A drug-target interaction prediction method based on multi-granularity representation

Abstract

The present application belongs to the technical field of biological information, and particularly relates to a drug-target interaction prediction method based on multi-granularity representation; the method comprises the following steps: extracting an enhanced molecular representation of a drug molecule by using a hierarchical network; modeling first-order information and second-order information in an amino acid sequence respectively based on adjacent residues in the amino acid sequence, so as to extract multi-order sequence features; processing multiple sequence information by using a Pconsc4 tool to obtain spatial structure information representation of a protein; splicing the enhanced molecular representation, the multi-order sequence information and the spatial structure information representation of the protein to obtain fusion features; inputting the fusion information into an interaction prediction network to obtain a drug-target interaction prediction result; the present application can not only solve the weakness of existing methods that only focus on single-granularity information, improve prediction accuracy, but also show interpretability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of biological information, and particularly relates to a drug-target interaction prediction method based on multi-granularity representation. BACKGROUND

[0002] Drug-target interaction (DTI) is a key part of drug discovery, which provides information about whether small molecules bind to the target and can effectively screen potential targets and drugs. With the accumulation of extensive drug data and more complex interactions, it is extremely expensive and impractical to obtain all possible drug-target pairs in traditional biological laboratories. Fortunately, computational models based on various intelligent algorithms bring new dawn to DTI.

[0003] Currently, the computational methods for determining DTI include similarity-based methods, molecular docking simulation and feature-based methods. The similarity-based method assumes that structurally similar drug molecules will act on similar receptor proteins, thereby producing similar efficacy, and the limitation of this method lies in the scarcity of known drug-target interaction data. Compared with the similarity-based method, molecular docking simulation can more accurately predict the interaction between drugs and targets, because it simulates the binding process of drug molecules and protein targets. However, for proteins with unknown structure, molecular docking simulation will be difficult to perform or cannot obtain reliable results, because molecular docking simulation relies on known protein structure and explicit binding site. Feature-based machine learning method, this method represents drug-target pairs as fixed-length vectors, i.e. feature vectors, which contain various features of drugs and targets (such as physical and chemical properties, molecular structure, etc.), and uses various intelligent algorithms to predict drug-target interaction. The advantage of deep learning method is that it can better identify, capture and infer complex patterns in biological data. Since biological data often has complex multi-level structure, the multi-level feature extraction mechanism of deep learning makes feature-based machine learning method perform well in processing complex data, especially in processing large-scale data, which not only improves the accuracy of prediction, but also reduces the dependence on manual feature extraction.

[0004] The main reasons affecting the wide application of deep learning in DTI include two points: (1) input representation. Correct input representation can significantly affect the overall architecture and final performance of the constructed model. The limitation of existing deep DTI methods lies in that one-dimensional sequence representation of drugs and proteins only contains limited information of single granularity level. (2) The improvement of deep DTI method performance is accompanied by obstacles in model explanation, that is, the operator does not understand the decision-making process of the model, which is not allowed in the field of drug design. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application proposes a drug-target interaction prediction method based on multi-granularity representation, which comprises:

[0006] S1: using a hierarchical network to extract an enhanced molecular representation of the drug molecule;

[0007] S2: modeling the first-order information and the second-order information in the amino acid sequence based on adjacent residues in the amino acid sequence to obtain multi-order sequence information;

[0008] S2: using the Pconsc4 tool to process the multi-order sequence information to obtain the spatial structure information representation of the protein;

[0009] S4: splicing the enhanced molecular representation, the multi-order sequence information and the spatial structure information representation of the protein to obtain the fusion feature;

[0010] S5: inputting the fusion information into the interaction prediction network to obtain the drug-target interaction prediction result.

[0011] Preferably, the process of extracting the enhanced molecular representation of the drug molecule comprises:

[0012] S11: obtaining the atomic feature matrix, the adjacency matrix and the bond energy matrix of the drug molecule;

[0013] S12: using the GNN network to process the atomic feature matrix and the adjacency matrix to obtain the node embedding matrix and the assignment matrix of the current level;

[0014] S13: calculating the normalized bond energy matrix of the drug molecule according to the bond energy matrix; and calculating the constraint of atomic interaction according to the normalized bond energy matrix and the adjacency matrix;

[0015] S14: calculating the coarse-grained representation of the next layer according to the constraint of atomic interaction, the node embedding matrix and the assignment matrix of the current layer;

[0016] S15: the atomic feature matrix and the adjacency matrix of the drug molecule input to the next layer GNN network are the coarse-grained representation of the previous layer, and the step S12 is returned;

[0017] S16: taking the coarse-grained representation calculated by the last layer of the hierarchical network as the final enhanced molecular representation.

[0018] Further, the formula for calculating the normalized bond energy matrix of the drug molecule is:

[0019]

[0020] wherein E norm represents the normalized bond energy matrix of the drug molecule, denotes the bond energy relation matrix, denotes the element in the i-th row and j-th column of the bond energy matrix.

[0021] Further, the formula for calculating the constraint of atomic interaction is:

[0022]

[0023] wherein C denotes the constraint of atomic interaction, A denotes the adjacency matrix, denotes the normalized bond energy matrix of the drug molecule, and denotes the matrix point multiplication operation.

[0024] Further, the formula for calculating the coarse-grained representation is:

[0025]

[0026]

[0027] wherein A (l+1) denotes the atomic feature matrix of the drug molecule at the l+1-th layer, X (l+1) denotes the adjacency matrix of the drug molecule at the l+1-th layer, S (l) denotes the assignment matrix at the l-th layer, Z (l) denotes the node embedding matrix at the l-th layer, C (l) denotes the constraint of atomic interaction at the l-th layer.

[0028] Preferably, the process of obtaining the multi-order sequence information comprises:

[0029] calculating the original hydrophobic value, the original hydrophilic value, and the original side chain mass of the residue;

[0030] calculating the amino acid coefficient according to the original hydrophobic value, the original hydrophilic value, and the original side chain mass of the residue;

[0031] calculating the first-order information and the second-order information of the residue according to the amino acid coefficient;

[0032] performing weighted summation on the first-order information and the second-order information of the residue to obtain the multi-order sequence information.

[0033] Further, the formula for calculating the first-order information and the second-order information of the residue is:

[0034]

[0035] wherein ψ λ (S i ) denotes the λ-th order information of the i-th residue, and λ is 1 or 2; S i denotes the i-th residue, S i+λ denotes the λ-th residue on the right side of the i-th residue, S i-λrepresents the λth residue to the left of the ith residue and respectively represent S i the amino acid coefficient between the right and left λth residues in the amino acid sequence.

[0036] Preferably, the interaction prediction network comprises a multi-layer perceptron and an activation function.

[0037] The beneficial effects of the present application are: first, learning multi-granularity drug molecule pattern information through a layered network module with constraints. Second, based on adjacent residues in the amino acid sequence, first-order information and second-order information in the amino acid sequence are respectively modeled, and then Pconsc4 is used to generate a residue contact map, and a multi-layer neural network combines each residue with its first-order distance and second-order distance information to finally obtain the spatial structure information of the protein. Finally, according to the characteristics of the drug-target pair, the actual binding position is selected as an important element for searching, and a reasonable explanation is provided for the prediction result from the perspective of multi-granularity. Compared with the existing method which only focuses on single-granularity information to obtain input representation, the present application proposes a new interpretable framework for predicting drug-target interactions. This framework integrates information at different levels of granularity, which can make up for the shortcomings of existing methods that only focus on single-granularity information, and solve the limitations in related research. Moreover, it not only performs excellently in prediction performance, but also has good interpretability. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flow chart of the drug-target interaction prediction method in the present application;

[0039] Figure 2 is an example diagram of multi-granularity representation of molecules in the present application;

[0040] Figure 3 is a comparison diagram of prediction results and actual interaction intensities in the present application;

[0041] Figure 4 is an AUC, precision and recall comparison diagram of the present application and the comparative method;

[0042] Figure 5 is a multi-granularity level explanation diagram of ligand and receptor interaction prediction in the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0044] The present application provides a drug-target interaction prediction method based on multi-granularity representation, as shown in the formula: Figure 1 The method comprises the following contents:

[0045] S1: using a hierarchical network to extract an enhanced molecular representation of a drug molecule.

[0046] The present application constructs multiple granularity levels of drug representation, and further enhances this granularity information in a hierarchical network to obtain an enhanced molecular representation; specifically:

[0047] S11: obtaining an atomic feature matrix, an adjacency matrix and a bond energy matrix of a drug molecule.

[0048] Suppose a drug molecule m, G m =(V,V), wherein V={v1,v2,…,v n} is a set of n vertices (atoms), and the feature matrix is described, and the node features generally include atomic types (C, N, O, S, etc.), whether the atom belongs to an aromatic ring, the valence of the atom and other information; E represents the relationship (bond) between the vertices, which is represented by the adjacency matrix . The bond energy information of the drug molecule is obtained, and the bond energy matrix

[0049] S12: using a GNN network to process the atomic feature matrix and the adjacency matrix to obtain a node embedding matrix and an assignment matrix of the current level.

[0050] The present application uses a GNN network to construct a hierarchical network, which is composed of two parts in each layer:

[0051] GNN assign (.) and GNN embed (.), the former is responsible for the natural conversion of drug representation of different granularity, and the latter can capture information of different granularity levels.

[0052] The node embedding matrix is obtained by GNN embed (.) processing:

[0053] Z (l) =GNN l，embed (A (l) ,X (l) )

[0054] The essence of the assignment matrix is a probability matrix, and in each layer, the assignment probability is obtained by applying the Softmax function row by row:

[0055] S (l) =Softmax(GNNl，assign (A (l) ,X (l) ))

[0056] wherein Z (l) denotes the node embedding matrix of the l-th layer, S (l) denotes the assignment matrix of the l-th layer.

[0057] The assignment matrix with constraints groups relevant nodes into substructures or special functional groups, generates a coarsened graph, and obtains a coarse-grained representation, specifically:

[0058] S13: Calculate the normalized bond energy matrix of the drug molecule according to the bond energy matrix; calculate the constraint of atomic interaction according to the normalized bond energy matrix and the adjacency matrix.

[0059] Molecules exhibit different levels of granularity, including atomic, substructure, and molecular levels, which are due to the formation and breaking of specific chemical bonds of functional groups. Different chemical bonds have different bond energies, and the shorter and stronger the bond is, the stronger the atomic interaction and the higher the reaction energy. These bonds are not easy to break, thereby forming substructures in the chemical reaction process. Therefore, the present application proposes a bond energy matrix of each drug molecule as a constraint of atomic interaction, which is then merged into the hierarchical module.

[0060] In order to avoid numerical overflow, the bond energy matrix is normalized using the following equation:

[0061]

[0062] wherein denotes the normalized bond energy matrix of the drug molecule, denotes the bond energy matrix, denotes the element in the i-th row and j-th column of the bond energy matrix, denotes the maximum value in the matrix .

[0063] The constraint C of atomic interaction is formed by the following formula:

[0064]

[0065] wherein C denotes the constraint of atomic interaction, A denotes the adjacency matrix, denotes the normalized bond energy matrix of the drug molecule, and denotes the matrix dot product operation.

[0066] S14: Calculate the coarse-grained representation of the next level according to the constraint of atomic interaction, the node embedding matrix of the current level, and the assignment matrix.

[0067]

[0068] S15: The atom feature matrix and adjacency matrix of the drug molecule input to the next layer GNN network are the coarse-grained representation of the previous layer, return to step S12.

[0069] S16: The coarse-grained representation calculated by the last layer of the hierarchical network is taken as the final enhanced molecular representation.

[0070] The present application extracts features by stacking multiple hierarchical network modules, and takes the representation of the last layer of the coarse graph as the final enhanced molecular representation. Specifically:

[0071] For a molecule, the most fine-grained representation corresponds to Figure 2 (a) Single atoms shown in (a) are characterized by initial atom features. These features are constructed based on unique atom symbols and their chemical environment (e.g. structure and implicit valence) within the molecule. In order to further construct the coarse-grained representation of the molecule, the assignment matrix combines specific atoms into different clusters, generating a coarse graph, where each node consists of a substructure. These substructures in the molecular graph, i.e. Figure 2 (b) Coarse-grained representation shown in (b), usually constitutes a functional group structure with specific chemical properties. It is worth noting that the present application extracts features by stacking multiple hierarchical modules, and takes the representation of the last layer of the coarse graph as the molecular representation.

[0072] S2: Based on the adjacent residues in the amino acid sequence, the first-order information and the second-order information in the amino acid sequence are modeled respectively, to obtain multi-order sequence information.

[0073] For the purpose of learning protein representation, the present application not only considers the constituent elements of amino acid residues, but also integrates multi-order sequence information by combining the chemical properties of adjacent residues. Specifically:

[0074] Modeling multi-order sequence information is achieved by integrating each residue and its relationship with adjacent residues. Let S2,S2,…,S n be a protein chain with l amino acid residues, R i and R j represent the amino acid residues corresponding to a certain position in the protein chain S. The twenty amino acids constitute important proteins: A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W and Y. The letter index of these amino acids is R1,R2,…,R 20 The correlation function simulates the multi-order sequence information of the protein by introducing three types of residue attributes: hydrophobicity, hydrophilicity and side chain mass.

[0075]

[0076]

[0077] in, and Residues R obtained from relevant chemical materials are respectively represented. i The original hydrophobicity, original hydrophilicity, and original side chain quality.

[0078] The amino acid coefficient is calculated as follows:

[0079]

[0080] The first-order information ψ1(S) of the i-th residue of an amino acid i ) and second-order information ψ2(S i Constructed from the following equation:

[0081]

[0082] Where, ψ λ (S i ) represents the λ-order information of the i-th residue, where λ is 1 or 2; S i S represents the i-th residue. i+λ S represents the λ-th residue to the right of the i-th residue. i-λ This represents the λ-th residue to the left of the i-th residue. and S represents i The amino acid coefficient between the λth residue on the right and left sides of the amino acid sequence.

[0083] The multi-order sequence information of each residue in a protein chain is ultimately formed by the following equation:

[0084] M(S i )=αψ1(S i )+(1-α)ψ2(S i ),α∈(0,1)

[0085] S3: The Pconsc4 tool was used to process the multiple sequence information to obtain the spatial structure information representation of the protein.

[0086] Integrating multi-order sequence information is a valuable approach to protein representation because it can effectively model the interactions and relationships between adjacent residues, which are fundamental to protein structure and function. Furthermore, the spatial structural information of proteins is crucial for predicting drug-target interactions. Protein contact maps contain valuable information about protein structure and function, providing detailed information on the connection distances and contact probabilities between residue pairs.

[0087] The application adopts Pconsc4 to calculate the protein contact map, thereby integrating the spatial structure information of the protein. HHbilts and HHfilter are adopted to realize protein multiple sequence comparison and filtering through a protein database; the generated comparison and filtering results are taken as the input of Pconsc4; Pconsc4 generates a distance threshold and The contact probability between the residue pairs is measured by the distance between the residue pairs through S-score, which is a distance-based scoring function used to reflect the relative distance between two residues:

[0088]

[0089] Based on the contact probability and contact distance described above, the contact map is generated. The contact map is a two-dimensional matrix used to represent the spatial contact of each residue pair in the protein sequence. After obtaining the contact map, the next step is to further process and extract features of the protein, and the protein multiple order sequence information and the spatial information of the contact map are encoded through a multi-layer neural network, so as to more accurately represent the spatial structure of the protein.

[0090] S4: splice the enhanced molecular representation, the multiple order sequence information and the spatial structure information representation of the protein to obtain the fusion feature.

[0091] The enhanced molecular representation, the multiple order sequence information and the spatial structure information representation of all proteins are connected into a one-dimensional vector.

[0092] S5: input the fusion information into the interaction prediction network to obtain the drug-target interaction prediction result.

[0093] The fusion information is input into the interaction prediction network built by the multi-layer perceptron and the activation function to generate the interaction probability, that is, the drug-target interaction prediction result. In order to enhance the generalization, a dropout function is adopted in each linear layer.

[0094] In the end-to-end training process of the interaction prediction network, the optimal parameters are found by minimizing the loss function so that the model can make more accurate predictions.

[0095]

[0096] wherein N is the number of samples, L is the loss function, and the binary cross-entropy loss (Binary Cross-entroy) L C and the mean square error loss (Mean-squard Error) L R are calculated:

[0097]

[0098] where y i represents the true case, represents the predicted value of the model.

[0099] After obtaining the drug-target interaction prediction results, inspired by the existing graph neural network explanation method, the model explanation involves searching for key subgraphs in the original graph for prediction, and drug-target interaction often occurs at the active site of the molecule, usually involving functional groups or special substructures with multiple atoms, so the explanation of the model can be explored by taking the actual action position as an important element.

[0100] Evaluation of the present application:

[0101] Drug-target interaction strength experiments were performed on the Kiba and Davis datasets. Kiba contains 2111 drugs, 229 proteins and 118254 drug-target interaction pairs, with interaction scores ranging from 0.0 to 17.2. The Davis dataset contains 68 drugs, 442 proteins and 30056 interaction pairs, with interaction scores ranging from 5.0 to 10.8.

[0102] Drug-target interaction strength prediction is shown in Figure 3 The example shows that the present application can learn more rich semantic information in the prediction process. The color gradient from dark to light symbolizes the density of a series of samples from low to high. The green line indicates that the predicted value is equal to the corresponding label for all samples. Figure 3 The large number of samples around the medium green line illustrates the incredible performance of the model.

[0103] Classification task experiments were performed on the Human and C.elegans datasets, both of which contain positive and negative samples. Positive samples refer to pairs of drugs and targets that have actual interactions between them, which are manually extracted from two authoritative databases, Matador and DrugBank. Negative samples refer to highly reliable drug-target pairs without interaction inferred by the system screening framework. The Human dataset covers 1052 compounds, 852 proteins and 3369 positive sample pairs, and the C.elegans dataset covers 1434 compounds, 2504 proteins and 4000 positive sample pairs.

[0104] The results of comparison with several state-of-the-art deep learning methods are shown in Figure 4As shown, the examples demonstrate that the present application can generate more accurate prediction results in DTI prediction. The application performs well in handling large-scale biological data sets, especially in precision and recall metrics, which further demonstrates the effectiveness of integrating multi-granularity information for DTI research, by capturing multi-level information in molecular structures, thus better handling drug-target interaction prediction, especially in screening drug molecules with high interaction, while its disadvantage in AUC may be caused by the noise of negative samples.

[0105] By identifying key substructures and functional groups in molecules, the understanding of DTI interaction prediction is improved as Figure 5 As shown, the examples demonstrate that the present application can effectively identify substructures and functional groups in the process of generating DTI prediction. As shown in Figure 5 (a), the ASP-546 residue forms a polar interaction with the drug molecule, while the ASN-545 residue forms a hydrophobic interaction. These drug-interacting atoms are identified by the hierarchical module with constraints, and are presented in the form of substructures or functional groups, similar to the case shown in Figure 5 (b). Therefore, the model provides an in-depth understanding of drug-target interaction by analyzing molecular structures from a multi-granularity perspective, which helps to better understand the binding mechanism of drugs and proteins, and has wide biomedical application potential.

[0106] The above examples further illustrate the objects, technical solutions and advantages of the present application. It should be understood that the above examples are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made to the present application within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A drug-target interaction prediction method based on multi-granularity representation, characterized in that, include: S1: Enhanced molecular representation of drug molecules is extracted using a hierarchical network; The process of extracting enhanced molecular representations of drug molecules includes: S11: Obtain the atomic feature matrix, adjacency matrix, and bond energy matrix of the drug molecule; S12: Use a GNN network to process the atomic feature matrix and adjacency matrix to obtain the node embedding matrix and assignment matrix of the current level; S13: Calculate the normalized bond energy matrix of the drug molecule based on the bond energy matrix; calculate the constraints of atomic interactions based on the normalized bond energy matrix and the adjacency matrix; the formula for calculating the normalized bond energy matrix of the drug molecule is: ； in, This represents the normalized bond energy matrix of a drug molecule. Represents the bond energy relation matrix. This represents the element in the i-th row and j-th column of the bond energy matrix; S14: Calculate the coarse-grained representation of the next layer based on the constraints of atomic interactions, the node embedding matrix of the current layer, and the allocation matrix; S15: The atomic feature matrix and adjacency matrix of the drug molecule input to the next layer of the GNN network are coarse-grained representations of the previous layer. Return to step S12. S16: Use the coarse-grained representation calculated from the last layer of the hierarchical network as the final enhanced molecular representation; S2: Based on adjacent residues in the amino acid sequence, the first-order and second-order information in the amino acid sequence are modeled to obtain multi-order sequence information; S3: The Pconsc4 tool was used to process the multiple sequence information to obtain the spatial structure information representation of the protein; S4: Splicing enhances molecular representation, multi-order sequence information, and protein spatial structure information to obtain fusion features; S5: Input the fusion information into the interaction prediction network to obtain the drug-target interaction prediction results.

2. The drug-target interaction prediction method based on multi-granularity representation according to claim 1, characterized in that, The formula for calculating the constraints of atomic interactions is: ； in, Represents constraints on atomic interactions. Represents the adjacency matrix. This represents the normalized bond energy matrix of a drug molecule. This represents the matrix dot product operation.

3. The drug-target interaction prediction method based on multi-granularity representation according to claim 1, characterized in that, The formula for calculating the coarse-grained representation is: ； ； in, Indicates the first The atomic feature matrix of drug molecules in the layer. Indicates the first The adjacency matrix of drug molecules in the layer. Indicates the first Layer allocation matrix, Indicates the first Layer node embedding matrix, Indicates the first Constraints on atomic interactions within the layers.

4. The drug-target interaction prediction method based on multi-granularity representation according to claim 1, characterized in that, The process of obtaining multi-order sequence information includes: Calculate the original hydrophobicity, original hydrophilicity, and original side chain mass of the residues; The amino acid coefficient is calculated based on the original hydrophobicity, original hydrophilicity, and original side chain mass of the residues; Calculate the first-order and second-order information of residues based on amino acid coefficients; The first-order and second-order information of residues are weighted and summed to obtain multi-order sequence information.

5. The drug-target interaction prediction method based on multi-granularity representation according to claim 4, characterized in that, The formulas for calculating the first-order and second-order information of residues are: ； in, Indicates the first residues Hierarchical information, Choose 1 or 2; Indicates the first One residue, Indicates the first The residue to the right of the first residue One residue, Indicates the first The first residue to the left of the residue residues and They represent In the amino acid sequence, the right and left sides are... The amino acid coefficients between residues.

6. The drug-target interaction prediction method based on multi-granularity representation according to claim 1, characterized in that, Interactive prediction networks consist of multilayer perceptrons and activation functions.

Images