CN115966266B - An anti-tumor molecule enhancement method based on graph neural network - Google Patents

Abstract

The present invention designs an anti-tumor molecule reinforcement learning method based on graph neural network. The method includes the following steps: Step 1: Divide molecules into anti-tumor (positive) and non-anti-tumor (negative) categories according to the molecular labels in the database. Step 2: Input the obtained graph into the proposed anti-tumor molecule enhancement model, learn the implicit representation of the graph according to the different properties of positive molecules and negative molecules, and obtain the local structural characteristics of the anti-tumor molecules for one-step generation and optimization of molecules. Step 3: Apply constraints and perform target optimization to ensure the drug-like properties of the molecules during the molecule enhancement process. Step 4: Substitute the obtained local molecular structure into the structural modification and optimization of the anti-tumor molecule. Step 5: Use the existing molecular properties to reasonably detect The tool determines the synthesizability, outputs reasonable molecules, and further reversely optimizes unreasonable molecules. Step 6: Obtain reasonable new anti-tumor molecules, and the task is over.

Description

An anti-tumor molecule enhancement method based on graph neural network

Technical field

The invention relates to a drug molecule reinforcement technology based on graph neural network, and belongs to the technical fields of anti-tumor drug molecular chemistry research and graph neural network reinforcement learning technology.

Background technique

Describe the state of the art and existing problems that are closest to the present invention.

The optimization of drug molecules and the research of new drugs are crucial to the treatment of tumors. The goal of drug molecule optimization is to enhance more desirable biological effects in a specific direction while ensuring biopharmaceutical acceptability. The current problem is the high cost of time and money in developing highly effective drugs. In the current field of drug research, the traditional strategy is to screen based on existing compound libraries. However, due to the limited structural diversity of molecules in existing compound libraries and the fact that research institutions have screened them many times, it has become increasingly challenging to discover and innovate highly effective drug molecules.

With the development of machine learning, its role in the field of drug discovery is growing. Compared with traditional screening strategies, the integration of machine learning and molecular research will be more effective and scalable in processing tasks.

Currently, relevant personnel in this field have developed a "de novo molecule design technology" - an atom-based, fragment-based, reaction-based molecular design method that generates new molecules with high effectiveness from scratch through analytical calculations. Its design idea is to abandon the traditional crude direct screening method, balance the global and local aspects, explore key features, and build target molecules based on clues. The disadvantage is that this method requires high-level manual generation and molecular architecture from scratch; at the same time, clear design goals and standardized design principles are also necessary for the interpretability of the results. At the same time, since de novo molecular design is based on fragments, the quality of the generated molecules is also inseparable from the previous operations. In short, the current outstanding challenges of de novo design methods are high computational complexity, low interpretability, and high dependence of generated results on early choices.

Contents of the invention

Technical problem: To address the above problems, the present invention proposes a reinforcement learning method for anti-tumor drug molecules based on graph neural network. Construct a molecular anti-tumor enhanced graph neural network model to perform feature extraction and molecular property classification based on drug molecules; observe the classification results by modifying the features to extract key features. Use the analysis of key features to chemically strengthen and modify the original molecular structure, and finally obtain new molecules with stronger target properties, thereby improving the efficiency and interpretability of molecule generation and reducing the difficulty and development of anti-tumor molecular drugs. Cycle time and cost.

Technical solution: In order to achieve the purpose of the present invention, the technical solution adopted by the present invention is: an anti-tumor molecule enhancement method based on graph neural network, which method includes the following steps:

Step 1: Classify molecules into anti-tumor (positive) and non-anti-tumor (negative) categories based on molecular tags in the database. Describe each input molecule as an undirected graph (graph matrix), where nodes and edges correspond to atoms and chemical bonds, respectively. Construct a graph neural network model, use the chemical stability and drug-like properties of the generated molecules as the loss function, and pre-train.

Step 2: Input the obtained graph into the proposed anti-tumor molecule enhancement model, learn the implicit representation of the graph based on the different properties of positive molecules and negative molecules, and obtain the local structural characteristics of the anti-tumor molecules for one-step generation and optimization of molecules.

Step 3: Apply constraints and perform target optimization to ensure the drug-like properties of the molecules during the molecule enhancement process.

Step 4: Substitute the obtained local molecular structure into the structural modification and optimization of anti-tumor molecules.

Step 5: Use existing molecular property reasonable detection tools to determine the synthesizability and output reasonable molecules. For unreasonable molecules, further reverse optimization is performed.

Step 6: Obtain reasonable new anti-tumor molecules and the task is over.

Further, in step (1), the molecules are classified into anti-tumor (positive) and non-anti-tumor (negative) categories according to the molecular tags in the database. Describe each input molecule as an undirected graph (graph matrix), where nodes and edges correspond to atoms and chemical bonds, respectively. Construct a graph neural network model and pre-train it as follows:

(101) Input molecules in the form of a graph, including G = (V, E, An n×n size adjacency matrix, X is the characteristics of each atom of the molecule, and is an n×n size matrix. Describe each input molecule as an undirected graph (graph matrix), which is G.

(102) The molecules are classified into anti-tumor (positive) and non-anti-tumor (negative) categories according to the molecular tags in the database, represented by G ₊ or G ₋ .

(103) Construct a graph neural network model, whose input is the original molecular undirected graph G, and the output is a binary classification probability matrix P. The chemical stability and drug-like properties of the generated molecules are used as the loss function. For this GNN The model is pre-trained.

Further, in step (2), the obtained graph is input into the proposed anti-tumor molecule enhancement model, and the implicit representation of the graph is learned based on the different properties of positive molecules and negative molecules to obtain the local structural characteristics of the anti-tumor molecule, which can be used for molecule identification. Generate and optimize in one step. Methods as below:

(201) Input the obtained molecule G into the feature extraction model f:G→F∈R ^n×h to extract the implicit feature F of the molecule

(202) The feature extraction model f:G→F∈R ^n×h includes: (1)

in Represents the characteristics of each molecule obtained at layer l-1,/> is matrix/> In the content of line j, N(v) is the neighbor node of v, UPDATE is the update function of each layer, AGG is the aggregation function, READOUT is the readout function, and the characteristics of the molecule are obtained after l iterations

(2) The obtained h _G is used as input to obtain the implicit feature F of the molecule through an MLP model.

(203) Input the obtained implicit feature F into the classification prediction network c:F→P _out ∈R ^n×2 to obtain the two-class probability result matrix P _u .

(204) Partially transform the obtained molecular structure and bring it into the feature extraction model to obtain the transformed implicit feature F′. The obtained implicit feature F′ is input into the classification prediction network to obtain the binary classification probability result matrix P _n .

(205) Input the probability results obtained in steps (203) and (204) into the probability fluctuation function PFF (Fluctuation probability function), bring it into the MEAS (measure) function, and analyze the calculation results of PFF,

PFF＝||p _u -p _n ||

S _out =MEAS{PFF(P _u ,P _n )}

Extract the local structure feature S _{out of} the molecule that has a large impact on the probability fluctuation after the change, and use this as the output.

Further, in step (3), constraints are imposed and target optimization is performed to ensure the drug-like properties of the molecules during the molecular strengthening process. The method is as follows:

(301) Utilizing QED values in the MEAS function when training anti-tumor molecule reinforcement models

to constrain

MEAS: S _out =RF{PFF(P _u ,P _n )+γQED}

Among them, QED is calculated using RDKit. The RF function maps the calculated value of the probability fluctuation function after QED constraints to the molecular structure to obtain the local molecular structure, thus ensuring the drug-likeness of the drug-likeness corresponding to the generated local features.

Further, in step (4), the obtained local molecular structure is substituted into the structure modification and optimization method of the anti-tumor molecule as follows:

(401) The input of existing positive molecules includes G=(V,E,X), where G represents the input molecule, V represents the atoms of the molecule, E represents the chemical bond of the molecule, and Get the input numerator.

(402) Use the local structural characteristics of the molecule obtained by the above-trained model to strengthen the molecule: through the automatic iterative optimization method, continuously find the most likely atoms or chemical bonds to connect, thereby modifying the structure of the molecule and converting the previously obtained local molecule Structural features are applied to molecules, enabling the progressive construction of molecules with better anti-tumor properties.

Further, in step (5), existing reasonable molecular property detection tools are used to determine the synthesizability and output reasonable molecules. For unreasonable molecules, further reverse optimization is performed as follows:

(501) Establish an existing reasonable molecular property detection tool p:G=(V,E,X)→V∈R to analyze the rationality of molecular properties and evaluate chemical feasibility.

(502) Model p mainly consists of state-action module A and reward module Q. At any time step t, the input of module A is the state and the output is the action. The action is defined in the feature representation space of all initial reactants. Tensor. Module Q calculates the optimal value of the state through the Q network; the Actor uses this optimal value to iteratively update the parameters of the policy function, then selects actions, and obtains feedback and new states. The environment uses status, best response templates, and actions as reference to calculate whether the turn is over. Finally, the evaluation value V is output through the Softmax layer to represent the feasibility score.

Further, in step (6), a reasonable new anti-tumor molecule is obtained and the task is completed. The method is as follows:

(601) As shown in claim 6, in the optimization of the feedback standard penalty, the generated molecules obviously cannot become realistically available drugs. This shows that the model with strong anti-tumor characteristics can be generated by strengthening the anti-tumor molecule, but the generation cannot be guaranteed. molecules can be made or stabilized in real life. This highlights the need to use multi-objective optimization for rewards and penalties when using reinforcement learning for anti-tumor molecule enhancement.

(602) Re-emphasis on variable definition:

X _v : Dimension R ^DV , feature vector of node v

h _v : Dimension R ^DV , state vector of node v

x _v1,v2 : dimension R ^DE , eigenvector of edge (V ₁ , V ₂ )

During the node operation process, it is necessary to define the node status update function This allows the node status to be iteratively stabilized. And for

The state transformation function of the node. For node V, the transformation of its state vector can be expressed as:

(603) Here the emphasis is on the backpropagation process:

In this step, we can find the gradient of the parameters by following the steps of backpropagation, and then use the gradient descent method.

optimization. The iterative process is:

…

Δw _ij =ηδ _j x _i

Then, the Backpropagation method is used, and the automatic backpropagation under the pytorch framework is used to continuously and iteratively optimize the molecular structure, so that the generated enhanced molecular model meets multiple requirements such as anti-tumor and molecular rationality.

Beneficial effects: Compared with the existing technology, the technical solution of this method has the following beneficial technical effects:

1. Learn the characteristics of drug molecules through graph neural networks, use graph neural networks and reinforcement learning methods to strengthen given molecules, and transfer the work of strengthening molecules to machines. Compared with traditional manual strengthening methods, it is greatly improved. Improved efficiency.

2. The method mentioned in this method of strengthening existing molecules based on graph neural network and reinforcement learning greatly improves the efficiency compared to other methods of generating molecules from scratch based on the model of a given molecule, and to a certain extent improves the efficiency accuracy.

3. The molecules generated by this method have higher interpretability and better visualization, which is conducive to understanding the characteristics of the generated enhanced drug molecules and the reasons for the enhancement, and provides more convenience for relevant drug researchers.

4. Use multiple constraints and property detection during the generation process to ensure the chemical rationality and pharmacological properties of the generated molecules.

Description of the drawings

Figure 1 is a flow chart of implementation steps of the present invention;

Figure 2 is a framework diagram of the anti-tumor molecule enhancement model;

Figure 3 is an example molecular structure diagram;

Figure 4 is a diagram of the structure optimization results of an example molecule.

Detailed ways

In order to deepen the understanding of the present invention, this embodiment will be described in detail below with reference to the accompanying drawings.

Examples: Taking the molecules in the ChEMBL database as an example, the technical solution of the present invention will be described in detail in conjunction with the accompanying drawings.

An anti-tumor molecule enhancement method based on graph neural network, the method includes the following steps:

Step (1): Classify molecules into anti-tumor (positive) and non-anti-tumor (negative) categories based on molecular tags in the database. Describe each input molecule as an undirected graph (graph matrix), where nodes and edges correspond to atoms and chemical bonds, respectively. Construct a graph neural network model, use the chemical stability and drug-like properties of the generated molecules as the loss function, and pre-train. The method is as follows:

(101) Input molecules in the form of a graph, including G = (V, E, An n×n size adjacency matrix, X is the characteristics of each atom of the molecule, and is an n×n size matrix. Describe each input molecule as an undirected graph (graph matrix), which is G.

(102) The molecules are classified into anti-tumor (positive) and non-anti-tumor (negative) categories according to the molecular tags in the database, represented by G ₊ or G ₋ .

(103) Construct a graph neural network model, using the chemical stability and drug-like properties of the generated molecules as the loss function. The model structure is shown in Figure 2. Its input is the original molecular undirected graph G, and the output is binary Classification

Probability matrix P, pre-train the GNN model.

Step (2): Input the obtained graph into the proposed anti-tumor molecule enhancement model, learn the implicit representation of the graph based on the different properties of positive molecules and negative molecules, and obtain the local structural characteristics of the anti-tumor molecules for one-step generation and optimization of molecules. . Methods as below:

(201) Input the obtained molecule G into the feature extraction model f:G→F∈R ^n×h to extract the implicit feature F of the molecule

(202) The feature extraction model f:G→F∈R ^n×h includes: (1)

in Represents the characteristics of each molecule obtained at layer l-1,/> is matrix/> In the content of line j, N(v) is the neighbor node of v, UPDATE is the update function of each layer, AGG is the aggregation function, READOUT is the readout function, and the characteristics of the molecule are obtained after l iterations

(2) The obtained h _G is used as input to obtain the implicit feature F of the molecule through an MLP model.

(203) Input the obtained implicit feature F into the classification prediction network c:F→P _out ∈R ^n×2 to obtain the two-class probability result matrix P _u .

(204) Partially transform the obtained molecular structure and bring it into the feature extraction model to obtain the transformed implicit feature F′. The obtained implicit feature F′ is input into the classification prediction network to obtain the binary classification probability result matrix P _n .

(205) Input the probability results obtained in steps (203) and (204) into the probability fluctuation function PFF (Fluctuation probability function), bring it into the MEAS (measure) function, and analyze the calculation results of PFF,

PFF＝||p _u -p _n ||

S _out =MEAS{PFF(P _u ,P _n )}

Extract the local structure feature S _{out of} the molecule that has a large impact on the probability fluctuation after the change, and use this as the output.

Step (3): Apply constraints and perform target optimization to ensure the drug-like properties of the molecules during the molecular strengthening process. The method is as follows:

(301) Utilizing QED values in the MEAS function when training anti-tumor molecule reinforcement models

to constrain

MEAS: S _out =RF{PFF(P _u ,P _n )+γQED}

Among them, QED is calculated using RDKit. The RF function maps the calculated value of the probability fluctuation function after QED constraints to the molecular structure to obtain the local molecular structure, thus ensuring the drug-likeness of the drug-likeness corresponding to the generated local features.

Step (4): Substitute the obtained local molecular structure into the structural modification and optimization of anti-tumor molecules. Methods as below:

(401) The input of existing positive molecules includes G=(V,E,X), where G represents the input molecule, V represents the atoms of the molecule, E represents the chemical bond of the molecule, and Get the input numerator.

(402) Use the local structural characteristics of the molecule obtained by the above-trained model to strengthen the molecule: through the automatic iterative optimization method, continuously find the most likely atoms or chemical bonds to connect, thereby modifying the structure of the molecule and converting the previously obtained local molecule Structural features are applied to molecules, enabling the progressive construction of molecules with better anti-tumor properties.

Step (5): Use existing molecular property reasonable detection tools to determine the synthesizability and output reasonable molecules. For unreasonable molecules, further reverse optimization is performed as follows:

(501) Establish a reasonable detection tool for existing molecular properties p:G=(V,E,X)→V∈R, analyze the rationality of molecular properties and evaluate

Evaluate chemical feasibility

(502) Model p mainly consists of state-action module A and reward module Q. At any time step t, the input of module A is the state and the output is the action. The action is defined in the feature representation space of all initial reactants. Tensor. Module Q calculates the optimal value of the state through the Q network; the Actor uses this optimal value to iteratively update the parameters of the policy function, then selects actions, and obtains feedback and new states. The environment uses status, best response templates, and actions as reference to calculate whether the turn is over. Finally, the evaluation value V is output through the Softmax layer to represent the feasibility score.

Step (6): Obtain reasonable new anti-tumor molecules. The task is over. The method is as follows:

(601) As shown in claim 6, in the optimization of the feedback standard penalty, the generated molecules obviously cannot become realistically available drugs. This shows that the model with strong anti-tumor characteristics can be generated by strengthening the anti-tumor molecule, but the generation cannot be guaranteed. molecules can be made or stabilized in real life. This highlights the need to use multi-objective optimization for rewards and penalties when using reinforcement learning for anti-tumor molecule enhancement.

(602) Re-emphasis on variable definition:

X _v : Dimension R ^DV , feature vector of node v

h _v : Dimension R ^DV , state vector of node v

x _v1,v2 : dimension R ^DE , eigenvector of edge (V ₁ , V ₂ )

During the node operation process, it is necessary to define the node status update function This allows the node status to be iteratively stabilized. As for the node's state transformation function, for node V, the transformation of its state vector can be expressed as:

(603) Here the emphasis is on the backpropagation process:

In this step, we can find the gradient of the parameters by following the steps of backpropagation, and then use the gradient descent method for optimization. The iterative process is:

…

Δw _ij = ηδ _j X _i

Then, the Backpropagation method is used, and the automatic backpropagation under the pytorch framework is used to continuously and iteratively optimize the molecular structure, so that the generated enhanced molecular model meets multiple requirements such as anti-tumor and molecular rationality. ChEMBL number

Figure 4 shows before and after optimization of a certain molecule in the database.

It should be noted that the above-mentioned embodiments are not used to limit the scope of protection of the present invention. Equivalent transformations or substitutions made on the basis of the above-mentioned technical solutions all fall within the scope of protection of the claims of the present invention.

Claims (5)

1. An anti-tumor molecule reinforcement learning method based on graph neural network, characterized in that the method includes the following steps: Step 1: Classify the molecules into anti-tumor (positive) and non-anti-tumor (negative) categories according to the molecule labels in the database. Describe each input molecule as an undirected graph, that is, a graph matrix, where nodes and edges correspond to atoms and chemical bonds respectively. Construct a graph neural network model, use the chemical stability and drug-like properties of the generated molecules as the loss function, and perform pre-training for the node classification task. Step 2: Input the obtained graph into the proposed anti-tumor molecule enhancement model, learn the implicit representation of the graph according to the different properties of positive molecules and negative molecules, and obtain the local structural characteristics of the anti-tumor molecules for one-step generation and optimization of molecules. Step 3: Apply constraints and perform target optimization to ensure the drug-like properties of the molecules during the molecular strengthening process. Step 4: Substitute the obtained local molecular structure into the structural modification and optimization of anti-tumor molecules. Step 5: Use existing molecular property reasonable detection tools to determine the synthesizability, output reasonable molecules, and further reverse optimization for unreasonable molecules. Step 6: Obtain reasonable new anti-tumor molecules, the task is over; In step 1, molecules are classified into anti-tumor (positive) and non-anti-tumor (negative) categories according to the molecule labels in the database. Each input molecule is described as an undirected graph, that is, a graph matrix, in which nodes and edges correspond to atoms and chemical bonds respectively. , build a graph neural network model and pre-train it. The specific method is as follows: (101) Input molecules in the form of a graph, including G = (V, E, An n×n size adjacency matrix. (102) According to the molecular tags in the database, molecules are divided into anti-tumor (positive) and non-anti-tumor (negative) categories, represented by G ₊ or G-; (103) Construct a graph neural network model, the input of which is the original molecular undirected graph G, and the output is the binary classification probability matrix P. Pre-train the GNN model; in step (2), input the resulting graph The proposed anti-tumor molecule enhancement model learns the implicit representation of the graph based on the different properties of positive molecules and negative molecules, and obtains the local structural characteristics of anti-tumor molecules for one-step generation and optimization of molecules. The method is as follows: (201) Input the obtained molecule G into the feature extraction model f:G→F∈R ^n×h Extract the implicit feature F of the molecule, (202) Feature extraction f:G→F∈R ^n×h model includes: (1) in Represents the characteristics of each molecule obtained at the l-1 layer, /> is matrix/> In the content of line j, N(v) is the neighbor node of v, UPDATE is the update function of each layer, AGG is the aggregation function, and READOUT is the readout function. After l iterations, the characteristics of the molecule are obtained (2) to obtain h _G is taken as input, and the implicit feature F of the molecule is obtained through an MLP model, (203) Input the obtained implicit feature F into the classification prediction network c:F→P _out ∈R ^n×2 Obtain the two-class probability result matrix P _u , (204) Partially transform the obtained molecular structure and bring it into the feature extraction model to obtain the transformed implicit feature F'. Enter the obtained implicit feature F' into the classification prediction network to obtain the binary classification probability result matrix P _n . (205) Enter the probability results obtained in steps (203) and (204) into the probability change function PFF (Fluctuation probability function), bring it into the MEAS (measure) function, and analyze the calculation results of PFF PFF＝||p _u -p _n || S _out =MEAS{PFF(P _u ,P _n )}, Extract the local structure feature S _{out of} the molecule that has a large impact on the probability fluctuation after the change, and use this as the output. 2. The anti-tumor molecule reinforcement learning method based on graph neural network according to claim 1, characterized in that in step 3, constraints are imposed to perform target optimization to ensure the drug-like properties of the molecules during the molecule reinforcement process. The method is as follows: (301) When training the anti-tumor molecule reinforcement model, use QED (quantitative estimate of drug-likeness) numerical constraints in the MEAS function. MEAS: S _out =RF{PFF(P _u ,P _n )+γQED} Among them, QED is calculated using RDKit, and the RF function maps the calculated value of the probability fluctuation function after QED constraints to the molecular structure to obtain the local molecular structure, thus ensuring the drug-likeness of the generated local features corresponding to the drug similarity. 3. The anti-tumor molecule reinforcement learning method based on graph neural network according to claim 1, characterized in that, in step 4, the obtained local molecular structure is substituted for structural modification and optimization of the anti-tumor molecule, Methods as below: (401) The input of existing positive molecules includes G=(V,E,X), where G represents the input molecule, V represents the atoms of the molecule, E represents the chemical bond of the molecule, and Get the input numerator, (402) Use the local structural characteristics of the molecule obtained by the anti-tumor molecule reinforcement model trained above to strengthen the molecule: through the automatic iterative optimization method, constantly find the most likely atoms or chemical bonds to connect, thereby modifying the structure of the molecule and replacing the previous ones. The obtained local structural characteristics of the molecules are applied to the molecules to gradually build molecules with better anti-tumor properties. 4. The anti-tumor molecule reinforcement learning method based on graph neural network according to claim 1, characterized in that, in step 5, existing reasonable molecular property detection tools are used to determine the synthesizeability and output reasonable molecules, while targeting different Reasonable molecule, further reverse optimization, the method is as follows: (501) Establish an existing reasonable molecular property detection tool p:G=(V,E,X)→V∈R, analyze the rationality of molecular properties and evaluate chemical feasibility, (502) Model p mainly consists of state-action module A and reward module Q. At any time step t, the input of module A is the state and the output is the action. The action is defined in the feature representation space of all initial reactants. Tensor, module Q calculates the optimal value of the state through the Q network; Actor uses this optimal value to iteratively update the parameters of the policy function, then selects actions, and gets feedback and new states. The environment combines the state, best response template and action As a reference, the calculation determines whether the round is over, and finally the evaluation value V is output through the Softmax layer to represent the feasibility score. 5. The anti-tumor molecule reinforcement learning method based on graph neural network according to claim 1, characterized in that existing reasonable molecular property detection tools are used to judge the synthesizeability, output reasonable molecules, and further react against unreasonable molecules. To optimize, it is characterized in that, in step (6), a reasonable new anti-tumor molecule is obtained, and the method is as follows: (601) In the optimization of feedback standard penalty, the generated molecules obviously cannot become practical drugs. This shows that strengthening the model with anti-tumor molecules can generate strong anti-tumor characteristics, but there is no guarantee that the generated molecules can be used in real life. can be manufactured or exist stably, which highlights the need to use multi-objective optimization for rewards and punishments when using reinforcement learning for anti-tumor molecule enhancement. (602) Re-emphasis on variable definition: X _v : Dimension R ^DV , feature vector of node v, h _v : Dimension R ^DV , state vector of node v, x _v1,v2 : dimension R ^DE , eigenvector of edge (V ₁ , V ₂ ), During the node operation process, it is necessary to define the node status update function This makes the node state iteratively stable, and for the node state transformation function, for node V, the transformation of its state vector can be expressed as: (603) Here the emphasis is on the backpropagation process: In this step, the gradient of the parameters is obtained according to the steps of backpropagation, and then the gradient descent method is used for optimization. The iterative process is: … Then, the Backpropagation method is used, and the automatic backpropagation under the pytorch framework is used to continuously and iteratively optimize the molecular structure, so that the generated enhanced molecular model meets the multiple requirements of anti-tumor and molecular rationality.