Methods for constructing multi-target drug molecule generation models and designing multi-target drugs

Abstract

This paper presents a method for constructing a multi-target drug molecule generation model and designing multi-target drugs, involving computer-aided drug design and bioinformatics. The method involves determining the corresponding two-dimensional molecular graph based on SMILES sequences; using a fully connected pharmacophore graph and the two-dimensional molecular graph as inputs to the GatedGCN module, where each atom node in the two-dimensional molecular graph is connected to all pharmacophore nodes in the fully connected pharmacophore graph during message passing, enabling information exchange between different nodes; using the masked SMILES sequence as input to the encoder, and the decoder generating SMILES sequences conforming to pharmacophore characteristics in an autoregressive manner; freezing the network parameters of the GatedGCN module and the encoder; training the multi-target drug molecule generation model using an elite multi-target molecule set, and back-optimizing and updating the decoder network parameters based on the output results. This method allows the model to fully learn the structural characteristics of multi-target molecules, improving the accuracy of multi-target drug molecule generation.

Description

Technical Field

[0001] This invention relates to the fields of computer-aided drug design and bioinformatics, and in particular to a method for constructing a multi-target drug molecule generation model and a method for multi-target drug design. Background Technology

[0002] Traditional drug design primarily targets a single biological target (usually a specific protein), emphasizing high specificity to avoid off-target effects and their associated adverse reactions. Clinical practice has shown that single-target drugs often have limited efficacy in treating complex diseases (such as neurodegenerative diseases and malignant tumors). While combination therapies using multiple drugs with different mechanisms of action can improve efficacy, this approach has significant limitations: it involves complex drug interactions, dose-limiting toxicities, unpredictable pharmacokinetic characteristics, and potential off-target risks, and also reduces patient adherence. Against this backdrop, multi-target drug design has become an important direction in modern innovative drug development. However, multi-target drug design faces two major challenges: first, the search space for drug molecules is extremely vast, and suitable multi-target drugs are relatively scarce; second, there is a high degree of dependence between molecular structure and its properties, making precise design of drug molecules difficult.

[0003] Current research on methods for generating multi-target drug molecules, both domestically and internationally, mainly focuses on using deep learning models (such as variational autoencoders and autoregressive models) for drug generation. These models employ variational autoencoders or autoregressive techniques, pre-trained on large-scale compound databases (such as the ChEMBL and ZINC databases) to extract chemical features and patterns. Then, multi-objective optimization algorithms (such as reinforcement learning and generative adversarial networks) are used to optimize the generated molecules, enabling them to act on multiple targets simultaneously. SMILES (Simplified Molecular Input Line Entry System) sequences are a commonly used molecular structure representation in computer-aided drug design, widely used for molecular feature extraction and model input. However, current techniques using SMILES sequences only encode and decode them without combining them with pharmacophore features for deep fusion.

[0004] However, existing methods for constructing multi-target drug molecule generation models still have shortcomings. For example, REINVENT, RationaleRL, MARS, AIxFuse, DLGN, and MTMol-GPT do not incorporate fully connected pharmacophore graphs, nor do they achieve fully connected message passing between atomic nodes and pharmacophore nodes in two-dimensional molecular graphs. This results in the models being unable to fully learn the complex relationship between molecular structure and biological activity, and the modeling of the correlation between pharmacophore features and molecular structure is not deep enough, leading to low success rates in multi-target drug molecule generation. The efficiency and quality of existing generation methods are insufficient to meet the needs of actual drug development.

[0005] Therefore, developing a method for constructing a multi-target drug molecule generation model that enables the model to fully learn the structural characteristics of multiple targets and improve the accuracy of molecule generation, as well as a multi-target drug design method based on the model, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to address the problems of existing multi-target drug molecule generation models failing to fully learn the complex relationship between molecular structure and biological activity, and having low generation success rates. This invention provides a method for constructing a multi-target drug molecule generation model and a method for designing multi-target drugs. Through deep fusion of pharmacophores and molecular features, multi-loss collaborative pre-training, parameter freeze-tuning, and dynamic elite molecule screening, the invention enhances the model's ability to learn the complex relationship between molecular structure and pharmacophores, improves the accuracy of the model in generating multi-target drug molecules, and provides technical support for the development of multi-target drugs for complex diseases.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a method for constructing a multi-target drug molecule generation model. The multi-target drug molecule generation model includes a GatedGCN module, an encoder, and a decoder. The graph embedding output by the GatedGCN module is concatenated with the latent variables output by the encoder and used as the input of the decoder. Before concatenation, position encoding is added to the latent variables, and fragment label vectors are introduced and concatenated with the output of the GatedGCN module and the latent variables respectively to avoid semantic confusion.

[0009] The method for constructing a multi-target drug molecule generation model includes a pre-training stage and a fine-tuning stage.

[0010] The pre-training phase includes:

[0011] SMILES sequences of sample molecules are obtained from a real ligand set. A corresponding two-dimensional molecular graph is determined based on the SMILES sequences. Pharmacophore features of the sample molecules are extracted from the SMILES sequences, and a fully connected pharmacophore graph is constructed based on these features. Specific pharmacophore features may include: hydrogen bond donors, hydrogen bond acceptors, hydrophobic centers, aromatic rings, positive charge centers, negative charge centers, etc. The fully connected pharmacophore graph uses pharmacophore features as nodes, and connections are established between any two pharmacophore nodes. Edge weights are determined based on the cumulative sum of the relative bond lengths of each chemical bond on the molecular topological shortest path.

[0012] A dual-graph fusion input mechanism is adopted, in which the fully connected pharmacophore graph and the two-dimensional molecular graph are jointly input into the GatedGCN module to obtain graph embedding. During message transmission, each atom node in the two-dimensional molecular graph is connected to all pharmacophore nodes in the fully connected pharmacophore graph to realize cross-type information exchange between atoms, atoms and pharmacophores, and pharmacophores. Random masking is performed on the original SMILES, and the masked SMILES sequence is used as the input of the encoder to obtain latent variables. Position encoding is added to the latent variables and concatenated. The fragment label vector is introduced and concatenated with the output of the GatedGCN module. The concatenated output of the GatedGCN module is concatenated with the concatenated encoder latent variables as the input of the decoder. The decoder generates SMILES sequences that conform to the pharmacophore characteristics step by step in an autoregressive manner.

[0013] A multi-loss collaborative optimization strategy is adopted. Based on the output of the decoder, the network parameters of the GatedGCN module, encoder and decoder are back-optimized and updated by combining the pre-training loss. The pre-training loss includes contrastive learning loss, KL divergence loss and mask modeling loss. The three losses work together to ensure that the model fully learns the correlation between molecular chemical features and pharmacophores.

[0014] The fine-tuning phase includes:

[0015] A parameter freezing mechanism is introduced to freeze the network parameters of the GatedGCN module and the encoder, while only the decoder network parameters are updatable, in order to avoid catastrophic forgetting of the model and effectively reuse the general features learned in the pre-training stage.

[0016] A dynamic elite molecule screening strategy is adopted to construct an elite multi-target molecule set. The elite multi-target molecule set is obtained from the candidate multi-target molecule set based on the molecular comprehensive score and the average probability of interaction of multiple target molecules, so as to ensure that the screened molecules have high-quality multi-target activity.

[0017] The multi-target drug molecule generation model is fine-tuned using an elite multi-target molecule set. Based on the output and fine-tuning loss during the decoder fine-tuning process, the network parameters of the decoder are back-optimized and updated to achieve accurate adaptation of the decoder to the features of multi-target molecules, thus completing the construction of the multi-target drug molecule generation model.

[0018] Furthermore, the fine-tuning stage is a multi-round mode. In each round of fine-tuning, the specific steps for constructing the elite multi-target molecule set include:

[0019] A number of molecules are randomly selected, and the pharmacophore features of these molecules are extracted to generate new molecules with the pharmacophore features. The first round of fine-tuning is performed by randomly selecting from the set of real ligands, and subsequent rounds of fine-tuning are performed by selecting from the set of candidate multi-target molecules. The generated new molecules are then deduplicated based on the Morgan fingerprint.

[0020] A comprehensive score is calculated for each new molecule. The comprehensive score is obtained by weighting the syntheticity score (SA score), the drug-likeness score (QED score), and the probability of the new molecule interacting with multiple target sites as predicted by a pre-trained random forest regression model, in order to comprehensively evaluate the molecular quality. The weighting coefficients correspond to the syntheticity score, drug-likeness score, and target interaction probability, respectively, and the sum of the weighting coefficients is 1.

[0021] Calculate the average probability of each new molecule interacting with multiple target sites, and construct a candidate multi-target molecule set from the top 50% of the new molecules with the highest average probability to preliminarily screen molecules with multi-target potential.

[0022] Within the candidate multi-target molecule set, all new molecules with a comprehensive score greater than the dynamic scoring threshold are constructed into an elite multi-target molecule set. Specifically, the dynamic scoring threshold is dynamically adjusted, increasing linearly with each fine-tuning round to gradually raise the screening criteria and further optimize the quality of generated molecules. Specific parameters can be flexibly set according to screening requirements.

[0023] Furthermore, the dimensions of the position encoding and fragment tag vector are consistent with the output dimension of the GatedGCN module and the encoder latent variable dimension.

[0024] Furthermore, the masking process of the SMILES sequence involves randomly masking a portion of the SMILES sequence, with the random masking ratio ranging from 10% to 40%. The portion of the masked sequence includes atomic symbols and chemical bond information. An identifier is also inserted at the beginning of the masked SMILES sequence, and the identifier is used to aggregate the global semantic representation.

[0025] Furthermore, the GatedGCN module includes 4 to 8 stacked GatedGCN layers. Each GatedGCN layer includes a message passing layer, a layer normalization layer, and a ReLU activation layer. The message passing layer controls the aggregation of neighbor node information through a gating mechanism and residual connections.

[0026] Furthermore, the encoder includes a Transformer encoder block, an attention pooling layer, and a multilayer perceptron connected in sequence;

[0027] The sampling steps for the latent variables of the encoder are as follows: the output of the Transformer encoder block is averaged by the attention pooling layer, the variational posterior distribution of the latent variables is calculated by the multilayer perceptron based on the output of the attention pooling layer, and the latent variables of the encoder are sampled from the variational posterior distribution.

[0028] Secondly, the present invention provides a multi-target drug design method, which uses the constructed multi-target drug molecule generation model to generate multi-target drug molecules. Specifically, it includes: extracting pharmacophore features corresponding to target targets, inputting the pharmacophore features and the target target ligand SMILES sequence into the constructed multi-target drug molecule generation model, and generating drug molecules that can interact with multiple target targets simultaneously and meet the constraints of pharmacophore features.

[0029] Furthermore, the design method may also include multiple verifications of the generated drug molecules. The verification steps include, but are not limited to, physicochemical property verification and target interaction verification. The physicochemical property verification can use cheminformatics toolkit (RDKit) tools to calculate drug-likeness score (QED), syntheticity score (SA), logP value, and molecular weight to ensure that the generated molecules meet the requirements of drug physicochemical properties. The target interaction verification can use molecular docking software AutoDock Vina to calculate the docking score between the molecule and each target site to verify the binding ability of the molecule to the target site.

[0030] Thirdly, the present invention provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for constructing a multi-target drug molecule generation model or a multi-target drug design method.

[0031] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the multi-target drug molecule generation model construction method or the multi-target drug design method.

[0032] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the multi-target drug molecule generation model construction method or the multi-target drug design method.

[0033] Compared with the prior art, the present invention has the following technical effects and outstanding advantages:

[0034] This invention introduces pharmacophore structure as prior information and guides the model to learn local and global conformational features related to drug-target interactions through a series of structure-aware generation tasks. The framework is based on the joint representation of pharmacophore and molecular SMILES sequences, employing a dual-graph fusion input mechanism to achieve deep adaptation between the two. It utilizes a Transformer architecture to learn the deep structural mapping between them and models the semantic relationship between pharmacophore and molecule through comparative learning. Combined with innovative designs such as multi-loss collaborative optimization strategies, a two-stage training architecture of "pre-training-fine-tuning," parameter freezing, and dynamic elite screening, the model fully learns the complex relationship between molecular structure and pharmacophore, significantly improving the success rate of generating multi-target drug molecules. The multi-target drug molecule generation model constructed in this invention can be directly applied to multi-target drug design. In the GSK3β|JNK3 dual-target molecule generation task, the success rate of generating dual-target molecules reached 37.84%, which is 17.5% higher than the second-best performing method (AIxFuse, 20.34%). This effectively meets the needs of actual drug development and provides efficient technical support for the development of multi-target drugs for complex diseases. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a method for constructing a multi-target drug molecule generation model in an embodiment of the present invention.

[0036] Figure 2 This is a flowchart illustrating the pre-training stage of the multi-target drug molecule generation model in an embodiment of the present invention. Wherein, a is a structural diagram of the GatedGCN module, and b is a structural diagram of the encoder.

[0037] Figure 3 This is a flowchart illustrating the fine-tuning stage of the multi-target drug molecule generation model in an embodiment of the present invention.

[0038] Figure 4 This is a schematic diagram of the fully connected pharmacophore diagram structure in an embodiment of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. Parameter settings, tool calls, data processing methods, etc., not detailed in the embodiments, can all adopt conventional technical means in the art.

[0040] The method for constructing a multi-target drug molecule generation model provided in this invention can be applied to terminals or servers. The terminal can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, etc. The server can be a standalone server, a server cluster consisting of multiple servers, or a cloud server.

[0041] Example 1: Method for constructing a multi-target drug molecule generation model

[0042] This embodiment provides a method for constructing a multi-target drug molecule generation model. This method is executed by a computer device, specifically by a computer device such as a terminal or server alone, or by a terminal and server together.

[0043] In this embodiment, the multi-target drug molecule generation model includes a GatedGCN module, an encoder, and a decoder, wherein both the encoder and decoder adopt the standard Transformer architecture; the output of the GatedGCN module and the latent variables of the encoder are concatenated and used as the input of the decoder. During the concatenation process, positional encoding is added to the latent variables, and then fragment label vectors are introduced and concatenated with the output of the GatedGCN module and the latent variables respectively to avoid semantic confusion; the feature dimension of both the positional encoding and the fragment label vector is 512.

[0044] The method for constructing multi-target drug molecule generation models includes a pre-training phase and a fine-tuning phase. Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention, wherein S110-S140 is the pre-training stage and S150-S170 is the fine-tuning stage, as detailed below:

[0045] 1. Pre-training phase

[0046] Step S110: Obtain the SMILES sequence of the sample molecule, determine the corresponding two-dimensional molecular graph based on the SMILES sequence, extract the pharmacophore features of the sample molecule from the SMILES sequence, and construct a fully connected pharmacophore graph based on the pharmacophore features; specifically:

[0047] The pre-training data came from the ChEMBL database. This database contains over 1.25 million compounds with clearly defined biological activities, and the molecules are stored in SMILES format. SMILES sequences of the sample molecules were obtained from the ChEMBL database. Based on the SMILES sequences, the corresponding two-dimensional molecular graphs were determined. Pharmacophore features of the sample molecules were extracted from the SMILES sequences, and a fully connected pharmacophore graph was constructed based on these features.

[0048] For example, SMILES sequences can be converted into corresponding two-dimensional molecular graphs using the cheminformatics toolkit RDKit. Nodes in the two-dimensional molecular graph represent atoms (e.g., C, O, N, F, etc.), and edges represent chemical bond types (e.g., single bonds are marked with "-", double bonds with "=", and triple bonds with "#"). Simultaneously, the Basefeatures.fdef module in RDKit is used to extract key pharmacophore features of the sample molecules corresponding to the SMILES sequences. These pharmacophore features include hydrogen bond donors, hydrogen bond acceptors, hydrophobic centers, aromatic rings, positive charge centers, and negative charge centers. This invention can use at least two of these types to construct a fully connected pharmacophore graph; the following explanation uses three types of features. In the constructed fully connected pharmacophore graph, nodes represent different types of pharmacophore features. Edge weights are determined based on the cumulative sum of the relative bond lengths of each chemical bond along the shortest path in the molecular topology. The relative bond length is the ratio of the actual bond length of each chemical bond to the standard C / D single bond length, and the edge weight is the cumulative value of the relative bond lengths of each chemical bond along the shortest path in the molecular topology.

[0049] See Figure 2 Taking the molecule with the SMILES sequence "CC(=O)OC1=CC=CC=C1C(=O)O" as an example, its pharmacophore features include one negative charge center (-COOH), two hydrogen bond acceptors (oxygen atoms in C=O and -O-), and one aromatic ring (benzene ring). A fully connected pharmacophore graph is constructed based on the extracted pharmacophore features. Nodes in the fully connected pharmacophore graph represent different types of chemical features, and edge weights are determined by calculating the shortest path between chemical features in the molecular topology. Specifically, this is achieved by summing the relative bond lengths of each chemical bond along the path. Figure 4 A schematic diagram of the fully connected pharmacophore structure of the above molecules is given. Figure 4 In the equation, the edge weights of the two "hydrogen bond acceptors" and "aromatic rings" are 3.7, the edge weights of the "aromatic rings" and "negative charge centers" are 4.8, the edge weights of the two "hydrogen bond acceptors" and "negative charge centers" are 6.2, and the edge weights of the two "hydrogen bond acceptors" are 7.9.

[0050] Step S120: Custom message passing connects each atom node in the two-dimensional molecular graph to all pharmacophore nodes in the fully connected pharmacophore graph, forming message passing between different types of nodes. Specifically:

[0051] In the message passing process within the GatedGCN module, message passing is implemented between atoms, atoms, and pharmacophores, as well as between pharmacophores. The core of this process is breaking down the node type barriers between the two graphs, enabling cross-graph interaction between structural features and pharmacophore features. This step is used to perform association mapping between the two-dimensional molecular graph and the fully connected pharmacophore graph. During neighborhood information aggregation, each atom node in the two-dimensional molecular graph not only aggregates local structural information from neighboring atom nodes but also integrates the semantic representation of its associated pharmacophore nodes. Taking the two-dimensional molecular graph “CC(=O)OC1=CC=CC=C1C(=O)O” as an example, it contains 10 C atom nodes and 4 O atom nodes. Each atom node is connected to three pharmacophore nodes: “hydrogen bond acceptor,” “aromatic ring,” and “negative charge center.”

[0052] In step S130, the fully connected pharmacophore map and the two-dimensional molecular map are used as inputs to the GatedGCN module, and the masked SMILES sequence is used as input to the encoder. The decoder generates SMILES sequences that conform to the pharmacophore characteristics step by step in an autoregressive manner.

[0053] In this step, the SMILES sequence needs to be masked. For example, the SMILES sequence masking process includes randomly masking a portion of the sequence, which includes atomic symbols and chemical bond information in the SMILES sequence, to construct a molecular-level self-supervised learning task; the masked SMILES sequence also has an identifier [CLS] inserted at the beginning position, which is used to aggregate the global semantic representation.

[0054] See Figure 2 In example b, the SMILES sequence masking process includes randomly masking 25% of the segments, some of which include atomic symbols (such as C, O, N, etc.) and chemical bond information (such as single bonds, double bonds, triple bonds, etc.) in the SMILES sequence to construct a molecular-level self-supervised learning task; the identifier [CLS] is also inserted at the beginning of the masked SMILES sequence to aggregate the global semantic representation. For the SMILES sequence “CC(=O)OC1=CC=CC=C1C(=O)O”, the masked result is “[CLS]CC(=O)OC1=CC=CC=[MASK]O”.

[0055] Reference Figure 2 Part a, specifically, the GatedGCN module includes multiple stacked GatedGCN layers (corresponding to...). Figure 2(Pharmacophore convolution and atomic graph convolution in GatedGCN). For fully connected pharmacophore graphs and two-dimensional molecular graphs, a multi-layer stacked GatedGCN is used as an encoder to extract node feature information from the graph. The GatedGCN module includes a message passing layer and a layer normalization & ReLU activation layer. The message passing layer controls the flow of neighbor node information in message passing through gating mechanisms and residual connections, including three types of information: atom-to-atom, atom-to-pharmacophore, and pharmacophore-to-pharmacophore. It can also simultaneously fuse node and edge attributes to generate graph embeddings. The graph embeddings are then processed by the normalization & activation layer (layer normalization & ReLU activation function) to generate the output of the GatedGCN module (dimensionality of 512). The number of stacked GatedGCN layers can be adjusted according to actual needs. Too few layers cannot fully extract features, while too many layers can easily lead to overfitting. Generally, 4 to 8 layers are set, and 6 layers are used in this embodiment.

[0056] In the message passing layer, let Represents graph nodes In the The feature vector of the layer, , Represents the d-dimensional real space, For nodes The set of neighboring nodes, and the update process of node features are as follows:

[0057]

[0058] in, Represents graph nodes In the The updated feature vector after +1 layer Represents the ReLU activation function. This indicates a batch normalization operation. It is the first Layer dimensional learnable parameter matrix, Represents the set of real numbers. This represents the Hadema product. It is the first Layer nodes With nodes The gating unit for the edge features is defined as follows:

[0059]

[0060]

[0061] in, It is the sigmoid activation function. Indicates the first Layer nodes With nodes The original gating value of the edge features between them. It is a fixed constant for maintaining numerical stability, with a value of 0.05. It is traversing nodes The index variable when using neighbor set. For nodes The set of neighboring nodes, Indicates the first Layer nodes With nodes The original gating value of the edge features between them. Represents the ReLU activation function. This indicates a batch normalization operation; Indicates the third Layer dimensional learnable parameter matrix; Indicates the first Layer diagram nodes eigenvectors, This is the index of the layer number in the graph convolution; , The index of the graph node; the encoded node features ,in The number of pharmacophore nodes. The number of convolutional layers is specified. The graph convolution method described above is used for all three relationships: atom-to-atom, atom-to-pharmacophore, and pharmacophore-to-pharmacophore, to achieve deep fusion of structural and functional information. To extract global feature information from the heterogeneous graph, an average pooling layer is used to calculate the mean of all node features in the graph, which is then used as the global feature of the heterogeneous graph (feature dimension 512).

[0062] Reference Figure 2 Specifically, part b of the encoder comprises a Transformer encoder block, an attention pooling layer, and a multilayer perceptron connected in sequence. The number of layers in the Transformer encoder block can be adjusted between 6 and 10 to accommodate different molecular sequence lengths. Too few layers cannot adequately capture sequence semantics, while too many layers increase computational costs. The sampling steps for the latent variables of the encoder include: average pooling the output of the Transformer encoder block through the attention pooling layer; calculating the variational posterior distribution of the latent variables using the multilayer perceptron based on the output of the attention pooling layer; and sampling the latent variables from the variational posterior distribution.

[0063] For the SMILES sequence, both the encoder and decoder employ a standard Transformer architecture. The encoder consists of eight stacked Transformer encoder blocks, followed by an attention pooling layer and a multilayer perceptron. The multilayer perceptron has three layers, each with 256 neurons. Each encoding block contains two sub-layers: a multi-head self-attention sub-layer and a feedforward fully connected sub-layer. Each sub-layer is followed by a layer normalization layer and a residual connection. The preprocessed SMILES sequence (masking and insertion of a start identifier [CLS]) is first encoded into high-dimensional sequence features (1024 dimensions) through an embedding layer, and after adding positional encoding, it is fed into the encoder. After passing through multiple Transformer encoder blocks, the molecular features are averaged through an attention pooling layer to obtain the final molecular representation. ; then Input two independent linear mapping layers (multilayer perceptron) to compute the mean of the variational posterior distribution. and variance Then from the normal distribution Medium-sampled latent variables (dimension size 512), normally distributed The variational posterior distribution of the latent variable is given.

[0064] The decoder employs a Transformer architecture, concatenating the output of the GatedGCN module with latent variables (positional encoding is added to the latent variables before concatenation, and fragment marker vectors are introduced to identify pharmacophore and sequence features to avoid semantic confusion). It then progressively generates SMILES strings that conform to pharmacophore features using an autoregressive approach. For the masked sequence "[CLS]CC(=O)OC1=CC=CC=[MASK]O", the decoder can accurately predict the characters at the mask positions, ultimately generating the complete SMILES string "CC(=O)OC1=CC=CC=C1C(=O)O", which is completely identical to the original sequence, achieving a 100% match. During the generation process, each step is filtered according to chemical rationality rules (such as atomic valence rules and chemical bond formation rules) to ensure that the output molecule satisfies both chemical rationality and possesses the expected pharmacophore features.

[0065] Step S140: Based on the decoder output, the network parameters of the GatedGCN module, encoder, and decoder are updated in reverse optimization.

[0066] In the decoding stage of molecular generation, the latent variable representation (output of the encoder) and the prior information of the pharmacophore (output of the GatedGCN module) are integrated by the decoder, thereby guiding the generation process toward the structural space with the target pharmacophore.

[0067] Specifically, latent variables sampled from the latent space serve as a compressed representation of the molecular global semantics. To enhance the model's ability to model sequence structure, positional encodings are added to the latent variables before they are input to the decoder. Simultaneously, pharmacophore features, as prior structural constraint information, are also fed into the decoder along with the latent variables. To avoid semantic confusion with the latent variables, fragment label vectors are introduced and concatenated with both the pharmacophore features and the latent variables. The feature dimensions of both the positional encodings and the fragment label vectors are 512.

[0068] The latent variables after introducing fragment label vectors are first modeled for features by the Transformer encoder block in the decoder; then, they are fused with the pharmacophore features after introducing fragment label vectors in the Transformer decoder block; finally, the hidden state output by the Transformer decoder block is mapped to the specific molecular generation result, i.e., the SMILES sequence that conforms to the pharmacophore features, through a multilayer perceptron.

[0069] The decoder generates SMILES sequences that conform to pharmacophore features stepwise in an autoregressive manner. Each generation step fully considers the degree of matching between the context information of the generated part and the target pharmacophore features, ensuring that the output molecule satisfies both chemical rationality and has the expected pharmacophore features.

[0070] The above describes the model input and operation process during pre-training. During training, it's also necessary to set an appropriate training loss to drive the model to back-optimize and update network parameters.

[0071] Specifically, the pre-training loss of multi-target drug molecule generation models Including contrastive learning loss KL divergence loss and mask modeling loss , represented as: .in:

[0072] Comparative learning loss for:

[0073]

[0074] in, It is a balance coefficient used to balance the loss ratio in both the sequence-to-pharmacophore and pharmacophore-to-sequence directions; Represents the mathematical expectation operator. Indicates the distribution of data All sample pairs Take the expected value; Indicates sample pairs From the joint probability distribution obtained from sampling, Encoder output of sample molecules, The output of the GatedGCN module for sample molecules. For sample pairs The joint probability distribution, The cross-entropy function, This represents the similarity label value from sequence features to pharmacophore features. This represents the similarity inference value from sequence features to pharmacophore features. This represents the similarity label value from pharmacophore features to sequence features. This represents the similarity inference value from pharmacophore features to sequence features.

[0075] For any pair of samples The inferred value of similarity is:

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082] in, This represents the similarity inference value from sequence features to pharmacophore features. This represents the similarity inference value from pharmacophore features to sequence features. It is an exponential function. For similarity function, The encoder output (sequence features) for the sample molecules. Output of the GatedGCN module for sample molecules (pharmacophore features). For the first The output of the GatedGCN module for the nth sample molecule (the nth sample molecule) (characteristics of individual pharmacophores) For the first The encoder output of the nth sample molecule (the nth sample molecule) (sequence features) A learnable temperature parameter (used for scaling adjustments of similarity). This represents the total number of sample molecules involved in a single loss calculation. For fully connected layers used to process sequence features, For processing pharmacophore features, a fully connected layer, This is a matrix transpose operation.

[0083] In calculating the contrastive learning loss, it is necessary to first calculate the loss for each pair of samples. Similarity inference value from sequence features to pharmacophore features Similarity inference value from pharmacophore features to sequence features In a training batch, positive sample pairs are designed as the SMILES sequence of the same molecule and its corresponding pharmacophore feature, aiming to maximize the similarity of cross-modal features; negative sample pairs are constructed using an intra-batch negative sampling strategy, that is, for any sample in the batch, its SMILES sequence is paired with the pharmacophore features of all other molecules in the batch as negative sample pairs. For any pair of samples... Similarity inference value from sequence features to pharmacophore features The calculation involves, on the one hand, calculating the encoder output of the sample molecule. Output to the GatedGCN module of this sample molecule similarity On the other hand, the encoder output of the sample molecule is calculated. To the respective The output of the GatedGCN module for each sample molecule similarity For any pair of samples Similarity inference value from pharmacophore features to sequence features The calculation involves, on the one hand, calculating the output of the GatedGCN module for the sample molecule. Encoder output to the sample molecule similarity On the other hand, the output of the GatedGCN module is calculated for this sample molecule. To the respective Encoder output of each sample molecule similarity The similarity value of negative sample pairs is 0, and the similarity value of positive sample pairs is 1.

[0084] The KL divergence loss is: ,in, It follows a standard normal distribution. The variational posterior distribution of the latent variable is given.

[0085] To enable the decoder to generate text, this embodiment uses language modeling loss (LM loss) as the mask modeling loss to train the model. It is an optimized cross-entropy loss that maximizes text likelihood in an autoregressive manner. Mask Modeling Loss for:

[0086]

[0087] in, The length of the SMILES sequence. For index variables, Indicates the SMILES sequence number 1 One word; Indicates the SMILES sequence number 1 All words preceding the given word, i.e., positions 1 to 1. -1 character set For the output of the decoder, for the SMILES sequence, the 1st... The word is The predicted probability. This metric directly reflects how well the model fits the real SMILES sequences; a smaller value indicates a more accurate prediction. The Adam optimizer was used to train the model, with a learning rate of 3e-4 and a weight decay rate of 1e-6. Cosine learning rate annealing was employed with a cycle length of four epochs. Gradient pruning was also used, with the maximum gradient set to 5. The pre-training run consisted of 20 epochs with a batch size of 512, saving the weights at which the model performed best on the validation set.

[0088] 2. Fine-tuning stage:

[0089] Step S150: Freeze the network parameters of the GatedGCN module and encoder.

[0090] During the fine-tuning phase, a parameter freezing mechanism is adopted, which does not update the network parameters of the GatedGCN module and the encoder, but only optimizes and updates the network parameters of the decoder, in order to avoid catastrophic forgetting of the model, while reusing the general features learned in the pre-training phase.

[0091] Step S160: Construct an elite multi-target molecule set, which is obtained by screening molecules that interact with multiple target molecules.

[0092] In this embodiment, the fine-tuning phase consists of 20 rounds. In each round, constructing the elite multi-target molecule set specifically includes: randomly selecting several molecules from the real ligand set, extracting the pharmacophore features of these molecules, and generating new molecules with pharmacophore features; calculating the comprehensive score of each new molecule, which is weighted by a syntheticity score, a drug-likeness score, and the probability of the new molecule interacting with multiple target targets as predicted by a pre-trained random forest regression model; calculating the average probability of each new molecule interacting with multiple target targets, and constructing the top 50% of the new molecules with the highest average probability as a candidate multi-target molecule set; and constructing the elite multi-target molecule set from all new molecules in the candidate multi-target molecule set whose comprehensive scores are greater than the dynamic scoring threshold. For example, it should be noted that, except for the first round of fine-tuning, several molecules can be randomly selected from the candidate multi-target molecule set generated in the previous round of fine-tuning. For the first round of fine-tuning, several molecules can be randomly selected from the real ligand set of multiple target targets.

[0093] It should be noted that the dynamic scoring threshold is dynamically adjusted, and the threshold increases linearly with the number of fine-tuning rounds. For example, the initial value of the dynamic scoring threshold is 0.6, and it increases by 0.05 every 5 fine-tuning rounds, with an upper limit of 0.85. This can gradually improve the quality of elite molecules while ensuring molecular diversity.

[0094] It should be noted that the weighting parameters are set differently in different multi-target molecule generation tasks. The weights are adjusted according to the quality adaptability of the generated molecules in each round, while ensuring that the sum of the weights equals 1. For example, the weighting coefficients are set as follows: syntheticity (SA) score weight 0.3, drug-likeness (QED) score weight 0.2, probability weight of new molecule interacting with multiple target sites weight 0.5, and the total weight is 1.

[0095] Step S170: Train the decoder of the multi-target drug molecule generation model using an elite multi-target molecule set, and back-optimize and update the decoder's network parameters based on the decoder's output.

[0096] Specifically: Using the SMILES sequences of each molecule in the elite multi-target molecule set as training data, the autoregressive generation process of the decoder in the pre-training stage is followed. The pharmacophore features of the elite molecules are input into the GatedGCN module (parameters frozen, used only for feature extraction), and the SMILES sequences of the elite molecules are input into the encoder after masking (parameters frozen, used only for generating latent variables). The features of the two are concatenated and then sent to the decoder for training. The training epochs are set to 5. During training, based on the deviation between the SMILES sequences generated by the decoder and the real SMILES sequences in the elite multi-target molecule set, backpropagation of loss modeling through masking is used to optimize and update only the network parameters of the decoder, without changing any parameters of the GatedGCN module and the encoder. It should be noted that the fine-tuning training process is consistent with the data processing process in the pre-training stage, only the data object is different. In the pre-training stage, the data processing object is the SMILES sequences of sample molecules, while in the fine-tuning training stage, the data processing object is the SMILES sequences of elite multi-target molecules.

[0097] Steps S160 and S170 constitute one round of fine-tuning, and multiple rounds of fine-tuning are used to optimize the network parameters of the decoder.

[0098] For example, multiple targets can be dual targets. Taking dual targets A and B as an example, before fine-tuning begins, two globally shared dynamic molecular sets can be constructed: a candidate dual-target molecular set and an elite dual-target molecular set. The candidate dual-target molecular set is used to store dual-target molecules generated in each round of fine-tuning, initially consisting of the true ligand set of the two target molecules; the elite dual-target molecular set is used to store high-quality, high-scoring dual-target molecules, initially empty.

[0099] Reference Figure 3 A round of fine-tuning can be broken down into the following sub-steps:

[0100] (1) Sampling. 10,000 molecules were randomly selected from the candidate dual-target molecule set (the first round of fine-tuning was performed to select the actual ligands for the two target molecules). Figure 3 The pharmacophore features of target A and target B are extracted to construct conditional inputs, which then guide the decoder to generate new molecules with specific pharmacophore distributions.

[0101] (2) Deduplication and scoring. For the generated new molecules, deduplication is performed based on the Morgan fingerprint. The fingerprint radius is set to 2 and the number of bits is set to 1024 to prevent duplicate samples from causing sorting bias and to ensure sample diversity in subsequent stages. Next, the comprehensive score of the generated new molecules is evaluated. A multi-objective weighted scoring function r is constructed, with a value range of 0 to 1, where 0 represents the lowest quality and 1 represents the best quality. r is composed of multiple sub-items, specifically including: syntheticity (SA) score, drug-likeness (QED) score, and the probability that the new molecule will interact with target A and target B, respectively, as predicted by the pre-trained random forest regression model (including target A interaction predictor and target B interaction predictor). and The QED score and SA score were calculated using the RDKit tool and obtained after half-Gaussian scaling. The tools for generating the multi-objective weighted scoring function r include target A interaction predictors, target B interaction predictors, and RDKit. This stage can then generate score information for each new molecule.

[0102] (3) Ranking. All new molecules are considered as candidates, and then the generated new molecules are ranked in descending order based on the mean of their probability values. If a new molecule has a high probability of interacting with both targets, then its ranking is higher; otherwise, its ranking is lower. Specifically, if a molecule has a probability of interacting with a target of not less than 0.5, then it has the potential to interact with that target. Similarly, if a molecule has a probability of interacting with both targets of not less than 0.5, then it has the potential to interact with both targets simultaneously. The top 50% of the ranked new molecules are considered as candidate dual-target molecules (corresponding to...). Figure 3 Candidate molecules were selected and added to the candidate dual-target molecule pool for subsequent sampling.

[0103] (4) Selection. To further improve molecular quality, this embodiment uses a dynamic scoring threshold t (exemplarily, the dynamic scoring threshold t is initialized to 0.6 and increased by 0.05 every 5 rounds) to perform secondary screening on candidate dual-target molecules. Those with a comprehensive score higher than the dynamic scoring threshold t are regarded as "elite dual-target molecules" and added to the elite dual-target molecule set.

[0104] The above are the specific steps for constructing an elite multi-target molecular set.

[0105] (5) Optimization. The decoder is back-optimized and trained using an elite multi-target molecular set.

[0106] The aforementioned fine-tuning aims to stabilize the generalized representation of the latent space and prevent catastrophic forgetting caused by over-updating. The decoder is repeatedly trained on an elite multi-target molecular set multiple times (e.g., 5 times), using a combination of mapping loss and language modeling loss (mask modeling loss) as the loss function. The training parameters can be set in the same way as in the pre-training phase.

[0107] It should be noted that the above example uses a dual-target target, but in reality, there can be more target targets. For example, for a triple-target target, the candidate dual-target molecules are the top 50% ranked in descending order of the mean probability of interacting with all three target targets simultaneously.

[0108] The loss during the fine-tuning phase can be specifically expressed as: .in, The language modeling loss (mask modeling loss) is the mask modeling loss during the pre-training stage, which can be referred to the explanation above. To further enhance the responsiveness of the model-generated molecules to the features of the target pharmacophore, a mapping loss between the pharmacophore and the SMILES sequence is introduced during model training. This loss term aims to explicitly constrain whether the generated molecules effectively meet the structural requirements of a given pharmacophore, thereby achieving precise alignment between the pharmacophore prior and the molecular structure. The mapping loss is defined as follows: for:

[0109]

[0110] in, and These are two learnable matrices, each with a dimension of 512×512, used to project two different embedding vectors into the same space. For the sigmoid function, This represents the output of the decoder. This indicates the output of the GatedGCN module.

[0111] Example 2: Multi-target drug design method

[0112] In one exemplary embodiment, a multi-target drug design method is also provided, comprising: obtaining a real ligand set of multiple target targets (data sourced from public databases such as the ZINC database, the ChEMBL database, and open-source research, statistically surveying target combinations with research value in the current field of multi-target drug design, and collecting drug molecule data that have been clinically used), using the real ligand set as input to a multi-target drug molecule generation model, and obtaining drug molecules that can interact with multiple target targets through the multi-target drug molecule generation model, wherein the multi-target drug molecule generation model is obtained by the multi-target drug molecule generation model construction method in Embodiment 1 above.

[0113] The generated drug molecules need to undergo multiple verifications:

[0114] Physicochemical property verification: The physicochemical properties of the generated molecules were calculated using the RDKit tool, and the distribution of four attribute indicators of the generated molecules, namely drug-likeness fraction (QED), syntheticity fraction (SA), logP, and molecular weight, were statistically analyzed.

[0115] Target-target interaction verification: The screened molecules were validated for target-target docking using the molecular docking software AutoDock Vina. The docking scores were calculated and compared with the actual binding complexes of ligands and targets.

[0116] Comparative verification: The model of this invention is compared with existing technologies (variational autoencoder VAE, traditional autoregressive model AR). Evaluation indicators include the effectiveness, uniqueness, novelty, diversity, drug-likeness score (QED), syntheticity score (SA), and generation success rate (the proportion of molecules with multi-target activity) of the generated molecules, to verify the technical advantages of the model of this invention.

[0117] Example 3: Experimental Verification

[0118] To verify the effectiveness and superiority of the multi-target drug molecule generation model construction method and multi-target drug design method provided by this invention, this embodiment uses GSK3β|JNK3 dual-target drug design as the experimental task and conducts the following comparative experiments:

[0119] Targeting the therapeutically significant combination of GSK3β and JNK3, 10,000 candidate molecules were generated for each comparative method. All experiments were conducted in the same computational environment, following the model construction method described in Example 1 and the drug design method described in Example 2 of this invention, to ensure the fairness of the comparison.

[0120] Table 1. Performance of different methods on the GSK3β|JNK3 dual-target drug design task.

[0121]

[0122] Bold text indicates that this method is the best among all methods.

[0123] The underscore indicates that this method is the second best among all methods.

[0124] First, the molecular generation was evaluated based on three key metrics: effectiveness, uniqueness, and diversity. As shown in Table 1, the effectiveness of molecules generated by most methods was greater than 98%. Regarding uniqueness, the molecules generated by this invention achieved 100% uniqueness, significantly outperforming other methods, indicating that there were absolutely no duplicate molecules in the generated results.

[0125] In terms of effectiveness indicators, MARS, AIxFuse, and the method of this invention all achieved 100% effectiveness, indicating that the molecules generated by this invention can well meet the requirements of chemical rationality and basic physicochemical properties.

[0126] In terms of uniqueness, both this invention and MTMol-GPT achieve 100% uniqueness, which is significantly better than other comparative methods. This indicates that there are no duplicate molecules among the 10,000 candidate molecules generated by this invention, which can effectively avoid the efficiency loss caused by repeated screening.

[0127] In terms of novelty index, the novelty of this invention compared with REINVENT, RationaleRL, MARS, and MTMol-GPT all reached 100%, indicating that the molecules generated by this invention are all novel structures, do not repeat known compounds in existing public databases, and have potential innovative drug development value.

[0128] In terms of diversity, the diversity of this invention reaches 71.22%, indicating that the molecular structures generated by this invention have large differences and can cover a wider range of molecular structure spaces, providing more diverse candidate molecules for subsequent drug screening.

[0129] In terms of the key success rate metric, the model constructed in this invention achieved an excellent performance of 37.84%, which is far higher than other comparative methods. Compared with the second-place AIxFuse (20.34%), it represents a relative improvement of 17.5 percentage points.

[0130] To further verify the practical application value of the molecules generated in this invention, target binding capability verification was conducted. Notably, the molecules generated in this invention also exhibit outstanding performance in target binding. In molecular docking evaluations targeting GSK3β and JNK3, the docking scores of the molecules generated in this invention ranked first and second, respectively, among all comparative methods; the average docking score also outperformed all baseline models. Particularly on the GSK3β target, the docking score of the molecules generated in this invention significantly surpassed all baseline methods, demonstrating stronger binding potential.

[0131] Experiments show that the model constructed in this invention can more stably generate molecules with potential activity against the GSK3β|JNK3 dual target sites.

[0132] Example 4: Computer Equipment

[0133] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0134] The computer has the following hardware configuration: Intel(R) Core(TM) i7-6850K CPU @3.60GHz processor, 128GB of memory, 2TB of hard drive, two NVIDIA GeForce RTX 3090 24GB graphics cards, and Ubuntu 23.04 operating system, supporting efficient operation of model training and drug design processes.

[0135] Example 5: Computer-readable storage medium

[0136] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0137] The storage medium can be any medium capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, facilitating the storage, migration, and deployment of model algorithms.

[0138] Example 6: Computer Program Product

[0139] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0140] The computer program product is packaged as an installation package and supports installation and deployment on operating systems such as Windows, macOS, and Linux. The computer program product contains three core modules: a model training module, a drug generation module, and a result verification module. The model training module implements parameter configuration, loss calculation, and parameter optimization for pre-training and fine-tuning. The drug generation module implements molecule generation and screening. The result verification module implements verification of the physicochemical properties of the generated molecules and their interactions with the target.

[0141] This invention provides a method for constructing multi-target drug molecule generation models and for designing multi-target drugs, and its applicable number of targets is not limited to two. In addition to the dual-target example mentioned above, this invention is also applicable to multi-target drug design scenarios with three or more targets.

[0142] When the number of target points is expanded to three or more, the method flow of this invention remains consistent with the dual-target case, with only corresponding expansions in the number of target points and related parameter settings. Specifically, let the number of target points be... N For any newly generated molecule, the pre-trained N Each random forest regression model predicts the probability of the molecule interacting with each target site, thus yielding the corresponding... N Each interaction probability value is defined as an interaction probability of the new molecule. The mean interaction probability is defined as the arithmetic mean of the interaction probabilities of the molecule at all target sites, reflecting the overall level of the molecule's interaction with multiple targets.

[0143] During the comprehensive score calculation, when the number of targets increases, only the interaction probability term of the corresponding targets needs to be included in the weighted calculation. The weight parameters can be set to the same or different weights according to the importance of the targets, and the sum of the weights remains 1. The calculation methods for syntheticity score and drug-likeness score remain unchanged and do not require additional adjustment.

[0144] In the candidate molecule screening stage, for application scenarios with three or more targets, all new molecules are also ranked according to the mean interaction probability, and the top 50% of the ranked molecules are selected to construct a candidate multi-target molecule set. Through this method, the method of the present invention can be naturally extended to any number of target targets without substantial modification to the model structure or generation process.

[0145] As explained above, the multi-target drug molecule generation model construction method in this embodiment introduces pharmacophore structure as prior information and guides the model to learn local and global conformational features related to drug-target interactions through a series of structure-aware generation tasks. This framework is based on the joint representation of pharmacophore and molecular SMILES sequences, utilizes the Transformer architecture to learn the deep structural mapping between them, and models the semantic relationship between pharmacophore and molecule through contrastive learning. This allows the model to fully learn the complex relationship between molecular structure and pharmacophore, thereby improving the accuracy of the model in generating multi-target drug molecules.

[0146] Furthermore, in the fine-tuning stage, the reward mechanism of reinforcement learning and the sample selection strategy of active learning are innovatively integrated: the candidate molecules generated by the pre-trained model are evaluated by a multi-objective weighted scoring function (covering indicators such as binding affinity and drug-likeness), and high-quality samples are selected from them for the parameter update of the model.

[0147] This invention overcomes the technical bottleneck of existing multi-target drug molecule generation models that lack deep integration of pharmacophore features and molecular structure features, resulting in insufficient modeling accuracy. Through an integrated design of "feature fusion - two-stage training - precise optimization," it achieves efficient and high-precision generation of multi-target drug molecules. The core innovation of this invention lies in:

[0148] 1. An innovative dual-graph fusion input mechanism of "fully connected pharmacophore graph - two-dimensional molecular graph" is constructed. The message passing logic of the GatedGCN module is customized to realize the full-connection information interaction between each atomic node in the two-dimensional molecular graph and all pharmacophore nodes in the fully connected pharmacophore graph. This breaks the limitation of existing technologies where atomic nodes only aggregate information of adjacent atoms and pharmacophores are disconnected from molecular structures. The model can capture the local and global topological structure and pharmacophore features of molecules at the same time, and deeply learn the complex relationship between molecular structure and biological activity, providing accurate structure-pharmacophore constraints for the generation of multi-target molecules.

[0149] 2. A multi-loss collaborative pre-training strategy is designed, which organically combines contrastive learning loss, KL divergence loss, and mask modeling loss. The contrastive learning loss innovatively models the bidirectional semantic mapping relationship between SMILES sequence features and pharmacophore features, the KL divergence loss constrains the rationality of the latent variable distribution, and the mask modeling loss ensures the accuracy of molecular sequence generation. The three work together to enable the model to fully learn the correlation between molecular chemical features and pharmacophores, thus solving the problem that existing pre-training only focuses on a single feature and has weak model generalization ability.

[0150] 3. A two-stage training architecture of "pre-training-fine-tuning" is proposed, and a parameter freezing mechanism and dynamic elite molecule screening strategy are innovatively introduced: In the pre-training stage, the parameters of the GatedGCN module, encoder and decoder are comprehensively optimized to consolidate the model's general feature learning ability; In the fine-tuning stage, the parameters of the GatedGCN module and encoder are frozen, and only the decoder is optimized to avoid catastrophic forgetting. At the same time, through multi-index weighted scoring (synthetizability, drug-likeness, target interaction probability) and dynamic scoring threshold adjustment, an elite multi-target molecule set is constructed to achieve accurate adaptation of the decoder to the features of multi-target molecules, significantly improving the targeting and quality of generated molecules.

[0151] 4. Innovatively, the model structure is deeply adapted to the needs of multiple targets. The pharmacophore feature output of the GatedGCN module is concatenated with the latent variables of the SMILES sequence of the encoder and then input into the decoder. Fragment label vectors and positional encoding are introduced to avoid semantic confusion. The decoder generates molecules in an autoregressive manner combined with chemical rationality rules. At the same time, mapping loss is introduced in the fine-tuning stage to enhance the accurate alignment between the pharmacophore and the molecular structure. Finally, the model fully learns the characteristics of multi-target molecular structures. In the dual-target molecule generation task, the success rate is improved by 17.5 percentage points compared with the existing best method (AIxFuse), breaking through the dilemma that the generation efficiency and quality of the existing technology cannot meet the actual research and development needs.

[0152] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations.

[0153] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0154] The databases involved in the various embodiments provided by this invention may include at least one type of relational database and non-relational database. The processors involved in the various embodiments provided by this invention may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited thereto.

[0155] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0156] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for constructing a multi-target drug molecule generation model, characterized in that, The constructed multi-target drug molecule generation model includes a GatedGCN module, an encoder, and a decoder; the method includes a pre-training phase and a fine-tuning phase, as detailed below: Pre-training phase: The SMILES sequence of the sample molecule is taken, and a two-dimensional molecular graph is determined based on the SMILES sequence. Based on at least two of the six chemical features identified from the SMILES sequence, namely hydrogen bond donor, hydrogen bond acceptor, hydrophobic center, aromatic ring, positive charge center and negative charge center, a fully connected pharmacophore graph is constructed, wherein the edge weight between any two pharmacophore nodes is determined based on the sum of the relative bond lengths of each chemical bond on the shortest path of molecular topology. The fully connected pharmacophore graph and the two-dimensional molecular graph are input into the GatedGCN module. The GatedGCN module includes 4 to 8 stacked GatedGCN layers. Each GatedGCN layer includes a message passing layer, a layer normalization layer, and a ReLU activation layer. The message passing layer controls the aggregation of neighbor node information through a gating mechanism and residual connections. The message passing layer also supports information exchange between three types of nodes: atom-atom, atom-pharmacophore, and pharmacophore-pharmacophore. During the message passing process of the GatedGCN module, each atom node in the two-dimensional molecular graph is connected to each pharmacophore node in the fully connected pharmacophore graph, realizing cross-type information exchange between three types of nodes: atom-atom, atom-pharmacophore, and pharmacophore-pharmacophore, to obtain a graph embedding that incorporates pharmacophore priors; After masking the SMILES sequence, the latent variables are input into the encoder. After position encoding and fragment label vector concatenation, the latent variables are input into the decoder along with the graph embedding. The decoder generates the SMILES sequence in an autoregressive manner. The network parameters of the GatedGCN module, encoder, and decoder are updated by combining the pre-training loss with reverse optimization. The pre-training loss is the sum of contrastive learning loss, KL divergence loss, and mask modeling loss. Fine-tuning phase: Freeze the network parameters of the pre-trained GatedGCN module and encoder, and only keep the decoder parameters that can be updated; A number of molecules are randomly selected, their pharmacophore features are extracted, and new molecules with the pharmacophore features are generated. The mean probability of each new molecule interacting with multiple target sites is calculated, and the new molecules with the top 50% of the mean probability are constructed as a candidate multi-target molecule set. Constructing an elite multi-target molecule set: From the candidate multi-target molecule set, molecules interacting with multiple target targets are initially ranked based on the interaction probabilities predicted by a pre-trained random forest regression model. All new molecules with a comprehensive score greater than a dynamic scoring threshold constitute the elite multi-target molecule set. The comprehensive score is a weighted sum of the molecule's syntheticity score, drug-likeness score, and interaction probabilities. The dynamic scoring threshold is dynamically adjusted, increasing linearly with each fine-tuning round. The model is fine-tuned using the elite multi-target molecule set, and the decoder network parameters are updated based on the fine-tuning loss to complete the construction of the multi-target drug molecule generation model.

2. The method for constructing a multi-target drug molecule generation model according to claim 1, characterized in that, During the pre-training phase, the masking process of the SMILES sequence includes randomly masking a portion of the SMILES sequence, the portion of which includes atomic symbols and chemical bond information in the SMILES sequence; an identifier is also inserted at the start position of the masked SMILES sequence, the identifier being used to aggregate the global semantic representation.

3. The method for constructing a multi-target drug molecule generation model according to claim 1, characterized in that, The encoder includes a Transformer encoder block, an attention pooling layer, and a multilayer perceptron connected in sequence. The sampling steps for the latent variables of the encoder are as follows: the output of the Transformer encoder block is averaged by the attention pooling layer, the mean and variance of the latent variables are calculated by the multilayer perceptron based on the output of the attention pooling layer to determine the variational posterior distribution, and the latent variables of the encoder are sampled from the variational posterior distribution.

4. The method for constructing a multi-target drug molecule generation model according to claim 1, characterized in that, The contrastive learning loss is: in, It is a balance coefficient used to balance the loss ratio in both the sequence-to-pharmacophore and pharmacophore-to-sequence directions; Represents the mathematical expectation operator. Indicates the distribution of data All sample pairs Take the expected value; Indicates sample pairs From the joint probability distribution obtained from sampling, The encoder output for the sample molecules, The output of the GatedGCN module for sample molecules. For sample pairs The joint probability distribution, The cross-entropy function, This represents the similarity label value from sequence features to pharmacophore features. This represents the similarity inference value from sequence features to pharmacophore features. This represents the similarity label value from pharmacophore features to sequence features. This represents the similarity inference value from pharmacophore features to sequence features; For any pair of samples The inferred value of similarity is: in, For similarity function, It is a learnable temperature parameter. The total number of sample molecules mentioned in a single loss calculation. Encoder output of sample molecules, For the first Encoder output for each sample molecule, The output of the GatedGCN module for sample molecules. For the first Output of the GatedGCN module for each sample molecule; It is an exponential function; The KL divergence loss for: in, It follows a standard normal distribution. The variational posterior distribution for latent variables; The mask modeling loss for: in, The length of the SMILES sequence. For index variables, Indicates the SMILES sequence number 1 One word; Indicates the SMILES sequence number 1 All words preceding the given word, i.e., positions 1 to 1. The character set of -1 For the output of the decoder, for the SMILES sequence, the 1st... The word is The predicted probability.

5. The method for constructing a multi-target drug molecule generation model according to claim 1, characterized in that, The fine-tuning phase has multiple rounds, and in each round of fine-tuning, the specific steps for constructing the elite multi-target molecule set include: The molecules for the first round of fine-tuning were randomly selected from the set of real ligands, and the molecules for subsequent rounds of fine-tuning were selected from the set of candidate multi-target molecules; the generated new molecules were deduplicated based on Morgan fingerprints. Calculate a comprehensive score for each new molecule, which is obtained by weighting a syntheticity score, a drug-likeness score, and the probability of the new molecule interacting with multiple target sites as predicted by a pre-trained random forest regression model. Each weighting coefficient is a value between 0 and 1, and the sum of all weighting coefficients is 1.

6. The method for constructing a multi-target drug molecule generation model according to claim 1, characterized in that, The fine-tuning loss is the sum of the mapping loss and the mask modeling loss; whereby the mapping loss is: in, and These are two learnable matrices used to project two different embedding vectors into the same space. For the sigmoid function, and These represent the outputs of the decoder and the GatedGCN module, respectively.

7. A multi-target drug design method, characterized in that, The multi-target drug molecule generation model constructed using the method described in any one of claims 1-6 shall be subjected to the following steps: Extract the pharmacophore features corresponding to the target site, and input the pharmacophore features and the target target ligand SMILES sequence into the model to generate a drug molecule that can interact with multiple target sites simultaneously and conforms to the pharmacophore feature constraints.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for constructing a multi-target drug molecule generation model according to any one of claims 1-6, or the method for designing a multi-target drug according to claim 7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for constructing a multi-target drug molecule generation model as described in any one of claims 1-6, or the method for designing a multi-target drug as described in claim 7.

Images