Liver microsome compound stability prediction method based on deep neural network

Abstract

The invention belongs to the technical field of biology, and particularly relates to a liver microsome compound stability prediction method based on a deep neural network, which comprises the following steps: acquiring liver microsome compound molecule data, and representing each liver microsome compound molecule as a corresponding SMILES character string; character group-to-ID processing is carried out on the liver microsome compound molecule data to obtain matrix vector data; constructing a deep neural network model; training a deep neural network model by adopting matrix vector data, calculating loss and carrying out back propagation to train model parameters; and predicting the stability of the liver microsome compound by adopting the trained deep neural network model. According to the invention, high-precision molecular stability prediction can be carried out.

Description

Technical Field

[0001] The present invention belongs to the field of biotechnology, and in particular relates to a method for predicting the stability of liver microsome compounds based on a deep neural network. Background Art

[0002] Preclinical research is an essential part of new drug development, and drug metabolism dynamics is an important part of preclinical pharmacological evaluation. After the drug is absorbed by the body, the process of chemical structure change under the action of various enzymes and body fluid environment in the body is called drug metabolism (Drug Metabolism), also known as biotransformation (Biotransformation). Drug metabolism mainly occurs in the liver, and also occurs in organs such as the intestines, kidneys, lungs, blood and skin. The biotransformation of drugs in the body is divided into phase I metabolic reactions and phase II metabolic reactions. The reaction of introducing new groups or removing the functional groups of the original small groups on the drug molecules is called phase I metabolism, including oxidation, reduction and hydrolysis reactions. The reaction of drugs or phase I metabolites binding to certain endogenous small molecules in the body is phase II metabolism, also known as binding reaction, such as glucuronic acid binding, phosphorylation, methylation, acetylation, glutathione binding and other reactions. After most lipophilic drugs are absorbed, they can be transformed into polar and water-soluble metabolites through phase I metabolism, which is conducive to the conduct of phase II metabolism and increases polarity. Phase II metabolism is the true "detoxification" pathway, and its metabolites are usually more water-soluble and more easily excreted from the body through urine and bile.

[0003] Hepatic microsomes are a unique structure in liver cells, located around the mitochondria of liver cells and arranged in a ring on their outer surface. Hepatic microsomes are the main source of metabolic enzymes, the most famous of which is the cytochrome P450 (CYP) enzyme system. The CYP enzyme system plays an extremely important role in drug metabolism, detoxification and the treatment of endogenous substances; in addition, it also includes UDP-glucuronyl transferase, etc. These enzymes participate in conjugation reactions and help convert toxic substances into water-soluble compounds, which are easy to excrete; therefore, hepatic microsomes are widely used in in vitro drug metabolism studies. Hepatic microsomes can not only provide qualitative information for systemic drug metabolism, but also can be used to screen and optimize the metabolic stability of a large number of compounds during drug development, and can also be used to evaluate potential drug-drug interactions.

[0004] Confirming the stability of liver microsomes is crucial for studying their role in drug metabolism and biological reactions. The stability of liver microsomes not only affects the reliability of experimental results, but is also directly related to the accurate prediction of drug activity and potential toxicity. There are currently a variety of methods for evaluating the stability of liver microsomes. A more direct one is the in vitro enzyme reaction method, which simulates the drug metabolism process by using known liver microsomal enzymes in the laboratory; this method can provide accurate data, but it has the problems of high cost and long time. Another is conventional biochemical and biochemical analysis, that is, using spectral analysis (such as UV-Vis spectroscopy), mass spectrometry, high-performance liquid chromatography (HPLC) and other methods to measure the change in concentration of specific compounds before and after the reaction to predict stability; although these methods are relatively fast and low-cost, they lack comprehensive capture of complex dynamic processes. There are also ways to use theoretical chemical principles, such as molecular docking and molecular dynamics simulation, to predict the possibility of drug binding to enzymes and their possible metabolic pathways; the advantage of this method is that it can provide a deep understanding and reduce the need for experimental operations to a certain extent, but it requires high computing resources and professional skills. In addition, there is a statistical learning model that builds regression or classification models (such as support vector machines, decision trees, etc.) based on historical data sets to predict the metabolic stability of new compounds through feature engineering; these methods can handle large amounts of data and automate the prediction process to a certain extent, but are affected by overfitting problems and specific dependence on training data.

[0005] Liver microsomes contain a variety of enzymes and other biomolecules, which have different structures, functions and regulatory mechanisms, making their stability assessment very complicated. At the same time, different types of CYP enzymes show differences in activity on different substrates, and there are significant differences in their responses to environmental factors (such as temperature and pH). This diversity of intrinsic properties increases the difficulty of stability prediction. In addition, the function of liver microsomes is closely related to the state and disease state of the individual. For example, enzyme activity and protein levels may change after liver disease or drug exposure, which affects the consistency and reliability of stability assessment. Based on this, it is very necessary to propose a stability prediction method for liver microsome compounds based on deep neural networks. Summary of the invention

[0006] To solve the above problems, the present invention provides a method for predicting the stability of liver microsome compounds based on a deep neural network, comprising the following steps:

[0007] S1. Obtain liver microsomal compound molecule data, wherein each liver microsomal compound molecule is represented by a corresponding SMILES string;

[0008] S2. Perform character group conversion on the molecular data of liver microsome compounds to obtain matrix vector data;

[0009] S3. Build a deep neural network model;

[0010] S4. Use matrix vector data to train the deep neural network model, calculate the loss and back-propagate the training model parameters;

[0011] S5. Use the trained deep neural network model to predict the stability of liver microsomal compounds.

[0012] Beneficial effects of the present invention:

[0013] Compared with conventional prediction methods, the present invention introduces deep neural network technology to predict the stability of liver microsome compounds; the deep neural network model constructed by the present invention can better obtain important characteristic information of liver microsome compound molecules, more comprehensively consider the influence of various factors, and thus improve the prediction accuracy. Compared with traditional prediction methods, the present invention does not need to use known enzymes for reactions, which reduces experimental costs and does not take too much time.

[0014] The application of deep neural network models can speed up the screening process of candidate drugs, which helps to predict the liver microsomal stability of compounds in the early stages of drug discovery and reduce the failure rate of compounds in subsequent trial stages, thereby reducing the waste of resources.

[0015] At the same time, the introduction of deep neural network models can reduce dependence on computing resources and achieve more efficient predictions.

[0016] At the same time, the deep neural network model has strong generalization ability and is not limited to the stability prediction of a single species. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the method of the present invention;

[0018] Figure 2 The molecular structure diagram of the embodiment of the present invention;

[0019] Figure 3 This is a molecular structure diagram of the group of embodiments of the present invention;

[0020] Figure 4 This is a structural diagram of the deep neural network model of the present invention;

[0021] Figure 5 This is a schematic diagram of the BRN module structure of the present invention;

[0022] Figure 6 It is a schematic diagram of the Stage module structure of the present invention. DETAILED DESCRIPTION

[0023] 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.

[0024] The present invention provides a method for predicting the stability of liver microsome compounds based on deep neural network, such as Figure 1 As shown, the following steps are included:

[0025] S1. Obtain liver microsomal compound molecule data, wherein each liver microsomal compound molecule is represented as a corresponding SMILES string.

[0026] Specifically, relevant liver microsomal compound molecular data are obtained through public databases, scientific literature, or experiments, and the liver microsomal compound molecular data undergo a series of preprocessing, including data cleaning, format unification, and removal of duplicate or invalid records.

[0027] S2. Perform character group to ID conversion on the molecular data of liver microsome compounds to obtain matrix vector data.

[0028] Preferably, in cheminformatics, SMILES (Simplified Molecular Input Line Entry System) is a linear symbol system for representing molecular structures, which uses ASCII characters to describe the atoms and chemical bonds of a molecule. A SMILES string can uniquely represent the structure of a molecule, including its stereochemical characteristics. When processing large chemical databases, it is often necessary to standardize and normalize SMILES strings to ensure consistency and comparability of data.

[0029] The process of converting character groups to IDs for each SMILES string in step S2 includes:

[0030] S21. Use the RDKit tool library to standardize the SMILES string.

[0031] Specifically, the SMILES string can be normalized using the RDKit tool library, which can convert the SMILES string into a standardized form. This usually involves removing unnecessary hydrogen atoms, eliminating stereochemical ambiguity, and converting the molecular structure into a standard representation method.

[0032] Meanwhile, the BRICS algorithm of the RDKit toolkit can be used to identify potential breakpoints in the molecule and split the molecule into smaller fragments (groups). These fragments can be used for further chemical analysis or as intermediates in chemical reactions.

[0033] S22. Split the standardized SMILES string character by character and assign a character position serial number to each character to obtain a character position information vector of size 1×N, where N is the number of characters.

[0034] S23. Assign a character ID to each character according to the character type to form a character ID vector of size 1×N.

[0035] Specifically, the SMILES string usually contains atomic symbols (such as C, N, O, etc.), numbers, chemical bond representations (such as -, =, #, etc.), representations of branches and rings (such as (, ), [, ], etc.), and other special characters (such as @ representing a chiral center). Assign a unique character ID to each type of character. This character ID can be a number or a string, used to identify a specific character in subsequent data processing. At the same time, assign the corresponding character position serial number according to the position of the character in the SMILES string.

[0036] S24. Split the standardized SMILES string into groups and assign a group position serial number to each group to obtain a group position vector of size 1×M, where M is the number of groups and M < N; S25. Assign a group ID to each group to form a group ID vector of size 1×M;

[0037] S26. Perform zero-padding on the group position vector and the group ID vector so that the sizes of the two vectors are expanded to 1×N;

[0038] S27. Stack the character position vector, the character ID vector, the group position vector, and the group ID vector to obtain a matrix vector of size 4×N.

[0039] In one embodiment, the present invention obtains the standardized SMILES strings of the related hepatic microsomal compound molecules as shown in Table 1.

[0040] Table 1 Related hepatic microsomal compound molecules

[0041] SMILES r1 CS(=O)(=O)c1ccccc1OC2CCN(CC2)C(=O)NCc3ccc(Cl)cc3Cl r2 Nc1nc(Nc2ccc(cc2)S(=O)(=O)N)n(n1)C(=S)Nc3c(F)cccc3F r3 CCOC(=O)N1CC(N)C(C1)c2ccc(Cl)cc2Cl r4 Ccc1ccc(NC(=O)c2cc3ncccn3n2)cc1

[0042] The molecular structure diagrams of the 4 hepatic microsomal compound molecules r1 - r4 in Table 1 are as Figure 2 shown in (a) - (d).

[0043] Using the RDKit toolkit for splitting, the results are shown in Table 2.

[0044] Table 2 Group separation results

[0045]

[0046] For better explanation, Figure 3 (a)-(e) show the molecular structures of the various groups of the liver microsomal compound molecule r4 in Table 2.

[0047] In the character group conversion ID processing of the present invention, the SMILES character conversion ID rules are shown in Table 3.

[0048] Table 3 SMILES character conversion rules

[0049] character # ％ ( ) + - . / 0 1 2 3 4 5 6 7 8 9 ＝ ID 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 character @ B C F H I N O P S [ \\ ] c l n o r s ID 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38

[0050] Taking the liver microsomal compound molecule r4 as an example, the results of character group conversion to ID processing are shown in Tables 4 and 5.

[0051] Table 4 SMILES character conversion to ID

[0052] char C C c 1 c c c ( N C ( ＝ O ) c char ID 22 22 33 10 33 33 33 3 23 22 3 19 27 4 33 Position number 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 char 2 c c 3 n c c c n 3 n 2 ) c c 1 char ID 11 33 33 12 35 33 33 33 35 12 35 11 4 33 33 10 Position number 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

[0053] Table 5 SMILES group transfer ID

[0054] Group [1*]C([6*])=O [14*]c1cc2ncccn2n1 [16*]c1ccc([16*])cc1 [5*]N[5*] [8*]CC Group ID 106 43 335 4 88 Group position number 1 2 3 4 5

[0055] S3. Construct a deep neural network model, which includes a feature extraction module, a feature fusion module and a feature prediction module.

[0056] Specifically, Figure 4 As shown, the feature extraction module includes an Embedding layer, a CBNR unit, a first BR2 module, a BR3 module, a BR5 module, a second BR2 module and a convolutional layer that are cascaded in sequence;

[0057] Among them, the Embedding layer is used to map the matrix vector data composed of ID data to the vector space; the first BR2 module, BR3 module, BR5 module and the second BR2 module are mainly used to perform multiple downsampling operations to extract multi-layer features; the first BR2 module performs downsampling calculations on the output of the CBNR unit, and the BR3 module performs downsampling calculations on the output of the first BR2 module; the BR5 module performs downsampling calculations on the output of the BR3 module; the second BR2 module performs downsampling calculations on the output of the BR5 module.

[0058] The feature fusion module includes a first CSS unit, a first upsampling layer, a second CSS unit, a second upsampling layer, a third CSS unit, and a third upsampling layer, which are cascaded in sequence;

[0059] The feature prediction module includes a Stage A module, a Stage B module, a Stage C module, a Stage D module Concat layer and a GES unit; the four stage modules constitute the multi-scale output of the model, and the outputs of the four Stage modules are spliced ​​through the Concat layer to obtain information that integrates features of various levels.

[0060] The input of the first upsampling layer includes the output of the convolution layer and the output of the first CSS unit; the input of the second upsampling layer includes the output of the second CSS unit and the sum of the output of the BR5 module and the first upsampling layer; the input of the third upsampling layer includes the output of the third CSS unit and the sum of the output of the BR3 module and the second upsampling layer;

[0061] The input of the Stage A module is the sum of the outputs of the first BR2 module and the third upsampling layer; the input of the Stage B module is the sum of the outputs of the BR3 module and the second upsampling layer; the input of the Stage C module is the sum of the outputs of the BR5 module and the first upsampling layer; the input of the Stage D module is the output of the convolution; the Concat layer is used to concatenate the outputs of the Stage A module, Stage B module, Stage C module, and Stage D module.

[0062] Specifically, Figure 4 As shown, the CBNR unit is the basic unit of the model, which includes a 1×1 convolutional layer, a BN layer, a ReLu layer, and a maximum pooling layer cascaded in sequence;

[0063] The first CSS unit, the second CSS unit and the third CSS unit have the same structure, and all include a Shape layer, a Slice layer and a Concat layer cascaded in sequence;

[0064] The GES unit includes a Gemm (General Matrix Multiplication) layer, an Elu (Exponential Linear Units) layer, and a Softmax layer which are cascaded in sequence.

[0065] Specifically, Figure 5 As shown, the first BR2 module has the same structure as the second BR2 module, both of which include 1 double-branch convolution residual block and 2 single-branch convolution residual blocks; the BR3 module includes 1 double-branch convolution residual block and 3 single-branch convolution residual blocks; the BR5 module includes 1 double-branch convolution residual block and 5 single-branch convolution residual blocks; among them,

[0066] The dual-branch convolution residual block includes a first convolution branch and a second convolution branch; the first convolution branch includes a convolution layer; the second convolution branch includes a stacked convolution block, the stacked convolution block includes 3 convolution layers, and a ReLu activation function is connected between every two convolution layers; the outputs of the first convolution branch and the second convolution branch are added and output;

[0067] A single convolution residual block includes a ReLu layer and a stacked convolution block connected in sequence, and the output of the ReLu layer is added to the output of the stacked convolution block and then output.

[0068] Specifically, Figure 6 As shown, the Stage A module includes 5 convolutional layers, 4 ReLu layers, 1 Reshape layer, 2 Gemm layers, and 2 Elu layers;

[0069] The Stage B module includes 4 convolutional layers, 3 ReLu layers, 1 Reshape layer, 2 Gemm layers, and 2 Elu layers;

[0070] The Stage C module includes 3 convolutional layers, 2 ReLu layers, 1 Reshape layer, 2 Gemm layers, and 2 Elu layers;

[0071] The Stage D module includes 1 convolution layer, 1 ReLu layer, 1 Reshape layer, 2 Gemm layers and 2 Elu layers.

[0072] S4. Use matrix-vector data to train the deep neural network model, calculate the loss and back-propagate the training model parameters.

[0073] Specifically, the classic loss function SmoothL1Loss is used, which combines the advantages of L1 loss and L2 loss, can reduce the sensitivity to outliers to a certain extent, and prevent gradient explosion in some cases.

[0074] S5. Use the trained deep neural network model to predict the stability of liver microsomal compounds.

[0075] The network structure proposed by the present invention can realize the prediction of the stability of liver microsome compounds under limited data volume, and the accuracy rate reaches 83.5%.

[0076] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "rotation" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0077] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the stability of liver microsome compounds based on deep neural networks, characterized in that: The following steps are involved: S1. Obtain liver microsomal compound molecule data, wherein each liver microsomal compound molecule is represented by a corresponding SMILES string; S2. Perform character group conversion on the molecular data of liver microsome compounds to obtain matrix vector data; S3. Constructing a deep neural network model; the deep neural network model includes a feature extraction module, a feature fusion module and a feature prediction module; S4. Use matrix vector data to train the deep neural network model, calculate the loss and back-propagate the training model parameters; S5. Use the trained deep neural network model to predict the stability of liver microsomal compounds.

2. The method for predicting the stability of liver microsome compounds based on deep neural network according to claim 1, characterized in that: The process of converting character groups to IDs for each SMILES string in step S2 includes: S21. Use RDKit tool library to standardize SMILES strings; S22. Segment the standardized SMILES string into characters, and assign a character position number to each character to obtain a character position information vector of size 1×N, where N is the number of characters; S23. Assign a character ID to each character according to the character type to form a character ID vector of size 1×N; S24. Segment the standardized SMILES string into groups and assign a group position number to each group to obtain a group position vector of size 1×M, where M is the number of groups and M <N; S25. assign a group ID to each group to form a group ID vector of size 1×M; S26. Fill the group position vector and the group ID vector with 0 so that the sizes of the two vectors are expanded to 1×N; S27. Stack the character position vector, character ID vector, group position vector and group ID vector to obtain a matrix vector of size 4×N.

3. The method for predicting the stability of liver microsome compounds based on deep neural network according to claim 1, characterized in that: The feature extraction module includes an Embedding layer, a CBNR unit, a first BR2 module, a BR3 module, a BR5 module, a second BR2 module and a convolutional layer that are cascaded in sequence; The feature fusion module includes a first CSS unit, a first upsampling layer, a second CSS unit, a second upsampling layer, a third CSS unit, and a third upsampling layer, which are cascaded in sequence; The feature prediction module includes a Stage A module, a Stage B module, a Stage C module, a Stage D module, a Concat layer and a GES unit; The input of the first upsampling layer includes the output of the convolution layer and the output of the first CSS unit; the input of the second upsampling layer includes the output of the second CSS unit and the sum of the output of the BR5 module and the first upsampling layer; the input of the third upsampling layer includes the output of the third CSS unit and the sum of the output of the BR3 module and the second upsampling layer; The input of the Stage A module is the sum of the outputs of the first BR2 module and the third upsampling layer; the input of the Stage B module is the sum of the outputs of the BR3 module and the second upsampling layer; the input of the Stage C module is the sum of the outputs of the BR5 module and the first upsampling layer; the input of the Stage D module is the output of the convolution; the Concat layer is used to concatenate the outputs of the Stage A module, Stage B module, Stage C module, and Stage D module.

4. The method for predicting the stability of liver microsome compounds based on deep neural network according to claim 3, characterized in that: The CBNR unit consists of a cascaded 1×1 convolutional layer, a BN layer, a ReLu layer, and a maximum pooling layer; The first CSS unit, the second CSS unit and the third CSS unit have the same structure, and all include a Shape layer, a Slice layer and a Concat layer cascaded in sequence; The GES unit consists of a Gemm layer, an Elu layer, and a Softmax layer cascaded in sequence.

5. The method for predicting the stability of liver microsome compounds based on deep neural network according to claim 3, characterized in that: The first BR2 module has the same structure as the second BR2 module, both of which include 1 double-branch convolution residual block and 2 single-branch convolution residual blocks; the BR3 module includes 1 double-branch convolution residual block and 3 single-branch convolution residual blocks; the BR5 module includes 1 double-branch convolution residual block and 5 single-branch convolution residual blocks; among them, The dual-branch convolution residual block includes a first convolution branch and a second convolution branch; the first convolution branch includes a convolution layer; the second convolution branch includes a stacked convolution block, the stacked convolution block includes 3 convolution layers, and a ReLu activation function is connected between every two convolution layers; the outputs of the first convolution branch and the second convolution branch are added and output; A single convolution residual block includes a ReLu layer and a stacked convolution block connected in sequence, and the output of the ReLu layer is added to the output of the stacked convolution block and then output.

6. The method for predicting the stability of liver microsome compounds based on deep neural network according to claim 3, characterized in that: The Stage A module includes 5 convolutional layers, 4 ReLu layers, 1 Reshape layer, 2 Gemm layers, and 2 Elu layers; The Stage B module includes 4 convolutional layers, 3 ReLu layers, 1 Reshape layer, 2 Gemm layers, and 2 Elu layers; The Stage C module includes 3 convolutional layers, 2 ReLu layers, 1 Reshape layer, 2 Gemm layers, and 2 Elu layers; The Stage D module includes 1 convolution layer, 1 ReLu layer, 1 Reshape layer, 2 Gemm layers and 2 Elu layers.

Images