Antibacterial peptide activity and mic value combined prediction framework based on cross-modal deep learning

Abstract

The application belongs to the technical field of antibacterial peptide activity recognition and evaluation, and relates to an antibacterial peptide activity and MIC value combined prediction framework based on cross-modal deep learning, which utilizes a protein language model ESM2 to respectively perform token-level semantic embedding coding on antibacterial peptide sequences and pathogen protein sequences; cross-modal feature extraction and fusion are performed through a multi-branch structure comprising a Mamba module, a multi-head self-attention mechanism and a DASM 1D convolution; a multi-task decoding structure is adopted to simultaneously realize antibacterial peptide activity classification and MIC value regression prediction; the application can effectively mine functional features in sequences, significantly improve the accuracy and generalization ability of antibacterial peptide activity and MIC value prediction, and provides a reliable calculation tool for high-throughput screening and rational design of antibacterial peptides.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of antibacterial peptide activity recognition and evaluation, and particularly relates to an antibacterial peptide activity and MIC value joint prediction framework based on cross-modal deep learning. BACKGROUND

[0002] Antibiotic resistance has become a major challenge in the field of global public health. Driven by the overuse of broad-spectrum antibiotics, the problem of bacterial drug resistance has risen significantly, leading to a continuous decline in the effectiveness of traditional antibiotic therapy. In such a severe situation, the clinical management of common infectious diseases is facing the dilemma of prolonged treatment period and skyrocketing medical expenses. Moreover, due to the cross-transmission of drug-resistant strains in medical institutions, it frequently causes outbreaks of nosocomial infections, seriously threatening the public health security system. Therefore, the development of antibacterial drugs with new mechanisms has become one of the important scientific propositions in the global biomedical field.

[0003] Natural antibacterial peptides are considered to be the most promising new generation of antibacterial drugs due to their unique membrane lysis mechanism and the difficulty of inducing drug resistance. However, the traditional development mode is limited by the long database construction period and the difficulty in analyzing the structure-activity relationship, resulting in high costs of new drug development. With the rapid development of artificial intelligence technology, it has brought a revolutionary change to this field - the deep learning-based generative model can efficiently mine antibacterial peptide sequence features and intelligently predict molecular activity and toxicity data. Through virtual screening technology, the candidate molecule discovery period is shortened from several years to several weeks, significantly reducing the cost of experimental verification.

[0004] In this process, the minimum inhibitory concentration (MIC) value plays a key role as an important indicator of the effectiveness of antibacterial drugs. The MIC value reflects the minimum drug concentration required to inhibit bacterial growth and is a direct quantitative standard for evaluating the inhibitory effect of antibacterial peptides on bacteria. By predicting the MIC value of antibacterial peptides, deep learning technology not only accelerates the development process of new antibacterial peptides, but also provides more accurate prediction and screening methods to help scientists identify potential candidate molecules in the early stages. This technological innovation not only accelerates the development of new antibacterial peptides, but also provides a new solution to the problem of drug-resistant strains. SUMMARY

[0005] The purpose of the present application is to overcome the deficiencies in the prior art and provide an antibacterial peptide activity and MIC value joint prediction framework based on cross-modal deep learning.

[0006] To achieve the purpose of the present application, the following technical solutions are adopted.

[0007] The antibacterial peptide activity and MIC value joint prediction framework based on cross-modal deep learning comprises:

[0008] The sequence encoding module uses a pre-trained protein language model to embed antimicrobial peptide sequences and pathogen protein sequences, respectively, and introduces learnable position encoding to inject position information into the antimicrobial peptide embedding sequence and the pathogen protein embedding sequence; wherein, the antimicrobial peptide sequence has an active MIC tag;

[0009] The feature extraction and fusion module includes a multi-layered stacked residual structure, which includes:

[0010] The Mamba module includes a parallel multi-head self-attention branch, a Mamba+GLU branch and a DASM 1D convolution branch for simultaneously capturing global long-range dependency features, gated weighted fusion features and local and global fusion features from antimicrobial peptide embedding sequences and pathogen protein embedding sequences; a learnable gating and SE unit for dynamically weighted fusion of global long-range dependency features, gated weighted fusion features and local and global fusion features and adaptively adjusting channel feature weights; and a feature mixing unit for explicitly injecting the dynamically weighted fusion features of the pathogen modality into the dynamically weighted fusion features of the antimicrobial peptide modality.

[0011] Feedforward neural networks, through position-by-position linear transformation of the structure, enable the features input from the Mamba module to have smooth gradients while possessing non-linear expressive power.

[0012] Residual connections and layer normalization units are used to simultaneously preserve the transformation results and original input information of the feedforward neural network;

[0013] The decoding and prediction module is used to decode the features input from the feature extraction and fusion module into global feature vectors for the activity classification task and the MIC value regression task. It predicts the probability distribution of the activity category and continuously predicts the MIC value through parallel classification and regression branches, respectively.

[0014] Furthermore, the active MIC tag is generated by logarithmic transformation of the experimentally determined minimum inhibitory concentration (MIC) data.

[0015] Furthermore, the antimicrobial peptide sequence and the pathogen protein sequence are obtained by preprocessing the original antimicrobial peptide sequence and the pathogen protein sequence in FASTA format, respectively; the preprocessing includes removing irrelevant characters, normalizing amino acid residue representation, and limiting the maximum sequence length; wherein: normalized amino acid residue representation is performed by formula (1), which is:

[0016] (1);

[0017] In the formula: Indicates the first The amino acid sequence of an antimicrobial peptide or pathogen. For the first one amino acid residue, a set of amino acid residues supported by the protein language model, is the actual length of the sequence, and , represents the maximum sequence length.

[0018] Further, the embedding encoding is represented as:

[0019] (2);

[0020] wherein: is the embedding representation matrix of the antibacterial peptide sequence, the number of rows is the actual length of the sequence , each row corresponds to a feature vector of an amino acid residue, is the output embedding dimension of the protein language model.

[0021] Further, the specific process of the embedding encoding includes the following steps:

[0022] S51, indexing the sequence input into the protein language model, mapping each residue character in the sequence to an integer number to generate an indexed sequence; wherein: the sequence is an antibacterial peptide sequence or a pathogen protein sequence;

[0023] S52, adding start token BOS and end token EOS at the beginning and end of the sequence to indicate the sequence boundary;

[0024] S53, using a learnable embedding matrix to map the indexed sequence into a dense word vector, and superimposing a learnable position encoding to provide residue position information;

[0025] S54, inputting the dense word vector with residue position information into a multi-layer Transformer encoder of the protein language model for encoding;

[0026] S55, after the encoding is completed, the vectors corresponding to the BOS / EOS tokens at the beginning and end are removed, and only the representation of the real residues is retained, to output a residue-level embedding matrix with a shape of , and binding with a unique ID of the sample to form an antibacterial peptide embedding feature or a pathogen embedding feature.

[0027] Further, the Mamba module includes a parallel branch unit, a learnable gating and SE unit, and a feature mixing unit introducing a cross-modal interaction attention mechanism, and the parallel branch unit includes a parallel multi-head self-attention mechanism branch, a Mamba+GLU branch, and a DASM 1D convolution branch; wherein:

[0028] Mamba+GLU branch, long-range dependency features are extracted by Mamba state space model, and functional fragments are screened by GLU gating mechanism, and the gated weighted fusion features are obtained by element-by-element multiplication operation;

[0029] Multi-head self-attention branch, global long-range dependency relationship features are extracted by multi-head self-attention mechanism;

[0030] DASM 1D convolution branch, local context semantic information is extracted by dynamic sampling and channel convolution, and global long-range dependency relationship features are supplemented, to obtain local and global fusion features;

[0031] Learnable gating and SE unit, branch gating weighting mechanism and channel adaptive mechanism SE are introduced, to dynamically weight fuse the locally and globally fused features, the gated weighted fusion features and the global long-range dependency relationship features, and to adaptively adjust the channel feature weights in the channel dimension, to obtain the dynamic weighted fusion features of the antimicrobial peptide mode and the pathogen mode;

[0032] Feature mixing unit, the antimicrobial peptide mode dynamic weighted fusion features are mapped to the cross-modal interaction dimension through the full connection layer to obtain the antimicrobial peptide projection features, and the antimicrobial peptide projection features are input as the query vector into the introduced cross-modal attention mechanism, and the pathogen mode dynamic weighted fusion features are input as the key vector and the value vector into the introduced cross-modal attention mechanism to match the query vector; cross-modal attention mechanism, the attention weight is calculated according to the similarity of the query vector and the key vector, and the value vector is weighted aggregated based on the calculated attention weight, to generate the cross-modal interaction fusion features of the antimicrobial peptide mode and the pathogen mode.

[0033] Further, the cross-modal interaction fusion features are input into the feedforward neural network, the feedforward neural network adopts the position-by-position linear transformation structure, independently transforms the features of each position without mixing the information between positions, so that the modal interaction fusion features have nonlinear expression ability while having smooth gradient, to generate the cross-modal interaction fusion enhanced features.

[0034] Further, the cross-modal interaction fusion enhanced features are input into the residual connection and layer normalization unit, through the residual connection and layer normalization, the output cross-modal interaction fusion enhanced features retain the feedforward transformation result and the original input information at the same time, to relieve the gradient vanishing, accelerate the convergence and deepen the network modeling depth.

[0035] Further, the global feature vector of the activity classification task is input into the full connection layer of the classification branch, and is projected into classification features after nonlinear transformation, normalization and regularization processing, and then the classification features are converted into the probability distribution of the activity class through the activation function, to realize the activity classification prediction.

[0036] Further, the global feature vector input regression branch of the MIC value regression task is combined with normalization and regularization to obtain the continuous prediction of the MIC value through two layers of full connection layers.

[0037] Compared with the prior art, the present application has the following beneficial effects:

[0038] The module structure provided by the present application has high scalability and universality, can flexibly adapt to different types of antibacterial peptide-pathogen data, significantly improves the prediction performance and generalization ability of the overall model, and provides strong technical support for efficient screening and mechanism analysis of new antibacterial peptides. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The framework diagram of the prediction model is shown in the figure;

[0040] Figure 2 The framework diagram of the Mamba module is shown in the figure;

[0041] Figure 3 The processing process diagram of the protein language model is shown in the figure;

[0042] Figure 4 The processing process diagram of the input sequence is shown in the figure;

[0043] Figure 5 The flowchart of the DASM 1D convolution branch is shown in the figure;

[0044] Figure 6 The processing process diagram of the DASM 1D branch is shown in the figure;

[0045] Figure 7 The framework diagram of the feature mixing and cross-modal multi-head attention module is shown in the figure;

[0046] Figure 8 The processing process diagram of the convolution-position encoding block is shown in the figure. DETAILED DESCRIPTION

[0047] The present application will be further described below in conjunction with the accompanying drawings. The following examples are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0048] I. Model Introduction

[0049] The antibacterial peptide and pathogen activity prediction framework provided by the present application is composed of three core modules, namely sequence encoding module, feature extraction and fusion module and decoding prediction module, as shown in the figure. Figure 1 The modules work together to automatically extract effective representations from multi-modal sequence information and jointly predict the activity and MIC value of antibacterial peptides.

[0050] The sequence encoding module takes the antibacterial peptide and pathogen protein sequences as input, both expressed in FASTA format. Since the MIC original values span multiple orders of magnitude and are highly skewed in distribution, direct use will lead to high-value samples dominating the model learning in training, so the MIC is first log-transformed before modeling to compress the value range, reduce the scale difference, and improve the distribution characteristics. Subsequently, the pre-trained protein language model ESM2 is used to embed the antibacterial peptide and pathogen sequences in high dimensions, fully extracting the context semantics and structural dependency information of the sequences, and automatically generating activity labels based on the processed MIC data, thereby providing reliable supervision signals for subsequent classification and regression multi-task joint modeling.

[0051] The feature extraction and fusion module, the embedded features of the antibacterial peptide and the pathogen are jointly input into the Mamba module in the feature extraction and fusion module, as shown in Figure 1 The specific structure of the Mamba module is shown in Figure 2 The internal adopts a parallel three-branch design: the first branch is a DASM 1D convolution branch that extracts local context patterns through dynamic sampling and channel convolution, and supplements the global cross-location dependency relationship of the multi-head self-attention branch, realizing the fusion of local and global features. The second branch is the Mamba+GLU branch, which directly receives the embedded features of the antibacterial peptide and the pathogen, regulates information flow through the GLU gating mechanism, and captures long-range dependencies and global sequence patterns by combining state space modeling structures, which has a significant advantage in long-distance information transmission and key state preservation. The third branch is the multi-head self-attention branch, which also directly receives the embedded features of the antibacterial peptide and the pathogen, establishes cross-location dependency relationships in the global range, and enhances the aggregation of long-distance information and the ability to integrate cross-segment semantics.

[0052] The multi-level features output by the three branches are dynamically weighted by learnable gating weights and combined with the SE module to realize channel adaptive enhancement, to ensure that the fused representation reaches an optimal balance between global, local, and long-distance information modeling. The dual-modal fusion features integrated by the hybrid layer are sent to the cross-modal interaction attention mechanism to establish explicit interaction encoding between the antibacterial peptide and pathogen sequences at the token level, effectively depicting their potential mechanisms of action in structure and function.

[0053] The feature extraction and fusion module is sequentially inputted into a multi-layer stacked residual structure, each layer of which is sequentially composed of a Mamba module and a position-by-position feed-forward network. The Mamba module is used to capture long-range dependencies and global patterns, and the feed-forward network is used to enhance the non-linear expression capability of the features. Both are supplemented by layer normalization and residual connection to stabilize gradient transmission and improve network modeling depth. After the multi-layer stacked residual structure, the fusion features enter the feature enhancement and pooling stage to compress the sequence dimension and retain key information, and then are sent to the multi-task prediction output head: the classification branch uses a softmax activation function to realize the prediction of the activity of the antibacterial peptide, and the regression branch uses a linear output unit to realize the accurate regression prediction of the MIC value, so as to realize high precision and high efficiency in the activity screening and functional quantitative evaluation of the candidate antibacterial peptide.

[0054] 1. Representation of input data

[0055] The antibacterial peptide activity and MIC value joint prediction framework based on cross-modal deep learning takes the antibacterial peptide (AMP) sequence and the pathogen protein sequence as input, representing the antibacterial peptide candidate molecule to be screened and the target pathogen. First, the original FASTA format antibacterial peptide and pathogen protein sequence are standardized and preprocessed, and the pre-trained protein language model ESM2 is used to extract high-dimensional token-level embedding features of the two types of molecules. High-dimensional token-level embedding features not only contain the context dependence and global structure relationship between amino acid residues in the sequence, but also can reflect the potential functional region and molecular biology semantics, providing a rich expression basis for subsequent end-to-end feature modeling and molecular-level interaction analysis.

[0056] In terms of label system construction, combined with the minimum inhibitory concentration (MIC) data determined by experiments, and according to the threshold division rule, a binary activity label (active is 1, inactive is 0) is given to each antibacterial peptide to support the multi-task output of the model, including activity classification and MIC value regression prediction. Finally, the high-dimensional token-level embedding features of the antibacterial peptide and the pathogen protein are input into the multi-branch feature extraction module and the cross-modal interaction fusion module, respectively, to realize multi-scale expression, information complementation and token-level fine alignment of molecular sequences, thereby laying a solid data foundation and information support for drug peptide activity recognition and MIC value accurate prediction.

[0057] 2. Sequence encoding module

[0058] The module is one of the three core modules of the application, located at the starting stage of the entire model process, aiming to realize the uniform high-dimensional embedding expression of antimicrobial peptides (AMP) and pathogen protein sequences, and provide consistent, fine and cross-modal associated input features for the subsequent feature extraction and fusion module and decoding prediction module. The module introduces an advanced pre-trained protein language model ESM2 to perform end-to-end token-level semantic embedding processing on the molecular sequence. Unlike the existing technology which only uses a global mean vector, the application can retain the local features of each amino acid residue and simultaneously capture the overall structure pattern and context dependency of the sequence, thereby significantly improving the expression ability and deep modeling effect of the molecular sequence. The obtained high-dimensional embedding features will work with the Mamba module, DASM-1D convolution unit and cross-modal attention mechanism in the subsequent processing stage to realize cross-modal deep feature fusion and multi-task function prediction. The specific implementation of antimicrobial peptide sequence encoding and pathogen sequence encoding will be introduced as follows:

[0059] (1) Antimicrobial peptide sequence encoding

[0060] The module first performs data preprocessing on the original antimicrobial peptide FASTA sequence, including removing irrelevant characters, standardizing amino acid residue representation, and limiting the maximum sequence length . Let the standardized sequence of the th antimicrobial peptide be:

[0061] (1);

[0062] Wherein, represents the amino acid sequence of the th antimicrobial peptide or pathogen, is the th amino acid residue, is the set of amino acid residues supported by ESM2 (usually containing 20 standard amino acids and special symbols), is the actual length of the sequence, and .

[0063] This formula converts the original character type amino acid sequence into a symbolic representation that can be processed mathematically, providing standardized input for subsequent feature extraction.

[0064] After obtaining the standardized sequence, in order to capture the local and global semantic dependency in the antimicrobial peptide sequence, the pre-trained protein language model ESM2 is used to perform high-dimensional token-level embedding mapping on , obtaining:

[0065] (2);

[0066] Wherein, is the embedding matrix for the antimicrobial peptide sequence, with the number of rows being the actual length of the sequence , where each row corresponds to a feature vector of an amino acid residue, is the output embedding dimension of the ESM2 model (1280 in the present application). This process can convert the symbolic amino acid sequence into a dense numerical vector representation, so that the deep learning model can directly calculate and capture patterns.

[0067] To facilitate the intuitive understanding of the input and output relationship of ESM2 in the present application, Figure 3 The data flow process after the antimicrobial peptide sequence enters the ESM2 model is shown. First, the preprocessed amino acid sequence is input, and each residue character is mapped to an integer number by indexing; then the start marker (BOS) and the end marker (EOS) are added at the beginning and end of the sequence to indicate the sequence boundary; then the index sequence is mapped to a dense word vector using a learnable embedding matrix, and a learnable position encoding is added to provide residue position information for the model; then the ESM2 multi-layer Transformer encoder is input to model the global dependence of the sequence and extract the context features; after the encoding is completed, the vectors corresponding to the BOS / EOS markers at the beginning and end are removed, and only the representation of the real residues is retained; finally, the residue-level embedding matrix with a shape of is output, which is bound with the unique ID of the sample and used as the input of the subsequent feature fusion and prediction module.

[0068] To more clearly show the entire processing flow, the input sequence "C-A-T" is visualized at each processing stage in the ESM2 model, as shown in Figure 4 .

[0069] To support batch storage and fast indexing, the embedding results of all antimicrobial peptide sequences and their unique identifiers (id) are combined into a set:

[0070] (3);

[0071] wherein, is the set of embedding results of all antimicrobial peptide sequences, is the total number of antimicrobial peptides. This set can be serialized into a JSON or binary file format, which facilitates efficient loading and calling when performing cross-modal fusion with pathogen protein embedding features in the subsequent step. In the data flow, it plays the role of information storage and index mapping, so that the subsequent multi-task prediction (activity classification + MIC value regression) can be calculated based on the unified antimicrobial peptide feature representation.

[0072] (2) Pathogen sequence encoding

[0073] For pathogen protein sequences, the original data adopts a JSON structure and is grouped by TaxID. For each TaxID, the longest sequence is selected as the representative sequence, and the sequence is preprocessed in the same way as the antimicrobial peptide sequence. The sequence is first indexed to map each residue character to an integer number; then the start marker (BOS) and the end marker (EOS) are added at the beginning and end of the sequence to indicate the sequence boundary; then the index sequence is mapped to a dense word vector using a learnable embedding matrix, and a learnable position encoding is added to provide residue position information for the model; then the ESM2 multi-layer Transformer encoder is input to model the global dependence of the sequence and extract the context features; after the encoding is completed, the vectors corresponding to the BOS / EOS markers at the beginning and end are removed, and only the representation of the real residues is retained; finally, the residue-level embedding matrix with a shape of representative protein sequences, and uniformly truncated to the maximum length. Let TaxID be , its protein sequence set is . Similarly, ESM2 is used for high-dimensional token-level embedding representation, specifically:

[0074] (4);

[0075] where, represents the th protein sequence of TaxID , the token-level embedding matrix after ESM2 model encoding, the number of rows is the sequence length , each row is a high-dimensional feature vector of an amino acid residue, is the embedding dimension. This formula maps the symbolic amino acid sequence to a dense vector space that can directly participate in deep learning calculations, thereby laying the foundation for cross-modal feature fusion.

[0076] Within TaxID, all sequence embedding results are aggregated into a set:

[0077] (5);

[0078] where, is the token-level embedding set of all representative protein sequences of TaxID , which can be used for batch processing and parallel acceleration. Organizing and centrally managing multiple protein sequence embeddings of the same pathogen facilitates subsequent feature aggregation and modeling based on the pathogen level.

[0079] Further, the embedding set of all pathogens is organized with TaxID as the key:

[0080] (6);

[0081] where, is the global pathogen embedding index table, each element is a two-tuple , which can be serialized into JSON or binary file, facilitating efficient reading and distributed loading of the model in the training and inference stage. This formula establishes a mapping relationship from the unique identifier of the pathogen to its feature set, enabling subsequent cross-modal feature matching, activity classification, and MIC value regression prediction of AMP and pathogens to be directly and efficiently called based on the index.

[0082] This sequence encoding module realizes the encoding of the antimicrobial peptide and the protein of the pathogen End-to-end high-dimensional token-level unified modeling of the sequence. Unlike existing technologies that only compress the sequence into a global mean vector, the present invention retains both the local segment features and the global dependency of the molecular sequence during the encoding process, and automatically learns the high-dimensional representation of the functional regions through the ESM2 model, thereby ensuring the consistency and alignment of the multi-modal features in the subsequent feature fusion stage.

[0083] High-dimensional embedding features of the encoding output With The direct input feature extraction and fusion module cooperates with the Mamba module, the DASM-1D convolution unit, and the cross-modal attention mechanism to realize deep multi-branch feature fusion, and finally supports multi-task decoding prediction of activity classification and MIC value regression. All encoding processes support batch GPU accelerated inference, fast indexing, and efficient data loading, greatly improving the processing efficiency in large-scale AMP and pathogen interaction prediction scenarios.

[0084] In addition, the sequence encoding module is compatible with other protein or molecular language models (such as ProtTrans, UniRep, etc.), and can be flexibly adapted to various tasks under the premise of unchanged symbol definition and data organization method, providing an extensible and solid foundation for deep learning research of multi-modal biological molecular data.

[0085] 3. Feature extraction and fusion module

[0086] This module takes the output of the sequence encoding module as input, and is responsible for deep feature extraction and cross-modal fusion of high-dimensional embedding features of the antimicrobial peptide and pathogen protein. The core structure of this module is composed of the Mamba module, the feedforward network, and the residual connection and layer normalization unit, combined with the learnable gating mechanism, the Squeeze-and-Excitation (SE) block, and the DASM-1D convolution unit, to realize the collaborative modeling of local structure information and global context dependency based on the multi-layer stacked residual structure. By introducing the cross-modal attention mechanism, an explicit association is established between the antimicrobial peptide and pathogen feature spaces, effectively enhancing the discriminability and task relevance of the multi-modal interaction features, thereby providing stable and high-expression feature support for downstream classification and regression prediction tasks.

[0087] 3.1. Mamba module

[0088] After the sequence encoding is completed, the high-dimensional representation features of the antimicrobial peptide embedding set and the pathogen embedding set are received and mapped into batch tensors and The module adopts a parallel three-branch feature extraction and dynamic fusion structure, combines a Mamba state space mechanism and a cross-modal attention mechanism, and realizes explicit interaction of AMP and pathogen modal information. Different from a single-path sequence modeling method, the parallel three-branch structure can simultaneously capture long-range dependence features, global correlation patterns, and deformable local patterns and channel dynamic relationships in the same stage, thereby significantly enhancing the expressiveness of features and multi-task adaptability.

[0089] To ensure the consistency of processing of the two types of inputs of AMP and pathogen, first, a learnable position encoding is introduced: The position information is injected into the sequence feature to obtain a position-aware representation:

[0090] (7);

[0091] wherein, is the length of the AMP sequence, is the length of the pathogen protein sequence; represents the first position encoding vector intercepted from the position encoding matrix according to the length of the AMP sequence , which is used to be added to ; represents the first position encoding vector intercepted according to the length of the pathogen sequence , which is used to be added to ; represents the AMP sequence feature after adding the position encoding, which retains the original semantic features and superimposes the position information; represents the pathogen sequence feature after adding the position encoding, which also fuses the position information. The formula adds the learnable position encoding to the original embedding , , so that the Mamba module can simultaneously perceive the order information and the local structure information when processing the sequence, laying a unified feature space foundation for cross-modal feature fusion.

[0092] (1) Multi-branch sequence feature extraction

[0093] On the basis of the features after position encoding and , a parallel three-branch structure is adopted: Mamba+GLU branch, Multi-Head Attention branch (MHA, global self-attention), and DASM 1D branch (deformable local + group attention). The parallel three-branch independently runs on each modality, respectively capturing long-range dependence, global semantics, and variable local patterns, thereby realizing complementary modeling of multi-scale features of AMP and pathogen sequences.

[0094] a) Mamba + GLU branch (long-range gating)

[0095] Modeling temporal dependencies through Mamba module, and then channel dynamic filtering through GLU:

[0096] (8);

[0097] where, represents the Mamba state space model operation, responsible for capturing long-range dependency features in the sequence, suitable for processing long sequence information flow and global relationship modeling; is a gated linear unit, which performs channel-level dynamic filtering on the input, selectively enhancing or suppressing the response of different feature channels through the gating mechanism; represents an element-wise multiplication operator, which multiplies the Mamba output and the GLU output element by element, realizing the gated weighted feature fusion. Extract long-range dependency features with Mamba, and use GLU to dynamically control by channel, so that the model can selectively emphasize information related to functional fragments (such as active sites, conserved motifs) while preserving global dependencies, thereby improving the alignment effect with pathogen features.

[0098] b) Multi-head self-attention branch (global dependency)

[0099] (9);

[0100] where, is used to capture global long-range dependencies between any positions in the sequence, which is particularly important for modeling the overall structural features of antimicrobial peptides and the remote active sites of pathogen proteins. This model uses a self-attention structure in this branch, so Q (Query), K (Key), and V (Value) in the formula are respectively mapped from the input features through trainable projection matrices , , to obtain , where is the input feature dimension, is the feature dimension of a single head of attention. The mapped Q, K, and V are used to calculate attention weights, measure matching degrees, and provide weighted information, respectively, to effectively model the global dependency relationships of antimicrobial peptide sequences.

[0101] c) DASM 1D branch (morphable local + group attention)

[0102] Figure 5As shown, the DASM 1D convolution branch first normalizes and linearly transforms the input features to make the numerical scales of each channel feature consistent. Then, the features are mapped to a groupable dimension through channel grouping and projection operations, so as to perform efficient grouped convolution operations. Next, the query, key and value are generated by Q, K and V mapping, and the dynamic offset and modulation weight are introduced for the convolution kernel position by the deformable sampling mechanism, so as to adaptively capture the spatial variation of local features. After the offset sampling, the features are further enhanced by the depth separable convolution to strengthen the local details, while the global dependency information extracted by the multi-head self-attention branch is supplemented. Finally, the output layer fuses the local patterns and global information, and retains the residual connection to ensure stable training and feature consistency.

[0103] To more specifically show the processing process of the DASM 1D branch, the actual calculation path of each sub-module is given as shown in the following table, taking the antibacterial peptide fragment "C-A-T" as an input example. Figure 6 Figure 6 The normalization of the residue input feature, the linear dimension transformation, the group splitting, the dynamic offset sampling, the local convolution calculation, the group attention weighting and the residual connection are shown, which intuitively reflects the effect of the variable local receptive field on feature extraction, and facilitates the understanding of the modeling ability of the DASM module in structural representation enhancement.

[0104] First, layer normalization and automatic dimension alignment are performed: if the channel dimension cannot be divided by the group number , it is mapped to the nearest divisible dimension :

[0105] (10)

[0106] wherein, is a layer normalization operator, which is used to perform mean-variance normalization on the feature vector of each sequence position, to improve the training stability and alleviate the gradient vanishing / explosion; is a linear projection operator, which maps the input feature from the original dimension to a new dimension that can be divided by the group number , which is calculated by the formula , wherein is a floor operator. is the original feature dimension, is the adjusted feature dimension, is the group number of the grouped convolution. is the feature tensor after layer normalization. is the feature tensor after linear projection mapping ​The feature matrix is obtained. Through this operation, the number of channels can be completely divided in subsequent group convolution operations, thereby improving the calculation efficiency and ensuring the rationality of feature grouping.

[0107] Q, K and V linear mapping are obtained:

[0108] (11);

[0109] Generate queries, keys and values from the aligned representation, forming the attention-based representation of this branch.

[0110] Use group convolution to form the deformation offset and the modulation weight :

[0111] (12);

[0112] is the two-dimensional sampling offset (normalized to the coordinate system of ); is the modulation weight of the channel / group. The representation is a group convolution operation with Query as input, and its output contains two channel branches: the sampling offset and the modulation weight . This group convolution can not only reduce the number of parameters, but also learn spatial transformation independently within each group. First, the value of is compressed to the interval , and multiplied by the scaling factor to control the offset range; then, the Sigmoid function activation is applied to , which is mapped to (0, 1) to adjust the response amplitude of each sampling position in the deformable convolution. This mechanism can adaptively generate variable sampling offsets to correct local alignment errors, and use to weight the importance of each group, thereby improving the flexibility and accuracy of feature extraction. According to the offset, a one-dimensional sampling network

[0113] is constructed to perform deformation sampling and depth separable convolution fusion on :

[0114] (13);

[0115] This formula describes the deformation sampling and fusion process based on the offset. First, the offset is used to construct the sampling grid , and the grid sampling function ​​In the input feature map Deformable sampling is performed on the above, so as to realize flexible adjustment of the convolution sampling position. Subsequently, the sampled features are input into the depth separable convolution Local feature fusion is independently performed in each channel, which not only preserves the spatial structure but also improves the calculation efficiency. Finally, the convolution result is multiplied with the modulation weight , so as to control the response amplitude of each sampling position, and obtain the output features after deformation enhancement , so as to improve the robustness of the model to local deformation, insertion and deletion and other structural abnormalities.

[0116] Subsequently, intra-group multi-head attention is performed and residual connection is added:

[0117] (14);

[0118] The above formula describes the process of performing multi-head attention in the group and adding residual connection: wherein respectively represent the Query, Key and Value features after deformation sampling enhancement in the group; represent the multi-head attention mechanism, which is used to calculate the correlation between the features in the group and complete information aggregation; is an output projection operator, which maps the spliced result of the multi-head attention back to the original channel dimension; represents the linear projection of the input antimicrobial peptide feature to make its dimension consistent with the attention output, so as to facilitate residual addition. Finally, the attention result is added to the projected original feature to obtain , which can not only stabilize the gradient and preserve the original position information, but also realize consistent fusion of the deformable local structure and the original semantic features.

[0119] (2) Branch gating weighting and SE channel self-adaptation

[0120] After completing the three-branch sequence feature extraction, three types of complementary features are obtained for each modality (antimicrobial peptide modality and pathogen modality). The output of the antimicrobial peptide modality includes the features extracted by the long-range dependence branch, the global dependence branch and the variable local branch, which are respectively denoted as . The pathogen modality also contains corresponding three-branch outputs, which are respectively denoted as . Among them, the long-range dependence branch mainly captures the long-range dependence relationship in the sequence, the global dependence branch focuses on the global correlation pattern of the sequence as a whole, and the variable local branch adaptively models the local structure of the sequence, so as to provide diversified inputs for subsequent cross-modal feature fusion in different scales and dependence ranges.

[0121] ​To further improve the discriminability and task adaptability of the fused features, the application introduces a branch gating weighting mechanism and an SE (Squeeze-and-Excitation) channel adaptive mechanism in the same modality, dynamically weights and fuses the three branch features, and adaptively adjusts the feature weights in the channel dimension. Unlike the fixed weight or mean fusion method in the prior art, this method can adaptively adjust the contribution of each branch in the fused features according to the differences in the input sample feature distribution, thereby improving the discriminability and task adaptability of the multi-modal features, and is especially suitable for the scene of antimicrobial peptide and pathogen multi-task prediction (activity classification + MIC regression).

[0122] Let the branch gating weight vector be:

[0123] (15);

[0124] The dynamic gating weights of the three branches are respectively: The weight proportion is ensured to be comparable and the sum is 1 through Softmax normalization calculation.

[0125] The branch fusion formula of the antimicrobial peptide modality is:

[0126] (16);

[0127] Wherein, is the unified feature representation after the fusion of the antimicrobial peptide modality, which mixes the input features and exchanges information with the pathogen modality feature. is a channel attention module, which adaptively adjusts the activation strength of different channels by modeling the inter-channel dependency, enhances the key channel features, and suppresses redundant information.

[0128] (3) Feature mixing and cross-modal interaction attention

[0129] After branch gating weighting and SE channel adaptive processing, the antimicrobial peptide modality obtains the fused feature , and the pathogen modality obtains the fused feature .As shown in Figure 7 , in order to realize the explicit interaction of cross-modal features, this module first inputs the antimicrobial peptide modality feature into a fully connected layer (FC projection), maps it to a unified dimension required for cross-modal interaction, and obtains the projected antimicrobial peptide feature representation :

[0130] (17);

[0131] The projected As Query input into the cross-modal interaction attention mechanism; at the same time, the pathogen modality feature As Key and Value input into the mechanism, so as to match with the query vector of antibacterial peptide.

[0132] The cross-modal interaction attention module (CrossAttn) calculates the attention weight according to the similarity of the query and the key, and generates the fused cross-modal feature representation based on the weighted aggregation of the value vector :

[0133] (18);

[0134] is the cross-modal interaction attention module. Query (Q) is the antibacterial peptide modality feature , Key (K) and Value (V) are both pathogen modality features ; is the fused feature representation after cross-modal interaction, which has explicitly injected pathogen feature information into the antibacterial peptide modality. In this process, the antibacterial peptide modality feature can directly aggregate the key information of the pathogen related to it, realizing the explicit interaction of information between modalities. Finally, the fused feature is output to the downstream prediction module, which captures the corresponding relationship between antibacterial peptides and pathogens in the feature space in the MIC regression and activity classification tasks, and improves the modeling ability of the mechanism and structure-function relationship.

[0135] 3.2, feedforward neural network

[0136] The fused feature output by the cross-modal interaction module is input into the feedforward network (FFN) as input to further improve the nonlinear expression ability and discriminability of the feature. The feedforward layer adopts a linear transformation structure for each position, and the feature of each position is independently transformed without introducing position information mixing, thereby maintaining the independence of the sequence structure. The calculation formula of the feedforward layer is:

[0137] (19);

[0138] wherein, is the cross-modal interaction output feature from formula (20). , is the weight matrix, which controls the projection direction and feature compression / expansion of the linear mapping. , is the bias term. is the activation function, which is used to improve the nonlinear expression ability of the network, and its expression is:

[0139] (20);

[0140] The activation function can adaptively adjust the passing ratio of input, has smoother gradient while keeping nonlinearity, and improves the expression ability and training stability of deep network.

[0141] 3.3, residual connection and layer normalization

[0142] In order to improve the training stability of deep structure, this module introduces residual connection (Residual Connection) and layer normalization (LayerNorm) at the output end of the feedforward network, so that the output features retain the feedforward transformation results and the original input information at the same time, thereby relieving gradient disappearance and accelerating convergence. The calculation formula is:

[0143] (21);

[0144] wherein, The normalization operation is performed on the feature dimension to improve the numerical stability. is the output of the feedforward network, is the residual input, The final output of this module is the fusion feature representation, which will be directly input into the multi-task decoder.

[0145] 4, decoding prediction module

[0146] This module is used for multi-task prediction of the context features output by the feature extraction and fusion module, supporting the two tasks of antimicrobial peptide activity classification and MIC value regression. The structure mainly includes sequence feature enhancement, attention pooling, parallel classification branch and regression branch, and the specific process is as follows.

[0147] 4.1, feature enhancement and attention pooling

[0148] The output features of the previous module are taken as input, and are sequentially passed through the feature enhancement and attention pooling unit to generate a global representation suitable for multi-task prediction. Among them, is the batch, is the sequence length, is the feature dimension. This unit includes two sub-modules of sequence feature enhancement and attention pooling.

[0149] (1) sequence feature enhancement

[0150] As shown in Figure 8 , this module receives the fused feature representation After that, it is input into the convolution-position encoding block (CNNPosBlock). This structure consists of multiple layers of one-dimensional convolution and learnable position encoding, which can capture the local patterns of adjacent residues while preserving the global order information of the sequence. Specifically, the input features are first passed through one-dimensional convolution to extract local context, and then fused with learnable position encoding, thereby enhancing the spatial pattern expression and sequence dependence modeling ability at the same time, and finally obtaining enhanced feature representation containing local structure and order information.

[0151] First, the fusion features are input, and batch tensors are received ; then, local patterns are extracted through one-dimensional convolution to model the short-range dependence relationship of adjacent residues; then, learnable position encoding is injected to preserve the global sequence order information; then, the convolution result and the position encoding are fused to integrate the local pattern and the position information; finally, the enhanced sequence features are output to obtain . The above process can be represented as:

[0152] (22);

[0153] The enhanced sequence features not only retain the global context, but also significantly enhance the expression ability of local patterns, which helps to capture the short-range dependence relationship in the characteristics of antimicrobial peptides or pathogens, thereby improving the performance of downstream tasks.

[0154] (2) Attention pooling

[0155] To obtain global vector representation suitable for downstream classification and regression tasks, attention weighted pooling (AttentionPooling) is performed on the enhanced feature sequence :

[0156] (23);

[0157] wherein, the global feature vector for the activity classification task, is the attention weight vector specific to the classification task; the global feature vector for the MIC value regression task, is the attention weight vector specific to the regression task. Here, the global features are calculated separately for the two tasks to ensure task specificity.

[0158] The AttentionPooling calculation form is:

[0159] (24)

[0160] In the formula: is the i-th position in the enhanced sequence feature vector of the i-th position; attention weight of the i-th position; attention weight of the i-th position; total length of the sequence; feature vector of the j-th position in the augmented sequence, with attention weight The calculation is normalized based on the relative contribution of all positions in the sequence; finally, AttentionPooling obtains the aggregated global feature representation by weighting and summing the feature vectors of each position in the sequence according to its corresponding attention weight Thus, the position information that is more important to the task is highlighted and the influence of irrelevant parts is suppressed.

[0161] 4.2, Classification and Regression Branch

[0162] This module adopts a parallel classification and regression output structure based on the task-specific global features obtained through feature augmentation and pooling, realizing multi-task decoding of antimicrobial peptide activity classification and MIC value prediction.

[0163] (1) Classification Branch (Activity Prediction)

[0164] The global feature corresponding to the classification task is input into the fully connected network, and after nonlinear transformation, normalization and regularization, the activity classification probability is finally output:

[0165] (25);

[0166] Equation (25) describes the calculation process of activity prediction (classification branch): where represents the global feature vector corresponding to the classification task; and are learnable weight matrices and bias terms used to map input features to a new feature space; first, calculate through the fully connected layer, and perform nonlinear transformation through the GELU activation function, then normalize the features using LayerNorm to obtain the normalized feature representation ; then, Dropout is used to randomly inactivate to prevent overfitting, and the regularized feature is obtained; finally, the Linear layer is used to project to , which is the classification logits, and through Softmax it is converted into the probability distribution of the activity category, thus realizing the mapping from the task global feature to the classification result.

[0167] ​(2) Regression branch (MIC value prediction)

[0168] The global feature vector corresponding to the regression task is denoted as Input two-layer fully connected network, combined with normalization and Dropout, to get the continuous prediction of MIC value:

[0169] (26);

[0170] Equation (26) describes the calculation process of the regression branch (MIC value prediction): where, represents the global feature vector corresponding to the regression task; and are the learnable weight matrix and bias term, used to map the input features to a new feature space; first, the linear transformation is performed , then the Dropout is randomly inactivated to prevent overfitting, followed by the GELU activation function for nonlinear mapping, and the LayerNorm is used for normalization of the features to get the normalized regression feature vector ; then, the linear transformation is performed to get the prediction result of MIC value, where, and are the learnable weights and bias of the regression output layer, respectively. The final output is a continuous real value, used for the prediction task of minimum inhibitory concentration (MIC).

[0171] 4.3, Output and loss

[0172] The final model output includes:

[0173] Antibacterial peptide activity classification probability, supporting binary classification or multi-classification tasks.

[0174] MIC value regression prediction, supporting continuous phenotype value prediction.

[0175] In the training stage, multi-task loss is used for joint optimization:

[0176] (27);

[0177] Equation (27) gives the joint optimization method of the multi-task loss function: where, represents the loss function of the antibacterial peptide activity classification task, using cross-entropy loss (Cross-Entropy Loss) to measure the difference between the predicted class probability and the true label; The loss function representing the MIC value regression task adopts mean squared error (MSE) to measure the deviation between the predicted value and the true MIC value. is a weight coefficient for balancing the contribution proportion of the classification task and the regression task in the total loss The final loss function The loss of the two tasks is weighted and summed, so that the model can optimize the performance of activity classification and MIC value prediction in the training process.

[0178] The decoding prediction module realizes the full extraction and expression of the context information of the input sequence through multi-level feature enhancement and adaptive pooling. Combined with the parallel classification and regression branch structure, it can simultaneously complete the accurate discrimination of the activity category of the antibacterial peptide and the fine regression prediction of the MIC value in a single framework to meet the needs of multi-task bioinformatics analysis. The proposed module structure has high scalability and universality, can be flexibly adapted to different types of antibacterial peptide-pathogen data, significantly improves the prediction performance and generalization ability of the overall model, and provides strong technical support for efficient screening and mechanism analysis of new antibacterial peptides.

[0179] II. Simulation verification

[0180] The antibacterial peptide data used in the application is derived from the internationally published GRAMPA database. Through strict data cleaning and preprocessing of the original GRAMPA data set, redundant and non-standard sequences are removed, and finally 3345 effective antibacterial peptide samples are retained, fully guaranteeing the scientificity and representativeness of the experimental data.

[0181] The pathogen protein data is obtained from UniProtKB-reviewed (manually annotated protein), UniProtKB-unreviewed, UniParc and NCBIEntrez Protein international mainstream public databases in turn by using an automatic retrieval tool according to the pathogen species involved in the antibacterial peptide data. For each pathogen, a multi-level fallback mechanism is adopted, that is, high-quality UniProtKB-reviewed protein sequences are preferentially obtained, and if not obtained, the next priority data source is switched in turn until the available protein sequence is obtained. The finally constructed pathogen data set covers 163 kinds of pathogens, with a maximum of 30 representative protein sequences for each pathogen and a minimum of 1, fully reflecting the diversity and sequence representativeness of the pathogens.

[0182] After standardization, all data are divided into training set, validation set and test set according to the ratio of 8:1:1, ensuring the fairness and scientificity of model training and evaluation. In terms of model performance evaluation, for the classification task of antimicrobial peptide activity, mainstream classification indexes such as accuracy (ACC), F1 score and area under curve (AUC) are used for quantitative evaluation; for the MIC value regression prediction task, indexes such as root mean square error (RMSE) and mean absolute error (MAE) are used to measure the regression accuracy of the model. The above evaluation methods are based on the prediction results of the validation set and the test set, which fully reflect the classification and regression generalization ability and practical application value of the model.

[0183] After calculation, the results of the model in the classification task and the regression task of the application are shown in Table 1.

[0184]

[0185] The preferred embodiments of the embodiments of the application are described above with reference to the accompanying drawings, and the scope of the right of the embodiments of the application is not limited by this. Any modification, equivalent replacement and improvement made by those skilled in the art without departing from the scope and essence of the embodiments of the application shall be within the scope of the right of the embodiments of the application.

Claims

1. A framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning, characterized in that: include: The sequence encoding module uses a pre-trained protein language model to embed antimicrobial peptide sequences and pathogen protein sequences, respectively, and introduces learnable position encoding to inject position information into the antimicrobial peptide embedding sequence and the pathogen protein embedding sequence; wherein, the antimicrobial peptide sequence has an active MIC tag; The feature extraction and fusion module includes a multi-layered stacked residual structure, which includes: The Mamba module includes a parallel multi-head self-attention branch, a Mamba+GLU branch and a DASM 1D convolution branch for simultaneously capturing global long-range dependency features, gated weighted fusion features and local and global fusion features from antimicrobial peptide embedding sequences and pathogen protein embedding sequences; a learnable gating and SE unit for dynamically weighted fusion of global long-range dependency features, gated weighted fusion features and local and global fusion features and adaptively adjusting channel feature weights; and a feature mixing unit for explicitly injecting the dynamically weighted fusion features of the pathogen modality into the dynamically weighted fusion features of the antimicrobial peptide modality. Feedforward neural networks, through position-by-position linear transformation of the structure, enable the features input from the Mamba module to have smooth gradients while possessing non-linear expressive power. Residual connections and layer normalization units are used to simultaneously preserve the transformation results and original input information of the feedforward neural network; The decoding and prediction module is used to decode the features input from the feature extraction and fusion module into global feature vectors for the activity classification task and the MIC value regression task. It predicts the probability distribution of the activity category and continuously predicts the MIC value through parallel classification and regression branches, respectively.

2. The framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning according to claim 1, characterized in that: The active MIC tag is generated by logarithmic transformation of experimentally determined minimum inhibitory concentration (MIC) data.

3. The framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning according to claim 2, characterized in that: The antimicrobial peptide sequence and pathogen protein sequence are obtained by preprocessing the original antimicrobial peptide sequence and pathogen protein sequence in FASTA format, respectively; the preprocessing includes removing irrelevant characters, normalizing amino acid residue representation, and limiting the maximum sequence length; wherein: normalized amino acid residue representation is performed by formula (1), which is: (1)； In the formula: Indicates the first The amino acid sequence of an antimicrobial peptide or pathogen. For the first One amino acid residue, A set of amino acid residues supported by a protein language model. This is the actual length of the sequence, and , This indicates the maximum sequence length.

4. The framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning according to claim 3, characterized in that: The embedding encoding is represented as follows: (2)； In the formula: This is the embedding representation matrix of the antimicrobial peptide sequence, with the number of rows equal to the actual length of the sequence. Each row corresponds to a feature vector of one amino acid residue. This represents the output embedding dimension of the protein language model.

5. The framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning according to claim 4, characterized in that: The specific process of the embedding encoding includes the following steps: S51. Index the sequence of the input protein language model, mapping each residue character in the sequence to an integer number to generate an index sequence; wherein: the sequence is an antimicrobial peptide sequence or a pathogen protein sequence; S52. Add a start marker BOS and an end marker EOS to the beginning and end of the sequence to indicate the sequence boundaries; S53. The index sequence is mapped to dense word vectors using a learnable embedding matrix, and learnable positional encoding is superimposed to provide residue position information. S54. Input the dense word vectors with residue position information into the multi-layer Transformer encoder of the protein language model for encoding. S55. After encoding, remove the vectors corresponding to the first and last BOS / EOS markers, retaining only the representations of the actual residues, with the output shape as... The residue-level embedding matrix is ​​generated and bound to the unique ID of the sample to form antimicrobial peptide embedding features or pathogen embedding features.

6. The framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning according to claim 5, characterized in that: The Mamba module includes parallel branching units, learnable gating and SE units, and feature fusion units. The parallel branching units include parallel multi-head self-attention mechanism branches, Mamba+GLU branches, and DASM 1D convolutional branches; wherein: The Mamba+GLU branch uses the Mamba state-space model to extract long-range dependency features, while the functional fragments selected by the GLU gating mechanism are obtained by element-wise multiplication to obtain gated weighted fusion features. Multi-head self-attention branch: Extracts global long-range dependency features through a multi-head self-attention mechanism; The DASM 1D convolution branch extracts local contextual semantic information through dynamic sampling and channel convolution, and supplements it with global long-range dependency features to obtain local and global fused features; Learnable gating and SE units are introduced, and a branch-gating weighted mechanism and a channel adaptive mechanism SE are introduced to dynamically weight and fuse synchronously captured local and global fusion features, gating weighted fusion features and global long-range dependency features. The channel feature weights are adaptively adjusted in the channel dimension to obtain dynamic weighted fusion features of antimicrobial peptide modality and pathogen modality. The feature fusion unit maps the dynamically weighted fusion features of the antimicrobial peptide modality to the cross-modal interaction dimension through a fully connected layer to obtain the antimicrobial peptide projection features. The antimicrobial peptide projection features are used as input to the query vector to introduce a cross-modal attention mechanism. At the same time, the dynamically weighted fusion features of the pathogen modality are used as input to the key vector and value vector to introduce a cross-modal attention mechanism to match the query vector. The cross-modal attention mechanism calculates attention weights based on the similarity between the query vector and the key vector, and aggregates the value vector based on the calculated attention weights to generate cross-modal interaction fusion features of the antimicrobial peptide modality and the pathogen modality.

7. The framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning according to claim 6, characterized in that: The cross-modal interactive fusion feature is used as the input of the feedforward neural network. The feedforward neural network adopts a position-by-position linear transformation structure, which independently transforms the features at each position without mixing information between positions. This allows the cross-modal interactive fusion feature to have a smooth gradient while having non-linear expressive ability, so as to generate cross-modal interactive fusion enhanced features.

8. The framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning according to claim 7, characterized in that: The cross-modal interactive fusion enhancement feature is used as the input to the residual connection and layer normalization unit. Through residual connection and layer normalization, the output modal interactive fusion enhancement feature retains both the feedforward transformation result and the original input information, thereby alleviating gradient vanishing, accelerating convergence, and deepening the network modeling depth.

9. The framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning according to claim 8, characterized in that: The global feature vector of the active classification task is input into the fully connected layer of the classification branch. After nonlinear transformation, normalization and regularization, it is projected into classification features. Then, the classification features are transformed into the probability distribution of active categories through the activation function to achieve active classification prediction.

10. The framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning according to claim 9, characterized in that: The global feature vector of the MIC value regression task is input into the two fully connected layers of the regression branch. By combining normalization and regularization, continuous prediction of the MIC value is obtained.

Images