High-precision drug screening method, system and equipment based on hierarchical multi-modal representation learning and readable storage medium

Abstract

The invention relates to the technical field of computational biology, and discloses a high-precision drug screening method and system based on hierarchical multi-modal representation learning, and the method comprises the steps: carrying out the feature extraction of an amino acid sequence of a target protein through a plurality of pre-trained protein language models; based on the three-dimensional structure of the target protein, multi-scale graph representation and ligand molecular graph representation of the protein are constructed; performing multi-modal feature learning on the multi-scale graph representation and the ligand molecular graph representation of the constructed protein by using a neural network based on message passing; and based on the multi-modal features, constructing a binding affinity prediction model for drug screening, and completing a drug screening task. A complex interaction relationship in a protein and ligand graph structure is effectively captured by using a multilayer graph neural network, and efficient coding of protein-ligand pair features is realized. The surface, the structure and the ligand embedding vector of the protein can be extracted, and the accuracy and the efficiency of protein-ligand interaction prediction are improved.

Description

Technical Field

[0001] The present invention relates to the field of computational biology technology, and specifically to a high-precision drug screening method, system, device and readable storage medium based on hierarchical multimodal representation learning. Background Art

[0002] Drug screening is an important step in new drug development. Its goal is to quickly screen out highly effective compounds that can bind to target proteins from a large library of compounds. The binding affinity between protein and ligand is a key indicator for measuring drug potential. Traditional drug screening methods mainly rely on high-throughput experimental techniques, such as enzyme inhibition experiments and surface plasmon resonance experiments. Although these methods are accurate, they are time-consuming and costly, limiting the feasibility of large-scale applications.

[0003] In recent years, computational virtual screening methods have received widespread attention. These methods can rapidly evaluate the binding affinity of compounds by simulating the physical and chemical interactions between proteins and ligands, significantly improving the efficiency of drug screening. However, traditional computational models (such as molecular force fields and molecular docking) have limitations in dealing with multi-scale properties in complex biological systems. For example, these methods often have difficulty simultaneously capturing the sequence, structure, and surface features of proteins, as well as the molecular properties of ligands.

[0004] With the development of artificial intelligence technology, deep learning methods have shown great potential in the field of drug screening. In particular, the application of message passing neural network (MPNN) in molecular graph representation learning makes the prediction of protein-ligand interaction more accurate. However, the current models based on single modal features have certain shortcomings in practical applications, such as the inability to effectively integrate the multi-scale features of proteins or ignoring the global information of ligand molecular graphs. Therefore, developing a multimodal deep learning method that can integrate protein sequence, structure, surface features and ligand molecular information has become one of the key challenges to improve the accuracy of drug screening.

[0005] In recent years, the rapid development of protein language models (PLM) and multimodal representation learning techniques has provided new tools for this field. For example, pre-trained models such as Ankh, ESM-1b, and ProtTrans can extract deep semantic and evolutionary information from protein sequences, while triangular mesh-based protein surface modeling tools (such as MSMS) lay the foundation for efficient representation of surface features. At the same time, the combination of MPNN's message passing mechanism and multilayer perceptron (MLP) regressor makes it possible to fuse multimodal features for binding affinity prediction. Summary of the invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by the present invention is: the feature encoding problem in protein-ligand interaction prediction, especially the extraction of efficient and accurate embedding vectors in complex graph structure data.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions: a high-precision drug screening method based on hierarchical multimodal representation learning, comprising:

[0009] Through multiple pre-trained protein language models, feature extraction is performed on the amino acid sequence of the target protein;

[0010] Based on the three-dimensional structure of the target protein, construct a multi-scale graph representation of the protein and a ligand molecular graph representation;

[0011] Use message passing neural networks to construct multi-scale graph representations of proteins and ligand molecular graph representations, and perform multimodal feature learning;

[0012] Based on multimodal features, a binding affinity prediction model for drug screening is constructed to complete the drug screening task.

[0013] As a preferred solution of the high-precision drug screening method based on hierarchical multimodal representation learning described in the present invention, the multiple pre-trained protein language models include Ankh, ESM-1b and ProtTrans; feature extraction of protein amino acid sequences includes using the above three models to perform feature representation on amino acid sequences of length L respectively; the three feature matrices are fused to generate a mixed embedding matrix M B :

[0014]

[0015] Among them, M A represents the amino acid sequence feature matrix extracted by the Ankh model; M E represents the amino acid sequence feature matrix extracted by the ESM-1b model, M P represents the amino acid sequence feature matrix extracted by the ProtTrans model; M B represents the mixed embedding matrix after fusing the three feature matrices, It means to connect two feature matrices "row by row".

[0016] As a preferred solution of the high-precision drug screening method based on hierarchical multimodal representation learning described in the present invention, wherein: the construction of the multi-scale graph representation of the protein and the ligand molecule graph representation includes constructing a residue-level structural graph representation based on the three-dimensional structure of the protein;

[0017] Represent the three-dimensional structure of a protein as a residue-level graph G c =(V c ,E c );

[0018] Among them, V c = {v 1 ,…,v L} and E c ={e 1 ,…,e N} respectively correspond to Figure G c The set of L nodes and N edges in ;

[0019] By judging whether the distance between the Cα atoms of any two amino acid residues is less than To determine each edge in the residue-level graph;

[0020] Among them, the i-th node v i For any sequence S p The i-th residue in , 1≤i≤L;

[0021] Node v i The eigenvector of The formula is expressed as:

[0022]

[0023] When node v i Yes K i When there are 1 neighbor node, the edge e ij Connect node v i and v j ;

[0024] Among them, v j ∈V c ,e ij ∈E c ,j≤K i <L;

[0025] Edge ij The eigenvector ef ij The formula is defined as:

[0026] ef ij =[d ij ,θ ij ]

[0027] Among them, dij Indicates amino acid v i and v j The distance between ij Indicates v i and v j The angle between the normal vectors of

[0028] Use triangulated mesh generation tools to construct surface representations of proteins;

[0029] For a protein p, its surface was constructed with a triangular mesh by using MSMS software;

[0030] Combine all the grids to get the surface map G F =(V F ,E F );

[0031] Among them, V F represents the set of vertices extracted from all meshes, E F represents the set of edges extracted from all meshes;

[0032] Surface G F Any vertex u in j ∈V F , the eigenvector nf of the surface graph uj Initialize to the closest structure graph G c Middle residue node v j The characteristic vector of is expressed as:

[0033]

[0034] The edge e on the triangle mesh jk The eigenvector ef jk The formula is expressed as:

[0035] ef jk =[d jk ,g jk ]

[0036] Among them, d jk Represents vertex u j and u k The distance between jk Represents a 6-dimensional vector, indicating the edge e jk The six geometric parameters between two adjacent meshes are as follows, including the dihedral angle θ between two adjacent faces, the internal angle (θ a ,θ b ), the ratio of the side length to the vertical height of each adjacent face (d / h a ,d / h b ), and the angle θ(nk ,n j );

[0037] Use the SMILES string of the ligand to construct a molecular graph representation;

[0038] For a given 1D ligand SMILES string, use RDKit to obtain the molecular structure, represented as a 2D ligand graph G L =(V L ,E L );

[0039] Among them, V L represents the set of vertices extracted from the ligand graph, E L represents the set of edges extracted from the ligand;

[0040] 2D ligand maps include, based on the embedding and physicochemical properties of atom types, connected by covalent bonds;

[0041] Atomic characteristics include atomic symbol, atomic degree, implicit valence, explicit valence, and "whether it belongs to an aromatic ring";

[0042] The characteristics of a bond mainly include the bond type, whether it is a conjugated bond, and whether it belongs to a ring.

[0043] As a preferred solution of the high-precision drug screening method based on hierarchical multimodal representation learning described in the present invention, wherein: multimodal feature learning includes inputting a protein structure map, a surface map and a ligand molecule map, and performing multi-layer message transmission;

[0044] Construct an N-layer message passing neural network, from the structure graph G c , surface graph G F and ligand G L Generate different protein-ligand pair signature codes in

[0045] An N-layer message passing neural network, whose input is defined as a graph G = (V, E);

[0046] The i-th node v in the graph G i The set of neighbor nodes of ∈V is denoted as NP(v i ), the formula for message initialization of this node is expressed as:

[0047]

[0048] Among them, "0" represents the initial state of the network model at the moment t=0, Represents node v i At level 0, the message vector Represents node v i The eigenvector of

[0049] Each layer node v of the message passing neural network i The message passing formula is expressed as:

[0050]

[0051] in, Represents node v i In the message vector of the kth layer, τ represents the activation function, U 1 represents the weight matrix, U 2 represents the weight matrix, NP(v i ) represents node v i The set of neighbor nodes of Represents node v i and its neighbor node v j Attention weights at layer k-1, Q represents the weight matrix, Represents node v i The eigenvector of and edge ij The eigenvector ef ij Splicing is performed, where N represents the total number of layers of message transmission, from layer 1 to layer N; Represents node v i The initial eigenvector, ef ij Represents edge e ij The feature vector of the connected node v i and v j a represents the scalar coefficient, W represents the weight matrix, ev represents the final feature embedding vector, W (1) ,W (2) ,W (3) represents the learnable weight matrix, ev f Embedding vector representing protein surface features, ev c The embedding vector representing the protein structural features, ev l Embedding vector representing the molecular features of the ligand.

[0052] As a preferred solution of the high-precision drug screening method based on hierarchical multimodal representation learning described in the present invention, the multimodal feature learning also includes, in the message transmission process, introducing an attention mechanism to dynamically adjust the information weights of different neighbor nodes, which is expressed as follows:

[0053]

[0054] in, Indicates that at the k-1 layer, node v i Its neighbor node v jThe attention weight between them; exp represents the natural exponential function, LReLU represents the LeakyReLU activation function, Represents node v i Its neighbor node v j The attention pre-value at the k-1th layer, NP(v i ) represents node v i The set of neighbor nodes of Represents node v i With node v q The previous attention value at the k-1th layer; Represents node v i and its neighbor node v j The attention pre-value at the k-1th layer, a represents the scalar coefficient, W represents the learnable weight matrix, Represents node v i The message vector at level k-1, Represents node v j The message vector at the k-1th layer; || represents the concatenation operation;

[0055] After N layers of message passing, the structural embedding vector, surface embedding vector and ligand embedding vector of the protein are extracted, and the formula is expressed as:

[0056]

[0057] Among them, ev represents the final feature embedding vector, NP(v i ) represents node v i The set of neighbor nodes, W (1) represents the learnable weight matrix, Represents node v j The message vector at level N, W (2) represents the learnable weight matrix, ef ij Represents edge e ij The eigenvector of (3) represents the learnable weight matrix, Represents node v i In the message vector at level N, V ​​represents the set of all nodes in the graph G;

[0058] For any protein-ligand pair, the MPNN module outputs three embedding vectors of length 200; the three embedding vectors of length 200 include surface vector, structure vector and ligand vector.

[0059] As a preferred scheme of the high-precision drug screening method based on hierarchical multimodal representation learning described in the present invention, constructing a binding affinity prediction model for drug screening includes splicing surface vectors, structure vectors and ligand vectors to generate a unified multimodal feature representation of protein-ligand pairs.

[0060] As a preferred solution of the high-precision drug screening method based on hierarchical multimodal representation learning described in the present invention, wherein: constructing a binding affinity prediction model for drug screening also includes inputting a unified multimodal feature representation of a protein-ligand pair into a multi-layer perceptron regression model to predict the binding affinity value between the protein and the ligand;

[0061] The multilayer perceptron regression model is trained by minimizing the mean square error. The formula is expressed as:

[0062]

[0063] in, Represents the predicted value, y i represents the true value, and MSE represents the multilayer perceptron regression model.

[0064] A high-precision drug screening system based on hierarchical multimodal representation learning, where:

[0065] The feature extraction module extracts features from the amino acid sequence of the target protein through multiple pre-trained protein language models;

[0066] The graph representation module constructs a multi-scale graph representation of the protein and a ligand molecular graph representation based on the three-dimensional structure of the target protein;

[0067] The multimodal feature module uses a message passing neural network to construct a multi-scale graph representation of proteins and a ligand molecular graph representation, and conducts multimodal feature learning;

[0068] The screening module builds a binding affinity prediction model for drug screening based on multimodal features to complete the drug screening task.

[0069] A computer device comprises: a memory and a processor; the memory stores a computer program, wherein the processor implements the steps of any one of the methods of the present invention when executing the computer program.

[0070] A computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of any one of the methods of the present invention.

[0071] Beneficial effects of the present invention: The high-precision drug screening method based on hierarchical multimodal representation learning provided by the present invention effectively captures the complex interaction relationship in the protein and ligand graph structure by using a multi-layer graph neural network (MPNN), thereby realizing efficient encoding of protein-ligand pair features. This method can extract the surface, structure and ligand embedding vectors of proteins, improving the accuracy and efficiency of protein-ligand interaction prediction. At the same time, by introducing the attention mechanism and graph convolutional network, the information transfer process is optimized, so that the model has stronger robustness and generalization ability when processing large-scale molecular graphs. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0073] Figure 1 An overall flow chart of a high-precision drug screening method based on hierarchical multimodal representation learning provided in the first embodiment of the present invention;

[0074] Figure 2 The overall framework diagram of a high-precision drug screening method HMRLBA based on hierarchical multimodal representation learning provided in the fourth embodiment of the present invention;

[0075] Figure 3 The ablation experiment results of HMRLBA of a high-precision drug screening method based on hierarchical multimodal representation learning provided in the fourth embodiment of the present invention;

[0076] Figure 4 Analysis of the contribution of different language models in a high-precision drug screening method HMRLBA based on hierarchical multimodal representation learning provided in the fourth embodiment of the present invention;

[0077] Figure 5 The experimental results of comparing a high-precision drug screening method HMRLBA based on hierarchical multimodal representation learning provided in the fourth embodiment of the present invention with mainstream methods;

[0078] Figure 6 The prediction results and combined visualization of difficult samples in the test set by a high-precision drug screening method HMRLBA based on hierarchical multimodal representation learning provided in the fourth embodiment of the present invention;

[0079] Figure 7A flowchart of the application of a high-precision drug screening method HMRLBA based on hierarchical multimodal representation learning in EGFR virtual drug screening provided in the fourth embodiment of the present invention. DETAILED DESCRIPTION

[0080] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0081] Example 1, reference Figure 1 , which is an embodiment of the present invention, provides a high-precision drug screening method based on hierarchical multimodal representation learning, comprising:

[0082] S1: Feature extraction of the amino acid sequence of the target protein is performed through multiple pre-trained protein language models.

[0083] For the PDB file of protein p, first use Biopython to extract its amino acid sequence S p , the sequence length is L. In the protein language model system, the protein sequence S can be transformed into p Represented as three embedding matrices M A ,M E , and M P The sizes of these three matrices are L×1536, L×1280, and L×1024 respectively.

[0084] (2) Concatenate the embedded feature matrices to generate a mixed embedding matrix

[0085] The above three embedding matrices are merged into a hybrid embedding matrix M B :

[0086]

[0087] Furthermore, by splicing the embedding feature matrices from different pre-trained protein language models (Ankh, ESM-1b and ProtTrans), the advantages of multiple models are integrated to fully utilize the different understanding and expression capabilities of different models for amino acid sequence features. Through this mixed embedding matrix, the multi-dimensional information of protein sequences can be captured more comprehensively, improving the accuracy and robustness of downstream tasks (such as protein structure prediction, functional annotation, etc.). At the same time, this method can integrate the learning capabilities of different models, enhance the diversity of feature expression, and improve the model's ability to recognize protein structure and function.

[0088] S2: Based on the three-dimensional structure of the target protein, construct a multi-scale graph representation of the protein and a ligand molecular graph representation.

[0089] Constructing a multi-scale graph representation of proteins can fully capture the structural features and geometric relationships inside and outside the protein. This information plays a key role in understanding the biological functions of proteins and their interactions with ligands. The residue-level structural graph representation can intuitively display the spatial position relationship and local geometric characteristics between amino acids, while the protein surface graph representation constructed using a triangular mesh generation tool depicts the spatial characteristics of the protein's exterior. This multi-scale representation provides a solid foundation for subsequent feature extraction and learning. In addition, combining the multi-scale features of proteins with the molecular graph representation of ligands can fully reflect the multimodal characteristics of the protein-ligand system, laying an important foundation for the fusion of multimodal features in affinity prediction.

[0090] The construction of residue-level structural diagram representation based on protein three-dimensional structure includes:

[0091] The three-dimensional structure of a protein provides internal details including distances, angles, torsion angles, and other high-order terms. Here, we represent the structure of a protein as a residue-level graph G c =(V c ,E c ), where V c = {v 1 ,…,v L} and E c ={e 1 ,…,e N} respectively correspond to a set of L nodes and N edges. c Each edge in the equation is determined by determining whether the distance between the Cα atoms of any two amino acid residues is less than To be determined.

[0092] The i-th node v i Corresponding to the sequence S p The i-th residue in , where 1≤i≤L. Therefore, the node v i The eigenvector of Initialize as:

[0093]

[0094] If node v i has K i neighbor nodes, then edge e ij connects node v i and v j (where v j ∈V c , e ij ∈E c , j ≤ K i <L). The eigenvector ef ij of edge e ij can be initialized as:

[0095] ef ij = [d ij , θ ij (3)

[0096] where d ij represents the distance between amino acids v i and v j ; in addition, θ ij represents the angle between the normal vectors of these two nodes.

[0097] Using a triangular mesh generation tool to construct the surface map representation of a protein includes:

[0098] The surface representation shows the external details of the protein. For a protein p, its surface is constructed using the MSMS software with a triangular mesh. Combining all the meshes, we can obtain a surface graph G F = (V F , E F ). Among them, V F and E F are the sets of vertices and edges extracted from all the meshes respectively. For any vertex u F in G j ∈V F , its eigenvector is initialized as the eigenvector of the residue node v c in its nearest structure graph G j :

[0099]

[0100] The eigenvector ef jk of the edge e jk on the triangular mesh can be initialized as:

[0101] ef jk=[d jk , g jk ] (5)

[0102] Among them, d jk is the vertex u j and u k The distance between jk Defined as a 6-dimensional vector, representing the edge e jk Six geometric parameters associated between two adjacent meshes.

[0103] The molecular graph representation constructed using the SMILES string of the ligand includes:

[0104] For a given 1D ligand SMILES string, HMRLBA uses RDKit to obtain the molecular structure and further represents it as a 2D ligand graph G L =(V L ,E L ). The graph contains embeddings and physical and chemical properties based on atom types, connected by covalent bonds. Atomic features include atomic symbol, atomic degree, implicit valence, explicit valence, and "whether it belongs to an aromatic ring"; while bond features mainly include bond type, "whether it is a conjugated bond", and "whether it belongs to a ring".

[0105] Furthermore, by constructing a multi-scale graph representation of the protein and a molecular graph representation of the ligand, it is possible to fully capture the internal and external structural features of the protein and its geometric relationship with the ligand, providing rich structural information for subsequent feature extraction and learning. The residue-level structural graph can show the spatial relationship between amino acids, and the surface graph provides the geometric features of the outside of the protein, thus providing a more accurate basis for understanding protein-ligand interactions. The molecular graph representation of the ligand fully reflects the chemical properties of the ligand through modeling of molecular structure and physicochemical properties. Combining these two multimodal features provides key multi-dimensional information support for subsequent protein-ligand affinity prediction and other biological tasks.

[0106] S3: Use message passing neural networks to construct multi-scale graph representations of proteins and ligand molecular graph representations for multimodal feature learning.

[0107] Input protein structure diagram, surface diagram and ligand molecule diagram for multi-layer message transmission

[0108] We construct a message passing neural network (MPNN) to transform the above structure graph G c , surface graph G F and ligand G L Different protein-ligand pair feature encodings are generated in . As a space-based convolutional network, MPNN is able to model the complex interactions between nodes and edges in the graph.

[0109] For the input graph G = (V, E), the processing steps of an N-layer MPNN model include the following. i ∈V, its neighbor node set is represented as NP(v i ), the message initialization of this node is performed according to the following formula:

[0110]

[0111] Next, the message passing from layer k-1 to layer k is calculated as follows:

[0112]

[0113] Where 1≤k≤N. τ is the τ activation function, U 1 , U 2 and Q is the weight matrix. In formula (7), from node v i to v j The attention score α ij The definition is as follows:

[0114]

[0115] The coefficient β in formula (8) ij It can be calculated by formula (9):

[0116]

[0117] Extracting protein structural embedding vectors, surface embedding vectors, and ligand embedding vectors includes:

[0118] Message aggregation of graph encoding vectors is implemented in the following way:

[0119]

[0120] where the operation ⊙ represents the Hadamard product, W (1) , W (2) and W (3) is a learnable weight matrix. For any protein-ligand pair, the MPNN module outputs three embedding vectors of length 200, which are the surface vector (ev f ), structure vector (ev c ) and the ligand vector (ev l ).

[0121] Furthermore, by introducing a multi-scale graph representation method based on a message passing neural network (MPNN), multimodal feature learning is effectively performed on protein structure graphs, surface graphs, and ligand molecular graphs. MPNN can model the complex interactions between nodes and edges in the graph, thereby extracting multi-level features of protein-ligand pairs. Through this method, embedded vectors of surfaces, structures, and ligands can be obtained, thereby improving the accuracy and robustness of protein-ligand interaction predictions. This design provides a more comprehensive and accurate feature representation for an in-depth understanding of protein-ligand interactions, which helps efficient drug screening and development in the biomedical field.

[0122] S4: Based on multimodal features, a binding affinity prediction model for drug screening is constructed to complete the drug screening task.

[0123] Generating unified multimodal feature representations

[0124] The surface vector (ev f ), structure vector (ev c ) and the ligand vector (ev l ) are spliced ​​to generate a unified multimodal feature representation of protein-ligand pairs.

[0125] The binding affinity value between protein and ligand is output through the multi-layer perceptron regression model for drug screening tasks

[0126] The multimodal feature representation of the above protein-ligand pair is input into the multi-layer perceptron (MLP) regression model to predict the binding affinity value between the protein and the ligand. The hidden layer of the MLP model contains 512 neuron units, and its output represents the binding strength between the protein and the corresponding ligand, which can be directly used for drug screening tasks. As a regression model, MLP is trained by minimizing the mean square error (MSE), and the specific formula is as follows:

[0127]

[0128] in, and i represent the predicted value and the true value respectively.

[0129] Furthermore, through a unified multimodal feature representation, the surface, structure and ligand embedding vector of the protein-ligand pair are combined to form a more comprehensive feature description, which improves the accuracy and generalization ability of the model in drug screening tasks. The prediction of binding affinity through the multi-layer perceptron (MLP) regression model can effectively quantify the interaction strength between protein and ligand, provide accurate affinity scores for drug screening, thereby accelerating the screening process of potential drug candidate molecules and reducing the cost and time of experimental screening.

[0130] Embodiment 2 is an embodiment of the present invention, which provides a high-precision drug screening system based on hierarchical multimodal representation learning, including:

[0131] The feature extraction module extracts features from the amino acid sequence of the target protein through multiple pre-trained protein language models.

[0132] The graph representation module constructs a multi-scale graph representation of the protein and a ligand molecular graph representation based on the three-dimensional structure of the target protein.

[0133] The multimodal feature module uses a message passing neural network to construct multi-scale graph representations of proteins and ligand molecular graph representations, and performs multimodal feature learning.

[0134] The screening module builds a binding affinity prediction model for drug screening based on multimodal features to complete the drug screening task.

[0135] Embodiment 3, an embodiment of the present invention, is different from the first two embodiments in that:

[0136] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0137] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0138] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0139] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0140] Example 4, the present invention proposes a high-precision drug screening method based on hierarchical multimodal representation learning, comprising the following steps: first, embedding the protein amino acid sequence through multiple pre-trained protein language models to generate a mixed embedding matrix, retaining the evolutionary information of different data sources; then, constructing a multi-scale graph representation of the protein (including a residue-level structure graph based on a three-dimensional structure and a surface graph generated by a triangular mesh) and a molecular graph representation of the ligand to capture the internal and external features of the protein and the chemical information of the ligand; then, using a message passing neural network (MPNN) to learn multimodal features of the protein structure graph, surface graph and ligand graph, extracting the embedding vectors of the protein and the ligand; finally, the multimodal features are unified and fused, input into a multi-layer perceptron (MLP) regression model, and the binding affinity value between the protein and the ligand is output for drug screening tasks. By combining the sequence, structure and surface features of the protein, and the multimodal information of the ligand molecule, this method significantly improves the accuracy and efficiency of drug screening and has broad application prospects.

[0141] Figure 1The overall framework diagram of the high-precision drug screening method HMRLBA described in the present invention. HMRLBA includes four main steps: protein amino acid sequence feature extraction, protein multi-scale graph representation and ligand molecular graph representation construction, multimodal feature learning, and the construction of a binding affinity prediction model. The framework diagram describes in detail the generation of protein sequence features based on a pre-trained language model, the construction of protein surface graphs, structure graphs and ligand molecular graphs, the multimodal feature learning process based on a message passing neural network, and how to fuse multimodal features and input them into a multi-layer perceptron regression model, ultimately achieving high-precision prediction of binding affinity.

[0142] Figure 2 The ablation experiment results of the method HMRLBA described in the present invention are shown, showing the ablation experiment results of the HMRLBA method under different module removal conditions. The ablation experiment evaluates the impact of protein structure graph representation, surface graph representation, and the attention mechanism in MPNN on the overall performance of the model.

[0143] Figure 3 This is an analysis of the contribution of different language models in the HMRLBA method described in the present invention, showing the contribution of the pre-trained language models (Ankh, ESM-1b and ProtTrans) used in the HMRLBA method to protein sequence feature extraction. Through experimental comparison, the figure shows the impact of the features generated by each model on the binding affinity prediction performance, reflecting the advantages of multi-language model fusion and the complementarity of each model in feature expression ability.

[0144] Figure 4 The comparative experimental results of the HMRLBA method described in the present invention and the mainstream methods are shown, including the sequence-based methods DeepDTA, MGraphDTA and PSICHC, the structure-based method IEConv, the surface-based method MaSIF, and the multi-scale models HoloProt and HaPPy. The root mean square error RMSE, mean absolute error MAE, Pearson correlation coefficient, Spearman correlation coefficient and determination coefficient R 2 The results show that HMRLBA outperforms other methods in all indicators, among which it has the lowest error in RMSE and MAE, indicating that it has higher prediction accuracy; it has the highest values ​​in Pearson and Spearman correlation coefficients, respectively, reflecting its superior performance in linear and nonlinear relationship prediction; at the same time, its R 2 The value is significantly higher than that of other methods, showing stronger explanatory power. These results verify that the HMRLBA method has excellent performance and reliability in drug screening tasks.

[0145] Figure 5 The prediction results and binding visualization of the method HMRLBA described in the present invention for difficult samples in the test set. The binding site visualization of four "difficult samples" (complexes 1Q7A, 4QGD, 4TKJ and 5HZ6) using PyMOL software is shown, as well as the actual binding affinity and predicted affinity of the HMRLBA model. As can be seen from the figure, these ligands bind near the protein surface, and the prediction error between the affinity predicted by HMRLBA and the actual affinity is minimal, achieving very satisfactory results.

[0146] Figure 6 The flowchart of the application of the method HMRLBA described in the present invention in EGFR virtual drug screening. We first retrieved the SMILES structures of 2,616 FDA-approved small molecule drugs from the DrugBank database, including 13 EGFR inhibitors. At the same time, the PDB structure file of EGFR was obtained from the PDB database. These data were then input into the trained HMRLBA model to obtain the binding affinity prediction value. Table 1 shows the parameter settings of the method HMRLBA described in the present invention.

[0147] Table 1 Parameter setting table

[0148]

[0149]

[0150] Table 2 shows the ablation experiment results of the method HMRLBA described in the present invention. Specifically, we analyzed the contribution of each component to the model performance by removing the protein surface representation (woSurface), structure representation (woStructure), and attention mechanism (woAttention) in the MPNN module respectively. Except for the ablated components, the rest of the model architecture remains unchanged. The experimental results show that the performance of the HMRLBA model is significantly better than the three ablation models, among which removing the attention mechanism has the greatest impact on the performance, followed by removing the surface representation or structure representation. This is because the attention mechanism enhances the ability of weight distribution during message passing, the structure representation provides information about the three-dimensional conformation of the protein, and the surface representation captures the shape and geometric distribution of the protein surface. Therefore, we conclude that protein multimodal feature learning and attention mechanism can improve the accuracy of the model in predicting protein-ligand binding affinity.

[0151] Table 2 Ablation experiment results

[0152] Ablation Model RMSE↓ MAE↓ Pearson↑ Spearman↑ <![CDATA[R 2 ↑]]> wAttention 1.512±0.011 1.221±0.009 0.478±0.007 0.475±0.006 0.196±0.014 woStructure 1.495±0.011 1.198±0.008 0.519±0.006 0.516±0.006 0.207±0.012 woSurface 1.473±0.009 1.182±0.007 0.492±0.005 0.487±0.006 0.234±0.011 HMRLBA 1.428±0.007 1.146±0.005 0.544±0.004 0.537±0.004 0.280±0.008

[0153] Table 3 shows the contribution analysis of different language models in the method HMRLBA described in the present invention, and the contribution of three different protein language models used in HMRLBA is studied. We compared the experimental effects of seven feature embeddings in the sequence embedding module, including three independent embeddings from ProtTrans, ESM-1b and Ankh, and four hybrid embeddings created by connecting these embeddings: ProtTrans+ESM-1b (PE), ProtTrans+Ankh (PA), ESM-1b+Ankh (EA), and the combined embedding of the three (PEA=HMRLBA). The experimental results compared the performance of the seven feature embeddings, and the results showed that PEA significantly outperformed the other embedding methods. Among the three independent embeddings, ESM-1b performed the best. In addition, when ESM-1b was combined with PA, the model performance was significantly improved. Based on these findings, two conclusions can be drawn. First, the complementarity between the three protein language models pre-trained on different protein sequence databases significantly improved the accuracy of the model in predicting protein-ligand binding affinity. Second, ESM-1b played a key role in these three protein language models.

[0154] Table 3 Analysis of the contribution of different large language models to feature embedding

[0155]

[0156] Table 4 shows the experimental results of the comparison between the method HMRLBA described in the present invention and the mainstream methods. The results show that HMRLBA is significantly better than other methods. In addition, HMRLBA shows comparable or smaller standard deviations than other methods in all evaluation indicators, indicating that the volatility of its test results is small and the model is more stable.

[0157] Table 4 Comparative experimental results

[0158] Comparison Models RMSE↓ MAE↓ Pearson↑ Spearman↑ <![CDATA[R 2 ↑]]> DeepDTA 1.774±0.019 1.433±0.012 0.463±0.007 0.457±0.008 0.086±0.017 MGraphDTA 1.692±0.012 1.354±0.008 0.353±0.009 0.336±0.010 0.107±0.012 PSICHIC 1.539±0.014 1.235±0.009 0.504±0.011 0.498±0.010 0.166±0.013 IEConv 1.569±0.006 1.257±0.005 0.426±0.016 0.434±0.014 0.158±0.008 MaSIF 1.487±0.009 1.195±0.006 0.476±0.008 0.465±0.007 0.221±0.007 Holoprot 1.524±0.011 1.228±0.008 0.485±0.007 0.473±0.006 0.182±0.013 HbPy 1.469±0.013 1.180±0.011 0.526±0.009 0.519±0.008 0.242±0.012 HMRLBA 1.428±0.007 1.146±0.005 0.544±0.004 0.537±0.004 0.280±0.008

[0159] Table 5 shows the prediction results of the method HMRLBA described in the present invention for the "difficult samples" in the test set. In the field of protein-ligand affinity prediction, protein sequence homology is one of the key factors affecting the generalization ability of the model. In order to accurately evaluate the performance of the model on low-homology samples, this study used the NW-align algorithm to calculate protein homology and introduced the concept of "difficult samples", which is defined as samples in the test set with protein homology less than 30% compared with the training set. In this experiment, we systematically compared the performance of four methods - HMRLBA, HaPPy, MaSIF and HoloProt - on 21 difficult samples.

[0160] Table 5 Prediction results of difficult samples

[0161]

[0162] Table 6 shows the experimental results of the HMRLBA method and the comparative method in EGFR virtual drug screening. After confirming the effectiveness and stability of the HMRLBA model, we applied it to virtual drug screening of epidermal growth factor receptor (EGFR), and the experimental results showed the ranking of different methods for predicting affinity of EGFR small molecule inhibitors.

[0163] Table 6 Results of virtual screening of EGFR inhibitors

[0164]

[0165] In summary, the present invention proposes a high-precision drug screening method based on hierarchical multimodal representation learning, which is widely applicable to the fields of drug screening and drug discovery. The input data format of this method is fixed and easy to obtain from existing databases, and the algorithm is simple to implement and has low computing resource requirements, so it has extremely high promotion potential.

[0166] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A high-precision drug screening method based on hierarchical multimodal representation learning, characterized in that: include: The hierarchical multimodal representation learning method establishes a cascade fusion strategy for protein multi-scale features. Through multiple pre-trained protein language models, feature extraction is performed on the amino acid sequence of the target protein; Constructing a multi-scale graph representation of proteins based on their three-dimensional structure Based on the ligand molecular formula, an atomic-level structure diagram of the ligand molecule (referred to as the ligand molecular diagram) is constructed; Construct a message passing neural network to achieve multi-modal feature learning and representation of multi-scale graphs of proteins and multi-modal feature representation of ligand atomic structure graphs; By integrating multimodal features, a computational model for predicting protein-ligand binding affinity was constructed to complete high-throughput drug virtual screening tasks.

2. The high-precision drug virtual screening method based on hierarchical multimodal representation learning according to claim 1, characterized in that: The multiple pre-trained protein language models include Ankh, ESM-1b and ProtTrans; extracting features from the protein amino acid sequence includes using the above three models to represent the amino acid sequence of length L respectively; fusing the three feature matrices to generate a mixed embedding matrix M B : Among them, M A represents the amino acid sequence feature matrix extracted by the Ankh model; M E represents the amino acid sequence feature matrix extracted by the ESM-1b model, M P represents the amino acid sequence feature matrix extracted by the ProtTrans model; M B represents the mixed embedding matrix after fusing the three feature matrices, Indicates that the two feature matrices are connected "row by row".

3. The high-precision drug virtual screening method based on hierarchical multimodal representation learning according to claim 1 or 2, characterized in that: The construction of the multi-scale graph representation of the protein and the ligand graph representation includes constructing a residue-level structural graph representation based on the three-dimensional structure of the protein; Represent the three-dimensional structure of a protein as a residue-level graph G c =(V c ,E c ); Among them, V c ={v1,…,v L } and E c ={e1,…,e N } respectively correspond to Figure G c The set of L nodes and N edges in ; By judging whether the distance between the Cα atoms of any two amino acid residues is less than To determine each edge in the residue-level graph; Among them, the i-th node v i For any sequence S p The i-th residue in , 1≤i≤L; Node v i The eigenvector of The formula is expressed as: When node v i Yes K i When there are 1 neighbor node, the edge e ij Connect node v i and v j ; Among them, v j ∈V c ,e ij ∈E c ,j≤K i <L; Edge ij The eigenvector ef ij The formula is defined as: ef ij =[d ij ,θ ij ] Among them, d ij Indicates amino acid v i and v j The distance between ij Indicates v i and v j The angle between the normal vectors of Use triangulated mesh generation tools to construct surface representations of proteins; For a protein p, its surface was constructed with a triangular mesh by using MSMS software; Combine all the grids to get the surface map G F =(V F ,E F ); Among them, V F represents the set of vertices extracted from all meshes, E F represents the set of edges extracted from all meshes; Surface G F Any vertex u in j ∈V F , the eigenvector of the surface plot Initialize to the closest structure graph G c Middle residue node v j The characteristic vector of is expressed as: The edge e on the triangle mesh jk The eigenvector ef jk The formula is expressed as: ef jk =[d jk ,g jk ] Among them, d jk Represents vertex u j and u k The distance between jk Represents a 6-dimensional vector, indicating the edge e jk The six geometric parameters between two adjacent meshes are as follows, including the dihedral angle θ between two adjacent faces, the internal angle (θ a ,θ b ), the ratio of the side length to the vertical height of each adjacent face (d / h a ,d / h b ), and the angle θ(n k ,n j ); Use the SMILES string of the ligand to construct a molecular graph representation; For a given 1D ligand SMILES string, use RDKit to obtain the molecular structure, represented as a 2D ligand graph G L =(V L ,E L ); Among them, V L represents the set of vertices extracted from the ligand graph, E L represents the set of edges extracted from the ligand; 2D ligand maps include, based on the embedding and physicochemical properties of atom types, connected by covalent bonds; Atomic features include atomic symbol, atomic degree, implicit valence, explicit valence, and "whether it belongs to an aromatic ring"; The characteristics of a bond mainly include the bond type, whether it is a conjugated bond, and whether it belongs to a ring.

4. The high-precision drug virtual screening method based on hierarchical multimodal representation learning as claimed in claim 3, characterized in that: Multimodal feature learning includes inputting protein structure graphs, surface graphs, and ligand molecular graphs, and performing multi-layer message transmission; Construct an N-layer message passing neural network, from the structure graph G c , surface graph G F and ligand G L Generate different protein-ligand pair signature codes in An N-layer message passing neural network, whose input is defined as a graph G = (V, E); The i-th node v in the graph G i The set of neighbor nodes of ∈V is denoted as NP(v i ), the formula for initializing the message of this node is expressed as: Among them, "0" represents the time when the network model starts at t=0. Represents node v i At level 0, the message vector Represents node v i The eigenvector of Each layer node v of the message passing neural network i The message passing formula is expressed as: in, Represents node v i In the message vector of the kth layer, τ represents the activation function, U1 represents the weight matrix, U2 represents the weight matrix, NP(v i ) represents node v i The set of neighbor nodes of Represents node v i and its neighbor node v j Attention weights at layer k-1, Q represents the weight matrix, Represents node v i The eigenvector of and edge ij The eigenvector ef ij Splicing is performed, where N represents the total number of layers of message transmission, from layer 1 to layer N; Represents node v i The initial eigenvector, ef ij Represents edge e ij The feature vector of the connected node v i and v j a represents the scalar coefficient, W represents the weight matrix, ev represents the final feature embedding vector, W (1) ,W (2) ,W (3) represents the learnable weight matrix, ev f Embedding vector representing protein surface features, ev c The embedding vector representing the protein structural features, ev l Embedding vector representing the molecular features of the ligand.

5. The high-precision drug screening method based on hierarchical multimodal representation learning as claimed in claim 4, characterized in that: Multimodal feature learning also includes introducing an attention mechanism to dynamically adjust the information weights of different neighbor nodes during message transmission. The formula is expressed as: in, Indicates that at the k-1 layer, node v i Its neighbor node v j The attention weight between them; exp represents the natural exponential function, LReLU represents the LeakyReLU activation function, Represents node v i Its neighbor node v j The attention pre-value at the k-1th layer, NP(v i ) represents node v i The set of neighbor nodes of Represents node v i With node v q The previous attention value at the k-1th layer; Represents node v i and its neighbor node v j The attention pre-value at the k-1th layer, a represents the scalar coefficient, W represents the learnable weight matrix, Represents node v i The message vector at level k-1, Represents node v j The message vector at the k-1th layer; || represents the concatenation operation; After N layers of message passing, the structural embedding vector, surface embedding vector and ligand embedding vector of the protein are extracted, and the formula is expressed as: Among them, ev represents the final feature embedding vector, NP(v i ) represents node v i The set of neighbor nodes, W (1) represents the learnable weight matrix, Represents node v j The message vector at level N, W (2) represents the learnable weight matrix, ef ij Represents edge e ij The eigenvector of (3) represents the learnable weight matrix, Represents node v i In the message vector at level N, V ​​represents the set of all nodes in the graph G; For any protein-ligand pair, the MPNN module outputs three embedding vectors of length 200; the three embedding vectors of length 200 include surface vector, structure vector and ligand vector.

6. The high-precision drug screening method based on hierarchical multimodal representation learning as described in claims 1, 2, 4, and 5, characterized in that: Constructing a binding affinity prediction model for drug screening involves concatenating surface vectors, structure vectors, and ligand vectors to generate a unified multimodal feature representation of the protein-ligand pair.

7. The high-precision drug screening method based on hierarchical multimodal representation learning according to claim 6, characterized in that: Constructing a binding affinity prediction model for drug screening also includes inputting a unified multimodal feature representation of a protein-ligand pair into a multi-layer perceptron regression model to predict the binding affinity value between the protein and the ligand; The multilayer perceptron regression model is trained by minimizing the mean square error. The formula is expressed as: in, Represents the predicted value, y i represents the true value, and MSE represents the multilayer perceptron regression model.

8. A high-precision drug screening system based on hierarchical multimodal representation learning using the method, characterized in that: The feature extraction module extracts features from the amino acid sequence of the target protein through multiple pre-trained protein language models; The graph representation module constructs a multi-scale graph representation of the protein and a ligand molecular graph representation based on the three-dimensional structure of the target protein; The multimodal feature module uses a message passing neural network to construct a multi-scale graph representation of proteins and a ligand molecular graph representation, and conducts multimodal feature learning; The screening module builds a binding affinity prediction model for drug screening based on multimodal features to complete the drug screening task.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the high-precision drug screening method based on hierarchical multimodal representation learning are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the high-precision drug screening method based on hierarchical multimodal representation learning are implemented.

Images