Compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitor

Abstract

The invention discloses a compound toxicity prediction method based on artificial intelligence and an HSD11b2 inhibitor, which comprises the following steps: constructing a training sample set, each training sample in the training sample set being a compound sample in which the HSD11b2 inhibitor is positive or negative; respectively extracting molecular descriptor features for each training sample to form feature vectors of the training samples; taking the feature vector of each training sample as input data, and respectively training a plurality of models of different types to obtain a plurality of toxicity prediction models; taking a compound of which the toxicity needs to be predicted as a test sample, and extracting molecular descriptor characteristics of the test sample to form a characteristic vector of the test sample; respectively inputting the feature vectors of the test sample into each toxicity prediction model, and judging whether the test sample has potential toxicity or not according to the output result of each toxicity prediction model. The method can rapidly and accurately predict the potential toxicity of the compound, and has the advantages of simple operation and high intelligent degree.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of compound toxicity prediction, in particular to a compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitor. BACKGROUND

[0002] Hydroxysteroid (11β) dehydrogenase 2 (HSD11b2) is a key enzyme involved in steroid hormone metabolism, mainly acting on the metabolic process of hormones such as cortisol, and affecting the body's response mechanism to physiological reactions such as stress and inflammation by regulating the active form of hormones.

[0003] Hydroxysteroid (11β) dehydrogenase 2 (hereinafter referred to as HSD11b2) inhibitor is a drug that can inhibit the activity of the enzyme. HSD11b2 is mainly distributed in the distal nephron of the kidney, responsible for inactivating glucocorticoids (such as cortisol), ensuring that aldosterone is the only agonist of the mineralocorticoid receptor. When HSD11b2 activity is inhibited, it can lead to excessive mineralocorticoids, and thus cause hypertension and other symptoms, so detecting HSD11b2 inhibitor activity is an important means of screening compound toxicity; the principle of detecting HSD11b2 inhibitor to determine the toxicity of the compound is to evaluate whether the compound inhibits the activity of HSD11b2 enzyme, leading to abnormal accumulation of cortisol and activation of the mineralocorticoid receptor (MR), thereby causing pathological effects similar to apparent mineralocorticoid excess syndrome (AME).

[0004] In the traditional technology, the detection method of HSD11b2 inhibitor mainly includes in vitro enzyme activity detection method, solid sterol metabolite ratio analysis method, etc., which has the problems of complicated operation, many interferences, complex sample processing, low sensitivity and accuracy, etc. SUMMARY

[0005] The first object of the present application is to overcome the shortcomings and deficiencies of the prior art, and to provide a compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitor, which can quickly and accurately predict the potential toxicity of the compound, and has the advantages of simple operation and high intelligence.

[0006] The second object of the present application is to provide a compound synergistic toxicity effect detection method.

[0007] The third object of the present application is to provide a storage medium.

[0008] The fourth object of the present application is to provide a computer program product.

[0009] The first object of the present application is achieved by the following technical scheme: a compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitor, comprising: constructing a training sample set, each training sample in the training sample set being a compound sample positive or negative for HSD11b2 inhibitor; extracting molecular descriptor features for each training sample to form a feature vector of the training sample; training a plurality of different types of models respectively by taking the feature vector of each training sample as input data, to obtain a plurality of toxicity prediction models; extracting molecular descriptor features of the test sample to form a feature vector of the test sample; inputting the feature vector of the test sample into each toxicity prediction model respectively, and determining whether the test sample is potentially toxic according to the output result of each toxicity prediction model.

[0010] Preferably, after inputting the feature vector of the test sample into each toxicity prediction model respectively, the method further comprises: determining whether the output results of the toxicity prediction models are the same, i.e., whether the predicted categories of the toxicity prediction models are all positive for HSD11b2 inhibitor or all negative for HSD11b2 inhibitor; if yes, determining whether the test sample is potentially toxic based on the predicted categories of the toxicity prediction models, if the predicted categories of the toxicity prediction models are all positive for HSD11b2 inhibitor, determining that the test sample is potentially toxic, and if the predicted categories of the toxicity prediction models are all negative for HSD11b2 inhibitor, determining that the test sample is not toxic; if no, discarding the output results of the toxicity prediction models.

[0011] Preferably, after inputting the feature vector of the test sample into each toxicity prediction model respectively, the method further comprises: calculating a final prediction probability P according to the prediction probabilities of the toxicity prediction models: P= ; wherein, wi is the weight of the i-th toxicity prediction model, Pi is the prediction probability of the i-th toxicity prediction model output for HSD11b2 inhibitor being positive when the test sample is tested; and N is the number of toxicity prediction models; determining whether the final prediction probability P satisfies: P P_high or P P_low; if no, i.e., P_low if yes, determining whether the test sample is potentially toxic according to the final prediction probability P, specifically: when P P_high, it is determined that the test sample HSD11b2 inhibitor is positive, and the potential toxicity of the test sample is determined, When P P_low, it is determined that the test sample HSD11b2 inhibitor is negative, and it is determined that the test sample has no toxicity. Wherein, P_high is a first set value, and P_low is a second set value.

[0012] Preferably, the process of constructing the training sample set comprises: Through a python script, compound data of human species type and HSD11b2 target point in the ChEMBL database are screened as training samples, and the data type of the compound is IC50.

[0013] Preferably, the feature vectors of the training samples and the test samples are constructed using the molecular descriptors defined in the rdkit.Chem.Descriptors module, and 206 molecular descriptors are screened for feature vector construction.

[0014] Preferably, the types of the model include two types, which are a first model and a second model. The feature vectors of each training sample are used as input data to train the first model to obtain a first toxicity prediction model. The feature vectors of each training sample are used as input data to train the second model to obtain a second toxicity prediction model. For the test sample, the feature vectors of the test sample are input into the first toxicity prediction model and the second toxicity prediction model respectively. According to the output results of the first toxicity prediction model and the second toxicity prediction model, it is determined whether the test sample has potential toxicity.

[0015] Further, the first model is a deep forest model, and the second model is a random forest model.

[0016] The second object of the application is achieved by the following technical scheme: a compound synergistic toxicity effect detection method, characterized in that it comprises: A first training sample set is constructed, and the first training sample set comprises compound samples with positive AR inhibitors and compound samples with negative AR inhibitors. For each training sample in the first training sample set, a molecular fingerprint feature is extracted to obtain a feature vector; the feature vectors of each training sample in the first training sample set are used as input data to train a plurality of different types of models to obtain a plurality of toxicity prediction models as a toxicity detection model A. Obtain the plurality of toxicity prediction models constructed in the toxicity prediction method of the first object of the present application as a toxicity detection model B; Take a compound for which toxicity needs to be predicted as a test sample, extract molecular fingerprint features and molecular descriptor features of the test sample respectively, and correspondingly obtain a first feature vector and a second feature vector respectively; Input the first feature vector of the test sample into the toxicity detection model A, and input the second feature vector of the test sample into the toxicity detection model B; Determine whether the test sample has a synergistic toxicity effect according to the output results of the toxicity detection model A and the toxicity detection model B.

[0017] The third object of the present application is achieved by the following technical solution: a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitors of the first object of the present application, and / or implement the compound synergistic toxicity effect detection method of the second object of the present application.

[0018] The fourth object of the present application is achieved by the following technical solution: a computer program product comprising a computer program, characterized in that the computer program is executed by a processor to implement the compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitors of the first object of the present application, and / or implement the compound synergistic toxicity effect detection method of the second object of the present application.

[0019] The present application has the following advantages and effects relative to the prior art: (1) The compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitors of the present application first constructs a training sample set, the training samples being compound samples with known HSD11b2 inhibitor properties; for each training sample, molecular descriptor features are extracted to form a feature vector; the feature vectors of the training samples are input as input data to train a plurality of different types of models to obtain a plurality of toxicity prediction models; in testing, the feature vector formed by the molecular descriptor features extracted from the test sample is input into each toxicity prediction model, and the output results of each toxicity prediction model are used to determine whether the compound is potentially toxic. The present application determines whether a compound is potentially toxic based on the molecular descriptor features of the compound and the output results of the trained plurality of machine learning models. Compared with the compound toxicity prediction method in the prior art, the present application combines an artificial intelligence model to quickly and accurately predict the potential toxicity of a compound.

[0020] (2) In the compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitors, after obtaining each toxicity prediction model through training, the feature vectors of the test samples are respectively input into each toxicity prediction model, when the output results of each toxicity prediction model are the same, then whether the HSD11b2 inhibitor of the test sample is positive or negative can be determined based on the prediction categories of each toxicity prediction model, so as to finally determine whether the corresponding compound of the test sample is potentially toxic, this toxicity prediction method guarantees the prediction classification accuracy, and is more simple and fast in operation, and has smaller calculation amount.

[0021] (3) In the compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitors, after obtaining each toxicity prediction model through training, the feature vectors of the test samples are respectively input into each toxicity prediction model, then the final prediction probability P is calculated through weighted average according to the prediction probabilities of each toxicity prediction model, and finally whether the HSD11b2 inhibitor is positive or negative is determined according to the relationship between P and the set value, so as to further obtain the toxicity prediction result; based on this, the differences of each toxicity prediction model are fully considered, and the toxicity prediction model with more performance advantages is fully utilized, so that the finally obtained toxicity prediction result can be more accurate.

[0022] (4) In the compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitors, a plurality of types of models are combined, for example, a first model constructed by a deep forest model, a second model constructed by a random forest model, to train each toxicity prediction model, and finally the final prediction result is determined based on the output of each toxicity prediction model, compared with a single type of model, different types of models used in the application can complement each other's defects, so as to further improve the prediction accuracy.

[0023] (5) In the compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitors, the molecular descriptor features of the compound are extracted to form the feature vector, compared with other features, the extraction speed of the molecular descriptor features can be faster, 206 molecular descriptors can complete single molecule calculation within seconds, and the molecular descriptor can be associated with the inhibition mechanism through hydrophobicity, charge and other descriptors, based on this, the prediction speed of the prediction model constructed by the application to the HSD11b2 inhibitor category can be higher.

[0024] (6) The compound synergistic toxicity effect detection method of the present application trains a plurality of toxicity prediction models as toxicity detection model A through AR inhibitor positive and negative compound training samples; at the same time, a plurality of toxicity prediction models constructed in the toxicity prediction method of the present application are obtained as toxicity detection model B; for a compound (test sample) that needs to be predicted for toxicity, molecular fingerprint features and molecular descriptor features are extracted to obtain first feature vectors and second feature vectors; finally, the first feature vectors and the second feature vectors of the test sample are input into the toxicity detection model A and the toxicity detection model B respectively, when the toxicity detection model A predicts that the test sample AR inhibitor is positive and the toxicity detection model B predicts that HSD11b2 is positive, it can be determined that the test sample corresponding compound has a dual inhibition effect, the compound interferes with the androgen receptor (AR) pathway and the key enzyme of androgen synthesis (HSD11b2) at the same time, which may cause synergistic toxicity effect, providing effective data support for the priority control of the compound in clinical. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is the toxicity prediction method flow chart of the compound of the present application.

[0026] Figure 2 is the test sample potential toxicity determination flow chart in the method of the present application.

[0027] Figure 3 is the test verification diagram of HSD11b2 inhibitor positive compound. DETAILED DESCRIPTION

[0028] The embodiments of the present application will be described below in conjunction with the drawings in the present application. It should be understood that the embodiments described below in conjunction with the drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions of the embodiments of the present application.

[0029] Example 1 The hydroxysteroid (11β) dehydrogenase 2 (hereinafter referred to as HSD11b2) inhibitor is a drug that can inhibit the activity of the enzyme. HSD11b2 is mainly distributed in the distal nephron of the kidney and is responsible for inactivating glucocorticoids (such as cortisol), ensuring that aldosterone is the only agonist of the mineralocorticoid receptor. When HSD11b2 activity is inhibited, it can lead to excessive mineralocorticoids, and thus cause hypertension and other symptoms, so detecting HSD11b2 inhibitor activity is an important means of screening compound toxicity. In the traditional technology, the detection method of HSD11b2 inhibitor mainly includes in vitro enzyme activity detection method, sterol metabolite ratio analysis method, etc., which has the problems of complicated operation, many interferences, complex sample processing, low sensitivity and accuracy, etc.

[0030] Based on the problems existing in the prior art, the embodiment provides a compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitors. The method determines whether a compound is potentially toxic by detecting the properties (positive or negative) of HSD11b2 inhibitors in the compound, as shown in Figure 1 The prediction method of the embodiment includes: S1, constructing a training sample set, each training sample in the training sample set being a compound sample with HSD11b2 inhibitors being positive or negative. The compound sample with HSD11b2 inhibitors being positive refers to a compound verified by experiment to have inhibitory activity on the HSD11b2 target point, and the compound sample with HSD11b2 inhibitors being negative refers to a compound verified by experiment to have no or weak inhibitory activity on the HSD11b2 target point.

[0031] In the embodiment, the python script is used to screen the compound data of ChEMBL database with the species type being human and the target point being HSD11b2 as the training sample, and the selected compound data type is IC50. The training sample set includes a plurality of first training samples and a plurality of second training samples; the first training sample is a training sample with HSD11b2 inhibitors being positive, and the second training sample is a training sample with HSD11b2 inhibitors being negative.

[0032] In the embodiment, after obtaining the compound data by the python script, the data is cleaned to delete some data with noise, defects, and abnormalities. Specifically, in the training sample set of the embodiment, the ratio of the training sample with HSD11b2 inhibitors being positive and the training sample with HSD11b2 inhibitors being negative can be controlled to be about 1:6, for example: including 54 training samples with HSD11b2 inhibitors being positive, 297 training samples with HSD11b2 inhibitors being negative, and part of the training samples are shown in Table 1.

[0033] Table 1

[0034] In the above table, active indicates that the HSD11b2 inhibitor in the compound is positive, and inactive indicates that the HSD11b2 inhibitor in the compound is negative.

[0035] S2, extracting molecular descriptor features for each training sample in the training sample set to form a feature vector of the training sample.

[0036] In this embodiment, the feature vectors of the training samples are constructed using the molecular descriptors defined in the rdkit.Chem.Descriptors module-desclist, from which 206 molecular descriptors suitable for the method of this embodiment are screened for feature vector construction.

[0037] Specifically, rdkit.Chem.Descriptors_deslist is a list of predefined molecular descriptor calculation functions in RDKit, containing 206 functions that can be directly used to calculate the numerical descriptors of physical and chemical properties, topological structure, electronic features, etc. In this embodiment, based on the SHAP value analysis of input feature importance, the top 10 important features are finally selected: fr_bicyclic, SMR_VSA10, PEOE_VSA5, NHOHCount, SMR_VSA4, PEOE_VSA1, VSA_Estate6, VSA_Estate4, VSA_Estate7 and Chi4n, which have a significant impact on the inhibitor effect of HSD11b2.

[0038] S3, the feature vectors of each training sample in the training sample set are taken as input data, and a plurality of different types of models are trained respectively to obtain a plurality of toxicity prediction models.

[0039] In this embodiment, the validation sample set can also be constructed by the method of step S1, and the sample quantity ratio in the training sample set and the validation sample set can be set to 8:2. In the above model training process, multiple cross-validation, for example, 5-fold cross-validation, can be performed to avoid model overfitting.

[0040] In this embodiment, different types of models refer to different artificial intelligence models, such as traditional machine learning models: linear models, tree-based models, support vector machines, clustering algorithms, etc., and deep learning models: feedforward neural network models, convolutional neural network models, recurrent neural network models, self-attention mechanism models (Transformer models), etc.

[0041] In this embodiment, two or more of the above different types of models can be selected, for example, two models are selected, which are the first model and the second model. In this embodiment, the feature vectors of each training sample are taken as input data to train the first model to obtain the first toxicity prediction model, and the feature vectors of each training sample are taken as input data to train the second model to obtain the second toxicity prediction model.

[0042] In this embodiment, the first model can be a deep forest model, and the second model can be a random deep forest model. The deep forest model is a kind of deep learning model, which automatically learns feature representation through cascading forest structure. The model is composed of multiple forest levels. Each level receives the output of the previous level and the original features. The input of the first level is the original features, and the input of each subsequent level is the original features plus the output of all forests in the previous level. In this embodiment, when the first toxicity prediction model constructed by the deep forest model is used for prediction, the average value of the probability vectors output by all forests in the last layer is taken, and then the category corresponding to the maximum probability is taken. Random forest is an ensemble learning method composed of multiple decision trees. The final decision is made by voting or averaging. When training the second model, n samples are randomly selected (with replacement) from the training sample set constructed in step S1 to form a training subset. This process is repeated to generate multiple training subsets. Each training subset is used to train a decision tree. The second toxicity prediction model trained by the random deep forest model makes a prediction by voting from each tree, and the final result is the category with the most votes.

[0043] S4, the compound to be predicted for toxicity is taken as a test sample, the molecular descriptor features of the test sample are extracted to form a feature vector of the test sample, and the specific steps are as shown in step S2.

[0044] S5, the feature vector of the test sample is input into each toxicity prediction model, and whether the test sample is potentially toxic is determined according to the output results of each toxicity prediction model. The output results of each toxicity prediction model include the prediction probability of HSD11b2 inhibitor being positive and the prediction probability of HSD11b2 inhibitor being negative.

[0045] In this embodiment, if the first toxicity prediction model and the second toxicity prediction model are trained in step S3, the feature vector of the test sample needs to be input into the first toxicity prediction model and the second toxicity prediction model respectively. The first output result is obtained by the first toxicity prediction model based on the input test sample feature, and the second output result is obtained by the second toxicity prediction model based on the input test sample feature. In step S5, whether the test sample is potentially toxic is determined according to the first output result and the second output result.

[0046] In this embodiment, after the feature vector of the test sample is input into each toxicity prediction model in step S5, whether the test sample is potentially toxic can be determined by the following steps: S51, whether the output results of each toxicity prediction model are the same, i.e., whether the predicted categories of each toxicity prediction model are all HSD11b2 inhibitor positive or all HSD11b2 inhibitor negative; If yes, go to step S52; If not, the output results of each toxicity prediction model are discarded, i.e., no judgment result is output.

[0047] S52, determining whether the test sample is potentially toxic based on the prediction categories of each toxicity prediction model, wherein if the prediction categories of each toxicity prediction model are all positive, it is determined that the test sample is potentially toxic, and if the prediction categories of each toxicity prediction model are all negative, it is determined that the test sample is not toxic.

[0048] In this embodiment, if two toxicity prediction models including the first toxicity prediction model and the second toxicity prediction model are included, the feature vector of the test sample is respectively input into the two toxicity prediction models, and if the prediction categories of the two prediction models are the same, it can be further determined whether the test sample is potentially toxic based on the prediction categories output by the two toxicity prediction models. Specifically, when the prediction categories of the two toxicity prediction models are both positive, it is determined that the test sample is potentially toxic, when the prediction categories of the two toxicity prediction models are both negative, it is determined that the test sample is not toxic, and if the prediction category of one of the toxicity prediction models is positive and the prediction category of the other toxicity prediction model is negative, the toxicity prediction of the test sample cannot be performed, and the result is discarded.

[0049] In this embodiment, after the feature vector of the test sample is respectively input into each toxicity prediction model in step S5, as shown in FIG. 5B, the test sample can also be determined whether it is potentially toxic through the following steps: Figure 2 S5a, calculating a final prediction probability P according to the prediction probabilities of each toxicity prediction model: P = ∑wiPi ; wherein, wi is the weight of the i th toxicity prediction model, Pi is the prediction probability of the i th toxicity prediction model that the HSD11b2 inhibitor is positive when the test sample is tested; and N is the number of toxicity prediction models. In this embodiment, the weight of each toxicity prediction model can be determined according to the performance of the model. Specifically, the weight of the model can be determined according to the accuracy or AUC (area Under the ROC Curve, indicating the area under the ROC curve) of the model in verifying the verification sample set. The model with higher accuracy or AUC can be given a larger weight, i.e., the model with higher performance is given a higher weight. In addition, the weight of the model can also be adjusted dynamically according to the prediction difference, so that the accuracy of the adjusted model in verifying the verification sample set is higher.

[0050] ​S5b, judging whether the final prediction probability P satisfies: P P_high or P P_low; If not, i.e. P_low < P < P_high, the output result of each toxicity prediction model is eliminated.

[0051] If yes, it is determined whether the test sample is potentially toxic according to the final prediction probability P, specifically: When P P_high, it is determined that the test sample HSD11b2 inhibitor is positive, and it is determined that the test sample is potentially toxic; When P P_low, it is determined that the test sample HSD11b2 inhibitor is negative, and it is determined that the test sample is not toxic; Wherein, P_high is a first set value, and P_low is a second set value.

[0052] In this embodiment, if the first toxicity prediction model constructed by the deep forest model and the second toxicity prediction model constructed by the random forest model are included, the prediction probabilities of the first toxicity prediction model and the second toxicity prediction model are weighted and averaged to calculate the final prediction probability. For example, the first toxicity prediction model and the second toxicity prediction model are respectively assigned weights W1 and W2. If the prediction probability of the first toxicity prediction model predicting that the HSD11b2 inhibitor is positive is P1, and the prediction probability of the second toxicity prediction model predicting that the HSD11b2 inhibitor is positive is P2, then the final prediction probability P = W1*P1 + W2*P2, W1 + W2 = 1. When the calculated P is greater than or equal to the first preset value P_high, it is determined that the test sample HSD11b2 inhibitor is positive, and it is determined that the test sample is potentially toxic. When the calculated P is less than or equal to the second preset value P_low, it is determined that the test sample HSD11b2 inhibitor is negative, and it is determined that the test sample is not toxic. When the calculated P is P_low < P < P_high, the result is eliminated. In this embodiment, the above-mentioned weights W1 and W2 can be set according to the performance of the first toxicity prediction model and the second toxicity prediction model. If the performance of the first toxicity prediction model is higher than that of the second toxicity prediction model, W1 can be set higher than W2. For example, W1 can be set to 0.6, and W2 can be set to 0.4. The specific values can be adjusted according to actual use.

[0053] In this embodiment, the first preset value P_high can be set to 0.7, and P_low can be set to 0.3. Of course, it can also be adjusted according to the actual test situation. Based on this, when the final prediction probability is greater than or equal to 0.7, it is determined that the toxicity prediction models reach a consensus, and the HSD11b2 inhibitor corresponding to the test sample is positive, and the test sample has potential toxicity. When the final prediction probability is less than or equal to 0.3, it is determined that the toxicity prediction models reach a consensus, and the HSD11b2 inhibitor corresponding to the test sample is negative, and the test sample has no toxicity. When 0.3 < P < 0.7, it is determined that the toxicity prediction models cannot reach a consensus.

[0054] In summary: In this embodiment, the molecular descriptor features extracted from the compound have a significant effect on the inhibition of HSD11b2. Compared with other features, the molecular descriptor features can be extracted faster. The 206 molecular descriptors can complete single molecule calculation within seconds, and the molecular descriptor can be associated with the inhibition mechanism through hydrophobicity, charge and other descriptors. Based on this, the prediction model constructed in this embodiment can have higher prediction speed and prediction accuracy for HSD11b2 inhibitor categories. Through 5-fold cross-validation test, the first toxicity prediction model trained by the deep forest model in this embodiment has an accuracy ACC of 0.972, an area AUC under the ROC curve of 0.975, a specificity of 1.000, a sensitivity of 0.818, and an F1 value (F1-score) of 0.900. Through 5-fold cross-validation test, the second toxicity prediction model trained by the random forest model in this embodiment has an accuracy ACC of 0.972, an area AUC under the ROC curve of 0.975, a specificity of 1, a sensitivity of 0.818, and an F1 value (F1-score) of 0.900. The overall accuracy of compound toxicity prediction has been significantly improved.

[0055] In this embodiment, the molecular descriptor extracted from the compound has a significant effect on the inhibition of HSD11b2. Compared with other features, the molecular descriptor features can be extracted faster. The 206 molecular descriptors can complete single molecule calculation within seconds, and the molecular descriptor can be associated with the inhibition mechanism through hydrophobicity, charge and other descriptors. Based on this, the prediction model constructed in this embodiment can have higher prediction speed and prediction accuracy for HSD11b2 inhibitor categories. Through 5-fold cross-validation test, the first toxicity prediction model trained by the deep forest model in this embodiment has an accuracy ACC of 0.972, an area AUC under the ROC curve of 0.975, a specificity of 1.000, a sensitivity of 0.818, and an F1 value (F1-score) of 0.900. Through 5-fold cross-validation test, the second toxicity prediction model trained by the random forest model in this embodiment has an accuracy ACC of 0.972, an area AUC under the ROC curve of 0.975, a specificity of 1, a sensitivity of 0.818, and an F1 value (F1-score) of 0.900. The overall accuracy of compound toxicity prediction has been significantly improved.

[0056] In this embodiment, after obtaining each toxicity prediction model, the final prediction probability P can be calculated by weighted average method according to the prediction probability of each toxicity prediction model, and finally the relationship between P and the set value is used to determine whether the HSD11b2 inhibitor is positive or negative to further obtain the toxicity prediction result; this embodiment fully considers the difference of each toxicity prediction model, and fully gives play to the toxicity prediction model with better performance, so that the final toxicity prediction result can be more accurate.

[0057] In this embodiment, the HSD11b2 inhibitor is positive. The compound is verified by experiment: for example, the HSD11b2 inhibitor is positive. The compound six (1H, 1H, 7H-perfluoroheptyloxy) phosphazene (CAS: 3830-74-8) is used as an example, and the zebrafish embryo exposure test is verified to prove that it has a potential HSD11b2 inhibition effect, as shown in Figure 3 In this embodiment, the HSD11b2 inhibitor is positive. The compound is verified by experiment: for example, the HSD11b2 inhibitor is positive. The compound six (1H, 1H, 7H-perfluoroheptyloxy) phosphazene (CAS: 3830-74-8) is used as an example, and the zebrafish embryo exposure test is verified to prove that it has a potential HSD11b2 inhibition effect, as shown in

[0058] Those skilled in the art can understand that all or part of the steps in the method of this embodiment can be completed by programs instructing related hardware, and the corresponding programs can be stored in a computer readable storage medium. It should be noted that although the method operations of this embodiment are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all of the shown operations must be performed to achieve the desired result. On the contrary, the steps can be changed in the order of execution, some steps can be executed simultaneously, additionally or alternatively, some steps can be omitted, a plurality of steps can be combined into one step for execution, and / or one step can be divided into a plurality of steps for execution.

[0059] Embodiment 2 The embodiment discloses a compound synergistic toxicity effect detection method, which is realized based on the method of embodiment 1 and can realize prediction of the double inhibition effect (AR inhibitor and HSD11b2) of the compound, so as to detect whether the compound can cause synergistic toxicity effect, and specifically includes: T1, constructing a first training sample set, the first training sample set including compound samples with positive AR inhibitors and compound samples with negative AR inhibitors.

[0060] The compound sample positive for AR inhibitor refers to a compound verified by experiment to have inhibitory activity on the AR target point, and the compound sample negative for AR inhibitor refers to a compound verified by experiment to have no or weak inhibitory activity on the AR target point. In this embodiment, by using a python script, compound data of human species and AR target point in the ChEMBL database are screened as training samples, and the selected compound data type is IC50. The training sample set includes first training samples and second training samples; the first training samples are training samples positive for AR inhibitor, and the second training samples are training samples negative for AR inhibitor.

[0061] T2, for each training sample in the first training sample set, a molecular fingerprint feature is extracted respectively to obtain a feature vector; the feature vectors of the training samples in the first training sample set are taken as input data, and a plurality of different types of models are trained respectively to obtain a plurality of toxicity prediction models as the toxicity detection model A.

[0062] In this embodiment, the molecular fingerprint extracted for the training sample includes the Morgan fingerprint feature MF and the structure statistical fingerprint feature CMF. The Morgan fingerprint feature MF belongs to the 2D connectivity fingerprint, and its essence is to convert the substructure into a fixed length vector by a hash algorithm based on the two-dimensional topological structure (atomic connectivity) of the molecule; the structure statistical fingerprint feature CMF belongs to the 2D fingerprint, which is a fingerprint (vector) that uses integer counting to represent the occurrence frequency of substructures / features, rather than only using 0 / 1 to represent the presence or absence.

[0063] For each training sample, after the Morgan fingerprint feature MF and the structural statistical fingerprint feature CMF are extracted, they can be combined into a multi-dimensional vector to form a feature vector. Specifically, the extracted MF and CMF features can be directly spliced. Before splicing, the high-dimensional feature, i.e., the CMF feature, needs to be dimensionally reduced, and then spliced with the low-dimensional feature, i.e., the MF feature. The spliced vector is used as the feature vector. Alternatively, the MF and CMF features can be directly spliced, and then dimensionally reduced to obtain the feature vector. The kernel functions of MF and CMF can also be designed respectively, and then fused to obtain a kernel matrix as the feature vector. In addition, other feature fusion methods can also be used. In this embodiment, the input feature importance is analyzed based on the SHAP value, and the top 10 important features include Col_2210, Col_790, Col_2118, Col_2298, Col_1274, Col_1384, Col_674, Col_2160, Col_2218 and Col_2105, wherein the Col_ prefix represents the sparse matrix index of the Morgan fingerprint, corresponding to a specific molecular substructure. In this embodiment, for each compound training sample, the molecular fingerprint feature is calculated from the SMILES code using RDKit through a python script, including the Morgan fingerprint feature and the structural statistical fingerprint feature.

[0064] In this embodiment, as shown in Embodiment 1, the different types of models in step T2 refer to different artificial intelligence models, such as traditional machine learning models: linear models, tree-based models, support vector machines, clustering algorithms, etc., and deep learning models: feedforward neural network models, convolutional neural network models, recurrent neural network models, self-attention mechanism models (Transformer models), etc. For example, when two models are selected, a first model and a second model, the first model can be a deep forest model, and the second model can be a random deep forest model.

[0065] T3, obtaining the plurality of toxicity prediction models constructed in step S3 of the toxicity prediction method of Embodiment 1 as a toxicity detection model B; T4, taking a compound whose toxicity needs to be predicted as a test sample, extracting the molecular fingerprint feature and the molecular descriptor feature of the test sample to obtain a first feature vector and a second feature vector, respectively. The extraction of the first feature vector of the test sample can be as shown in step T2 above, and the extraction of the second feature vector of the test sample can be as shown in step S2 of Embodiment 1.

[0066] T5, inputting the first feature vector of the test sample into the toxicity detection model A and inputting the second feature vector of the test sample into the toxicity detection model B; T6, determine whether the test sample has a synergistic toxicity effect according to the output results of the toxicity detection model A and the toxicity detection model B. Specifically as follows: T61, for the same test sample, input the first feature vector of the test sample into each toxicity prediction model of the toxicity detection model A, and determine whether the test sample is potentially toxic according to the output results of each toxicity prediction model. Specifically, whether the test sample is potentially toxic can be determined by any of the following ways: After obtaining the output results of each toxicity prediction model, it is determined whether the prediction categories output by each toxicity prediction model are the same. If they are the same, it means that each toxicity prediction model has reached a consensus, and the above prediction category is taken as the final prediction result. For example, if the prediction categories of each toxicity prediction model are all AR inhibitor positive, then AR inhibitor positive is taken as the final prediction result, and the test sample is determined to be potentially toxic.

[0067] After obtaining the prediction probability of AR inhibitor positive output by each toxicity prediction model, the final prediction probability Q is calculated according to the prediction probability output by each toxicity prediction model: Q= ; Wherein: The above is the weight of the jth toxicity prediction model in the toxicity detection model A, and is the normalized weight, which satisfies The model performance can be set according to the corresponding model performance, wherein the higher the model performance, the greater the weight can be set. M is the number of toxicity prediction models in the toxicity detection model A. is the prediction probability of AR inhibitor positive output by the jth toxicity prediction model in the toxicity detection model A when testing the test sample.

[0068] In this embodiment, when satisfies: Q Q_high, it is determined that the test sample is AR inhibitor positive and potentially toxic; when satisfies: Q Q_low, it is determined that the test sample is AR inhibitor negative and non-toxic. Q_high is a first preset value, and Q_low is a second preset value, which can be adjusted according to the actual test situation. For example, Q_high is 0.8 and Q_low is 0.2.

[0069] T62, after inputting the first feature vector of the test sample into each toxicity prediction model of the toxicity detection model B, it is determined whether the test sample is potentially toxic according to the output results of each toxicity prediction model. The specific determination process is shown in steps S51-S52 or steps S5a-S5b in embodiment 1, which will not be repeated here.

[0070] T63, when the test sample is determined to be positive for AR inhibitor based on toxicity detection model A and the test sample is determined to be positive for HSD11b2 inhibitor based on toxicity detection model B, it is determined that the test sample has a synergistic toxicity effect.

[0071] The toxicity prediction method of the embodiment is suitable for different types of compounds and can be used as a general compound toxicity screening tool. For example, it can be applied to organic compounds such as fluorocarbons. The method can be used to detect whether the fluorocarbon has a dual inhibition effect on AR and HSD11b2. Since HSD11b2 plays an important role in catalyzing androgen biosynthesis, these fluorocarbons with dual inhibition effect not only affect the AR receptor pathway, but also further inhibit androgen biosynthesis.

[0072] In summary, the embodiment trains a plurality of toxicity prediction models as toxicity detection model A by using the compound training samples of AR inhibitor positive and negative. Meanwhile, the plurality of toxicity prediction models constructed in step S3 of embodiment 1 are obtained as toxicity detection model B. For a test sample that needs to be predicted for toxicity, the molecular fingerprint features and molecular descriptor features are extracted to obtain the first feature vector and the second feature vector. Finally, the first feature vector and the second feature vector of the test sample are input into the toxicity detection model A and the toxicity detection model B, respectively. When the toxicity detection model A predicts that the test sample is positive for AR inhibitor and the toxicity detection model B predicts that HSD11b2 is positive, it is determined that the test sample has a dual inhibition effect. The compound interferes with the androgen receptor (AR) pathway and the key enzyme (HSD11b2) of androgen synthesis at the same time, which may cause a synergistic toxicity effect. This provides effective data support for the priority control of the compound in clinical practice.

[0073] Those skilled in the art can understand that all or part of the steps in the method of the embodiment can be completed by programs instructing related hardware. The corresponding programs can be stored in a computer readable storage medium. It should be noted that although the method operations of the embodiment are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all of the shown operations must be performed to achieve the desired result. On the contrary, the steps can be changed in the order of execution, some steps can be executed simultaneously, additionally or alternatively, some steps can be omitted, a plurality of steps can be combined into one step for execution, and / or one step can be divided into a plurality of steps for execution; for example, step T3 can be executed before or together with step T1, steps T61 and T62 can be executed simultaneously, or step T62 can be executed before T61.

[0074] Embodiment 3 The embodiment discloses a storage medium, which stores a program. When the program is executed by a processor, a compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitors described in embodiment 1 and / or a compound synergistic toxicity effect detection method described in embodiment 2 are implemented. Wherein: The compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitors described in embodiment 1 comprises: A training sample set is constructed, each training sample in the training sample set is a compound sample with HSD11b2 inhibitor being positive or negative; Molecular descriptor features are extracted for each training sample to form a feature vector of the training sample; The feature vectors of each training sample are used as input data to train multiple different types of models to obtain multiple toxicity prediction models; A compound to be predicted for toxicity is used as a test sample, and molecular descriptor features of the test sample are extracted to form a feature vector of the test sample; The feature vector of the test sample is input into each toxicity prediction model, and whether the test sample is potentially toxic is determined according to the output result of each toxicity prediction model.

[0075] The specific implementation process of each step can be as described in embodiment 1.

[0076] The compound synergistic toxicity effect detection method described in embodiment 2 comprises: A first training sample set is constructed, which includes compound samples with AR inhibitors being positive and compound samples with AR inhibitors being negative; For each training sample in the first training sample set, molecular fingerprint features are extracted to obtain a feature vector; the feature vectors of each training sample in the first training sample set are used as input data to train multiple different types of models to obtain multiple toxicity prediction models as toxicity detection model A; The multiple toxicity prediction models constructed in the toxicity prediction method of embodiment 1 are obtained as toxicity detection model B; A compound to be predicted for toxicity is used as a test sample, and molecular fingerprint features and molecular descriptor features of the test sample are extracted to obtain first and second feature vectors respectively; The first feature vector of the test sample is input into the toxicity detection model A, and the second feature vector of the test sample is input into the toxicity detection model B; According to the output results of the toxicity detection model A and the toxicity detection model B, it is determined whether the test sample has synergistic toxicity effect.

[0077] The specific implementation process of each step can be as described in embodiment 2.

[0078] In the embodiment, the storage medium can be a disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), a U disk, a mobile hard disk, and the like.

[0079] Embodiment 4 The embodiment discloses a computer program product comprising a computer program which, when executed by a processor, implements the compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitor described in embodiment 1, and / or the compound synergistic toxicity effect detection method described in embodiment 2. Wherein: The compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitor described in embodiment 1 comprises: A training sample set is constructed, each training sample in the training sample set being a compound sample with HSD11b2 inhibitor being positive or negative; Molecular descriptor features are extracted for each training sample to form a feature vector of the training sample; The feature vectors of each training sample are taken as input data, and a plurality of different types of models are trained respectively to obtain a plurality of toxicity prediction models; A compound to be predicted for toxicity is taken as a test sample, and molecular descriptor features of the test sample are extracted to form a feature vector of the test sample; The feature vector of the test sample is input into each toxicity prediction model, and whether the test sample is potentially toxic is determined according to the output result of each toxicity prediction model.

[0080] The specific implementation process of each step can be as described in embodiment 1.

[0081] The compound synergistic toxicity effect detection method described in embodiment 2 comprises: A first training sample set is constructed, and the first training sample set comprises compound samples with AR inhibitor being positive and compound samples with AR inhibitor being negative; For each training sample in the first training sample set, a molecular fingerprint feature is extracted to obtain a feature vector; the feature vectors of each training sample in the first training sample set are taken as input data, and a plurality of different types of models are trained respectively to obtain a plurality of toxicity prediction models as toxicity detection model A; The plurality of toxicity prediction models constructed in the toxicity prediction method of embodiment 1 are obtained as toxicity detection model B; A compound to be predicted for toxicity is taken as a test sample, and molecular fingerprint features and molecular descriptor features of the test sample are extracted respectively to obtain first and second feature vectors respectively; inputting the first feature vector of the test sample into the toxicity detection model A and inputting the second feature vector of the test sample into the toxicity detection model B; determining whether the test sample has a synergistic toxicity effect according to the output results of the toxicity detection model A and the toxicity detection model B.

[0082] The specific implementation process of each step can be described in Embodiment 2.

[0083] The computer program product includes one or more computer instructions, which, when loaded and executed on a computer, generate all or part of the processes or functions described in the embodiments of the present application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted by the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server, data center, etc. that contains one or more available media sets. The available media can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a digital versatile disc (DVD)), or a semiconductor medium (for example, a solid state disk (SSD)) and the like.

[0084] Embodiment 5 The embodiment discloses an electronic device, including a memory, a processor and a computer program stored in the memory, the processor executes the above computer program to realize the compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitor described in embodiment 1, and / or the compound synergistic toxicity effect detection method described in embodiment 2. Wherein: The compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitor described in embodiment 1 comprises: A training sample set is constructed, each training sample in the training sample set is a compound sample with HSD11b2 inhibitor positive or negative; Molecular descriptor features are extracted for each training sample to form a feature vector of the training sample; The feature vectors of each training sample are used as input data to train multiple different types of models to obtain multiple toxicity prediction models; The compound to be predicted for toxicity is taken as a test sample, molecular descriptor features of the test sample are extracted, and a feature vector of the test sample is constructed; The feature vector of the test sample is input into each toxicity prediction model, and it is determined whether the test sample is potentially toxic according to the output results of the toxicity prediction models.

[0085] The specific implementation process of each step can be as described in Embodiment 1.

[0086] The compound synergistic toxicity effect detection method described in Embodiment 2 is implemented, including: A first training sample set is constructed, and the first training sample set includes compound samples positive for AR inhibitors and compound samples negative for AR inhibitors; For each training sample in the first training sample set, a molecular fingerprint feature is extracted to obtain a feature vector; the feature vectors of the training samples in the first training sample set are taken as input data, and a plurality of different types of models are trained to obtain a plurality of toxicity prediction models as a toxicity detection model A; The plurality of toxicity prediction models constructed in the toxicity prediction method of Embodiment 1 are obtained as a toxicity detection model B; The compound to be predicted for toxicity is taken as a test sample, and molecular fingerprint features and molecular descriptor features of the test sample are extracted to correspondingly obtain a first feature vector and a second feature vector; The first feature vector of the test sample is input into the toxicity detection model A, and the second feature vector of the test sample is input into the toxicity detection model B; It is determined whether the test sample has a synergistic toxicity effect according to the output results of the toxicity detection model A and the toxicity detection model B.

[0087] The specific implementation process of each step can be as described in Embodiment 2.

[0088] In this embodiment, the electronic device can be a terminal device such as a mobile phone, a computer, or a tablet computer.

[0089] In this embodiment, the electronic device includes a processor, a memory, a bus, and a communication interface, and the processor, the communication interface, and the memory are connected through the bus; the processor is configured to execute executable modules stored in the memory, such as a computer program.

[0090] The above is only an optional implementation of some implementation scenarios of the present application. It should be noted that, for those skilled in the art, other similar implementation means based on the technical concept of the present application can be used without departing from the technical concept of the present application, and such implementation also falls within the protection scope of the embodiments of the present application.

Claims

1. A method for predicting the toxicity of compounds based on artificial intelligence and HSD11b2 inhibitors, characterized in that, Comprising: Construct a training sample set, where each training sample in the training sample set is a compound sample that is positive or negative for HSD11b2 inhibitor; Extract molecular descriptor features for each training sample respectively to form a feature vector of the training sample; Use the feature vectors of each training sample as input data to train multiple different types of models respectively to obtain multiple toxicity prediction models; Use the compound whose toxicity needs to be predicted as a test sample, extract the molecular descriptor features of the test sample to form a feature vector of the test sample; Input the feature vector of the test sample into each toxicity prediction model respectively, and determine whether the test sample has potential toxicity according to the output results of each toxicity prediction model.

2. The method for predicting compound toxicity based on artificial intelligence and HSD11b2 inhibitors according to claim 1, characterized in that, After inputting the feature vector of the test sample into each toxicity prediction model respectively, it further includes: Judge whether the output results of each toxicity prediction model are the same, that is, whether the prediction categories of each toxicity prediction model are all that the HSD11b2 inhibitor is positive or the HSD11b2 inhibitor is negative; If so, judge whether the test sample has potential toxicity based on the prediction categories of each toxicity prediction model. If the prediction categories of each toxicity prediction model are all that the HSD11b2 inhibitor is positive, it is determined that the test sample has potential toxicity. If the prediction categories of each toxicity prediction model are all that the HSD11b2 inhibitor is negative, it is determined that the test sample has no toxicity; If not,剔除 the output results of each toxicity prediction model.

3. The method for predicting compound toxicity based on artificial intelligence and HSD11b2 inhibitors according to claim 1, characterized in that, After inputting the feature vector of the test sample into each toxicity prediction model respectively, it further includes: Calculate the final prediction probability P according to the prediction probabilities of each toxicity prediction model: P= ； in, Let i be the weight of the i-th toxicity prediction model. The probability that the HSD11b2 inhibitor is positive when tested on the test sample is given by the i-th toxicity prediction model; N is the number of toxicity prediction models. Determine whether the final predicted probability P satisfies: P P_high or P P_low; If not, that is, P_low < P < P_high, then剔除 the output results of each toxicity prediction model; If so, judge whether the test sample has potential toxicity according to the final prediction probability P. Specifically: When P If P_high, the test sample is considered positive for HSD11b2 inhibitor, thus determining the potential toxicity of the test sample. When P If P_low, the test sample is determined to be negative for HSD11b2 inhibitor, indicating that the test sample is not toxic. Where, P_high is the first set value, and P_low is the second set value.

4. The method for predicting compound toxicity based on artificial intelligence and HSD11b2 inhibitors according to claim 1, characterized in that, The process of constructing the training sample set includes: Use a python script to screen compound data with the species type of human and the target of HSD11b2 in the ChEMBL database as training samples, and the data type of the compound is IC50.

5. The method for predicting compound toxicity based on artificial intelligence and HSD11b2 inhibitors according to claim 1, characterized in that, The feature vectors of the training samples and test samples are constructed using the molecular descriptors defined by -desclist in the rdkit.Chem.Descriptors module, and 206 molecular descriptors are selected from them for feature vector construction.

6. The method for predicting compound toxicity based on artificial intelligence and HSD11b2 inhibitors according to claim 1, characterized in that, The types of the models include two kinds, namely the first model and the second model; Use the feature vectors of each training sample as input data to train the first model to obtain the first toxicity prediction model; Use the feature vectors of each training sample as input data to train the second model to obtain the second toxicity prediction model; For the test sample, input the feature vector of the test sample into the first toxicity prediction model and the second toxicity prediction model respectively; Judge whether the test sample has potential toxicity according to the output results of the first toxicity prediction model and the second toxicity prediction model.

7. The method for predicting compound toxicity based on artificial intelligence and HSD11b2 inhibitors according to claim 6, characterized in that, The first model is a deep forest model, and the second model is a random forest model.

8. A method for detecting the synergistic toxicity effect of a compound, characterized in that, Comprising: It should be noted that the "剔除" in the original text seems to be a wrong or incomplete expression. I translated it as "剔除" as it is, but it may need to be corrected or further clarified in the actual context. Construct a first training sample set, which includes compound samples that are positive for AR inhibitors and compound samples that are negative for AR inhibitors; For each training sample in the first training sample set, molecular fingerprint features are extracted to obtain feature vectors; The feature vectors of each training sample in the first training sample set are used as input data to train multiple different types of models to obtain multiple toxicity prediction models, which are used as toxicity detection model A. Obtain multiple toxicity prediction models constructed in the toxicity prediction method according to any one of claims 1 to 7, as toxicity detection model B; The compounds whose toxicity needs to be predicted are used as test samples. The molecular fingerprint features and molecular descriptor features of the test samples are extracted respectively, and the first feature vector and the second feature vector are obtained respectively. The first feature vector of the test sample is input into the toxicity detection model A, and the second feature vector of the test sample is input into the toxicity detection model B. The test sample is determined to have a synergistic toxic effect based on the output results of toxicity detection model A and toxicity detection model B.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitor as described in any one of claims 1 to 7, and / or implements the compound synergistic toxicity effect detection method as described in claim 8.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the compound toxicity prediction method based on artificial intelligence and HSD11b2 inhibitors as described in claims 1-7, and / or implements the compound synergistic toxicity effect detection method as described in claim 8.

Images