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

Abstract

The invention belongs to the technical field of antibacterial peptide activity identification and evaluation, and relates to an antibacterial peptide activity and MIC value combined prediction framework based on cross-modal deep learning, and the framework uses a protein language model ESM2 to respectively carry out token-level semantic embedding coding on an antibacterial peptide sequence and a pathogen protein sequence; performing cross-modal feature extraction and fusion through a multi-branch structure comprising a Mama module, a multi-head self-attention mechanism and DASM 1D convolution; a multi-task decoding structure is adopted to realize antibacterial peptide activity classification and MIC value regression prediction at the same time; according to the method, functional characteristics in the sequence can be effectively mined, the accuracy and generalization ability of antibacterial peptide activity and MIC value prediction are remarkably improved, and a reliable calculation tool is provided for high-throughput screening and rational design of the antibacterial peptide.

Description

Technical Field

[0001] This invention belongs to the field of antimicrobial peptide activity identification and evaluation technology, specifically involving a framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning. Background Technology

[0002] Antibiotic resistance has become a major challenge in global public health. Driven by the overuse of broad-spectrum antibiotics, bacterial resistance has risen significantly, leading to a continuous decline in the effectiveness of traditional antibiotic treatments. Under these severe circumstances, the clinical management of common infectious diseases faces challenges such as prolonged treatment cycles and soaring medical expenditures. Furthermore, the cross-transmission of drug-resistant strains within healthcare institutions frequently triggers nosocomial infection outbreaks, seriously threatening public health security systems. Therefore, the development of antimicrobial drugs with novel mechanisms of action has become one of the most important scientific issues in the global biomedical field.

[0003] Natural antimicrobial peptides are considered the most promising next-generation antimicrobial drugs due to their unique membrane cleavage mechanism and low likelihood of inducing drug resistance. However, traditional research and development models are limited by bottlenecks such as long database construction cycles and difficulties in structure-activity relationship analysis, resulting in high costs for new drug development. The rapid development of artificial intelligence technology has brought a revolutionary change to this field—deep learning-based generative models can efficiently mine antimicrobial peptide sequence features, intelligently predict molecular activity and toxicity data, and shorten the candidate molecule discovery cycle from years to weeks using virtual screening technology, while significantly reducing experimental verification costs.

[0004] In this process, the minimum inhibitory concentration (MIC) plays a crucial role as an important indicator for measuring the efficacy of antimicrobial drugs. The MIC reflects the minimum drug concentration required to inhibit bacterial growth and is a direct quantitative standard for evaluating the inhibitory effect of antimicrobial peptides on bacteria. By predicting the MIC values ​​of antimicrobial peptides, deep learning technology not only accelerates the research and development of novel antimicrobial peptides but also provides more accurate prediction and screening methods, helping scientists identify potential candidate molecules in the early stages. This technological innovation not only accelerates the development of novel antimicrobial peptides but also provides new solutions for combating drug-resistant strains. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a framework for joint prediction of antimicrobial peptide activity and MIC value based on cross-modal deep learning.

[0006] To achieve the objectives of this invention, the following technical solutions are adopted.

[0007] A framework for jointly predicting antimicrobial peptide activity and MIC values ​​based on cross-modal deep learning includes: 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.

[0008] Furthermore, the active MIC tag is generated by logarithmic transformation of the experimentally determined minimum inhibitory concentration (MIC) data.

[0009] 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: (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.

[0010] Furthermore, the embedding encoding is represented as: (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.

[0011] Furthermore, 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.

[0012] Furthermore, the Mamba module includes parallel branching units, learnable gating and SE units, and feature fusion units that introduce cross-modal interactive attention mechanisms. 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 using the GLU gating mechanism to filter functional fragments and perform element-wise multiplication operations 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.

[0013] Furthermore, 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 modal interactive fusion feature to have a smooth gradient while possessing nonlinear expressive power, so as to generate cross-modal interactive fusion enhanced features.

[0014] Furthermore, 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 cross-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.

[0015] Furthermore, the global feature vector of the active classification task is input into the fully connected layer of the classification branch, and 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.

[0016] Furthermore, the global feature vector of the MIC value regression task is input into the two fully connected layers of the regression branch, and combined with normalization and regularization, continuous prediction of the MIC value is obtained.

[0017] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: The modular structure proposed in this invention has high scalability and versatility, and can be flexibly adapted to different types of antimicrobial peptide-pathogen data, significantly improving the overall model's predictive performance and generalization ability, and providing strong technical support for the efficient screening and mechanism analysis of novel antimicrobial peptides. Attached Figure Description

[0018] Figure 1A framework diagram of the prediction model; Figure 2 This is a diagram of the framework of the Mamba module. Figure 3 This is a diagram illustrating the processing steps of the protein language model. Figure 4 This is a diagram illustrating the processing of the input sequence. Figure 5 A flowchart of the DASM 1D convolution branch; Figure 6 This is a flowchart illustrating the processing steps of a DASM 1D branch. Figure 7 A framework diagram for the feature fusion and cross-modal multi-head attention module; Figure 8 This is a diagram illustrating the processing of convolutional-positional coding blocks. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0020] I. Model Introduction The antimicrobial peptide and pathogen activity prediction framework proposed in this invention consists of three core modules: a sequence encoding module, a feature extraction and fusion module, and a decoding and prediction module, such as... Figure 1 As shown, the modules work together to automatically extract effective representations from multimodal sequence information and jointly predict the activity and MIC value of antimicrobial peptides.

[0021] The sequence encoding module takes antimicrobial peptide and pathogen protein sequences as input, both expressed in FASTA format. Since the raw MIC values ​​span multiple orders of magnitude and are highly skewed, direct use could lead to high-value samples excessively dominating model learning during training. Therefore, a logarithmic transformation of the MIC is performed before modeling to compress the numerical range, reduce scale differences, and improve distribution characteristics. Subsequently, a pre-trained protein language model, ESM2, is used to perform high-dimensional embeddings of the antimicrobial peptide and pathogen sequences, fully extracting the contextual semantics and structural dependencies of the sequences. Activity labels are automatically generated based on the processed MIC data, thus providing reliable supervision signals for subsequent multi-task joint modeling of classification and regression.

[0022] In the feature extraction and fusion module, the embedded features of antimicrobial peptides and pathogens are jointly input into the Mamba module within the feature extraction and fusion module, such as... Figure 1 As shown, this enables multi-level feature extraction, including local detail capture, global semantic modeling, and long-distance dependency awareness. The specific structure of the Mamba module is as follows: Figure 2As shown, the internal architecture employs a parallel three-branch design: The first branch is a DASM 1D convolution branch, which extracts local contextual patterns through dynamic sampling and channel convolution, and supplements the global cross-positional dependencies of the multi-head self-attention branch to achieve the fusion of local and global features. The second branch is a Mamba+GLU branch, which directly receives the embedding features of antimicrobial peptides and pathogens, regulates information flow through the GLU gating mechanism, and combines a state-space modeling structure to capture long-range dependencies and global sequence patterns, demonstrating significant advantages in long-distance information transmission and key state preservation. The third branch is a multi-head self-attention branch, which also directly receives the embedding features of antimicrobial peptides and pathogens, establishes cross-positional dependencies globally, and strengthens the aggregation of long-distance information and the ability to integrate semantics across fragments.

[0023] The multi-level features output from the three branches are dynamically weighted using learnable gating weights and enhanced with an SE module to achieve adaptive channel enhancement, ensuring that the fused representation achieves an optimal balance between global, local, and long-range information modeling. The bimodal fused features, integrated by the hybrid layer, are fed into a cross-modal interactive attention mechanism to establish explicit interactive encoding between antimicrobial peptides and pathogen sequences at the token level, thereby effectively characterizing their potential mechanisms of action at the structural and functional levels.

[0024] The feature extraction and fusion module sequentially feeds cross-modal fused features into a multi-layered stacked residual structure. Each layer consists of a Mamba module and a position-wise feed-forward network. The Mamba module captures long-range dependencies and global patterns, while the feed-forward network enhances the non-linear expressive power of the features. Both are supplemented with layer normalization and residual connections to stabilize gradient propagation and increase the network modeling depth. After passing through the multi-layered stacked residual structure, the fused features enter the feature enhancement and pooling stage to compress the sequence dimension and retain key information. They are then fed into the multi-task prediction output head: the classification branch uses a softmax activation function to predict antimicrobial peptide activity, while the regression branch uses a linear output unit to accurately predict the MIC value, thus achieving both high accuracy and efficiency in the activity screening and functional quantitative evaluation of candidate antimicrobial peptides.

[0025] 1. Representation of input data The cross-modal deep learning-based framework for jointly predicting antimicrobial peptide activity and MIC values ​​uses antimicrobial peptide (AMP) sequences and pathogen protein sequences as inputs, representing candidate antimicrobial peptide molecules and target pathogens, respectively. First, the original FASTA-formatted antimicrobial peptide and pathogen protein sequences are standardized and preprocessed. Then, a pre-trained protein language model, ESM2, is used to extract high-dimensional token-level embedding features for both types of molecules. These high-dimensional token-level embedding features not only include contextual dependencies and global structural relationships between amino acid residues in the sequence but also reflect potential functional regions and molecular biological semantics, providing a rich expressive foundation for subsequent end-to-end feature modeling and molecular-level interaction analysis.

[0026] In terms of tagging system construction, based on experimentally determined minimum inhibitory concentration (MIC) data and according to threshold classification rules, each antimicrobial peptide is assigned a binary activity tag (activity = 1, inactivity = 0) to support the model's multi-task output, including activity classification and MIC value regression prediction. Finally, the high-dimensional token-level embedding features of antimicrobial peptides and pathogen proteins are input into the multi-branch feature extraction module and the cross-modal interaction fusion module, respectively, to achieve multi-scale expression of molecular sequences, information complementarity, and token-level fine alignment. This lays a solid data foundation and information support for the identification of drug peptide activity and accurate prediction of MIC values.

[0027] 2. Sequence Encoding Module This module, one of the three core modules of this invention, is located at the beginning of the entire model flow. Its aim is to achieve unified high-dimensional embedding expression of antimicrobial peptides (AMP) and pathogen protein sequences, providing consistent, refined, and cross-modal associative input features for subsequent feature extraction and fusion modules and decoding and prediction modules. This module introduces the advanced pre-trained protein language model ESM2 to perform end-to-end token-level semantic embedding processing on molecular sequences. Unlike existing technologies that only use global mean vectors, this invention can retain the local features of each amino acid residue while simultaneously capturing the overall structural pattern and contextual dependencies of the sequence, thereby significantly improving the expressive power and deep modeling effect of the molecular sequence. The obtained high-dimensional embedding features will work collaboratively with the Mamba module, DASM-1D convolutional units, and cross-modal attention mechanisms in subsequent processing stages to achieve cross-modal deep feature fusion and multi-task functional prediction. The specific implementation of antimicrobial peptide sequence encoding and pathogen sequence encoding will be described below: (1) Antimicrobial peptide sequence encoding This 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 first... The normalized sequence of the antimicrobial peptide is as follows: (1); in, Indicates the first The amino acid sequence of an antimicrobial peptide or pathogen. For the first One amino acid residue, It is a set of amino acid residues supported by ESM2 (usually containing 20 standard amino acids and special symbols). This is the actual length of the sequence, and .

[0028] This formula converts the original character-type amino acid sequence into a mathematically processable symbolic representation, providing standardized input for subsequent feature extraction.

[0029] After obtaining the normalized sequence, in order to capture the local and global semantic dependencies in the antimicrobial peptide sequence, the pre-trained protein language model ESM2 was used to analyze the semantic dependencies. By performing high-dimensional token-level embedding mapping, we obtain: (2); in, 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. The output embedding dimension of the ESM2 model is 1280 (in this invention). This process converts symbolic amino acid sequences into dense numerical vector representations, enabling deep learning models to directly perform computations and pattern capture.

[0030] To facilitate a clear and intuitive understanding of the input-output relationship of ESM2 in this invention, Figure 3 The data flow process of an antimicrobial peptide sequence after entering the ESM2 model is presented. First, a preprocessed amino acid sequence is input, and each residue character is mapped to an integer number after indexing. Then, start markers (BOS) and end markers (EOS) are added to the beginning and end of the sequence to indicate sequence boundaries. Next, a learnable embedding matrix is ​​used to map the indexed sequence into dense word vectors, and learnable positional encodings are superimposed to provide residue position information for the model. Subsequently, the sequence is input into the ESM2 multilayer Transformer encoder for global dependency modeling and contextual feature extraction. After encoding, the vectors corresponding to the start and end BOS / EOS markers are removed, retaining only the representations of the actual residues. The final output shape is... The residue-level embedding matrix is ​​generated and bound to the unique ID of the sample, serving as the input for subsequent feature fusion and prediction modules.

[0031] To more clearly illustrate the entire processing flow, the various processing stages of the input sequence "CAT" in the ESM2 model are visualized step by step, such as... Figure 4 As shown: To support batch storage and fast indexing, all embedding results of antimicrobial peptide sequences are grouped together with their unique identifiers (id): (3); in, This is a collection of all antimicrobial peptide sequence embedding results. This represents the total number of antimicrobial peptides. This set can be serialized into JSON or binary files, facilitating efficient loading and retrieval during subsequent cross-modal fusion with pathogen protein embedding features. It serves as information storage and index mapping in the data flow, enabling subsequent multi-task predictions (activity classification + MIC value regression) to be calculated based on a unified antimicrobial peptide feature representation.

[0032] (2) Pathogen sequence encoding For pathogen protein sequences, the raw data is in JSON format, grouped by TaxID. For each TaxID, the most frequently selected... A representative protein sequence was obtained, and all sequences were uniformly truncated to their maximum length. Let TaxID be... At that time, its proteome was Similarly, ESM2 is used for high-dimensional token-level embedding representation, specifically as follows: (4); in, Indicates TaxID is The A protein sequence, encoded using the ESM2 model, is represented by a token-level embedding matrix, with the number of rows equal to the sequence length. Each row is a high-dimensional feature vector of an amino acid residue. For the embedding dimension, this formula maps symbolic amino acid sequences to a dense vector space that can directly participate in deep learning computations, thus laying the foundation for cross-modal feature fusion.

[0033] Within TaxID, all sequence embedding results are aggregated into a set: (5); in, TaxID is This is a token-level embedding set of all representative protein sequences, which can be used for batch processing and parallel acceleration. Multiple protein sequence embeddings of the same pathogen are organized and centrally managed, facilitating subsequent feature aggregation and modeling at the pathogen level.

[0034] Furthermore, the embedding sets of all pathogens are uniformly organized using TaxID as the key as follows: (6); in, A global pathogen embedding index table, where each element is a tuple. It supports serialization to JSON or binary files, facilitating efficient reading and distributed loading of models during training and inference. This formula establishes a mapping from the pathogen's unique identifier to its feature set, enabling subsequent cross-modal feature matching between AMP and pathogens, activity classification, and MIC value regression prediction to be efficiently invoked directly based on this index.

[0035] This sequence encoding module implements antimicrobial peptides. With pathogen proteins This invention employs end-to-end high-dimensional token-level unified modeling. Unlike existing technologies that only compress sequences into a global mean vector, this invention preserves both local segment features and global dependencies of the molecular sequence during the encoding process. It also automatically learns high-dimensional representations of functional regions through the ESM2 model, thereby ensuring the consistency and alignment of multimodal features in the subsequent feature fusion stage.

[0036] High-dimensional embedding features of the encoded output and The direct input feature extraction and fusion module works in conjunction with the Mamba module, DASM-1D convolutional units, and cross-modal attention mechanisms to achieve deep multi-branch feature fusion, ultimately supporting multi-task decoding and prediction for activity classification and MIC value regression. All encoding processes support batch GPU-accelerated inference, fast indexing, and efficient data loading, significantly improving processing efficiency in large-scale AMP and pathogen interaction prediction scenarios.

[0037] Furthermore, this sequence encoding module is compatible with other protein or molecular language models (such as ProtTrans, UniRep, etc.). Under the premise that the symbol definition and data organization method remain unchanged, it can be flexibly adapted to a variety of tasks, providing a scalable and solid foundation for deep learning research on multimodal biomolecular data.

[0038] 3. Feature Extraction and Fusion Module This module takes the output of the sequence encoding module as input and is responsible for deep feature extraction and cross-modal fusion of the high-dimensional embedding features of antimicrobial peptides and pathogen proteins. The core structure of this module consists of a Mamba module, a feedforward network, residual connections, and layer normalization units. Combined with a learnable gating mechanism, a Squeeze-and-Excitation (SE) block, and DASM-1D convolutional units, it achieves collaborative modeling of local structural information and global contextual dependencies based on a multi-layered stacked residual structure. By introducing a cross-modal attention mechanism, an explicit association is established between the antimicrobial peptide and pathogen feature spaces, effectively enhancing the discriminative power and task relevance of multimodal interaction features, thereby providing stable and highly expressive feature support for downstream classification and regression prediction tasks.

[0039] 3.1 Mamba module After sequence encoding is complete, the antimicrobial peptide embedding set is received. With pathogen embedding set High-dimensional representation features are then mapped to batch tensors. and This module employs a parallel three-branch feature extraction and dynamic fusion structure, combining the Mamba state-space mechanism and cross-modal attention mechanism to achieve explicit interaction between AMP and pathogen modal information. Unlike single-path sequence modeling methods, the parallel three-branch structure of this invention can simultaneously capture long-range dependent features, globally relevant patterns, and deformable local patterns and channel dynamic relationships at the same stage, thereby significantly enhancing the expressiveness and multi-task adaptability of features.

[0040] To ensure consistent processing of AMP and pathogen inputs, learnable positional encoding is first introduced: By injecting location information into sequence features, a location-aware representation is obtained: (7); in, The length of the AMP sequence. The length of the pathogen protein sequence; Indicates the length of the AMP sequence from the position encoding matrix. Cut off the front Each positional encoding vector is used to... Add; Indicated by pathogen sequence length Cut off the front Each positional encoding vector is used to... Add; This indicates that the AMP sequence features, after adding positional encoding, retain the original semantic features and overlay positional information; This indicates that the pathogen sequence features, after incorporating positional encoding, also integrate positional information. The formula uses learnable positional encoding... With the original embedding , The addition of these features enables the Mamba module to simultaneously perceive sequence information and local structural information when processing sequences, laying a unified feature space foundation for cross-modal feature fusion.

[0041] (1) Feature extraction of multi-branch sequences Features after position encoding and Based on this, a parallel three-branch structure is adopted: the Mamba+GLU branch, the Multi-Head Attention branch (MHA, global self-attention), and the DASM 1D branch (deformable local + group attention). The three parallel branches operate independently on their respective modalities, capturing long-range dependencies, global semantics, and variable local patterns respectively, thereby achieving complementary modeling of multi-scale features of AMP and pathogen sequences.

[0042] a) Mamba+GLU branch (long-range gating) Timing dependencies are modeled using the Mamba module, and then channel dynamic filtering is performed using GLU: (8); in, This represents the Mamba state-space model operation, which is responsible for capturing long-range dependency features in the sequence and is suitable for handling long sequence information flow and global relationship modeling. As a gated linear unit, it performs channel-level dynamic filtering of the input, selectively enhancing or suppressing the response of different characteristic channels through a gating mechanism; This represents the element-wise multiplication operator, which multiplies the Mamba output and the GLU output element-wise to achieve gated weighted feature fusion. Mamba is used to extract long-range dependency features, and GLU is used for dynamic control on a channel-by-channel basis. This allows the model to selectively emphasize information from functionally relevant segments (such as action sites and conserved motifs) while preserving global dependencies, thereby improving the alignment with pathogen features.

[0043] b) Multi-head self-attention branch (global dependency) (9); in, This method is used to capture global long-range dependencies between arbitrary positions in a sequence, which is particularly important for modeling the overall structural features of antimicrobial peptides and the long-range action sites of pathogen proteins. This model employs a self-attention structure in this branch; therefore, Q (Query), K (Key), and V (Value) in the formula are derived from the input features. Through trainable projection matrix , , The mapping yields, where, For the input feature dimension, The feature dimensions for single-head attention are defined. The mapped Q, K, and V are used to calculate attention weights, measure the degree of matching, and provide weighted information, respectively, to achieve effective modeling of the global dependency relationship of antimicrobial peptide sequences.

[0044] c) DASM 1D branch (deformable local + group attention) like Figure 5 As shown, the DASM 1D convolutional branch first normalizes and linearly transforms the input features to ensure consistent numerical scale across all channels. Then, through channel grouping and projection operations, the features are mapped to a divisible grouping dimension for efficient grouped convolution operations. Next, query, key, and value are generated using Q, K, and V mappings. A deformable sampling mechanism introduces dynamic offsets and modulation weights to the convolutional kernel positions, adaptively capturing spatial variations in local features. The offset-sampled features are then enhanced with depthwise separable convolutions to improve local details while supplementing global dependency information extracted by the multi-head self-attention branch. Finally, the output layer fuses local patterns with global information and preserves residual connections to ensure stable training and feature consistency.

[0045] To illustrate the processing flow of the DASM 1D branch more concretely, using the antimicrobial peptide fragment "CAT" as an input example, the actual computational paths of each submodule are given, such as... Figure 6 As shown. Figure 6 The process of normalizing residue input features, linear dimension transformation, group splitting, dynamic offset sampling, local convolution calculation, intra-group attention weighting and residual connection is demonstrated, which intuitively reflects the role of variable local receptive fields in feature extraction and facilitates understanding of the DASM module's modeling capabilities in structural representation enhancement.

[0046] First, perform layer normalization and automatic dimension alignment: if the channel dimension Number of groups that cannot be grouped If divisible, it is mapped to the nearest divisible dimension. : (10); in, This is a layer normalization operator used to normalize the mean-variance of the feature vector at each sequence position, improving training stability and mitigating gradient vanishing / exploding. It is a linear projection operator that projects input features from their original dimensions. Mapping to the number of groups A New Dimension of Divisibility This dimension is derived from the formula. Calculation, where This is the floor operator. For the original feature dimension, For the adjusted feature dimensions, This represents the number of groups in the grouped convolution. This is the feature tensor after layer normalization. For linear projection mapping The resulting feature matrix. This operation ensures that the number of channels is completely divisible in subsequent group convolution operations, thereby improving computational efficiency and ensuring the rationality of feature grouping.

[0047] The linear mapping of Q, K, and V yields: (11); Generate queries, keys, and values ​​from the aligned representation to form the attention-based representation for this branch.

[0048] Using grouped convolution in upper deformation offset With modulation weights : (12); Two-dimensional sampling offset (normalized to) (coordinate system) The modulation weights for channels / groups. This represents a group convolution operation with a query as input, whose output includes two channel branches: sampling offset. and modulation weights This grouped convolution reduces the number of parameters and allows each group to learn spatial transformations independently. Firstly, through... Will The value is compressed to Range, and multiply by the scaling factor Control the offset range; subsequently, for Applying a sigmoid activation function, mapping it to the range (0,1), adjusts the response amplitude at each sampling location in the deformable convolution. This mechanism adaptively generates variable sampling offsets, corrects local alignment errors, and utilizes... The importance of each group is weighted to improve the flexibility and accuracy of feature extraction.

[0049] Construct a one-dimensional sampling network based on the offset. ,right Perform deformation sampling and depthwise separable convolution fusion: (13); This formula describes the offset-based deformable sampling and fusion process. First, the offset is used... Constructing a sampling grid and through the grid sampling function Input feature map Deformation sampling is performed on the top layer, allowing for flexible adjustment of the convolution sampling position. Subsequently, the sampled features are input into a depthwise separable convolution. Local feature fusion is performed independently within each channel, preserving spatial structure while improving computational efficiency. Finally, the convolution results are combined with modulation weights. Perform element-wise multiplication ( This allows for the control of the response amplitude at each sampling location, resulting in enhanced output characteristics. This improves the model's robustness to structural anomalies such as local deformation, insertion, and missing parts.

[0050] Then, multi-head attention within the group is performed and residuals are added: (14); The above formula describes the process of performing multi-head attention and residual connections within a group: where These represent the Query, Key, and Value features within the group, respectively, enhanced by deformable sampling. This refers to a multi-head attention mechanism, used to compute the correlation between features within a group and to aggregate information. It is the output projection operator, which maps the splicing result of multi-head attention back to the original channel dimension; Indicates the characteristics of the input antimicrobial peptide A linear projection is performed to match the dimensions of the attention output, facilitating the summation of residuals. Finally, the attention result is added to the projected original features to obtain the final result. This operation can stabilize the gradient and preserve the original position information, and achieve consistent fusion of deformable local structures and original semantic features.

[0051] (2) Branch gated weighting and SE channel adaptive After completing the three-branch sequence feature extraction, three types of complementary features were obtained for each modality (antimicrobial peptide modality and pathogen modality). The output of the antimicrobial peptide modality includes the features extracted from the long-range dependent branch, the global dependent branch, and the variable local branch, denoted as follows: The pathogen modality also contains a corresponding three-branch output, denoted as follows: Among them, the long-range dependency branch mainly captures the long-range dependencies in the sequence, the global dependency branch focuses on the overall features of the sequence and the global association pattern, while the variable local branch adaptively models the local structure of the sequence, thus providing diverse inputs for subsequent cross-modal feature fusion at different scales and dependency ranges.

[0052] To further enhance the discriminative power and task adaptability of fused features, this invention introduces a branch-gated weighting mechanism and a Squeeze-and-Excitation (SE) channel adaptive mechanism within the same modality. This dynamically weights and fuses the three-branch features, adaptively adjusting the feature weights along the channel dimension. Unlike existing methods that use fixed weights or mean fusion, this method adaptively adjusts the contribution of each branch to the fused features based on the differences in the feature distribution of the input samples, thereby improving the discriminative power and task adaptability of multimodal features. It is particularly suitable for scenarios involving multi-task prediction (activity classification + MIC regression) of antimicrobial peptides and pathogens.

[0053] Let the branch gating weight vector be: (15); These are dynamic gating weights for the three branches, reflecting the relative importance of the current input on the three feature types; the weights The calculation is performed using Softmax normalization to ensure that the weight ratios are comparable and that the sum is 1.

[0054] The branching fusion formula for the antimicrobial peptide modality is: (16); in, To achieve a unified feature representation after antimicrobial peptide modality fusion, the input features are mixed with a cross-modal interactive attention module for alignment and information exchange with pathogen modality features. The channel attention module models the dependencies between channels and adaptively adjusts the activation intensity of different channels to enhance key channel features and suppress redundant information.

[0055] (3) Feature mixing and cross-modal interactive attention After branch-gated weighting and SE channel adaptive processing, the antimicrobial peptide modality was fused. Pathogen modalities are fused to obtain features .like Figure 7 As shown, in order to achieve explicit interaction of cross-modal features, this module first integrates the antimicrobial peptide modality feature. Input a fully connected layer (FC projection) and map it to the unified dimension required for cross-modal interaction. The projected antimicrobial peptide features were obtained. : (17); The projection As a cross-modal interaction attention mechanism for query input; simultaneously, pathogen modality features This mechanism takes the input as both a key and a value, and then matches it with the antimicrobial peptide query vector.

[0056] The Cross-Modal Interaction Attention Module (CrossAttn) calculates attention weights based on the similarity between the query and the key, and then aggregates the value vectors based on these weights to generate the fused cross-modal feature representation. : (18); This is a cross-modal interactive attention module. The query (Q) represents the antimicrobial peptide modality feature. Both the key (K) and the value (V) are pathogen modal characteristics. ; To represent the fused features after cross-modal interaction, pathogen characteristic information has been explicitly injected into the antimicrobial peptide modality. In this process, the antimicrobial peptide modality features can directly aggregate key pathogen information related to them, achieving explicit information interaction between modalities. Ultimately, the fused features... The output to the downstream prediction module captures the correspondence between antimicrobial peptides and pathogens in the feature space in MIC regression and activity classification tasks, thereby improving the ability to model the mechanism of action and structure-function relationship.

[0057] 3.2 Feedforward Neural Network The fusion features output by the cross-modal interaction module As input, the features are fed into a feedforward network (FFN) to further enhance the nonlinear expressiveness and discriminative power of the features. The feedforward layer employs a position-by-position linear transformation structure, where features at each position are transformed independently without introducing information mixing between positions, thus maintaining the independence of the sequence structure. The calculation formula for the feedforward layer is as follows: (19); in, Cross-modal interaction output features derived from formula (20). , The weight matrix controls the projection direction of the linear mapping and the feature compression / expansion. , This is a bias term. The activation function is used to enhance the nonlinear expressive power of the network, and its expression is: (20); This activation function can adaptively adjust the input throughput ratio, maintaining nonlinearity while providing a smoother gradient, thus improving the expressive power and training stability of deep networks.

[0058] 3.3 Residual Connectivity and Layer Normalization To improve the training stability of deep structures, this module introduces residual connections and layer normalization at the output of the feedforward network. This allows the output features to retain both the feedforward transformation results and the original input information, thereby mitigating gradient vanishing and accelerating convergence. The calculation formula is as follows: (twenty one); in, Normalization is performed on the feature dimension to improve numerical stability. For the output of the feedforward network, For residual input, The final fused feature representation output by this module will be directly input into the multi-task decoder.

[0059] 4. Decoding and Prediction Module This module is used for multi-task prediction of the contextual features output by the feature extraction and fusion module, supporting two main tasks: antimicrobial peptide activity classification and MIC value regression. The structure mainly includes sequence feature enhancement, attention pooling, parallel classification branches, and regression branches, with the specific process as follows.

[0060] 4.1 Feature Enhancement and Attention Pooling Output features of the previous module As input, the data sequentially passes through feature enhancement and attention pooling units to generate a global representation suitable for multi-task prediction. For batch processing, For sequence length, The feature dimension is defined as follows. This unit comprises two sub-modules: sequence feature enhancement and attention pooling.

[0061] (1) Sequence feature enhancement like Figure 8 As shown, this module receives the fused feature representation. (Obtained from formula (21)) and then input into the convolutional-positional encoding block (CNNPosBlock). This structure consists of multiple layers of one-dimensional convolutions and learnable positional encodings, which can capture the local patterns of adjacent residues while preserving the global sequence information. Specifically, the input features are first processed by one-dimensional convolutions to extract the local context, and then fused with the learnable positional encodings, thereby simultaneously enhancing the spatial pattern representation and sequence dependency modeling capabilities, and finally obtaining an enhanced feature representation containing local structure and sequence information.

[0062] First, input the fused features, then receive a batch of tensors. Next, local patterns are extracted through one-dimensional convolution to model short-range dependencies between neighboring residues. Then, learnable positional encodings are injected to preserve global sequence order information. The convolution results are then fused with the positional encodings to integrate local patterns and positional information. Finally, enhanced sequence features are output, resulting in... The above process can be represented as: (twenty two); This enhanced sequence feature not only preserves the global context but also significantly strengthens the expression of local patterns, which helps to capture short-range dependencies in antimicrobial peptides or pathogen features, thereby improving the performance of downstream tasks.

[0063] (2) Attention pooling To obtain a global vector representation suitable for downstream classification and regression tasks, the enhanced feature sequence is... Perform attention-weighted pooling: (twenty three); in, Global feature vectors used for live classification tasks. It is an attention weight vector specifically for classification tasks; Global feature vectors used for MIC value regression tasks. This is an attention weight vector specifically for regression tasks. Global features are calculated separately for each of the two tasks to ensure task specificity.

[0064] The calculation method for AttentionPooling is as follows: (twenty four) In the formula: To enhance the first sequence Feature vectors at each position; For the first Attention weights for each position; Indicates the total length of the sequence; To enhance the feature vector at position j in the sequence, the attention weights are adjusted accordingly. The calculation is based on normalizing the relative contributions of all positions in the sequence; finally, AttentionPooling normalizes the feature vectors of each position in the sequence. According to its corresponding attention weight Weighted summation yields an aggregated global feature representation. This highlights location information that is more relevant to the task and suppresses the influence of irrelevant parts.

[0065] 4.2 Classification and Regression Branches This module utilizes task-specific global features obtained through feature enhancement and pooling. and Based on this, a parallel classification and regression output structure is adopted to realize multi-task decoding of antimicrobial peptide activity classification and MIC value prediction.

[0066] (1) Taxonomic branching (activity prediction) Global features corresponding to the classification task The input is a fully connected network, which undergoes nonlinear transformation, normalization, and regularization to finally output the liveness classification probability: (25); Formula (25) describes the calculation process of activity prediction (category branching): where, This represents the global feature vector corresponding to the classification task; and The learnable weight matrix and bias terms are used to map the input features to a new feature space; first, they are calculated through a fully connected layer. The features are then subjected to a nonlinear transformation using the GELU activation function, and subsequently normalized using LayerNorm to obtain the normalized feature representation. Next, use Dropout to... Random deactivation is performed to prevent overfitting, resulting in regularized features. Finally, use a Linear layer to... Projection as It is a classification logits, which is transformed into a probability distribution of active classes through Softmax, thereby realizing the mapping from global features of the task to classification results.

[0067] (2) Regression branch (MIC value prediction) Global features corresponding to the regression task By inputting a two-layer fully connected network and combining normalization and Dropout, continuous predictions of the MIC value are obtained: (26); Formula (26) describes the calculation process of the regression branch (MIC value prediction): where, This represents the global feature vector corresponding to the regression task; and The weight matrix and bias terms are learnable and used to map the input features to a new feature space; firstly, for Perform linear transformation Then, Dropout is used to randomly deactivate the features to prevent overfitting. Next, the GELU activation function is used for non-linear mapping, and LayerNorm is used to normalize the features, resulting in a normalized regression feature vector. Subsequently, through linear transformation The predicted MIC values ​​are obtained, where, and These represent the learnable weights and biases of the regression output layer. The final output... These are continuous real values ​​used for the prediction of minimum inhibitory concentration (MIC).

[0068] 4.3 Output and Loss The final model output includes: Antimicrobial peptide activity classification probability, supporting binary or multi-class classification tasks.

[0069] MIC value regression prediction supports numerical prediction of continuous phenotypic values.

[0070] During the training phase, multi-task loss joint optimization is employed: (27); Formula (27) gives the joint optimization method for the multi-task loss function: where, The loss function for the antimicrobial peptide activity classification task is represented by cross-entropy loss, which measures the difference between the predicted class probability and the true label. The loss function for the MIC value regression task uses mean squared error (MSE) to measure the deviation between the predicted value and the true MIC value. These are weighting coefficients used to balance the classification and regression tasks in the total loss. The contribution ratio in the calculation. The final loss function. By weighting and summing the losses of the two types of tasks, the model can simultaneously optimize the performance of liveness classification and MIC value prediction during training.

[0071] This decoding and prediction module achieves full extraction and expression of contextual information from the input sequence through multi-level feature enhancement and adaptive pooling. Combined with a parallel classification and regression branch structure, it can simultaneously and accurately identify the activity category of antimicrobial peptides and perform fine regression prediction of MIC values ​​within a single framework, meeting the needs of multi-task bioinformatics analysis. The proposed module structure possesses high scalability and versatility, flexibly adapting to different types of antimicrobial peptide-pathogen data, significantly improving the overall model's predictive performance and generalization ability, and providing strong technical support for the efficient screening and mechanistic analysis of novel antimicrobial peptides.

[0072] II. Simulation Verification The antimicrobial peptide data used in this invention are derived from the internationally available GRAMPA database. Through rigorous data cleaning and preprocessing of the original GRAMPA dataset, redundant and non-standard sequences were removed, ultimately retaining 3345 valid antimicrobial peptide samples, thus fully ensuring the scientific validity and representativeness of the experimental data.

[0073] Pathogen protein data were retrieved from major international public databases, including UniProtKB-reviewed (artificially annotated protein), UniProtKB-unreviewed, UniParc, and NCBI Entrez Protein, using automated search tools based on the pathogen species involved in the antimicrobial peptide data. For each pathogen, a multi-level fallback mechanism was employed—first attempting to obtain high-quality UniProtKB-reviewed protein sequences; if not found, the search would switch to the next higher priority data source until usable protein sequences were obtained. The final pathogen dataset covered 163 pathogens, retaining a maximum of 30 representative protein sequences and a minimum of 1 for each pathogen, comprehensively reflecting pathogen diversity and sequence representativeness.

[0074] After standardization, all data were divided into training, validation, and test sets in an 8:1:1 ratio to ensure the fairness and scientific rigor of model training and evaluation. For model performance evaluation, mainstream classification metrics such as accuracy (ACC), F1 score, and area under the curve (AUC) were used for quantitative evaluation of the antimicrobial peptide activity classification task. For the MIC value regression prediction task, the root mean square error (RMSE) and mean absolute error (MAE) were used to measure the model's regression accuracy. All evaluation methods were based on the prediction results of the validation and test sets, comprehensively reflecting the model's classification and regression generalization capabilities and practical application value.

[0075] The results of the model in this invention on classification and regression tasks are shown in Table 1 below.

[0076]

[0077] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. An antimicrobial peptide activity and MIC value combined prediction framework based on cross-modal deep learning, characterized in that: The method comprises the following steps: A sequence encoding module is used to respectively embed and encode the antibacterial peptide sequence and the pathogen protein sequence by using a pre-trained protein language model, and learnable position encoding is introduced to inject position information into the antibacterial peptide embedding sequence and the pathogen protein embedding sequence; wherein the antibacterial peptide sequence has an active MIC label; A feature extraction and fusion module comprises a multi-layer stacked residual structure, and the residual structure comprises: A Mamba module comprises parallel multi-head self-attention branches, Mamba+GLU branches and DASM 1D convolution branches for simultaneously capturing global long-range dependency features, gate-weighted fusion features and local-global fusion features from the antibacterial peptide embedding sequence and the pathogen protein embedding sequence, learnable gating and SE units for dynamically weighting and fusing the global long-range dependency features, the gate-weighted fusion features and the local-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 antibacterial peptide modality; A feedforward neural network is used to make the features input from the Mamba module have smooth gradients while having nonlinear expression capability through a position-by-position linear transformation structure; A residual connection and layer normalization unit is used to simultaneously retain the transformation result of the feedforward neural network and the original input information; A decoding and prediction module is used to decode the features input from the feature extraction and fusion module into global feature vectors for an activity classification task and an MIC value regression task, and predict the probability distribution of the activity category and the continuous predicted MIC value through parallel classification branches and regression branches, respectively.

2. The cross-modality deep learning based framework for the combined prediction of antimicrobial peptide activity and MIC value according to claim 1, characterized in that: The active MIC label is obtained by logarithmic conversion of the minimum inhibitory concentration MIC data determined by experiments.

3. The cross-modality deep learning based framework for combined prediction of antimicrobial peptide activity and MIC value according to claim 2, characterized in that: The antibacterial peptide sequence and the pathogen protein sequence are obtained by preprocessing the original antibacterial peptide sequence and the pathogen protein sequence in FASTA format; the preprocessing comprises removing irrelevant characters, standardizing amino acid residue representation and limiting the maximum sequence length; wherein the standardization of the amino acid residue representation is performed by formula (1), and formula (1) is as follows: (1)； wherein: represents the number of antimicrobial peptides or pathogen, is the number of amino acid residues, is the set of amino acid residues supported by the protein language model, is the actual length of the sequence, and , represents the maximum sequence length.

4. The cross-modality deep learning based framework for combined prediction of antimicrobial peptide activity and MIC value according to claim 3, characterized in that: The embedding encoding is represented as: (2)； wherein: is the embedding of the antibacterial peptide sequence, with the number of rows being the actual length of the sequence , each row corresponding to a feature vector of an amino acid residue, is the output embedding dimension of the protein language model.

5. The cross-modality deep learning based framework for combined prediction of antimicrobial peptide activity and MIC value according to claim 4, characterized in that: The specific process of the embedding encoding comprises the following steps: S51, indexing the sequence input into the protein language model to map 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; S52, adding a start token BOS and an end token EOS at the beginning and end of the sequence to indicate the sequence boundary; S53, mapping the indexed sequence to a dense word vector by using a learnable embedding matrix, and superimposing learnable position encoding to provide residue position information; S54, inputting the dense word vector with residue position information into a multi-layer Transformer encoder of the protein language model for encoding; S55、After the encoding is completed, the vectors corresponding to the beginning and end BOS / EOS markers are removed, and only the representation of the real residues is retained to output a residue-level embedding matrix with a shape of and bound with the unique ID of the sample to form an antibacterial peptide embedding feature or a pathogen embedding feature.

6. The cross-modality deep learning based framework for combined prediction of antimicrobial peptide activity and MIC value according to claim 5, characterized in that: The Mamba module comprises parallel branch units, learnable gating and SE units, and a feature mixing unit, and the parallel branch units comprise parallel multi-head self-attention mechanism branches, Mamba+GLU branches and DASM 1D convolution branches; wherein: Mamba+GLU branch, using Mamba state space model to extract long-range dependence features, and using GLU gating mechanism to filter out functional fragments, and through element-by-element multiplication operation to obtain gated weighted fusion features; Multi-head self-attention branch, extracting global long-range dependence relationship features through multi-head self-attention mechanism; DASM 1D convolution branch, extracting local context semantic information through dynamic sampling and channel convolution, and supplementing global long-range dependence relationship features to obtain local and global fusion features; Learnable gating and SE unit, introducing branch gating weighting mechanism and channel adaptive mechanism SE to dynamically weight fuse the locally and globally fused features, the gated weighted fusion features and the global long-range dependence relationship features, and adaptively adjusting the channel feature weights in the channel dimension to obtain the dynamic weighted fusion features of the antimicrobial peptide mode and the pathogen mode; Feature mixing unit, mapping the antimicrobial peptide mode dynamic weighted fusion features to the cross-modal interaction dimension through the fully connected layer to obtain the antimicrobial peptide projection features, and inputting the antimicrobial peptide projection features as the query vector into the introduced cross-modal attention mechanism, and at the same time, inputting the pathogen mode dynamic weighted fusion features as the key vector and the value vector into the introduced cross-modal attention mechanism to match the query vector; cross-modal attention mechanism, calculating the attention weight according to the similarity of the query vector and the key vector, and weighting and aggregating the value vector based on the calculated attention weight to generate the cross-modal interaction fusion features of the antimicrobial peptide mode and the pathogen mode.

7. The cross-modality deep learning based framework for combined prediction of antimicrobial peptide activity and MIC value according to claim 6, characterized in that: The cross-modal interaction fusion features are input into the feedforward neural network, which adopts a position-by-position linear transformation structure to independently transform the features of each position without mixing the information between positions, so that the cross-modal interaction fusion features have nonlinear expression ability while having smooth gradients to generate cross-modal interaction fusion enhanced features.

8. The cross-modality deep learning based framework for combined prediction of antimicrobial peptide activity and MIC value according to claim 7, characterized in that: The cross-modal interaction fusion enhanced features are input into the residual connection and layer normalization unit, and through the residual connection and layer normalization, the output modal interaction fusion enhanced features retain the feedforward transformation results and the original input information at the same time, so as to relieve gradient vanishing, accelerate convergence and deepen network modeling depth.

9. The cross-modality deep learning based framework for combined prediction of antimicrobial peptide activity and MIC value according to claim 8, characterized in that: The global feature vector of the activity classification task is input into the fully connected layer of the classification branch, which is projected into classification features after nonlinear transformation, normalization and regularization processing, and then the classification features are converted into probability distribution of activity class through activation function to realize activity classification prediction.

10. The cross-modality deep learning based framework for combined prediction of antimicrobial peptide activity and MIC value according to claim 9, characterized in that: The global feature vector of the MIC value regression task is input into the two-layer fully connected layer of the regression branch, which is combined with normalization and regularization to obtain continuous prediction of MIC value.

Images