Prediction model of anti-HBV small molecule drugs based on representation learning and its construction method

Abstract

The present invention provides an anti-HBV small molecule drug prediction model based on representation learning and a construction method thereof, which relates to the technical field of drug research. By processing SMILES coding, the SMILES coding is represented and learned using an NLP method. A supervised classification model of a spatial vector of SMILES coding is used using a machine learning method, and the activity of small molecules on specific targets is classified and predicted in the model. Compared with the prior art, it shows better performance than traditional Word2vec on different target data sets. At the same time, in the downstream tasks after the training vector, it has a deterministic advantage in predicting the HBV inhibition rate and hepatocellular toxicity of the compound, and has a good ability to screen potential anti-HBV drugs. The method can be used for the prediction of drug-like properties of other different target compounds, thereby simplifying the drug development process.

Description

Technical Field

[0001] The present invention relates to the technical field of drug research and development, and in particular to a method for constructing an anti-HBV small molecule drug prediction model based on representation learning. Background Art

[0002] In the process of drug development, designing and synthesizing highly active compounds against specific targets is the primary task of medicinal chemistry. Conventional drug design relies more on various types of high-throughput virtual screening, fragment- or structure-based drug design, and the experience of drug developers, and lacks a convenient, fast, and accurate way to evaluate the pharmacological activity of drug molecules against specific targets.

[0003] Limited by the conventional drug development process, large-scale drug molecule design, synthesis and anti-HBV activity experiments consume a lot of manpower, material resources, time and money, which slows down the speed of drug development. Through artificial intelligence drug screening prediction models, we can effectively speed up drug development, save a lot of cost output, and successfully predict and screen potential active drug molecules, which is also an inevitable trend in future drug development. Summary of the invention

[0004] The purpose of the present invention is to solve the problems in the prior art.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for constructing an anti-HBV small molecule drug prediction model based on representation learning comprises the following steps:

[0007] S1: Construct a sparse matrix W based on the data structure of the Ising Model using the compound structure token data extracted from SMILES. ISM ;

[0008] S2: W ISM After SLEP's large-scale sparse data processing, the correlation matrix W is obtained. WCM ;

[0009] S3: Divide the input text data entity into n batches to process the data. We take the i-th batch as the batch. i , Batch i The entity in (Vm i1 ,Vm i2 ,…,Vm ij ) T ;

[0010] S4: Input Batch i Chemical structural entity (Vm i1,Vm i2 ,…,Vm ij ) T , initialize the Embedding (Ve i1 ,Ve i2 ,…,Ve ij ) T ;

[0011] S5: Batch i After pairing, the Skip-Gram relationship is generated to obtain the upper and lower structures Con(Vm ij ) and negative sampling word space NEG (Vm ij ), continue to calculate its predicted probability u∈NEG(Vm ij );

[0012] Based on Batch i Medium Ve ij Count contextCon(Vm ij ) and negative sampling word space NEG (Vm ij ) and calculate the loss function (6):

[0013]

[0014] S6: According to the loss function obtained by formula (6), a variety of optimization schemes can be adopted to obtain the gradient in order to optimize the parameters;

[0015] S7: Based on the pairing relationship between entities, WCM The local relationship matrix W corresponding to the entity-entity pairing relationship is obtained by the index SCM And the relationship matrix W with the negative sampling pairing relationship NCM ;

[0016] S8: W SCM , W NCM After combining with the gradient, redistributing the weights instead of averaging, back propagation updates Vm ij Word embedding And the auxiliary vector corresponding to word u

[0017] Preferably, in S2, Logistic Regression is used to solve W ISM The regularized logic problem of . The formula is as follows:

[0018]

[0019] W WCM=(X1X2,…,X all )+(X1X2,…,X all ) T (2).

[0020] Preferably, S4 includes:

[0021] (Ve i1 ,Ve i2 ,…,Ve ij ) T =(Vm i1 ,Vm i2 ,…,Vm ij ) T ·W Ve (4)

[0022] Vm ij In the above code, j stands for Batch. i The jth entity in , correspondingly, Ve ij is the embedding corresponding to the jth entity, W Ve It is the weight matrix in the vector generation process.

[0023] Preferably, in S5:

[0024]

[0025] In the above formula (5), u∈NEG(Vm ij ), σ is the sigmoid activation function, The parameters corresponding to word u are to be trained.

[0026] Preferably, the optimization scheme in S6 is:

[0027] set up:

[0028]

[0029] According to formula (7), we can calculate the u (8) and Ve ij The gradient of (9):

[0030]

[0031] Preferably, in S8:

[0032]

[0033]

[0034] The present application also provides an anti-HBV small molecule drug prediction model based on representation learning, which is constructed using the above-mentioned method for constructing an anti-HBV small molecule drug prediction model based on representation learning.

[0035] The above-mentioned anti-HBV small molecule drug prediction model based on representation learning shows better performance than traditional Word2vec on different target data sets. At the same time, in the downstream tasks after training the vector, it has a deterministic advantage in predicting the HBV inhibition rate and liver cell toxicity of the compound, and has a good ability to screen potential anti-HBV drugs. In addition, in the process of generating Tokens, multiple methods were compared, and the token extraction parameters that can enable the model to achieve optimal performance were tried out. Not only did the model show its superior performance in compound representation learning on public data sets, but the accurate prediction ability of the classification model based on the Ising-Word2vec pre-training results was also confirmed through cell experiments, demonstrating the application potential of a Word2vec combined with Ising gradient correction model in screening new potential drug-like compounds. This method can be widely used in the prediction of drug-likeness of other different target compounds, thereby simplifying the drug development process. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic flow chart of a method for constructing an anti-HBV small molecule drug prediction model based on representation learning is proposed in the present invention;

[0037] Figure 2 A schematic diagram of a web tool page for an anti-HBV small molecule drug prediction model based on representation learning is proposed in the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with specific embodiments.

[0039] Partial English interpretation:

[0040] SMILES: Simplified Molecular Input Line Entry System, a chemical symbol system designed for modern chemical information processing. Based on the principles of molecular graph theory, SMILES allows for strict structural descriptions through a very small and natural syntax. The SMILES symbol system is also suitable for computer recognition and processing.

[0041] Because it is easy to use by researchers and readable by computers, it is suitable for use in models designed to produce unique digital symbols, continuous database searches, smooth substructure searches, and performance prediction models.

[0042] ECFP: Molecular fingerprints are representations of chemical structures that were originally designed to assist substructure searches in chemical databases, but were later used for analytical tasks such as similarity searches, clustering, and classification. Currently, commonly used molecular fingerprints include MACCS, PubChem, Tree fingerprint, MolPrint2D, ECFP, FCFP, UNITY 2D, MP-MFP, etc. ECFP (Extended Connectivity Fingerprint) is a fingerprint methodology explicitly designed to capture molecular features associated with molecular activity. Although not designed for substructure searches, they are well suited for tasks related to prediction and insight into drug activity. In addition, ECFP can be used in methods like other fingerprints, such as similarity searches, clustering, and virtual screening. Since 2000, ECFP has been applied to a wide range of scientifically relevant problems using a variety of analytical methods.

[0043] Word2Vec: A mainstream word embedding model with two different structures: CBOW and Skip-Gram. CBOW and Skip-Gram are similar overall. The main difference between the two structures is that CBOW uses context to predict words, while Skip-Gram uses the target word to predict the context. There are many schemes for the gradient of neural networks, such as stochastic gradient descent (SGD), Momentum, Adagrad, etc. Most current word embedding models adopt a simple stochastic gradient optimization scheme, that is, for different words in the bag of words, the gradient is equally contributed to each word vector. The most commonly used scheme in Word2vec is the stochastic gradient descent scheme.

[0044] ChEMBL: https: / / chembl.gitbook.io / chembl-interface-documentation / downloadsIt is an online free database developed by EBI (European Bioinformatics Institute), which contains a large number of binding, functional and ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) information of pharmaceutical compounds. These data are manually extracted from major published literature on a regular basis, and then further organized and standardized. At present, the database contains 2,105,464 compounds, 14,554 targets, and 1,383,553 drug activity tests. Through this database, you can quickly query the compounds and their activity information that have been reported for a certain target. The data in this database are all from various reported literatures, the data are relatively reliable, and can be traced to the source of the data. Through this database, you can quickly obtain accurate compounds and their biological data, further accelerating the speed of drug design and drug development. This study will screen potential HBV (Hepatitis B Virus) inhibitory drugs, while taking into account their liver toxicity, so compounds targeting HBV and HepG2 were screened as training data.

[0045] W NCM :Sub Correlation Matrix

[0046] W ISM :Ising Sparse Matrix

[0047] SLEP: Sparse Learning with Efficient Projections

[0048] W WCM :Whole Correlation Matrix

[0049] Token:

[0050] Logistic Regression

[0051] Batch: Group

[0052] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments of the following disclosure.

[0053] Usually, word segmentation is required before the text is subjected to the next step of pre-training. Figure 1As shown in Part A, for chemical structure encoding SMILES, character segmentation (SC, Splitted Character) and molecular fingerprint (ECFP) can be used:

[0054] SMILES to SC

[0055] 1. Perform sliding window shift cutting with a specific window size in the order of SMILES, with each shift being one character.

[0056] 2. Remove the “(” and “)” used to represent the branch at the branch, and use a single letter “R” to retain the branch information here.

[0057] 3. For double characters such as "Cl" and "Fe" that represent single atoms, replace them with single characters such as "L" and "E" that cannot be confused with other characters.

[0058] 4. Each SMILES is concatenated in the order of the main chain and the branch chain to form a linear word list.

[0059] like Figure 1 As shown in Part A, the compound with SMILES expression Nc1cc(OCCOCP(=O)(O)O)nc(N)n1 was processed by SMILES to SC to form 17 tokens as shown below:

[0060] Nc1c / c1cc / 1ccR / ccRn / cRnc / RncR / ncRn / cRn1 / OCCO / CCOC / COCP / OCPR / CPRR / PRRO / N / =O / O

[0061] SMILES to ECFP

[0062] Morgan FP (Extended Connectivity Fingerprints, ECFP) is generated by generating all substructures around all heavy atoms in a molecule within a defined radius and assigning a unique identifier (called Morgan identifier). Through these identifiers, these identifiers are usually hashed into vectors with a fixed length.

[0063] For each compound encoded by SMILES, SMILES to ECFP generates identifiers for all atoms with a fixed radius, and then arranges the identifiers of each atom into a "molecular sentence" in the order of the atoms in the SMILES encoding. At the same time, if an atom has multiple identifiers, the atomic identifiers are arranged from small to large according to their identification radius.

[0064] SMILES encoded text is composed of different atoms and atomic structures. It expresses various chemical structures such as atoms, chemical bonds, branched chains, ring structures and ionic bonds through encoding. These chemical structures are interconnected and can be distinguished from each other, forming a similar entity meaning. The concept of entity refers to an object or concept that exists in the objective world and can be distinguished from each other. The Ising model was proposed to explain the phase change of ferromagnetic materials. This transition between magnetic and non-magnetic phases can also be reflected in the entity sparse matrix constructed from text data as the transition between the existence and non-existence of the entity. After verification, it can also describe the transition between the existence and non-existence of chemical structure entities. There will be long-range associations between different components of the Ising model system. This phenomenon of long-range connections caused by local interactions is exactly what is needed in the process of constructing a global correlation matrix.

[0065] See also Figure 1 , Figure 1 The module shown in Part B is a new Ising-Word2vec designed in combination with the Ising model, which is used to flexibly capture global and local connections in the process of generating word embeddings. The tokens extracted from SMILES are input, and the corresponding global gradient-corrected word embedding model can be obtained through the model.

[0066] At this time, the chemical structure entity vector will no longer be evenly distributed to update the Embedding, but will be combined with W ISM Redistribute and combine to form new chemical structure entity Embedding. Use negative sampling function to repeatedly train compound chemical structure Embedding, and finally obtain a stable compound chemical structure Embedding representation.

[0067] The present application provides a method for constructing an anti-HBV small molecule drug prediction model based on representation learning, comprising the following steps:

[0068] S1: Construct a sparse matrix W based on the data structure of the Ising Model using the compound structure token data extracted from SMILES. ISM .

[0069] S2: W ISM After SLEP's large-scale sparse data processing, the correlation matrix W is obtained. WCM .

[0070] In one embodiment, Logistic Regression is used to solve W ISM The regularized logic problem of . The formula is as follows:

[0071]

[0072] W WCM =(X1X2,…,X all )+(X1X2,…,X all ) T (2)

[0073] The above formula (1) is to solve W WCM The calculation formula in the process, where w t is the weight of the t-th entity among all chemical structure entities, z t is our input W ISM The tth column, z kt is the entity in the kth row and tth column extracted from the sparse matrix, used to solve z kt Relevance to other entities, X k is z kt The corresponding correlation solution, λ is the l1 norm regularization parameter, and ρ is the squared 2 norm regularization parameter.

[0074] S3: Divide the input text data entity into n batches to process the data. We take the i-th batch as the batch. i . Batch i The entity in (Vm i1 , Vm i2 ,…,Vm ij ) T :

[0075]

[0076] S4: Input Batch i Chemical structural entity (Vm i1 ,Vm i2 , …, Vm ij ) T , initialize the Embedding (Ve i1 ,Ve i2 ,…,Ve ij ) T :

[0077] (Ve i1 ,Ve i2 ,…,Ve ij ) T =(Vm i1 ,Vm i2 ,…,Vm ij ) T ·W Ve (4)

[0078] Vm ij In the above code, j stands for Batch. i The jth entity in , correspondingly, Ve ij is the corresponding Embedding of the jth entity. Ve It is the weight matrix in the vector generation process.

[0079] S5: Batch i After pairing, the Skip-Gram relationship is generated to obtain the upper and lower structures Con(Vm ij ) and negative sampling word space NEG (Vm ij ), continue to calculate its predicted probability u∈NEG(Vm ij ),

[0080] in:

[0081]

[0082] In the above formula (5), u∈NEG(Vm ij ), σ is the sigmoid activation function, The parameters corresponding to word u are to be trained.

[0083] Based on Batch i Medium Ve ij Count contextCon(Vm ij ) and negative sampling word space NEG (Vm ij ) and calculate the loss function (6):

[0084]

[0085] S6: According to the loss function obtained by formula (6), a variety of optimization schemes can be adopted to obtain the gradient in order to optimize the parameters. In one embodiment, in order to facilitate understanding, the gradient descent method is adopted for detailed analysis. Assume:

[0086]

[0087] According to formula (7), we can calculate the u (8) and Ve ij The gradient of (9):

[0088]

[0089]

[0090] S7: Batch iIn S5, the Skip-Gram pairing relationship and negative sampling pairing relationship between entities have been constructed. Based on the pairing relationship between entities, from W WCM The local relationship matrix W corresponding to the entity-entity pairing relationship is obtained by the index SCM And the relationship matrix W with the negative sampling pairing relationship NCM .

[0091] S8: W SCM , W NCM After combining with the gradient, redistributing the weights instead of averaging, back propagation updates Vm ij Word embedding And the auxiliary vector corresponding to word u

[0092]

[0093]

[0094] Example:

[0095] 1. Evaluation of anti-HBV activity and cytotoxicity: Real-time quantitative PCR and MTT assay were used to evaluate the in vitro anti-HBV activity and cytotoxicity of all compounds in HepG2 2.2.15 cells. The compound concentration required for 50% inhibition of DNA replication was defined as IC50, and the compound concentration that induced 50% cell death in HepG2 2.2.15 cell culture was defined as CC50.

[0096] Cytotoxicity assay: To evaluate the in vitro cytotoxic effects of target compounds, HepG2 2.2.15 cells were treated with different doses of compounds for 72 h, and then the cell viability was determined by MTT assay.

[0097] Inhibition of HBV DNA replication: The anti-HBV activity of compounds in the HepG2 2.2.15 cell line depends on their inhibitory effect on HBV DNA replication. After treating cells with different concentrations of compounds for 6 days, extracellular HBV DNA levels were quantified by real-time PCR.

[0098] 2. Biological Analysis

[0099] HepG2 cell line was obtained from the 302 Military Hospital of China. The cell line was cultured in DMEM medium supplemented with 10% fetal bovine serum, 100 U / mL penicillin, 100 mg / mL streptomycin, and 0.38 mg / mL G418. The cytotoxicity, anti-HBV activity, and effects on HBV viral gene expression of the compounds were evaluated on the HepG2 2.2.15 cell line (a stably transfected cell line containing the HBV genome on a plasmid).

[0100] 3. Cytotoxicity Analysis

[0101] The cytotoxicity induced by the target compounds in HepG2 2.2.15 cell cultures was determined. Briefly, HepG22.2.15 cells were grown to confluence in 96-well culture plates and then treated with the test compounds (in 0.2 mL of culture medium / well). Untreated control cultures were kept on each 96-well plate. Cytotoxicity was determined by measuring the MTT (3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide) dye uptake, which was determined based on the absorbance at 490 nm relative to untreated cells 24 h after 72 h of treatment. Inhibitory effect on HBV-DNA replication, in order to detect extracellular replication intermediates, HepG2 2.2.15 cells were plated at 2*10 per well. 5 Cells were seeded at a density of 100 cells / mL in 24-well culture plates and allowed to recover for 24 h. The culture medium was changed every 2 days. After 6 days, intracellular HBV DNA was extracted using previously reported methods and quantified by real-time PCR.

[0102] The study compared the performance differences of the Word2vec model before and after combining with the Ising model.

[0103] 4. Dataset Selection

[0104] In order to reduce the computational dimension required for training and generate a unique vector space with the chemical structure distribution of specific target compounds to accurately represent the spatial distribution relationship of chemical structures, all compounds targeting HBV and HepG2 were screened, and compound structure word embedding models for HBV and HepG2 were generated respectively. From Table 1, a total of 6,705 experimental activity values ​​were obtained for the screening of HBV as the target, of which 4,411 were half-inhibitory concentration values. A total of 182,550 experimental results were obtained for HepG2 as the target, of which 2,270 were half-toxic concentrations.

[0105] Table 1 ChEMBL database screening results

[0106] Target Name Standard Type Target Organism Assays Number Hepatitis B virus ALL Hepatitis B virus 6,705 Hepatitis B virus EC50 / IC50 Hepatitis B virus 4,411 HepG2 ALL Homo sapiens 182,550 HepG2 CC50 Homo sapiens 2,270

[0107] Table 2 Performance comparison of different segmentation lengths in the SMILES to SC method

[0108]

[0109]

[0110] Considering the impact of split char length on the embedding effect in the SMILES to SC method, we use different lengths for Embedding Generation. When the split char length is 1 and 2, the number of tokens generated is less than 100, and the performance of Word2vec cannot be brought into play. When the split char length is 5, a large number of SMILES cannot generate enough tokens. Therefore, we chose 3 and 4 for comparative experiments. As shown in Table 2, when the length of the token extracted characters is 4, better performance can be obtained than when the length of the extracted characters is 3. Therefore, in the subsequent SMILES to SC method, the extraction length defaults to 4.

[0111] Table 3 Comparison of classification effects of ECFP models with different sampling radii (HBV IC50)

[0112]

[0113]

[0114] Regarding the extraction of tokens for the ECFP model, we took HBV as an example to compare the model classification effects under different ECFP sampling radii. As shown in Table 3, in the Word2vec and Ising-Word2vec models, when the sampling radius is 1, there is no significant disadvantage compared with the larger sampling radius (Radii=2 and 3). Considering that a smaller number of tokens are generated when the sampling radius is low (when Radii is 1, 2, 3, N-Token is 1,474, 7,187, 17,668 respectively), it avoids the generation of overly sparse matrices and greatly increases the computational requirements. In subsequent experiments, the default sampling radius is 1 (Radii=1).

[0115] To verify whether the anti-HBV compound has potential research prospects, it is necessary to verify its ability to inhibit HBV (measured by IC 50 present) and hepatotoxicity (CC 50 Rendering).

[0116] The embedding model was validated in 4,411 HBV inhibition tests and 2,270 HepG2 toxicity tests. Eight machine learning models were used to build compound classification models. In the early biological data prediction work of compounds targeting HBV and HepG2, we used a threshold to set the HBV half-effective inhibitory concentration (IC 50 / EC 50 ) was set at 1 μM, and the half-toxic concentration of HepG2 (CC 50) was set to 100 μM to convert the activity data of the compound against the target into binary data for prediction. All experimental conclusions with inconsistent units in the data set were converted to a unified unit in μM. Ten-fold cross validation was used to obtain model evaluation parameters.

[0117] Obtain the latest version of the drug activity database from ChEMBL. Filter fixed targets based on fields such as "Target Name" and "Target Organism". Considering the inhibition rate of HBV IC 50 and EC 50 have the same effect, so in order to retain as many samples as possible, we retain EC 50 and IC 50 The data is used for training and testing the classification model.

[0118] Table 4 SMILES to ECFP (Radii=1)

[0119]

[0120]

[0121] Table 5 SMILES to SC

[0122]

[0123] Tables 4 and 5 are the comparisons of the classification performance of the Word2vec model and the Ising-Word2vec model under two token extraction methods. Table 4 uses the SMILES segmentation to generate tokens, and Table 5 uses the Morgan algorithm to generate tokens. The conclusion proves that Word2vec combined with the Ising model has better classification results in molecular activity prediction than the native Word2vec, and RandonForest is the best classifier. Comparing the two different token extraction methods, the SMILES segmentation (Accuracy: 0.841Precision: 0.843Recall: 0.881F1: 0.853AUC: 0.92) has better classification results than the Embedding (ECFP) generated by the Morgan algorithm (Accuracy: 0.837Precision: 0.835Recall: 0.878F1: 0.845AUC: 0.90), which may be because the SMILES segmentation method has a better representation of cyclic and branched chemical structures than the Morgan algorithm.

[0124] HepG2 toxicity (CC50)

[0125] Table 6 SMILES to ECFP

[0126]

[0127] Table 7 SMILES to SC

[0128]

[0129]

[0130] Similarly, we used the same method to make a classification model and conclusion statistics for HepG2 toxicity. In Tables 6 and 7, we can see that whether it is directly using SMILES to SC or Smile to ECFP for Token extraction, the Word2vec model with Ising has a slight improvement in performance compared to directly using Word2vec (SMILES to SC: Accuracy = 0.902, Precision = 0.908, Recall = 0.972, F1 = 0.935, AUC = 0.91; SMILES to ECFP: Accuracy = 0.912, Precision = 0.974, Recall = 0.973, F1 = 0.943, AUC = 0.94). At the same time, the AUC value in the ROC curve is also excellent, which means that by adding the Word2vec word embedding model of the Ising model, the SMILES encoding of the compound molecule can be accurately represented and learned. From this, we can see that the Ising-Word2vec model can be used to train and accurately predict parameter indicators such as drug-likeness of multiple compounds. This will help pharmaceutical personnel avoid compounds with no medicinal potential in advance during drug screening, reduce drug screening investment and improve drug screening efficiency.

[0131] Drug experimental validation

[0132] In order to verify the reliability of the model, we used in vitro cell experiments to test the compound inhibition rate prediction model for HBV and the compound-specific prediction model for HepG2. During the entire experimental process, we tested the drug activity data of 56 compounds, including the HBV half-inhibition rate concentration (IC 50 and EC 50 ) and HepG2 half toxic concentration (CC 50 ).

[0133] After unifying the units of the data, the SMILES corresponding to the compound is input into the model to obtain the binary classification result of the model for the compound for the specified threshold. The post-prediction confusion matrix is ​​established, the accuracy, precision, F1 value and recall rate are calculated, and the ROC curve is drawn to calculate the area under the curve AUC value.

[0134] In actual compound activity tests, the criteria for judging the activity of a compound are not fixed values. It is necessary to form a threshold change range that can respond in real time based on the overall activity distribution of the current set of tested compounds. 50 The threshold range from 1 to 20 μM and HepG2 CC 50 The threshold value ranges from 10 to 100 μM.

[0135] In the compound IC 50 and CC 50 A relative threshold ratio is established between them, so that IC 50 and CC 50 The ability to maintain a large concentration difference to maintain a reasonable activity-toxicity ratio depends entirely on the drug testing personnel's screening needs for the drug. In this experiment, we are more concerned about the drug's ability to inhibit HBV, so IC 50 The threshold was set relatively low; in contrast, the toxicity threshold for HepG2 was set more leniently at 30 μM.

[0136] Table 8 HBV drug experimental compound model prediction verification

[0137]

[0138]

[0139] Table 9 HepG2 drug experimental compound model prediction verification

[0140]

[0141] As shown in Table 8 and Table 9, they are HBV drug compound IC 50 At the threshold of 1 μM and HepG2 drug compound CC 50 The prediction model was validated at a threshold of 30 μM. It can be seen that in the actual in vitro anti-HBV experiment, the model can still obtain a relatively high prediction accuracy (IC 50 The highest accuracy was 0.732 obtained using the XGBoost classification model, with an AUC value of 0.962; CC 50 The highest accuracy was 0.893 obtained using the SVM classification model, with an AUC value of 0.943).

[0142] This application also provides an anti-HBV small molecule drug prediction model based on representation learning, which is constructed using the above construction method and packaged into a tool after tuning and published on the Internet. Please refer to Figure 2 This web tool supports the prediction of the inhibitory ability and liver toxicity of compounds against HBV through compound SMILES coding, and determines whether they have the potential for further development of anti-HBV drugs.

[0143] In summary, the anti-HBV small molecule drug prediction model based on representation learning provided in this application shows better performance than traditional Word2vec on different target data sets. At the same time, in the downstream tasks after training the vector, it has a deterministic advantage in predicting the HBV inhibition rate and liver cell toxicity of the compound, and has a good ability to screen potential anti-HBV drugs. In addition, in the process of generating Tokens, a variety of methods were compared, and the Token extraction parameters that can enable the model to achieve optimal performance were tried out. Not only did the model show its superior performance in compound representation learning on public data sets, but the accurate prediction ability of the classification model based on the Ising-Word2vec pre-training results was also confirmed through cell experiments, demonstrating the application potential of a Word2vec combined with Ising gradient correction model in screening new potential drug-like compounds. This method can be widely used in the prediction of drug-likeness of other different target compounds, thereby simplifying the drug development process.

[0144] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the claims of the present invention.

Claims

1. A method for constructing an anti-HBV small molecule drug prediction model based on representation learning, characterized in that: The following steps are involved: S1: Construct a sparse matrix W based on the data structure of IsingModel using the compound structure token data extracted from SMILES ISM ; S2: W ISM After SLEP's large-scale sparse data processing, the correlation matrix W is obtained. WCM ; S3: Divide the input text data entity into n batches to process the data. We take the i-th batch as the batch. i , Batch i The entity in (Vm i1 , Vm i2 , ..., Vm ij ) T ; S4: Input Batch i Chemical structural entity (Vm i1 , Vm i2 , ..., Vm ij ) T , initialize the Embedding (Ve i1 ,Ve i2 , ..., Ve ij ) T ; S5: Batch i After pairing, the Skip-Gram relationship is generated to obtain the upper and lower structures Con(Vm ij ) and negative sampling word space NEG (Vm ij ), continue to calculate its predicted probability u∈NEG(Vm ij ); Based on Batch i Medium Ve ij Count contextCon(Vm ij ) and negative sampling word space NEG (Vm ij ) and calculate the loss function (6): S6: According to the loss function obtained by formula (6), a variety of optimization schemes can be adopted to obtain the gradient in order to optimize the parameters; S7: Based on the pairing relationship between entities, WCM The local relationship matrix W corresponding to the entity-entity pairing relationship is obtained by the index SCM And the relationship matrix W with the negative sampling pairing relationship NCM ; S8: W SCM , W NCM After combining with the gradient, redistributing the weights instead of averaging, back propagation updates Vm ij Word embedding And the auxiliary vector corresponding to word u Before the text is pre-trained, it needs to be segmented. For the chemical structure encoding SMILES, character segmentation (SC) and molecular fingerprint (ECFP) can be used: SMILES to SC 1. Perform sliding window shift cutting with a specific window size in the order of SMILES, with each shift being one character; 2. Remove the "(", ")" used to represent the branch at the branch, and use a single letter "R" to retain the branch information here; 3. For "Cl" and "Fe", the double characters represent single atoms and are replaced with single characters "L" and "E" that cannot be confused with other characters; 4. Each SMILES is concatenated in the order of the main chain and the branch chain to form a linear word list; SMILEStoECFP MorganFP is generated by generating,all substructures around all heavy atoms in a molecule within a defined radius,and assigning a unique identifier, which is usually hashed into a,vector with a fixed length; For each SMILES-encoded compound, SMILEStoECFP generates identifiers for all atoms of a fixed radius, and then arranges the identifiers of each atom into a "molecular sentence" in the order of the atoms encoded in the SMILES. For an atom with multiple identifiers, the atomic identifiers are arranged from small to large according to their identification radius.

2. The method for constructing an anti-HBV small molecule drug prediction model based on representation learning according to claim 1, characterized in that: In S2, Logistic Regression is used to solve W ISM The regularized logic problem is as follows: W WCM =(X1X2,...,X all )+(X1X2,...,X all ) T (2) The above formula (1) is to solve W WCM The calculation formula in the process, where w t is the weight of the t-th entity among all chemical structure entities, z t is our input W ISM The tth column, z kt is the entity in the kth row and tth column extracted from the sparse matrix, used to solve z kt Relevance to other entities, X k is z kt The corresponding correlation solution, λ is the l1 norm regularization parameter, and ρ is the squared 2 norm regularization parameter.

3. The method for constructing an anti-HBV small molecule drug prediction model based on representation learning according to claim 2, characterized in that: In S4, there are: (Ve i1 ,Ve i2 ,...,Ve ij ) T =(Vm i1 ,Vm i2 ,...,Vm ij ) T ·W Ve (4) The above formula (4) Vm ij In the above code, j stands for Batch. i The jth entity in , correspondingly, Ve ij is the embedding corresponding to the jth entity, W Ve It is the weight matrix in the vector generation process.

4. The method for constructing an anti-HBV small molecule drug prediction model based on representation learning according to claim 3, characterized in that: In S5: In the above formula (5), u∈NEG(Vm ij ), σ is the sigmoid activation function, The parameters corresponding to word u are to be trained.

5. The method for constructing an anti-HBV small molecule drug prediction model based on representation learning according to claim 4, characterized in that: The optimization scheme in S6 is: set up: According to formula (7), we can calculate the u (8) and Ve ij The gradient of (9):

6. The method for constructing an anti-HBV small molecule drug prediction model based on representation learning according to claim 5, characterized in that: In said S8:

7. A prediction model for anti-HBV small molecule drugs based on representation learning, characterized by: The method for constructing an anti-HBV small molecule drug prediction model based on representation learning as described in any one of claims 1-6 is used for construction.

Images