Molecular generation method and system driven by interaction perception comparison coding and multi-modal pocket characteristics

Abstract

The invention discloses an interaction perception comparison coding and multi-mode pocket feature driven molecule generation method and system. The method comprises the following steps: firstly, constructing a pre-training and fine-tuning data set; the method comprises the following steps: acquiring spatial interaction, structure and sequence features of protein pocket-ligand compound data, optimizing the sequence features through cross-modal contrast learning, and fusing the optimized sequence features with the structure features to obtain first multi-modal fusion features; constructing a molecule generation model, pre-training the model by using a pre-training data set, and then finely adjusting the model by using a second multi-modal fusion feature fused with the first multi-modal fusion feature and a corresponding ligand molecule sequence feature; for a specific target protein, the model can generate a molecule adapted to the specific target protein according to the multi-modal fusion feature of the pocket. The invention also discloses a system for realizing the method. The method and the system provided by the invention can effectively generate the candidate molecule which is highly matched with the binding site of the target protein, has high affinity and is excellent in synthesis accessibility.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics and medicine, and specifically relates to a molecular generation method and system driven by interaction-aware contrastive coding and multimodal pocket features. Background Technology

[0002] Targeted drug design, a core component of drug development, aims to design and screen candidate molecules with high biological activity, precise binding to target proteins, and potential therapeutic effects based on the structural and functional information of specific biological targets. However, traditional drug design suffers from problems such as long development cycles, high costs, low success rates, and difficulty in fully exploring large-scale chemical spaces, limiting the efficiency and innovation of new drug development. Against this backdrop, combining data-driven deep learning technology with molecular generation technology can significantly improve the efficiency of molecular design and screening, generate new molecules that bind more tightly to target proteins, and, by generating novel structures from scratch, potentially alleviate problems such as drug resistance. Therefore, it has significant application value and practical significance.

[0003] Deep learning-based drug design methods, through systematic learning of large-scale molecular structure data, chemical property information, and bioactivity data, can automatically extract potential structural patterns and functional correlations, enabling the rapid generation and property prediction of new molecules. Structure-based drug design is a widely used approach. Its basic idea is to design and generate small molecule compounds that are complementary to the binding pocket, based on the three-dimensional structural information of the target protein. This method, based on well-defined target structural features, can improve the targeting and rationality of molecular design to a certain extent. However, this method is highly dependent on high-quality target structures and requires substantial computational resources. Furthermore, most existing methods fail to fully utilize pocket-ligand interaction information, thus affecting the accuracy and reliability of the design. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a molecular generation method and system driven by interaction-aware contrastive coding and multimodal pocket features. Specifically, it proposes a method and system for constraining molecular generation by optimizing the amino acid sequence features of protein pockets through contrastive learning and integrating multimodal features, including the multi-scale structural features of protein pockets, amino acid sequence features and molecular sequence features of corresponding ligands.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] This invention proposes a molecular generation method driven by interaction-aware contrastive coding and multimodal pocket features, comprising the following steps:

[0007] Construct a pre-training dataset P-Drug and a fine-tuning dataset F-PocketLig. The pre-training dataset P-Drug contains drug small molecule data, and the fine-tuning dataset F-PocketLig contains protein pocket-ligand complex data.

[0008] For protein pocket-ligand complex data in the fine-tuning dataset F-PocketLig, spatial interaction features of protein pocket-ligand, structural features of protein pockets, and amino acid sequence features are obtained respectively. Through interaction-aware cross-modal contrastive learning, the amino acid sequence features of protein pockets are optimized using the spatial interaction features of protein pocket-ligand. The optimized amino acid sequence features are then fused with the structural features of protein pockets to obtain the first multimodal fusion feature of protein pockets.

[0009] A molecular generation model was constructed and pre-trained on the pre-training dataset P-Drug to enable it to learn the topological rules and potential chemical rules in the structure of drug molecules.

[0010] The first multimodal fusion feature of the protein pocket is further fused with the molecular sequence feature of the corresponding ligand to obtain the second multimodal fusion feature of the protein pocket. The second multimodal fusion feature is then used to fine-tune the pre-trained molecular generation model so that the model learns the interaction rules between the protein pocket and the ligand.

[0011] For a specific target protein, the structural features and amino acid sequence features of the target protein pocket are obtained and fused to obtain the multimodal fusion features of the specific target protein pocket;

[0012] The multimodal fusion features of the specific target protein pocket are input into the fine-tuned molecular generation model to generate molecules that are adapted to the specific target protein.

[0013] Preferably, the acquisition of the spatial interaction features of the protein pocket-ligand, the structural features of the protein pocket, and the amino acid sequence features includes:

[0014] The protein pocket-ligand complex is represented as a spatial interaction graph of the protein pocket-ligand with atoms as nodes, and its feature embedding representation is extracted using a multi-layer interaction embedding module. ;

[0015] Protein pockets are represented as amino acid hypergraphs with amino acid nodes, and their feature embeddings are extracted using a hypergraph neural network. ;

[0016] The protein pocket is represented as a graph with atoms as nodes, and its feature embedding is extracted using a multi-channel neural weighted message passing combined with a skip knowledge aggregation network. ;

[0017] Amino acid sequence feature embeddings of protein pockets were extracted using a pre-trained language model. ;

[0018] The structural features of the protein pocket include the amino acid hypergraph features and atomic graph features of the protein pocket.

[0019] Preferably, the optimization of the amino acid sequence features of the protein pocket using the spatial interaction features of the protein pocket-ligand through interaction-aware cross-modal contrastive learning includes:

[0020] The feature embedding representation of the protein pocket-ligand space interaction graph is used. Amino acid sequence feature embedding representation Project the data onto a unified low-dimensional common space and calculate the similarity matrix between the two projections.

[0021] We introduce losses from the sequence to the spatial interaction direction and losses from the spatial interaction to the sequence direction, and use the average of the two losses as the final comparative loss.

[0022] By minimizing the contrast loss, the optimized protein pocket amino acid sequence feature embedding representation is obtained. .

[0023] Preferably, the step of fusing the optimized amino acid sequence features with the structural features of the protein pocket to obtain the first multimodal fusion feature of the protein pocket includes:

[0024] First, embed the amino acid hypergraph features of the protein pocket into the representation. With atomic graph feature embedding representation By performing fusion, structural fusion features are obtained. ;

[0025] Then integrate structural features With optimized protein pocket amino acid sequence feature embedding representation Further fusion yields a fusion feature representation of the protein pocket structure and sequence. This is the first multimodal fusion feature of the protein pocket.

[0026] Preferably, the embedding of amino acid hypergraph features of the protein pocket is represented as... With atomic graph feature embedding representation By performing fusion, structural fusion features are obtained. ,include:

[0027] The amino acid hypergraph features of the protein pocket are embedded in the representation. With atomic graph feature embedding representation As two primitive modal features, they are respectively mapped to a unified low-dimensional common space;

[0028] Three interactive methods—forward cross-fusion, reverse cross-fusion, and element-by-element interactive fusion—are used to fuse the relationship between two original modal features from different perspectives, resulting in three interactive features.

[0029] The two original modal features are concatenated with the three interaction features to obtain the structural fusion feature. .

[0030] Preferably, the structural fusion feature With optimized protein pocket amino acid sequence feature embedding representation The fusion was performed to obtain the fusion feature representation of the protein pocket structure and sequence. ,include:

[0031] Fusing protein pocket structure features With optimized amino acid sequence feature embedding representation The data is then concatenated and input into a gating network to calculate the gating weights.

[0032] Based on the gating weights, the two modal features are weighted and fused to generate weighted fused features;

[0033] A weak residual factor is introduced into the weighted fusion features to obtain the final fusion feature representation of the protein pocket structure and sequence. .

[0034] Preferably, the molecular generation model includes a word embedding layer, a position embedding layer, and multiple stacked GPT modules, and the loss function of the model during pre-training is shown in formula (2):

[0035]

[0036] Where B represents the number of samples in the training batch, and K represents the total number of classes. Indicates the first Does the true category of each sample belong to a category? , belongs to category The value is 1 if it is true, and 0 otherwise. Indicates the first Each sample belongs to category The predicted probability.

[0037] Preferably, the further fusion of the first multimodal fusion feature of the protein pocket with the molecular sequence feature of the corresponding ligand to obtain the second multimodal fusion feature of the protein pocket includes:

[0038] The first multimodal fusion feature of the protein pocket is introduced as a constraint condition of the pre-trained molecular generation model and concatenated with the molecular sequence features of the corresponding ligands obtained by word embedding layer and position embedding layer to obtain the second multimodal fusion feature of the protein pocket.

[0039] Preferably, the loss function of the molecular generation model in the fine-tuning stage is as shown in formula (3):

[0040]

[0041] in, It is the molecular generation loss function defined in formula (2). It is a contrastive loss function. It is the cosine similarity loss function. and These are learnable parameters. Used to prevent The calculation anomaly caused by the value approaching zero.

[0042] In another aspect, this invention proposes a molecular generation system driven by interaction-aware contrastive coding and multimodal pocket features, for implementing the aforementioned molecular generation method driven by interaction-aware contrastive coding and multimodal pocket features, comprising:

[0043] The dataset construction module is used to construct the pre-training dataset P-Drug and the fine-tuning dataset F-PocketLig. The pre-training dataset P-Drug contains drug small molecule data, and the fine-tuning dataset F-PocketLig contains protein pocket-ligand complex data.

[0044] The first multimodal fusion feature acquisition module is used to acquire the spatial interaction features of protein pockets and ligands, the structural features of protein pockets, and the amino acid sequence features of protein pockets for the protein pocket-ligand complex data in the fine-tuning dataset F-PocketLig. Through interaction-aware cross-modal comparative learning, the spatial interaction features of the protein pockets are used to optimize the amino acid sequence features of the protein pockets. The optimized amino acid sequence features are then fused with the structural features of the protein pockets to obtain the first multimodal fusion feature of the protein pockets.

[0045] The model building and pre-training module is used to build a molecular generation model and pre-train the model based on the pre-training dataset P-Drug, so that it learns the topological rules and potential chemical rules in the structure of drug molecules.

[0046] The model fine-tuning module is used to further fuse the first multimodal fusion feature of the protein pocket with the molecular sequence feature of the corresponding ligand to obtain the second multimodal fusion feature of the protein pocket, and to fine-tune the pre-trained molecular generation model using the second multimodal fusion feature so that the model learns the interaction rules between the protein pocket and the ligand.

[0047] The target multimodal fusion feature acquisition module is used to acquire and fuse the structural features and amino acid sequence features of a target protein pocket for a specific target protein, and obtain the multimodal fusion features of that specific target protein pocket.

[0048] The molecule generation module is used to input the multimodal fusion features of the specific target protein pocket into the fine-tuned molecule generation model to generate molecules that are adapted to the specific target protein.

[0049] The present invention has the following beneficial effects:

[0050] (1) By constructing an atomic-level protein pocket-ligand spatial interaction diagram, this invention can intuitively depict the spatial conformation and atomic-level interaction relationship within the binding region, capture the potential interaction mode between the pocket and the ligand, thereby retaining key binding information during the molecular design process, making the generated molecule more structurally consistent with the geometric and chemical characteristics of the pocket, and helping to improve the binding affinity and specificity of the generated molecule to the target.

[0051] (2) This invention optimizes and corrects the amino acid sequence features of the pocket by using real spatial interaction information between the protein pocket and the ligand through cross-modal comparative learning of interaction perception, which significantly improves the semantic clarity and functional guidance of sequence features in molecular generation, enabling them to more accurately reflect binding requirements.

[0052] (3) This invention integrates the structural features of protein pockets (including amino acid hypergraph features and atomic graph features) with the optimized amino acid sequence features through a hierarchical fusion strategy, forming a comprehensive multimodal pocket feature, which enhances the characterization ability of protein pocket features, provides more sufficient condition constraints for molecular generation models, avoids the problem of poor compatibility between generated molecules and targets caused by traditional single features, and helps to significantly improve the quality of generated molecules and R&D efficiency. Attached Figure Description

[0053] Figure 1 This is a basic flowchart of the interaction-aware contrastive coding and multimodal pocket feature-driven molecular generation method of Embodiment 1 of the present invention;

[0054] Figure 2 This is a flowchart of the interaction-aware contrastive coding and multimodal pocket feature-driven molecular generation method of Embodiment 1 of the present invention.

[0055] Figure 3 This is a schematic diagram of the cross-modal contrastive learning process in Embodiment 1 of the present invention;

[0056] Figure 4 This is a schematic diagram of the process for obtaining the first multimodal fusion feature of the protein pocket in Embodiment 1 of the present invention;

[0057] Figure 5 This is a schematic diagram of the bond lengths of the molecules generated in Example 1 of the present invention;

[0058] Figure 6 This is a schematic diagram of the chemical spatial distribution of the generated molecules in Example 1 of the present invention;

[0059] Figure 7 This is a schematic diagram of the docking of the generated molecule with poly(ADP-ribose) polymerase-13 in Example 1 of the present invention;

[0060] Figure 8 This is a schematic diagram of the binding of the generated molecule to the poly(ADP-ribose) polymerase-13 pocket in Example 1 of the present invention;

[0061] Figure 9 This is a diagram showing the interaction between the generated molecule and poly(ADP-ribose) polymerase-13 in Example 1 of the present invention. Detailed Implementation

[0062] To make the technical means, creative features, objectives and effects of the present invention easier to understand, the preferred embodiments of the present invention are described below in conjunction with examples. However, it should be understood that these descriptions are only for further illustrating the features and advantages of the present invention, and not for limiting the claims of the present invention.

[0063] Terminology Explanation:

[0064] Protein pocket: refers to a cavity or depression on or inside a protein surface that is suitable for binding with ligands;

[0065] Hypergraph: An extended form of graph in which hyperedges can connect two or more nodes at the same time, used to describe complex multi-dimensional relationships;

[0066] SMILES (Simplified Molecular-Input Line-Entry System): A descriptive language that represents molecular structures in string form, used to record the atoms, bonds, and topological relationships of molecules, and to convert the structural information of chemical molecules into resolvable linear sequences;

[0067] Hypergraph Neural Networks (HGNN) is a deep learning model used to process hypergraph data. By propagating and aggregating features on nodes and hyperedges of the hypergraph, it can effectively capture high-order relationships and multi-dimensional dependencies between nodes.

[0068] FLAG: Fragment-based 3D molecular generation framework.

[0069] DrugGPS: A method for molecular generation by learning sub-pocket prototypes and combining them with molecular motifs.

[0070] PMDM: A method for efficiently generating 3D molecules and optimizing lead compounds based on target protein pocket information.

[0071] GCDM: A method that integrates molecular chirality and geometric completeness biases to achieve 3D molecular generation and optimization, and can be extended to protein pocket condition generation.

[0072] Multi-Channel Neural Weighted Message Passing with Jumping Knowledge Aggregation Network (MC-NWMP-JKN) is a graph neural network architecture that uses multiple channels to weighted message passing of node features and integrates information from each layer using a jumping knowledge aggregation mechanism, thereby effectively capturing high-order relationships and global structural features between nodes.

[0073] Recurrent Neural Network (RNN): A type of neural network suitable for sequential data that captures contextual dependencies through cyclic updates of hidden states;

[0074] Content Attention: An attention mechanism that determines the importance of each input element by calculating the similarity between the query vector and the input content, thereby weighting and aggregating information in a sequence or graph.

[0075] Multi-Head Attention (MHA): An extended attention mechanism that aggregates features of an input sequence through multiple parallel attention heads to capture dependencies between different subspaces and enhance information representation;

[0076] Feed-Forward Network (FFN): A basic neural network structure that achieves layer-by-layer feature mapping and representation learning by processing the input vector through linear transformation and non-linear activation functions;

[0077] Vocabulary: A vocabulary built on the word segmentation table, containing all possible tokens and converting them into integer indices;

[0078] Bonding relationships: The way atoms in a molecule are connected by chemical bonds, including single bonds, double bonds, triple bonds, and aromatic bonds;

[0079] database:

[0080] ChEMBL database: Target and Bioactivity Database, specifically used to store data on the bioactivity of small molecule drugs. Its official website is https: / / www.ebi.ac.uk / chembl / .

[0081] CrossDocked database: A large-scale structural database dedicated to protein-ligand binding research, containing information on protein pockets and ligand complexes generated through cross-docking;

[0082] RCSB PDB: A global scientific resource repository that primarily collects and provides three-dimensional structural data of biological macromolecules such as proteins, nucleic acids, and some polysaccharides. Access URL: https: / / www.rcsb.org / .

[0083] Example 1

[0084] See Figures 1 to 2 This embodiment provides a molecular generation method driven by interaction-aware contrastive coding and multimodal pocket features. The specific steps of this method are as follows:

[0085] S1. Construct a pre-training dataset P-Drug and a fine-tuning dataset F-PocketLig. The pre-training dataset P-Drug contains drug small molecule data, and the fine-tuning dataset F-PocketLig contains protein pocket-ligand complex data.

[0086] Specifically, step S1 further includes:

[0087] S101, Data Acquisition. Includes:

[0088] (1) Obtain the pre-training dataset P-Drug data: Extract drug small molecules from the database (such as ChEMBL) and represent the small molecules using SMILES sequences.

[0089] (2) Obtain protein pocket-ligand data: Obtain protein-ligand complex data from a database (such as CrossDocked), where protein pockets are stored in PDB format and ligands are stored in SDF format.

[0090] S102, Data Preprocessing. Includes:

[0091] (1) Preprocessing of drug small molecule data: Screening of drug small molecules represented by SMILES sequences, selecting molecules with a length between 20 and 100 characters; cleaning and validity verification of the screened drug small molecules, removing molecular sequences that cannot be parsed or do not conform to the basic chemical structure of the molecules, and performing desalting and destereotyping on the valid SMILES sequences; splitting the cleaned SMILES sequences into ordered character fragments, constructing a drug small molecule vocabulary, and mapping the character fragment sequence of each sequence to the corresponding integer index to form the drug small molecule dataset P-Drug for pre-training of the molecular generation model.

[0092] All preprocessed drug small molecule data were used for model pre-training so that the model could learn the topological relationships and underlying chemical laws of molecular structures.

[0093] (2) Preprocessing of protein-ligand complex data: convert the SDF format data of ligands into SMILES representation, and perform desalting, destereotyping and validity verification; split the cleaned SMILES sequence into character fragments, construct a ligand vocabulary, and map each character fragment to an integer index; at the same time, extract information such as amino acid residues, atomic properties and spatial coordinates from the PDB file of the protein pocket.

[0094] After preprocessing, the ligand sequence data and the corresponding protein pocket information are integrated to form a protein pocket-ligand dataset, namely the fine-tuning dataset F-PocketLig, which includes a training set and a test set for subsequent fine-tuning of the molecular generation model and targeting evaluation.

[0095] It should be noted that in this embodiment, 2,474,590 small molecule drug data were initially obtained from the ChEMBL database. After preprocessing, 1,127,080 SMILES data were obtained for pre-training. Simultaneously, the initial number of protein pocket-ligand complexes was approximately 160,000. The proteins were clustered using the mmseqs2 tool with 30% sequence identity. Then, 100,000 protein pocket-ligand pairs were randomly selected for training, and 100 protein pocket-ligand pairs were randomly selected for testing. After further preprocessing, a final total of 15,770 protein pocket-ligand pairs were obtained for training, and 52 protein pocket-ligand pairs were used for testing.

[0096] S2, for the protein pocket-ligand complex data in the fine-tuning dataset F-PocketLig, the spatial interaction features of the protein pocket-ligand, the structural features of the protein pocket, and the amino acid sequence features are obtained respectively; through interaction-aware cross-modal contrastive learning, the amino acid sequence features of the protein pocket are optimized using the interaction features of the protein pocket-ligand; and the optimized amino acid sequence features are fused with the structural features of the protein pocket to obtain the first multimodal fusion feature of the protein pocket. Further, this step specifically includes:

[0097] S201 transforms the protein pocket-ligand complex data in the fine-tuning dataset F-PocketLig into a graph representation, including a protein pocket-ligand spatial interaction graph for characterizing the atomic-level interaction between protein pockets and ligands in three-dimensional space, a protein pocket amino acid hypergraph for characterizing the global structural features of protein pockets, and a protein pocket atomic graph for characterizing the local structural features of protein pockets.

[0098] (1) Represent the protein pocket-ligand complex as a graph with atoms as nodes and construct a protein pocket-ligand spatial interaction graph.

[0099] Specifically, for a given protein pocket-ligand complex dataset, the protein pocket and ligand are represented as sets of atoms, each atom vectorized by its class characteristics (e.g., carbon, oxygen, nitrogen) and three-dimensional spatial coordinates. Then, the set of all atoms in the protein pocket is merged with the set of all atoms in the ligand molecule to form the overall atomic set of the complex, which together serve as the node set of the spatial interaction graph. Next, an edge set is constructed based on the spatial proximity relationships between nodes (atoms). All atom pairs are traversed, and their spatial Euclidean distance is calculated. If the distance between any two atoms is less than or equal to a preset distance threshold (preferably 5 Å), an edge is established between them to characterize potential van der Waals interactions, hydrogen bonds, and other non-bonded interactions. Optionally, to reduce the impact of redundant edges on computational efficiency, a "k-nearest neighbor strategy" can be adopted, where for each atom node, only connections to its k nearest atoms (e.g., k=10) are retained, ensuring that the edge set focuses on the most significant spatial proximity relationships. Through the above process, each atom is connected to several nearest neighbor atoms, ultimately forming an atomic-level spatial interaction diagram of the protein pocket and ligand, which is used to characterize the interaction between the protein pocket and the ligand in three-dimensional space.

[0100] (2) Represent the protein pocket as a hypergraph with amino acid nodes and construct the amino acid hypergraph of the protein pocket.

[0101] Specifically, using protein pocket data from the fine-tuned dataset F-PocketLig, the coordinates of the central carbon atom (Cα) of each amino acid in the protein pocket are extracted as the amino acid node position. A spatial distance threshold (e.g., 5 Å) is set for each amino acid node, and all neighboring amino acid nodes within this threshold range (including the central node itself) are grouped into the same set, which constitutes a hyperedge. This process is repeated for all amino acid nodes to construct a complete set of hyperedges, ultimately forming the amino acid hypergraph of the protein pocket, used to characterize the overall shape and topology of the protein pocket.

[0102] (3) Represent the protein pocket as a graph with atoms as nodes and construct the atomic graph of the protein pocket.

[0103] Specifically, using protein pocket data from the fine-tuning dataset F-PocketLig, non-hydrogen atoms in the protein pockets are used as nodes. The node features are the same as the atomic features in step S201 (1), namely, the atom category and three-dimensional coordinates. The construction of edges is based on the following two rules:

[0104] a. Bonding connection: Based on the explicit chemical bond information in the protein structure data, an edge is established between all pairs of atoms connected by covalent bonds;

[0105] b. Spatial Proximity Connection: Calculate the Euclidean distance between any two atoms inside the pocket. If the distance is less than or equal to a preset threshold (e.g., 5 Å), then establish an edge between them to characterize potential nonbonded interactions.

[0106] The above process ultimately forms an atomic map of the protein pocket, which is used to depict the local structural details inside the protein pocket.

[0107] This step involves the targeted construction of atomic-level spatial interaction graphs, protein pocket amino acid hypergraphs, and protein pocket atomic graphs. These graphs lay a multi-dimensional feature foundation for the precise generation of high-affinity candidate drug molecules from three dimensions: intermolecular interaction patterns, global functional topology of pockets, and local structural details of pockets.

[0108] S202, extracts and represents the features of the protein pocket-ligand spatial interaction diagram, protein pocket amino acid hypergraph, protein pocket atomic diagram, and protein pocket amino acid sequence, respectively.

[0109] (1) Extract and represent the features of the protein pocket-ligand spatial interaction map. .

[0110] This embodiment employs a multi-layer interaction embedding module to extract features from the spatial interaction graph, preferably with 6 layers. Each layer of the interaction embedding module consists of a multi-head attention (MHA) module and a feedforward network layer. The feature extraction process is as follows:

[0111] First, for any atomic node in the spatial interaction graph, such as the first... The initial features of each atomic node (including atomic type, 3D spatial coordinates, and physicochemical properties) are linearly transformed to obtain a low-dimensional embedding vector. This vector is then iteratively updated through a 6-layer interactive embedding module to obtain a feature vector for that atomic node that includes contextual information. Finally, the feature vectors of all atomic nodes are stacked row-wise to obtain the atomic feature matrix. The feature embedding representation of the spatial interaction graph is obtained after scattering mean processing. .

[0112] (2) Extract and represent the amino acid hypergraph features of the protein pocket. .

[0113] This embodiment uses Hypergraph Neural Networks (HGNN) to encode the pocket hypergraph. The HGNN consists of two hypergraph convolutional layers, used to extract direct association features of local residues and structural dependency features across regions, respectively. The feature extraction process is as follows:

[0114] For any amino acid node in the protein pocket amino acid hypergraph, its initial features (including the physicochemical properties of the amino acid, amino acid type, and 3D coordinates) are extracted and concatenated into a comprehensive feature vector. Then, the feature vectors of all amino acid nodes are stacked row-wise to obtain an initial node feature matrix. The data is then input into an HGNN and processed through two hypergraph convolutional layers. Finally, Set2Set pooling is used to dynamically aggregate the features of all nodes to obtain the overall representation of the protein pocket. The Set2Set mechanism combines a recurrent neural network (RNN) and content attention, using multiple iterations to dynamically aggregate the features of all nodes, ultimately yielding the hypergraph feature embedding representation of the protein pocket. .

[0115] (3) Extract and represent the atomic map features of the protein pocket. .

[0116] This embodiment employs a multi-channel neural weighted message passing network combined with a skip-knowledge aggregation network to encode the atomic graph features of protein pockets. Nodes in this graph represent pocket atoms, and edges are constructed based on bonding relationships or spatial proximity between atoms. When an edge exists between two atoms, they are adjacent nodes (or neighboring nodes). The feature extraction process is as follows:

[0117] The protein pocket atomic diagram The initial features of the atomic nodes and their neighboring nodes j (as described in step (1) of S202, and will not be repeated here) are input into the graph neural network. The network preferably contains 6 message passing layers, each layer capturing topological semantic information at different scales through neural weighted message passing, wherein the first layer... The layer update is defined as shown in formula (1):

[0118]

[0119] in, and Representing nodes respectively In the Layer and first The feature embedding vector of the layer, Representing neighboring nodes In the The feature embedding vector of the layer, Let (i,j) represent the learnable weights of edge (i,j). Represents a node The set of adjacent points, Indicates the first Learnable residual coefficients of the layer This represents a non-linear activation function.

[0120] Furthermore, this embodiment introduces a skip knowledge aggregation block to aggregate node representations from all layers, effectively integrating multi-scale node features and minimizing information loss during deep message passing to generate the final node features. All node features are then integrated to generate a node feature matrix. The feature matrix After Set2Set pooling, the final atomic graph feature embedding representation is obtained. .

[0121] (4) Extract and represent the amino acid sequence characteristics of the protein pocket. .

[0122] In this embodiment, for the protein pocket data in the fine-tuning dataset F-PocketLig, the pre-trained language model ProteinBert is used to encode the amino acid sequences of the protein pockets. Specifically, the amino acid sequences (represented by single-letter code strings for amino acids) are input into the model, and a 512-dimensional global embedding vector is extracted from the intermediate layer during forward propagation and projected. After projection transformation, the sequence feature embedding representation of the protein pocket is finally obtained. .

[0123] S203 optimizes the amino acid sequence features of protein pockets by utilizing the spatial interaction features of protein pockets and ligands through cross-modal contrastive learning based on interaction perception.

[0124] Specifically, such as Figure 3 As shown, this embodiment first identifies the amino acid sequence characteristics of the protein pocket. Features of atomic interaction diagrams of protein pockets and ligand molecules Each feature is projected onto a unified low-dimensional common space, and then the similarity matrix after projection is calculated. During training, sequence features are optimized using a contrastive loss function, which introduces losses from the sequence to the spatial interaction direction and from the spatial interaction to the sequence direction. The average of the two losses is used as the final contrastive loss. By minimizing the contrastive loss during training. This approach makes the sequence features more closely resemble their corresponding interaction features in the embedding space, thereby enhancing the expressive power of the sequence features and resulting in an optimized amino acid sequence feature embedding representation. .

[0125] S204, the amino acid sequence features of the optimized protein pocket are fused with the structural features of the protein pocket to obtain the first multimodal fusion feature of the protein pocket.

[0126] It should be noted that, in this embodiment, the structural features of the protein pocket include the amino acid hypergraph features and the atomic graph features of the protein pocket. In specific operation, such as... Figure 4 As shown, the amino acid hypergraph features and atomic graph features are first fused to obtain the structural fusion features; then the structural fusion features are further fused with the optimized protein pocket amino acid sequence features to finally obtain the first multimodal fusion feature of protein pocket structure and sequence fusion.

[0127] (1) The amino acid hypermap features of the protein pocket are fused with the atomic map features to obtain the structural fusion features.

[0128] First, the amino acid hypergraph features of the protein pocket are embedded in the representation. and atomic graph embedding feature embedding representation The original modal features are mapped to a unified low-dimensional common space to eliminate differences in scale and distribution between the different modal features, laying the foundation for subsequent fusion. To fully capture the complex relationship between the two modal features, three complementary interactive modeling methods are used to fuse them from different perspectives:

[0129] Forward cross-fusion: Based on hypergraph features and atomic graph features, a forward cross matrix is ​​constructed and flattened;

[0130] Reverse cross-fusion: Construct a reverse cross matrix and flatten it;

[0131] Element-wise interactive fusion: Multiply the two original modal features element-wise.

[0132] Finally, the two mapped original modal features are concatenated with the three interaction features mentioned above to form a protein pocket structure-level fusion feature. .

[0133] (2) Further fuse the amino acid sequence features of the optimized protein pocket with the structural fusion features of the protein pocket.

[0134] This embodiment designs a gated fusion module to fuse the structural features of protein pockets. Amino acid sequence characteristics of the optimized protein pocket Adaptive fusion is performed, and the contributions of both structure and sequence are adaptively adjusted through gating weights to measure their relative importance:

[0135] First, integrate structural features With sequence features The concatenated vector is then passed through a gating network, where gating weights are calculated using an activation function and two layers of nonlinear gating networks. Subsequently, these gating weights are used to weightedly fuse structural and sequence features, and a weak residual factor is introduced into the weighted fusion result, which is essentially the original fused structural features. Regression is performed with relatively small weights (e.g., 0.1) to ensure that the fused features do not lose key structural and semantic information. Furthermore, to improve the efficiency of the gated fusion module, cosine similarity loss is introduced during its training process. As a regularization term, it is used to improve the module's perception of the differences in features between the two modalities.

[0136] Through the synergistic effect of the above-mentioned gating, weighting, residual and regularization, the fusion characteristics of the protein pocket structure and sequence are finally obtained. .

[0137] S3. Construct a molecular generation model and pre-train the model based on the pre-training dataset P-Drug to enable it to learn the topological rules and potential chemical rules in molecular structures.

[0138] The molecular generation model in this embodiment Main package It includes a word embedding layer, a position embedding layer, and eight stacked GPT (Generative Pre-trained Transformer) modules.

[0139] To enable the model to generate reasonable small drug molecules, the P-Drug pre-training dataset was used to pre-train the model, learning the topological and chemical rules in the structure of small drug molecules. During the pre-training phase, the SMILES sequences of molecules in the P-Drug dataset were first segmented into character fragments by word embedding, and each character fragment was mapped to a corresponding integer index representation. These character fragments, represented by integer indices, were encoded by word embedding and position embedding layers and converted into continuous vector representations in high-dimensional space. These vector representations were then input into the aforementioned stacked GPT module to predict the conditional probability distribution of the next character fragment, i.e., the probability of each possible subsequent character fragment given the preceding character fragment.

[0140] Furthermore, to measure the predictive performance of the model, a cross-entropy loss function is defined to measure the difference between the probability distribution predicted by the model and the true distribution. The loss function is shown in formula (2):

[0141]

[0142] Where B represents the number of samples in the training batch, and K represents the total number of character segment categories (i.e., the total number of all different character segment types that may appear in the SMILES sequence). Indicates the first Does the true category of each sample belong to a category? The value is 1 if it belongs to category c, and 0 otherwise. Indicates the first Each sample belongs to category The predicted probability.

[0143] S4, the first multimodal fusion feature of the protein pocket is further fused with the molecular sequence feature of the corresponding ligand to obtain the second multimodal fusion feature of the protein pocket, and the pre-trained molecular generation model is fine-tuned using the second multimodal fusion feature so that the model learns the interaction rules between the protein pocket and the ligand.

[0144] This step uses protein pocket-ligand complex data from the F-PocketLig fine-tuning dataset to fine-tune the pre-trained model. Specifically, the protein pocket-ligand complex data and individual protein pocket data are processed in step S2 to form the first multimodal fusion feature of the protein pockets. , which serves as a constraint input to the preprocessed molecular generation model.

[0145] Furthermore, the ligand data corresponding to the protein pockets are extracted separately from the fine-tuning dataset F-PocketLig. After undergoing the same word segmentation and character fragment mapping processing as in step S3 for drug small molecules, the data is input into the preprocessed molecular generation model in integer index representation. The ligand molecules are then encoded through the model's word embedding layer and position embedding layer to obtain the molecular sequence features of the ligands. .

[0146] Furthermore, the first multimodal fusion features of the input protein pocket are... Molecular sequence characteristics of corresponding ligands Further fusion yields the second multimodal fusion feature of the protein pocket, which serves as input to the GPT module of the model, enabling the model to fully utilize the interaction between the protein pocket and the ligand during the molecular generation process.

[0147] Furthermore, the molecular generation model employs a loss function during the fine-tuning phase. To optimize the molecule generation process, two auxiliary loss functions are introduced. and Strengthening the guiding role of pocket features in molecule generation, total loss function As shown in formula (3):

[0148]

[0149] in, It is the molecular generation loss function defined in formula (2). It is a contrastive loss function. It is the cosine similarity loss function. and These are learnable parameters. Used to prevent The calculation anomaly caused by the value approaching zero.

[0150] S5. For a specific target protein, obtain and fuse the structural features and amino acid sequence features of the target protein pocket to obtain the multimodal fusion features of the specific target protein pocket.

[0151] It should be noted that the structural features of the target protein pocket include amino acid hypergraph features and atomic graph features. For a specific target protein, its protein pocket data is first converted into a graph structure representation, including the amino acid hypergraph and atomic graph of the protein pocket. Subsequently, features of the amino acid hypergraph, atomic graph, and amino acid sequence are extracted separately. Then, the amino acid hypergraph features and atomic graph features are first fused to obtain structural fusion features, and then the structural fusion features are further fused with the amino acid sequence features to finally obtain the multimodal fusion features of the specific target protein pocket.

[0152] For specific instructions, please refer to step S2 for the feature extraction and fusion processing method of the fine-tuning dataset F-PocketLig, which will not be elaborated here.

[0153] In this embodiment, a core domain (PDB ID: 7KZH) of poly(ADPribose) polymerase-13 (PARP13) was selected as the target protein. The PDB file (PDB ID: 7KZH) of poly(ADPribose) polymerase-13 was first downloaded from the RCSB PDB database. After removing water molecules, its multimodal fusion characteristics were obtained using the method described above.

[0154] S6, input the multimodal fusion features of the specific target protein pocket into the fine-tuned molecular generation model to generate a molecule that is compatible with the specific target protein.

[0155] Specifically, this embodiment uses a fine-tuned molecular generation model, taking the multimodal fusion features of the poly(ADP-ribose) polymerase-13 pocket as the input of the model, and generating a batch of potential drug molecules based on the protein pocket features. Figure 5 The figure shows the distribution of bond lengths on different bonds in molecules generated by different methods (including FLAG, DrugGPS, PMDM, and GCDM). As can be seen from the figure, the bond length distribution of molecules generated by the method (Our) in this embodiment is concentrated and closer to the distribution of real molecules in the F-PocketLig dataset. This indicates that compared with other methods, the method in this embodiment can more accurately capture the typical characteristics of different chemical bond lengths and simulate or design molecules with specific structures and functions. Figure 6 The chemical spatial distribution of the generated molecules and the dataset molecules is shown. As can be seen from the figure, the generated molecules are densely and uniformly distributed in space, indicating that the generated molecules have good cohesion. At the same time, the generated molecules have a wide spatial distribution range, covering almost all regions of the real molecule distribution, indicating that this method can fully learn the potential chemical rules of molecules and generate molecules with diverse structures.

[0156] Furthermore, a molecular docking method was used to dock the generated molecule with poly(ADP-ribose) polymerase-13 to assess their interaction at the active site. Figure 7 The distances between the generated molecules and the pocket residues were shown to be between 1.8 and 2.5 Å, all less than 5 Å, indicating an interaction between the generated molecules and the pocket. Figure 8 The spatial relationship between the generated molecule and the target pocket is shown, demonstrating that the generated molecule's shape fits the pocket cavity perfectly, exhibiting a stable binding conformation. Furthermore, Figure 9 The interaction between the generated molecule and the target pocket was demonstrated. The molecule binds to the protein through hydrogen bonds, halogen bonds, and conjugation effects. The oxygen on the pyran ring and the nitrogen on the pyrrole ring form hydrogen bonds with ARG-588 and ARG-682, respectively, while the terminal trifluoromethyl group forms a halogen bond with SER-662. The two benzene rings exhibit conjugation effects with ARG-588 and TRP-609, respectively. The results indicate that the generated molecule has a high degree of target matching with the target protein and can form stable interactions at the binding site.

[0157] Example 2

[0158] This embodiment provides a molecular generation system driven by interaction-aware contrastive coding and multimodal pocket features, used to implement the molecular generation method described in Embodiment 1, including:

[0159] The dataset construction module is used to construct the pre-training dataset P-Drug and the fine-tuning dataset F-PocketLig. The pre-training dataset P-Drug contains drug small molecule data, and the fine-tuning dataset F-PocketLig contains protein pocket-ligand complex data.

[0160] The first multimodal fusion feature acquisition module is used to acquire the spatial interaction features of protein pockets and ligands, the structural features of protein pockets, and the amino acid sequence features of protein pockets for the protein pocket-ligand complex data in the fine-tuning dataset F-PocketLig. Through interaction-aware cross-modal comparative learning, the spatial interaction features of the protein pockets are used to optimize the amino acid sequence features of the protein pockets. The optimized amino acid sequence features are then fused with the structural features of the protein pockets to obtain the first multimodal fusion feature of the protein pockets.

[0161] The model building and pre-training module is used to build a molecular generation model and pre-train the model based on the pre-training dataset P-Drug, so that it learns the topological rules and potential chemical rules in the structure of drug molecules.

[0162] The model fine-tuning module is used to further fuse the first multimodal fusion feature of the protein pocket with the molecular sequence feature of the corresponding ligand to obtain the second multimodal fusion feature of the protein pocket, and to fine-tune the pre-trained molecular generation model using the second multimodal fusion feature so that the model learns the interaction rules between the protein pocket and the ligand.

[0163] The target multimodal fusion feature acquisition module is used to acquire and fuse the structural features and amino acid sequence features of a target protein pocket for a specific target protein, and obtain the multimodal fusion features of that specific target protein pocket.

[0164] The molecule generation module is used to input the multimodal fusion features of the specific target protein pocket into the fine-tuned molecule generation model to generate molecules that are adapted to the specific target protein.

[0165] For detailed descriptions of each module in the system, please refer to the relevant descriptions in the above method embodiments, which will not be repeated here.

[0166] It should be understood that the above embodiments are only one specific embodiment of the present invention. For those skilled in the art, various modifications and variations made based on the above description should be considered within the scope of protection of the present invention.

Claims

1. A molecular generation method driven by interaction-aware contrastive coding and multimodal pocket features, characterized in that, Includes the following steps: Construct a pre-training dataset P-Drug and a fine-tuning dataset F-PocketLig. The pre-training dataset P-Drug contains drug small molecule data, and the fine-tuning dataset F-PocketLig contains protein pocket-ligand complex data. For protein pocket-ligand complex data in the fine-tuning dataset F-PocketLig, we obtained the spatial interaction features of protein pocket-ligand, the structural features of protein pockets, and the amino acid sequence features. Through interaction-aware cross-modal contrastive learning, the amino acid sequence features of the protein pocket are optimized using the spatial interaction features of the protein pocket-ligand; and the optimized amino acid sequence features are fused with the structural features of the protein pocket to obtain the first multimodal fusion feature of the protein pocket. A molecular generation model was constructed and pre-trained on the pre-training dataset P-Drug to enable it to learn the topological rules and potential chemical rules in the structure of drug molecules. The first multimodal fusion feature of the protein pocket is further fused with the molecular sequence feature of the corresponding ligand to obtain the second multimodal fusion feature of the protein pocket. The second multimodal fusion feature is then used to fine-tune the pre-trained molecular generation model so that the model learns the interaction rules between the protein pocket and the ligand. For a specific target protein, the structural features and amino acid sequence features of the target protein pocket are obtained and fused to obtain the multimodal fusion features of the specific target protein pocket; The multimodal fusion features of the specific target protein pocket are input into the fine-tuned molecular generation model to generate molecules that are adapted to the specific target protein.

2. The molecular generation method according to claim 1, characterized in that, The acquisition of the spatial interaction features of the protein pocket-ligand, the structural features of the protein pocket, and the amino acid sequence features includes: The protein pocket-ligand complex is represented as a spatial interaction graph of the protein pocket-ligand with atoms as nodes, and its feature embedding representation is extracted using a multi-layer interaction embedding module. ; Protein pockets are represented as amino acid hypergraphs with amino acid nodes, and their feature embeddings are extracted using a hypergraph neural network. ; The protein pocket is represented as a graph with atoms as nodes, and its feature embedding is extracted using a multi-channel neural weighted message passing combined with a skip knowledge aggregation network. ; Amino acid sequence feature embeddings of protein pockets were extracted using a pre-trained language model. ; The structural features of the protein pocket include the amino acid hypergraph features and atomic graph features of the protein pocket.

3. The molecular generation method according to claim 2, characterized in that, The optimization of the amino acid sequence features of the protein pocket using the spatial interaction features of the protein pocket-ligand through interaction-aware cross-modal contrastive learning includes: The feature embedding representation of the protein pocket-ligand space interaction graph is used. Amino acid sequence feature embedding representation Project the data onto a unified low-dimensional common space and calculate the similarity matrix between the two projections. We introduce losses from the sequence to the spatial interaction direction and losses from the spatial interaction to the sequence direction, and use the average of the two losses as the final comparative loss. By minimizing the contrast loss, the optimized protein pocket amino acid sequence feature embedding representation is obtained. .

4. The molecular generation method according to claim 3, characterized in that, The process of fusing the optimized amino acid sequence features with the structural features of the protein pocket to obtain the first multimodal fusion feature of the protein pocket includes: First, embed the amino acid hypergraph features of the protein pocket into the representation. With atomic graph feature embedding representation By performing fusion, structural fusion features are obtained. ; Then integrate structural features With optimized protein pocket amino acid sequence feature embedding representation Further fusion yields a fusion feature representation of the protein pocket structure and sequence. This is the first multimodal fusion feature of the protein pocket.

5. The molecular generation method according to claim 4, characterized in that, The method of embedding amino acid hypergraph features of protein pockets is represented. With atomic graph feature embedding representation By performing fusion, structural fusion features are obtained. ,include: The amino acid hypergraph features of the protein pocket are embedded in the representation. With atomic graph feature embedding representation As two primitive modal features, they are respectively mapped to a unified low-dimensional common space; Three interactive methods—forward cross-fusion, reverse cross-fusion, and element-by-element interactive fusion—are used to fuse the relationship between two original modal features from different perspectives, resulting in three interactive features. The two original modal features are concatenated with the three interaction features to obtain the structural fusion feature. .

6. The molecular generation method according to claim 4, characterized in that, The structural fusion feature With optimized protein pocket amino acid sequence feature embedding representation The fusion was performed to obtain the fusion feature representation of the protein pocket structure and sequence. ,include: Fusing protein pocket structure features With optimized amino acid sequence feature embedding representation The data is then concatenated and input into a gating network to calculate the gating weights. Based on the gating weights, the two modal features are weighted and fused to generate weighted fused features; A weak residual factor is introduced into the weighted fusion features to obtain the final fusion feature representation of the protein pocket structure and sequence. .

7. The molecular generation method according to claim 1, characterized in that, The molecular generation model includes a word embedding layer, a position embedding layer, and multiple stacked GPT modules. The loss function of the model during pre-training is shown in Equation (2): , Where B represents the number of samples in the training batch, and K represents the total number of classes. Indicates the first Does the true category of each sample belong to a category? , belongs to category The value is 1 if it is true, and 0 otherwise. Indicates the first Each sample belongs to category The predicted probability.

8. The molecular generation method according to claim 1, characterized in that, The process of further fusing the first multimodal fusion feature of the protein pocket with the molecular sequence feature of the corresponding ligand to obtain the second multimodal fusion feature of the protein pocket includes: The first multimodal fusion feature of the protein pocket is introduced as a constraint condition of the pre-trained molecular generation model and concatenated with the molecular sequence features of the corresponding ligands obtained by word embedding layer and position embedding layer to obtain the second multimodal fusion feature of the protein pocket.

9. The molecular generation method according to claim 7, characterized in that, The loss function of the molecular generation model during the fine-tuning stage is shown in Equation (3): , in, It is the molecular generation loss function defined in formula (2). It is a contrastive loss function. It is the cosine similarity loss function. and These are learnable parameters. Used to prevent The calculation anomaly caused by the value approaching zero.

10. A molecular generation system driven by interaction-aware contrastive coding and multimodal pocket features, used to implement the molecular generation method driven by interaction-aware contrastive coding and multimodal pocket features as described in any one of claims 1-9, characterized in that, include: The dataset construction module is used to construct the pre-training dataset P-Drug and the fine-tuning dataset F-PocketLig. The pre-training dataset P-Drug contains drug small molecule data, and the fine-tuning dataset F-PocketLig contains protein pocket-ligand complex data. The first multimodal fusion feature acquisition module is used to acquire the spatial interaction features of protein pockets and ligands, the structural features of protein pockets, and the amino acid sequence features of protein pockets for the protein pocket-ligand complex data in the fine-tuning dataset F-PocketLig. Through interaction-aware cross-modal contrastive learning, the amino acid sequence features of the protein pocket are optimized using the spatial interaction features of the protein pocket-ligand; and the optimized amino acid sequence features are fused with the structural features of the protein pocket to obtain the first multimodal fusion feature of the protein pocket. The model building and pre-training module is used to build a molecular generation model and pre-train the model based on the pre-training dataset P-Drug, so that it learns the topological rules and potential chemical rules in the structure of drug molecules. The model fine-tuning module is used to further fuse the first multimodal fusion feature of the protein pocket with the molecular sequence feature of the corresponding ligand to obtain the second multimodal fusion feature of the protein pocket, and to fine-tune the pre-trained molecular generation model using the second multimodal fusion feature so that the model learns the interaction rules between the protein pocket and the ligand. The target multimodal fusion feature acquisition module is used to acquire and fuse the structural features and amino acid sequence features of a target protein pocket for a specific target protein, and obtain the multimodal fusion features of that specific target protein pocket. The molecule generation module is used to input the multimodal fusion features of the specific target protein pocket into the fine-tuned molecule generation model to generate molecules that are adapted to the specific target protein.

Images