RNA drug binding prediction method based on subgraph matching and domain adaptation

Abstract

The application discloses an RNA drug binding prediction method based on subgraph matching and field adaptation, and the method comprises the following steps: constructing a prediction model in an offline stage, pretraining the prediction model by adopting a drug-protein binding task, freezing part of model parameters, and fine-tuning the prediction model by adopting a drug-RNA binding training task; and in an online stage, the fine-tuned prediction model is used for real-time RNA drug binding prediction. The training mode combining subgraph matching and field adaptation can effectively avoid shortcut learning and enhance the generalization effect.

Description

Technical Field

[0001] This invention relates to a technology in the field of RNA and small drug molecules, specifically an RNA drug binding prediction method based on subgraph matching and domain adaptation. Background Technology

[0002] Deep learning-based drug-RNA interaction prediction currently suffers from potential shortcut learning, leading to poor performance of existing methods for novel RNAs and drugs not encountered during model training. Meanwhile, the lack of drug-RNA complex data is a key issue in training a deep learning model. In contrast, drug-protein complex data is abundant, which can support the training of deep learning models with better generalization capabilities. Summary of the Invention

[0003] This invention addresses the problem of prediction models struggling to predict newly discovered RNAs due to a lack of drug-RNA complex data. It proposes an RNA-drug binding prediction method based on subgraph matching and domain adaptation. The training method, which combines subgraph matching and domain adaptation, effectively avoids shortcut learning and enhances generalization performance.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to an RNA drug binding prediction method based on subgraph matching and domain adaptation. The method involves constructing a prediction model offline and pre-training it using a drug-protein binding task. Then, it freezes some model parameters and fine-tunes them using a drug-RNA binding training task. Finally, it uses the fine-tuned prediction model online to perform real-time RNA drug binding prediction.

[0006] The pre-training, based on protein structure, drug small molecule structure, and their binding energy values, establishes a drug-protein binding training task, specifically including:

[0007] S1.1 Collect numerical data on protein structure, drug small molecule structure and their binding energy.

[0008] S1.2. Read the protein PDB structure to obtain the chemical information and three-dimensional coordinates of each amino acid residue. Set each amino acid residue as a node in a graph network and calculate the Euclidean distance between them using the three-dimensional coordinates of each amino acid. At this point, there is an edge between the nodes corresponding to the two amino acids, where (x1, y1, z1) and (c2, y2, z2) are the three-dimensional coordinates of the two amino acid residues, respectively. The ESM model is then used to encode the protein residue information, ultimately yielding the protein graph structure and the initial characterization of the graph nodes.

[0009] S1.3. Use the RDKit software package to convert the drug small molecule SMILES string into a graph network structure, where small molecule atoms are graph network nodes and intermolecular chemical bonds are graph network edges. Use the UniMol model to encode the drug small molecule node features to obtain the initial representation of the small molecule graph nodes.

[0010] The prediction model includes a domain adaptation module, a feature extraction module, a subgraph matching module, and a prediction module. Specifically: the domain adaptation module, constructed using a multilayer perceptron, processes the initial representation of protein graph nodes, transferring the drug-protein binding knowledge learned by the model to the drug-RNA binding task; the feature extraction module extracts features from the protein graph structure using an isovariant graph neural network unit and from the small molecule graph structure using a graph convolutional neural network unit; the subgraph matching module, based on the multilayer perceptron, selects protein nodes to form key protein subgraphs and generates key protein subgraph representations, ensuring consistency between the matched key subgraphs and the actual binding pockets; and the prediction module, constructed using the multilayer perceptron, predicts the binding state between the key protein subgraph representations and the drug small molecule representations, obtaining the model prediction value.

[0011] The knowledge transfer refers to mapping the initial representation of the generated protein map and the initial representation of the generated RNA map to the same sample space before model training. For the initial representation of protein map nodes: Among them: MLP P For a multilayer perceptron, a p Initial characterization of protein graph nodes. These are the parameters of a multilayer perceptron network.

[0012] The aforementioned feature extraction of the protein graph structure refers to: while maintaining node symmetry and isovariability, updating the representation by aggregating the features of neighboring nodes, and then combining the protein graph node representations into a protein graph representation, specifically: Protein characterization was obtained. in: These are node features, coordinate features, and edge features, respectively, φ e φ h and φ x Each is a two-layer MLP, where C is the standardization constant, calculated using the following formula: p i For the i-th protein node, and To calculate intermediate products.

[0013] The aforementioned feature extraction of small molecule graph structures refers to: small molecule graph characterization ,in: Where: ReLU is the activation function. Let D be the adjacency matrix, and let D be the degree matrix of the adjacency matrix. Let W represent the features of layer l nodes, and W be the parameters of the neural network. This represents the number of nodes in the graph structure.

[0014] The protein key subgraph representation is obtained as follows: a protein subgraph node matching model is established using a multilayer perceptron; the protein graph representation and the drug small molecule graph representation are concatenated and input into the protein subgraph node matching model to obtain the node allocation matrix. This leads to the protein keystone graph representation G. sub =Z T G p [0], where: MLP is the protein subgraph node matching model, G P For protein mapping, G m For small molecule graph characterization, Softmax is used as the activation function, and a loss function that promotes node selection is set. I² is the identity matrix, A is the protein graph adjacency matrix, and minimizing... This allows the node allocation matrix Z to be close to [0,1] / [1,0].

[0015] The model predicted values Among them: G m For small molecule characterization of drugs, G sub For protein keystone characterization, θ cls Here are the parameters of the multilayer perceptron network, and the classification loss function of this multilayer perceptron. y represents the true value; and the loss function for predicting the task is then used.

[0016] The aforementioned frozen model parameters refer to saving the model parameters of the feature extraction module, subgraph matching module, and prediction module after pre-training.

[0017] The aforementioned drug-RNA binding training task refers to: reading the RNA PDB structure, obtaining the chemical information and three-dimensional coordinates of each RNA base, setting each base as a node in a graph network, and calculating the Euclidean distance between the bases based on their three-dimensional coordinates. At this point, an edge exists between the nodes corresponding to the two bases, where (x1, y1, z1) and (x2, y2, z2) are the three-dimensional coordinates of the two bases, respectively. The RNA-FM model is used to encode the RNA base information, ultimately obtaining the RNA graph structure and the initial representation of the RNA graph nodes. Then, the RDKit software package is used to convert the drug small molecule SMILES string into a graph network structure, where small molecule atoms serve as graph network nodes, and intermolecular chemical bonds serve as graph network edges. The UniMol model is used to encode the drug small molecule node features, obtaining the initial representation of the small molecule graph nodes.

[0018] The loading of the pre-trained prediction model refers to loading the model parameters of the feature extraction module, subgraph matching module and prediction module in the pre-trained model, and re-initializing the domain adaptation module.

[0019] The fine-tuning process involves establishing a drug-RNA binding training task based on RNA structure, drug small molecule structure, and their binding energy values. This is followed by loading a pre-trained prediction model, re-initializing the domain adaptation module, and fine-tuning the pre-trained model using the drug-RNA binding task to obtain a drug-RNA binding prediction model. The model's performance is then evaluated. Specifically, this includes:

[0020] a) The domain adaptation module is used to process the initial representation of RNA graph nodes in the drug-RNA binding training task. Initial representation of RNA graph nodes: Among them: MLP R For a multilayer perceptron, a R Initial characterization of protein graph nodes. These are the parameters of a multilayer perceptron network.

[0021] b) Use pre-trained isovariant graph neural network units to process RNA graph structures, and use pre-trained graph convolutional neural network units to process small molecule graph structures.

[0022] The aforementioned isomorphic graph neural network unit has four identical layers. Each layer includes: an edge feature processing MLP subunit, a dimension mapping MLP subunit, and a node feature processing MLP subunit. Specifically: the edge feature processing MLP subunit takes the node features and edge features of two nodes as input, and obtains the edge embedding vector after passing through the MLP; the dimension mapping MLP subunit takes the edge embedding vector as input, obtains the position weight after passing through the MLP, and obtains the coordinate features of the next layer by weighted summation of the relative distance between the coordinates of the two nodes; the node feature processing MLP subunit sums all edge embedding vectors related to the node as the edge feature of the node, and obtains the node features of the next layer by passing through the MLP with the node features and edge features as input.

[0023] The graph convolutional neural network unit has three identical layers. Each layer includes an MLP subunit, which takes node features and the graph adjacency matrix and its degree matrix as input, and obtains the node features of the next layer after passing through the MLP.

[0024] Step b specifically includes:

[0025] b.1) Feature extraction of RNA graph node representations is performed using an isovariant graph neural network. While maintaining symmetry and isovariance, the node representations are updated by aggregating features from neighboring nodes, thus combining the RNA graph node representations into a single RNA graph representation. in:

[0026] These are node features, coordinate features, and edge features, respectively, φ e φ h and φ x Each is a two-layer MLP, where C is the standardization constant, calculated using the following formula: R i For the i-th RNA node, and To calculate intermediate products.

[0027] b.2) After extracting features from the initial representations of small molecule graph nodes using a graph convolutional neural network, the small molecule graph node representations are combined into a small molecule graph representation. in: ReLU is the activation function. Let D be the adjacency matrix, and let D be the degree matrix of the adjacency matrix. Let W represent the features of layer l nodes, and W be the parameters of the neural network. This represents the number of nodes in the graph structure.

[0028] c) Based on the representations of each node in the RNA graph, while ensuring consistency between the matched key subgraphs and the actual binding pockets, a pre-trained multilayer perceptron-based subgraph matching module is used to select RNA nodes to form RNA key subgraphs, thereby obtaining RNA key subgraph representations. Specifically, an RNA subgraph node matching model is established using a multilayer perceptron, concatenating the RNA graph representation with the drug small molecule graph representation, and then inputting it into the RNA subgraph node matching model to obtain the node allocation matrix. Then, the RNA key subgraph representation G is obtained based on the node assignment matrix Z. sub =Z T G R [0], where: MLP is a protein subgraph node matching model. G R For RNA mapping, G mThis is used for small molecule characterization. Softmax is the activation function. A loss function that promotes node selection is set. Where: I2 is the identity 2D matrix, and A is the adjacency matrix of the RNA graph. Minimize This allows the node allocation matrix Z to be close to [0,1] / [1,0].

[0029] The consistency between the matched key subgraph and the true combined pocket refers to the consistency between the first column of the node assignment matrix Z and the true combined pocket Z. Y Applying MSE loss between them: The true combination of pocket vector Z Y The first column of the node allocation matrix Z has the same dimension. When the corresponding base belongs to the real binding pocket, the value is 1, otherwise it is 0.

[0030] This invention relates to a system for implementing the above-mentioned method, comprising: a data preprocessing unit, a model building unit, a model pre-training unit, and a model testing unit, wherein: the data preprocessing unit reads the three-dimensional coordinates of the protein to build a graph structure based on protein structure information, and encodes the protein graph structure using an ESM model to obtain the protein graph structure and node representation; based on RNA structure information, reads the three-dimensional coordinates of the RNA to build a graph structure, and encodes the RNA graph structure using an RNA-FM model to obtain the RNA graph structure and node representation; based on drug small molecule structure information, uses rdkit to convert the drug small molecule structure into a graph structure, and uses a UniMol model to encode the drug small molecule node features to obtain the initial representation of the small molecule graph nodes; according to... The protein-drug small molecule binding information is used for threshold screening to obtain drug-protein binding and non-binding binary labels. Similarly, RNA-drug small molecule binding information is used for threshold screening to obtain drug-RNA binding and non-binding binary labels. The model building unit establishes graph neural networks and multilayer perceptrons based on the number of network layers and dimensions. The model pre-training unit performs model pre-training based on the protein-drug small molecule graph structure and node representation information obtained from the data preprocessing unit, resulting in a pre-trained model. The model testing unit fine-tunes the pre-trained model based on the RNA-drug small molecule graph structure and node representation information obtained from the data preprocessing unit, resulting in a drug-RNA binding prediction model. Finally, predictions are made on the test set, and evaluation results are obtained.

[0031] Technical effect

[0032] This invention employs a pre-training-fine-tuning paradigm for model training, effectively alleviating the problem of insufficient drug-RNA complex data. Considering that molecules bind to binding pockets within RNA rather than the entire RNA strand, after feature extraction, a subgraph matching module is applied to identify which nodes correspond to binding pockets on the RNA backbone. Since binding affinity largely depends on RNA pocket features, a well-learned subgraph matching module can locate RNA pockets even without real pocket information, thus providing a better estimate of binding affinity. Attached Figure Description

[0033] Figure 1 This is a flowchart of the present invention;

[0034] Figure 2 This is a block diagram of the modules of the present invention;

[0035] Figure 3 This is a diagram showing the effect of a drug screening simulation experiment according to the present invention. Detailed Implementation

[0036] like Figure 1 As shown, this embodiment illustrates an RNA drug binding prediction method based on subgraph matching and domain adaptation, specifically including:

[0037] Step 1, establish as follows Figure 2 The prediction model shown includes a domain adaptation module, a feature extraction module, a subgraph matching module, and a prediction module.

[0038] Step 2: Establish a drug-protein binding training task based on the protein structure, drug small molecule structure and binding energy value, as a pre-training task for the prediction model established in Step 1.

[0039] Step 3: Construct initial characterizations of protein graph structures and drug small molecule graph structures based on protein structure data and drug small molecule structure data, respectively.

[0040] Step 4: Use the pre-training task established in Step 2 to pre-train the prediction model and save the model parameters of the feature extraction module, subgraph matching module, and prediction module.

[0041] Step 5: Establish a drug-RNA binding training task based on RNA structure, drug small molecule structure and binding energy values, as a fine-tuning task for the prediction model.

[0042] Step 6: Construct initial characterizations of RNA graph structures and drug small molecule graph structures based on RNA structure data and drug small molecule structure data, respectively.

[0043] Step 7: Load the pre-trained model parameters for the feature extraction module, subgraph matching module, and prediction module. Further train the prediction model using a fine-tuning task to obtain the drug-RNA binding prediction model. Use this model to predict the binding effect of RNA with small drug molecules.

[0044] Through specific practical experiments, the above method was run in a Linux environment with GPU training settings and random initialization parameters. Detailed information on the protein pre-training set, RNA training set, and RNA test set is shown in Table 1.

[0045] Table 1

[0046]

[0047] The accuracy of the experimental data reached 90%, with the drug screening simulation experiment showing the following results: Figure 2 The diagram illustrates the ranking of real drugs in various drug screening procedures. Drug screening simulation experiments demonstrate the effectiveness of this invention by comparing the performance of this method and others in drug screening. In drug screening, it is generally desirable to find the most promising compound from a set of compounds. For validation, the experiment hides a known-bound drug within a set of inactive compounds. Predictions are made for this set of RNAs and all drugs using this method and others. The effectiveness of the method is measured by the ranking of the predicted scores of naturally bound compounds. Therefore, a successful predictor will rank the naturally bound drug higher, while a random predictor will result in an average ranking of 50%. This embodiment uses RNA molecules from an RNA test set, selects drugs bound to them as candidates for screening, and constructs 36 inactive compounds for each drug using the DecoyFinder method to preserve common chemical properties. Each RNA and all other compounds are screened using this method and others.

[0048] Compared to existing technologies, this invention employs a pre-training-fine-tuning paradigm for model training, effectively alleviating the problem of insufficient drug-RNA complex data. Considering that molecules bind to binding pockets within RNA rather than the entire RNA strand, after the feature extraction module, a subgraph matching module is applied to identify which nodes correspond to binding pockets on the RNA backbone. Since binding affinity largely depends on RNA pocket features, a well-learned subgraph matching module can locate RNA pockets even without real pocket information, thus providing a better estimate of binding affinity.

[0049] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. An RNA drug binding prediction method based on subgraph matching and domain adaptation, characterized in that, After constructing a prediction model offline and pre-training it using a drug-protein binding task, the model parameters were frozen and fine-tuned using a drug-RNA binding training task. In the online stage, the fine-tuned prediction model was used to perform real-time RNA drug binding prediction. The aforementioned pre-training establishes a drug-protein binding training task based on protein structure, drug small molecule structure, and their binding energy values. The prediction model includes a domain adaptation module, a feature extraction module, a subgraph matching module, and a prediction module. Specifically: the domain adaptation module, built using a multilayer perceptron, processes the initial representation of protein graph nodes, transferring the drug-protein binding knowledge learned by the model to the drug-RNA binding task; the feature extraction module extracts features from the protein graph structure using an isovariant graph neural network unit and from the small molecule graph structure using a graph convolutional neural network unit; the subgraph matching module, based on the multilayer perceptron, selects protein nodes to form key protein subgraphs and generates key protein subgraph representations, ensuring consistency between the matched key subgraphs and the actual binding pockets; and the prediction module, also built using the multilayer perceptron, predicts the binding state between the key protein subgraph representations and the drug small molecule representations, obtaining the model prediction value. The aforementioned freezing of some model parameters refers to saving the model parameters of the feature extraction module, subgraph matching module, and prediction module after pre-training is completed; The fine-tuning refers to: after establishing a drug-RNA binding training task based on RNA structure, drug small molecule structure and binding energy values, loading the pre-trained prediction model and re-initializing the domain adaptation module, using the drug-RNA binding task to fine-tune the pre-trained model to obtain a drug-RNA binding prediction model, and evaluating the model performance. Loading the pre-trained prediction model means: loading the model parameters of the feature extraction module, subgraph matching module and prediction module in the pre-trained model, and re-initializing the domain adaptation module.

2. The RNA drug binding prediction method based on subgraph matching and domain adaptation according to claim 1, characterized in that, The aforementioned pre-training specifically includes: S1.1 Collect numerical data on protein structure, drug small molecule structure and their binding energy; S1.

2. Read the protein PDB structure to obtain the chemical information and three-dimensional coordinates of each amino acid residue. Set each amino acid residue as a node in a graph network and calculate the Euclidean distance between them using the three-dimensional coordinates of each amino acid. At that time, there is an edge between the nodes corresponding to the two amino acids, where: ( , , ), ( , , The three-dimensional coordinates of the two amino acid residues are shown below; then the ESM model is used to encode the protein residue information, and finally the protein graph structure and the initial characterization of the graph nodes are obtained. S1.

3. Use the RDKit software package to convert the drug small molecule SMILES string into a graph network structure, where: small molecule atoms are graph network nodes and intermolecular chemical bonds are graph network edges; use the UniMol model to encode the drug small molecule node features to obtain the initial representation of the small molecule graph nodes.

3. The RNA drug binding prediction method based on subgraph matching and domain adaptation according to claim 1, characterized in that, The knowledge transfer refers to mapping the initial representation of the generated protein map and the initial representation of the generated RNA map to the same sample space before model training. For the initial representation of protein map nodes: ,in: It is a multilayer perceptron. Initial characterization of protein graph nodes. These are the parameters of a multilayer perceptron network.

4. The RNA drug binding prediction method based on subgraph matching and domain adaptation according to claim 1, characterized in that, The aforementioned feature extraction of the protein graph structure refers to: while maintaining node symmetry and isovariability, updating the representation by aggregating the features of neighboring nodes, and then combining the protein graph node representations into a protein graph representation, specifically: , , , Protein characterization was obtained. ,in: , , These are node features, coordinate features, and edge features. , and Each is a two-layer MLP, where C is the standardization constant, calculated using the following formula: , For the i-th protein node, and To calculate intermediate products; The aforementioned feature extraction of small molecule graph structures refers to: small molecule graph characterization ,in: For activation function, It is an adjacency matrix. Let be the degree matrix of the adjacency matrix. for Layer node features For neural network parameters, This represents the number of nodes in the graph structure.

5. The RNA drug binding prediction method based on subgraph matching and domain adaptation according to claim 1, characterized in that, The protein key subgraph representation is obtained as follows: a protein subgraph node matching model is established using a multilayer perceptron; the protein graph representation and the drug small molecule graph representation are concatenated and input into the protein subgraph node matching model to obtain the node allocation matrix. This leads to the protein keystone characterization. ,in: This is a protein subgraph node matching model. For protein mapping, For small molecule characterization, Set a loss function to promote node selection as the activation function. , It is a unit second-order matrix. Given the protein graph adjacency matrix, minimize Allows for the allocation of nodes in a matrix. Approaching [0,1] / [1,0].

6. The RNA drug binding prediction method based on subgraph matching and domain adaptation according to claim 1, characterized in that, The aforementioned drug-RNA binding training task refers to: reading the RNA PDB structure, obtaining the chemical information and three-dimensional coordinates of each RNA base, setting each base as a node in a graph network, and calculating the Euclidean distance between the bases based on their three-dimensional coordinates. At that time, there is an edge between the nodes corresponding to the two bases, where: ( , , ), ( , , The three-dimensional coordinates of the two bases are given. The RNA-FM model is used to encode the RNA base information, and finally the RNA graph structure and the initial characterization of the RNA graph nodes are obtained. Then, the RDKit software package is used to convert the drug small molecule SMILES string into a graph network structure, in which: small molecule atoms are used as graph network nodes, intermolecular chemical bonds are used as graph network edges, and the UniMol model is used to encode the drug small molecule node features to obtain the initial characterization of the small molecule graph nodes.

7. The RNA drug binding prediction method based on subgraph matching and domain adaptation according to claim 1, characterized in that, The aforementioned fine-tuning specifically includes: a) Use the domain adaptation module to process the initial representation of RNA graph nodes in the drug-RNA binding training task; initial representation of RNA graph nodes: ,in: It is a multilayer perceptron. Initial characterization of protein graph nodes. These are the parameters of the multilayer perceptron network; b) Process RNA graph structures using pre-trained isotropic graph neural network units, and process small molecule graph structures using pre-trained graph convolutional neural network units, specifically including: b.1) Feature extraction of RNA graph node representations is performed using an isovariant graph neural network; while maintaining symmetry and isovariance, the representations of nodes are updated by aggregating features from neighboring nodes, and the RNA graph node representations are combined into an RNA graph representation. ,in: , , , , , , These are node features, coordinate features, and edge features. , and Each is a two-layer MLP, where C is the standardization constant, calculated using the following formula: , For the i-th RNA node, and To calculate intermediate products; b.2) After extracting features from the initial representations of small molecule graph nodes using a graph convolutional neural network, the small molecule graph node representations are combined into a small molecule graph representation. , For activation function, It is an adjacency matrix. Let be the degree matrix of the adjacency matrix. for Layer node features For neural network parameters, This represents the number of nodes in the graph structure. c) Based on the representations of each node in the RNA graph, while ensuring consistency between the matched key subgraphs and the actual binding pockets, a pre-trained multilayer perceptron-based subgraph matching module is used to select RNA nodes to form RNA key subgraphs, thereby obtaining RNA key subgraph representations. Specifically, an RNA subgraph node matching model is established using a multilayer perceptron, concatenating the RNA graph representation with the drug small molecule graph representation, and then inputting it into the RNA subgraph node matching model to obtain the node allocation matrix. Then, assign the matrix according to the nodes. RNA key component characterization ,in: A protein subgraph node matching model; RNA mapping, Characterization of small molecules; Set the activation function; set the loss function to promote node selection. ,in: It is a unit second-order matrix. For the adjacency matrix of the RNA graph; minimize Allows for the allocation of nodes in a matrix. Approaching [0,1] / [1,0].

8. The RNA drug binding prediction method based on subgraph matching and domain adaptation according to claim 1, characterized in that, The aforementioned isomorphic graph neural network unit has a four-layer identical structure. Each layer includes: an edge feature processing MLP subunit, a dimension mapping MLP subunit, and a node feature processing MLP subunit. Specifically: the edge feature processing MLP subunit takes the node features and edge features of two nodes as input, and obtains the edge embedding vector after passing through the MLP; the dimension mapping MLP subunit takes the edge embedding vector as input, obtains the position weight after passing through the MLP, and obtains the coordinate features of the next layer by weighted summation of the relative distance between the coordinates of two nodes; the node feature processing MLP subunit sums all edge embedding vectors related to the node as the edge feature of the node, and obtains the node features of the next layer by passing through the MLP with the node features and edge features as input. The graph convolutional neural network unit has three identical layers. Each layer includes an MLP subunit, which takes node features and the graph adjacency matrix and its degree matrix as input, and obtains the node features of the next layer after passing through the MLP.

9. The RNA drug binding prediction method based on subgraph matching and domain adaptation according to claim 7, characterized in that, The consistency between the matched key subgraph and the true combined pocket refers to: the matrix allocation at the nodes. The first column combines with the real pocket Applying MSE loss between them: The actual combination of pocket vectors With node allocation matrix The first column has the same dimensions. When the corresponding base belongs to the actual binding pocket, the value is 1, otherwise it is 0.

10. An RNA drug binding prediction system based on subgraph matching and domain adaptation that implements the method of any one of claims 1-9, characterized in that, include: The system comprises a data preprocessing unit, a model building unit, a model pretraining unit, and a model testing unit. Specifically: The data preprocessing unit reads the three-dimensional coordinates of proteins to build a graph structure based on protein structure information, and encodes the protein graph structure using an ESM model to obtain the protein graph structure and node representations. Based on RNA structure information, it reads the three-dimensional coordinates of RNA to build a graph structure, and encodes the RNA graph structure using an RNA-FM model to obtain the RNA graph structure and node representations. Based on drug small molecule structure information, it uses rdkit to convert the drug small molecule structure into a graph structure, and uses a UniMol model to encode the drug small molecule node features to obtain the initial representation of the small molecule graph nodes. Based on the protein and drug small molecule structure... The data preprocessing unit performs threshold screening based on the combined information to obtain binary labels for drug-protein binding and non-binding. Similarly, it performs threshold screening based on RNA-drug small molecule binding information to obtain binary labels for drug-RNA binding and non-binding. The model building unit establishes graph neural networks and multilayer perceptrons based on the number of network layers and dimensions. The model pre-training unit performs model pre-training based on the protein and drug small molecule graph structure and node representation information obtained from the data preprocessing unit to obtain a pre-trained model. The model testing unit fine-tunes the pre-trained model based on the RNA and drug small molecule graph structure and node representation information obtained from the data preprocessing unit to obtain a drug-RNA binding prediction model. Finally, it predicts on the test set and obtains the evaluation results.

Images