A molecular optimization method and system based on protein structure and affinity guidance

Abstract

This application relates to a molecular optimization method and system based on protein structure and affinity guidance. The method includes: inputting structural data of protein pockets and the molecular backbone into a diffusion model; in the forward process, noise is progressively introduced from a Gaussian distribution at each time step, and prior guidance features of the R-group are extracted using a pre-trained prior guidance model, with the mean shift of the atomic coordinates of the R-group calculated and added to the forward process; in the reverse process, sampling from the Gaussian distribution is used as a starting point, and the R-group matching the molecular backbone is generated through progressive denoising at each time step, based on the protein pockets and the molecular backbone; simultaneously, at each time step, a guidance correction term is calculated based on the prior guidance features to guide the correction of the atomic coordinates of the R-group generated at the current time step; after denoising, the molecular backbone and the R-group are assembled into a complete molecule and output. This invention can improve the binding affinity between molecules and protein pockets.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a molecular optimization method and system based on protein structure and affinity guidance. Background Technology

[0002] Diffusion models have achieved state-of-the-art (SOTA) performance in various molecular generation tasks in recent years. Among them, diffusion models based on the three-dimensional structure of protein receptor pockets can achieve scaffold decoration by learning the joint three-dimensional spatial distribution of ligand molecules and protein pockets, targeting specific pockets and existing ligand scaffolds. This type of model has become a research hotspot in structure-based drug design and molecular optimization. A typical example is DiffDec.

[0003] However, existing methods rely solely on the structural information of the protein pocket or ligand molecule backbone during molecule generation, lacking explicit guidance mechanisms for pocket-ligand non-covalent interactions. This results in low binding specificity and insufficient affinity between the generated candidate molecules and the protein pocket, or inappropriate anchor selection during molecular modification, leading to unreasonable or even illegal molecular structures. Therefore, current technologies are insufficient to meet the high requirements for molecular binding activity and specificity in drug design, and further improvement and optimization are urgently needed. Summary of the Invention

[0004] Based on this, the purpose of this invention is to address the above-mentioned technical problems by providing a molecular optimization method and system based on protein structure and affinity, which can effectively improve the binding affinity between the modified molecule and the protein pocket, thereby improving the targeting specificity and structural rationality of the generated molecule.

[0005] To achieve the aforementioned objectives, the first aspect of this application provides a molecular optimization method based on protein structure and affinity guidance, comprising: Input the structural data of the protein pocket and the molecular backbone into the diffusion model; In the forward noise addition process of the diffusion model, the atomic coordinates and atomic types of the R-group are treated as continuous random variables, and noise is gradually introduced from the Gaussian distribution at each time step; At each time step in the forward noise addition process, the pocket-ligand interaction prior guidance features of the R-group group are extracted using a pre-trained prior guidance model, and the mean shift of the atomic coordinates of the R-group group is calculated based on the pocket-ligand interaction prior guidance features and added as a guidance correction term in the forward noise addition process. In the reverse denoising process of the diffusion model, samples are taken from the Gaussian distribution, and the atomic coordinates and atomic types of R-group groups that match the molecular skeleton are generated by progressive denoising at each time step, based on the protein pocket and molecular skeleton. At each time step in the reverse denoising process, the guiding correction term is calculated based on the prior guiding features of the pocket-ligand interaction to guide the correction of the atomic coordinates of the R-group group generated at the current time step; After noise reduction, the molecular backbone and the guided R-group groups are assembled into a complete molecular structure and output.

[0006] Preferably, the diffusion model includes a 6-layer SE(3) equivariant neural network and a 4-layer MLP output module, wherein the 6-layer SE(3) equivariant neural network serves as the feature extraction backbone network and the 4-layer MLP output module is used for noise prediction; The training process of the diffusion model includes: Obtain a dataset of protein pocket-ligand complex structures and filter the dataset using specific criteria, including: Samples with RMSD < 1 Å for ligand-protein docking conformations were retained to ensure conformation quality. Protein sequence identity is limited to <30% to ensure protein diversity in the dataset; For the selected complex samples, the protein substructures within a radius of 10 Å centered on the ligand are extracted as the effective pocket region, and the three-dimensional coordinates and atom types of all atoms in the pocket are preserved. Subsequently, based on LibINVENT's 37 reaction rules, each ligand was decomposed into a molecular backbone and R-group groups, and the processed dataset was divided into a training set and a test set, with the test set samples covering 10 different protein families. The diffusion model is trained using a training set, with the optimization objective being the mean squared error loss function for noise reconstruction. After training, the diffusion model is used to generate and evaluate on the test set.

[0007] Preferably, the pre-trained prior guidance model uses an SE(3) equivariant neural network as the feature extraction backbone network to extract atomic-level feature vectors between the protein pocket and the ligand; based on the atomic-level feature vectors, a cross-attention mechanism is used to establish the interaction feature matrix between the protein pocket and the ligand atoms; finally, the binding affinity value of the complex is predicted by a two-layer MLP network. The prior-guided model is pre-trained using the training set for affinity prediction task training; After training, the interaction feature matrix between protein pockets and ligand atoms output by the prior-guided model is used as the prior-guided feature of the pocket-ligand interaction.

[0008] Preferably, the forward noise addition process employs a signal-to-noise ratio parameterized cosine noise scheduling strategy, wherein the cosine noise scheduling strategy is defined as follows:

[0009]

[0010]

[0011] in, The logarithmic signal-to-noise ratio is used to describe the relative proportion of the signal to the noise, making noise regulation smoother and more stable. The noise intensity coefficient is introduced at time step t during the diffusion process to control the rate at which the original data transitions to the noise distribution. This represents the amount of effective signal retained in the data at time step t, which gradually decreases as the forward noise is added. This represents the noise amplitude at time step t, and... Together they constitute the normalized decomposition of signal and noise; This represents the normalization function.

[0012] Preferably, the molecular backbone and protein pocket are fixed conditions that remain unchanged during forward noise addition and backward noise removal, i.e.:

[0013] in, Indicates a protein pocket; Represents the molecular skeleton; Indicates the R-group; , , These represent the number of atoms in the protein pocket, molecular backbone, and R-group, respectively. Represents the three-dimensional coordinates of an atom; Representing atom type characteristics and using One-hot vector encoding; This represents a Gaussian distribution.

[0014] Preferably, the guiding correction item is:

[0015] in, It is an MLP network; Indicates at time step The cumulative mean offset of atomic coordinates over time; The predefined retention factor is expressed as follows:

[0016] After adding the aforementioned guidance correction term to the forward noise addition process, the noise addition process with fusion guidance is obtained, defined as follows:

[0017]

[0018] in:

[0019] .

[0020] Preferably, in the reverse denoising process, the goal is to learn how to reverse the forward denoising process, and the corresponding posterior distribution is represented as:

[0021] Among them, the mean term The definition of is:

[0022] variance term The definition of is: .

[0023] Preferably, a denoiser is used to predict the injected Gaussian noise. :

[0024] in, This indicates that the MLP-based neural network is trained as a denoiser; The training objective is to minimize the real noise. With prediction noise Mean square error between:

[0025] The predicted noise can be used to obtain the true R-group groups. Estimate :

[0026] Due to the truth It is unknown at each time step, so the previous time step is used. The estimated value To replace the real Calculate the prior guiding features of the pocket-ligand interaction:

[0027] in, This is a pre-trained prior guidance model. These are the prior guiding features of pocket-ligand interactions in protein pockets, molecular backbones, and R-group groups, respectively. Using the prior guidance features of pocket-ligand interactions, the mean shift of the R-group groups, i.e., the guidance correction term, is calculated:

[0028] In the next step In the middle, use the current time step The obtained estimated value To calculate At the initial time step Interaction features Initialize to null; Will , , This information is further input into the denoiser to assist in prediction. :

[0029] Therefore, the transfer nucleus of back-diffusion is defined as:

[0030] in: .

[0031] To achieve the purpose of the invention, a second aspect of this application provides a molecular optimization system based on protein structure and affinity, applying the molecular optimization method based on protein structure and affinity described above. The system includes: The input module is used to acquire structural data of protein pockets and molecular backbones and input them into the diffusion model module; The diffusion model module includes a diffusion model for generating R-group groups by progressively adding and denoising, based on the structural data provided by the input module. The prior guidance module includes a pre-trained prior guidance model for providing pocket-ligand interaction prior guidance features at each noisy and denoising time step of the diffusion model module; The guided correction calculation module is used to calculate the mean shift of atomic coordinates of R-group groups based on the prior guided features of pocket-ligand interactions and use it as a guided correction term; The output module is used to assemble the molecular backbone and the guided correction R-group groups into a complete molecular structure and output it after the noise reduction is completed.

[0032] Preferably, in the forward noise addition process of the diffusion model, the atomic coordinates and atomic types of the R-group are treated as continuous random variables, and noise is gradually introduced from the Gaussian distribution at each time step; At each time step in the forward noise addition process, the pocket-ligand interaction prior guidance features of the R-group group are extracted using a pre-trained prior guidance model, and the mean shift of the atomic coordinates of the R-group group is calculated based on the pocket-ligand interaction prior guidance features and added as a guidance correction term in the forward noise addition process. In the reverse denoising process of the diffusion model, samples are taken from the Gaussian distribution, and the atomic coordinates and atomic types of R-group groups that match the molecular skeleton are generated by progressive denoising at each time step, based on the protein pocket and molecular skeleton. At each time step in the reverse denoising process, the guiding correction term is calculated based on the prior guiding features of the pocket-ligand interaction to guide the correction of the atomic coordinates of the R-group groups generated at the current time step.

[0033] Compared with the prior art, the beneficial effects of this invention are: This invention introduces a pre-trained prior guidance model for interactions, providing pocket-ligand interaction prior guidance features at each time step during forward and backward diffusion. This enables more precise explicit guidance of the binding site space during the R-group generation stage, effectively improving the binding affinity between the modified molecule and the protein pocket, thereby enhancing the targeting specificity and structural rationality of the generated molecule. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating the steps of a molecular optimization method based on protein structure and affinity in an embodiment of this application. Figure 2 This is a schematic diagram of the process for generating the complete molecular structure in the embodiments of this application; Figure 3 This is a schematic diagram of a molecular optimization system based on protein structure and affinity guided by an embodiment of this application. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. The following embodiments are used to illustrate the invention but are not intended to limit its scope.

[0036] Example 1 Example 1 of this application provides a molecular optimization method based on protein structure and affinity guidance, such as... Figure 1 As shown, it includes the following steps: S1: Input the structural data of the protein pocket and the molecular backbone into the diffusion model; S2: In the forward noise addition process of the diffusion model, the atomic coordinates and atomic types of the R-group are treated as continuous random variables, and noise is gradually introduced from the Gaussian distribution at each time step; S3: At each time step in the forward noise addition process, the pocket-ligand interaction prior guidance features of the R-group group are extracted using a pre-trained prior guidance model, and the atomic coordinate mean offset of the R-group group is calculated based on the pocket-ligand interaction prior guidance features and added as a guidance correction term in the forward noise addition process. S4: In the reverse denoising process of the diffusion model, samples are taken from the Gaussian distribution, and the atomic coordinates and atomic types of R-group groups that match the molecular skeleton are generated by progressive denoising at each time step, based on the protein pocket and molecular skeleton. S5: At each time step in the reverse denoising process, the guiding correction term is calculated based on the prior guiding features of the pocket-ligand interaction to guide the correction of the atomic coordinates of the R-group group generated at the current time step; S6: After noise reduction, the molecular skeleton and the guided modified R-group groups are assembled into a complete molecular structure and output.

[0037] In this application, the following symbols are used for definition: Let the protein pocket be... The compound molecule is ,in Represents the molecular skeleton. This indicates the R-group. , , These represent the number of atoms in the protein pocket, molecular backbone, and R-group, respectively. For atomic three-dimensional coordinates, For atomic type features and using One-hot vector encoding. Define a pre-trained prior-guided model. When given a protein pocket Molecular skeleton and R-group groups At that time, prior guidance model Output its corresponding pocket-ligand interaction prior guiding features and represent them as follows: .

[0038] like Figure 2 As shown in this embodiment, the process of generating the complete molecular structure includes: inputting the molecular backbone and protein pockets, then the diffusion model, under the guidance of a priori model and MLP, performs T loading and denoising operations step by step according to the time step, finally generating the R-group. The generated R-group is assembled with the molecular backbone to form a complete molecule and added to the candidate molecule set.

[0039] (1) The forward noise addition process, including S2-S3, is as follows: In this application, the forward diffusion process (i.e., the forward noise addition process) of the diffusion model employs a cosine noise scheduling scheme, which is used to allocate noise at each time step. Define a set of noise Unlike directly using a fixed noise scheduling method, this embodiment redefines the noise scheduling scheme using the signal-to-noise ratio (SNR), as follows:

[0040]

[0041]

[0042] in, The logarithmic signal-to-noise ratio is used to describe the relative proportion of the signal to the noise, making noise regulation smoother and more stable. The noise intensity coefficient is introduced at time step t during the diffusion process to control the rate at which the original data transitions to the noise distribution. This represents the amount of effective signal retained in the data at time step t, which gradually decreases as the forward noise is added. This represents the noise amplitude at time step t, and... Together they constitute the normalized decomposition of signal and noise; This represents the normalization function.

[0043] During the forward diffusion process, the atom coordinates and atom types of the R-group are treated as continuous random variables, and at each time step... From Gaussian distribution Noise is gradually introduced. Molecular backbone. With protein pocket As a fixed condition, it remains unchanged in both the forward and reverse processes, that is:

[0044] Then, prior guiding features of pocket-ligand interactions of R-group groups were extracted from the prior guiding model. At this point, a learning neural network is introduced. The atom coordinate mean offset was used as a guiding correction to calculate the R-group based on the prior guided features of pocket-ligand interactions.

[0045] in, It is an MLP. Indicates at time step Cumulative mean shift. Predefined retention coefficient. The expression is:

[0046] After incorporating the aforementioned mean offset term into the forward diffusion process, we can obtain the fusion-guided noise addition process, defined as follows: ,

[0047] in:

[0048]

[0049] This forward noise-addition process combines signal-to-noise ratio parameterization with a priori guided model, enabling the diffusion model to dynamically control the R-group atomic coordinates based on protein pocket and molecular backbone information during generation. This improves the structural rationality and affinity of the generated molecules while maintaining generation stability.

[0050] (2) The reverse denoising process, including S4-S5, is as follows: In the reverse denoising process, the goal is to learn how to reverse the forward denoising process. The corresponding posterior distribution can be expressed as:

[0051] The mean and variance terms are defined as follows:

[0052]

[0053] However, due to the time step It is impossible to directly obtain the actual R-group. Therefore, a neural network model based on MLP is trained. To approximate this value, the model is referred to as the denoiser in this application. Specifically, the denoiser is used to predict the injected Gaussian noise. :

[0054] The training objective of the model is to minimize the true noise. With prediction noise Mean square error between:

[0055] The predicted noise can be used to obtain the true R-group groups. Estimate:

[0056] At the same time, due to the truth It is unknown at each time step, therefore the present invention utilizes the previous time step. The estimated value To replace the real Calculate prior guiding features:

[0057] in, For the pre-trained prior-guided model, the output will be the prior-guided features of the protein pocket, molecular backbone, and R-group. Using these features, the mean shift of the R-group can be calculated.

[0058] Similarly, in the next time step In the middle, the current time step will be used. The obtained estimated value To calculate At the initial time step Interaction features Initialized to null (None), sampled from a Gaussian distribution. .

[0059] In order to maximize the use of protein-ligand prior features in the prior-guided model, this invention also uses prior features. , , As a condition, the output is further fed into the denoiser to assist in prediction:

[0060] Therefore, the transfer nucleus of back-diffusion is defined as:

[0061] in:

[0062] By introducing a priori guidance mechanism, the diffusion model can fully utilize the spatial interaction features between proteins and ligands in the reverse process (denoising generation), effectively guiding the model to generate potential molecular structures from Gaussian noise.

[0063] Example 2 Example 2 of this application, based on Example 1, provides a molecular optimization system guided by protein structure and affinity, such as... Figure 3 As shown, it includes an input module 201, a diffusion model module 202, a priori guidance module 203, a guidance correction calculation module 204, and an output module 205; wherein: Input module 201 is used to acquire structural data of protein pockets and molecular backbones and input them into diffusion model module 202; Diffusion model module 202 includes a diffusion model for generating R-group groups by progressive noise addition and progressive noise reduction based on the structural data provided by the input module. The prior guidance module 203 includes a pre-trained prior guidance model for providing pocket-ligand interaction prior guidance features at each noisy and denoising time step of the diffusion model module; Guided correction calculation module 204 is used to calculate the mean shift of atomic coordinates of R-group groups based on the prior guided features of pocket-ligand interactions and use it as a guided correction term; The output module 205 is used to assemble the molecular backbone and the guided modified R-group groups into a complete molecular structure and output it after the noise is removed.

[0064] In the forward noise addition process of the diffusion model in diffusion model module 202, the atomic coordinates and atomic types of R-group groups are treated as continuous random variables, and noise is gradually introduced from the Gaussian distribution at each time step; At each time step in the forward noise addition process, the pocket-ligand interaction prior guidance features of the R-group group are extracted using a pre-trained prior guidance model, and the mean shift of the atomic coordinates of the R-group group is calculated based on the pocket-ligand interaction prior guidance features and added as a guidance correction term in the forward noise addition process. Given the protein pocket and molecular backbone, the atomic coordinates and atom types of R-group groups matching the molecular backbone are generated by progressively denoising at each time step. At each time step in the reverse denoising process, the guiding correction term is calculated based on the prior guiding features of the pocket-ligand interaction to guide the correction of the atomic coordinates of the R-group groups generated at the current time step.

[0065] Example 3 This Example 3 is based on Examples 1 and 2, and involves an experiment to generate complete molecules, as detailed below: (1) Hardware and software environment The operating system uses Linux, and the main software packages it relies on include Python, PyTorch, RDKit, QVina, and LiBINVENT, which are used for data processing, model training, and inference deployment.

[0066] The hardware environment consists of an NVIDIA RTX 4090 GPU (24GB VRAM) used for model training and inference.

[0067] (2) Dataset processing Example 3 uses the CrossDocked dataset, which contains approximately 2.25 × 10⁻⁶ cells. 7 To ensure sample quality and structural diversity, the specific screening criteria for each protein-ligand complex structure are as follows: 1. Retain samples with RMSD < 1 Å for ligand-protein docking conformation to ensure conformation quality; 2. Limit protein sequence identity to <30% to ensure protein diversity in the dataset.

[0068] For the selected complex samples, PyMOL was used to extract the protein substructures within a radius of 10 Å centered on the ligand as the effective pocket region, preserving the three-dimensional coordinates and atom types of all atoms within the pocket.

[0069] Subsequently, based on LibinVENT's 37 reaction rules (such as amide bond and ester bond splitting rules), each ligand was decomposed into a molecular backbone and R-group groups. The processed data was divided into training and testing sets, with the testing set covering 10 different protein families (such as kinases and proteases) to ensure the comprehensiveness of model generalization validation.

[0070] (3) Pre-trained prior guidance model To achieve guided generation based on protein-ligand interactions, this example uses a pre-trained IPNet model as a priori guiding model.

[0071] IPNet uses an SE (3) equivariant neural network as the backbone network for feature extraction to ensure the model’s equivariance to translation and rotation in the three-dimensional space of molecules; then it extracts the atomic-level feature vectors of protein pockets and ligands (including molecular skeleton and R-group groups); it uses a cross-attention mechanism to establish the interaction feature matrix between protein and ligand atoms; finally, it uses a two-layer MLP network to predict the binding affinity of the complex.

[0072] After training through this affinity prediction task, the interaction feature matrix output by IPNet is used as the prior guiding feature of the interaction. In the diffusion model of this invention, it is used to guide the generation process of R-group atoms.

[0073] (4) Diffusion model The diffusion model in this embodiment is named DiffDiG (Diffusion model with Interaction Guidance).

[0074] The overall structure of the model includes: A 6-layer SE(3) equivariant neural network is used as the backbone network for feature extraction; A 4-layer MLP output module is used for noise prediction; The time step is set to T = 1000, and a cosine noise scheduling strategy is adopted.

[0075] The model is trained using the training set processed in (2), and the optimization objective is the mean squared error loss function (MSE) for noise reconstruction.

[0076] After training, generation and evaluation are performed on the test set. The specific process is as follows: 100 candidate R-group groups are generated for each sample; invalid molecules (such as those with unreasonable chemical bonds or abnormal charges) are filtered out using RDKit; molecules that conform to medicinal chemistry rules are retained, and the following metrics are evaluated: Validity: The percentage of molecules that RDKit successfully resolved and preserved the original backbone structure, calculated as: number of valid molecules / total number of generated molecules × 100%; Uniqueness: The percentage of molecules after removing duplicates, calculated as: number of unique molecules / number of effective molecules × 100%; Vina Score: QVina is used to perform molecular docking between the generated molecule and the target protein pocket, and outputs the binding affinity (the lower the value, the stronger the affinity). High Affinity Ratio: The proportion of generated molecules with a Vina Score superior to the complete molecules corresponding to the original R groups in the test set; Number of non-covalent interactions: The total number of non-covalent interactions (such as hydrogen bonds and hydrophobic interactions) between the generated molecule and the pocket residues is calculated using Schrödinger, and the average value of the samples is taken.

[0077] (5) Analysis of evaluation results Experimental results show that the DiffDiG model proposed in this application has the following significant advantages over the baseline diffusion model without prior guidance: Enhanced affinity: The average Vina Score in this embodiment is -8.19 kcal / mol, which is better than the baseline model's -8.11 kcal / mol; the High Affinity ratio is increased by 1.7 percentage points, indicating that prior guidance can effectively enhance the pocket binding strength of the generated molecules.

[0078] Enhanced molecular diversity and effectiveness: The uniqueness of this embodiment reached 67.3% (higher than the baseline model's 53.3%), while the effectiveness remained at a high level of 90.1%. This demonstrates that introducing prior guidance not only improves binding activity but also does not sacrifice molecular diversity and chemical rationality.

[0079] Enhanced molecule-pocket interactions: The R-group generated in this embodiment formed an average of 4.5 non-covalent interactions with protein residues, which is significantly higher than the 3.2 interactions in the baseline model. This verifies that the prior guidance provided by IPNet can effectively guide molecules to form more stable binding modes during the generation process.

[0080] In summary, this invention provides a molecular optimization method and system based on protein structure and affinity guidance. By introducing a pre-trained prior guidance model, the pocket-ligand interaction prior is used as a guiding feature at each time step during the diffusion process, thereby achieving more refined guidance of the binding site space during the R-group generation stage.

Claims

1. A molecular optimization method based on protein structure and affinity guidance, characterized in that, Includes the following steps: Input the structural data of the protein pocket and the molecular backbone into the diffusion model; In the forward noise addition process of the diffusion model, the atomic coordinates and atomic types of the R-group are treated as continuous random variables, and noise is gradually introduced from the Gaussian distribution at each time step; At each time step in the forward noise addition process, the pocket-ligand interaction prior guidance features of the R-group group are extracted using a pre-trained prior guidance model, and the mean shift of the atomic coordinates of the R-group group is calculated based on the pocket-ligand interaction prior guidance features and added as a guidance correction term in the forward noise addition process. In the reverse denoising process of the diffusion model, samples are taken from the Gaussian distribution, and the atomic coordinates and atomic types of R-group groups that match the molecular skeleton are generated by progressive denoising at each time step, based on the protein pocket and molecular skeleton. At each time step in the reverse denoising process, the guiding correction term is calculated based on the prior guiding features of the pocket-ligand interaction to guide the correction of the atomic coordinates of the R-group group generated at the current time step; After noise reduction, the molecular backbone and the guided R-group groups are assembled into a complete molecular structure and output.

2. The method according to claim 1, characterized in that, The diffusion model includes a 6-layer SE(3) equivariant neural network and a 4-layer MLP output module. The 6-layer SE(3) equivariant neural network serves as the feature extraction backbone network, and the 4-layer MLP output module is used for noise prediction. The training process of the diffusion model includes: Obtain a dataset of protein pocket-ligand complex structures and filter the dataset using specific criteria, including: Samples with RMSD < 1 Å for ligand-protein docking conformations were retained to ensure conformation quality. Protein sequence identity is limited to <30% to ensure protein diversity in the dataset; For the selected complex samples, the protein substructures within a radius of 10 Å centered on the ligand are extracted as the effective pocket region, and the three-dimensional coordinates and atom types of all atoms in the pocket are preserved. Subsequently, based on LibINVENT's 37 reaction rules, each ligand was decomposed into a molecular backbone and R-group groups, and the processed dataset was divided into a training set and a test set, with the test set samples covering 10 different protein families. The diffusion model is trained using a training set, with the optimization objective being the mean squared error loss function for noise reconstruction. After training, the diffusion model is used to generate and evaluate on the test set.

3. The method according to claim 2, characterized in that, The pre-trained prior guidance model uses an SE(3) equivariant neural network as the feature extraction backbone network to extract atomic-level feature vectors between protein pockets and ligands; based on the atomic-level feature vectors, a cross-attention mechanism is used to establish the interaction feature matrix between protein pockets and ligand atoms; finally, the binding affinity value of the complex is predicted by a two-layer MLP network. The prior-guided model is pre-trained using the training set for affinity prediction task training; After training, the interaction feature matrix between protein pockets and ligand atoms output by the prior-guided model is used as the prior-guided feature of the pocket-ligand interaction.

4. The method according to claim 1, characterized in that, The forward noise addition process employs a signal-to-noise ratio parameterized cosine noise scheduling strategy, which is defined as follows: in, The logarithmic signal-to-noise ratio is used to describe the relative proportion of the signal to the noise, making noise regulation smoother and more stable. The noise intensity coefficient is introduced at time step t during the diffusion process to control the rate at which the original data transitions to the noise distribution. This represents the amount of effective signal retained in the data at time step t, which gradually decreases as the forward noise is added. This represents the noise amplitude at time step t, and... Together they constitute the normalized decomposition of signal and noise; This represents the normalization function.

5. The method according to claim 4, characterized in that, The molecular backbone and protein pocket, as fixed conditions, remain unchanged during forward noise addition and backward noise removal, i.e.: in, Indicates a protein pocket; Represents the molecular skeleton; Indicates the R-group; , , These represent the number of atoms in the protein pocket, molecular backbone, and R-group, respectively. Represents the three-dimensional coordinates of an atom; Representing atom type characteristics and using One-hot vector encoding; This represents a Gaussian distribution.

6. The method according to claim 5, characterized in that, The guidance correction item is: in, It is an MLP network; Indicates at time step The cumulative mean offset of atomic coordinates over time; The predefined retention factor is expressed as follows: After adding the aforementioned guidance correction term to the forward noise addition process, the noise addition process with fusion guidance is obtained, defined as follows: in: 。 7. The method according to claim 5, characterized in that, In the reverse denoising process, the goal is to learn how to reverse the forward denoising process, and the corresponding posterior distribution is represented as: Among them, the mean term The definition of is: variance term The definition of is: 。 8. The method according to claim 7, characterized in that, Use a denoiser to predict the injected Gaussian noise. : in, This indicates that the MLP-based neural network is trained as a denoiser; The training objective is to minimize the real noise. With prediction noise Mean square error between: The predicted noise can be used to obtain the true R-group groups. Estimate : Due to the truth It is unknown at each time step, so the previous time step is used. The estimated value To replace the real Calculate the prior guiding features of the pocket-ligand interaction: in, This is a pre-trained prior guidance model. These are the prior guiding features of pocket-ligand interactions in protein pockets, molecular backbones, and R-group groups, respectively. Using the prior guidance features of pocket-ligand interactions, the mean shift of the R-group groups, i.e., the guidance correction term, is calculated: In the next step In the middle, use the current time step The obtained estimated value To calculate At the initial time step Interaction features Initialize to null; Will , , This information is further input into the denoiser to assist in prediction. : Therefore, the transfer nucleus of back-diffusion is defined as: in: 。 9. A molecular optimization system based on protein structure and affinity, used to implement the molecular optimization method based on protein structure and affinity as described in any one of claims 1-8, characterized in that, The system includes: The input module is used to acquire structural data of protein pockets and molecular backbones and input them into the diffusion model module; The diffusion model module includes a diffusion model for generating R-group groups by progressively adding and denoising, based on the structural data provided by the input module. The prior guidance module includes a pre-trained prior guidance model for providing pocket-ligand interaction prior guidance features at each noisy and denoising time step of the diffusion model module; The guided correction calculation module is used to calculate the mean shift of atomic coordinates of R-group groups based on the prior guided features of pocket-ligand interactions and use it as a guided correction term; The output module is used to assemble the molecular backbone and the guided correction R-group groups into a complete molecular structure and output it after the noise reduction is completed.

10. The system according to claim 9, characterized in that, In the forward noise addition process of the diffusion model, the atomic coordinates and atomic types of the R-group are treated as continuous random variables, and noise is gradually introduced from the Gaussian distribution at each time step; At each time step in the forward noise addition process, the pocket-ligand interaction prior guidance features of the R-group group are extracted using a pre-trained prior guidance model, and the mean shift of the atomic coordinates of the R-group group is calculated based on the pocket-ligand interaction prior guidance features and added as a guidance correction term in the forward noise addition process. In the reverse denoising process of the diffusion model, samples are taken from the Gaussian distribution, and the atomic coordinates and atomic types of R-group groups that match the molecular skeleton are generated by progressive denoising at each time step, based on the protein pocket and molecular skeleton. At each time step in the reverse denoising process, the guiding correction term is calculated based on the prior guiding features of the pocket-ligand interaction to guide the correction of the atomic coordinates of the R-group groups generated at the current time step.

Images