A method for screening liver toxicity of chemicals based on high content cell imaging and convolutional neural network

Abstract

The application provides a method for screening liver toxicity of chemicals based on high-content cell imaging and convolutional neural network, comprising the following steps: constructing a high-content cell image data set of chemicals; randomly selecting a plurality of images, training a nucleus and cell segmentation model suitable for the high-content cell image data set of chemicals by using Cellpose2.0, performing batch processing on the images by using the trained segmentation model, and saving the processed Mask images; extracting phenotype parameters based on the Mask images and performing pretreatment; taking the pretreated phenotype parameters as input, taking the presence or absence of liver toxicity as the endpoint of the model, and constructing a classification model based on convolutional neural network; and screening liver toxicity by using the constructed classification model to obtain a prediction effect. The method effectively realizes high-throughput screening of liver toxicity of chemicals, the screening model has good fitting ability, robustness and prediction ability, provides an effective tool for risk assessment and management of chemicals, and has important application value.

Description

Technical Field

[0001] This invention relates to the field of screening technology for chemical hepatotoxicity risk assessment, and more particularly to a method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural networks. Background Technology

[0002] The number of chemicals with hepatotoxicity data is currently far lower than the existing number of chemicals, indicating a huge gap in hepatotoxicity data for chemicals.

[0003] Traditional methods for evaluating the hepatotoxicity of chemicals mainly rely on animal and clinical trials. These methods are not only time-consuming, labor-intensive, costly, and difficult to operate, but also involve ethical issues and cannot perform high-throughput screening. Therefore, developing a time-saving, labor-saving, low-cost, simple, and high-throughput screening technology has become an urgent need for current chemical hepatotoxicity screening.

[0004] Convolutional Neural Networks (CNNs), as powerful deep learning models, can delve deeper into complex patterns and structures in data through automatic feature extraction and hierarchical feature representation, thus processing high-dimensional data more accurately and efficiently. High-content imaging techniques capture numerous different microscopic descriptors by imaging stained cells, encompassing the phenotypic characteristics of cells and their organelles. These phenotypic features not only provide more information than traditional assays but also reveal the cell's response to chemicals from a broader perspective. Using phenotypic parameters as input to construct a convolutional neural network screening model for chemical hepatotoxicity holds promise for filling the gaps in the market for chemical hepatotoxicity data.

[0005] Currently, existing literature utilizes high-content phenotypic parameters to construct machine learning screening models for toxicological assessment. The literature "J Appl Toxicol. 2014, 34(3), 281-8." uses phenotypic parameters from cell imaging as input to construct a random forest classification model for screening drug hepatotoxicity; the literature "Chem. Res. Toxicol. 2020, 33, 2261-2275." constructs a drug hepatotoxicity prediction model based on high-content phenotypes and random forests; and the literature "Chem Res Toxicol. 2023, 18, 36(9), 1456–1470." utilizes high-content phenotypes to develop three machine learning methods—random forest, linear support vector machine, and elastic network—to predict compounds with potential drug-induced liver injury risk. However, most of these models employ traditional machine learning methods, are not based on deep learning, and do not use cells that can characterize the liver, such as HepG2 cells (human hepatocellular carcinoma cells). Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural networks. This invention selects high-content images of HepG2 cells, extracts phenotypic parameters from them, and applies convolutional neural networks to construct a chemical hepatotoxicity screening model. This method saves time, effort, and cost, avoids ethical issues, and effectively achieves high-throughput screening for chemical hepatotoxicity, providing technical support for filling the gaps in chemical hepatotoxicity data.

[0007] The technical means employed in this invention are as follows:

[0008] A method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural networks includes:

[0009] S1. Construct a high-content cell image dataset of chemicals;

[0010] S2. Randomly select several images, use Cellpose 2.0 to train a cell nucleus and cell segmentation model suitable for chemical high-content cell image datasets, use the trained segmentation model to batch process the images, and save the processed Mask images.

[0011] S3. Based on the saved Mask image, extract phenotypic parameters and preprocess the extracted phenotypic parameters;

[0012] S4. Using the preprocessed phenotypic parameters as input and the presence or absence of hepatotoxicity as the model endpoint, construct a classification model based on a convolutional neural network.

[0013] S5. The constructed classification model was used for liver toxicity screening to obtain the predictive results.

[0014] Further, step S1 specifically includes:

[0015] S11. Select the data with the number idr0088 in the IDR (Image Data Resource) database, and refer to the DILIst list to filter out the chemical high-content cell images contained in the DILIst list;

[0016] S12. Label chemicals with hepatotoxicity as 1 and chemicals without hepatotoxicity as 0.

[0017] Furthermore, in step S11, a total of 34,390 images were screened. These images came from 400 384-well plates, including 22,400 images of DMSO treatment and 11,990 images of chemical treatment.

[0018] Further, step S2 specifically includes:

[0019] S21. Randomly select 2 images from every 10 plates, for a total of 80 images, to train a segmentation model suitable for chemical high-content cell image datasets.

[0020] S22. Open the Cellpose 2.0 visualization interface, select a suitable base model, and use the "human-in-the-loop" method to modify the segmented region of interest, thereby training a cell and nucleus segmentation model suitable for the dataset.

[0021] S23. Use Python code to call the trained segmentation model to perform batch segmentation of the dataset and save the segmented Mask images.

[0022] Furthermore, in step S22, when training the cell nucleus segmentation model using Cellpose 2.0, the channel is set to [0,0] and the cell diameter is 25 μm. When training the cell segmentation model, the channel is set to [3,0] and the cell diameter is 37 μm.

[0023] Further, step S3 specifically includes:

[0024] S31. Using Python code, extract phenotypic parameters from the Mask image, including the area, perimeter, longest axis, shortest axis, roughness, roundness, average, maximum, minimum, extreme values, variance, median of blue fluorescence intensity of cells and nuclei, as well as the number of cells and nuclei, for a total of 80 phenotypic parameters.

[0025] S32. Perform preprocessing operations on the extracted parameters, as follows:

[0026] S321, Remove missing values;

[0027] S322. Use the nearZeroVar function in the Caret package to identify and remove phenotypic parameters with variance of 0 or close to 0.

[0028] S323. Separate the data for DMSO and chemicals, calculate the median absolute deviation score for each phenotypic parameter in each group, set the threshold to 3.5, and consider MAD_score greater than 3.5 as an outlier and remove it;

[0029] S324. Subtract the corresponding DMSO background value from the chemical parameter value to perform batch effect correction, using the following formula:

[0030]

[0031] Where 0.6745 represents a constant used to convert MAD to a scale approximating the standard deviation; X iThis represents the i-th data point in the dataset; median(X) represents the median of the dataset.

[0032] Further, in step S4, the convolutional neural network includes three convolutional layers. The first layer uses 32 3x3 filters, the second layer uses 64 3x3 filters, and the third layer uses 128 3x3 filters, all of which use the ReLU activation function. Each convolutional layer is followed by a 2x2 max pooling layer, and each pooling layer is followed by a Dropout layer with a Dropout rate of 0.5. The number of training samples per batch is 64, and the total number of iterations is 500.

[0033] Further, step S5 specifically includes:

[0034] S51. Divide the dataset into training and test sets in a 7:3 ratio. Use 10-fold cross-validation for internal validation and the test set for external validation.

[0035] S52, Using the accuracy R of the training set A(train) Sensitivity R SE(train) Specific R SP(train) The area under the ROC curve of the receiver operating characteristic curve A ROC(train) Characterize the goodness of fit of the model; where:

[0036]

[0037] Among them, R A R represents the proportion of samples that the model correctly predicted out of the total number of samples; SE R represents the proportion of samples that are actually positive but were correctly predicted as positive by the model; SP This represents the proportion of samples that were actually negative but were correctly predicted as negative by the model; TP represents true positives, indicating the number of samples correctly predicted as positive by the model; TN represents true negatives, indicating the number of samples correctly predicted as negative by the model; FP represents false positives, indicating the number of samples incorrectly predicted as positive by the model; FN represents false negatives, indicating the number of samples incorrectly predicted as negative by the model; A ROC This represents the area under the ROC curve, with values ​​ranging from 0 to 1. The closer it is to 1, the better the classification model's ability to distinguish between positive and negative classes.

[0038] S53, R using a validation set A(test) R SE(test) R SP(test) A ROC(test) Characterize the predictive power of the model;

[0039] S54. The final prediction results of the classification model are as follows:

[0040] R A(train) =0.793

[0041] R SE(train) =0.794

[0042] R SP(train) =0.793

[0043] A ROC(train) =0.881

[0044] R A(test) =0.671

[0045] R SE(test) =0.721

[0046] R SP(test) =0.590

[0047] A ROC(test) =0.732.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] 1. This invention provides a method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural networks. The method uses cell phenotypic parameters obtained from high-content imaging as input to construct a convolutional neural network classification model, and performs internal and external validation. The performance of the model is evaluated using indicators such as accuracy, which can effectively achieve high-throughput screening of chemical hepatotoxicity.

[0050] 2. The present invention provides a method for screening the hepatotoxicity of chemicals based on high-content cell imaging and convolutional neural networks. The established screening model has good fitting ability, robustness and predictive ability, providing an effective tool for chemical risk assessment and management, and has important application value.

[0051] Based on the above reasons, this invention can be widely applied in fields such as screening for chemical hepatotoxicity risk assessment. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of the method of the present invention.

[0054] Figure 2 This is a graph showing the loss change during the model training process provided in an embodiment of the present invention.

[0055] Figure 3 This is a graph showing the changes in accuracy during the model training process, as provided in an embodiment of the present invention. Detailed Implementation

[0056] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0057] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0058] like Figure 1 As shown, this invention provides a method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural networks, comprising:

[0059] S1. Construct a high-content cell image dataset of chemicals;

[0060] S2. Randomly select several images, use Cellpose 2.0 to train a cell nucleus and cell segmentation model suitable for chemical high-content cell image datasets, use the trained segmentation model to batch process the images, and save the processed Mask images.

[0061] S3. Based on the saved Mask image, extract phenotypic parameters and preprocess the extracted phenotypic parameters;

[0062] S4. Using the preprocessed phenotypic parameters as input and the presence or absence of hepatotoxicity as the model endpoint, construct a classification model based on a convolutional neural network.

[0063] S5. The constructed classification model was used for liver toxicity screening to obtain the predictive results.

[0064] In a specific implementation, as a preferred embodiment of the present invention, step S1 specifically includes:

[0065] S11. Select the data with the number idr0088 in the IDR (Image Data Resource) database, and refer to the DILIst list (a list of drug-induced liver injury classifications) to filter out the chemical high-content cell images contained in the DILIst list;

[0066] S12. Label chemicals with hepatotoxicity as 1 and chemicals without hepatotoxicity as 0.

[0067] In a specific implementation, as a preferred embodiment of the present invention, in step S11, a total of 34,390 images are screened. These images are from 400 384-well plates, including 22,400 images treated with DMSO (Dimethyl sulfoxide) (negative control) and 11,990 images treated with chemicals.

[0068] In a specific implementation, as a preferred embodiment of the present invention, step S2 specifically includes:

[0069] S21. Randomly select 2 images from every 10 plates, for a total of 80 images, to train a segmentation model suitable for chemical high-content cell image datasets.

[0070] S22. Open the Cellpose 2.0 visualization interface, select a suitable base model, and use the "human-in-the-loop" method to modify the segmented regions of interest, thereby training a cell and nucleus segmentation model suitable for the dataset; for example... Figure 2 , 3 The figures shown are the loss and accuracy changes during the model training process.

[0071] S23. Use Python code to call the trained segmentation model to perform batch segmentation of the dataset and save the segmented Mask images.

[0072] In a specific implementation, as a preferred embodiment of the present invention, in step S22, when training the cell nucleus segmentation model using Cellpose 2.0, the channel is set to [0,0] and the cell diameter is 25μm; when training the cell segmentation model, the channel is set to [3,0] and the cell diameter is 37μm.

[0073] In a specific implementation, as a preferred embodiment of the present invention, step S3 specifically includes:

[0074] S31. Using Python code, extract phenotypic parameters from the Mask image, including the area, perimeter, longest axis, shortest axis, roughness, roundness, average, maximum, minimum, extreme values, variance, and median (obtained after removing 1% of the measurements before and after each image to prevent abnormal results due to cell and segmentation abnormalities) of cells and nuclei, as well as the number of cells and nuclei, for a total of 80 phenotypic parameters;

[0075] S32. Perform preprocessing operations on the extracted parameters, as follows:

[0076] S321, Remove missing values;

[0077] S322. Use the nearZeroVar function in the Caret package to identify and remove phenotypic parameters with variance of 0 or close to 0.

[0078] S323. Separate the data for DMSO and chemicals, and calculate the median absolute deviation score (MAD_score) for each phenotypic parameter in each group (DMSO on the same plate is considered as one group, and the same drug is considered as one group). Set the threshold to 3.5. A MAD_score greater than 3.5 is considered an outlier and removed.

[0079] S324. Subtract the corresponding DMSO background value from the chemical parameter value to perform batch effect correction, using the following formula:

[0080]

[0081] Where 0.6745 represents a constant used to convert MAD to a scale approximating the standard deviation; X i This represents the i-th data point in the dataset; median(X) represents the median of the dataset.

[0082] In a preferred embodiment of the present invention, in step S4, the convolutional neural network includes three convolutional layers. The first layer uses 32 3x3 filters, the second layer uses 64 3x3 filters, and the third layer uses 128 3x3 filters, all of which use the ReLU activation function. Each convolutional layer is followed by a 2x2 max pooling layer, and each pooling layer is followed by a Dropout layer with a Dropout rate of 0.5. The number of samples trained per batch (batch_size) is 64, and the total number of iterations (epochs) is 500.

[0083] In a specific implementation, as a preferred embodiment of the present invention, step S5 specifically includes:

[0084] S51. Divide the dataset into training and test sets in a 7:3 ratio. Use 10-fold cross-validation for internal validation and the test set for external validation.

[0085] S52, Using the accuracy R of the training set A(train) Sensitivity R SE(train) Specific R SP(train) The area under the ROC curve of the receiver operating characteristic curve A ROC(train) Characterize the goodness of fit of the model; where:

[0086]

[0087]

[0088] Among them, R A R represents the proportion of samples that the model correctly predicted out of the total number of samples; SE R represents the proportion of samples that are actually positive but were correctly predicted as positive by the model; SP This represents the proportion of samples that were actually negative but were correctly predicted as negative by the model; TP represents true positives, indicating the number of samples correctly predicted as positive by the model; TN represents true negatives, indicating the number of samples correctly predicted as negative by the model; FP represents false positives, indicating the number of samples incorrectly predicted as positive by the model; FN represents false negatives, indicating the number of samples incorrectly predicted as negative by the model; A ROC This represents the area under the ROC curve, with values ​​ranging from 0 to 1. The closer it is to 1, the better the classification model's ability to distinguish between positive and negative classes.

[0089] S53, R using a validation set A(test) R SE(test) R SP(test) A ROC(test) Characterize the predictive power of the model;

[0090] S54. The final prediction results of the classification model are as follows:

[0091] R A(train) =0.793

[0092] R SE(train) =0.794

[0093] R SP(train) =0.793

[0094] A ROC(train) =0.881

[0095] R A(test) =0.671

[0096] R SE(test) =0.721

[0097] R SP(test) =0.590

[0098] A ROC(test) =0.732.

[0099] Example 1

[0100] Given a chemical procarbazine, the goal is to screen for its hepatotoxicity. First, relevant high-content images are found in cell image databases or literature. Cellpose 2.0 is used to segment the cell nuclei and cells, and the segmented mask images are saved. Phenotypic parameters are extracted using Python code. The obtained parameters undergo preprocessing operations such as removing missing values, removing phenotypic parameters with variance of 0 or close to 0, removing outliers, and batch effect correction. The model constructed in this invention is then used for prediction. The result shows an activity of 1, consistent with experimental results.

[0101] Example 2

[0102] Given a chemical called phenformin, the goal is to screen for its hepatotoxicity. First, relevant high-content images are found in cell image databases or literature. Cellpose 2.0 is used to segment the cell nuclei and cells, and the segmented mask images are saved. Phenotypic parameters are extracted using Python code. The obtained parameters undergo preprocessing operations such as removing missing values, removing phenotypic parameters with variance of 0 or close to 0, removing outliers, and batch effect correction. The model constructed in this invention is then used for prediction. The result shows an activity of 1, consistent with experimental results.

[0103] Example 3

[0104] Given a chemical called amoxapine, the goal is to screen for its hepatotoxicity. First, relevant high-content images are found in cell image databases or literature. Cellpose 2.0 is used to segment the cell nuclei and cells, and the segmented mask images are saved. Phenotypic parameters are extracted using Python code. The obtained parameters undergo preprocessing operations such as removing missing values, removing phenotypic parameters with variance of 0 or close to 0, removing outliers, and batch effect correction. The model constructed in this invention is then used for prediction. The result shows that the activity is 0, consistent with experimental results.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural networks, characterized in that, include: S1. Construct a high-content cell image dataset of chemicals, including: S11. Select the data with the number idr0088 in the IDR (Image Data Resource) database, and refer to the DILIst list to filter out the chemical high-content cell images contained in the DILIst list; S12. Label chemicals with hepatotoxicity as 1 and chemicals without hepatotoxicity as 0; S2. Randomly select several images, and use Cellpose 2.0 to train a cell nucleus and cell segmentation model suitable for chemical high-content cell image datasets. Use the trained segmentation model to batch process the images and save the processed mask images, including: S21. Randomly select 2 images from every 10 plates, for a total of 80 images, to train a segmentation model suitable for chemical high-content cell image datasets. S22. Open the Cellpose 2.0 visualization interface, select a suitable base model, and use the "human-in-the-loop" method to modify the segmented regions of interest, thereby training a cell and nucleus segmentation model suitable for the dataset; when training the cell nucleus segmentation model using Cellpose 2.0, the channels are set to [0, 0], and the cell diameter is 25. When training the cell segmentation model, the channels were set to [3, 0], and the cell diameter was 37. ; S23. Use Python code to call the trained segmentation model to perform batch segmentation of the dataset and save the segmented Mask images; S3. Based on the saved Mask image, extract phenotypic parameters and preprocess the extracted phenotypic parameters, including: S31. Using Python code, extract phenotypic parameters from the Mask image, including the area, perimeter, longest axis, shortest axis, roughness, roundness, average, maximum, minimum, extreme values, variance, median of blue fluorescence intensity of cells and nuclei, as well as the number of cells and nuclei, for a total of 80 phenotypic parameters. S32. Perform preprocessing operations on the extracted parameters, as follows: S321, Remove missing values; S322. Use the nearZeroVar function in the Caret package to identify and remove phenotypic parameters with variance of 0 or close to 0. S323. Separate the data for DMSO and chemicals, calculate the median absolute deviation score for each phenotypic parameter in each group, set the threshold to 3.5, and consider MAD_score greater than 3.5 as an outlier and remove it; S324. Subtract the corresponding DMSO background value from the chemical parameter value to perform batch effect correction, using the following formula: in, Represents a constant used to... Converted to a scale approximating the standard deviation; Represents the first in the dataset i One data point; This indicates the median of the dataset; S4. Using the preprocessed phenotypic parameters as input and the presence or absence of hepatotoxicity as the model endpoint, construct a classification model based on a convolutional neural network. The convolutional neural network consists of three convolutional layers: the first layer uses 32 3x3 filters, the second layer uses 64 3x3 filters, and the third layer uses 128 3x3 filters, all using the ReLU activation function. Each convolutional layer is followed by a 2x2 max pooling layer, and each pooling layer is followed by a Dropout layer with a Dropout rate of 0.

5. The number of training samples per batch is 64, and the total number of iterations is 500. S5. The constructed classification model was used for hepatotoxicity screening to obtain predictive results, including: S51. Divide the dataset into training and test sets in a 7:3 ratio. Use 10-fold cross-validation for internal validation and the test set for external validation. S52, Accuracy using the training set Sensitivity Specificity Area under the ROC curve of the receiver operating characteristic curve Characterize the goodness of fit of the model; where: in, This represents the proportion of samples whose predictions were correct out of the total number of samples. This represents the proportion of samples that are actually positive but were correctly predicted as positive by the model. This represents the proportion of samples that are actually negative but were correctly predicted as negative by the model. For true examples, it represents the number of samples that the model correctly predicts as positive. A true negative example represents the number of samples that the model correctly predicted as negative. False positives represent the number of samples that the model incorrectly predicted as positive. False negatives represent the number of samples that the model incorrectly predicted as negative. This represents the area under the ROC curve, with values ​​ranging from 0 to 1. The closer it is to 1, the better the classification model's ability to distinguish between positive and negative classes. S53, Using a validation set , , , Characterize the predictive power of the model; S54. The final prediction results of the classification model are as follows: 。 2. The method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural networks according to claim 1, characterized in that, In step S11, a total of 34,390 images were selected. These images came from 400 384-well plates, including 22,400 images of DMSO treatment and 11,990 images of chemical treatment.

Images