A drug target prediction method based on heterogeneous multi-encoder

Abstract

The application discloses a drug target prediction method based on a heterogeneous multi-encoder, and relates to the technical fields of bioinformatics and artificial intelligence, and comprises the following steps: obtaining multiple modal data of a drug molecule and multiple modal data of a target protein; performing feature extraction on the multiple modal data of the drug molecule through multiple parallel drug encoders respectively to obtain multiple drug feature vectors; performing feature extraction on the multiple modal data of the target protein through multiple parallel protein encoders respectively to obtain multiple protein feature vectors; performing feature splicing on the multiple drug feature vectors and the multiple protein feature vectors through an attention mechanism to obtain a fusion feature vector, and inputting the fusion feature vector into a classifier to obtain a prediction result of an activation-inhibition interaction. The application solves the problem of insufficient prediction accuracy caused by limited single-modal feature representation capability and simple multi-modal fusion strategy in the prior art, and realizes accurate prediction of the activation-inhibition interaction.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and artificial intelligence, and in particular to a drug target prediction method based on heterogeneous multi-encoders. Background Technology

[0002] Predicting activation-inhibition interactions and accurately determining the activating or inhibiting effects of compounds on proteins are core tasks in pharmacology, therapeutics development, and systems biology. With advancements in sequencing technology and cheminformatics, the accessibility of biological and chemical data continues to expand, and computational methods are playing an increasingly important role in inferring mechanistic activation-inhibition relationships.

[0003] However, most current models rely on single-modal feature representations, making it difficult to fully capture the complex biochemical properties inherent in drug structures and protein sequences. More importantly, activation and inhibition mechanisms often depend on subtle patterns across multiple feature domains, which makes related prediction tasks particularly challenging.

[0004] Current computational methods still have limitations in capturing these complex mechanisms. Early approaches typically employed traditional machine learning algorithms, such as Support Vector Machines and Random Forests. These methods rely on manually designed feature descriptors and often struggle to generalize to novel molecular structures. With the development of deep learning, models like DeepDTA and GraphDTA have emerged, automatically extracting features using convolutional neural networks or graph neural networks. However, these methods still rely on single-viewpoint representations. While some recent research has attempted to integrate multimodal data, it primarily employs simple fusion strategies, such as linear concatenation of feature vectors. This late-stage fusion strategy processes features from different perspectives independently before the final classification layer, failing to effectively model the intrinsic cross-modal interactions between drug substructures and protein domains.

[0005] Therefore, designing a prediction method that can effectively integrate the multimodal features of drug molecules and target proteins and capture complex cross-modal interactions is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to provide a drug target prediction method based on heterogeneous multi-encoders, so as to solve the problem of insufficient prediction accuracy caused by the limited single-modal feature representation capability and the simple multi-modal fusion strategy in the current technology, and to achieve accurate prediction of activation-inhibition interaction.

[0007] To achieve the above objectives, this invention provides a drug target prediction method based on heterogeneous multi-encoders, comprising the following steps: Acquire multimodal data of drug molecules and multimodal data of target proteins; Multiple parallel drug encoders are used to extract features from various modal data of drug molecules to obtain multiple drug feature vectors; the multiple drug feature vectors include: SMILES feature vector, molecular graph feature vector, fingerprint feature vector, and physicochemical property feature vector; Multiple protein feature vectors are obtained by extracting features from the multimodal data of the target protein using multiple parallel protein encoders; the multiple protein feature vectors include: protein fusion feature vector and triplet frequency vector. By using an attention mechanism, multiple drug feature vectors and multiple protein feature vectors are concatenated to obtain a fused feature vector. The fused feature vector is input into the classifier to obtain the prediction results of the interaction between activation and inhibition.

[0008] Preferably, the modal data of the drug molecule includes at least two of the following: SMILES sequence, molecular graph structure, molecular fingerprint, and physicochemical property descriptor; The plurality of parallel drug encoders include at least two of the following: a drug SMILES recurrent neural network encoder for extracting SMILES sequence features, a graph neural network encoder for extracting molecular graph features, a first multilayer perceptron encoder for extracting molecular fingerprint features, and a second multilayer perceptron encoder for extracting physicochemical property descriptors. The molecular fingerprint is a spliced ​​vector of MACCS fingerprint, ErG fingerprint, PubChem fingerprint, and ECFP2 fingerprint; the physicochemical property descriptors include: molecular weight, octanol / water partition coefficient, number of hydrogen bond donors, number of hydrogen bond acceptors, number of rotatable bonds, number of aromatic rings, topological polar surface area, number of heavy atoms, total number of atoms, total number of rings, and sp... 3 The number of hybridized carbon atoms, the number of aliphatic rings, the number of saturated rings, the number of heteroatoms, the number of aromatic heterocycles, the number of saturated heterocycles, the number of aliphatic heterocycles, and the Labute approximate surface area.

[0009] Preferably, the multiple modal data of the target protein include at least two of the following: one-hot coding matrix of amino acid sequence, integer index sequence of amino acid sequence, and triplet frequency vector; The plurality of parallel protein encoders include at least two of the following: a convolutional neural network encoder for extracting features of a one-hot encoded matrix, a protein sequence recurrent neural network encoder for extracting features of an integer index sequence, and a third multilayer perceptron encoder for extracting the frequency vector of a triplet.

[0010] Preferably, feature extraction of multiple modalities of the target protein is performed using multiple parallel protein encoders to obtain multiple protein feature vectors, the specific contents of which include: The first intermediate feature vector is obtained by extracting features from multiple modal data of the target protein using a convolutional neural network encoder. The second intermediate feature vector is obtained by extracting features from multiple modal data of the target protein using a recurrent neural network encoder. The first intermediate feature vector and the second intermediate feature vector are fused to obtain the protein fusion feature vector; The triplet frequency vector is obtained by extracting features from the multimodal data of the target protein using a third multilayer perceptron encoder. Protein fusion feature vectors and triplet frequency vectors are essentially multiple protein feature vectors.

[0011] Preferably, the method for generating the triplet frequency vector includes: classifying 20 standard amino acids according to their physicochemical properties into hydrophilic amino acids, hydrophobic amino acids, polar amino acids, aromatic amino acids, basic amino acids, acidic amino acids, and sulfur-containing amino acids; performing frequency statistics on all triplets composed of three consecutive amino acid residues in the protein sequence to obtain a 343-dimensional original vector; and performing normalization processing to generate the triplet frequency vector.

[0012] Preferably, the fused feature vector is obtained by concatenating multiple drug feature vectors and multiple protein feature vectors through an attention mechanism, specifically including: The SMILES feature vector, molecular graph feature vector, fingerprint feature vector, physicochemical property feature vector, protein fusion feature vector, and triplet frequency vector are arranged in a preset order and sequentially pass through a cross-modal feature enhancement module to obtain multiple enhanced feature vectors. The cross-attention module uses the feature vector of the current modality as the query vector and the feature vector of the next modality as the key vector and value vector. After multi-head attention calculation, residual connection, and layer normalization, it outputs the enhanced feature vector of the current modality, and the enhanced feature vector of the current modality has the same dimension as the feature vector of the current modality. Multiple enhanced feature vectors are concatenated along the sequence dimension to obtain a concatenated feature sequence; The concatenated feature sequence is input into the multi-head self-attention module to extract high-order interaction features between modalities. The extracted results are then processed by a convolutional layer and flattened to obtain a fused feature vector.

[0013] Preferably, inputting the fused feature vector into the classifier to obtain the prediction result of the interaction between activation and inhibition specifically includes: The fused feature vector is input into a multilayer perceptron classifier, and the output layer of the multilayer perceptron classifier outputs a scalar logit. The scalar logit is converted into a probability value between 0 and 1 using the Sigmoid function; The judgment is based on a probability value of 0.5: if the probability value is greater than 0.5, the prediction result is an activation relationship; if the probability value is less than or equal to 0.5, the prediction result is an inhibition relationship.

[0014] Preferably, the drug target prediction method based on heterogeneous multi-encoders further includes: Obtain a labeled training dataset; each sample in the training dataset includes: multiple modal data of drug molecules, multiple modal data of target proteins, and corresponding interaction labels; The training dataset is input into multiple parallel drug encoders, multiple parallel protein encoders, and a multilayer perceptron classifier. The trainable parameters of these multiple parallel drug encoders, multiple parallel protein encoders, and a multilayer perceptron classifier are optimized with the goal of minimizing the binary cross-entropy loss function.

[0015] Preferably, the multiple modal data of the drug molecule include: SMILES sequence, molecular graph structure, molecular fingerprint, and physicochemical property descriptor; the multiple parallel drug encoders include: drug SMILES recurrent neural network encoder, graph neural network encoder, first multilayer perceptron encoder, and second multilayer perceptron encoder.

[0016] Preferably, the multiple modal data of the target protein include: a one-hot encoding matrix of amino acid sequences, an integer index sequence of amino acid sequences, and a triplet frequency vector; the multiple parallel protein encoders include: a convolutional neural network encoder, a protein sequence recurrent neural network encoder, and a third multilayer perceptron encoder.

[0017] In summary, the drug target prediction method based on heterogeneous multi-encoders provided by this invention has the following advantages compared with traditional technologies: By constructing multiple parallel encoders to extract multiple feature vectors of drugs and proteins respectively, and adopting an optimized feature fusion strategy based on attention mechanism, accurate prediction of activation-inhibition interactions is achieved; by using multiple parallel drug encoders to extract multiple drug feature vectors of drug molecules and multiple parallel protein encoders to extract multiple protein feature vectors of target proteins, the complex biochemical characteristics of drugs and proteins can be comprehensively captured, overcoming the problem of incomplete single-modal characterization information; through feature fusion based on attention mechanism, the intrinsic cross-modal interactions between drug substructures and protein domains can be effectively fused, significantly improving prediction accuracy compared with the simple feature splicing in traditional technologies.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1This is a flowchart of a drug target prediction method based on heterogeneous multi-encoders in this invention; Figure 2 This is a diagram illustrating the architecture for extracting multiple drug feature vectors in this invention. Figure 3 This is a diagram illustrating the architecture for extracting multiple protein feature vectors in this invention. Figure 4 This is a diagram illustrating the feature splicing architecture in this invention. Figure 5 This is a diagram illustrating the predictive architecture for the activation-inhibition interaction in this invention. Figure 6 This is a diagram showing the ablation experiment results in an embodiment of the present invention; Figure 6 (a) in the figure is a bar chart of the ablation experiment results; Figure 6 (b) in the figure is a specific representation of the view selection for each combination in the ablation experiment. Detailed Implementation

[0020] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0021] The following description of at least one exemplary embodiment is merely illustrative and is not intended to limit the scope of this application or its application or use.

[0022] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.

[0023] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0025] This invention provides a drug target prediction method based on heterogeneous multi-encoders, such as... Figure 1 As shown, it includes the following steps: S1. Obtain multiple modal data of the drug molecule and multiple modal data of the target protein. The multiple modal data of the drug molecule includes at least two of the following: SMILES sequences, molecular graph structures, molecular fingerprints, and physicochemical property descriptors. The multiple modal data of the target protein includes at least two of the following: one-hot coding matrices of amino acid sequences, integer index sequences of amino acid sequences, and triplet frequency vectors.

[0026] The molecular fingerprint is a spliced ​​vector of the MACCS fingerprint, ErG fingerprint, PubChem fingerprint, and ECFP2 fingerprint. Physicochemical property descriptors include: molecular weight, octanol / water partition coefficient, number of hydrogen bond donors, number of hydrogen bond acceptors, number of rotatable bonds, number of aromatic rings, topological polar surface area, number of heavy atoms, total number of atoms, total number of rings, and sp... 3 The number of hybridized carbon atoms, the number of aliphatic rings, the number of saturated rings, the number of heteroatoms, the number of aromatic heterocycles, the number of saturated heterocycles, the number of aliphatic heterocycles, and the Labute approximate surface area.

[0027] The method for generating the triplet frequency vector includes: classifying 20 standard amino acids into seven categories according to their physicochemical properties (hydrophilic amino acids, hydrophobic amino acids, polar amino acids, aromatic amino acids, basic amino acids, acidic amino acids, and sulfur-containing amino acids); performing frequency statistics on all triplets composed of three consecutive amino acid residues in the protein sequence to obtain a 343-dimensional original vector; and then normalizing the vector to generate the triplet frequency vector.

[0028] S2. Multiple parallel drug encoders are used to extract features from various modalities of drug molecules, resulting in multiple drug feature vectors. The architecture for extracting multiple drug feature vectors is shown in the diagram below. Figure 2 As shown. The multiple drug feature vectors include: SMILES feature vector, molecular graph feature vector, fingerprint feature vector, and physicochemical property feature vector. The multiple parallel drug encoders include at least two of the following: a drug SMILES recurrent neural network encoder for extracting SMILES sequence features, a graph neural network encoder for extracting molecular graph features, a first multilayer perceptron encoder for extracting molecular fingerprint features, and a second multilayer perceptron encoder for extracting physicochemical property descriptors. Figure 2 In h graph The feature vectors are generated after extraction by the molecular graph neural network. h fp The feature vector is generated after fingerprint extraction of the molecular formula. h pc This is the feature vector generated after extracting the physicochemical properties of the molecular formula. h seq The feature vector is generated after the molecular formula is extracted by a recurrent neural network.

[0029] S3. Multiple protein encoders are used to extract features from the multimodal data of the target protein to obtain multiple protein feature vectors.

[0030] Specifically, step S3 includes the following steps: S301. Feature extraction is performed on the multimodal data of the target protein using a convolutional neural network encoder to obtain the first intermediate feature vector.

[0031] S302. Feature extraction is performed on the multimodal data of the target protein using a recurrent neural network encoder to obtain a second intermediate feature vector.

[0032] S303. Merge the first intermediate feature vector and the second intermediate feature vector to obtain the protein fusion feature vector.

[0033] S304. Feature extraction is performed on the multimodal data of the target protein using the third multilayer perceptron encoder to obtain the triplet frequency vector.

[0034] S305, protein fusion feature vector, and triplet frequency vector are multiple protein feature vectors.

[0035] An architecture diagram for extracting multiple protein feature vectors is shown below. Figure 3 As shown. The multiple protein feature vectors include: protein feature vectors and triplet frequency vectors. The multiple parallel protein encoders include at least two of the following: a convolutional neural network encoder for extracting features from the one-hot encoding matrix, a protein sequence recurrent neural network encoder for extracting features from the integer index sequence, and a third multilayer perceptron encoder for extracting triplet frequency vectors. Figure 3 middle h proteinseq The feature vector is generated after extracting protein sequences using recurrent neural networks and convolutional neural networks. h ctf This is a feature vector generated after extracting the protein sequence using a combined triplet fingerprint.

[0036] In an exemplary embodiment of this invention, drug-target relationship data from the DrugBank public database was used. A total of 11,229 relationships were obtained, of which 3,322 were activation relationships and 7,907 were inhibition relationships. These were divided into training, validation, and test sets in a 7:2:1 ratio.

[0037] For each drug molecule's SMILES string, the RDKit toolkit was used to parse the molecular structure and generate four complementary modes.

[0038] Multiple modal data of drug molecules include: SMILES sequences, molecular graph structures, molecular fingerprints, and physicochemical property descriptors.

[0039] SMILES sequence: Splits the SMILES string into sub-units by character, maps them to integer indices in a predefined vocabulary, and aligns them to a uniform length using padding.

[0040] Molecular Graph: Molecules are converted into graph structures using DGL (Deep Graph Library). Node features include atom type, degree, charge, chirality, number of hydrogen atoms, hybridization type, aromaticity, and atomic mass; edge features are bond types.

[0041] Molecular fingerprinting: By splicing MACCS fingerprint, ErG fingerprint, PubChem fingerprint and ECFP2 fingerprint, a 2513-dimensional binary vector is obtained.

[0042] Physicochemical property descriptors: Calculate 18 RDKit descriptors, including molecular weight, octanol / water partition coefficient, number of hydrogen bond donors, number of hydrogen bond acceptors, number of rotatable bonds, number of aromatic rings, topological polar surface area, number of heavy atoms, total number of atoms, total number of rings, and sp... 3 Hybridized carbon atom ratio, number of aliphatic rings, number of saturated rings, number of heteroatoms, number of aromatic heterocycles, number of saturated heterocycles, number of aliphatic heterocycles, and Labute approximate surface area.

[0043] Multiple modal data of the target protein include: one-hot coding matrix of amino acid sequence, integer index sequence of amino acid sequence, and triplet frequency vector.

[0044] For each protein amino acid sequence, three encoding methods are used: The input (one-hot encoding matrix) of the Convolutional Neural Network (CNN) is: the sequence is one-hot encoded according to 20 common amino acids to obtain an L×20 matrix, where L is the sequence length.

[0045] The input (integer index sequence) of the Recurrent Neural Network (RNN) is: mapping amino acids to integer indices from 1 to 20, padding positions with less than 3000 with 0, and recording the actual length of the sequence.

[0046] Triplet Frequencies (CTF): Twenty standard amino acids are divided into seven categories according to their physicochemical properties. The frequency of occurrence of all consecutive tripeptides in the sequence is counted and normalized to obtain a 343-dimensional vector.

[0047] Multiple parallel drug encoders include: a drug SMILES recurrent neural network encoder, a graph neural network encoder, a first multilayer perceptron encoder, and a second multilayer perceptron encoder.

[0048] Drug SMILES recurrent neural network encoder: The embedding layer is combined with a bidirectional long short-term memory (LSTM) network, which concatenates the hidden states at the last time step of the forward and backward passes, and outputs a 384-dimensional feature vector through a fully connected layer and LayerNorm.

[0049] Graph Neural Network Encoder: 2-layer Neural Network Convolution (NNConv), activated by the ReLU function, then the global graph representation is read out using Set2Set, and then passed through a fully connected layer and LayerNorm to output a 384-dimensional feature vector.

[0050] The first multilayer perceptron encoder: Multilayer Perceptron (MLP), outputs a 384-dimensional feature vector.

[0051] The second multilayer perceptron encoder is a three-layer MLP, each with BatchNorm and Dropout, containing residual connections, and outputting a 384-dimensional feature vector.

[0052] Multiple parallel protein encoders include: a convolutional neural network encoder, a protein sequence recurrent neural network encoder, and a third multilayer perceptron encoder.

[0053] CNN encoder: Input L×20 one-hot matrix, after 1D convolution and pooling operations, fully connected to 343 dimensions, and then mapped to 128 dimensions through linear layers (as intermediate features).

[0054] Protein sequence RNN encoder: The embedding layer is combined with a bidirectional LSTM, and the hidden states at the last time step of the forward and backward directions are spliced ​​together and fully connected to 128 dimensions (as intermediate features).

[0055] After concatenating the two 128-dimensional intermediate features, the result is mapped to 384 dimensions through a linear layer, yielding the fused protein feature vector.

[0056] The third multilayer perceptron encoder (triple frequency encoder): a two-layer MLP, each layer with BatchNorm and Dropout, inputting a 343-dimensional triple frequency vector and outputting a 384-dimensional triple frequency vector.

[0057] S4. Using an attention mechanism, multiple drug feature vectors and multiple protein feature vectors are concatenated to obtain a fused feature vector. The architecture diagram for feature concatenation is shown below. Figure 4 As shown. Figure 4 middle The feature vectors generated after extraction by the molecular graph neural network hgraph The vector enhanced by the cross-attention mechanism The feature vector generated after fingerprint extraction of the molecular formula h fp The vector enhanced by the cross-attention mechanism The feature vector generated after extracting the physicochemical properties of the molecular formula h pc The vector enhanced by the cross-attention mechanism The feature vector generated after extracting the molecular formula using a recurrent neural network. h seq The vector enhanced by the cross-attention mechanism The feature vectors generated after protein sequences are extracted using recurrent neural networks and convolutional neural networks. h proteinseq The vector enhanced by the cross-attention mechanism.

[0058] Furthermore, step S4 specifically includes the following steps: S401. Arrange the SMILES feature vector, molecular graph feature vector, fingerprint feature vector, physicochemical property feature vector, protein feature vector, and triplet frequency vector in a preset order, and sequentially perform cross-modal feature enhancement through the cross-attention module to obtain multiple enhanced feature vectors. Specifically, the cross-attention module uses the feature vector of the current modality as the query vector and the feature vector of the next modality as the key and value vectors. After multi-head attention calculation, residual connection, and layer normalization, it outputs the enhanced feature vector of the current modality, and the enhanced feature vector of the current modality has the same dimension as the feature vector of the current modality.

[0059] S402. Concatenate multiple enhanced feature vectors along the sequence dimension to obtain a concatenated feature sequence.

[0060] S403. Input the concatenated feature sequence into the multi-head self-attention module to extract high-order interaction features between each modality. Process the extracted results through a convolutional layer and flatten them to obtain the fused feature vector.

[0061] In an exemplary embodiment of the present invention, the feature concatenation of multiple drug feature vectors and multiple protein feature vectors through an attention mechanism specifically includes: Cross-attention: The SMILES feature vector, molecular graph feature vector, fingerprint feature vector, physicochemical property feature vector, protein feature vector, and triplet frequency vector are arranged sequentially and then passed through a cross-attention module for cross-modal feature enhancement, resulting in multiple enhanced feature vectors. Each cross-attention module includes multi-head attention, residual connections, and LayerNorm. The multiple enhanced feature vectors still maintain a 384-dimensional dimension.

[0062] Multi-head self-attention and convolution fusion: Multiple (6) enhanced feature vectors are concatenated in the sequence dimension to obtain a 6×384 concatenated feature sequence, which is then input into the multi-head self-attention module and then extracted through a 2D convolutional layer to obtain high-order interactive features. After flattening, the fused feature vector is obtained.

[0063] S5. Input the fused feature vector into the classifier to obtain the prediction result of the interaction between activation and inhibition. The prediction architecture diagram is as follows. Figure 5 As shown.

[0064] Furthermore, step S5 is replaced by the following steps: S501. Input the fused feature vector into the multilayer perceptron classifier. The output layer of the multilayer perceptron classifier outputs a scalar logit.

[0065] S502. Convert the scalar logit to a probability value between 0 and 1 using the Sigmoid function.

[0066] S503. Use a probability value of 0.5 as a threshold for judgment: if the probability value is greater than 0.5, the prediction result is an activation relationship; if the probability value is less than or equal to 0.5, the prediction result is an inhibition relationship.

[0067] In an exemplary embodiment of the present invention, the drug target prediction method based on heterogeneous multi-encoders further includes: Obtain a labeled training dataset. Each sample in the training dataset includes: multiple modalities of drug molecules, multiple modalities of target proteins, and corresponding interaction labels.

[0068] The training dataset is input into multiple parallel drug encoders, multiple parallel protein encoders, and a multilayer perceptron classifier. The trainable parameters of these multiple parallel drug encoders, multiple parallel protein encoders, and a multilayer perceptron classifier are optimized with the goal of minimizing the binary cross-entropy loss function.

[0069] Using the strategy of this invention, K-nearest neighbor algorithm, support vector machine, deep neural network and random forest iterative five times, the mean and standard deviation of accuracy, precision, recall, F1 score (obtained by harmonic average of precision and recall), area under the receiver operating characteristic curve (AUROC), and area under the precision-recall curve (AUPR) of each method were calculated. The specific data are shown in Tables 1 and 2.

[0070] Table 1 shows the average accuracy, precision, recall, F1 score, AUROC, and AUPR metrics for each method.

[0071] Table 2 shows the standard deviations of accuracy, precision, recall, F1 score, AUROC, and AUPR metrics for each method.

[0072] Compared with the K-nearest neighbors algorithm, support vector machine, deep neural network and random forest methods, the present invention has the best overall performance, with significant improvements in accuracy, precision, recall, F1 score, AUROC and AUPR metrics compared with traditional methods.

[0073] Meanwhile, to verify the superiority of the present invention, ablation experiments were conducted, and ten combinations were designed (combination a includes SMILES sequence, molecular map, and attention fusion mechanism; combination b includes SMILES sequence, molecular map, physicochemical properties, target sequence features, and attention fusion mechanism; combination c includes SMILES sequence, molecular map, molecular fingerprint, target sequence features, and attention fusion mechanism; combination d includes molecular map, molecular fingerprint, target sequence features, and attention fusion mechanism; combination e includes SMILES sequence, molecular map, physicochemical properties, target sequence features, triplet frequency vector, and attention fusion mechanism; combination f includes SMILES sequence, molecular fingerprint, physicochemical properties, and attention fusion mechanism). The model incorporates target sequence features and an attention fusion mechanism. Combination g includes SMILES sequences, molecular fingerprints, target sequence features, triplet frequency vectors, and an attention fusion mechanism. Combination h includes SMILES sequences, molecular graphs, molecular fingerprints, triplet frequency vectors, and an attention fusion mechanism. Combination i includes SMILES sequences, molecular graphs, molecular fingerprints, physicochemical properties, target sequence features, and triplet frequency vectors. Combination j includes SMILES sequences, molecular graphs, molecular fingerprints, physicochemical properties, target sequence features, triplet frequency vectors, and an attention fusion mechanism. Each combination under the same conditions underwent five iterations, demonstrating that multiple modalities play a crucial role in model prediction. Specific data is provided by […]. Figure 6 As shown. According to Figure 6 (a) and Figure 6 As shown in (b), overall, combination j has the highest index and the best prediction effect, which demonstrates the superiority of the drug target prediction method based on heterogeneous multi-encoder in this invention.

[0074] An exemplary embodiment of the present invention is illustrated using a drug-target pair as an example. The drug is dexamethasone palmitate, and the target is steroid hormone receptor ERR1.

[0075] The drug target prediction method based on heterogeneous multi-encoders in this invention performs data loading, preprocessing, feature extraction, feature fusion, and classification prediction. The molecular formula of dexamethasone palmitate is read and processed to generate four drug features: SMILES sequence, molecular map, molecular fingerprint, and physicochemical properties. The protein sequence of the target is read and processed to generate target sequence features and triplet frequency vectors. These features are then processed through a recurrent neural network, a convolutional neural network, fingerprint processing, physicochemical property processing, and triplet frequency processing to extract drug and target features. These features are then fed into an attention fusion module, processed through cross-attention and multi-head self-attention fusion mechanisms, and then fed into a multilayer perceptron, ultimately outputting a scalar as the prediction probability value. The final prediction probability value is 0.999994, which is greater than the threshold of 0.5, therefore the prediction result is an activation relationship, consistent with the actual label.

[0076] The heterogeneous multi-encoder-based drug target prediction method in this invention, by fusing multimodal heterogeneous features with an attention mechanism, can accurately and efficiently predict the activation or inhibition relationship between drugs and targets, and has broad application prospects and commercial value. Furthermore, experiments have demonstrated that this invention achieves superior prediction performance compared to traditional methods on the DrugBank dataset, providing reliable technical support for drug screening, drug relocation, and mechanism of action research.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A drug target prediction method based on heterogeneous multi-encoders, characterized in that, Includes the following steps: Acquire multimodal data of drug molecules and multimodal data of target proteins; Multiple parallel drug encoders are used to extract features from various modal data of drug molecules to obtain multiple drug feature vectors; the multiple drug feature vectors include: SMILES feature vector, molecular graph feature vector, fingerprint feature vector, and physicochemical property feature vector; Multiple protein feature vectors are obtained by extracting features from the multimodal data of the target protein using multiple parallel protein encoders; the multiple protein feature vectors include: protein fusion feature vector and triplet frequency vector. By using an attention mechanism, multiple drug feature vectors and multiple protein feature vectors are concatenated to obtain a fused feature vector. The fused feature vector is input into the classifier to obtain the prediction results of the interaction between activation and inhibition.

2. The drug target prediction method based on heterogeneous multi-encoder according to claim 1, characterized in that, The drug molecule's modal data includes at least two of the following: SMILES sequence, molecular graph structure, molecular fingerprint, and physicochemical property descriptors; The plurality of parallel drug encoders include at least two of the following: a drug SMILES recurrent neural network encoder for extracting SMILES sequence features, a graph neural network encoder for extracting molecular graph features, a first multilayer perceptron encoder for extracting molecular fingerprint features, and a second multilayer perceptron encoder for extracting physicochemical property descriptors. The molecular fingerprint is a spliced ​​vector of MACCS fingerprint, ErG fingerprint, PubChem fingerprint, and ECFP2 fingerprint; the physicochemical property descriptors include: molecular weight, octanol / water partition coefficient, number of hydrogen bond donors, number of hydrogen bond acceptors, number of rotatable bonds, number of aromatic rings, topological polar surface area, number of heavy atoms, total number of atoms, total number of rings, and sp... 3 The number of hybridized carbon atoms, the number of aliphatic rings, the number of saturated rings, the number of heteroatoms, the number of aromatic heterocycles, the number of saturated heterocycles, the number of aliphatic heterocycles, and the Labute approximate surface area.

3. The drug target prediction method based on heterogeneous multi-encoders according to claim 1, characterized in that, The target protein's multiple modal data include at least two of the following: one-hot coding matrix of amino acid sequences, integer index sequence of amino acid sequences, and triplet frequency vector; The plurality of parallel protein encoders include at least two of the following: a convolutional neural network encoder for extracting features of a one-hot encoded matrix, a protein sequence recurrent neural network encoder for extracting features of an integer index sequence, and a third multilayer perceptron encoder for extracting the frequency vector of a triplet.

4. The drug target prediction method based on heterogeneous multi-encoder according to claim 3, characterized in that, Feature extraction of multiple protein feature vectors was performed on various modalities of the target protein using multiple parallel protein encoders. The specific contents of the resulting protein feature vectors include: The first intermediate feature vector is obtained by extracting features from multiple modal data of the target protein using a convolutional neural network encoder. The second intermediate feature vector is obtained by extracting features from multiple modal data of the target protein using a recurrent neural network encoder. The first intermediate feature vector and the second intermediate feature vector are fused to obtain the protein fusion feature vector; The triplet frequency vector is obtained by extracting features from the multimodal data of the target protein using a third multilayer perceptron encoder. Protein fusion feature vectors and triplet frequency vectors are essentially multiple protein feature vectors.

5. The drug target prediction method based on heterogeneous multi-encoders according to claim 4, characterized in that, The method for generating the triplet frequency vector includes: classifying 20 standard amino acids into seven categories according to their physicochemical properties, performing frequency statistics on all triplets composed of three consecutive amino acid residues in the protein sequence to obtain a 343-dimensional original vector, and then performing normalization processing to generate the triplet frequency vector.

6. The drug target prediction method based on heterogeneous multi-encoder according to claim 1, characterized in that, By employing an attention mechanism, multiple drug feature vectors and multiple protein feature vectors are concatenated to obtain a fused feature vector, which specifically includes: The SMILES feature vector, molecular graph feature vector, fingerprint feature vector, physicochemical property feature vector, protein fusion feature vector, and triplet frequency vector are arranged in a preset order and sequentially pass through a cross-modal feature enhancement module to obtain multiple enhanced feature vectors. The cross-attention module uses the feature vector of the current modality as the query vector and the feature vector of the next modality as the key vector and value vector. After multi-head attention calculation, residual connection, and layer normalization, it outputs the enhanced feature vector of the current modality, and the enhanced feature vector of the current modality has the same dimension as the feature vector of the current modality. Multiple enhanced feature vectors are concatenated along the sequence dimension to obtain a concatenated feature sequence; The concatenated feature sequence is input into the multi-head self-attention module to extract high-order interaction features between modalities. The extracted results are then processed by a convolutional layer and flattened to obtain a fused feature vector.

7. The drug target prediction method based on heterogeneous multi-encoder according to claim 1, characterized in that, The fused feature vector is input into the classifier to obtain the prediction results of the interaction between activation and inhibition, specifically including: The fused feature vector is input into a multilayer perceptron classifier, and the output layer of the multilayer perceptron classifier outputs a scalar logit. The scalar logit is converted into a probability value between 0 and 1 using the Sigmoid function; The judgment is based on a probability value of 0.5: if the probability value is greater than 0.5, the prediction result is an activation relationship; if the probability value is less than or equal to 0.5, the prediction result is an inhibition relationship.

8. The drug target prediction method based on heterogeneous multi-encoder according to claim 7, characterized in that, The drug target prediction method based on heterogeneous multi-encoders also includes: Obtain a labeled training dataset; each sample in the training dataset includes: multiple modal data of drug molecules, multiple modal data of target proteins, and corresponding interaction labels; The training dataset is input into multiple parallel drug encoders, multiple parallel protein encoders, and a multilayer perceptron classifier. The trainable parameters of these multiple parallel drug encoders, multiple parallel protein encoders, and a multilayer perceptron classifier are optimized with the goal of minimizing the binary cross-entropy loss function.

9. The drug target prediction method based on heterogeneous multi-encoder according to claim 2, characterized in that, The drug molecule's multiple modal data include: SMILES sequence, molecular graph structure, molecular fingerprint, and physicochemical property descriptors; the multiple parallel drug encoders include: a drug SMILES recurrent neural network encoder, a graph neural network encoder, a first multilayer perceptron encoder, and a second multilayer perceptron encoder.

10. The drug target prediction method based on heterogeneous multi-encoders according to claim 3, characterized in that, The target protein's multiple modal data include: one-hot encoding matrix of amino acid sequences, integer index sequence of amino acid sequences, and triplet frequency vector; the multiple parallel protein encoders include: convolutional neural network encoder, protein sequence recurrent neural network encoder, and third multilayer perceptron encoder.

Images