System and method for predicting binding affinity of radiopharmaceutical

Abstract

A system for predicting the binding affinity of a radiopharmaceutical, according to the present invention, comprises a feature extraction model for extracting a feature related to a radiopharmaceutical connected to a linker-drug structure with respect to a nuclide present in a radioactive decay chain by using chemical properties of an atom and physical properties of radiation related to the radiopharmaceutical as inputs, wherein the physical properties of the radiation include at least any one among the charge amount, mass, energy, and radioisotope decay constant of the radiation.

Description

System and method for predicting binding affinity of radiopharmaceuticals

[0001] The present invention relates to a system and method for predicting the binding affinity of a radiopharmaceutical, and more particularly, to a system and method for predicting the binding affinity of a radiopharmaceutical that predicts the binding affinity of a radiopharmaceutical used for the diagnosis and treatment of diseases using radiation.

[0002] Recent advancements in artificial intelligence (AI) technology have led to the development of AI models that assist in the screening and development of general drugs. For example, AI models applied in the early stages of drug development can predict drug responses to specific cell lines and binding affinity for target proteins. In other words, AI models provide predictions of a compound's binding affinity for its target and drug response during the new drug development process. Furthermore, analysis of the learned parameters within the prediction model can provide information on structural features and binding sites that critically influence binding. Therefore, AI models offer significant time and cost advantages in the new drug development process.

[0003] Meanwhile, radiopharmaceuticals are substances labeled with radioactive isotopes that emit radiation, enabling the use of radiation for the diagnosis and treatment of diseases. These radiopharmaceuticals can destroy or visualize cells in a patient's body based on radiation. However, research and development in virtual screening models or artificial intelligence (AI) models for new drug development in the field of radiopharmaceuticals is lacking. Therefore, there is a growing need for AI deep learning models that can predict drug response and binding of radiopharmaceuticals to target proteins.

[0004] The purpose of the present invention is to provide a system and method for predicting the binding affinity of a radiopharmaceutical by predicting structural features and binding positions through drug response prediction for the radiopharmaceutical.

[0005] The system for predicting the binding affinity of a radiopharmaceutical according to the present invention includes a feature extraction model that extracts features of the radiopharmaceutical linked to a linker-drug structure for a nuclide existing in a radioactive decay chain by inputting chemical characteristics of an atom of the radiopharmaceutical and physical characteristics of radiation, and the physical characteristics of the radiation include at least one of charge, mass, energy, and decay constant of a radioactive isotope of the radiation.

[0006] The chemical characteristics of the above atoms and the physical characteristics of the above radiation are input into the feature extraction model as node embedding, and the feature extraction model can extract features connected to the linker-drug structure for all nuclides existing in the radioactive decay chain, including a graph neural network (GNN).

[0007] The above feature extraction model is configured with GCN (Graph Convolution Network), ReLU (Rectified Linear Unit), BatchNorm, Dropout, GCN, ReLU, BatchNorm, Dropout, GCN, ReLU, BatchNorm, Dropout, Global pool, and Linear layers, and can extract the features of the radiopharmaceutical.

[0008] The above feature extraction model can be applied to a general drug response prediction model and a general drug affinity prediction model, so that the general drug response prediction model and the general drug affinity prediction model can perform processing on the radiopharmaceutical.

[0009] The above radiopharmaceutical binding force prediction system may further include a radiopharmaceutical response prediction model that predicts a reaction between the radiopharmaceutical drug and the cell based on the features of the radiopharmaceutical extracted from the feature extraction model and multi-omics of the cell line.

[0010] The above radiopharmaceutical response prediction model can extract features for the cell line by performing processing on mutation representation, gene expression representation, and DNA methylation representation for the cell line.

[0011] The above radiopharmaceutical response prediction model can estimate the half maximal inhibitory concentration (IC50) by connecting the characteristics of the radiopharmaceutical and the characteristics of the cell line and inputting the connected characteristics into a multilayer perceptron.

[0012] In the mutation expression for the above cell line, the mutation tensor is input into a 1-dimensional convolutional neural network (1D CNN) to extract features for the mutation, and the 1-dimensional convolutional neural network may include 1D Conv, Tanh, Max pooling, 1D Conv, ReLU, Max pooing, Flatten, Linear, ReLU, and Dropout layers.

[0013] In the gene expression level expression and DNA methylation level expression for the above cell line, the gene expression level tensor and the DNA methylation tensor are input into the gene expression level expression model and the DNA methylation level expression model, respectively, to extract the gene expression level feature and the DNA methylation level feature, and the gene expression level expression model and the DNA methylation level expression model may include a multilayer perceptron.

[0014] The above multilayer perceptron may include Linear, Tanh, BatchNorm, Dropout, Linear, and ReLU layers.

[0015] The above radiopharmaceutical binding affinity prediction system may further include a radiopharmaceutical affinity prediction model that predicts binding affinity between the radiopharmaceutical and the protein based on the features of the radiopharmaceutical extracted from the feature extraction model and the features of the protein.

[0016] The above radiopharmaceutical affinity prediction model can predict binding affinity including at least one of an inhibition constant (Ki) or a dissociation constant (Kd) between the radiopharmaceutical and the protein.

[0017] The above radiopharmaceutical affinity prediction model applies the radiopharmaceutical features and protein embedding to an attention mechanism (Attention map) to extract weighted radiopharmaceutical features and weighted protein features, and connects the weighted radiopharmaceutical features and the weighted protein features to estimate the binding affinity between the radiopharmaceutical and the protein.

[0018] The above protein embedding can be extracted by preprocessing the amino acid sequence of the protein, applying the amino acid to an embedding layer to express it as a vector, and applying the expressed vector to a long short-term memory (LSTM) model.

[0019] In the vector representation of the above protein embedding, the amino acid sequence of the protein can be data-ized, and the amino acids can be mapped to random integers to be expressed as a 32-dimensional vector through the embedding layer.

[0020] In the estimation of the above binding affinity, the connected representation of the weighted radiopharmaceutical features and the weighted protein features can be input into a multilayer perceptron to estimate the inhibition constant or dissociation constant between the radiopharmaceutical and the protein.

[0021] Meanwhile, the system for predicting the binding affinity of a radiopharmaceutical according to the present invention includes a feature extraction model that extracts features of the radiopharmaceutical linked to a linker-drug structure for a nuclide existing in a radioactive decay chain by inputting chemical characteristics of an atom of the radiopharmaceutical and physical characteristics of radiation, and a radiopharmaceutical reaction prediction model that predicts a reaction between the radiopharmaceutical drug and the cell based on multi-omics of a cell line, and a radiopharmaceutical affinity prediction model that predicts binding affinity between the radiopharmaceutical and the protein based on characteristics of a protein, wherein the physical characteristics of the radiation include at least one of a charge, a mass, an energy, and a decay constant of a radioisotope of the radiation.

[0022] Meanwhile, the method for predicting the binding force of a radiopharmaceutical according to the present invention includes a step of inputting chemical characteristics of atoms of the radiopharmaceutical and physical characteristics of radiation into a feature extraction model to extract features of the radiopharmaceutical linked to a linker-drug structure for nuclides existing in a radioactive decay chain, wherein the physical characteristics of the radiation include at least one of charge, mass, energy, and decay constant of the radioactive isotope of the radiation.

[0023] The method for predicting the binding affinity of the radiopharmaceutical may further include a step in which a radiopharmaceutical response prediction model predicts a reaction between the radiopharmaceutical drug and the cell based on the features of the radiopharmaceutical extracted from the feature extraction model and multi-omics of the cell line.

[0024] The method for predicting the binding affinity of the radiopharmaceutical may further include a step in which a radiopharmaceutical affinity prediction model predicts the binding affinity between the radiopharmaceutical and the protein based on the features of the radiopharmaceutical and the features of the protein extracted from the feature extraction model.

[0025] The system and method for predicting the binding force of a radioactive pharmaceutical according to the present invention have the following effects.

[0026] First, the present invention has the effect of solving the problem that conventional drug-cell line response prediction models and drug-protein binding affinity prediction models are difficult to apply to radiopharmaceuticals.

[0027] Second, the present invention is capable of predicting cell line responses to general drugs and binding affinity to proteins based on a graph neural network, and includes the effect of distinguishing radiopharmaceuticals to which isotopes are bound by efficiently reflecting the structure of radiopharmaceuticals that changes through a radioactive decay chain.

[0028] The technical effects of the present invention are not limited to the effects mentioned above, and other technical effects not mentioned will be clearly understood by those skilled in the art from the description below.

[0029] Figure 1 is a conceptual diagram schematically illustrating a system for predicting the binding power of a radioactive pharmaceutical according to this embodiment.

[0030] Figure 2 is a conceptual diagram schematically showing a reaction prediction model of a radiopharmaceutical binding force prediction system according to this embodiment.

[0031] Figure 3 is a conceptual diagram schematically showing an affinity prediction model of a radiopharmaceutical binding force prediction system according to this embodiment.

[0032] Figure 4 is a conceptual diagram showing the chemical change of a radiopharmaceutical according to the radioactive decay chain process and the node-level embedding extraction process for the same in the feature extraction model of the radiopharmaceutical binding force prediction system according to the present embodiment.

[0033] Figure 5 is a conceptual diagram showing a radiopharmaceutical reaction prediction model of a radiopharmaceutical binding force prediction system according to this embodiment.

[0034] Figure 6 is a conceptual diagram showing a radiopharmaceutical affinity prediction model of a radiopharmaceutical binding affinity prediction system according to this embodiment.

[0035] Figure 7 is a test data scatter plot of a radiopharmaceutical reaction prediction model installed in a radiopharmaceutical prediction system according to the present embodiment.

[0036] Figure 8 is a test data scatter plot of a radiopharmaceutical affinity prediction model installed in a radiopharmaceutical binding force prediction system according to the present embodiment.

[0037] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings. However, the present embodiments are not limited to the embodiments disclosed below and may be implemented in various forms. These embodiments are provided only to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention. The shapes of elements in the drawings may be exaggerated for clearer explanation, and elements indicated by the same reference numerals in the drawings represent the same elements.

[0038] Figure 1 is a conceptual diagram schematically illustrating a system for predicting the binding strength of a radioactive pharmaceutical according to the present embodiment.

[0039] As illustrated in Fig. 1, the radiopharmaceutical binding affinity prediction system (1000, hereinafter referred to as the prediction system) according to the present embodiment can predict the drug response of the radiopharmaceutical and the binding affinity with the target protein.

[0040] Typically, artificial intelligence models for predicting drug response and binding affinity are regression models that estimate labels such as drug response or binding affinity through the molecular structural features of the drug and the chemical properties of the elements.

[0041] However, radiopharmaceuticals consist of a structure consisting of a conventional drug, a radioisotope, and a linker for labeling the drug. Therefore, predicting drug response and binding affinity for radiopharmaceuticals requires considering the structure of the radiopharmaceutical, the radioactive material, and the physical properties of the emitted radiation.

[0042] In particular, the chemical properties of radiopharmaceuticals can change depending on the radioactive decay of the radioisotope. Therefore, conventional AI deep learning models applied to general pharmaceuticals have limitations in predicting the binding affinity of radiopharmaceuticals.

[0043] However, the prediction system (1000) according to the present embodiment can predict the structural characteristics and binding location of the radiopharmaceutical by considering the characteristics of the radiopharmaceutical.

[0044] For example, the prediction system (1000) may include a storage device (100) in which an artificial intelligence deep learning model is stored, and a processor (200) that performs a function of predicting structural characteristics and binding locations of a radiopharmaceutical based on the artificial intelligence deep learning model. The storage device (100) and the processor (200) may be implemented as a single computer system or multiple systems, and the types of the storage device (100) and the processor (200) are not limited.

[0045] Meanwhile, the prediction system (1000) may use artificial intelligence deep learning models for predicting structural features and binding positions of radiopharmaceuticals, such as a reaction prediction model (M100), an affinity prediction model (M200), a feature extraction model (M300), a radiopharmaceutical reaction prediction model (M400), and a radiopharmaceutical affinity prediction model (M500).

[0046] Figure 2 is a conceptual diagram schematically illustrating a reaction prediction model of a radiopharmaceutical binding force prediction system according to this embodiment.

[0047] As illustrated in Fig. 2, the reaction prediction model (M100) according to the present embodiment is a deep learning model capable of predicting the reaction between a drug and a cell.

[0048] For example, the response prediction model (M100) may be a deep learning model that predicts the reaction between a general drug and cells. This response prediction model (M100) may be a regression model that estimates the half maximal inhibitory concentration (IC50) using the molecular structural features of the drug and the multi-omics features of the cell line, and a graph neural network (GNN) may be used. Here, the multi-omics features are a technology that simultaneously analyzes and integrates various types of biological data, allowing for understanding the molecular structural features of the drug and the interaction between the cell line. The half maximal inhibitory concentration (IC50) is an indicator used to measure the efficacy of a drug and refers to the concentration required for a specific drug to take effect biologically.

[0049] The response prediction model (M100) performs drug representation for drugs to extract drug features, and performs mutation representation, gene expression representation, and DNA methylation representation for cell lines to predict the response between drugs and cells.

[0050] The response prediction model (M100) uses graph data that reflects the molecular structural characteristics of drugs in drug representation. To reflect the molecular structural characteristics of drugs, the graph data uses the atoms that compose the drug as nodes, and the molecular bonds between atoms as edges. For example, the model for drug representation (M110) is configured with a Graph Convolution Network (GCN), Rectified Linear Unit (ReLU), BatchNorm, Dropout, GCN, ReLU, BatchNorm, Dropout, GCN, ReLU, BatchNorm, Dropout, Global Pool, and Linear layers, enabling it to estimate drug characteristics.

[0051] And the response prediction model (M100) performs mutation expression, gene expression level expression, and DNA methylation expression for cell lines based on the multi-omics features of the cell lines.

[0052] For mutation representation of cell lines, the mutation tensor can be input into a 1-dimensional convolutional neural network (1D CNN). For example, a model for mutation representation (M120) can estimate mutation features by constructing 1D Convolution, Tanh, Max pooling, 1D Convolution, ReLU, Max pooling, Flatten, Linear, ReLU, and Dropout layers.

[0053] And in the gene expression level expression and DNA methylation level expression for cell lines, the gene expression level tensor and the DNA methylation tensor are applied to the gene expression level expression model (M130) and the DNA methylation level expression model (M140). The gene expression level expression model (M130) and the DNA methylation level expression model (M140) may include a multilayer perceptron (MLP). For example, the multilayer perceptron model is configured with each of the Linear, Tanh, BatchNorm, Dropout, Linear, and ReLU layers, respectively, and can estimate the gene expression level features and DNA methylation features for the cell line.

[0054] Afterwards, the response prediction model (M100) can estimate the half-maximal inhibitory concentration (IC50) by connecting drug features, mutation features, gene expression level features, and DNA methylation features and inputting the connected features into a multilayer perceptron.

[0055] Meanwhile, in training the response prediction model (M100), sensitivity data (IC50) between existing drugs and cell lines and multi-omics features for cell lines can be utilized. For example, in training the response prediction model (M100), sensitivity data (IC50) can be obtained from the Genomics of Drug Sensitivity in Cancer (GDSC) library, and multi-omics features for 549 cell lines in the Cancer Cell Line Encyclopedia (CCLE) library can be obtained. In addition, for drugs, the RDKit library was utilized to structure a graph for 461 drugs that have PubChem compound IDs in the GDSC drug information.

[0056] Figure 3 is a conceptual diagram schematically illustrating an affinity prediction model of a radiopharmaceutical binding force prediction system according to this embodiment.

[0057] As illustrated in FIG. 3, the affinity prediction model (M200) according to the present embodiment is a deep learning model capable of predicting the binding affinity between a drug and a protein.

[0058] For example, an affinity prediction model (M200) may be a deep learning model that predicts the binding affinity between a common drug and a protein. This affinity prediction model (M200) can predict binding affinity, such as the inhibition constant (Ki) or dissociation constant (Kd), between two substances based on the structural characteristics of the drug and the protein. Graph neural networks may be used.

[0059] The affinity prediction model (M200) can predict binding affinity by performing drug representation for drugs and protein representation for proteins.

[0060] The affinity prediction model (M200) can use graph data that reflects molecular structural features in drug representation, similar to the response prediction model (M100). Graph data uses the atoms that make up the drug as nodes and the molecular bonds between atoms as edges to reflect the molecular structural features of the drug. For example, the model for drug representation (M210) is configured with GCN, ReLU, BatchNorm, Dropout, GCN, ReLU, BatchNorm, Dropout, GCN, ReLU, BatchNorm, Dropout, Global pool, and Linear layers to estimate drug characteristics.

[0061] And the affinity prediction model (M200) performs protein expression by considering the protein structure.

[0062] In protein representation, the amino acid sequence of a protein is digitized. Each amino acid mapped to a random integer is input into a protein representation model (M220), which then uses an embedding layer to represent it as a 32-dimensional vector. The resulting vector is then applied to a Long Short-Term Memory (LSTM) model to extract protein features.

[0063] Afterwards, the affinity prediction model (M200) can estimate the binding affinity (Ki, Kd) such as the inhibition constant or dissociation constant between the drug and the protein by connecting the drug features and protein features and inputting the connected expression into a multilayer perceptron.

[0064] Meanwhile, the affinity prediction model (M200) can be trained using existing data on drug-protein binding affinities. For example, 2,407,381 drug-protein binding affinity pairs from the BindingDB library can be used for training the affinity prediction model (M200). Furthermore, training can be performed on 522,770 pairs labeled with inhibition constants to model the affinity prediction model (M200).

[0065] Meanwhile, the prediction system (1000) according to the present embodiment must predict the binding affinity of a radiopharmaceutical. The prediction system (1000) can efficiently reflect the structure of a radiopharmaceutical that changes through the radioactive decay chain, thereby distinguishing radiopharmaceuticals with bound isotopes.

[0066] Figure 4 is a conceptual diagram showing the chemical change of a radiopharmaceutical according to the radioactive decay chain process and the node-level embedding extraction process for the same in a feature extraction model of a radiopharmaceutical binding force prediction system according to this embodiment.

[0067] As illustrated in Fig. 4, the feature extraction model (M300) according to the present embodiment extracts features of radioactive pharmaceuticals.

[0068] Radiopharmaceuticals are compounds composed of a radioisotope, a linker, and a drug. The primary difference between radiopharmaceuticals and conventional drugs lies in the presence or absence of a radioisotope, and they possess two distinct characteristics. First, the interaction between radiation and matter must be considered. Second, because radioisotopes decay into other elements through radioactivity, the chemical structure of the drug changes over time.

[0069] Accordingly, in conventional drug-related prediction models, the physical properties of the radiation emitted are used as additional inputs for the node embedding of radioisotopes, along with the chemical characteristics of each atom used as node embedding. These physical properties may include the charge, mass, energy, and decay constant of the radiation, as well as the radioisotope's decay constant.

[0070] And by applying node embedding that reflects the chemical characteristics of atoms and the physical properties of radiation to a graph neural network, features of radiopharmaceuticals connected to the linker-drug structure for all nuclides existing in the radioactive decay chain are extracted. For example, the feature extraction model for radiopharmaceuticals (M300) is composed of GCN (Graph Convolution Network), ReLU (Rectified Linear Unit), BatchNorm, Dropout, GCN, ReLU, BatchNorm, Dropout, GCN, ReLU, BatchNorm, Dropout, Global pool, and Linear layers, and can estimate the features of radiopharmaceuticals.

[0071] Meanwhile, the feature extraction model (M300) can be applied to the response prediction model (M100) and the affinity prediction model (M200), so that the response prediction model (M100) and the affinity prediction model (M200), which perform response prediction and affinity prediction for general drugs, can perform response prediction and affinity prediction for radiopharmaceuticals.

[0072] Figure 5 is a conceptual diagram showing a radiopharmaceutical reaction prediction model of a radiopharmaceutical binding force prediction system according to this embodiment.

[0073] As illustrated in FIG. 5, the radiopharmaceutical reaction prediction model (M400) according to the present embodiment can be prepared by adjusting the reaction prediction model (M100) of FIG. 2.

[0074] For example, a radiopharmaceutical response prediction model (M400) may be a deep learning model that predicts the interaction between a radiopharmaceutical and cells. This radiopharmaceutical response prediction model (M400) can estimate the half-maximal inhibitory concentration (HMIC) using the radiopharmaceutical features estimated from the feature extraction model (M300) and the multi-omics features of the cell line.

[0075] That is, the radiopharmaceutical characteristics can be estimated by applying the feature extraction model (M300) to the model (M110) for extracting pharmaceutical characteristics from the response prediction model (M100) of Fig. 2. In addition, the radiopharmaceutical response prediction model (M) performs mutation expression, gene expression level expression, and DNA methylation expression for the cell line based on the multi-omics characteristics of the cell line.

[0076] In the mutation expression for a cell line, the mutation tensor can be input into a mutation expression model (M120) that can be prepared as a 1-dimensional convolutional neural network (1D CNN). And in the gene expression level expression and DNA methylation level expression for a cell line, the gene expression tensor and DNA methylation tensor can be input into a gene expression level expression model (M130) and a DNA methylation expression model (M140) that are prepared as a multilayer perceptron (MLP).

[0077] Next, the radiopharmaceutical response prediction model (M400) connects radiopharmaceutical features, gene expression features, and DNA methylation features. The radiopharmaceutical response prediction model (M400) then inputs the connected features into a multilayer perceptron to estimate the half-maximal inhibitory concentration (IC50), thereby predicting the response between the radiopharmaceutical and the cell line.

[0078] Figure 6 is a conceptual diagram showing a radiopharmaceutical affinity prediction model of a radiopharmaceutical binding force prediction system according to the present embodiment.

[0079] As illustrated in FIG. 6, the radiopharmaceutical affinity prediction model (M500) according to the present embodiment can be prepared by adjusting the affinity prediction model (M200) of FIG. 3.

[0080] For example, a radiopharmaceutical affinity prediction model (M500) may be a deep learning model that predicts the binding affinity between a radiopharmaceutical and a protein. This radiopharmaceutical affinity prediction model (M500) can predict binding affinity, such as the inhibition constant (Ki) or dissociation constant (Kd), between two substances by using the radiopharmaceutical features estimated from the feature extraction model (M300) and the structural features of the protein.

[0081] That is, in the affinity prediction model (M200) of FIG. 3, the feature extraction model (M300) can be applied to the model (M210) for extracting pharmaceutical features to estimate radiopharmaceutical features. Furthermore, the radiopharmaceutical affinity prediction model (M500) can extract weighted radiopharmaceutical features and weighted protein features by applying radiopharmaceutical features and protein embedding to an attention mechanism (Attention map). Here, protein embedding digitizes the amino acid sequence of a protein. Furthermore, each amino acid mapped to a random integer can be expressed as a 32-dimensional vector through an embedding layer, and the expressed vector can be applied to a short-term and long-term model to extract protein embedding. Furthermore, the attention weights between atoms and amino acids according to the attention mechanism can be used to evaluate the influence on the binding affinity of radioisotopes and predict binding positions through dimension-wise sum operations.

[0082] Afterwards, the radiopharmaceutical affinity prediction model (M500) connects the weighted radiopharmaceutical features and the weighted protein features, and inputs the connected expression into a multilayer perceptron to estimate the binding affinity (Ki, Kd) such as the inhibition constant or dissociation constant between the radiopharmaceutical and the protein.

[0083] Meanwhile, the performance of the radiopharmaceutical response prediction model (M400) and the radiopharmaceutical affinity prediction model (M500) applied with the feature extraction model (M300) was investigated.

[0084] FIG. 7 is a test data scatter plot of a radiopharmaceutical reaction prediction model installed in a radiopharmaceutical prediction system according to the present embodiment, and FIG. 8 is a test data scatter plot of a radiopharmaceutical affinity prediction model installed in a radiopharmaceutical binding force prediction system according to the present embodiment.

[0085] As shown in FIGS. 7 and 8, tests were conducted on a radiopharmaceutical reaction prediction model (M400) and a radiopharmaceutical affinity prediction model (M500) to which a feature extraction model (M300) according to the present embodiment was applied.

[0086] In testing the radiopharmaceutical response prediction model (M400), training was performed using data from the GDSC library. The evaluation results showed that the coefficient of determination (R²) for the normalized ln(IC50), used as a label during the training process of the radiopharmaceutical response prediction model (M400), was 0.7016, and the mean absolute error (MAE) was 0.02729. This confirms that the radiopharmaceutical response prediction model (M400) can predict the response between radiopharmaceuticals and cell lines with high performance.

[0087] And in the test for the radiopharmaceutical affinity prediction model (M500), training was conducted using data from the BindingDB library. As a result of the evaluation, the radiopharmaceutical affinity prediction model (M500) showed an R² of approximately 0.3863 for ln(Ki), which was used as a label during the training process, and an average absolute error of approximately 2.324, confirming that the radiopharmaceutical affinity prediction model (M500) can predict the binding affinity between radiopharmaceuticals and proteins with high performance.

[0088] In this way, the system and method for predicting the binding force of a radiopharmaceutical according to the present invention have the following effects.

[0089] First, the present invention has the effect of solving the problem that conventional drug-cell line response prediction models and drug-protein binding affinity prediction models are difficult to apply to radiopharmaceuticals.

[0090] Second, the present invention is capable of predicting cell line responses to general drugs and binding affinity to proteins based on a graph neural network, and includes the effect of distinguishing radiopharmaceuticals to which isotopes are bound by efficiently reflecting the structure of radiopharmaceuticals that changes through a radioactive decay chain.

[0091] The embodiments of the present invention described above and illustrated in the drawings should not be construed as limiting the technical concept of the present invention. The scope of protection of the present invention is limited only by the matters set forth in the claims, and those skilled in the art will be able to make various improvements and modifications to the technical concept of the present invention. Accordingly, such improvements and modifications, as long as they are obvious to those skilled in the art, will fall within the scope of protection of the present invention.

Claims

1. Includes a feature extraction model that extracts features of the radiopharmaceutical linked to a linker-drug structure for nuclides existing in a radioactive decay chain by inputting chemical characteristics of atoms and physical characteristics of radiation for the radiopharmaceutical. The physical properties of the above radiation are A system for predicting the binding force of a radiopharmaceutical, wherein the system comprises at least one of charge, mass, energy, and decay constant of a radioactive isotope of the radiation.

2. In paragraph 1, The chemical characteristics of the above atoms and the physical properties of the above radiation It is input into the above feature extraction model as node embedding, The above feature extraction model A system for predicting the binding affinity of a radiopharmaceutical, characterized in that it extracts features associated with a linker-drug structure for all nuclides present in the radioactive decay chain, including a graph neural network (GNN).

3. In paragraph 1, The above feature extraction model A system for predicting the binding affinity of a radiopharmaceutical, characterized in that it extracts the features of the radiopharmaceutical by providing a GCN (Graph Convolution Network), ReLU (Rectified Linear Unit), BatchNorm, Dropout, GCN, ReLU, BatchNorm, Dropout, GCN, ReLU, BatchNorm, Dropout, Global pool, and Linear layer.

4. In paragraph 1, The above feature extraction model A system for predicting the binding affinity of a radiopharmaceutical, characterized in that the system is applied to a reaction prediction model of a general drug and an affinity prediction model of a general drug, and the reaction prediction model of the general drug and the affinity prediction model of the general drug perform processing for the radiopharmaceutical.

5. In paragraph 1, A radiopharmaceutical binding force prediction system further comprising a radiopharmaceutical response prediction model that predicts a reaction between the radiopharmaceutical drug and the cell based on the features of the radiopharmaceutical extracted from the feature extraction model and multi-omics of the cell line.

6. In paragraph 5, The above radiopharmaceutical reaction prediction model A system for predicting binding affinity of a radiopharmaceutical, characterized in that it extracts features for the cell line by performing processing on mutation representation, gene expression representation, and DNA methylation representation for the cell line.

7. In paragraph 6, The above radiopharmaceutical reaction prediction model Linking the characteristics of the above radioactive pharmaceutical with the characteristics of the above cell line, A radiopharmaceutical binding affinity prediction system characterized by estimating the half maximal inhibitory concentration (IC50) by inputting connected features into a multilayer perceptron.

8. In paragraph 6, In the expression of mutations for the above cell lines, The mutation tensor is input into a 1-dimensional convolutional neural network (1D CNN) to extract features for the mutation. The above one-dimensional convolutional neural network A system for predicting the binding affinity of a radiopharmaceutical, characterized by including 1D Conv, Tanh, Max pooling, 1D Conv, ReLU, Max pooing, Flatten, Linear, ReLU and Dropout layers.

9. In paragraph 6, In the expression of gene expression level and DNA methylation expression for the above cell lines, Gene expression tensor and DNA methylation tensor are input into the gene expression expression model and DNA methylation expression model, respectively, to extract the gene expression features and the DNA methylation features. The above gene expression level expression model and the above DNA methylation expression model A system for predicting binding affinity of a radiopharmaceutical, characterized by including a multilayer perceptron.

10. In paragraph 9, The above multilayer perceptron A radiopharmaceutical binding affinity prediction system characterized by including Linear, Tanh, BatchNorm, Dropout, Linear and ReLU layers.

11. In paragraph 1, A radiopharmaceutical binding affinity prediction system further comprising a radiopharmaceutical affinity prediction model that predicts binding affinity between the radiopharmaceutical and the protein based on features of the radiopharmaceutical extracted from the feature extraction model and features of the protein.

12. In paragraph 11, The above radiopharmaceutical affinity prediction model A system for predicting binding affinity of a radiopharmaceutical, characterized in that it predicts binding affinity including at least one of an inhibition constant (Ki) or a dissociation constant (Kd) between the radiopharmaceutical and the protein.

13. In paragraph 11, The above radiopharmaceutical affinity prediction model By applying the above radiopharmaceutical features and protein embedding to the attention mechanism (Attention map), weighted radiopharmaceutical features and weighted protein features are extracted. A system for predicting binding affinity of a radiopharmaceutical, characterized in that the binding affinity between the radiopharmaceutical and the protein is estimated by connecting the radiopharmaceutical features to which the weights are applied and the protein features to which the weights are applied.

14. In paragraph 13, The above protein embedding is Preprocessing the amino acid sequence of the above protein, The above amino acids are applied to the embedding layer and expressed as vectors, A radiopharmaceutical binding force prediction system characterized in that the vector expressed above is extracted by applying it to a long short term memory (LSTM) model.

15. In paragraph 14, In the vector representation of the above protein embedding, Data the amino acid sequence of the above protein, A system for predicting the binding affinity of a radiopharmaceutical, characterized in that the above amino acids are mapped to random integers and expressed as a 32-dimensional vector through the embedding layer.

16. In paragraph 11, In the above estimation of binding affinity, A system for predicting binding affinity of a radiopharmaceutical, characterized in that the linked representation of the weighted radiopharmaceutical features and the weighted protein features are input into a multilayer perceptron to estimate an inhibition constant or dissociation constant between the radiopharmaceutical and the protein.

17. A feature extraction model that extracts features of the radiopharmaceutical linked to a linker-drug structure for nuclides existing in a radioactive decay chain by inputting chemical characteristics of atoms and physical characteristics of radiation for the radiopharmaceutical; A radiopharmaceutical response prediction model that predicts the reaction between the radiopharmaceutical drug and the cell based on the features of the radiopharmaceutical extracted from the feature extraction model and the multi-omics of the cell line; and A radiopharmaceutical affinity prediction model is included that predicts the binding affinity between the radiopharmaceutical and the protein based on the features of the radiopharmaceutical extracted from the feature extraction model and the features of the protein. The physical properties of the above radiation are A system for predicting binding strength of a radiopharmaceutical, characterized in that it includes at least one of charge, mass, energy, and decay constant of a radioactive isotope of the radiation.

18. A step of inputting chemical characteristics of atoms and physical characteristics of radiation for a radiopharmaceutical into a feature extraction model and extracting features for the radiopharmaceutical linked to a linker-drug structure for nuclides existing in a radioactive decay chain, The physical properties of the above radiation are A method for predicting the binding force of a radiopharmaceutical, wherein the radiopharmaceutical comprises at least one of charge, mass, energy, and decay constant of a radioisotope.

19. In Article 18, A method for predicting binding affinity of a radiopharmaceutical, characterized in that it further includes a step of predicting a reaction between the radiopharmaceutical drug and the cell by a radiopharmaceutical response prediction model based on features of the radiopharmaceutical extracted from the feature extraction model and multi-omics of a cell line.

20. In paragraph 18, A method for predicting binding affinity of a radiopharmaceutical, characterized in that it further includes a step of predicting binding affinity between the radiopharmaceutical and the protein by a radiopharmaceutical affinity prediction model based on features of the radiopharmaceutical and features of the protein extracted from the feature extraction model.

Images