A molecular optimization method, system, device, and storage medium based on neural networks.

Abstract

This invention relates to the field of biomedical engineering technology, specifically to a molecular optimization method, system, device, and storage medium based on neural networks. The method includes: acquiring contextual information, raw noise data, and affinity information of a target protein; inputting these information into a diffusion model; processing the diffusion model according to preset rules to select the drug molecule structure corresponding to the affinity-sensitive R-chain; and outputting the target drug molecule structure. This invention utilizes the affinity-sensitive R-chain and the original scaffold of the drug molecule to generate the target drug molecule structure, improving the molecular docking accuracy between the original scaffold and the R-chain of the drug molecule.

Description

Technical Field

[0001] This invention relates to the field of biomedical engineering technology, and in particular to a molecular optimization method, system, device and storage medium based on neural networks. Background Technology

[0002] In the field of drug discovery, computer-aided drug design (CADD) and artificial intelligence-driven drug design (AIDD) have been widely applied to molecular optimization.

[0003] Molecular optimization is a crucial step in drug discovery and materials design. It refers to the process of improving the structure of candidate molecules through chemical modification or computational simulation to enhance their biological activity, druggability (such as ADMET properties), or specific functions (such as catalytic performance). Its core objective is to optimize molecular properties to meet multi-dimensional requirements while maintaining the integrity of the pharmacophore.

[0004] Currently, commonly used molecular optimization techniques mainly include ligand-based drug design (LBDD) and structure-based drug design (SBDD). Specifically:

[0005] Ligand-based drug design methods include LibinVENT (which uses chemical reaction templates and RNNs to generate R chains) and DeepScaffold (which applies graph variational autoencoders to optimize scaffold topology), generating molecules through pharmacophore modeling or structure-activity relationship (SAR) analysis.

[0006] Structure-based drug design methods include DiffDec (based on the E(3)-isovariant diffusion model), which can model the joint distribution of ligands and targets to achieve end-to-end molecule generation.

[0007] Current techniques mostly employ standard diffusion model conditions for R-chain generation within scaffolds and pockets, which falls short in terms of generation quality, affinity enhancement, and structural realism. Specific drawbacks are as follows:

[0008] 1. Inability to model affinity-driven optimization: Existing technologies (such as LibinVENT, DeepScaffold, and DiffDec) do not use ligand-target binding affinity as a conditional input when generating the R chain, thus failing to capture the quantitative relationship between affinity and structure. This results in molecules that may have a reasonable structure but low binding efficiency.

[0009] 2. Insufficient geometric feature compatibility: Existing technologies (such as Pocket2Mol and FLAG) employ de novo generation strategies, which often lead to molecular geometric distortion (e.g., large bond angle deviations) and poor synthetic feasibility. Furthermore, spatial conflicts easily occur when molecules bind to the target, resulting in poor Vina scores. These shortcomings stem from the fact that existing technologies have not introduced affinity guidance mechanisms and local coordinate system alignment strategies, which can be effectively addressed through the technical means of this patent.

[0010] Current techniques mostly employ standard diffusion model conditions for R-chain generation within scaffolds and pockets, but neglect explicit guidance based on affinity and lack dedicated affinity-driven attention mechanisms. This results in shortcomings in terms of generation quality, affinity enhancement, and structural realism. Summary of the Invention

[0011] This application aims to at least solve the technical problems existing in the prior art, and to provide a molecular optimization method, system, device and storage medium based on neural networks.

[0012] In a first aspect, the present invention provides a molecular optimization method based on a neural network, the method comprising:

[0013] Obtain contextual information, raw noise data, and affinity information of the target protein;

[0014] The context information, raw noise data, and affinity information of the target protein are input into the diffusion model. The diffusion model processes the context information, raw noise data, and affinity information of the target protein according to preset rules, filters out the drug molecule structure corresponding to the affinity-sensitive R chain, and outputs the target drug molecule structure.

[0015] The diffusion model's processing steps for the target protein's contextual information, raw noise data, and affinity information include:

[0016] Determine the spatial relationship between the target pocket and the drug molecule based on the contextual information of the target protein;

[0017] Based on the original noise data and the spatial relationship between the target pocket and the drug molecule, noise is added to the R chain of the drug molecule to obtain the noisy molecular structure;

[0018] Based on affinity information, noise in the denoised molecular structure is removed iteratively, and the affinity-sensitive R chain is recovered from the noise to obtain the target drug molecular structure.

[0019] By employing the above technical solution, the spatial relationship between the target pocket of the target protein and the drug molecule is analyzed in real time using the context information of the target protein, enabling dynamic matching between the conformational adjustment of the R chain and the protein binding site. Noise is added to the R chain of the drug molecule based on the original noise data to generate random perturbations. During iterative denoising, the noise in the denoised molecular structure is adjusted based on affinity information, and the noise-adjusted R chain is then denoised to obtain an affinity-sensitive R chain. The target drug molecule structure is generated using the affinity-sensitive R chain and the original scaffold of the drug molecule. Combining the R chain obtained after affinity-sensitive screening significantly improves the molecular docking accuracy between the original scaffold of the drug molecule and the R chain, thereby increasing the molecular generation rate of the target drug molecule.

[0020] Optionally, the contextual information of the target protein includes the pocket atoms of the target protein and the drug molecule scaffold; the contextual information of the target protein is used to extract features by using an isomorphic graph neural network to obtain the spatial relationship feature representation of the target pocket and the drug molecule, and the drug molecule structure is generated by gradually denoising based on affinity information and the spatial relationship feature representation of the target pocket and the drug molecule.

[0021] By employing the above technical solution, the specific content of the contextual information of the target protein was clarified. Furthermore, the spatial relationship features between the target pocket and the drug molecule obtained by the isomorphic graph neural network can automatically adapt to the rotation or translation of the protein pocket, reducing conformational bias caused by neglecting molecular symmetry.

[0022] Optionally, determining the spatial relationship between the target pocket and the drug molecule based on the contextual information of the target protein includes:

[0023] The active site information of the target protein is determined based on the target pocket. The active site information of the target protein represents the region that can bind to the target drug molecule.

[0024] The spatial relationship between the target pocket and the drug molecule is represented by information on the active sites of the drug molecule scaffold and the target protein.

[0025] By adopting the above technical solution, the extraction steps for the spatial relationship between the target pocket and drug molecules were clarified.

[0026] Optionally, the step of adding noise to the R chain of the drug molecule based on the original noise data to obtain the noisy molecular structure includes:

[0027] For each training iteration, the initial R chain on the drug molecule scaffold is denoted as... ,make Indicates t ( The R-chain after adding noise at the next time step;

[0028] The formula for adding noise to the R chain of a drug molecule in a diffusion model is as follows:

[0029] ;

[0030] in, The initial R chain on the drug molecule scaffold , This represents the noise parameter added to the drug molecule scaffold at time step t. , This is the variance scheduling parameter at time step i, used to control the amount of noise added; variance scheduling parameter at time step i Acquired from raw noise data; It is random standard Gaussian noise; This represents the R-chain with noise added in step t;

[0031] The noisy molecular structure was determined based on pocket atoms, drug molecule scaffolds, and the t-step noisy R-chain.

[0032] By adopting the above technical solution, the specific method of adding noise to the R chain of the drug molecule based on the original noise data is clarified. The noise value is added to the R chain, and the pocket atoms of the target protein and the drug molecule scaffold remain fixed without noise. This can preserve the structural information of the target pocket and the drug molecule scaffold and reduce the possibility of geometric distortion of the target drug molecule generated subsequently.

[0033] Optionally, the step of iteratively removing noise from the noisy molecular structure based on affinity information to recover the affinity-sensitive R chain from the noise and obtain the target drug molecule structure includes:

[0034] Using an isomorphic graph neural network, node feature noise and coordinate noise of the R chain in a noisy molecular structure are predicted;

[0035] The affinity characteristics of the target pocket and R chain were determined based on affinity information;

[0036] The node feature noise and affinity feature of the R chain are fused through a cross-attention mechanism to obtain the first fused feature;

[0037] The coordinate noise and affinity features of the R-chain are fused using a cross-attention mechanism to obtain a second fused feature;

[0038] The adjusted node feature noise is obtained based on the first fusion feature, and the adjusted coordinate noise is obtained based on the second fusion feature;

[0039] By iteratively removing the adjusted node feature noise and adjusted coordinate noise corresponding to the molecular structure R chain, the molecular structure R chain is reconstructed, and the reconstructed R chain is obtained.

[0040] The target drug molecule structure is generated based on the drug molecule scaffold and the reconstructed R chain.

[0041] By adopting the above technical solution, the node feature noise and coordinate noise of the R chain in the noisy molecular structure are predicted by the isomorphic graph neural network, and the node feature noise and coordinate noise of the R chain are adjusted by the affinity cross-attention mechanism, so as to gradually restore the structurally reasonable R chain and improve the Vina score and structural authenticity.

[0042] Optionally, before removing the adjusted node feature noise and adjusted coordinate noise of the molecular structure R chain, the method further includes:

[0043] The current length of the molecular structure R chain is identified and compared with the reference R chain length. If the current length of the R chain is less than the reference R chain length, pseudo-atoms are used to fill the R chain so that the length of the molecular structure R chain is the reference R chain length.

[0044] By adopting the above technical solution and standardizing the R chain structure through pseudo-atom filling, it is helpful to unify the chain length of different samples and eliminate the influence of length differences on molecular docking analysis.

[0045] Optionally, the generation of the target drug molecule structure based on the drug molecule scaffold and the reconstructed R chain includes:

[0046] Analyze the reconstructed R chain and determine whether there are pseudo atoms in the reconstructed R chain:

[0047] If so, remove the pseudo atoms in the reconstructed R chain to obtain the processed R chain, and generate the target drug molecule structure based on the drug molecule scaffold and the processed R chain.

[0048] If not, the target drug molecule structure is generated based on the drug molecule scaffold and the reconstructed R chain.

[0049] Secondly, the present invention provides a molecular optimization system based on a neural network, the system comprising:

[0050] The acquisition module is used to acquire contextual information, raw noise data, and affinity information of the target protein;

[0051] The processing module is used to input the context information, raw noise data and affinity information of the target protein into the diffusion model. The diffusion model processes the context information, raw noise data and affinity information of the target protein according to preset rules to screen out the drug molecule structure corresponding to the affinity-sensitive R chain and output the target drug molecule structure.

[0052] The diffusion model's processing steps for the target protein's contextual information, raw noise data, and affinity information include:

[0053] Determine the spatial relationship between the target pocket and the drug molecule based on the contextual information of the target protein;

[0054] Based on the original noise data and the spatial relationship between the target pocket and the drug molecule, noise is added to the R chain of the drug molecule to obtain the noisy molecular structure;

[0055] Based on affinity information, noise in the denoised molecular structure is removed iteratively, and the affinity-sensitive R chain is recovered from the noise to obtain the target drug molecular structure.

[0056] Thirdly, the present invention provides an electronic device, the electronic device comprising:

[0057] At least one processor; and,

[0058] A memory communicatively connected to the at least one processor; wherein,

[0059] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the neural network-based molecular optimization method described above.

[0060] Fourthly, the present invention also provides a computer-readable storage medium storing at least one computer program, which is executed by a processor in an electronic device to implement the aforementioned neural network-based molecular optimization method.

[0061] In summary, this application includes the following beneficial technical effects:

[0062] By analyzing the spatial relationship between the target protein's target pocket and the drug molecule in real time using the target protein's contextual information, the conformational adjustment of the R-chain is dynamically matched with the protein binding site. Random perturbations are generated by adding noise to the R-chain of the drug molecule based on the original noise data. During iterative denoising, the noise in the denoised molecular structure is adjusted based on affinity information, and the noise-adjusted R-chain is then denoised to obtain an affinity-sensitive R-chain. The target drug molecule structure is generated using the affinity-sensitive R-chain and the original scaffold of the drug molecule. Combining the R-chain obtained after affinity-sensitive screening significantly improves the molecular docking accuracy between the original scaffold and the R-chain of the drug molecule, thereby increasing the molecular generation rate of the target drug molecule. Attached Figure Description

[0063] Figure 1This is a schematic flowchart of a neural network-based molecular optimization method provided in an embodiment of the present invention.

[0064] Figure 2 This is a flowchart illustrating the processing of contextual information, raw noise data, and affinity information by an affinity-driven diffusion model according to an embodiment of the present invention.

[0065] Figure 3 This is a schematic diagram showing the drug molecule optimization results of the diffusion model of this application at three therapeutic targets: dopamine transporter, SARS-CoV-2 main protease, and vascular endothelial growth factor receptor.

[0066] Figure 4 This is a schematic diagram of an electronic device for implementing the neural network-based molecular optimization method according to an embodiment of the present invention.

[0067] Reference numerals: 10, processor; 11, memory; 12, communication bus; 13, communication interface.

[0068] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0069] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0070] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0071] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0072] Reference Figure 1The diagram shown is a flowchart illustrating a neural network-based molecular optimization method according to an embodiment of the present invention. In this embodiment, the neural network-based molecular optimization method includes:

[0073] S1. Obtain contextual information, raw noise data, and affinity information for the target protein.

[0074] S2. Input the context information, raw noise data and affinity information of the target protein into the diffusion model. The diffusion model processes the context information, raw noise data and affinity information of the target protein according to preset rules to screen out the drug molecule structure corresponding to the affinity-sensitive R chain and output the target drug molecule structure.

[0075] Diffusion models are neural network models primarily used for generative AI tasks such as image generation, audio generation, and molecular design. Specifically, the contextual information of the target protein includes the target protein's pocket atoms and the drug molecule scaffold.

[0076] Let the targeting pocket structure of the drug molecule and the targeting protein be represented by an atomic-level diagram, denoted as P for the pocket atom and S for the scaffold atom corresponding to the drug molecule scaffold. The pocket atom P and the scaffold atom S together constitute the context information c. The objective of this application is to generate the R chain of the target drug molecule from the context information c and the affinity A (affinity information), with the coordinates of the R chain anchored to the scaffold anchor atoms.

[0077] The contextual information of the target protein can be obtained from existing datasets or constructed by researchers; this application does not impose any restrictions. In a preferred embodiment of this application, affinity information represents the affinity between the pocket atoms of the target protein and the drug molecule scaffold, i.e., the affinity between the reconstructed R chain and the pocket atoms. The affinity information is obtained by collecting protein-ligand complexes and performing SAR analysis on the protein-ligand complexes. SAR stands for Structure-Activity Relationship, which in protein-ligand complex analysis refers to the structure-activity relationship to obtain affinity information. In this embodiment, a large dataset from BindingNet is used for the protein-ligand complexes. This dataset contains 75,348 entries, each containing: protein pocket, backbone, R chain, and corresponding ligand-target binding affinity value, thus obtaining affinity information.

[0078] In this embodiment, the standardized affinity value is denoted as pAffi, and high / low affinity is classified with a threshold of 8.0. Samples are generated (the samples contain pocket atoms, scaffold atoms, R chains, and affinity tags), and the diffusion model is trained using the samples to optimize the model parameters. The trained diffusion model can directly process the original noise data, affinity information, and contextual information of the target protein.

[0079] Reference Figure 2 The diffusion model's processing steps for the target protein's contextual information, raw noise data, and affinity information include:

[0080] S21. Determine the spatial relationship between the target pocket and the drug molecule based on the contextual information of the target protein.

[0081] Specifically, the spatial relationship between the target pocket and the drug molecule is determined based on the contextual information of the target protein, including:

[0082] S211. Determine the active site information of the target protein based on the target pocket. The active site information of the target protein represents the region that can bind to the target drug molecule.

[0083] S212. Determine the spatial relationship characteristics of the target pocket and drug molecule based on the active site information of the drug molecule scaffold and the target protein.

[0084] S22. Based on the original noise data and the spatial relationship between the target pocket and the drug molecule, noise is added to the R chain of the drug molecule to obtain the noise-added molecular structure.

[0085] Specifically, noise is added to the R chain of the drug molecule based on the original noise data to obtain the noisy molecular structure, including:

[0086] For each training iteration, the initial R chain on the drug molecule scaffold is denoted as... ,make Indicates t ( The R-chain after adding noise at the next time step;

[0087] The formula for adding noise to the R chain of a drug molecule in a diffusion model is as follows:

[0088] ;

[0089] in, The initial R chain on the drug molecule scaffold , This represents the noise parameter added to the drug molecule scaffold at time step t. , This is the variance scheduling parameter at time step i, used to control the amount of noise added; variance scheduling parameter at time step i Acquired from raw noise data; It is random standard Gaussian noise; `r` represents the R-chain with noise added at step `t`; `c` is a user-defined parameter. Here, noise is only added to the coordinates and features of the R-chain `x`, while the context `c` remains fixed and no noise is added to preserve the structural information of the pockets and scaffolds. This process is a conditionally independent Markov chain. The total distribution is: Finally, training pairs are generated. (The parameters here have the same meaning as in the noise addition formula for the diffusion model above; z represents the R-chain at the i-th noise reduction step.) Note: Sampling At that time, the coordinate part applies anchoring transformation to ensure In the anchor point coordinate system, chemical effectiveness and accuracy can be improved.

[0090] The noisy molecular structure was determined based on the pocket atoms, drug molecule scaffold, and the t-step noisy R-chain. It should be noted that noise was only added to the R-chain; the pocket atoms of the target protein and the drug molecule scaffold remained fixed without noise addition to preserve the structural information of the pocket and scaffold.

[0091] S23. Based on the affinity information, the noise in the denoised molecular structure is removed iteratively, and the affinity-sensitive R chain is recovered from the noise to obtain the target drug molecular structure.

[0092] Specifically, based on affinity information, noise in the noise-added molecular structure is iteratively removed, and the affinity-sensitive R chain is recovered from the noise to obtain the target drug molecule structure, including:

[0093] S231. Use an equivariant graph neural network to predict the node feature noise and coordinate noise of the R chain in the noisy molecular structure.

[0094] S232. Determine the affinity characteristics of the target pocket and R chain based on affinity information.

[0095] Affinity features are extracted from affinity information using an affinity cross-attention module. The affinity cross-attention module is equipped with an encoder, which is used to extract affinity features from affinity information.

[0096] S233. The node feature noise and affinity feature of the R chain are fused through a cross-attention mechanism to obtain the first fused feature.

[0097] S234. The coordinate noise and affinity features of the R chain are fused through a cross-attention mechanism to obtain the second fused feature.

[0098] S235. The adjusted node feature noise is obtained based on the first fusion feature, and the adjusted coordinate noise is obtained based on the second fusion feature.

[0099] S236. By iteratively removing the adjusted node feature noise and adjusted coordinate noise corresponding to the molecular structure R chain, the molecular structure R chain is reconstructed to obtain the reconstructed R chain.

[0100] Equivariant Graph Neural Networks (EGNN) are used to extract features from the contextual information of the target protein, resulting in a spatial relationship feature representation between the target pocket and the drug molecule. Based on affinity information and the spatial relationship feature representation between the target pocket and the drug molecule, the denoising process is gradually performed to generate the drug molecule structure.

[0101] This diffusion-based three-model injects affinity into the EGNN layer through an affinity cross-attention module, enabling joint learning of affinity and molecular structure, and improving the generation rate of high-affinity R chains.

[0102] Before using the isomorphic graph neural network, initialize the relevant parameters of the EGNN backbone (6-layer EGCL), including the noise predictor for node features h and coordinates r. The overall function of EGNN is: Output noise

[0103] in, Here, R is the R-chain at step t for denoising, c is the intermediate variable of the joint features of pocket atoms and scaffold atoms after being extracted by the E(3)-equivariant graph neural network, and t is the current denoising step number. Affinity information (affinity features) is injected through the affinity cross-attention module. For the output of EGNN, To address the node characteristic noise of the previous node and This represents the coordinate noise of the current step node (the noise of the coordinates of each node atom).

[0104] Then, the affinity cross-attention module encoder was initialized. Projecting scalar affinity A into an intermediate representation ;

[0105] Finally, initialize the affinity cross-attention module projection matrix. , , .

[0106] The training steps for the diffusion model are as follows:

[0107] S301. Data Sampling: Sample batch data and extract (x,c,A) from the dataset, where x is the R chain, c is the pocket protein and ligand scaffold (target pocket and drug molecule scaffold structure), and A is the affinity tag.

[0108] The atomic coordinates of chain x are locally transformed and anchored to the anchor atoms of the scaffold to ensure translation invariance and chemical validity. ,in These are the original coordinates of atom i. These are the coordinates of the scaffold anchor points. This step is applied to all samples to ensure that the generated R chain is aligned with the attachment site.

[0109] S302, Forward Diffusion: For each training iteration, set... (Initial R chain), let Indicates t ( The R-chain after adding noise at the next time step.

[0110] S303, Model Forward Propagation and Noise Prediction: Input Noise Samples Contextual information c, time step t, and affinity label A.

[0111] Using the E(3)-Equivariant Graphical Neural Network (EGNN) function To predict noise ,in It is node characteristic noise. It's coordinate noise. The specific steps are as follows:

[0112] Constructing a joint graph Initial node features , express Based on atomic characteristics With conditional features Composed of multiple parts; coordinates ,in Anchored These represent the coordinates of atoms in the R chain of the molecule. Coordinates of protein pocket atoms Time step embedding Or cosine variant, spliced ​​to Time step embedding Used to concatenate with node features after encoding, so that the model can perceive the current noise reduction progress.

[0113] Then, the adjusted node feature noise and adjusted coordinate noise corresponding to the molecular structure R chain are removed iteratively. In this embodiment, since the isotropic graph neural network has 6 layers, 6 iterations are performed. Among them, the steps of iterating the EGCL layer (l=1~6, iterating 6 times) include:

[0114] 1. Compute inter-node messages: (i, j) represent any two nodes, and the model learns the local or global association information between any two nodes), where MLP stands for Multilayer Perceptron.

[0115] 2. Update node characteristics:

[0116] 3. Update coordinates (ensure E(3) is isovariant): ,in It is an MLP function.

[0117] Add ADCA to a specified layer of EGNN (e.g., after the third or sixth layer) to inject affinity A into the current feature. The specific steps are as follows:

[0118] 1. Affinity coding: Where M is the total number of nodes. It is a linear projection layer.

[0119] 2. Calculate the attention matrix: ,in It is the flat node feature extracted from the current EGNN layer; , , It is the science department projection matrix for each layer i; K is the query vector, K is the key vector, and V is the value vector.

[0120] 3. Calculate attention score: , The attention score is represented by the Softmax function, which is a normalized exponential function that transforms any real vector into a probability distribution.

[0121] 4. Unify features: Inject affinity guidance into the structural context to ensure that noise prediction is biased towards high affinity optimization.

[0122] 5. Output prediction noise: ,in and It is the initial noisy input used to isolate and denoise residuals.

[0123] S304: Loss Calculation: Calculate the simplified denoising loss (including feature noise and coordinate noise): This is equivalent to the diffusion loss term of the variational lower bound, focusing on noise matching. This represents the square of the second normal form.

[0124] S305: Backpropagation updates all parameters. In this embodiment, the Adam optimizer is used to update all parameters.

[0125] S306: Iteration and Convergence: Repeat the above steps until the performance of the validation set is stable, and the final diffusion model is obtained.

[0126] Before removing the node feature noise and coordinate noise after adjusting the molecular structure R chain, the method further includes:

[0127] The current length of the molecular structure R chain is identified and compared with the reference R chain length. If the current length of the R chain is less than the reference R chain length, pseudo-atoms are used to fill the R chain so that the length of the molecular structure R chain is the reference R chain length.

[0128] In this embodiment, the reference R-chain length is set to 10. For samples with an R-chain length < 10, pseudo-atom node features are added, such as... Pseudo-atomic node features are added using standardized input dimensions.

[0129] S237. Generate the target drug molecule structure based on the drug molecule scaffold and the reconstructed R chain.

[0130] In a preferred embodiment of this example, generating the target drug molecule structure based on the drug molecule scaffold and the reconstructed R chain includes:

[0131] Analyze the reconstructed R chain and determine whether there are pseudo atoms in the reconstructed R chain:

[0132] If so, remove the pseudo atoms in the reconstructed R chain to obtain the processed R chain, and generate the target drug molecule structure based on the drug molecule scaffold and the processed R chain.

[0133] If not, the target drug molecule structure is generated based on the drug molecule scaffold and the reconstructed R chain.

[0134] Based on the above, by using an isovariant graph neural network and an affinity cross-attention module to predict noise, the structurally sound R-chain is gradually restored, thereby improving the Vina score and structural authenticity.

[0135] In this embodiment, the noise prediction steps are as follows:

[0136] S401. Initialize noise samples: Randomly sample noise samples of length 10 from a standard Gaussian distribution. And anchor the coordinates to the scaffold anchor point atom.

[0137] S402, Inverse Denoising Iteration (t = T~1):

[0138] Using the EGNN function Predicted noise Calculate the noise-free estimate: ,

[0139] in .

[0140] Update one step: .

[0141] S403, R-chain generation: From the final noise reduction result The pseudo-atoms are removed, and the R chain is generated by reduction.

[0142] Existing de novo generation methods are prone to geometric distortion. This scheme uses a coordinate anchoring strategy and pseudo-atom filling to ensure that the generation process takes place in a local coordinate system, avoiding spatial conflicts. Combined with SAR supervision on a proprietary dataset, the generated molecules achieve a 0.137 reduction in SAS score and a 0.025 improvement in similarity, demonstrating higher realism and feasibility.

[0143] To evaluate the experimental results, the diffusion model of this application was validated on three key therapeutic targets: dopamine transporter (DAT), SARS-CoV-2 main protease (Mpro), and vascular endothelial growth factor receptor (VEGFR). These three targets represent three major categories of diseases: neurological disorders, viral infections, and cancer.

[0144] Dopamine transporter (DAT): DAT is an important target for the treatment of neuropsychiatric disorders such as Parkinson's disease, depression, and ADHD.

[0145] SARS-CoV-2 main protease (Mpro): Mpro is a key enzyme in the replication and transcription of the novel coronavirus and a priority target for the development of antiviral drugs.

[0146] Vascular endothelial growth factor receptor (VEGFR): The VEGFR family is an important target for the development of anti-tumor and anti-angiogenic drugs.

[0147] Each target has a different structure and functional type, resulting in the following: Figure 3 exhibit.

[0148] For each given scaffold-pocket pair, the model underwent 100 independent modification runs, generating 100 candidate molecules. The molecule with the highest Vina docking score was selected for molecular dynamics simulations to assess the stability of ligand-protein binding. Docking results showed that ADOptDiff generated molecules with higher binding affinity in all three systems compared to the reference ligand. Furthermore, all selected candidates had an overall reachability score below 3, which is generally considered to indicate good feasibility. Molecular dynamics simulations further demonstrated that the generated ligands maintained stable binding modes at the DAT and SARS-CoV-2019 Mpro targets throughout the simulation, crucial for ensuring sustained drug action in vivo. Despite high conformational fluctuations in the VEGFR complex due to the inherent flexibility of its loop region, the ligand binding pocket remained structurally stable. This indicates that ADOptDiff can handle complex, somewhat flexible targets and successfully generate ligands that bind stably to them. These results demonstrate that ADOptDiff can generate chemically feasible and functionally optimized ligands for different protein classes, confirming its potential use in the drug discovery scenario.

[0149] Based on the same inventive concept, one embodiment of the present invention provides a molecular optimization system based on a neural network.

[0150] The neural network-based molecular optimization system described in this invention can be installed in an electronic device. According to its functions, the neural network-based molecular optimization system includes an acquisition module and a processing module. The acquisition module acquires the contextual information, raw noise data, and affinity information of the target protein. The processing module inputs the contextual information, raw noise data, and affinity information of the target protein into a diffusion model. The diffusion model processes the contextual information, raw noise data, and affinity information of the target protein according to preset rules, filters out the drug molecule structure corresponding to the affinity-sensitive R chain, and outputs the target drug molecule structure.

[0151] The diffusion model's processing steps for the target protein's contextual information, raw noise data, and affinity information include:

[0152] Determine the spatial relationship between the target pocket and the drug molecule based on the contextual information of the target protein;

[0153] Based on the original noise data and the spatial relationship between the target pocket and the drug molecule, noise is added to the R chain of the drug molecule to obtain the noisy molecular structure;

[0154] Based on affinity information, noise in the denoised molecular structure is removed iteratively, and the affinity-sensitive R chain is recovered from the noise to obtain the target drug molecular structure.

[0155] The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0156] The various variations and specific examples of the neural network-based molecular optimization method provided in the above embodiments are also applicable to the neural network-based molecular optimization system of this embodiment. Through the foregoing detailed description of the neural network-based molecular optimization method, those skilled in the art can clearly understand the implementation method of the neural network-based molecular optimization system in this embodiment. For the sake of brevity, it will not be described in detail here.

[0157] This application also discloses an electronic device, such as Figure 4 The diagram shown is a schematic representation of an electronic device for a neural network-based molecular optimization method according to an embodiment of the present invention. The electronic device may include at least one processor 10, a memory 11 communicatively connected to the at least one processor, a communication bus 12, and a communication interface 13. It may also include a computer program, such as a program for a neural network-based molecular optimization method, stored in the memory 11 and executable on the processor 10.

[0158] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It performs various functions of the electronic device and processes data by running or executing programs or modules stored in the memory 11 (e.g., executing methods for molecular optimization based on neural networks), and by calling data stored in the memory 11.

[0159] The memory 11 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of an electronic device, such as a portable hard drive. In other embodiments, the memory 11 can be an external storage device of the electronic device, such as a plug-in portable hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc. Furthermore, the memory 11 can include both internal and external storage units of the electronic device. The memory 11 can be used not only to store application software and various types of data installed on the electronic device, such as code for molecular optimization methods based on neural networks, but also to temporarily store data that has been output or will be output.

[0160] The communication bus 12 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc. The bus is configured to enable communication between the memory 11 and at least one processor 10, etc.

[0161] Communication interface 13 is used for communication between the aforementioned electronic device and other devices, including a network interface and a user interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, Bluetooth interface, etc.), typically used to establish communication connections between the electronic device and other electronic devices. The user interface may be a display, an input unit (such as a keyboard), or optionally, a standard wired or wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device and to display a visual user interface.

[0162] Figure 4 Only electronic devices with components are shown; it will be understood by those skilled in the art that... Figure 4The structure shown does not constitute a limitation on the electronic device and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0163] For example, although not shown, the electronic device may also include a power supply (such as a battery) to power various components. Preferably, the power supply can be logically connected to at least one processor 10 via a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0164] It should be understood that the embodiments are for illustrative purposes only and are not limited to this structure in the scope of the patent application.

[0165] Furthermore, if the modules / units integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile.

[0166] This application provides a computer-readable storage medium, including, for example, any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM). The computer-readable storage medium stores a computer program capable of being loaded by a processor and executing the neural network-based molecular optimization method of the above embodiments.

[0167] In the description of this specification, the references to terms such as "an embodiment," "some embodiments," "example," "specific example," "a implementation," "a preferred implementation," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0168] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A molecular optimization method based on neural networks, characterized in that, The method includes: Obtain contextual information, raw noise data, and affinity information of the target protein; The context information, raw noise data, and affinity information of the target protein are input into the diffusion model. The diffusion model processes the context information, raw noise data, and affinity information of the target protein according to preset rules, filters out the drug molecule structure corresponding to the affinity-sensitive R chain, and outputs the target drug molecule structure. The diffusion model's processing steps for the target protein's contextual information, raw noise data, and affinity information include: Determine the spatial relationship between the target pocket and the drug molecule based on the contextual information of the target protein; Based on the original noise data and the spatial relationship between the target pocket and the drug molecule, noise is added to the R chain of the drug molecule to obtain the noisy molecular structure; Based on affinity information, noise in the denoised molecular structure is removed iteratively, and the affinity-sensitive R chain is recovered from the noise to obtain the target drug molecular structure. The process of iteratively removing noise from the denoised molecular structure based on affinity information, and recovering the affinity-sensitive R chain from the noise to obtain the target drug molecule structure, includes: Using an isomorphic graph neural network, node feature noise and coordinate noise of the R chain in a noisy molecular structure are predicted; The affinity characteristics of the target pocket and R chain were determined based on affinity information; The node feature noise and affinity feature of the R chain are fused using a cross-attention mechanism to obtain the first fused feature; The coordinate noise and affinity features of the R-chain are fused using a cross-attention mechanism to obtain a second fused feature; The adjusted node feature noise is obtained based on the first fusion feature, and the adjusted coordinate noise is obtained based on the second fusion feature; By iteratively removing the adjusted node feature noise and adjusted coordinate noise corresponding to the molecular structure R chain, the molecular structure R chain is reconstructed, and the reconstructed R chain is obtained. The target drug molecule structure is generated based on the drug molecule scaffold and the reconstructed R chain; Input noise samples, context information, time steps, and affinity labels; Using the E(3)-equivariant graph neural network function to predict noise. ,in It is node characteristic noise. This is coordinate noise, where t represents the time step. The specific steps are as follows: Constructing a joint graph Initial node features , express Based on atomic characteristics With conditional features Composed of multiple parts; coordinates ,in Anchored These represent the coordinates of atoms in the R chain of the molecule. Coordinates of protein pocket atoms Time step embedding Or cosine variant, spliced ​​to Time step embedding Used to concatenate with node features after encoding; iteratively removes the adjusted node feature noise and adjusted coordinate noise corresponding to the molecular structure R chain; The steps for noise prediction are as follows: Initialize noise samples: Randomly sample noise samples of length 10 from a standard Gaussian distribution. And anchor the coordinates to the scaffold anchor point atom; Reverse denoising iteration; The pseudo-atoms are removed from the final denoising result, restoring the R chain.

2. The molecular optimization method based on neural networks as described in claim 1, characterized in that, The contextual information of the target protein includes the target pocket and the drug molecule scaffold. The contextual information of the target protein is extracted using an isomorphic graph neural network to obtain the spatial relationship feature representation of the target pocket and the drug molecule. Based on the affinity information and the spatial relationship feature representation of the target pocket and the drug molecule, the noise is gradually removed to generate the drug molecule structure.

3. The molecular optimization method based on neural networks as described in claim 2, characterized in that, Determining the spatial relationship between the target pocket and the drug molecule based on the contextual information of the target protein includes: The active site information of the target protein is determined based on the target pocket. The active site information of the target protein represents the region that can bind to the target drug molecule. The spatial relationship between the target pocket and the drug molecule is represented by information on the active sites of the drug molecule scaffold and the target protein.

4. The molecular optimization method based on neural networks as described in claim 2, characterized in that, The step of adding noise to the R chain of the drug molecule based on the original noise data and the spatial relationship between the target pocket and the drug molecule to obtain the noisy molecular structure includes: For each training iteration, the initial R chain on the drug molecule scaffold is denoted as... ,make Indicates t ( The R-chain after adding noise at the next time step; The formula for adding noise to the R chain of a drug molecule in a diffusion model is as follows: ； in, The initial R chain on the drug molecule scaffold , This represents the noise parameter added to the drug molecule scaffold at time step t. , This is the variance scheduling parameter at time step i, used to control the amount of noise added; variance scheduling parameter at time step i Acquired from raw noise data; It is random standard Gaussian noise; This represents the R-chain with noise added in step t; The noisy molecular structure was determined based on pocket atoms, drug molecule scaffolds, and the t-step noisy R-chain.

5. The molecular optimization method based on neural networks as described in claim 1, characterized in that, Before removing the node feature noise and coordinate noise after adjusting the molecular structure R chain, the method further includes: The current length of the molecular structure R chain is identified and compared with the reference R chain length. If the current length of the R chain is less than the reference R chain length, pseudo-atoms are used to fill the R chain so that the length of the molecular structure R chain is the reference R chain length.

6. The molecular optimization method based on neural networks as described in claim 5, characterized in that, The generation of the target drug molecule structure based on the drug molecule scaffold and the reconstructed R chain includes: Analyze the reconstructed R chain and determine whether there are pseudo atoms in the reconstructed R chain: If so, remove the pseudo atoms in the reconstructed R chain to obtain the processed R chain, and generate the target drug molecule structure based on the drug molecule scaffold and the processed R chain. If not, the target drug molecule structure is generated based on the drug molecule scaffold and the reconstructed R chain.

7. A neural network-based molecular optimization system for implementing the neural network-based molecular optimization method according to any one of claims 1 to 6, characterized in that, include: The acquisition module is used to acquire contextual information, raw noise data, and affinity information of the target protein; The processing module is used to input the context information, raw noise data and affinity information of the target protein into the diffusion model. The diffusion model processes the context information, raw noise data and affinity information of the target protein according to preset rules to screen out the drug molecule structure corresponding to the affinity-sensitive R chain and output the target drug molecule structure. The diffusion model's processing steps for the target protein's contextual information, raw noise data, and affinity information include: Determine the spatial relationship between the target pocket and the drug molecule based on the contextual information of the target protein; Based on the original noise data and the spatial relationship between the target pocket and the drug molecule, noise is added to the R chain of the drug molecule to obtain the noisy molecular structure; Based on affinity information, noise in the denoised molecular structure is removed iteratively, and the affinity-sensitive R chain is recovered from the noise to obtain the target drug molecular structure.

8. An electronic device, characterized in that, The electronic device includes: At least one processor (10); and, A memory (11) communicatively connected to the at least one processor (10); The memory (11) stores a computer program that can be executed by the at least one processor (10) to enable the at least one processor (10) to perform the neural network-based molecular optimization method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program; when executed by a processor, the computer program implements the neural network-based molecular optimization method as described in any one of claims 1 to 6.

Images