Drug design model based on structure-integrated Bayesian flow networks and diffusion models

Abstract

The present invention belongs to the field of bioinformatics technology, and in particular to a drug design model based on a structure-integrated Bayesian flow network and a diffusion model. The drug design model integrates the advantages of the Bayesian flow network and the diffusion model through stochastic differential equations. Atomic coordinates are sampled by stochastic differential equations, and the Bayesian flow samples the atomic type, thereby shortening the sampling time and improving the properties of the molecule. In addition, a continuous-time loss function with an analytical solution is introduced to improve the stability of the stochastic differential equation by achieving a smoother transition to the target distribution. By comparing with other models on the commonly used data set CrossDocked2020 in this field, the drug design model has achieved good results in generating the binding affinity, drug-likeness index and conformational stability of molecules, and the sampling efficiency has been improved by 25%.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, and specifically relates to a drug design model based on structure-integrated Bayesian flow networks and diffusion models. Background Technology

[0002] Structure-based drug design (SBDD) plays a crucial role in drug discovery by using structural information from target protein pockets to guide the generation of three-dimensional molecules. This approach helps design molecules with tight binding affinity and favorable drug properties.

[0003] Recent advances in deep generative modeling have provided powerful tools for tackling this complex task. Early studies employed autoregressive models to generate molecules at the atom or fragment level. However, these methods produce unnatural sequences and fail to account for the overall molecular structure. In contrast, diffusion models (DM), which start with a noisy prior and reconstruct the complete molecule through iterative denoising, have become the dominant approach in the field of deep generative modeling (SBDD). These methods can simultaneously model both local and global interactions between molecular atoms, resulting in significantly better performance.

[0004] However, molecules simultaneously contain continuous atomic coordinates and discrete atomic types. While diffusion models excel at handling continuous variables, they still face challenges when dealing with mixed modes, potentially leading to suboptimal or even chemically invalid molecular structures. Furthermore, diffusion models suffer from excessively long sampling processes, typically requiring 1000 denoising steps to generate 100 molecules, which can ultimately delay the discovery of viable drug candidates. Recently, the MolCRAFT model employs Bayesian Flow Networks (BFNs) to generate molecules in parameter space, effectively capturing the multimodal joint distribution of molecules, providing a new perspective on molecule generation and accelerating sampling speed—generating 100 molecules in just 100 steps. Despite the excellent performance of the MolCRAFT model, there is still room for improvement in the docking affinity of its generated molecules, and the sampling efficiency requires further optimization. Summary of the Invention

[0005] This invention proposes a drug design model based on a structure-integrated Bayesian flow network and diffusion model, which solves the problems of insufficient molecular docking affinity and slow sampling efficiency in existing MolCRAFT models.

[0006] The technical solution of this invention is implemented as follows:

[0007] A drug design model based on a structured integrated Bayesian flow network and diffusion model, named MolSolver, comprises the following components:

[0008] Feature extraction module: Reads the three-dimensional coordinates and atom type features of protein pockets and ligand molecules, and encodes and fuses the features;

[0009] Feature update and molecular prediction module: The features extracted by the feature extraction module are input into the SE(3) equal-variable graph neural network to predict the atomic coordinates of the molecule. and atomic type In the In this prediction, the prediction and And the MolSolver model is trained using a set continuous-time loss;

[0010] Parameter Update Module: This module uses the predicted atomic coordinates and atom types from the feature update and molecular prediction modules to update parameters. It employs a hybrid sampling method and a stochastic differential equation solver to update the atomic coordinate distribution parameters. The parameters of the atom type distribution are updated using Bayesian flow. In the In this update, the parameters are updated. and .

[0011] Through the above technical solutions, the MolSolver model fully and efficiently represents and integrates the three-dimensional coordinate features and atom type features of molecules and proteins, and can sample molecules in a continuous parameter space while maintaining low variance. SE(3) and other variable graph neural networks can maintain the translational and rotational invariance of molecules and proteins during feature updates. Continuous-time loss has analytical solutions, which can improve molecular conformational stability. A hybrid sampling method based on SDE solvers and Bayesian flow can improve the molecular quality and sampling speed.

[0012] Optionally, after the feature extraction module reads the features of the protein pocket and ligand molecules, it stitches together the three-dimensional coordinates of the protein pocket and ligand molecules to construct a K-nearest neighbor graph. (KNN graph), K=32, then construct edge features through the KNN graph, including edge type and edge distance, etc.; and combine the atom type of ligand molecules and time. The data are concatenated and input into a single-layer linear fully connected network to obtain the embedded representations of atom type and time features. Similarly, the atom type features of the protein pocket are also input into a single-layer linear fully connected network to obtain the embedded representations of the atom type features of the protein pocket. Finally, the embedded representations of the atom type features of the molecule and the protein are concatenated for unified processing. The concatenated atom coordinate feature embedded representation, the atom type feature embedded representation, and the constructed edge features are used as the initial features input into the feature update and molecule prediction modules.

[0013] Using the above technical solution, the MolSolver model successfully extracted the atomic coordinate features, atomic type features, and edge features of the constructed KNN graph of molecules and proteins, and encoded them. These features can be used as inputs to the equivariant graph neural network of the feature update and molecular prediction modules.

[0014] Optionally, in the feature update and molecular prediction modules, the SE(3) equivalent variable graph neural network is used in the first... +1 layer to the first Layer Atom coordinate features and type features The update method is as follows:

[0015] ;

[0016] ;

[0017] ;

[0018] in, It is the first +1 layer of atoms Coordinate characteristics, It is time. It is the first Atoms of the layer Type characteristics, It is the first +1 layer of atoms Type characteristics, It is the first Atoms of the layer Type characteristics, It is the first +1 layer of atoms Type characteristics, It is an update to atoms The change in coordinates, It is the first The atom in the first Layer coordinate characteristics, Represents a node In K-nearest neighbor graph The set of neighbors in the middle, and There are two attention modules; Represents atoms and The Euclidean distance between them; It is an additional feature used to distinguish atoms. and atoms The edge between them belongs to protein-protein, ligand-ligand, or protein-ligand; It is a ligand molecule mask used to ensure that only the ligand coordinates are updated without modifying the protein coordinates;

[0019] After iterative updates, the SE(3) and other variable graph neural networks obtain the atomic coordinate features and type features of the current ligand molecules.

[0020] Through the above technical solutions, the SE(3) and other variable graph neural networks can update the atomic and type features of molecules and proteins layer by layer, and predict molecules based on the current features. The graph neural network can maintain the translation and rotation invariance of coordinate features in three-dimensional space, which is crucial for molecular generation.

[0021] Optionally, in the feature update and molecular prediction modules, the continuous-time loss function is implemented by calculating the KL divergence between two noisy sample distributions; considering that both atomic coordinates and noise are modeled as Gaussian distributions, the loss of atomic coordinates... Written as:

[0022] ;

[0023] in, Indicates hyperparameters, Parameters representing the atomic coordinate distribution of ligand molecules. Represents the actual atomic coordinates, Indicates the predicted atomic coordinates, Indicates time, Indicates protein;

[0024] Similarly, for discrete atom types, the loss of its parameters... The analytical solution is obtained by calculating the KL divergence between the two Gaussian distributions.

[0025] ;

[0026] in, The number of types of atoms. Indicates time, ,in, For category indexing of One-hot code vector, Yes The prediction yes Noise scheduling during the event;

[0027] Ultimately, the overall training objective is the sum of the atomic coordinate loss and the atomic type loss:

[0028] .

[0029] Using the above technical solution, the MolSolver model can be trained using this loss. The continuous-time loss with analytical solutions helps to make the transition from the data distribution to the target distribution smoother and helps to improve the stability of molecular conformation.

[0030] Optionally, in the parameter update module, a second-order multi-step fast SDE solver is used to update the molecular coordinate parameters. The formula is as follows:

[0031] ;

[0032] in, Indicates the first The parameters of the coordinate distribution during each sampling step; ; ; Indicates the first Second and third The predicted coordinate values ​​of the second time and The finite difference slope between them; It is noise scheduling, and ; It is standard Gaussian noise. ,and .

[0033] Optionally, in the parameter update module, the molecular coordinates are first processed from the standard normal distribution. Mid-sampling, therefore Initialized as a vector In each iteration, the SDE solver simultaneously references the first... Second and third Next, the estimated coordinates are used to update... Gaussian noise is added to maintain the randomness of the process; when the parameters of the coordinate distribution converge to the final value... Then, it is fed into a neural network to predict the final molecular coordinates. .

[0034] Through the above technical solutions, the MolSolver model can efficiently update molecular coordinate parameters, improve the properties of sampled molecules, and increase sampling efficiency by using this second-order multi-step fast SDE solver.

[0035] Optionally, in the parameter update module, given the atom type predicted by the SE(3) equivalent graph neural network. and noise scheduling function The MolSolver model uses Bayesian flow distribution. Update : ;

[0036] in, Indicates the predicted atom type; Expressing expectations; Represents a normal distribution; Indicates a protein pocket; Indicates time; It is a characteristic of the atomic type of the molecule. For category indexing of One-hot code vector; Represents a unit vector; Represents the Dirichlet function; This represents the parameters used to obtain the updated atom type distribution through a formula; Indicates noise variables;

[0037] Real Atom Types The distribution model is a classification distribution. ,in, This represents the total number of atoms in a molecule. The matrix represents the number of types of atoms. Each line gives the probability that a given atom belongs to each of the possible types.

[0038] Optionally, in the parameter update module, the atom type is initially sampled from a uniform distribution, therefore, Initialize to When a neural network gives an atomic type prediction Then, the MolSolver model first samples noise variables from a Gaussian distribution. Then apply the softmax function to it to obtain ;

[0039] The MolSolver model updates both atomic coordinate parameters and atomic type parameters simultaneously; in the... In the next iteration, the parameters obtained in the previous step are... , Protein pocket and time Input into a neural network to obtain molecular predictions , which includes coordinates and type Subsequently, the MolSolver model simultaneously performs two updates: refining the coordinate parameters using a second-order multistep stochastic differential equation solver; and updating the type parameters using a Bayesian flow distribution. This simultaneous update cycle performs N=80 steps in total; the final molecular prediction is... .

[0040] Through the above technical solution, MolSolver updates the parameters of the atomic type distribution of molecules through Bayesian flow and updates the parameters of the atomic coordinate distribution of molecules through SDE solver. This hybrid sampling method improves the properties of sampled molecules and the sampling efficiency.

[0041] After adopting the above technical solution, the beneficial effects of the present invention are:

[0042] This invention designs a drug design model, MolSolver, that integrates a Bayesian flow network and a diffusion model for protein structure. Specifically, the MolSolver model employs a hybrid sampling method, utilizing an SDE (Stochastic Differential Equation) solver to update atomic coordinate parameters and a Bayesian flow distribution to update atomic type parameters. Furthermore, the MolSolver model uses a continuous-time loss function with analytical solutions to ensure a smooth transition from the data distribution to the target distribution. Comprehensive experimental evaluations have been conducted, demonstrating that the drug design model not only improves binding affinity, drug-likeness indices, and conformational stability but also increases sampling efficiency by 25%.

[0043] While Bayesian flow network-based methods have achieved excellent performance, their sampling strategies remain to be explored, which is crucial for efficient molecule generation. To address this issue, this invention proposes a model called MolSolver. The MolSolver model uses a Bayesian flow network as its backbone and samples in the parameter space to ensure both the quality and efficiency of molecule generation. It employs a hybrid sampling strategy, utilizing an SDE solver to update atomic coordinates and leveraging Bayesian flow to optimize atom types. To enhance the stability of the SDE solver, the MolSolver model uses a continuous loss function with analytical solutions. This improvement facilitates a smoother transition from the data distribution to the target distribution, thereby enhancing sampling stability. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A diagram of the MolSovler model;

[0046] Figure 2 Flowchart for coordinate prediction;

[0047] Figure 3A schematic diagram showing the strain energy of molecules generated by different methods, the proportion of generated molecules whose root mean square deviation from their optimal conformation is less than 1 Å, and the number of atomic collisions.

[0048] Figure 4 The sampling efficiency graphs for different methods are shown.

[0049] Figure 5 This is a schematic diagram of the generated molecules. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] This application discloses a drug design model based on a structure-integrated Bayesian flow network and diffusion model.

[0052] Example: According to Figures 1 to 5 As shown, a drug design model based on structured integrated Bayesian flow networks and diffusion models is presented, with the following details:

[0053] 1. Task Description

[0054] In structure-based drug design, this invention treats both protein binding sites and small molecules as attributed three-dimensional point clouds: protein pockets are used... It means that among them It's in my pocket The three-dimensional coordinates of each atom, These are the corresponding atomic feature vectors; the goal is to generate small molecules. ,in and They are small molecules The coordinates and characteristics of each atom. The generated small molecule must fit precisely into the pocket in space, while exhibiting high binding affinity, good drug-like properties, and a stable binding conformation. In other words, it must not only "fit in" but also "stick firmly" and meet all the indicators of druggability.

[0055] 2. Dataset

[0056] This invention utilizes the widely used public dataset CrossDocked2020 to evaluate the performance of the MolSolver model. This dataset contains PDF files of proteins and SDF files of ligand molecules, including 3D coordinates and feature information of atoms. First, the PDB structure of the ligand binding pockets defined by Pocketome is calibrated; then, the smina tool is used to perform comprehensive cross-docking of each ligand with each pocket receptor; finally, the 22.5 million generated conformations are annotated using PDBBind 2017 affinity data. After applying RMSD benchmark filtering and a 30% sequence homology threshold, 100,000 protein-ligand pairs are selected for training, while 100 proteins are retained for testing.

[0057] 3. Overall Architecture of the MolSolver Model

[0058] The overall architecture of the MolSolver model combined with Figure 1 The model skeleton adopts the Bayesian flow model. Based on this, the advantages of the diffusion model are combined with stochastic differential equations to achieve better model performance. The generated molecules have improved binding affinity, drug-like properties and conformational stability, and the sampling efficiency has also been improved. The MolSolver model does not operate on the data, but on the parameters of the data distribution. For a given real molecule, its coordinate distribution and type distribution are characterized by certain parameters. It is assumed that the initially given molecule follows a certain distribution, which is called the input distribution. During the training process, the loss is reduced by calculating the KL divergence of the sender distribution and the receiver distribution, so that the model can converge. The sender distribution is obtained by adding noise to the input distribution. For the receiver distribution, the input distribution is first passed through the SE(3) and other variable graph neural network to obtain the output distribution. The same noise as the input distribution is applied to the output distribution to obtain the receiver distribution. For the parameter update, the real initial molecule information, including atomic coordinates and atomic types, is predicted by the SE(3) and other variable graph neural network. Then, the parameters are updated by the SDE solver and Bayesian flow respectively, so as to gradually reach the real parameters.

[0059] 4. Continuous time loss

[0060] The continuous-time loss function is implemented by calculating the KL divergence between two noisy sample distributions; considering that both atomic coordinates and noise are modeled as Gaussian distributions, the loss of atomic coordinates... Written as:

[0061] ;

[0062] in, Indicates hyperparameters, Parameters representing the atomic coordinate distribution of ligand molecules. Represents the actual atomic coordinates, Indicates the predicted atomic coordinates, Indicates time, It refers to protein.

[0063] Similarly, for discrete atom types, the parameter loss... The analytical solution is obtained by calculating the KL divergence between the two Gaussian distributions.

[0064] ;

[0065] in, The number of types of atoms. Indicates time, ,in, For category indexing of One-hot code vector, Yes The prediction yes Noise scheduling during the event;

[0066] Ultimately, the overall training objective is the sum of the atomic coordinate loss and the atomic type loss:

[0067] .

[0068] 5. Feature extraction and SE(3) equivalent graph neural network

[0069] This section will combine Figure 2 The following explanation is provided. Regarding feature representation, for ligands, atom types are encoded using one-hot vectors, which simultaneously capture chemical element types (C, N, O, F, P, S, Cl) and aromaticity information. For protein pockets, each atom is characterized by three complementary features: (1) the one-hot encoding of the atomic element (H, C, N, O, S, Se), (2) a 20-dimensional one-hot vector indicating its corresponding amino acid residue type, and (3) a binary flag indicating whether the atom belongs to the protein backbone. Atoms in both molecules and proteins are accompanied by 3D spatial coordinates and include a binary indicator to clearly distinguish between molecule and protein atoms. These coordinates are then concatenated to construct a K-nearest neighbor graph, where K=32, and based on this, edge types (ligand-to-ligand, ligand-to-protein, protein-to-ligand, protein-to-protein) and other edge features are constructed. RDkit is used to process ligand and protein files during feature extraction.

[0070] Molecular type features are linked to time and then processed through a single-layer linear neural network to obtain a fused embedding representation. Protein type features are also processed through a single-layer linear neural network to obtain an embedding representation. The two sets of embeddings are then merged and input into an SE(3) variable graph neural network.

[0071] To fully capture ligand-pocket interactions, the network backbone employs PosNet3D with SE(3) isovariability, effectively utilizing molecular geometry information while maintaining invariance to overall rotation and translation. The SE(3) isovariable graph neural network comprises nine layers, each containing two attention blocks. The layers are updated to obtain the atom type features and coordinates, where the first layer is... The characteristic of the first atom is obtained through the first Layer type characteristics and coordinates Update:

[0072] ;

[0073] ;

[0074] ;

[0075] in, It is the first +1 layer of atoms Coordinate characteristics, It is the first The type characteristics of atoms in layer +1 It is time; It is the first Atoms of the layer Type characteristics, It is the first +1 layer of atoms Type characteristics, It is the first Atoms of the layer Type characteristics, It is the first +1 layer of atoms Type characteristics; It is an update to atoms The change in coordinates; It is the first The atom in the first Layer coordinate characteristics; Represents a node In K-nearest neighbor graph The set of neighbors in the middle, and There are two attention modules; Represents atoms and The Euclidean distance between them; It is an additional feature used to distinguish atoms. and atoms The edge between them belongs to protein-protein, ligand-ligand, or protein-ligand; It is a ligand molecule mask used to ensure that only the ligand coordinates are updated without modifying the protein coordinates.

[0076] Finally, the prediction network outputs the predicted atomic coordinates and atomic types at each iteration step; in the post-processing stage, the MolSolver model uses OpenBabel to add chemical bonds between the molecular atoms.

[0077] 6. Atomic coordinate parameter update

[0078] To leverage the advantages of diffusion models (DM) within Bayesian flow networks (BFNs), this invention models atomic coordinate parameter updates as stochastic differential equations (SDEs) and utilizes an SDE solver for sampling, thereby enhancing the performance of existing Bayesian flow-based networks. This is the first time this strategy has been applied in structure-based drug design (SBDD).

[0079] For molecular coordinates ( (where the number of atoms in the molecule is the number of atoms), this invention considers it to follow a Gaussian distribution. ,in, It is the mean. It is variance. It's about precision. It is a unit vector. Therefore, the parameter set of the coordinate distribution is... Due to accuracy By noise factor Pre-set and meet , It is the first Step accuracy value, It is the first Step accuracy value, It is the first noise factor of step The value of . Therefore, during training, only the mean parameter is considered. Update.

[0080] The original update strategy originates from the Bayesian update function. Existing research has shown that this update is actually equivalent to the first-order discretization of a specific inverse stochastic differential equation.

[0081] ;

[0082] in, Represents noise, satisfying Therefore, it is possible to put The update process is considered to follow a reverse-time stochastic differential equation.

[0083] in, Indicates the first The parameters of the coordinate distribution during each sampling step; These are the predicted atomic coordinates; Is Noise scheduling in time It is in the The time of the step, and , Is Noise scheduling in time, and ; It's a hyperparameter; It is standard Gaussian noise.

[0084] To solve the equation while ensuring sampling efficiency and stability, this invention introduces a fast second-order multistep SDE solver with the sampling formula shown below.

[0085] ;

[0086] in, Indicates the first The parameters of the coordinate distribution during each sampling step; ; ; Indicates the first Second and third The predicted coordinate values ​​of the second time and The finite difference slope between them; It is noise scheduling, and ; It is standard Gaussian noise. ,and ; yes of Power of 1 yes of Power of 1.

[0087] Specifically, molecular coordinates are first derived from the standard normal distribution. Mid-sampling, therefore Initialized as a vector In each iteration, the SDE solver simultaneously references the current step and the previous step (the first step). Second and third Update the coordinates obtained from the (time) estimation. Gaussian noise is added to maintain the randomness of the process; when the parameters of the coordinate distribution converge to the final value... Then, it is fed into a neural network to predict the final molecular coordinates. .

[0088] 7. Atom type parameter update

[0089] Discrete atom types Distribution modeling as a classification distribution ,in, This represents the total number of atoms in a molecule. The number of types of atoms. Matrix Each line gives the probability that a given atom belongs to each of the possible types. Given the atom types predicted by the neural network... and noise scheduling function The MolSolver model uses Bayesian flow distribution. Update :

[0090] ;

[0091] in, Indicates the predicted atom type; Expressing expectations; Represents a normal distribution; Indicates a protein pocket; Indicates time; It is a characteristic of the atomic type of the molecule. For category indexing of One-hot code vector; Represents a unit vector; Represents the Dirichlet function; This represents the parameters used to obtain the updated atom type distribution through a formula; This represents a noise variable.

[0092] Specifically, the atom types are initially sampled from a uniform distribution; therefore, the initial atom type distribution parameters... Initialize to When a neural network gives an atom type prediction Then, the MolSolver model first samples noise variables from a Gaussian distribution. Then apply the softmax function to it to obtain .

[0093] The MolSolver model updates both atom coordinate parameters and atom type parameters simultaneously. In the... In the next iteration, the present invention will... The parameters of the coordinate distribution obtained by the step Parameters of type distribution Protein pocket and time Inputting the data into a neural network yields predicted molecules. This includes atomic coordinates. and atomic type Subsequently, the MolSolver model simultaneously performs two updates: refining the coordinate parameters using a second-order multistep stochastic differential equation solver; and updating the type parameters using a Bayesian flow distribution. This simultaneous update cycle consists of N=80 steps. The final molecule (containing atomic coordinates and atom types) is obtained as follows: , Represents the final atomic coordinates. Indicates the final atom type.

[0094] 8. Performance comparison of the MolSolver model with other prediction models

[0095] To verify the performance of the MolSolver model, following domain convention, this invention was tested on the CrossDocked2020 dataset. Using the same partitioning method as other models, a subset of 100,000 protein-ligand pairs was selected for training, and 100 proteins were reserved for testing. This invention compared the MolSolver model with six recent representative models and test set molecules. AR, Pocket2Mol, and FLAG are autoregressive models; the first two sample molecules atom-by-atom, while FLAG generates molecules fragment-by-fragment. TargetDiff and DecompDiff are state-of-the-art diffusion-based models. MolCRAFT is a state-of-the-art Bayesian flow network-based model.

[0096] This invention comprehensively evaluates these models in terms of binding affinity, molecular properties, conformational stability, and sampling efficiency. Binding affinity includes docking score, docking score minimization, and re-docking score. The docking score directly estimates the binding affinity based on the generated three-dimensional molecules; docking score minimization first performs local energy minimization before evaluation; the re-docking score involves the re-docking process and reflects the optimal binding affinity. Molecular properties include three indicators: synthetic accessibility, drug-likeness, and diversity. Synthetic accessibility represents the difficulty of ligand synthesis; drug-likeness is a quantitative estimate of drug similarity; diversity is assessed by calculating the average pairwise differences between molecules generated from each binding pocket. This invention also sets a success rate indicator, which is the proportion of molecules with high affinity and reasonable molecular properties (re-docking score < -8.18, drug-likeness > 0.25, synthetic accessibility > 0.59). Conformational stability includes strain energy, number of atomic collisions, and the proportion of molecules with optimal conformation among the generated molecules. Sampling efficiency includes sampling speed and generation success rate. Generation success rate refers to the proportion of molecules that simultaneously meet the requirements of chemical validity and structural integrity.

[0097] As shown in Table 1, the MolSolver model outperforms other strong baseline models on almost all affinity metrics. Taking docking score minimization as an example, the MolSolver model achieved an average score of -7.42 and a median score of -7.52, representing improvements of 2.1% and 3.6% respectively compared to the strongest baseline, the MolCRAFT model. On the re-docking score metric, the MolSolver model achieved an average score of -8.21 and a median score of -8.39, exceeding the MolCRAFT model by 3.7% and 4.7% respectively. These improvements demonstrate the MolSolver model's excellent ability to explore high-affinity candidate molecules and modulate protein structures for tight binding. In terms of molecular properties such as synthetic accessibility and diversity, the MolSolver model performs competitively with other methods, achieving a synthetic accessibility score of 0.66, second only to the MolCRAFT model among non-autoregressive methods. It should be noted that the higher synthetic accessibility values ​​of the Pocket2Mol and FLAG models are partly due to the smaller molecular sizes they generate. For the drug-likeness metric, the MolSolver model achieved the highest score of 0.58, outperforming all autoregressive and non-autoregressive methods. This demonstrates that the MolSolver model can effectively generate molecules with good drug-like properties. The overall success rate of the MolSolver model reached 27.8%, the highest among all benchmark methods, showcasing its superior ability to generate synthetically usable molecules while maintaining high affinity. Notably, except for synthetic accessibility, the MolSolver model outperformed the reference molecule in all metrics, further highlighting its effectiveness in capturing molecular-protein interactions and generating compounds with excellent drug-like properties.

[0098] Table 1. Overall results of reference molecules, molecules generated by various models, and molecules in the test set.

[0099] Model and test set docking score docking score Minimize docking score Minimize docking score Reconnecting scores Reconnecting scores Synthesis accessibility Drug-like properties diversity Success rate average median average median average median average average average test set -6.36 -6.46 -6.71 -6.49 -7.45 -7.26 0.73 0.48 — 25.0 AR -5.75 -5.64 -6.18 -5.88 -6.75 -6.62 0.64 0.51 0.70 7.1 Pocket2Mol -5.14 -4.70 -6.42 -5.82 -7.15 -6.79 0.76 0.57 0.69 24.4 FLAG 16.48 4.53 1.21 -4.04 -5.63 -6.61 0.70 0.49 0.70 14.1 TargetDiff -5.47 -6.30 -6.64 -6.83 -7.80 -7.91 0.58 0.48 0.72 10.5 DecompDiff -5.19 -5.27 -6.03 -6.00 -7.03 -7.16 0.66 0.51 0.73 14.9 MolCRAFT -6.59 -7.04 -7.27 -7.26 -7.92 -8.01 0.69 0.50 0.72 26.8 MolSolver -6.45 -7.18 -7.42 -7.52 -8.21 -8.39 0.66 0.58 0.71 27.8

[0100] like Figure 3 As shown, in terms of strain energy, the MolSolver model performs consistent with the strongest baseline MolCRAFT model, with a median of 208. Regarding the optimal conformation ratio and the number of atomic collisions, the MolSolver model achieves 42.6% and 6.83 respectively, both being the best values ​​among all models, highlighting the MolSolver model's ability to generate structurally stable conformations.

[0101] Figure 4 Scatter plots are provided to illustrate the relationship between generation success rate (horizontal axis) and corresponding sampling efficiency metrics for four diffusion models (DM) and Bayesian flow networks—TargetDiff, DecompDiff, MolCRAFT, and the MolSolver model of this invention. Generation success rate refers to the proportion of generated samples that simultaneously satisfy molecular chemical validity and structural integrity. Specifically, the MolSolver model achieves a generation success rate of 98.89%, ranking first among the four, surpassing MolCRAFT's 96.67%, TargetDiff's 90.36%, and DecompDiff's 82.92%. In terms of sampling speed, the MolSolver model also boasts the fastest sampling speed, approximately 1.69 (59 seconds to generate 100 molecules), significantly higher than TargetDiff's 0.127, DecompDiff's 0.091, and MolCRAFT's 1.35. Compared to the best baseline model, MolCRAFT, the MolSolver model improves sampling speed by approximately 25% (1.69 - 1.35) / 1.35. These results demonstrate that the MolSolver model combines high efficiency with superior performance in molecular generation tasks.

[0102] This invention randomly selected two proteins—3daf and 14gs—and, Figure 5 The paper presents a reference molecule, molecules generated by the MolCRAFT model, and molecules generated by the MolSolver model. For the generated molecules, this invention first prioritizes them according to docking score, synthetic accessibility, and drug-likeness, then selects the top two molecules for visualization. The results show that molecules generated by the MolSolver model perform better.

[0103] This invention proposes a novel method, the first in the field of structure-based drug design to combine Bayesian flow networks and diffusion models through stochastic differential equations. The MolSolver model employs a hybrid sampling strategy: updating atom coordinates using a higher-order SDE solver and updating atom types using a Bayesian flow distribution. Furthermore, the MolSolver model introduces a continuous loss function with closed-form solutions, enabling a smoother transition between the data distribution and the target distribution, thereby mitigating the instability problems commonly found in stochastic differential equations. Experiments show that the MolSolver model not only improves binding affinity, drug-likeness indices, and conformational stability, but also increases sampling efficiency by 25%.

[0104] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the technical solution of the present invention should be included within the protection scope of the present invention.

Claims

1. A drug design model based on structured integrated Bayesian flow networks and diffusion models, characterized in that, The model is named MolSolver, and the MolSolver model consists of the following parts: Feature extraction module: Reads the three-dimensional coordinates and atom type features of protein pockets and ligand molecules, and encodes and fuses the features; Feature update and molecular prediction module: The features extracted by the feature extraction module are input into the SE(3) equal-variable graph neural network to predict the atomic coordinates of the molecule. and atomic type , in In this prediction, the prediction and And the MolSolver model is trained using a set continuous-time loss; Parameter Update Module: This module uses the predicted atomic coordinates and atom types from the feature update and molecular prediction modules to update parameters, employing a stochastic differential equation solver to update the parameters of the atomic coordinate distribution. The parameters of the atom type distribution are updated using Bayesian flow. , in In this update, the parameters are updated. and ; In the feature update and molecular prediction modules, the continuous-time loss function is implemented by calculating the KL divergence between two noisy sample distributions; considering that both atomic coordinates and noise are modeled as Gaussian distributions, the loss of atomic coordinates... Written as: ； in, Indicates hyperparameters, Parameters representing the atomic coordinate distribution of ligand molecules. Represents the actual atomic coordinates, Indicates the predicted atomic coordinates, Indicates time, Indicates protein; Similarly, for discrete atom types, the loss of its parameters... The analytical solution is obtained by calculating the KL divergence between the two Gaussian distributions. ； in, The number of types of atoms. Indicates time, ,in, For category indexing of One-hot code vector, Yes The prediction yes Noise scheduling during the event; Ultimately, the overall training objective is the sum of the atomic coordinate loss and the atomic type loss: ； In the parameter update module, a second-order multi-step fast SDE solver is used to update the molecular coordinate parameters. The formula is as follows: ； in, Indicates the first The parameters of the coordinate distribution during each sampling step; ; ; Indicates the first Second and third The predicted coordinate values ​​of the second time and The finite difference slope between them; It is noise scheduling, and ; It is standard Gaussian noise. ,and .

2. The drug design model based on structure-integrated Bayesian flow network and diffusion model according to claim 1, characterized in that, In the feature extraction module, after reading the features of the protein pocket and ligand molecules, the three-dimensional coordinates of the protein pocket and ligand molecules are stitched together to construct a KNN graph, K=32, and edge features are constructed through the KNN graph. The atomic types and time of the ligand molecules The data are concatenated and input into a single-layer linear fully connected network to obtain the embedded representations of atom type and time features. The atom type features of the protein pocket are also input into a single-layer linear fully connected network to obtain the embedded representations of the atom type features of the protein pocket. Finally, the embedded representations of the atom type features of the molecule and the protein are concatenated. The concatenated atom coordinate feature embedded representation, the atom type feature embedded representation, and the constructed edge features are used as the initial features input into the feature update and molecule prediction modules.

3. The drug design model based on structure-integrated Bayesian flow network and diffusion model according to claim 1, characterized in that, In the feature update and molecular prediction modules, the SE(3) equivalent graph neural network is used in the first phase. +1 layer to the first Layer Atom coordinate features and type features The update method is as follows: ； ； ； in, It is +1 layer of atoms Coordinate characteristics, It is time. It is Atoms of the layer Type characteristics, It is +1 layer of atoms Type characteristics, It is Atoms of the layer Type characteristics, It is +1 layer of atoms Type characteristics, It is an update to atoms The change in coordinates, It is The atom in the first Layer coordinate characteristics, Represents a node In K-nearest neighbor graph The set of neighbors in the middle, and There are two attention modules; Represents atoms and The Euclidean distance between them; It is an additional feature used to distinguish atoms. and atoms The edge between them belongs to protein-protein, ligand-ligand, or protein-ligand; It is a ligand molecule mask used to ensure that only the ligand coordinates are updated without modifying the protein coordinates; After iterative updates, the SE(3) and other variable graph neural networks obtain the atomic coordinate features and type features of the current ligand molecules.

4. The drug design model based on structure-integrated Bayesian flow network and diffusion model according to claim 1, characterized in that, In the parameter update module, the molecular coordinates are first updated from the standard normal distribution. Mid-sampling, therefore Initialized as a vector In each iteration, the SDE solver simultaneously references the first... Second and third Next, the estimated coordinates are used to update Gaussian noise is added to maintain the randomness of the process; when the parameters of the coordinate distribution converge to the final value... Then, it is fed into a neural network to predict the final molecular coordinates. .

5. The drug design model based on structure-integrated Bayesian flow network and diffusion model according to claim 1, characterized in that, In the parameter update module, given the atom type predicted by the SE(3) equal-variable graph neural network... and noise scheduling function The MolSolver model uses Bayesian flow distribution. Update : ; in, Indicates the predicted atom type; Expressing expectations; Represents a normal distribution; Indicates a protein pocket; Indicates time; It is a characteristic of the atomic type of the molecule. For category indexing of One-hot code vector; Represents a unit vector; Represents the Dirichlet function; This represents the parameters used to obtain the updated atom type distribution through a formula; Indicates noise variables; Real Atom Types The distribution model is a classification distribution. ,in, This represents the total number of atoms in a molecule. The matrix represents the number of types of atoms. Each line gives the probability that a given atom belongs to each of the possible types.

6. The drug design model based on structure-integrated Bayesian flow network and diffusion model according to claim 5, characterized in that, In the parameter update module, the atom type is initially sampled from a uniform distribution, therefore, Initialize to When a neural network gives an atomic type prediction Then, the MolSolver model first samples noise variables from a Gaussian distribution. Then apply the softmax function to it to obtain ; The MolSolver model updates both atomic coordinate parameters and atomic type parameters simultaneously; in the... In the next iteration, the parameters obtained in the previous step are... , Protein pocket and time Input into a neural network to obtain molecular predictions , which includes coordinates and type Subsequently, the MolSolver model simultaneously performs two updates: refining the coordinate parameters using a second-order multistep stochastic differential equation solver; and updating the type parameters using a Bayesian flow distribution. This simultaneous update cycle performs N=80 steps in total; the final molecular prediction is... .

Images