A drug target association identification method based on spatial atlas and isometric network

Abstract

This invention relates to the field of bioinformatics and proposes a drug target association identification method based on spatial maps and isovariant networks. First, a three-dimensional spatial geometric map of the target protein and a microscopic chemical topological map of the drug molecule are respectively deduced and constructed to preserve the physical conformation of the receptor and the local pharmacophores of small molecules without loss. Second, a three-dimensional isovariant graph neural network and a graph message passing mechanism are jointly used for deep feature characterization. Finally, an affinity force strategy based on hyperbolic space is introduced to project features without loss onto a Poincaré sphere model with constant negative curvature, and the association probability between the drug and the target is predicted using non-Euclidean geometric distance. This invention effectively overcomes the feature congestion problem of traditional spatial mapping, fully explores the hidden hierarchical information within complex biomolecules, and significantly improves the accuracy and generalization robustness of drug target association analysis, providing effective technical support for new drug development, drug retargeting, and precision medicine.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics and relates to a binding tendency inference technique that combines protein three-dimensional spatial information with drug molecule characteristics. It is mainly applied to virtual drug screening, potential target verification and new drug development. Background Technology

[0002] Accurately assessing the association between drug molecules and target proteins is a core aspect of modern drug development. Traditional methods for identifying drug-target associations often rely on the protein's primary sequence information, neglecting its spatial conformational characteristics. While experimentally determined 3D structural PDB data offers high accuracy, its quantity is limited and acquisition costs are high. Existing computational models often suffer from information loss or dimensionality mismatch when fusing sequence features with spatial geometric features. To address these issues, this invention proposes a novel approach that utilizes coordinate-hot encoding technology to deeply fuse protein spatial coordinates with sequence features, combined with multi-task comparative learning to enhance the performance of molecular interaction assessment. Summary of the Invention

[0003] This invention proposes a drug target association identification method based on spatial maps and isovariant networks. Its aim is to overcome the information loss bottleneck caused by the fusion of traditional sequence features and conventional features, improve the accuracy of molecular interaction assessment, enhance the efficiency of high-throughput virtual screening, and provide more accurate computational support for new drug development. It also introduces a geometric deep learning architecture and hyperbolic space metric technology. This method first constructs a spatial topological map using the three-dimensional atomic coordinates of proteins, and then directly encodes the spatial geometric coordinates of the target using a three-dimensional rotation and translation isovariant graph neural network. Next, for drug molecules, a topological subgraph message passing mechanism is used to generate drug characterization. Then, the protein data is fed into the isovariant neural network. Finally, the fusion features of the drug and target are mapped to hyperbolic space, and the hyperbolic distance is calculated using a Poincaré sphere model. This perfectly fits the complex tree-like hierarchical structure and nonlinear correlation forces between drugs and protein families with a very small number of dimensions, thereby achieving highly sensitive and robust prediction of novel interaction associations. The specific technical process is as follows:

[0004] Step 1: First, construct a spatial topological map of the protein's three-dimensional atomic coordinates. For the amino acid sequence of the target protein, use the protein folding prediction system model of the large evolutionary scale to perform end-to-end conformational deduction. Without experimental determination, obtain the true three-dimensional spatial coordinates of each amino acid node in the global coordinate system, and obtain the encoding vectors representing the direction, orientation, and orientation information. Concatenate these three encoding vectors to obtain the coordinate hot encoding representation of this amino acid.

[0005] Step 2: Next, a spatial topological graph is constructed for the drug data, based on a topological message passing mechanism for drug molecule graph construction and feature extraction. To accurately capture the microscopic chemical environment and local pharmacophores of candidate drugs, this invention transforms the input drug molecule sequence into a microscopic chemical topological graph, in which the physicochemical properties of heavy atoms are extracted as node features, and covalent bond properties are extracted as edge features. Subsequently, a graph message passing neural network is used to perform multi-layer information iteration on the topological graph, enabling each atomic node to capture deep local topological subgraph information by continuously aggregating the features of neighboring nodes and connected edges. Finally, through a global feature pooling operation with permutation invariance, the deep features of all atoms are spatially aggregated, ultimately generating a global embedding vector of the drug containing global topological semantics and local pharmacophore features, completing the high-dimensional spatial feature mapping of the drug molecule.

[0006] Step 3: Input the target protein encoding vector from Step 1 into a three-dimensional isotropic graph neural network for deep characterization. This network utilizes the physicochemical scalar features of amino acid nodes and molecular nodes, as well as the relative distance between their three-dimensional coordinates, for isotropic message passing. While strictly ensuring the isotropic nature of macroscopic physical space rotation and translation, it synchronously updates the three-dimensional coordinates and high-dimensional features of the nodes. Through multi-layer network iteration, it fully captures the extremely complex spatial folding geometry and nonlinear dependencies of amino acids within macromolecules. Finally, through a global self-attention pooling mechanism, the residue features containing rich topological and spatial information are weighted and aggregated to generate a global embedding vector of the target protein that comprehensively characterizes the receptor's microscopic binding environment and global conformation.

[0007] Step 4: Input the protein global embedding vector from Step 3 and the drug global embedding vector from Step 2 into the Poincaré sphere model. First, the Euclidean features are losslessly projected into a hyperbolic space with constant negative curvature using a manifold index mapping. Then, the hyperbolic distance between drug and target features is accurately calculated using the Möbius summation method. Finally, the hyperbolic distance is nonlinearly transformed into the drug-target association binding probability using the Fermi-Dirac distribution function, thereby achieving a novel action association prediction with both high resolution and high generalization ability.

[0008] A drug target association identification method based on spatial graphs and isovariant networks is proposed. Step 1 is implemented as follows: In the target protein preprocessing stage, this method directly constructs a spatial topological graph based on a three-dimensional coordinate system. First, the amino acid sequence of the receptor protein is input into a large-scale structure prediction model of a large-scale evolutionary model protein folding prediction system for end-to-end deduction to obtain the three-dimensional spatial coordinates of each key residue and core atom. Then, each amino acid in the protein sequence is defined as a node in the graph network, and its inherent physicochemical properties are extracted as basic scalar features. These are then deeply integrated with the geometric vector features of three-dimensional spatial coordinates to construct a spatially aware composite node feature. Based on this, according to the actual distribution of amino acid nodes in three-dimensional space, the Euclidean distance between any two nodes is calculated, and a reasonable spatial proximity threshold is set using the K-nearest neighbor algorithm. When the actual distance between nodes is less than this preset threshold, a topological connection edge is established between them to accurately characterize and represent the potential physical contact or chemical bonding interactions of amino acids in the three-dimensional folding conformation of macromolecules.

[0009] A drug target association identification method based on spatial graphs and isovariant networks is described in step 2 as follows: While completing the spatial preprocessing of the target protein, this invention requires high-dimensional spatial feature mapping of the input candidate drug molecules. In order to accurately capture the microscopic chemical environment and local pharmacophores of drug molecules, this step uses cheminformatics tools to parse the input drug molecule sequence and convert it into a microscopic chemical topology graph. Each heavy atom in the molecule is regarded as a graph node and its physicochemical properties such as atom type, degree, formal charge, and hybrid orbital are extracted as initial node features. At the same time, the covalent bonds between heavy atoms are defined as connected edges and the bond type, stereoconfiguration, and ring state are extracted as initial edge features. Subsequently, a graph message-passing neural network is used to perform multi-layer feature aggregation on the constructed topological graph. In each network iteration, each atomic node actively receives and aggregates feature information from its first-order neighboring nodes and corresponding connected edges using a learnable message extraction function. The received neighboring pharmacophore information is then non-linearly fused with the current atomic state using a state update function. After multiple deep message-passing iterations, each atomic node successfully captures multi-order deep local topological subgraph information centered on itself. Finally, since different drug molecules contain varying numbers of atoms, this invention introduces a global feature pooling operation with graph node arrangement invariance. Through global average pooling, max pooling, or self-attention weighting, the deep features of all atoms are spatially aggregated, ultimately generating a fixed-dimensional global drug embedding vector. This vector highly condenses the global topological semantics and local pharmacophore features of the drug molecule, thus successfully completing the high-dimensional spatial feature mapping of the drug molecule and providing input for the subsequent target association strength module.

[0010] A drug target association identification method based on spatial graphs and isovariant networks is proposed. Step 3 is implemented as follows: First, drug molecule data with microscopic chemical topology and target protein geometric maps are simultaneously input into a three-dimensional isovariant graph neural network. Multi-level isovariant message passing is performed using the physicochemical scalar features of amino acid nodes and their relative distances in the three-dimensional coordinate system. Second, while maintaining the macroscopic physical space rotation and translation isovariance, the three-dimensional physical coordinates and high-dimensional semantic features of the receptor macromolecular nodes are updated synchronously using dynamic momentum adjustment. Then, through iterative iteration of a multi-layer deep network, the complex microscopic spatial folding geometric features within the macromolecule and the nonlinear dependence of depth between amino acid residues are accurately fitted. Finally, through a global self-attention pooling mechanism, the extracted deep residue features are weighted and aggregated to output a global embedding vector of the target protein that can faithfully characterize the receptor's microscopic binding environment and macroscopic dynamic conformation.

[0011] A drug target association identification method based on spatial graphs and isovariant networks is proposed. Step 4 is implemented as follows: The previously acquired global embedding vectors of drugs and targets are losslessly projected from a flat Euclidean space to a hyperbolic space with a constant negative curvature Poincaré sphere model via a manifold exponential mapping, giving the features non-Euclidean geometric properties that adapt to the tree-like hierarchical structure of biomolecules. Next, the Möbius summation method is introduced to calculate the hyperbolic geometric distance between drug and target features in hyperbolic space with high fidelity, accurately reflecting their deep structural similarity and true correlation. Finally, the Fermi-Dirac distribution function is used as a nonlinear classifier to transform the calculated hyperbolic distance into the final drug-target interaction probability, thus achieving a combined association prediction with both high resolution and high generalization ability.

[0012] This method can significantly reduce the time and resource consumption of blind trial and error in traditional R&D, and provide a powerful computing engine and technical support for high-throughput virtual drug screening, drug repositioning, and new target discovery in precision medicine. Attached Figure Description

[0013] Figure 1 This is a flowchart of a drug target association identification method based on spatial maps and isovariant networks.

[0014] Figure 2 This is a flowchart for constructing a spatial topological map of a protein using its three-dimensional atomic coordinates.

[0015] Figure 3 It is a flowchart for generating drug characterization using a topological subgraph message passing mechanism.

[0016] Figure 4 This is a structural diagram of an equivariant neural network.

[0017] Figure 5 This is a flowchart for predicting associated force quantities using hyperbolic space. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and examples.

[0019] This invention proposes a drug target association identification method based on spatial maps and isovariant networks, the specific process of which is as follows: Figure 1 The process involves four steps: constructing a spatial topological map of the protein using three-dimensional atomic coordinates, generating drug characterization using a topological subgraph message passing mechanism, inputting the data into an isovariant neural network for training, and predicting associated force magnitudes through hyperbolic space. The specific steps are as follows:

[0020] Step 1: Construct a spatial topological map of the protein using its three-dimensional atomic coordinates, as shown below. Figure 2 As shown, in the target protein pretreatment stage, in order to preserve the amino acid sequence and its true physical spatial conformation features without loss, a spatial topology map is constructed based on a three-dimensional coordinate system. First, the amino acid sequence of the receptor protein is input into a large-scale protein folding prediction system based on an evolutionary scale model for end-to-end deduction. This allows for the rapid acquisition of key residues without relying on costly experimental methods. Precise three-dimensional coordinates of atoms Subsequently, each amino acid in the sequence is defined as a graph node. Extract its physicochemical properties as scalar features. and with three-dimensional coordinates This geometric vector feature is deeply bound to form a composite node feature. Simultaneously, based on the actual distribution of nodes in three-dimensional space, the Euclidean distance between any two nodes is calculated. The K-nearest neighbor algorithm is used to set the spatial proximity threshold. Establish topological connections between nodes whose distance is less than the threshold. This allows for precise characterization of the potential physical or chemical interactions of amino acids in their three-dimensional folded conformations.

[0021] Step 2: Generate drug characterization using a topological subgraph message passing mechanism, such as... Figure 3 As shown: First, while constructing the spatial geometry map of the target protein, multi-level feature extraction is performed on the input candidate drug molecules. To accurately capture the local pharmacophores and global chemical environment of the drug molecules, this step abandons the traditional linear sequence extraction network based on one-dimensional strings and instead adopts a molecular topology map construction and message passing mechanism. First, using cheminformatics analysis tools, the input drug molecule sequence is transformed into a microscopic topology map that conforms to its true chemical structure. In this process, each heavy atom in the molecule is defined as a graph node. The initial node feature vector is constructed by extracting intrinsic physicochemical properties, including atom type, atomic degree, formal charge, hybridization type, and whether it is in an aromatic ring; at the same time, the covalent bonds between atoms are defined as the topological connected edges of the graph. The covalent bond attributes, including bond type, stereochemical configuration, and cyclic state, are extracted as edge feature vectors. Subsequently, to deeply explore the complex chemical environment and local functional groups within the drug, this invention employs a graph message-passing neural network to analyze the molecular graph. Multi-level iterative updates are performed. In each iteration, each atomic node receives and aggregates chemical topological information from its first-order neighboring atoms and connecting edges. Then, the aggregated neighborhood pharmacophore information is non-linearly fused with the current atom's features through a state update function. After multi-level deep message-passing iterations, each atomic node can capture the topological subgraph structure information within a multi-order range centered on itself. Finally, a global pooling operation with permutation invariance is used to spatially aggregate the deepest features of all atomic nodes, ultimately generating a fixed-dimensional drug global embedding vector that contains global topological semantics and local pharmacophore features. This allows for the precise mapping of candidate drug molecules into high-dimensional vectors rich in microscopic chemical structural information.

[0022] Step 3: Input the protein vector into the equivariant neural network for training. Figure 4 As shown: First, based on spatial distance and scalar features, equivariant message passing is performed. In the first layer of the EGNN network, for any two amino acid residues in the protein map, the connected nodes... and Extract its current scalar features , and the square of the relative distance in three-dimensional space. These three elements are concatenated and then input into the message extraction function constructed by the multilayer perceptron. In the context of communication messages between computing nodes : This message item relies only on relative distances rather than absolute coordinates, thus ensuring physical invariance to spatial rotations and translations. Next, to capture potential conformational fine-tuning and dynamic changes in macromolecules within feature space, this network performs isovariant updates on the three-dimensional coordinates of amino acid residues by introducing a coordinate update function. Based on the messages exchanged between nodes For nodes Momentum adjustment based on spatial position: ,in, For nodes The set of neighboring nodes, the update rule is based on the difference between the original coordinate vectors. Using the reference direction, it is ensured that the new coordinates maintain strict isotropy even when the entire macroscopic system undergoes rotation or translation. Subsequently, nonlinear fusion and updating of node scalar features are performed; after obtaining the aggregated message, a feature update function is used. Node Its own historical characteristics The scalar features of the amino acid are updated by performing a non-linear mapping with the aggregated messages received from all its neighboring nodes: ,go through The isomorphic graph network of the layers iterates, with each amino acid node... scalar characteristics All methods deeply integrate the physicochemical properties and three-dimensional spatial topological geometric information of their higher-order neighborhoods. Finally, target global spatial feature pooling is performed. After extracting the high-level isovariant spatial features of each amino acid residue, a global self-attention pooling mechanism is introduced to assign greater weights to key residues that contribute significantly to the binding pocket. The deep features of all nodes are weighted and aggregated, ultimately outputting a high-dimensional, dense target global embedding vector that contains the protein's global three-dimensional conformation and folding state. .

[0023] Step 4: The process of predicting associated force quantities through hyperbolic space is as follows. Figure 5 As shown: First, a manifold mapping is performed from Euclidean space to hyperbolic space, and the negative curvature parameter of the Poincaré sphere is set to... The characteristics of drugs in Euclidean tangent space With target features Hyperbolic feature representations are generated by projecting the data onto the hyperbolic space manifold using an exponential mapping function. and Taking the exponential mapping at the origin 0 as an example, the mapping formula is as follows: ,in, The Euclidean features representing the input drug or target are mapped to the non-Euclidean geometry of hyperbolic space while preserving local geometric properties. A high-fidelity distance metric is then performed in hyperbolic space. Since the space capacity grows exponentially with the radius in hyperbolic space, it is naturally suitable for accommodating hierarchical molecular distributions. The Möbius summation method is introduced to calculate drug features under the Poincaré sphere model. With target features hyperbolic distance between : ,in, This represents the Möbius vector addition on a hyperbolic manifold, and the distance accurately reflects the true correlation between drug molecules and target proteins in deep biochemical logic and phylogenetic trees. Finally, a nonlinear transformation of the binding affinity probability is performed, introducing the Fermi-Dirac distribution as a nonlinear classifier to transform the calculated hyperbolic distance. Predicted probability of drug-target interaction :

[0024] ,in, and These are learnable hyperparameters, representing the distance decision threshold and the smoothness of the probability distribution, respectively. Finally, the predicted probability is output. This represents the final combined association prediction result.

[0025] To comprehensively evaluate the practical performance of the interaction correlation prediction model constructed in this invention in drug-target interaction tasks, systematic validation was conducted on the authoritative BIOSNAP benchmark dataset. Experimental results show that the method of this invention exhibits excellent discriminative ability on the test set, with its AUPR index reaching a breakthrough of 0.942. Compared with the Hyperattention framework based on deep graph neural networks and multi-head attention, which has an AUPR of 0.931, this represents an improvement of approximately 1.2%. This performance leap strongly confirms that by deeply fusing high-dimensional physical space features and nonlinear topological semantics, this invention can capture the complex microscopic binding patterns between drugs and targets with extremely high accuracy. Especially in cold-start screening scenarios where known interaction data is lacking, the model demonstrates extremely strong generalization robustness, providing a high-value and highly reliable computational engine for accelerating the discovery of high-throughput lead compounds and the accurate mining of potential new targets.

[0026] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A drug target association identification method based on spatial graphs and isovariant networks, characterized in that, The specific steps are as follows: Step 1: Construct a spatial topological map of the protein using its three-dimensional atomic coordinates. Utilize a protein folding prediction system based on a large evolutionary-scale model to obtain the true three-dimensional spatial coordinates of each amino acid node in the global coordinate system, thus representing... direction, direction and The directional information is encoded by a vector. Concatenating these three encoded vectors gives the coordinate-hot encoded representation of this amino acid. Step 2: Construct a spatial topology map of the drug data, and use a graph message passing neural network to perform multi-layer information iteration on the topology map; Step 3: Input the target protein coding vector into a three-dimensional isovariant graph neural network for deep representation, and generate a global embedding vector of the target protein through a global self-attention pooling mechanism. Step 4: Input the global embedding vector of the protein and the global embedding vector of the drug into hyperbolic space, use the Möbius summation method to accurately calculate the hyperbolic distance between the drug and target features, and calculate the drug-target association binding probability.

2. The drug target association identification method based on spatial map and isovariant network according to claim 1, characterized in that, Based on the actual distribution of nodes in three-dimensional space, the distance to drug targets is calculated, and the process of constructing a spatial topological map of proteins based on a three-dimensional coordinate system is as follows: First, the amino acid sequence of the receptor protein is input into a large-scale protein folding prediction system based on an evolutionary scale model for end-to-end deduction to obtain key residues and... Precise three-dimensional coordinates of atoms Subsequently, each amino acid in the sequence is defined as a graph node. Extract its physicochemical properties as scalar features. and with three-dimensional coordinates This geometric vector feature is deeply bound to form a composite node feature. Simultaneously, based on the actual distribution of nodes in three-dimensional space, the Euclidean distance between any two nodes is calculated. The K-nearest neighbor algorithm is used to set the spatial proximity threshold. Establish topological connections between nodes whose distance is less than the threshold. This allows for precise characterization of the potential physical or chemical interactions of amino acids in their three-dimensional folded conformations.

3. The drug target association identification method based on spatial map and isovariant network according to claim 1, characterized in that, Equivariant message passing for protein vectors based on spatial distance and scalar features: For any two amino acid residues in the protein map, connect nodes... and Extract its current scalar features , and the square of the relative distance in three-dimensional space. These three elements are concatenated and then input into the message extraction function constructed by the multilayer perceptron. In the context of communication messages between computing nodes : Next, in order to capture the potential conformational fine-tuning and dynamic changes of macromolecules in the feature space, this network performs isovariant updates on the three-dimensional coordinates of amino acid residues by introducing a coordinate update function. Based on the messages exchanged between nodes For nodes Momentum adjustment based on spatial position: ,in, For nodes The set of neighboring nodes is then used for nonlinear fusion and updating of node scalar features. After obtaining the aggregated message, a feature update function is used. Node Its own historical characteristics The scalar features of the amino acid are updated by performing a non-linear mapping with the aggregated messages received from all its neighboring nodes: ,go through The isomorphic graph network of the layers iterates, with each amino acid node... scalar characteristics All of them deeply integrate the physical and chemical properties of their higher-order neighborhoods with the topological and geometric information of three-dimensional space.

Images