Protein drug response prediction method based on em fusion learning

Abstract

The present application relates to a protein drug reaction prediction method based on EM fusion learning, comprising: 1) constructing a data set containing protein nodes, drug nodes and intermediate nodes into a text attribute graph; 2) embedding the protein amino acid sequence and the drug molecule SMILES sequence to obtain corresponding sequence representations; 3) constructing a transformer-based protein-drug reaction prediction model; 4) constructing a GNN-based protein-drug reaction prediction model; 5) using a variational EM framework to alternately update the above two models to obtain an EM fusion protein-drug reaction prediction model; 6) inputting the to-be-tested protein amino acid sequence and the to-be-tested drug SMILES sequence into the above prediction model to output the reaction prediction result of the to-be-tested protein-drug pair. The present application learns the topological information in the protein-drug network to realize protein-drug reaction prediction.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, combining the principles and methods of biology and informatics, and using computer and mathematical tools to process, analyze and interpret biological data, specifically involving response prediction of drug-protein interaction problems. Background Technology

[0002] In recent years, with the rapid development of computer technology and the increasing demand for computer and artificial intelligence technologies from various industries, computer technologies, represented by machine learning and deep learning algorithms, have gradually penetrated into various fields such as medicine, biology, and chemistry. At the same time, the massive amounts of data in these emerging fields have also stimulated the enormous potential of artificial intelligence.

[0003] In the medical field, new drug development is an expensive and time-consuming process. The total development cost of a new drug ranges from $2 billion to $3 billion, and the development time is typically 13-15 years. Therefore, drug retargeting methods have become a research hotspot in the field of drug development. A growing body of drug retargeting research indicates that drug-target interactions are crucial, but it usually takes 2-3 years to validate the accuracy of drug-protein interaction (DPI) methods through expensive, large-scale biochemical experiments. The DPI problem, as an important issue in the biopharmaceutical field, has been extensively studied. Traditional methods typically employ wet laboratory drug target characterization to explore protein-drug response relationships. While accurate, this method is generally expensive, cumbersome, and slow.

[0004] Currently, drug-protein interaction research plays an extremely important role in drug development and design, drug optimization and side effect prediction, disease treatment and personalized medicine, drug safety assessment, drug reuse and repositioning.

[0005] In conclusion, drug-protein interaction (DPI) research has significant implications for real life. It plays a crucial role in drug development, drug optimization, disease treatment and personalized medicine, drug safety assessment, and drug reuse and repositioning. By gaining a deeper understanding of drug-protein interactions, DPI research holds the potential to accelerate the drug development process, improve drug efficacy and safety, provide more effective treatment options for patients, and reduce the costs and risks of drug development. Therefore, DPI research has important and far-reaching practical significance for promoting progress in the medical field and improving human health and quality of life.

[0006] With the rapid development of internet technology, data on drug compounds, targets, and interactions are accumulating rapidly, which can be used to calculate drug-protein interaction (DPI) research methods. The internet has further promoted the development of drug retargeting. Drug-protein interaction (DPI) prediction reduces the cost and time of drug development and plays a crucial role in drug discovery. Machine learning and deep learning are increasingly involved in DPI research, and DPI research is gradually entering the era of artificial intelligence.

[0007] Based on the above research, this invention finds that current research on the DPI problem mainly focuses on two aspects: analyzing the protein-drug network topology and learning the drug molecule structure and protein characteristics. This invention aims to construct a model that simultaneously learns both aspects of information and allows for the mutual extraction of information learned by the two methods. Summary of the Invention

[0008] This invention aims to predict drug-protein interactions. It uses the EM algorithm to fuse two models, resulting in a new model that combines the advantages of both. This model can learn from the textual information of the nodes themselves, as well as from the topological information in the protein-drug network. Finally, the EM algorithm effectively extracts information from both models.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a protein drug response prediction method based on EM fusion learning, characterized by comprising the following steps: Step 1) Treat the entire single dataset as a text attribute graph, as follows: Step 1.1) In the model of this invention, the entire single dataset is treated as a text attribute graph.

[0010] Step 1.2) Each protein node, drug node, and intermediate node in the dataset can be regarded as a node V in the graph.

[0011] Step 1.3) In the dataset, each piece of information can be regarded as a data point in the adjacency matrix A, with a value of 1 or 0.

[0012] Step 2) Embed the protein amino acid sequence and the drug molecule SMILES sequence to obtain the protein amino acid sequence representation and the drug molecule sequence representation.

[0013] Step 3) Construct a DPI prediction method based on the Transformer model. The specific method is as follows: Step 3.1) Construct a method consisting of three parts: protein data embedding, drug data embedding, and DPI prediction method based on Transformer model.

[0014] Step 3.2) To achieve the acquisition, training, and prediction of protein sequence information and drug molecular structure information in the protein-drug response prediction model.

[0015] Step 3.3) Modify the transformer model by inputting the embedded protein sequence features into the Encoder and the drug molecule features into the Decoder for interaction.

[0016] Step 4) Construct a GNN-based protein-drug response prediction system, the specific method of which is as follows: Step 4.1) Encode the protein sequence using a one-hot code, and then extract the protein features using a convolutional neural network and k-mer statistical methods.

[0017] Step 4.2) Extract drug molecule features using convolutional neural networks and molecular fingerprinting methods.

[0018] Step 4.3) Train the GNN for protein-drug interaction prediction.

[0019] Step 5) Use a variational EM framework to alternately update the two models. Through alternating E-step and M-step iterations, the two models learn alternately, thereby achieving the goal of the EMPDPR model jointly learning from protein and drug molecule information and protein-drug network topology information; the specific method is as follows: Step 5.1) Use the variational EM algorithm to fuse the transformer-based protein-drug response prediction and the GNN-based protein-drug response prediction, so that they can learn from each other.

[0020] Step 5.2) During the model mutual optimization process, the E-step and M-step are performed iteratively. In the E-step, the pseudo-labels predicted by the GNN and the gold labels in the actual database are used to train the transformer. In the M-step, the pseudo-labels predicted by the transformer and the actual labels in the actual database are used to train the GNN.

[0021] Step 5.3) By alternating between the E-step and the M-step, the two models learn alternately, thereby achieving the goal of the EMPDPR model learning from protein and drug molecule information and the topological information of the protein-drug network.

[0022] Step 6) After the model training is completed, the transformer model or GNN model is used to predict the node labels.

[0023] Step 6.1) Input the amino acid sequence of the protein to be tested and the SMILES sequence of the drug to be tested into the trained EM fusion protein-drug response prediction model to obtain the predicted probability of the response between the protein to be tested and the drug to be tested.

[0024] Step 6.2) Output the node label prediction results based on the reaction prediction results.

[0025] The beneficial effects of this invention are as follows: This invention takes drug-protein interaction prediction as its research objective, and utilizes the EM algorithm to fuse Transformer and GNN models. The resulting new model combines the advantages of both, learning from the textual information of the nodes themselves while also utilizing the topological information of the protein-drug network. The EMPDPR model significantly improves the accuracy of both the GNN and Transformer models. The mutual distillation learning between the two models effectively helps each learn new knowledge, thus better refining the model. Attached Figure Description

[0026] Figure 1 This is a diagram of the protein-drug response prediction framework based on transformers in this invention; Figure 2 This is a structural diagram of the protein-drug response prediction based on the GNN model of this invention; Figure 3 This is a framework diagram of the protein-drug response prediction model based on EM fusion learning of the present invention; Figure 4 This is a schematic diagram of a sliding window for protein sequences. Figure 5 Embed structural diagrams into protein data; Figure 6 Flowchart for protein feature extraction Figure 7 Embedding structural diagrams into drug molecule data; Figure 8 This is a flowchart of drug feature extraction. Figure 9 Example diagram of Conv1D operation; Figure 10 Example diagram of transformer DPI; Figure 11 This is a schematic diagram of an intermediate node; Figure 12 Here is a diagram of the GNNDPI framework; Figure 13 This is a framework diagram of a protein-drug response prediction model based on EM fusion learning. Figure 14 EM framework diagram; Figure 15 For the overall framework of EMDPI; Figure 16 A bar chart showing the performance comparison of different models on the C.elegans dataset - AUC; Figure 17 A bar chart showing the performance comparison of different models on the Human_data dataset - AUC; Figure 18 EMDPI C. elegans trend chart - AUC; Figure 19 EMDPI Human_data trend chart - AUC; Figure 20 EMPDPR GPCR trend plot - AUC; Figure 21 Comparison of EMDPI ablation experimental performance - AUC; Figure 22 Comparison chart of EMDPI ablation experimental performance - ACC. Detailed Implementation

[0027] The technical solution of this invention will be clearly and completely described below with reference to process pictures recorded during the creation of this invention, making it easier to understand.

[0028] The protein-drug response prediction framework based on transformers in this invention is shown in the diagram below. Figure 1 As shown, the protein-drug response prediction structure diagram based on the GNN model is as follows: Figure 2 As shown in the diagram, the framework of the protein-drug response prediction model based on EM fusion learning is as follows: Figure 3 As shown.

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation process.

[0030] Protein drug response prediction method based on EM fusion learning: 1) Construct a text attribute graph.

[0031] In this implementation, the entire single dataset is first constructed as a text attribute graph (TAG). The TAG is denoted as G=(V,A,S), where V represents the nodes in the graph, A is the adjacency matrix, and S represents the text features contained in each node. According to the definition in the paper, in each node V, A and S can be learned and predicted using graph neural networks and language models, respectively.

[0032] Furthermore, in the model of this invention, since the dataset is in the form of a bipartite graph, n intermediate nodes are introduced to update surrounding nodes during the EM algorithm. Each protein node, drug node, and intermediate node in the dataset can be regarded as a node V in the graph; the protein sequence statement contained in each protein or drug molecule and the SMILES sequence of the drug molecule are regarded as the text features S contained in that node; each piece of information in the dataset can be regarded as a data point in the adjacency matrix A, with a value of 1 or 0. Thus, the original protein-drug data can be uniformly represented as a text graph structure that can be processed by both language models and graph models.

[0033] 2) The model shares the input and embedding parts.

[0034] 2.1) Protein Data Embedding

[0035] In the transformer model input, the amino acid sequence of the protein is used as the input sequence representation. Then, a pre-trained word2vec model is used to convert the protein sequence into an embedding sequence. In practice, the sliding window for the amino acid sequence is set to w=3, and the string is split into stacked sequences of length three. Figure 4 An example of this splitting is given in the document. To standardize the input length, protein sequences shorter than the maximum protein sequence length are padded.

[0036] After obtaining the sequence embedding, a one-dimensional convolutional neural network is used to extract local protein features. The extracted features are then transformed using a two-layer feedforward neural network to obtain the local features of the protein sequence. Let the protein sequence be represented as... Let the local features of protein i be... We can obtain the following formula:

[0037] For the overall sequence characteristics of the protein, the k-mer statistical method with k=1, 2, 3 was used to manipulate the protein amino acid sequence, as shown in Table 1: Table 1. Schematic table of protein sequence k-mers (GTAGAC)

[0038] An 8420-dimensional embedding is constructed to obtain the global features of the protein. Let the k-mer embedding be represented as... Let the global features of protein i be... We can obtain the following formula:

[0039] Finally, local protein features With global features Adding them together yields a characteristic representation of the protein sequence. The overall structure is shown below. Figure 5 See the overall process. Figure 6 The formula is shown below:

[0040] 2.2) Drug Data Embedding

[0041] For drug information, the RDKit library functions in Python are used to process drug data in the transformer model. Drug features include both individual atomic characteristics such as all atom types and charges within the drug molecule, and various chemical bonds between these atoms. Both play a crucial role in representing drug information. Therefore, a `compounds` list is used to represent the atomic information within the drug molecule, and an `adjacencies` list is used to represent the adjacency matrix formed by the various chemical bonds within the drug. For the molecule itself, information is mainly represented using its common atoms, degrees, formal charges, free radical electrons, hybridization, aromaticity, and chirality. These vectors are concatenated to form a 34-dimensional representation vector. For molecular chemical bonds, the `MolfromSmiles` function in RDKit is primarily used to complete the representation and store it in a list.

[0042] Next, the vectors from the `compounds` list and the adjacency matrix `adjacencies` are fed into a simple GCN model to integrate the features of adjacent atoms to learn the representation of individual atoms. For the input graph data, assuming there are N nodes, each node has its own features, let these features form an N×D matrix `X`, and the relationships between the nodes also form an N×N matrix `A`, also called the adjacency matrix. `X` and `A` are the inputs to the model. The formula for the propagation between layers in GCN is as follows:

[0043] Where I is the identity matrix, D is the degree matrix of A, H is the feature of each layer, and H is the non-linear activation function.

[0044] A single-layer GCN is used to process the 3D tensor `input` and the 3D adjacency matrix tensor `adj` of the graph data structure. The shape of `input` is [batch, num_node, atom_dim], representing the node features, while the shape of `adj` is [batch, num_node, num_node], representing the connections between nodes in the graph. For the input tensor, a learnable weight tensor `self.weight` is used, and the node features are linearly transformed using matrix multiplication `torch.matmul(input, self.weight)`. Then, the node features are updated using batch matrix multiplication `torch.bmm(adj, support)`. Finally, the drug molecule expression is output.

[0045] In the GNN model, we use molecular fingerprinting to extract features of atoms and chemical bonds in drug molecules.

[0046] Four lists are constructed for drug molecule features: drug molecule data (dMolData), drug sequence data (dSeqData), drug feature data (dFeaData), and drug fingerprint data (dFinData). Each drug molecule is labeled and stored; for drug molecule sequences shorter than dSeqMaxLen, padding is performed. For data added to the drug fingerprint data list, molecular fingerprints are used to encode the drug sequence into a series of binary bit strings containing the structural features and topological information of the chemical molecules. Then, a feedforward neural network is used to extract the global features of the drug molecules. Let the drug molecule sequence be represented as... Let the global characteristics of drug molecule i be... We can obtain the following formula:

[0047] The drug molecule is encoded as a 75-dimensional vector. Then, the drug sequence is processed using the same method as for protein sequences. First, a convolutional neural network is used to extract features from the drug molecule sequence vector, and these features are then fed into a two-layer feedforward neural network to obtain the final local features of the drug molecule. Let the drug molecule sequence be represented as... Let the local features of drug molecule i be... We can obtain the following formula:

[0048] After obtaining the local and global features of the drug molecule, adding the two together yields the feature representation of the drug sequence. The formula is as follows:

[0049] After the above operations, protein amino acid sequence representations and drug molecule sequence representations were obtained. The overall structure is shown below. Figure 7 See the overall process. Figure 8 The paper presents a structural diagram of drug molecule data embedding and a flowchart of drug feature extraction.

[0050] 3) Construct a protein-drug response prediction model based on transformer.

[0051] The overall framework of the DPI prediction method based on the Transformer model is as follows: Figure 1 As shown, this section consists of three parts: protein data embedding, drug data embedding, and a DPI prediction method based on the Transformer model. It mainly realizes the acquisition, training, and prediction of protein sequence information and drug molecular structure information.

[0052] After input embedding, the feature vector of the protein amino acid sequence is fed into the encoder to learn more abstract protein features; the drug molecule sequence representation is fed into the decoder to learn the interaction features between proteins and drugs. The paper points out that, since the protein database is much smaller than the corpus required for normal use of traditional transformer models, to prevent overfitting during training on the protein database, a simple gated convolutional network is used in the encoder instead of the traditional self-attention mechanism. The gate includes Conv1D and gated linear units (GLUs). Figure 9 An example diagram of Conv1D operation is provided.

[0053] In the model, the `Conv1d` function is used to convolve the input feature vector of dimension `hid_dim`, employing convolutions with an odd `kernel_size` to ensure kernel center alignment. Then, in the `forward` method, the convolutional layers in `convs` are iterated through, and convolution operations are performed on the input. The `F.glu` function is then used to process the convolution result using the activation function of a gated linear unit to obtain a new result. The convolution result is then further processed through residual connections and a scaling factor `self.scale`. Finally, the processed result is normalized using `LayerNorm`. The remaining parts are the same as the encoder in a traditional transformer model. In the model, the output of the gated convolutional network is the final representation vector of the protein sequence. The model is designed with 3 hidden layers, and the input and hidden feature dimensions are 64 and 128, respectively.

[0054] After process 2), the drug molecule sequence representation is obtained. Next, the interaction features between the encoder and decoder are learned through the decoder in the transformer model. The protein amino acid sequence representation serves as the input to the encoder, while the drug molecule sequence serves as the input to the decoder. In the decoder part, a multi-head self-attention layer is used to extract the interaction information between the encoder and decoder. The multi-head self-attention layer accepts three inputs, Q, K, and V, and is calculated according to the following formula:

[0055] Since this task requires directly determining whether the protein reacts with the drug, the sequence mask mechanism in the traditional transformer decoder was modified. After removing the sequence mask to obtain the interaction feature vector, its weights were calculated using the softmax function, as follows:

[0056] Where n is the maximum value of the interaction sequence. The modulus of each vector is given. Next, we weight each vector using its weights and sum them to obtain the final interaction feature vector. This interaction feature vector is then fed into a fully connected layer to obtain the probability of whether the drug and protein interact. During prediction, this probability is reduced to 0 or 1 to obtain the predicted structure of whether the drug and protein react. During training, for this traditional binary classification task, the binary cross-entropy loss function is used to train the model. The binary cross-entropy loss is calculated as follows:

[0057] Where y is the true value. These are predicted values. This completes the transformer model portion for predicting drug-protein responses. The overall process is as follows: Figure 10 As shown.

[0058] 4) Construct a protein-drug response prediction model based on GNN.

[0059] In constructing the GNN model, a supervised drug-protein association network was built. To learn the implicit relationship between protein and drug molecules, and to facilitate the EM algorithm's update operations on surrounding nodes later, this invention introduces intermediate nodes between protein and drug nodes. For the GNN model itself, in learning protein... With drugs In determining whether an interaction exists, the model learns not only explicit path information but also implicit path information. By learning both explicit and implicit path information simultaneously, the model can determine whether a drug and protein interact without directly relying on similarity scores, avoiding the error of directly linking similar proteins and drugs together.

[0060] In the construction of the graph neural network, m intermediate nodes are introduced. These intermediate nodes have the same feature dimension as the protein and drug nodes, and their initial feature representations are randomly initialized from a standard normal distribution, with values ​​limited to between 0 and 1. Therefore, we can calculate their cosine similarity with the protein or drug nodes to obtain path association information, such as... Figure 11 As shown, the cosine similarity between intermediate nodes and protein or drug nodes is calculated to obtain path association information. Assume the intermediate node is... Protein nodes are Drug node is Then the cosine similarities between the intermediate node and the protein node, and between the intermediate node and the drug node, are respectively:

[0061]

[0062] Using intermediate nodes as a medium, a network consisting of one drug node, one protein node, and m intermediate nodes is constructed for each protein node and drug node pair. The path information of the drug and protein reaction can be obtained from this network structure.

[0063] To capture path information using graph neural networks, we first calculate the cosine similarity between nodes and construct a drug-node-protein weighted adjacency matrix. Obtain the adjacency matrix. Then, use the following formula to filter out negative values ​​and normalize them:

[0064] in, for The degree matrix, This is used to filter out negative values ​​in the adjacency matrix. Then, a three-layer graph neural network is constructed to capture path information from the adjacency matrix. The initialization information of the graph is embedded as follows:

[0065] The graph neural network layer is represented as:

[0066] Furthermore, the three-layer propagation process is as follows:

[0067]

[0068]

[0069] in, , , , , , These are the hidden parameters of the GNN. (·) is the activation function. Finally, the protein information and drug information of Z are selected, i.e., the first and second rows of the matrix, and then summed to obtain the final protein information vector and drug information vector:

[0070]

[0071] After obtaining protein characteristics Drug characteristics Next, the following method is used to determine whether the drug-protein pair will interact. The element-wise product of the two vectors is used to simulate whether they interact, and the result is passed through a feedforward neural network to obtain the probability of their interaction. During training, the model is trained using the following binary cross-entropy loss function:

[0072] The above steps complete the training process for the GNN part of protein-drug interaction prediction. Compared with the Transformer part, its method is less effective in utilizing textual information such as drug molecule information and protein amino acid sequence information. However, it can better utilize and learn the topological information of the protein-drug interaction network, and thus predict the DPI problem from different perspectives. Its GNN model structure is as follows: Figure 2 As shown, the GNN framework is as follows Figure 12 As shown.

[0073] 5) Perform fusion training and prediction based on the variational EM framework.

[0074] After constructing the two basic models, the variational EM framework is used to fuse and train them, ultimately completing node label prediction. The overall framework is as follows: Figure 13 As shown.

[0075] 5.1) Variational EM Framework

[0076] This invention employs a variational EM framework to alternately update two models: a transformer model and a GNN model. The transformer model learns from the text information of each text graph node, thus modeling the label distribution based on the node's text attributes. The GNN model learns from the text encoding of each graph node and the labels of its neighboring nodes, thus modeling the label distribution based on global conditions. The GLEM algorithm optimizes both the transformer and GNN models by maximizing the variational lower bound of the log-likelihood function, updating one model while keeping the other fixed through E-steps and M-steps. Based on this theoretical foundation, this invention uses the variational EM algorithm to fuse transformer-based and GNN-based protein-drug response prediction, enabling them to learn from each other.

[0077] This method is based on the pseudo-likelihood variational framework. Within this framework, the aim is to maximize the log-likelihood function of the observed node labels, i.e. Because of the existence of node labels that are difficult to obtain directly. It is usually difficult to directly optimize this function, so the following formula is used to optimize the lower bound of the evidence:

[0078] in For variational distributions, ELBO can be optimized by alternately optimizing distribution p (E-step) and distribution q (M-step). In the E-step, the goal is to update distribution q to minimize the KL divergence between q and p, thus tightening the ELBO lower bound. In the M-step, p is updated according to the following maximum pseudo-likelihood function to maximize the data likelihood:

[0079] The distribution q aims to define the distribution of node labels using textual information. Let q be the node labels learned by the transformer model. Then, for the overall transformer model, we can obtain the following formula:

[0080] Similarly, a conditional distribution is defined for distribution p. Wherein, Let A be the node features, A be the graph adjacency matrix, and y be the node labels. This structural form can be learned by a graph neural network model. Similarly, we can obtain:

[0081] 5.1.1) Expectation step

[0082] In the E-step, the transformer model is instantiated, the GNN model is fixed, and the transformer model is updated, thereby distilling the global information learned from the graph nodes into the transformer model.

[0083] Maximizing the lower bound of transformer evidence is equivalent to minimizing the KL divergence between the posterior distribution and the variational distribution, i.e. Since KL divergence depends on The entropy follows the wake-sleep algorithm, minimizing the inverse KL divergence according to the formula to maximize the transformer. :

[0084] In a GNN model with a defined distribution, it's necessary to predict the label distribution of a node itself based on the labels of its surrounding nodes. However, in the distribution described above, only some node labels can be observed, while the labels of other nodes cannot be determined. To address this issue, pseudo-labels predicted by the transformer model are used to replace the unlabeled nodes in the graph, resulting in the following approximate distribution:

[0085] Combining the previous objective function, we can obtain the objective function for training the transformer as follows:

[0086] Here These are hyperparameters used to control the relative importance of training the two models. In this objective function, the first term can be seen as forcing the transformer model to compare its predictions of surrounding information with the GNN model's predicted labels, thereby gradually learning the information distilled by the GNN model. The second term, on the other hand, uses a supervised model to train the given labeled nodes normally. By combining the two, the transformer model gradually learns from the GNN model to master the global feature information of the learning object.

[0087] 5.1.2) Maximization step

[0088] In the M-step, the GNN model is instantiated. In the GNN phase, unlike the E-step, the transformer model is fixed to optimize the GNN model to maximize the pseudo-likelihood. The objective function is rewritten as follows: Pseudo-labels are generated for unlabeled nodes using the transformer model and fed into the GNN model.

[0089] Here These are hyperparameters used to control the relative importance of training the two models. In this objective function, the first term can be seen as forcing the GNN model to predict surrounding information and compare it with the transformer model's predicted labels, thereby gradually learning the information distilled by the transformer model. The second term, on the other hand, uses a supervised model to train the given labeled nodes normally. By adding the two together, the GNN model gradually learns from the transformer model to master the global feature information of the learning object.

[0090] 5.2) Implementation of the EM Fusion Learning Mechanism

[0091] In terms of data structure, intermediate nodes are added to the bipartite graph, thereby updating the surrounding nodes of each protein and drug node. Through the connections of the intermediate nodes, pseudo-labels can be added to some nodes, thus enabling distillation training of the target node. The EM framework diagram is as follows: Figure 14 As shown. Next, by incorporating the transformer-based protein-drug response prediction model and the GNN-based protein-drug response prediction model into the framework, we can obtain the overall EMPDPR framework, as follows. Figure 15 As shown.

[0092] During the mutual optimization process, E-steps and M-steps are performed iteratively. In the E-step, the pseudo-labels predicted by the GNN and the gold labels in the actual database are used for transformer training. In the M-step, the pseudo-labels predicted by the transformer and the actual labels in the actual database are used for GNN training. Through the alternating iterations of E-steps and M-steps, the two models learn alternately, thereby achieving the goal of the EMDPI model jointly learning from protein and drug molecule information and the topological information of the protein-drug network.

[0093] 6) Comparative validation of protein-drug response prediction results

[0094] To verify the effectiveness of the protein-drug response prediction method based on EM fusion learning proposed in this invention, three datasets—C. elegans, Human_data, and GPCR—were selected for testing the model, and their data composition is shown in Table 2. To comprehensively evaluate the model's performance in the node label prediction task, this invention uses six metrics: accuracy (ACC), precision (Precision), recall, F1 score, AUC, and PRC. These evaluation metrics can reflect the model's predictive performance from multiple perspectives, including overall classification accuracy, positive sample recognition ability, model balance performance, and ranking ability.

[0095] Table 2 Data Composition Table

[0096] In protein-drug response prediction research, traditional machine learning methods, transformer-based methods, and GNN-based methods are all representative to some extent. To verify the effectiveness of the method proposed in this invention, the model of this invention was compared with existing methods using the same dataset and a unified evaluation metric. The comparison results show that the EMPDPR model proposed in this invention outperforms basic models using transformers or GNNs alone, and also outperforms some existing prediction methods. This result indicates that this invention achieves the joint utilization of protein and drug molecular structure information and protein-drug network topology information through the EM fusion learning mechanism, thereby effectively improving prediction accuracy and model generalization ability.

[0097] During the testing of the model in this invention, the trained transformer model, GNN model, and the model after EM fusion learning were validated on the three datasets mentioned above. Experimental results show that the EMPDPR model proposed in this invention exhibits good prediction performance on multiple datasets. Specifically, on the C. elegans dataset, the performance of different models on the test set is as follows: Figure 16 The experimental results show that the model of this invention also achieved superior results on the Human_data dataset, as shown in the following figures. Figure 17 As shown in Table 3, the EM-GNN fusion model maintained good stability on both the validation and test sets on the GPCR dataset. Overall, the results demonstrate that EM-GNN and EM-transformer outperform traditional machine learning methods and some existing transformer and GNN models on multiple evaluation metrics, indicating that this invention has good predictive ability in protein-drug response prediction tasks.

[0098] Table 3 Performance of the GPCR dataset on the EMDPI validation and test sets

[0099] 7) Validation and parameter stability analysis of the EM fusion learning mechanism

[0100] To further illustrate the role of the EM fusion learning mechanism in improving prediction results, this invention conducted convergence experiments, ablation experiments, and hyperparameter experiments.

[0101] In the convergence experiments, both the transformer and GNN models showed significant or near-convergence trends during multiple rounds of EM distillation. Specifically, on the C.elegans dataset, both models showed significant AUC convergence in the fourth round of EM distillation; on the Human_data dataset, their AUCs nearly converged in the sixth round; and on the GPCR dataset, both models exhibited relatively stable convergence trends. The results can be further analyzed by... Figure 18 , Figure 19 , Figure 20 The results above demonstrate that the EM fusion learning mechanism can gradually stabilize the two base models during alternating training and continuously improve node label prediction performance.

[0102] In the ablation experiments, to verify the independent contribution of the EM fusion learning mechanism, the unfused base model was compared with the model after incorporating the EM framework. Experimental results show that after introducing the EM framework, the AUC of the GNN-DPI model increased from 0.993 to 0.997, and the AUC of the transformer-DPI model increased from approximately 0.98 to 0.994; the ACC of the GNN-DPI model increased from 0.97 to 0.996, and the ACC of the transformer-DPI model increased from approximately 0.945 to 0.975. The results are as follows... Figure 21 and Figure 22 As shown, the EM fusion learning mechanism enables two models to receive feature information from each other through pseudo-label distillation, thereby effectively improving the accuracy of node label prediction.

[0103] Furthermore, to analyze the model's sensitivity to parameter selection, hyperparameter experiments were conducted. Experimental results show that the model exhibits good robustness to the values ​​of hyperparameters α and β, with α=0.5 and β=0.5 ultimately chosen. Simultaneously, the optimal learning rate for both the GNN and transformer models is 0.001. These results further demonstrate that, under reasonable parameter settings, this invention can stably output node label prediction results and possesses good engineering applicability.

Claims

1. A protein drug response prediction method based on EM fusion learning, characterized in that, Includes the following steps: Step 1) Construct a text attribute graph from the dataset containing protein nodes, drug nodes, and intermediate nodes; Step 2) The protein amino acid sequence and the drug molecule SMILES sequence are embedded to obtain the protein amino acid sequence representation and the drug molecule sequence representation; Step 3) Construct a transformer-based protein-drug response prediction model; Step 4) Construct a protein-drug response prediction model based on GNN; Step 5) The variational EM framework is used to alternately update the transformer-based protein-drug response prediction model and the GNN-based protein-drug response prediction model to obtain the trained EM fusion protein-drug response prediction model. Step 6) Input the amino acid sequence of the protein to be tested and the SMILES sequence of the drug to be tested into the trained EM fusion protein-drug response prediction model, and output the response prediction results of the protein-drug pair.

2. The protein drug response prediction method based on EM fusion learning according to claim 1, characterized in that, The specific method for step 1) is as follows: Step 1.1) Treat the entire single dataset as a text attribute graph; Step 1.2) Each protein node, drug node, and intermediate node in the dataset is considered as a node V in the graph; Step 1.3) The protein sequence statement contained in each protein or drug molecule and the SMILES sequence of the drug molecule are used as the text feature S contained in that node; Step 1.4) Each piece of information in the dataset is regarded as a data point in the adjacency matrix A, with a value of 1 or 0.

3. The protein drug response prediction method based on EM fusion learning according to claim 1, characterized in that, The specific method for step 2) is as follows: Step 2.1) Express the amino acid sequence in the protein as the input sequence, and then use a pre-trained word2vec model to convert the protein sequence into an embedding sequence; Step 2.2) Use RDKit to convert drug molecule data into atomic feature vectors and atomic adjacency matrices, and obtain drug molecule expression feature vectors through the GCN model; Step 2.3) Obtain the protein amino acid sequence representation and the drug molecule sequence representation.

4. The protein drug response prediction method based on EM fusion learning according to claim 1, characterized in that, The specific method for step 3) is as follows: Step 3.1) Input the feature vector of the protein amino acid sequence into the encoder to continue learning more abstract features of the protein; Step 3.2) The drug molecule sequence representation is fed into the decoder to learn the interaction characteristics between the protein and the drug; Step 3.3) Input the interaction features into the fully connected layer to obtain the protein-drug response probability and construct a protein-drug response prediction model based on transformer.

5. The protein drug response prediction method based on EM fusion learning according to claim 1, characterized in that, The specific method for step 4) is as follows: Step 4.1) Encode the protein sequence using a one-hot code, and then extract the protein features using a convolutional neural network and k-mer statistical methods; Step 4.2) Extract drug molecule features using convolutional neural networks and molecular fingerprinting methods; Step 4.3) Introduce an intermediate node between the protein node and the drug node, calculate the path information in the adjacency matrix formed by the drug-protein network, and input the regularized path information into a three-layer graph neural network for training to obtain the protein-drug response probability and construct a protein-drug response prediction model based on GNN.

6. The protein drug response prediction method based on EM fusion learning according to claim 1, characterized in that, The specific method for step 5) is as follows: Step 5.1) The variational EM framework is used to alternately update the transformer-based protein-drug response prediction model and the GNN-based protein-drug response prediction model; Step 5.2) In step E, the pseudo-labels predicted by GNN and the gold labels in the actual database are used to train the transformer; Step 5.3) In the M-step, GNN is trained using the pseudo-labels predicted by the transformer and the actual labels in the actual database; Step 5.4) Repeat the E step and the M step until the preset iteration termination condition is met to obtain the trained EM fusion protein-drug response prediction model, which includes the alternately updated transformer model and GNN model.

7. The protein drug response prediction method based on EM fusion learning according to claim 1, characterized in that, The specific method of step 6) is as follows: Step 6.1) Input the amino acid sequence of the protein to be tested and the SMILES sequence of the drug to be tested into the trained EM fusion protein-drug response prediction model to obtain the predicted probability of the response between the protein to be tested and the drug to be tested. Step 6.2) Output the node label prediction results based on the reaction prediction results.

Images