A bioactivity fingerprint-based molecular similarity calculation method and system

Abstract

The application discloses a molecular similarity calculation method based on a biological activity fingerprint, and comprises the following steps: obtaining a graph neural network for representing a molecule based on the biological activity data of the molecule and a target protein, and obtaining a pre-training vector mapping of each atom in the molecule based on the graph neural network; decomposing the molecule into a target molecule and a query molecule to obtain corresponding subgraphs; summing the pre-training vector mappings of all atoms in each node of the subgraph to obtain the representation s i Of the node in the corresponding subgraph; traversing the maximum common subgraph possibility of the two subgraphs, and calculating the valley base similarity coefficient between the nodes of the query molecule and the target molecule in the two subgraphs, wherein the maximum common subgraph valley base similarity coefficient is used as the similarity of the two molecules. Corresponding systems and applications are also disclosed. The application brings great convenience to ligand-based drug discovery, reduces the dependence of related staff on domain knowledge, improves the efficiency of drug design and development, and shortens the research and development period.

Description

Technical Field

[0001] This invention belongs to the field of biogene technology, and in particular relates to a method and system for calculating molecular similarity based on bioactive fingerprints. Background Technology

[0002] Predicting the biological activity of ligands is a key challenge in drug discovery. Quantitative structure-activity relationship (QSAR) is a fundamental method, but it requires a large amount of target-specific data. Early drug discovery processes typically begin with one or a few active compounds. A basic assumption is that structurally similar molecules have similar biological activities. Therefore, finding compounds most similar to known active reference molecules is a key challenge based on ligand virtual screening (LBVS). The Tanimoto coefficient of molecular fingerprints is the most traditional method for assessing molecular similarity. However, molecular fingerprint similarity decreases sharply with changes in the skeletal structure because all changes in the molecular fingerprint are equidistant; therefore, it is difficult to achieve skeletal transitions using only molecular fingerprint similarity.

[0003] In recent years, some scholars have proposed molecular similarity calculation methods based on three-dimensional structures. Among them, the Flexi-LS-align method can perform atomic-level ligand structure similarity calculations for flexible molecular structures. The EViS method further considers the structure of protein-ligand complexes, using protein-ligand templates from structural biology databases to calculate similarity. However, these similarity calculation methods all involve the three-dimensional structure of ligands or even complexes, resulting in high computational costs.

[0004] Recent research employs deep learning techniques, such as graph neural networks (GNNs), to learn information representations of molecules. Most of these methods focus on unlabeled data containing little biological activity information, primarily due to a shortage of labeled molecules for supervised training. However, with the development of publicly available large-scale biological activity databases and research data, valuable labeled data is increasing. Biologically meaningful tags derived from the interaction of a compound with a set of targets have reached a usable scale.

[0005] Therefore, there is indeed a need to propose better solutions for the aforementioned existing technologies. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, and application for calculating molecular similarity based on bioactivity fingerprints. First, bioactivity data of molecules and target proteins are collected. Then, a graph neural network (GNN) for molecular characterization is pre-trained using millions of activity data. Through the above process, a pre-trained graph neural network for characterizing molecules can be obtained. Then, the similarity between any two molecules can be calculated through the maximum common subgraph of the reduced graph. These bio-information representations are used to align the molecular graphs to calculate similarity, thereby achieving more efficient skeletal transitions.

[0007] This invention provides a method for calculating molecular similarity based on bioactive fingerprints, comprising:

[0008] S1, Based on the bioactivity data of molecules and target proteins, the molecular graph is pre-trained to obtain a graph neural network for characterizing molecules, and the pre-trained vectorized mapping (embedding) of each atom in the molecule is obtained based on the graph neural network.

[0009] S2, decompose the molecule into target molecule and query molecule to obtain the corresponding subgraph (reduced graph);

[0010] S3 sums the pre-trained vectorized embeddings of all atoms in each node of the reduced graph, which serves as the representation s of that node in the corresponding subgraph. i ;

[0011] S4, traverse the maximum common subgraph probability of the two subgraphs (reduced graphs), and calculate the Tanimoto similarity coefficient between each node of the query molecule and the target molecule in the two subgraphs based on the probability and the representation of the node in the corresponding subgraph, where the maximum common subgraph Tanimoto similarity coefficient is used as the similarity between the two molecules.

[0012] Preferably, the graph neural network used to characterize the molecule is GAT, and the process of pre-training the molecular graph based on bioactivity data of the molecule and target protein to obtain the graph neural network for characterizing the molecule includes:

[0013] S11 aggregates atomic information in molecules;

[0014] S12 utilizes a recurrent neural network (GRU) to represent node propagation in a graph network;

[0015] S13 uses a multilayer fully connected network (MLP) to predict the biological activity of molecules.

[0016] Preferably, S11 includes:

[0017] Information about atomic polymerization bonds is obtained as shown in equation (1):

[0018]

[0019] Among them, v i Represents an atom, u i，j Indicates the bond between atoms, v si The atoms following the information about the bonding of the aggregates, Indicates feature splicing;

[0020] Based on the aggregation of atomic information in the molecule, as shown in equation (2):

[0021]

[0022] Where, v′ si Atomic representation after aggregating atomic information. W represents the attention weights of atoms i and j in the k-th head. k Let σ represent the learnable weights of the k-th head, and let σ represent the non-linear activation function.

[0023] Preferably, in step S12, the node propagation in the graph network is represented by equation (3) using a recurrent neural network (GRU):

[0024]

[0025] Where z represents the final molecular characterization.

[0026] Preferably, in step S13, the bioactivity of molecules is predicted using a multilayer fully connected network (MLP), as expressed by equation (4):

[0027] y = MLP(z) (4);

[0028] y represents biological activity.

[0029] Preferably, S2 includes:

[0030] S21, decompose the molecule into target molecule and query molecule, and take a single atom as the center, and take the atoms whose connection distance is less than or equal to 2 as the elements of the subgraph;

[0031] S22, gather all the elements to form corresponding subgraphs, wherein there are no duplicate atoms between subgraphs, and the number of atoms in each subgraph is greater than 2.

[0032] Preferably, S3 includes:

[0033] S31, the target molecule and the query molecule are input into the trained graph neural network to obtain the target molecule representation and the query molecule representation, which together form a pre-trained vectorized mapping (embedding) of all atoms.

[0034] S32, sum the pre-trained vectorized embeddings of all atoms in each node of the reduced graph, and use this sum as the representation s of that node in the corresponding subgraph. i ; Equation (5) is used to represent:

[0035]

[0036] Among them, z j This represents the pre-trained vectorized mapping (embedding) of atoms.

[0037] Preferably, S4 includes:

[0038] S41, use nexworkx to traverse the maximum common subgraph probability of two subgraphs (reduced graphs) to obtain multiple pairs of common subgraphs;

[0039] S42, based on equation (6), calculate the Tanimoto similarity coefficient between each node of the query molecule and the target molecule in each pair of common subgraphs:

[0040]

[0041] Where, N t It is the number of atoms in the target molecule, s qi It is the representation of the i-th node of the query molecule, s ti It is the i-th node representation of the target molecule;

[0042] S43, take the maximum common subgraph similarity as the similarity between the two molecules, as shown in equation (7).

[0043]

[0044] A second aspect of the present invention provides a molecular similarity calculation system based on bioactive fingerprints, comprising:

[0045] The neural network training module (101) is used to pre-train the molecular graph based on the bioactivity data of the molecule and the target protein to obtain a graph neural network for characterizing the molecule, and to obtain the pre-trained vectorized mapping (embedding) of each atom in the molecule based on the graph neural network.

[0046] The subgraph building module (102) is used to decompose the molecule into target molecule and query molecule to obtain the corresponding subgraph (reduced graph);

[0047] The representation module (103) is used to sum the pre-trained vectorized embeddings of all atoms in each node of the reduced graph as the representation s of that node in the corresponding subgraph. i ;

[0048] The molecular similarity calculation module (104) is used to traverse the maximum common subgraph probability of two subgraphs and calculate the Tanimoto similarity coefficient between each node of the query molecule and the target molecule in the two subgraphs based on the probability and the representation of the node in the corresponding subgraph, wherein the maximum common subgraph Tanimoto similarity coefficient is used as the similarity between the two molecules.

[0049] A third aspect of the present invention is to provide a molecular similarity calculation method based on bioactive fingerprints for the discovery of replaceable molecular fragments for scaffold transitions.

[0050] A fourth aspect of the present invention is to provide an application of a molecular similarity calculation method based on bioactive fingerprints in molecular drug development.

[0051] The system, method, and application provided by this invention have the following beneficial technical effects:

[0052] This invention innovatively proposes a molecular similarity calculation method based on bioactive fingerprints, overcoming the limitation of traditional molecular fingerprints which cannot achieve scaffold transitions due to the lack of biological information. By matching the maximum common subgraph of the reduced graph, a highly efficient and more generalizable similarity calculation method is achieved, greatly facilitating ligand-based drug discovery, reducing the reliance of researchers on domain knowledge, improving the efficiency of drug design and development, and shortening the R&D cycle, thus possessing high practical value. Attached Figure Description

[0053] Figure 1 This is a schematic flowchart illustrating a molecular similarity calculation method based on bioactive fingerprints according to a preferred embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram illustrating the process of training molecular characterization using a graph neural network according to a preferred embodiment of the present invention.

[0055] Figure 3 The flowchart illustrates the calculation of intermolecular bioactivity fingerprint similarity using a pre-trained graph neural network according to a preferred embodiment of the present invention.

[0056] Figure 4 This is a schematic diagram of a molecular similarity calculation system based on bioactive fingerprints according to a preferred embodiment of the present invention. Detailed Implementation

[0057] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0058] Example 1

[0059] like Figure 1 As shown, a method for calculating molecular similarity based on bioactive fingerprints is provided, including:

[0060] S1, Based on the bioactivity data of molecules and target proteins, the molecular graph is pre-trained to obtain a graph neural network for characterizing molecules, and the pre-trained vectorized mapping (embedding) of each atom in the molecule is obtained based on the graph neural network.

[0061] In this embodiment, the graph neural network used to characterize molecules is GAT, wherein:

[0062] like Figure 2 As shown, the process of pre-training a molecular graph based on bioactivity data of molecules and target proteins to obtain a graph neural network for characterizing molecules includes:

[0063] S11, aggregating atomic information in the molecule; specifically, in this embodiment, it includes:

[0064] Information about atomic polymerization bonds is obtained as shown in equation (1):

[0065]

[0066] Among them, v i Represents an atom, u i，j Indicates the bond between atoms, v si The atoms following the information about the bonding of the aggregates, Indicates feature splicing;

[0067] Based on the aggregation of atomic information in the molecule, as shown in equation (2):

[0068]

[0069] Where, v′ si Atomic representation after aggregating atomic information. W represents the attention weights of atoms i and j in the k-th head. kLet σ represent the learnable weights of the k-th head, and let σ represent the non-linear activation function.

[0070] S12 utilizes a recurrent neural network (GRU) to represent node propagation in a graph network;

[0071] In this embodiment, specifically, equation (3) is used:

[0072]

[0073] Where z represents the final molecular characterization;

[0074] S13, the biological activity of molecules is predicted using a multilayer fully connected network (MLP), and is expressed by equation (4):

[0075] y = MLP(z) (4);

[0076] y represents biological activity.

[0077] S2, decompose the molecule into target molecule and query molecule to obtain the corresponding subgraph (reduced graph);

[0078] In this embodiment, specifically, it includes:

[0079] S21, decompose the molecule into target molecule and query molecule, and take a single atom as the center, and take the atoms whose connection distance is less than or equal to 2 as the elements of the subgraph;

[0080] S22, gather all the elements to form corresponding subgraphs, wherein there are no duplicate atoms between subgraphs, and the number of atoms in each subgraph is greater than 2.

[0081] S3 sums the pre-trained vectorized embeddings of all atoms in each node of the reduced graph, which serves as the representation s of that node in the corresponding subgraph. i ;include:

[0082] S31, the target molecule and the query molecule are input into the trained graph neural network to obtain the target molecule representation and the query molecule representation, which together form a pre-trained vectorized mapping (embedding) of all atoms.

[0083] S32, sum the pre-trained vectorized embeddings of all atoms in each node of the reduced graph, and use this sum as the representation s of that node in the corresponding subgraph. i ;

[0084] In this embodiment, equation (5) is specifically used:

[0085]

[0086] Among them, z j This represents the pre-trained vectorized mapping (embedding) of atoms.

[0087] In graph embedding algorithms based on graph subgraphs, the subgraph is essentially a substructure. The algorithm process is analogous to word2vec and doc2vec, using substructures within a compound as basic embedding units to obtain the final molecule embedding. In graph2vec, substructures update their neighborhood information through the substructures of neighboring nodes. In mol2vec, a substructure is merely a hash value after a single search at a specified step size (lower-order substructure), while in graph2vec, the substructure is the true structure within the graph (higher-order substructure).

[0088] S4, traverse the maximum common subgraph probability of the two subgraphs (reduced graphs), and calculate the Tanimoto similarity coefficient between each node of the query molecule and the target molecule in the two subgraphs based on the probability and the representation of the node in the corresponding subgraph, where the maximum common subgraph Tanimoto similarity coefficient is used as the similarity between the two molecules.

[0089] like Figure 3 As shown, this embodiment specifically includes:

[0090] S41, use nexworkx to traverse the maximum common subgraph probability of two subgraphs (reduced graphs) to obtain multiple pairs of common subgraphs;

[0091] S42, based on equation (6), calculate the Tanimoto similarity coefficient between each node of the query molecule and the target molecule in each pair of common subgraphs:

[0092]

[0093] Where, N t It is the number of atoms in the target molecule, s qi It is the representation of the i-th node of the query molecule, s ti It is the i-th node representation of the target molecule;

[0094] The Tanimoto coefficient is an extension of the Jaccard coefficient and is also known as the generalized Jaccard similarity coefficient. It is calculated as follows: where A and B represent two vectors, and each element in the set represents a dimension of the vector. In each dimension, the value is usually between [0,1] (if the value is a binary vector of 0 or 1, then the Tanimoto coefficient is equivalent to the Jaccard distance).

[0095] S43, take the maximum common subgraph similarity as the similarity between the two molecules, as shown in equation (7).

[0096]

[0097] Example 2

[0098] like Figure 4 As shown, a molecular similarity calculation system based on bioactive fingerprints is provided, comprising:

[0099] The neural network training module 101 is used to pre-train a molecular graph based on bioactivity data of molecules and target proteins to obtain a graph neural network for characterizing molecules, and to obtain a pre-trained vectorized mapping (embedding) of each atom in the molecule based on the graph neural network.

[0100] The subgraph building module 102 is used to decompose the molecule into target molecules and query molecules to obtain the corresponding subgraphs (reduced graphs);

[0101] Representation module 103 is used to sum the pre-trained vectorized embeddings of all atoms in each node of the reduced graph, as the representation s of that node in the corresponding subgraph. i ;

[0102] The molecular similarity calculation module 104 is used to traverse the maximum common subgraph probability of two subgraphs (reduced graphs), and calculate the Tanimoto similarity coefficient between each node of the query molecule and the target molecule in the two subgraphs based on the probability and the representation of the node in the corresponding subgraph, wherein the maximum common subgraph Tanimoto similarity coefficient is used as the similarity between the two molecules.

[0103] Example 3

[0104] This paper presents an application of a molecular similarity calculation method based on bioactive fingerprints in the discovery of replaceable molecular fragments for backbone transitions.

[0105] Example 4

[0106] This paper presents an application of a molecular similarity calculation method based on bioactive fingerprints in molecular drug development.

[0107] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for calculating molecular similarity based on bioactive fingerprints, characterized in that, include: S1, Based on the bioactivity data of molecules and target proteins, the molecular graph is pre-trained to obtain a graph neural network for characterizing molecules, and based on the graph neural network, a pre-trained vectorized mapping of each atom in the molecule is obtained. S2, decompose the molecule into target molecule and query molecule to obtain the corresponding sub-graph; S3 sums the pre-trained vectorized mappings of all atoms in each node of the subgraph, which serve as the representation of that node in the corresponding subgraph. S4, traverse the maximum common subgraph probability of the two subgraphs, and calculate the Tanimoto similarity coefficient between each node of the query molecule and the target molecule in the two subgraphs based on the probability and the representation of the node in the corresponding subgraph, wherein the maximum common subgraph Tanimoto similarity coefficient is used as the similarity between the two molecules. The graph neural network used to characterize molecules is GAT, and the process of pre-training the molecular graph based on bioactivity data of molecules and target proteins to obtain the graph neural network for characterizing molecules includes: S11 aggregates atomic information in molecules; S12 utilizes a recurrent neural network (GRU) to represent node propagation in the graph network; S13, using a multilayer fully connected network (MLP) to predict the biological activity of molecules; S11 includes: Information about atomic polymerization bonds is obtained as shown in equation (1): （1）； in, Represents atoms, Represents the bonds between atoms. The atoms following the information about the bonding of the aggregates, Indicates feature splicing; Based on the aggregation of atomic information in the molecule, as shown in equation (2): （2） in, Atomic representation after aggregating atomic information. This represents the attention weights of atoms i and j in the k-th head. This represents the learnable weight of the k-th head. Represents a non-linear activation function; S12, using the recurrent neural network GRU, represents the node propagation in the graph network as expressed by equation (3): （3）； in, This indicates the final molecular characterization; S3 includes: S31, the target molecule and the query molecule are input into the trained graph neural network to obtain the target molecule representation and the query molecule representation, which together form a pre-trained vectorized mapping of all atoms. S32, sum the pre-trained vectorized mappings of all atoms in each node of the subgraph, as the representation of that node in the corresponding subgraph. ; expressed using equation (5): （5） in, Represents the pre-trained vectorized mapping of atoms.

2. The molecular similarity calculation method based on bioactive fingerprints according to claim 1, characterized in that, S13 uses a multilayer fully connected network (MLP) to predict the bioactivity of molecules, expressed by equation (4): （4）； y represents biological activity.

3. The molecular similarity calculation method based on bioactive fingerprints according to claim 2, characterized in that, S2 includes: S21, decompose the molecule into target molecule and query molecule, and take a single atom as the center, and take the atoms whose connection distance is less than or equal to 2 as the elements of the subgraph; S22, gather all the elements to form corresponding subgraphs, wherein there are no duplicate atoms between subgraphs, and the number of atoms in each subgraph is greater than 2.

4. The molecular similarity calculation method based on bioactive fingerprints according to claim 3, characterized in that, S4 includes: S41, use nexworkx to traverse the maximum common subgraph probability of two subgraphs to obtain multiple pairs of common subgraphs; S42, based on equation (6), calculate the Tanimoto similarity coefficient between each node of the query molecule and the target molecule in each pair of common subgraphs: （6）； in, It is the number of atoms in the target molecule. It represents the i-th node of the queried molecule. It is the i-th node representation of the target molecule; S43, take the maximum common subgraph similarity as the similarity between the two molecules, as shown in equation (7); （7）。 5. A molecular similarity calculation system based on bioactive fingerprints, used to implement the method according to any one of claims 1-4, characterized in that, include: The neural network training module (101) is used to pre-train the molecular graph based on the bioactivity data of the molecule and the target protein to obtain a graph neural network for characterizing the molecule, and to obtain a pre-trained vectorized mapping of each atom in the molecule based on the graph neural network. The subgraph building module (102) is used to decompose the molecule into target molecule and query molecule to obtain the corresponding subgraph; The representation module (103) is used to sum the pre-trained vectorized mappings of all atoms in each node of the subgraph as the representation of that node in the corresponding subgraph. ; The molecular similarity calculation module (104) is used to traverse the probability of the maximum common subgraph between two subgraphs and calculate the Tanimoto similarity coefficient between each node of the query molecule and the target molecule in the two subgraphs, where the Tanimoto similarity coefficient of the maximum common subgraph is used as the similarity between the two molecules.

6. The application of the molecular similarity calculation method based on bioactive fingerprints as described in any one of claims 1-4 in the discovery of alternative molecular fragments for scaffold transitions and / or in molecular drug development.

Images