Compound toxicity prediction method and device based on artificial intelligence and AR inhibitor

Abstract

The invention discloses a compound toxicity prediction method and device based on artificial intelligence and an androgen receptor (AR) inhibitor, and the method comprises the following steps: firstly, constructing a training sample set, training samples being compound samples with known AR inhibitor properties (positive or negative); aiming at each training sample, respectively extracting molecular fingerprint features to form feature vectors; 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; during testing, molecular fingerprint features extracted from a test sample serve as feature vectors to be input into each toxicity prediction model, whether a consensus is achieved or not is determined based on output results of each toxicity prediction model, and whether a compound corresponding to the test sample has potential toxicity or not is further judged based on a consensus result. The method can rapidly and accurately predict the potential toxicity of the compound.

Description

Technical Field

[0001] This application relates to the field of endocrine-disrupting compound toxicity prediction technology, and in particular to a method and device for predicting compound toxicity based on artificial intelligence and androgen receptor (AR) inhibitors. Background Technology

[0002] Endocrine-disrupting chemicals (EDCs) are a class of exogenous chemical substances that can interfere with the normal function of an organism's endocrine system. They affect reproductive, developmental, metabolic, and immune functions by mimicking, blocking, or disrupting the synthesis, release, metabolism, or receptor binding of natural hormones.

[0003] Androgen receptors (ARs) are ligand-dependent transcription factors in the nuclear receptor superfamily, playing a central role in the androgen signaling pathway. Abnormal AR function is closely associated with diseases such as prostate cancer, androgenetic alopecia, and polycystic ovary syndrome. Androgen receptor (AR) inhibitors, also known as anti-androgen drugs, are a class of drugs that block the binding of androgens (mainly testosterone and dihydrotestosterone) to their receptors. They do not lower androgen levels themselves, but rather prevent androgens from exerting their effects on target tissues (such as the prostate and hair follicles). AR inhibitors are an important class of endocrine disruptors (EDCs), specifically interfering with androgen-dependent physiological processes. Abnormal interference with the AR signaling pathway can lead to multi-organ developmental disorders, reproductive toxicity, metabolic disorders, and even carcinogenic risks. Therefore, detecting AR inhibitor activity is an important method for screening compound toxicity.

[0004] Detection methods for AR inhibitors generally include reporter gene assays, molecular docking techniques, SRM / MRM mass spectrometry, CTC-AR-V7 detection, GC-MS / MS, LC-MS / MS, etc. These methods generally suffer from problems such as cumbersome operation, low sensitivity, and strong equipment dependence, thus limiting the accuracy of toxicity prediction and clinical translation. Summary of the Invention

[0005] The primary objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a method for predicting the toxicity of compounds based on artificial intelligence and AR inhibitors. This method can quickly and accurately predict the potential toxicity of compounds.

[0006] A second objective of this invention is to provide a compound toxicity prediction device based on artificial intelligence and AR inhibitors.

[0007] A third objective of this invention is to provide a storage medium.

[0008] The fourth objective of this invention is to provide a computer program product.

[0009] A fourth objective of this invention is to provide a compound toxicity prediction system.

[0010] The first objective of this invention is achieved through the following technical solution: a method for predicting the toxicity of compounds based on artificial intelligence and AR inhibitors, comprising: A training sample set is constructed, which includes compound samples that are positive for AR inhibitors and compound samples that are negative for AR inhibitors. For each training sample, molecular fingerprint features are extracted to form a feature vector; the molecular fingerprint features include Morgan fingerprint features and structural statistical fingerprint features, and the feature vector is obtained by fusing the Morgan fingerprint features and structural statistical fingerprint features; 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 compounds whose toxicity needs to be predicted are used as test samples, and the molecular fingerprint features of the test samples are extracted to form a feature vector. Input the feature vectors of the test samples into each toxicity prediction model; Determine whether a consensus has been reached in the predictions based on the output results of each toxicity prediction model; If so, the potential toxicity of the test sample is determined based on the output results of each toxicity prediction model.

[0011] Preferably, the process of constructing the training sample set includes: Using a Python script, compound data with human species and AR target were selected from the ChEMBL database as training samples. The data type of the compounds was IC50. The training sample set includes multiple first training samples and multiple second training samples; The first training sample is a training sample that is positive for AR inhibitors, and the second training sample is a training sample that is negative for AR inhibitors.

[0012] Furthermore, the construction of the feature vectors for the training and test samples includes: Molecular fingerprint features, including Morgan fingerprint features and structural statistical fingerprint features, are calculated from SMILES encoding using RDKit via Python scripts.

[0013] Preferably, the types of models include two types, namely a first model and a second model; The first toxicity prediction model is obtained by using the feature vectors of each training sample in the training sample set as input data to train the first model. The second toxicity prediction model is obtained by using the feature vectors of each training sample in the training sample set as input data to train the second model. For a test sample, input the feature vectors of the test sample into the first toxicity prediction model and the second toxicity prediction model respectively; Determine whether the outputs of the first toxicity prediction model and the second toxicity prediction model reach a consensus; If so, determine whether the test sample has potential toxicity based on the output results of the first toxicity prediction model and the second toxicity prediction model.

[0014] Preferably, determining whether each toxicity prediction model reaches a consensus according to the output results of each toxicity prediction model includes: For the same test sample, after inputting the feature vectors of the test sample into each toxicity prediction model respectively, determine whether the output results of each toxicity prediction model meet the preset conditions; If so, determine that each toxicity prediction model reaches a consensus, otherwise determine that each toxicity prediction model does not reach a consensus; The preset condition is that the predicted categories output by each toxicity prediction model are the same, or the final predicted probability calculated based on the predicted probabilities of each toxicity prediction model meets the set value.

[0015] Furthermore, the final predicted probability P calculated based on the predicted probabilities of each toxicity prediction model is: P = ; Wherein, is the weight of the i-th toxicity prediction model, is the predicted probability that the AR inhibitor is positive output by the i-th toxicity prediction model when testing the test sample; N is the number of toxicity prediction models; When satisfies: P P_high or P P_low, it is determined that the final predicted probability meets the set value, that is, a consensus is reached; When P_low < P < P_high, it is determined that the final predicted probability does not meet the set value, that is, a consensus cannot be reached; Where P_high is the first set value and P_low is the second set value.

[0016] The second object of the present invention is achieved by the following technical solution: A compound toxicity prediction device based on artificial intelligence and AR inhibitors, comprising: A training sample set construction module for constructing a training sample set, where the training sample set includes compound samples with positive AR inhibitors and compound samples with negative AR inhibitors; A first feature extraction module for extracting molecular fingerprint features for each training sample respectively to form feature vectors; The training module is used to train multiple different types of models by taking the feature vectors of each training sample as input data, so as to obtain multiple toxicity prediction models. The second feature extraction module is used to extract the molecular fingerprint features of the test sample to form a feature vector, where the test sample is the compound whose toxicity needs to be predicted. The judgment module is used to input the feature vectors of the test samples into each toxicity prediction model, and judge whether the prediction has reached a consensus based on the output results of each toxicity prediction model. The prediction module is used to determine whether the test sample has potential toxicity based on the judgment result of the judgment module and the output results of each toxicity prediction model.

[0017] The third objective of this invention is achieved through the following technical solution: a computer-readable storage medium storing a computer program thereon, wherein when the computer program is executed by a processor, it implements the compound toxicity prediction method based on artificial intelligence and AR inhibitors as described in the first objective of this invention.

[0018] The fourth objective of this invention is achieved through the following technical solution: a computer program product, comprising a computer program, characterized in that, when the computer program is executed by a processor, it implements the compound toxicity prediction method based on artificial intelligence and AR inhibitors as described in the first objective of this invention.

[0019] The fifth objective of this invention is achieved through the following technical solution: a compound toxicity prediction device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the compound toxicity prediction method based on artificial intelligence and AR inhibitors as described in the first objective of this invention.

[0020] The present invention has the following advantages and effects compared with the prior art: (1) The compound toxicity prediction method based on artificial intelligence and AR inhibitors of the present invention first constructs a training sample set, in which the training samples are compound samples with known AR inhibitor properties (positive or negative); for each training sample, molecular fingerprint features are extracted to form a feature vector; 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; during testing, the molecular fingerprint features extracted from the test sample are used as feature vectors to input into each toxicity prediction model, and the consensus is determined based on the output results of each toxicity prediction model, so as to further determine whether the compound corresponding to the test sample has potential toxicity based on the consensus results. The present invention uses the molecular fingerprint features of the compound as a consensus model after training multiple machine learning models. When each trained model reaches a consensus, the potential toxicity of the compound is determined based on the consensus results. Compared with the compound toxicity prediction methods in the prior art, the method of the present invention, combined with the trained artificial intelligence model, can quickly and accurately predict the potential toxicity of the compound.

[0021] (2) In the compound toxicity prediction method based on artificial intelligence and AR inhibitors of the present invention, when constructing the training sample set, the compound data with human species and AR target are selected from the ChEMBL database as training samples by using a Python script, and the data type of the compound is IC50. The method of the present invention uses human source data to construct the training sample set, which can effectively avoid species bias, and the unified data type can improve data quality and avoid the standardization challenge of multiple types of data, so as to train a more accurate toxicity prediction model.

[0022] (3) In the compound toxicity prediction method based on artificial intelligence and AR inhibitors of the present invention, in order to achieve toxicity detection, the molecular fingerprint features of the compound are extracted to form a feature vector, specifically including Morgan fingerprint features and structural statistical fingerprint features. Based on the above feature vector, the prediction model constructed by the present invention can achieve a higher accuracy (ACC) for predicting AR inhibitor categories, thus effectively improving the accuracy of toxicity prediction by the method of the present invention.

[0023] (4) In the compound toxicity prediction method based on artificial intelligence and AR inhibitors of the present invention, multiple types of models (e.g., the first model and the second model) are combined to train each toxicity prediction model. Finally, the final prediction result is determined based on the consensus results of each toxicity prediction model. Compared with a single type of model, the present invention uses different types of models to complement each other's defects, so as to further improve the accuracy of prediction.

[0024] (5) In the compound toxicity prediction method based on artificial intelligence and AR inhibitors of the present invention, consensus can be determined based on whether the prediction categories output by each toxicity prediction model are the same. This method makes the determination algorithm simpler and faster. In addition, the present invention can also calculate the final prediction probability by weighted averaging the prediction probabilities output by each toxicity prediction model, and determine whether the toxicity prediction models have reached a consensus based on whether the final prediction probability meets the set value. Based on this, the present invention fully considers the differences between each toxicity prediction model, gives full play to the toxicity prediction model with better performance, and makes the final toxicity prediction result more accurate. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention.

[0026] Figure 2 This is a diagram illustrating the experimental verification of AR inhibitors as positive compounds.

[0027] Figure 3 This is a flowchart of the apparatus of the present invention. Detailed Implementation

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

[0029] Example 1 AR inhibitors are a core subset of EDC (androgen-dependent hormone) pathways, specifically interfering with androgen-dependent physiological processes. Abnormal interference with the AR signaling pathway can lead to multi-organ developmental disorders, reproductive toxicity, metabolic disturbances, and even carcinogenic risks. Currently, detecting the toxicity of AR inhibitors is a crucial method in modern toxicology and drug safety evaluation, particularly for assessing endocrine disruption, reproductive toxicity, and organ-specific toxicity. However, current detection methods for AR inhibitors generally suffer from cumbersome procedures, low sensitivity, and strong equipment dependence, thus limiting the accuracy of toxicity prediction and its clinical translation.

[0030] To address the problems existing in the prior art, this embodiment provides a compound toxicity prediction method based on artificial intelligence and AR inhibitors. This method determines the potential toxicity of a compound by detecting whether it contains an AR inhibitor. Figure 1 As shown, the prediction method in this embodiment includes: S1. Construct a training sample set, which includes compound samples that are positive for AR inhibitors and compound samples that are negative for AR inhibitors. Compound samples that are positive for AR inhibitors refer to compounds that have been experimentally verified to have inhibitory activity against the AR target, while compound samples that are negative for AR inhibitors refer to compounds that have been experimentally verified to have no or weak inhibitory activity against the AR target.

[0031] In this embodiment, a Python script is used to screen compound data from the ChEMBL database that are of human species type and target AR as training samples, wherein the selected compound data type is IC50. The training sample set includes multiple first training samples and multiple second training samples; the first training samples are training samples that are positive for AR inhibitors, and the second training samples are training samples that are negative for AR inhibitors.

[0032] In this embodiment, after obtaining the compound data through a Python script, data cleaning is performed to remove data with noise, defects, anomalies, or other issues. Specifically, in this embodiment, the ratio of training samples positive for AR inhibitors to training samples negative for AR inhibitors in the training sample set can be controlled to approximately 3:1. For example, it includes 1355 training samples positive for AR inhibitors and 472 training samples negative for AR inhibitors. Some of the training samples are shown in Table 1 below.

[0033] Table 1

[0034] In the table above, "active" indicates that the compound is positive for AR inhibitors, and "inactive" indicates that the compound is negative for AR inhibitors.

[0035] S2. For each training sample in the training sample set, extract the molecular fingerprint features to form a feature vector.

[0036] In this embodiment, the molecular fingerprint features extracted from the training samples include Morgan fingerprint features (MF) and structural statistical fingerprint features (CMF). The Morgan fingerprint feature (MF) is a 2D connectivity fingerprint, essentially based on the two-dimensional topological structure (atomic connectivity) of molecules, using a hash algorithm to transform substructures into fixed-length vectors. The structural statistical fingerprint feature (CMF) is also a 2D fingerprint, using integer counts to represent the frequency of occurrence of substructures / features, rather than simply using 0 / 1 to indicate presence or absence.

[0037] For each training sample in the training sample set, after extracting the Morgan fingerprint features (MF) and structural statistical fingerprint features (CMF), they can be merged into a multi-dimensional vector to form a feature vector. Specifically, the extracted MF and CMF features can be directly concatenated. Before concatenation, the high-dimensional features (CMF features) need to be reduced in dimensionality, and then concatenated with the low-dimensional features (MF features), with the concatenated vector serving as the feature vector. Alternatively, the MF and CMF features can be directly concatenated, followed by dimensionality reduction to obtain the feature vector. Kernel functions can also be designed for MF and CMF respectively, and then fused to obtain a kernel matrix, which can also serve as the feature vector. Other feature fusion methods can also be used. In this embodiment, the importance of input features is analyzed based on SHAP values, and the top 10 important features are identified as Col_2210, Col_790, Col_2118, Col_2298, Col_1274, Col_1384, Col_674, Col_2160, Col_2218 and Col_2105, where the Col_ prefix represents the sparse matrix index of the Morgan fingerprint, corresponding to a specific molecular substructure.

[0038] In this embodiment, for each compound training sample, molecular fingerprint features, including Morgan fingerprint features and structural statistical fingerprint features, are calculated from SMILES encoding using a Python script and RDKit.

[0039] S3. Use the feature vectors of each training sample as input data to train multiple different types of models to obtain multiple toxicity prediction models.

[0040] In this embodiment, a validation sample set can also be constructed through step S1. The ratio of the number of samples in the training sample set to the number of samples in the validation sample set can be set to 8:2. During the model training process, multiple cross-validation, such as 5 times, can be performed to avoid model overfitting.

[0041] 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.

[0042] In this embodiment, two or more different types of models can be selected. For example, if two models are selected, namely the first model and the second model, then in this embodiment, the feature vectors of each training sample are used as input data to train the first model to obtain the first toxicity prediction model; the feature vectors of each training sample are used as input data to train the second model to obtain the second toxicity prediction model.

[0043] In this embodiment, the first model can be a deep forest model, and the second model can be a random forest model. The deep forest model is a deep learning model that automatically learns feature representations through a cascaded forest structure. The model consists of multiple forest layers, each receiving the output of the previous layer and the original features. The input of the first layer is the original features, and the input of each subsequent layer is the original features plus the outputs of all the forests in the previous layer. In this embodiment, the first toxicity prediction model built using the deep forest model takes the average of the probability vectors of all forest outputs in the last layer during prediction, and then selects the category corresponding to the highest probability. The random forest model is an ensemble learning method composed of multiple decision trees, making the final decision through voting or averaging. When training the second model, n samples (with replacement) are randomly selected from the training sample set constructed in step S1 to form a training subset. This process is repeated to generate multiple training subsets, each used to train a decision tree. The second toxicity prediction model trained using the random forest model uses voting by each tree during prediction, and the final result is the category with the most votes.

[0044] S4. Use the compound whose toxicity needs to be predicted as a test sample, extract the molecular fingerprint features of the test sample, and construct a feature vector.

[0045] In this embodiment, the method for extracting molecular fingerprint features from the test sample can be as shown in step S2. The molecular fingerprint features extracted from the test sample include Morgan fingerprint features (MF) and structural statistical fingerprint features (CMF). The method for constructing the feature vector of the test sample can be as shown in step S2.

[0046] S5. Input the feature vectors of the test samples into each toxicity prediction model.

[0047] In this embodiment, if a first toxicity prediction model and a 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 toxicity prediction model obtains a first output result based on the input test sample features, and the second toxicity prediction model obtains a second output result based on the input test sample features.

[0048] S6. Determine whether a consensus has been reached on the predictions based on the output results of each toxicity prediction model. If so, determine whether the test sample has potential toxicity based on the output results of each toxicity prediction model; otherwise, do not adopt the output results of each toxicity prediction model, i.e., the prediction has failed. The output results of each toxicity prediction model include the predicted probability of AR inhibitors being positive and the predicted probability of AR inhibitors being negative.

[0049] In this embodiment, for the same test sample, the feature vector of the test sample is input into each toxicity prediction model, and it is determined whether the output results of each toxicity prediction model meet preset conditions. If so, it is determined that the toxicity prediction models have reached a consensus; otherwise, it is determined that the toxicity prediction models have not reached a consensus. In this embodiment, if there are two toxicity prediction models, a first toxicity prediction model and a second toxicity prediction model, it is necessary to determine whether the first toxicity prediction model and the second toxicity prediction model have reached a consensus. If so, it is determined whether the test sample has potential toxicity based on the output results of the first toxicity prediction model and the second toxicity prediction model.

[0050] In this embodiment, the aforementioned preset condition can be that the prediction categories output by each toxicity prediction model are the same. Specifically, if the prediction categories output by each toxicity prediction model are all positive for AR inhibitors, then it is determined that the toxicity prediction models have reached a consensus, and the potential toxicity of the test sample is judged. If the prediction categories output by each toxicity prediction model are all negative for AR inhibitors, then it is determined that the toxicity prediction models have reached a consensus, and the test sample is judged to be non-toxic. If the prediction categories output by one or more toxicity prediction models are all positive for AR inhibitors, and the prediction categories output by other toxicity prediction models are all negative for AR inhibitors, then it is determined that the toxicity prediction models have not reached a consensus, and the corresponding results are discarded.

[0051] In this embodiment, the aforementioned preset conditions can also be based on the final prediction probability calculated from the prediction probabilities of each toxicity prediction model satisfying the set value.

[0052] The final prediction probability P calculated based on the prediction probabilities of each toxicity prediction model is: P= ; in: The above Let be the weight of the i-th toxicity prediction model, and be the normalized weight, satisfying . N represents the number of toxicity prediction models. In this embodiment, the weight of each toxicity prediction model can be determined based on the model's performance. Specifically, it can be determined based on the model's accuracy in validating the validation sample set or its AUC (area under the ROC curve). Models with higher validation accuracy or AUC can be assigned greater weights, meaning models with better performance can be assigned higher weights. In addition, the model weights can be dynamically adjusted based on the prediction differences to make the adjusted models more accurate in validating the validation sample set.

[0053] The above This represents the predicted probability that the AR inhibitor will be positive when the i-th toxicity prediction model is used to test the test sample.

[0054] In this embodiment, when the following condition is satisfied: P ≥ P_high or P ≤ P_low, it is determined that the final predicted probability meets the set value, that is, a consensus is reached. Here, P_high is the first set value and P_low is the second set value.

[0055] In this embodiment, when the first toxicity prediction model constructed by a deep forest model and the second toxicity prediction model constructed by a random forest model are included, the final predicted probability is calculated by weighted averaging the predicted probabilities of the first toxicity prediction model and the second toxicity prediction model. For example, weights W1 and W2 are respectively assigned to the first toxicity prediction model and the second toxicity prediction model. If the predicted probability that the first toxicity prediction model predicts the AR inhibitor as positive is P1, and the probability that the second toxicity prediction model predicts the AR inhibitor as positive is P2, then the final predicted probability P = W1 * P1 + W2 * P2, where W1 + W2 = 1. When the calculated P is greater than or equal to the first preset value P_high, or less than the second preset value P_low, it is determined that the first toxicity prediction model and the second toxicity prediction model can reach a consensus; otherwise (that is, P_low < P < P_high), it is determined that a consensus cannot be reached, and this result will be excluded.

[0056] In this embodiment, the 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, then W1 can be set higher than W2. For example, W1 can be set to 0.6 and W2 can be set to 0.4, and specific adjustments can be made according to the actual usage situation.

[0057] In this embodiment, the above first preset value P_high can be set to 0.7, and P_low can be set to 0.3. Of course, corresponding adjustments can also be made according to the actual test situation. Based on this, when the final predicted probability is greater than or equal to 0.7, it is determined that each toxicity prediction model reaches a consensus, and the AR inhibitor of the corresponding test sample is positive, indicating that the test sample has potential toxicity; when the final predicted probability is less than or equal to 0.3, it is determined that each toxicity prediction model reaches a consensus, and the AR inhibitor of the corresponding test sample is negative, indicating that the test sample has no toxicity. When 0.3 < P < 0.7, it is determined that each toxicity prediction model cannot reach a consensus.

[0058] In summary: In this embodiment, a compound toxicity prediction method based on artificial intelligence and AR inhibitors uses the molecular fingerprint features of training samples as feature vectors to train various types of models, resulting in multiple toxicity prediction models. During testing, the feature vectors formed by the molecular fingerprint features extracted from the test samples are input into each toxicity prediction model. The output results of each toxicity prediction model are used to determine whether a consensus has been reached, and further, the consensus result is used to determine whether the corresponding compound of the test sample has potential toxicity. This invention, based on the molecular fingerprint features of compounds, uses multiple trained machine learning models as consensus models. When the trained models reach a consensus, the consensus result is used to determine whether the compound has potential toxicity. Compared to traditional compound toxicity prediction methods in the prior art, this embodiment's method, combined with trained artificial intelligence models, can quickly and accurately predict the potential toxicity of compounds.

[0059] In this embodiment, the molecular fingerprint features extracted from the compound include Morgan fingerprint features (MF) and structural statistical fingerprint features (CMF). Since the Morgan fingerprint feature (MF) captures the local molecular environment through radius iteration, it can encode atom types, bond types, functional groups, and their surrounding topological environment. The structural statistical fingerprint feature (CMF) encodes molecular fragments, ring systems, hydrogen bond donors / acceptors, and other statistical features directly correlated with key properties of AR suppression, such as LogP and polar surface area. Combining the two allows for the learning of subtle structural differences (e.g., the sulfonyl group of Bicalutamide vs. the hydroxyl group of DHT), improving the model's generalization ability and making it more robust. Therefore, the feature vectors formed by the above feature combination enable the training of various toxicity prediction models with optimal performance and higher classification accuracy. After 5-fold cross-validation and averaging, the first toxicity prediction model trained by the deep forest model in this embodiment achieved an accuracy (ACC) of 0.912, an area under the ROC curve (AUC) of 0.960, a specificity of 0.860, a sensitivity of 0.949, and an F1 score of 0.942. After 5-fold cross-validation and averaging, the second toxicity prediction model trained by the random forest model in this embodiment achieved an accuracy (ACC) of 0.904, an area under the ROC curve (AUC) of 0.955, a specificity of 0.841, a sensitivity of 0.935, and an F1 score of 0.936. Overall, the accuracy of compound toxicity prediction has been significantly improved.

[0060] In this embodiment, when determining whether consensus has been reached on various toxicity predictions, the final prediction probability obtained by weighted averaging of the prediction probabilities of each toxicity prediction model can be used as the criterion. Consensus is only considered reached when the final prediction probability meets a set value. Therefore, the method in this embodiment considers the differences between various toxicity prediction models, fully utilizing the more superior toxicity prediction model to ensure more accurate final toxicity prediction results. This toxicity prediction method is applicable to different types of compounds and can serve as a general compound toxicity screening tool, such as for organic compounds like fluorocarbons.

[0061] In this embodiment, experimental verification was conducted based on compounds that showed positive results for AR inhibitors: taking hexa(1H,1H,7H-perfluoroheptoxy)phosphazene (CAS: 3830-74-8), a compound that showed positive results for AR inhibitors, as an example, its potential AR inhibitory effect was demonstrated through zebrafish embryo exposure assays. Figure 2 As shown, zebrafish embryos were exposed to hexa(1H,1H,7H-perfluoroheptoxy)phosphazene at concentrations of 0.01 ng / L, 0.1 ng / L, 1 ng / L, 10 ng / L, and 100 ng / L for 7 days. Hexa(1H,1H,7H-perfluoroheptoxy)phosphazene not only reduced vitellogenin (VTG) levels but also reduced androgen AR receptor levels, indicating that it has an AR inhibitory effect.

[0062] Those skilled in the art will understand that all or part of the steps in the method of this embodiment can be implemented by a program instructing related hardware, and the corresponding program 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 accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Rather, the described steps may be performed in a different order, some steps may be performed simultaneously, additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0063] Example 2 This embodiment discloses a compound toxicity prediction device based on artificial intelligence and AR inhibitors, such as Figure 3 As shown, it includes: The training sample set construction module is used to construct a training sample set, which includes compound samples that are positive for AR inhibitors and compound samples that are negative for AR inhibitors. In this embodiment, the ratio of the number of positive AR inhibitor compound samples to the number of negative AR inhibitor compound samples in the training sample set is approximately 3:1. The specific method for obtaining the samples can be as shown in step S1 of Embodiment 1, and will not be repeated here.

[0064] The first feature extraction module is used to extract molecular fingerprint features for each training sample to form a feature vector. In this embodiment, the first feature extraction module extracts features for the training samples. The extracted molecular fingerprint features include the Morgan fingerprint features (MF) and structural statistical fingerprint features (CMF) of each training sample. The two features are fused to obtain the feature vector of the training sample. The specific fusion method is as shown in step S2 of embodiment 1, and will not be repeated here. In this embodiment, the top 10 important features can be selected based on feature importance, such as: Col_2210, Col_790, Col_2118, Col_2298, Col_1274, Col_1384, Col_674, Col_2160, Col_2218 and Col_2105. The Col_ prefix indicates the sparse matrix index of the Morgan fingerprint, corresponding to a specific molecular substructure.

[0065] The training module is used to train multiple different types of models using the feature vectors of each training sample as input data, thereby obtaining multiple toxicity prediction models. In this embodiment, the different types of models can be selected from traditional machine learning models, such as linear models, tree-based models, support vector machines, clustering algorithms, etc., or from deep learning models, such as feedforward neural network models, convolutional neural network models, recurrent neural network models, self-attention mechanism models, etc. In this embodiment, two models are trained using the training sample set: a first model constructed using a deep forest model and a second model constructed using a random forest model. After training, a first toxicity prediction model and a second toxicity prediction model are obtained.

[0066] The second feature extraction module is used to extract molecular fingerprint features of the test sample to form a feature vector, where the test sample is a compound whose toxicity needs to be predicted. In this embodiment, the second feature extraction module performs feature extraction on the test sample. The extracted molecular fingerprint features include the Morgan fingerprint feature (MF) and the structural statistical fingerprint feature (CMF) of the test sample. The two features are fused to obtain the feature vector of the test sample. The specific fusion method is as shown in step S2 of Embodiment 1, and will not be repeated here.

[0067] The judgment module is used to input the feature vector of the test sample into each toxicity prediction model, and determine whether the prediction has reached a consensus based on the output results of each toxicity prediction model. In this embodiment, for the same test sample, after inputting the feature vector of the test sample into each toxicity prediction model, the judgment module determines whether the output results of each toxicity prediction model meet preset conditions; if so, it is determined that the toxicity prediction models have reached a consensus; otherwise, it is determined that the toxicity prediction models have not reached a consensus.

[0068] In this embodiment, the aforementioned preset conditions may include a first preset condition and a second preset condition, wherein: The first preset condition is that the predicted categories output by each toxicity prediction model are the same. Specifically, if the predicted categories output by each toxicity prediction model are all positive for AR inhibitors or all negative for AR inhibitors, it is determined that the toxicity prediction models reach a consensus; otherwise, it is determined that the toxicity prediction models do not reach a consensus, and the corresponding results are excluded.

[0069] The second preset condition is that the final predicted probability calculated based on the predicted probabilities of each toxicity prediction model meets the set value. Among them, the final predicted probability P calculated based on the predicted probabilities of each toxicity prediction model is: P = ; Among them: The above is the weight of the i-th toxicity prediction model, and is the normalized weight, satisfying ; N is the number of toxicity prediction models; in this embodiment, the weights of each toxicity prediction model can be determined according to the performance of the model. For models with higher verification accuracy or AUC, larger weights can be assigned, that is, models with better performance are assigned higher weights; in addition, the weights of the models can also be dynamically adjusted according to the prediction differences to make the adjusted models have higher verification accuracy for the verification sample set. is the predicted probability that the AR inhibitor is positive output by the i-th toxicity prediction model when testing the test sample.

[0070] In this embodiment: When satisfies: P P_high or P P_low, it is determined that the final predicted probability meets the set value, that is, the toxicity prediction models reach a consensus; When P_low < P < P_high, it is determined that the final predicted probability does not meet the set value, that is, the toxicity prediction models do not reach a consensus; The above P_high is the first set value, and P_low is the second set value.

[0071] The prediction module is used to determine whether the test sample is potentially toxic according to the determination result of the judgment module and the output results of each toxicity prediction model.

[0072] In this embodiment, when the judgment result of the judgment module is that the toxicity prediction models reach a consensus, the prediction module determines whether the test sample is potentially toxic according to the output results of the toxicity prediction models. When the preset condition is the first preset condition, if the prediction categories of all the toxicity prediction models are positive for AR inhibitors, it is determined that the test sample is potentially toxic; if the prediction categories of all the toxicity prediction models are negative for AR inhibitors, it is determined that the test sample has no toxicity. When the preset condition is the second preset condition, if P > P_high, it is determined that the final prediction category is positive for AR inhibitors, and it is determined that the test sample is potentially toxic; if P < P_low, it is determined that the final prediction category is negative for AR inhibitors, and it is determined that the test sample has no toxicity.

[0073] In this embodiment, the above device further includes a verification sample set construction module for constructing a verification sample set. The number of verification samples in the constructed verification sample set and the number of training samples in the training sample set are 2:8. When the training module is training, 5-fold cross-validation is simultaneously performed based on the verification sample set.

[0074] The device in this embodiment can execute the method provided in Embodiment 1, and its implementation principle is similar. The actions performed by each module in the device in this embodiment correspond to the steps in the method in Embodiment 1. For the detailed functional descriptions of each module of the device, reference can specifically be made to the descriptions in the corresponding method shown above, and details are not described herein again.

[0075] Embodiment 3 This embodiment discloses a storage medium storing a program, which when executed by a processor, implements the method for predicting the toxicity of a compound based on artificial intelligence and AR inhibitors described in Embodiment 1, including: Constructing a training sample set, where the training sample set includes compound samples that are positive for AR inhibitors and compound samples that are negative for AR inhibitors; For each training sample, molecular fingerprint features are respectively extracted to form a feature vector; Taking the feature vectors of each training sample as input data, training multiple different types of models respectively to obtain multiple toxicity prediction models; Taking the compound for which toxicity needs to be predicted as a test sample, extracting the molecular fingerprint features of the test sample to form a feature vector; Inputting the feature vector of the test sample into each toxicity prediction model respectively; Judging whether the predictions reach a consensus according to the output results of each toxicity prediction model; If so, determining whether the test sample is potentially toxic based on the output results of each toxicity prediction model.

[0076] The specific implementation processes of the above steps can be as described in Embodiment 1.

[0077] In this embodiment, the storage medium can be a disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), USB flash drive, portable hard drive, etc.

[0078] Example 4 This embodiment discloses a computer program product, including a computer program that, when executed by a processor, implements the compound toxicity prediction method based on artificial intelligence and AR inhibitors described in Embodiment 1, including: A training sample set is constructed, which includes compound samples that are positive for AR inhibitors and compound samples that are negative for AR inhibitors. For each training sample, molecular fingerprint features are extracted to form a feature vector; 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 compounds whose toxicity needs to be predicted are used as test samples, and the molecular fingerprint features of the test samples are extracted to form a feature vector. Input the feature vectors of the test samples into each toxicity prediction model; Determine whether a consensus has been reached in the predictions based on the output results of each toxicity prediction model; If so, the potential toxicity of the test sample is determined based on the output results of each toxicity prediction model.

[0079] The specific implementation process of each of the above steps can be as described in Example 1.

[0080] The computer program product includes one or more computer instructions that, when loaded and executed on a computer, generate, in whole or in part, the processes or functions described in the embodiments of this application. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in or transmitted through a computer-readable storage medium. The computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital versatile disc (DVD)), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0081] Example 5 This embodiment discloses a compound toxicity prediction device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the compound toxicity prediction method based on artificial intelligence and AR inhibitors described in Embodiment 1, including: A training sample set is constructed, which includes compound samples that are positive for AR inhibitors and compound samples that are negative for AR inhibitors. For each training sample, molecular fingerprint features are extracted to form a feature vector; 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 compounds whose toxicity needs to be predicted are used as test samples, and the molecular fingerprint features of the test samples are extracted to form a feature vector. Input the feature vectors of the test samples into each toxicity prediction model; Determine whether a consensus has been reached in the predictions based on the output results of each toxicity prediction model; If so, the potential toxicity of the test sample is determined based on the output results of each toxicity prediction model.

[0082] The specific implementation process of each of the above steps can be as described in Example 1.

[0083] In this embodiment, the compound toxicity prediction device can be implemented by terminal devices such as mobile phones, computers, and tablets.

[0084] In this embodiment, the compound toxicity prediction system includes a processor, a memory, a bus, and a communication interface, wherein the processor, the communication interface, and the memory are connected via the bus; the processor is configured to execute executable modules, such as computer programs, stored in the memory.

[0085] The above description is only an optional implementation method for some implementation scenarios of this application. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this application without departing from the technical concept of this application also fall within the protection scope of the embodiments of this application.

Claims

1. A method for predicting the toxicity of compounds based on artificial intelligence and AR inhibitors, characterized in that, include: A training sample set is constructed, which includes compound samples that are positive for AR inhibitors and compound samples that are negative for AR inhibitors. For each training sample, molecular fingerprint features are extracted to form a feature vector; the molecular fingerprint features include Morgan fingerprint features and structural statistical fingerprint features, and the feature vector is obtained by fusing the Morgan fingerprint features and structural statistical fingerprint features; 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 compounds whose toxicity needs to be predicted are used as test samples, and the molecular fingerprint features of the test samples are extracted to form a feature vector. Input the feature vectors of the test samples into each toxicity prediction model; Determine whether a consensus has been reached in the predictions based on the output results of each toxicity prediction model; If so, the potential toxicity of the test sample is determined based on the output results of each toxicity prediction model.

2. The compound toxicity prediction method based on artificial intelligence and AR inhibitors according to claim 1, characterized in that, The process of constructing the training sample set includes: Using a Python script, compound data with human species and AR target were selected from the ChEMBL database as training samples to form a training sample set. The data type of the compounds was IC50. The training sample set includes multiple first training samples and multiple second training samples; The first training sample is a training sample that is positive for AR inhibitors, and the second training sample is a training sample that is negative for AR inhibitors. Based on artificial intelligence and AR inhibitors.

3. The compound toxicity prediction method based on artificial intelligence and AR inhibitors according to claim 1, characterized in that, The construction of the feature vectors for the training and testing samples includes: Molecular fingerprint features, including Morgan fingerprint features and structural statistical fingerprint features, are calculated from SMILES encoding using RDKit via Python scripts.

4. The compound toxicity prediction method based on artificial intelligence and AR inhibitors according to claim 1, characterized in that, The models include two types: the first model and the second model. The first toxicity prediction model is obtained by using the feature vectors of each training sample in the training sample set as input data to train the first model. The second toxicity prediction model is obtained by using the feature vectors of each training sample in the training sample set as input data to train the second model. For the test sample, the feature vector of the test sample is input into the first toxicity prediction model and the second toxicity prediction model respectively; Determine whether the outputs of the first toxicity prediction model and the second toxicity prediction model reach a consensus; If so, the potential toxicity of the test sample is determined based on the output results of the first toxicity prediction model and the second toxicity prediction model.

5. The method for predicting compound toxicity based on artificial intelligence and AR inhibitors according to any one of claims 1 to 4, characterized in that, Determine whether the toxicity prediction models have reached a consensus based on their output results, including: For the same test sample, the feature vector of the test sample is input into each toxicity prediction model, and then it is determined whether the output results of each toxicity prediction model meet the preset conditions. If so, it is determined that the toxicity prediction models have reached a consensus; otherwise, it is determined that the toxicity prediction models have not reached a consensus. The preset condition is that the prediction categories output by each toxicity prediction model are the same, or the final prediction probability calculated based on the prediction probabilities of each toxicity prediction model meets the set value.

6. The compound toxicity prediction method based on artificial intelligence and AR inhibitors according to claim 5, characterized in that, The final prediction probability P calculated based on the prediction probabilities of each toxicity prediction model is: P= ； in, Let i be the weight of the i-th toxicity prediction model. When testing a test sample, N represents the predicted probability that the AR inhibitor output by the i-th toxicity prediction model is positive; N is the number of toxicity prediction models. when Satisfy: P P_high or P If P_low is reached, it is determined that the final predicted probability meets the set value, that is, a consensus is reached. When P_low < P < P_high, it is determined that the final predicted probability does not meet the set value, that is, consensus cannot be reached; where P_high is the first set value and P_low is the second set value.

7. A compound toxicity prediction device based on artificial intelligence and AR inhibitors, characterized in that, It includes: A training sample set construction module for constructing a training sample set, which includes compound samples with positive AR inhibitors and compound samples with negative AR inhibitors; A first feature extraction module for extracting molecular fingerprint features for each training sample respectively to form a feature vector; A training module for using the feature vectors of each training sample as input data to train multiple different types of models respectively to obtain multiple toxicity prediction models; A second feature extraction module for extracting the molecular fingerprint features of the test sample to form a feature vector, where the test sample is a compound for which toxicity needs to be predicted; A judgment module for inputting the feature vectors of the test sample into each toxicity prediction model respectively, and judging whether the prediction reaches a consensus according to the output results of each toxicity prediction model; A prediction module for determining whether the test sample has potential toxicity according to the determination result of the judgment module and the output results of each toxicity prediction model.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the compound toxicity prediction method based on artificial intelligence and AR inhibitors according to any one of claims 1 to 7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the compound toxicity prediction method based on artificial intelligence and AR inhibitors according to claims 1 to 7.

10. A compound toxicity prediction device, comprising: At least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the compound toxicity prediction method based on artificial intelligence and AR inhibitors according to any one of claims 1 to 7.

Images