometric constraint embedded in the parameter space essentially constructs a cross-task feature transfer mechanism, enabling long-tail attribute tasks with scarce samples to effectively transfer and utilize key discriminative patterns of head attributes. This mechanism shows that the CLMP-AMP model can still fundamentally enhance the biological consistency and generalization ability of multi-attribute predictions without relying on hard constraint rules.

[0153] Clarifying the interpretability of deep learning models is crucial for understanding the biophysical mechanisms behind their decisions and is a prerequisite for establishing the credibility of the CLMP-AMP model in the screening of novel antimicrobial peptides. To explore the decision-making basis of CLMP-AMP in antimicrobial peptide identification, this invention extracts cross-attention weights from the classification readout stage and uses these weights to quantify the importance of residue levels. Figure 9 As shown, this invention generates a feature saliency heatmap for three representative antimicrobial peptide sequences, where the color intensity at each position is proportional to the contribution of the corresponding residue to the classification decision of the CLMP-AMP model.

[0154] like Figure 18-20The weight distributions shown demonstrate that the CLMP-AMP model can accurately capture key sequence features that determine the bioactivity of antimicrobial peptides. Taking the sequences IHFKWRRWKFHI, RFLVCWKQKIWGKARPSMCTRRAR, and IPCGESCVWIPCISGMFGCSCKDKVCYS as examples, the CLMP-AMP model exhibits a high degree of focus on tryptophan (W) and is extremely sensitive to cationic-hydrophobic mixed regions. In particular, in IHFKWRRWKFHI, the core region of tryptophan (W) and positively charged arginine (R) together constitute a highly significant hot zone, a distribution feature highly consistent with the electrostatic adsorption mediated by cationic residues and the membrane anchoring function specific to tryptophan. Furthermore, in IPCGESCVWIPCISGMFGCSCKDKVCYS, consecutive hydrophobic residues (such as the VWI region) exhibit a highly significant response, reflecting the key driving role of the hydrophobic core in the membrane permeation and interaction process. This indicates that the CLMP-AMP model has effectively captured the intrinsic relationship between sequence features and biological activity, providing reliable interpretability support for the rational design of antimicrobial peptides.

[0155] The above are merely preferred embodiments of the invention and are not intended to limit the invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. A deep learning-based method for antimicrobial peptide identification and multi-attribute prediction, characterized in that, Includes the following steps: S10: Obtain the original amino acid sequence and active tag, and perform standardized preprocessing on the original amino acid sequence to generate a standardized short peptide sequence; S20: Perform three-dimensional conformation prediction and encoding on the standardized short peptide sequence, and calculate physicochemical features through a five-perspective physicochemical prior mechanism to construct a structural contact diagram, a semantic embedding matrix, and a physicochemical feature matrix; wherein, the five-perspective physicochemical prior mechanism includes AAIndex perspective, PAAC perspective, PC6 perspective, BLOSUM62 perspective, and one-hot perspective. S30: The physicochemical feature matrix and the semantic embedding matrix are concatenated along the channel dimension to generate a multimodal fusion representation matrix. The multimodal fusion representation matrix is ​​then optimized. The optimization method is as follows: The multimodal fusion representation matrix is ​​input into a BiLSTM network to capture long program column dependencies. The BiLSTM network is set to 2 layers with a hidden layer dimension of 512. Dropout regularization and adaptive channel gating mechanisms are used to reweight the features output by the BiLSTM network. Then, residual connections and layer normalization operations are combined to optimize the feature representation and generate a robust residue node representation matrix. S40: Based on the robust residue node characterization matrix and the structural contact graph, a hybrid graph of main chain sequence edges and structural contact edges is constructed. After classification head operation, an attribute correlation graph is constructed in combination with the active tag and cross-attribute association is performed. A prediction model of the original amino acid sequence is obtained through training. Specifically, the robust residue node representation matrix is ​​used as input, and the sequence-level shared representation matrix is ​​extracted through MHQPooling. An attribute correlation graph is constructed based on the multi-attribute labels in the active labels, and the correlation coefficient matrix between any two different active labels is calculated as the edge weight of the attribute correlation graph. A graph Laplacian matrix is ​​constructed based on the attribute correlation graph with edge weights, and graph Laplacian regularization based on the graph Laplacian matrix is ​​applied to the multi-attribute prediction task header parameters. Combined with the homoscedasticity uncertainty weighting strategy, the loss weights corresponding to each multi-attribute label are balanced. The parameters of the shared encoder corresponding to the sequence-level shared representation matrix are frozen and extracted. A lightweight dedicated branch is constructed for each multi-attribute active label in the active label. Local features of the sequence are extracted through the lightweight dedicated branch. The fusion weights are generated by using the sigmoid activation function, and then the sequence-level shared representation matrix and the sequence local features are fused using the fusion weights. A resampling strategy is adopted to balance the distribution of training data corresponding to each of the multi-attribute labels. A binary cross-entropy loss function combined with L2 regularization is used to optimize the parameters of each of the lightweight dedicated branches, so as to obtain the multi-attribute prediction model of the original amino acid sequence after training. S50: Input the robust residue node characterization matrix into the prediction model and output the antimicrobial peptide recognition probability result and multi-attribute activity prediction result of the original amino acid sequence.

2. The method for antimicrobial peptide identification and multi-attribute prediction based on deep learning according to claim 1, characterized in that, The specific steps for standardizing the original amino acid sequence to generate a standardized short peptide sequence are as follows: All the original amino acid sequences were converted to uppercase single-letter codes, and non-standard amino acid residues were deleted; then, amino acid sequences with 10-200 residues were selected to generate the standardized short peptide sequences.

3. The method for antimicrobial peptide identification and multi-attribute prediction based on deep learning according to claim 1, characterized in that, The specific steps for encoding the standardized short peptide sequence are as follows: The normalized short peptide sequences were encoded using the ProtT5 model to generate the original semantic embedding matrix; The original semantic embedding matrix is ​​standardized with zero mean and unit variance to generate the semantic embedding matrix.

4. The method for antimicrobial peptide identification and multi-attribute prediction based on deep learning according to claim 1, characterized in that, The AAIndex perspective is used to extract multidimensional physicochemical indicators of amino acid residues; The PAAC perspective is used to calculate pseudo-amino acid composition features that take into account both amino acid sequence composition and local sequence dependence. The PC6 perspective is used to perform principal component analysis on high-dimensional physicochemical indicators and extract six core principal components; The BLOSUM62 perspective is used to extract amino acid evolutionary substitution preference features; The one-hot perspective is used to identify standard amino acids.

5. The method for antimicrobial peptide identification and multi-attribute prediction based on deep learning according to claim 4, characterized in that, The specific steps for calculating the physicochemical characteristics of the standardized short peptide sequence and generating the physicochemical characteristic matrix using a five-perspective physicochemical prior mechanism are as follows: The physicochemical features extracted by the five-perspective physicochemical prior mechanism are spliced ​​and integrated to generate the original physicochemical feature matrix; The original physicochemical feature matrix is ​​then standardized with zero mean and unit variance to generate the physicochemical feature matrix.

6. The method for antimicrobial peptide identification and multi-attribute prediction based on deep learning according to claim 4, characterized in that, The specific steps for concatenating the physicochemical feature matrix and the semantic embedding matrix along the channel dimension to generate a multimodal fusion representation matrix are as follows: A 1D convolution operator is used to perform local phantom aggregation on the physicochemical features of each viewpoint. The kernel size of the 1D convolution operator is set to 3, the stride is set to 1, and the padding method is set to same. Each viewpoint corresponds to independent convolution parameters. After convolution, the feature scale is stabilized by layer normalization operation to obtain the adapted physicochemical feature matrix. The adapted physicochemical feature matrix and the semantic embedding matrix are concatenated along the third channel dimension to generate a multimodal fusion representation matrix.

7. The method for antimicrobial peptide identification and multi-attribute prediction based on deep learning according to claim 6, characterized in that, Based on the robust residue node characterization and the structural contact graph, the specific steps for constructing a hybrid graph of main chain sequence edges and structural contact edges are as follows: The amino acid residues of the standardized short peptide sequence are used as nodes in the hybrid graph, and the feature of the hybrid graph node is the feature vector of the corresponding amino acid residue in the robust residue node characterization matrix. The main chain sequence edge is constructed using any two amino acid residues in the standardized short peptide sequence with a sequence distance of no more than 3. The edge weight of the main chain sequence edge is the reciprocal of the sequence distance between the corresponding two amino acid residues in the standardized short peptide sequence. The spatial distance in the contact diagram of the structure is no greater than Any two amino acid residues are used to construct the contact edge of the structure, and the edge weight of the contact edge is the contact strength of the corresponding two amino acid residues in the contact graph.

8. The method for antimicrobial peptide identification and multi-attribute prediction based on deep learning according to claim 7, characterized in that, The specific steps for outputting the recognition probability result of antimicrobial peptides through classification head operation are as follows: A differentiated DropEdge regularization strategy is applied to the hybrid graph, setting different drop probabilities for the main chain sequential edges and the structural contact edges respectively; The sequence distance encoding of the main chain sequential edges and the spatial contact strength encoding of the structural contact edges are used as the attention scoring function of the graph attention mechanism with inductive bias. The features of the hybrid graph nodes are aggregated by 8-head attention and generated after residual connection and layer normalization operations. The full-graph node features of the hybrid graph are aggregated through the CLS active readout mechanism to obtain a graph-level representation matrix; The graph-level representation matrix is ​​input into a fully connected classification head, mapped to probability values ​​between 0 and 1 by a Sigmoid activation function, and jointly optimized using a binary cross-entropy loss function combined with L2 regularization to output the recognition probability result of the antimicrobial peptide.

Images