A method and system for directed production of pharmaceutical molecules

Abstract

The application relates to the technical field of drug molecule design, in particular to a drug molecule directional generation method and system.The method comprises the following steps: firstly, a molecular representation base model is constructed, chemical space information of a drug molecule structure is transformed, updated, pooled and integrated, and a one-dimensional molecular code vector capable of reflecting molecular structure and pharmacodynamic attribute information is generated; a molecular structure generation model and a molecular attribute prediction model are constructed, and the one-dimensional molecular code vector is segmented for training, the former is subjected to dimension expansion, feature conversion and decoding to restore the vector into a drug molecule structural formula, and the latter predicts the molecular attribute; according to the molecular attribute prediction result, a new drug molecule structure with a target pharmacodynamic attribute can be directionally optimized and generated. Through accurate modeling of the distribution law of the drug chemical space, the drug research and development efficiency and success rate are effectively improved, and strong technical support is provided for the field.

Description

Technical Field

[0001] This invention relates to the field of drug molecule design technology, and in particular to a method and system for the directed generation of drug molecules. Background Technology

[0002] In the field of drug development, the efficient and precise design of drug molecules with specific structures and properties has always been a goal pursued by researchers. With the continuous development of computer technology and machine learning algorithms, more and more molecular design methods have been proposed to assist in the drug development process.

[0003] In early molecular design methods, REINVENT emerged as a framework based on reinforcement learning and SMILES generation. It uses the SMILES expression of a molecule as the core optimization target, leveraging reinforcement learning to optimize the structure of the input molecule and generate new molecules. However, this method has a significant drawback: its optimization target is a complete, independent molecular structure. This prevents it from achieving directed, continuous optimization in the encoding space, resulting in low search efficiency and difficulty in quickly finding molecular structures that meet specific requirements when facing large-scale molecular search tasks.

[0004] LatentGAN and MDM, two molecular generation frameworks, also have some applications in molecular design. They mainly achieve molecular generation by constructing latent space vectors. However, these latent space vectors are only designed for molecular generation tasks and lack effective characterization of molecular pharmacological properties. This means that in the drug design process, molecules generated solely based on these methods are unlikely to have good efficacy and cannot meet the actual functional requirements of drug development.

[0005] ScaffoldGVAE is also a molecular generative framework, but its entire molecular generation process is fragmented by multiple steps. This means it must rely on specific combinations of input molecules to define the direction of molecular generation, thus preventing targeted molecular optimization. In practical applications, when molecules need to be designed according to specific drug targets or efficacy requirements, this limitation of ScaffoldGVAE makes it difficult to flexibly and accurately generate the required molecular structures.

[0006] RegressionTransformer, as a conditional molecule generation framework, has its own characteristics in molecular design. However, this method requires explicitly modeling the molecular properties within the generative model, which leads to poor generalization and versatility. In specific drug design tasks, especially with small datasets, fine-tuning the model often fails to achieve ideal results, limiting its application in practical drug development.

[0007] In summary, existing molecular design methods have many shortcomings in areas such as molecular structure optimization, pharmacodynamic characterization, targeted molecule generation, and model generalization. Therefore, there is an urgent need for a new molecular design technology that can effectively address these issues, improve the efficiency and accuracy of molecular design, and provide stronger support for drug development. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method and system for the directed generation of drug molecules, the method comprising the following steps:

[0009] Step S1: Obtain the drug molecular structure, construct a molecular characterization base model, and based on the molecular characterization base model, perform information transformation, information updating, information pooling and information integration operations on the chemical spatial information of the drug molecular structure to generate a one-dimensional molecular encoding vector that can characterize molecular structure information and pharmacodynamic attribute information. Then, optimize the molecular characterization base model according to the intermolecular similarity in the one-dimensional molecular encoding vector.

[0010] Step S2: Construct a molecular structure generation model and a molecular property prediction model. Based on the one-dimensional molecular encoding vector, perform segmented training on the molecular structure generation model and the molecular property prediction model to obtain the trained molecular structure generation model and the trained molecular property prediction model, including:

[0011] The trained molecular structure generation model is used to expand the dimension, transform the features, and decode the one-dimensional molecular encoding vector to restore the one-dimensional molecular encoding vector to the drug molecule structure formula, so as to ensure that the one-dimensional molecular encoding vector and the drug molecule structure formula have a one-to-one correspondence; and based on the one-dimensional molecular encoding vector, the trained molecular property prediction model is used to obtain one or more molecular property prediction results.

[0012] Step S3: Based on the molecular attribute prediction results of the one-dimensional molecular encoding vector, optimize the one-dimensional molecular encoding vector through backpropagation, and restore the optimized one-dimensional molecular encoding vector to the molecular structure to achieve targeted optimization of drug molecules.

[0013] In one embodiment of the present invention, the method for generating a one-dimensional molecular encoding vector capable of representing molecular structure information and pharmacodynamic property information in step S1 is as follows:

[0014] S11: Convert the drug molecule structure into a SMILES string, and unify the format of the SMILES string to obtain a preprocessed SMILES string;

[0015] S12: Convert the preprocessed SMILES string into a molecular graph, and convert it into an isomorphic graph composed of nodes representing atoms and edges representing chemical bonds according to the atomic and chemical bond symbols in the molecular graph;

[0016] S13: Update the node and edge representations in the isomorphic graph to obtain the updated molecular graph representation;

[0017] S14: Input each node representation in the updated molecular graph representation into a multilayer perceptron, predict the attention weight of each node, and input the attention weight into the Softmax activation function for normalization to obtain the normalized attention weight;

[0018] S15: Multiply the normalized attention weights by the corresponding node representations and then perform sum pooling to obtain a one-dimensional molecular encoding vector.

[0019] In one embodiment of the present invention, the method for updating the node representation and edge representation in the isomorphic graph in step S13 is as follows:

[0020] S131: For the initial feature representation of each node, the initial feature representation is projected to a higher-dimensional space through a linear transformation to obtain the input node;

[0021] S132: For each edge connecting any two input nodes, the current edge features are iteratively updated by concatenating the features of any node with the features of the edge to obtain the updated edge.

[0022] S133: Based on the updated edges, for each node, aggregate the features of all its connected edges, and fuse the aggregation result with the features of the node to iteratively update the current node features, thereby obtaining the currently updated node;

[0023] S134: For any neighboring node of the currently updated node, multiply the features of the neighboring node and the updated edge features element-wise to obtain the integrated features of the neighboring node. Calculate the integrated features of all neighboring nodes of the updated node and aggregate them. Multiply the aggregated neighboring node features element-wise with the features generated by the self-loop attention mechanism to obtain the final updated node features.

[0024] In one embodiment of the present invention, the method for updating the information of the node representation and edge representation in the isomorphic graph further includes: introducing information such as molecular 3D conformation, conformational dynamics, conformational space, pharmacophore features, and molecular shape to update the features of each node and each edge in the isomorphic graph.

[0025] In one embodiment of the present invention, the method for optimizing the molecular representation pedestal model based on the inter-molecular similarity in the one-dimensional molecular encoding vector in step S1 is as follows:

[0026] Using a multilayer perceptron as a projection head, any two sets of one-dimensional molecular coding vectors are concatenated and their shapes adjusted before being input into the projection head to obtain a prediction similarity matrix between the two sets of molecular coding vectors.

[0027] The similarity between different molecules in the drug molecular structure is calculated to obtain the true similarity matrix;

[0028] The error between the predicted similarity matrix and the true similarity matrix is ​​calculated, and backpropagation is performed based on the error to adjust the parameters of the molecular characterization pedestal model.

[0029] In one embodiment of the present invention, any element in the true similarity matrix represents one or more similarities between different molecules.

[0030] In one embodiment of the present invention, the method for obtaining the trained molecular structure generation model is as follows:

[0031] Based on the molecular structure generation model, the one-dimensional encoded vector is autoregressively decoded to calculate the predicted probability of each character representing different molecules at the current position. Based on the predicted probability, a prediction vocabulary is obtained.

[0032] The SMILES string, transformed according to the drug molecule structure, is divided into multiple character tokens. Based on the actual position information of each character token in the original SMILES string, all character tokens are summarized to form a real vocabulary.

[0033] Calculate the character error at each corresponding position in the predicted vocabulary and the real vocabulary, and adjust the parameters of the molecular structure generation model based on the character error through backpropagation until the character error is within a preset range, thereby obtaining the trained molecular structure generation model.

[0034] In one embodiment of the present invention, the method for restoring the one-dimensional molecular encoding vector to a drug molecule structural formula using the trained molecular structure generation model in step S2 is as follows:

[0035] The one-dimensional molecular encoding vector is used as a memory vector, and the memory vector is autoregressively decoded through the Transformer decoder structure to obtain a feature vector sequence.

[0036] The feature vector sequence is decoded using a multilayer perceptron to obtain the prediction score for each character;

[0037] Based on the predicted score, the probability distribution value of each character appearing at the current position is calculated. According to the probability distribution value, the output character at the current position is predicted. After the first character is generated, the character is mapped to a low-dimensional embedding vector and concatenated with the previous feature vector. The concatenated vector is then input into the multilayer perceptron to predict the next character. This prediction process is repeated until the generated string meets the specified length, and finally a SMILES string composed of multiple characters is obtained.

[0038] The SMILES string is converted into the drug molecular structure formula.

[0039] In one embodiment of the present invention, the method for obtaining one or more molecular attribute prediction results through the trained molecular attribute prediction model is as follows: the one-dimensional molecular encoding vector is transformed into molecular attribute prediction values ​​by performing multi-level dimensional expansion, batch standardization, and activation function through a deep neural network.

[0040] Based on the same inventive concept, the present invention also provides a drug molecule-directed generation system for implementing the aforementioned drug molecule-directed generation method, the drug molecule-directed generation system comprising the following modules:

[0041] The molecular characterization module is used to acquire the drug molecular structure, construct a molecular characterization base model, and perform information transformation, information updating, information pooling and information integration operations on the chemical spatial information of the drug molecular structure based on the molecular characterization base model to generate a one-dimensional molecular encoding vector that can characterize molecular structure information and pharmacodynamic attribute information. Furthermore, the molecular characterization base model is optimized based on the intermolecular similarity in the one-dimensional molecular encoding vector.

[0042] The decoding module is used to construct a molecular structure generation model and a molecular property prediction model. Based on the one-dimensional molecular encoding vector, the molecular structure generation model and the molecular property prediction model are trained in segments to obtain the trained molecular structure generation model and the trained molecular property prediction model. The molecular structure formula and the molecular property prediction result are obtained through the trained molecular structure generation model and the trained molecular property prediction model, respectively.

[0043] And a molecular orientation optimization module, used to optimize the one-dimensional molecular encoding vector through backpropagation based on the molecular attribute prediction results of the one-dimensional molecular encoding vector, and restore the optimized one-dimensional molecular encoding vector to the molecular structure, so as to achieve the orientation optimization of drug molecules;

[0044] The decoding module includes a molecular structure generation submodule and a molecular property prediction submodule. The molecular structure generation submodule restores the one-dimensional molecular code to a molecular structure formula through the trained molecular structure generation model. The molecular property prediction submodule obtains one or more molecular property prediction results based on the one-dimensional molecular code and through the trained molecular property prediction model.

[0045] The technical solution of the present invention has the following advantages over the prior art:

[0046] First, information integration and efficient characterization: By constructing a molecular characterization base model, the chemical spatial information of drug molecule structures can be comprehensively transformed, updated, pooled, and integrated to generate a one-dimensional molecular encoding vector. This approach effectively integrates molecular structure information and pharmacodynamic attribute information, efficiently representing the key features of drug molecules in a concise one-dimensional vector form, facilitating subsequent model processing and improving computational and information transfer efficiency.

[0047] Second, multi-task processing based on one-dimensional encoding vectors obtained from molecular characterization pedestal models can reconstruct molecular structures or predict various molecular properties. This multi-task capability not only helps to comprehensively understand the characteristics of drug molecules but also provides a rich data foundation for drug development. Based on the reconstructed molecular structures and property prediction results, new drug molecule structures with target pharmacodynamic properties can be optimized and generated in a targeted manner, expanding the application scope of the technology in the field of drug development and improving development efficiency and success rate. Attached Figure Description

[0048] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein...

[0049] Figure 1 This is a schematic diagram of a drug molecule-directed generation method provided in an embodiment of the present invention;

[0050] Figure 2 This is a schematic flowchart of a method for generating a one-dimensional molecular coding vector that can represent molecular structure information and pharmacodynamic properties, provided in an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of a molecular characterization pedestal model structure provided in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the structure decoder of the molecular characterization vector obtained in an embodiment of the present invention;

[0053] Figure 5 This is a performance comparison of the molecular characterization method of the present invention in virtual screening for similarity.

[0054] Figure 6 It involves testing the ability of a molecular structure decoder to generate new molecules by increasing the temperature and adding noise.

[0055] Figure 7 This invention demonstrates the performance of the Ouroboros encoder combined with a multilayer perceptron in molecular property prediction.

[0056] Figure 8 These are the migration results of two PDE4B inhibitors in the Ouroboros chemical coding space of this invention;

[0057] Figure 9 This is a schematic diagram of a drug molecule directed generation system provided in an embodiment of the present invention;

[0058] Explanation of the reference numerals in the accompanying drawings: 100, Molecular characterization module; 200, Decoding module; 200a, Molecular structure generation submodule; 200b, Molecular property prediction submodule; 300, Molecular orientation optimization module. Detailed Implementation

[0059] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0060] Example 1

[0061] See Figure 1 As shown, this invention provides a method for the targeted generation of drug molecules based on the basic architecture of the Ouroboros snake. The method includes the following steps:

[0062] Step S1: Obtain the drug molecular structure, construct a molecular characterization base model, and based on the molecular characterization base model, perform information transformation, information updating, information pooling and information integration operations on the chemical spatial information of the drug molecular structure to generate a one-dimensional molecular encoding vector that can characterize molecular structure information and pharmacodynamic attribute information.

[0063] Step S2: Construct a molecular structure generation model and a molecular property prediction model. Based on the one-dimensional molecular encoding vector, perform segmented training on the molecular structure generation model and the molecular property prediction model to obtain the trained molecular structure generation model and the trained molecular property prediction model. This includes: using the trained molecular structure generation model to perform dimensional expansion, feature transformation, and decoding on the one-dimensional molecular encoding vector, restoring the one-dimensional molecular encoding vector to a drug molecule structural formula, to ensure a one-to-one correspondence between the one-dimensional molecular encoding vector and its drug molecule structural formula; and based on the one-dimensional molecular encoding vector, using the trained molecular property prediction model to obtain one or more molecular property prediction results.

[0064] Step S3: Based on the molecular attribute prediction results of the one-dimensional molecular encoding vector, optimize the one-dimensional molecular encoding vector through backpropagation, and restore the optimized one-dimensional molecular encoding vector to the molecular structure to achieve targeted optimization of drug molecules.

[0065] Furthermore, the main architecture of the molecular characterization pedestal model includes: transformation of drug molecule structure, information updating, information pooling and information integration, which can be implemented in ways such as chemical language models and molecular diagram models.

[0066] Taking the molecular graph model as an example, the molecular representation base model includes the generation of molecular graphs, a graph neural network for extracting molecular information, a pooling network for compressing the updated graph data to one dimension, and an information integration network.

[0067] In step S1, as Figure 3 As shown, based on the molecular characterization pedestal model, the method for generating a one-dimensional molecular encoding vector that can represent molecular structure information and pharmacodynamic properties by performing information transformation, information updating, information pooling, and information integration operations on the chemical spatial information of the drug molecule structure is as follows:

[0068] S11: Convert the drug molecule structure into a SMILES string, and unify the format of the SMILES string to obtain a preprocessed SMILES string;

[0069] S12: The preprocessed SMILES string is divided into multiple tokens using a byte-pair tokenizer. Each token represents a single atom (e.g., "C" represents carbon), a chemical bond (e.g., "=" represents a double bond), or a substructure fragment (e.g., "Cl" represents a chlorine atom). Each token is assigned an identifier. All tokens are converted into a molecular graph. Based on the atomic and chemical bond symbols in the molecular graph, it is converted into an isomorphic graph G=(V,E) composed of nodes representing atoms and edges representing chemical bonds, where V represents the set of atoms and E represents the set of chemical bonds.

[0070] S13: The node and edge representations in the isomorphic graph are updated using a Weisfeller-Lyman (WLN) network of a graph neural network (GNN) to obtain an updated molecular graph representation; the node representation includes, but is not limited to, atom type, atom hybridization, atom charge, atom chirality, whether it is located inside a ring, and whether it has aromaticity; the edge representation includes, but is not limited to, the bond order of chemical bonds, whether it belongs to a conjugated system, and whether it is located inside a ring;

[0071] S14: Use a global self-attention pooling module to represent each node h in the updated molecular graph representation. i Given a multilayer perceptron (MLP), predict the attention weight 'a' for each node. i To ensure that the sum of all attention weights equals 1, the attention weight a is... i The input is passed to the Softmax activation function for normalization, resulting in the normalized attention weights w. i :

[0072]

[0073] S15: Adjust the normalized attention weights w i The corresponding node representation h i After multiplication, sum-pooling is performed to generate the graph-level node encoding H. graph That is, a one-dimensional molecular encoding vector:

[0074] H graph =∑ {i∈V} h i ·w i .

[0075] Furthermore, in step S13 above, the Weisfeller-Lyman (WLN) network of the graph neural network (GNN) updates the information of the node (atom) representation and edge (chemical bond) representation in the isomorphic graph through message passing as follows:

[0076] S131: Node Feature Projection: Initial feature representation h for each atom v , the initial atomic features h v The input WLN is projected to a higher-dimensional space through a linear transformation, and then the ReLU activation function is used to obtain the atomic features.

[0077]

[0078] In the formula, W in Let b represent the trainable weight matrix. in It is a bias parameter.

[0079] S132: Edge Feature Update: For each chemical bond connecting atoms u and v, update the edge feature h of atom u by... u and the characteristics of this chemical bond Perform a splicing operation to update the chemical bond features. The updated chemical bond characteristics were obtained.

[0080]

[0081] In the formula, t represents the number of message passes. This represents the chemical bond characteristics after (t+1) updates. Denotes the chemical bond characteristics after t updates, || denotes the linking operation, and W contact Let b represent the trainable weight matrix in the join operation. contact Indicates the bias parameter in the join operation;

[0082] S133: Node Feature Aggregation and Update: The current atomic features are iteratively updated by aggregating all adjacent bond features of each atom and fusing the aggregation result with the atomic features. Get the currently updated atomic features

[0083]

[0084] In the formula, N(v) is the set of neighboring atoms of atom v, and t represents the number of message passes; W aggregate b represents the trainable weight matrix in the aggregation operation; aggregate This represents the bias parameter in the aggregation operation;

[0085] S134: Attention Mechanism and Final Feature Update: For any neighboring atom u of atom v, the neighboring atom features h are updated. uv and updated key features Element-wise multiplication is performed to obtain the integrated features of the adjacent atoms. These integrated features are then used to calculate and aggregate the integrated features of all adjacent atoms of atom v. The aggregated adjacent atom features are then multiplied element-wise with the features generated through the self-loop attention mechanism to obtain the final updated atom features.

[0086]

[0087] In the formula, ⊙ represents element-wise multiplication, and W self Let W represent the self-loop attention weight matrix. bond W represents the key feature weight matrix. atom h represents the atomic characteristic weight matrix. uv This represents the characteristic representation of the neighboring atom u of the connected atom v.

[0088] Furthermore, the method for optimizing the molecular representation pedestal model based on the intermolecular similarity in the one-dimensional molecular encoding vector is as follows:

[0089] Using a multilayer perceptron as a projection head, any two sets of one-dimensional molecular coding vectors are concatenated and their shapes adjusted before being input into the projection head to obtain a prediction similarity matrix between the two sets of molecular coding vectors.

[0090] The true similarity matrix is ​​calculated based on the similarity of different molecules in the drug molecule structure. Each element in the true similarity matrix represents one or more similarities between different molecules, including but not limited to molecular fingerprint similarity and conformational space similarity. Molecular fingerprint similarity includes, but is not limited to, the similarity between molecular fingerprints such as topological molecular fingerprints (AtomPairs), circular molecular fingerprints (ECFP4), and substructure fingerprints (MACCS). Conformational space similarity refers to the characteristic values ​​(including maximum, average, minimum, etc.) of the similarity distribution between conformational ensembles of molecules under different energy windows.

[0091] The error between the predicted similarity matrix and the true similarity matrix is ​​calculated using mean squared error loss, and the parameters of the molecular characterization pedestal model are adjusted based on backpropagation using the error.

[0092] For 3D molecular conformation information, the geometric features can be encoded into vector form by calculating the three-dimensional distances and angles between atoms in the molecule, and then concatenated or fused with the vectors after word embedding. For example, the distance information between atoms can be added as a new feature dimension to the vector representation of the molecular sequence, allowing the model to learn the structural features of the molecule in three-dimensional space. For pharmacophore features, the pharmacophore component in the molecule can be identified based on known pharmacophore models, and its features can be integrated into the vector representation of the corresponding label, enabling the model to learn information related to pharmacophore efficacy.

[0093] Furthermore, in step S2, it is possible to use, as follows: Figure 4 The training process shown below yields the trained molecular structure generation model as follows:

[0094] The Transformer decoder based on the molecular structure generation model performs autoregressive decoding on the one-dimensional encoded vector, calculates the predicted probability of each character representing different molecules at the current position, and constructs a prediction vocabulary based on the predicted probability.

[0095] The SMILES string, which is transformed based on the drug molecule structure, is divided into multiple character tokens using a word segmenter. Based on the actual position information of each character token in the original SMILES string, all character tokens are summarized to form a real vocabulary.

[0096] The error between corresponding character tokens in the predicted vocabulary and the real vocabulary is calculated using weighted cross-entropy loss. Backpropagation is then performed based on the character error to adjust the parameters of the molecular structure generation model until the character error is within a preset range, thus obtaining the trained molecular structure generation model.

[0097] During the inference phase, the method for restoring the one-dimensional molecular encoding vector into a drug molecule structural formula using the trained molecular structure generation model is as follows:

[0098] The one-dimensional molecular encoding vector is used as a memory vector, and the memory vector is autoregressively decoded through the Transformer decoder structure to obtain a feature vector sequence.

[0099] The feature vector sequence is decoded using a multilayer perceptron to obtain a prediction score for each character; wherein the multilayer perceptron consists of multiple fully connected layers, and the number of neurons in the last layer is equal to the number of characters in the molecular structure formula;

[0100] Based on the predicted scores, the probability distribution value of each character appearing at the current position is calculated. The output character at the current position is predicted based on the maximum value of all probability distribution values, or by sampling with temperature-related noise. After generating the first character, it is mapped to a low-dimensional embedding vector and concatenated with the previous feature vector. The concatenated vector is then input back into the multilayer perceptron to predict the next character. During generation, empty tokens are ignored. When a termination token is encountered, the generation process ends, ultimately yielding a chemical language representation composed of a series of tokens, i.e., the complete SMILES string. The SMILES string is then used to reconstruct the drug molecule structure, thus achieving the goal of converting a one-dimensional molecular encoding vector into a drug molecule structure.

[0101] Furthermore, before performing molecular attribute prediction, the structure of the drug molecule to be predicted needs to be encoded into a one-dimensional molecular encoder vector through the molecular characterization pedestal model, and the one-dimensional molecular encoder vector is transformed into a 2048-dimensional molecular encoding vector. This vector contains key information such as the chemical structure and pharmacophore features of the molecule, providing a data foundation for subsequent attribute prediction.

[0102] The molecular property prediction model consists of three neural network components. The aforementioned 2048-dimensional molecular encoding vector is input into the trained molecular property prediction model.

[0103] 1) The first component includes a dropout layer with a dropout rate of 0.5, which encodes the drug molecule structure to be predicted into a one-dimensional molecular encoder vector using the molecular representation pedestal model, and then transforms this one-dimensional molecular encoder vector into a 2048-dimensional molecular encoding vector. The molecular encoding is projected from 2048 dimensions onto an 8192-dimensional feature space. After dimensional projection, a sigmoid activation function is applied to perform a nonlinear transformation on the features, thereby introducing nonlinear characteristics and enhancing the model's expressive power.

[0104] 2) The second component concatenates the 8192-dimensional molecular code generated by the first component with the original 2048-dimensional molecular code, resulting in a 10240-dimensional vector. This concatenated 10240-dimensional vector is then processed through two linear layers for feature transformation. Throughout these two linear layers, the feature dimension remains at 10240. To further enhance the model's learning ability and stability, the LeakyReLU activation function and batch normalization are applied between the two linear layers.

[0105] 3) The third component element-wise adds the input and output features of the second component to achieve feature fusion, resulting in the fused features. The fused features are first transformed through a linear layer, then normalized using batch normalization, followed by the introduction of non-linearity using the SiLU (Sigmoid-weighted Linear Unit) activation function. Next, the features are projected into a 1024-dimensional feature space, undergoing another round of batch normalization and SiLU activation. Finally, the features are projected through another linear layer with SiLU activation to the dimension of the target attribute. In the benchmark scenario, all target attributes are one-dimensional.

[0106] After processing by the molecular property prediction model, the final output is the predicted molecular property value. This value represents the model's prediction result for a specific property of the molecule (such as solubility, membrane permeability, lipophilicity, etc.), which can be used to assess the molecule's potential in drug development, to optimize and generate new drug molecule structures with target pharmacodynamic properties, and to help researchers screen and design more active drug molecules.

[0107] like Figure 5As shown, the molecular representation pedestal model (i.e., the Ouroboros molecular encoder) in this embodiment exhibits superior performance in benchmark tests after training. The benchmark sets involved include DUD-E (an artificially constructed virtual screening benchmark set, considered the gold standard in the field of virtual screening) and LIT-PCBA (a real-world virtual screening benchmark set derived from high-throughput experimental data), which are constructed in different ways. During the study, the PhaseShape method (using experimental structures as reference conformations), the GeminiMol method (using a similar pre-training scheme), and three traditional molecular fingerprinting methods were compared. The results show that the Ouroboros molecular encoder performs best in both benchmarks. This fully demonstrates that the molecular representation pedestal model in this invention possesses good generalization performance and is more powerful in characterization than traditional molecular fingerprinting, especially when processing the LIT-PCBA dataset from real-world data. The GeminiMol method is a molecular representation method that uses different molecular similarity descriptors for pre-training, but it lacks a corresponding molecular generation decoder.

[0108] Typically, the Transformer decoder structure used in this invention has between 16 and 64 attention unit heads, a hidden feature dimension between 1024 and 4096, and at least 4 network layers. The token set covers all common characters in SMILES. During training, a greedy matching mechanism is used to distinguish elements that are linguistically similar but chemically different, such as C and Cl, B and Br, thereby enabling different elements to be represented by different tokens.

[0109] The molecular decoder can be trained using synthetic molecules from public datasets or commercially available compound libraries. In this embodiment, synthetic compounds from various sources were used, with 128 different molecules trained in each step. The model converged after 500 million training steps. During molecule generation, the effectiveness of the molecular structure decoder in reconstructing the molecular structure encoded by the molecule and in generating new molecules was evaluated using two metrics: Similarity and Validity. Figure 6 As shown, during molecular structure decoding, by increasing the temperature and applying different levels of noise, the molecular structure decoder in this embodiment can generate more new molecules with low similarity to the original molecules (high diversity) and reasonable structures while maintaining high molecular effectiveness. However, when the temperature exceeds 1.50 (where token sampling is too aggressive) or the noise exceeds 0.005 (outside the numerical distribution range of the molecular encoding features themselves), the effectiveness of the molecular decoder decreases.

[0110] Since the encoding and decoding of molecular structures are trained independently, the molecular property prediction model in this invention uses a one-dimensional molecular vector to predict molecular properties. The implementation methods include, but are not limited to, multiple linear regression, decision trees, gradient boosting machines, and multilayer perceptrons (deep neural networks).

[0111] In this embodiment, a set of molecular property prediction models was trained using a multilayer perceptron to predict molecular properties such as water solubility, membrane permeability, clearance rate, half-life, toxicity, and tissue distribution. Figure 7 The results shown indicate that the method described in this invention ( Figure 7 Ouroboros outperforms the GeminiMol model and the combined approach of multiple molecular fingerprinting methods (ECFP4, FCFP6, AtomPairs, TopologicalTorsion) and automated machine learning (AutoGluon) in property prediction. Based on the SPEARMAN rank correlation coefficient between predicted and true values, the Ouroboros prediction model not only performs better in predicting properties like water, free energy, and lipid solubility (which CombineFP excels at), but also shows improved prediction performance in tasks where CombineFP performs poorly, such as half-life, clearance rate, and tissue distribution.

[0112] Because the molecular encoding in this invention exhibits excellent performance in drug virtual screening tasks, even independently of the molecular decoder and property predictor, the Ouroboros framework enables the optimization of molecular encoding vectors and their application in pharmacologically significant directed molecular structure migration. This capability allows molecular encoding vectors to migrate from a starting molecular structure to a target molecular structure, capturing structural changes throughout the migration path, thereby discovering novel chemical skeletons and uncovering new molecules with the potential to possess pharmacological properties of both skeletons.

[0113] In this embodiment, two PDE4B inhibitors with different backbones were selected, and their molecular encoding vectors were optimized using the AdamW optimizer. The loss function was set to the similarity between the molecular encoding vector and the target molecule, and the molecular vector was decoded in each optimization step. Figure 8 As shown, as the migration process progresses, the molecular structure gradually deviates from the starting structure, undergoing significant skeletal changes, and the substructures gradually transform into the target molecule. Towards the end of optimization, a large number of structures similar to the target molecule are generated. During the migration process, other molecular property optimization items can be added to the loss function, thereby achieving optimization of the drug-like properties of the molecule during the migration process.

[0114] Example 2

[0115] Based on the same inventive concept as in Example 1, this invention also provides a drug molecule-directed generation system for implementing the drug molecule-directed generation method described in Example 1. See also... Figure 9 As shown, the drug molecule-directed generation system includes: a molecular characterization module 100, a decoding module 200, and a molecular direction optimization module 300. Among them,

[0116] The molecular characterization module 100 is used to construct a molecular characterization base model. Based on the molecular characterization base model, the molecular structure is read, information is updated, information is pooled and information is integrated to generate a one-dimensional molecular encoding vector that can represent molecular structure information and pharmacodynamic attribute information.

[0117] The decoding module 200 is used to construct a molecular structure generation model and a molecular property prediction model. Based on the one-dimensional molecular encoding vector, the molecular structure generation model and the molecular property prediction model are trained in segments to obtain molecular structure formulas or molecular property prediction results.

[0118] The molecular orientation optimization module 300 is used to evaluate the interaction between drug molecules and different targets based on the reduced drug molecule structure and molecular property prediction results, and to develop multi-target drugs.

[0119] Specifically, the decoding module 200 includes a molecular structure generation submodule 200a and a molecular property prediction submodule 200b. The molecular structure generation submodule 200a performs dimensional expansion, feature transformation, and decoding on the one-dimensional molecular code using the molecular structure generation model, restoring the one-dimensional molecular code to a molecular structural formula. The molecular property prediction submodule 200b predicts one or more molecular properties based on the one-dimensional molecular code using the molecular property prediction model.

[0120] This embodiment proposes a drug molecule-directed generation system for implementing the aforementioned drug molecule-directed generation method. Therefore, the specific implementation of the drug molecule-directed generation system can be found in the embodiment section of the aforementioned drug molecule-directed generation method. For example, the molecular characterization module 100, the decoding module 200, and the molecular orientation optimization module 300 are respectively used to implement steps S1, S2, and S3 in the drug molecule-directed generation method described in Embodiment 1. Therefore, the specific implementation can be referred to the description of the corresponding embodiments. To avoid redundancy, it will not be repeated here.

[0121] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0122] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0123] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0124] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0125] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for the directed generation of drug molecules, characterized in that, Includes the following steps: Step S1: Obtain the drug molecular structure, construct a molecular characterization base model, and based on the molecular characterization base model, perform information transformation, information updating, information pooling and information integration operations on the chemical spatial information of the drug molecular structure to generate a one-dimensional molecular encoding vector that can characterize molecular structure information and pharmacodynamic attribute information. Then, optimize the molecular characterization base model according to the intermolecular similarity in the one-dimensional molecular encoding vector. Step S2: Construct a molecular structure generation model and a molecular property prediction model. Based on the one-dimensional molecular encoding vector, perform segmented training on the molecular structure generation model and the molecular property prediction model to obtain the trained molecular structure generation model and the trained molecular property prediction model, including: The trained molecular structure generation model is used to restore the one-dimensional molecular encoding vector to the drug molecule structure formula to ensure that the one-dimensional molecular encoding vector and the drug molecule structure formula have a one-to-one correspondence; and based on the one-dimensional molecular encoding vector, the trained molecular property prediction model is used to obtain one or more molecular property prediction results. Step S3: Based on the molecular attribute prediction results of the one-dimensional molecular encoding vector, optimize the one-dimensional molecular encoding vector through backpropagation, and restore the optimized one-dimensional molecular encoding vector to the molecular structure to achieve targeted optimization of drug molecules.

2. The method for targeted generation of drug molecules according to claim 1, characterized in that, In step S1, the method for generating a one-dimensional molecular encoding vector that can represent molecular structure information and pharmacodynamic properties is as follows: S11: Convert the drug molecule structure into a SMILES string, and unify the format of the SMILES string to obtain a preprocessed SMILES string; S12: Convert the preprocessed SMILES string into a molecular graph, and convert it into an isomorphic graph composed of nodes representing atoms and edges representing chemical bonds according to the atomic and chemical bond symbols in the molecular graph; S13: Update the node and edge representations in the isomorphic graph to obtain the updated molecular graph representation; S14: Input each node representation in the updated molecular graph representation into a multilayer perceptron, predict the attention weight of each node, and input the attention weight into the Softmax activation function for normalization to obtain the normalized attention weight; S15: Multiply the normalized attention weights by the corresponding node representations and then perform sum pooling to obtain a one-dimensional molecular encoding vector.

3. The method for targeted generation of drug molecules according to claim 2, characterized in that, In step S13, the method for updating the information of the node representations and edge representations in the isomorphic graph includes: S131: For the initial feature representation of each node, the initial feature representation is projected to a higher-dimensional space through a linear transformation to obtain the input node; S132: For each edge connecting any two input nodes, the current edge features are iteratively updated by concatenating the features of any node with the features of the edge to obtain the updated edge. S133: Based on the updated edges, for each node, aggregate the features of all its connected edges, and fuse the aggregation result with the features of the node to iteratively update the current node features, thereby obtaining the currently updated node; S134: For any neighboring node of the currently updated node, multiply the features of the neighboring node and the updated edge features element-wise to obtain the integrated features of the neighboring node. Calculate the integrated features of all neighboring nodes of the updated node and aggregate them. Multiply the aggregated neighboring node features element-wise with the features generated by the self-loop attention mechanism to obtain the final updated node features.

4. The method for targeted generation of drug molecules according to claim 3, characterized in that, The method for updating the node and edge representations in the isomorphic graph further includes: introducing information such as molecular 3D conformation, conformational dynamics, conformational space, pharmacophore features, and molecular shape to update the features of each node and each edge in the isomorphic graph.

5. The method for targeted generation of drug molecules according to claim 1, characterized in that, The method for optimizing the molecular representation pedestal model based on the intermolecular similarity in the one-dimensional molecular encoding vector is as follows: Using a multilayer perceptron as a projection head, any two sets of one-dimensional molecular coding vectors are concatenated and their shapes adjusted before being input into the projection head to obtain a prediction similarity matrix between the two sets of molecular coding vectors. The similarity between different molecules in the drug molecular structure is calculated to obtain the true similarity matrix; The error between the predicted similarity matrix and the true similarity matrix is ​​calculated, and backpropagation is performed based on the error to adjust the parameters of the molecular characterization pedestal model.

6. The method for targeted generation of drug molecules according to claim 5, characterized in that, Any element in the true similarity matrix represents one or more similarities between different molecules.

7. The method for targeted generation of drug molecules according to claim 1, characterized in that, The method for obtaining the trained molecular structure generation model is as follows: Based on the molecular structure generation model, the one-dimensional encoding vector is autoregressively decoded to calculate the predicted probability of each character token representing different molecules at the current position. Based on the predicted probability, a prediction vocabulary is constructed. The SMILES string, transformed according to the drug molecule structure, is divided into multiple character tokens. Based on the actual position information of each character token in the original SMILES string, all character tokens are summarized to form a real vocabulary. Calculate the error between the corresponding character tokens in the predicted vocabulary and the real vocabulary, and adjust the parameters of the molecular structure generation model based on the character error through backpropagation until the character error is within a preset range, thereby obtaining the trained molecular structure generation model.

8. The method for targeted generation of drug molecules according to claim 1, characterized in that, In step S2, the method for restoring the one-dimensional molecular encoding vector to the drug molecule structural formula using the trained molecular structure generation model is as follows: The one-dimensional molecular encoding vector is used as a memory vector, and the memory vector is autoregressively decoded through the Transformer decoder structure to obtain a feature vector sequence. The feature vector sequence is decoded using a multilayer perceptron to obtain the prediction score for each character; Based on the predicted score, the probability distribution value of each character appearing at the current position is calculated. According to the probability distribution value, the output character at the current position is predicted. After the first character is generated, the character is mapped to a low-dimensional embedding vector and concatenated with the previous feature vector. The concatenated vector is then input into the multilayer perceptron to predict the next character. This prediction process is repeated until the generated string meets the specified length, and finally a SMILES string composed of multiple characters is obtained. The SMILES string is converted into the drug molecular structure formula.

9. The method for targeted generation of drug molecules according to claim 1, characterized in that, The method for obtaining one or more molecular attribute prediction results through the trained molecular attribute prediction model is as follows: the one-dimensional molecular encoding vector is transformed into molecular attribute prediction values ​​by multi-level dimensional expansion, batch standardization, and activation function through a deep neural network.

10. A drug molecule directed generation system, characterized in that, For implementing the drug molecule-directed generation method as described in any one of claims 1 to 9, the drug molecule-directed generation system comprises the following modules: The molecular characterization module is used to acquire the drug molecular structure, construct a molecular characterization base model, and perform information transformation, information updating, information pooling and information integration operations on the chemical spatial information of the drug molecular structure based on the molecular characterization base model to generate a one-dimensional molecular encoding vector that can characterize molecular structure information and pharmacodynamic attribute information. Furthermore, the molecular characterization base model is optimized based on the intermolecular similarity in the one-dimensional molecular encoding vector. The decoding module is used to construct a molecular structure generation model and a molecular property prediction model. Based on the one-dimensional molecular encoding vector, the molecular structure generation model and the molecular property prediction model are trained in segments to obtain the trained molecular structure generation model and the trained molecular property prediction model. The molecular structure formula and the molecular property prediction result are obtained through the trained molecular structure generation model and the trained molecular property prediction model, respectively. And a molecular orientation optimization module, used to optimize the one-dimensional molecular encoding vector through backpropagation based on the molecular attribute prediction results of the one-dimensional molecular encoding vector, and restore the optimized one-dimensional molecular encoding vector to the molecular structure, so as to achieve the orientation optimization of drug molecules; The decoding module includes a molecular structure generation submodule and a molecular property prediction submodule. The molecular structure generation submodule restores the one-dimensional molecular code to a molecular structure formula through the trained molecular structure generation model. The molecular property prediction submodule obtains one or more molecular property prediction results based on the one-dimensional molecular code and through the trained molecular property prediction model.

Images