Drug-target interaction prediction method based on multi-granularity expression

Abstract

The invention belongs to the technical field of biological information, and particularly relates to a drug-target interaction prediction method based on multi-granularity expression. The method comprises: extracting an enhanced molecular representation of a drug molecule using a layered network; modeling first-order information and second-order information in the amino acid sequence based on adjacent residues in the amino acid sequence so as to extract multi-order sequence features; processing the multiple sequence information by adopting a Pconsc4 tool to obtain spatial structure information representation of the 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; according to the method, the weakness that an existing method only pays attention to single-granularity information can be overcome, the prediction accuracy is improved, and interpretability is shown.

Description

Technical Field

[0001] The invention belongs to the field of bioinformatics, and in particular relates to a drug-target interaction prediction method based on multi-granularity representation. Background Art

[0002] Drug-target interaction (DTI) is a key part of drug discovery, which provides information about whether small molecules bind to targets 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 have brought new hope to DTI.

[0003] Currently, there are three computational methods for determining DTI: similarity-based methods, molecular docking simulations, and feature-based methods. Similarity-based methods assume that drug molecules with similar structures will act on similar receptor proteins to produce similar pharmacodynamics. The limitation of this method is the scarcity of known drug-target interaction data. Compared with similarity-based methods, molecular docking simulations can more accurately predict the interaction between drugs and targets because it simulates the binding process of drug molecules to protein targets. However, for proteins with unknown structures, molecular docking simulations will be difficult to perform or will not produce reliable results because molecular docking simulations rely on known protein structures and clear binding sites. Feature-based machine learning methods, which represent drug-target pairs as vectors of fixed length, namely feature vectors, contain various features of drugs and targets (such as physicochemical properties, molecular structures, etc.), and use various intelligent algorithms to predict drug-target interactions. The advantage of deep learning methods is that they can better identify, capture, and infer complex patterns in biological data. Since biological data often have complex multi-level structures, the multi-layer feature extraction mechanism of deep learning enables feature-based machine learning methods to perform well when processing complex data, especially when processing large-scale data. It can not only improve the accuracy of predictions, but also reduce dependence on manual feature extraction.

[0004] There are two main reasons that affect the widespread application of deep learning in DTI: (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 is that the one-dimensional sequence representation of drugs and proteins only contains limited information at a single granularity level. (2) The improvement in the performance of deep DTI methods is accompanied by obstacles in model interpretation, 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 of the invention

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

[0006] S1: Extraction of enhanced molecular representation of drug molecules using hierarchical networks;

[0007] S2: Based on the adjacent residues in the amino acid sequence, the first-order information and second-order information in the amino acid sequence are modeled separately to obtain multi-order sequence information;

[0008] S2: Use Pconsc4 tool to process multiple sequence information to obtain the spatial structure information representation of the protein;

[0009] S4: Splicing enhances the molecular representation, multi-order sequence information and spatial structure information of proteins to obtain fusion features;

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

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

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

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

[0014] 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;

[0015] 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 assignment matrix;

[0016] S15: The atomic feature matrix and adjacency matrix of the drug molecules input by the next layer of GNN network are the coarse-grained representations of the previous layer, and return to step S12;

[0017] S16: The coarse-grained representation calculated by the last layer of the hierarchical network is used as the final enhanced molecule representation.

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

[0019]

[0020] Among them, E norm represents the normalized bond energy matrix of the drug molecule, represents the bond energy relationship matrix, Represents the element in the i-th row and j-th column of the bond energy matrix.

[0021] Furthermore, the formula for calculating the constraints of atomic interactions is:

[0022]

[0023] Where C represents the constraints of atomic interactions, A represents the adjacency matrix, represents the normalized bond energy matrix of the drug molecule, and ⊙ represents the matrix dot multiplication operation.

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

[0025]

[0026]

[0027] Among them, A (l+1) represents the atomic feature matrix of the drug molecule at the l+1th layer, X (l+1) represents the adjacency matrix of drug molecules at the l+1th layer, S (l) represents the allocation matrix of the lth layer, Z (l) represents the node embedding matrix of the lth layer, C (l) Represents the constraints on the atomic interactions in the lth layer.

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

[0029] Calculate the original hydrophobicity value, original hydrophilicity value, and original side chain mass of the residue;

[0030] The amino acid coefficients were calculated based on the original hydrophobicity value, original hydrophilicity value, and original side chain mass of the residues;

[0031] Calculate the first-order and second-order information of residues based on amino acid coefficients;

[0032] The first-order information and second-order information of the residues are weighted summed to obtain multi-order sequence information.

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

[0034]

[0035] Among them, ψ λ (S i ) represents the λ-order information of the i-th residue, λ is 1 or 2; S i represents the ith residue, S i+λ represents the λth residue to the right of the ith 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 lambda residues on the right and the lambda residues on the left in the amino acid sequence.

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

[0037] The beneficial effects of the present invention are as follows: first, multi-granular drug molecular pattern information is learned through a constrained hierarchical network module. Secondly, based on adjacent residues in the amino acid sequence, the first-order information and second-order information in the amino acid sequence are modeled respectively, and then Pconsc4 is used to generate a residue contact map. The multi-layer neural network merges each residue with its first-order distance and second-order distance information to finally obtain the spatial structure information of the protein. Finally, predictions are made based on the characteristics of the drug-target pair, and the actual binding position is selected as an important element for search, providing a reasonable explanation for the prediction results from a multi-granular perspective. Compared with the existing methods that only focus on single-granularity information to obtain input representations, the present invention proposes a new interpretable framework for predicting drug-target interactions. This framework integrates information at different granularity levels, which can make up for the shortcomings of the existing methods that only focus on single-granularity information, and solve the limitations in related research, and not only performs well in prediction performance, but also has good interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0039] Figure 2 An exemplary diagram of the multi-granularity representation of molecules in the present invention;

[0040] Figure 3 A comparison diagram of the predicted results and the actual interaction strength in the present invention;

[0041] Figure 4 AUC, precision and recall comparison chart of the present invention and the comparison method;

[0042] Figure 5 A diagram explaining the multiple granularity levels of the prediction of ligand-receptor interactions in the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] The present invention proposes a drug-target interaction prediction method based on multi-granularity representation, such as Figure 1 As shown, the method includes the following contents:

[0045] S1: Extracting enhanced molecular representations of drug molecules using hierarchical networks.

[0046] The present invention constructs drug representations at multiple granularity levels, and further enhances this granularity information in a hierarchical network to obtain enhanced molecular representations; specifically:

[0047] S11: Obtain the atomic feature matrix, adjacency matrix and bond energy matrix of the drug molecule.

[0048] Assume that drug molecule m, G m =(V,V), where V = {v1,v2,…,v n} is a set of n vertices (atoms), represented by the characteristic matrix Description: Node features generally include atomic type (C, N, O, S, etc.), whether the atom belongs to an aromatic ring, the valence state of the atom, etc.; E represents the relationship (bond) between vertices, which is represented by the adjacency matrix Get the bond energy information of drug molecules, and get the bond energy matrix based on the bond energy information and adjacency matrix

[0049] S12: Use the GNN network to process the atomic feature matrix and adjacency matrix to obtain the node embedding matrix and allocation matrix of the current level.

[0050] The present invention uses the GNN network to construct a hierarchical network, and each layer consists of two parts:

[0051] GNN assign (.) and GNN embed (.), the former is responsible for the natural conversion of drug representations of different granularities, while the latter is able to capture information at different granularity levels.

[0052] The node embedding matrix is ​​obtained by GNN embed (.) is processed to obtain:

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

[0054] The essence of the assignment matrix is ​​a probability matrix. 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] Among them, Z (l) represents the node embedding matrix of the lth layer, S (l) represents the allocation matrix of the lth layer.

[0057] The constrained allocation matrix groups related nodes into substructures or special function groups, generates a coarsened graph, and obtains a coarse-grained representation. Specifically:

[0058] 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.

[0059] Molecules exhibit different levels of granularity, including atomic, substructure, and molecular levels, which are generated by the formation and breaking of specific chemical bonds of functional groups. Different chemical bonds have different bond energies, and the shorter and stronger the bonds, the stronger the atomic interactions and the higher the reaction energy. These bonds are not easy to break, thus forming substructures during chemical reactions. Therefore, the present invention proposes a bond energy matrix for each drug molecule as an atomic interaction constraint, which is then incorporated into the hierarchical module.

[0060] To avoid numerical overflow, the bond energy matrix Use the following equation for normalization:

[0061]

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

[0063] The constraints C for atomic interactions are formed by the following formula:

[0064]

[0065] Where C represents the constraints of atomic interactions, A represents the adjacency matrix, represents the normalized bond energy matrix of the drug molecule, and ⊙ represents the matrix dot multiplication operation.

[0066] S14: Calculate the coarse-grained representation of the next level based on the constraints of atomic interactions, the node embedding matrix of the current level, and the assignment matrix.

[0067]

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

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

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

[0071] For a molecule, its finest-grained representation corresponds to Figure 2 The individual atoms shown in (a) are characterized by initial atomic features. These features are constructed based on the unique atomic symbols and the chemical environment they exist in within the molecule (such as structure and implicit valence). To further build a coarse-grained representation of the molecule, the assignment matrix groups specific atoms into different clusters, generating a coarsened graph where each node consists of a substructure. These substructures in the molecular graph, i.e. Figure 2 The coarse-grained representation shown in (b) generally constitutes a functional group structure with specific chemical properties. It is worth noting that the present invention extracts features by stacking multiple hierarchical modules and adopts the representation of the last layer of coarsened graph as the molecular representation.

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

[0073] For the purpose of learning protein representation, the present invention not only considers the constituent elements of amino acid residues, but also considers integrating 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 neighboring residues. Let S2,S2,…,S n is a protein chain with 1 amino acid residue, R i and R j Represents the amino acid residue corresponding to a position in the protein chain S. Twenty amino acids make up important proteins: A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, and Y. The alphabetical indexes of these amino acids are R1, R2, …, R 20 The correlation coefficient function models the multi-order sequence information of proteins by introducing three types of residue attributes: hydrophobicity, hydrophilicity, and side chain quality.

[0075]

[0076]

[0077] in, and Represent the residues R obtained from relevant chemical materials i The original hydrophobicity value, original hydrophilicity value, and original side chain mass.

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

[0079]

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

[0081]

[0082] Among them, ψ λ (S i ) represents the λ-order information of the i-th residue, λ is 1 or 2; S i represents the ith residue, S i+λ represents the λth residue to the right of the ith residue, S i-λ represents the λth residue to the left of the ith residue, and Respectively represent S i The amino acid coefficient between the lambda residues on the right and the lambda residues on the left in the amino acid sequence.

[0083] The multi-order sequence information of each residue in the 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 is 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 for protein representation because it can effectively model the interactions and relationships between neighboring residues, which are the basis of protein structure and function. In addition, the spatial structural information of proteins is crucial for drug-target interaction prediction. Protein contact maps contain valuable information about protein structure and function, providing detailed information about the connection distance and contact probability between residue pairs.

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

[0088]

[0089] Based on the above contact probability and contact distance, a contact map is generated. A contact map is a two-dimensional matrix that represents the spatial contact of each residue pair in a protein sequence. After obtaining the contact map, the next step is to further process and extract features from the protein. The multi-layer neural network encodes the multi-order sequence information of the protein and the spatial information of the contact map together, thereby more accurately representing the spatial structure of the protein.

[0090] S4: Splicing enhances the molecular representation, multi-order sequence information and spatial structure information representation of proteins to obtain fusion features.

[0091] The enhanced molecular representation, multi-order sequence information, and spatial structural information representation of all proteins are concatenated into a one-dimensional vector.

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

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

[0094] The interaction prediction network finds the optimal parameters by minimizing the loss function during the end-to-end training process. This enables the model to make more accurate predictions.

[0095]

[0096] Among them, N is the number of samples, L is the loss function, which is composed of binary cross-entroy L C And mean square error loss (Mean-squard Error) L R calculate:

[0097]

[0098] Among them, y i Indicates the true situation, Represents the predicted value of the model.

[0099] After obtaining the drug-target interaction prediction results, inspired by the existing graph neural network interpretation methods, the model interpretation involves searching for key subgraphs in the original graph for prediction. Drug-target interactions often occur at the active sites of molecules, usually involving functional groups or special substructures with multiple atoms. Therefore, the interpretation of the model can be explored by taking the actual action site as an important element.

[0100] Evaluation of the present invention:

[0101] Drug-target interaction strength experiments were performed on the Kiba and Davis datasets. The Kiba dataset contains 2,111 drugs, 229 proteins, and 118,254 drug-target interaction pairs, with interaction scores ranging from 0.0 to 17.2. The Davis dataset contains 68 drugs, 442 proteins, and 30,056 interaction pairs, with interaction scores ranging from 5.0 to 10.8.

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

[0103] Classification task experiments were conducted on the Human and C.elegans datasets, both of which contain positive and negative samples. Positive samples refer to pairs with actual interactions between drugs and targets, which are manually extracted from two authoritative databases, Matador and DrugBank. Negative samples refer to highly credible drug-target pairs without interactions inferred by the systematic 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 example shows that the present invention can generate more accurate prediction results in DTI prediction. This application performs well in processing large-scale biological data sets, especially in terms of precision and recall, which further proves the effectiveness of integrating multi-granularity information for DTI research. By capturing multi-level information in molecular structure, it can better handle the interaction prediction between drugs and targets, especially in screening drug molecules with high interactivity. Its disadvantage in AUC may be caused by the noise of negative samples.

[0105] Understanding DTI interaction predictions by identifying key substructures and functional groups in molecules Figure 5 As shown in the example, the present invention can effectively identify substructures and functional groups in the process of generating DTI predictions. Figure 5 In the protein cocrystal complex shown in (a), the ASP-546 residue forms a polar interaction with the drug molecule, while the ASN-545 residue forms a hydrophobic interaction. These atoms interacting with the drug are identified through a constrained hierarchical module and presented in the form of substructures or functional groups. Figure 5 (b) Therefore, the model provides a deep understanding of drug-target interaction by analyzing the molecular structure from a multi-granular perspective. This capability helps to gain a deeper understanding of the binding mechanism of drugs to proteins and has the potential for a wide range of biomedical applications.

[0106] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A drug-target interaction prediction method based on multi-granularity representation, characterized in that: include: S1: Extraction of enhanced molecular representation of drug molecules using hierarchical networks; S2: Based on the adjacent residues in the amino acid sequence, the first-order information and second-order information in the amino acid sequence are modeled separately to obtain multi-order sequence information; S3: Pconsc4 tool is used to process multiple sequence information to obtain the spatial structure information representation of the protein; S4: Splicing enhances the molecular representation, multi-order sequence information and spatial structure information of proteins to obtain fusion features; S5: Input the fusion information into the interaction prediction network to obtain the drug-target interaction prediction results.

2. A drug-target interaction prediction method based on multi-granularity representation according to claim 1, characterized in that: The process of extracting an enhanced molecular representation of a drug molecule involves: S11: Obtain the atomic feature matrix, adjacency matrix and bond energy matrix of the drug molecule; S12: Use the GNN network to process the atomic feature matrix and the adjacency matrix to obtain the node embedding matrix and the allocation 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; 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 assignment matrix; S15: The atomic feature matrix and adjacency matrix of the drug molecules input by the next layer of GNN network are the coarse-grained representations of the previous layer, and return to step S12; S16: The coarse-grained representation calculated by the last layer of the hierarchical network is used as the final enhanced molecule representation.

3. A drug-target interaction prediction method based on multi-granularity representation according to claim 2, characterized in that: The formula for calculating the normalized bond energy matrix of a drug molecule is: in, represents the normalized bond energy matrix of the drug molecule, represents the bond energy relationship matrix, Represents the element in the i-th row and j-th column of the bond energy matrix.

4. The method for predicting drug-target interaction based on multi-granularity representation according to claim 2, characterized in that: The formula for calculating the constraints of atomic interactions is: Where C represents the constraints of atomic interactions, A represents the adjacency matrix, represents the normalized bond energy matrix of the drug molecule, and ⊙ represents the matrix dot multiplication operation.

5. The drug-target interaction prediction method based on multi-granularity representation according to claim 2, characterized in that: The formula for calculating the coarse-grained representation is: A (l+1) =S (l)T C (l) S (l) X (l+1) =S (l)T Z (l) Among them, A (l+1) represents the atomic feature matrix of the drug molecule at the l+1th layer, X (l+1) represents the adjacency matrix of drug molecules at the l+1th layer, S (l) represents the allocation matrix of the lth layer, Z (l) represents the node embedding matrix of the lth layer, C (l) Represents the constraints on the atomic interactions in the lth layer.

6. The method for predicting drug-target interaction 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 value, original hydrophilicity value, and original side chain mass of the residue; The amino acid coefficients were calculated based on the original hydrophobicity value, original hydrophilicity value, 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 information and second-order information of the residues are weighted summed to obtain multi-order sequence information.

7. A drug-target interaction prediction method based on multi-granularity representation according to claim 6, characterized in that: The formula for calculating the first-order information and second-order information of residues is: Among them, ψ λ (S i ) represents the λ-order information of the i-th residue, λ is 1 or 2; S i represents the ith residue, S i+λ represents the λth residue to the right of the ith residue, S i-λ represents the λth residue to the left of the ith residue and Respectively represent S i The amino acid coefficient between the lambda residues on the right and the lambda residues on the left in the amino acid sequence.

8. The method for predicting drug-target interaction based on multi-granularity representation according to claim 1, characterized in that: The interaction prediction network includes multi-layer perceptrons and activation functions.

Images