A cancer-targeted drug generation method based on differential geometry and reinforcement learning

Abstract

The application discloses a cancer targeting drug generation method based on differential geometry and reinforcement learning, belongs to the intelligent control technical field, and accurately describes the geometric characteristics of a molecular surface through differential geometry, generates diversified and chemically reasonable molecular conformations through DDPM, expands a molecular library in combination with a genetic algorithm, and optimizes the binding affinity of a molecule and a mutant target protein through reinforcement learning. The cancer targeting drug generation method based on differential geometry and reinforcement learning provided by the application does not need the complex post-processing of traditional methods, has a simple structure, is high in efficiency, and provides a brand-new technical paradigm for cancer targeting drug design.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug design, and in particular to a cancer targeting drug generation method based on differential geometry and reinforcement learning. BACKGROUND

[0002] In the field of cancer treatment, targeted drugs exert their efficacy by binding to specific target proteins (such as EGFR). However, the high mutation rate of target proteins (such as EGFR T790M mutation) often leads to drug off-targeting or decreased affinity, making the original drug (such as gefitinib) ineffective. To address this challenge, molecular generation techniques are widely used to design new molecules that bind to mutant target proteins. Existing technologies include rule-based combinatorial chemistry, molecular docking, and machine learning models (such as variational autoencoder VAE and generative adversarial network GAN). However, existing technologies have the shortcomings of insufficient generation quality, inadequate three-dimensional interaction modeling, strong data dependency, low synthesis feasibility, imperfect evaluation indicators, and insufficient interpretability.

[0003] In view of the problems of insufficient geometric accuracy, poor chemical rationality, limited conformational diversity, and insufficient affinity optimization in existing cancer targeting drug molecular generation techniques, the present application aims to provide a molecular generation and optimization method based on differential geometry, denoising diffusion probability model, and reinforcement learning. The challenge faced by existing technologies is that the generated drug molecules often have problems of inaccurate geometric structure, unreasonable chemical structure, and insufficient affinity optimization, which seriously affect the effectiveness and clinical application of the drugs. SUMMARY

[0004] The present application aims to provide a cancer targeting drug generation method based on differential geometry and reinforcement learning to solve the problems in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides a cancer targeting drug generation method based on differential geometry and reinforcement learning, comprising the following steps:

[0006] S1, accurately capturing the geometric features of the molecular surface curvature and local structure using differential geometry;

[0007] S2, generating diversified and chemically reasonable candidate molecules by combining the denoising diffusion probability model DDPM with chemical bond constraints;

[0008] S3, generating a diversified molecular library by mutating and crossing the candidate molecules using a genetic algorithm, and optimizing the candidate molecules and screening high-efficiency candidate molecules using reinforcement learning with the binding affinity of the mutant target protein as the reward.

[0009] Preferably, S1 includes loading reference molecules, defining the molecular surface and sampling point clouds, PCA dimension reduction, and calculating the first fundamental form. The content of loading reference molecules is as follows:

[0010] Load molecule from SMILES string of failed drug, add explicit hydrogen atoms using RDKit's Chem.AddHs(), generate initial 3D conformation by AllChem.EmbedMolecule, and optimize structure based on MMFF force field with AllChem.MMFFOptimizeMolecule, MMFF optimization by minimizing potential energy function E MMFF = E bond + E angle + E torsion + E non-bonded Adjust coordinates to physically and chemically stable state, and cache optimized coordinates.

[0011] The contents of defining the molecular surface and sampling the point cloud are preferably as follows:

[0012] Define the molecular surface as the union of atom van der Waals spheres, two-dimensional manifold Generate point cloud with adaptive sampling Sampling points are based on spherical coordinates:

[0013] p j = r i + R vdw,i (sinφ j cosθ j , sinφ j sinθ j , cosφ j );

[0014] where θ j ~ Unif(0, 2π), φ j = arccos(2v j -1), v j ~ Unif(0, 1); p j denotes the three-dimensional position vector of the jth point; r i denotes the three-dimensional position vector of the i-th base point; R vdw,i denotes the van der Waals radius of the i-th base point; θ denotes the azimuthal angle; φ denotes the polar angle.

[0015] The number of sampling points M = min(1000, 50N).

[0016] The contents of PCA dimension reduction are preferably as follows:

[0017] Apply principal component analysis (PCA) to the point cloud, extract the first two principal components v1, v2, and v2 as the base vectors of the two-dimensional coordinate system, and the third principal component v3 as the normal vector; the point cloud covariance matrix is:

[0018]

[0019] where M represents the total number of points in the point cloud; represents the centroid of the point cloud;

[0020] The characteristic decomposition obtains v1, v2, v3; the atomic projection coordinates are:

[0021]

[0022] The height is

[0023] Preferably, the first fundamental form content is calculated as follows:

[0024] The height field h(u, v) is fitted by Gaussian Process Regression GPR, modeled by RBF kernel:

[0025]

[0026] where σ 2 is the signal variance; l is the length scale parameter;

[0027] The coefficients E, F, G of the first fundamental form are calculated:

[0028]

[0029]

[0030] The surface distance metric is:

[0031] ds 2 = E du 2 + 2F du dv + G dv 2 ;

[0032] The projection distance is:

[0033]

[0034] where d ij is the Euclidean distance between atoms i and j.

[0035] Preferably, in S2, within the spherical region centered at the molecular centroid, a new conformation is generated using Transformer-based DDPM; the forward process is:

[0036]

[0037] where q t represents the latent variable at time step t; α t is a predefined diffusion coefficient; ∈ is a standard Gaussian noise;

[0038] The reverse process predicts the noise ∈θ (q t , t ), generate new coordinates; spherical constraint is achieved by projected gradient method:

[0039]

[0040] where ∈ θ is the noise prediction function modeled by neural network parameters θ; c is the center of the sphere; R is the radius of the sphere; represents the unit vector from the center of the sphere to p i ;

[0041] The total loss function is:

[0042] L = L DDPM + λ1L bond + λ2L geom ;

[0043] where L DDPM is the standard diffusion model loss; L bond is the bond constraint loss; L geom is the geometric constraint loss; λ1, λ2 are weight parameters to balance different loss terms.

[0044] Preferably, the content of the bond constraint is as follows:

[0045] Extract the bond set of the reference molecule Calculate the three-dimensional bond distance d ij = |r i -r j | and the projection distance Define the loss function:

[0046]

[0047] Preferably, the fitness function in the genetic algorithm in S3 is:

[0048]

[0049] ChemValid(m);

[0050] where m represents the molecular structure to be evaluated; p i and p j represent the three-dimensional coordinates of atoms i and j in the molecular structure.

[0051] Preferably, the content of affinity optimization in S3 is as follows:

[0052] First, define MDP, the state is the molecule (SMILES and coordinates), the action is the molecule editing, and the reward is the binding affinity:

[0053] R(m) = -BindingEnergy(m, P mut ) + lambda * ChemValid(m) ;

[0054] wherein, R(m) represents the score value of molecule m; BindingEnergy(m, P mut ) represents the binding energy of small molecule m and mutant protein P mut ; lambda represents a weight coefficient; and ChemValid(m) represents a chemical validity index;

[0055] The REINFORCE algorithm is used for optimization:

[0056]

[0057] wherein, represents the gradient of the policy parameter theta to the objective function J(theta); pi θ (a t |s t ) represents a policy function with parameters theta, and the probability of taking action a t in state s t ; R(tau) represents the cumulative reward of trajectory tau; and b(s t ) represents a baseline function; represents the expectation of trajectory tau based on the policy pi θ sampling.

[0058] Therefore, the cancer targeting drug generation method based on differential geometry and reinforcement learning has the following beneficial effects:

[0059] (1) Differential geometry provides a mathematical framework for accurately capturing the geometric features of the molecular surface curvature and local structure, so that the generated molecules have more reasonable spatial configurations, thereby improving the accuracy and stability of the molecular binding with the target protein; and overcoming the inaccuracy problem in the generation of three-dimensional structure in the traditional method;

[0060] (2) The denoising diffusion probability model generates different molecular conformations by gradually adding noise and performing reverse denoising, and can maintain certain physical and chemical rationality during the generation process; ensures the diversity of the conformations, and also ensures the rationality of the generated molecules in the aspect of molecular dynamics, thereby providing more potential drug candidate molecules;

[0061] (3) By introducing chemical bond constraints (such as covalent bonds, cyclic structures, and other chemical rules), it is ensured that the generated molecules not only meet the spatial geometric requirements of physics, but also meet the chemical rationality; effectively avoids the problems of unnatural structure and chemical instability that may occur during the generation process, greatly improves the synthesis feasibility of the generated molecules;

[0062] (4) By introducing the reinforcement learning framework, the model can intelligently optimize the molecules according to the target function (such as binding affinity), thereby improving the binding ability of the molecules to the mutant target protein, and can gradually discover the molecular structure with high affinity through self-learning and exploration, thereby realizing more efficient targeted drug design.

[0063] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 is a flowchart of an embodiment of the present application;

[0065] Figure 2 is an existing VAE / GAN generation process based only on SMILES, ignoring three-dimensional geometry, of an embodiment of the present application;

[0066] Figure 3 is a schematic diagram of the conversion of atomic coordinates to two-dimensional projection of an embodiment of the present application; wherein (a) represents a schematic diagram of molecular surface point cloud sampling and PCA dimension reduction, and (b) represents a two-dimensional schematic diagram after conversion;

[0067] Figure 4 is a generation process of a high-level field and metric tensor E, F, G of an embodiment of the present application;

[0068] Figure 5 is a DDPM generation process of an embodiment of the present application, showing a reverse denoising schematic diagram from noise to molecular conformation;

[0069] Figure 6 is an interaction schematic diagram of a policy network and a reward function of an embodiment of the present application. DETAILED DESCRIPTION

[0070] The following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0071] Example 1:

[0072] Please refer to Figures 1-6 A cancer targeted drug generation method based on differential geometry and reinforcement learning, taking ineffective drugs as reference, comprising the following steps:

[0073] S1, the geometric features of the molecular surface curvature and local structure are accurately captured by using differential geometry.

[0074] S1 includes loading reference molecules, defining molecular surface and sampling point cloud, PCA dimension reduction and calculating the first fundamental form; the content of loading reference molecules is as follows:

[0075] Load molecules from SMILES string of failed drugs, add explicit hydrogen atoms using RDKit's Chem.AddHs(), generate initial three-dimensional conformation by AllChem.EmbedMolecule, and optimize the structure based on MMFF force field by AllChem.MMFFOptimizeMolecule, MMFF optimization is performed by minimizing the potential energy function E MMFF = E bond + E angle + E torsion + E non-bonded Adjust the coordinates to a physically and chemically stable state and cache the optimized coordinates. Provide an accurate three-dimensional geometric starting point for subsequent differential geometric analysis, and ensure the chemical consistency of the molecular representation.

[0076] The content of defining the molecular surface and sampling the point cloud is as follows:

[0077] The molecular surface is defined as the union of atomic van der Waals spheres, and the two-dimensional manifold Adaptive sampling is used to generate point cloud The sampling points are based on spherical coordinates:

[0078] p j = r i + R vdw,i (sinφ j cosθ j , sinφ j sinθ j , cosφ j );

[0079] Where θ j ~ Unif(0, 2π), φ j = arccos(2v j -1), v j ~ Unif(0, 1); p j represents the three-dimensional position vector of the jth point; r i represents the three-dimensional position vector of the i-th base point; R vdw,i represents the van der Waals radius of the i-th base point; θ represents the azimuth angle; φ represents the polar angle; the number of sampling points M = min(1000, 50N).

[0080] Point cloud is a discrete representation of molecular surface manifold, uniform sampling ensures complete capture of geometric information. Adaptive sampling dynamically adjusts the number of points according to the size of the molecule, balancing computational efficiency and accuracy. Provide high-quality surface data, provide input for differential geometry analysis, solve the problem of ignoring three-dimensional geometry in the prior art.

[0081] GPU parallel sampling is implemented using PyTorch's torch.rand, and the point cloud is cached (pointcloud_cache.npy).

[0082] The PCA dimensionality reduction content is as follows:

[0083] Apply principal component analysis (PCA) to the point cloud, extract the first two principal components v1, v2, and v2 as the base vectors of the two-dimensional coordinate system, and the third principal component v3 as the normal vector; The point cloud covariance matrix is:

[0084]

[0085] where M represents the total number of points in the point cloud; represents the centroid of the point cloud;

[0086] Eigen decomposition obtains v1, v2, v3; Atomic projection coordinates are:

[0087]

[0088] The height is

[0089] PCA preserves the main geometric information by maximizing the projection variance, reducing the three-dimensional manifold problem to a two-dimensional parameter space, which is convenient for surface analysis. It is used to simplify the surface parameterization and reduce the computational complexity, and provides the basis for the first fundamental form calculation.

[0090] The PCA implementation of Scikit-learn is used, and the result is cached (pca_cache.npy).

[0091] The first fundamental form calculation content is as follows:

[0092] Fit the height field h(u, v) by Gaussian process regression GPR, modeled with RBF kernel:

[0093]

[0094] where σ 2 is the signal variance; l is the length scale parameter;

[0095] Calculate the coefficients E, F, G of the first fundamental form:

[0096]

[0097] The curved surface distance metric is:

[0098] ds 2 = Edu 2 + 2Fdudv + Gdv 2 ;

[0099] The projected distance is:

[0100]

[0101] where d ij is the Euclidean distance between atoms i and j.

[0102] The first fundamental form defines the intrinsic geometry on the curved surface, and the non-parametric modeling of GPR ensures the smoothness and accuracy of the height field. It is used to precisely control the molecular surface geometry, solving the problem of low matching degree of three-dimensional structure in the prior art.

[0103] GPR with GPU acceleration is implemented using GPyTorch, and E, F, G (metric_cache.npy) are cached.

[0104] S2, adopt denoising diffusion probability model DDPM combined with chemical bond constraint to generate diversified and chemically reasonable candidate molecules.

[0105] In the spherical region centered on the molecular centroid, a Transformer-based DDPM is used to generate new conformations; the forward process is:

[0106]

[0107] where q t represents the latent variable at time step t; α t is a pre-defined diffusion coefficient; ∈ is a standard Gaussian noise;

[0108] The reverse process predicts the noise ∈ θ (q t , t), generates new coordinates; the spherical constraint is realized by the projected gradient method:

[0109]

[0110] where ∈ θ is a noise prediction function modeled by neural network parameters θ; c is the spherical center coordinate; R is the radius of the sphere; represents the unit vector pointing from the center of the sphere to p i ;

[0111] The total loss function is:

[0112] L = L DDPM + λ1Lbond + λ2L geom ;

[0113] where L DDPM is the standard diffusion model loss; L bond is the bond constraint loss; L geom is the geometric constraint loss; λ1, λ2 are weight parameters to balance different loss terms.

[0114] DDPM generates diversified conformations through multi-step denoising, the self-attention mechanism of Transformer captures long-range interactions between atoms, and spherical constraints limit the generation range. It is used to generate diversified and geometrically accurate molecules, solving the problem of single conformation in existing technologies.

[0115] The Transformer is implemented using PyTorch, GPU acceleration is used for training, and 100 conformations are generated in batches.

[0116] The content of the bond constraint is as follows:

[0117] Extract the bond set of the reference molecule Calculate the three-dimensional bond distance d ij = |r i - r j | and the projection distance Define the loss function:

[0118]

[0119] The three-dimensional bond distance constraint the topological structure of the molecule, the projection distance combined with the first fundamental form embedded the surface geometry, and the loss function was optimized through gradient descent. It is used to ensure that the generated molecules are chemically reasonable and reduce the chemical defects problem of existing technologies.

[0120] Use PyTorch automatic differentiation to calculate the gradient, and cache the bond distance (bond_cache.npy).

[0121] Post-processing and verification:

[0122] Optimize the generated molecules through MMFF, and call RDKit's Chem.SanitizeMol to verify the chemical reasonableness. MMFF optimization minimizes E MMFF , and RDKit checks valence and bond order based on a chemical rule base. It is used to improve the practicality of molecules and ensure chemical and physical stability.

[0123] Parallel optimization (multi-threaded RDKit), cache valid molecules (valid_molecules.npy).

[0124] S3, mutate and crossover candidate molecules through genetic algorithm to generate a diverse molecular library, and use reinforcement learning to optimize candidate molecules and screen high-efficiency candidate molecules by taking the binding affinity of the mutant target protein as the reward.

[0125] First, initialization is performed to take the initial candidate molecules as the population (size 1000), and the fitness function is used to balance geometric similarity and chemical reasonableness:

[0126]

[0127] Then, mutation and crossover are performed, mutation is random replacement of atoms (C / N / O, probability 0.1), adjustment of bond type (single bond / double bond, probability 0.05). Crossover is based on subgraph matching to exchange fragments, and finally selection is performed.

[0128] Genetic algorithm expands chemical space by simulating natural selection, mutation and crossover increase molecular diversity, and fitness function balances geometric similarity and chemical reasonableness. It is used to generate a diverse molecular library to make up for the problem of insufficient coverage of chemical space in the prior art.

[0129] Mutation and crossover are performed in parallel, and the population is cached (generation_cache.npy).

[0130] Use reinforcement learning to optimize affinity, first perform MDP definition, state is molecule (SMILES and coordinates), action is molecule editing, reward is binding affinity:

[0131] R(m) = -BindingEnergy(m, P mut ) + λ·ChemValid(m);

[0132] Wherein, R(m) represents the score value of molecule m; BindingEnergy(m, P mut ) represents the binding energy of small molecule m and mutant protein P mut ; λ represents the weight coefficient; ChemValid(m) represents the chemical validity index;

[0133] Use REINFORCE algorithm for optimization:

[0134]

[0135] Wherein, represents the gradient of policy parameter θ to objective function J(θ); π θ (a t |s t ) represents the policy function with parameter θ, which takes action a t under state s tprobability of the trajectory τ; R(τ) represents the cumulative reward of the trajectory τ; b(s t ) represents the baseline function; represents the expected value of the sampled trajectory τ based on the policy π θ .

[0136] Reinforcement learning dynamically optimizes molecules through reward feedback, and the policy gradient method ensures efficient convergence. It is used to optimize the binding affinity of mutant target proteins, solving the problem of insufficient efficacy of existing technologies.

[0137] GPU-accelerated training, cache state-action pairs (rl_cache.npy).

[0138] MMFF optimization and RDKit verification are performed on the optimized molecules, and the top 10 molecules with the highest binding affinity are selected. It is used to screen high-efficiency target molecules and ensure practicality.

[0139] Parallel verification, store the final candidate molecules (final_molecules.npy).

[0140] Example Two:

[0141] Based on the Riemannian manifold and the GAN generation method, the Riemannian manifold is used to replace the first fundamental form to describe the molecular surface geometry, and the GAN is used to generate the molecular conformation. The generator of the GAN outputs the molecular coordinates, and the discriminator evaluates the geometric similarity and chemical rationality.

[0142] Implementation steps:

[0143] Load reference molecules, optimize three-dimensional conformation; sample point cloud, calculate Riemannian metric tensor; train GAN, generator takes noise as input, outputs molecular coordinates, discriminator evaluates geometric and chemical loss; generate molecules through MMFF optimization and RDKit verification.

[0144] Instead of differential geometry and DDPM, it retains geometric accuracy but has slightly lower computational complexity and diversity.

[0145] Example Three:

[0146] Based on the optimization method of simulated annealing and molecular docking, the reference molecule is taken as the starting point, the molecular editing is performed through simulated annealing (Simulated Annealing), and the molecular docking (AutoDockVina) is combined to directly optimize the binding affinity.

[0147] Implementation steps:

[0148] Load reference molecules, initialize molecule pool; adjust atom type, bond type or conformation randomly by simulated annealing, accept probability based on affinity change; calculate binding affinity using molecular docking, screen high-affinity molecules; MMFF optimization and RDKit verification.

[0149] Instead of genetic algorithm and reinforcement learning, simplify optimization process, but limited exploration ability.

[0150] Embodiment four:

[0151] Generation method based on graph neural network (GNN), input is molecular graph (node is atom, edge is chemical bond), use GNN to generate new molecules, combine molecular docking to optimize affinity.

[0152] Implementation steps:

[0153] Convert reference molecules to graph representation; train GNN to predict node and edge addition / deletion operations, generate new molecules; use molecular docking to evaluate affinity, iterate optimization; MMFF optimization and RDKit verification.

[0154] Instead of DDPM and differential geometry, simplify geometric modeling, but geometric accuracy is poor.

[0155] Therefore, the cancer targeted drug generation method based on differential geometry and reinforcement learning provided by the present application solves the key problems in the design of cancer targeted drugs in the prior art through accurate mathematical modeling, efficient generation model and dynamic optimization mechanism, and provides an implementable solution for efficient and accurate drug molecule generation.

[0156] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application and not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or replace the technical solutions of the present application, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for generating a cancer-targeting drug based on differential geometry and reinforcement learning, characterized in that, Comprising the following steps: S1, accurately capture the surface curvature of the molecule, the geometric features of the local structure by differential geometry; S2, generate diverse and chemically reasonable candidate molecules by combining denoising diffusion probability model DDPM with chemical bond constraints; In the spherical region centered on the molecular centroid, use DDPM based on Transformer to generate new conformations; the forward process is as follows: ； wherein, denotes the time step of the latent variable; is a predefined diffusion coefficient; is a standard Gaussian noise; Inverse process predicts noise New coordinates are generated; spherical constraints are achieved by projection gradient method: ； wherein, is a noise prediction function modeled by neural network parameters ; is a spherical center coordinate; is a spherical radius; denotes a unit vector pointing from the spherical center to ; The total loss function is as follows: ； wherein, is a standard diffusion model loss; is a chemical bond constraint loss; is a geometric constraint loss; , is a weight parameter balancing different loss terms; The content of the chemical bond constraint is as follows: Extracting a set of chemical bonds of a reference molecule , calculating three-dimensional bond distances and projected distances , defining a loss function: ； ； S3, generate a diverse molecular library by mutation and crossover of candidate molecules through genetic algorithm, and use reinforcement learning to optimize candidate molecules and screen high-efficiency candidate molecules by taking the binding affinity of the mutant target protein as the reward.

2. The method of claim 1, wherein the method is characterized by: S1 includes loading reference molecules, defining molecular surfaces and sampling point clouds, PCA dimensionality reduction and calculating the first fundamental form; the content of loading reference molecules is as follows: Load the molecule from the SMILES string of the failed drug, add explicit hydrogen atoms using RDKit's Chem.AddHs(), generate an initial three-dimensional conformation by AllChem.EmbedMolecule, and optimize the structure based on MMFF force field with AllChem.MMFFOptimizeMolecule, MMFF optimization by minimizing the potential energy function Adjust the coordinates to a physically and chemically stable state and cache the optimized coordinates. 3.The method of claim 2, wherein, The content of defining molecular surfaces and sampling point clouds is as follows: Defining the molecular surface as the union of the van der Waals spheres of the atoms, two-dimensional manifold Adaptive sampling to generate point cloud Sampling points based on spherical coordinates: ； wherein , , ; represents a three-dimensional position vector of the th point; represents a three-dimensional position vector of the base point, is the th base point; represents the van der Waals radius of the th base point; represents an azimuthal angle; represents a polar angle; number of sampling points . 4.The method of claim 2, wherein, The content of PCA dimensionality reduction is as follows: Apply principal component analysis (PCA) to the point cloud, extract the first two principal components , , The third principal component as a two-dimensional coordinate system base vector The covariance matrix of the point cloud is: ； wherein, denotes the total number of points in the point cloud; denotes the centroid of the point cloud; The eigen decomposition gives , , ; the atomic projection coordinates are: ； height is . 5.The method for generating cancer targeting drugs based on differential geometry and reinforcement learning according to claim 4, characterized in that, The content of calculating the first fundamental form is as follows: fitting height fields by gaussian process regression GPR with RBF kernel ； wherein, is the signal variance; is the length scale parameter; Calculate the coefficients E, F, G of the first fundamental form: ； ； ； The surface distance metric is as follows: ； The projection distance is as follows: ； wherein is the Euclidean distance between atoms and and 6. The method of claim 5, wherein the method is characterized by: The fitness function in the genetic algorithm in S3 is as follows: ； wherein represents a molecular structure to be evaluated; and represents the three-dimensional coordinates of an atom in the molecular structure and in the molecular structure.

7. The method of claim 5, wherein the method is characterized by: The content of affinity optimization in S3 is as follows: First, define MDP, the state is molecule (SMILES and coordinates), the action is molecule editing, and the reward is binding affinity: ； wherein, represents a score value of a molecule ; represents a binding energy of a small molecule to a mutant protein ; represents a weight coefficient; represents a chemical effectiveness index; Optimize using REINFORCE algorithm: ； where, denotes a policy parameter denotes a target function a gradient of the target function denotes a policy function with parameters a probability of taking action in state denotes a trajectory a cumulative reward of the trajectory denotes a baseline function denotes an expectation based on trajectories sampled by the policy denotes an expectation based on trajectories sampled by the policy

Images