A method for predicting the properties of antibody-drug conjugates based on multimodal fusion

Abstract

This invention proposes a method for predicting the properties of antibody-drug conjugates based on multimodal fusion, belonging to the field of bioinformatics. First, sequence position information is explicitly modeled using sinusoidal positional encoding. Second, a bidirectional cross-attention mechanism is introduced to establish light chain and heavy chain interactions and antigen-antibody alignment. Third, an integrated graph neural network is used to reconstruct adjacency relationships and extract topological features based on attention weights. Finally, a two-stage adaptive refinement module is combined to achieve alignment and refinement in two stages, adaptively adjusting the contribution of each modality's features. Simultaneously, mask-aware feature extraction enhances sequence robustness, and attention weights enable interpretability analysis of key binding sites. This method significantly improves prediction accuracy and can be widely applied in cancer targeted therapy and drug design optimization.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of bioinformatics, and relates to an antibody-drug conjugate (ADC) property prediction method based on multi-modal fusion. The method realizes high-precision prediction of ADC effectiveness and safety by fusing sequence encoding, graph neural network, bidirectional cross-attention mechanism and adaptive multi-modal refinement technology. BACKGROUND

[0002] ADC is an important class of targeted anti-cancer drugs, which consists of three core components: antibody, linker and toxic payload. The antibody includes heavy chain (HC) and light chain (LC), which is responsible for specific recognition of tumor cell surface antigens; the linker connects the chemical bond between the antibody and the toxic payload, affecting the stability and release characteristics of the drug; the toxic payload is a small molecule drug with cytotoxicity, which is responsible for killing tumor cells. The design and optimization of ADC need to consider the antibody-antigen binding specificity, linker stability, toxic payload release efficiency and overall pharmacokinetic characteristics, which has traditionally been highly dependent on experimental verification, with defects such as high cost, long cycle and low throughput.

[0003] To break through the experimental bottleneck, computational prediction methods have gradually become a research hotspot. Early methods were mainly based on sequence similarity analysis or simple feature splicing, and such methods could only capture linear sequence features, ignoring the three-dimensional conformation dynamics, topological dependence and multi-modal synergistic effects between the components of ADC. For example, the antibody light chain and heavy chain affect the antigen binding ability through the synergistic formation mechanism of the complementarity determining region (CDR), and a simple splicing model cannot model such inter-chain interaction; antigen-antibody binding involves the spatial complementarity of epitopes and CDR regions, and sequence similarity models cannot capture specific pairing relationships; the topological structure of the linker and the toxic payload (such as atom-bond connection mode, flexible length) directly affects the stability of the drug, and sequence encoding cannot fully utilize such structural information.

[0004] In recent years, deep learning techniques have been introduced into the field of ADC property prediction, especially the Transformer architecture and graph neural networks (GNN). In the prior art, methods such as ADCNet use a basic Transformer Encoder to encode SMILES (Simplified Molecular Input Line Entry System) sequences and fuse multi-modal information through simple feature concatenation. However, such methods still have the following technical defects: insufficient sequence modeling: lack of position encoding mechanism, unable to effectively model the position information in the SMILES sequence, resulting in insufficient model perception of sequence order; missing inter-chain interaction: light chain and heavy chain features are processed independently, without explicitly modeling inter-chain synergistic relationships, unable to capture the synergistic formation mechanism of CDR regions; simplified antigen-antibody interaction: simple concatenation is used for antigen-antibody interaction, unable to capture specific pairing relationships and identify key binding sites; missing graph structure modeling: lack of graph neural networks, unable to fully utilize the topological information of small molecules and capture the atom-bond topological properties of linkers and toxic payloads; inefficient multi-modal fusion: multi-modal feature fusion is too simple, without considering the importance differences of different modalities, unable to adaptively adjust the contribution of each modality feature.

[0005] At the algorithm optimization level, the prior art has the following shortcomings: single loss function: most models only use a single loss function, such as binary cross-entropy (BCE), making it difficult to optimize both classification accuracy and global ranking ability; rigid optimization strategy: traditional optimizers do not design a collaborative update mechanism for multi-modal fusion, resulting in conflicts in the gradient direction of different modal features, slow model convergence speed, and easy to fall into local optimum. In summary, existing ADC property prediction techniques face multiple challenges such as sequence modeling bias, inter-chain interaction missing, antigen-antibody interaction simplification, graph structure modeling missing, and inefficient multi-modal fusion. Developing a new prediction framework that integrates multi-modal feature representation, bidirectional cross-attention, graph neural networks, and adaptive refinement mechanisms has become an urgent need to promote the development of computational drug design. SUMMARY

[0006] To address the core issues of sequence modeling bias, inter-chain interaction missing, antigen-antibody interaction simplification, graph structure modeling missing, and inefficient multi-modal fusion identified in the background art, the present invention proposes an antibody drug conjugate property prediction method based on multi-modal fusion. The technical solution achieves breakthrough through the following innovative designs:

[0007] Sinusoidal Positional Encoding System: Integrate sinusoidal positional encoding in the Transformer Encoder to explicitly model the position information in SMILES sequences, improving the perception of sequence order; Chain-level vector structured representation: Map the chain-level vectors of antibody light chains, heavy chains, and antigens to fixed-length token sequences through the ChainEmbedding module to unify the feature representation of different modalities and provide standard input for subsequent cross-attention; Bidirectional Cross-Attention Mechanism: Light Chain-Heavy Chain Bidirectional Cross-Attention: Establish explicit inter-chain interaction through the AntibodyFusion module to simulate the CDR cooperation formation process; Antigen-Antibody Bidirectional Cross-Attention: Identify key binding sites through the AntigenAntibodyCrossFusion module to capture specific pairing relationships; Graph Neural Network Integration: Through the GraphConvLayer module, reconstruct the adjacency relationship based on attention weights, extract topological features of linkers and toxic payloads, and capture structural signals such as atomic bond connection patterns, lengths, and flexibility; Dual-Stage Adaptive Refinement (DAR): Through the DualStageAdaptiveRefinement module, implement two-stage processing of the alignment stage and the adaptive refinement stage, adaptively adjust the contribution of each modality feature, and improve the multi-modal fusion effect; Through these designs, improve the accuracy, interpretability, and robustness of ADC property prediction, and the invention specifically includes the following steps:

[0008] Step 1, data acquisition and multi-modal feature extraction, obtain the SMILES sequences of toxic payloads and linkers, ESM-2 embedding vectors of heavy chains / light chains / antigens, and DAR feature scalars from the ADC dataset; process the data by removing noise, length standardization, and multi-modal feature extraction;

[0009] Step 2, sequence encoding and structure feature representation, use the Transformer encoder to perform deep representation of the SMILES sequences of toxic payloads and linkers, and add sinusoidal position encoding to retain key relative position information in the structure. Antibody and antigen sequences are mapped to fixed-length representations that can be used for cross-chain interaction through chain-level vectors;

[0010] Step 3, chain-level vector structured representation and bidirectional cross-attention fusion, construct a bidirectional attention mechanism between antibody light chains and heavy chains to realize the fusion of antibody internal features; further construct a bidirectional attention interaction between antibody and antigen sequences to align key residue regions, capture potential binding sites and action areas, and enable effective fusion of multi-modal features at the chain level;

[0011] Step 4, graph neural network topology feature extraction, based on the attention weight output of SMILES encoding and the initialized bond adjacency matrix, the molecular structure of the toxicity load and the linker is topologically reconstructed; then the graph convolution network is used to extract the atomic level and bond level graph structure features, realizing the joint expression of sequence features and graph structure features;

[0012] Step 5, DAR multi-modal fusion, construct a cross-modal alignment mechanism to map sequence features, graph features and chain level semantic representations to a unified space; introduce DAR scalar to adaptively adjust the weight of each modal feature, and filter key modal features through the gating method to improve the stability and pertinence of feature expression;

[0013] Step 6, feature aggregation and classification prediction, pooling the effective token; concatenating the DAR refined features and the DAR scalar, and performing binary classification prediction through the fully connected layer;

[0014] Step 7, model training and verification, using cross-validation strategy to evaluate the model performance with AUC index; calculate the results with statistical significance through multiple independent training (different random seeds).

[0015] Precise prediction of ADC properties is crucial for cancer targeted therapy, drug design optimization and clinical translation. Based on the residue level analysis of bidirectional cross attention weight, the key binding site of antibody-antigen is accurately located, which provides direct guidance for ADC optimization design. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is a flow chart of an antibody drug conjugate property prediction method based on multi-modal fusion.

[0017] Figure 2 It is a flow chart of a bidirectional cross attention mechanism.

[0018] Figure 3 It is a flow chart of antigen-antibody bidirectional cross attention alignment.

[0019] Figure 4 It is a flow chart of graph neural network topology feature extraction. DETAILED DESCRIPTION

[0020] The implementation steps of the present application are described in detail below in combination with the drawings and examples. The present method is realized in Python 3.7 environment based on TensorFlow 2.3 framework, and the specific process includes seven steps of data acquisition and multi-modal feature extraction, sequence encoding and structure feature representation, chain level vector structured representation and bidirectional cross attention fusion, graph neural network topology feature extraction, DAR multi-modal fusion, feature aggregation and classification prediction, and model training and verification, as shown in Figure 1A flowchart of a method for predicting the properties of antibody-drug conjugates based on multimodal fusion is shown below, with details of each part:

[0021] Step 1: Data Acquisition and Multimodal Feature Extraction. This step begins with data acquisition and cleaning. The data source is the ADC dataset, which contains the following information: toxicity load. Figure 2 SMILES Sequence: A string representation of the SMILES sequence of the toxic payload small molecule, with variable length; Linker SMILES Sequence: A string representation of the SMILES sequence of the linker small molecule, with variable length; Heavy Chain Feature Vector: A heavy chain embedding vector extracted using the ESM-2 pre-trained model, with a dimension of 1280; Light Chain Feature Vector: A light chain embedding vector extracted using the ESM-2 pre-trained model, with a dimension of 1280; Antigen Feature Vector: An antigen embedding vector extracted using the ESM-2 pre-trained model, with a dimension of 1280; DAR Feature Scalar: A numerical feature of the drug-to-antibody ratio. Denoising is then performed, removing samples containing non-standard characters from the SMILES sequence; removing samples with abnormal lengths: the SMILES sequence length is limited to 3-200 characters; removing duplicate records: deduplication is performed based on the MD5 hash value of the Payload-Linker-heavy chain-light chain-antigen combination to avoid training data redundancy. The second part of this step is sequence standardization and padding. The SMILES sequences are standardized in length and padded to the maximum length using a special marker [PAD], generating a corresponding mask matrix to identify the positions of valid tokens. The third part of this step involves constructing the adjacency matrix. Based on the atomic-bond connections of the SMILES sequences, a toxic vector is constructed. Figure 2 The adjacency matrix of the connectors is used for subsequent graph convolution operations.

[0022] Step 2: Sequence Encoding and Structural Feature Representation. This step performs deep encoding of the SMILES sequence using a Transformer Encoder architecture and explicitly models sequence position information. A 6-layer encoder is used, with each layer containing 8 independent attention heads. The model dimension is... =256, feedforward network dimensions =512, such as Figure 3 The flowchart of the bidirectional cross-attention mechanism is shown. Position encoding uses a sinusoidal form, and the angle function is defined as: The position code is in the first peacekeeping The dimensions are defined as follows: ,in : Sequence position index (starting from 0); Encoding dimension index; Channel index; Model dimension. For models with a length of... sequence, construct position encoding matrix . When Encoder forward propagates, add position encoding to token embedding: where is the embedding matrix of SMILES token. In implementation, first construct independent vocabulary for toxicity payload and linker respectively, map SMILES character to integer token, then map to dimension through embedding layer; then generate position index according to actual length of sequence, look up table to get corresponding sinusoidal position encoding vector, add to embedding result element by element as input tensor of Transformer Encoder, shape is . Encoder outputs at last layer: encoded feature ; multi-head attention weight where is the batch size of 32, . When return_attention=True, return and at the same time, which are used for graph structure reconstruction later. Call Encoder for toxicity payload and linker SMILES respectively, input includes token sequence, mask matrix and adjacency matrix, and set return_attention=True, output corresponding encoded feature and attention weight.

[0023] Step 3, chain level vector structured representation and bidirectional cross attention fusion. This step converts ESM-2 chain level vector to fixed length token sequence, and models antibody internal and antigen-antibody interaction through bidirectional cross attention, as shown in the antigen-antibody bidirectional cross attention alignment flow chart. Figure 4 ChainEmbedding module: linear projection and normalization on heavy chain, light chain and antigen chain level vector to get token representation. Mathematical form is: where: : linear projection layer, map 1280 dimension to ; : learnable position embedding; : scaling factor; and are standard normalization operators. Output dimension is token sequence of . In implementation, ChainEmbedding submodule with the same structure can be instantiated for heavy chain, light chain and antigen respectively, input is chain vector with shape Reshape to After adding learnable position embeddings, it is followed by Dropout and LayerNorm. This can unify the token representation of three chains while ensuring the clarity of parameter sharing strategy. AntibodyFusion module: This module performs two cross-attention, heavy chain is enhanced by light chain, heavy chain token as Query, light chain token as Key / Value; light chain is enhanced by heavy chain, light chain token as Query, updated heavy chain as Key / Value. Both updates use residual connection and LayerNorm, finally concatenate the two in the sequence dimension to get the fusion antibody token tensor AntigenAntibodyCrossFusion module: Two-way cross-attention is used to align the antibody and antigen, and then global average pooling is performed on the sequence, and then two fully connected networks are used to map to a 256-dimensional interaction feature. This interaction feature is used as the input of the DAR module.

[0024] Step 4, Topological feature extraction of graph neural network. This step extracts the topological features of toxic carriers Figure 4 and linkers through the GraphConvLayer module, such as Figure 4 as shown in the flowchart of the graph neural network topology feature extraction. Adjacency matrix reconstruction: First, the multi-head attention weight of the last layer of the Encoder is averaged in the attention head dimension to obtain a fused attention matrix. Then normalize the matrix and inverse sigmoid to map it to 0-1. In order to consider the structure prior and attention information, the normalized attention matrix and the original adjacency matrix are fused according to the weight of 1:1 to obtain the final reconstructed adjacency matrix. This fusion operation is realized by simple linear weighting. Symmetric normalization: In order to avoid the scaling instability caused by the difference in node degree, symmetric normalization needs to be performed on the reconstructed adjacency matrix: where each diagonal element of the degree matrix D represents the sum of all edge weights of the corresponding node. Graph convolution and pooling: Let the node feature matrix be , and the learnable weight matrix be . The graph convolution is defined as: where the residual connection preserves the original features. Global average pooling is performed on the convolved node features: , respectively, to obtain the corresponding graph-level topological features. In specific implementation, the GraphConvLayer can be designed to accept three types of inputs: the original adjacency matrix, the attention weight, and the node feature. First, the attention head average and Sigmoid normalization are performed on each sample in the batch dimension, and then the original adjacency matrix is linearly fused and the degree matrix and its square are calculated, with numerical protection for nodes with a degree of 0 (for example, setting the corresponding diagonal element to 1). Subsequently, batch matrix multiplication is used to complete the normalization and graph convolution operation, and finally the average is taken in the node dimension to obtain the graph-level feature vector of each molecule, which can be directly concatenated with other modal features.

[0025] Step 5, DAR multi-modal fusion. In this step, the DAR module is used to align and adaptively refine the antibody-antigen interaction features, antigen features, Payload graph features, and Linker graph features. Alignment stage: linear projection is performed on each modal feature: , where: : antibody-antigen interaction feature, : antigen feature, : toxicity load ​ graph feature, : linker graph feature, : corresponding projection matrix, and the output dimension is . The four modal features are expanded into four tokens and concatenated: , cross-modal alignment is performed through multi-head self-attention: , . Adaptive refinement stage: introduce learnable DAR scalar , which corresponds to the four modalities respectively. The four aligned tokens are split by modality and multiplied by their respective DAR scalar to achieve dynamic weighting. The four weighted vectors are concatenated again and fed into the gated network. The gated network performs element-wise filtering on the input features to suppress noise channels and highlight key modalities. Then, two-layer feedforward networks are used for final fusion: , output unified multi-modal feature .

[0026] Step 6, feature aggregation and classification prediction. Mask-aware pooling: mask-aware global average pooling is used for the encoding features of the toxicity load and linker. Let the feature of the i-th token be , the corresponding mask be , and the pooled feature be: , where: : valid token; ​​: padding token. In implementation, the original mask (0 for padding, 1 for valid) should be first converted to a float weight matrix, and the number of valid tokens should be counted for each sample in the batch dimension to avoid division by 0 (e.g., force the sample with 0 count to be 1). Then the features at padding positions are set to 0 by element-wise multiplication, and the sum is taken in the sequence dimension and divided by the number of valid tokens to get the pooling vector independent of sequence length. Feature concatenation and classification head: DAR output features with DAR scalar concatenated to get classification head input: The classification head is a two-layer fully connected network: , , with the first layer output dimension 256 and the second layer output dimension 2 to get binary classification probability by Softmax. In implementation, the classification head can be regarded as a simple MLP head: the first linear layer maps dimension input to 256 dimension and applies LeakyReLU activation, then Dropout is applied to suppress overfitting, and the second linear layer outputs logits of length 2. Finally, in the training phase, the cross-entropy loss function is fed with the labels, and in the inference phase, the probabilities belonging to the positive and negative classes are obtained by Softmax.

[0027] Step 7, model training and validation. This step uses a cross-validation strategy to divide the dataset into a training set and a test set. In training, the binary cross-entropy loss function and the Adam optimizer are used, with the learning rate set to , the batch size is 32, and the Dropout rate is 0.1. Multiple independent training: use different random seeds for completely independent training, and save the best model weights each time. After training, the average and standard deviation of all indicators are calculated to provide statistically significant results. Result output: output CSV file containing: toxicity payload SMILES, linker SMILES, heavy chain / light chain / antigen identifier; predicted probability, predicted label (1 / 0).

[0028] After multiple rounds of experimental verification, the AUC value obtained by this method on the ADC dataset is 0.9509. Compared with the AUC value 0.9293 obtained by the ADCNet method, the increase of this method is 2.16%. By introducing the attention guiding mechanism, this method can clearly locate the key binding sites in the antibody-antigen interaction, providing direct guidance for ADC optimization design.

[0029] The above is further detailed description of the present application in combination with specific preferred embodiments, and cannot be deemed as limitation of the specific implementation of the present application to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, and all of them shall be deemed as falling within the protection scope of the present application.

Claims

1. A method for predicting properties of an antibody drug conjugate based on multi-modal fusion, characterized in that, Fusion sequence encoding, graph neural network, bidirectional cross attention mechanism and adaptive multi-modal refinement technology, antibody drug conjugate is abbreviated as ADC, two-stage adaptive refinement is abbreviated as DAR, including the following steps: Step 1, data acquisition and multi-modal feature extraction, obtaining SMILES sequence of toxicity load and linker, ESM-2 embedding vector of heavy chain / light chain / antigen and DAR characteristic scalar from ADC data set; The data is processed by denoising, length standardization and multi-modal feature extraction; Step 2, sequence encoding and structure feature representation, using Transformer encoder to deeply represent the SMILES sequence of toxicity load and linker, while adding sinusoidal position encoding to preserve the key relative position information in the structure; specifically, the SMILES sequence is deeply encoded by the Transformer Encoder architecture, and the sequence position information is explicitly modeled, 6 layers of Encoder are adopted, each layer contains 8 independent attention heads, the model dimension = 256, the feedforward network dimension = 512; the position encoding adopts a sinusoidal form, and the angle function is defined as: , the position encoding is defined in the th dimension and the th dimension, respectively: , , where : sequence position index; : encoding dimension index; : channel index; : model dimension; for a sequence of length , the position encoding matrix is constructed; when the Encoder is forward propagated, the position encoding is added to the token embedding: , where is the embedding matrix of SMILES token; in the specific implementation, first, an independent vocabulary is constructed for the toxicity load and linker respectively, mapping the SMILES characters to integer tokens, and then mapping to dimension through the embedding layer; then generate the position index according to the actual length of the sequence, look up the corresponding sinusoidal position encoding vector, and add it to the embedding result element by element to serve as the input tensor of the Transformer Encoder, with a shape of ; the Encoder outputs at the last layer: encoding features ; multi-head attention weights , where is the batch size of 32, ; when return_attention=True, both and are returned, which are used for graph structure reconstruction later; the Encoder is called for the toxicity load and linker SMILES respectively, the input includes token sequence, mask matrix and adjacency matrix, and return_attention=True is set, the output corresponding encoding features and attention weights; Step 3, chain level vector structured representation and bidirectional cross attention fusion, construct bidirectional attention mechanism between antibody light chain and heavy chain, realize the fusion of antibody internal features; Further construct the bidirectional attention interaction between antibody and antigen sequence, align the key residue region, capture the potential binding site and action area, so that the multi-modal features are effectively fused at the chain level; Step 4, graph neural network topology feature extraction, based on the attention weight output by SMILES encoding and the initialized bond adjacency matrix, the molecular structure of toxicity load and linker is reconstructed; Then graph convolution network is used to extract atomic level and bond level graph structure features, realizing the joint expression of sequence features and graph structure features; Step 5, DAR multi-modal fusion, construct cross-modal alignment mechanism, map sequence features, graph features and chain level semantic representation to unified space; Introduce DAR scalar to adaptively adjust the weight of each modal feature, filter key modal features through gating, and improve the stability and pertinence of feature expression; Step 6, feature aggregation and classification prediction, pool the effective token; The features after DAR refinement and DAR scalar are spliced, and the two-class prediction is made through the full connection layer; Step 7, model training and verification, cross-validation strategy is adopted, and AUC index is used to evaluate the model performance.

2. The method according to claim 1, wherein, Step 3 converts the ESM-2 chain-level vectors into a fixed-length token sequence and models the intra-antibody and antigen-antibody interactions through bidirectional cross-attention; ChainEmbedding module: chain-level vectors for heavy chain, light chain, and antigen Linear projection and normalization are performed to obtain the token representation; the mathematical form is: wherein: : linear projection layer that maps 1280 dimensions to ; : learnable position embedding; : scaling factor; and are standard regularization operators; the output dimension is a token sequence of ; in implementation, the ChainEmbedding submodule of the same structure can be instantiated for heavy chain, light chain, and antigen respectively, with the input being a chain vector of shape , which is first reduced in dimension by a fully connected layer to , then reshaped to , and after adding a learnable position embedding, sequentially passes through Dropout and LayerNorm; this can unify the token representation of the three chains while ensuring a clear parameter sharing strategy; AntibodyFusion module: this module performs cross-attention twice, with the heavy chain being enhanced by the light chain and the heavy chain token serving as the Query and the light chain token serving as the Key / Value; the light chain is enhanced by the heavy chain, with the light chain token serving as the Query and the updated heavy chain serving as the Key / Value; both updates use residual connection and LayerNorm, and finally the two are concatenated in the sequence dimension to obtain the fused antibody token tensor ; AntigenAntibodyCrossFusion module: bidirectional cross-attention is used twice to align the antibody and antigen with each other, and then global average pooling is performed on the sequence, followed by a two-layer fully connected network to map to a 256-dimensional interaction feature; this interaction feature is used as the input to the DAR module. 3.The method of claim 1, wherein the method is characterized by, Step 4 extracts the topological features of the toxicity load and linker through the GraphConvLayer module; Adjacency matrix reconstruction: first, take the average of the multi-head attention weight of the last layer of the Encoder in the attention head dimension to obtain a fused attention matrix; Then map the matrix and inverse sigmoid normalization to 0-1; In order to consider the structure prior and attention information, the normalized attention matrix and the original adjacency matrix are fused according to the weight of 1:1 to obtain the final reconstructed adjacency matrix; The fusion operation is achieved by simple linear weighting; symmetric normalization: to avoid scaling instability caused by node degree difference, symmetric normalization is needed for the reconstructed adjacency matrix: where each diagonal element of the degree matrix D represents the sum of all edge weights of the corresponding node; graph convolution and pooling: let the node feature matrix be , and the learnable weight matrix be ; the graph convolution is defined as: where the residual connection preserves the original features; global average pooling is performed on the convolved node features: The above process is performed on the toxic load and linker, respectively, to obtain the corresponding graph-level topological features; in specific implementation, the GraphConvLayer can be designed to accept three types of input: the original adjacency matrix, attention weight, and node feature; first, the attention head average and Sigmoid normalization are performed on each sample in the batch dimension, then linearly fused with the original adjacency matrix and the degree matrix and its power are calculated, and numerical protection is performed on nodes with a degree of 0; then, batch matrix multiplication is used to complete the normalization and graph convolution operation, and finally, the average is taken in the node dimension to obtain the graph-level feature vector of each molecule, which can be directly concatenated with other modal features.

4. The method according to claim 1, wherein, Step 5 aligns and adaptively refines the antibody-antigen interaction features, antigen features, Payload graph features and Linker graph features through the DAR module; Alignment stage: linear projection is performed on each modal feature: wherein: : antibody-antigen interaction feature, : antigen feature, : toxic load profile feature, : linker profile feature, : corresponding projection matrix, the output dimensions are ; four modal features are expanded into 4 tokens and spliced: , cross-modal alignment is performed through multi-head self-attention: , ; Adaptive refining stage: introducing learnable DAR scalars , which correspond to the four modalities respectively; the aligned four tokens are split by modality and multiplied by their respective DAR scalars to achieve dynamic weighting; the four weighted vectors are spliced again and fed into the gating network; the gating network performs element-by-element screening on the input features to suppress noise channels and highlight key modalities; then the final fusion is performed through two layers of feedforward network: , output unified multi-modal feature .

5. The method according to claim 1, wherein, Step 6: Implement Mask-Aware Pooling: Apply mask-aware global average pooling to the encoded features of toxicity payloads and connectors; let the... The characteristics of a token are: The corresponding mask is The pooling characteristics are: ,in: : Valid token; Padding token; In implementation, the original mask should first be converted into a floating-point weight matrix, and the number of valid tokens for each sample should be counted in the batch dimension to avoid division by zero; then, the features at the padding positions are set to zero through element-wise multiplication, and then summed in the sequence dimension and divided by the number of valid tokens to obtain a pooling vector that is independent of the sequence length; Feature concatenation and classification head: DAR output features With DAR scalar By concatenating the data, we obtain the classification header input: The classification head is a two-layer fully connected network: , , The first layer has an output dimension of 256, and the second layer has an output dimension of 2. Binary classification probabilities are obtained through Softmax. In implementation, the classification head can be considered a simple MLP head: the first linear layer will... The input is mapped to 256 dimensions and LeakyReLU activation is applied. Dropout is then applied to suppress overfitting. The second linear layer outputs logits of length 2. Finally, during the training phase, the logits are fed into the cross-entropy loss function along with the labels. During the inference phase, the probabilities of belonging to the positive and negative classes are obtained through Softmax.

Images