Cancer targeted drug generation method based on differential geometry and reinforcement learning

Abstract

The invention discloses a cancer targeted drug generation method based on differential geometry and reinforcement learning, and belongs to the technical field of intelligent control. Molecular surface geometric characteristics are accurately described through differential geometry, diversified and chemically reasonable molecular conformations are generated by using DDPM, and a molecular library is expanded in combination with a genetic algorithm; and the binding affinity of the molecule and the mutation target protein is optimized through reinforcement learning. According to the cancer targeted drug generation method based on differential geometry and reinforcement learning, complex post-processing of a traditional method is not needed, the structure is simple, efficiency is high, and a brand new technical normal form is provided for cancer targeted drug design.

Description

Technical Field

[0001] The present invention relates to the technical field of drug design, and in particular to a method for generating cancer-targeted drugs based on differential geometry and reinforcement learning. Background Art

[0002] In the field of cancer treatment, targeted drugs exert their therapeutic effects 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-target or decreased affinity, making the original drugs (such as gefitinib) ineffective. To meet this challenge, molecular generation technology is 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 autoencoders (VAE) and generative adversarial networks (GAN). However, existing technologies have shortcomings such as insufficient generation quality, insufficient three-dimensional interaction modeling, strong data dependence, low synthetic feasibility, imperfect evaluation indicators, and insufficient interpretability.

[0003] To address the challenges of existing cancer-targeted drug molecule generation technologies, which suffer from insufficient geometric precision, poor chemical rationality, limited conformational diversity, and insufficient affinity optimization, the present invention provides a molecule generation and optimization method based on differential geometry, a denoised diffusion probability model, and reinforcement learning. The challenges faced by existing technologies are that the generated drug molecules often suffer from geometric inaccuracies, irrational chemical structures, and insufficient affinity optimization, which seriously impact drug efficacy and clinical application. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for generating cancer targeted drugs based on differential geometry and reinforcement learning to solve the problems existing in the background technology.

[0005] To achieve the above objectives, the present invention provides a method for generating cancer-targeted drugs based on differential geometry and reinforcement learning, comprising the following steps:

[0006] S1. Use differential geometry to accurately capture the geometric features of molecular surface curvature and local structure;

[0007] S2. Use the denoised diffusion probability model (DDPM) combined with chemical bond constraints to generate diverse and chemically plausible candidate molecules.

[0008] S3. Genetic algorithms are used to mutate and crossover candidate molecules to generate a diverse molecular library, and reinforcement learning is used to optimize candidate molecules and screen for efficient candidate molecules with the binding affinity of the mutant target protein as a reward.

[0009] Preferably, S1 includes loading a reference molecule, defining a molecular surface and sampling a point cloud, PCA dimensionality reduction, and calculating a first basic form; the contents of loading a reference molecule are as follows:

[0010] The molecule is loaded from the SMILES string of the invalid drug, explicit hydrogen atoms are added using RDKit's Chem.AddHs(), an initial 3D conformation is generated using AllChem.EmbedMolecule, and the structure is optimized based on the MMFF force field using AllChem.MMFFOptimizeMolecule. MMFF optimization minimizes 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.

[0011] Preferably, the molecular surface is defined and the content of the sampled point cloud is as follows:

[0012] The molecular surface is defined as the union of atomic van der Waals spheres, a two-dimensional manifold Generate point cloud using adaptive sampling The 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] Among them, 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 j-th point; r i Represents the three-dimensional position vector of the base point, i is 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;

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

[0016] Preferably, the PCA dimensionality reduction content is as follows:

[0017] Apply principal component analysis (PCA) to the point cloud and extract the first two principal components v1 and v2. v2 is used as the basis vector of the two-dimensional coordinate system, and the third principal component v3 is used 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] Characteristic decomposition yields v1, v2, and v3; the atomic projection coordinates are:

[0021]

[0022] Height is

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

[0024] The height field h(u,v) is fitted by Gaussian process regression GPR and modeled with RBF kernel:

[0025]

[0026] Among them, σ 2 is the signal variance; l is the length scale parameter;

[0027] Calculate the coefficients E, F, G of the first fundamental form:

[0028]

[0029]

[0030] The surface distance metric is:

[0031] ds 2 =Edu 2 +2Fdudv+Gdv 2 ;

[0032] The projection distance is:

[0033]

[0034] Among them, d ij is the Euclidean distance between atoms i and j.

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

[0036]

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

[0038] The inverse process predicts noise ∈θ (q t , t ), generate new coordinates; spherical constraints are implemented by projected gradient method:

[0039]

[0040] Among them, ∈ θ is the noise prediction function modeled by the neural network parameter θ; c is the coordinate of the sphere center; R is the radius of the sphere; Indicates the direction from the center of the sphere to p i The unit vector of

[0041] The total loss function is:

[0042] L=L DDPM +λ1L bond +λ2L geom ;

[0043] Among them, L DDPM is the standard diffusion model loss; L bond is the chemical bond constraint loss; L geom is the geometric constraint loss; λ1 and λ2 are weight parameters for balancing different loss terms.

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

[0045] Extract the chemical bond set of a reference molecule Calculate the three-dimensional bond distance d ij =|r i -r j | and 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 Represents the three-dimensional coordinates of atoms i and j in the molecular structure.

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

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

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

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

[0055] Optimization using the REINFORCE algorithm:

[0056]

[0057] in, represents the gradient of the policy parameter θ to the objective function J(θ); π θ (a t |s t ) represents the policy function with parameter θ, in state s t Next take action a t The probability of ; R(τ) represents the cumulative reward of trajectory τ; b(s t ) represents the baseline function; Represents the strategy based on π θ The expectation of the sampled trajectory τ.

[0058] Therefore, the present invention adopts the above-mentioned cancer targeted drug generation method based on differential geometry and reinforcement learning, which has the following beneficial effects:

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

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

[0061] (3) By introducing chemical bond constraints (such as covalent bonds, ring structures and other chemical rules), the generated molecules are ensured to meet not only the spatial geometry requirements of physics but also chemical rationality; this effectively avoids the problems of unnatural structures and chemical instability that may occur during the generation process, greatly improving the synthetic feasibility of the generated molecules;

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

[0063] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Schematic diagram of a flow chart of an embodiment of the present invention;

[0065] Figure 2 The existing VAE / GAN generation process of the embodiment of the present invention is based only on SMILES and ignores three-dimensional geometry;

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

[0067] Figure 4 The generation process of the height field and metric tensors E, F, and G is shown in the embodiment of the present invention;

[0068] Figure 5 The DDPM generation process of an embodiment of the present invention shows a schematic diagram of reverse denoising from noise to molecular conformation;

[0069] Figure 6 A schematic diagram showing the interaction between the policy network and the reward function according to an embodiment of the present invention. DETAILED DESCRIPTION

[0070] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort shall fall within the scope of protection of the present invention.

[0071] Example 1:

[0072] See also Figures 1-6 A method for generating cancer-targeted drugs based on differential geometry and reinforcement learning, using failed drugs as a reference, includes the following steps:

[0073] S1. Use differential geometry to accurately capture the geometric characteristics of molecular surface curvature and local structure.

[0074] S1 includes loading a reference molecule, defining the molecular surface and sampling the point cloud, PCA dimensionality reduction, and calculating the first basic form. The contents of loading a reference molecule are as follows:

[0075] The molecule is loaded from the SMILES string of the invalid drug, explicit hydrogen atoms are added using RDKit's Chem.AddHs(), an initial 3D conformation is generated using AllChem.EmbedMolecule, and the structure is optimized based on the MMFF force field using AllChem.MMFFOptimizeMolecule. MMFF optimization minimizes 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. This provides an accurate 3D geometric starting point, laying the foundation for subsequent differential geometry analysis and ensuring 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, a two-dimensional manifold Generate point cloud using adaptive sampling 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] Among them, 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 j-th point; r i Represents the three-dimensional position vector of the base point, i is 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] The point cloud is a discrete representation of the molecular surface manifold. Uniform sampling ensures complete capture of geometric information. Adaptive sampling dynamically adjusts the number of points based on molecule size, balancing computational efficiency and accuracy. This provides high-quality surface data, serving as input for differential geometry analysis and addressing the problem of existing techniques neglecting 3D geometry.

[0081] Use PyTorch's torch.rand to implement GPU parallel sampling and cache point cloud (pointcloud_cache.npy).

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

[0083] Apply principal component analysis (PCA) to the point cloud and extract the first two principal components v1 and v2. v2 is used as the basis vector of the two-dimensional coordinate system, and the third principal component v3 is used 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] Characteristic decomposition yields v1, v2, and v3; the atomic projection coordinates are:

[0087]

[0088] Height is

[0089] PCA maximizes the projected variance to preserve key geometric information, reducing the dimensionality of a three-dimensional manifold problem to a two-dimensional parameter space, making it easier to analyze surfaces. It is used to simplify surface parameterization, reduce computational complexity, and provide a foundation for first fundamental form calculations.

[0090] Use Scikit-learn’s PCA implementation and cache the results (pca_cache.npy).

[0091] The first basic form of calculation is as follows:

[0092] The height field h(u,v) is fitted by Gaussian process regression GPR and modeled with RBF kernel:

[0093]

[0094] Among them, 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 surface distance metric is:

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

[0099] The projection distance is:

[0100]

[0101] Among them, d ij is the Euclidean distance between atoms i and j.

[0102] The first fundamental form defines the intrinsic geometric metric on the surface, and the non-parametric modeling of GPR ensures the smoothness and accuracy of the height field. It is used to precisely control the geometry of molecular surfaces and solve the problem of low 3D structure matching in existing technologies.

[0103] Implement GPU-accelerated GPR using GPyTorch, caching E, F, G (metric_cache.npy).

[0104] S2. The denoised diffusion probability model (DDPM) is used in combination with chemical bond constraints to generate diverse and chemically plausible candidate molecules.

[0105] A Transformer-based DDPM is used to generate new conformations within a spherical region centered at the molecular center of mass; the forward process is:

[0106]

[0107] Among them, q t represents the latent variable at time step t; α t is the predefined diffusion coefficient; ∈ is the standard Gaussian noise;

[0108] The inverse process predicts noise ∈ θ (q t ,t), generate new coordinates; spherical constraints are implemented by the projected gradient method:

[0109]

[0110] Among them, ∈ θ is the noise prediction function modeled by the neural network parameter θ; c is the coordinate of the sphere center; R is the radius of the sphere; Indicates the direction from the center of the sphere to p i The unit vector of

[0111] The total loss function is:

[0112] L=L DDPM +λ1Lbond +λ2L geom ;

[0113] Among them, L DDPM is the standard diffusion model loss; L bond is the chemical bond constraint loss; L geom is the geometric constraint loss; λ1 and λ2 are weight parameters for balancing different loss terms.

[0114] DDPM generates diverse conformations through multi-step denoising. The Transformer's self-attention mechanism captures long-range interactions between atoms, and spherical constraints limit the generation range. It is used to generate diverse and geometrically precise molecules, addressing the single conformational limitations of existing technologies.

[0115] Use PyTorch to implement Transformer, GPU-accelerated training, and batch generate 100 conformations.

[0116] The chemical bond constraints are as follows:

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

[0118]

[0119] The three-dimensional bond distances constrain the molecular topology, the projected distances are embedded in the surface geometry in conjunction with the first fundamental form, and the loss function is optimized via gradient descent. This ensures that the generated molecules are chemically sound and mitigates chemical defects found in existing technologies.

[0120] Use PyTorch automatic differentiation to compute gradients and cache bond distances (bond_cache.npy).

[0121] Perform post-processing and verification:

[0122] The molecules generated by MMFF optimization are verified by calling Chem.SanitizeMol of RDKit. MMFF RDKit checks valence states and bond orders based on a chemical rule library to improve molecular practicality and ensure chemical and physical stability.

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

[0124] S3. Genetic algorithms are used to mutate and crossover candidate molecules to generate a diverse molecular library, and reinforcement learning is used to optimize candidate molecules and screen for efficient candidate molecules with the binding affinity of the mutant target protein as a reward.

[0125] First, we initialize the candidate molecules with a population size of 1000 and use a fitness function to balance geometric similarity and chemical plausibility:

[0126]

[0127] Then, mutation and crossover are performed. Mutation is to randomly replace atoms (C / N / O, probability 0.1) and adjust bond types (single bond / double bond, probability 0.05). Crossover is to exchange fragments based on subgraph matching, and finally selection is performed.

[0128] Genetic algorithms expand chemical space by simulating natural selection, increasing molecular diversity through mutation and crossover, and balancing geometric similarity and chemical rationality through fitness functions. They are used to generate diverse molecular libraries, addressing the inadequate chemical space coverage of existing technologies.

[0129] Perform mutation and crossover in parallel, caching the population (generation_cache.npy).

[0130] To use reinforcement learning for affinity optimization, we first define the MDP, where the state is the molecule (SMILES and coordinates), the action is the molecule edit, and the reward is the binding affinity:

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

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

[0133] Optimization using the REINFORCE algorithm:

[0134]

[0135] in, represents the gradient of the policy parameter θ to the objective function J(θ); π θ (a t |s t ) represents the policy function with parameter θ, in state s t Next take action a tThe probability of ; R(τ) represents the cumulative reward of trajectory τ; b(s t ) represents the baseline function; Represents the strategy based on π θ The expectation of the sampled trajectory τ.

[0136] Reinforcement learning dynamically optimizes molecules through reward feedback, and strategy gradient methods ensure efficient convergence. It is used to optimize binding affinity for mutant target proteins, addressing the insufficient efficacy of existing technologies.

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

[0138] The optimized molecules were subjected to MMFF optimization and RDKit validation, and the 10 molecules with the highest binding affinity were selected for screening highly effective targeting molecules to ensure practicality.

[0139] Verify in parallel and store the final candidate molecules (final_molecules.npy).

[0140] Example 2:

[0141] Based on the generative method of Riemann manifolds and GANs, the Riemann manifold is used to replace the first fundamental form to describe the molecular surface geometry, and combined with GAN to generate molecular conformations. The GAN generator outputs molecular coordinates, and the discriminator evaluates geometric similarity and chemical plausibility.

[0142] Implementation steps:

[0143] Load a reference molecule and optimize its 3D conformation; sample the point cloud and calculate the Riemann metric tensor; train a GAN, where the generator takes noise as input and outputs molecular coordinates, and the discriminator is based on geometric and chemical losses; generate molecules through MMFF optimization and RDKit verification.

[0144] An alternative to differential geometry and DDPM, retaining geometric accuracy but with slightly lower computational complexity and diversity.

[0145] Example 3:

[0146] The optimization method based on simulated annealing and molecular docking takes the reference molecule as the starting point, performs molecular editing through simulated annealing (Simulated Annealing), and combines molecular docking (AutoDockVina) to directly optimize the binding affinity.

[0147] Implementation steps:

[0148] Load reference molecules and initialize the molecule pool; randomly adjust atom types, bond types, or conformations through simulated annealing, with acceptance probability based on affinity changes; calculate binding affinity using molecular docking and screen for high-affinity molecules; MMFF optimization and RDKit validation.

[0149] It replaces genetic algorithms and reinforcement learning, simplifies the optimization process, but has limited exploration capabilities.

[0150] Example 4:

[0151] A generation method based on graph neural network (GNN) takes molecular graphs (nodes are atoms and edges are chemical bonds) as input, uses GNN to generate new molecules, and combines molecular docking to optimize affinity.

[0152] Implementation steps:

[0153] Convert reference molecules to graph representations; train GNNs, predict node and edge addition / deletion operations, and generate new molecules; assess affinity using molecular docking and iterative optimization; MMFF optimization and RDKit validation.

[0154] It replaces DDPM and differential geometry, simplifies geometric modeling, but has poor geometric accuracy.

[0155] Therefore, the present invention adopts the above-mentioned cancer-targeted drug generation method based on differential geometry and reinforcement learning. Through precise mathematical modeling, efficient generation model and dynamic optimization mechanism, it solves the key problems of the existing technology in cancer-targeted drug design and provides a feasible 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 invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for generating cancer-targeted drugs based on differential geometry and reinforcement learning, characterized in that: The following steps are involved: S1. Use differential geometry to accurately capture the geometric features of molecular surface curvature and local structure; S2. Use the denoised diffusion probability model (DDPM) combined with chemical bond constraints to generate diverse and chemically plausible candidate molecules. S3. Genetic algorithms are used to mutate and crossover candidate molecules to generate a diverse molecular library, and reinforcement learning is used to optimize candidate molecules and screen for efficient candidate molecules with the binding affinity of the mutant target protein as a reward.

2. The method for generating cancer-targeted drugs based on differential geometry and reinforcement learning according to claim 1, characterized in that: S1 includes loading a reference molecule, defining the molecular surface and sampling the point cloud, PCA dimensionality reduction, and calculating the first basic form. The contents of loading a reference molecule are as follows: The molecule is loaded from the SMILES string of the invalid drug, explicit hydrogen atoms are added using RDKit's Chem.AddHs(), an initial 3D conformation is generated using AllChem.EmbedMolecule, and the structure is optimized based on the MMFF force field using AllChem.MMFFOptimizeMolecule. MMFF optimization minimizes 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.

3. The method for generating cancer targeted drugs based on differential geometry and reinforcement learning according to claim 2, characterized in that: The content of defining the molecular surface and sampling the point cloud is as follows: The molecular surface is defined as the union of atomic van der Waals spheres, a two-dimensional manifold Generate point cloud using adaptive sampling The sampling points are based on spherical coordinates: p j =r i +R vdw,i (sinφ j cosθ j ,sinφ j sinθ j ,cosφ j ); Among them, 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 j-th point; r i Represents the three-dimensional position vector of the base point, i is 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; Number of sampling points M = min(1000,50N).

4. The method for generating cancer targeted drugs based on differential geometry and reinforcement learning according to claim 2, characterized in that: The PCA dimensionality reduction content is as follows: Apply principal component analysis (PCA) to the point cloud and extract the first two principal components v1 and v2. v2 is used as the basis vector of the two-dimensional coordinate system, and the third principal component v3 is used as the normal vector. The point cloud covariance matrix is: Where M represents the total number of points in the point cloud; represents the centroid of the point cloud; Characteristic decomposition yields v1, v2, and v3; the atomic projection coordinates are: Height is 5. The method for generating cancer targeted drugs based on differential geometry and reinforcement learning according to claim 4, characterized in that: The first basic form of calculation is as follows: The height field h(u,v) is fitted by Gaussian process regression GPR and modeled with RBF kernel: Among them, σ 2 is the signal variance; l is the length scale parameter; Calculate the coefficients E, F, G of the first fundamental form: The surface distance metric is: ds 2 =Edu 2 +2Fdudv+Gdv 2 ; The projection distance is: Among them, d ij is the Euclidean distance between atoms i and j.

6. The method for generating cancer-targeted drugs based on differential geometry and reinforcement learning according to claim 5, characterized in that: In S2, a new conformation is generated using Transformer-based DDPM within a spherical region centered on the molecular center of mass; the forward process is: Among them, q t represents the latent variable at time step t; α t is the predefined diffusion coefficient; ∈ is the standard Gaussian noise; The inverse process predicts noise ∈ θ (q t ,t), generate new coordinates; spherical constraints are implemented by the projected gradient method: Among them, ∈ θ is the noise prediction function modeled by the neural network parameter θ; c is the coordinate of the sphere center; R is the radius of the sphere; Indicates the direction from the center of the sphere to p i The unit vector of The total loss function is: L=L DDPM +λ1L bond +λ2L geom ; Among them, L DDPM is the standard diffusion model loss; L bond is the chemical bond constraint loss; L geom is the geometric constraint loss; λ1 and λ2 are weight parameters for balancing different loss terms.

7. The method for generating cancer targeted drugs based on differential geometry and reinforcement learning according to claim 5, characterized in that: The chemical bond constraints are as follows: Extract the chemical bond set of a reference molecule Calculate the three-dimensional bond distance d ij =|r i -r j | and projection distance Define the loss function:

8. The method for generating cancer-targeted drugs based on differential geometry and reinforcement learning according to claim 5, characterized in that: The fitness function in the genetic algorithm in S3 is: ChemValid(m); Where m represents the molecular structure to be evaluated; p i and p j Represents the three-dimensional coordinates of atoms i and j in the molecular structure.

9. The method for generating cancer-targeted drugs based on differential geometry and reinforcement learning according to claim 5, characterized in that: The affinity optimization content in S3 is as follows: First, define the MDP, where the state is the molecule (SMILES and coordinates), the action is the molecule edit, and the reward is the binding affinity: R(m)=-BindingEnergy(m,P mut )+λChemValid(m); Among them, R(m) represents the score value of molecule m; BindingEnergy(m,P mut ) represents the small molecule m and the mutant protein P mut The binding energy of ;λ represents the weight coefficient;ChemValid(m) represents the chemical validity index; Optimization using the REINFORCE algorithm: in, represents the gradient of the policy parameter θ to the objective function J(θ); π θ (a t |s t ) represents the policy function with parameter θ, in state s t Next take action a t The probability of ; R(τ) represents the cumulative reward of trajectory τ; b(s t ) represents the baseline function; Represents the strategy based on π θ The expectation of the sampled trajectory τ.

Images