A method and device for predicting drug-target association

Abstract

This application belongs to the field of data processing technology, and specifically relates to a method and apparatus for predicting drug-target association. The method includes: first, acquiring samples containing the compound SMILES string, protein amino acid sequence, and its structural attribute sequence; second, extracting the compound's sequence, molecular graph, molecular fingerprint, and pre-trained language model representation, as well as the protein's sequence, structural attributes, and pre-trained language model representation; then, fusing the compound-side representations in a unified latent space to obtain a compound fusion vector, and performing the same operation on the protein-side to obtain a protein fusion vector; finally, concatenating the two fusion vectors and inputting them into a prediction head composed of a KAN linear layer, outputting the interaction probability or binding affinity value. This application fully utilizes the multi-source features of drugs and proteins, enhances the ability to model complex nonlinear relationships, and improves the accuracy and generalization ability in drug target prediction tasks.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and apparatus for predicting the correlation between drugs and targets. Background Technology

[0002] Drug-target interaction prediction is a crucial step in drug discovery. Traditional experimental methods are costly and time-consuming; therefore, virtual screening using machine learning or deep learning models has become an important research direction. Existing methods mainly suffer from the following shortcomings:

[0003] 1. Limited single-modal modeling capabilities. Models based solely on SMILES, protein amino acid sequences, or molecular diagrams are insufficient to fully express the complex interactions between drugs and targets.

[0004] 2. Insufficient utilization of complementary information between different modalities. Although many existing methods introduce multiple features, they usually only perform simple splicing or coarse-grained fusion, which easily introduces redundant information.

[0005] 3. There is an imbalance in representational capabilities between the compound and protein sides. Compounds can extract information from graph structures and fingerprints, while proteins are typically encoded only by sequence, resulting in underutilization of structural properties.

[0006] 4. Traditional fully connected prediction heads have limited expressive power. When performing the final nonlinear mapping on fused features, conventional MLPs struggle to fully characterize complex high-order relationships.

[0007] 5. Limited model generalization ability. Existing methods often require specific modifications when faced with different datasets, classification tasks, and regression tasks, resulting in insufficient generalizability.

[0008] Therefore, there is a need for a DTI / DTA prediction scheme that can unify and integrate multimodal drug and protein information while taking into account both expression capacity and engineering feasibility. Summary of the Invention

[0009] To address the aforementioned issues, this application provides a method and apparatus for predicting drug-target correlation. This method comprehensively utilizes multi-source information such as compound molecular maps, molecular sequences, protein sequences, protein structural attribute sequences, and molecular fingerprints to establish a unified prediction model, thereby improving the prediction accuracy, generalization ability, and interpretability in drug screening tasks.

[0010] The first aspect of this application provides a method for predicting the correlation between a drug and a target, mainly including:

[0011] S1. Obtain the data to be predicted, which includes the SMILES string of the drug compound, the amino acid sequence of the target protein, and the structural property sequence of the target protein.

[0012] S2. Encode the SMILES string, amino acid sequence, and structural attribute sequence to obtain the corresponding compound digital tensor, protein digital tensor, and structural attribute digital tensor.

[0013] S3. Based on the SMILES string, generate the molecular map and molecular fingerprint of the drug compound;

[0014] S4. Input the compound digital tensor into the first sequence feature encoding network branch of the pre-trained prediction model to obtain the compound sequence vector; input the molecular graph into the graph message passing network branch of the prediction model to obtain the molecular graph vector; input the molecular fingerprint into the projection network branch of the prediction model to obtain the molecular fingerprint vector; input the SMILES string into the first language analysis model branch of the prediction model to obtain the compound model encoding vector; input the protein digital tensor into the second sequence feature encoding network branch of the prediction model to obtain the protein sequence vector; input the structural attribute digital tensor into the structural attribute encoding network branch of the prediction model to obtain the structural attribute vector; input the amino acid sequence into the second language analysis model branch of the prediction model to obtain the protein model encoding vector.

[0015] S5. The compound sequence vector, molecular graph vector, molecular fingerprint vector, and compound model encoding vector on the compound side are fused in the latent space of the prediction model to obtain the compound fusion vector; the protein sequence vector, structural attribute vector, and protein model encoding vector on the protein side are fused in the latent space of the prediction model to obtain the protein fusion vector.

[0016] S6. Concatenate the compound fusion vector with the protein fusion vector to obtain the drug target joint vector;

[0017] S7. Input the drug target joint vector into the prediction head of the prediction model, which consists of at least one KAN linear layer, and output the predicted probability of drug target interaction or the predicted value of binding affinity.

[0018] Preferably, in step S2, the SMILES string and amino acid sequence are encoded using the BPE vocabulary, and the structural attribute sequence is encoded using a vocabulary constructed from a predefined set of structural attribute lexicons, which includes structural, exposure, polarity, and fragment length attributes.

[0019] Preferably, step S2 further includes:

[0020] Fixed-length padding or truncation is performed on compound digital tensors, protein digital tensors, and structural attribute digital tensors. The vocabulary constructed from the BPE vocabulary and the predefined set of structural attribute lexicons contains padding markers for supplementing each digital tensor to a uniform length, as well as unknown markers for sequence attributes that are not present or cannot be identified in the vocabulary.

[0021] Preferably, step S3 further includes:

[0022] The SMILES string was converted into a molecular graph with atoms as nodes and chemical bonds as edges using a cheminformatics toolkit.

[0023] A fixed-length circular fingerprint of Morgan was calculated using a cheminformatics toolkit and used as a molecular fingerprint.

[0024] Preferably, in step S7, the prediction head includes multiple KAN linear layers, which achieve nonlinear mapping through sine and cosine basis expansion and learnable Fourier coefficients.

[0025] The second aspect of this application provides a drug-target correlation prediction device, mainly comprising:

[0026] The data acquisition module is used to acquire the data to be predicted, which includes the SMILES string of the drug compound, the amino acid sequence of the target protein, and the structural property sequence of the target protein.

[0027] The data encoding module is used to encode the SMILES string, amino acid sequence, and structural attribute sequence respectively to obtain the corresponding compound digital tensor, protein digital tensor, and structural attribute digital tensor.

[0028] The compound characterization module is used to generate molecular maps and molecular fingerprints of drug compounds based on the SMILES string.

[0029] A multi-vector generation module is used to input the compound digital tensor into the first sequence feature encoding network branch of the pre-trained prediction model to obtain a compound sequence vector; input the molecular graph into the graph message passing network branch of the prediction model to obtain a molecular graph vector; input the molecular fingerprint into the projection network branch of the prediction model to obtain a molecular fingerprint vector; input the SMILES string into the first language analysis model branch of the prediction model to obtain a compound model encoding vector; input the protein digital tensor into the second sequence feature encoding network branch of the prediction model to obtain a protein sequence vector; input the structural attribute digital tensor into the structural attribute encoding network branch of the prediction model to obtain a structural attribute vector; and input the amino acid sequence into the second language analysis model branch of the prediction model to obtain a protein model encoding vector.

[0030] The vector fusion module is used to fuse the compound sequence vector, molecular graph vector, molecular fingerprint vector, and compound model encoding vector on the compound side in the latent space of the prediction model to obtain the compound fusion vector; and to fuse the protein sequence vector, structural attribute vector, and protein model encoding vector on the protein side in the latent space of the prediction model to obtain the protein fusion vector.

[0031] The vector splicing module is used to splice the compound fusion vector with the protein fusion vector to obtain a drug target joint vector;

[0032] The correlation prediction module is used to input the drug target joint vector into the prediction head of the prediction model, which consists of at least one KAN linear layer, and output the predicted probability of drug target interaction or the predicted value of binding affinity.

[0033] Preferably, in the data encoding module, the SMILES string and amino acid sequence are encoded using the BPE vocabulary, and the structural attribute sequence is encoded using a vocabulary constructed from a predefined set of structural attribute lexicons, which includes structural, exposure, polarity, and fragment length attributes.

[0034] Preferably, the data encoding module further includes:

[0035] The length adjustment unit is used to perform fixed-length padding or truncation on compound digital tensors, protein digital tensors, and structural attribute digital tensors. The vocabulary constructed from the BPE vocabulary and the predefined set of structural attribute lexicons contains padding markers for supplementing each digital tensor to a uniform length, as well as unknown markers for sequence attributes that are not present or cannot be identified in the vocabulary.

[0036] Preferably, the compound characteristic extraction module includes:

[0037] The molecular graph generation unit is used to convert the SMILES string into a molecular graph with atoms as nodes and chemical bonds as edges using a cheminformatics toolkit.

[0038] The molecular fingerprint generation unit is used to calculate a fixed-length Morgan circular fingerprint as a molecular fingerprint using a cheminformatics toolkit.

[0039] Preferably, in the correlation prediction module, the prediction head includes multiple KAN linear layers, which achieve nonlinear mapping through sine and cosine basis expansion and learnable Fourier coefficients.

[0040] This application fully utilizes the multi-source characteristics of drugs and proteins to enhance the ability to model complex nonlinear relationships, thereby improving the accuracy and generalization ability in drug target prediction tasks. Attached Figure Description

[0041] Figure 1 This is a prediction model framework diagram of a preferred embodiment of the drug-target correlation prediction method of this application. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are only some, not all, of the embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0043] The first aspect of this application provides a method for predicting the correlation between drugs and targets, such as... Figure 1 As shown, it mainly includes:

[0044] S1. Obtain the data to be predicted, which includes the SMILES string of the drug compound, the amino acid sequence of the target protein, and the structural property sequence of the target protein.

[0045] S2. Encode the SMILES string, amino acid sequence, and structural attribute sequence to obtain the corresponding compound digital tensor, protein digital tensor, and structural attribute digital tensor.

[0046] S3. Based on the SMILES string, generate the molecular map and molecular fingerprint of the drug compound;

[0047] S4. Input the compound digital tensor into the first sequence feature encoding network branch of the pre-trained prediction model to obtain the compound sequence vector; input the molecular graph into the graph message passing network branch of the prediction model to obtain the molecular graph vector; input the molecular fingerprint into the projection network branch of the prediction model to obtain the molecular fingerprint vector; input the SMILES string into the first language analysis model branch of the prediction model to obtain the compound model encoding vector; input the protein digital tensor into the second sequence feature encoding network branch of the prediction model to obtain the protein sequence vector; input the structural attribute digital tensor into the structural attribute encoding network branch of the prediction model to obtain the structural attribute vector; input the amino acid sequence into the second language analysis model branch of the prediction model to obtain the protein model encoding vector.

[0048] S5. The compound sequence vector, molecular graph vector, molecular fingerprint vector, and compound model encoding vector on the compound side are fused in the latent space of the prediction model to obtain the compound fusion vector; the protein sequence vector, structural attribute vector, and protein model encoding vector on the protein side are fused in the latent space of the prediction model to obtain the protein fusion vector.

[0049] S6. Concatenate the compound fusion vector with the protein fusion vector to obtain the drug target joint vector;

[0050] S7. Input the drug target joint vector into the prediction head of the prediction model, which consists of at least one KAN linear layer, and output the predicted probability of drug target interaction or the predicted value of binding affinity.

[0051] This project addresses the challenges in drug-target interaction prediction, such as "multiple feature sources with weak fusion, strong sequence information but insufficient structural information, and limited model expressive power," by proposing a unified multimodal modeling scheme. Instead of simply concatenating multiple features and inputting them into a classifier, this scheme first maps different modalities to a unified latent space, then performs same-dimensional fusion on both the drug and protein sides, and finally performs cross-modal joint prediction.

[0052] In step S1, the data to be predicted, i.e., the data input into the model, is usually in CSV format. (See reference...) Figure 1On the compound side, the input data includes the string "SMILES" A; on the protein side, the input data includes the amino acid sequence B and the structural attribute sequence C. Using the trained prediction model, in step S7, the prediction result is output. The prediction result can be the predicted probability of drug-target interaction or the predicted value of drug-target binding affinity. Correspondingly, during model training, the input data should include the aforementioned supervision labels, i.e., the model's expected output. These include labels for classification tasks, indicating whether there is an interaction between the drug and the target (e.g., a label of 1 indicates interaction, and 0 indicates no interaction); and pka labels for regression tasks, indicating the magnitude of the binding affinity between the drug and the target (a higher pka value generally indicates stronger binding ability). During the training phase, the prediction model compares the prediction results with the true supervision labels. If labels are used, the classification error is calculated; if pka is used, the regression error is calculated. Subsequently, the program performs backpropagation based on the error, automatically adjusting the weight parameters in the embedding layer, sequence encoder, graph message passing network, convolutional network, and prediction head, so that the model's prediction results gradually approach the true labels. The model learns the relationship between input features A / B / C and the expected output label or pka to predict the interaction or affinity of new samples.

[0053] During training, classification tasks can use metrics such as AUC, PRC-AUC, and Accuracy; regression tasks can use metrics such as MSE, RMSE, R2, and CI. Early stopping and optimal model saving are performed based on validation set metrics. The early stopping logic allows training to be terminated prematurely if there is no improvement for a specified number of consecutive rounds. Furthermore, this application can train multiple models. If multiple models are used for prediction, the outputs of the multiple models are averaged to obtain the ensemble prediction result.

[0054] Steps S2-S6 provide the data transfer relationships in the prediction model. In steps S2-S3, multi-source information of the input features is constructed. In step S4, the independent branch features are fused in the same latent space by element-wise addition. In step S5, the compound-side fused representation and the protein-side fused representation are concatenated. In step S6, the concatenated features are input into the prediction head composed of multiple KAN linear layers, and the classification probability or regression value is output. The following is a detailed explanation.

[0055] In step S1, SMILES is a specification that explicitly describes the molecular structure using ASCII strings. The SMILES string A can be generated by processing the compound using conventional methods. Similarly, the amino acid sequence B is the most basic structural information of a protein, composed of 20 common amino acids linked by peptide bonds, and can also be generated using conventional methods. The structural attribute sequence C of the target protein usually refers to its primary and secondary structures, i.e., the linear sequence of amino acids, and can also be obtained through conventional methods.

[0056] Subsequently, in step S2, the above-mentioned input sequence is encoded.

[0057] In some optional implementations, in step S2, the SMILES string and amino acid sequence are encoded using the BPE vocabulary, and the structural attribute sequence is encoded using a vocabulary constructed from a predefined set of structural attribute lexicons, which includes structural, exposure, polarity, and fragment length attributes.

[0058] In this embodiment, the vocabulary constructed from the predefined set of structural attribute lexical units is shown in Table 1.

[0059] Table 1 Examples of word representations constructed from predefined structural attribute lexical sets

[0060] mark coding Structural attribute lexicon set ANKM 2 α-helix, unexposed, basic, medium-length fragment BETS 3 β-sheet, exposure, polarity, short fragment CEDL 4 Curled structure, exposure, acidity, long segments BNGM 5 β-sheet, unexposed, nonpolar, medium-length fragment AEKL 6 α-helix, exposure, alkaline, long fragment

[0061] In some alternative implementations, step S2 further includes:

[0062] Fixed-length padding or truncation is performed on compound digital tensors, protein digital tensors, and structural attribute digital tensors. The vocabulary constructed from the BPE vocabulary and the predefined set of structural attribute lexicons contains padding markers for supplementing each digital tensor to a uniform length, as well as unknown markers for sequence attributes that are not present or cannot be identified in the vocabulary.

[0063] In this embodiment, a completion marker is set. <pad>Length processing is performed, its encoding is 0, and an unknown flag is set. <unk>An exception is handled, coded as 1. The above method can obtain fixed-length tensors. For example, the maximum length of a protein digital tensor is `sequence_length=1500`; the maximum length of a structural attribute digital tensor is `sps_length=128`; and the maximum length of a compound digital tensor is `smiles_length=99`. The program truncates and pads with zeros according to the maximum length. If the maximum length is 100, the last bit of a tensor with a length of 99 is padded with 0s; a tensor with a length of 101 is truncated to 100.

[0064] Step S3 is used to construct multi-source information of the input features, including molecular maps and molecular fingerprints.

[0065] In some alternative implementations, step S3 further includes:

[0066] The SMILES string was converted into a molecular graph with atoms as nodes and chemical bonds as edges using a cheminformatics toolkit.

[0067] A fixed-length circular fingerprint of Morgan was calculated using a cheminformatics toolkit and used as a molecular fingerprint.

[0068] like Figure 1 As shown, this application uses RDKit to automatically convert SMILES into two-dimensional molecular diagrams for use by D-MPNN or KA-DMPNN; simultaneously, it generates a 2048-character Morgan fingerprint for the ECFP branch. RDKit is an open-source cheminformatics toolkit that can be used to process compound structural information, such as reading and parsing compound strings, generating molecular diagrams, calculating molecular fingerprints, and extracting molecular structural features.

[0069] Step S4 is used to process the above multi-source information separately. The processing uses trainable parameters within the prediction model, which are explained below.

[0070] First, multiple feature extraction branches are performed on the compound side.

[0071] First, the compound's numerical tensor is input into the first sequence feature encoding network branch, which is the compound sequence encoder. This encoder first converts each numerical identifier into a vector through an embedding layer, and then adds positional encoding, enabling the model to distinguish the order of characters or fragments in the sequence. Subsequently, the sequence vector is input into the sequence feature encoding network to calculate the contextual features at each position. Finally, the features from all positions are averaged to obtain a fixed-length compound sequence vector, such as... Figure 1 of Branch.

[0072] Second, the program uses the cheminformatics toolkit Rdkit to convert the SMILES string into a two-dimensional molecular graph. This step is a standard cheminformatics processing technique. In the molecular graph, atoms are nodes and chemical bonds are edges. Subsequently, the molecular graph is input into a graph message passing network branch. The model transmits information along the chemical bonds between atoms, gradually aggregating local molecular structural features, and finally obtaining a molecular graph vector, such as... Figure 1 of Branch.

[0073] Third, the program also generates a molecular fingerprint based on the compound structure. A molecular fingerprint can be understood as a set of numerical features describing the presence or absence of local substructures within a molecule. This fingerprint, after linear transformation, normalization, activation function processing, and random deactivation of the projection network branches, is converted into a molecular fingerprint vector with dimensions consistent with other branches, such as... Figure 1 of Branch.

[0074] Fourth, the SMILES string is encoded using the pre-trained language model of the compound, i.e., the first language analysis model branch, to obtain the compound model encoding vector, such as... Figure 1 of Branch.

[0075] Then, multiple feature extraction branches on the protein side are performed.

[0076] First, the protein digital tensor is input into the second sequence feature encoding network branch, which is the protein sequence encoder. Its processing method is similar to that of the compound sequence: it is first converted into a vector through an embedding layer, then positional encoding is added, then it is input into the sequence feature encoding network, and finally, average pooling is used to obtain the protein sequence vector, such as... Figure 1 of Branch.

[0077] Second, structural attribute digital tensors are input into the structural attribute encoding network branch for structural attribute convolution computation. The model first converts each structural attribute label into a vector through an embedding layer, then extracts local patterns between adjacent structural attribute segments through a one-dimensional convolutional layer. The convolution results are then subjected to gated activation and residual connections to enhance effective structural information and reduce information loss during training. Finally, the convolutional features are unfolded and input into a linear layer to obtain the structural attribute vector, such as... Figure 1 of Branch.

[0078] Third, the protein pre-trained language model, i.e., the second language analysis model branch, is used to encode the amino acid sequence to obtain the protein model encoding vector, such as... Figure 1 of Branch.

[0079] Subsequently, in step S5, vector fusion is performed on the compound side and the protein side respectively, wherein the compound fusion vector = compound sequence vector + molecular graph vector + molecular fingerprint vector + compound model encoding vector; the protein fusion vector = protein sequence vector + structural attribute vector + protein model encoding vector.

[0080] Next, in step S6, the two fused vectors are concatenated. Finally, in step S7, the joint vector is input into the prediction head, and the prediction result is output.

[0081] In some alternative implementations, in step S7, the prediction head includes multiple KAN linear layers, which achieve nonlinear mapping through sine and cosine basis expansion and learnable Fourier coefficients.

[0082] The prediction head in this application is not a typical multi-layer fully connected network, but rather employs a nonlinear mapping layer based on function expansion. This layer performs more flexible nonlinear computations on the input vector through learnable weights and Fourier-form function coefficients. The dimensionality change of the prediction head is approximately as follows:

[0083] Drug-target joint vector - first prediction layer - second prediction layer - third prediction layer - output layer - prediction result. For classification tasks, the output represents the probability of drug-target interaction; for regression tasks, the output represents the predicted binding affinity.

[0084] Comparative analysis revealed that this application demonstrates advantages over various comparative methods in metrics such as AUROC, AUPRC, ACC, Recall, and Precision. Taking the results recorded in model_compare_results.csv as an example, the AUROC of our method, LKAU-DTI, is 0.9121, AUPRC is 0.9319, ACC is 0.8383, and Precision is 0.8724. Compared to models such as TransformerCPI, HyperAttentionDTI, GraphDTA, Moltrans, and PerceiverCPI, this application improves discriminative ability through multimodal fusion and KAN nonlinear mapping. Specifically, LKAU-DTI improves the AUROC of GraphDTA from 0.9065 to 0.9121 and the AUROC of Moltrans from 0.8525 to 0.9121, demonstrating performance advantages in complex drug target prediction tasks.

[0085] Compared with the prior art, this application has at least the following beneficial effects:

[0086] Multimodal modeling is more comprehensive. On the compound side, sequence, graph structure, fingerprint, and optional pre-trained semantics are utilized simultaneously; on the protein side, sequence and structural attribute information (SPS) are utilized simultaneously. Compared with existing technologies, this application additionally introduces molecular graph and fingerprint information, which can capture the topological structure and local substructure features of compounds. At the same time, the additional introduction of structural attribute sequences can supplement protein structure-level information in the absence of a complete three-dimensional structure.

[0087] The fusion method is simple and effective. By adding elements one by one in the unified latent space, low-redundancy fusion of independent information from multiple branches is achieved, reducing the risk of parameter inflation and overfitting.

[0088] The predictor head's expressive power is enhanced. By introducing a KAN linear layer to replace the traditional fully connected layer, the ability to model complex nonlinear relationships is improved.

[0089] Enhanced graph representation capabilities. By introducing KAN into the message-passing network through KA-DMPNN, molecular graph feature extraction becomes more flexible, enabling the model to learn more flexible nonlinear function mappings through Fourier basis expansion. Compared to ordinary D-MPNN, this application also enhances molecular graph encoding capabilities by introducing a linear KAN layer into the message-passing process through KA-DMPNN.

[0090] Highly versatile. Provides unified support for DTI classification and DTA regression tasks.

[0091] Highly scalable. Modules such as SPS, D-MPNN, KA-DMPNN, ECFP, and PLM can be enabled or disabled as needed, facilitating ablation experiments and scenario adaptation.

[0092] It is highly practical for engineering implementation. It is implemented based on the PyTorch and Chemprop frameworks, and has a complete training, verification, testing, and inference process.

[0093] This application proposes a multimodal deep learning method for DTI / DTA tasks. By jointly modeling compound sequences, molecular graphs, molecular fingerprints, protein sequences, and SPS structural properties, and further combining the KAN prediction head with the optional KA-DMPNN, it achieves efficient prediction of drug-target interactions and affinity.

[0094] A second aspect of this application provides a drug-target correlation prediction device corresponding to the above method, mainly comprising:

[0095] The data acquisition module is used to acquire the data to be predicted, which includes the SMILES string of the drug compound, the amino acid sequence of the target protein, and the structural property sequence of the target protein.

[0096] The data encoding module is used to encode the SMILES string, amino acid sequence, and structural attribute sequence respectively to obtain the corresponding compound digital tensor, protein digital tensor, and structural attribute digital tensor.

[0097] The compound characterization module is used to generate molecular maps and molecular fingerprints of drug compounds based on the SMILES string.

[0098] A multi-vector generation module is used to input the compound digital tensor into the first sequence feature encoding network branch of the pre-trained prediction model to obtain a compound sequence vector; input the molecular graph into the graph message passing network branch of the prediction model to obtain a molecular graph vector; input the molecular fingerprint into the projection network branch of the prediction model to obtain a molecular fingerprint vector; input the SMILES string into the first language analysis model branch of the prediction model to obtain a compound model encoding vector; input the protein digital tensor into the second sequence feature encoding network branch of the prediction model to obtain a protein sequence vector; input the structural attribute digital tensor into the structural attribute encoding network branch of the prediction model to obtain a structural attribute vector; and input the amino acid sequence into the second language analysis model branch of the prediction model to obtain a protein model encoding vector.

[0099] The vector fusion module is used to fuse the compound sequence vector, molecular graph vector, molecular fingerprint vector, and compound model encoding vector on the compound side in the latent space of the prediction model to obtain the compound fusion vector; and to fuse the protein sequence vector, structural attribute vector, and protein model encoding vector on the protein side in the latent space of the prediction model to obtain the protein fusion vector.

[0100] The vector splicing module is used to splice the compound fusion vector with the protein fusion vector to obtain a drug target joint vector;

[0101] The correlation prediction module is used to input the drug target joint vector into the prediction head of the prediction model, which consists of at least one KAN linear layer, and output the predicted probability of drug target interaction or the predicted value of binding affinity.

[0102] In some optional implementations, in the data encoding module, the SMILES string and amino acid sequence are encoded using a BPE vocabulary, and the structural attribute sequence is encoded using a vocabulary constructed from a predefined set of structural attribute lexicons, which includes structural, exposure, polarity, and fragment length attributes.

[0103] In some optional implementations, the data encoding module further includes:

[0104] The length adjustment unit is used to perform fixed-length padding or truncation on compound digital tensors, protein digital tensors, and structural attribute digital tensors. The vocabulary constructed from the BPE vocabulary and the predefined set of structural attribute lexicons contains padding markers for supplementing each digital tensor to a uniform length, as well as unknown markers for sequence attributes that are not present or cannot be identified in the vocabulary.

[0105] In some optional embodiments, the compound characterization module includes:

[0106] The molecular graph generation unit is used to convert the SMILES string into a molecular graph with atoms as nodes and chemical bonds as edges using a cheminformatics toolkit.

[0107] The molecular fingerprint generation unit is used to calculate a fixed-length Morgan circular fingerprint as a molecular fingerprint using a cheminformatics toolkit.

[0108] In some alternative implementations, in the correlation prediction module, the prediction head includes multiple KAN linear layers, which achieve nonlinear mapping through sine and cosine basis expansion and learnable Fourier coefficients.

[0109] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.< / unk> < / pad>

Claims

1. A method for predicting the correlation between a drug and a target, characterized in that, include: S1. Obtain the data to be predicted, which includes the SMILES string of the drug compound, the amino acid sequence of the target protein, and the structural property sequence of the target protein. S2. Encode the SMILES string, amino acid sequence, and structural attribute sequence to obtain the corresponding compound digital tensor, protein digital tensor, and structural attribute digital tensor. S3. Based on the SMILES string, generate the molecular map and molecular fingerprint of the drug compound; S4. Input the compound digital tensor into the first sequence feature encoding network branch of the pre-trained prediction model to obtain the compound sequence vector; input the molecular graph into the graph message passing network branch of the prediction model to obtain the molecular graph vector; input the molecular fingerprint into the projection network branch of the prediction model to obtain the molecular fingerprint vector; input the SMILES string into the first language analysis model branch of the prediction model to obtain the compound model encoding vector; input the protein digital tensor into the second sequence feature encoding network branch of the prediction model to obtain the protein sequence vector; input the structural attribute digital tensor into the structural attribute encoding network branch of the prediction model to obtain the structural attribute vector; input the amino acid sequence into the second language analysis model branch of the prediction model to obtain the protein model encoding vector. S5. The compound sequence vector, molecular graph vector, molecular fingerprint vector, and compound model encoding vector on the compound side are fused in the latent space of the prediction model to obtain the compound fusion vector; the protein sequence vector, structural attribute vector, and protein model encoding vector on the protein side are fused in the latent space of the prediction model to obtain the protein fusion vector. S6. Concatenate the compound fusion vector with the protein fusion vector to obtain the drug target joint vector; S7. Input the drug target joint vector into the prediction head of the prediction model, which consists of at least one KAN linear layer, and output the predicted probability of drug target interaction or the predicted value of binding affinity.

2. The method for predicting drug-target correlation according to claim 1, characterized in that, In step S2, the SMILES string and amino acid sequence are encoded using the BPE vocabulary, and the structural attribute sequence is encoded using a vocabulary constructed from a predefined set of structural attribute lexicons, which includes structural, exposure, polarity, and fragment length attributes.

3. The method for predicting drug-target correlation according to claim 2, characterized in that, Step S2 further includes: Fixed-length padding or truncation is performed on compound digital tensors, protein digital tensors, and structural attribute digital tensors. The vocabulary constructed from the BPE vocabulary and the predefined set of structural attribute lexicons contains padding markers for supplementing each digital tensor to a uniform length, as well as unknown markers for sequence attributes that are not present or cannot be identified in the vocabulary.

4. The method for predicting drug-target correlation according to claim 1, characterized in that, Step S3 further includes: The SMILES string was converted into a molecular graph with atoms as nodes and chemical bonds as edges using a cheminformatics toolkit. A fixed-length circular fingerprint of Morgan was calculated using a cheminformatics toolkit and used as a molecular fingerprint.

5. The method for predicting drug-target correlation according to claim 1, characterized in that, In step S7, the prediction head includes multiple KAN linear layers, which achieve nonlinear mapping through sine and cosine basis expansion and learnable Fourier coefficients.

6. A device for predicting the correlation between a drug and a target, characterized in that, include: The data acquisition module is used to acquire the data to be predicted, which includes the SMILES string of the drug compound, the amino acid sequence of the target protein, and the structural property sequence of the target protein. The data encoding module is used to encode the SMILES string, amino acid sequence, and structural attribute sequence respectively to obtain the corresponding compound digital tensor, protein digital tensor, and structural attribute digital tensor. The compound characterization module is used to generate molecular maps and molecular fingerprints of drug compounds based on the SMILES string. A multi-vector generation module is used to input the compound digital tensor into the first sequence feature encoding network branch of the pre-trained prediction model to obtain a compound sequence vector; input the molecular graph into the graph message passing network branch of the prediction model to obtain a molecular graph vector; input the molecular fingerprint into the projection network branch of the prediction model to obtain a molecular fingerprint vector; input the SMILES string into the first language analysis model branch of the prediction model to obtain a compound model encoding vector; input the protein digital tensor into the second sequence feature encoding network branch of the prediction model to obtain a protein sequence vector; input the structural attribute digital tensor into the structural attribute encoding network branch of the prediction model to obtain a structural attribute vector; and input the amino acid sequence into the second language analysis model branch of the prediction model to obtain a protein model encoding vector. The vector fusion module is used to fuse the compound sequence vector, molecular graph vector, molecular fingerprint vector, and compound model encoding vector on the compound side in the latent space of the prediction model to obtain the compound fusion vector; and to fuse the protein sequence vector, structural attribute vector, and protein model encoding vector on the protein side in the latent space of the prediction model to obtain the protein fusion vector. The vector splicing module is used to splice the compound fusion vector with the protein fusion vector to obtain a drug target joint vector; The correlation prediction module is used to input the drug target joint vector into the prediction head of the prediction model, which consists of at least one KAN linear layer, and output the predicted probability of drug target interaction or the predicted value of binding affinity.

7. The drug-target correlation prediction device according to claim 6, characterized in that, In the data encoding module, the SMILES string and amino acid sequence are encoded using the BPE vocabulary, and the structural attribute sequence is encoded using a vocabulary constructed from a predefined set of structural attribute lexicons, which includes structural, exposure, polarity, and fragment length attributes.

8. The drug-target correlation prediction device according to claim 7, characterized in that, The data encoding module further includes: The length adjustment unit is used to perform fixed-length padding or truncation on compound digital tensors, protein digital tensors, and structural attribute digital tensors. The vocabulary constructed from the BPE vocabulary and the predefined set of structural attribute lexicons contains padding markers for supplementing each digital tensor to a uniform length, as well as unknown markers for sequence attributes that are not present or cannot be identified in the vocabulary.

9. The drug-target correlation prediction device according to claim 6, characterized in that, The compound characteristic extraction module includes: The molecular graph generation unit is used to convert the SMILES string into a molecular graph with atoms as nodes and chemical bonds as edges using a cheminformatics toolkit. The molecular fingerprint generation unit is used to calculate a fixed-length Morgan circular fingerprint as a molecular fingerprint using a cheminformatics toolkit.

10. The drug-target correlation prediction device according to claim 6, characterized in that, In the correlation prediction module, the prediction head includes multiple KAN linear layers, which achieve nonlinear mapping through sine and cosine basis expansion and learnable Fourier coefficients.

Images